anv/allocator: Support softpin in the BO cache
[mesa.git] / src / intel / vulkan / anv_allocator.c
index a27af4eccc473d76e8c54c0100af990ba0073934..117851f7593ba31a16d31ac81b454f27c3b689b0 100644 (file)
@@ -243,6 +243,7 @@ anv_block_pool_expand_range(struct anv_block_pool *pool,
 VkResult
 anv_block_pool_init(struct anv_block_pool *pool,
                     struct anv_device *device,
+                    uint64_t start_address,
                     uint32_t initial_size,
                     uint64_t bo_flags)
 {
@@ -250,6 +251,8 @@ anv_block_pool_init(struct anv_block_pool *pool,
 
    pool->device = device;
    pool->bo_flags = bo_flags;
+   pool->start_address = gen_canonical_address(start_address);
+
    anv_bo_init(&pool->bo, 0, 0);
 
    pool->fd = memfd_create("block pool", MFD_CLOEXEC);
@@ -402,6 +405,10 @@ anv_block_pool_expand_range(struct anv_block_pool *pool,
     * hard work for us.
     */
    anv_bo_init(&pool->bo, gem_handle, size);
+   if (pool->bo_flags & EXEC_OBJECT_PINNED) {
+      pool->bo.offset = pool->start_address + BLOCK_POOL_MEMFD_CENTER -
+         center_bo_offset;
+   }
    pool->bo.flags = pool->bo_flags;
    pool->bo.map = map;
 
@@ -508,12 +515,12 @@ anv_block_pool_grow(struct anv_block_pool *pool, struct anv_block_state *state)
       assert(center_bo_offset >= back_used);
 
       /* Make sure we don't shrink the back end of the pool */
-      if (center_bo_offset < pool->back_state.end)
-         center_bo_offset = pool->back_state.end;
+      if (center_bo_offset < back_required)
+         center_bo_offset = back_required;
 
       /* Make sure that we don't shrink the front end of the pool */
-      if (size - center_bo_offset < pool->state.end)
-         center_bo_offset = size - pool->state.end;
+      if (size - center_bo_offset < front_required)
+         center_bo_offset = size - front_required;
    }
 
    assert(center_bo_offset % PAGE_SIZE == 0);
@@ -610,16 +617,18 @@ anv_block_pool_alloc_back(struct anv_block_pool *pool,
 VkResult
 anv_state_pool_init(struct anv_state_pool *pool,
                     struct anv_device *device,
+                    uint64_t start_address,
                     uint32_t block_size,
                     uint64_t bo_flags)
 {
    VkResult result = anv_block_pool_init(&pool->block_pool, device,
+                                         start_address,
                                          block_size * 16,
                                          bo_flags);
    if (result != VK_SUCCESS)
       return result;
 
-   assert(util_is_power_of_two(block_size));
+   assert(util_is_power_of_two_or_zero(block_size));
    pool->block_size = block_size;
    pool->back_alloc_free_list = ANV_FREE_LIST_EMPTY;
    for (unsigned i = 0; i < ANV_STATE_BUCKETS; i++) {
@@ -814,7 +823,7 @@ done:
 static void
 anv_state_pool_free_no_vg(struct anv_state_pool *pool, struct anv_state state)
 {
-   assert(util_is_power_of_two(state.alloc_size));
+   assert(util_is_power_of_two_or_zero(state.alloc_size));
    unsigned bucket = anv_state_pool_get_bucket(state.alloc_size);
 
    if (state.offset < 0) {
@@ -1041,7 +1050,7 @@ anv_bo_pool_free(struct anv_bo_pool *pool, const struct anv_bo *bo_in)
    struct bo_pool_bo_link *link = bo.map;
    VG_NOACCESS_WRITE(&link->bo, bo);
 
-   assert(util_is_power_of_two(bo.size));
+   assert(util_is_power_of_two_or_zero(bo.size));
    const unsigned size_log2 = ilog2_round_up(bo.size);
    const unsigned bucket = size_log2 - 12;
    assert(bucket < ARRAY_SIZE(pool->free_list));
@@ -1228,11 +1237,20 @@ anv_bo_cache_lookup(struct anv_bo_cache *cache, uint32_t gem_handle)
    return bo ? &bo->bo : NULL;
 }
 
+#define ANV_BO_CACHE_SUPPORTED_FLAGS \
+   (EXEC_OBJECT_WRITE | \
+    EXEC_OBJECT_ASYNC | \
+    EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
+    EXEC_OBJECT_PINNED)
+
 VkResult
 anv_bo_cache_alloc(struct anv_device *device,
                    struct anv_bo_cache *cache,
-                   uint64_t size, struct anv_bo **bo_out)
+                   uint64_t size, uint64_t bo_flags,
+                   struct anv_bo **bo_out)
 {
+   assert(bo_flags == (bo_flags & ANV_BO_CACHE_SUPPORTED_FLAGS));
+
    struct anv_cached_bo *bo =
       vk_alloc(&device->alloc, sizeof(struct anv_cached_bo), 8,
                VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
@@ -1250,6 +1268,16 @@ anv_bo_cache_alloc(struct anv_device *device,
       return result;
    }
 
+   bo->bo.flags = bo_flags;
+
+   if (!anv_vma_alloc(device, &bo->bo)) {
+      anv_gem_close(device, bo->bo.gem_handle);
+      vk_free(&device->alloc, bo);
+      return vk_errorf(device->instance, NULL,
+                       VK_ERROR_OUT_OF_DEVICE_MEMORY,
+                       "failed to allocate virtual address for BO");
+   }
+
    assert(bo->bo.gem_handle);
 
    pthread_mutex_lock(&cache->mutex);
@@ -1267,8 +1295,11 @@ anv_bo_cache_alloc(struct anv_device *device,
 VkResult
 anv_bo_cache_import(struct anv_device *device,
                     struct anv_bo_cache *cache,
-                    int fd, struct anv_bo **bo_out)
+                    int fd, uint64_t bo_flags,
+                    struct anv_bo **bo_out)
 {
+   assert(bo_flags == (bo_flags & ANV_BO_CACHE_SUPPORTED_FLAGS));
+
    pthread_mutex_lock(&cache->mutex);
 
    uint32_t gem_handle = anv_gem_fd_to_handle(device, fd);
@@ -1279,6 +1310,47 @@ anv_bo_cache_import(struct anv_device *device,
 
    struct anv_cached_bo *bo = anv_bo_cache_lookup_locked(cache, gem_handle);
    if (bo) {
+      /* We have to be careful how we combine flags so that it makes sense.
+       * Really, though, if we get to this case and it actually matters, the
+       * client has imported a BO twice in different ways and they get what
+       * they have coming.
+       */
+      uint64_t new_flags = 0;
+      new_flags |= (bo->bo.flags | bo_flags) & EXEC_OBJECT_WRITE;
+      new_flags |= (bo->bo.flags & bo_flags) & EXEC_OBJECT_ASYNC;
+      new_flags |= (bo->bo.flags & bo_flags) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+      new_flags |= (bo->bo.flags | bo_flags) & EXEC_OBJECT_PINNED;
+
+      /* It's theoretically possible for a BO to get imported such that it's
+       * both pinned and not pinned.  The only way this can happen is if it
+       * gets imported as both a semaphore and a memory object and that would
+       * be an application error.  Just fail out in that case.
+       */
+      if ((bo->bo.flags & EXEC_OBJECT_PINNED) !=
+          (bo_flags & EXEC_OBJECT_PINNED)) {
+         pthread_mutex_unlock(&cache->mutex);
+         return vk_errorf(device->instance, NULL,
+                          VK_ERROR_INVALID_EXTERNAL_HANDLE,
+                          "The same BO was imported two different ways");
+      }
+
+      /* It's also theoretically possible that someone could export a BO from
+       * one heap and import it into another or to import the same BO into two
+       * different heaps.  If this happens, we could potentially end up both
+       * allowing and disallowing 48-bit addresses.  There's not much we can
+       * do about it if we're pinning so we just throw an error and hope no
+       * app is actually that stupid.
+       */
+      if ((new_flags & EXEC_OBJECT_PINNED) &&
+          (bo->bo.flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) !=
+          (bo_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) {
+         return vk_errorf(device->instance, NULL,
+                          VK_ERROR_INVALID_EXTERNAL_HANDLE,
+                          "The same BO was imported on two different heaps");
+      }
+
+      bo->bo.flags = new_flags;
+
       __sync_fetch_and_add(&bo->refcount, 1);
    } else {
       off_t size = lseek(fd, 0, SEEK_END);
@@ -1299,6 +1371,16 @@ anv_bo_cache_import(struct anv_device *device,
       bo->refcount = 1;
 
       anv_bo_init(&bo->bo, gem_handle, size);
+      bo->bo.flags = bo_flags;
+
+      if (!anv_vma_alloc(device, &bo->bo)) {
+         anv_gem_close(device, bo->bo.gem_handle);
+         pthread_mutex_unlock(&cache->mutex);
+         vk_free(&device->alloc, bo);
+         return vk_errorf(device->instance, NULL,
+                          VK_ERROR_OUT_OF_DEVICE_MEMORY,
+                          "failed to allocate virtual address for BO");
+      }
 
       _mesa_hash_table_insert(cache->bo_map, (void *)(uintptr_t)gem_handle, bo);
    }
@@ -1381,6 +1463,8 @@ anv_bo_cache_release(struct anv_device *device,
    if (bo->bo.map)
       anv_gem_munmap(bo->bo.map, bo->bo.size);
 
+   anv_vma_free(device, &bo->bo);
+
    anv_gem_close(device, bo->bo.gem_handle);
 
    /* Don't unlock until we've actually closed the BO.  The whole point of