Allow for accounting N pages at once instead of one page at a time.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: David Howells <dhowells@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
When batching events (such as writing back N pages in a single I/O), it
is better to do one flex_proportion operation instead of N. There is
only one caller of __fprop_inc_percpu_max(), and it's the one we're
going to change in the next patch, so rename it instead of adding a
compatibility wrapper.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Make this look like the newly renamed vmstat functions.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Howells <dhowells@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
This is the folio equivalent of migrate_page_copy(), which is retained
as a wrapper for filesystems which are not yet converted to folios.
Also convert copy_huge_page() to folio_copy().
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Turn migrate_page_states() into a wrapper around folio_migrate_flags().
Also convert two functions only called from folio_migrate_flags() to
be folio-based. ksm_migrate_page() becomes folio_migrate_ksm() and
copy_page_owner() becomes folio_copy_owner(). folio_migrate_flags()
alone shrinks by two thirds -- 1967 bytes down to 642 bytes.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: David Howells <dhowells@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Reimplement migrate_page_move_mapping() as a wrapper around
folio_migrate_mapping(). Saves 193 bytes of kernel text.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Howells <dhowells@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Transform page_mkclean() into folio_mkclean() and add a page_mkclean()
wrapper around folio_mkclean().
folio_mkclean is 15 bytes smaller than page_mkclean, but the kernel
is enlarged by 33 bytes due to inlining page_folio() into each caller.
This will go away once the callers are converted to use folio_mkclean().
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Howells <dhowells@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Convert mark_page_accessed() to folio_mark_accessed(). It already
operated on the entire compound page, but now we can avoid calling
compound_head quite so many times. Shrinks the function from 424 bytes
to 295 bytes (shrinking by 129 bytes). The compatibility wrapper is 30
bytes, plus the 8 bytes for the exported symbol means the kernel shrinks
by 91 bytes.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: David Howells <dhowells@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
This replaces activate_page() and eliminates lots of calls to
compound_head(). Saves net 118 bytes of kernel text. There are still
some redundant calls to page_folio() here which will be removed when
pagevec_lru_move_fn() is converted to use folios.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Howells <dhowells@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Idle page tracking is handled through page_ext on 32-bit architectures.
Add folio equivalents for 32-bit and move all the page compatibility
parts to common code.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: William Kucharski <william.kucharski@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: David Howells <dhowells@redhat.com>
As a default implementation, call arch_make_page_accessible n times.
If an architecture can do better, it can override this.
Also move the default implementation of arch_make_page_accessible()
from gfp.h to mm.h.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: David Howells <dhowells@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
This allows us to map a portion of a folio. Callers can only expect
to access up to the next page boundary.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
This is a default implementation which calls flush_dcache_page() on
each page in the folio. If architectures can do better, they should
implement their own version of it.
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
We've had an x86-specific SG-buffer handling code, but now it can be
merged gracefully with the standard non-contiguous DMA pages.
After the migration, SNDRV_DMA_TYPE_DMA_SG becomes identical with
SNDRV_DMA_TYPE_NONCONTIG on x86, while others still fall back to
SNDRV_DMA_TYPE_DEV.
The remaining problem is about the SG-buffer with WC pages: the DMA
core stuff on x86 doesn't treat it well, so we still need some special
handling to manipulate the page attribute manually. The mmap handler
for SNDRV_DMA_TYPE_DEV_SG_WC still returns -ENOENT intentionally for
the fallback to the default handler.
Link: https://lore.kernel.org/r/20211017074859.24112-4-tiwai@suse.de
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Following to the addition of non-contiguous pages, this patch adds the
new contiguous non-coherent page allocation to the standard memalloc
helper. Like the previous non-contig type, this non-coherent type is
also directional and requires the explicit sync, too. Hence the
driver using this type of buffer may need to set
SNDRV_PCM_INFO_EXPLICIT_SYNC flag to the PCM hardware.info as well,
unless it's set up in the managed mode.
Link: https://lore.kernel.org/r/20211017074859.24112-3-tiwai@suse.de
Signed-off-by: Takashi Iwai <tiwai@suse.de>
This patch adds the support for allocation of non-contiguous DMA pages
in the common memalloc helper. It's another SG-buffer type, but
unlike the existing one, this is directional and requires the explicit
sync / invalidation of dirty pages on non-coherent architectures.
For this enhancement, the following points are changed:
- snd_dma_device stores the DMA direction.
- snd_dma_device stores need_sync flag indicating whether the explicit
sync is required or not.
- A new variant of helper functions, snd_dma_alloc_dir_pages() and
*_all() are introduced; the old snd_dma_alloc_pages() and *_all()
kept as just wrappers with DMA_BIDIRECTIONAL.
- A new helper snd_dma_buffer_sync() is introduced; this gets called
in the appropriate places.
- A new allocation type, SNDRV_DMA_TYPE_NONCONTIG, is introduced.
When the driver allocates pages with this new type, and it may require
the SNDRV_PCM_INFO_EXPLICIT_SYNC flag set to the PCM hardware.info for
taking the full control of PCM applptr and hwptr changes (that implies
disabling the mmap of control/status data). When the buffer
allocation is managed by snd_pcm_set_managed_buffer(), this flag is
automatically set depending on the result of dma_need_sync()
internally. Otherwise, if the buffer is managed manually, the driver
has to set the flag explicitly, too.
The explicit sync between CPU and device for non-coherent memory is
performed at the points before and after read/write transfer as well
as the applptr/hwptr syncptr ioctl. In the case of mmap mode,
user-space is supposed to call the syncptr ioctl with the hwptr flag
to update and fetch the status at first; this corresponds to CPU-sync.
Then user-space advances the applptr via syncptr ioctl again with
applptr flag, and this corresponds to the device sync with flushing.
Other than the DMA direction and the explicit sync, the usage of this
new buffer type is almost equivalent with the existing
SNDRV_DMA_TYPE_DEV_SG; you can get the page and the address via
snd_sgbuf_get_page() and snd_sgbuf_get_addr(), also calculate the
continuous pages via snd_sgbuf_get_chunk_size().
For those SG-page handling, the non-contig type shares the same ops
with the vmalloc handler. As we do always vmap the SG pages at first,
the actual address can be deduced from the vmapped address easily
without iterating the SG-list.
Link: https://lore.kernel.org/r/20211017074859.24112-2-tiwai@suse.de
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Mapping something twice should be possible as long as,
DMA_ATTR_SKIP_CPU_SYNC is passed to the strictly speaking second relevant
mapping operation (that attempts to map the same thing). So, don't issue a
warning if the specified condition is met in add_dma_entry().
Signed-off-by: Hamza Mahfooz <someguy@effective-light.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
The __domain_mapping() always removes the pages in the range from
'iov_pfn' to 'end_pfn', but the 'end_pfn' is always the last pfn
of the range that the caller wants to map.
This would introduce too many duplicated removing and leads the
map operation take too long, for example:
Map iova=0x100000,nr_pages=0x7d61800
iov_pfn: 0x100000, end_pfn: 0x7e617ff
iov_pfn: 0x140000, end_pfn: 0x7e617ff
iov_pfn: 0x180000, end_pfn: 0x7e617ff
iov_pfn: 0x1c0000, end_pfn: 0x7e617ff
iov_pfn: 0x200000, end_pfn: 0x7e617ff
...
it takes about 50ms in total.
We can reduce the cost by recalculate the 'end_pfn' and limit it
to the boundary of the end of this pte page.
Map iova=0x100000,nr_pages=0x7d61800
iov_pfn: 0x100000, end_pfn: 0x13ffff
iov_pfn: 0x140000, end_pfn: 0x17ffff
iov_pfn: 0x180000, end_pfn: 0x1bffff
iov_pfn: 0x1c0000, end_pfn: 0x1fffff
iov_pfn: 0x200000, end_pfn: 0x23ffff
...
it only need 9ms now.
This also removes a meaningless BUG_ON() in __domain_mapping().
Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
Tested-by: Liujunjie <liujunjie23@huawei.com>
Link: https://lore.kernel.org/r/20211008000433.1115-1-longpeng2@huawei.com
Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
Link: https://lore.kernel.org/r/20211014053839.727419-10-baolu.lu@linux.intel.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>
update_pasid() and its call chain are currently unused in the tree because
Thomas disabled the ENQCMD feature. The feature will be re-enabled shortly
using a different approach and update_pasid() and its call chain will not
be used in the new approach.
Remove the useless functions.
Signed-off-by: Fenghua Yu <fenghua.yu@intel.com>
Reviewed-by: Tony Luck <tony.luck@intel.com>
Link: https://lore.kernel.org/r/20210920192349.2602141-1-fenghua.yu@intel.com
Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
Link: https://lore.kernel.org/r/20211014053839.727419-8-baolu.lu@linux.intel.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>
The IOMMU VT-d implementation uses the first level for GPA->HPA translation
by default. Although both the first level and the second level could handle
the DMA translation, they're different in some way. For example, the second
level translation has separate controls for the Access/Dirty page tracking.
With the first level translation, there's no such control. On the other
hand, the second level translation has the page-level control for forcing
snoop, but the first level only has global control with pasid granularity.
This uses the second level for GPA->HPA translation so that we can provide
a consistent hardware interface for use cases like dirty page tracking for
live migration.
Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Link: https://lore.kernel.org/r/20210926114535.923263-1-baolu.lu@linux.intel.com
Link: https://lore.kernel.org/r/20211014053839.727419-6-baolu.lu@linux.intel.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>
An iommu domain could be allocated and mapped before it's attached to any
device. This requires that in scalable mode, when the domain is allocated,
the format (FL or SL) of the page table must be determined. In order to
achieve this, the platform should support consistent SL or FL capabilities
on all IOMMU's. This adds a check for this and aborts IOMMU probing if it
doesn't meet this requirement.
Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Link: https://lore.kernel.org/r/20210926114535.923263-1-baolu.lu@linux.intel.com
Link: https://lore.kernel.org/r/20211014053839.727419-5-baolu.lu@linux.intel.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>
When the dmar translation fault happens, the kernel prints a single line
fault reason with corresponding hexadecimal code defined in the Intel VT-d
specification.
Currently, when user wants to debug the translation fault in detail,
debugfs is used for dumping the dmar_translation_struct, which is not
available when the kernel failed to boot.
Dump the DMAR translation structure, pagewalk the IO page table and print
the page table entry when the fault happens.
This takes effect only when CONFIG_DMAR_DEBUG is enabled.
Signed-off-by: Kyung Min Park <kyung.min.park@intel.com>
Link: https://lore.kernel.org/r/20210815203845.31287-1-kyung.min.park@intel.com
Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
Link: https://lore.kernel.org/r/20211014053839.727419-3-baolu.lu@linux.intel.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Handling of intel_iommu kernel command line option should return "true" to
indicate option is valid and so avoid logging it as unknown by the core
parsing code.
Also log unknown sub-options at the notice level to let user know of
potential typos or similar.
Reported-by: Eero Tamminen <eero.t.tamminen@intel.com>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://lore.kernel.org/r/20210831112947.310080-1-tvrtko.ursulin@linux.intel.com
Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
Link: https://lore.kernel.org/r/20211014053839.727419-2-baolu.lu@linux.intel.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Replace the unconditional clflush() with drm_clflush_virt_range()
which does the wbinvd() fallback when clflush is not available.
This time no justification is given for the clflush in the
offending commit.
Cc: stable@vger.kernel.org
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Fixes: 2c8ab3339e ("drm/i915: Pin timeline map after first timeline pin, v4.")
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211014090941.12159-4-ville.syrjala@linux.intel.com
Reviewed-by: Dave Airlie <airlied@redhat.com>
This one is apparently a "clflush for good measure", so bit more
justification (if you can call it that) than some of the others.
Convert to drm_clflush_virt_range() again so that machines without
clflush will survive the ordeal.
Cc: stable@vger.kernel.org
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Thomas Hellström <thomas.hellstrom@intel.com> #v1
Fixes: 12ca695d2c ("drm/i915: Do not share hwsp across contexts any more, v8.")
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211014090941.12159-3-ville.syrjala@linux.intel.com
Reviewed-by: Dave Airlie <airlied@redhat.com>
Not all machines have clflush, so don't go assuming they do.
Not really sure why the clflush is even here since hwsp
is supposed to get snooped I thought.
Although in my case we're talking about a i830 machine where
render/blitter snooping is definitely busted. But it might
work for the hswp perhaps. Haven't really reverse engineered
that one fully.
Cc: stable@vger.kernel.org
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Fixes: b436a5f8b6 ("drm/i915/gt: Track all timelines created using the HWSP")
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211014090941.12159-2-ville.syrjala@linux.intel.com
Reviewed-by: Dave Airlie <airlied@redhat.com>
Adds support for using FF-A [1] as transport to the OP-TEE driver.
Introduces struct optee_msg_param_fmem which carries all information
needed when OP-TEE is calling FFA_MEM_RETRIEVE_REQ to get the shared
memory reference mapped by the hypervisor in S-EL2. Register usage is
also updated to include the information needed.
The FF-A part of this driver is enabled if CONFIG_ARM_FFA_TRANSPORT is
enabled.
[1] https://developer.arm.com/documentation/den0077/latest
Acked-by: Sumit Garg <sumit.garg@linaro.org>
Signed-off-by: Jens Wiklander <jens.wiklander@linaro.org>
Isolate the ABI based on raw SMCs. Code specific to the raw SMC ABI is
moved into smc_abi.c. This makes room for other ABIs with a clear
separation.
The driver changes to use module_init()/module_exit() instead of
module_platform_driver(). The platform_driver_register() and
platform_driver_unregister() functions called directly to keep the same
behavior. This is needed because module_platform_driver() is based on
module_driver() which can only be used once in a module.
A function optee_rpc_cmd() is factored out from the function
handle_rpc_func_cmd() to handle the ABI independent part of RPC
processing.
This patch is not supposed to change the driver behavior, it's only a
matter of reorganizing the code.
Reviewed-by: Sumit Garg <sumit.garg@linaro.org>
Signed-off-by: Jens Wiklander <jens.wiklander@linaro.org>
To avoid checking if fence exists multipled times, changed fence
handling to depend only on the fence status field:
Busy, which means CS still did not completed :
Add its QID so multi CS wait on its completion.
Finished, which means CS completed and fence exists:
Raise its completion bit if it finished mcs handling and
update if necessary the earliest timestamp.
Gone, which means CS already completed and fence deleted:
Update multi CS data to ignore timestamp and raise its
completion bit.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reported-by: kernel test robot <lkp@intel.com>
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
No need to check the return value if the following action is the same for
both cases. In addition, now that hl_ctx_free() doesn't print if the
context is not released, its name can be misleading as the context might
stay alive after it is executed with no indication for that.
Hence we can discard it and simply put the refcount.
Signed-off-by: Omer Shpigelman <oshpigelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Remove the flag that determines whether to take a timestamp once the
interrupt arrives.
Instead, always take the timestamp once per interrupt.
This is a must for the user-space to measure its graph operations
to evaluate the graph computation time.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
When adding a new node to the hpriv list, the driver should
initialize its fields before adding the new node.
Otherwise, there may be some small chance of another thread traversing
that list and accessing the new node's fields without them being
initialized.
Signed-off-by: Moti Haimovski <mhaimovski@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Make the frequency set/get functionality common to all ASICs.
This makes more code reusable when adding support for newer ASICs.
Signed-off-by: Rajaravi Krishna Katta <rkatta@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Fix the following build/link error by adding a dependency on the CRC32
routines:
ld: drivers/misc/habanalabs/common/firmware_if.o: in function `hl_fw_dynamic_request_descriptor':
firmware_if.c:(.text.unlikely+0xc89): undefined reference to `crc32_le'
Fixes: 8a43c83fec ("habanalabs: load boot fit to device")
Signed-off-by: Vegard Nossum <vegard.nossum@oracle.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Implement the calls to the dma-buf kernel api to create a dma-buf
object backed by FD.
We block the option to mmap the DMA-BUF object because we don't support
DIRECT_IO and implicit P2P. We only implement support for explicit P2P
through importing the FD of the DMA-BUF.
In the export phase, we provide to the DMA-BUF object an array of pages
that represent the device's memory area. During the map callback,
we convert the array of pages into an SGT. We split/merge the pages
according to the dma max segment size of the importer.
To get the DMA address of the PCI bar, we use the dma_map_resources()
kernel API, because our device memory is not backed by page struct
and this API doesn't need page struct to map the physical address to
a DMA address.
We set the orig_nents member of the SGT to be 0, to indicate to other
drivers that we don't support CPU mappings.
Note that in Habanalabs's ASICs, the device memory is pinned and
immutable. Therefore, there is no need for dynamic mappings and pinning
callbacks.
Also note that in GAUDI we don't have an MMU towards the device memory
and the user works on physical addresses. Therefore, the user doesn't
pass through the kernel driver to allocate memory there. As a result,
only for GAUDI we receive from the user a device memory physical address
(instead of a handle) and a size.
We check the p2p distance using pci_p2pdma_distance_many() and refusing
to map dmabuf in case the distance doesn't allow p2p.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Gal Pressman <galpress@amazon.com>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
User process might want to share the device memory with another
driver/device, and to allow it to access it over PCIe (P2P).
To enable this, we utilize the dma-buf mechanism and add a dma-buf
exporter support, so the other driver can import the device memory and
access it.
The device memory is allocated using our existing allocation uAPI,
where the user will get a handle that represents the allocation.
The user will then need to call the new
uAPI (HL_MEM_OP_EXPORT_DMABUF_FD) and give the handle as a parameter.
The driver will return a FD that represents the DMA-BUF object that
was created to match that allocation.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
When polling fences for multi CS, it is possible that fence is
no longer exists (its corresponding CS completed and the fence was
deleted) but we still accessing its parameters, causing NULL pointer
dereference.
Fixed by checking if fence exits before accessing its parameters.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Race condition occurs when CS fence completes and multi CS did not
completed yet, while waiting for multi CS ends and returns indication
to user that the CS completed. Next wait for multi CS may be triggered
by previous multi CS completion without any current CS completed,
causing an error.
Example scenario :
1. User do multi CS wait for CSs 1 and 2 on master QID 0
2. CS 1 and 2 reached the "cs release" code. The thread of CS 1
completed both the CS and multi CS handling but the completion
thread of CS 2 completed the CS but still did not executed
complete_multi_cs (note that in CS completion the sequence is to
first do complete all for the CS and then another complete all to
signal the multi_cs)
3. User received indication that CS 1 and 2 completed (since we check
the CS fence and both indicated as completed) and immediately waits
on CS 3 and 4, also on master QID 0.
4. Completion thread of CS2 executed complete_multi_cs before
completion of CS 3 and 4 and so will trigger the multi CS wait of
CSs 3 and 4 as they wait on master QID 0.
This will trigger multi CS completion although none of its
current CS has been completed.
Fixed by adding multi CS complete handling indication for each CS.
CS will be marked to the user as completed only if its fence completed
and multi CS handling is done.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
There may be a situation where drivers receives continuous fatal H/W
error events from FW immediately post reset cycle.
This may be due to some fault on the silicon itself.
In such case its better to bypass reset cycle so we won't be stuck in
endless loop of resets.
This commit bypasses reset request in case driver received two back to
back FW fatal error before first occurrence of heartbeat event.
Signed-off-by: Bharat Jauhari <bjauhari@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Taking an accurate timestamp in a close proximity of the interrupt is
required for user side statistics management.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
The driver allows only a single process to open a device's FD at any
single time. This is done by checking "hdev->compute_ctx" under mutex.
Therefore, to prevent a race between the moment a user closes it's FD
and when another user tries to open the device, we need to make sure
that clearing this variable is the very last thing that is done in the
code of the FD's release.
I'm moving the idle check before clearing this variable and the
"reset on device release". btw, if the reset happens it will prevent
any other user from opening the device until the reset is finished.
An important thing to note is that we need to remove the user process
that is closing the device from the process list BEFORE calling the
reset function. That is to prevent a case where the reset code will
try to kill that user process and it is unnecessary as the process
doesn't hold any device/driver resources anymore.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reset to the device is not necessarily due to an error, so print it
as info instead of error.
In addition, print the type of reset we are doing:
- reset of the entire device (aka hard reset)
- reset of the device after user have released it (less than hard reset)
- lighter reset of an inference device (aka soft reset)
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Soft-reset is the procedure where we reset only the compute/DMA engines
of the device, without requiring the current user-space process to
release the device.
This type of reset can happen if TDR event occurred (a workload got
stuck) or by a root request through sysfs.
This is only relevant for inference ASICs, as there is no real-world
use-case to do that in training, because training runs on multiple
devices.
In addition, we also do (in certain ASICs) a reset upon device release.
That reset uses the same code as the soft-reset.
Therefore, to better differentiate between the two resets, it is better
to rename the soft-reset support as "inference soft-reset", to make
the code more self-explanatory.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
