From 03f48dd5d2085b978353a804635b6d34e77a5635 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Christian=20K=C3=B6nig?= Date: Mon, 15 Aug 2016 17:00:22 +0200 Subject: drm/amdgpu: add AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS flag v3 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Add a flag noting that a BO must be created using linear VRAM and set this flag on all in kernel users where appropriate. Hopefully I haven't missed anything. v2: add it in a few more places, fix CPU mapping. v3: rename to VRAM_CONTIGUOUS, fix typo in CS code. Signed-off-by: Christian König Reviewed-by: Edward O'Callaghan Tested-by: Mike Lothian Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index dcaf691f56b5..a06cf9988912 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -1119,7 +1119,8 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) r = amdgpu_bo_create(adev, 256 * 1024, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, - AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED | + AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS, NULL, NULL, &adev->stollen_vga_memory); if (r) { return r; -- cgit v1.2.1 From 8892f153c83e521aff2ee7ac620856e2983cd7a0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Christian=20K=C3=B6nig?= Date: Wed, 17 Aug 2016 10:46:52 +0200 Subject: drm/amdgpu: enable amdgpu_move_blit to handle multiple MM nodes v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This allows us to move scattered buffers around. v2: fix a couple of typos, handle scattered to scattered moves as well. Signed-off-by: Christian König Reviewed-by: Edward O'Callaghan Tested-by: Mike Lothian Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 118 +++++++++++++++++++++++--------- 1 file changed, 85 insertions(+), 33 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index a06cf9988912..1427c40c70bb 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -260,64 +260,116 @@ static void amdgpu_move_null(struct ttm_buffer_object *bo, new_mem->mm_node = NULL; } -static int amdgpu_move_blit(struct ttm_buffer_object *bo, - bool evict, bool no_wait_gpu, - struct ttm_mem_reg *new_mem, - struct ttm_mem_reg *old_mem) +static int amdgpu_mm_node_addr(struct ttm_buffer_object *bo, + struct drm_mm_node *mm_node, + struct ttm_mem_reg *mem, + uint64_t *addr) { - struct amdgpu_device *adev; - struct amdgpu_ring *ring; - uint64_t old_start, new_start; - struct fence *fence; int r; - adev = amdgpu_get_adev(bo->bdev); - ring = adev->mman.buffer_funcs_ring; - - switch (old_mem->mem_type) { + switch (mem->mem_type) { case TTM_PL_TT: - r = amdgpu_ttm_bind(bo, old_mem); + r = amdgpu_ttm_bind(bo, mem); if (r) return r; case TTM_PL_VRAM: - old_start = (u64)old_mem->start << PAGE_SHIFT; - old_start += bo->bdev->man[old_mem->mem_type].gpu_offset; + *addr = mm_node->start << PAGE_SHIFT; + *addr += bo->bdev->man[mem->mem_type].gpu_offset; break; default: - DRM_ERROR("Unknown placement %d\n", old_mem->mem_type); + DRM_ERROR("Unknown placement %d\n", mem->mem_type); return -EINVAL; } - switch (new_mem->mem_type) { - case TTM_PL_TT: - r = amdgpu_ttm_bind(bo, new_mem); - if (r) - return r; - case TTM_PL_VRAM: - new_start = (u64)new_mem->start << PAGE_SHIFT; - new_start += bo->bdev->man[new_mem->mem_type].gpu_offset; - break; - default: - DRM_ERROR("Unknown placement %d\n", old_mem->mem_type); - return -EINVAL; - } + return 0; +} + +static int amdgpu_move_blit(struct ttm_buffer_object *bo, + bool evict, bool no_wait_gpu, + struct ttm_mem_reg *new_mem, + struct ttm_mem_reg *old_mem) +{ + struct amdgpu_device *adev = amdgpu_get_adev(bo->bdev); + struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; + + struct drm_mm_node *old_mm, *new_mm; + uint64_t old_start, old_size, new_start, new_size; + unsigned long num_pages; + struct fence *fence = NULL; + int r; + + BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0); + if (!ring->ready) { DRM_ERROR("Trying to move memory with ring turned off.\n"); return -EINVAL; } - BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0); + old_mm = old_mem->mm_node; + r = amdgpu_mm_node_addr(bo, old_mm, old_mem, &old_start); + if (r) + return r; + old_size = old_mm->size; + - r = amdgpu_copy_buffer(ring, old_start, new_start, - new_mem->num_pages * PAGE_SIZE, /* bytes */ - bo->resv, &fence, false); + new_mm = new_mem->mm_node; + r = amdgpu_mm_node_addr(bo, new_mm, new_mem, &new_start); if (r) return r; + new_size = new_mm->size; + + num_pages = new_mem->num_pages; + while (num_pages) { + unsigned long cur_pages = min(old_size, new_size); + struct fence *next; + + r = amdgpu_copy_buffer(ring, old_start, new_start, + cur_pages * PAGE_SIZE, + bo->resv, &next, false); + if (r) + goto error; + + fence_put(fence); + fence = next; + + num_pages -= cur_pages; + if (!num_pages) + break; + + old_size -= cur_pages; + if (!old_size) { + r = amdgpu_mm_node_addr(bo, ++old_mm, old_mem, + &old_start); + if (r) + goto error; + old_size = old_mm->size; + } else { + old_start += cur_pages * PAGE_SIZE; + } + + new_size -= cur_pages; + if (!new_size) { + r = amdgpu_mm_node_addr(bo, ++new_mm, new_mem, + &new_start); + if (r) + goto error; + + new_size = new_mm->size; + } else { + new_start += cur_pages * PAGE_SIZE; + } + } r = ttm_bo_pipeline_move(bo, fence, evict, new_mem); fence_put(fence); return r; + +error: + if (fence) + fence_wait(fence, false); + fence_put(fence); + return r; } static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo, -- cgit v1.2.1 From 6a7f76e70fac0b1d41e7c5e193c33b1ee7e902c6 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Christian=20K=C3=B6nig?= Date: Wed, 24 Aug 2016 15:51:49 +0200 Subject: drm/amdgpu: add VRAM manager v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Split VRAM allocations into 4MB blocks. v2: fix typo in comment, some suggested cleanups v3: document how to disable the feature, fix rebase issue Signed-off-by: Christian König Reviewed-by: Edward O'Callaghan Tested-by: Mike Lothian Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index 1427c40c70bb..f757f68418b7 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -168,7 +168,7 @@ static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type, break; case TTM_PL_VRAM: /* "On-card" video ram */ - man->func = &ttm_bo_manager_func; + man->func = &amdgpu_vram_mgr_func; man->gpu_offset = adev->mc.vram_start; man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_MAPPABLE; -- cgit v1.2.1 From a7d64de659946e852cd8f2a9691a21ddbb4ebc86 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Christian=20K=C3=B6nig?= Date: Thu, 15 Sep 2016 14:58:48 +0200 Subject: drm/amdgpu: remove adev pointer from struct amdgpu_bo v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit It's completely pointless to have two pointers to the device in the same structure. v2: rename function to amdgpu_ttm_adev, fix typos Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 43 +++++++++++++-------------------- 1 file changed, 17 insertions(+), 26 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index f757f68418b7..414fa8fa8488 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -51,16 +51,6 @@ static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev); static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev); -static struct amdgpu_device *amdgpu_get_adev(struct ttm_bo_device *bdev) -{ - struct amdgpu_mman *mman; - struct amdgpu_device *adev; - - mman = container_of(bdev, struct amdgpu_mman, bdev); - adev = container_of(mman, struct amdgpu_device, mman); - return adev; -} - /* * Global memory. @@ -150,7 +140,7 @@ static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type, { struct amdgpu_device *adev; - adev = amdgpu_get_adev(bdev); + adev = amdgpu_ttm_adev(bdev); switch (type) { case TTM_PL_SYSTEM: @@ -195,6 +185,7 @@ static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type, static void amdgpu_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *placement) { + struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); struct amdgpu_bo *abo; static struct ttm_place placements = { .fpfn = 0, @@ -213,7 +204,7 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo, abo = container_of(bo, struct amdgpu_bo, tbo); switch (bo->mem.mem_type) { case TTM_PL_VRAM: - if (abo->adev->mman.buffer_funcs_ring->ready == false) { + if (adev->mman.buffer_funcs_ring->ready == false) { amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU); } else { amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT); @@ -229,7 +220,7 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo, * allocating address space for the BO. */ abo->placements[i].lpfn = - abo->adev->mc.gtt_size >> PAGE_SHIFT; + adev->mc.gtt_size >> PAGE_SHIFT; } } break; @@ -290,7 +281,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem, struct ttm_mem_reg *old_mem) { - struct amdgpu_device *adev = amdgpu_get_adev(bo->bdev); + struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; struct drm_mm_node *old_mm, *new_mm; @@ -384,7 +375,7 @@ static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo, struct ttm_placement placement; int r; - adev = amdgpu_get_adev(bo->bdev); + adev = amdgpu_ttm_adev(bo->bdev); tmp_mem = *new_mem; tmp_mem.mm_node = NULL; placement.num_placement = 1; @@ -431,7 +422,7 @@ static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo, struct ttm_place placements; int r; - adev = amdgpu_get_adev(bo->bdev); + adev = amdgpu_ttm_adev(bo->bdev); tmp_mem = *new_mem; tmp_mem.mm_node = NULL; placement.num_placement = 1; @@ -474,7 +465,7 @@ static int amdgpu_bo_move(struct ttm_buffer_object *bo, if (WARN_ON_ONCE(abo->pin_count > 0)) return -EINVAL; - adev = amdgpu_get_adev(bo->bdev); + adev = amdgpu_ttm_adev(bo->bdev); /* remember the eviction */ if (evict) @@ -527,7 +518,7 @@ memcpy: static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) { struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; - struct amdgpu_device *adev = amdgpu_get_adev(bdev); + struct amdgpu_device *adev = amdgpu_ttm_adev(bdev); mem->bus.addr = NULL; mem->bus.offset = 0; @@ -659,7 +650,7 @@ release_pages: /* prepare the sg table with the user pages */ static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm) { - struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev); + struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev); struct amdgpu_ttm_tt *gtt = (void *)ttm; unsigned nents; int r; @@ -691,7 +682,7 @@ release_sg: static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm) { - struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev); + struct amdgpu_device *adev = amdgpu_ttm_adev(ttm->bdev); struct amdgpu_ttm_tt *gtt = (void *)ttm; struct sg_page_iter sg_iter; @@ -851,7 +842,7 @@ static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev, struct amdgpu_device *adev; struct amdgpu_ttm_tt *gtt; - adev = amdgpu_get_adev(bdev); + adev = amdgpu_ttm_adev(bdev); gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL); if (gtt == NULL) { @@ -895,7 +886,7 @@ static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm) return 0; } - adev = amdgpu_get_adev(ttm->bdev); + adev = amdgpu_ttm_adev(ttm->bdev); #ifdef CONFIG_SWIOTLB if (swiotlb_nr_tbl()) { @@ -941,7 +932,7 @@ static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm) if (slave) return; - adev = amdgpu_get_adev(ttm->bdev); + adev = amdgpu_ttm_adev(ttm->bdev); #ifdef CONFIG_SWIOTLB if (swiotlb_nr_tbl()) { @@ -1064,7 +1055,7 @@ uint32_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm, static void amdgpu_ttm_lru_removal(struct ttm_buffer_object *tbo) { - struct amdgpu_device *adev = amdgpu_get_adev(tbo->bdev); + struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev); unsigned i, j; for (i = 0; i < AMDGPU_TTM_LRU_SIZE; ++i) { @@ -1081,7 +1072,7 @@ static void amdgpu_ttm_lru_removal(struct ttm_buffer_object *tbo) static struct amdgpu_mman_lru *amdgpu_ttm_lru(struct ttm_buffer_object *tbo) { - struct amdgpu_device *adev = amdgpu_get_adev(tbo->bdev); + struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev); unsigned log2_size = min(ilog2(tbo->num_pages), AMDGPU_TTM_LRU_SIZE - 1); @@ -1370,7 +1361,7 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo, struct reservation_object *resv, struct fence **fence) { - struct amdgpu_device *adev = bo->adev; + struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); struct amdgpu_job *job; struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; -- cgit v1.2.1 From a2ab19fed9d1dc5a7a2ced44f4b289885c522a8f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Christian=20K=C3=B6nig?= Date: Tue, 30 Aug 2016 17:26:04 +0200 Subject: drm/ttm: make eviction decision a driver callback v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This way the driver can decide if it is valuable to evict a BO or not. The current implementation is added as default to all existing drivers. v2: fix some typos found during internal testing Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 1 + 1 file changed, 1 insertion(+) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index 414fa8fa8488..fb646c44faae 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -1109,6 +1109,7 @@ static struct ttm_bo_driver amdgpu_bo_driver = { .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate, .invalidate_caches = &amdgpu_invalidate_caches, .init_mem_type = &amdgpu_init_mem_type, + .eviction_valuable = ttm_bo_eviction_valuable, .evict_flags = &amdgpu_evict_flags, .move = &amdgpu_bo_move, .verify_access = &amdgpu_verify_access, -- cgit v1.2.1 From 9982ca681e430907cb7f4e19e2ae4845031d5d59 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Christian=20K=C3=B6nig?= Date: Wed, 19 Oct 2016 14:44:22 +0200 Subject: drm/amdgpu: add amdgpu_ttm_bo_eviction_valuable callback MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This way we can correctly check split VRAM buffers as well. Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 26 +++++++++++++++++++++++++- 1 file changed, 25 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index fb646c44faae..f1a206df9823 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -1103,13 +1103,37 @@ static struct list_head *amdgpu_ttm_swap_lru_tail(struct ttm_buffer_object *tbo) return res; } +static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, + const struct ttm_place *place) +{ + if (bo->mem.mem_type == TTM_PL_VRAM && + bo->mem.start == AMDGPU_BO_INVALID_OFFSET) { + unsigned long num_pages = bo->mem.num_pages; + struct drm_mm_node *node = bo->mem.mm_node; + + /* Check each drm MM node individually */ + while (num_pages) { + if (place->fpfn < (node->start + node->size) && + !(place->lpfn && place->lpfn <= node->start)) + return true; + + num_pages -= node->size; + ++node; + } + + return false; + } + + return ttm_bo_eviction_valuable(bo, place); +} + static struct ttm_bo_driver amdgpu_bo_driver = { .ttm_tt_create = &amdgpu_ttm_tt_create, .ttm_tt_populate = &amdgpu_ttm_tt_populate, .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate, .invalidate_caches = &amdgpu_invalidate_caches, .init_mem_type = &amdgpu_init_mem_type, - .eviction_valuable = ttm_bo_eviction_valuable, + .eviction_valuable = amdgpu_ttm_bo_eviction_valuable, .evict_flags = &amdgpu_evict_flags, .move = &amdgpu_bo_move, .verify_access = &amdgpu_verify_access, -- cgit v1.2.1