When the DRM scheduler times out, it's possible that the GPU isn't hung;
instead, a job just took unusually long (longer than the timeout) but is
still running, and there is, thus, no reason to reset the hardware. This
can occur in two scenarios:
1. The job is taking longer than the timeout, but the driver determined
through a GPU-specific mechanism that the hardware is still making
progress. Hence, the driver would like the scheduler to skip the
timeout and treat the job as still pending from then onward. This
happens in v3d, Etnaviv, and Xe.
2. Timeout has fired before the free-job worker. Consequently, the
scheduler calls `sched->ops->timedout_job()` for a job that isn't
timed out.
These two scenarios are problematic because the job was removed from the
`sched->pending_list` before calling `sched->ops->timedout_job()`, which
means that when the job finishes, it won't be freed by the scheduler
though `sched->ops->free_job()` - leading to a memory leak.
To solve these problems, create a new `drm_gpu_sched_stat`, called
DRM_GPU_SCHED_STAT_NO_HANG, which allows a driver to skip the reset. The
new status will indicate that the job must be reinserted into
`sched->pending_list`, and the hardware / driver will still complete that
job.
Reviewed-by: Philipp Stanner <phasta@kernel.org>
Link: https://lore.kernel.org/r/20250714-sched-skip-reset-v6-2-5c5ba4f55039@igalia.com
Signed-off-by: Maíra Canal <mcanal@igalia.com>
Among the scheduler's statuses, the only one that indicates an error is
DRM_GPU_SCHED_STAT_ENODEV. Any status other than DRM_GPU_SCHED_STAT_ENODEV
signifies that the operation succeeded and the GPU is in a nominal state.
However, to provide more information about the GPU's status, it is needed
to convey more information than just "OK".
Therefore, rename DRM_GPU_SCHED_STAT_NOMINAL to
DRM_GPU_SCHED_STAT_RESET, which better communicates the meaning of this
status. The status DRM_GPU_SCHED_STAT_RESET indicates that the GPU has
hung, but it has been successfully reset and is now in a nominal state
again.
Reviewed-by: Philipp Stanner <phasta@kernel.org>
Link: https://lore.kernel.org/r/20250714-sched-skip-reset-v6-1-5c5ba4f55039@igalia.com
Signed-off-by: Maíra Canal <mcanal@igalia.com>
Fix the compile-time warnings
drivers/gpu/drm/scheduler/sched_entity.c: warning: EXPORT_SYMBOL() is used, but #include <linux/export.h> is missing
drivers/gpu/drm/scheduler/sched_fence.c: warning: EXPORT_SYMBOL() is used, but #include <linux/export.h> is missing
drivers/gpu/drm/scheduler/sched_main.c: warning: EXPORT_SYMBOL() is used, but #include <linux/export.h> is missing
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Fixes: a934a57a42 ("scripts/misc-check: check missing #include <linux/export.h> when W=1")
Reviewed-by: André Almeida <andrealmeid@igalia.com>
Cc: Masahiro Yamada <masahiroy@kernel.org>
Cc: Nathan Chancellor <nathan@kernel.org>
Link: https://lore.kernel.org/r/20250612121633.229222-9-tzimmermann@suse.de
This will be used in a later commit to trace the drm client_id in
some of the gpu_scheduler trace events.
This requires changing all the users of drm_sched_job_init to
add an extra parameter.
The newly added drm_client_id field in the drm_sched_fence is a bit
of a duplicate of the owner one. One suggestion I received was to
merge those 2 fields - this can't be done right now as amdgpu uses
some special values (AMDGPU_FENCE_OWNER_*) that can't really be
translated into a client id. Christian is working on getting rid of
those; when it's done we should be able to squash owner/drm_client_id
together.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Signed-off-by: Philipp Stanner <phasta@kernel.org>
Link: https://lore.kernel.org/r/20250526125505.2360-3-pierre-eric.pelloux-prayer@amd.com
This reverts commit 44d2f310f0.
The function drm_sched_job_arm() is indeed the point of no return. The
background is that it is nearly impossible for the driver to correctly
retract the fence and signal it in the order enforced by the dma_fence
framework.
The code in drm_sched_job_cleanup() is for the purpose to cleanup after
the job was armed through drm_sched_job_arm() *and* processed by the
scheduler.
We can certainly improve the documentation, but removing the warning is
clearly not a good idea.
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Philipp Stanner <phasta@kernel.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20250312134400.2176393-1-christian.koenig@amd.com
drm_sched_backend_ops.run_job() returns a dma_fence for the scheduler.
That fence is signalled by the driver once the hardware completed the
associated job. The scheduler does not increment the reference count on
that fence, but implicitly expects to inherit this fence from run_job().
This is relatively subtle and prone to misunderstandings.
This implies that, to keep a reference for itself, a driver needs to
call dma_fence_get() in addition to dma_fence_init() in that callback.
It's further complicated by the fact that the scheduler even decrements
the refcount in drm_sched_run_job_work() since it created a new
reference in drm_sched_fence_scheduled(). It does, however, still use
its pointer to the fence after calling dma_fence_put() - which is safe
because of the aforementioned new reference, but actually still violates
the refcounting rules.
Move the call to dma_fence_put() to the position behind the last usage
of the fence.
Suggested-by: Danilo Krummrich <dakr@kernel.org>
Signed-off-by: Philipp Stanner <pstanner@redhat.com>
Reviewed-by: Danilo Krummrich <dakr@kernel.org>
Signed-off-by: Philipp Stanner <phasta@kernel.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20250305130551.136682-4-phasta@kernel.org
drm_sched_job_cleanup()'s documentation claims that calling
drm_sched_job_arm() is a "point of no return", implying that afterwards
a job cannot be cancelled anymore.
This is not correct, as proven by the function's code itself, which
takes a previous call to drm_sched_job_arm() into account. In truth, the
decisive factors are whether fences have been shared (e.g., with other
processes) and if the job has been submitted to an entity already.
Correct the wrong docstring.
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
Signed-off-by: Philipp Stanner <phasta@kernel.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20250304141346.102683-2-phasta@kernel.org
Idea is to add helpers for peeking and popping jobs from entities with
the goal of decoupling the hidden assumption in the code that queue_node
is the first element in struct drm_sched_job.
That assumption usually comes in the form of:
while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue))))
Which breaks if the queue_node is re-positioned due to_drm_sched_job
being implemented with a container_of.
This also allows us to remove duplicate definitions of to_drm_sched_job.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: Danilo Krummrich <dakr@kernel.org>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Philipp Stanner <phasta@kernel.org>
Signed-off-by: Philipp Stanner <phasta@kernel.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20250221105038.79665-2-tvrtko.ursulin@igalia.com
If jobs are still enqueued in struct drm_gpu_scheduler.pending_list
when drm_sched_fini() gets called, those jobs will be leaked since that
function stops both job-submission and (automatic) job-cleanup. It is,
thus, up to the driver to take care of preventing leaks.
The related function drm_sched_wqueue_stop() also prevents automatic job
cleanup.
Those pitfals are not reflected in the documentation, currently.
Explicitly inform about the leak problem in the docstring of
drm_sched_fini().
Additionally, detail the purpose of drm_sched_wqueue_{start,stop} and
hint at the consequences for automatic cleanup.
Signed-off-by: Philipp Stanner <pstanner@redhat.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241105143137.71893-2-pstanner@redhat.com
drm_sched_start()'s and drm_sched_stop()'s names suggest that those
functions might be intended for actively starting and stopping the
scheduler on initialization and teardown.
They are, however, only used on timeout handling (reset recovery). The
docstrings should reflect that to prevent confusion.
Document those functions' purpose.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Philipp Stanner <pstanner@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241029133819.78696-2-pstanner@redhat.com
drm_sched_job_init()'s name suggests that after the function succeeded,
parameter "job" will be fully initialized. This is not the case; some
members are only later set, notably drm_sched_job.sched by
drm_sched_job_arm().
Document that drm_sched_job_init() does not set all struct members.
Document the lifetime of drm_sched_job.sched.
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Philipp Stanner <pstanner@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241023141530.113370-2-pstanner@redhat.com
drm_sched_job_init() has no control over how users allocate struct
drm_sched_job. Unfortunately, the function can also not set some struct
members such as job->sched.
This could theoretically lead to UB by users dereferencing the struct's
pointer members too early.
It is easier to debug such issues if these pointers are initialized to
NULL, so dereferencing them causes a NULL pointer exception.
Accordingly, drm_sched_entity_init() does precisely that and initializes
its struct with memset().
Initialize parameter "job" to 0 in drm_sched_job_init().
Signed-off-by: Philipp Stanner <pstanner@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241021105028.19794-2-pstanner@redhat.com
Reviewed-by: Christian König <christian.koenig@amd.com>
Having removed one re-lock cycle on the entity->lock in a patch titled
"drm/sched: Optimise drm_sched_entity_push_job", with only a tiny bit
larger refactoring we can do the same optimisation on the rq->lock.
(Currently both drm_sched_rq_add_entity() and
drm_sched_rq_update_fifo_locked() take and release the same lock.)
To achieve this we make drm_sched_rq_update_fifo_locked() and
drm_sched_rq_add_entity() expect the rq->lock to be held.
We also align drm_sched_rq_update_fifo_locked(),
drm_sched_rq_add_entity() and
drm_sched_rq_remove_fifo_locked() function signatures, by adding rq as a
parameter to the latter.
v2:
* Fix after rebase of the series.
* Avoid naming inconsistency between drm_sched_rq_add/remove. (Christian)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Luben Tuikov <ltuikov89@gmail.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Philipp Stanner <pstanner@redhat.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Philipp Stanner <pstanner@redhat.com>
Signed-off-by: Philipp Stanner <pstanner@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241016122013.7857-6-tursulin@igalia.com
drm-misc-next for v6.12:
UAPI Changes:
- Add panthor/DEV_QUERY_TIMESTAMP_INFO query.
Cross-subsystem Changes:
- Updated dt bindings.
- Add documentation explaining default errnos for fences.
- Mark dma-buf heaps creation functions as __init.
Core Changes:
- Split DSC helpers from DP helpers.
- Clang build fixes for drm/mm test.
- Remove simple pipeline support for gem-vram,
no longer any users left after converting bochs.
- Add erno to drm_sched_start to distinguish between GPU and queue
reset.
- Add drm_framebuffer testcases.
- Fix uninitialized spinlock acquisition with CONFIG_DRM_PANIC=n.
- Use read_trylock instead of read_lock in dma_fence_begin_signalling to
quiesce lockdep.
Driver Changes:
- Assorted small fixes and updates for tegra, host1x, imagination,
nouveau, panfrost, panthor, panel/ili9341, mali, exynos,
panel/samsung-s6e3fa7, ast, bridge/ti-sn65dsi86, panel/himax-hx83112a,
bridge/tc358767, bridge/imx8mp-hdmi-tx, panel/khadas-ts050,
panel/nt36523, panel/sony-acx565akm, kmb, accel/qaic, omap, v3d.
- Add bridge/TI TDP158.
- Assorted documentation updates.
- Convert bochs from simple drm to gem shmem, and check modes
against available memory.
- Many VC4 fixes, most related to scaling and YUV support.
- Convert some drivers to use SYSTEM_SLEEP_PM_OPS and RUNTIME_PM_OPS.
- Rockchip 4k@60 support.
Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/445713a6-2427-4c53-8ec2-3a894ec62405@linux.intel.com
The current implementation of drm_sched_start uses a hardcoded
-ECANCELED to dispose of a job when the parent/hw fence is NULL.
This results in drm_sched_job_done being called with -ECANCELED for
each job with a NULL parent in the pending list, making it difficult
to distinguish between recovery methods, whether a queue reset or a
full GPU reset was used.
To improve this, we first try a soft recovery for timeout jobs and
use the error code -ENODATA. If soft recovery fails, we proceed with
a queue reset, where the error code remains -ENODATA for the job.
Finally, for a full GPU reset, we use error codes -ECANCELED or
-ETIME. This patch adds an error code parameter to drm_sched_start,
allowing us to differentiate between queue reset and GPU reset
failures. This enables user mode and test applications to validate
the expected correctness of the requested operation. After a
successful queue reset, the only way to continue normal operation is
to call drm_sched_job_done with the specific error code -ENODATA.
v1: Initial implementation by Jesse utilized amdgpu_device_lock_reset_domain
and amdgpu_device_unlock_reset_domain to allow user mode to track
the queue reset status and distinguish between queue reset and
GPU reset.
v2: Christian suggested using the error codes -ENODATA for queue reset
and -ECANCELED or -ETIME for GPU reset, returned to
amdgpu_cs_wait_ioctl.
v3: To meet the requirements, we introduce a new function
drm_sched_start_ex with an additional parameter to set
dma_fence_set_error, allowing us to handle the specific error
codes appropriately and dispose of bad jobs with the selected
error code depending on whether it was a queue reset or GPU reset.
v4: Alex suggested using a new name, drm_sched_start_with_recovery_error,
which more accurately describes the function's purpose.
Additionally, it was recommended to add documentation details
about the new method.
v5: Fixed declaration of new function drm_sched_start_with_recovery_error.(Alex)
v6 (chk): rebase on upstream changes, cleanup the commit message,
drop the new function again and update all callers,
apply the errno also to scheduler fences with hw fences
v7 (chk): rebased
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
Signed-off-by: Vitaly Prosyak <vitaly.prosyak@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240826122541.85663-1-christian.koenig@amd.com
Thomas Zimmermann asked to backmerge -rc6 for drm-misc branches,
there's a few same-area-changed conflicts (xe and amdgpu mostly) that
are getting a bit too annoying.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Currently, job flow control is implemented simply by limiting the number
of jobs in flight. Therefore, a scheduler is initialized with a credit
limit that corresponds to the number of jobs which can be sent to the
hardware.
This implies that for each job, drivers need to account for the maximum
job size possible in order to not overflow the ring buffer.
However, there are drivers, such as Nouveau, where the job size has a
rather large range. For such drivers it can easily happen that job
submissions not even filling the ring by 1% can block subsequent
submissions, which, in the worst case, can lead to the ring run dry.
In order to overcome this issue, allow for tracking the actual job size
instead of the number of jobs. Therefore, add a field to track a job's
credit count, which represents the number of credits a job contributes
to the scheduler's credit limit.
Signed-off-by: Danilo Krummrich <dakr@redhat.com>
Reviewed-by: Luben Tuikov <ltuikov89@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231110001638.71750-1-dakr@redhat.com
