/* * Copyright © 2011 Marek Olšák * Copyright © 2015 Advanced Micro Devices, Inc. * * SPDX-License-Identifier: MIT */ #include #include "amdgpu_cs.h" #include "util/os_drm.h" #include "util/hash_table.h" #include "util/os_time.h" #include "util/u_hash_table.h" #include "util/u_process.h" #include "frontend/drm_driver.h" #include "drm-uapi/amdgpu_drm.h" #include "drm-uapi/dma-buf.h" #include "sid.h" #include #include #include #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 bool amdgpu_bo_fence_wait(struct amdgpu_winsys *aws, struct pipe_fence_handle **fence, uint64_t timeout, int64_t abs_timeout) { if (timeout == 0) { bool idle = amdgpu_fence_wait(*fence, 0, false); if (!idle) { simple_mtx_unlock(&aws->bo_fence_lock); return false; /* busy */ } /* It's idle. Remove it from the ring to skip checking it again later. */ amdgpu_fence_reference(fence, NULL); } else { struct pipe_fence_handle *tmp_fence = NULL; amdgpu_fence_reference(&tmp_fence, *fence); /* While waiting, unlock the mutex. */ simple_mtx_unlock(&aws->bo_fence_lock); bool idle = amdgpu_fence_wait(tmp_fence, abs_timeout, true); if (!idle) { amdgpu_fence_reference(&tmp_fence, NULL); return false; /* busy */ } simple_mtx_lock(&aws->bo_fence_lock); /* It's idle. Remove it from the ring to skip checking it again later. */ if (tmp_fence == *fence) amdgpu_fence_reference(fence, NULL); amdgpu_fence_reference(&tmp_fence, NULL); } return true; } static bool amdgpu_bo_wait(struct radeon_winsys *rws, struct pb_buffer_lean *_buf, uint64_t timeout, unsigned usage) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf); int64_t abs_timeout = 0; assert(p_atomic_read(&bo->num_active_ioctls) >= 0); if (timeout == 0) { if (p_atomic_read(&bo->num_active_ioctls)) return false; } else { abs_timeout = os_time_get_absolute_timeout(timeout); /* Wait if any ioctl is being submitted with this buffer. */ if (!os_wait_until_zero_abs_timeout(&bo->num_active_ioctls, abs_timeout)) return false; } if (is_real_bo(bo) && (get_real_bo(bo)->is_shared || get_real_bo(bo)->slab_has_busy_alt_fences)) { /* We can't use user fences for shared buffers, because user fences are local to this * process only. If we want to wait for all buffer uses in all processes, we have to * use amdgpu_bo_wait_for_idle. * * Additionally, if this is a slab buffer and one of the slab entries has non-NULL * alt_fence, we can't easily wait for that here. Instead, use the kernel ioctl to wait * for the buffer. */ bool buffer_busy = true; int r; /* The GEM_WAIT_IDLE ioctl with timeout=0 can take up to 1 ms to return. This is a kernel * inefficiency. This flag indicates whether it's better to return busy than wait for 1 ms. */ if (timeout == 0 && usage & RADEON_USAGE_DISALLOW_SLOW_REPLY) return false; r = ac_drm_bo_wait_for_idle(aws->dev, get_real_bo(bo)->bo, timeout, &buffer_busy); if (r) fprintf(stderr, "%s: amdgpu_bo_wait_for_idle failed %i\n", __func__, r); if (!buffer_busy) get_real_bo(bo)->slab_has_busy_alt_fences = false; return !buffer_busy; } simple_mtx_lock(&aws->bo_fence_lock); u_foreach_bit(i, bo->fences.valid_fence_mask) { struct pipe_fence_handle **fence = get_fence_from_ring(aws, &bo->fences, i); if (fence) { /* This also unlocks the mutex on failure. */ if (!amdgpu_bo_fence_wait(aws, fence, timeout, abs_timeout)) return false; } bo->fences.valid_fence_mask &= ~BITFIELD_BIT(i); /* remove the fence from the BO */ } /* Also wait for alt_fence. */ if (bo->alt_fence) { /* This also unlocks the mutex on failure. */ if (!amdgpu_bo_fence_wait(aws, &bo->alt_fence, timeout, abs_timeout)) return false; } simple_mtx_unlock(&aws->bo_fence_lock); return true; /* idle */ } static void amdgpu_bo_get_syncobjs(struct amdgpu_winsys *aws, struct amdgpu_winsys_bo *bo, uint32_t *syncobj, uint32_t *num_fences) { if (p_atomic_read(&bo->num_active_ioctls)) os_wait_until_zero(&bo->num_active_ioctls, OS_TIMEOUT_INFINITE); simple_mtx_lock(&aws->bo_fence_lock); u_foreach_bit(queue_index, bo->fences.valid_fence_mask) { struct pipe_fence_handle **fence = get_fence_from_ring(aws, &bo->fences, queue_index); if (fence) { if (!amdgpu_fence_wait(*fence, 0, 0)) { syncobj[(*num_fences)++] = ((struct amdgpu_fence*)*fence)->syncobj; } else { amdgpu_fence_reference(fence, NULL); /* remove the fence from the BO */ bo->fences.valid_fence_mask &= ~BITFIELD_BIT(queue_index); } } } if (bo->alt_fence) { if (!amdgpu_fence_wait(bo->alt_fence, 0, 0)) syncobj[(*num_fences)++] = ((struct amdgpu_fence*)bo->alt_fence)->syncobj; else amdgpu_fence_reference(&bo->alt_fence, NULL); } simple_mtx_unlock(&aws->bo_fence_lock); } static int amdgpu_bo_va_op_common(struct amdgpu_winsys *aws, struct amdgpu_winsys_bo *bo, uint32_t bo_handle, bool send_input_fence, uint64_t *vm_timeline_point, uint64_t offset, uint64_t size, uint64_t addr, uint64_t flags, uint32_t ops) { int r; if (aws->info.use_userq) { uint32_t syncobj_arr[AMDGPU_MAX_QUEUES + 1]; uint32_t num_fences = 0; if (send_input_fence) amdgpu_bo_get_syncobjs(aws, bo, &syncobj_arr[0], &num_fences); /* The lock guarantees that the execution ordering of the vm ioctls match the timeline * sequence number ordering. */ simple_mtx_lock(&aws->vm_ioctl_lock); aws->vm_timeline_seq_num++; if (vm_timeline_point) { /* Sparse buffers can be updated concurrently by another thread so we use atomic operation * to get a valid seqno. */ p_atomic_set(vm_timeline_point, aws->vm_timeline_seq_num); } r = ac_drm_bo_va_op_raw2(aws->dev, bo_handle, offset, size, addr, flags, ops, aws->vm_timeline_syncobj, aws->vm_timeline_seq_num, (uintptr_t)&syncobj_arr, num_fences); simple_mtx_unlock(&aws->vm_ioctl_lock); } else { r = ac_drm_bo_va_op_raw(aws->dev, bo_handle, offset, size, addr, flags, ops); } return r; } static inline unsigned get_slab_entry_offset(struct amdgpu_winsys_bo *bo) { struct amdgpu_bo_slab_entry *slab_entry_bo = get_slab_entry_bo(bo); struct amdgpu_bo_real_reusable_slab *slab_bo = (struct amdgpu_bo_real_reusable_slab *)get_slab_entry_real_bo(bo); unsigned entry_index = slab_entry_bo - slab_bo->entries; return slab_bo->slab.entry_size * entry_index; } static enum radeon_bo_domain amdgpu_bo_get_initial_domain( struct pb_buffer_lean *buf) { return ((struct amdgpu_winsys_bo*)buf)->base.placement; } static enum radeon_bo_flag amdgpu_bo_get_flags( struct pb_buffer_lean *buf) { return ((struct amdgpu_winsys_bo*)buf)->base.usage; } static void amdgpu_bo_remove_fences(struct amdgpu_winsys_bo *bo) { bo->fences.valid_fence_mask = 0; amdgpu_fence_reference(&bo->alt_fence, NULL); } void amdgpu_bo_destroy(struct amdgpu_winsys *aws, struct pb_buffer_lean *_buf) { struct amdgpu_bo_real *bo = get_real_bo(amdgpu_winsys_bo(_buf)); struct amdgpu_screen_winsys *sws_iter; simple_mtx_lock(&aws->bo_export_table_lock); /* amdgpu_bo_from_handle might have revived the bo */ if (p_atomic_read(&bo->b.base.reference.count)) { simple_mtx_unlock(&aws->bo_export_table_lock); return; } _mesa_hash_table_remove_key(aws->bo_export_table, bo->bo.abo); if (bo->b.base.placement & RADEON_DOMAIN_VRAM_GTT) { amdgpu_bo_va_op_common(aws, amdgpu_winsys_bo(_buf), bo->kms_handle, true, NULL, 0, bo->b.base.size, amdgpu_va_get_start_addr(bo->va_handle), AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE | AMDGPU_VM_PAGE_EXECUTABLE, AMDGPU_VA_OP_UNMAP); ac_drm_va_range_free(bo->va_handle); } simple_mtx_unlock(&aws->bo_export_table_lock); if (!bo->is_user_ptr && bo->cpu_ptr) { bo->cpu_ptr = NULL; amdgpu_bo_unmap(&aws->dummy_sws.base, &bo->b.base); } assert(bo->is_user_ptr || bo->map_count == 0); ac_drm_bo_free(aws->dev, bo->bo); #if MESA_DEBUG if (aws->debug_all_bos) { simple_mtx_lock(&aws->global_bo_list_lock); list_del(&bo->global_list_item); aws->num_buffers--; simple_mtx_unlock(&aws->global_bo_list_lock); } #endif /* Close all KMS handles retrieved for other DRM file descriptions */ simple_mtx_lock(&aws->sws_list_lock); for (sws_iter = aws->sws_list; sws_iter; sws_iter = sws_iter->next) { struct hash_entry *entry; if (!sws_iter->kms_handles) continue; entry = _mesa_hash_table_search(sws_iter->kms_handles, bo); if (entry) { struct drm_gem_close args = { .handle = (uintptr_t)entry->data }; drm_ioctl(sws_iter->fd, DRM_IOCTL_GEM_CLOSE, &args); _mesa_hash_table_remove(sws_iter->kms_handles, entry); } } simple_mtx_unlock(&aws->sws_list_lock); amdgpu_bo_remove_fences(&bo->b); if (bo->b.base.placement & RADEON_DOMAIN_VRAM) aws->allocated_vram -= align64(bo->b.base.size, aws->info.gart_page_size); else if (bo->b.base.placement & RADEON_DOMAIN_GTT) aws->allocated_gtt -= align64(bo->b.base.size, aws->info.gart_page_size); simple_mtx_destroy(&bo->map_lock); FREE(bo); } static void amdgpu_bo_destroy_or_cache(struct radeon_winsys *rws, struct pb_buffer_lean *_buf) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf); assert(is_real_bo(bo)); /* slab buffers have a separate vtbl */ if (bo->type >= AMDGPU_BO_REAL_REUSABLE) pb_cache_add_buffer(&aws->bo_cache, &((struct amdgpu_bo_real_reusable*)bo)->cache_entry); else amdgpu_bo_destroy(aws, _buf); } static void amdgpu_clean_up_buffer_managers(struct amdgpu_winsys *aws) { pb_slabs_reclaim(&aws->bo_slabs); pb_cache_release_all_buffers(&aws->bo_cache); } static bool amdgpu_bo_do_map(struct radeon_winsys *rws, struct amdgpu_bo_real *bo, void **cpu) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); assert(!bo->is_user_ptr); *cpu = NULL; int r = ac_drm_bo_cpu_map(aws->dev, bo->bo, cpu); if (r) { /* Clean up buffer managers and try again. */ amdgpu_clean_up_buffer_managers(aws); r = ac_drm_bo_cpu_map(aws->dev, bo->bo, cpu); if (r) return false; } if (p_atomic_inc_return(&bo->map_count) == 1) { if (bo->b.base.placement & RADEON_DOMAIN_VRAM) aws->mapped_vram += bo->b.base.size; else if (bo->b.base.placement & RADEON_DOMAIN_GTT) aws->mapped_gtt += bo->b.base.size; aws->num_mapped_buffers++; } return true; } void *amdgpu_bo_map(struct radeon_winsys *rws, struct pb_buffer_lean *buf, struct radeon_cmdbuf *rcs, enum pipe_map_flags usage) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf; struct amdgpu_bo_real *real; struct amdgpu_cs *cs = rcs ? amdgpu_cs(rcs) : NULL; assert(bo->type != AMDGPU_BO_SPARSE); /* If it's not unsynchronized bo_map, flush CS if needed and then wait. */ if (!(usage & PIPE_MAP_UNSYNCHRONIZED)) { /* DONTBLOCK doesn't make sense with UNSYNCHRONIZED. */ if (usage & PIPE_MAP_DONTBLOCK) { if (!(usage & PIPE_MAP_WRITE)) { /* Mapping for read. * * Since we are mapping for read, we don't need to wait * if the GPU is using the buffer for read too * (neither one is changing it). * * 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_START_NEXT_GFX_IB_NOW, NULL); return NULL; } if (!amdgpu_bo_wait(rws, (struct pb_buffer_lean*)bo, 0, RADEON_USAGE_WRITE)) { return NULL; } } else { if (cs && amdgpu_bo_is_referenced_by_cs(cs, bo)) { cs->flush_cs(cs->flush_data, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL); return NULL; } if (!amdgpu_bo_wait(rws, (struct pb_buffer_lean*)bo, 0, RADEON_USAGE_READWRITE)) { return NULL; } } } else { uint64_t time = os_time_get_nano(); if (!(usage & PIPE_MAP_WRITE)) { /* Mapping for read. * * Since we are mapping for read, we don't need to wait * if the GPU is using the buffer for read too * (neither one is changing it). * * Only check whether the buffer is being used for write. */ if (cs) { if (amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo, RADEON_USAGE_WRITE)) { 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)) amdgpu_cs_sync_flush(rcs); } } amdgpu_bo_wait(rws, (struct pb_buffer_lean*)bo, OS_TIMEOUT_INFINITE, RADEON_USAGE_WRITE); } else { /* Mapping for write. */ if (cs) { if (amdgpu_bo_is_referenced_by_cs(cs, bo)) { 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)) amdgpu_cs_sync_flush(rcs); } } amdgpu_bo_wait(rws, (struct pb_buffer_lean*)bo, OS_TIMEOUT_INFINITE, RADEON_USAGE_READWRITE); } aws->buffer_wait_time += os_time_get_nano() - time; } } /* Buffer synchronization has been checked, now actually map the buffer. */ void *cpu = NULL; uint64_t offset = 0; if (is_real_bo(bo)) { real = get_real_bo(bo); } else { real = get_slab_entry_real_bo(bo); offset = get_slab_entry_offset(bo); } if (usage & RADEON_MAP_TEMPORARY) { if (real->is_user_ptr) { cpu = real->cpu_ptr; } else { if (!amdgpu_bo_do_map(rws, real, &cpu)) return NULL; } } else { cpu = p_atomic_read(&real->cpu_ptr); if (!cpu) { simple_mtx_lock(&real->map_lock); /* Must re-check due to the possibility of a race. Re-check need not * be atomic thanks to the lock. */ cpu = real->cpu_ptr; if (!cpu) { if (!amdgpu_bo_do_map(rws, real, &cpu)) { simple_mtx_unlock(&real->map_lock); return NULL; } p_atomic_set(&real->cpu_ptr, cpu); } simple_mtx_unlock(&real->map_lock); } } return (uint8_t*)cpu + offset; } void amdgpu_bo_unmap(struct radeon_winsys *rws, struct pb_buffer_lean *buf) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf; struct amdgpu_bo_real *real; assert(bo->type != AMDGPU_BO_SPARSE); real = is_real_bo(bo) ? get_real_bo(bo) : get_slab_entry_real_bo(bo); if (real->is_user_ptr) return; assert(real->map_count != 0 && "too many unmaps"); if (p_atomic_dec_zero(&real->map_count)) { assert(!real->cpu_ptr && "too many unmaps or forgot RADEON_MAP_TEMPORARY flag"); if (real->b.base.placement & RADEON_DOMAIN_VRAM) aws->mapped_vram -= real->b.base.size; else if (real->b.base.placement & RADEON_DOMAIN_GTT) aws->mapped_gtt -= real->b.base.size; aws->num_mapped_buffers--; } assert(aws->dev); ac_drm_bo_cpu_unmap(aws->dev, real->bo); } static void amdgpu_add_buffer_to_global_list(struct amdgpu_winsys *aws, struct amdgpu_bo_real *bo) { #if MESA_DEBUG if (aws->debug_all_bos) { simple_mtx_lock(&aws->global_bo_list_lock); list_addtail(&bo->global_list_item, &aws->global_bo_list); aws->num_buffers++; simple_mtx_unlock(&aws->global_bo_list_lock); } #endif } static unsigned amdgpu_get_optimal_alignment(struct amdgpu_winsys *aws, uint64_t size, unsigned alignment) { /* Increase the alignment for faster address translation and better memory * access pattern. */ if (size >= aws->info.pte_fragment_size) { alignment = MAX2(alignment, aws->info.pte_fragment_size); } else if (size) { unsigned msb = util_last_bit(size); alignment = MAX2(alignment, 1u << (msb - 1)); } return alignment; } static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *aws, uint64_t size, unsigned alignment, enum radeon_bo_domain initial_domain, unsigned flags, int heap) { struct amdgpu_bo_alloc_request request = {0}; ac_drm_bo buf_handle; uint64_t va = 0; struct amdgpu_bo_real *bo; amdgpu_va_handle va_handle = NULL; int r; /* VRAM or GTT must be specified, but not both at the same time. */ assert(util_bitcount(initial_domain & (RADEON_DOMAIN_VRAM_GTT | RADEON_DOMAIN_GDS | RADEON_DOMAIN_OA | RADEON_DOMAIN_DOORBELL)) == 1); alignment = amdgpu_get_optimal_alignment(aws, size, alignment); if (heap >= 0 && flags & RADEON_FLAG_NO_INTERPROCESS_SHARING) { struct amdgpu_bo_real_reusable *new_bo; bool slab_backing = flags & RADEON_FLAG_WINSYS_SLAB_BACKING; if (slab_backing) new_bo = (struct amdgpu_bo_real_reusable *)CALLOC_STRUCT(amdgpu_bo_real_reusable_slab); else new_bo = CALLOC_STRUCT(amdgpu_bo_real_reusable); if (!new_bo) return NULL; bo = &new_bo->b; pb_cache_init_entry(&aws->bo_cache, &new_bo->cache_entry, &bo->b.base, heap); bo->b.type = slab_backing ? AMDGPU_BO_REAL_REUSABLE_SLAB : AMDGPU_BO_REAL_REUSABLE; } else { bo = CALLOC_STRUCT(amdgpu_bo_real); if (!bo) return NULL; bo->b.type = AMDGPU_BO_REAL; } request.alloc_size = size; request.phys_alignment = alignment; if (initial_domain & RADEON_DOMAIN_VRAM) { request.preferred_heap |= AMDGPU_GEM_DOMAIN_VRAM; /* 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 (!aws->info.has_dedicated_vram) request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT; } if (initial_domain & RADEON_DOMAIN_GTT) request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT; if (initial_domain & RADEON_DOMAIN_GDS) request.preferred_heap |= AMDGPU_GEM_DOMAIN_GDS; if (initial_domain & RADEON_DOMAIN_OA) request.preferred_heap |= AMDGPU_GEM_DOMAIN_OA; if (initial_domain & RADEON_DOMAIN_DOORBELL) request.preferred_heap |= AMDGPU_GEM_DOMAIN_DOORBELL; 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 (aws->info.has_local_buffers && initial_domain & (RADEON_DOMAIN_VRAM_GTT | RADEON_DOMAIN_DOORBELL) && flags & RADEON_FLAG_NO_INTERPROCESS_SHARING) request.flags |= AMDGPU_GEM_CREATE_VM_ALWAYS_VALID; if (flags & RADEON_FLAG_DISCARDABLE && aws->info.drm_minor >= 47) request.flags |= AMDGPU_GEM_CREATE_DISCARDABLE; if ((flags & RADEON_FLAG_CLEAR_VRAM) || (aws->zero_all_vram_allocs && (request.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM))) request.flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED; if ((flags & RADEON_FLAG_ENCRYPTED) && aws->info.has_tmz_support) { request.flags |= AMDGPU_GEM_CREATE_ENCRYPTED; if (!(flags & RADEON_FLAG_DRIVER_INTERNAL)) { struct amdgpu_screen_winsys *sws_iter; simple_mtx_lock(&aws->sws_list_lock); for (sws_iter = aws->sws_list; sws_iter; sws_iter = sws_iter->next) { *((bool*) &sws_iter->base.uses_secure_bos) = true; } simple_mtx_unlock(&aws->sws_list_lock); } } if (flags & RADEON_FLAG_GFX12_ALLOW_DCC) request.flags |= AMDGPU_GEM_CREATE_GFX12_DCC; /* Set AMDGPU_GEM_CREATE_VIRTIO_SHARED if the driver didn't disable buffer sharing. */ if (aws->info.is_virtio && (initial_domain & RADEON_DOMAIN_VRAM_GTT) && (flags & (RADEON_FLAG_DRIVER_INTERNAL | RADEON_FLAG_NO_INTERPROCESS_SHARING)) == 0) request.flags |= AMDGPU_GEM_CREATE_VIRTIO_SHARED; r = ac_drm_bo_alloc(aws->dev, &request, &buf_handle); if (r) { fprintf(stderr, "amdgpu: Failed to allocate a buffer:\n"); fprintf(stderr, "amdgpu: size : %"PRIu64" bytes\n", size); fprintf(stderr, "amdgpu: alignment : %u bytes\n", alignment); fprintf(stderr, "amdgpu: domains : %u\n", initial_domain); fprintf(stderr, "amdgpu: flags : %" PRIx64 "\n", request.flags); goto error_bo_alloc; } uint32_t kms_handle = 0; ac_drm_bo_export(aws->dev, buf_handle, amdgpu_bo_handle_type_kms, &kms_handle); if (initial_domain & RADEON_DOMAIN_VRAM_GTT) { unsigned va_gap_size = aws->check_vm ? MAX2(4 * alignment, 64 * 1024) : 0; r = ac_drm_va_range_alloc(aws->dev, amdgpu_gpu_va_range_general, 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; unsigned vm_flags = AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE | AMDGPU_VM_PAGE_EXECUTABLE; if (flags & RADEON_FLAG_GL2_BYPASS) vm_flags |= AMDGPU_VM_MTYPE_UC; r = amdgpu_bo_va_op_common(aws, NULL, kms_handle, false, &bo->vm_timeline_point, 0, size, va, vm_flags, AMDGPU_VA_OP_MAP); if (r) goto error_va_map; } simple_mtx_init(&bo->map_lock, mtx_plain); pipe_reference_init(&bo->b.base.reference, 1); bo->b.base.placement = initial_domain; bo->b.base.alignment_log2 = util_logbase2(alignment); bo->b.base.usage = flags; bo->b.base.size = size; bo->b.unique_id = __sync_fetch_and_add(&aws->next_bo_unique_id, 1); bo->bo = buf_handle; bo->va_handle = va_handle; bo->kms_handle = kms_handle; if (initial_domain & RADEON_DOMAIN_VRAM) aws->allocated_vram += align64(size, aws->info.gart_page_size); else if (initial_domain & RADEON_DOMAIN_GTT) aws->allocated_gtt += align64(size, aws->info.gart_page_size); amdgpu_add_buffer_to_global_list(aws, bo); return &bo->b; error_va_map: ac_drm_va_range_free(va_handle); error_va_alloc: ac_drm_bo_free(aws->dev, buf_handle); error_bo_alloc: FREE(bo); return NULL; } bool amdgpu_bo_can_reclaim(struct amdgpu_winsys *aws, struct pb_buffer_lean *_buf) { return amdgpu_bo_wait(&aws->dummy_sws.base, _buf, 0, RADEON_USAGE_READWRITE); } bool amdgpu_bo_can_reclaim_slab(void *priv, struct pb_slab_entry *entry) { struct amdgpu_bo_slab_entry *bo = container_of(entry, struct amdgpu_bo_slab_entry, entry); return amdgpu_bo_can_reclaim(priv, &bo->b.base); } static unsigned get_slab_wasted_size(struct amdgpu_winsys *aws, struct amdgpu_bo_slab_entry *bo) { assert(bo->b.base.size <= bo->entry.slab->entry_size); assert(bo->b.base.size < (1 << bo->b.base.alignment_log2) || bo->b.base.size < 1 << aws->bo_slabs.min_order || bo->b.base.size > bo->entry.slab->entry_size / 2); return bo->entry.slab->entry_size - bo->b.base.size; } static void amdgpu_bo_slab_destroy(struct radeon_winsys *rws, struct pb_buffer_lean *_buf) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_bo_slab_entry *bo = get_slab_entry_bo(amdgpu_winsys_bo(_buf)); if (bo->b.base.placement & RADEON_DOMAIN_VRAM) aws->slab_wasted_vram -= get_slab_wasted_size(aws, bo); else aws->slab_wasted_gtt -= get_slab_wasted_size(aws, bo); pb_slab_free(&aws->bo_slabs, &bo->entry); } /* Return the power of two size of a slab entry matching the input size. */ static unsigned get_slab_pot_entry_size(struct amdgpu_winsys *aws, unsigned size) { unsigned entry_size = util_next_power_of_two(size); unsigned min_entry_size = 1 << aws->bo_slabs.min_order; return MAX2(entry_size, min_entry_size); } /* Return the slab entry alignment. */ static unsigned get_slab_entry_alignment(struct amdgpu_winsys *aws, unsigned size) { unsigned entry_size = get_slab_pot_entry_size(aws, size); if (size <= entry_size * 3 / 4) return entry_size / 4; return entry_size; } struct pb_slab *amdgpu_bo_slab_alloc(void *priv, unsigned heap, unsigned entry_size, unsigned group_index) { struct amdgpu_winsys *aws = priv; enum radeon_bo_domain domains = radeon_domain_from_heap(heap); enum radeon_bo_flag flags = radeon_flags_from_heap(heap); /* Determine the slab buffer size. */ unsigned max_entry_size = 1 << (aws->bo_slabs.min_order + aws->bo_slabs.num_orders - 1); assert(entry_size <= max_entry_size); /* The slab size is twice the size of the largest possible entry. */ unsigned slab_size = max_entry_size * 2; if (!util_is_power_of_two_nonzero(entry_size)) { assert(util_is_power_of_two_nonzero(entry_size * 4 / 3)); /* If the entry size is 3/4 of a power of two, we would waste space and not gain * anything if we allocated only twice the power of two for the backing buffer: * 2 * 3/4 = 1.5 usable with buffer size 2 * * Allocating 5 times the entry size leads us to the next power of two and results * in a much better memory utilization: * 5 * 3/4 = 3.75 usable with buffer size 4 */ if (entry_size * 5 > slab_size) slab_size = util_next_power_of_two(entry_size * 5); } /* The largest slab should have the same size as the PTE fragment * size to get faster address translation. */ slab_size = MAX2(slab_size, aws->info.pte_fragment_size); flags |= RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_NO_SUBALLOC | RADEON_FLAG_WINSYS_SLAB_BACKING; struct amdgpu_bo_real_reusable_slab *slab_bo = (struct amdgpu_bo_real_reusable_slab*)amdgpu_bo_create(aws, slab_size, slab_size, domains, flags); if (!slab_bo) return NULL; /* The slab is not suballocated. */ assert(is_real_bo(&slab_bo->b.b.b)); assert(slab_bo->b.b.b.type == AMDGPU_BO_REAL_REUSABLE_SLAB); /* We can get a buffer from pb_cache that is slightly larger. */ slab_size = slab_bo->b.b.b.base.size; slab_bo->slab.num_entries = slab_size / entry_size; slab_bo->slab.num_free = slab_bo->slab.num_entries; slab_bo->slab.group_index = group_index; slab_bo->slab.entry_size = entry_size; slab_bo->entries = os_malloc_aligned(slab_bo->slab.num_entries * sizeof(*slab_bo->entries), CACHE_LINE_SIZE); if (!slab_bo->entries) goto fail; memset(slab_bo->entries, 0, slab_bo->slab.num_entries * sizeof(*slab_bo->entries)); list_inithead(&slab_bo->slab.free); for (unsigned i = 0; i < slab_bo->slab.num_entries; ++i) { struct amdgpu_bo_slab_entry *bo = &slab_bo->entries[i]; bo->b.base.placement = domains; bo->b.base.alignment_log2 = util_logbase2(get_slab_entry_alignment(aws, entry_size)); bo->b.base.size = entry_size; bo->b.type = AMDGPU_BO_SLAB_ENTRY; bo->entry.slab = &slab_bo->slab; list_addtail(&bo->entry.head, &slab_bo->slab.free); } /* Wasted alignment due to slabs with 3/4 allocations being aligned to a power of two. */ assert(slab_bo->slab.num_entries * entry_size <= slab_size); if (domains & RADEON_DOMAIN_VRAM) aws->slab_wasted_vram += slab_size - slab_bo->slab.num_entries * entry_size; else aws->slab_wasted_gtt += slab_size - slab_bo->slab.num_entries * entry_size; return &slab_bo->slab; fail: amdgpu_winsys_bo_reference(aws, (struct amdgpu_winsys_bo**)&slab_bo, NULL); return NULL; } void amdgpu_bo_slab_free(struct amdgpu_winsys *aws, struct pb_slab *slab) { struct amdgpu_bo_real_reusable_slab *bo = get_bo_from_slab(slab); unsigned slab_size = bo->b.b.b.base.size; assert(bo->slab.num_entries * bo->slab.entry_size <= slab_size); if (bo->b.b.b.base.placement & RADEON_DOMAIN_VRAM) aws->slab_wasted_vram -= slab_size - bo->slab.num_entries * bo->slab.entry_size; else aws->slab_wasted_gtt -= slab_size - bo->slab.num_entries * bo->slab.entry_size; for (unsigned i = 0; i < bo->slab.num_entries; ++i) amdgpu_bo_remove_fences(&bo->entries[i].b); os_free_aligned(bo->entries); amdgpu_winsys_bo_reference(aws, (struct amdgpu_winsys_bo**)&bo, NULL); } #if DEBUG_SPARSE_COMMITS static void sparse_dump(struct amdgpu_bo_sparse *bo, const char *func) { fprintf(stderr, "%s: %p (size=%"PRIu64", num_va_pages=%u) @ %s\n" "Commitments:\n", __func__, bo, bo->b.base.size, bo->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->num_va_pages) { backing = bo->commitments[va_page].backing; backing_page = bo->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->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->backing, list) { fprintf(stderr, " %p (size=%"PRIu64")\n", backing, backing->bo->b.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 *aws, struct amdgpu_bo_sparse *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->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_lean *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->num_backing_pages < DIV_ROUND_UP(bo->b.base.size, RADEON_SPARSE_PAGE_SIZE)); size = MIN3(bo->b.base.size / 16, 8 * 1024 * 1024, bo->b.base.size - (uint64_t)bo->num_backing_pages * RADEON_SPARSE_PAGE_SIZE); size = MAX2(size, RADEON_SPARSE_PAGE_SIZE); buf = amdgpu_bo_create(aws, size, RADEON_SPARSE_PAGE_SIZE, bo->b.base.placement, (bo->b.base.usage & ~RADEON_FLAG_SPARSE & /* Set the interprocess sharing flag to disable pb_cache because * amdgpu_bo_wait doesn't wait for active CS jobs. */ ~RADEON_FLAG_NO_INTERPROCESS_SHARING) | 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 = get_real_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->backing); bo->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 *aws, struct amdgpu_bo_sparse *bo, struct amdgpu_sparse_backing *backing) { bo->num_backing_pages -= backing->bo->b.base.size / RADEON_SPARSE_PAGE_SIZE; /* Add fences from bo to backing->bo. */ simple_mtx_lock(&aws->bo_fence_lock); u_foreach_bit(i, bo->b.fences.valid_fence_mask) { add_seq_no_to_list(aws, &backing->bo->b.fences, i, bo->b.fences.seq_no[i]); } simple_mtx_unlock(&aws->bo_fence_lock); list_del(&backing->list); amdgpu_winsys_bo_reference(aws, (struct amdgpu_winsys_bo**)&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 *aws, struct amdgpu_bo_sparse *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->b.base.size / RADEON_SPARSE_PAGE_SIZE) sparse_free_backing_buffer(aws, bo, backing); return true; } static void amdgpu_bo_sparse_destroy(struct radeon_winsys *rws, struct pb_buffer_lean *_buf) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_bo_sparse *bo = get_sparse_bo(amdgpu_winsys_bo(_buf)); int r; r = amdgpu_bo_va_op_common(aws, amdgpu_winsys_bo(_buf), 0, true, NULL, 0, (uint64_t)bo->num_va_pages * RADEON_SPARSE_PAGE_SIZE, amdgpu_va_get_start_addr(bo->va_handle), 0, AMDGPU_VA_OP_CLEAR); if (r) { fprintf(stderr, "amdgpu: clearing PRT VA region on destroy failed (%d)\n", r); } while (!list_is_empty(&bo->backing)) { sparse_free_backing_buffer(aws, bo, container_of(bo->backing.next, struct amdgpu_sparse_backing, list)); } ac_drm_va_range_free(bo->va_handle); FREE(bo->commitments); simple_mtx_destroy(&bo->commit_lock); FREE(bo); } static struct pb_buffer_lean * amdgpu_bo_sparse_create(struct amdgpu_winsys *aws, uint64_t size, enum radeon_bo_domain domain, enum radeon_bo_flag flags) { struct amdgpu_bo_sparse *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_bo_sparse); if (!bo) return NULL; simple_mtx_init(&bo->commit_lock, mtx_plain); pipe_reference_init(&bo->b.base.reference, 1); bo->b.base.placement = domain; bo->b.base.alignment_log2 = util_logbase2(RADEON_SPARSE_PAGE_SIZE); bo->b.base.usage = flags; bo->b.base.size = size; bo->b.unique_id = __sync_fetch_and_add(&aws->next_bo_unique_id, 1); bo->b.type = AMDGPU_BO_SPARSE; bo->num_va_pages = DIV_ROUND_UP(size, RADEON_SPARSE_PAGE_SIZE); bo->commitments = CALLOC(bo->num_va_pages, sizeof(*bo->commitments)); if (!bo->commitments) goto error_alloc_commitments; list_inithead(&bo->backing); /* For simplicity, we always map a multiple of the page size. */ map_size = align64(size, RADEON_SPARSE_PAGE_SIZE); va_gap_size = aws->check_vm ? 4 * RADEON_SPARSE_PAGE_SIZE : 0; uint64_t gpu_address; r = ac_drm_va_range_alloc(aws->dev, amdgpu_gpu_va_range_general, map_size + va_gap_size, RADEON_SPARSE_PAGE_SIZE, 0, &gpu_address, &bo->va_handle, AMDGPU_VA_RANGE_HIGH); if (r) goto error_va_alloc; r = amdgpu_bo_va_op_common(aws, NULL, 0, false, &bo->vm_timeline_point, 0, map_size, gpu_address, AMDGPU_VM_PAGE_PRT, AMDGPU_VA_OP_MAP); if (r) goto error_va_map; return &bo->b.base; error_va_map: ac_drm_va_range_free(bo->va_handle); error_va_alloc: FREE(bo->commitments); error_alloc_commitments: simple_mtx_destroy(&bo->commit_lock); FREE(bo); return NULL; } static bool amdgpu_bo_sparse_commit(struct radeon_winsys *rws, struct pb_buffer_lean *buf, uint64_t offset, uint64_t size, bool commit) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_bo_sparse *bo = get_sparse_bo(amdgpu_winsys_bo(buf)); struct amdgpu_sparse_commitment *comm; uint32_t va_page, end_va_page; bool ok = true; int r; assert(offset % RADEON_SPARSE_PAGE_SIZE == 0); assert(offset <= bo->b.base.size); assert(size <= bo->b.base.size - offset); assert(size % RADEON_SPARSE_PAGE_SIZE == 0 || offset + size == bo->b.base.size); comm = bo->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->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(aws, bo, &backing_start, &backing_size); if (!backing) { ok = false; goto out; } r = amdgpu_bo_va_op_common(aws, amdgpu_winsys_bo(buf), backing->bo->kms_handle, true, &bo->vm_timeline_point, (uint64_t)backing_start * RADEON_SPARSE_PAGE_SIZE, (uint64_t)backing_size * RADEON_SPARSE_PAGE_SIZE, amdgpu_va_get_start_addr(bo->va_handle) + (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(aws, 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_common(aws, amdgpu_winsys_bo(buf), 0, true, &bo->vm_timeline_point, 0, (uint64_t)(end_va_page - va_page) * RADEON_SPARSE_PAGE_SIZE, amdgpu_va_get_start_addr(bo->va_handle) + (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(aws, 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->commit_lock); return ok; } static unsigned amdgpu_bo_find_next_committed_memory(struct pb_buffer_lean *buf, uint64_t range_offset, unsigned *range_size) { struct amdgpu_bo_sparse *bo = get_sparse_bo(amdgpu_winsys_bo(buf)); struct amdgpu_sparse_commitment *comm; uint32_t va_page, end_va_page; uint32_t span_va_page, start_va_page; unsigned uncommitted_range_prev, uncommitted_range_next; if (*range_size == 0) return 0; assert(*range_size + range_offset <= bo->b.base.size); uncommitted_range_prev = uncommitted_range_next = 0; comm = bo->commitments; start_va_page = va_page = range_offset / RADEON_SPARSE_PAGE_SIZE; end_va_page = (*range_size + range_offset) / RADEON_SPARSE_PAGE_SIZE; simple_mtx_lock(&bo->commit_lock); /* Lookup the first committed page with backing physical storage */ while (va_page < end_va_page && !comm[va_page].backing) va_page++; /* Fisrt committed page lookup failed, return early. */ if (va_page == end_va_page && !comm[va_page].backing) { uncommitted_range_prev = *range_size; *range_size = 0; simple_mtx_unlock(&bo->commit_lock); return uncommitted_range_prev; } /* Lookup the first uncommitted page without backing physical storage */ span_va_page = va_page; while (va_page < end_va_page && comm[va_page].backing) va_page++; simple_mtx_unlock(&bo->commit_lock); /* Calc byte count that need to skip before committed range */ if (span_va_page != start_va_page) uncommitted_range_prev = span_va_page * RADEON_SPARSE_PAGE_SIZE - range_offset; /* Calc byte count that need to skip after committed range */ if (va_page != end_va_page || !comm[va_page].backing) { uncommitted_range_next = *range_size + range_offset - va_page * RADEON_SPARSE_PAGE_SIZE; } /* Calc size of first committed part */ *range_size = *range_size - uncommitted_range_next - uncommitted_range_prev; return *range_size ? uncommitted_range_prev : uncommitted_range_prev + uncommitted_range_next; } static void amdgpu_buffer_get_metadata(struct radeon_winsys *rws, struct pb_buffer_lean *_buf, struct radeon_bo_metadata *md, struct radeon_surf *surf) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_bo_real *bo = get_real_bo(amdgpu_winsys_bo(_buf)); struct amdgpu_bo_info info = {0}; uint32_t md_version, md_flags; enum amd_gfx_level gfx_level = aws->info.gfx_level; int r; r = ac_drm_bo_query_info(aws->dev, bo->kms_handle, &info); if (r) return; md->size_metadata = info.metadata.size_metadata; memcpy(md->metadata, info.metadata.umd_metadata, sizeof(md->metadata)); md_version = md->metadata[0] & 0xffff; if (md_version >= 3 && md->size_metadata > 4) { md_flags = md->metadata[0] >> 16; if (md_flags & (1u << AC_SURF_METADATA_FLAG_FAMILY_OVERRIDEN_BIT)) { /* The overriden gfx_level is always the last dword. */ gfx_level = md->metadata[md->size_metadata / 4 - 1]; /* Fallback to the default value if the value we got is incorrect. */ if (gfx_level < GFX6 || gfx_level >= NUM_GFX_VERSIONS) gfx_level = aws->info.gfx_level; } } ac_surface_apply_bo_metadata(gfx_level, surf, info.metadata.tiling_info, &md->mode); } static void amdgpu_buffer_set_metadata(struct radeon_winsys *rws, struct pb_buffer_lean *_buf, struct radeon_bo_metadata *md, struct radeon_surf *surf) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf); struct amdgpu_bo_real *real = is_real_bo(bo) ? get_real_bo(bo) : get_slab_entry_real_bo(bo); struct amdgpu_bo_metadata metadata = {0}; ac_surface_compute_bo_metadata(&aws->info, surf, &metadata.tiling_info); metadata.size_metadata = md->size_metadata; memcpy(metadata.umd_metadata, md->metadata, sizeof(md->metadata)); ac_drm_bo_set_metadata(aws->dev, real->kms_handle, &metadata); } struct pb_buffer_lean * amdgpu_bo_create(struct amdgpu_winsys *aws, uint64_t size, unsigned alignment, enum radeon_bo_domain domain, enum radeon_bo_flag flags) { struct amdgpu_winsys_bo *bo; radeon_canonicalize_bo_flags(&domain, &flags); /* Handle sparse buffers first. */ if (flags & RADEON_FLAG_SPARSE) { assert(RADEON_SPARSE_PAGE_SIZE % alignment == 0); return amdgpu_bo_sparse_create(aws, size, domain, flags); } unsigned max_slab_entry_size = 1 << (aws->bo_slabs.min_order + aws->bo_slabs.num_orders - 1); int heap = radeon_get_heap_index(domain, flags); /* Sub-allocate small buffers from slabs. */ if (heap >= 0 && size <= max_slab_entry_size) { struct pb_slab_entry *entry; unsigned alloc_size = size; /* Always use slabs for sizes less than 4 KB because the kernel aligns * everything to 4 KB. */ if (size < alignment && alignment <= 4 * 1024) alloc_size = alignment; if (alignment > get_slab_entry_alignment(aws, alloc_size)) { /* 3/4 allocations can return too small alignment. Try again with a power of two * allocation size. */ unsigned pot_size = get_slab_pot_entry_size(aws, alloc_size); if (alignment <= pot_size) { /* This size works but wastes some memory to fulfil the alignment. */ alloc_size = pot_size; } else { goto no_slab; /* can't fulfil alignment requirements */ } } entry = pb_slab_alloc(&aws->bo_slabs, alloc_size, heap); if (!entry) { /* Clean up buffer managers and try again. */ amdgpu_clean_up_buffer_managers(aws); entry = pb_slab_alloc(&aws->bo_slabs, alloc_size, heap); } if (!entry) return NULL; struct amdgpu_bo_slab_entry *slab_bo = container_of(entry, struct amdgpu_bo_slab_entry, entry); pipe_reference_init(&slab_bo->b.base.reference, 1); slab_bo->b.base.size = size; slab_bo->b.unique_id = __sync_fetch_and_add(&aws->next_bo_unique_id, 1); assert(alignment <= 1 << slab_bo->b.base.alignment_log2); if (domain & RADEON_DOMAIN_VRAM) aws->slab_wasted_vram += get_slab_wasted_size(aws, slab_bo); else aws->slab_wasted_gtt += get_slab_wasted_size(aws, slab_bo); return &slab_bo->b.base; } no_slab: /* Align size to page size. This is the minimum alignment for normal * BOs. Aligning this here helps the cached bufmgr. Especially small BOs, * like constant/uniform buffers, can benefit from better and more reuse. */ if (domain & RADEON_DOMAIN_VRAM_GTT) { size = align64(size, aws->info.gart_page_size); alignment = align(alignment, aws->info.gart_page_size); } bool use_reusable_pool = !(domain & RADEON_DOMAIN_DOORBELL) && (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING) && !(flags & (RADEON_FLAG_DISCARDABLE | RADEON_FLAG_CLEAR_VRAM)); if (use_reusable_pool) { /* RADEON_FLAG_NO_SUBALLOC is irrelevant for the cache. */ heap = radeon_get_heap_index(domain, flags & ~RADEON_FLAG_NO_SUBALLOC); assert(heap >= 0 && heap < RADEON_NUM_HEAPS); /* Get a buffer from the cache. */ bo = (struct amdgpu_winsys_bo*) pb_cache_reclaim_buffer(&aws->bo_cache, size, alignment, 0, heap); if (bo) { /* If the buffer is amdgpu_bo_real_reusable, but we need amdgpu_bo_real_reusable_slab, * keep the allocation but make the structure bigger. */ if (flags & RADEON_FLAG_WINSYS_SLAB_BACKING && bo->type == AMDGPU_BO_REAL_REUSABLE) { const unsigned orig_size = sizeof(struct amdgpu_bo_real_reusable); const unsigned new_size = sizeof(struct amdgpu_bo_real_reusable_slab); struct amdgpu_winsys_bo *new_bo = (struct amdgpu_winsys_bo*)REALLOC(bo, orig_size, new_size); if (!new_bo) { amdgpu_winsys_bo_reference(aws, &bo, NULL); return NULL; } memset((uint8_t*)new_bo + orig_size, 0, new_size - orig_size); bo = new_bo; bo->type = AMDGPU_BO_REAL_REUSABLE_SLAB; } return &bo->base; } } /* Create a new one. */ bo = amdgpu_create_bo(aws, size, alignment, domain, flags, heap); if (!bo) { /* Clean up buffer managers and try again. */ amdgpu_clean_up_buffer_managers(aws); bo = amdgpu_create_bo(aws, size, alignment, domain, flags, heap); if (!bo) return NULL; } return &bo->base; } static struct pb_buffer_lean * amdgpu_buffer_create(struct radeon_winsys *rws, uint64_t size, unsigned alignment, enum radeon_bo_domain domain, enum radeon_bo_flag flags) { struct pb_buffer_lean * res = amdgpu_bo_create(amdgpu_winsys(rws), size, alignment, domain, flags); return res; } static struct pb_buffer_lean *amdgpu_bo_from_handle(struct radeon_winsys *rws, struct winsys_handle *whandle, unsigned vm_alignment, bool is_prime_linear_buffer) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); struct amdgpu_bo_real *bo = NULL; enum amdgpu_bo_handle_type type; struct ac_drm_bo_import_result result = {0}; uint64_t va; amdgpu_va_handle va_handle = NULL; struct amdgpu_bo_info info = {0}; enum radeon_bo_domain initial = 0; enum radeon_bo_flag flags = 0; int r; switch (whandle->type) { case WINSYS_HANDLE_TYPE_SHARED: assert(!aws->info.is_virtio); /* Legacy-path, not handled */ 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; } r = ac_drm_bo_import(aws->dev, type, whandle->handle, &result); if (r) return NULL; simple_mtx_lock(&aws->bo_export_table_lock); bo = util_hash_table_get(aws->bo_export_table, result.bo.abo); /* If the amdgpu_winsys_bo instance already exists, bump the reference * counter and return it. */ if (bo) { p_atomic_inc(&bo->b.base.reference.count); simple_mtx_unlock(&aws->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. */ ac_drm_bo_free(aws->dev, result.bo); return &bo->b.base; } uint32_t kms_handle; ac_drm_bo_export(aws->dev, result.bo, amdgpu_bo_handle_type_kms, &kms_handle); /* Get initial domains. */ r = ac_drm_bo_query_info(aws->dev, kms_handle, &info); if (r) goto error; r = ac_drm_va_range_alloc(aws->dev, amdgpu_gpu_va_range_general, result.alloc_size, amdgpu_get_optimal_alignment(aws, result.alloc_size, vm_alignment), 0, &va, &va_handle, AMDGPU_VA_RANGE_HIGH); if (r) goto error; bo = CALLOC_STRUCT(amdgpu_bo_real); if (!bo) goto error; r = amdgpu_bo_va_op_common(aws, NULL, kms_handle, false, &bo->vm_timeline_point, 0, result.alloc_size, va, AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE | AMDGPU_VM_PAGE_EXECUTABLE | (is_prime_linear_buffer ? AMDGPU_VM_MTYPE_UC : 0), AMDGPU_VA_OP_MAP); if (r) 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; if (info.alloc_flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS) flags |= RADEON_FLAG_NO_CPU_ACCESS; if (info.alloc_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC) flags |= RADEON_FLAG_GTT_WC; if (info.alloc_flags & AMDGPU_GEM_CREATE_ENCRYPTED) { /* Imports are always possible even if the importer isn't using TMZ. * For instance libweston needs to import the buffer to be able to determine * if it can be used for scanout. */ flags |= RADEON_FLAG_ENCRYPTED; *((bool*)&rws->uses_secure_bos) = true; } if (info.alloc_flags & AMDGPU_GEM_CREATE_GFX12_DCC) flags |= RADEON_FLAG_GFX12_ALLOW_DCC; /* Initialize the structure. */ pipe_reference_init(&bo->b.base.reference, 1); bo->b.base.placement = initial; bo->b.base.alignment_log2 = util_logbase2(info.phys_alignment ? info.phys_alignment : aws->info.gart_page_size); bo->b.base.usage = flags; bo->b.base.size = result.alloc_size; bo->b.type = AMDGPU_BO_REAL; bo->b.unique_id = __sync_fetch_and_add(&aws->next_bo_unique_id, 1); simple_mtx_init(&bo->map_lock, mtx_plain); bo->bo = result.bo; bo->va_handle = va_handle; bo->kms_handle = kms_handle; bo->is_shared = true; if (bo->b.base.placement & RADEON_DOMAIN_VRAM) aws->allocated_vram += align64(bo->b.base.size, aws->info.gart_page_size); else if (bo->b.base.placement & RADEON_DOMAIN_GTT) aws->allocated_gtt += align64(bo->b.base.size, aws->info.gart_page_size); amdgpu_add_buffer_to_global_list(aws, bo); _mesa_hash_table_insert(aws->bo_export_table, bo->bo.abo, bo); simple_mtx_unlock(&aws->bo_export_table_lock); return &bo->b.base; error: simple_mtx_unlock(&aws->bo_export_table_lock); if (bo) FREE(bo); if (va_handle) ac_drm_va_range_free(va_handle); ac_drm_bo_free(aws->dev, result.bo); return NULL; } static bool amdgpu_bo_get_handle(struct radeon_winsys *rws, struct pb_buffer_lean *buffer, struct winsys_handle *whandle) { struct amdgpu_screen_winsys *sws = amdgpu_screen_winsys(rws); struct amdgpu_winsys *aws = amdgpu_winsys(rws); enum amdgpu_bo_handle_type type; struct hash_entry *entry; int r; /* Don't allow exports of slab entries and sparse buffers. */ if (!is_real_bo(amdgpu_winsys_bo(buffer))) return false; struct amdgpu_bo_real *bo = get_real_bo(amdgpu_winsys_bo(buffer)); /* This removes the REUSABLE enum if it's set. */ bo->b.type = AMDGPU_BO_REAL; switch (whandle->type) { case WINSYS_HANDLE_TYPE_SHARED: /* This is a legacy code-path, not supported by virtio. */ assert(!aws->info.is_virtio); type = amdgpu_bo_handle_type_gem_flink_name; break; case WINSYS_HANDLE_TYPE_KMS: if (sws->fd == aws->fd) { /* For virtio we can't return kms_handle, because it's not a GEM handle, * but a resource ID. Instead, repurpose the deprecated type * amdgpu_bo_handle_type_kms_noimport to request a GEM handle. */ if (aws->info.is_virtio) ac_drm_bo_export(aws->dev, bo->bo, amdgpu_bo_handle_type_kms_noimport, &whandle->handle); else whandle->handle = bo->kms_handle; if (bo->is_shared) return true; goto hash_table_set; } simple_mtx_lock(&aws->sws_list_lock); entry = _mesa_hash_table_search(sws->kms_handles, bo); simple_mtx_unlock(&aws->sws_list_lock); if (entry) { whandle->handle = (uintptr_t)entry->data; return true; } FALLTHROUGH; case WINSYS_HANDLE_TYPE_FD: type = amdgpu_bo_handle_type_dma_buf_fd; break; default: return false; } r = ac_drm_bo_export(aws->dev, bo->bo, type, &whandle->handle); if (r) return false; #if defined(DMA_BUF_SET_NAME_B) if (whandle->type == WINSYS_HANDLE_TYPE_FD && !bo->is_shared) { char dmabufname[32]; snprintf(dmabufname, 32, "%d-%s", getpid(), util_get_process_name()); r = ioctl(whandle->handle, DMA_BUF_SET_NAME_B, (uint64_t)(uintptr_t)dmabufname); } #endif if (whandle->type == WINSYS_HANDLE_TYPE_KMS) { int dma_fd = whandle->handle; r = drmPrimeFDToHandle(sws->fd, dma_fd, &whandle->handle); close(dma_fd); if (r) return false; simple_mtx_lock(&aws->sws_list_lock); _mesa_hash_table_insert_pre_hashed(sws->kms_handles, bo->kms_handle, bo, (void*)(uintptr_t)whandle->handle); simple_mtx_unlock(&aws->sws_list_lock); } hash_table_set: simple_mtx_lock(&aws->bo_export_table_lock); _mesa_hash_table_insert(aws->bo_export_table, bo->bo.abo, bo); simple_mtx_unlock(&aws->bo_export_table_lock); bo->is_shared = true; return true; } static struct pb_buffer_lean *amdgpu_bo_from_ptr(struct radeon_winsys *rws, void *pointer, uint64_t size, enum radeon_bo_flag flags) { struct amdgpu_winsys *aws = amdgpu_winsys(rws); ac_drm_bo buf_handle; struct amdgpu_bo_real *bo; uint64_t va; amdgpu_va_handle va_handle; /* Avoid failure when the size is not page aligned */ uint64_t aligned_size = align64(size, aws->info.gart_page_size); bo = CALLOC_STRUCT(amdgpu_bo_real); if (!bo) return NULL; if (ac_drm_create_bo_from_user_mem(aws->dev, pointer, aligned_size, &buf_handle)) goto error; if (ac_drm_va_range_alloc(aws->dev, amdgpu_gpu_va_range_general, aligned_size, amdgpu_get_optimal_alignment(aws, aligned_size, aws->info.gart_page_size), 0, &va, &va_handle, AMDGPU_VA_RANGE_HIGH)) goto error_va_alloc; uint32_t kms_handle; ac_drm_bo_export(aws->dev, buf_handle, amdgpu_bo_handle_type_kms, &kms_handle); if (amdgpu_bo_va_op_common(aws, NULL, kms_handle, false, &bo->vm_timeline_point, 0, aligned_size, va, AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE | AMDGPU_VM_PAGE_EXECUTABLE, AMDGPU_VA_OP_MAP)) goto error_va_map; /* Initialize it. */ bo->is_user_ptr = true; pipe_reference_init(&bo->b.base.reference, 1); bo->b.base.placement = RADEON_DOMAIN_GTT; bo->b.base.alignment_log2 = 0; bo->b.base.size = size; bo->b.type = AMDGPU_BO_REAL; bo->b.unique_id = __sync_fetch_and_add(&aws->next_bo_unique_id, 1); simple_mtx_init(&bo->map_lock, mtx_plain); bo->bo = buf_handle; bo->cpu_ptr = pointer; bo->va_handle = va_handle; bo->kms_handle = kms_handle; aws->allocated_gtt += aligned_size; amdgpu_add_buffer_to_global_list(aws, bo); return (struct pb_buffer_lean*)bo; error_va_map: ac_drm_va_range_free(va_handle); error_va_alloc: ac_drm_bo_free(aws->dev, buf_handle); error: FREE(bo); return NULL; } static bool amdgpu_bo_is_user_ptr(struct pb_buffer_lean *buf) { struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf; return is_real_bo(bo) ? get_real_bo(bo)->is_user_ptr : false; } static bool amdgpu_bo_is_suballocated(struct pb_buffer_lean *buf) { struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf; return bo->type == AMDGPU_BO_SLAB_ENTRY; } uint64_t amdgpu_bo_get_va(struct pb_buffer_lean *buf) { struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(buf); if (bo->type == AMDGPU_BO_SLAB_ENTRY) { struct amdgpu_bo_real_reusable_slab *slab_bo = (struct amdgpu_bo_real_reusable_slab *)get_slab_entry_real_bo(bo); return amdgpu_va_get_start_addr(slab_bo->b.b.va_handle) + get_slab_entry_offset(bo); } else if (bo->type == AMDGPU_BO_SPARSE) { return amdgpu_va_get_start_addr(get_sparse_bo(bo)->va_handle); } else { return amdgpu_va_get_start_addr(get_real_bo(bo)->va_handle); } } static void amdgpu_buffer_destroy(struct radeon_winsys *rws, struct pb_buffer_lean *buf) { struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(buf); if (bo->type == AMDGPU_BO_SLAB_ENTRY) amdgpu_bo_slab_destroy(rws, buf); else if (bo->type == AMDGPU_BO_SPARSE) amdgpu_bo_sparse_destroy(rws, buf); else amdgpu_bo_destroy_or_cache(rws, buf); } void amdgpu_bo_init_functions(struct amdgpu_screen_winsys *sws) { sws->base.buffer_set_metadata = amdgpu_buffer_set_metadata; sws->base.buffer_get_metadata = amdgpu_buffer_get_metadata; sws->base.buffer_map = amdgpu_bo_map; sws->base.buffer_unmap = amdgpu_bo_unmap; sws->base.buffer_wait = amdgpu_bo_wait; sws->base.buffer_create = amdgpu_buffer_create; sws->base.buffer_destroy = amdgpu_buffer_destroy; sws->base.buffer_from_handle = amdgpu_bo_from_handle; sws->base.buffer_from_ptr = amdgpu_bo_from_ptr; sws->base.buffer_is_user_ptr = amdgpu_bo_is_user_ptr; sws->base.buffer_is_suballocated = amdgpu_bo_is_suballocated; sws->base.buffer_get_handle = amdgpu_bo_get_handle; sws->base.buffer_commit = amdgpu_bo_sparse_commit; sws->base.buffer_find_next_committed_memory = amdgpu_bo_find_next_committed_memory; sws->base.buffer_get_virtual_address = amdgpu_bo_get_va; sws->base.buffer_get_initial_domain = amdgpu_bo_get_initial_domain; sws->base.buffer_get_flags = amdgpu_bo_get_flags; }