]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
drm/gpu-sched: fix force APP kill hang(v4)
authorEmily Deng <Emily.Deng@amd.com>
Mon, 16 Apr 2018 02:07:02 +0000 (10:07 +0800)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 15 May 2018 18:43:17 +0000 (13:43 -0500)
issue:
there are VMC page fault occurred if force APP kill during
3dmark test, the cause is in entity_fini we manually signal
all those jobs in entity's queue which confuse the sync/dep
mechanism:

1)page fault occurred in sdma's clear job which operate on
shadow buffer, and shadow buffer's Gart table is cleaned by
ttm_bo_release since the fence in its reservation was fake signaled
by entity_fini() under the case of SIGKILL received.

2)page fault occurred in gfx' job because during the lifetime
of gfx job we manually fake signal all jobs from its entity
in entity_fini(), thus the unmapping/clear PTE job depend on those
result fence is satisfied and sdma start clearing the PTE and lead
to GFX page fault.

fix:
1)should at least wait all jobs already scheduled complete in entity_fini()
if SIGKILL is the case.

2)if a fence signaled and try to clear some entity's dependency, should
set this entity guilty to prevent its job really run since the dependency
is fake signaled.

v2:
splitting drm_sched_entity_fini() into two functions:
1)The first one is does the waiting, removes the entity from the
runqueue and returns an error when the process was killed.
2)The second one then goes over the entity, install it as
completion signal for the remaining jobs and signals all jobs
with an error code.

v3:
1)Replace the fini1 and fini2 with better name
2)Call the first part before the VM teardown in
amdgpu_driver_postclose_kms() and the second part
after the VM teardown
3)Keep the original function drm_sched_entity_fini to
refine the code.

v4:
1)Rename entity->finished to entity->last_scheduled;
2)Rename drm_sched_entity_fini_job_cb() to
drm_sched_entity_kill_jobs_cb();
3)Pass NULL to drm_sched_entity_fini_job_cb() if -ENOENT;
4)Replace the type of entity->fini_status with "int";
5)Remove the check about entity->finished.

Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Signed-off-by: Emily Deng <Emily.Deng@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu.h
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
drivers/gpu/drm/scheduler/gpu_scheduler.c
include/drm/gpu_scheduler.h

index c25ee750c362b6e59088990dd0af4b2a684d0eb9..ea1b28536bfcb2d4e2c29793ca54392197c309bf 100644 (file)
@@ -681,6 +681,8 @@ int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx, unsigned ring_id);
 
 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr);
+void amdgpu_ctx_mgr_entity_cleanup(struct amdgpu_ctx_mgr *mgr);
+void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr);
 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr);
 
 
index 09d35051fdd68689ac00d69504765166e738076f..eb80edfb1b0a7a33e16f4dbc2db5a4795727b9d7 100644 (file)
@@ -111,8 +111,9 @@ failed:
        return r;
 }
 
-static void amdgpu_ctx_fini(struct amdgpu_ctx *ctx)
+static void amdgpu_ctx_fini(struct kref *ref)
 {
+       struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount);
        struct amdgpu_device *adev = ctx->adev;
        unsigned i, j;
 
@@ -125,13 +126,11 @@ static void amdgpu_ctx_fini(struct amdgpu_ctx *ctx)
        kfree(ctx->fences);
        ctx->fences = NULL;
 
-       for (i = 0; i < adev->num_rings; i++)
-               drm_sched_entity_fini(&adev->rings[i]->sched,
-                                     &ctx->rings[i].entity);
-
        amdgpu_queue_mgr_fini(adev, &ctx->queue_mgr);
 
        mutex_destroy(&ctx->lock);
+
+       kfree(ctx);
 }
 
 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
@@ -170,12 +169,15 @@ static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
 static void amdgpu_ctx_do_release(struct kref *ref)
 {
        struct amdgpu_ctx *ctx;
+       u32 i;
 
        ctx = container_of(ref, struct amdgpu_ctx, refcount);
 
-       amdgpu_ctx_fini(ctx);
+       for (i = 0; i < ctx->adev->num_rings; i++)
+               drm_sched_entity_fini(&ctx->adev->rings[i]->sched,
+                       &ctx->rings[i].entity);
 
-       kfree(ctx);
+       amdgpu_ctx_fini(ref);
 }
 
 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id)
@@ -435,16 +437,62 @@ void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
        idr_init(&mgr->ctx_handles);
 }
 
+void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
+{
+       struct amdgpu_ctx *ctx;
+       struct idr *idp;
+       uint32_t id, i;
+
+       idp = &mgr->ctx_handles;
+
+       idr_for_each_entry(idp, ctx, id) {
+
+               if (!ctx->adev)
+                       return;
+
+               for (i = 0; i < ctx->adev->num_rings; i++)
+                       if (kref_read(&ctx->refcount) == 1)
+                               drm_sched_entity_do_release(&ctx->adev->rings[i]->sched,
+                                                 &ctx->rings[i].entity);
+                       else
+                               DRM_ERROR("ctx %p is still alive\n", ctx);
+       }
+}
+
+void amdgpu_ctx_mgr_entity_cleanup(struct amdgpu_ctx_mgr *mgr)
+{
+       struct amdgpu_ctx *ctx;
+       struct idr *idp;
+       uint32_t id, i;
+
+       idp = &mgr->ctx_handles;
+
+       idr_for_each_entry(idp, ctx, id) {
+
+               if (!ctx->adev)
+                       return;
+
+               for (i = 0; i < ctx->adev->num_rings; i++)
+                       if (kref_read(&ctx->refcount) == 1)
+                               drm_sched_entity_cleanup(&ctx->adev->rings[i]->sched,
+                                       &ctx->rings[i].entity);
+                       else
+                               DRM_ERROR("ctx %p is still alive\n", ctx);
+       }
+}
+
 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
 {
        struct amdgpu_ctx *ctx;
        struct idr *idp;
        uint32_t id;
 
+       amdgpu_ctx_mgr_entity_cleanup(mgr);
+
        idp = &mgr->ctx_handles;
 
        idr_for_each_entry(idp, ctx, id) {
-               if (kref_put(&ctx->refcount, amdgpu_ctx_do_release) != 1)
+               if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1)
                        DRM_ERROR("ctx %p is still alive\n", ctx);
        }
 
index bd9e723dbb2b62f8c8bad4ec1add7f82a1425880..1ed379524117f33d18426bd58eb3b059e60686ee 100644 (file)
@@ -913,8 +913,7 @@ void amdgpu_driver_postclose_kms(struct drm_device *dev,
                return;
 
        pm_runtime_get_sync(dev->dev);
-
-       amdgpu_ctx_mgr_fini(&fpriv->ctx_mgr);
+       amdgpu_ctx_mgr_entity_fini(&fpriv->ctx_mgr);
 
        if (adev->asic_type != CHIP_RAVEN) {
                amdgpu_uvd_free_handles(adev, file_priv);
@@ -935,6 +934,8 @@ void amdgpu_driver_postclose_kms(struct drm_device *dev,
        pd = amdgpu_bo_ref(fpriv->vm.root.base.bo);
 
        amdgpu_vm_fini(adev, &fpriv->vm);
+       amdgpu_ctx_mgr_fini(&fpriv->ctx_mgr);
+
        if (pasid)
                amdgpu_pasid_free_delayed(pd->tbo.resv, pasid);
        amdgpu_bo_unref(&pd);
index 310275eaf128d32bda2994d07bd86df49f900608..44d21981bf3bc123e133de4e91883ca32fe37558 100644 (file)
@@ -136,6 +136,8 @@ int drm_sched_entity_init(struct drm_gpu_scheduler *sched,
        entity->rq = rq;
        entity->sched = sched;
        entity->guilty = guilty;
+       entity->fini_status = 0;
+       entity->last_scheduled = NULL;
 
        spin_lock_init(&entity->rq_lock);
        spin_lock_init(&entity->queue_lock);
@@ -197,19 +199,30 @@ static bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
        return true;
 }
 
+static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
+                                   struct dma_fence_cb *cb)
+{
+       struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
+                                                finish_cb);
+       drm_sched_fence_finished(job->s_fence);
+       WARN_ON(job->s_fence->parent);
+       dma_fence_put(&job->s_fence->finished);
+       job->sched->ops->free_job(job);
+}
+
+
 /**
  * Destroy a context entity
  *
  * @sched       Pointer to scheduler instance
  * @entity     The pointer to a valid scheduler entity
  *
- * Cleanup and free the allocated resources.
+ * Splitting drm_sched_entity_fini() into two functions, The first one is does the waiting,
+ * removes the entity from the runqueue and returns an error when the process was killed.
  */
-void drm_sched_entity_fini(struct drm_gpu_scheduler *sched,
+void drm_sched_entity_do_release(struct drm_gpu_scheduler *sched,
                           struct drm_sched_entity *entity)
 {
-       int r;
-
        if (!drm_sched_entity_is_initialized(sched, entity))
                return;
        /**
@@ -217,13 +230,28 @@ void drm_sched_entity_fini(struct drm_gpu_scheduler *sched,
         * queued IBs or discard them on SIGKILL
        */
        if ((current->flags & PF_SIGNALED) && current->exit_code == SIGKILL)
-               r = -ERESTARTSYS;
+               entity->fini_status = -ERESTARTSYS;
        else
-               r = wait_event_killable(sched->job_scheduled,
+               entity->fini_status = wait_event_killable(sched->job_scheduled,
                                        drm_sched_entity_is_idle(entity));
        drm_sched_entity_set_rq(entity, NULL);
-       if (r) {
+}
+EXPORT_SYMBOL(drm_sched_entity_do_release);
+
+/**
+ * Destroy a context entity
+ *
+ * @sched       Pointer to scheduler instance
+ * @entity     The pointer to a valid scheduler entity
+ *
+ * The second one then goes over the entity and signals all jobs with an error code.
+ */
+void drm_sched_entity_cleanup(struct drm_gpu_scheduler *sched,
+                          struct drm_sched_entity *entity)
+{
+       if (entity->fini_status) {
                struct drm_sched_job *job;
+               int r;
 
                /* Park the kernel for a moment to make sure it isn't processing
                 * our enity.
@@ -241,13 +269,26 @@ void drm_sched_entity_fini(struct drm_gpu_scheduler *sched,
                        struct drm_sched_fence *s_fence = job->s_fence;
                        drm_sched_fence_scheduled(s_fence);
                        dma_fence_set_error(&s_fence->finished, -ESRCH);
-                       drm_sched_fence_finished(s_fence);
-                       WARN_ON(s_fence->parent);
-                       dma_fence_put(&s_fence->finished);
-                       sched->ops->free_job(job);
+                       r = dma_fence_add_callback(entity->last_scheduled, &job->finish_cb,
+                                                       drm_sched_entity_kill_jobs_cb);
+                       if (r == -ENOENT)
+                               drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
+                       else if (r)
+                               DRM_ERROR("fence add callback failed (%d)\n", r);
                }
+
+               dma_fence_put(entity->last_scheduled);
+               entity->last_scheduled = NULL;
        }
 }
+EXPORT_SYMBOL(drm_sched_entity_cleanup);
+
+void drm_sched_entity_fini(struct drm_gpu_scheduler *sched,
+                               struct drm_sched_entity *entity)
+{
+       drm_sched_entity_do_release(sched, entity);
+       drm_sched_entity_cleanup(sched, entity);
+}
 EXPORT_SYMBOL(drm_sched_entity_fini);
 
 static void drm_sched_entity_wakeup(struct dma_fence *f, struct dma_fence_cb *cb)
@@ -530,6 +571,10 @@ void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
                spin_unlock(&sched->job_list_lock);
                fence = sched->ops->run_job(s_job);
                atomic_inc(&sched->hw_rq_count);
+
+               dma_fence_put(s_job->entity->last_scheduled);
+               s_job->entity->last_scheduled = dma_fence_get(&s_fence->finished);
+
                if (fence) {
                        s_fence->parent = dma_fence_get(fence);
                        r = dma_fence_add_callback(fence, &s_fence->cb,
@@ -556,6 +601,7 @@ int drm_sched_job_init(struct drm_sched_job *job,
                       void *owner)
 {
        job->sched = sched;
+       job->entity = entity;
        job->s_priority = entity->rq - sched->sched_rq;
        job->s_fence = drm_sched_fence_create(entity, owner);
        if (!job->s_fence)
@@ -669,6 +715,9 @@ static int drm_sched_main(void *param)
                fence = sched->ops->run_job(sched_job);
                drm_sched_fence_scheduled(s_fence);
 
+               dma_fence_put(entity->last_scheduled);
+               entity->last_scheduled = dma_fence_get(&s_fence->finished);
+
                if (fence) {
                        s_fence->parent = dma_fence_get(fence);
                        r = dma_fence_add_callback(fence, &s_fence->cb,
index c053a32341bfe6156415f424e45516d4eb6881d6..350a62c26b296f628ec1463800c01c853b3af2e5 100644 (file)
@@ -65,6 +65,8 @@ struct drm_sched_entity {
        struct dma_fence                *dependency;
        struct dma_fence_cb             cb;
        atomic_t                        *guilty; /* points to ctx's guilty */
+       int            fini_status;
+       struct dma_fence    *last_scheduled;
 };
 
 /**
@@ -119,6 +121,7 @@ struct drm_sched_job {
        uint64_t                        id;
        atomic_t                        karma;
        enum drm_sched_priority         s_priority;
+       struct drm_sched_entity  *entity;
 };
 
 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
@@ -186,6 +189,10 @@ int drm_sched_entity_init(struct drm_gpu_scheduler *sched,
                          struct drm_sched_entity *entity,
                          struct drm_sched_rq *rq,
                          uint32_t jobs, atomic_t *guilty);
+void drm_sched_entity_do_release(struct drm_gpu_scheduler *sched,
+                          struct drm_sched_entity *entity);
+void drm_sched_entity_cleanup(struct drm_gpu_scheduler *sched,
+                          struct drm_sched_entity *entity);
 void drm_sched_entity_fini(struct drm_gpu_scheduler *sched,
                           struct drm_sched_entity *entity);
 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,