return NULL;
}
-/* All pointers in the ptr_free_list are assumed to be page-aligned. This
- * means that the bottom 12 bits should all be zero.
- */
-#define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
-#define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
-#define PFL_PACK(ptr, count) ({ \
- (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
-})
-
-static bool
-anv_ptr_free_list_pop(void **list, void **elem)
-{
- void *current = *list;
- while (PFL_PTR(current) != NULL) {
- void **next_ptr = PFL_PTR(current);
- void *new_ptr = VG_NOACCESS_READ(next_ptr);
- unsigned new_count = PFL_COUNT(current) + 1;
- void *new = PFL_PACK(new_ptr, new_count);
- void *old = __sync_val_compare_and_swap(list, current, new);
- if (old == current) {
- *elem = PFL_PTR(current);
- return true;
- }
- current = old;
- }
-
- return false;
-}
-
-static void
-anv_ptr_free_list_push(void **list, void *elem)
-{
- void *old, *current;
- void **next_ptr = elem;
-
- /* The pointer-based free list requires that the pointer be
- * page-aligned. This is because we use the bottom 12 bits of the
- * pointer to store a counter to solve the ABA concurrency problem.
- */
- assert(((uintptr_t)elem & 0xfff) == 0);
-
- old = *list;
- do {
- current = old;
- VG_NOACCESS_WRITE(next_ptr, PFL_PTR(current));
- unsigned new_count = PFL_COUNT(current) + 1;
- void *new = PFL_PACK(elem, new_count);
- old = __sync_val_compare_and_swap(list, current, new);
- } while (old != current);
-}
-
static VkResult
anv_block_pool_expand_range(struct anv_block_pool *pool,
uint32_t center_bo_offset, uint32_t size);
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)
+ uint32_t initial_size)
{
VkResult result;
pool->device = device;
- pool->bo_flags = bo_flags;
+ pool->use_softpin = device->instance->physicalDevice.use_softpin;
pool->nbos = 0;
pool->size = 0;
pool->center_bo_offset = 0;
pool->start_address = gen_canonical_address(start_address);
pool->map = NULL;
- if (!(pool->bo_flags & EXEC_OBJECT_PINNED)) {
+ if (pool->use_softpin) {
+ pool->bo = NULL;
+ pool->fd = -1;
+ } else {
/* Just make it 2GB up-front. The Linux kernel won't actually back it
* with pages until we either map and fault on one of them or we use
* userptr and send a chunk of it off to the GPU.
if (pool->fd == -1)
return vk_error(VK_ERROR_INITIALIZATION_FAILED);
- anv_bo_init(&pool->wrapper_bo, 0, 0);
- pool->wrapper_bo.is_wrapper = true;
+ pool->wrapper_bo = (struct anv_bo) {
+ .refcount = 1,
+ .offset = -1,
+ .is_wrapper = true,
+ };
pool->bo = &pool->wrapper_bo;
- } else {
- /* This pointer will always point to the first BO in the list */
- anv_bo_init(&pool->bos[0], 0, 0);
- pool->bo = &pool->bos[0];
-
- pool->fd = -1;
}
if (!u_vector_init(&pool->mmap_cleanups,
fail_mmap_cleanups:
u_vector_finish(&pool->mmap_cleanups);
fail_fd:
- if (!(pool->bo_flags & EXEC_OBJECT_PINNED))
+ if (pool->fd >= 0)
close(pool->fd);
return result;
munmap(cleanup->map, cleanup->size);
u_vector_finish(&pool->mmap_cleanups);
- if (!(pool->bo_flags & EXEC_OBJECT_PINNED))
+ if (pool->fd >= 0)
close(pool->fd);
}
anv_block_pool_expand_range(struct anv_block_pool *pool,
uint32_t center_bo_offset, uint32_t size)
{
- const bool use_softpin = !!(pool->bo_flags & EXEC_OBJECT_PINNED);
-
/* Assert that we only ever grow the pool */
assert(center_bo_offset >= pool->back_state.end);
assert(size - center_bo_offset >= pool->state.end);
/* Assert that we don't go outside the bounds of the memfd */
assert(center_bo_offset <= BLOCK_POOL_MEMFD_CENTER);
- assert(use_softpin ||
+ assert(pool->use_softpin ||
size - center_bo_offset <=
BLOCK_POOL_MEMFD_SIZE - BLOCK_POOL_MEMFD_CENTER);
- if (use_softpin) {
- uint32_t newbo_size = size - pool->size;
- uint32_t gem_handle = anv_gem_create(pool->device, newbo_size);
- void *map = anv_gem_mmap(pool->device, gem_handle, 0, newbo_size, 0);
- if (map == MAP_FAILED) {
- anv_gem_close(pool->device, gem_handle);
- return vk_errorf(pool->device->instance, pool->device,
- VK_ERROR_MEMORY_MAP_FAILED, "gem mmap failed: %m");
- }
+ /* For state pool BOs we have to be a bit careful about where we place them
+ * in the GTT. There are two documented workarounds for state base address
+ * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
+ * which state that those two base addresses do not support 48-bit
+ * addresses and need to be placed in the bottom 32-bit range.
+ * Unfortunately, this is not quite accurate.
+ *
+ * The real problem is that we always set the size of our state pools in
+ * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
+ * likely significantly smaller. We do this because we do not no at the
+ * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
+ * the pool during command buffer building so we don't actually have a
+ * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
+ * overflows 48 bits, the GPU appears to treat all accesses to the buffer
+ * as being out of bounds and returns zero. For dynamic state, this
+ * usually just leads to rendering corruptions, but shaders that are all
+ * zero hang the GPU immediately.
+ *
+ * The easiest solution to do is exactly what the bogus workarounds say to
+ * do: restrict these buffers to 32-bit addresses. We could also pin the
+ * BO to some particular location of our choosing, but that's significantly
+ * more work than just not setting a flag. So, we explicitly DO NOT set
+ * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
+ * hard work for us. When using softpin, we're in control and the fixed
+ * addresses we choose are fine for base addresses.
+ */
+ enum anv_bo_alloc_flags bo_alloc_flags = 0;
+ if (!pool->use_softpin)
+ bo_alloc_flags |= ANV_BO_ALLOC_32BIT_ADDRESS;
- /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
- * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
- * always created as I915_CACHING_CACHED, which on non-LLC means
- * snooped.
- *
- * On platforms that support softpin, we are not going to use userptr
- * anymore, but we still want to rely on the snooped states. So make
- * sure everything is set to I915_CACHING_CACHED.
- */
- if (!pool->device->info.has_llc)
- anv_gem_set_caching(pool->device, gem_handle, I915_CACHING_CACHED);
+ uint64_t bo_flags = 0;
+ if (pool->device->instance->physicalDevice.has_exec_capture)
+ bo_flags |= EXEC_OBJECT_CAPTURE;
+
+ if (pool->use_softpin) {
+ uint32_t new_bo_size = size - pool->size;
+ struct anv_bo *new_bo;
+ VkResult result = anv_device_alloc_bo(pool->device, new_bo_size,
+ bo_alloc_flags |
+ ANV_BO_ALLOC_FIXED_ADDRESS |
+ ANV_BO_ALLOC_MAPPED |
+ ANV_BO_ALLOC_SNOOPED,
+ &new_bo);
+ if (result != VK_SUCCESS)
+ return result;
assert(center_bo_offset == 0);
- struct anv_bo *bo = &pool->bos[pool->nbos++];
- anv_bo_init(bo, gem_handle, newbo_size);
- bo->offset = pool->start_address + pool->size;
- bo->flags = pool->bo_flags;
- bo->map = map;
+ new_bo->offset = pool->start_address + pool->size;
+ pool->bos[pool->nbos++] = new_bo;
+
+ /* This pointer will always point to the first BO in the list */
+ pool->bo = pool->bos[0];
} else {
/* Just leak the old map until we destroy the pool. We can't munmap it
* without races or imposing locking on the block allocate fast path. On
return vk_errorf(pool->device->instance, pool->device,
VK_ERROR_MEMORY_MAP_FAILED, "mmap failed: %m");
- uint32_t gem_handle = anv_gem_userptr(pool->device, map, size);
- if (gem_handle == 0) {
+ struct anv_bo *new_bo;
+ VkResult result = anv_device_import_bo_from_host_ptr(pool->device,
+ map, size,
+ bo_alloc_flags,
+ &new_bo);
+ if (result != VK_SUCCESS) {
munmap(map, size);
- return vk_errorf(pool->device->instance, pool->device,
- VK_ERROR_TOO_MANY_OBJECTS, "userptr failed: %m");
+ return result;
}
struct anv_mmap_cleanup *cleanup = u_vector_add(&pool->mmap_cleanups);
if (!cleanup) {
munmap(map, size);
- anv_gem_close(pool->device, gem_handle);
+ anv_device_release_bo(pool->device, new_bo);
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
}
cleanup->map = map;
pool->center_bo_offset = center_bo_offset;
pool->map = map + center_bo_offset;
- struct anv_bo *bo = &pool->bos[pool->nbos++];
- anv_bo_init(bo, gem_handle, size);
- bo->flags = pool->bo_flags;
- pool->wrapper_bo.map = bo;
+ pool->bos[pool->nbos++] = new_bo;
+ pool->wrapper_bo.map = new_bo;
}
assert(pool->nbos < ANV_MAX_BLOCK_POOL_BOS);
void*
anv_block_pool_map(struct anv_block_pool *pool, int32_t offset)
{
- if (pool->bo_flags & EXEC_OBJECT_PINNED) {
+ if (pool->use_softpin) {
struct anv_bo *bo = NULL;
int32_t bo_offset = 0;
anv_block_pool_foreach_bo(iter_bo, pool) {
if (state.next + block_size <= state.end) {
return state.next;
} else if (state.next <= state.end) {
- if (pool->bo_flags & EXEC_OBJECT_PINNED && state.next < state.end) {
+ if (pool->use_softpin && state.next < state.end) {
/* We need to grow the block pool, but still have some leftover
* space that can't be used by that particular allocation. So we
* add that as a "padding", and return it.
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)
+ uint32_t block_size)
{
VkResult result = anv_block_pool_init(&pool->block_pool, device,
start_address,
- block_size * 16,
- bo_flags);
+ block_size * 16);
if (result != VK_SUCCESS)
return result;
return state;
}
-struct bo_pool_bo_link {
- struct bo_pool_bo_link *next;
- struct anv_bo bo;
-};
-
void
anv_bo_pool_init(struct anv_bo_pool *pool, struct anv_device *device,
uint64_t bo_flags)
{
pool->device = device;
pool->bo_flags = bo_flags;
- memset(pool->free_list, 0, sizeof(pool->free_list));
+ for (unsigned i = 0; i < ARRAY_SIZE(pool->free_list); i++) {
+ util_sparse_array_free_list_init(&pool->free_list[i],
+ &device->bo_cache.bo_map, 0,
+ offsetof(struct anv_bo, free_index));
+ }
VG(VALGRIND_CREATE_MEMPOOL(pool, 0, false));
}
anv_bo_pool_finish(struct anv_bo_pool *pool)
{
for (unsigned i = 0; i < ARRAY_SIZE(pool->free_list); i++) {
- struct bo_pool_bo_link *link = PFL_PTR(pool->free_list[i]);
- while (link != NULL) {
- struct bo_pool_bo_link link_copy = VG_NOACCESS_READ(link);
-
- anv_gem_munmap(link_copy.bo.map, link_copy.bo.size);
- anv_vma_free(pool->device, &link_copy.bo);
- anv_gem_close(pool->device, link_copy.bo.gem_handle);
- link = link_copy.next;
+ while (1) {
+ struct anv_bo *bo =
+ util_sparse_array_free_list_pop_elem(&pool->free_list[i]);
+ if (bo == NULL)
+ break;
+
+ /* anv_device_release_bo is going to "free" it */
+ VG(VALGRIND_MALLOCLIKE_BLOCK(bo->map, bo->size, 0, 1));
+ anv_device_release_bo(pool->device, bo);
}
}
}
VkResult
-anv_bo_pool_alloc(struct anv_bo_pool *pool, struct anv_bo *bo, uint32_t size)
+anv_bo_pool_alloc(struct anv_bo_pool *pool, uint32_t size,
+ struct anv_bo **bo_out)
{
- VkResult result;
-
const unsigned size_log2 = size < 4096 ? 12 : ilog2_round_up(size);
const unsigned pow2_size = 1 << size_log2;
const unsigned bucket = size_log2 - 12;
assert(bucket < ARRAY_SIZE(pool->free_list));
- void *next_free_void;
- if (anv_ptr_free_list_pop(&pool->free_list[bucket], &next_free_void)) {
- struct bo_pool_bo_link *next_free = next_free_void;
- *bo = VG_NOACCESS_READ(&next_free->bo);
- assert(bo->gem_handle);
- assert(bo->map == next_free);
- assert(size <= bo->size);
-
+ struct anv_bo *bo =
+ util_sparse_array_free_list_pop_elem(&pool->free_list[bucket]);
+ if (bo != NULL) {
VG(VALGRIND_MEMPOOL_ALLOC(pool, bo->map, size));
-
+ *bo_out = bo;
return VK_SUCCESS;
}
- struct anv_bo new_bo;
-
- result = anv_bo_init_new(&new_bo, pool->device, pow2_size);
+ VkResult result = anv_device_alloc_bo(pool->device,
+ pow2_size,
+ ANV_BO_ALLOC_MAPPED |
+ ANV_BO_ALLOC_SNOOPED,
+ &bo);
if (result != VK_SUCCESS)
return result;
- new_bo.flags = pool->bo_flags;
-
- if (!anv_vma_alloc(pool->device, &new_bo))
- return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
-
- assert(new_bo.size == pow2_size);
-
- new_bo.map = anv_gem_mmap(pool->device, new_bo.gem_handle, 0, pow2_size, 0);
- if (new_bo.map == MAP_FAILED) {
- anv_gem_close(pool->device, new_bo.gem_handle);
- anv_vma_free(pool->device, &new_bo);
- return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
- }
-
- /* We are removing the state flushes, so lets make sure that these buffers
- * are cached/snooped.
- */
- if (!pool->device->info.has_llc) {
- anv_gem_set_caching(pool->device, new_bo.gem_handle,
- I915_CACHING_CACHED);
- }
-
- *bo = new_bo;
-
+ /* We want it to look like it came from this pool */
+ VG(VALGRIND_FREELIKE_BLOCK(bo->map, 0));
VG(VALGRIND_MEMPOOL_ALLOC(pool, bo->map, size));
+ *bo_out = bo;
+
return VK_SUCCESS;
}
void
-anv_bo_pool_free(struct anv_bo_pool *pool, const struct anv_bo *bo_in)
+anv_bo_pool_free(struct anv_bo_pool *pool, struct anv_bo *bo)
{
- /* Make a copy in case the anv_bo happens to be storred in the BO */
- struct anv_bo bo = *bo_in;
+ VG(VALGRIND_MEMPOOL_FREE(pool, bo->map));
- VG(VALGRIND_MEMPOOL_FREE(pool, bo.map));
-
- struct bo_pool_bo_link *link = bo.map;
- VG_NOACCESS_WRITE(&link->bo, bo);
-
- assert(util_is_power_of_two_or_zero(bo.size));
- const unsigned size_log2 = ilog2_round_up(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));
- anv_ptr_free_list_push(&pool->free_list[bucket], link);
+ assert(util_sparse_array_get(&pool->device->bo_cache.bo_map,
+ bo->gem_handle) == bo);
+ util_sparse_array_free_list_push(&pool->free_list[bucket],
+ &bo->gem_handle, 1);
}
// Scratch pool
{
for (unsigned s = 0; s < MESA_SHADER_STAGES; s++) {
for (unsigned i = 0; i < 16; i++) {
- struct anv_scratch_bo *bo = &pool->bos[i][s];
- if (bo->exists > 0) {
- anv_vma_free(device, &bo->bo);
- anv_gem_close(device, bo->bo.gem_handle);
- }
+ if (pool->bos[i][s] != NULL)
+ anv_device_release_bo(device, pool->bos[i][s]);
}
}
}
unsigned scratch_size_log2 = ffs(per_thread_scratch / 2048);
assert(scratch_size_log2 < 16);
- struct anv_scratch_bo *bo = &pool->bos[scratch_size_log2][stage];
+ struct anv_bo *bo = p_atomic_read(&pool->bos[scratch_size_log2][stage]);
- /* We can use "exists" to shortcut and ignore the critical section */
- if (bo->exists)
- return &bo->bo;
-
- pthread_mutex_lock(&device->mutex);
-
- __sync_synchronize();
- if (bo->exists) {
- pthread_mutex_unlock(&device->mutex);
- return &bo->bo;
- }
+ if (bo != NULL)
+ return bo;
const struct anv_physical_device *physical_device =
&device->instance->physicalDevice;
uint32_t size = per_thread_scratch * max_threads[stage];
- anv_bo_init_new(&bo->bo, device, size);
-
/* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
* are still relative to the general state base address. When we emit
* STATE_BASE_ADDRESS, we set general state base address to 0 and the size
*
* so nothing will ever touch the top page.
*/
- assert(!(bo->bo.flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS));
-
- if (device->instance->physicalDevice.has_exec_async)
- bo->bo.flags |= EXEC_OBJECT_ASYNC;
-
- if (device->instance->physicalDevice.use_softpin)
- bo->bo.flags |= EXEC_OBJECT_PINNED;
-
- anv_vma_alloc(device, &bo->bo);
-
- /* Set the exists last because it may be read by other threads */
- __sync_synchronize();
- bo->exists = true;
-
- pthread_mutex_unlock(&device->mutex);
+ VkResult result = anv_device_alloc_bo(device, size,
+ ANV_BO_ALLOC_32BIT_ADDRESS, &bo);
+ if (result != VK_SUCCESS)
+ return NULL; /* TODO */
- return &bo->bo;
+ struct anv_bo *current_bo =
+ p_atomic_cmpxchg(&pool->bos[scratch_size_log2][stage], NULL, bo);
+ if (current_bo) {
+ anv_device_release_bo(device, bo);
+ return current_bo;
+ } else {
+ return bo;
+ }
}
VkResult
/* The kernel is going to give us whole pages anyway */
size = align_u64(size, 4096);
- struct anv_bo new_bo;
- VkResult result = anv_bo_init_new(&new_bo, device, size);
- if (result != VK_SUCCESS)
- return result;
+ uint32_t gem_handle = anv_gem_create(device, size);
+ if (gem_handle == 0)
+ return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
- new_bo.flags = bo_flags;
- new_bo.is_external = (alloc_flags & ANV_BO_ALLOC_EXTERNAL);
+ struct anv_bo new_bo = {
+ .gem_handle = gem_handle,
+ .refcount = 1,
+ .offset = -1,
+ .size = size,
+ .flags = bo_flags,
+ .is_external = (alloc_flags & ANV_BO_ALLOC_EXTERNAL),
+ };
if (alloc_flags & ANV_BO_ALLOC_MAPPED) {
new_bo.map = anv_gem_mmap(device, new_bo.gem_handle, 0, size, 0);
}
__sync_fetch_and_add(&bo->refcount, 1);
} else {
- struct anv_bo new_bo;
- anv_bo_init(&new_bo, gem_handle, size);
- new_bo.map = host_ptr;
- new_bo.flags = bo_flags;
- new_bo.is_external = true;
- new_bo.from_host_ptr = true;
+ struct anv_bo new_bo = {
+ .gem_handle = gem_handle,
+ .refcount = 1,
+ .offset = -1,
+ .size = size,
+ .map = host_ptr,
+ .flags = bo_flags,
+ .is_external = true,
+ .from_host_ptr = true,
+ };
if (!anv_vma_alloc(device, &new_bo)) {
anv_gem_close(device, new_bo.gem_handle);
return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
}
- struct anv_bo new_bo;
- anv_bo_init(&new_bo, gem_handle, size);
- new_bo.flags = bo_flags;
- new_bo.is_external = true;
+ struct anv_bo new_bo = {
+ .gem_handle = gem_handle,
+ .refcount = 1,
+ .offset = -1,
+ .size = size,
+ .flags = bo_flags,
+ .is_external = true,
+ };
if (!anv_vma_alloc(device, &new_bo)) {
anv_gem_close(device, new_bo.gem_handle);
if (!bo->has_fixed_address)
anv_vma_free(device, bo);
- anv_gem_close(device, bo->gem_handle);
+ uint32_t gem_handle = bo->gem_handle;
+
+ /* Memset the BO just in case. The refcount being zero should be enough to
+ * prevent someone from assuming the data is valid but it's safer to just
+ * stomp to zero just in case. We explicitly do this *before* we close the
+ * GEM handle to ensure that if anyone allocates something and gets the
+ * same GEM handle, the memset has already happen and won't stomp all over
+ * any data they may write in this BO.
+ */
+ memset(bo, 0, sizeof(*bo));
+
+ anv_gem_close(device, gem_handle);
/* Don't unlock until we've actually closed the BO. The whole point of
* the BO cache is to ensure that we correctly handle races with creating
projects / mesa.git / blobdiff