1 files changed, 338 insertions, 269 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
index e104d7ef3c3d..1bbd39b3b0fc 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
@@ -64,11 +64,51 @@ static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p,
 	return 0;
 }
 
+static int amdgpu_cs_job_idx(struct amdgpu_cs_parser *p,
+			     struct drm_amdgpu_cs_chunk_ib *chunk_ib)
+{
+	struct drm_sched_entity *entity;
+	unsigned int i;
+	int r;
+
+	r = amdgpu_ctx_get_entity(p->ctx, chunk_ib->ip_type,
+				  chunk_ib->ip_instance,
+				  chunk_ib->ring, &entity);
+	if (r)
+		return r;
+
+	/*
+	 * Abort if there is no run queue associated with this entity.
+	 * Possibly because of disabled HW IP.
+	 */
+	if (entity->rq == NULL)
+		return -EINVAL;
+
+	/* Check if we can add this IB to some existing job */
+	for (i = 0; i < p->gang_size; ++i)
+		if (p->entities[i] == entity)
+			return i;
+
+	/* If not increase the gang size if possible */
+	if (i == AMDGPU_CS_GANG_SIZE)
+		return -EINVAL;
+
+	p->entities[i] = entity;
+	p->gang_size = i + 1;
+	return i;
+}
+
 static int amdgpu_cs_p1_ib(struct amdgpu_cs_parser *p,
 			   struct drm_amdgpu_cs_chunk_ib *chunk_ib,
 			   unsigned int *num_ibs)
 {
-	++(*num_ibs);
+	int r;
+
+	r = amdgpu_cs_job_idx(p, chunk_ib);
+	if (r < 0)
+		return r;
+
+	++(num_ibs[r]);
 	return 0;
 }
 
@@ -142,11 +182,12 @@ static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
 			   union drm_amdgpu_cs *cs)
 {
 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	unsigned int num_ibs[AMDGPU_CS_GANG_SIZE] = { };
 	struct amdgpu_vm *vm = &fpriv->vm;
 	uint64_t *chunk_array_user;
 	uint64_t *chunk_array;
-	unsigned size, num_ibs = 0;
 	uint32_t uf_offset = 0;
+	unsigned int size;
 	int ret;
 	int i;
 
@@ -209,7 +250,7 @@ static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
 			if (size < sizeof(struct drm_amdgpu_cs_chunk_ib))
 				goto free_partial_kdata;
 
-			ret = amdgpu_cs_p1_ib(p, p->chunks[i].kdata, &num_ibs);
+			ret = amdgpu_cs_p1_ib(p, p->chunks[i].kdata, num_ibs);
 			if (ret)
 				goto free_partial_kdata;
 			break;
@@ -246,17 +287,28 @@ static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
 		}
 	}
 
-	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
-	if (ret)
-		goto free_all_kdata;
+	if (!p->gang_size)
+		return -EINVAL;
+
+	for (i = 0; i < p->gang_size; ++i) {
+		ret = amdgpu_job_alloc(p->adev, num_ibs[i], &p->jobs[i], vm);
+		if (ret)
+			goto free_all_kdata;
+
+		ret = drm_sched_job_init(&p->jobs[i]->base, p->entities[i],
+					 &fpriv->vm);
+		if (ret)
+			goto free_all_kdata;
+	}
+	p->gang_leader = p->jobs[p->gang_size - 1];
 
-	if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
+	if (p->ctx->vram_lost_counter != p->gang_leader->vram_lost_counter) {
 		ret = -ECANCELED;
 		goto free_all_kdata;
 	}
 
 	if (p->uf_entry.tv.bo)
-		p->job->uf_addr = uf_offset;
+		p->gang_leader->uf_addr = uf_offset;
 	kvfree(chunk_array);
 
 	/* Use this opportunity to fill in task info for the vm */
@@ -278,93 +330,69 @@ free_chunk:
 	return ret;
 }
 
-static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
-			     struct amdgpu_cs_parser *parser)
+static int amdgpu_cs_p2_ib(struct amdgpu_cs_parser *p,
+			   struct amdgpu_cs_chunk *chunk,
+			   unsigned int *ce_preempt,
+			   unsigned int *de_preempt)
 {
-	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
+	struct drm_amdgpu_cs_chunk_ib *chunk_ib = chunk->kdata;
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 	struct amdgpu_vm *vm = &fpriv->vm;
-	int r, ce_preempt = 0, de_preempt = 0;
 	struct amdgpu_ring *ring;
-	int i, j;
-
-	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
-		struct amdgpu_cs_chunk *chunk;
-		struct amdgpu_ib *ib;
-		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
-		struct drm_sched_entity *entity;
-
-		chunk = &parser->chunks[i];
-		ib = &parser->job->ibs[j];
-		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
-
-		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
-			continue;
-
-		if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
-		    chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
-			if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
-				ce_preempt++;
-			else
-				de_preempt++;
+	struct amdgpu_job *job;
+	struct amdgpu_ib *ib;
+	int r;
 
-			/* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
-			if (ce_preempt > 1 || de_preempt > 1)
-				return -EINVAL;
-		}
+	r = amdgpu_cs_job_idx(p, chunk_ib);
+	if (r < 0)
+		return r;
 
-		r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
-					  chunk_ib->ip_instance, chunk_ib->ring,
-					  &entity);
-		if (r)
-			return r;
+	job = p->jobs[r];
+	ring = amdgpu_job_ring(job);
+	ib = &job->ibs[job->num_ibs++];
 
-		if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
-			parser->job->preamble_status |=
-				AMDGPU_PREAMBLE_IB_PRESENT;
+	/* MM engine doesn't support user fences */
+	if (p->uf_entry.tv.bo && ring->funcs->no_user_fence)
+		return -EINVAL;
 
-		if (parser->entity && parser->entity != entity)
-			return -EINVAL;
+	if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
+	    chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
+		if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
+			(*ce_preempt)++;
+		else
+			(*de_preempt)++;
 
-		/* Return if there is no run queue associated with this entity.
-		 * Possibly because of disabled HW IP*/
-		if (entity->rq == NULL)
+		/* Each GFX command submit allows only 1 IB max
+		 * preemptible for CE & DE */
+		if (*ce_preempt > 1 || *de_preempt > 1)
 			return -EINVAL;
+	}
 
-		parser->entity = entity;
-
-		ring = to_amdgpu_ring(entity->rq->sched);
-		r =  amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
-				   chunk_ib->ib_bytes : 0,
-				   AMDGPU_IB_POOL_DELAYED, ib);
-		if (r) {
-			DRM_ERROR("Failed to get ib !\n");
-			return r;
-		}
-
-		ib->gpu_addr = chunk_ib->va_start;
-		ib->length_dw = chunk_ib->ib_bytes / 4;
-		ib->flags = chunk_ib->flags;
+	if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
+		job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
 
-		j++;
+	r =  amdgpu_ib_get(p->adev, vm, ring->funcs->parse_cs ?
+			   chunk_ib->ib_bytes : 0,
+			   AMDGPU_IB_POOL_DELAYED, ib);
+	if (r) {
+		DRM_ERROR("Failed to get ib !\n");
+		return r;
 	}
 
-	/* MM engine doesn't support user fences */
-	ring = to_amdgpu_ring(parser->entity->rq->sched);
-	if (parser->job->uf_addr && ring->funcs->no_user_fence)
-		return -EINVAL;
-
+	ib->gpu_addr = chunk_ib->va_start;
+	ib->length_dw = chunk_ib->ib_bytes / 4;
+	ib->flags = chunk_ib->flags;
 	return 0;
 }
 
-static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
-				       struct amdgpu_cs_chunk *chunk)
+static int amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser *p,
+				     struct amdgpu_cs_chunk *chunk)
 {
+	struct drm_amdgpu_cs_chunk_dep *deps = chunk->kdata;
 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 	unsigned num_deps;
 	int i, r;
-	struct drm_amdgpu_cs_chunk_dep *deps;
 
-	deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
 	num_deps = chunk->length_dw * 4 /
 		sizeof(struct drm_amdgpu_cs_chunk_dep);
 
@@ -402,7 +430,7 @@ static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
 			dma_fence_put(old);
 		}
 
-		r = amdgpu_sync_fence(&p->job->sync, fence);
+		r = amdgpu_sync_fence(&p->gang_leader->sync, fence);
 		dma_fence_put(fence);
 		if (r)
 			return r;
@@ -410,9 +438,9 @@ static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
 	return 0;
 }
 
-static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
-						 uint32_t handle, u64 point,
-						 u64 flags)
+static int amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser *p,
+					 uint32_t handle, u64 point,
+					 u64 flags)
 {
 	struct dma_fence *fence;
 	int r;
@@ -424,25 +452,23 @@ static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
 		return r;
 	}
 
-	r = amdgpu_sync_fence(&p->job->sync, fence);
+	r = amdgpu_sync_fence(&p->gang_leader->sync, fence);
 	dma_fence_put(fence);
 
 	return r;
 }
 
-static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
-					    struct amdgpu_cs_chunk *chunk)
+static int amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser *p,
+				   struct amdgpu_cs_chunk *chunk)
 {
-	struct drm_amdgpu_cs_chunk_sem *deps;
+	struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
 	unsigned num_deps;
 	int i, r;
 
-	deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
 	num_deps = chunk->length_dw * 4 /
 		sizeof(struct drm_amdgpu_cs_chunk_sem);
 	for (i = 0; i < num_deps; ++i) {
-		r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
-							  0, 0);
+		r = amdgpu_syncobj_lookup_and_add(p, deps[i].handle, 0, 0);
 		if (r)
 			return r;
 	}
@@ -450,21 +476,19 @@ static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
 	return 0;
 }
 
-static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
-						     struct amdgpu_cs_chunk *chunk)
+static int amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser *p,
+					      struct amdgpu_cs_chunk *chunk)
 {
-	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
+	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
 	unsigned num_deps;
 	int i, r;
 
-	syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
 	num_deps = chunk->length_dw * 4 /
 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
 	for (i = 0; i < num_deps; ++i) {
-		r = amdgpu_syncobj_lookup_and_add_to_sync(p,
-							  syncobj_deps[i].handle,
-							  syncobj_deps[i].point,
-							  syncobj_deps[i].flags);
+		r = amdgpu_syncobj_lookup_and_add(p, syncobj_deps[i].handle,
+						  syncobj_deps[i].point,
+						  syncobj_deps[i].flags);
 		if (r)
 			return r;
 	}
@@ -472,14 +496,13 @@ static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
 	return 0;
 }
 
-static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
-					     struct amdgpu_cs_chunk *chunk)
+static int amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser *p,
+				    struct amdgpu_cs_chunk *chunk)
 {
-	struct drm_amdgpu_cs_chunk_sem *deps;
+	struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
 	unsigned num_deps;
 	int i;
 
-	deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
 	num_deps = chunk->length_dw * 4 /
 		sizeof(struct drm_amdgpu_cs_chunk_sem);
 
@@ -507,15 +530,13 @@ static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
 	return 0;
 }
 
-
-static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
-						      struct amdgpu_cs_chunk *chunk)
+static int amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser *p,
+						struct amdgpu_cs_chunk *chunk)
 {
-	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
+	struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
 	unsigned num_deps;
 	int i;
 
-	syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
 	num_deps = chunk->length_dw * 4 /
 		sizeof(struct drm_amdgpu_cs_chunk_syncobj);
 
@@ -552,9 +573,9 @@ static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p
 	return 0;
 }
 
-static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
-				  struct amdgpu_cs_parser *p)
+static int amdgpu_cs_pass2(struct amdgpu_cs_parser *p)
 {
+	unsigned int ce_preempt = 0, de_preempt = 0;
 	int i, r;
 
 	for (i = 0; i < p->nchunks; ++i) {
@@ -563,29 +584,34 @@ static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
 		chunk = &p->chunks[i];
 
 		switch (chunk->chunk_id) {
+		case AMDGPU_CHUNK_ID_IB:
+			r = amdgpu_cs_p2_ib(p, chunk, &ce_preempt, &de_preempt);
+			if (r)
+				return r;
+			break;
 		case AMDGPU_CHUNK_ID_DEPENDENCIES:
 		case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
-			r = amdgpu_cs_process_fence_dep(p, chunk);
+			r = amdgpu_cs_p2_dependencies(p, chunk);
 			if (r)
 				return r;
 			break;
 		case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
-			r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
+			r = amdgpu_cs_p2_syncobj_in(p, chunk);
 			if (r)
 				return r;
 			break;
 		case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
-			r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
+			r = amdgpu_cs_p2_syncobj_out(p, chunk);
 			if (r)
 				return r;
 			break;
 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
-			r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
+			r = amdgpu_cs_p2_syncobj_timeline_wait(p, chunk);
 			if (r)
 				return r;
 			break;
 		case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
-			r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
+			r = amdgpu_cs_p2_syncobj_timeline_signal(p, chunk);
 			if (r)
 				return r;
 			break;
@@ -830,6 +856,7 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 	struct amdgpu_vm *vm = &fpriv->vm;
 	struct amdgpu_bo_list_entry *e;
 	struct list_head duplicates;
+	unsigned int i;
 	int r;
 
 	INIT_LIST_HEAD(&p->validated);
@@ -913,16 +940,6 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 		e->bo_va = amdgpu_vm_bo_find(vm, bo);
 	}
 
-	/* Move fence waiting after getting reservation lock of
-	 * PD root. Then there is no need on a ctx mutex lock.
-	 */
-	r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entity);
-	if (unlikely(r != 0)) {
-		if (r != -ERESTARTSYS)
-			DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
-		goto error_validate;
-	}
-
 	amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
 					  &p->bytes_moved_vis_threshold);
 	p->bytes_moved = 0;
@@ -943,124 +960,138 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 	if (r)
 		goto error_validate;
 
-	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
-				     p->bytes_moved_vis);
-
-	amdgpu_job_set_resources(p->job, p->bo_list->gds_obj,
-				 p->bo_list->gws_obj, p->bo_list->oa_obj);
-
-	if (!r && p->uf_entry.tv.bo) {
+	if (p->uf_entry.tv.bo) {
 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
 
 		r = amdgpu_ttm_alloc_gart(&uf->tbo);
-		p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
+		if (r)
+			goto error_validate;
+
+		p->gang_leader->uf_addr += amdgpu_bo_gpu_offset(uf);
 	}
 
+	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
+				     p->bytes_moved_vis);
+
+	for (i = 0; i < p->gang_size; ++i)
+		amdgpu_job_set_resources(p->jobs[i], p->bo_list->gds_obj,
+					 p->bo_list->gws_obj,
+					 p->bo_list->oa_obj);
+	return 0;
+
 error_validate:
-	if (r)
-		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
+	ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 
 out_free_user_pages:
-	if (r) {
-		amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
-			struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
+	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
+		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 
-			if (!e->user_pages)
-				continue;
-			amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
-			kvfree(e->user_pages);
-			e->user_pages = NULL;
-		}
-		mutex_unlock(&p->bo_list->bo_list_mutex);
+		if (!e->user_pages)
+			continue;
+		amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
+		kvfree(e->user_pages);
+		e->user_pages = NULL;
 	}
 	return r;
 }
 
-static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *parser)
+static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *p)
 {
-	int i;
+	int i, j;
 
 	if (!trace_amdgpu_cs_enabled())
 		return;
 
-	for (i = 0; i < parser->job->num_ibs; i++)
-		trace_amdgpu_cs(parser, i);
+	for (i = 0; i < p->gang_size; ++i) {
+		struct amdgpu_job *job = p->jobs[i];
+
+		for (j = 0; j < job->num_ibs; ++j)
+			trace_amdgpu_cs(p, job, &job->ibs[j]);
+	}
 }
 
-static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
+static int amdgpu_cs_patch_ibs(struct amdgpu_cs_parser *p,
+			       struct amdgpu_job *job)
 {
-	struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
-	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
-	struct amdgpu_device *adev = p->adev;
-	struct amdgpu_vm *vm = &fpriv->vm;
-	struct amdgpu_bo_list_entry *e;
-	struct amdgpu_bo_va *bo_va;
-	struct amdgpu_bo *bo;
+	struct amdgpu_ring *ring = amdgpu_job_ring(job);
+	unsigned int i;
 	int r;
 
 	/* Only for UVD/VCE VM emulation */
-	if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
-		unsigned i, j;
-
-		for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
-			struct drm_amdgpu_cs_chunk_ib *chunk_ib;
-			struct amdgpu_bo_va_mapping *m;
-			struct amdgpu_bo *aobj = NULL;
-			struct amdgpu_cs_chunk *chunk;
-			uint64_t offset, va_start;
-			struct amdgpu_ib *ib;
-			uint8_t *kptr;
-
-			chunk = &p->chunks[i];
-			ib = &p->job->ibs[j];
-			chunk_ib = chunk->kdata;
-
-			if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
-				continue;
+	if (!ring->funcs->parse_cs && !ring->funcs->patch_cs_in_place)
+		return 0;
 
-			va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
-			r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
-			if (r) {
-				DRM_ERROR("IB va_start is invalid\n");
-				return r;
-			}
+	for (i = 0; i < job->num_ibs; ++i) {
+		struct amdgpu_ib *ib = &job->ibs[i];
+		struct amdgpu_bo_va_mapping *m;
+		struct amdgpu_bo *aobj;
+		uint64_t va_start;
+		uint8_t *kptr;
 
-			if ((va_start + chunk_ib->ib_bytes) >
-			    (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
-				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
-				return -EINVAL;
-			}
+		va_start = ib->gpu_addr & AMDGPU_GMC_HOLE_MASK;
+		r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
+		if (r) {
+			DRM_ERROR("IB va_start is invalid\n");
+			return r;
+		}
 
-			/* the IB should be reserved at this point */
-			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
-			if (r) {
-				return r;
-			}
+		if ((va_start + ib->length_dw * 4) >
+		    (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
+			DRM_ERROR("IB va_start+ib_bytes is invalid\n");
+			return -EINVAL;
+		}
 
-			offset = m->start * AMDGPU_GPU_PAGE_SIZE;
-			kptr += va_start - offset;
-
-			if (ring->funcs->parse_cs) {
-				memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
-				amdgpu_bo_kunmap(aobj);
-
-				r = amdgpu_ring_parse_cs(ring, p, p->job, ib);
-				if (r)
-					return r;
-			} else {
-				ib->ptr = (uint32_t *)kptr;
-				r = amdgpu_ring_patch_cs_in_place(ring, p, p->job, ib);
-				amdgpu_bo_kunmap(aobj);
-				if (r)
-					return r;
-			}
+		/* the IB should be reserved at this point */
+		r = amdgpu_bo_kmap(aobj, (void **)&kptr);
+		if (r) {
+			return r;
+		}
+
+		kptr += va_start - (m->start * AMDGPU_GPU_PAGE_SIZE);
 
-			j++;
+		if (ring->funcs->parse_cs) {
+			memcpy(ib->ptr, kptr, ib->length_dw * 4);
+			amdgpu_bo_kunmap(aobj);
+
+			r = amdgpu_ring_parse_cs(ring, p, job, ib);
+			if (r)
+				return r;
+		} else {
+			ib->ptr = (uint32_t *)kptr;
+			r = amdgpu_ring_patch_cs_in_place(ring, p, job, ib);
+			amdgpu_bo_kunmap(aobj);
+			if (r)
+				return r;
 		}
 	}
 
-	if (!p->job->vm)
-		return 0;
+	return 0;
+}
+
+static int amdgpu_cs_patch_jobs(struct amdgpu_cs_parser *p)
+{
+	unsigned int i;
+	int r;
+
+	for (i = 0; i < p->gang_size; ++i) {
+		r = amdgpu_cs_patch_ibs(p, p->jobs[i]);
+		if (r)
+			return r;
+	}
+	return 0;
+}
+
+static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
+{
+	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_job *job = p->gang_leader;
+	struct amdgpu_device *adev = p->adev;
+	struct amdgpu_vm *vm = &fpriv->vm;
+	struct amdgpu_bo_list_entry *e;
+	struct amdgpu_bo_va *bo_va;
+	struct amdgpu_bo *bo;
+	unsigned int i;
+	int r;
 
 	r = amdgpu_vm_clear_freed(adev, vm, NULL);
 	if (r)
@@ -1070,7 +1101,7 @@ static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
 	if (r)
 		return r;
 
-	r = amdgpu_sync_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
+	r = amdgpu_sync_fence(&job->sync, fpriv->prt_va->last_pt_update);
 	if (r)
 		return r;
 
@@ -1081,7 +1112,7 @@ static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
 		if (r)
 			return r;
 
-		r = amdgpu_sync_fence(&p->job->sync, bo_va->last_pt_update);
+		r = amdgpu_sync_fence(&job->sync, bo_va->last_pt_update);
 		if (r)
 			return r;
 	}
@@ -1100,7 +1131,7 @@ static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
 		if (r)
 			return r;
 
-		r = amdgpu_sync_fence(&p->job->sync, bo_va->last_pt_update);
+		r = amdgpu_sync_fence(&job->sync, bo_va->last_pt_update);
 		if (r)
 			return r;
 	}
@@ -1113,11 +1144,18 @@ static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
 	if (r)
 		return r;
 
-	r = amdgpu_sync_fence(&p->job->sync, vm->last_update);
+	r = amdgpu_sync_fence(&job->sync, vm->last_update);
 	if (r)
 		return r;
 
-	p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
+	for (i = 0; i < p->gang_size; ++i) {
+		job = p->jobs[i];
+
+		if (!job->vm)
+			continue;
+
+		job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
+	}
 
 	if (amdgpu_vm_debug) {
 		/* Invalidate all BOs to test for userspace bugs */
@@ -1138,7 +1176,9 @@ static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 {
 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
+	struct amdgpu_job *leader = p->gang_leader;
 	struct amdgpu_bo_list_entry *e;
+	unsigned int i;
 	int r;
 
 	list_for_each_entry(e, &p->validated, tv.head) {
@@ -1148,12 +1188,23 @@ static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
 
 		sync_mode = amdgpu_bo_explicit_sync(bo) ?
 			AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
-		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, sync_mode,
+		r = amdgpu_sync_resv(p->adev, &leader->sync, resv, sync_mode,
 				     &fpriv->vm);
 		if (r)
 			return r;
 	}
-	return 0;
+
+	for (i = 0; i < p->gang_size - 1; ++i) {
+		r = amdgpu_sync_clone(&leader->sync, &p->jobs[i]->sync);
+		if (r)
+			return r;
+	}
+
+	r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entities[p->gang_size - 1]);
+	if (r && r != -ERESTARTSYS)
+		DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
+
+	return r;
 }
 
 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
@@ -1177,20 +1228,28 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
 			    union drm_amdgpu_cs *cs)
 {
 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
-	struct drm_sched_entity *entity = p->entity;
+	struct amdgpu_job *leader = p->gang_leader;
 	struct amdgpu_bo_list_entry *e;
-	struct amdgpu_job *job;
+	unsigned int i;
 	uint64_t seq;
 	int r;
 
-	job = p->job;
-	p->job = NULL;
+	for (i = 0; i < p->gang_size; ++i)
+		drm_sched_job_arm(&p->jobs[i]->base);
 
-	r = drm_sched_job_init(&job->base, entity, &fpriv->vm);
-	if (r)
-		goto error_unlock;
+	for (i = 0; i < (p->gang_size - 1); ++i) {
+		struct dma_fence *fence;
 
-	drm_sched_job_arm(&job->base);
+		fence = &p->jobs[i]->base.s_fence->scheduled;
+		r = amdgpu_sync_fence(&leader->sync, fence);
+		if (r)
+			goto error_cleanup;
+	}
+
+	if (p->gang_size > 1) {
+		for (i = 0; i < p->gang_size; ++i)
+			amdgpu_job_set_gang_leader(p->jobs[i], leader);
+	}
 
 	/* No memory allocation is allowed while holding the notifier lock.
 	 * The lock is held until amdgpu_cs_submit is finished and fence is
@@ -1201,6 +1260,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
 	/* If userptr are invalidated after amdgpu_cs_parser_bos(), return
 	 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
 	 */
+	r = 0;
 	amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
 
@@ -1208,62 +1268,65 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
 	}
 	if (r) {
 		r = -EAGAIN;
-		goto error_abort;
+		goto error_unlock;
 	}
 
-	p->fence = dma_fence_get(&job->base.s_fence->finished);
+	p->fence = dma_fence_get(&leader->base.s_fence->finished);
+	list_for_each_entry(e, &p->validated, tv.head) {
 
-	seq = amdgpu_ctx_add_fence(p->ctx, entity, p->fence);
+		/* Everybody except for the gang leader uses READ */
+		for (i = 0; i < (p->gang_size - 1); ++i) {
+			dma_resv_add_fence(e->tv.bo->base.resv,
+					   &p->jobs[i]->base.s_fence->finished,
+					   DMA_RESV_USAGE_READ);
+		}
+
+		/* The gang leader is remembered as writer */
+		e->tv.num_shared = 0;
+	}
+
+	seq = amdgpu_ctx_add_fence(p->ctx, p->entities[p->gang_size - 1],
+				   p->fence);
 	amdgpu_cs_post_dependencies(p);
 
-	if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
+	if ((leader->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
 	    !p->ctx->preamble_presented) {
-		job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
+		leader->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
 		p->ctx->preamble_presented = true;
 	}
 
 	cs->out.handle = seq;
-	job->uf_sequence = seq;
-
-	amdgpu_job_free_resources(job);
+	leader->uf_sequence = seq;
 
-	trace_amdgpu_cs_ioctl(job);
 	amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
-	drm_sched_entity_push_job(&job->base);
+	for (i = 0; i < p->gang_size; ++i) {
+		amdgpu_job_free_resources(p->jobs[i]);
+		trace_amdgpu_cs_ioctl(p->jobs[i]);
+		drm_sched_entity_push_job(&p->jobs[i]->base);
+		p->jobs[i] = NULL;
+	}
 
 	amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
-
-	/* Make sure all BOs are remembered as writers */
-	amdgpu_bo_list_for_each_entry(e, p->bo_list)
-		e->tv.num_shared = 0;
-
 	ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
+
 	mutex_unlock(&p->adev->notifier_lock);
 	mutex_unlock(&p->bo_list->bo_list_mutex);
-
 	return 0;
 
-error_abort:
-	drm_sched_job_cleanup(&job->base);
+error_unlock:
 	mutex_unlock(&p->adev->notifier_lock);
 
-error_unlock:
-	amdgpu_job_free(job);
+error_cleanup:
+	for (i = 0; i < p->gang_size; ++i)
+		drm_sched_job_cleanup(&p->jobs[i]->base);
 	return r;
 }
 
 /* Cleanup the parser structure */
-static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
-				  bool backoff)
+static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser)
 {
 	unsigned i;
 
-	if (error && backoff) {
-		ttm_eu_backoff_reservation(&parser->ticket,
-					   &parser->validated);
-		mutex_unlock(&parser->bo_list->bo_list_mutex);
-	}
-
 	for (i = 0; i < parser->num_post_deps; i++) {
 		drm_syncobj_put(parser->post_deps[i].syncobj);
 		kfree(parser->post_deps[i].chain);
@@ -1280,8 +1343,10 @@ static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
 	for (i = 0; i < parser->nchunks; i++)
 		kvfree(parser->chunks[i].kdata);
 	kvfree(parser->chunks);
-	if (parser->job)
-		amdgpu_job_free(parser->job);
+	for (i = 0; i < parser->gang_size; ++i) {
+		if (parser->jobs[i])
+			amdgpu_job_free(parser->jobs[i]);
+	}
 	if (parser->uf_entry.tv.bo) {
 		struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
 
@@ -1293,7 +1358,6 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 {
 	struct amdgpu_device *adev = drm_to_adev(dev);
 	struct amdgpu_cs_parser parser;
-	bool reserved_buffers = false;
 	int r;
 
 	if (amdgpu_ras_intr_triggered())
@@ -1306,22 +1370,16 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	if (r) {
 		if (printk_ratelimit())
 			DRM_ERROR("Failed to initialize parser %d!\n", r);
-		goto out;
+		return r;
 	}
 
 	r = amdgpu_cs_pass1(&parser, data);
 	if (r)
-		goto out;
+		goto error_fini;
 
-	r = amdgpu_cs_ib_fill(adev, &parser);
+	r = amdgpu_cs_pass2(&parser);
 	if (r)
-		goto out;
-
-	r = amdgpu_cs_dependencies(adev, &parser);
-	if (r) {
-		DRM_ERROR("Failed in the dependencies handling %d!\n", r);
-		goto out;
-	}
+		goto error_fini;
 
 	r = amdgpu_cs_parser_bos(&parser, data);
 	if (r) {
@@ -1329,25 +1387,36 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 			DRM_ERROR("Not enough memory for command submission!\n");
 		else if (r != -ERESTARTSYS && r != -EAGAIN)
 			DRM_ERROR("Failed to process the buffer list %d!\n", r);
-		goto out;
+		goto error_fini;
 	}
 
-	reserved_buffers = true;
-
-	trace_amdgpu_cs_ibs(&parser);
+	r = amdgpu_cs_patch_jobs(&parser);
+	if (r)
+		goto error_backoff;
 
 	r = amdgpu_cs_vm_handling(&parser);
 	if (r)
-		goto out;
+		goto error_backoff;
 
 	r = amdgpu_cs_sync_rings(&parser);
 	if (r)
-		goto out;
+		goto error_backoff;
+
+	trace_amdgpu_cs_ibs(&parser);
 
 	r = amdgpu_cs_submit(&parser, data);
-out:
-	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
+	if (r)
+		goto error_backoff;
+
+	amdgpu_cs_parser_fini(&parser);
+	return 0;
+
+error_backoff:
+	ttm_eu_backoff_reservation(&parser.ticket, &parser.validated);
+	mutex_unlock(&parser.bo_list->bo_list_mutex);
 
+error_fini:
+	amdgpu_cs_parser_fini(&parser);
 	return r;
 }