This series addresses several issues in the v3d scheduler and submission code. Most of the fixes were motivated by feedback in the vc4 scheduler series [1], which inherited issues from v3d. Based on the issues found there, this series addresses the issues in the v3d driver as well.
This series has cleanup patches, fixes, and finally, a refactoring of the submission code, which allowed us to fix the atomicity of a submission. - Cleanups and small improvements: - PATCH 1/12: "drm/v3d: Drop unused drm_encoder.h include from v3d_drv.h" - Fixes: - PATCH 2/12: "drm/v3d: Clear queue->active_job when v3d_fence_create() fails" - PATCH 3/12: "drm/v3d: Extract v3d_job_add_syncobjs() helper" - PATCH 4/12: "drm/v3d: Reject invalid syncobj handles in submit ioctls" - PATCH 11/12: "drm/v3d: Reject invalid out_sync handles in submit ioctls" - Submission refactoring: - PATCH 5/12: "drm/v3d: Migrate BO reservation locking to DRM exec" - PATCH 6/12: "drm/v3d: Introduce struct v3d_submit and convert CL/TFU/CSD ioctls" - PATCH 7/12: "drm/v3d: Make v3d_get_cpu_indirect_csd_params() a pure parser" - PATCH 8/12: "drm/v3d: Convert submit helpers to operate on struct v3d_submit" - PATCH 9/12: "drm/v3d: Refactor CPU ioctl into unified submission chain" - PATCH 10/12: "drm/v3d: Split BO fence attach from syncobj output handling" - PATCH 12/12: "drm/v3d: Ensure atomic submissions in v3d_submit_jobs()" (Also a fix) During the refactoring, I tried to break the conversion into small steps. On the one hand, this helps reviewability; on the other hand, some things may look unfinished at the end of a given commit. I tried to balance the trade-off, but I'm happy to take split or squash requests during review. [1] https://lore.kernel.org/dri-devel/[email protected]/T/ Best regards, - Maíra --- v1 -> v2: https://lore.kernel.org/r/[email protected] - [2/14, 3/14, 5/14, 6/14] Add Tvrtko's R-b tag (Tvrtko Ursulin) - [4/14] Use spin_lock() instead of spin_lock_irq() and change the commit message (Tvrtko Ursulin) - [4/14] Use scoped_guard() instead of open-coding spin_(un)lock() - [6/14] s/NULL/zero in the comments (Tvrtko Ursulin) - [8/14] s/kcalloc(1, size, GFP_KERNEL)/kzalloc (Tvrtko Ursulin) - [8/14] Save one atomic in v3d_attach_perfmon_to_jobs() by adding a conditional in the loop (Tvrtko Ursulin) - [8/14] Zero is implied when initializing the struct v3d_submit (Tvrtko Ursulin) - [8/14] Create v3d_submit_put_jobs() and v3d_submit_cleanup_jobs() (Tvrtko Ursulin) - [9/14, 10/14, 11/14] NEW PATCHES: Trying to make the original patch "[PATCH 09/10] drm/v3d: Refactor CPU ioctl into unified submission chain" a bit easier to review. - [10/14] Wrap drm_exec_fini() in a function v3d_submit_unlock_reservations() (Tvrtko Ursulin) - [12/14] NEW PATCH: "drm/v3d: Split BO fence attach from syncobj output handling" - Split v3d_attach_fences_and_unlock_reservation() into different functions and fold parts of it into v3d_submit_jobs (Tvrtko Ursulin) - [13/14] NEW PATCH: "drm/v3d: Reject invalid out_sync handles in submit ioctls" - Stop silently ignoring invalid syncobj handles (Tvrtko Ursulin) - [14/14] Move the error path to a goto - [14/14] Skip fence attachment and syncobj exportation to user-space if the submission has failed (Tvrtko Ursulin) v2 -> v3: https://lore.kernel.org/r/[email protected] - Rebased on top of drm-misc-next. - [1/14] Add Iago's R-b tag (Iago Toral) - [4/14, 7/14, 9/14] Add Tvrtko's R-b tag (Tvrtko Ursulin) - [5/14] Create a job->queue variable to be used by v3d_job_add_syncobjs() - [8/14] v3d_submit_add_job() returns struct v3d_job * instead of using void **container as argument (Tvrtko Ursulin) - [10/14] Pull v3d_submit_lock_reservations() out of v3d_setup_csd_jobs_and_bos() (Tvrtko Ursulin) - [11/14] Assert that jobs[1] is actually a CSD job (Tvrtko Ursulin) - [11/14] v3d_submit_lock_reservations() is now called unconditionally in v3d_submit_cpu_ioctl() (Tvrtko Ursulin) - [12/14, 13/14] Fold v3d_submit_process_post_deps() and v3d_submit_put_jobs() into v3d_submit_jobs() (Tvrtko Ursulin) v3 -> v4: https://lore.kernel.org/r/[email protected] - Dropped "[PATCH v3 03/14] drm/v3d: Use inline lock for dma fence initialization" and "[PATCH v3 04/14] drm/v3d: Replace spin_lock_irqsave() with spin_lock()". - As pointed out by Sashiko, decoupling the DMA fence lock from the queue lock would break implicit synchronization, causing a UAF of `bin_job` in `v3d_overflow_mem_work()`. - I'll need to rework this in the future. - [6/12, 8/12, 9/12, 10/12, 11/12] Add Tvrtko's R-b - [2/12] Propagate the error pointer when v3d_fence_create() fails instead of treating the job as successfully completed (Sashiko) - [6/12] Don't initialize the variables: render, clean_job and tfu_job (Tvrtko Ursulin) - [11/12] Add an assert that sync_obj and multisync are mutually exclusive (Tvrtko Ursulin) - [11/12] Document that multisync is prioritized over sync_obj (Tvrtko Ursulin) - [12/12] Change the loop to for (i = 1; i < count; i++) (Tvrtko Ursulin) --- Maíra Canal (12): drm/v3d: Drop unused drm_encoder.h include from v3d_drv.h drm/v3d: Clear queue->active_job when v3d_fence_create() fails drm/v3d: Extract v3d_job_add_syncobjs() helper drm/v3d: Reject invalid syncobj handles in submit ioctls drm/v3d: Migrate BO reservation locking to DRM exec drm/v3d: Introduce struct v3d_submit and convert CL/TFU/CSD ioctls drm/v3d: Make v3d_get_cpu_indirect_csd_params() a pure parser drm/v3d: Convert submit helpers to operate on struct v3d_submit drm/v3d: Refactor CPU ioctl into unified submission chain drm/v3d: Split BO fence attach from syncobj output handling drm/v3d: Reject invalid out_sync handles in submit ioctls drm/v3d: Ensure atomic submissions in v3d_submit_jobs() drivers/gpu/drm/v3d/Kconfig | 1 + drivers/gpu/drm/v3d/v3d_drv.h | 35 +- drivers/gpu/drm/v3d/v3d_sched.c | 66 ++-- drivers/gpu/drm/v3d/v3d_submit.c | 770 +++++++++++++++++++-------------------- 4 files changed, 445 insertions(+), 427 deletions(-) --- base-commit: 86925ba9da20857b7d4525256a405fe789d4aa33 change-id: 20260407-v3d-sched-misc-fixes-623739017e53
