diff options
| author | Karol Wachowski <karol.wachowski@linux.intel.com> | 2023-03-31 13:36:02 +0200 |
|---|---|---|
| committer | Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com> | 2023-04-05 09:06:38 +0200 |
| commit | 774e7cb50359eff7e966f4e2e80c9486014d3a23 (patch) | |
| tree | d071f22d288e3717a50465a920724aa93ed37199 | |
| parent | 764a2ab9eb56e1200083e771aab16186836edf1d (diff) | |
| download | nvme-774e7cb50359eff7e966f4e2e80c9486014d3a23.tar.gz | |
accel/ivpu: Add dma fence to command buffers only
Currently job->done_fence is added to every BO handle within a job. If job
handle (command buffer) is shared between multiple submits, KMD will add
the fence in each of them. Then bo_wait_ioctl() executed on command buffer
will exit only when all jobs containing that handle are done.
This creates deadlock scenario for user mode driver in case when job handle
is added as dependency of another job, because bo_wait_ioctl() of first job
will wait until second job finishes, and second job can not finish before
first one.
Having fences added only to job buffer handle allows user space to execute
bo_wait_ioctl() on the job even if it's handle is submitted with other job.
Fixes: cd7272215c44 ("accel/ivpu: Add command buffer submission logic")
Signed-off-by: Karol Wachowski <karol.wachowski@linux.intel.com>
Signed-off-by: Stanislaw Gruszka <stanislaw.gruszka@linux.intel.com>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230331113603.2802515-2-stanislaw.gruszka@linux.intel.com
| -rw-r--r-- | drivers/accel/ivpu/ivpu_job.c | 18 |
1 files changed, 7 insertions, 11 deletions
diff --git a/drivers/accel/ivpu/ivpu_job.c b/drivers/accel/ivpu/ivpu_job.c index 910301fae4351..3c6f1e16cf2ff 100644 --- a/drivers/accel/ivpu/ivpu_job.c +++ b/drivers/accel/ivpu/ivpu_job.c @@ -461,26 +461,22 @@ ivpu_job_prepare_bos_for_submit(struct drm_file *file, struct ivpu_job *job, u32 job->cmd_buf_vpu_addr = bo->vpu_addr + commands_offset; - ret = drm_gem_lock_reservations((struct drm_gem_object **)job->bos, buf_count, - &acquire_ctx); + ret = drm_gem_lock_reservations((struct drm_gem_object **)job->bos, 1, &acquire_ctx); if (ret) { ivpu_warn(vdev, "Failed to lock reservations: %d\n", ret); return ret; } - for (i = 0; i < buf_count; i++) { - ret = dma_resv_reserve_fences(job->bos[i]->base.resv, 1); - if (ret) { - ivpu_warn(vdev, "Failed to reserve fences: %d\n", ret); - goto unlock_reservations; - } + ret = dma_resv_reserve_fences(bo->base.resv, 1); + if (ret) { + ivpu_warn(vdev, "Failed to reserve fences: %d\n", ret); + goto unlock_reservations; } - for (i = 0; i < buf_count; i++) - dma_resv_add_fence(job->bos[i]->base.resv, job->done_fence, DMA_RESV_USAGE_WRITE); + dma_resv_add_fence(bo->base.resv, job->done_fence, DMA_RESV_USAGE_WRITE); unlock_reservations: - drm_gem_unlock_reservations((struct drm_gem_object **)job->bos, buf_count, &acquire_ctx); + drm_gem_unlock_reservations((struct drm_gem_object **)job->bos, 1, &acquire_ctx); wmb(); /* Flush write combining buffers */ |