Complete reset-on-init sequence before sysfs interfaces are created.
Devices get properly initiaized only after reset, and then only sysfs
interfaces should be made available.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Asad Kamal <asad.kamal@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Decouple the amdgpu fence from the amdgpu_job structure.
This lets us clean up the separate fence ops for the embedded
fence and other fences. This also allows us to allocate the
vm fence up front when we allocate the job.
v2: Additional cleanup suggested by Christian
v3: Additional cleanups suggested by Christian
v4: Additional cleanups suggested by David and
vm fence fix
v5: cast seqno (David)
Cc: David.Wu3@amd.com
Cc: christian.koenig@amd.com
Tested-by: David (Ming Qiang) Wu <David.Wu3@amd.com>
Reviewed-by: David (Ming Qiang) Wu <David.Wu3@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
VCN and VPE userqs are not yet supported and this code is
not correct. Userspace should provide the correct
doorbell offset with in their doorbell page for the IP.
Adjusting it here will not work as expected as userspace
and the queue itself will have different offsets.
We need to add a INFO IOCTL query to get the offset and
range for each IP within the doorbell page to handle this
properly.
Cc: Saleemkhan Jamadar <saleemkhan.jamadar@amd.com>
Reviewed-by: Saleemkhan Jamadar <saleemkhan.jamadar@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If a userq failed to suspend the rest of the suspend sequence may
have problems. Pass the error code up to the caller for a decision
on what to do.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If multiple userq IDR are in use and there is an error handling one
at suspend or resume it will be silently discarded.
Switch the suspend/resume() code to use guards and return immediately.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_device_ip_suspend_phase2() calls amdgpu_ip_block_suspend()
which already sets HW block status to false when succeeding with
IP suspend. Remove the explicit call in
amdgpu_device_ip_suspend_phase2() so that the status is accurate.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Error handling was introduced in commit e095026f00 ("drm/amdgpu:
validate suspend before function call") so the comment about TODO is no
longer needed.
Fixes: e095026f00 ("drm/amdgpu: validate suspend before function call")
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
The shutdown() callback uses amdgpu_ip_suspend() which doesn't notify
drm clients during shutdown. This could lead to hangs.
[How]
Change amdgpu_pci_shutdown() to call the same sequence as suspend/resume.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When a user unmaps a userq VA, the driver must ensure
the queue has no in-flight jobs. If there is pending work,
the kernel should wait for the attached eviction (bookkeeping)
fence to signal before deleting the mapping.
Suggested-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Prike Liang <Prike.Liang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Keeping waiting the userq fence infinitely until
hang detection, and then suspend the hang queue and
set the fence error.
Signed-off-by: Prike Liang <Prike.Liang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
There has been multiple complains that 10 seconds are usually to long.
The original requirement for longer timeout came from compute tests on
AMDVLK, since that is no longer a topic reduce the timeout back to 2
seconds for all queues.
While at it also remove any special handling for compute queues under
SRIOV or pass through.
v2: fix checkpatch warning.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The firmware limits the max vmid, but align the
settings with the hw limits as well just to be safe.
Reviewed-by: Shaoyun liu <Shaoyun.liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This lock was part of the SDMA workaround originally implemented in
gmc_v10_0_flush_gpu_tlb (a70cb2176f moved it to
amdgpu_gmc_flush_gpu_tlb).
This means this lock is useless and be safely dropped.
Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_ttm_copy_mem_to_mem has a single caller, make sure the out
fence is non-NULL to simplify the code.
Since none of the pointers should be NULL, we can enable
__attribute__((nonnull))__.
While at it make the function static since it's only used from
amdgpuu_ttm.c.
Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
By design the MES will return an array result that is twice the number
of hung doorbells it can report.
i.e. if up k reported doorbells are supported, then the
second half of the array, also of length k, holds the HQD information
(type/queue/pipe) where queue 1 corresponds to index 0 and k,
queue 2 corresponds to index 1 and k + 1 etc ...
The driver will use the HDQ info to target queue/pipe reset for
hardware scheduled user compute queues.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Initialized doorbells should be set to invalid rather than 0 to prevent
driver from over counting hung doorbells since it checks against the
invalid value to begin with.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Clean up the amdgpu hmm range functions for clearer
definition of each.
a. Split amdgpu_ttm_tt_get_user_pages_done into two:
1. amdgpu_hmm_range_valid: To check if the user pages
are valid and update seq num
2. amdgpu_hmm_range_free: Clean up the hmm range
and pfn memory.
b. amdgpu_ttm_tt_get_user_pages_done and
amdgpu_ttm_tt_discard_user_pages are similar function so remove
discard and directly use amdgpu_hmm_range_free to clean up the
hmm range and pfn memory.
Suggested-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Sunil Khatri <sunil.khatri@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
update the amdgpu_ttm_tt_get_user_pages and all dependent function
along with it callers to use a user allocated hmm_range buffer instead
hmm layer allocates the buffer.
This is a need to get hmm_range pointers easily accessible
without accessing the bo and that is a requirement for the
userqueue to lock the userptrs effectively.
Signed-off-by: Sunil Khatri <sunil.khatri@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GFX12 MES uses low 32 bits of status return for success (1 or 0)
and high bits for debug information if low bits are 0.
GFX11 MES doesn't do this so checking full 64-bit status return
for 1 or 0 is still valid.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Previously, APU platforms (and other scenarios with uninitialized VRAM managers)
triggered a NULL pointer dereference in `ttm_resource_manager_usage()`. The root
cause is not that the `struct ttm_resource_manager *man` pointer itself is NULL,
but that `man->bdev` (the backing device pointer within the manager) remains
uninitialized (NULL) on APUs—since APUs lack dedicated VRAM and do not fully
set up VRAM manager structures. When `ttm_resource_manager_usage()` attempts to
acquire `man->bdev->lru_lock`, it dereferences the NULL `man->bdev`, leading to
a kernel OOPS.
1. **amdgpu_cs.c**: Extend the existing bandwidth control check in
`amdgpu_cs_get_threshold_for_moves()` to include a check for
`ttm_resource_manager_used()`. If the manager is not used (uninitialized
`bdev`), return 0 for migration thresholds immediately—skipping VRAM-specific
logic that would trigger the NULL dereference.
2. **amdgpu_kms.c**: Update the `AMDGPU_INFO_VRAM_USAGE` ioctl and memory info
reporting to use a conditional: if the manager is used, return the real VRAM
usage; otherwise, return 0. This avoids accessing `man->bdev` when it is
NULL.
3. **amdgpu_virt.c**: Modify the vf2pf (virtual function to physical function)
data write path. Use `ttm_resource_manager_used()` to check validity: if the
manager is usable, calculate `fb_usage` from VRAM usage; otherwise, set
`fb_usage` to 0 (APUs have no discrete framebuffer to report).
This approach is more robust than APU-specific checks because it:
- Works for all scenarios where the VRAM manager is uninitialized (not just APUs),
- Aligns with TTM's design by using its native helper function,
- Preserves correct behavior for discrete GPUs (which have fully initialized
`man->bdev` and pass the `ttm_resource_manager_used()` check).
v4: use ttm_resource_manager_used(&adev->mman.vram_mgr.manager) instead of checking the adev->gmc.is_app_apu flag (Christian)
Reviewed-by: Christian König <christian.koenig@amd.com>
Suggested-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The atomic variable vm_fault_info_updated is used to synchronize access to
adev->gmc.vm_fault_info between the interrupt handler and
get_vm_fault_info().
The default atomic functions like atomic_set() and atomic_read() do not
provide memory barriers. This allows for CPU instruction reordering,
meaning the memory accesses to vm_fault_info and the vm_fault_info_updated
flag are not guaranteed to occur in the intended order. This creates a
race condition that can lead to inconsistent or stale data being used.
The previous implementation, which used an explicit mb(), was incomplete
and inefficient. It failed to account for all potential CPU reorderings,
such as the access of vm_fault_info being reordered before the atomic_read
of the flag. This approach is also more verbose and less performant than
using the proper atomic functions with acquire/release semantics.
Fix this by switching to atomic_set_release() and atomic_read_acquire().
These functions provide the necessary acquire and release semantics,
which act as memory barriers to ensure the correct order of operations.
It is also more efficient and idiomatic than using explicit full memory
barriers.
Fixes: b97dfa27ef ("drm/amdgpu: save vm fault information for amdkfd")
Cc: stable@vger.kernel.org
Signed-off-by: Gui-Dong Han <hanguidong02@gmail.com>
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When we backup ring contents to reemit after a queue reset,
we don't backup ring contents from the bad context. When
we signal the fences, we should set an error on those
fences as well.
v2: misc cleanups
v3: add locking for fence error, fix comment (Christian)
v4: fix wrap around, locking (Christian)
Fixes: 77cc0da39c ("drm/amdgpu: track ring state associated with a fence")
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Chips which use the IP discovery firmware loaded by the driver
reported incorrect harvesting information in the ip discovery
table in sysfs because the driver only uses the ip discovery
firmware for populating sysfs and not for direct parsing for the
driver itself as such, the fields that are used to print the
harvesting info in sysfs report incorrect data for some IPs. Populate
the relevant fields for this case as well.
Fixes: 514678da56 ("drm/amdgpu/discovery: fix fw based ip discovery")
Acked-by: Tom St Denis <tom.stdenis@amd.com>
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The Constant Engine found on gfx6-gfx10 HW has been a notorious source of
problems.
RADV never used it in the first place, radeonsi only used it for a few
releases around 2017 for gfx6-gfx9 before dropping support for it as
well.
While investigating another problem I just recently found that submitting
to the CE seems to be completely broken on gfx9 for quite a while.
Since nobody complained about that problem it most likely means that
nobody is using any of the affected radeonsi versions on current Linux
kernels any more.
So to potentially phase out the support for the CE and eliminate another
source of problems block submitting CE IBs unless it is enabled again
using a debug flag.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Timur Kristóf <timur.kristof@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Enabling ASPM causes randoms hangs on Tahiti and Oland on Zen4.
It's unclear if this is a platform-specific or GPU-specific issue.
Disable ASPM on SI for the time being.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Pull more drm fixes from Dave Airlie:
"Just the follow up fixes for rc1 from the next branch, amdgpu and xe
mostly with a single v3d fix in there.
amdgpu:
- DC DCE6 fixes
- GPU reset fixes
- Secure diplay messaging cleanup
- MES fix
- GPUVM locking fixes
- PMFW messaging cleanup
- PCI US/DS switch handling fix
- VCN queue reset fix
- DC FPU handling fix
- DCN 3.5 fix
- DC mirroring fix
amdkfd:
- Fix kfd process ref leak
- mmap write lock handling fix
- Fix comments in IOCTL
xe:
- Fix build with clang 16
- Fix handling of invalid configfs syntax usage and spell out the
expected syntax in the documentation
- Do not try late bind firmware when running as VF since it shouldn't
handle firmware loading
- Fix idle assertion for local BOs
- Fix uninitialized variable for late binding
- Do not require perfmon_capable to expose free memory at page
granularity. Handle it like other drm drivers do
- Fix lock handling on suspend error path
- Fix I2C controller resume after S3
v3d:
- fix fence locking"
* tag 'drm-next-2025-10-11-1' of https://gitlab.freedesktop.org/drm/kernel: (34 commits)
drm/amd/display: Incorrect Mirror Cositing
drm/amd/display: Enable Dynamic DTBCLK Switch
drm/amdgpu: Report individual reset error
drm/amdgpu: partially revert "revert to old status lock handling v3"
drm/amd/display: Fix unsafe uses of kernel mode FPU
drm/amd/pm: Disable VCN queue reset on SMU v13.0.6 due to regression
drm/amdgpu: Fix general protection fault in amdgpu_vm_bo_reset_state_machine
drm/amdgpu: Check swus/ds for switch state save
drm/amdkfd: Fix two comments in kfd_ioctl.h
drm/amd/pm: Avoid interface mismatch messaging
drm/amdgpu: Merge amdgpu_vm_set_pasid into amdgpu_vm_init
drm/amd/amdgpu: Fix the mes version that support inv_tlbs
drm/amd: Check whether secure display TA loaded successfully
drm/amdkfd: Fix mmap write lock not release
drm/amdkfd: Fix kfd process ref leaking when userptr unmapping
drm/amdgpu: Fix for GPU reset being blocked by KIQ I/O.
drm/amd/display: Disable scaling on DCE6 for now
drm/amd/display: Properly disable scaling on DCE6
drm/amd/display: Properly clear SCL_*_FILTER_CONTROL on DCE6
drm/amd/display: Add missing DCE6 SCL_HORZ_FILTER_INIT* SRIs
...
If reinitialization of one of the GPUs fails after reset, it logs
failure on all subsequent GPUs eventhough they have resumed
successfully.
A sample log where only device at 0000:95:00.0 had a failure -
amdgpu 0000:15:00.0: amdgpu: GPU reset(19) succeeded!
amdgpu 0000:65:00.0: amdgpu: GPU reset(19) succeeded!
amdgpu 0000:75:00.0: amdgpu: GPU reset(19) succeeded!
amdgpu 0000:85:00.0: amdgpu: GPU reset(19) succeeded!
amdgpu 0000:95:00.0: amdgpu: GPU reset(19) failed
amdgpu 0000:e5:00.0: amdgpu: GPU reset(19) failed
amdgpu 0000:f5:00.0: amdgpu: GPU reset(19) failed
amdgpu 0000:05:00.0: amdgpu: GPU reset(19) failed
amdgpu 0000:15:00.0: amdgpu: GPU reset end with ret = -5
To avoid confusion, report the error for each device
separately and return the first error as the overall result.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Asad Kamal <asad.kamal@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The CI systems are pointing out list corruptions, so we still need to
fix something here.
Keep the asserts, but revert the lock changes for now.
Fixes: 59e4405e9e ("drm/amdgpu: revert to old status lock handling v3")
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
As KFD no longer uses a separate PASID, the global amdgpu_vm_set_pasid()function is no longer necessary.
Merge its functionality directly intoamdgpu_vm_init() to simplify code flow and eliminate redundant locking.
v2: remove superflous check
adjust amdgpu_vm_fin and remove amdgpu_vm_set_pasid (Chritian)
v3: drop amdgpu_vm_assert_locked (Chritian)
Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/4614
Fixes: 59e4405e9e ("drm/amdgpu: revert to old status lock handling v3")
Reviewed-by: Christian König <christian.koenig@amd.com>
Suggested-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
MES version 0x83 is not stable to use the inv_tlbs API. Defer it to 0x84 vertsion.
Fixes: 85442bac84 ("drm/amd/amdgpu: Fix the mes version that support inv_tlbs")
Signed-off-by: Shaoyun Liu <shaoyun.liu@amd.com>
Reviewed-by: Michael Chen <michael.chen@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Not all renoir hardware supports secure display. If the TA is present
but the feature isn't supported it will fail to load or send commands.
This shows ERR messages to the user that make it seems like there is
a problem.
[How]
Check the resp_status of the context to see if there was an error
before trying to send any secure display commands.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/1415
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
kfd_lookup_process_by_pid hold the kfd process reference to ensure it
doesn't get destroyed while sending the segfault event to user space.
Calling kfd_lookup_process_by_pid as function parameter leaks the kfd
process refcount and miss the NULL pointer check if app process is
already destroyed.
Fixes: 2d274bf709 ("amd/amdkfd: Trigger segfault for early userptr unmmapping")
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
