drm/amdgpu: add VM generation token

Instead of using the VRAM lost counter add a 64bit token which indicates
if a context or job is still valid to use.

Should the VRAM be lost or the page tables need re-creation the token will
change indicating that userspace needs to act and re-create the contexts
and re-submit the work.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Luben Tuikov <luben.tuikov@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c

@@ -309,7 +309,7 @@ static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
 	}
 	p->gang_leader = p->jobs[p->gang_leader_idx];
 
-	if (p->ctx->vram_lost_counter != p->gang_leader->vram_lost_counter) {
+	if (p->ctx->generation != p->gang_leader->generation) {
 		ret = -ECANCELED;
 		goto free_all_kdata;
 	}

drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c

@@ -333,7 +333,7 @@ static int amdgpu_ctx_init(struct amdgpu_ctx_mgr *mgr, int32_t priority,
 
 	ctx->reset_counter = atomic_read(&mgr->adev->gpu_reset_counter);
 	ctx->reset_counter_query = ctx->reset_counter;
-	ctx->vram_lost_counter = atomic_read(&mgr->adev->vram_lost_counter);
+	ctx->generation = amdgpu_vm_generation(mgr->adev, &fpriv->vm);
 	ctx->init_priority = priority;
 	ctx->override_priority = AMDGPU_CTX_PRIORITY_UNSET;
 
@@ -586,7 +586,7 @@ static int amdgpu_ctx_query2(struct amdgpu_device *adev,
 	if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter))
 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET;
 
-	if (ctx->vram_lost_counter != atomic_read(&adev->vram_lost_counter))
+	if (ctx->generation != amdgpu_vm_generation(adev, &fpriv->vm))
 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST;
 
 	if (atomic_read(&ctx->guilty))

drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h

@@ -47,7 +47,7 @@ struct amdgpu_ctx {
 	struct amdgpu_ctx_mgr		*mgr;
 	unsigned			reset_counter;
 	unsigned			reset_counter_query;
-	uint32_t			vram_lost_counter;
+	uint64_t			generation;
 	spinlock_t			ring_lock;
 	struct amdgpu_ctx_entity	*entities[AMDGPU_HW_IP_NUM][AMDGPU_MAX_ENTITY_NUM];
 	bool				preamble_presented;

drivers/gpu/drm/amd/amdgpu/amdgpu_job.c

@@ -109,7 +109,7 @@ int amdgpu_job_alloc(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	(*job)->vm = vm;
 
 	amdgpu_sync_create(&(*job)->explicit_sync);
-	(*job)->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
+	(*job)->generation = amdgpu_vm_generation(adev, vm);
 	(*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
 
 	if (!entity)
@@ -295,7 +295,7 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
 	trace_amdgpu_sched_run_job(job);
 
 	/* Skip job if VRAM is lost and never resubmit gangs */
-	if (job->vram_lost_counter != atomic_read(&adev->vram_lost_counter) ||
+	if (job->generation != amdgpu_vm_generation(adev, job->vm) ||
 	    (job->job_run_counter && job->gang_submit))
 		dma_fence_set_error(finished, -ECANCELED);
 

drivers/gpu/drm/amd/amdgpu/amdgpu_job.h

@@ -61,7 +61,7 @@ struct amdgpu_job {
 	uint32_t		gds_base, gds_size;
 	uint32_t		gws_base, gws_size;
 	uint32_t		oa_base, oa_size;
-	uint32_t		vram_lost_counter;
+	uint64_t		generation;
 
 	/* user fence handling */
 	uint64_t		uf_addr;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

@@ -405,6 +405,30 @@ static void amdgpu_vm_fini_entities(struct amdgpu_vm *vm)
 	drm_sched_entity_destroy(&vm->delayed);
 }
 
+/**
+ * amdgpu_vm_generation - return the page table re-generation counter
+ * @adev: the amdgpu_device
+ * @vm: optional VM to check, might be NULL
+ *
+ * Returns a page table re-generation token to allow checking if submissions
+ * are still valid to use this VM. The VM parameter might be NULL in which case
+ * just the VRAM lost counter will be used.
+ */
+uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm)
+{
+	uint64_t result = (u64)atomic_read(&adev->vram_lost_counter) << 32;
+
+	if (!vm)
+		return result;
+
+	result += vm->generation;
+	/* Add one if the page tables will be re-generated on next CS */
+	if (drm_sched_entity_error(&vm->delayed))
+		++result;
+
+	return result;
+}
+
 /**
  * amdgpu_vm_validate_pt_bos - validate the page table BOs
  *
@@ -428,6 +452,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	int r;
 
 	if (drm_sched_entity_error(&vm->delayed)) {
+		++vm->generation;
 		amdgpu_vm_bo_reset_state_machine(vm);
 		amdgpu_vm_fini_entities(vm);
 		r = amdgpu_vm_init_entities(adev, vm);
@@ -2134,6 +2159,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	vm->last_update = dma_fence_get_stub();
 	vm->last_unlocked = dma_fence_get_stub();
 	vm->last_tlb_flush = dma_fence_get_stub();
+	vm->generation = 0;
 
 	mutex_init(&vm->eviction_lock);
 	vm->evicting = false;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h

@@ -295,6 +295,9 @@ struct amdgpu_vm {
 	atomic64_t		tlb_seq;
 	struct dma_fence	*last_tlb_flush;
 
+	/* How many times we had to re-generate the page tables */
+	uint64_t		generation;
+
 	/* Last unlocked submission to the scheduler entities */
 	struct dma_fence	*last_unlocked;
 
@@ -397,6 +400,7 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
 			 struct list_head *validated,
 			 struct amdgpu_bo_list_entry *entry);
 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
+uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 			      int (*callback)(void *p, struct amdgpu_bo *bo),
 			      void *param);