X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fwinsys%2Famdgpu%2Fdrm%2Famdgpu_bo.c;h=3ee38b8a79f8a19f5b11d142d1d35570d60912fd;hb=82aa07f81fcc5ed696eea16f48cec7e39c3cd3d1;hp=c805da8a18ab52b78a0840bc722e5c8b4d046af3;hpb=f4ab7a5415bb1e54c19bf57f923c47c2c3d4f408;p=mesa.git diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c index c805da8a18a..3ee38b8a79f 100644 --- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c +++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c @@ -24,20 +24,32 @@ * next paragraph) shall be included in all copies or substantial portions * of the Software. */ -/* - * Authors: - * Marek Olšák - */ #include "amdgpu_cs.h" -#include "os/os_time.h" +#include "util/os_time.h" +#include "util/u_hash_table.h" #include "state_tracker/drm_driver.h" #include #include #include #include +#ifndef AMDGPU_GEM_CREATE_VM_ALWAYS_VALID +#define AMDGPU_GEM_CREATE_VM_ALWAYS_VALID (1 << 6) +#endif + +#ifndef AMDGPU_VA_RANGE_HIGH +#define AMDGPU_VA_RANGE_HIGH 0x2 +#endif + +/* Set to 1 for verbose output showing committed sparse buffer ranges. */ +#define DEBUG_SPARSE_COMMITS 0 + +struct amdgpu_sparse_backing_chunk { + uint32_t begin, end; +}; + static struct pb_buffer * amdgpu_bo_create(struct radeon_winsys *rws, uint64_t size, @@ -83,7 +95,7 @@ static bool amdgpu_bo_wait(struct pb_buffer *_buf, uint64_t timeout, unsigned idle_fences; bool buffer_idle; - mtx_lock(&ws->bo_fence_lock); + simple_mtx_lock(&ws->bo_fence_lock); for (idle_fences = 0; idle_fences < bo->num_fences; ++idle_fences) { if (!amdgpu_fence_wait(bo->fences[idle_fences], 0, false)) @@ -99,13 +111,13 @@ static bool amdgpu_bo_wait(struct pb_buffer *_buf, uint64_t timeout, bo->num_fences -= idle_fences; buffer_idle = !bo->num_fences; - mtx_unlock(&ws->bo_fence_lock); + simple_mtx_unlock(&ws->bo_fence_lock); return buffer_idle; } else { bool buffer_idle = true; - mtx_lock(&ws->bo_fence_lock); + simple_mtx_lock(&ws->bo_fence_lock); while (bo->num_fences && buffer_idle) { struct pipe_fence_handle *fence = NULL; bool fence_idle = false; @@ -113,12 +125,12 @@ static bool amdgpu_bo_wait(struct pb_buffer *_buf, uint64_t timeout, amdgpu_fence_reference(&fence, bo->fences[0]); /* Wait for the fence. */ - mtx_unlock(&ws->bo_fence_lock); + simple_mtx_unlock(&ws->bo_fence_lock); if (amdgpu_fence_wait(fence, abs_timeout, true)) fence_idle = true; else buffer_idle = false; - mtx_lock(&ws->bo_fence_lock); + simple_mtx_lock(&ws->bo_fence_lock); /* Release an idle fence to avoid checking it again later, keeping in * mind that the fence array may have been modified by other threads. @@ -132,7 +144,7 @@ static bool amdgpu_bo_wait(struct pb_buffer *_buf, uint64_t timeout, amdgpu_fence_reference(&fence, NULL); } - mtx_unlock(&ws->bo_fence_lock); + simple_mtx_unlock(&ws->bo_fence_lock); return buffer_idle; } @@ -157,13 +169,20 @@ static void amdgpu_bo_remove_fences(struct amdgpu_winsys_bo *bo) void amdgpu_bo_destroy(struct pb_buffer *_buf) { struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf); + struct amdgpu_winsys *ws = bo->ws; assert(bo->bo && "must not be called for slab entries"); - mtx_lock(&bo->ws->global_bo_list_lock); - LIST_DEL(&bo->u.real.global_list_item); - bo->ws->num_buffers--; - mtx_unlock(&bo->ws->global_bo_list_lock); + 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); @@ -172,16 +191,16 @@ void amdgpu_bo_destroy(struct pb_buffer *_buf) amdgpu_bo_remove_fences(bo); if (bo->initial_domain & RADEON_DOMAIN_VRAM) - bo->ws->allocated_vram -= align64(bo->base.size, bo->ws->info.gart_page_size); + ws->allocated_vram -= align64(bo->base.size, ws->info.gart_page_size); else if (bo->initial_domain & RADEON_DOMAIN_GTT) - bo->ws->allocated_gtt -= align64(bo->base.size, bo->ws->info.gart_page_size); + ws->allocated_gtt -= align64(bo->base.size, ws->info.gart_page_size); if (bo->u.real.map_count >= 1) { if (bo->initial_domain & RADEON_DOMAIN_VRAM) - bo->ws->mapped_vram -= bo->base.size; + ws->mapped_vram -= bo->base.size; else if (bo->initial_domain & RADEON_DOMAIN_GTT) - bo->ws->mapped_gtt -= bo->base.size; - bo->ws->num_mapped_buffers--; + ws->mapped_gtt -= bo->base.size; + ws->num_mapped_buffers--; } FREE(bo); @@ -200,7 +219,7 @@ static void amdgpu_bo_destroy_or_cache(struct pb_buffer *_buf) } static void *amdgpu_bo_map(struct pb_buffer *buf, - struct radeon_winsys_cs *rcs, + struct radeon_cmdbuf *rcs, enum pipe_transfer_usage usage) { struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf; @@ -210,6 +229,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf, void *cpu = NULL; uint64_t offset = 0; + assert(!bo->sparse); + /* If it's not unsynchronized bo_map, flush CS if needed and then wait. */ if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) { /* DONTBLOCK doesn't make sense with UNSYNCHRONIZED. */ @@ -224,7 +245,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf, * Only check whether the buffer is being used for write. */ if (cs && amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo, RADEON_USAGE_WRITE)) { - cs->flush_cs(cs->flush_data, RADEON_FLUSH_ASYNC, NULL); + cs->flush_cs(cs->flush_data, + RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL); return NULL; } @@ -234,7 +256,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf, } } else { if (cs && amdgpu_bo_is_referenced_by_cs(cs, bo)) { - cs->flush_cs(cs->flush_data, RADEON_FLUSH_ASYNC, NULL); + cs->flush_cs(cs->flush_data, + RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL); return NULL; } @@ -257,7 +280,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf, if (cs) { if (amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo, RADEON_USAGE_WRITE)) { - cs->flush_cs(cs->flush_data, 0, NULL); + cs->flush_cs(cs->flush_data, + RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL); } else { /* Try to avoid busy-waiting in amdgpu_bo_wait. */ if (p_atomic_read(&bo->num_active_ioctls)) @@ -271,7 +295,8 @@ static void *amdgpu_bo_map(struct pb_buffer *buf, /* Mapping for write. */ if (cs) { if (amdgpu_bo_is_referenced_by_cs(cs, bo)) { - cs->flush_cs(cs->flush_data, 0, NULL); + cs->flush_cs(cs->flush_data, + RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL); } else { /* Try to avoid busy-waiting in amdgpu_bo_wait. */ if (p_atomic_read(&bo->num_active_ioctls)) @@ -322,6 +347,8 @@ static void amdgpu_bo_unmap(struct pb_buffer *buf) struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf; struct amdgpu_winsys_bo *real; + assert(!bo->sparse); + if (bo->user_ptr) return; @@ -349,19 +376,20 @@ static void amdgpu_add_buffer_to_global_list(struct amdgpu_winsys_bo *bo) assert(bo->bo); - mtx_lock(&ws->global_bo_list_lock); - LIST_ADDTAIL(&bo->u.real.global_list_item, &ws->global_bo_list); - ws->num_buffers++; - mtx_unlock(&ws->global_bo_list_lock); + if (ws->debug_all_bos) { + simple_mtx_lock(&ws->global_bo_list_lock); + LIST_ADDTAIL(&bo->u.real.global_list_item, &ws->global_bo_list); + ws->num_buffers++; + simple_mtx_unlock(&ws->global_bo_list_lock); + } } static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws, uint64_t size, unsigned alignment, - unsigned usage, enum radeon_bo_domain initial_domain, unsigned flags, - unsigned pb_cache_bucket) + int heap) { struct amdgpu_bo_alloc_request request = {0}; amdgpu_bo_handle buf_handle; @@ -371,14 +399,18 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws, unsigned va_gap_size; int r; - assert(initial_domain & RADEON_DOMAIN_VRAM_GTT); + /* VRAM or GTT must be specified, but not both at the same time. */ + assert(util_bitcount(initial_domain & RADEON_DOMAIN_VRAM_GTT) == 1); + bo = CALLOC_STRUCT(amdgpu_winsys_bo); if (!bo) { return NULL; } - pb_cache_init_entry(&ws->bo_cache, &bo->u.real.cache_entry, &bo->base, - pb_cache_bucket); + if (heap >= 0) { + pb_cache_init_entry(&ws->bo_cache, &bo->u.real.cache_entry, &bo->base, + heap); + } request.alloc_size = size; request.phys_alignment = alignment; @@ -387,12 +419,24 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws, if (initial_domain & RADEON_DOMAIN_GTT) request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT; - if (flags & RADEON_FLAG_CPU_ACCESS) - request.flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; + /* Since VRAM and GTT have almost the same performance on APUs, we could + * just set GTT. However, in order to decrease GTT(RAM) usage, which is + * shared with the OS, allow VRAM placements too. The idea is not to use + * VRAM usefully, but to use it so that it's not unused and wasted. + */ + if (!ws->info.has_dedicated_vram) + request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT; + if (flags & RADEON_FLAG_NO_CPU_ACCESS) request.flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS; if (flags & RADEON_FLAG_GTT_WC) request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC; + if (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING && + ws->info.has_local_buffers) + request.flags |= AMDGPU_GEM_CREATE_VM_ALWAYS_VALID; + if (ws->zero_all_vram_allocs && + (request.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM)) + request.flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED; r = amdgpu_bo_alloc(ws->dev, &request, &buf_handle); if (r) { @@ -404,18 +448,29 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws, } va_gap_size = ws->check_vm ? MAX2(4 * alignment, 64 * 1024) : 0; + if (size > ws->info.pte_fragment_size) + alignment = MAX2(alignment, ws->info.pte_fragment_size); r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general, - size + va_gap_size, alignment, 0, &va, &va_handle, 0); + size + va_gap_size, alignment, 0, &va, &va_handle, + (flags & RADEON_FLAG_32BIT ? AMDGPU_VA_RANGE_32_BIT : 0) | + AMDGPU_VA_RANGE_HIGH); if (r) goto error_va_alloc; - r = amdgpu_bo_va_op(buf_handle, 0, size, va, 0, AMDGPU_VA_OP_MAP); + unsigned vm_flags = AMDGPU_VM_PAGE_READABLE | + AMDGPU_VM_PAGE_EXECUTABLE; + + if (!(flags & RADEON_FLAG_READ_ONLY)) + vm_flags |= AMDGPU_VM_PAGE_WRITEABLE; + + r = amdgpu_bo_va_op_raw(ws->dev, buf_handle, 0, size, va, vm_flags, + AMDGPU_VA_OP_MAP); if (r) goto error_va_map; pipe_reference_init(&bo->base.reference, 1); bo->base.alignment = alignment; - bo->base.usage = usage; + bo->base.usage = 0; bo->base.size = size; bo->base.vtbl = &amdgpu_winsys_bo_vtbl; bo->ws = ws; @@ -424,12 +479,15 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws, bo->u.real.va_handle = va_handle; bo->initial_domain = initial_domain; bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1); + bo->is_local = !!(request.flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID); if (initial_domain & RADEON_DOMAIN_VRAM) ws->allocated_vram += align64(size, ws->info.gart_page_size); else if (initial_domain & RADEON_DOMAIN_GTT) ws->allocated_gtt += align64(size, ws->info.gart_page_size); + amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle); + amdgpu_add_buffer_to_global_list(bo); return bo; @@ -484,33 +542,16 @@ struct pb_slab *amdgpu_bo_slab_alloc(void *priv, unsigned heap, { struct amdgpu_winsys *ws = priv; struct amdgpu_slab *slab = CALLOC_STRUCT(amdgpu_slab); - enum radeon_bo_domain domains; - enum radeon_bo_flag flags = 0; + enum radeon_bo_domain domains = radeon_domain_from_heap(heap); + enum radeon_bo_flag flags = radeon_flags_from_heap(heap); uint32_t base_id; if (!slab) return NULL; - if (heap & 1) - flags |= RADEON_FLAG_GTT_WC; - if (heap & 2) - flags |= RADEON_FLAG_CPU_ACCESS; - - switch (heap >> 2) { - case 0: - domains = RADEON_DOMAIN_VRAM; - break; - default: - case 1: - domains = RADEON_DOMAIN_VRAM_GTT; - break; - case 2: - domains = RADEON_DOMAIN_GTT; - break; - } - + unsigned slab_size = 1 << AMDGPU_SLAB_BO_SIZE_LOG2; slab->buffer = amdgpu_winsys_bo(amdgpu_bo_create(&ws->base, - 64 * 1024, 64 * 1024, + slab_size, slab_size, domains, flags)); if (!slab->buffer) goto fail; @@ -566,6 +607,463 @@ void amdgpu_bo_slab_free(void *priv, struct pb_slab *pslab) FREE(slab); } +#if DEBUG_SPARSE_COMMITS +static void +sparse_dump(struct amdgpu_winsys_bo *bo, const char *func) +{ + fprintf(stderr, "%s: %p (size=%"PRIu64", num_va_pages=%u) @ %s\n" + "Commitments:\n", + __func__, bo, bo->base.size, bo->u.sparse.num_va_pages, func); + + struct amdgpu_sparse_backing *span_backing = NULL; + uint32_t span_first_backing_page = 0; + uint32_t span_first_va_page = 0; + uint32_t va_page = 0; + + for (;;) { + struct amdgpu_sparse_backing *backing = 0; + uint32_t backing_page = 0; + + if (va_page < bo->u.sparse.num_va_pages) { + backing = bo->u.sparse.commitments[va_page].backing; + backing_page = bo->u.sparse.commitments[va_page].page; + } + + if (span_backing && + (backing != span_backing || + backing_page != span_first_backing_page + (va_page - span_first_va_page))) { + fprintf(stderr, " %u..%u: backing=%p:%u..%u\n", + span_first_va_page, va_page - 1, span_backing, + span_first_backing_page, + span_first_backing_page + (va_page - span_first_va_page) - 1); + + span_backing = NULL; + } + + if (va_page >= bo->u.sparse.num_va_pages) + break; + + if (backing && !span_backing) { + span_backing = backing; + span_first_backing_page = backing_page; + span_first_va_page = va_page; + } + + va_page++; + } + + fprintf(stderr, "Backing:\n"); + + list_for_each_entry(struct amdgpu_sparse_backing, backing, &bo->u.sparse.backing, list) { + fprintf(stderr, " %p (size=%"PRIu64")\n", backing, backing->bo->base.size); + for (unsigned i = 0; i < backing->num_chunks; ++i) + fprintf(stderr, " %u..%u\n", backing->chunks[i].begin, backing->chunks[i].end); + } +} +#endif + +/* + * Attempt to allocate the given number of backing pages. Fewer pages may be + * allocated (depending on the fragmentation of existing backing buffers), + * which will be reflected by a change to *pnum_pages. + */ +static struct amdgpu_sparse_backing * +sparse_backing_alloc(struct amdgpu_winsys_bo *bo, uint32_t *pstart_page, uint32_t *pnum_pages) +{ + struct amdgpu_sparse_backing *best_backing; + unsigned best_idx; + uint32_t best_num_pages; + + best_backing = NULL; + best_idx = 0; + best_num_pages = 0; + + /* This is a very simple and inefficient best-fit algorithm. */ + list_for_each_entry(struct amdgpu_sparse_backing, backing, &bo->u.sparse.backing, list) { + for (unsigned idx = 0; idx < backing->num_chunks; ++idx) { + uint32_t cur_num_pages = backing->chunks[idx].end - backing->chunks[idx].begin; + if ((best_num_pages < *pnum_pages && cur_num_pages > best_num_pages) || + (best_num_pages > *pnum_pages && cur_num_pages < best_num_pages)) { + best_backing = backing; + best_idx = idx; + best_num_pages = cur_num_pages; + } + } + } + + /* Allocate a new backing buffer if necessary. */ + if (!best_backing) { + struct pb_buffer *buf; + uint64_t size; + uint32_t pages; + + best_backing = CALLOC_STRUCT(amdgpu_sparse_backing); + if (!best_backing) + return NULL; + + best_backing->max_chunks = 4; + best_backing->chunks = CALLOC(best_backing->max_chunks, + sizeof(*best_backing->chunks)); + if (!best_backing->chunks) { + FREE(best_backing); + return NULL; + } + + assert(bo->u.sparse.num_backing_pages < DIV_ROUND_UP(bo->base.size, RADEON_SPARSE_PAGE_SIZE)); + + size = MIN3(bo->base.size / 16, + 8 * 1024 * 1024, + bo->base.size - (uint64_t)bo->u.sparse.num_backing_pages * RADEON_SPARSE_PAGE_SIZE); + size = MAX2(size, RADEON_SPARSE_PAGE_SIZE); + + buf = amdgpu_bo_create(&bo->ws->base, size, RADEON_SPARSE_PAGE_SIZE, + bo->initial_domain, + bo->u.sparse.flags | RADEON_FLAG_NO_SUBALLOC); + if (!buf) { + FREE(best_backing->chunks); + FREE(best_backing); + return NULL; + } + + /* We might have gotten a bigger buffer than requested via caching. */ + pages = buf->size / RADEON_SPARSE_PAGE_SIZE; + + best_backing->bo = amdgpu_winsys_bo(buf); + best_backing->num_chunks = 1; + best_backing->chunks[0].begin = 0; + best_backing->chunks[0].end = pages; + + list_add(&best_backing->list, &bo->u.sparse.backing); + bo->u.sparse.num_backing_pages += pages; + + best_idx = 0; + best_num_pages = pages; + } + + *pnum_pages = MIN2(*pnum_pages, best_num_pages); + *pstart_page = best_backing->chunks[best_idx].begin; + best_backing->chunks[best_idx].begin += *pnum_pages; + + if (best_backing->chunks[best_idx].begin >= best_backing->chunks[best_idx].end) { + memmove(&best_backing->chunks[best_idx], &best_backing->chunks[best_idx + 1], + sizeof(*best_backing->chunks) * (best_backing->num_chunks - best_idx - 1)); + best_backing->num_chunks--; + } + + return best_backing; +} + +static void +sparse_free_backing_buffer(struct amdgpu_winsys_bo *bo, + struct amdgpu_sparse_backing *backing) +{ + struct amdgpu_winsys *ws = backing->bo->ws; + + bo->u.sparse.num_backing_pages -= backing->bo->base.size / RADEON_SPARSE_PAGE_SIZE; + + simple_mtx_lock(&ws->bo_fence_lock); + amdgpu_add_fences(backing->bo, bo->num_fences, bo->fences); + simple_mtx_unlock(&ws->bo_fence_lock); + + list_del(&backing->list); + amdgpu_winsys_bo_reference(&backing->bo, NULL); + FREE(backing->chunks); + FREE(backing); +} + +/* + * Return a range of pages from the given backing buffer back into the + * free structure. + */ +static bool +sparse_backing_free(struct amdgpu_winsys_bo *bo, + struct amdgpu_sparse_backing *backing, + uint32_t start_page, uint32_t num_pages) +{ + uint32_t end_page = start_page + num_pages; + unsigned low = 0; + unsigned high = backing->num_chunks; + + /* Find the first chunk with begin >= start_page. */ + while (low < high) { + unsigned mid = low + (high - low) / 2; + + if (backing->chunks[mid].begin >= start_page) + high = mid; + else + low = mid + 1; + } + + assert(low >= backing->num_chunks || end_page <= backing->chunks[low].begin); + assert(low == 0 || backing->chunks[low - 1].end <= start_page); + + if (low > 0 && backing->chunks[low - 1].end == start_page) { + backing->chunks[low - 1].end = end_page; + + if (low < backing->num_chunks && end_page == backing->chunks[low].begin) { + backing->chunks[low - 1].end = backing->chunks[low].end; + memmove(&backing->chunks[low], &backing->chunks[low + 1], + sizeof(*backing->chunks) * (backing->num_chunks - low - 1)); + backing->num_chunks--; + } + } else if (low < backing->num_chunks && end_page == backing->chunks[low].begin) { + backing->chunks[low].begin = start_page; + } else { + if (backing->num_chunks >= backing->max_chunks) { + unsigned new_max_chunks = 2 * backing->max_chunks; + struct amdgpu_sparse_backing_chunk *new_chunks = + REALLOC(backing->chunks, + sizeof(*backing->chunks) * backing->max_chunks, + sizeof(*backing->chunks) * new_max_chunks); + if (!new_chunks) + return false; + + backing->max_chunks = new_max_chunks; + backing->chunks = new_chunks; + } + + memmove(&backing->chunks[low + 1], &backing->chunks[low], + sizeof(*backing->chunks) * (backing->num_chunks - low)); + backing->chunks[low].begin = start_page; + backing->chunks[low].end = end_page; + backing->num_chunks++; + } + + if (backing->num_chunks == 1 && backing->chunks[0].begin == 0 && + backing->chunks[0].end == backing->bo->base.size / RADEON_SPARSE_PAGE_SIZE) + sparse_free_backing_buffer(bo, backing); + + return true; +} + +static void amdgpu_bo_sparse_destroy(struct pb_buffer *_buf) +{ + struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf); + int r; + + assert(!bo->bo && bo->sparse); + + r = amdgpu_bo_va_op_raw(bo->ws->dev, NULL, 0, + (uint64_t)bo->u.sparse.num_va_pages * RADEON_SPARSE_PAGE_SIZE, + bo->va, 0, AMDGPU_VA_OP_CLEAR); + if (r) { + fprintf(stderr, "amdgpu: clearing PRT VA region on destroy failed (%d)\n", r); + } + + while (!list_empty(&bo->u.sparse.backing)) { + struct amdgpu_sparse_backing *dummy = NULL; + sparse_free_backing_buffer(bo, + container_of(bo->u.sparse.backing.next, + dummy, list)); + } + + amdgpu_va_range_free(bo->u.sparse.va_handle); + simple_mtx_destroy(&bo->u.sparse.commit_lock); + FREE(bo->u.sparse.commitments); + FREE(bo); +} + +static const struct pb_vtbl amdgpu_winsys_bo_sparse_vtbl = { + amdgpu_bo_sparse_destroy + /* other functions are never called */ +}; + +static struct pb_buffer * +amdgpu_bo_sparse_create(struct amdgpu_winsys *ws, uint64_t size, + enum radeon_bo_domain domain, + enum radeon_bo_flag flags) +{ + struct amdgpu_winsys_bo *bo; + uint64_t map_size; + uint64_t va_gap_size; + int r; + + /* We use 32-bit page numbers; refuse to attempt allocating sparse buffers + * that exceed this limit. This is not really a restriction: we don't have + * that much virtual address space anyway. + */ + if (size > (uint64_t)INT32_MAX * RADEON_SPARSE_PAGE_SIZE) + return NULL; + + bo = CALLOC_STRUCT(amdgpu_winsys_bo); + if (!bo) + return NULL; + + pipe_reference_init(&bo->base.reference, 1); + bo->base.alignment = RADEON_SPARSE_PAGE_SIZE; + bo->base.size = size; + bo->base.vtbl = &amdgpu_winsys_bo_sparse_vtbl; + bo->ws = ws; + bo->initial_domain = domain; + bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1); + bo->sparse = true; + bo->u.sparse.flags = flags & ~RADEON_FLAG_SPARSE; + + bo->u.sparse.num_va_pages = DIV_ROUND_UP(size, RADEON_SPARSE_PAGE_SIZE); + bo->u.sparse.commitments = CALLOC(bo->u.sparse.num_va_pages, + sizeof(*bo->u.sparse.commitments)); + if (!bo->u.sparse.commitments) + goto error_alloc_commitments; + + simple_mtx_init(&bo->u.sparse.commit_lock, mtx_plain); + LIST_INITHEAD(&bo->u.sparse.backing); + + /* For simplicity, we always map a multiple of the page size. */ + map_size = align64(size, RADEON_SPARSE_PAGE_SIZE); + va_gap_size = ws->check_vm ? 4 * RADEON_SPARSE_PAGE_SIZE : 0; + r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general, + map_size + va_gap_size, RADEON_SPARSE_PAGE_SIZE, + 0, &bo->va, &bo->u.sparse.va_handle, + AMDGPU_VA_RANGE_HIGH); + if (r) + goto error_va_alloc; + + r = amdgpu_bo_va_op_raw(bo->ws->dev, NULL, 0, size, bo->va, + AMDGPU_VM_PAGE_PRT, AMDGPU_VA_OP_MAP); + if (r) + goto error_va_map; + + return &bo->base; + +error_va_map: + amdgpu_va_range_free(bo->u.sparse.va_handle); +error_va_alloc: + simple_mtx_destroy(&bo->u.sparse.commit_lock); + FREE(bo->u.sparse.commitments); +error_alloc_commitments: + FREE(bo); + return NULL; +} + +static bool +amdgpu_bo_sparse_commit(struct pb_buffer *buf, uint64_t offset, uint64_t size, + bool commit) +{ + struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(buf); + struct amdgpu_sparse_commitment *comm; + uint32_t va_page, end_va_page; + bool ok = true; + int r; + + assert(bo->sparse); + assert(offset % RADEON_SPARSE_PAGE_SIZE == 0); + assert(offset <= bo->base.size); + assert(size <= bo->base.size - offset); + assert(size % RADEON_SPARSE_PAGE_SIZE == 0 || offset + size == bo->base.size); + + comm = bo->u.sparse.commitments; + va_page = offset / RADEON_SPARSE_PAGE_SIZE; + end_va_page = va_page + DIV_ROUND_UP(size, RADEON_SPARSE_PAGE_SIZE); + + simple_mtx_lock(&bo->u.sparse.commit_lock); + +#if DEBUG_SPARSE_COMMITS + sparse_dump(bo, __func__); +#endif + + if (commit) { + while (va_page < end_va_page) { + uint32_t span_va_page; + + /* Skip pages that are already committed. */ + if (comm[va_page].backing) { + va_page++; + continue; + } + + /* Determine length of uncommitted span. */ + span_va_page = va_page; + while (va_page < end_va_page && !comm[va_page].backing) + va_page++; + + /* Fill the uncommitted span with chunks of backing memory. */ + while (span_va_page < va_page) { + struct amdgpu_sparse_backing *backing; + uint32_t backing_start, backing_size; + + backing_size = va_page - span_va_page; + backing = sparse_backing_alloc(bo, &backing_start, &backing_size); + if (!backing) { + ok = false; + goto out; + } + + r = amdgpu_bo_va_op_raw(bo->ws->dev, backing->bo->bo, + (uint64_t)backing_start * RADEON_SPARSE_PAGE_SIZE, + (uint64_t)backing_size * RADEON_SPARSE_PAGE_SIZE, + bo->va + (uint64_t)span_va_page * RADEON_SPARSE_PAGE_SIZE, + AMDGPU_VM_PAGE_READABLE | + AMDGPU_VM_PAGE_WRITEABLE | + AMDGPU_VM_PAGE_EXECUTABLE, + AMDGPU_VA_OP_REPLACE); + if (r) { + ok = sparse_backing_free(bo, backing, backing_start, backing_size); + assert(ok && "sufficient memory should already be allocated"); + + ok = false; + goto out; + } + + while (backing_size) { + comm[span_va_page].backing = backing; + comm[span_va_page].page = backing_start; + span_va_page++; + backing_start++; + backing_size--; + } + } + } + } else { + r = amdgpu_bo_va_op_raw(bo->ws->dev, NULL, 0, + (uint64_t)(end_va_page - va_page) * RADEON_SPARSE_PAGE_SIZE, + bo->va + (uint64_t)va_page * RADEON_SPARSE_PAGE_SIZE, + AMDGPU_VM_PAGE_PRT, AMDGPU_VA_OP_REPLACE); + if (r) { + ok = false; + goto out; + } + + while (va_page < end_va_page) { + struct amdgpu_sparse_backing *backing; + uint32_t backing_start; + uint32_t span_pages; + + /* Skip pages that are already uncommitted. */ + if (!comm[va_page].backing) { + va_page++; + continue; + } + + /* Group contiguous spans of pages. */ + backing = comm[va_page].backing; + backing_start = comm[va_page].page; + comm[va_page].backing = NULL; + + span_pages = 1; + va_page++; + + while (va_page < end_va_page && + comm[va_page].backing == backing && + comm[va_page].page == backing_start + span_pages) { + comm[va_page].backing = NULL; + va_page++; + span_pages++; + } + + if (!sparse_backing_free(bo, backing, backing_start, span_pages)) { + /* Couldn't allocate tracking data structures, so we have to leak */ + fprintf(stderr, "amdgpu: leaking PRT backing memory\n"); + ok = false; + } + } + } +out: + + simple_mtx_unlock(&bo->u.sparse.commit_lock); + + return ok; +} + static unsigned eg_tile_split(unsigned tile_split) { switch (tile_split) { @@ -684,36 +1182,27 @@ amdgpu_bo_create(struct radeon_winsys *rws, { struct amdgpu_winsys *ws = amdgpu_winsys(rws); struct amdgpu_winsys_bo *bo; - unsigned usage = 0, pb_cache_bucket; + int heap = -1; + + /* VRAM implies WC. This is not optional. */ + assert(!(domain & RADEON_DOMAIN_VRAM) || flags & RADEON_FLAG_GTT_WC); + + /* NO_CPU_ACCESS is valid with VRAM only. */ + assert(domain == RADEON_DOMAIN_VRAM || !(flags & RADEON_FLAG_NO_CPU_ACCESS)); + + /* Sparse buffers must have NO_CPU_ACCESS set. */ + assert(!(flags & RADEON_FLAG_SPARSE) || flags & RADEON_FLAG_NO_CPU_ACCESS); /* Sub-allocate small buffers from slabs. */ - if (!(flags & RADEON_FLAG_HANDLE) && + if (!(flags & (RADEON_FLAG_NO_SUBALLOC | RADEON_FLAG_SPARSE)) && size <= (1 << AMDGPU_SLAB_MAX_SIZE_LOG2) && alignment <= MAX2(1 << AMDGPU_SLAB_MIN_SIZE_LOG2, util_next_power_of_two(size))) { struct pb_slab_entry *entry; - unsigned heap = 0; + int heap = radeon_get_heap_index(domain, flags); - if (flags & RADEON_FLAG_GTT_WC) - heap |= 1; - if (flags & RADEON_FLAG_CPU_ACCESS) - heap |= 2; - if (flags & ~(RADEON_FLAG_GTT_WC | RADEON_FLAG_CPU_ACCESS)) + if (heap < 0 || heap >= RADEON_MAX_SLAB_HEAPS) goto no_slab; - switch (domain) { - case RADEON_DOMAIN_VRAM: - heap |= 0 * 4; - break; - case RADEON_DOMAIN_VRAM_GTT: - heap |= 1 * 4; - break; - case RADEON_DOMAIN_GTT: - heap |= 2 * 4; - break; - default: - goto no_slab; - } - entry = pb_slab_alloc(&ws->bo_slabs, size, heap); if (!entry) { /* Clear the cache and try again. */ @@ -733,8 +1222,14 @@ amdgpu_bo_create(struct radeon_winsys *rws, } no_slab: + if (flags & RADEON_FLAG_SPARSE) { + assert(RADEON_SPARSE_PAGE_SIZE % alignment == 0); + + return amdgpu_bo_sparse_create(ws, size, domain, flags); + } + /* This flag is irrelevant for the cache. */ - flags &= ~RADEON_FLAG_HANDLE; + flags &= ~RADEON_FLAG_NO_SUBALLOC; /* Align size to page size. This is the minimum alignment for normal * BOs. Aligning this here helps the cached bufmgr. Especially small BOs, @@ -743,45 +1238,31 @@ no_slab: size = align64(size, ws->info.gart_page_size); alignment = align(alignment, ws->info.gart_page_size); - /* Only set one usage bit each for domains and flags, or the cache manager - * might consider different sets of domains / flags compatible - */ - if (domain == RADEON_DOMAIN_VRAM_GTT) - usage = 1 << 2; - else - usage = domain >> 1; - assert(flags < sizeof(usage) * 8 - 3); - usage |= 1 << (flags + 3); - - /* Determine the pb_cache bucket for minimizing pb_cache misses. */ - pb_cache_bucket = 0; - if (domain & RADEON_DOMAIN_VRAM) /* VRAM or VRAM+GTT */ - pb_cache_bucket += 1; - if (flags == RADEON_FLAG_GTT_WC) /* WC */ - pb_cache_bucket += 2; - assert(pb_cache_bucket < ARRAY_SIZE(ws->bo_cache.buckets)); - - /* Get a buffer from the cache. */ - bo = (struct amdgpu_winsys_bo*) - pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment, usage, - pb_cache_bucket); - if (bo) - return &bo->base; + bool use_reusable_pool = flags & RADEON_FLAG_NO_INTERPROCESS_SHARING; + + if (use_reusable_pool) { + heap = radeon_get_heap_index(domain, flags); + assert(heap >= 0 && heap < RADEON_MAX_CACHED_HEAPS); + + /* Get a buffer from the cache. */ + bo = (struct amdgpu_winsys_bo*) + pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment, 0, heap); + if (bo) + return &bo->base; + } /* Create a new one. */ - bo = amdgpu_create_bo(ws, size, alignment, usage, domain, flags, - pb_cache_bucket); + bo = amdgpu_create_bo(ws, size, alignment, domain, flags, heap); if (!bo) { /* Clear the cache and try again. */ pb_slabs_reclaim(&ws->bo_slabs); pb_cache_release_all_buffers(&ws->bo_cache); - bo = amdgpu_create_bo(ws, size, alignment, usage, domain, flags, - pb_cache_bucket); + bo = amdgpu_create_bo(ws, size, alignment, domain, flags, heap); if (!bo) return NULL; } - bo->u.real.use_reusable_pool = true; + bo->u.real.use_reusable_pool = use_reusable_pool; return &bo->base; } @@ -791,56 +1272,78 @@ static struct pb_buffer *amdgpu_bo_from_handle(struct radeon_winsys *rws, unsigned *offset) { struct amdgpu_winsys *ws = amdgpu_winsys(rws); - struct amdgpu_winsys_bo *bo; + struct amdgpu_winsys_bo *bo = NULL; enum amdgpu_bo_handle_type type; struct amdgpu_bo_import_result result = {0}; uint64_t va; - amdgpu_va_handle va_handle; + amdgpu_va_handle va_handle = NULL; struct amdgpu_bo_info info = {0}; enum radeon_bo_domain initial = 0; int r; - /* Initialize the structure. */ - bo = CALLOC_STRUCT(amdgpu_winsys_bo); - if (!bo) { - return NULL; - } - switch (whandle->type) { - case DRM_API_HANDLE_TYPE_SHARED: + case WINSYS_HANDLE_TYPE_SHARED: type = amdgpu_bo_handle_type_gem_flink_name; break; - case DRM_API_HANDLE_TYPE_FD: + 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) - goto error; + 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); + + /* Release the buffer handle, because we don't need it anymore. + * This function is returning an existing buffer, which has its own + * handle. + */ + amdgpu_bo_free(result.buf_handle); + return &bo->base; + } /* Get initial domains. */ r = amdgpu_bo_query_info(result.buf_handle, &info); if (r) - goto error_query; + goto error; r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general, - result.alloc_size, 1 << 20, 0, &va, &va_handle, 0); + result.alloc_size, 1 << 20, 0, &va, &va_handle, + AMDGPU_VA_RANGE_HIGH); if (r) - goto error_query; + goto error; + + bo = CALLOC_STRUCT(amdgpu_winsys_bo); + if (!bo) + goto error; r = amdgpu_bo_va_op(result.buf_handle, 0, result.alloc_size, va, 0, AMDGPU_VA_OP_MAP); if (r) - goto error_va_map; + goto error; if (info.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM) initial |= RADEON_DOMAIN_VRAM; if (info.preferred_heap & AMDGPU_GEM_DOMAIN_GTT) initial |= RADEON_DOMAIN_GTT; - + /* Initialize the structure. */ pipe_reference_init(&bo->base.reference, 1); bo->base.alignment = info.phys_alignment; bo->bo = result.buf_handle; @@ -853,28 +1356,27 @@ static struct pb_buffer *amdgpu_bo_from_handle(struct radeon_winsys *rws, 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); + amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle); - return &bo->base; + amdgpu_add_buffer_to_global_list(bo); -error_va_map: - amdgpu_va_range_free(va_handle); + util_hash_table_set(ws->bo_export_table, bo->bo, bo); + simple_mtx_unlock(&ws->bo_export_table_lock); -error_query: - amdgpu_bo_free(result.buf_handle); + return &bo->base; error: - FREE(bo); + 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; } @@ -884,24 +1386,24 @@ static bool amdgpu_bo_get_handle(struct pb_buffer *buffer, 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; - if (!bo->bo) { - offset += bo->va - bo->u.slab.real->va; - bo = bo->u.slab.real; - } + /* 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) { - case DRM_API_HANDLE_TYPE_SHARED: + case WINSYS_HANDLE_TYPE_SHARED: type = amdgpu_bo_handle_type_gem_flink_name; break; - case DRM_API_HANDLE_TYPE_FD: + case WINSYS_HANDLE_TYPE_FD: type = amdgpu_bo_handle_type_dma_buf_fd; break; - case DRM_API_HANDLE_TYPE_KMS: + case WINSYS_HANDLE_TYPE_KMS: type = amdgpu_bo_handle_type_kms; break; default: @@ -912,6 +1414,10 @@ static bool amdgpu_bo_get_handle(struct pb_buffer *buffer, 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; @@ -927,19 +1433,23 @@ static struct pb_buffer *amdgpu_bo_from_ptr(struct radeon_winsys *rws, struct amdgpu_winsys_bo *bo; uint64_t va; amdgpu_va_handle va_handle; + /* Avoid failure when the size is not page aligned */ + uint64_t aligned_size = align64(size, ws->info.gart_page_size); bo = CALLOC_STRUCT(amdgpu_winsys_bo); if (!bo) return NULL; - if (amdgpu_create_bo_from_user_mem(ws->dev, pointer, size, &buf_handle)) + if (amdgpu_create_bo_from_user_mem(ws->dev, pointer, + aligned_size, &buf_handle)) goto error; if (amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general, - size, 1 << 12, 0, &va, &va_handle, 0)) + aligned_size, 1 << 12, 0, &va, &va_handle, + AMDGPU_VA_RANGE_HIGH)) goto error_va_alloc; - if (amdgpu_bo_va_op(buf_handle, 0, size, va, 0, AMDGPU_VA_OP_MAP)) + if (amdgpu_bo_va_op(buf_handle, 0, aligned_size, va, 0, AMDGPU_VA_OP_MAP)) goto error_va_map; /* Initialize it. */ @@ -955,10 +1465,12 @@ static struct pb_buffer *amdgpu_bo_from_ptr(struct radeon_winsys *rws, bo->initial_domain = RADEON_DOMAIN_GTT; bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1); - ws->allocated_gtt += align64(bo->base.size, ws->info.gart_page_size); + ws->allocated_gtt += aligned_size; amdgpu_add_buffer_to_global_list(bo); + amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle); + return (struct pb_buffer*)bo; error_va_map: @@ -977,6 +1489,13 @@ static bool amdgpu_bo_is_user_ptr(struct pb_buffer *buf) return ((struct amdgpu_winsys_bo*)buf)->user_ptr != NULL; } +static bool amdgpu_bo_is_suballocated(struct pb_buffer *buf) +{ + struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf; + + return !bo->bo && !bo->sparse; +} + static uint64_t amdgpu_bo_get_va(struct pb_buffer *buf) { return ((struct amdgpu_winsys_bo*)buf)->va; @@ -993,7 +1512,9 @@ void amdgpu_bo_init_functions(struct amdgpu_winsys *ws) ws->base.buffer_from_handle = amdgpu_bo_from_handle; ws->base.buffer_from_ptr = amdgpu_bo_from_ptr; ws->base.buffer_is_user_ptr = amdgpu_bo_is_user_ptr; + ws->base.buffer_is_suballocated = amdgpu_bo_is_suballocated; ws->base.buffer_get_handle = amdgpu_bo_get_handle; + ws->base.buffer_commit = amdgpu_bo_sparse_commit; ws->base.buffer_get_virtual_address = amdgpu_bo_get_va; ws->base.buffer_get_initial_domain = amdgpu_bo_get_initial_domain; }