linux/drivers/nvme/host/pci.c

4331 lines
117 KiB

nvme-pci: add support for ACPI StorageD3Enable property This patch implements a solution for a BIOS hack used on some currently shipping Intel systems to change driver power management policy for PCIe NVMe drives. Some newer Intel platforms, like some Comet Lake systems, require that PCIe devices use D3 when doing suspend-to-idle in order to allow the platform to realize maximum power savings. This is particularly needed to support ATX power supply shutdown on desktop systems. In order to ensure this happens for root ports with storage devices, Microsoft apparently created this ACPI _DSD property as a way to influence their driver policy. To my knowledge this property has not been discussed with the NVME specification body. Though the solution is not ideal, it addresses a problem that also affects Linux since the NVMe driver's default policy of using NVMe APST during suspend-to-idle prevents the PCI root port from going to D3 and leads to higher power consumption for these platforms. The power consumption difference may be negligible on laptop systems, but many watts on desktop systems when the ATX power supply is blocked from powering down. The patch creates a new nvme_acpi_storage_d3 function to check for the StorageD3Enable property during probe and enables D3 as a quirk if set. It also provides a 'noacpi' module parameter to allow skipping the quirk if needed. Tested with: - PM961 NVMe SED Samsung 512GB - INTEL SSDPEKKF512G8 Link: https://docs.microsoft.com/en-us/windows-hardware/design/component-guidelines/power-management-for-storage-hardware-devices-intro Signed-off-by: David E. Box <david.e.box@linux.intel.com> Reviewed-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2020-07-09 11:43:33 -07:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
NVMe: Metadata format support Adds support for NVMe metadata formats and exposes block devices for all namespaces regardless of their format. Namespace formats that are unusable will have disk capacity set to 0, but a handle to the block device is created to simplify device management. A namespace is not usable when the format requires host interleave block and metadata in single buffer, has no provisioned storage, or has better data but failed to register with blk integrity. The namespace has to be scanned in two phases to support separate metadata formats. The first establishes the sector size and capacity prior to invoking add_disk. If metadata is required, the capacity will be temporarilly set to 0 until it can be revalidated and registered with the integrity extenstions after add_disk completes. The driver relies on the integrity extensions to provide the metadata buffer. NVMe requires this be a single physically contiguous region, so only one integrity segment is allowed per command. If the metadata is used for T10 PI, the driver provides mappings to save and restore the reftag physical block translation. The driver provides no-op functions for generate and verify if metadata is not used for protection information. This way the setup is always provided by the block layer. If a request does not supply a required metadata buffer, the command is failed with bad address. This could only happen if a user manually disables verify/generate on such a disk. The only exception to where this is okay is if the controller is capable of stripping/generating the metadata, which is possible on some types of formats. The metadata scatter gather list now occupies the spot in the nvme_iod that used to be used to link retryable IOD's, but we don't do that anymore, so the field was unused. Signed-off-by: Keith Busch <keith.busch@intel.com>
2015-02-19 13:39:03 -07:00
asm-generic: architecture independent readq/writeq for 32bit environment This provides unified readq()/writeq() helper functions for 32-bit drivers. For some cases, readq/writeq without atomicity is harmful, and order of io access has to be specified explicitly. So in this patch, new two header files which contain non-atomic readq/writeq are added. - <asm-generic/io-64-nonatomic-lo-hi.h> provides non-atomic readq/ writeq with the order of lower address -> higher address - <asm-generic/io-64-nonatomic-hi-lo.h> provides non-atomic readq/ writeq with reversed order This allows us to remove some readq()s that were added drivers when the default non-atomic ones were removed in commit dbee8a0affd5 ("x86: remove 32-bit versions of readq()/writeq()") The drivers which need readq/writeq but can do with the non-atomic ones must add the line: #include <asm-generic/io-64-nonatomic-lo-hi.h> /* or hi-lo.h */ But this will be nop in 64-bit environments, and no other #ifdefs are required. So I believe that this patch can solve the problem of 1. driver-specific readq/writeq 2. atomicity and order of io access This patch is tested with building allyesconfig and allmodconfig as ARCH=x86 and ARCH=i386 on top of tip/master. Cc: Kashyap Desai <Kashyap.Desai@lsi.com> Cc: Len Brown <lenb@kernel.org> Cc: Ravi Anand <ravi.anand@qlogic.com> Cc: Vikas Chaudhary <vikas.chaudhary@qlogic.com> Cc: Matthew Garrett <mjg@redhat.com> Cc: Jason Uhlenkott <juhlenko@akamai.com> Cc: James Bottomley <James.Bottomley@parallels.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Roland Dreier <roland@purestorage.com> Cc: James Bottomley <jbottomley@parallels.com> Cc: Alan Cox <alan@lxorguk.ukuu.org.uk> Cc: Matthew Wilcox <matthew.r.wilcox@intel.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Hitoshi Mitake <h.mitake@gmail.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-02-07 11:45:33 +09:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: make sure write/poll_queues less or equal then cpu count Check module parameter write/poll_queues before using it to catch too large values. Reproducer: modprobe -r nvme modprobe nvme write_queues=`nproc` echo $((`nproc`+1)) > /sys/module/nvme/parameters/write_queues echo 1 > /sys/block/nvme0n1/device/reset_controller [ 657.069000] ------------[ cut here ]------------ [ 657.069022] WARNING: CPU: 10 PID: 1163 at kernel/irq/affinity.c:390 irq_create_affinity_masks+0x47c/0x4a0 [ 657.069056] dm_region_hash dm_log dm_mod [ 657.069059] CPU: 10 PID: 1163 Comm: kworker/u193:9 Kdump: loaded Tainted: G W 5.6.0+ #8 [ 657.069060] Hardware name: Inspur SA5212M5/YZMB-00882-104, BIOS 4.0.9 08/27/2019 [ 657.069064] Workqueue: nvme-reset-wq nvme_reset_work [nvme] [ 657.069066] RIP: 0010:irq_create_affinity_masks+0x47c/0x4a0 [ 657.069067] Code: fe ff ff 48 c7 c0 b0 89 14 95 48 89 46 20 e9 e9 fb ff ff 31 c0 e9 90 fc ff ff 0f 0b 48 c7 44 24 08 00 00 00 00 e9 e9 fc ff ff <0f> 0b e9 87 fe ff ff 48 8b 7c 24 28 e8 33 a0 80 00 e9 b6 fc ff ff [ 657.069068] RSP: 0018:ffffb505ce1ffc78 EFLAGS: 00010202 [ 657.069069] RAX: 0000000000000060 RBX: ffff9b97921fe5c0 RCX: 0000000000000000 [ 657.069069] RDX: ffff9b67bad80000 RSI: 00000000ffffffa0 RDI: 0000000000000000 [ 657.069070] RBP: 0000000000000000 R08: 0000000000000000 R09: ffff9b97921fe718 [ 657.069070] R10: ffff9b97921fe710 R11: 0000000000000001 R12: 0000000000000064 [ 657.069070] R13: 0000000000000060 R14: 0000000000000000 R15: 0000000000000001 [ 657.069071] FS: 0000000000000000(0000) GS:ffff9b67c0880000(0000) knlGS:0000000000000000 [ 657.069072] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 657.069072] CR2: 0000559eac6fc238 CR3: 000000057860a002 CR4: 00000000007606e0 [ 657.069073] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 657.069073] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 657.069073] PKRU: 55555554 [ 657.069074] Call Trace: [ 657.069080] __pci_enable_msix_range+0x233/0x5a0 [ 657.069085] ? kernfs_put+0xec/0x190 [ 657.069086] pci_alloc_irq_vectors_affinity+0xbb/0x130 [ 657.069089] nvme_reset_work+0x6e6/0xeab [nvme] [ 657.069093] ? __switch_to_asm+0x34/0x70 [ 657.069094] ? __switch_to_asm+0x40/0x70 [ 657.069095] ? nvme_irq_check+0x30/0x30 [nvme] [ 657.069098] process_one_work+0x1a7/0x370 [ 657.069101] worker_thread+0x1c9/0x380 [ 657.069102] ? max_active_store+0x80/0x80 [ 657.069103] kthread+0x112/0x130 [ 657.069104] ? __kthread_parkme+0x70/0x70 [ 657.069105] ret_from_fork+0x35/0x40 [ 657.069106] ---[ end trace f4f06b7d24513d06 ]--- [ 657.077110] nvme nvme0: 95/1/0 default/read/poll queues Signed-off-by: Weiping Zhang <zhangweiping@didiglobal.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2020-05-09 14:22:08 +08:00
nvme-pci: make sure write/poll_queues less or equal then cpu count Check module parameter write/poll_queues before using it to catch too large values. Reproducer: modprobe -r nvme modprobe nvme write_queues=`nproc` echo $((`nproc`+1)) > /sys/module/nvme/parameters/write_queues echo 1 > /sys/block/nvme0n1/device/reset_controller [ 657.069000] ------------[ cut here ]------------ [ 657.069022] WARNING: CPU: 10 PID: 1163 at kernel/irq/affinity.c:390 irq_create_affinity_masks+0x47c/0x4a0 [ 657.069056] dm_region_hash dm_log dm_mod [ 657.069059] CPU: 10 PID: 1163 Comm: kworker/u193:9 Kdump: loaded Tainted: G W 5.6.0+ #8 [ 657.069060] Hardware name: Inspur SA5212M5/YZMB-00882-104, BIOS 4.0.9 08/27/2019 [ 657.069064] Workqueue: nvme-reset-wq nvme_reset_work [nvme] [ 657.069066] RIP: 0010:irq_create_affinity_masks+0x47c/0x4a0 [ 657.069067] Code: fe ff ff 48 c7 c0 b0 89 14 95 48 89 46 20 e9 e9 fb ff ff 31 c0 e9 90 fc ff ff 0f 0b 48 c7 44 24 08 00 00 00 00 e9 e9 fc ff ff <0f> 0b e9 87 fe ff ff 48 8b 7c 24 28 e8 33 a0 80 00 e9 b6 fc ff ff [ 657.069068] RSP: 0018:ffffb505ce1ffc78 EFLAGS: 00010202 [ 657.069069] RAX: 0000000000000060 RBX: ffff9b97921fe5c0 RCX: 0000000000000000 [ 657.069069] RDX: ffff9b67bad80000 RSI: 00000000ffffffa0 RDI: 0000000000000000 [ 657.069070] RBP: 0000000000000000 R08: 0000000000000000 R09: ffff9b97921fe718 [ 657.069070] R10: ffff9b97921fe710 R11: 0000000000000001 R12: 0000000000000064 [ 657.069070] R13: 0000000000000060 R14: 0000000000000000 R15: 0000000000000001 [ 657.069071] FS: 0000000000000000(0000) GS:ffff9b67c0880000(0000) knlGS:0000000000000000 [ 657.069072] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 657.069072] CR2: 0000559eac6fc238 CR3: 000000057860a002 CR4: 00000000007606e0 [ 657.069073] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 657.069073] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 657.069073] PKRU: 55555554 [ 657.069074] Call Trace: [ 657.069080] __pci_enable_msix_range+0x233/0x5a0 [ 657.069085] ? kernfs_put+0xec/0x190 [ 657.069086] pci_alloc_irq_vectors_affinity+0xbb/0x130 [ 657.069089] nvme_reset_work+0x6e6/0xeab [nvme] [ 657.069093] ? __switch_to_asm+0x34/0x70 [ 657.069094] ? __switch_to_asm+0x40/0x70 [ 657.069095] ? nvme_irq_check+0x30/0x30 [nvme] [ 657.069098] process_one_work+0x1a7/0x370 [ 657.069101] worker_thread+0x1c9/0x380 [ 657.069102] ? max_active_store+0x80/0x80 [ 657.069103] kthread+0x112/0x130 [ 657.069104] ? __kthread_parkme+0x70/0x70 [ 657.069105] ret_from_fork+0x35/0x40 [ 657.069106] ---[ end trace f4f06b7d24513d06 ]--- [ 657.077110] nvme nvme0: 95/1/0 default/read/poll queues Signed-off-by: Weiping Zhang <zhangweiping@didiglobal.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2020-05-09 14:22:08 +08:00
nvme-pci: make sure write/poll_queues less or equal then cpu count Check module parameter write/poll_queues before using it to catch too large values. Reproducer: modprobe -r nvme modprobe nvme write_queues=`nproc` echo $((`nproc`+1)) > /sys/module/nvme/parameters/write_queues echo 1 > /sys/block/nvme0n1/device/reset_controller [ 657.069000] ------------[ cut here ]------------ [ 657.069022] WARNING: CPU: 10 PID: 1163 at kernel/irq/affinity.c:390 irq_create_affinity_masks+0x47c/0x4a0 [ 657.069056] dm_region_hash dm_log dm_mod [ 657.069059] CPU: 10 PID: 1163 Comm: kworker/u193:9 Kdump: loaded Tainted: G W 5.6.0+ #8 [ 657.069060] Hardware name: Inspur SA5212M5/YZMB-00882-104, BIOS 4.0.9 08/27/2019 [ 657.069064] Workqueue: nvme-reset-wq nvme_reset_work [nvme] [ 657.069066] RIP: 0010:irq_create_affinity_masks+0x47c/0x4a0 [ 657.069067] Code: fe ff ff 48 c7 c0 b0 89 14 95 48 89 46 20 e9 e9 fb ff ff 31 c0 e9 90 fc ff ff 0f 0b 48 c7 44 24 08 00 00 00 00 e9 e9 fc ff ff <0f> 0b e9 87 fe ff ff 48 8b 7c 24 28 e8 33 a0 80 00 e9 b6 fc ff ff [ 657.069068] RSP: 0018:ffffb505ce1ffc78 EFLAGS: 00010202 [ 657.069069] RAX: 0000000000000060 RBX: ffff9b97921fe5c0 RCX: 0000000000000000 [ 657.069069] RDX: ffff9b67bad80000 RSI: 00000000ffffffa0 RDI: 0000000000000000 [ 657.069070] RBP: 0000000000000000 R08: 0000000000000000 R09: ffff9b97921fe718 [ 657.069070] R10: ffff9b97921fe710 R11: 0000000000000001 R12: 0000000000000064 [ 657.069070] R13: 0000000000000060 R14: 0000000000000000 R15: 0000000000000001 [ 657.069071] FS: 0000000000000000(0000) GS:ffff9b67c0880000(0000) knlGS:0000000000000000 [ 657.069072] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 657.069072] CR2: 0000559eac6fc238 CR3: 000000057860a002 CR4: 00000000007606e0 [ 657.069073] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 657.069073] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 657.069073] PKRU: 55555554 [ 657.069074] Call Trace: [ 657.069080] __pci_enable_msix_range+0x233/0x5a0 [ 657.069085] ? kernfs_put+0xec/0x190 [ 657.069086] pci_alloc_irq_vectors_affinity+0xbb/0x130 [ 657.069089] nvme_reset_work+0x6e6/0xeab [nvme] [ 657.069093] ? __switch_to_asm+0x34/0x70 [ 657.069094] ? __switch_to_asm+0x40/0x70 [ 657.069095] ? nvme_irq_check+0x30/0x30 [nvme] [ 657.069098] process_one_work+0x1a7/0x370 [ 657.069101] worker_thread+0x1c9/0x380 [ 657.069102] ? max_active_store+0x80/0x80 [ 657.069103] kthread+0x112/0x130 [ 657.069104] ? __kthread_parkme+0x70/0x70 [ 657.069105] ret_from_fork+0x35/0x40 [ 657.069106] ---[ end trace f4f06b7d24513d06 ]--- [ 657.077110] nvme nvme0: 95/1/0 default/read/poll queues Signed-off-by: Weiping Zhang <zhangweiping@didiglobal.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2020-05-09 14:22:08 +08:00
nvme-pci: make sure write/poll_queues less or equal then cpu count Check module parameter write/poll_queues before using it to catch too large values. Reproducer: modprobe -r nvme modprobe nvme write_queues=`nproc` echo $((`nproc`+1)) > /sys/module/nvme/parameters/write_queues echo 1 > /sys/block/nvme0n1/device/reset_controller [ 657.069000] ------------[ cut here ]------------ [ 657.069022] WARNING: CPU: 10 PID: 1163 at kernel/irq/affinity.c:390 irq_create_affinity_masks+0x47c/0x4a0 [ 657.069056] dm_region_hash dm_log dm_mod [ 657.069059] CPU: 10 PID: 1163 Comm: kworker/u193:9 Kdump: loaded Tainted: G W 5.6.0+ #8 [ 657.069060] Hardware name: Inspur SA5212M5/YZMB-00882-104, BIOS 4.0.9 08/27/2019 [ 657.069064] Workqueue: nvme-reset-wq nvme_reset_work [nvme] [ 657.069066] RIP: 0010:irq_create_affinity_masks+0x47c/0x4a0 [ 657.069067] Code: fe ff ff 48 c7 c0 b0 89 14 95 48 89 46 20 e9 e9 fb ff ff 31 c0 e9 90 fc ff ff 0f 0b 48 c7 44 24 08 00 00 00 00 e9 e9 fc ff ff <0f> 0b e9 87 fe ff ff 48 8b 7c 24 28 e8 33 a0 80 00 e9 b6 fc ff ff [ 657.069068] RSP: 0018:ffffb505ce1ffc78 EFLAGS: 00010202 [ 657.069069] RAX: 0000000000000060 RBX: ffff9b97921fe5c0 RCX: 0000000000000000 [ 657.069069] RDX: ffff9b67bad80000 RSI: 00000000ffffffa0 RDI: 0000000000000000 [ 657.069070] RBP: 0000000000000000 R08: 0000000000000000 R09: ffff9b97921fe718 [ 657.069070] R10: ffff9b97921fe710 R11: 0000000000000001 R12: 0000000000000064 [ 657.069070] R13: 0000000000000060 R14: 0000000000000000 R15: 0000000000000001 [ 657.069071] FS: 0000000000000000(0000) GS:ffff9b67c0880000(0000) knlGS:0000000000000000 [ 657.069072] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 657.069072] CR2: 0000559eac6fc238 CR3: 000000057860a002 CR4: 00000000007606e0 [ 657.069073] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 657.069073] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 657.069073] PKRU: 55555554 [ 657.069074] Call Trace: [ 657.069080] __pci_enable_msix_range+0x233/0x5a0 [ 657.069085] ? kernfs_put+0xec/0x190 [ 657.069086] pci_alloc_irq_vectors_affinity+0xbb/0x130 [ 657.069089] nvme_reset_work+0x6e6/0xeab [nvme] [ 657.069093] ? __switch_to_asm+0x34/0x70 [ 657.069094] ? __switch_to_asm+0x40/0x70 [ 657.069095] ? nvme_irq_check+0x30/0x30 [nvme] [ 657.069098] process_one_work+0x1a7/0x370 [ 657.069101] worker_thread+0x1c9/0x380 [ 657.069102] ? max_active_store+0x80/0x80 [ 657.069103] kthread+0x112/0x130 [ 657.069104] ? __kthread_parkme+0x70/0x70 [ 657.069105] ret_from_fork+0x35/0x40 [ 657.069106] ---[ end trace f4f06b7d24513d06 ]--- [ 657.077110] nvme nvme0: 95/1/0 default/read/poll queues Signed-off-by: Weiping Zhang <zhangweiping@didiglobal.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2020-05-09 14:22:08 +08:00
nvme-pci: add support for ACPI StorageD3Enable property This patch implements a solution for a BIOS hack used on some currently shipping Intel systems to change driver power management policy for PCIe NVMe drives. Some newer Intel platforms, like some Comet Lake systems, require that PCIe devices use D3 when doing suspend-to-idle in order to allow the platform to realize maximum power savings. This is particularly needed to support ATX power supply shutdown on desktop systems. In order to ensure this happens for root ports with storage devices, Microsoft apparently created this ACPI _DSD property as a way to influence their driver policy. To my knowledge this property has not been discussed with the NVME specification body. Though the solution is not ideal, it addresses a problem that also affects Linux since the NVMe driver's default policy of using NVMe APST during suspend-to-idle prevents the PCI root port from going to D3 and leads to higher power consumption for these platforms. The power consumption difference may be negligible on laptop systems, but many watts on desktop systems when the ATX power supply is blocked from powering down. The patch creates a new nvme_acpi_storage_d3 function to check for the StorageD3Enable property during probe and enables D3 as a quirk if set. It also provides a 'noacpi' module parameter to allow skipping the quirk if needed. Tested with: - PM961 NVMe SED Samsung 512GB - INTEL SSDPEKKF512G8 Link: https://docs.microsoft.com/en-us/windows-hardware/design/component-guidelines/power-management-for-storage-hardware-devices-intro Signed-off-by: David E. Box <david.e.box@linux.intel.com> Reviewed-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2020-07-09 11:43:33 -07:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
nvme-pci: Use u32 for nvme_dev.q_depth and nvme_queue.q_depth Recently nvme_dev.q_depth was changed from an int to u16 type. This falls over for the queue depth calculation in nvme_pci_enable(), where NVME_CAP_MQES(dev->ctrl.cap) + 1 may overflow as a u16, as NVME_CAP_MQES() is a 16b number also. That happens for me, and this is the result: root@ubuntu:/home/john# [148.272996] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000010 Mem abort info: ESR = 0x96000004 EC = 0x25: DABT (current EL), IL = 32 bits SET = 0, FnV = 0 EA = 0, S1PTW = 0 Data abort info: ISV = 0, ISS = 0x00000004 CM = 0, WnR = 0 user pgtable: 4k pages, 48-bit VAs, pgdp=00000a27bf3c9000 [0000000000000010] pgd=0000000000000000, p4d=0000000000000000 Internal error: Oops: 96000004 [#1] PREEMPT SMP Modules linked in: nvme nvme_core CPU: 56 PID: 256 Comm: kworker/u195:0 Not tainted 5.8.0-next-20200812 #27 Hardware name: Huawei D06 /D06, BIOS Hisilicon D06 UEFI RC0 - V1.16.01 03/15/2019 Workqueue: nvme-reset-wq nvme_reset_work [nvme] pstate: 80c00009 (Nzcv daif +PAN +UAO BTYPE=--) pc : __sg_alloc_table_from_pages+0xec/0x238 lr : __sg_alloc_table_from_pages+0xc8/0x238 sp : ffff800013ccbad0 x29: ffff800013ccbad0 x28: ffff0a27b3d380a8 x27: 0000000000000000 x26: 0000000000002dc2 x25: 0000000000000dc0 x24: 0000000000000000 x23: 0000000000000000 x22: ffff800013ccbbe8 x21: 0000000000000010 x20: 0000000000000000 x19: 00000000fffff000 x18: ffffffffffffffff x17: 00000000000000c0 x16: fffffe289eaf6380 x15: ffff800011b59948 x14: ffff002bc8fe98f8 x13: ff00000000000000 x12: ffff8000114ca000 x11: 0000000000000000 x10: ffffffffffffffff x9 : ffffffffffffffc0 x8 : ffff0a27b5f9b6a0 x7 : 0000000000000000 x6 : 0000000000000001 x5 : ffff0a27b5f9b680 x4 : 0000000000000000 x3 : ffff0a27b5f9b680 x2 : 0000000000000000 x1 : 0000000000000001 x0 : 0000000000000000 Call trace: __sg_alloc_table_from_pages+0xec/0x238 sg_alloc_table_from_pages+0x18/0x28 iommu_dma_alloc+0x474/0x678 dma_alloc_attrs+0xd8/0xf0 nvme_alloc_queue+0x114/0x160 [nvme] nvme_reset_work+0xb34/0x14b4 [nvme] process_one_work+0x1e8/0x360 worker_thread+0x44/0x478 kthread+0x150/0x158 ret_from_fork+0x10/0x34 Code: f94002c3 6b01017f 540007c2 11000486 (f8645aa5) ---[ end trace 89bb2b72d59bf925 ]--- Fix by making onto a u32. Also use u32 for nvme_dev.q_depth, as we assign this value from nvme_dev.q_depth, and nvme_dev.q_depth will possibly hold 65536 - this avoids the same crash as above. Fixes: 61f3b8963097 ("nvme-pci: use unsigned for io queue depth") Signed-off-by: John Garry <john.garry@huawei.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-14 23:34:25 +08:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme-pci: Use u32 for nvme_dev.q_depth and nvme_queue.q_depth Recently nvme_dev.q_depth was changed from an int to u16 type. This falls over for the queue depth calculation in nvme_pci_enable(), where NVME_CAP_MQES(dev->ctrl.cap) + 1 may overflow as a u16, as NVME_CAP_MQES() is a 16b number also. That happens for me, and this is the result: root@ubuntu:/home/john# [148.272996] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000010 Mem abort info: ESR = 0x96000004 EC = 0x25: DABT (current EL), IL = 32 bits SET = 0, FnV = 0 EA = 0, S1PTW = 0 Data abort info: ISV = 0, ISS = 0x00000004 CM = 0, WnR = 0 user pgtable: 4k pages, 48-bit VAs, pgdp=00000a27bf3c9000 [0000000000000010] pgd=0000000000000000, p4d=0000000000000000 Internal error: Oops: 96000004 [#1] PREEMPT SMP Modules linked in: nvme nvme_core CPU: 56 PID: 256 Comm: kworker/u195:0 Not tainted 5.8.0-next-20200812 #27 Hardware name: Huawei D06 /D06, BIOS Hisilicon D06 UEFI RC0 - V1.16.01 03/15/2019 Workqueue: nvme-reset-wq nvme_reset_work [nvme] pstate: 80c00009 (Nzcv daif +PAN +UAO BTYPE=--) pc : __sg_alloc_table_from_pages+0xec/0x238 lr : __sg_alloc_table_from_pages+0xc8/0x238 sp : ffff800013ccbad0 x29: ffff800013ccbad0 x28: ffff0a27b3d380a8 x27: 0000000000000000 x26: 0000000000002dc2 x25: 0000000000000dc0 x24: 0000000000000000 x23: 0000000000000000 x22: ffff800013ccbbe8 x21: 0000000000000010 x20: 0000000000000000 x19: 00000000fffff000 x18: ffffffffffffffff x17: 00000000000000c0 x16: fffffe289eaf6380 x15: ffff800011b59948 x14: ffff002bc8fe98f8 x13: ff00000000000000 x12: ffff8000114ca000 x11: 0000000000000000 x10: ffffffffffffffff x9 : ffffffffffffffc0 x8 : ffff0a27b5f9b6a0 x7 : 0000000000000000 x6 : 0000000000000001 x5 : ffff0a27b5f9b680 x4 : 0000000000000000 x3 : ffff0a27b5f9b680 x2 : 0000000000000000 x1 : 0000000000000001 x0 : 0000000000000000 Call trace: __sg_alloc_table_from_pages+0xec/0x238 sg_alloc_table_from_pages+0x18/0x28 iommu_dma_alloc+0x474/0x678 dma_alloc_attrs+0xd8/0xf0 nvme_alloc_queue+0x114/0x160 [nvme] nvme_reset_work+0xb34/0x14b4 [nvme] process_one_work+0x1e8/0x360 worker_thread+0x44/0x478 kthread+0x150/0x158 ret_from_fork+0x10/0x34 Code: f94002c3 6b01017f 540007c2 11000486 (f8645aa5) ---[ end trace 89bb2b72d59bf925 ]--- Fix by making onto a u32. Also use u32 for nvme_dev.q_depth, as we assign this value from nvme_dev.q_depth, and nvme_dev.q_depth will possibly hold 65536 - this avoids the same crash as above. Fixes: 61f3b8963097 ("nvme-pci: use unsigned for io queue depth") Signed-off-by: John Garry <john.garry@huawei.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-14 23:34:25 +08:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme-pci: Fix slab-out-of-bounds in nvme_dbbuf_set dev->online_queues is a count incremented in nvme_init_queue. Thus, valid indices are 0 through dev->online_queues − 1. This patch fixes the loop condition to ensure the index stays within the valid range. Index 0 is excluded because it is the admin queue. KASAN splat: ================================================================== BUG: KASAN: slab-out-of-bounds in nvme_dbbuf_free drivers/nvme/host/pci.c:377 [inline] BUG: KASAN: slab-out-of-bounds in nvme_dbbuf_set+0x39c/0x400 drivers/nvme/host/pci.c:404 Read of size 2 at addr ffff88800592a574 by task kworker/u8:5/74 CPU: 0 UID: 0 PID: 74 Comm: kworker/u8:5 Not tainted 6.19.0-dirty #10 PREEMPT(voluntary) Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/01/2014 Workqueue: nvme-reset-wq nvme_reset_work Call Trace: <TASK> __dump_stack lib/dump_stack.c:94 [inline] dump_stack_lvl+0xea/0x150 lib/dump_stack.c:120 print_address_description mm/kasan/report.c:378 [inline] print_report+0xce/0x5d0 mm/kasan/report.c:482 kasan_report+0xdc/0x110 mm/kasan/report.c:595 __asan_report_load2_noabort+0x18/0x20 mm/kasan/report_generic.c:379 nvme_dbbuf_free drivers/nvme/host/pci.c:377 [inline] nvme_dbbuf_set+0x39c/0x400 drivers/nvme/host/pci.c:404 nvme_reset_work+0x36b/0x8c0 drivers/nvme/host/pci.c:3252 process_one_work+0x956/0x1aa0 kernel/workqueue.c:3257 process_scheduled_works kernel/workqueue.c:3340 [inline] worker_thread+0x65c/0xe60 kernel/workqueue.c:3421 kthread+0x41a/0x930 kernel/kthread.c:463 ret_from_fork+0x6f8/0x8c0 arch/x86/kernel/process.c:158 ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:246 </TASK> Allocated by task 34 on cpu 1 at 4.241550s: kasan_save_stack+0x2c/0x60 mm/kasan/common.c:57 kasan_save_track+0x1c/0x70 mm/kasan/common.c:78 kasan_save_alloc_info+0x3c/0x50 mm/kasan/generic.c:570 poison_kmalloc_redzone mm/kasan/common.c:398 [inline] __kasan_kmalloc+0xb5/0xc0 mm/kasan/common.c:415 kasan_kmalloc include/linux/kasan.h:263 [inline] __do_kmalloc_node mm/slub.c:5657 [inline] __kmalloc_node_noprof+0x2bf/0x8d0 mm/slub.c:5663 kmalloc_array_node_noprof include/linux/slab.h:1075 [inline] nvme_pci_alloc_dev drivers/nvme/host/pci.c:3479 [inline] nvme_probe+0x2f1/0x1820 drivers/nvme/host/pci.c:3534 local_pci_probe+0xef/0x1c0 drivers/pci/pci-driver.c:324 pci_call_probe drivers/pci/pci-driver.c:392 [inline] __pci_device_probe drivers/pci/pci-driver.c:417 [inline] pci_device_probe+0x743/0x920 drivers/pci/pci-driver.c:451 call_driver_probe drivers/base/dd.c:583 [inline] really_probe+0x29b/0xb70 drivers/base/dd.c:661 __driver_probe_device+0x3b0/0x4a0 drivers/base/dd.c:803 driver_probe_device+0x56/0x1f0 drivers/base/dd.c:833 __driver_attach_async_helper+0x155/0x340 drivers/base/dd.c:1159 async_run_entry_fn+0xa6/0x4b0 kernel/async.c:129 process_one_work+0x956/0x1aa0 kernel/workqueue.c:3257 process_scheduled_works kernel/workqueue.c:3340 [inline] worker_thread+0x65c/0xe60 kernel/workqueue.c:3421 kthread+0x41a/0x930 kernel/kthread.c:463 ret_from_fork+0x6f8/0x8c0 arch/x86/kernel/process.c:158 ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:246 The buggy address belongs to the object at ffff88800592a000 which belongs to the cache kmalloc-2k of size 2048 The buggy address is located 244 bytes to the right of allocated 1152-byte region [ffff88800592a000, ffff88800592a480) The buggy address belongs to the physical page: page: refcount:0 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x5928 head: order:3 mapcount:0 entire_mapcount:0 nr_pages_mapped:0 pincount:0 anon flags: 0xfffffc0000040(head|node=0|zone=1|lastcpupid=0x1fffff) page_type: f5(slab) raw: 000fffffc0000040 ffff888001042000 0000000000000000 dead000000000001 raw: 0000000000000000 0000000000080008 00000000f5000000 0000000000000000 head: 000fffffc0000040 ffff888001042000 0000000000000000 dead000000000001 head: 0000000000000000 0000000000080008 00000000f5000000 0000000000000000 head: 000fffffc0000003 ffffea0000164a01 00000000ffffffff 00000000ffffffff head: ffffffffffffffff 0000000000000000 00000000ffffffff 0000000000000008 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff88800592a400: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ffff88800592a480: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc >ffff88800592a500: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc ^ ffff88800592a580: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc ffff88800592a600: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc ================================================================== Fixes: 0f0d2c876c96 (nvme: free sq/cq dbbuf pointers when dbbuf set fails) Acked-by: Chao Shi <cshi008@fiu.edu> Acked-by: Weidong Zhu <weizhu@fiu.edu> Acked-by: Dave Tian <daveti@purdue.edu> Signed-off-by: Sungwoo Kim <iam@sung-woo.kim> Signed-off-by: Keith Busch <kbusch@kernel.org>
2026-03-08 14:20:59 -04:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme-pci: add a memory barrier to nvme_dbbuf_update_and_check_event In many architectures loads may be reordered with older stores to different locations. In the nvme driver the following two operations could be reordered: - Write shadow doorbell (dbbuf_db) into memory. - Read EventIdx (dbbuf_ei) from memory. This can result in a potential race condition between driver and VM host processing requests (if given virtual NVMe controller has a support for shadow doorbell). If that occurs, then the NVMe controller may decide to wait for MMIO doorbell from guest operating system, and guest driver may decide not to issue MMIO doorbell on any of subsequent commands. This issue is purely timing-dependent one, so there is no easy way to reproduce it. Currently the easiest known approach is to run "Oracle IO Numbers" (orion) that is shipped with Oracle DB: orion -run advanced -num_large 0 -size_small 8 -type rand -simulate \ concat -write 40 -duration 120 -matrix row -testname nvme_test Where nvme_test is a .lun file that contains a list of NVMe block devices to run test against. Limiting number of vCPUs assigned to given VM instance seems to increase chances for this bug to occur. On test environment with VM that got 4 NVMe drives and 1 vCPU assigned the virtual NVMe controller hang could be observed within 10-20 minutes. That correspond to about 400-500k IO operations processed (or about 100GB of IO read/writes). Orion tool was used as a validation and set to run in a loop for 36 hours (equivalent of pushing 550M IO operations). No issues were observed. That suggest that the patch fixes the issue. Fixes: f9f38e33389c ("nvme: improve performance for virtual NVMe devices") Signed-off-by: Michal Wnukowski <wnukowski@google.com> Reviewed-by: Keith Busch <keith.busch@intel.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> [hch: updated changelog and comment a bit] Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-08-15 15:51:57 -07:00
nvme-pci: add a memory barrier to nvme_dbbuf_update_and_check_event In many architectures loads may be reordered with older stores to different locations. In the nvme driver the following two operations could be reordered: - Write shadow doorbell (dbbuf_db) into memory. - Read EventIdx (dbbuf_ei) from memory. This can result in a potential race condition between driver and VM host processing requests (if given virtual NVMe controller has a support for shadow doorbell). If that occurs, then the NVMe controller may decide to wait for MMIO doorbell from guest operating system, and guest driver may decide not to issue MMIO doorbell on any of subsequent commands. This issue is purely timing-dependent one, so there is no easy way to reproduce it. Currently the easiest known approach is to run "Oracle IO Numbers" (orion) that is shipped with Oracle DB: orion -run advanced -num_large 0 -size_small 8 -type rand -simulate \ concat -write 40 -duration 120 -matrix row -testname nvme_test Where nvme_test is a .lun file that contains a list of NVMe block devices to run test against. Limiting number of vCPUs assigned to given VM instance seems to increase chances for this bug to occur. On test environment with VM that got 4 NVMe drives and 1 vCPU assigned the virtual NVMe controller hang could be observed within 10-20 minutes. That correspond to about 400-500k IO operations processed (or about 100GB of IO read/writes). Orion tool was used as a validation and set to run in a loop for 36 hours (equivalent of pushing 550M IO operations). No issues were observed. That suggest that the patch fixes the issue. Fixes: f9f38e33389c ("nvme: improve performance for virtual NVMe devices") Signed-off-by: Michal Wnukowski <wnukowski@google.com> Reviewed-by: Keith Busch <keith.busch@intel.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> [hch: updated changelog and comment a bit] Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-08-15 15:51:57 -07:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
nvme-pci: use blk-mq mapping for unmanaged irqs If a device is providing a single IRQ vector, the IO queue will share that vector with the admin queue. This is an unmanaged vector, so does not have a valid PCI IRQ affinity. Avoid trying to extract a managed affinity in this case and let blk-mq set up the cpu:queue mapping instead. Otherwise we'd hit the following warning when the device is using MSI: WARNING: CPU: 4 PID: 7 at drivers/pci/msi.c:1272 pci_irq_get_affinity+0x66/0x80 Modules linked in: nvme nvme_core serio_raw CPU: 4 PID: 7 Comm: kworker/u16:0 Tainted: G W 5.2.0-rc1+ #494 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.1-0-ga5cab58e9a3f-prebuilt.qemu.org 04/01/2014 Workqueue: nvme-reset-wq nvme_reset_work [nvme] RIP: 0010:pci_irq_get_affinity+0x66/0x80 Code: 0b 31 c0 c3 83 e2 10 48 c7 c0 b0 83 35 91 74 2a 48 8b 87 d8 03 00 00 48 85 c0 74 0e 48 8b 50 30 48 85 d2 74 05 39 70 14 77 05 <0f> 0b 31 c0 c3 48 63 f6 48 8d 04 76 48 8d 04 c2 f3 c3 48 8b 40 30 RSP: 0000:ffffb5abc01d3cc8 EFLAGS: 00010246 RAX: ffff9536786a39c0 RBX: 0000000000000000 RCX: 0000000000000080 RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9536781ed000 RBP: ffff95367346a008 R08: ffff95367d43f080 R09: ffff953678c07800 R10: ffff953678164800 R11: 0000000000000000 R12: 0000000000000000 R13: ffff9536781ed000 R14: 00000000ffffffff R15: ffff95367346a008 FS: 0000000000000000(0000) GS:ffff95367d400000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fdf814a3ff0 CR3: 000000001a20f000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: blk_mq_pci_map_queues+0x37/0xd0 nvme_pci_map_queues+0x80/0xb0 [nvme] blk_mq_alloc_tag_set+0x133/0x2f0 nvme_reset_work+0x105d/0x1590 [nvme] process_one_work+0x291/0x530 worker_thread+0x218/0x3d0 ? process_one_work+0x530/0x530 kthread+0x111/0x130 ? kthread_park+0x90/0x90 ret_from_fork+0x1f/0x30 ---[ end trace 74587339d93c83c0 ]--- Fixes: 22b5560195bd6 ("nvme-pci: Separate IO and admin queue IRQ vectors") Reported-by: Iván Chavero <ichavero@chavero.com.mx> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Keith Busch <keith.busch@intel.com>
2019-05-21 10:56:43 -06:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
nvme-pci: convert the data mapping to blk_rq_dma_map Use the blk_rq_dma_map API to DMA map requests instead of scatterlists. This removes the need to allocate a scatterlist covering every segment, and thus the overall transfer length limit based on the scatterlist allocation. Instead the DMA mapping is done by iterating the bio_vec chain in the request directly. The unmap is handled differently depending on how we mapped: - when using an IOMMU only a single IOVA is used, and it is stored in iova_state - for direct mappings that don't use swiotlb and are cache coherent, unmap is not needed at all - for direct mappings that are not cache coherent or use swiotlb, the physical addresses are rebuild from the PRPs or SGL segments The latter unfortunately adds a fair amount of code to the driver, but it is code not used in the fast path. The conversion only covers the data mapping path, and still uses a scatterlist for the multi-segment metadata case. I plan to convert that as soon as we have good test coverage for the multi-segment metadata path. Thanks to Chaitanya Kulkarni for an initial attempt at a new DMA API conversion for nvme-pci, Kanchan Joshi for bringing back the single segment optimization, Leon Romanovsky for shepherding this through a gazillion rebases and Nitesh Shetty for various improvements. Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Leon Romanovsky <leonro@nvidia.com> Link: https://lore.kernel.org/r/20250625113531.522027-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-06-25 13:35:03 +02:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
NVMe: Metadata format support Adds support for NVMe metadata formats and exposes block devices for all namespaces regardless of their format. Namespace formats that are unusable will have disk capacity set to 0, but a handle to the block device is created to simplify device management. A namespace is not usable when the format requires host interleave block and metadata in single buffer, has no provisioned storage, or has better data but failed to register with blk integrity. The namespace has to be scanned in two phases to support separate metadata formats. The first establishes the sector size and capacity prior to invoking add_disk. If metadata is required, the capacity will be temporarilly set to 0 until it can be revalidated and registered with the integrity extenstions after add_disk completes. The driver relies on the integrity extensions to provide the metadata buffer. NVMe requires this be a single physically contiguous region, so only one integrity segment is allowed per command. If the metadata is used for T10 PI, the driver provides mappings to save and restore the reftag physical block translation. The driver provides no-op functions for generate and verify if metadata is not used for protection information. This way the setup is always provided by the block layer. If a request does not supply a required metadata buffer, the command is failed with bad address. This could only happen if a user manually disables verify/generate on such a disk. The only exception to where this is okay is if the controller is capable of stripping/generating the metadata, which is possible on some types of formats. The metadata scatter gather list now occupies the spot in the nvme_iod that used to be used to link retryable IOD's, but we don't do that anymore, so the field was unused. Signed-off-by: Keith Busch <keith.busch@intel.com>
2015-02-19 13:39:03 -07:00
nvme-pci: fix controller reset hang when racing with nvme_timeout reset_work() in nvme-pci may hang forever in the following scenario: 1) A reset caused by a command timeout occurs due to a controller being temporarily irresponsive. 2) nvme_reset_work() restarts admin queue at nvme_alloc_admin_tags(). At the same time, a user-submitted admin command is queued and waiting for completion. Then, reset_work() changes its state to CONNECTING, and submits an identify command. 3) However, the controller does still not respond to any command, causing a timeout being fired at the user-submitted command. Unfortunately, nvme_timeout() does not see the completion on cq, and any timeout that takes place under CONNECTING state causes a controller shutdown. 4) Normally, the identify command in reset_work() would be canceled with SC_HOST_ABORTED by nvme_dev_disable(), then reset_work can tear down the controller accordingly. But the controller happens to return online and respond the identify command before nvme_dev_disable() should have been reaped it off. 5) reset_work() continues to setup_io_queues() as it observes no error in init_identify(). However, the admin queue has already been quiesced in dev_disable(). Thus, any following commands would be blocked forever in blk_execute_rq(). This can be fixed by restricting usercmd commands when controller is not in a LIVE state in nvme_queue_rq(), as what has been done previously in fabrics. ``` nvme_reset_work(): | nvme_alloc_admin_tags() | | nvme_submit_user_cmd(): nvme_init_identify(): | ... __nvme_submit_sync_cmd(): | ... | ... ---------------------------------------> nvme_timeout(): (Controller starts reponding commands) | nvme_dev_disable(, true): nvme_setup_io_queues(): | __nvme_submit_sync_cmd(): | (hung in blk_execute_rq | since run_hw_queue sees | queue quiesced) | ``` Signed-off-by: Tao Chiu <taochiu@synology.com> Signed-off-by: Cody Wong <codywong@synology.com> Reviewed-by: Leon Chien <leonchien@synology.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-04-26 10:53:55 +08:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
block: change blk_mq_add_to_batch() third argument type to bool Commit 1f47ed294a2b ("block: cleanup and fix batch completion adding conditions") modified the evaluation criteria for the third argument, 'ioerror', in the blk_mq_add_to_batch() function. Initially, the function had checked if 'ioerror' equals zero. Following the commit, it started checking for negative error values, with the presumption that such values, for instance -EIO, would be passed in. However, blk_mq_add_to_batch() callers do not pass negative error values. Instead, they pass status codes defined in various ways: - NVMe PCI and Apple drivers pass NVMe status code - virtio_blk driver passes the virtblk request header status byte - null_blk driver passes blk_status_t These codes are either zero or positive, therefore the revised check fails to function as intended. Specifically, with the NVMe PCI driver, this modification led to the failure of the blktests test case nvme/039. In this test scenario, errors are artificially injected to the NVMe driver, resulting in positive NVMe status codes passed to blk_mq_add_to_batch(), which unexpectedly processes the failed I/O in a batch. Hence the failure. To correct the ioerror check within blk_mq_add_to_batch(), make all callers to uniformly pass the argument as boolean. Modify the callers to check their specific status codes and pass the boolean value 'is_error'. Also describe the arguments of blK_mq_add_to_batch as kerneldoc. Fixes: 1f47ed294a2b ("block: cleanup and fix batch completion adding conditions") Signed-off-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com> Link: https://lore.kernel.org/r/20250311104359.1767728-3-shinichiro.kawasaki@wdc.com [axboe: fold in documentation update] Signed-off-by: Jens Axboe <axboe@kernel.dk>
2025-03-11 19:43:59 +09:00
nvme-pci: Fix race bug in nvme_poll_irqdisable() In the following scenario, pdev can be disabled between (1) and (3) by (2). This sets pdev->msix_enabled = 0. Then, pci_irq_vector() will return MSI-X IRQ(>15) for (1) whereas return INTx IRQ(<=15) for (2). This causes IRQ warning because it tries to enable INTx IRQ that has never been disabled before. To fix this, save IRQ number into a local variable and ensure disable_irq() and enable_irq() operate on the same IRQ number. Even if pci_free_irq_vectors() frees the IRQ concurrently, disable_irq() and enable_irq() on a stale IRQ number is still valid and safe, and the depth accounting reamins balanced. task 1: nvme_poll_irqdisable() disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector)) ...(1) enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector)) ...(3) task 2: nvme_reset_work() nvme_dev_disable() pdev->msix_enable = 0; ...(2) crash log: ------------[ cut here ]------------ Unbalanced enable for IRQ 10 WARNING: kernel/irq/manage.c:753 at __enable_irq+0x102/0x190 kernel/irq/manage.c:753, CPU#1: kworker/1:0H/26 Modules linked in: CPU: 1 UID: 0 PID: 26 Comm: kworker/1:0H Not tainted 6.19.0-dirty #9 PREEMPT(voluntary) Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/01/2014 Workqueue: kblockd blk_mq_timeout_work RIP: 0010:__enable_irq+0x107/0x190 kernel/irq/manage.c:753 Code: ff df 48 89 fa 48 c1 ea 03 0f b6 14 02 48 89 f8 83 e0 07 83 c0 03 38 d0 7c 04 84 d2 75 79 48 8d 3d 2e 7a 3f 05 41 8b 74 24 2c <67> 48 0f b9 3a e8 ef b9 21 00 5b 41 5c 5d e9 46 54 66 03 e8 e1 b9 RSP: 0018:ffffc900001bf550 EFLAGS: 00010046 RAX: 0000000000000007 RBX: 0000000000000000 RCX: ffffffffb20c0e90 RDX: 0000000000000000 RSI: 000000000000000a RDI: ffffffffb74b88f0 RBP: ffffc900001bf560 R08: ffff88800197cf00 R09: 0000000000000001 R10: 0000000000000003 R11: 0000000000000003 R12: ffff8880012a6000 R13: 1ffff92000037eae R14: 000000000000000a R15: 0000000000000293 FS: 0000000000000000(0000) GS:ffff8880b49f7000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000555da4a25fa8 CR3: 00000000208e8000 CR4: 00000000000006f0 Call Trace: <TASK> enable_irq+0x121/0x1e0 kernel/irq/manage.c:797 nvme_poll_irqdisable+0x162/0x1c0 drivers/nvme/host/pci.c:1494 nvme_timeout+0x965/0x14b0 drivers/nvme/host/pci.c:1744 blk_mq_rq_timed_out block/blk-mq.c:1653 [inline] blk_mq_handle_expired+0x227/0x2d0 block/blk-mq.c:1721 bt_iter+0x2fc/0x3a0 block/blk-mq-tag.c:292 __sbitmap_for_each_set include/linux/sbitmap.h:269 [inline] sbitmap_for_each_set include/linux/sbitmap.h:290 [inline] bt_for_each block/blk-mq-tag.c:324 [inline] blk_mq_queue_tag_busy_iter+0x969/0x1e80 block/blk-mq-tag.c:536 blk_mq_timeout_work+0x627/0x870 block/blk-mq.c:1763 process_one_work+0x956/0x1aa0 kernel/workqueue.c:3257 process_scheduled_works kernel/workqueue.c:3340 [inline] worker_thread+0x65c/0xe60 kernel/workqueue.c:3421 kthread+0x41a/0x930 kernel/kthread.c:463 ret_from_fork+0x6f8/0x8c0 arch/x86/kernel/process.c:158 ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:246 </TASK> irq event stamp: 74478 hardirqs last enabled at (74477): [<ffffffffb5720a9c>] __raw_spin_unlock_irq include/linux/spinlock_api_smp.h:159 [inline] hardirqs last enabled at (74477): [<ffffffffb5720a9c>] _raw_spin_unlock_irq+0x2c/0x60 kernel/locking/spinlock.c:202 hardirqs last disabled at (74478): [<ffffffffb57207b5>] __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:108 [inline] hardirqs last disabled at (74478): [<ffffffffb57207b5>] _raw_spin_lock_irqsave+0x85/0xa0 kernel/locking/spinlock.c:162 softirqs last enabled at (74304): [<ffffffffb1e9466c>] __do_softirq kernel/softirq.c:656 [inline] softirqs last enabled at (74304): [<ffffffffb1e9466c>] invoke_softirq kernel/softirq.c:496 [inline] softirqs last enabled at (74304): [<ffffffffb1e9466c>] __irq_exit_rcu+0xdc/0x120 kernel/softirq.c:723 softirqs last disabled at (74287): [<ffffffffb1e9466c>] __do_softirq kernel/softirq.c:656 [inline] softirqs last disabled at (74287): [<ffffffffb1e9466c>] invoke_softirq kernel/softirq.c:496 [inline] softirqs last disabled at (74287): [<ffffffffb1e9466c>] __irq_exit_rcu+0xdc/0x120 kernel/softirq.c:723 ---[ end trace 0000000000000000 ]--- Fixes: fa059b856a59 (nvme-pci: Simplify nvme_poll_irqdisable) Acked-by: Chao Shi <cshi008@fiu.edu> Acked-by: Weidong Zhu <weizhu@fiu.edu> Acked-by: Dave Tian <daveti@purdue.edu> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Sungwoo Kim <iam@sung-woo.kim> Signed-off-by: Keith Busch <kbusch@kernel.org>
2026-03-07 14:46:36 -05:00
nvme-pci: Fix race bug in nvme_poll_irqdisable() In the following scenario, pdev can be disabled between (1) and (3) by (2). This sets pdev->msix_enabled = 0. Then, pci_irq_vector() will return MSI-X IRQ(>15) for (1) whereas return INTx IRQ(<=15) for (2). This causes IRQ warning because it tries to enable INTx IRQ that has never been disabled before. To fix this, save IRQ number into a local variable and ensure disable_irq() and enable_irq() operate on the same IRQ number. Even if pci_free_irq_vectors() frees the IRQ concurrently, disable_irq() and enable_irq() on a stale IRQ number is still valid and safe, and the depth accounting reamins balanced. task 1: nvme_poll_irqdisable() disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector)) ...(1) enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector)) ...(3) task 2: nvme_reset_work() nvme_dev_disable() pdev->msix_enable = 0; ...(2) crash log: ------------[ cut here ]------------ Unbalanced enable for IRQ 10 WARNING: kernel/irq/manage.c:753 at __enable_irq+0x102/0x190 kernel/irq/manage.c:753, CPU#1: kworker/1:0H/26 Modules linked in: CPU: 1 UID: 0 PID: 26 Comm: kworker/1:0H Not tainted 6.19.0-dirty #9 PREEMPT(voluntary) Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/01/2014 Workqueue: kblockd blk_mq_timeout_work RIP: 0010:__enable_irq+0x107/0x190 kernel/irq/manage.c:753 Code: ff df 48 89 fa 48 c1 ea 03 0f b6 14 02 48 89 f8 83 e0 07 83 c0 03 38 d0 7c 04 84 d2 75 79 48 8d 3d 2e 7a 3f 05 41 8b 74 24 2c <67> 48 0f b9 3a e8 ef b9 21 00 5b 41 5c 5d e9 46 54 66 03 e8 e1 b9 RSP: 0018:ffffc900001bf550 EFLAGS: 00010046 RAX: 0000000000000007 RBX: 0000000000000000 RCX: ffffffffb20c0e90 RDX: 0000000000000000 RSI: 000000000000000a RDI: ffffffffb74b88f0 RBP: ffffc900001bf560 R08: ffff88800197cf00 R09: 0000000000000001 R10: 0000000000000003 R11: 0000000000000003 R12: ffff8880012a6000 R13: 1ffff92000037eae R14: 000000000000000a R15: 0000000000000293 FS: 0000000000000000(0000) GS:ffff8880b49f7000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000555da4a25fa8 CR3: 00000000208e8000 CR4: 00000000000006f0 Call Trace: <TASK> enable_irq+0x121/0x1e0 kernel/irq/manage.c:797 nvme_poll_irqdisable+0x162/0x1c0 drivers/nvme/host/pci.c:1494 nvme_timeout+0x965/0x14b0 drivers/nvme/host/pci.c:1744 blk_mq_rq_timed_out block/blk-mq.c:1653 [inline] blk_mq_handle_expired+0x227/0x2d0 block/blk-mq.c:1721 bt_iter+0x2fc/0x3a0 block/blk-mq-tag.c:292 __sbitmap_for_each_set include/linux/sbitmap.h:269 [inline] sbitmap_for_each_set include/linux/sbitmap.h:290 [inline] bt_for_each block/blk-mq-tag.c:324 [inline] blk_mq_queue_tag_busy_iter+0x969/0x1e80 block/blk-mq-tag.c:536 blk_mq_timeout_work+0x627/0x870 block/blk-mq.c:1763 process_one_work+0x956/0x1aa0 kernel/workqueue.c:3257 process_scheduled_works kernel/workqueue.c:3340 [inline] worker_thread+0x65c/0xe60 kernel/workqueue.c:3421 kthread+0x41a/0x930 kernel/kthread.c:463 ret_from_fork+0x6f8/0x8c0 arch/x86/kernel/process.c:158 ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:246 </TASK> irq event stamp: 74478 hardirqs last enabled at (74477): [<ffffffffb5720a9c>] __raw_spin_unlock_irq include/linux/spinlock_api_smp.h:159 [inline] hardirqs last enabled at (74477): [<ffffffffb5720a9c>] _raw_spin_unlock_irq+0x2c/0x60 kernel/locking/spinlock.c:202 hardirqs last disabled at (74478): [<ffffffffb57207b5>] __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:108 [inline] hardirqs last disabled at (74478): [<ffffffffb57207b5>] _raw_spin_lock_irqsave+0x85/0xa0 kernel/locking/spinlock.c:162 softirqs last enabled at (74304): [<ffffffffb1e9466c>] __do_softirq kernel/softirq.c:656 [inline] softirqs last enabled at (74304): [<ffffffffb1e9466c>] invoke_softirq kernel/softirq.c:496 [inline] softirqs last enabled at (74304): [<ffffffffb1e9466c>] __irq_exit_rcu+0xdc/0x120 kernel/softirq.c:723 softirqs last disabled at (74287): [<ffffffffb1e9466c>] __do_softirq kernel/softirq.c:656 [inline] softirqs last disabled at (74287): [<ffffffffb1e9466c>] invoke_softirq kernel/softirq.c:496 [inline] softirqs last disabled at (74287): [<ffffffffb1e9466c>] __irq_exit_rcu+0xdc/0x120 kernel/softirq.c:723 ---[ end trace 0000000000000000 ]--- Fixes: fa059b856a59 (nvme-pci: Simplify nvme_poll_irqdisable) Acked-by: Chao Shi <cshi008@fiu.edu> Acked-by: Weidong Zhu <weizhu@fiu.edu> Acked-by: Dave Tian <daveti@purdue.edu> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Sungwoo Kim <iam@sung-woo.kim> Signed-off-by: Keith Busch <kbusch@kernel.org>
2026-03-07 14:46:36 -05:00
nvme-pci: Fix race bug in nvme_poll_irqdisable() In the following scenario, pdev can be disabled between (1) and (3) by (2). This sets pdev->msix_enabled = 0. Then, pci_irq_vector() will return MSI-X IRQ(>15) for (1) whereas return INTx IRQ(<=15) for (2). This causes IRQ warning because it tries to enable INTx IRQ that has never been disabled before. To fix this, save IRQ number into a local variable and ensure disable_irq() and enable_irq() operate on the same IRQ number. Even if pci_free_irq_vectors() frees the IRQ concurrently, disable_irq() and enable_irq() on a stale IRQ number is still valid and safe, and the depth accounting reamins balanced. task 1: nvme_poll_irqdisable() disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector)) ...(1) enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector)) ...(3) task 2: nvme_reset_work() nvme_dev_disable() pdev->msix_enable = 0; ...(2) crash log: ------------[ cut here ]------------ Unbalanced enable for IRQ 10 WARNING: kernel/irq/manage.c:753 at __enable_irq+0x102/0x190 kernel/irq/manage.c:753, CPU#1: kworker/1:0H/26 Modules linked in: CPU: 1 UID: 0 PID: 26 Comm: kworker/1:0H Not tainted 6.19.0-dirty #9 PREEMPT(voluntary) Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/01/2014 Workqueue: kblockd blk_mq_timeout_work RIP: 0010:__enable_irq+0x107/0x190 kernel/irq/manage.c:753 Code: ff df 48 89 fa 48 c1 ea 03 0f b6 14 02 48 89 f8 83 e0 07 83 c0 03 38 d0 7c 04 84 d2 75 79 48 8d 3d 2e 7a 3f 05 41 8b 74 24 2c <67> 48 0f b9 3a e8 ef b9 21 00 5b 41 5c 5d e9 46 54 66 03 e8 e1 b9 RSP: 0018:ffffc900001bf550 EFLAGS: 00010046 RAX: 0000000000000007 RBX: 0000000000000000 RCX: ffffffffb20c0e90 RDX: 0000000000000000 RSI: 000000000000000a RDI: ffffffffb74b88f0 RBP: ffffc900001bf560 R08: ffff88800197cf00 R09: 0000000000000001 R10: 0000000000000003 R11: 0000000000000003 R12: ffff8880012a6000 R13: 1ffff92000037eae R14: 000000000000000a R15: 0000000000000293 FS: 0000000000000000(0000) GS:ffff8880b49f7000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000555da4a25fa8 CR3: 00000000208e8000 CR4: 00000000000006f0 Call Trace: <TASK> enable_irq+0x121/0x1e0 kernel/irq/manage.c:797 nvme_poll_irqdisable+0x162/0x1c0 drivers/nvme/host/pci.c:1494 nvme_timeout+0x965/0x14b0 drivers/nvme/host/pci.c:1744 blk_mq_rq_timed_out block/blk-mq.c:1653 [inline] blk_mq_handle_expired+0x227/0x2d0 block/blk-mq.c:1721 bt_iter+0x2fc/0x3a0 block/blk-mq-tag.c:292 __sbitmap_for_each_set include/linux/sbitmap.h:269 [inline] sbitmap_for_each_set include/linux/sbitmap.h:290 [inline] bt_for_each block/blk-mq-tag.c:324 [inline] blk_mq_queue_tag_busy_iter+0x969/0x1e80 block/blk-mq-tag.c:536 blk_mq_timeout_work+0x627/0x870 block/blk-mq.c:1763 process_one_work+0x956/0x1aa0 kernel/workqueue.c:3257 process_scheduled_works kernel/workqueue.c:3340 [inline] worker_thread+0x65c/0xe60 kernel/workqueue.c:3421 kthread+0x41a/0x930 kernel/kthread.c:463 ret_from_fork+0x6f8/0x8c0 arch/x86/kernel/process.c:158 ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:246 </TASK> irq event stamp: 74478 hardirqs last enabled at (74477): [<ffffffffb5720a9c>] __raw_spin_unlock_irq include/linux/spinlock_api_smp.h:159 [inline] hardirqs last enabled at (74477): [<ffffffffb5720a9c>] _raw_spin_unlock_irq+0x2c/0x60 kernel/locking/spinlock.c:202 hardirqs last disabled at (74478): [<ffffffffb57207b5>] __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:108 [inline] hardirqs last disabled at (74478): [<ffffffffb57207b5>] _raw_spin_lock_irqsave+0x85/0xa0 kernel/locking/spinlock.c:162 softirqs last enabled at (74304): [<ffffffffb1e9466c>] __do_softirq kernel/softirq.c:656 [inline] softirqs last enabled at (74304): [<ffffffffb1e9466c>] invoke_softirq kernel/softirq.c:496 [inline] softirqs last enabled at (74304): [<ffffffffb1e9466c>] __irq_exit_rcu+0xdc/0x120 kernel/softirq.c:723 softirqs last disabled at (74287): [<ffffffffb1e9466c>] __do_softirq kernel/softirq.c:656 [inline] softirqs last disabled at (74287): [<ffffffffb1e9466c>] invoke_softirq kernel/softirq.c:496 [inline] softirqs last disabled at (74287): [<ffffffffb1e9466c>] __irq_exit_rcu+0xdc/0x120 kernel/softirq.c:723 ---[ end trace 0000000000000000 ]--- Fixes: fa059b856a59 (nvme-pci: Simplify nvme_poll_irqdisable) Acked-by: Chao Shi <cshi008@fiu.edu> Acked-by: Weidong Zhu <weizhu@fiu.edu> Acked-by: Dave Tian <daveti@purdue.edu> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Sungwoo Kim <iam@sung-woo.kim> Signed-off-by: Keith Busch <kbusch@kernel.org>
2026-03-07 14:46:36 -05:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
nvme-pci: cancel nvme device request before disabling This patch addresses an irq free warning and null pointer dereference error problem when nvme devices got timeout error during initialization. This problem happens when nvme_timeout() function is called while nvme_reset_work() is still in execution. This patch fixed the problem by setting flag of the problematic request to NVME_REQ_CANCELLED before calling nvme_dev_disable() to make sure __nvme_submit_sync_cmd() returns an error code and let nvme_submit_sync_cmd() fail gracefully. The following is console output. [ 62.472097] nvme nvme0: I/O 13 QID 0 timeout, disable controller [ 62.488796] nvme nvme0: could not set timestamp (881) [ 62.494888] ------------[ cut here ]------------ [ 62.495142] Trying to free already-free IRQ 11 [ 62.495366] WARNING: CPU: 0 PID: 7 at kernel/irq/manage.c:1751 free_irq+0x1f7/0x370 [ 62.495742] Modules linked in: [ 62.495902] CPU: 0 PID: 7 Comm: kworker/u4:0 Not tainted 5.8.0+ #8 [ 62.496206] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.13.0-48-gd9c812dda519-p4 [ 62.496772] Workqueue: nvme-reset-wq nvme_reset_work [ 62.497019] RIP: 0010:free_irq+0x1f7/0x370 [ 62.497223] Code: e8 ce 49 11 00 48 83 c4 08 4c 89 e0 5b 5d 41 5c 41 5d 41 5e 41 5f c3 44 89 f6 48 c70 [ 62.498133] RSP: 0000:ffffa96800043d40 EFLAGS: 00010086 [ 62.498391] RAX: 0000000000000000 RBX: ffff9b87fc458400 RCX: 0000000000000000 [ 62.498741] RDX: 0000000000000001 RSI: 0000000000000096 RDI: ffffffff9693d72c [ 62.499091] RBP: ffff9b87fd4c8f60 R08: ffffa96800043bfd R09: 0000000000000163 [ 62.499440] R10: ffffa96800043bf8 R11: ffffa96800043bfd R12: ffff9b87fd4c8e00 [ 62.499790] R13: ffff9b87fd4c8ea4 R14: 000000000000000b R15: ffff9b87fd76b000 [ 62.500140] FS: 0000000000000000(0000) GS:ffff9b87fdc00000(0000) knlGS:0000000000000000 [ 62.500534] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 62.500816] CR2: 0000000000000000 CR3: 000000003aa0a000 CR4: 00000000000006f0 [ 62.501165] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 62.501515] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 62.501864] Call Trace: [ 62.501993] pci_free_irq+0x13/0x20 [ 62.502167] nvme_reset_work+0x5d0/0x12a0 [ 62.502369] ? update_load_avg+0x59/0x580 [ 62.502569] ? ttwu_queue_wakelist+0xa8/0xc0 [ 62.502780] ? try_to_wake_up+0x1a2/0x450 [ 62.502979] process_one_work+0x1d2/0x390 [ 62.503179] worker_thread+0x45/0x3b0 [ 62.503361] ? process_one_work+0x390/0x390 [ 62.503568] kthread+0xf9/0x130 [ 62.503726] ? kthread_park+0x80/0x80 [ 62.503911] ret_from_fork+0x22/0x30 [ 62.504090] ---[ end trace de9ed4a70f8d71e2 ]--- [ 123.912275] nvme nvme0: I/O 12 QID 0 timeout, disable controller [ 123.914670] nvme nvme0: 1/0/0 default/read/poll queues [ 123.916310] BUG: kernel NULL pointer dereference, address: 0000000000000000 [ 123.917469] #PF: supervisor write access in kernel mode [ 123.917725] #PF: error_code(0x0002) - not-present page [ 123.917976] PGD 0 P4D 0 [ 123.918109] Oops: 0002 [#1] SMP PTI [ 123.918283] CPU: 0 PID: 7 Comm: kworker/u4:0 Tainted: G W 5.8.0+ #8 [ 123.918650] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.13.0-48-gd9c812dda519-p4 [ 123.919219] Workqueue: nvme-reset-wq nvme_reset_work [ 123.919469] RIP: 0010:__blk_mq_alloc_map_and_request+0x21/0x80 [ 123.919757] Code: 66 0f 1f 84 00 00 00 00 00 41 55 41 54 55 48 63 ee 53 48 8b 47 68 89 ee 48 89 fb 8b4 [ 123.920657] RSP: 0000:ffffa96800043d40 EFLAGS: 00010286 [ 123.920912] RAX: ffff9b87fc4fee40 RBX: ffff9b87fc8cb008 RCX: 0000000000000000 [ 123.921258] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9b87fc618000 [ 123.921602] RBP: 0000000000000000 R08: ffff9b87fdc2c4a0 R09: ffff9b87fc616000 [ 123.921949] R10: 0000000000000000 R11: ffff9b87fffd1500 R12: 0000000000000000 [ 123.922295] R13: 0000000000000000 R14: ffff9b87fc8cb200 R15: ffff9b87fc8cb000 [ 123.922641] FS: 0000000000000000(0000) GS:ffff9b87fdc00000(0000) knlGS:0000000000000000 [ 123.923032] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 123.923312] CR2: 0000000000000000 CR3: 000000003aa0a000 CR4: 00000000000006f0 [ 123.923660] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 123.924007] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 123.924353] Call Trace: [ 123.924479] blk_mq_alloc_tag_set+0x137/0x2a0 [ 123.924694] nvme_reset_work+0xed6/0x12a0 [ 123.924898] process_one_work+0x1d2/0x390 [ 123.925099] worker_thread+0x45/0x3b0 [ 123.925280] ? process_one_work+0x390/0x390 [ 123.925486] kthread+0xf9/0x130 [ 123.925642] ? kthread_park+0x80/0x80 [ 123.925825] ret_from_fork+0x22/0x30 [ 123.926004] Modules linked in: [ 123.926158] CR2: 0000000000000000 [ 123.926322] ---[ end trace de9ed4a70f8d71e3 ]--- [ 123.926549] RIP: 0010:__blk_mq_alloc_map_and_request+0x21/0x80 [ 123.926832] Code: 66 0f 1f 84 00 00 00 00 00 41 55 41 54 55 48 63 ee 53 48 8b 47 68 89 ee 48 89 fb 8b4 [ 123.927734] RSP: 0000:ffffa96800043d40 EFLAGS: 00010286 [ 123.927989] RAX: ffff9b87fc4fee40 RBX: ffff9b87fc8cb008 RCX: 0000000000000000 [ 123.928336] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9b87fc618000 [ 123.928679] RBP: 0000000000000000 R08: ffff9b87fdc2c4a0 R09: ffff9b87fc616000 [ 123.929025] R10: 0000000000000000 R11: ffff9b87fffd1500 R12: 0000000000000000 [ 123.929370] R13: 0000000000000000 R14: ffff9b87fc8cb200 R15: ffff9b87fc8cb000 [ 123.929715] FS: 0000000000000000(0000) GS:ffff9b87fdc00000(0000) knlGS:0000000000000000 [ 123.930106] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 123.930384] CR2: 0000000000000000 CR3: 000000003aa0a000 CR4: 00000000000006f0 [ 123.930731] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 123.931077] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Co-developed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Tong Zhang <ztong0001@gmail.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
2020-08-28 10:17:08 -04:00
nvme-pci: cancel nvme device request before disabling This patch addresses an irq free warning and null pointer dereference error problem when nvme devices got timeout error during initialization. This problem happens when nvme_timeout() function is called while nvme_reset_work() is still in execution. This patch fixed the problem by setting flag of the problematic request to NVME_REQ_CANCELLED before calling nvme_dev_disable() to make sure __nvme_submit_sync_cmd() returns an error code and let nvme_submit_sync_cmd() fail gracefully. The following is console output. [ 62.472097] nvme nvme0: I/O 13 QID 0 timeout, disable controller [ 62.488796] nvme nvme0: could not set timestamp (881) [ 62.494888] ------------[ cut here ]------------ [ 62.495142] Trying to free already-free IRQ 11 [ 62.495366] WARNING: CPU: 0 PID: 7 at kernel/irq/manage.c:1751 free_irq+0x1f7/0x370 [ 62.495742] Modules linked in: [ 62.495902] CPU: 0 PID: 7 Comm: kworker/u4:0 Not tainted 5.8.0+ #8 [ 62.496206] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.13.0-48-gd9c812dda519-p4 [ 62.496772] Workqueue: nvme-reset-wq nvme_reset_work [ 62.497019] RIP: 0010:free_irq+0x1f7/0x370 [ 62.497223] Code: e8 ce 49 11 00 48 83 c4 08 4c 89 e0 5b 5d 41 5c 41 5d 41 5e 41 5f c3 44 89 f6 48 c70 [ 62.498133] RSP: 0000:ffffa96800043d40 EFLAGS: 00010086 [ 62.498391] RAX: 0000000000000000 RBX: ffff9b87fc458400 RCX: 0000000000000000 [ 62.498741] RDX: 0000000000000001 RSI: 0000000000000096 RDI: ffffffff9693d72c [ 62.499091] RBP: ffff9b87fd4c8f60 R08: ffffa96800043bfd R09: 0000000000000163 [ 62.499440] R10: ffffa96800043bf8 R11: ffffa96800043bfd R12: ffff9b87fd4c8e00 [ 62.499790] R13: ffff9b87fd4c8ea4 R14: 000000000000000b R15: ffff9b87fd76b000 [ 62.500140] FS: 0000000000000000(0000) GS:ffff9b87fdc00000(0000) knlGS:0000000000000000 [ 62.500534] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 62.500816] CR2: 0000000000000000 CR3: 000000003aa0a000 CR4: 00000000000006f0 [ 62.501165] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 62.501515] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 62.501864] Call Trace: [ 62.501993] pci_free_irq+0x13/0x20 [ 62.502167] nvme_reset_work+0x5d0/0x12a0 [ 62.502369] ? update_load_avg+0x59/0x580 [ 62.502569] ? ttwu_queue_wakelist+0xa8/0xc0 [ 62.502780] ? try_to_wake_up+0x1a2/0x450 [ 62.502979] process_one_work+0x1d2/0x390 [ 62.503179] worker_thread+0x45/0x3b0 [ 62.503361] ? process_one_work+0x390/0x390 [ 62.503568] kthread+0xf9/0x130 [ 62.503726] ? kthread_park+0x80/0x80 [ 62.503911] ret_from_fork+0x22/0x30 [ 62.504090] ---[ end trace de9ed4a70f8d71e2 ]--- [ 123.912275] nvme nvme0: I/O 12 QID 0 timeout, disable controller [ 123.914670] nvme nvme0: 1/0/0 default/read/poll queues [ 123.916310] BUG: kernel NULL pointer dereference, address: 0000000000000000 [ 123.917469] #PF: supervisor write access in kernel mode [ 123.917725] #PF: error_code(0x0002) - not-present page [ 123.917976] PGD 0 P4D 0 [ 123.918109] Oops: 0002 [#1] SMP PTI [ 123.918283] CPU: 0 PID: 7 Comm: kworker/u4:0 Tainted: G W 5.8.0+ #8 [ 123.918650] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.13.0-48-gd9c812dda519-p4 [ 123.919219] Workqueue: nvme-reset-wq nvme_reset_work [ 123.919469] RIP: 0010:__blk_mq_alloc_map_and_request+0x21/0x80 [ 123.919757] Code: 66 0f 1f 84 00 00 00 00 00 41 55 41 54 55 48 63 ee 53 48 8b 47 68 89 ee 48 89 fb 8b4 [ 123.920657] RSP: 0000:ffffa96800043d40 EFLAGS: 00010286 [ 123.920912] RAX: ffff9b87fc4fee40 RBX: ffff9b87fc8cb008 RCX: 0000000000000000 [ 123.921258] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9b87fc618000 [ 123.921602] RBP: 0000000000000000 R08: ffff9b87fdc2c4a0 R09: ffff9b87fc616000 [ 123.921949] R10: 0000000000000000 R11: ffff9b87fffd1500 R12: 0000000000000000 [ 123.922295] R13: 0000000000000000 R14: ffff9b87fc8cb200 R15: ffff9b87fc8cb000 [ 123.922641] FS: 0000000000000000(0000) GS:ffff9b87fdc00000(0000) knlGS:0000000000000000 [ 123.923032] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 123.923312] CR2: 0000000000000000 CR3: 000000003aa0a000 CR4: 00000000000006f0 [ 123.923660] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 123.924007] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 123.924353] Call Trace: [ 123.924479] blk_mq_alloc_tag_set+0x137/0x2a0 [ 123.924694] nvme_reset_work+0xed6/0x12a0 [ 123.924898] process_one_work+0x1d2/0x390 [ 123.925099] worker_thread+0x45/0x3b0 [ 123.925280] ? process_one_work+0x390/0x390 [ 123.925486] kthread+0xf9/0x130 [ 123.925642] ? kthread_park+0x80/0x80 [ 123.925825] ret_from_fork+0x22/0x30 [ 123.926004] Modules linked in: [ 123.926158] CR2: 0000000000000000 [ 123.926322] ---[ end trace de9ed4a70f8d71e3 ]--- [ 123.926549] RIP: 0010:__blk_mq_alloc_map_and_request+0x21/0x80 [ 123.926832] Code: 66 0f 1f 84 00 00 00 00 00 41 55 41 54 55 48 63 ee 53 48 8b 47 68 89 ee 48 89 fb 8b4 [ 123.927734] RSP: 0000:ffffa96800043d40 EFLAGS: 00010286 [ 123.927989] RAX: ffff9b87fc4fee40 RBX: ffff9b87fc8cb008 RCX: 0000000000000000 [ 123.928336] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9b87fc618000 [ 123.928679] RBP: 0000000000000000 R08: ffff9b87fdc2c4a0 R09: ffff9b87fc616000 [ 123.929025] R10: 0000000000000000 R11: ffff9b87fffd1500 R12: 0000000000000000 [ 123.929370] R13: 0000000000000000 R14: ffff9b87fc8cb200 R15: ffff9b87fc8cb000 [ 123.929715] FS: 0000000000000000(0000) GS:ffff9b87fdc00000(0000) knlGS:0000000000000000 [ 123.930106] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 123.930384] CR2: 0000000000000000 CR3: 000000003aa0a000 CR4: 00000000000006f0 [ 123.930731] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 123.931077] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Co-developed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Tong Zhang <ztong0001@gmail.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
2020-08-28 10:17:08 -04:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
nvme-pci: rerun irq setup on IO queue init errors If the driver is unable to create a subset of IO queues for any reason, the read/write and polled queue sets will not match the actual allocated hardware contexts. This leaves gaps in the CPU affinity mappings and causes the following kernel panic after blk_mq_map_queue_type() returns a NULL hctx. BUG: unable to handle kernel NULL pointer dereference at 0000000000000198 #PF error: [normal kernel read fault] PGD 0 P4D 0 Oops: 0000 [#1] SMP CPU: 64 PID: 1171 Comm: kworker/u259:1 Not tainted 4.20.0+ #241 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-2.fc27 04/01/2014 Workqueue: nvme-wq nvme_scan_work [nvme_core] RIP: 0010:blk_mq_init_allocated_queue+0x2d9/0x440 RSP: 0018:ffffb1bf0abc3cd0 EFLAGS: 00010286 RAX: 000000000000001f RBX: ffff8ea744cf0718 RCX: 0000000000000000 RDX: 0000000000000002 RSI: 000000000000007c RDI: ffffffff9109a820 RBP: ffff8ea7565f7008 R08: 000000000000001f R09: 000000000000003f R10: ffffb1bf0abc3c00 R11: 0000000000000000 R12: 000000000001d008 R13: ffff8ea7565f7008 R14: 000000000000003f R15: 0000000000000001 FS: 0000000000000000(0000) GS:ffff8ea757200000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000198 CR3: 0000000013058000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: blk_mq_init_queue+0x35/0x60 nvme_validate_ns+0xc6/0x7c0 [nvme_core] ? nvme_identify_ctrl.isra.56+0x7e/0xc0 [nvme_core] nvme_scan_work+0xc8/0x340 [nvme_core] ? __wake_up_common+0x6d/0x120 ? try_to_wake_up+0x55/0x410 process_one_work+0x1e9/0x3d0 worker_thread+0x2d/0x3d0 ? process_one_work+0x3d0/0x3d0 kthread+0x111/0x130 ? kthread_park+0x90/0x90 ret_from_fork+0x1f/0x30 Modules linked in: nvme nvme_core serio_raw CR2: 0000000000000198 Fix by re-running the interrupt vector setup from scratch using a reduced count that may be successful until the created queues matches the irq affinity plus polling queue sets. Signed-off-by: Keith Busch <keith.busch@intel.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2019-01-04 15:04:33 -07:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
nvme: improve performance for virtual NVMe devices This change provides a mechanism to reduce the number of MMIO doorbell writes for the NVMe driver. When running in a virtualized environment like QEMU, the cost of an MMIO is quite hefy here. The main idea for the patch is provide the device two memory location locations: 1) to store the doorbell values so they can be lookup without the doorbell MMIO write 2) to store an event index. I believe the doorbell value is obvious, the event index not so much. Similar to the virtio specification, the virtual device can tell the driver (guest OS) not to write MMIO unless you are writing past this value. FYI: doorbell values are written by the nvme driver (guest OS) and the event index is written by the virtual device (host OS). The patch implements a new admin command that will communicate where these two memory locations reside. If the command fails, the nvme driver will work as before without any optimizations. Contributions: Eric Northup <digitaleric@google.com> Frank Swiderski <fes@google.com> Ted Tso <tytso@mit.edu> Keith Busch <keith.busch@intel.com> Just to give an idea on the performance boost with the vendor extension: Running fio [1], a stock NVMe driver I get about 200K read IOPs with my vendor patch I get about 1000K read IOPs. This was running with a null device i.e. the backing device simply returned success on every read IO request. [1] Running on a 4 core machine: fio --time_based --name=benchmark --runtime=30 --filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32 --direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4 --rw=randread --blocksize=4k --randrepeat=false Signed-off-by: Rob Nelson <rlnelson@google.com> [mlin: port for upstream] Signed-off-by: Ming Lin <mlin@kernel.org> [koike: updated for upstream] Signed-off-by: Helen Koike <helen.koike@collabora.co.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <keith.busch@intel.com>
2017-04-10 12:51:07 -03:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
nvme-pci: fix use-after-free in nvme_free_host_mem() nvme_free_host_mem() frees dev->hmb_sgt via dma_free_noncontiguous() but never clears the pointer afterward. This leads to a use-after-free if nvme_free_host_mem() is called twice in the same error path. This can happen during nvme_probe() when nvme_setup_host_mem() succeeds in allocating the HMB (setting dev->hmb_sgt) but nvme_set_host_mem() fails with an I/O error: nvme_setup_host_mem() nvme_alloc_host_mem_single() -> sets dev->hmb_sgt nvme_set_host_mem() -> fails with -EIO nvme_free_host_mem() -> frees hmb_sgt, but does NOT NULL it return error nvme_probe() error path: nvme_free_host_mem() -> dev->hmb_sgt is stale, use-after-free The second call dereferences the freed sgt, causing a NULL pointer dereference in iommu_dma_free_noncontiguous() when it accesses sgt->sgl->dma_address (the backing memory has been freed and zeroed). This is reproducible on Thunderbolt-attached NVMe devices (e.g., OWC Envoy Express behind a Dell WD22TB4 dock) where the device intermittently returns I/O errors during HMB setup due to PCIe link instability. BUG: kernel NULL pointer dereference, address: 0000000000000010 RIP: 0010:iommu_dma_free_noncontiguous+0x22/0x80 Call Trace: <TASK> dma_free_noncontiguous+0x3b/0x130 nvme_free_host_mem+0x30/0xf0 [nvme] nvme_probe.cold+0xcc/0x275 [nvme] local_pci_probe+0x43/0xa0 pci_device_probe+0xeea/0x290 really_probe+0xf9/0x3b0 __driver_probe_device+0x8b/0x170 driver_probe_device+0x24/0xd0 __driver_attach_async_helper+0x6b/0x110 async_run_entry_fn+0x37/0x170 process_one_work+0x1ac/0x3d0 worker_thread+0x1b8/0x360 kthread+0xf7/0x130 ret_from_fork+0x2d8/0x3a0 ret_from_fork_asm+0x1a/0x30 </TASK> Fix this by setting dev->hmb_sgt to NULL after freeing it, so the second call takes the multi-descriptor path which safely handles the already-cleaned-up state. Fixes: 63a5c7a4b4c4 ("nvme-pci: use dma_alloc_noncontigous if possible") Signed-off-by: Chia-Lin Kao (AceLan) <acelan.kao@canonical.com> Signed-off-by: Keith Busch <kbusch@kernel.org>
2026-04-29 16:11:16 +08:00
nvme-pci: fix use-after-free in nvme_free_host_mem() nvme_free_host_mem() frees dev->hmb_sgt via dma_free_noncontiguous() but never clears the pointer afterward. This leads to a use-after-free if nvme_free_host_mem() is called twice in the same error path. This can happen during nvme_probe() when nvme_setup_host_mem() succeeds in allocating the HMB (setting dev->hmb_sgt) but nvme_set_host_mem() fails with an I/O error: nvme_setup_host_mem() nvme_alloc_host_mem_single() -> sets dev->hmb_sgt nvme_set_host_mem() -> fails with -EIO nvme_free_host_mem() -> frees hmb_sgt, but does NOT NULL it return error nvme_probe() error path: nvme_free_host_mem() -> dev->hmb_sgt is stale, use-after-free The second call dereferences the freed sgt, causing a NULL pointer dereference in iommu_dma_free_noncontiguous() when it accesses sgt->sgl->dma_address (the backing memory has been freed and zeroed). This is reproducible on Thunderbolt-attached NVMe devices (e.g., OWC Envoy Express behind a Dell WD22TB4 dock) where the device intermittently returns I/O errors during HMB setup due to PCIe link instability. BUG: kernel NULL pointer dereference, address: 0000000000000010 RIP: 0010:iommu_dma_free_noncontiguous+0x22/0x80 Call Trace: <TASK> dma_free_noncontiguous+0x3b/0x130 nvme_free_host_mem+0x30/0xf0 [nvme] nvme_probe.cold+0xcc/0x275 [nvme] local_pci_probe+0x43/0xa0 pci_device_probe+0xeea/0x290 really_probe+0xf9/0x3b0 __driver_probe_device+0x8b/0x170 driver_probe_device+0x24/0xd0 __driver_attach_async_helper+0x6b/0x110 async_run_entry_fn+0x37/0x170 process_one_work+0x1ac/0x3d0 worker_thread+0x1b8/0x360 kthread+0xf7/0x130 ret_from_fork+0x2d8/0x3a0 ret_from_fork_asm+0x1a/0x30 </TASK> Fix this by setting dev->hmb_sgt to NULL after freeing it, so the second call takes the multi-descriptor path which safely handles the already-cleaned-up state. Fixes: 63a5c7a4b4c4 ("nvme-pci: use dma_alloc_noncontigous if possible") Signed-off-by: Chia-Lin Kao (AceLan) <acelan.kao@canonical.com> Signed-off-by: Keith Busch <kbusch@kernel.org>
2026-04-29 16:11:16 +08:00
nvme-pci: fix use-after-free in nvme_free_host_mem() nvme_free_host_mem() frees dev->hmb_sgt via dma_free_noncontiguous() but never clears the pointer afterward. This leads to a use-after-free if nvme_free_host_mem() is called twice in the same error path. This can happen during nvme_probe() when nvme_setup_host_mem() succeeds in allocating the HMB (setting dev->hmb_sgt) but nvme_set_host_mem() fails with an I/O error: nvme_setup_host_mem() nvme_alloc_host_mem_single() -> sets dev->hmb_sgt nvme_set_host_mem() -> fails with -EIO nvme_free_host_mem() -> frees hmb_sgt, but does NOT NULL it return error nvme_probe() error path: nvme_free_host_mem() -> dev->hmb_sgt is stale, use-after-free The second call dereferences the freed sgt, causing a NULL pointer dereference in iommu_dma_free_noncontiguous() when it accesses sgt->sgl->dma_address (the backing memory has been freed and zeroed). This is reproducible on Thunderbolt-attached NVMe devices (e.g., OWC Envoy Express behind a Dell WD22TB4 dock) where the device intermittently returns I/O errors during HMB setup due to PCIe link instability. BUG: kernel NULL pointer dereference, address: 0000000000000010 RIP: 0010:iommu_dma_free_noncontiguous+0x22/0x80 Call Trace: <TASK> dma_free_noncontiguous+0x3b/0x130 nvme_free_host_mem+0x30/0xf0 [nvme] nvme_probe.cold+0xcc/0x275 [nvme] local_pci_probe+0x43/0xa0 pci_device_probe+0xeea/0x290 really_probe+0xf9/0x3b0 __driver_probe_device+0x8b/0x170 driver_probe_device+0x24/0xd0 __driver_attach_async_helper+0x6b/0x110 async_run_entry_fn+0x37/0x170 process_one_work+0x1ac/0x3d0 worker_thread+0x1b8/0x360 kthread+0xf7/0x130 ret_from_fork+0x2d8/0x3a0 ret_from_fork_asm+0x1a/0x30 </TASK> Fix this by setting dev->hmb_sgt to NULL after freeing it, so the second call takes the multi-descriptor path which safely handles the already-cleaned-up state. Fixes: 63a5c7a4b4c4 ("nvme-pci: use dma_alloc_noncontigous if possible") Signed-off-by: Chia-Lin Kao (AceLan) <acelan.kao@canonical.com> Signed-off-by: Keith Busch <kbusch@kernel.org>
2026-04-29 16:11:16 +08:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
genirq/affinity: Store interrupt sets size in struct irq_affinity The interrupt affinity spreading mechanism supports to spread out affinities for one or more interrupt sets. A interrupt set contains one or more interrupts. Each set is mapped to a specific functionality of a device, e.g. general I/O queues and read I/O queus of multiqueue block devices. The number of interrupts per set is defined by the driver. It depends on the total number of available interrupts for the device, which is determined by the PCI capabilites and the availability of underlying CPU resources, and the number of queues which the device provides and the driver wants to instantiate. The driver passes initial configuration for the interrupt allocation via a pointer to struct irq_affinity. Right now the allocation mechanism is complex as it requires to have a loop in the driver to determine the maximum number of interrupts which are provided by the PCI capabilities and the underlying CPU resources. This loop would have to be replicated in every driver which wants to utilize this mechanism. That's unwanted code duplication and error prone. In order to move this into generic facilities it is required to have a mechanism, which allows the recalculation of the interrupt sets and their size, in the core code. As the core code does not have any knowledge about the underlying device, a driver specific callback will be added to struct affinity_desc, which will be invoked by the core code. The callback will get the number of available interupts as an argument, so the driver can calculate the corresponding number and size of interrupt sets. To support this, two modifications for the handling of struct irq_affinity are required: 1) The (optional) interrupt sets size information is contained in a separate array of integers and struct irq_affinity contains a pointer to it. This is cumbersome and as the maximum number of interrupt sets is small, there is no reason to have separate storage. Moving the size array into struct affinity_desc avoids indirections and makes the code simpler. 2) At the moment the struct irq_affinity pointer which is handed in from the driver and passed through to several core functions is marked 'const'. With the upcoming callback to recalculate the number and size of interrupt sets, it's necessary to remove the 'const' qualifier. Otherwise the callback would not be able to update the data. Implement #1 and store the interrupt sets size in 'struct irq_affinity'. No functional change. [ tglx: Fixed the memcpy() size so it won't copy beyond the size of the source. Fixed the kernel doc comments for struct irq_affinity and de-'This patch'-ed the changelog ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.423723127@linutronix.de
2019-02-16 18:13:08 +01:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: rerun irq setup on IO queue init errors If the driver is unable to create a subset of IO queues for any reason, the read/write and polled queue sets will not match the actual allocated hardware contexts. This leaves gaps in the CPU affinity mappings and causes the following kernel panic after blk_mq_map_queue_type() returns a NULL hctx. BUG: unable to handle kernel NULL pointer dereference at 0000000000000198 #PF error: [normal kernel read fault] PGD 0 P4D 0 Oops: 0000 [#1] SMP CPU: 64 PID: 1171 Comm: kworker/u259:1 Not tainted 4.20.0+ #241 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-2.fc27 04/01/2014 Workqueue: nvme-wq nvme_scan_work [nvme_core] RIP: 0010:blk_mq_init_allocated_queue+0x2d9/0x440 RSP: 0018:ffffb1bf0abc3cd0 EFLAGS: 00010286 RAX: 000000000000001f RBX: ffff8ea744cf0718 RCX: 0000000000000000 RDX: 0000000000000002 RSI: 000000000000007c RDI: ffffffff9109a820 RBP: ffff8ea7565f7008 R08: 000000000000001f R09: 000000000000003f R10: ffffb1bf0abc3c00 R11: 0000000000000000 R12: 000000000001d008 R13: ffff8ea7565f7008 R14: 000000000000003f R15: 0000000000000001 FS: 0000000000000000(0000) GS:ffff8ea757200000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000198 CR3: 0000000013058000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: blk_mq_init_queue+0x35/0x60 nvme_validate_ns+0xc6/0x7c0 [nvme_core] ? nvme_identify_ctrl.isra.56+0x7e/0xc0 [nvme_core] nvme_scan_work+0xc8/0x340 [nvme_core] ? __wake_up_common+0x6d/0x120 ? try_to_wake_up+0x55/0x410 process_one_work+0x1e9/0x3d0 worker_thread+0x2d/0x3d0 ? process_one_work+0x3d0/0x3d0 kthread+0x111/0x130 ? kthread_park+0x90/0x90 ret_from_fork+0x1f/0x30 Modules linked in: nvme nvme_core serio_raw CR2: 0000000000000198 Fix by re-running the interrupt vector setup from scratch using a reduced count that may be successful until the created queues matches the irq affinity plus polling queue sets. Signed-off-by: Keith Busch <keith.busch@intel.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2019-01-04 15:04:33 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: rerun irq setup on IO queue init errors If the driver is unable to create a subset of IO queues for any reason, the read/write and polled queue sets will not match the actual allocated hardware contexts. This leaves gaps in the CPU affinity mappings and causes the following kernel panic after blk_mq_map_queue_type() returns a NULL hctx. BUG: unable to handle kernel NULL pointer dereference at 0000000000000198 #PF error: [normal kernel read fault] PGD 0 P4D 0 Oops: 0000 [#1] SMP CPU: 64 PID: 1171 Comm: kworker/u259:1 Not tainted 4.20.0+ #241 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-2.fc27 04/01/2014 Workqueue: nvme-wq nvme_scan_work [nvme_core] RIP: 0010:blk_mq_init_allocated_queue+0x2d9/0x440 RSP: 0018:ffffb1bf0abc3cd0 EFLAGS: 00010286 RAX: 000000000000001f RBX: ffff8ea744cf0718 RCX: 0000000000000000 RDX: 0000000000000002 RSI: 000000000000007c RDI: ffffffff9109a820 RBP: ffff8ea7565f7008 R08: 000000000000001f R09: 000000000000003f R10: ffffb1bf0abc3c00 R11: 0000000000000000 R12: 000000000001d008 R13: ffff8ea7565f7008 R14: 000000000000003f R15: 0000000000000001 FS: 0000000000000000(0000) GS:ffff8ea757200000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000198 CR3: 0000000013058000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: blk_mq_init_queue+0x35/0x60 nvme_validate_ns+0xc6/0x7c0 [nvme_core] ? nvme_identify_ctrl.isra.56+0x7e/0xc0 [nvme_core] nvme_scan_work+0xc8/0x340 [nvme_core] ? __wake_up_common+0x6d/0x120 ? try_to_wake_up+0x55/0x410 process_one_work+0x1e9/0x3d0 worker_thread+0x2d/0x3d0 ? process_one_work+0x3d0/0x3d0 kthread+0x111/0x130 ? kthread_park+0x90/0x90 ret_from_fork+0x1f/0x30 Modules linked in: nvme nvme_core serio_raw CR2: 0000000000000198 Fix by re-running the interrupt vector setup from scratch using a reduced count that may be successful until the created queues matches the irq affinity plus polling queue sets. Signed-off-by: Keith Busch <keith.busch@intel.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2019-01-04 15:04:33 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: rerun irq setup on IO queue init errors If the driver is unable to create a subset of IO queues for any reason, the read/write and polled queue sets will not match the actual allocated hardware contexts. This leaves gaps in the CPU affinity mappings and causes the following kernel panic after blk_mq_map_queue_type() returns a NULL hctx. BUG: unable to handle kernel NULL pointer dereference at 0000000000000198 #PF error: [normal kernel read fault] PGD 0 P4D 0 Oops: 0000 [#1] SMP CPU: 64 PID: 1171 Comm: kworker/u259:1 Not tainted 4.20.0+ #241 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-2.fc27 04/01/2014 Workqueue: nvme-wq nvme_scan_work [nvme_core] RIP: 0010:blk_mq_init_allocated_queue+0x2d9/0x440 RSP: 0018:ffffb1bf0abc3cd0 EFLAGS: 00010286 RAX: 000000000000001f RBX: ffff8ea744cf0718 RCX: 0000000000000000 RDX: 0000000000000002 RSI: 000000000000007c RDI: ffffffff9109a820 RBP: ffff8ea7565f7008 R08: 000000000000001f R09: 000000000000003f R10: ffffb1bf0abc3c00 R11: 0000000000000000 R12: 000000000001d008 R13: ffff8ea7565f7008 R14: 000000000000003f R15: 0000000000000001 FS: 0000000000000000(0000) GS:ffff8ea757200000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000198 CR3: 0000000013058000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: blk_mq_init_queue+0x35/0x60 nvme_validate_ns+0xc6/0x7c0 [nvme_core] ? nvme_identify_ctrl.isra.56+0x7e/0xc0 [nvme_core] nvme_scan_work+0xc8/0x340 [nvme_core] ? __wake_up_common+0x6d/0x120 ? try_to_wake_up+0x55/0x410 process_one_work+0x1e9/0x3d0 worker_thread+0x2d/0x3d0 ? process_one_work+0x3d0/0x3d0 kthread+0x111/0x130 ? kthread_park+0x90/0x90 ret_from_fork+0x1f/0x30 Modules linked in: nvme nvme_core serio_raw CR2: 0000000000000198 Fix by re-running the interrupt vector setup from scratch using a reduced count that may be successful until the created queues matches the irq affinity plus polling queue sets. Signed-off-by: Keith Busch <keith.busch@intel.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2019-01-04 15:04:33 -07:00
nvme-pci: fix multiple races in nvme_setup_io_queues Below two paths could overlap each other if we power off a drive quickly after powering it on. There are multiple races in nvme_setup_io_queues() because of shutdown_lock missing and improper use of NVMEQ_ENABLED bit. nvme_reset_work() nvme_remove() nvme_setup_io_queues() nvme_dev_disable() ... ... A1 clear NVMEQ_ENABLED bit for admin queue lock retry: B1 nvme_suspend_io_queues() A2 pci_free_irq() admin queue B2 nvme_suspend_queue() admin queue A3 pci_free_irq_vectors() nvme_pci_disable() A4 nvme_setup_irqs(); B3 pci_free_irq_vectors() ... unlock A5 queue_request_irq() for admin queue set NVMEQ_ENABLED bit ... nvme_create_io_queues() A6 result = queue_request_irq(); set NVMEQ_ENABLED bit ... fail to allocate enough IO queues: A7 nvme_suspend_io_queues() goto retry If B3 runs in between A1 and A2, it will crash if irqaction haven't been freed by A2. B2 is supposed to free admin queue IRQ but it simply can't fulfill the job as A1 has cleared NVMEQ_ENABLED bit. Fix: combine A1 A2 so IRQ get freed as soon as the NVMEQ_ENABLED bit gets cleared. After solved #1, A2 could race with B3 if A2 is freeing IRQ while B3 is checking irqaction. A3 also could race with B2 if B2 is freeing IRQ while A3 is checking irqaction. Fix: A2 and A3 take lock for mutual exclusion. A3 could race with B3 since they could run free_msi_irqs() in parallel. Fix: A3 takes lock for mutual exclusion. A4 could fail to allocate all needed IRQ vectors if A3 and A4 are interrupted by B3. Fix: A4 takes lock for mutual exclusion. If A5/A6 happened after B2/B1, B3 will crash since irqaction is not NULL. They are just allocated by A5/A6. Fix: Lock queue_request_irq() and setting of NVMEQ_ENABLED bit. A7 could get chance to pci_free_irq() for certain IO queue while B3 is checking irqaction. Fix: A7 takes lock. nvme_dev->online_queues need to be protected by shutdown_lock. Since it is not atomic, both paths could modify it using its own copy. Co-developed-by: Yuanyuan Zhong <yzhong@purestorage.com> Signed-off-by: Casey Chen <cachen@purestorage.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-07-07 14:14:31 -07:00
nvme-pci: Use u32 for nvme_dev.q_depth and nvme_queue.q_depth Recently nvme_dev.q_depth was changed from an int to u16 type. This falls over for the queue depth calculation in nvme_pci_enable(), where NVME_CAP_MQES(dev->ctrl.cap) + 1 may overflow as a u16, as NVME_CAP_MQES() is a 16b number also. That happens for me, and this is the result: root@ubuntu:/home/john# [148.272996] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000010 Mem abort info: ESR = 0x96000004 EC = 0x25: DABT (current EL), IL = 32 bits SET = 0, FnV = 0 EA = 0, S1PTW = 0 Data abort info: ISV = 0, ISS = 0x00000004 CM = 0, WnR = 0 user pgtable: 4k pages, 48-bit VAs, pgdp=00000a27bf3c9000 [0000000000000010] pgd=0000000000000000, p4d=0000000000000000 Internal error: Oops: 96000004 [#1] PREEMPT SMP Modules linked in: nvme nvme_core CPU: 56 PID: 256 Comm: kworker/u195:0 Not tainted 5.8.0-next-20200812 #27 Hardware name: Huawei D06 /D06, BIOS Hisilicon D06 UEFI RC0 - V1.16.01 03/15/2019 Workqueue: nvme-reset-wq nvme_reset_work [nvme] pstate: 80c00009 (Nzcv daif +PAN +UAO BTYPE=--) pc : __sg_alloc_table_from_pages+0xec/0x238 lr : __sg_alloc_table_from_pages+0xc8/0x238 sp : ffff800013ccbad0 x29: ffff800013ccbad0 x28: ffff0a27b3d380a8 x27: 0000000000000000 x26: 0000000000002dc2 x25: 0000000000000dc0 x24: 0000000000000000 x23: 0000000000000000 x22: ffff800013ccbbe8 x21: 0000000000000010 x20: 0000000000000000 x19: 00000000fffff000 x18: ffffffffffffffff x17: 00000000000000c0 x16: fffffe289eaf6380 x15: ffff800011b59948 x14: ffff002bc8fe98f8 x13: ff00000000000000 x12: ffff8000114ca000 x11: 0000000000000000 x10: ffffffffffffffff x9 : ffffffffffffffc0 x8 : ffff0a27b5f9b6a0 x7 : 0000000000000000 x6 : 0000000000000001 x5 : ffff0a27b5f9b680 x4 : 0000000000000000 x3 : ffff0a27b5f9b680 x2 : 0000000000000000 x1 : 0000000000000001 x0 : 0000000000000000 Call trace: __sg_alloc_table_from_pages+0xec/0x238 sg_alloc_table_from_pages+0x18/0x28 iommu_dma_alloc+0x474/0x678 dma_alloc_attrs+0xd8/0xf0 nvme_alloc_queue+0x114/0x160 [nvme] nvme_reset_work+0xb34/0x14b4 [nvme] process_one_work+0x1e8/0x360 worker_thread+0x44/0x478 kthread+0x150/0x158 ret_from_fork+0x10/0x34 Code: f94002c3 6b01017f 540007c2 11000486 (f8645aa5) ---[ end trace 89bb2b72d59bf925 ]--- Fix by making onto a u32. Also use u32 for nvme_dev.q_depth, as we assign this value from nvme_dev.q_depth, and nvme_dev.q_depth will possibly hold 65536 - this avoids the same crash as above. Fixes: 61f3b8963097 ("nvme-pci: use unsigned for io queue depth") Signed-off-by: John Garry <john.garry@huawei.com> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-14 23:34:25 +08:00
nvme: Suspend all queues before deletion When nvme_delete_queue fails in the first pass of the nvme_disable_io_queues() loop, we return early, failing to suspend all of the IO queues. Later, on the nvme_pci_disable path, this causes us to disable MSI without actually having freed all the IRQs, which triggers the BUG_ON in free_msi_irqs(), as show below. This patch refactors nvme_disable_io_queues to suspend all queues before start submitting delete queue commands. This way, we ensure that we have at least returned every IRQ before continuing with the removal path. [ 487.529200] kernel BUG at ../drivers/pci/msi.c:368! cpu 0x46: Vector: 700 (Program Check) at [c0000078c5b83650] pc: c000000000627a50: free_msi_irqs+0x90/0x200 lr: c000000000627a40: free_msi_irqs+0x80/0x200 sp: c0000078c5b838d0 msr: 9000000100029033 current = 0xc0000078c5b40000 paca = 0xc000000002bd7600 softe: 0 irq_happened: 0x01 pid = 1376, comm = kworker/70:1H kernel BUG at ../drivers/pci/msi.c:368! Linux version 4.7.0.mainline+ (root@iod76) (gcc version 5.3.1 20160413 (Ubuntu/IBM 5.3.1-14ubuntu2.1) ) #104 SMP Fri Jul 29 09:20:17 CDT 2016 enter ? for help [c0000078c5b83920] d0000000363b0cd8 nvme_dev_disable+0x208/0x4f0 [nvme] [c0000078c5b83a10] d0000000363b12a4 nvme_timeout+0xe4/0x250 [nvme] [c0000078c5b83ad0] c0000000005690e4 blk_mq_rq_timed_out+0x64/0x110 [c0000078c5b83b40] c00000000056c930 bt_for_each+0x160/0x170 [c0000078c5b83bb0] c00000000056d928 blk_mq_queue_tag_busy_iter+0x78/0x110 [c0000078c5b83c00] c0000000005675d8 blk_mq_timeout_work+0xd8/0x1b0 [c0000078c5b83c50] c0000000000e8cf0 process_one_work+0x1e0/0x590 [c0000078c5b83ce0] c0000000000e9148 worker_thread+0xa8/0x660 [c0000078c5b83d80] c0000000000f2090 kthread+0x110/0x130 [c0000078c5b83e30] c0000000000095f0 ret_from_kernel_thread+0x5c/0x6c Signed-off-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com> Cc: Brian King <brking@linux.vnet.ibm.com> Cc: Keith Busch <keith.busch@intel.com> Cc: linux-nvme@lists.infradead.org Signed-off-by: Jens Axboe <axboe@fb.com>
2016-08-11 09:35:57 -06:00
nvme-pci: rerun irq setup on IO queue init errors If the driver is unable to create a subset of IO queues for any reason, the read/write and polled queue sets will not match the actual allocated hardware contexts. This leaves gaps in the CPU affinity mappings and causes the following kernel panic after blk_mq_map_queue_type() returns a NULL hctx. BUG: unable to handle kernel NULL pointer dereference at 0000000000000198 #PF error: [normal kernel read fault] PGD 0 P4D 0 Oops: 0000 [#1] SMP CPU: 64 PID: 1171 Comm: kworker/u259:1 Not tainted 4.20.0+ #241 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-2.fc27 04/01/2014 Workqueue: nvme-wq nvme_scan_work [nvme_core] RIP: 0010:blk_mq_init_allocated_queue+0x2d9/0x440 RSP: 0018:ffffb1bf0abc3cd0 EFLAGS: 00010286 RAX: 000000000000001f RBX: ffff8ea744cf0718 RCX: 0000000000000000 RDX: 0000000000000002 RSI: 000000000000007c RDI: ffffffff9109a820 RBP: ffff8ea7565f7008 R08: 000000000000001f R09: 000000000000003f R10: ffffb1bf0abc3c00 R11: 0000000000000000 R12: 000000000001d008 R13: ffff8ea7565f7008 R14: 000000000000003f R15: 0000000000000001 FS: 0000000000000000(0000) GS:ffff8ea757200000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000198 CR3: 0000000013058000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: blk_mq_init_queue+0x35/0x60 nvme_validate_ns+0xc6/0x7c0 [nvme_core] ? nvme_identify_ctrl.isra.56+0x7e/0xc0 [nvme_core] nvme_scan_work+0xc8/0x340 [nvme_core] ? __wake_up_common+0x6d/0x120 ? try_to_wake_up+0x55/0x410 process_one_work+0x1e9/0x3d0 worker_thread+0x2d/0x3d0 ? process_one_work+0x3d0/0x3d0 kthread+0x111/0x130 ? kthread_park+0x90/0x90 ret_from_fork+0x1f/0x30 Modules linked in: nvme nvme_core serio_raw CR2: 0000000000000198 Fix by re-running the interrupt vector setup from scratch using a reduced count that may be successful until the created queues matches the irq affinity plus polling queue sets. Signed-off-by: Keith Busch <keith.busch@intel.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2019-01-04 15:04:33 -07:00
nvme-pci: don't WARN_ON in nvme_reset_work if ctrl.state is not RESETTING Followling process: nvme_probe nvme_reset_ctrl nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING) queue_work(nvme_reset_wq, &ctrl->reset_work) --------------> nvme_remove nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING) worker_thread process_one_work nvme_reset_work WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING) , which will trigger WARN_ON in nvme_reset_work(): [ 127.534298] WARNING: CPU: 0 PID: 139 at drivers/nvme/host/pci.c:2594 [ 127.536161] CPU: 0 PID: 139 Comm: kworker/u8:7 Not tainted 5.13.0 [ 127.552518] Call Trace: [ 127.552840] ? kvm_sched_clock_read+0x25/0x40 [ 127.553936] ? native_send_call_func_single_ipi+0x1c/0x30 [ 127.555117] ? send_call_function_single_ipi+0x9b/0x130 [ 127.556263] ? __smp_call_single_queue+0x48/0x60 [ 127.557278] ? ttwu_queue_wakelist+0xfa/0x1c0 [ 127.558231] ? try_to_wake_up+0x265/0x9d0 [ 127.559120] ? ext4_end_io_rsv_work+0x160/0x290 [ 127.560118] process_one_work+0x28c/0x640 [ 127.561002] worker_thread+0x39a/0x700 [ 127.561833] ? rescuer_thread+0x580/0x580 [ 127.562714] kthread+0x18c/0x1e0 [ 127.563444] ? set_kthread_struct+0x70/0x70 [ 127.564347] ret_from_fork+0x1f/0x30 The preceding problem can be easily reproduced by executing following script (based on blktests suite): test() { pdev="$(_get_pci_dev_from_blkdev)" sysfs="/sys/bus/pci/devices/${pdev}" for ((i = 0; i < 10; i++)); do echo 1 > "$sysfs/remove" echo 1 > /sys/bus/pci/rescan done } Since the device ctrl could be updated as an non-RESETTING state by repeating probe/remove in userspace (which is a normal situation), we can replace stack dumping WARN_ON with a warnning message. Fixes: 82b057caefaff ("nvme-pci: fix multiple ctrl removal schedulin") Signed-off-by: Zhihao Cheng <chengzhihao1@huawei.com>
2021-07-05 21:38:29 +08:00
nvme-pci: allocate device queues storage space at probe It may cause race by setting 'nvmeq' in nvme_init_request() because .init_request is called inside switching io scheduler, which may happen when the NVMe device is being resetted and its nvme queues are being freed and created. We don't have any sync between the two pathes. This patch changes the nvmeq allocation to occur at probe time so there is no way we can dereference it at init_request. [ 93.268391] kernel BUG at drivers/nvme/host/pci.c:408! [ 93.274146] invalid opcode: 0000 [#1] SMP [ 93.278618] Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc ipmi_ssif vfat fat intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul crc32_pclmul ghash_clmulni_intel iTCO_wdt intel_cstate ipmi_si iTCO_vendor_support intel_uncore mxm_wmi mei_me ipmi_devintf intel_rapl_perf pcspkr sg ipmi_msghandler lpc_ich dcdbas mei shpchp acpi_power_meter wmi dm_multipath ip_tables xfs libcrc32c sd_mod mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops ttm drm ahci libahci nvme libata crc32c_intel nvme_core tg3 megaraid_sas ptp i2c_core pps_core dm_mirror dm_region_hash dm_log dm_mod [ 93.349071] CPU: 5 PID: 1842 Comm: sh Not tainted 4.15.0-rc2.ming+ #4 [ 93.356256] Hardware name: Dell Inc. PowerEdge R730xd/072T6D, BIOS 2.5.5 08/16/2017 [ 93.364801] task: 00000000fb8abf2a task.stack: 0000000028bd82d1 [ 93.371408] RIP: 0010:nvme_init_request+0x36/0x40 [nvme] [ 93.377333] RSP: 0018:ffffc90002537ca8 EFLAGS: 00010246 [ 93.383161] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000008 [ 93.391122] RDX: 0000000000000000 RSI: ffff880276ae0000 RDI: ffff88047bae9008 [ 93.399084] RBP: ffff88047bae9008 R08: ffff88047bae9008 R09: 0000000009dabc00 [ 93.407045] R10: 0000000000000004 R11: 000000000000299c R12: ffff880186bc1f00 [ 93.415007] R13: ffff880276ae0000 R14: 0000000000000000 R15: 0000000000000071 [ 93.422969] FS: 00007f33cf288740(0000) GS:ffff88047ba80000(0000) knlGS:0000000000000000 [ 93.431996] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 93.438407] CR2: 00007f33cf28e000 CR3: 000000047e5bb006 CR4: 00000000001606e0 [ 93.446368] Call Trace: [ 93.449103] blk_mq_alloc_rqs+0x231/0x2a0 [ 93.453579] blk_mq_sched_alloc_tags.isra.8+0x42/0x80 [ 93.459214] blk_mq_init_sched+0x7e/0x140 [ 93.463687] elevator_switch+0x5a/0x1f0 [ 93.467966] ? elevator_get.isra.17+0x52/0xc0 [ 93.472826] elv_iosched_store+0xde/0x150 [ 93.477299] queue_attr_store+0x4e/0x90 [ 93.481580] kernfs_fop_write+0xfa/0x180 [ 93.485958] __vfs_write+0x33/0x170 [ 93.489851] ? __inode_security_revalidate+0x4c/0x60 [ 93.495390] ? selinux_file_permission+0xda/0x130 [ 93.500641] ? _cond_resched+0x15/0x30 [ 93.504815] vfs_write+0xad/0x1a0 [ 93.508512] SyS_write+0x52/0xc0 [ 93.512113] do_syscall_64+0x61/0x1a0 [ 93.516199] entry_SYSCALL64_slow_path+0x25/0x25 [ 93.521351] RIP: 0033:0x7f33ce96aab0 [ 93.525337] RSP: 002b:00007ffe57570238 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 93.533785] RAX: ffffffffffffffda RBX: 0000000000000006 RCX: 00007f33ce96aab0 [ 93.541746] RDX: 0000000000000006 RSI: 00007f33cf28e000 RDI: 0000000000000001 [ 93.549707] RBP: 00007f33cf28e000 R08: 000000000000000a R09: 00007f33cf288740 [ 93.557669] R10: 00007f33cf288740 R11: 0000000000000246 R12: 00007f33cec42400 [ 93.565630] R13: 0000000000000006 R14: 0000000000000001 R15: 0000000000000000 [ 93.573592] Code: 4c 8d 40 08 4c 39 c7 74 16 48 8b 00 48 8b 04 08 48 85 c0 74 16 48 89 86 78 01 00 00 31 c0 c3 8d 4a 01 48 63 c9 48 c1 e1 03 eb de <0f> 0b 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 85 f6 53 48 89 [ 93.594676] RIP: nvme_init_request+0x36/0x40 [nvme] RSP: ffffc90002537ca8 [ 93.602273] ---[ end trace 810dde3993e5f14e ]--- Reported-by: Yi Zhang <yi.zhang@redhat.com> Signed-off-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Christoph Hellwig <hch@lst.de>
2018-01-14 12:39:01 +02:00
nvme-pci: add support for ACPI StorageD3Enable property This patch implements a solution for a BIOS hack used on some currently shipping Intel systems to change driver power management policy for PCIe NVMe drives. Some newer Intel platforms, like some Comet Lake systems, require that PCIe devices use D3 when doing suspend-to-idle in order to allow the platform to realize maximum power savings. This is particularly needed to support ATX power supply shutdown on desktop systems. In order to ensure this happens for root ports with storage devices, Microsoft apparently created this ACPI _DSD property as a way to influence their driver policy. To my knowledge this property has not been discussed with the NVME specification body. Though the solution is not ideal, it addresses a problem that also affects Linux since the NVMe driver's default policy of using NVMe APST during suspend-to-idle prevents the PCI root port from going to D3 and leads to higher power consumption for these platforms. The power consumption difference may be negligible on laptop systems, but many watts on desktop systems when the ATX power supply is blocked from powering down. The patch creates a new nvme_acpi_storage_d3 function to check for the StorageD3Enable property during probe and enables D3 as a quirk if set. It also provides a 'noacpi' module parameter to allow skipping the quirk if needed. Tested with: - PM961 NVMe SED Samsung 512GB - INTEL SSDPEKKF512G8 Link: https://docs.microsoft.com/en-us/windows-hardware/design/component-guidelines/power-management-for-storage-hardware-devices-intro Signed-off-by: David E. Box <david.e.box@linux.intel.com> Reviewed-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2020-07-09 11:43:33 -07:00
nvme-pci: add support for ACPI StorageD3Enable property This patch implements a solution for a BIOS hack used on some currently shipping Intel systems to change driver power management policy for PCIe NVMe drives. Some newer Intel platforms, like some Comet Lake systems, require that PCIe devices use D3 when doing suspend-to-idle in order to allow the platform to realize maximum power savings. This is particularly needed to support ATX power supply shutdown on desktop systems. In order to ensure this happens for root ports with storage devices, Microsoft apparently created this ACPI _DSD property as a way to influence their driver policy. To my knowledge this property has not been discussed with the NVME specification body. Though the solution is not ideal, it addresses a problem that also affects Linux since the NVMe driver's default policy of using NVMe APST during suspend-to-idle prevents the PCI root port from going to D3 and leads to higher power consumption for these platforms. The power consumption difference may be negligible on laptop systems, but many watts on desktop systems when the ATX power supply is blocked from powering down. The patch creates a new nvme_acpi_storage_d3 function to check for the StorageD3Enable property during probe and enables D3 as a quirk if set. It also provides a 'noacpi' module parameter to allow skipping the quirk if needed. Tested with: - PM961 NVMe SED Samsung 512GB - INTEL SSDPEKKF512G8 Link: https://docs.microsoft.com/en-us/windows-hardware/design/component-guidelines/power-management-for-storage-hardware-devices-intro Signed-off-by: David E. Box <david.e.box@linux.intel.com> Reviewed-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2020-07-09 11:43:33 -07:00
nvme-pci: make PRP list DMA pools per-NUMA-node NVMe commands with over 8 KB of discontiguous data allocate PRP list pages from the per-nvme_device dma_pool prp_page_pool or prp_small_pool. Each call to dma_pool_alloc() and dma_pool_free() takes the per-dma_pool spinlock. These device-global spinlocks are a significant source of contention when many CPUs are submitting to the same NVMe devices. On a workload issuing 32 KB reads from 16 CPUs (8 hypertwin pairs) across 2 NUMA nodes to 23 NVMe devices, we observed 2.4% of CPU time spent in _raw_spin_lock_irqsave called from dma_pool_alloc and dma_pool_free. Ideally, the dma_pools would be per-hctx to minimize contention. But that could impose considerable resource costs in a system with many NVMe devices and CPUs. As a compromise, allocate per-NUMA-node PRP list DMA pools. Map each nvme_queue to the set of DMA pools corresponding to its device and its hctx's NUMA node. This reduces the _raw_spin_lock_irqsave overhead by about half, to 1.2%. Preventing the sharing of PRP list pages across NUMA nodes also makes them cheaper to initialize. Link: https://lore.kernel.org/linux-nvme/CADUfDZqa=OOTtTTznXRDmBQo1WrFcDw1hBA7XwM7hzJ-hpckcA@mail.gmail.com/T/#u Signed-off-by: Caleb Sander Mateos <csander@purestorage.com> Reviewed-by: Jens Axboe <axboe@kernel.dk> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Kanchan Joshi <joshi.k@samsung.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-05-12 15:50:40 +02:00
pci-v4.13-changes -----BEGIN PGP SIGNATURE----- Version: GnuPG v1 iQIcBAABAgAGBQJZYAFUAAoJEFmIoMA60/r8cFQP/A4fpdjhd42WRNQXGTpZieop i40lBQtGdBn/UY97U6BoutcS1ygDi9OiSzg+IR6I90iMgidqyUHFhe4hGWgVHD2g Tg0KLzd+lKKfQ6Gqt1P6t4dLGLvyEj5NUbCeFE4XYODAUkkiBaOndax6DK1GvU54 Vjuj63rHtMKFR/tG/4iFTigObqyI8QE6O9JVxwuvIyEX6RXKbJe+wkulv5taSnWt Ne94950i10MrELtNreVdi8UbCbXiqjg0r5sKI/WTJ7Bc7WsC7X5PhWlhcNrbHyBT Ivhoypkui3Ky8gvwWqL0KBG+cRp8prBXAdabrD9wRbz0TKnfGI6pQzseCGRnkE6T mhlSJpsSNIHaejoCjk93yPn5oRiTNtPMdVhMpEQL9V/crVRGRRmbd7v2TYvpMHVR JaPZ8bv+C2aBTY8uL3/v/rgrjsMKOYFeaxeNklpErxrknsbgb6BgubmeZXDvTBVv YUIbAkvveonUKisv+kbD8L7tp1+jdbRUT0AikS0NVgAJQhfArOmBcDpTL9YC51vE feFhkVx4A32vvOm7Zcg9A7IMXNjeSfccKGw3dJOAvzgDODuJiaCG6S0o7B5Yngze axMi87ixGT4QM98z/I4MC8E9rDrJdIitlpvb6ZBgiLzoO3kmvsIZZKt8UxWqf5r8 w3U2HoyKH13Qbkn1xkum =mkyb -----END PGP SIGNATURE----- Merge tag 'pci-v4.13-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci Pull PCI updates from Bjorn Helgaas: - add sysfs max_link_speed/width, current_link_speed/width (Wong Vee Khee) - make host bridge IRQ mapping much more generic (Matthew Minter, Lorenzo Pieralisi) - convert most drivers to pci_scan_root_bus_bridge() (Lorenzo Pieralisi) - mutex sriov_configure() (Jakub Kicinski) - mutex pci_error_handlers callbacks (Christoph Hellwig) - split ->reset_notify() into ->reset_prepare()/reset_done() (Christoph Hellwig) - support multiple PCIe portdrv interrupts for MSI as well as MSI-X (Gabriele Paoloni) - allocate MSI/MSI-X vector for Downstream Port Containment (Gabriele Paoloni) - fix MSI IRQ affinity pre/post/min_vecs issue (Michael Hernandez) - test INTx masking during enumeration, not at run-time (Piotr Gregor) - avoid using device_may_wakeup() for runtime PM (Rafael J. Wysocki) - restore the status of PCI devices across hibernation (Chen Yu) - keep parent resources that start at 0x0 (Ard Biesheuvel) - enable ECRC only if device supports it (Bjorn Helgaas) - restore PRI and PASID state after Function-Level Reset (CQ Tang) - skip DPC event if device is not present (Keith Busch) - check domain when matching SMBIOS info (Sujith Pandel) - mark Intel XXV710 NIC INTx masking as broken (Alex Williamson) - avoid AMD SB7xx EHCI USB wakeup defect (Kai-Heng Feng) - work around long-standing Macbook Pro poweroff issue (Bjorn Helgaas) - add Switchtec "running" status flag (Logan Gunthorpe) - fix dra7xx incorrect RW1C IRQ register usage (Arvind Yadav) - modify xilinx-nwl IRQ chip for legacy interrupts (Bharat Kumar Gogada) - move VMD SRCU cleanup after bus, child device removal (Jon Derrick) - add Faraday clock handling (Linus Walleij) - configure Rockchip MPS and reorganize (Shawn Lin) - limit Qualcomm TLP size to 2K (hardware issue) (Srinivas Kandagatla) - support Tegra MSI 64-bit addressing (Thierry Reding) - use Rockchip normal (not privileged) register bank (Shawn Lin) - add HiSilicon Kirin SoC PCIe controller driver (Xiaowei Song) - add Sigma Designs Tango SMP8759 PCIe controller driver (Marc Gonzalez) - add MediaTek PCIe host controller support (Ryder Lee) - add Qualcomm IPQ4019 support (John Crispin) - add HyperV vPCI protocol v1.2 support (Jork Loeser) - add i.MX6 regulator support (Quentin Schulz) * tag 'pci-v4.13-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci: (113 commits) PCI: tango: Add Sigma Designs Tango SMP8759 PCIe host bridge support PCI: Add DT binding for Sigma Designs Tango PCIe controller PCI: rockchip: Use normal register bank for config accessors dt-bindings: PCI: Add documentation for MediaTek PCIe PCI: Remove __pci_dev_reset() and pci_dev_reset() PCI: Split ->reset_notify() method into ->reset_prepare() and ->reset_done() PCI: xilinx: Make of_device_ids const PCI: xilinx-nwl: Modify IRQ chip for legacy interrupts PCI: vmd: Move SRCU cleanup after bus, child device removal PCI: vmd: Correct comment: VMD domains start at 0x10000, not 0x1000 PCI: versatile: Add local struct device pointers PCI: tegra: Do not allocate MSI target memory PCI: tegra: Support MSI 64-bit addressing PCI: rockchip: Use local struct device pointer consistently PCI: rockchip: Check for clk_prepare_enable() errors during resume MAINTAINERS: Remove Wenrui Li as Rockchip PCIe driver maintainer PCI: rockchip: Configure RC's MPS setting PCI: rockchip: Reconfigure configuration space header type PCI: rockchip: Split out rockchip_pcie_cfg_configuration_accesses() PCI: rockchip: Move configuration accesses into rockchip_pcie_cfg_atu() ...
2017-07-08 15:51:57 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
NVMe: Convert to blk-mq This converts the NVMe driver to a blk-mq request-based driver. The NVMe driver is currently bio-based and implements queue logic within itself. By using blk-mq, a lot of these responsibilities can be moved and simplified. The patch is divided into the following blocks: * Per-command data and cmdid have been moved into the struct request field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id maintenance are now handled by blk-mq through the rq->tag field. * The logic for splitting bio's has been moved into the blk-mq layer. The driver instead notifies the block layer about limited gap support in SG lists. * blk-mq handles timeouts and is reimplemented within nvme_timeout(). This both includes abort handling and command cancelation. * Assignment of nvme queues to CPUs are replaced with the blk-mq version. The current blk-mq strategy is to assign the number of mapped queues and CPUs to provide synergy, while the nvme driver assign as many nvme hw queues as possible. This can be implemented in blk-mq if needed. * NVMe queues are merged with the tags structure of blk-mq. * blk-mq takes care of setup/teardown of nvme queues and guards invalid accesses. Therefore, RCU-usage for nvme queues can be removed. * IO tracing and accounting are handled by blk-mq and therefore removed. * Queue suspension logic is replaced with the logic from the block layer. Contributions in this patch from: Sam Bradshaw <sbradshaw@micron.com> Jens Axboe <axboe@fb.com> Keith Busch <keith.busch@intel.com> Robert Nelson <rlnelson@google.com> Acked-by: Keith Busch <keith.busch@intel.com> Acked-by: Jens Axboe <axboe@fb.com> Updated for new ->queue_rq() prototype. Signed-off-by: Jens Axboe <axboe@fb.com>
2014-11-04 08:20:14 -07:00
nvme-pci: Allow PCI bus-level PM to be used if ASPM is disabled One of the modifications made by commit d916b1be94b6 ("nvme-pci: use host managed power state for suspend") was adding a pci_save_state() call to nvme_suspend() so as to instruct the PCI bus type to leave devices handled by the nvme driver in D0 during suspend-to-idle. That was done with the assumption that ASPM would transition the device's PCIe link into a low-power state when the device became inactive. However, if ASPM is disabled for the device, its PCIe link will stay in L0 and in that case commit d916b1be94b6 is likely to cause the energy used by the system while suspended to increase. Namely, if the device in question works in accordance with the PCIe specification, putting it into D3hot causes its PCIe link to go to L1 or L2/L3 Ready, which is lower-power than L0. Since the energy used by the system while suspended depends on the state of its PCIe link (as a general rule, the lower-power the state of the link, the less energy the system will use), putting the device into D3hot during suspend-to-idle should be more energy-efficient that leaving it in D0 with disabled ASPM. For this reason, avoid leaving NVMe devices with disabled ASPM in D0 during suspend-to-idle. Instead, shut them down entirely and let the PCI bus type put them into D3. Fixes: d916b1be94b6 ("nvme-pci: use host managed power state for suspend") Link: https://lore.kernel.org/linux-pm/2763495.NmdaWeg79L@kreacher/T/#t Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Keith Busch <keith.busch@intel.com>
2019-08-08 23:58:38 +02:00
nvme-pci: Allow PCI bus-level PM to be used if ASPM is disabled One of the modifications made by commit d916b1be94b6 ("nvme-pci: use host managed power state for suspend") was adding a pci_save_state() call to nvme_suspend() so as to instruct the PCI bus type to leave devices handled by the nvme driver in D0 during suspend-to-idle. That was done with the assumption that ASPM would transition the device's PCIe link into a low-power state when the device became inactive. However, if ASPM is disabled for the device, its PCIe link will stay in L0 and in that case commit d916b1be94b6 is likely to cause the energy used by the system while suspended to increase. Namely, if the device in question works in accordance with the PCIe specification, putting it into D3hot causes its PCIe link to go to L1 or L2/L3 Ready, which is lower-power than L0. Since the energy used by the system while suspended depends on the state of its PCIe link (as a general rule, the lower-power the state of the link, the less energy the system will use), putting the device into D3hot during suspend-to-idle should be more energy-efficient that leaving it in D0 with disabled ASPM. For this reason, avoid leaving NVMe devices with disabled ASPM in D0 during suspend-to-idle. Instead, shut them down entirely and let the PCI bus type put them into D3. Fixes: d916b1be94b6 ("nvme-pci: use host managed power state for suspend") Link: https://lore.kernel.org/linux-pm/2763495.NmdaWeg79L@kreacher/T/#t Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Keith Busch <keith.busch@intel.com>
2019-08-08 23:58:38 +02:00
nvme-pci: Allow PCI bus-level PM to be used if ASPM is disabled One of the modifications made by commit d916b1be94b6 ("nvme-pci: use host managed power state for suspend") was adding a pci_save_state() call to nvme_suspend() so as to instruct the PCI bus type to leave devices handled by the nvme driver in D0 during suspend-to-idle. That was done with the assumption that ASPM would transition the device's PCIe link into a low-power state when the device became inactive. However, if ASPM is disabled for the device, its PCIe link will stay in L0 and in that case commit d916b1be94b6 is likely to cause the energy used by the system while suspended to increase. Namely, if the device in question works in accordance with the PCIe specification, putting it into D3hot causes its PCIe link to go to L1 or L2/L3 Ready, which is lower-power than L0. Since the energy used by the system while suspended depends on the state of its PCIe link (as a general rule, the lower-power the state of the link, the less energy the system will use), putting the device into D3hot during suspend-to-idle should be more energy-efficient that leaving it in D0 with disabled ASPM. For this reason, avoid leaving NVMe devices with disabled ASPM in D0 during suspend-to-idle. Instead, shut them down entirely and let the PCI bus type put them into D3. Fixes: d916b1be94b6 ("nvme-pci: use host managed power state for suspend") Link: https://lore.kernel.org/linux-pm/2763495.NmdaWeg79L@kreacher/T/#t Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Keith Busch <keith.busch@intel.com>
2019-08-08 23:58:38 +02:00
nvme-pci: 512 byte aligned dma pool segment quirk We initially introduced a quick fix limiting the queue depth to 1 as experimentation showed that it fixed data corruption on 64GB steamdecks. Further experimentation revealed corruption only happens when the last PRP data element aligns to the end of the page boundary. The device appears to treat this as a PRP chain to a new list instead of the data element that it actually is. This implementation is in violation of the spec. Encountering this errata with the Linux driver requires the host request a 128k transfer and coincidently be handed the last small pool dma buffer within a page. The QD1 quirk effectly works around this because the last data PRP always was at a 248 byte offset from the page start, so it never appeared at the end of the page, but comes at the expense of throttling IO and wasting the remainder of the PRP page beyond 256 bytes. Also to note, the MDTS on these devices is small enough that the "large" prp pool can hold enough PRP elements to never reach the end, so that pool is not a problem either. Introduce a new quirk to ensure the small pool is always aligned such that the last PRP element can't appear a the end of the page. This comes at the expense of wasting 256 bytes per small pool page allocated. Link: https://lore.kernel.org/linux-nvme/20241113043151.GA20077@lst.de/T/#u Fixes: 83bdfcbdbe5d ("nvme-pci: qdepth 1 quirk") Cc: Paweł Anikiel <panikiel@google.com> Signed-off-by: Robert Beckett <bob.beckett@collabora.com> Signed-off-by: Keith Busch <kbusch@kernel.org>
2024-11-12 19:50:00 +00:00
nvme-pci: add the DISABLE_WRITE_ZEROES quirk for a Samsung PM1725a This adds a quirk for Samsung PM1725a drive which fixes timeouts and I/O errors due to the fact that the controller does not properly handle the Write Zeroes command, dmesg log: nvme nvme0: I/O 528 QID 10 timeout, aborting nvme nvme0: I/O 529 QID 10 timeout, aborting nvme nvme0: I/O 530 QID 10 timeout, aborting nvme nvme0: I/O 531 QID 10 timeout, aborting nvme nvme0: I/O 532 QID 10 timeout, aborting nvme nvme0: I/O 533 QID 10 timeout, aborting nvme nvme0: I/O 534 QID 10 timeout, aborting nvme nvme0: I/O 535 QID 10 timeout, aborting nvme nvme0: Abort status: 0x0 nvme nvme0: Abort status: 0x0 nvme nvme0: Abort status: 0x0 nvme nvme0: Abort status: 0x0 nvme nvme0: Abort status: 0x0 nvme nvme0: Abort status: 0x0 nvme nvme0: Abort status: 0x0 nvme nvme0: Abort status: 0x0 nvme nvme0: I/O 528 QID 10 timeout, reset controller nvme nvme0: controller is down; will reset: CSTS=0x3, PCI_STATUS=0x10 nvme nvme0: Device not ready; aborting reset, CSTS=0x3 nvme nvme0: Device not ready; aborting reset, CSTS=0x3 nvme nvme0: Removing after probe failure status: -19 nvme0n1: detected capacity change from 6251233968 to 0 blk_update_request: I/O error, dev nvme0n1, sector 32776 op 0x1:(WRITE) flags 0x3000 phys_seg 6 prio class 0 blk_update_request: I/O error, dev nvme0n1, sector 113319936 op 0x9:(WRITE_ZEROES) flags 0x800 phys_seg 0 prio class 0 Buffer I/O error on dev nvme0n1p2, logical block 1, lost async page write blk_update_request: I/O error, dev nvme0n1, sector 113319680 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 Buffer I/O error on dev nvme0n1p2, logical block 2, lost async page write blk_update_request: I/O error, dev nvme0n1, sector 113319424 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 Buffer I/O error on dev nvme0n1p2, logical block 3, lost async page write blk_update_request: I/O error, dev nvme0n1, sector 113319168 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 Buffer I/O error on dev nvme0n1p2, logical block 4, lost async page write blk_update_request: I/O error, dev nvme0n1, sector 113318912 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 Buffer I/O error on dev nvme0n1p2, logical block 5, lost async page write blk_update_request: I/O error, dev nvme0n1, sector 113318656 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 Buffer I/O error on dev nvme0n1p2, logical block 6, lost async page write blk_update_request: I/O error, dev nvme0n1, sector 113318400 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 blk_update_request: I/O error, dev nvme0n1, sector 113318144 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 blk_update_request: I/O error, dev nvme0n1, sector 113317888 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 Signed-off-by: Dmitry Monakhov <dmtrmonakhov@yandex-team.ru> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-03-10 12:06:41 +00:00
nvme-pci: add quirks for WDC Blue SN550 15b7:5009 Add two quirks for the WDC Blue SN550 (PCI ID 15b7:5009) based on user reports and hardware analysis: - NVME_QUIRK_NO_DEEPEST_PS: liaozw talked to me the problem and solved with nvme_core.default_ps_max_latency_us=0, so add the quirk. I also found some reports in the following link. - NVME_QUIRK_BROKEN_MSI: after get the lspci from Jack Rio. I think that the disk also have NVME_QUIRK_BROKEN_MSI. described in commit d5887dc6b6c0 ("nvme-pci: Add quirk for broken MSIs") as sean said in link which match the MSI 1/32 and MSI-X 17. Log: lspci -nn | grep -i memory 03:00.0 Non-Volatile memory controller [0108]: Sandisk Corp SanDisk Ultra 3D / WD PC SN530, IX SN530, Blue SN550 NVMe SSD (DRAM-less) [15b7:5009] (rev 01) lspci -v -d 15b7:5009 03:00.0 Non-Volatile memory controller: Sandisk Corp SanDisk Ultra 3D / WD PC SN530, IX SN530, Blue SN550 NVMe SSD (DRAM-less) (rev 01) (prog-if 02 [NVM Express]) Subsystem: Sandisk Corp WD Blue SN550 NVMe SSD Flags: bus master, fast devsel, latency 0, IRQ 35, IOMMU group 10 Memory at fe800000 (64-bit, non-prefetchable) [size=16K] Memory at fe804000 (64-bit, non-prefetchable) [size=256] Capabilities: [80] Power Management version 3 Capabilities: [90] MSI: Enable- Count=1/32 Maskable- 64bit+ Capabilities: [b0] MSI-X: Enable+ Count=17 Masked- Capabilities: [c0] Express Endpoint, MSI 00 Capabilities: [100] Advanced Error Reporting Capabilities: [150] Device Serial Number 00-00-00-00-00-00-00-00 Capabilities: [1b8] Latency Tolerance Reporting Capabilities: [300] Secondary PCI Express Capabilities: [900] L1 PM Substates Kernel driver in use: nvme dmesg | grep nvme [ 0.000000] Command line: BOOT_IMAGE=/vmlinuz-6.12.20-amd64-desktop-rolling root=UUID= ro splash quiet nvme_core.default_ps_max_latency_us=0 DEEPIN_GFXMODE= [ 0.059301] Kernel command line: BOOT_IMAGE=/vmlinuz-6.12.20-amd64-desktop-rolling root=UUID= ro splash quiet nvme_core.default_ps_max_latency_us=0 DEEPIN_GFXMODE= [ 0.542430] nvme nvme0: pci function 0000:03:00.0 [ 0.560426] nvme nvme0: allocated 32 MiB host memory buffer. [ 0.562491] nvme nvme0: 16/0/0 default/read/poll queues [ 0.567764] nvme0n1: p1 p2 p3 p4 p5 p6 p7 p8 p9 [ 6.388726] EXT4-fs (nvme0n1p7): mounted filesystem ro with ordered data mode. Quota mode: none. [ 6.893421] EXT4-fs (nvme0n1p7): re-mounted r/w. Quota mode: none. [ 7.125419] Adding 16777212k swap on /dev/nvme0n1p8. Priority:-2 extents:1 across:16777212k SS [ 7.157588] EXT4-fs (nvme0n1p6): mounted filesystem r/w with ordered data mode. Quota mode: none. [ 7.165021] EXT4-fs (nvme0n1p9): mounted filesystem r/w with ordered data mode. Quota mode: none. [ 8.036932] nvme nvme0: using unchecked data buffer [ 8.096023] block nvme0n1: No UUID available providing old NGUID Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=d5887dc6b6c054d0da3cd053afc15b7be1f45ff6 Link: https://lore.kernel.org/all/20240422162822.3539156-1-sean.anderson@linux.dev/ Reported-by: liaozw <hedgehog-002@163.com> Closes: https://bbs.deepin.org.cn/post/286300 Reported-by: rugk <rugk+github@posteo.de> Closes: https://bugzilla.kernel.org/show_bug.cgi?id=208123 Signed-off-by: Wentao Guan <guanwentao@uniontech.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2025-04-24 10:40:10 +08:00
nvme-pci: add NVME_QUIRK_BOGUS_NID for ADATA XPG SX6000LNP (AKA SPECTRIX S40G) ADATA XPG SPECTRIX S40G drives report bogus eui64 values that appear to be the same across drives in one system. Quirk them out so they are not marked as "non globally unique" duplicates. Before: [ 2.258919] nvme nvme1: pci function 0000:06:00.0 [ 2.264898] nvme nvme2: pci function 0000:05:00.0 [ 2.323235] nvme nvme1: failed to set APST feature (2) [ 2.326153] nvme nvme2: failed to set APST feature (2) [ 2.333935] nvme nvme1: allocated 64 MiB host memory buffer. [ 2.336492] nvme nvme2: allocated 64 MiB host memory buffer. [ 2.339611] nvme nvme1: 7/0/0 default/read/poll queues [ 2.341805] nvme nvme2: 7/0/0 default/read/poll queues [ 2.346114] nvme1n1: p1 [ 2.347197] nvme nvme2: globally duplicate IDs for nsid 1 After: [ 2.427715] nvme nvme1: pci function 0000:06:00.0 [ 2.427771] nvme nvme2: pci function 0000:05:00.0 [ 2.488154] nvme nvme2: failed to set APST feature (2) [ 2.489895] nvme nvme1: failed to set APST feature (2) [ 2.498773] nvme nvme2: allocated 64 MiB host memory buffer. [ 2.500587] nvme nvme1: allocated 64 MiB host memory buffer. [ 2.504113] nvme nvme2: 7/0/0 default/read/poll queues [ 2.507026] nvme nvme1: 7/0/0 default/read/poll queues [ 2.509467] nvme nvme2: Ignoring bogus Namespace Identifiers [ 2.512804] nvme nvme1: Ignoring bogus Namespace Identifiers [ 2.513698] nvme1n1: p1 Signed-off-by: Pablo Greco <pgreco@centosproject.org> Reviewed-by: Keith Busch <kbusch@kernel.org> Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com> Cc: <stable@vger.kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2022-06-25 09:15:02 -03:00
nvme-pci: add the DISABLE_WRITE_ZEROES quirk for a SPCC device This adds a quirk for SPCC 256GB NVMe 1.3 drive which fixes timeouts and I/O errors due to the fact that the controller does not properly handle the Write Zeroes command: [ 2745.659527] CPU: 2 PID: 0 Comm: swapper/2 Tainted: G E 5.10.6-BET #1 [ 2745.659528] Hardware name: System manufacturer System Product Name/PRIME X570-P, BIOS 3001 12/04/2020 [ 2776.138874] nvme nvme1: I/O 414 QID 3 timeout, aborting [ 2776.138886] nvme nvme1: I/O 415 QID 3 timeout, aborting [ 2776.138891] nvme nvme1: I/O 416 QID 3 timeout, aborting [ 2776.138895] nvme nvme1: I/O 417 QID 3 timeout, aborting [ 2776.138912] nvme nvme1: Abort status: 0x0 [ 2776.138921] nvme nvme1: I/O 428 QID 3 timeout, aborting [ 2776.138922] nvme nvme1: Abort status: 0x0 [ 2776.138925] nvme nvme1: Abort status: 0x0 [ 2776.138974] nvme nvme1: Abort status: 0x0 [ 2776.138977] nvme nvme1: Abort status: 0x0 [ 2806.346792] nvme nvme1: I/O 414 QID 3 timeout, reset controller [ 2806.363566] nvme nvme1: 15/0/0 default/read/poll queues [ 2836.554298] nvme nvme1: I/O 415 QID 3 timeout, disable controller [ 2836.672064] blk_update_request: I/O error, dev nvme1n1, sector 16350 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672072] blk_update_request: I/O error, dev nvme1n1, sector 16093 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672074] blk_update_request: I/O error, dev nvme1n1, sector 15836 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672076] blk_update_request: I/O error, dev nvme1n1, sector 15579 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672078] blk_update_request: I/O error, dev nvme1n1, sector 15322 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672080] blk_update_request: I/O error, dev nvme1n1, sector 15065 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672082] blk_update_request: I/O error, dev nvme1n1, sector 14808 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672083] blk_update_request: I/O error, dev nvme1n1, sector 14551 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672085] blk_update_request: I/O error, dev nvme1n1, sector 14294 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672087] blk_update_request: I/O error, dev nvme1n1, sector 14037 op 0x9:(WRITE_ZEROES) flags 0x0 phys_seg 0 prio class 0 [ 2836.672121] nvme nvme1: failed to mark controller live state [ 2836.672123] nvme nvme1: Removing after probe failure status: -19 [ 2836.689016] Aborting journal on device dm-0-8. [ 2836.689024] Buffer I/O error on dev dm-0, logical block 25198592, lost sync page write [ 2836.689027] JBD2: Error -5 detected when updating journal superblock for dm-0-8. Reported-by: Bradley Chapman <chapman6235@comcast.net> Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com> Tested-by: Bradley Chapman <chapman6235@comcast.net> Reviewed-by: Keith Busch <kbusch@kernel.org> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-01-25 21:19:16 -08:00
nvme-pci: avoid the deepest sleep state on Kingston A2000 SSDs Some Kingston A2000 NVMe SSDs sooner or later get confused and stop working when they use the deepest APST sleep while running Linux. The system then crashes and one has to cold boot it to get the SSD working again. Kingston seems to known about this since at least mid-September 2020: https://bbs.archlinux.org/viewtopic.php?pid=1926994#p1926994 Someone working for a German company representing Kingston to the German press confirmed to me Kingston engineering is aware of the issue and investigating; the person stated that to their current knowledge only the deepest APST sleep state causes trouble. Therefore, make Linux avoid it for now by applying the NVME_QUIRK_NO_DEEPEST_PS to this SSD. I have two such SSDs, but it seems the problem doesn't occur with them. I hence couldn't verify if this patch really fixes the problem, but all the data in front of me suggests it should. This patch can easily be reverted or improved upon if a better solution surfaces. FWIW, there are many reports about the issue scattered around the web; most of the users disabled APST completely to make things work, some just made Linux avoid the deepest sleep state: https://bugzilla.kernel.org/show_bug.cgi?id=195039#c65 https://bugzilla.kernel.org/show_bug.cgi?id=195039#c73 https://bugzilla.kernel.org/show_bug.cgi?id=195039#c74 https://bugzilla.kernel.org/show_bug.cgi?id=195039#c78 https://bugzilla.kernel.org/show_bug.cgi?id=195039#c79 https://bugzilla.kernel.org/show_bug.cgi?id=195039#c80 https://askubuntu.com/questions/1222049/nvmekingston-a2000-sometimes-stops-giving-response-in-ubuntu-18-04dell-inspir https://community.acer.com/en/discussion/604326/m-2-nvme-ssd-aspire-517-51g-issue-compatibility-kingston-a2000-linux-ubuntu For the record, some data from 'nvme id-ctrl /dev/nvme0' NVME Identify Controller: vid : 0x2646 ssvid : 0x2646 mn : KINGSTON SA2000M81000G fr : S5Z42105 [...] ps 0 : mp:9.00W operational enlat:0 exlat:0 rrt:0 rrl:0 rwt:0 rwl:0 idle_power:- active_power:- ps 1 : mp:4.60W operational enlat:0 exlat:0 rrt:1 rrl:1 rwt:1 rwl:1 idle_power:- active_power:- ps 2 : mp:3.80W operational enlat:0 exlat:0 rrt:2 rrl:2 rwt:2 rwl:2 idle_power:- active_power:- ps 3 : mp:0.0450W non-operational enlat:2000 exlat:2000 rrt:3 rrl:3 rwt:3 rwl:3 idle_power:- active_power:- ps 4 : mp:0.0040W non-operational enlat:15000 exlat:15000 rrt:4 rrl:4 rwt:4 rwl:4 idle_power:- active_power:- Cc: stable@vger.kernel.org # 4.14+ Signed-off-by: Thorsten Leemhuis <linux@leemhuis.info> Signed-off-by: Christoph Hellwig <hch@lst.de>
2021-01-29 06:24:42 +01:00
nvme-pci: quirk Acer FA100 for non-uniqueue identifiers In order for two Acer FA100 SSDs to work in one PC (in the case of myself, a Lenovo Legion T5 28IMB05), and not show one drive and not the other, and sometimes mix up what drive shows up (randomly), these two lines of code need to be added, and then both of the SSDs will show up and not conflict when booting off of one of them. If you boot up your computer with both SSDs installed without this patch, you may also randomly get into a kernel panic (if the initrd is not set up) or stuck in the initrd "/init" process, it is set up, however, if you do apply this patch, there should not be problems with booting or seeing both contents of the drive. Tested with the btrfs filesystem with a RAID configuration of having the root drive '/' combined to make two 256GB Acer FA100 SSDs become 512GB in total storage. Kernel Logs with patch applied (`dmesg -t | grep -i nvm`): ``` ... nvme 0000:04:00.0: platform quirk: setting simple suspend nvme nvme0: pci function 0000:04:00.0 nvme 0000:05:00.0: platform quirk: setting simple suspend nvme nvme1: pci function 0000:05:00.0 nvme nvme1: missing or invalid SUBNQN field. nvme nvme1: allocated 64 MiB host memory buffer. nvme nvme0: missing or invalid SUBNQN field. nvme nvme0: allocated 64 MiB host memory buffer. nvme nvme1: 8/0/0 default/read/poll queues nvme nvme1: Ignoring bogus Namespace Identifiers nvme nvme0: 8/0/0 default/read/poll queues nvme nvme0: Ignoring bogus Namespace Identifiers nvme0n1: p1 p2 ... ``` Kernel Logs with patch not applied (`dmesg -t | grep -i nvm`): ``` ... nvme 0000:04:00.0: platform quirk: setting simple suspend nvme nvme0: pci function 0000:04:00.0 nvme 0000:05:00.0: platform quirk: setting simple suspend nvme nvme1: pci function 0000:05:00.0 nvme nvme0: missing or invalid SUBNQN field. nvme nvme1: missing or invalid SUBNQN field. nvme nvme0: allocated 64 MiB host memory buffer. nvme nvme1: allocated 64 MiB host memory buffer. nvme nvme0: 8/0/0 default/read/poll queues nvme nvme1: 8/0/0 default/read/poll queues nvme nvme1: globally duplicate IDs for nsid 1 nvme nvme1: VID:DID 1dbe:5216 model:Acer SSD FA100 256GB firmware:1.Z.J.2X nvme0n1: p1 p2 ... ``` Signed-off-by: Christopher Lentocha <christopherericlentocha@gmail.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Keith Busch <kbusch@kernel.org>
2025-02-18 08:59:29 -05:00
nvme-pci: mark Lexar NM760 as IGNORE_DEV_SUBNQN A system with more than one of these SSDs will only have one usable. The kernel fails to detect more than one nvme device due to duplicate cntlids. before: [ 9.395229] nvme 0000:01:00.0: platform quirk: setting simple suspend [ 9.395262] nvme nvme0: pci function 0000:01:00.0 [ 9.395282] nvme 0000:03:00.0: platform quirk: setting simple suspend [ 9.395305] nvme nvme1: pci function 0000:03:00.0 [ 9.409873] nvme nvme0: Duplicate cntlid 1 with nvme1, subsys nqn.2022-07.com.siliconmotion:nvm-subsystem-sn- , rejecting [ 9.409982] nvme nvme0: Removing after probe failure status: -22 [ 9.427487] nvme nvme1: allocated 64 MiB host memory buffer. [ 9.445088] nvme nvme1: 16/0/0 default/read/poll queues [ 9.449898] nvme nvme1: Ignoring bogus Namespace Identifiers after: [ 1.161890] nvme 0000:01:00.0: platform quirk: setting simple suspend [ 1.162660] nvme nvme0: pci function 0000:01:00.0 [ 1.162684] nvme 0000:03:00.0: platform quirk: setting simple suspend [ 1.162707] nvme nvme1: pci function 0000:03:00.0 [ 1.191354] nvme nvme0: allocated 64 MiB host memory buffer. [ 1.193378] nvme nvme1: allocated 64 MiB host memory buffer. [ 1.211044] nvme nvme1: 16/0/0 default/read/poll queues [ 1.211080] nvme nvme0: 16/0/0 default/read/poll queues [ 1.216145] nvme nvme0: Ignoring bogus Namespace Identifiers [ 1.216261] nvme nvme1: Ignoring bogus Namespace Identifiers Adding the NVME_QUIRK_IGNORE_DEV_SUBNQN quirk to resolves the issue. Signed-off-by: Juraj Pecigos <kernel@juraj.dev> Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com> Signed-off-by: Christoph Hellwig <hch@lst.de>
2023-03-26 11:29:49 +02:00
nvme-pci: Simplify interrupt allocation The NVME PCI driver contains a tedious mechanism for interrupt allocation, which is necessary to adjust the number and size of interrupt sets to the maximum available number of interrupts which depends on the underlying PCI capabilities and the available CPU resources. It works around the former short comings of the PCI and core interrupt allocation mechanims in combination with interrupt sets. The PCI interrupt allocation function allows to provide a maximum and a minimum number of interrupts to be allocated and tries to allocate as many as possible. This worked without driver interaction as long as there was only a single set of interrupts to handle. With the addition of support for multiple interrupt sets in the generic affinity spreading logic, which is invoked from the PCI interrupt allocation, the adaptive loop in the PCI interrupt allocation did not work for multiple interrupt sets. The reason is that depending on the total number of interrupts which the PCI allocation adaptive loop tries to allocate in each step, the number and the size of the interrupt sets need to be adapted as well. Due to the way the interrupt sets support was implemented there was no way for the PCI interrupt allocation code or the core affinity spreading mechanism to invoke a driver specific function for adapting the interrupt sets configuration. As a consequence the driver had to implement another adaptive loop around the PCI interrupt allocation function and calling that with maximum and minimum interrupts set to the same value. This ensured that the allocation either succeeded or immediately failed without any attempt to adjust the number of interrupts in the PCI code. The core code now allows drivers to provide a callback to recalculate the number and the size of interrupt sets during PCI interrupt allocation, which in turn allows the PCI interrupt allocation function to be called in the same way as with a single set of interrupts. The PCI code handles the adaptive loop and the interrupt affinity spreading mechanism invokes the driver callback to adapt the interrupt set configuration to the current loop value. This replaces the adaptive loop in the driver completely. Implement the NVME specific callback which adjusts the interrupt sets configuration and remove the adaptive allocation loop. [ tglx: Simplify the callback further and restore the dropped adjustment of number of sets ] Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Marc Zyngier <marc.zyngier@arm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Bjorn Helgaas <helgaas@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: Sagi Grimberg <sagi@grimberg.me> Cc: linux-nvme@lists.infradead.org Cc: linux-pci@vger.kernel.org Cc: Keith Busch <keith.busch@intel.com> Cc: Sumit Saxena <sumit.saxena@broadcom.com> Cc: Kashyap Desai <kashyap.desai@broadcom.com> Cc: Shivasharan Srikanteshwara <shivasharan.srikanteshwara@broadcom.com> Link: https://lkml.kernel.org/r/20190216172228.602546658@linutronix.de
2019-02-16 18:13:10 +01:00