From: Marek Olšák <marek.ol...@amd.com> Dependencies between rings are inserted correctly if a buffer is represented by only one unique amdgpu_winsys_bo instance. Use a hash table keyed by amdgpu_bo_handle to have exactly one amdgpu_winsys_bo per amdgpu_bo_handle.
v2: return offset and stride properly --- src/gallium/winsys/amdgpu/drm/amdgpu_bo.c | 36 ++++++++++++++++--- src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c | 5 +++ src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h | 5 +++ 3 files changed, 41 insertions(+), 5 deletions(-) diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c index d9192c209e2..80563d3df98 100644 --- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c +++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c @@ -21,20 +21,21 @@ * USE OR OTHER DEALINGS IN THE SOFTWARE. * * The above copyright notice and this permission notice (including the * next paragraph) shall be included in all copies or substantial portions * of the Software. */ #include "amdgpu_cs.h" #include "util/os_time.h" +#include "util/u_hash_table.h" #include "state_tracker/drm_driver.h" #include <amdgpu_drm.h> #include <xf86drm.h> #include <stdio.h> #include <inttypes.h> #ifndef AMDGPU_GEM_CREATE_VM_ALWAYS_VALID #define AMDGPU_GEM_CREATE_VM_ALWAYS_VALID (1 << 6) #endif @@ -172,20 +173,24 @@ void amdgpu_bo_destroy(struct pb_buffer *_buf) assert(bo->bo && "must not be called for slab entries"); if (ws->debug_all_bos) { simple_mtx_lock(&ws->global_bo_list_lock); LIST_DEL(&bo->u.real.global_list_item); ws->num_buffers--; simple_mtx_unlock(&ws->global_bo_list_lock); } + simple_mtx_lock(&ws->bo_export_table_lock); + util_hash_table_remove(ws->bo_export_table, bo->bo); + simple_mtx_unlock(&ws->bo_export_table_lock); + amdgpu_bo_va_op(bo->bo, 0, bo->base.size, bo->va, 0, AMDGPU_VA_OP_UNMAP); amdgpu_va_range_free(bo->u.real.va_handle); amdgpu_bo_free(bo->bo); amdgpu_bo_remove_fences(bo); if (bo->initial_domain & RADEON_DOMAIN_VRAM) ws->allocated_vram -= align64(bo->base.size, ws->info.gart_page_size); else if (bo->initial_domain & RADEON_DOMAIN_GTT) ws->allocated_gtt -= align64(bo->base.size, ws->info.gart_page_size); @@ -1278,24 +1283,41 @@ static struct pb_buffer *amdgpu_bo_from_handle(struct radeon_winsys *rws, case WINSYS_HANDLE_TYPE_SHARED: type = amdgpu_bo_handle_type_gem_flink_name; break; case WINSYS_HANDLE_TYPE_FD: type = amdgpu_bo_handle_type_dma_buf_fd; break; default: return NULL; } + if (stride) + *stride = whandle->stride; + if (offset) + *offset = whandle->offset; + r = amdgpu_bo_import(ws->dev, type, whandle->handle, &result); if (r) return NULL; + simple_mtx_lock(&ws->bo_export_table_lock); + bo = util_hash_table_get(ws->bo_export_table, result.buf_handle); + + /* If the amdgpu_winsys_bo instance already exists, bump the reference + * counter and return it. + */ + if (bo) { + p_atomic_inc(&bo->base.reference.count); + simple_mtx_unlock(&ws->bo_export_table_lock); + return &bo->base; + } + /* Get initial domains. */ r = amdgpu_bo_query_info(result.buf_handle, &info); if (r) goto error; r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general, result.alloc_size, 1 << 20, 0, &va, &va_handle, AMDGPU_VA_RANGE_HIGH); if (r) goto error; @@ -1319,49 +1341,49 @@ static struct pb_buffer *amdgpu_bo_from_handle(struct radeon_winsys *rws, bo->bo = result.buf_handle; bo->base.size = result.alloc_size; bo->base.vtbl = &amdgpu_winsys_bo_vtbl; bo->ws = ws; bo->va = va; bo->u.real.va_handle = va_handle; bo->initial_domain = initial; bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1); bo->is_shared = true; - if (stride) - *stride = whandle->stride; - if (offset) - *offset = whandle->offset; - if (bo->initial_domain & RADEON_DOMAIN_VRAM) ws->allocated_vram += align64(bo->base.size, ws->info.gart_page_size); else if (bo->initial_domain & RADEON_DOMAIN_GTT) ws->allocated_gtt += align64(bo->base.size, ws->info.gart_page_size); amdgpu_add_buffer_to_global_list(bo); + util_hash_table_set(ws->bo_export_table, bo->bo, bo); + simple_mtx_unlock(&ws->bo_export_table_lock); + return &bo->base; error: + simple_mtx_unlock(&ws->bo_export_table_lock); if (bo) FREE(bo); if (va_handle) amdgpu_va_range_free(va_handle); amdgpu_bo_free(result.buf_handle); return NULL; } static bool amdgpu_bo_get_handle(struct pb_buffer *buffer, unsigned stride, unsigned offset, unsigned slice_size, struct winsys_handle *whandle) { struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(buffer); + struct amdgpu_winsys *ws = bo->ws; enum amdgpu_bo_handle_type type; int r; /* Don't allow exports of slab entries and sparse buffers. */ if (!bo->bo) return false; bo->u.real.use_reusable_pool = false; switch (whandle->type) { @@ -1375,20 +1397,24 @@ static bool amdgpu_bo_get_handle(struct pb_buffer *buffer, type = amdgpu_bo_handle_type_kms; break; default: return false; } r = amdgpu_bo_export(bo->bo, type, &whandle->handle); if (r) return false; + simple_mtx_lock(&ws->bo_export_table_lock); + util_hash_table_set(ws->bo_export_table, bo->bo, bo); + simple_mtx_unlock(&ws->bo_export_table_lock); + whandle->stride = stride; whandle->offset = offset; whandle->offset += slice_size * whandle->layer; bo->is_shared = true; return true; } static struct pb_buffer *amdgpu_bo_from_ptr(struct radeon_winsys *rws, void *pointer, uint64_t size) { diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c b/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c index db7a4d7033c..882f500bc69 100644 --- a/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c +++ b/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c @@ -85,21 +85,23 @@ static void amdgpu_winsys_destroy(struct radeon_winsys *rws) if (ws->reserve_vmid) amdgpu_vm_unreserve_vmid(ws->dev, 0); if (util_queue_is_initialized(&ws->cs_queue)) util_queue_destroy(&ws->cs_queue); simple_mtx_destroy(&ws->bo_fence_lock); pb_slabs_deinit(&ws->bo_slabs); pb_cache_deinit(&ws->bo_cache); + util_hash_table_destroy(ws->bo_export_table); simple_mtx_destroy(&ws->global_bo_list_lock); + simple_mtx_destroy(&ws->bo_export_table_lock); do_winsys_deinit(ws); FREE(rws); } static void amdgpu_winsys_query_info(struct radeon_winsys *rws, struct radeon_info *info) { *info = ((struct amdgpu_winsys *)rws)->info; } @@ -307,22 +309,25 @@ amdgpu_winsys_create(int fd, const struct pipe_screen_config *config, ws->base.cs_request_feature = amdgpu_cs_request_feature; ws->base.query_value = amdgpu_query_value; ws->base.read_registers = amdgpu_read_registers; ws->base.get_chip_name = amdgpu_get_chip_name; amdgpu_bo_init_functions(ws); amdgpu_cs_init_functions(ws); amdgpu_surface_init_functions(ws); LIST_INITHEAD(&ws->global_bo_list); + ws->bo_export_table = util_hash_table_create(hash_pointer, compare_pointers); + (void) simple_mtx_init(&ws->global_bo_list_lock, mtx_plain); (void) simple_mtx_init(&ws->bo_fence_lock, mtx_plain); + (void) simple_mtx_init(&ws->bo_export_table_lock, mtx_plain); if (!util_queue_init(&ws->cs_queue, "cs", 8, 1, UTIL_QUEUE_INIT_RESIZE_IF_FULL)) { amdgpu_winsys_destroy(&ws->base); simple_mtx_unlock(&dev_tab_mutex); return NULL; } /* Create the screen at the end. The winsys must be initialized * completely. diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h b/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h index 8079255e4cf..c355eff5262 100644 --- a/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h +++ b/src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h @@ -78,20 +78,25 @@ struct amdgpu_winsys { bool check_vm; bool debug_all_bos; bool reserve_vmid; bool zero_all_vram_allocs; /* List of all allocated buffers */ simple_mtx_t global_bo_list_lock; struct list_head global_bo_list; unsigned num_buffers; + + /* For returning the same amdgpu_winsys_bo instance for exported + * and re-imported buffers. */ + struct util_hash_table *bo_export_table; + simple_mtx_t bo_export_table_lock; }; static inline struct amdgpu_winsys * amdgpu_winsys(struct radeon_winsys *base) { return (struct amdgpu_winsys*)base; } void amdgpu_surface_init_functions(struct amdgpu_winsys *ws); -- 2.17.1 _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev