drm/amdgpu: fix context switch

Properly protect the state and also handle submission failures.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>

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

@@ -893,6 +893,7 @@ struct amdgpu_ib {
 	struct amdgpu_fence		*fence;
 	struct amdgpu_user_fence        *user;
 	struct amdgpu_vm		*vm;
+	struct amdgpu_ctx		*ctx;
 	struct amdgpu_sync		sync;
 	uint32_t			gds_base, gds_size;
 	uint32_t			gws_base, gws_size;
@@ -943,9 +944,7 @@ struct amdgpu_ring {
 	unsigned		wptr_offs;
 	unsigned		next_rptr_offs;
 	unsigned		fence_offs;
-	struct drm_file		*current_filp;
-	unsigned		current_ctx;
-	bool			need_ctx_switch;
+	struct amdgpu_ctx	*current_ctx;
 	enum amdgpu_ring_type	type;
 	char			name[16];
 };
@@ -1236,7 +1235,7 @@ struct amdgpu_cs_chunk {
 struct amdgpu_cs_parser {
 	struct amdgpu_device	*adev;
 	struct drm_file		*filp;
-	uint32_t ctx_id;
+	struct amdgpu_ctx	*ctx;
 	struct amdgpu_bo_list *bo_list;
 	/* chunks */
 	unsigned		nchunks;

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

@@ -138,7 +138,11 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 	if (!cs->in.num_chunks)
 		goto out;
 
-	p->ctx_id = cs->in.ctx_id;
+	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
+	if (!p->ctx) {
+		r = -EINVAL;
+		goto out;
+	}
 	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
 
 	/* get chunks */
@@ -445,6 +449,8 @@ static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bo
 					   &parser->validated);
 	}
 
+	if (parser->ctx)
+		amdgpu_ctx_put(parser->ctx);
 	if (parser->bo_list)
 		amdgpu_bo_list_put(parser->bo_list);
 	drm_free_large(parser->vm_bos);
@@ -639,13 +645,7 @@ static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 		ib->length_dw = chunk_ib->ib_bytes / 4;
 
 		ib->flags = chunk_ib->flags;
-
-		if ((ib->ring->current_filp != parser->filp) ||
-		    (ib->ring->current_ctx != parser->ctx_id)) {
-			ib->ring->need_ctx_switch = true;
-			ib->ring->current_ctx = parser->ctx_id;
-			ib->ring->current_filp = parser->filp;
-		}
+		ib->ctx = parser->ctx;
 
 		ib_bo = &parser->ib_bos[j];
 		ib_bo->robj = aobj;

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

@@ -140,6 +140,7 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 {
 	struct amdgpu_ib *ib = &ibs[0];
 	struct amdgpu_ring *ring;
+	struct amdgpu_ctx *ctx, *old_ctx;
 	struct amdgpu_vm *vm;
 	unsigned i;
 	int r = 0;
@@ -148,6 +149,7 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 		return -EINVAL;
 
 	ring = ibs->ring;
+	ctx = ibs->ctx;
 	vm = ibs->vm;
 
 	if (!ring->ready) {
@@ -189,19 +191,23 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 	if (ring->funcs->emit_hdp_flush)
 		amdgpu_ring_emit_hdp_flush(ring);
 
+	old_ctx = ring->current_ctx;
 	for (i = 0; i < num_ibs; ++i) {
 		ib = &ibs[i];
 
-		if (ib->ring != ring) {
+		if (ib->ring != ring || ib->ctx != ctx || ib->vm != vm) {
+			ring->current_ctx = old_ctx;
 			amdgpu_ring_unlock_undo(ring);
 			return -EINVAL;
 		}
 		amdgpu_ring_emit_ib(ring, ib);
+		ring->current_ctx = ctx;
 	}
 
 	r = amdgpu_fence_emit(ring, owner, &ib->fence);
 	if (r) {
 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
+		ring->current_ctx = old_ctx;
 		amdgpu_ring_unlock_undo(ring);
 		return r;
 	}

drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c

@@ -2516,19 +2516,20 @@ static bool gfx_v7_0_ring_emit_semaphore(struct amdgpu_ring *ring,
 static void gfx_v7_0_ring_emit_ib(struct amdgpu_ring *ring,
 				  struct amdgpu_ib *ib)
 {
+	bool need_ctx_switch = ring->current_ctx != ib->ctx;
 	u32 header, control = 0;
 	u32 next_rptr = ring->wptr + 5;
 
 	/* drop the CE preamble IB for the same context */
 	if ((ring->type == AMDGPU_RING_TYPE_GFX) &&
 	    (ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
-	    !ring->need_ctx_switch)
+	    !need_ctx_switch)
 		return;
 
 	if (ring->type == AMDGPU_RING_TYPE_COMPUTE)
 		control |= INDIRECT_BUFFER_VALID;
 
-	if (ring->need_ctx_switch && ring->type == AMDGPU_RING_TYPE_GFX)
+	if (need_ctx_switch && ring->type == AMDGPU_RING_TYPE_GFX)
 		next_rptr += 2;
 
 	next_rptr += 4;
@@ -2539,10 +2540,9 @@ static void gfx_v7_0_ring_emit_ib(struct amdgpu_ring *ring,
 	amdgpu_ring_write(ring, next_rptr);
 
 	/* insert SWITCH_BUFFER packet before first IB in the ring frame */
-	if (ring->need_ctx_switch && ring->type == AMDGPU_RING_TYPE_GFX) {
+	if (need_ctx_switch && ring->type == AMDGPU_RING_TYPE_GFX) {
 		amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
 		amdgpu_ring_write(ring, 0);
-		ring->need_ctx_switch = false;
 	}
 
 	if (ib->flags & AMDGPU_IB_FLAG_CE)

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

@@ -3645,19 +3645,20 @@ static void gfx_v8_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
 static void gfx_v8_0_ring_emit_ib(struct amdgpu_ring *ring,
 				  struct amdgpu_ib *ib)
 {
+	bool need_ctx_switch = ring->current_ctx != ib->ctx;
 	u32 header, control = 0;
 	u32 next_rptr = ring->wptr + 5;
 
 	/* drop the CE preamble IB for the same context */
 	if ((ring->type == AMDGPU_RING_TYPE_GFX) &&
 	    (ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
-	    !ring->need_ctx_switch)
+	    !need_ctx_switch)
 		return;
 
 	if (ring->type == AMDGPU_RING_TYPE_COMPUTE)
 		control |= INDIRECT_BUFFER_VALID;
 
-	if (ring->need_ctx_switch && ring->type == AMDGPU_RING_TYPE_GFX)
+	if (need_ctx_switch && ring->type == AMDGPU_RING_TYPE_GFX)
 		next_rptr += 2;
 
 	next_rptr += 4;
@@ -3668,10 +3669,9 @@ static void gfx_v8_0_ring_emit_ib(struct amdgpu_ring *ring,
 	amdgpu_ring_write(ring, next_rptr);
 
 	/* insert SWITCH_BUFFER packet before first IB in the ring frame */
-	if (ring->need_ctx_switch && ring->type == AMDGPU_RING_TYPE_GFX) {
+	if (need_ctx_switch && ring->type == AMDGPU_RING_TYPE_GFX) {
 		amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
 		amdgpu_ring_write(ring, 0);
-		ring->need_ctx_switch = false;
 	}
 
 	if (ib->flags & AMDGPU_IB_FLAG_CE)