- Return vga_get_uninterruptible() back to userspace in the
/dev/vga_arbiter path so user can tell whether VGA routing was updated
(Simon Richter)
- Make pci_set_vga_state() fail if bridge doesn't support VGA routing,
i.e., PCI_BRIDGE_CTL_VGA is not writable, and return errors up to
vga_get() callers (Simon Richter)
* pci/vga:
PCI/VGA: Fail pci_set_vga_state() if VGA decoding not supported
PCI/VGA: Pass errors from pci_set_vga_state() up
PCI/VGA: Pass vga_get_uninterruptible() errors to userspace
- Prevent assigning space to unimplemented bridge windows; previously we
mistakenly assumed prefetchable window existed and assigned space and put
a BAR there (Ahmed Naseef)
- Avoid shrinking bridge windows to fit in the initial Root Port window;
this fixes one problem with devices with large BARs connected via
switches, e.g., Thunderbolt (Ilpo Järvinen)
- Retain information about optional resources to make assignment during
rescan more likely to succeed (Ilpo Järvinen)
- Add __resource_contains_unbound() for use in finding space for resources
with no address assigned (Ilpo Järvinen)
- Pass full extent of empty space, not just the aligned space, to
resource_alignf callback so free space before the requested alignment can
be used (Ilpo Järvinen)
- Remove unnecessary second alignment from ARM, m68k, MIPS (Ilpo Järvinen)
- Place small resources before larger ones for better utilization of
address space (Ilpo Järvinen)
- Fix alignment calculation for resource size larger than align, e.g.,
bridge windows larger than the 1MB required alignment (Ilpo Järvinen)
* pci/resource:
PCI: Fix alignment calculation for resource size larger than align
PCI: Align head space better
PCI: Rename window_alignment() to pci_min_window_alignment()
parisc/PCI: Clean up align handling
MIPS: PCI: Remove unnecessary second application of align
m68k/PCI: Remove unnecessary second application of align
ARM/PCI: Remove unnecessary second application of align
resource: Rename 'tmp' variable to 'full_avail'
resource: Pass full extent of empty space to resource_alignf callback
resource: Add __resource_contains_unbound() for internal contains checks
PCI: Fix premature removal from realloc_head list during resource assignment
PCI: Prevent shrinking bridge window from its required size
PCI: Prevent assignment to unsupported bridge windows
- Update slot handling so all ARI functions are treated as being in the
same slot. They're all reset by Secondary Bus Reset, but previously
drivers of ARI functions that appeared to be on a non-zero device weren't
notified and fatal hardware errors could result (Keith Busch)
- Make sysfs reset_subordinate hotplug safe to avoid spurious hotplug
events (Keith Busch)
- Consolidate bus iteration across the _lock(), _unlock(), and _trylock()
functions for pci_bus and pci_slot (Ilpo Järvinen)
- Hide Secondary Bus Reset ('bus') from sysfs reset_methods if masked by
CXL because it has no effect (Vidya Sagar)
* pci/reset:
PCI/CXL: Hide SBR from reset_methods if masked by CXL
PCI: Consolidate pci_bus/slot_lock/unlock/trylock()
PCI: Make reset_subordinate hotplug safe
PCI: Allow all bus devices to use the same slot
PCI: Rename __pci_bus_reset() and __pci_slot_reset()
- Rename 'slot' driver to 'generic' since it can handle any device with
individual power control as well as slots (Neil Armstrong)
- Add UPD720201/UPD720202 USB 3.0 xHCI Host Controller .compatible so
generic pwrctrl driver can control it (Neil Armstrong)
* pci/pwrctrl:
PCI/pwrctrl: generic: Add UPD720201/UPD720202 USB 3.0 xHCI Host Controller support
PCI/pwrctrl: generic: Simplify dev_err_probe() usage
PCI/pwrctrl: generic: Rename pci-pwrctrl-slot as generic
- Leave Precision Time Measurement disabled until a driver enables it to
avoid PCIe errors (Mika Westerberg)
* pci/ptm:
PCI/PTM: Do not enable PTM automatically for Root and Switch Upstream Ports
PCI/PTM: Drop pci_enable_ptm() granularity parameter
- Allow wildcards in list of host bridges that support peer-to-peer DMA
between hierarchy domains and add all Google SoCs (Jacob Moroni)
* pci/p2pdma:
PCI/P2PDMA: Add Google SoCs to the P2P DMA host bridge list
PCI/P2PDMA: Allow wildcard Device IDs in host bridge list
- Use for_each_child_of_node_scoped() to simplify iteration over OF
children (Krzysztof Kozlowski)
- Set LED_HW_PLUGGABLE for NPEM hotplug-capable ports so LED core doesn't
complain when setting brightness fails because the endpoint is gone
(Richard Cheng)
* pci/hotplug:
PCI/NPEM: Set LED_HW_PLUGGABLE for hotplug-capable ports
PCI: rpaphp: Simplify with scoped for each OF child loop
PCI: pnv_php: Simplify with scoped for each OF child loop
- Allow TPH to be enabled for RCiEPs (George Abraham P)
- Remove the pc110pad since 486 CPU support is being removed (Dmitry
Torokhov)
- Remove no_pci_devices() since pc110pad was the last remaining user
(Heiner Kallweit)
* pci/enumeration:
PCI: Remove no_pci_devices()
Input: pc110pad - remove driver
PCI/TPH: Allow TPH enable for RCiEPs
- Hold a pci_dev reference during error recovery (Sizhe Liu)
- Initialize ratelimit info so DPC and EDR paths log AER error information
(Kuppuswamy Sathyanarayanan)
* pci/dpc:
PCI/DPC: Log AER error info for DPC/EDR uncorrectable errors
PCI/DPC: Hold pci_dev reference during error recovery
- Don't enable AtomicOps by RCiEPs since none of them need Atomic Ops and
we can't tell whether the Root Complex would support them (Gerd Bayer)
- Enable AtomicOps only if we know the Root Port supports them (Gerd Bayer)
* pci/atomics:
PCI: Update PCIe spec references for AtomicOps
PCI: Enable AtomicOps only if Root Port supports them
PCI: Do not enable AtomicOps by RCiEPs
- Clear only error bits in PCIe Device Status to avoid accidentally
clearing Emergency Power Reduction Detected (Shuai Xue)
- Check for AER errors even in devices without drivers (Lukas Wunner)
* pci/aer:
PCI/AER: Stop ruling out unbound devices as error source
PCI/AER: Clear only error bits in PCIe Device Status
All Google SoCs support peer-to-peer DMA between Root Ports, so add a
wildcard rule to the host bridge list.
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Tested-by: David Hu <xuehaohu@google.com>
Reviewed-by: Logan Gunthorpe <logang@deltatee.com>
Link: https://patch.msgid.link/20260409150123.3538444-2-jmoroni@google.com
Currently, the pci_p2pdma_whitelist array requires an exact match for both
Vendor and Device ID. Some hardware vendors support cross bridge
peer-to-peer DMA across their entire silicon lineup, so add support for
wildcard device IDs to avoid the need to continuously update this array.
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Logan Gunthorpe <logang@deltatee.com>
Link: https://patch.msgid.link/20260409150123.3538444-1-jmoroni@google.com
Palm Top PC 110 is a handheld personal computer with 80486SX CPU that was
released exclusively in Japan in September 1995.
While the kernel still supports 486 CPU it is highly unlikely that anyone
is using this device with the latest kernel.
Remove the driver.
[bhelgaas: since this was posted, "x86/cpu: Remove M486/M486SX/ELAN
support" has been queued for v7.1, so pc110pad is no longer relevant:
https://lore.kernel.org/all/20251214084710.3606385-2-mingo@kernel.org/]
Signed-off-by: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20240808172733.1194442-4-dmitry.torokhov@gmail.com
NPEM registers LED classdevs on PCI endpoint that may be behind
hotplug-capable ports. During hot-removal, led_classdev_unregister() calls
led_set_brightness(LED_OFF) which leads to a PCI config read to a
disconnected device, which fails and returns -ENODEV (topology details in
msgid.link below):
leds 0003:01:00.0:enclosure:ok: Setting an LED's brightness failed (-19)
The LED core already suppresses this for devices with LED_HW_PLUGGABLE set,
but NPEM never sets it. Add the flag since NPEM LEDs are on hot-pluggable
hardware by nature.
Fixes: 4e893545ef ("PCI/NPEM: Add Native PCIe Enclosure Management support")
Signed-off-by: Richard Cheng <icheng@nvidia.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Lukas Wunner <lukas@wunner.de>
Acked-by: Kai-Heng Feng <kaihengf@nvidia.com>
Link: https://patch.msgid.link/20260402093850.23075-1-icheng@nvidia.com
Point to the relevant sections in the most recent release 7.0 of the PCIe
spec. Text has mostly just moved around without any semantic change.
Signed-off-by: Gerd Bayer <gbayer@linux.ibm.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260330-fix_pciatops-v7-3-f601818417e8@linux.ibm.com
When inspecting the config space of a Connect-X physical function in an
s390 system after it was initialized by the mlx5_core device driver, we
found the function to be enabled to request AtomicOps despite the Root Port
lacking support for completing them:
00:00.1 Ethernet controller: Mellanox Technologies MT2894 Family [ConnectX-6 Lx]
Subsystem: Mellanox Technologies Device 0002
DevCtl2: Completion Timeout: 50us to 50ms, TimeoutDis-
AtomicOpsCtl: ReqEn+
On s390 and many virtualized guests, the Endpoint is visible but the Root
Port is not. In this case, pci_enable_atomic_ops_to_root() previously
enabled AtomicOps in the Endpoint even though it can't tell whether the
Root Port supports them as a completer.
Change pci_enable_atomic_ops_to_root() to fail if there's no Root Port or
the Root Port doesn't support AtomicOps.
Fixes: 430a23689d ("PCI: Add pci_enable_atomic_ops_to_root()")
Reported-by: Alexander Schmidt <alexs@linux.ibm.com>
Signed-off-by: Gerd Bayer <gbayer@linux.ibm.com>
[bhelgaas: commit log, check RP first to simplify flow]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260330-fix_pciatops-v7-2-f601818417e8@linux.ibm.com
Since Root Complex Integrated Endpoints (RCiEPs) attach to a bus that has
no bridge device describing the Root Port, the capability to complete
AtomicOps requests cannot be determined with PCIe methods.
Change default of pci_enable_atomic_ops_to_root() to not enable AtomicOps
requests on RCiEPs.
As far as we know, there are no RCiEPs that need AtomicOps (see Link
below).
Signed-off-by: Gerd Bayer <gbayer@linux.ibm.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260330-fix_pciatops-v7-1-f601818417e8@linux.ibm.com
aer_print_error() skips printing if ratelimit_print[i] is not set. In the
native AER path, ratelimit_print is initialized by add_error_device()
during source device discovery, and is set to 1 for fatal errors to bypass
rate limiting since fatal errors should always be logged.
The DPC/EDR path uses the DPC-capable port as the error source and reads
its AER uncorrectable error status registers directly in
dpc_get_aer_uncorrect_severity(). Since it does not go through
add_error_device(), ratelimit_print[0] is left uninitialized and zero. As
a result, aer_print_error() silently drops all AER error messages for
DPC/EDR triggered events.
Set ratelimit_print[0] to 1 to bypass rate limiting and always print AER
logs for uncorrectable errors detected by the DPC port.
Fixes: a57f2bfb4a ("PCI/AER: Ratelimit correctable and non-fatal error logging")
Co-developed-by: Goudar Manjunath Ramanagouda <manjunath.ramanagouda.goudar@intel.com>
Signed-off-by: Goudar Manjunath Ramanagouda <manjunath.ramanagouda.goudar@intel.com>
Signed-off-by: Kuppuswamy Sathyanarayanan <sathyanarayanan.kuppuswamy@linux.intel.com>
[bhelgaas: commit log]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260318170449.2733581-1-sathyanarayanan.kuppuswamy@linux.intel.com
When searching for the error source, the AER driver rules out devices whose
enable_cnt is zero. This was introduced in 2009 by commit 28eb27cf08
("PCI AER: support invalid error source IDs") without providing a
rationale.
Drivers typically call pci_enable_device() on probe, hence the enable_cnt
check essentially filters out unbound devices. At the time of the commit,
drivers had to opt in to AER by calling pci_enable_pcie_error_reporting()
and so any AER-enabled device could be assumed to be bound to a driver.
The check thus made sense because it allowed skipping config space accesses
to devices which were known not to be the error source.
But since 2022, AER is universally enabled on all devices when they are
enumerated, cf. commit f26e58bf6f ("PCI/AER: Enable error reporting when
AER is native").
Errors may very well be reported by unbound devices, e.g. due to link
instability. By ruling them out as error source, errors reported by them
are neither logged nor cleared. When they do get bound and another error
occurs, the earlier error is reported together with the new error, which
may confuse users. Stop doing so.
Fixes: f26e58bf6f ("PCI/AER: Enable error reporting when AER is native")
Signed-off-by: Lukas Wunner <lukas@wunner.de>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Stefan Roese <stefan.roese@mailbox.org>
Cc: stable@vger.kernel.org # v6.0+
Link: https://patch.msgid.link/734338c2e8b669db5a5a3b45d34131b55ffebfca.1774605029.git.lukas@wunner.de
PCI bridges are allowed to refuse activating VGA decoding, by simply
ignoring attempts to set the bit that enables it, so after setting the bit,
read it back to verify.
One example of such a bridge is the root bridge in IBM PowerNV, but this is
also useful for GPU passthrough into virtual machines, where it is
difficult to set up routing for legacy IO through IOMMU.
Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
[bhelgaas: subject, add comment about VGA Enable writability]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260307173538.763188-5-Simon.Richter@hogyros.de
The commit bc75c8e507 ("PCI: Rewrite bridge window head alignment
function") did not use if (r_size <= align) check from pbus_size_mem() for
the new head alignment bookkeeping structure (aligns2[]). In some
configurations, this can result in producing a gap into the bridge window
which the resource larger than its alignment cannot fill.
The old alignment calculation algorithm was removed by the subsequent
commit 3958bf16e2 ("PCI: Stop over-estimating bridge window size") which
renamed the aligns2[] array leaving only aligns[] array.
Add the if (r_size <= align) check back to avoid this problem.
Fixes: bc75c8e507 ("PCI: Rewrite bridge window head alignment function")
Reported-by: Guenter Roeck <linux@roeck-us.net>
Closes: https://lore.kernel.org/all/b05a6f14-979d-42c9-924c-d8408cb12ae7@roeck-us.net/
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Tested-by: Xifer <xiferdev@gmail.com>
Link: https://patch.msgid.link/20260324165633.4583-11-ilpo.jarvinen@linux.intel.com
When a bridge window contains big and small resource(s), the small
resource(s) may not amount to the half of the size of the big resource
which would allow calculate_head_align() to shrink the head alignment.
This results in always placing the small resource(s) after the big
resource.
In general, it would be good to be able to place the small resource(s)
before the big resource to achieve better utilization of the address space.
In the cases where the large resource can only fit at the end of the
window, it is even required.
However, carrying the information over from pbus_size_mem() and
calculate_head_align() to __pci_assign_resource() and
pcibios_align_resource() is not easy with the current data structures.
A somewhat hacky way to move the non-aligning tail part to the head is
possible within pcibios_align_resource(). The free space between the start
of the free space span and the aligned start address can be compared with
the non-aligning remainder of the size. If the free space is larger than
the remainder, placing the remainder before the start address is possible.
This relocation should generally work, because PCI resources consist only
power-of-2 atoms.
Various arch requirements may still need to override the relocation, so the
relocation is only applied selectively in such cases.
Closes: https://bugzilla.kernel.org/show_bug.cgi?id=221205
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Tested-by: Xifer <xiferdev@gmail.com>
Link: https://patch.msgid.link/20260324165633.4583-10-ilpo.jarvinen@linux.intel.com
window_alignment() lacks prefix. Rename it to pci_min_window_alignment() in
order to include the prefix and also add min to indicate the returned
window alignment is the minimum PCI spec and arch allows.
Also make it available in drivers/pci/pci.h as upcoming changes will need
to call it from outside of setup-bus.c.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Tested-by: Xifer <xiferdev@gmail.com>
Link: https://patch.msgid.link/20260324165633.4583-9-ilpo.jarvinen@linux.intel.com
Caller of pcibios_align_resource() (__find_resource_space()) already aligns
the start address by 'alignment' so aligning is only necessary if align >
alignment.
Change also to use ALIGN() instead of open-coding.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260324165633.4583-8-ilpo.jarvinen@linux.intel.com
Aligning res->start by align inside pcibios_align_resource() is unnecessary
because caller of pcibios_align_resource() is __find_resource_space() that
aligns res->start with align before calling pcibios_align_resource().
Aligning by align in case of IORESOURCE_IO && start & 0x300 cannot ever
result in changing start either because 0x300 bits would have not survived
the earlier alignment if align was large enough to have an impact.
Thus, remove the duplicated aligning from pcibios_align_resource().
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260324165633.4583-7-ilpo.jarvinen@linux.intel.com
Aligning res->start by align inside pcibios_align_resource() is unnecessary
because caller of pcibios_align_resource() is __find_resource_space() that
aligns res->start with align before calling pcibios_align_resource().
Aligning by align in case of IORESOURCE_IO && start & 0x300 cannot ever
result in changing start either because 0x300 bits would have not survived
the earlier alignment if align was large enough to have an impact.
Thus, remove the duplicated aligning from pcibios_align_resource().
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Acked-by: Greg Ungerer <gerg@linux-m68k.org>
Link: https://patch.msgid.link/20260324165633.4583-6-ilpo.jarvinen@linux.intel.com
Aligning res->start by align inside pcibios_align_resource() is unnecessary
because caller of pcibios_align_resource() is __find_resource_space() that
aligns res->start with align before calling pcibios_align_resource().
Aligning by align in case of IORESOURCE_IO && start & 0x300 cannot ever
result in changing start either because 0x300 bits would have not survived
the earlier alignment if align was large enough to have an impact.
Thus, remove the duplicated aligning from pcibios_align_resource().
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260324165633.4583-5-ilpo.jarvinen@linux.intel.com
__find_resource_space() has variable called 'tmp'. Rename it to
'full_avail' to better indicate its purpose.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Tested-by: Xifer <xiferdev@gmail.com>
Link: https://patch.msgid.link/20260324165633.4583-4-ilpo.jarvinen@linux.intel.com
__find_resource_space() calculates the full extent of empty space but only
passes the aligned space to resource_alignf callback. In some situations,
the callback may choose take advantage of the free space before the
requested alignment.
Pass the full extent of the calculated empty space to resource_alignf
callback as an additional parameter.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Tested-by: Xifer <xiferdev@gmail.com>
Link: https://patch.msgid.link/20260324165633.4583-3-ilpo.jarvinen@linux.intel.com
__find_resource_space() currently uses resource_contains() but for
tentative resources that are not yet crafted into the resource tree. As
resource_contains() checks that IORESOURCE_UNSET is not set for either of
the resources, the caller has to hack around this problem by clearing the
IORESOURCE_UNSET flag (essentially lying to resource_contains()).
Instead of the hack, introduce __resource_contains_unbound() for cases like
this.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Tested-by: Xifer <xiferdev@gmail.com>
Link: https://patch.msgid.link/20260324165633.4583-2-ilpo.jarvinen@linux.intel.com
reassign_resources_sorted() checks for two things:
a) Resource assignment failures for mandatory resources by checking if the
resource remains unassigned, which are known to always repeat, and does
not attempt to assign them again.
b) That resource is not among the ones being processed/assigned at this
stage, leading to skip processing such resources in
reassign_resources_sorted() as well (resource assignment progresses
one PCI hierarchy level at a time).
The problem here is that a) is checked before b), but b) also implies the
resource is not being assigned yet, making also a) true. As a) only skips
resource assignment but still removes the resource from realloc_head, the
later stages that would need to process the information in realloc_head
cannot obtain the optional size information anymore. This leads to
considering only non-optional part for bridge windows deeper in the PCI
hierarchy.
This problem has been observed during rescan (add_size is not considered
while attempting assignment for 0000:e2:00.0 indicating the corresponding
entry was removed from realloc_head while processing resource assignments
for 0000:e1):
pci_bus 0000:e1: scanning bus
...
pci 0000:e3:01.0: bridge window [mem 0x800000000-0x1000ffffff 64bit pref] to [bus e4] add_size 60c000000 add_align 800000000
pci 0000:e3:01.0: bridge window [mem 0x00100000-0x000fffff] to [bus e4] add_size 200000 add_align 200000
pci 0000:e3:02.0: disabling bridge window [mem 0x00000000-0x000fffff 64bit pref] to [bus e5] (unused)
pci 0000:e2:00.0: bridge window [mem 0x800000000-0x1000ffffff 64bit pref] to [bus e3-e5] add_size 60c000000 add_align 800000000
pci 0000:e2:00.0: bridge window [mem 0x00100000-0x001fffff] to [bus e3-e5] add_size 200000 add_align 200000
pcieport 0000:e1:02.0: bridge window [io size 0x2000]: can't assign; no space
pcieport 0000:e1:02.0: bridge window [io size 0x2000]: failed to assign
pcieport 0000:e1:02.0: bridge window [io 0x1000-0x2fff]: resource restored
pcieport 0000:e1:02.0: bridge window [io 0x1000-0x2fff]: resource restored
pcieport 0000:e1:02.0: bridge window [io size 0x2000]: can't assign; no space
pcieport 0000:e1:02.0: bridge window [io size 0x2000]: failed to assign
pci 0000:e2:00.0: bridge window [mem 0x28f000000000-0x28f800ffffff 64bit pref]: assigned
Fixes: 96336ec702 ("PCI: Perform reset_resource() and build fail list in sync")
Reported-by: Peter Nisbet <peter.nisbet@intel.com>
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Tested-by: Peter Nisbet <peter.nisbet@intel.com>
Link: https://patch.msgid.link/20260313084551.1934-1-ilpo.jarvinen@linux.intel.com
Steve reported an eGPU (either Radeon Instinct MI50 32GB or NVIDIA 3080
10GB) connected via Thunderbolt was not assigned sufficient BAR space in
v6.11, so the amdgpu and nvidia drivers were unable to initialize the
device.
pci_bridge_distribute_available_resources() -> ... ->
adjust_bridge_window() is called between __pci_bus_size_bridges()
and assigning the resources. Since the commit 948675736a ("PCI: Allow
adjust_bridge_window() to shrink resource if necessary")
adjust_bridge_window() can also shrink the bridge window. The shrunken
size, however, conflicts with what __pci_bus_size_bridges() ->
pbus_size_mem() calculated as the required bridge window size. By shrinking
the size, adjust_bridge_window() prevents the rest of the resource fitting
algorithm from working as intended. Resource fitting logic is expecting
assignment failures when bridge windows need resizing, but there are cases
where failures are no longer happening after the commit 948675736a ("PCI:
Allow adjust_bridge_window() to shrink resource if necessary").
The commit 948675736a ("PCI: Allow adjust_bridge_window() to shrink
resource if necessary") justifies the change by the extra reservation
made due to hpmemsize parameter, however, the kernel code contradicts
that statement. (For simplicity, finer-grained hpmmiosize and hpmmiopref
parameters that can be used to the same effect as hpmemsize are ignored in
this description.)
pbus_size_mem() calls calculate_memsize() twice. First with add_size=0
to find out the minimal required resource size. The second call occurs
with add_size=hpmemsize (effectively) but the result does not directly
affect the resource size only resulting in an entry on the realloc_head
list (a.k.a. add_list). Yet, adjust_bridge_window() directly changes
the resource size which does not include what is reserved due to
hpmemsize. Also, if the required size for the bridge window exceeds
hpmemsize, the parameter does not have any effect even on the second
size calculation made by pbus_size_mem(); from calculate_memsize():
size = max(size, add_size) + children_add_size;
The commit ae4611f1d7 ("PCI: Set resource size directly in
adjust_bridge_window()") that precedes the commit 948675736a ("PCI:
Allow adjust_bridge_window() to shrink resource if necessary") is also
related to causing this problem. Its changelog explicitly states
adjust_bridge_window() wants to "guarantee" allocation success.
Guaranteed allocations, however, are incompatible with how the other
parts of the resource fitting algorithm work. The given justification
fails to explain why guaranteed allocations at this stage are required
nor why forcing window to a smaller value than what was calculated by
pbus_size_mem() is correct. While the change might have worked by chance
in some test scenario, too small bridge window does not "guarantee"
success from the point of view of the endpoint device resource
assignments. No issue is mentioned within the changelog so it's unclear
if the change was made to fix some observed issue nor and what that
issue was.
The unwanted shrinking of a bridge window occurs, e.g., when a device with
large BARs such as eGPU is attached using Thunderbolt and the Root Port
holds less than enough resource space for the eGPU. The GPU resources are
in order of GBs and the default hotplug allocation is a mere 2MB
(DEFAULT_HOTPLUG_MMIO_PREF_SIZE). The problem is illustrated by this log
(filtered to the relevant content only):
pci 0000:00:07.0: PCI bridge to [bus 03-2c]
pci 0000:00:07.0: bridge window [mem 0x6000000000-0x601bffffff 64bit pref]
pci 0000:03:00.0: PCI bridge to [bus 00]
pci 0000:03:00.0: bridge window [mem 0x00000000-0x000fffff 64bit pref]
pci 0000:03:00.0: bridge configuration invalid ([bus 00-00]), reconfiguring
pci 0000:03:00.0: PCI bridge to [bus 04-2c]
pcieport 0000:00:07.0: Assigned bridge window [mem 0x6000000000-0x601bffffff 64bit pref] to [bus 03-2c] cannot fit 0xc00000000 required for 0000:03:00.0 bridging to [bus 04-2c]
pci 0000:03:00.0: bridge window [mem 0x800000000-0x10003fffff 64bit pref] to [bus 04-2c] add_size 100000 add_align 100000
pcieport 0000:00:07.0: distributing available resources
pci 0000:03:00.0: bridge window [mem 0x800000000-0x10003fffff 64bit pref] shrunken by 0x00000007e4400000
pci 0000:03:00.0: bridge window [mem 0x6000000000-0x601bffffff 64bit pref]: assigned
The initial size of the Root Port's window is 448MB (0x601bffffff -
0x6000000000). __pci_bus_size_bridges() -> pbus_size_mem() calculates the
required size to be 32772 MB (0x10003fffff - 0x800000000) which would fit
the eGPU resources. adjust_bridge_window() then shrinks the bridge window
down to what is guaranteed to fit into the Root Port's bridge window. The
bridge window for 03:00.0 is also eliminated from the add_list (a.k.a.
realloc_head) list by adjust_bridge_window().
After adjustment, the resources are assigned and as the bridge window for
03:00.0 is assigned successfully, no failure is recorded. Without a
failure, no attempt to resize the window of the Root Port is required. The
end result is eGPU not having large enough resources to work.
The commit 948675736a ("PCI: Allow adjust_bridge_window() to shrink
resource if necessary") also claims nested bridge windows are sized the
same, which is false. pbus_size_mem() calculates the size for the parent
bridge window by summing all the downstream resources so the resource
fitting calculates larger bridge window for the parent to accommodate the
childen. That is, hpmemsize does not result the same size for the case
where there are nested bridge windows.
In order to fix the most immediate problem, don't shrink the resource size
in adjust_bridge_window() as hpmemsize had nothing to do with it. When
considering add_size, only reduce it up to what is added due to hpmemsize
(if required size is larger than hpmemsize, the parameter has no impact,
see calculate_memsize()). Unfortunately, if the tail of the bridge window
was aligned in calculate_memsize() from below hpmemsize to above it, the
size check will falsely match but the check at least errs to the side of
caution. There's not enough information available in adjust_bridge_window()
to know the calculated size precisely.
This is not exactly a revert of the commits e4611f1d7e9 ("PCI: Set resource
size directly in adjust_bridge_window()") and 948675736a ("PCI: Allow
adjust_bridge_window() to shrink resource if necessary") as shrinking still
remains in place but is implemented differently, and the end result behaves
very differently.
It is possible that those two commits fixed some other issue that is not
described with enough detail in the changelog and undoing parts of them
results in another regression due to behavioral change. Nonetheless, as
described above, the solution by those two commits was flawed and the
issue, if one exists, should be solved in a way that is compatible with the
rest of the resource fitting algorithm instead of working against it.
Besides shrinking, the case where adjust_bridge_window() expands the bridge
window is likely somewhat wrong as well because it removes the entry from
add_list (a.k.a. realloc_head), but it is less damaging as that only
impacts optional resources and may have no impact if expanding by hpmemsize
is larger than what add_size was. Fixing it is left as further work.
Fixes: 948675736a ("PCI: Allow adjust_bridge_window() to shrink resource if necessary")
Fixes: ae4611f1d7 ("PCI: Set resource size directly in adjust_bridge_window()")
Reported-by: Steve Oswald <stevepeter.oswald@gmail.com>
Closes: https://lore.kernel.org/linux-pci/CAN95MYEaO8QYYL=5cN19nv_qDGuuP5QOD17pD_ed6a7UqFVZ-g@mail.gmail.com/
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260219153951.68869-1-ilpo.jarvinen@linux.intel.com
Previously, pci_read_bridge_io() and pci_read_bridge_mmio_pref()
unconditionally set resource type flags (IORESOURCE_IO or IORESOURCE_MEM |
IORESOURCE_PREFETCH) when reading bridge window registers. For windows that
are not implemented in hardware, this may cause the allocator to assign
space for a window that doesn't exist.
For example, the EcoNET EN7528 SoC Root Port doesn't support the
prefetchable window, but since a downstream device had a prefetchable BAR,
the allocator mistakenly assigned a prefetchable window:
pci 0001:00:01.0: [14c3:0811] type 01 class 0x060400 PCIe Root Port
pci 0001:00:01.0: PCI bridge to [bus 01-ff]
pci 0001:00:01.0: bridge window [mem 0x28000000-0x280fffff]: assigned
pci 0001:00:01.0: bridge window [mem 0x28100000-0x282fffff pref]: assigned
pci 0001:01:00.0: BAR 0 [mem 0x28100000-0x281fffff 64bit pref]: assigned
pci_read_bridge_windows() already detects unsupported windows by testing
register writability and sets dev->io_window/pref_window accordingly.
Check dev->io_window/pref_window so we don't set the resource flags for
unsupported windows, which prevents the allocator from assigning space to
them.
After this commit, the prefetchable BAR is correctly allocated from the
non-prefetchable window:
pci 0001:00:01.0: bridge window [mem 0x28000000-0x281fffff]: assigned
pci 0001:01:00.0: BAR 0 [mem 0x28000000-0x280fffff 64bit pref]: assigned
Suggested-by: Bjorn Helgaas <helgaas@kernel.org>
Link: https://lore.kernel.org/all/20260113210259.GA715789@bhelgaas/
Signed-off-by: Ahmed Naseef <naseefkm@gmail.com>
Signed-off-by: Caleb James DeLisle <cjd@cjdns.fr>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260312165332.569772-4-cjd@cjdns.fr
pci_set_vga_state() returns an error code, which so far has been ignored by
the only caller, __vga_tryget(), so forward it to the caller. As the return
type of __vga_tryget() is a pointer, wrap the error in ERR_PTR().
Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260307173538.763188-3-Simon.Richter@hogyros.de
If VGA routing cannot be established, vga_get_uninterruptible() returns an
error and does not increment the lock count. Return the error to the
caller.
Return before incrementing uc->io_cnt/mem_cnt so vga_arb_release() won't
call vga_put() when userspace closes the handle.
Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
[bhelgaas: commit log]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260307173538.763188-2-Simon.Richter@hogyros.de
Enable the generic pwrctrl driver to control power of PCIe
UPD720201/UPD720202 USB 3.0 xHCI Host Controller.
Signed-off-by: Neil Armstrong <neil.armstrong@linaro.org>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Manivannan Sadhasivam <mani@kernel.org>
Reviewed-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>
Link: https://patch.msgid.link/20260220-topic-sm8650-ayaneo-pocket-s2-base-v5-4-1ad79caa1efa@linaro.org
dev_err_probe() returns the error code passed to it, so this:
dev_err_probe(dev, ret, "Failed to get slot regulators\n");
return ret;
is equivalent to this:
return dev_err_probe(dev, ret, "Failed to get slot regulators\n");
Simplify by using the latter.
Suggested-by: Bartosz Golaszewski <brgl@kernel.org>
Signed-off-by: Neil Armstrong <neil.armstrong@linaro.org>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>
Reviewed-by: Manivannan Sadhasivam <mani@kernel.org>
Link: https://patch.msgid.link/20260220-topic-sm8650-ayaneo-pocket-s2-base-v5-2-1ad79caa1efa@linaro.org
The driver is pretty generic and would fit for either PCI Slots or
endpoints connected to PCI ports, so rename the driver and module as
pci-pwrctrl-generic.
Suggested-by: Manivannan Sadhasivam <mani@kernel.org>
Signed-off-by: Neil Armstrong <neil.armstrong@linaro.org>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Manivannan Sadhasivam <mani@kernel.org>
Reviewed-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>
Link: https://patch.msgid.link/20260220-topic-sm8650-ayaneo-pocket-s2-base-v5-3-1ad79caa1efa@linaro.org
Use scoped for-each loop when iterating over device nodes to make code a
bit simpler.
Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@oss.qualcomm.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Reviewed-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Link: https://patch.msgid.link/20260317133322.266102-8-krzysztof.kozlowski@oss.qualcomm.com
Use scoped for-each loop when iterating over device nodes to make code a
bit simpler.
Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@oss.qualcomm.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Reviewed-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Link: https://patch.msgid.link/20260317133322.266102-7-krzysztof.kozlowski@oss.qualcomm.com
Per CXL r3.1, sec 8.1.5.2, the Secondary Bus Reset (SBR) bit in the Bridge
Control register of a CXL port has no effect unless the "Unmask SBR" bit in
the Port Control Extensions Register is set.
After b1956e2d07 ("PCI/CXL: Fail bus reset if upstream CXL Port has SBR
masked"), Linux checks the "Unmask SBR" bit in pci_reset_bus_function().
But when probe==true, it previously returned 0, incorrectly indicating that
SBR is a viable reset method for the device.
As a result, "bus" is listed in the device's "reset_method" attribute even
though the hardware is incapable of performing it. If a user writes "bus"
to "reset_method" or triggers a reset that falls back to SBR, the operation
fails with "write error: Inappropriate ioctl for device".
If the link is operating in CXL mode (pcie_is_cxl()), return -ENOTTY
immediately unless "Unmask SBR" is set, regardless of the probe argument.
This ensures that "bus" is not advertised in "reset_methods" when the
hardware prevents it, improving clarity for users and aligning the sysfs
capability report with actual hardware behavior.
Signed-off-by: Vidya Sagar <vidyas@nvidia.com>
[bhelgaas: commit log, use pcie_is_cxl()]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://patch.msgid.link/20260225133801.30231-1-vidyas@nvidia.com
The AER and EDR error handling paths hold a reference on the pci_dev during
recovery. Hold a reference during the DPC recovery path as well.
Signed-off-by: Sizhe Liu <liusizhe5@huawei.com>
[bhelgaas: split to separate patch]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
https://patch.msgid.link/20260214081130.1878424-1-liusizhe5@huawei.com
pci_bus/slot_lock/unlock/trylock() largely duplicate the bus iteration loop
with variation only due to slot filter handling. The only differences in
the loops is where the struct bus is found (directly in the argument vs in
slot->bus) and whether slot filter is applied. Those differences are simple
to handle using function parameters.
Consolidate the bus iteration loop to one place by creating
__pci_bus_{lock,unlock,trylock}() and call them from the non-underscore
locking functions.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Link: https://patch.msgid.link/20260304122139.1479-1-ilpo.jarvinen@linux.intel.com
Use the slot reset method when resetting the bridge if the bus contains
hot plug slots. This fixes spurious hot plug events that are triggered
by the secondary bus reset that bypasses the slot's detection disabling.
Resetting a bridge's subordinate bus can be done like this:
# echo 1 > /sys/bus/pci/devices/0000:50:01.0/reset_subordinate
Prior to this patch, an example kernel message may show something like:
pcieport 0000:50:01.0: pciehp: Slot(40): Link Down
With this change, the pciehp driver ignores the link event during the
reset, so may show this message instead:
pcieport 0000:50:01.0: pciehp: Slot(40): Link Down/Up ignored
Signed-off-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Link: https://patch.msgid.link/20260217160836.2709885-4-kbusch@meta.com
A PCIe hotplug slot applies to the entire secondary bus. Thus, pciehp only
allocates a single hotplug_slot for the bridge to that bus. The existing
PCI slot, though, would only match to functions on device 0, meaning any
devices beyond that, e.g., ARI functions, are not matched to any slot even
though they share it. A slot reset will break all the missing devices
because the handling skips them.
For example, ARI devices with more than 8 functions fail because their
state is not properly handled, nor is the attached driver notified of the
reset. In the best case, the device will appear unresponsive to the driver,
resulting in unexpected errors. A worse possibility may panic the kernel if
in-flight transactions trigger hardware reported errors like this real
observation:
vfio-pci 0000:01:00.0: resetting
vfio-pci 0000:01:00.0: reset done
{1}[Hardware Error]: Error 1, type: fatal
{1}[Hardware Error]: section_type: PCIe error
{1}[Hardware Error]: port_type: 0, PCIe end point
{1}[Hardware Error]: version: 0.2
{1}[Hardware Error]: command: 0x0140, status: 0x0010
{1}[Hardware Error]: device_id: 0000:01:01.0
{1}[Hardware Error]: slot: 0
{1}[Hardware Error]: secondary_bus: 0x00
{1}[Hardware Error]: vendor_id: 0x1d9b, device_id: 0x0207
{1}[Hardware Error]: class_code: 020000
{1}[Hardware Error]: bridge: secondary_status: 0x0000, control: 0x0000
{1}[Hardware Error]: aer_cor_status: 0x00008000, aer_cor_mask: 0x00002000
{1}[Hardware Error]: aer_uncor_status: 0x00010000, aer_uncor_mask: 0x00100000
{1}[Hardware Error]: aer_uncor_severity: 0x006f6030
{1}[Hardware Error]: TLP Header: 0a412800 00192080 60000004 00000004
GHES: Fatal hardware error but panic disabled
Kernel panic - not syncing: GHES: Fatal hardware error
Allow a slot to be created to claim all devices on a bus, not just a
matching device. This is done by introducing a sentinel value, named
PCI_SLOT_ALL_DEVICES, which then has the PCI slot match to any device on
the bus. This fixes slot resets for pciehp.
Since 0xff already has special meaning, the chosen value for this new
feature is 0xfe. This will not clash with any actual slot number since they
are limited to 5 bits.
Signed-off-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Link: https://patch.msgid.link/20260217160836.2709885-3-kbusch@meta.com
