kernel/linux.git/drivers/gpu/drm/amd/amdgpu/amdgpu_eviction_fence.h, branch linux-7.0.y

drm/amdgpu: simplify eviction fence suspend/resume

2025-04-08T20:48:19+00:00

The basic idea in this redesign is to add an eviction fence only in UQ resume path. When userqueue is not present, keep ev_fence as NULL Main changes are: - do not create the eviction fence during evf_mgr_init, keeping evf_mgr->ev_fence=NULL until UQ get active. - do not replace the ev_fence in evf_resume path, but replace it only in uq_resume path, so remove all the unnecessary code from ev_fence_resume. - add a new helper function (amdgpu_userqueue_ensure_ev_fence) which will do the following: - flush any pending uq_resume work, so that it could create an eviction_fence - if there is no pending uq_resume_work, add a uq_resume work and wait for it to execute so that we always have a valid ev_fence - call this helper function from two places, to ensure we have a valid ev_fence: - when a new uq is created - when a new uq completion fence is created v2: Worked on review comments by Christian. v3: Addressed few more review comments by Christian. v4: Move mutex lock outside of the amdgpu_userqueue_suspend() function (Christian). v5: squash in build fix (Alex) Cc: Alex Deucher Reviewed-by: Christian König Signed-off-by: Arvind Yadav Signed-off-by: Shashank Sharma Signed-off-by: Alex Deucher

drm/amdgpu: resume gfx userqueues

2025-04-08T20:48:18+00:00

This patch adds support for userqueue resume. What it typically does is this: - adds a new delayed work for resuming all the queues. - schedules this delayed work from the suspend work. - validates the BOs and replaces the eviction fence before resuming all the queues running under this instance of userq manager. V2: Addressed Christian's review comments: - declare local variables like ret at the bottom. - lock all the object first, then start attaching the new fence. - dont replace old eviction fence, just attach new eviction fence. - no error logs for drm_exec_lock failures - no need to reserve bos after drm_exec_locked - schedule the resume worker immediately (not after 100 ms) - check for NULL BO (Arvind) V5: Rebased wrt changes in suspend patch - moved amdgpu_userqueue_validate_vm_bo in this patch - initialized ret in resume_all V6: Rebase V7: Addressed review comments from Christian - Do not use list_for_each_safe() with vm->invalidated, its not correct way V8: Fixed the race condition between suspend/close/fence Cc: Alex Deucher Cc: Christian Koenig Acked-by: Christian Koenig Signed-off-by: Shashank Sharma Signed-off-by: Arvind Yadav Signed-off-by: Alex Deucher

drm/amdgpu: suspend gfx userqueues

2025-04-08T20:48:18+00:00

This patch adds suspend support for gfx userqueues. It typically does the following: - adds an enable_signaling function for the eviction fence, so that it can trigger the userqueue suspend, - adds a delayed work to handle suspending of the eviction_fence - adds a suspend function to handle suspending of userqueues which suspends all the queues under this userq manager and signals the eviction fence, - adds a function to replace the old eviction fence with a new one and attach it to each of the objects, - adds reference of userq manager in the eviction fence container so that it can be used in the suspend function. V2: Addressed Christian's review comments: - schedule suspend work immediately V4: Addressed Christian's review comments: - wait for pending uq fences before starting suspend, added queue->last_fence for the same - accommodate ev_fence_mgr into existing code - some bug fixes and NULL checks V5: Addressed Christian's review comments (gitlab) - Wait for eviction fence to get signaled in destroy, don't signal it - Wait for eviction fence to get signaled in replace fence, don't signal it V6: Addressed Christian's review comments - Do not destroy the old eviction fence until we have it replaced - Change the sequence of fence replacement sub-tasks - reusing the ev_fence delayed work for userqueue suspend as well (Shashank). V7: Addressed Christian's review comments - give evf_mgr as argument (instead of fpriv) to replace_fence() - save ptr to evf_mgr in ev_fence (instead of uq_mgr) - modify suspend_all_queues logic to reflect error properly - remove the garbage drm_exec_lock section in wait_for_signal - grab the userqueue mutex before starting the wait for fence - remove the unrelated gobj check from signal_ioctl V8: Added race condition fixes Cc: Alex Deucher Cc: Christian Koenig Acked-by: Christian Koenig Signed-off-by: Shashank Sharma Signed-off-by: Arvind Yadav Signed-off-by: Alex Deucher

drm/amdgpu: add gfx eviction fence helpers

2025-04-08T20:48:18+00:00

This patch adds basic eviction fence framework for the gfx buffers. The idea is to: - One eviction fence is created per gfx process, at kms_open. - This fence is attached to all the gem buffers created by this process. - This fence is detached to all the gem buffers at postclose_kms. This framework will be further used for usermode queues. V2: Addressed review comments from Christian - keep fence_ctx and fence_seq directly in fpriv - evcition_fence should be dynamically allocated - do not save eviction fence instance in BO, there could be many such fences attached to one BO - use dma_resv_replace_fence() in detach V3: Addressed review comments from Christian - eviction fence create and destroy functions should be called only once from fpriv create/destroy - use dma_fence_put() in eviction_fence_destroy V4: Addressed review comments from Christian: - create a separate ev_fence_mgr structure - cleanup fence init part - do not add a domain for fence owner KGD V5: Addressed review comments from Christian: - drop the dma_fence_is_signaled check - use a local variable to access evf_mgr->ev_fence under the spin_lock() multiple places - remove the vm->is_compute_ctx check to attach gfx eviction fence, in gem_object_open V6: Addressed review comments from Christian: - drop the return value from eviction_fence_signal - reserve_fence should be the first thing inside the attach_eviction_fence function, also keep the resv_add_fence inside the lock - remove the unwanted ev_fence check inside detach function - fix wrong variable check in eviction_fence_init function - return the error value of eviction_fence_init to the caller, dont keep it void. - fail gem_object_open if attaching of eviction_fence fails - detach the eviction fence only when amdgpu_vm_is_bo_always_valid is not true. V7: Addressed review comments from Christian: - Do not add a uq_mgr ptr in ev_fence, rather add evf_mgr V8: Move eviction fence enabling into separate patch for CI Cc: Christian Koenig Cc: Alex Deucher Reviewed-by: Christian Koenig Signed-off-by: Shashank Sharma Signed-off-by: Arvind Yadav Signed-off-by: Alex Deucher