1 files changed, 197 insertions, 37 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 9b61c8ba7aaf..dfb12237a6b0 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -34,6 +34,7 @@
 #include <ttm/ttm_placement.h>
 #include <ttm/ttm_module.h>
 #include <ttm/ttm_page_alloc.h>
+#include <ttm/ttm_memory.h>
 #include <drm/drmP.h>
 #include <drm/amdgpu_drm.h>
 #include <linux/seq_file.h>
@@ -74,7 +75,7 @@ static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
 	ttm_mem_global_release(ref->object);
 }
 
-static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
+int amdgpu_ttm_global_init(struct amdgpu_device *adev)
 {
 	struct drm_global_reference *global_ref;
 	struct amdgpu_ring *ring;
@@ -88,10 +89,10 @@ static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
 	global_ref->init = &amdgpu_ttm_mem_global_init;
 	global_ref->release = &amdgpu_ttm_mem_global_release;
 	r = drm_global_item_ref(global_ref);
-	if (r != 0) {
+	if (r) {
 		DRM_ERROR("Failed setting up TTM memory accounting "
 			  "subsystem.\n");
-		return r;
+		goto error_mem;
 	}
 
 	adev->mman.bo_global_ref.mem_glob =
@@ -102,26 +103,30 @@ static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
 	global_ref->init = &ttm_bo_global_init;
 	global_ref->release = &ttm_bo_global_release;
 	r = drm_global_item_ref(global_ref);
-	if (r != 0) {
+	if (r) {
 		DRM_ERROR("Failed setting up TTM BO subsystem.\n");
-		drm_global_item_unref(&adev->mman.mem_global_ref);
-		return r;
+		goto error_bo;
 	}
 
 	ring = adev->mman.buffer_funcs_ring;
 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_KERNEL];
 	r = amd_sched_entity_init(&ring->sched, &adev->mman.entity,
 				  rq, amdgpu_sched_jobs);
-	if (r != 0) {
+	if (r) {
 		DRM_ERROR("Failed setting up TTM BO move run queue.\n");
-		drm_global_item_unref(&adev->mman.mem_global_ref);
-		drm_global_item_unref(&adev->mman.bo_global_ref.ref);
-		return r;
+		goto error_entity;
 	}
 
 	adev->mman.mem_global_referenced = true;
 
 	return 0;
+
+error_entity:
+	drm_global_item_unref(&adev->mman.bo_global_ref.ref);
+error_bo:
+	drm_global_item_unref(&adev->mman.mem_global_ref);
+error_mem:
+	return r;
 }
 
 static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
@@ -196,6 +201,7 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 		.lpfn = 0,
 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
 	};
+	unsigned i;
 
 	if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
 		placement->placement = &placements;
@@ -207,10 +213,25 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 	rbo = container_of(bo, struct amdgpu_bo, tbo);
 	switch (bo->mem.mem_type) {
 	case TTM_PL_VRAM:
-		if (rbo->adev->mman.buffer_funcs_ring->ready == false)
+		if (rbo->adev->mman.buffer_funcs_ring->ready == false) {
 			amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU);
-		else
+		} else {
 			amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_GTT);
+			for (i = 0; i < rbo->placement.num_placement; ++i) {
+				if (!(rbo->placements[i].flags &
+				      TTM_PL_FLAG_TT))
+					continue;
+
+				if (rbo->placements[i].lpfn)
+					continue;
+
+				/* set an upper limit to force directly
+				 * allocating address space for the BO.
+				 */
+				rbo->placements[i].lpfn =
+					rbo->adev->mc.gtt_size >> PAGE_SHIFT;
+			}
+		}
 		break;
 	case TTM_PL_TT:
 	default:
@@ -255,22 +276,26 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 	new_start = new_mem->start << PAGE_SHIFT;
 
 	switch (old_mem->mem_type) {
-	case TTM_PL_VRAM:
-		old_start += adev->mc.vram_start;
-		break;
 	case TTM_PL_TT:
-		old_start += adev->mc.gtt_start;
+		r = amdgpu_ttm_bind(bo->ttm, old_mem);
+		if (r)
+			return r;
+
+	case TTM_PL_VRAM:
+		old_start += bo->bdev->man[old_mem->mem_type].gpu_offset;
 		break;
 	default:
 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 		return -EINVAL;
 	}
 	switch (new_mem->mem_type) {
-	case TTM_PL_VRAM:
-		new_start += adev->mc.vram_start;
-		break;
 	case TTM_PL_TT:
-		new_start += adev->mc.gtt_start;
+		r = amdgpu_ttm_bind(bo->ttm, new_mem);
+		if (r)
+			return r;
+
+	case TTM_PL_VRAM:
+		new_start += bo->bdev->man[new_mem->mem_type].gpu_offset;
 		break;
 	default:
 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
@@ -285,7 +310,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 
 	r = amdgpu_copy_buffer(ring, old_start, new_start,
 			       new_mem->num_pages * PAGE_SIZE, /* bytes */
-			       bo->resv, &fence);
+			       bo->resv, &fence, false);
 	if (r)
 		return r;
 
@@ -314,7 +339,7 @@ static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
 	placement.num_busy_placement = 1;
 	placement.busy_placement = &placements;
 	placements.fpfn = 0;
-	placements.lpfn = 0;
+	placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 			     interruptible, no_wait_gpu);
@@ -335,7 +360,7 @@ static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
 	if (unlikely(r)) {
 		goto out_cleanup;
 	}
-	r = ttm_bo_move_ttm(bo, true, interruptible, no_wait_gpu, new_mem);
+	r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, new_mem);
 out_cleanup:
 	ttm_bo_mem_put(bo, &tmp_mem);
 	return r;
@@ -361,14 +386,14 @@ static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
 	placement.num_busy_placement = 1;
 	placement.busy_placement = &placements;
 	placements.fpfn = 0;
-	placements.lpfn = 0;
+	placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 			     interruptible, no_wait_gpu);
 	if (unlikely(r)) {
 		return r;
 	}
-	r = ttm_bo_move_ttm(bo, true, interruptible, no_wait_gpu, &tmp_mem);
+	r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, &tmp_mem);
 	if (unlikely(r)) {
 		goto out_cleanup;
 	}
@@ -435,8 +460,7 @@ static int amdgpu_bo_move(struct ttm_buffer_object *bo,
 
 	if (r) {
 memcpy:
-		r = ttm_bo_move_memcpy(bo, evict, interruptible,
-				       no_wait_gpu, new_mem);
+		r = ttm_bo_move_memcpy(bo, interruptible, no_wait_gpu, new_mem);
 		if (r) {
 			return r;
 		}
@@ -524,6 +548,7 @@ struct amdgpu_ttm_tt {
 	spinlock_t              guptasklock;
 	struct list_head        guptasks;
 	atomic_t		mmu_invalidations;
+	struct list_head        list;
 };
 
 int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
@@ -641,7 +666,6 @@ static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
 				   struct ttm_mem_reg *bo_mem)
 {
 	struct amdgpu_ttm_tt *gtt = (void*)ttm;
-	uint32_t flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
 	int r;
 
 	if (gtt->userptr) {
@@ -651,7 +675,7 @@ static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
 			return r;
 		}
 	}
-	gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
+	gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
 	if (!ttm->num_pages) {
 		WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
 		     ttm->num_pages, bo_mem, ttm);
@@ -662,14 +686,62 @@ static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
 	    bo_mem->mem_type == AMDGPU_PL_OA)
 		return -EINVAL;
 
+	return 0;
+}
+
+bool amdgpu_ttm_is_bound(struct ttm_tt *ttm)
+{
+	struct amdgpu_ttm_tt *gtt = (void *)ttm;
+
+	return gtt && !list_empty(&gtt->list);
+}
+
+int amdgpu_ttm_bind(struct ttm_tt *ttm, struct ttm_mem_reg *bo_mem)
+{
+	struct amdgpu_ttm_tt *gtt = (void *)ttm;
+	uint32_t flags;
+	int r;
+
+	if (!ttm || amdgpu_ttm_is_bound(ttm))
+		return 0;
+
+	flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
 	r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
 		ttm->pages, gtt->ttm.dma_address, flags);
 
 	if (r) {
-		DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
-			  ttm->num_pages, (unsigned)gtt->offset);
+		DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
+			  ttm->num_pages, gtt->offset);
 		return r;
 	}
+	spin_lock(&gtt->adev->gtt_list_lock);
+	list_add_tail(&gtt->list, &gtt->adev->gtt_list);
+	spin_unlock(&gtt->adev->gtt_list_lock);
+	return 0;
+}
+
+int amdgpu_ttm_recover_gart(struct amdgpu_device *adev)
+{
+	struct amdgpu_ttm_tt *gtt, *tmp;
+	struct ttm_mem_reg bo_mem;
+	uint32_t flags;
+	int r;
+
+	bo_mem.mem_type = TTM_PL_TT;
+	spin_lock(&adev->gtt_list_lock);
+	list_for_each_entry_safe(gtt, tmp, &adev->gtt_list, list) {
+		flags = amdgpu_ttm_tt_pte_flags(gtt->adev, &gtt->ttm.ttm, &bo_mem);
+		r = amdgpu_gart_bind(adev, gtt->offset, gtt->ttm.ttm.num_pages,
+				     gtt->ttm.ttm.pages, gtt->ttm.dma_address,
+				     flags);
+		if (r) {
+			spin_unlock(&adev->gtt_list_lock);
+			DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
+				  gtt->ttm.ttm.num_pages, gtt->offset);
+			return r;
+		}
+	}
+	spin_unlock(&adev->gtt_list_lock);
 	return 0;
 }
 
@@ -677,6 +749,9 @@ static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
 {
 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
 
+	if (!amdgpu_ttm_is_bound(ttm))
+		return 0;
+
 	/* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
 	if (gtt->adev->gart.ready)
 		amdgpu_gart_unbind(gtt->adev, gtt->offset, ttm->num_pages);
@@ -684,6 +759,10 @@ static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
 	if (gtt->userptr)
 		amdgpu_ttm_tt_unpin_userptr(ttm);
 
+	spin_lock(&gtt->adev->gtt_list_lock);
+	list_del_init(&gtt->list);
+	spin_unlock(&gtt->adev->gtt_list_lock);
+
 	return 0;
 }
 
@@ -720,6 +799,7 @@ static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev,
 		kfree(gtt);
 		return NULL;
 	}
+	INIT_LIST_HEAD(&gtt->list);
 	return &gtt->ttm.ttm;
 }
 
@@ -950,6 +1030,8 @@ static struct list_head *amdgpu_ttm_lru_tail(struct ttm_buffer_object *tbo)
 	struct list_head *res = lru->lru[tbo->mem.mem_type];
 
 	lru->lru[tbo->mem.mem_type] = &tbo->lru;
+	while ((++lru)->lru[tbo->mem.mem_type] == res)
+		lru->lru[tbo->mem.mem_type] = &tbo->lru;
 
 	return res;
 }
@@ -960,6 +1042,8 @@ static struct list_head *amdgpu_ttm_swap_lru_tail(struct ttm_buffer_object *tbo)
 	struct list_head *res = lru->swap_lru;
 
 	lru->swap_lru = &tbo->swap;
+	while ((++lru)->swap_lru == res)
+		lru->swap_lru = &tbo->swap;
 
 	return res;
 }
@@ -987,10 +1071,6 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 	unsigned i, j;
 	int r;
 
-	r = amdgpu_ttm_global_init(adev);
-	if (r) {
-		return r;
-	}
 	/* No others user of address space so set it to 0 */
 	r = ttm_bo_device_init(&adev->mman.bdev,
 			       adev->mman.bo_global_ref.ref.object,
@@ -1011,6 +1091,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 		lru->swap_lru = &adev->mman.bdev.glob->swap_lru;
 	}
 
+	for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
+		adev->mman.guard.lru[j] = NULL;
+	adev->mman.guard.swap_lru = NULL;
+
 	adev->mman.initialized = true;
 	r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
 				adev->mc.real_vram_size >> PAGE_SHIFT);
@@ -1151,7 +1235,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring,
 		       uint64_t dst_offset,
 		       uint32_t byte_count,
 		       struct reservation_object *resv,
-		       struct fence **fence)
+		       struct fence **fence, bool direct_submit)
 {
 	struct amdgpu_device *adev = ring->adev;
 	struct amdgpu_job *job;
@@ -1195,8 +1279,79 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring,
 
 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 	WARN_ON(job->ibs[0].length_dw > num_dw);
+	if (direct_submit) {
+		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
+				       NULL, NULL, fence);
+		job->fence = fence_get(*fence);
+		if (r)
+			DRM_ERROR("Error scheduling IBs (%d)\n", r);
+		amdgpu_job_free(job);
+	} else {
+		r = amdgpu_job_submit(job, ring, &adev->mman.entity,
+				      AMDGPU_FENCE_OWNER_UNDEFINED, fence);
+		if (r)
+			goto error_free;
+	}
+
+	return r;
+
+error_free:
+	amdgpu_job_free(job);
+	return r;
+}
+
+int amdgpu_fill_buffer(struct amdgpu_bo *bo,
+		uint32_t src_data,
+		struct reservation_object *resv,
+		struct fence **fence)
+{
+	struct amdgpu_device *adev = bo->adev;
+	struct amdgpu_job *job;
+	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
+
+	uint32_t max_bytes, byte_count;
+	uint64_t dst_offset;
+	unsigned int num_loops, num_dw;
+	unsigned int i;
+	int r;
+
+	byte_count = bo->tbo.num_pages << PAGE_SHIFT;
+	max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
+	num_loops = DIV_ROUND_UP(byte_count, max_bytes);
+	num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
+
+	/* for IB padding */
+	while (num_dw & 0x7)
+		num_dw++;
+
+	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
+	if (r)
+		return r;
+
+	if (resv) {
+		r = amdgpu_sync_resv(adev, &job->sync, resv,
+				AMDGPU_FENCE_OWNER_UNDEFINED);
+		if (r) {
+			DRM_ERROR("sync failed (%d).\n", r);
+			goto error_free;
+		}
+	}
+
+	dst_offset = bo->tbo.mem.start << PAGE_SHIFT;
+	for (i = 0; i < num_loops; i++) {
+		uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
+
+		amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
+				dst_offset, cur_size_in_bytes);
+
+		dst_offset += cur_size_in_bytes;
+		byte_count -= cur_size_in_bytes;
+	}
+
+	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
+	WARN_ON(job->ibs[0].length_dw > num_dw);
 	r = amdgpu_job_submit(job, ring, &adev->mman.entity,
-			      AMDGPU_FENCE_OWNER_UNDEFINED, fence);
+			AMDGPU_FENCE_OWNER_UNDEFINED, fence);
 	if (r)
 		goto error_free;
 
@@ -1387,3 +1542,8 @@ static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
 
 #endif
 }
+
+u64 amdgpu_ttm_get_gtt_mem_size(struct amdgpu_device *adev)
+{
+	return ttm_get_kernel_zone_memory_size(adev->mman.mem_global_ref.object);
+}