2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
30 #include "anv_private.h"
32 #include "util/simple_mtx.h"
33 #include "util/anon_file.h"
36 #define VG_NOACCESS_READ(__ptr) ({ \
37 VALGRIND_MAKE_MEM_DEFINED((__ptr), sizeof(*(__ptr))); \
38 __typeof(*(__ptr)) __val = *(__ptr); \
39 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr)));\
42 #define VG_NOACCESS_WRITE(__ptr, __val) ({ \
43 VALGRIND_MAKE_MEM_UNDEFINED((__ptr), sizeof(*(__ptr))); \
45 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr))); \
48 #define VG_NOACCESS_READ(__ptr) (*(__ptr))
49 #define VG_NOACCESS_WRITE(__ptr, __val) (*(__ptr) = (__val))
53 #define MAP_POPULATE 0
58 * - Lock free (except when resizing underlying bos)
60 * - Constant time allocation with typically only one atomic
62 * - Multiple allocation sizes without fragmentation
64 * - Can grow while keeping addresses and offset of contents stable
66 * - All allocations within one bo so we can point one of the
67 * STATE_BASE_ADDRESS pointers at it.
69 * The overall design is a two-level allocator: top level is a fixed size, big
70 * block (8k) allocator, which operates out of a bo. Allocation is done by
71 * either pulling a block from the free list or growing the used range of the
72 * bo. Growing the range may run out of space in the bo which we then need to
73 * grow. Growing the bo is tricky in a multi-threaded, lockless environment:
74 * we need to keep all pointers and contents in the old map valid. GEM bos in
75 * general can't grow, but we use a trick: we create a memfd and use ftruncate
76 * to grow it as necessary. We mmap the new size and then create a gem bo for
77 * it using the new gem userptr ioctl. Without heavy-handed locking around
78 * our allocation fast-path, there isn't really a way to munmap the old mmap,
79 * so we just keep it around until garbage collection time. While the block
80 * allocator is lockless for normal operations, we block other threads trying
81 * to allocate while we're growing the map. It sholdn't happen often, and
82 * growing is fast anyway.
84 * At the next level we can use various sub-allocators. The state pool is a
85 * pool of smaller, fixed size objects, which operates much like the block
86 * pool. It uses a free list for freeing objects, but when it runs out of
87 * space it just allocates a new block from the block pool. This allocator is
88 * intended for longer lived state objects such as SURFACE_STATE and most
89 * other persistent state objects in the API. We may need to track more info
90 * with these object and a pointer back to the CPU object (eg VkImage). In
91 * those cases we just allocate a slightly bigger object and put the extra
92 * state after the GPU state object.
94 * The state stream allocator works similar to how the i965 DRI driver streams
95 * all its state. Even with Vulkan, we need to emit transient state (whether
96 * surface state base or dynamic state base), and for that we can just get a
97 * block and fill it up. These cases are local to a command buffer and the
98 * sub-allocator need not be thread safe. The streaming allocator gets a new
99 * block when it runs out of space and chains them together so they can be
103 /* Allocations are always at least 64 byte aligned, so 1 is an invalid value.
104 * We use it to indicate the free list is empty. */
105 #define EMPTY UINT32_MAX
107 #define PAGE_SIZE 4096
109 struct anv_mmap_cleanup
{
114 static inline uint32_t
115 ilog2_round_up(uint32_t value
)
118 return 32 - __builtin_clz(value
- 1);
121 static inline uint32_t
122 round_to_power_of_two(uint32_t value
)
124 return 1 << ilog2_round_up(value
);
127 struct anv_state_table_cleanup
{
132 #define ANV_STATE_TABLE_CLEANUP_INIT ((struct anv_state_table_cleanup){0})
133 #define ANV_STATE_ENTRY_SIZE (sizeof(struct anv_free_entry))
136 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
);
139 anv_state_table_init(struct anv_state_table
*table
,
140 struct anv_device
*device
,
141 uint32_t initial_entries
)
145 table
->device
= device
;
147 /* Just make it 2GB up-front. The Linux kernel won't actually back it
148 * with pages until we either map and fault on one of them or we use
149 * userptr and send a chunk of it off to the GPU.
151 table
->fd
= os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE
, "state table");
152 if (table
->fd
== -1) {
153 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
157 if (!u_vector_init(&table
->cleanups
,
158 round_to_power_of_two(sizeof(struct anv_state_table_cleanup
)),
160 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
164 table
->state
.next
= 0;
165 table
->state
.end
= 0;
168 uint32_t initial_size
= initial_entries
* ANV_STATE_ENTRY_SIZE
;
169 result
= anv_state_table_expand_range(table
, initial_size
);
170 if (result
!= VK_SUCCESS
)
176 u_vector_finish(&table
->cleanups
);
184 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
)
187 struct anv_state_table_cleanup
*cleanup
;
189 /* Assert that we only ever grow the pool */
190 assert(size
>= table
->state
.end
);
192 /* Make sure that we don't go outside the bounds of the memfd */
193 if (size
> BLOCK_POOL_MEMFD_SIZE
)
194 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
196 cleanup
= u_vector_add(&table
->cleanups
);
198 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
200 *cleanup
= ANV_STATE_TABLE_CLEANUP_INIT
;
202 /* Just leak the old map until we destroy the pool. We can't munmap it
203 * without races or imposing locking on the block allocate fast path. On
204 * the whole the leaked maps adds up to less than the size of the
205 * current map. MAP_POPULATE seems like the right thing to do, but we
206 * should try to get some numbers.
208 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
209 MAP_SHARED
| MAP_POPULATE
, table
->fd
, 0);
210 if (map
== MAP_FAILED
) {
211 return vk_errorf(table
->device
->instance
, table
->device
,
212 VK_ERROR_OUT_OF_HOST_MEMORY
, "mmap failed: %m");
216 cleanup
->size
= size
;
225 anv_state_table_grow(struct anv_state_table
*table
)
227 VkResult result
= VK_SUCCESS
;
229 uint32_t used
= align_u32(table
->state
.next
* ANV_STATE_ENTRY_SIZE
,
231 uint32_t old_size
= table
->size
;
233 /* The block pool is always initialized to a nonzero size and this function
234 * is always called after initialization.
236 assert(old_size
> 0);
238 uint32_t required
= MAX2(used
, old_size
);
239 if (used
* 2 <= required
) {
240 /* If we're in this case then this isn't the firsta allocation and we
241 * already have enough space on both sides to hold double what we
242 * have allocated. There's nothing for us to do.
247 uint32_t size
= old_size
* 2;
248 while (size
< required
)
251 assert(size
> table
->size
);
253 result
= anv_state_table_expand_range(table
, size
);
260 anv_state_table_finish(struct anv_state_table
*table
)
262 struct anv_state_table_cleanup
*cleanup
;
264 u_vector_foreach(cleanup
, &table
->cleanups
) {
266 munmap(cleanup
->map
, cleanup
->size
);
269 u_vector_finish(&table
->cleanups
);
275 anv_state_table_add(struct anv_state_table
*table
, uint32_t *idx
,
278 struct anv_block_state state
, old
, new;
284 state
.u64
= __sync_fetch_and_add(&table
->state
.u64
, count
);
285 if (state
.next
+ count
<= state
.end
) {
287 struct anv_free_entry
*entry
= &table
->map
[state
.next
];
288 for (int i
= 0; i
< count
; i
++) {
289 entry
[i
].state
.idx
= state
.next
+ i
;
293 } else if (state
.next
<= state
.end
) {
294 /* We allocated the first block outside the pool so we have to grow
295 * the pool. pool_state->next acts a mutex: threads who try to
296 * allocate now will get block indexes above the current limit and
297 * hit futex_wait below.
299 new.next
= state
.next
+ count
;
301 result
= anv_state_table_grow(table
);
302 if (result
!= VK_SUCCESS
)
304 new.end
= table
->size
/ ANV_STATE_ENTRY_SIZE
;
305 } while (new.end
< new.next
);
307 old
.u64
= __sync_lock_test_and_set(&table
->state
.u64
, new.u64
);
308 if (old
.next
!= state
.next
)
309 futex_wake(&table
->state
.end
, INT_MAX
);
311 futex_wait(&table
->state
.end
, state
.end
, NULL
);
318 anv_free_list_push(union anv_free_list
*list
,
319 struct anv_state_table
*table
,
320 uint32_t first
, uint32_t count
)
322 union anv_free_list current
, old
, new;
323 uint32_t last
= first
;
325 for (uint32_t i
= 1; i
< count
; i
++, last
++)
326 table
->map
[last
].next
= last
+ 1;
331 table
->map
[last
].next
= current
.offset
;
333 new.count
= current
.count
+ 1;
334 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
335 } while (old
.u64
!= current
.u64
);
339 anv_free_list_pop(union anv_free_list
*list
,
340 struct anv_state_table
*table
)
342 union anv_free_list current
, new, old
;
344 current
.u64
= list
->u64
;
345 while (current
.offset
!= EMPTY
) {
346 __sync_synchronize();
347 new.offset
= table
->map
[current
.offset
].next
;
348 new.count
= current
.count
+ 1;
349 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
350 if (old
.u64
== current
.u64
) {
351 struct anv_free_entry
*entry
= &table
->map
[current
.offset
];
352 return &entry
->state
;
360 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
361 * means that the bottom 12 bits should all be zero.
363 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
364 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
365 #define PFL_PACK(ptr, count) ({ \
366 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
370 anv_ptr_free_list_pop(void **list
, void **elem
)
372 void *current
= *list
;
373 while (PFL_PTR(current
) != NULL
) {
374 void **next_ptr
= PFL_PTR(current
);
375 void *new_ptr
= VG_NOACCESS_READ(next_ptr
);
376 unsigned new_count
= PFL_COUNT(current
) + 1;
377 void *new = PFL_PACK(new_ptr
, new_count
);
378 void *old
= __sync_val_compare_and_swap(list
, current
, new);
379 if (old
== current
) {
380 *elem
= PFL_PTR(current
);
390 anv_ptr_free_list_push(void **list
, void *elem
)
393 void **next_ptr
= elem
;
395 /* The pointer-based free list requires that the pointer be
396 * page-aligned. This is because we use the bottom 12 bits of the
397 * pointer to store a counter to solve the ABA concurrency problem.
399 assert(((uintptr_t)elem
& 0xfff) == 0);
404 VG_NOACCESS_WRITE(next_ptr
, PFL_PTR(current
));
405 unsigned new_count
= PFL_COUNT(current
) + 1;
406 void *new = PFL_PACK(elem
, new_count
);
407 old
= __sync_val_compare_and_swap(list
, current
, new);
408 } while (old
!= current
);
412 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
413 uint32_t center_bo_offset
, uint32_t size
);
416 anv_block_pool_init(struct anv_block_pool
*pool
,
417 struct anv_device
*device
,
418 uint64_t start_address
,
419 uint32_t initial_size
,
424 pool
->device
= device
;
425 pool
->bo_flags
= bo_flags
;
428 pool
->center_bo_offset
= 0;
429 pool
->start_address
= gen_canonical_address(start_address
);
432 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
)) {
433 /* Just make it 2GB up-front. The Linux kernel won't actually back it
434 * with pages until we either map and fault on one of them or we use
435 * userptr and send a chunk of it off to the GPU.
437 pool
->fd
= os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE
, "block pool");
439 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
441 anv_bo_init(&pool
->wrapper_bo
, 0, 0);
442 pool
->wrapper_bo
.is_wrapper
= true;
443 pool
->bo
= &pool
->wrapper_bo
;
445 /* This pointer will always point to the first BO in the list */
446 anv_bo_init(&pool
->bos
[0], 0, 0);
447 pool
->bo
= &pool
->bos
[0];
452 if (!u_vector_init(&pool
->mmap_cleanups
,
453 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
455 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
459 pool
->state
.next
= 0;
461 pool
->back_state
.next
= 0;
462 pool
->back_state
.end
= 0;
464 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
465 if (result
!= VK_SUCCESS
)
466 goto fail_mmap_cleanups
;
468 /* Make the entire pool available in the front of the pool. If back
469 * allocation needs to use this space, the "ends" will be re-arranged.
471 pool
->state
.end
= pool
->size
;
476 u_vector_finish(&pool
->mmap_cleanups
);
478 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
485 anv_block_pool_finish(struct anv_block_pool
*pool
)
487 anv_block_pool_foreach_bo(bo
, pool
) {
489 anv_gem_munmap(bo
->map
, bo
->size
);
490 anv_gem_close(pool
->device
, bo
->gem_handle
);
493 struct anv_mmap_cleanup
*cleanup
;
494 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
)
495 munmap(cleanup
->map
, cleanup
->size
);
496 u_vector_finish(&pool
->mmap_cleanups
);
498 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
503 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
504 uint32_t center_bo_offset
, uint32_t size
)
506 const bool use_softpin
= !!(pool
->bo_flags
& EXEC_OBJECT_PINNED
);
508 /* Assert that we only ever grow the pool */
509 assert(center_bo_offset
>= pool
->back_state
.end
);
510 assert(size
- center_bo_offset
>= pool
->state
.end
);
512 /* Assert that we don't go outside the bounds of the memfd */
513 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
514 assert(use_softpin
||
515 size
- center_bo_offset
<=
516 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
519 uint32_t newbo_size
= size
- pool
->size
;
520 uint32_t gem_handle
= anv_gem_create(pool
->device
, newbo_size
);
521 void *map
= anv_gem_mmap(pool
->device
, gem_handle
, 0, newbo_size
, 0);
522 if (map
== MAP_FAILED
) {
523 anv_gem_close(pool
->device
, gem_handle
);
524 return vk_errorf(pool
->device
->instance
, pool
->device
,
525 VK_ERROR_MEMORY_MAP_FAILED
, "gem mmap failed: %m");
528 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
529 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
530 * always created as I915_CACHING_CACHED, which on non-LLC means
533 * On platforms that support softpin, we are not going to use userptr
534 * anymore, but we still want to rely on the snooped states. So make
535 * sure everything is set to I915_CACHING_CACHED.
537 if (!pool
->device
->info
.has_llc
)
538 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_CACHED
);
540 assert(center_bo_offset
== 0);
542 struct anv_bo
*bo
= &pool
->bos
[pool
->nbos
++];
543 anv_bo_init(bo
, gem_handle
, newbo_size
);
544 bo
->offset
= pool
->start_address
+ pool
->size
;
545 bo
->flags
= pool
->bo_flags
;
548 /* Just leak the old map until we destroy the pool. We can't munmap it
549 * without races or imposing locking on the block allocate fast path. On
550 * the whole the leaked maps adds up to less than the size of the
551 * current map. MAP_POPULATE seems like the right thing to do, but we
552 * should try to get some numbers.
554 void *map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
555 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
556 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
557 if (map
== MAP_FAILED
)
558 return vk_errorf(pool
->device
->instance
, pool
->device
,
559 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
561 uint32_t gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
562 if (gem_handle
== 0) {
564 return vk_errorf(pool
->device
->instance
, pool
->device
,
565 VK_ERROR_TOO_MANY_OBJECTS
, "userptr failed: %m");
568 struct anv_mmap_cleanup
*cleanup
= u_vector_add(&pool
->mmap_cleanups
);
571 anv_gem_close(pool
->device
, gem_handle
);
572 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
575 cleanup
->size
= size
;
577 /* Now that we mapped the new memory, we can write the new
578 * center_bo_offset back into pool and update pool->map. */
579 pool
->center_bo_offset
= center_bo_offset
;
580 pool
->map
= map
+ center_bo_offset
;
582 struct anv_bo
*bo
= &pool
->bos
[pool
->nbos
++];
583 anv_bo_init(bo
, gem_handle
, size
);
584 bo
->flags
= pool
->bo_flags
;
585 pool
->wrapper_bo
.map
= bo
;
588 assert(pool
->nbos
< ANV_MAX_BLOCK_POOL_BOS
);
594 /** Returns current memory map of the block pool.
596 * The returned pointer points to the map for the memory at the specified
597 * offset. The offset parameter is relative to the "center" of the block pool
598 * rather than the start of the block pool BO map.
601 anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
)
603 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
) {
604 struct anv_bo
*bo
= NULL
;
605 int32_t bo_offset
= 0;
606 anv_block_pool_foreach_bo(iter_bo
, pool
) {
607 if (offset
< bo_offset
+ iter_bo
->size
) {
611 bo_offset
+= iter_bo
->size
;
614 assert(offset
>= bo_offset
);
616 return bo
->map
+ (offset
- bo_offset
);
618 return pool
->map
+ offset
;
622 /** Grows and re-centers the block pool.
624 * We grow the block pool in one or both directions in such a way that the
625 * following conditions are met:
627 * 1) The size of the entire pool is always a power of two.
629 * 2) The pool only grows on both ends. Neither end can get
632 * 3) At the end of the allocation, we have about twice as much space
633 * allocated for each end as we have used. This way the pool doesn't
634 * grow too far in one direction or the other.
636 * 4) If the _alloc_back() has never been called, then the back portion of
637 * the pool retains a size of zero. (This makes it easier for users of
638 * the block pool that only want a one-sided pool.)
640 * 5) We have enough space allocated for at least one more block in
641 * whichever side `state` points to.
643 * 6) The center of the pool is always aligned to both the block_size of
644 * the pool and a 4K CPU page.
647 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
649 VkResult result
= VK_SUCCESS
;
651 pthread_mutex_lock(&pool
->device
->mutex
);
653 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
655 /* Gather a little usage information on the pool. Since we may have
656 * threadsd waiting in queue to get some storage while we resize, it's
657 * actually possible that total_used will be larger than old_size. In
658 * particular, block_pool_alloc() increments state->next prior to
659 * calling block_pool_grow, so this ensures that we get enough space for
660 * which ever side tries to grow the pool.
662 * We align to a page size because it makes it easier to do our
663 * calculations later in such a way that we state page-aigned.
665 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
666 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
667 uint32_t total_used
= front_used
+ back_used
;
669 assert(state
== &pool
->state
|| back_used
> 0);
671 uint32_t old_size
= pool
->size
;
673 /* The block pool is always initialized to a nonzero size and this function
674 * is always called after initialization.
676 assert(old_size
> 0);
678 /* The back_used and front_used may actually be smaller than the actual
679 * requirement because they are based on the next pointers which are
680 * updated prior to calling this function.
682 uint32_t back_required
= MAX2(back_used
, pool
->center_bo_offset
);
683 uint32_t front_required
= MAX2(front_used
, old_size
- pool
->center_bo_offset
);
685 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
686 /* If we're in this case then this isn't the firsta allocation and we
687 * already have enough space on both sides to hold double what we
688 * have allocated. There's nothing for us to do.
693 uint32_t size
= old_size
* 2;
694 while (size
< back_required
+ front_required
)
697 assert(size
> pool
->size
);
699 /* We compute a new center_bo_offset such that, when we double the size
700 * of the pool, we maintain the ratio of how much is used by each side.
701 * This way things should remain more-or-less balanced.
703 uint32_t center_bo_offset
;
704 if (back_used
== 0) {
705 /* If we're in this case then we have never called alloc_back(). In
706 * this case, we want keep the offset at 0 to make things as simple
707 * as possible for users that don't care about back allocations.
709 center_bo_offset
= 0;
711 /* Try to "center" the allocation based on how much is currently in
712 * use on each side of the center line.
714 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
716 /* Align down to a multiple of the page size */
717 center_bo_offset
&= ~(PAGE_SIZE
- 1);
719 assert(center_bo_offset
>= back_used
);
721 /* Make sure we don't shrink the back end of the pool */
722 if (center_bo_offset
< back_required
)
723 center_bo_offset
= back_required
;
725 /* Make sure that we don't shrink the front end of the pool */
726 if (size
- center_bo_offset
< front_required
)
727 center_bo_offset
= size
- front_required
;
730 assert(center_bo_offset
% PAGE_SIZE
== 0);
732 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
735 pthread_mutex_unlock(&pool
->device
->mutex
);
737 if (result
== VK_SUCCESS
) {
738 /* Return the appropriate new size. This function never actually
739 * updates state->next. Instead, we let the caller do that because it
740 * needs to do so in order to maintain its concurrency model.
742 if (state
== &pool
->state
) {
743 return pool
->size
- pool
->center_bo_offset
;
745 assert(pool
->center_bo_offset
> 0);
746 return pool
->center_bo_offset
;
754 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
755 struct anv_block_state
*pool_state
,
756 uint32_t block_size
, uint32_t *padding
)
758 struct anv_block_state state
, old
, new;
760 /* Most allocations won't generate any padding */
765 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
766 if (state
.next
+ block_size
<= state
.end
) {
768 } else if (state
.next
<= state
.end
) {
769 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
&& state
.next
< state
.end
) {
770 /* We need to grow the block pool, but still have some leftover
771 * space that can't be used by that particular allocation. So we
772 * add that as a "padding", and return it.
774 uint32_t leftover
= state
.end
- state
.next
;
776 /* If there is some leftover space in the pool, the caller must
779 assert(leftover
== 0 || padding
);
782 state
.next
+= leftover
;
785 /* We allocated the first block outside the pool so we have to grow
786 * the pool. pool_state->next acts a mutex: threads who try to
787 * allocate now will get block indexes above the current limit and
788 * hit futex_wait below.
790 new.next
= state
.next
+ block_size
;
792 new.end
= anv_block_pool_grow(pool
, pool_state
);
793 } while (new.end
< new.next
);
795 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
796 if (old
.next
!= state
.next
)
797 futex_wake(&pool_state
->end
, INT_MAX
);
800 futex_wait(&pool_state
->end
, state
.end
, NULL
);
807 anv_block_pool_alloc(struct anv_block_pool
*pool
,
808 uint32_t block_size
, uint32_t *padding
)
812 offset
= anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
, padding
);
817 /* Allocates a block out of the back of the block pool.
819 * This will allocated a block earlier than the "start" of the block pool.
820 * The offsets returned from this function will be negative but will still
821 * be correct relative to the block pool's map pointer.
823 * If you ever use anv_block_pool_alloc_back, then you will have to do
824 * gymnastics with the block pool's BO when doing relocations.
827 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
830 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
833 /* The offset we get out of anv_block_pool_alloc_new() is actually the
834 * number of bytes downwards from the middle to the end of the block.
835 * We need to turn it into a (negative) offset from the middle to the
836 * start of the block.
839 return -(offset
+ block_size
);
843 anv_state_pool_init(struct anv_state_pool
*pool
,
844 struct anv_device
*device
,
845 uint64_t start_address
,
849 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
853 if (result
!= VK_SUCCESS
)
856 result
= anv_state_table_init(&pool
->table
, device
, 64);
857 if (result
!= VK_SUCCESS
) {
858 anv_block_pool_finish(&pool
->block_pool
);
862 assert(util_is_power_of_two_or_zero(block_size
));
863 pool
->block_size
= block_size
;
864 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
865 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
866 pool
->buckets
[i
].free_list
= ANV_FREE_LIST_EMPTY
;
867 pool
->buckets
[i
].block
.next
= 0;
868 pool
->buckets
[i
].block
.end
= 0;
870 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
876 anv_state_pool_finish(struct anv_state_pool
*pool
)
878 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
879 anv_state_table_finish(&pool
->table
);
880 anv_block_pool_finish(&pool
->block_pool
);
884 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
885 struct anv_block_pool
*block_pool
,
890 struct anv_block_state block
, old
, new;
893 /* We don't always use anv_block_pool_alloc(), which would set *padding to
894 * zero for us. So if we have a pointer to padding, we must zero it out
895 * ourselves here, to make sure we always return some sensible value.
900 /* If our state is large, we don't need any sub-allocation from a block.
901 * Instead, we just grab whole (potentially large) blocks.
903 if (state_size
>= block_size
)
904 return anv_block_pool_alloc(block_pool
, state_size
, padding
);
907 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
909 if (block
.next
< block
.end
) {
911 } else if (block
.next
== block
.end
) {
912 offset
= anv_block_pool_alloc(block_pool
, block_size
, padding
);
913 new.next
= offset
+ state_size
;
914 new.end
= offset
+ block_size
;
915 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
916 if (old
.next
!= block
.next
)
917 futex_wake(&pool
->block
.end
, INT_MAX
);
920 futex_wait(&pool
->block
.end
, block
.end
, NULL
);
926 anv_state_pool_get_bucket(uint32_t size
)
928 unsigned size_log2
= ilog2_round_up(size
);
929 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
930 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
931 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
932 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
936 anv_state_pool_get_bucket_size(uint32_t bucket
)
938 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
939 return 1 << size_log2
;
942 /** Helper to push a chunk into the state table.
944 * It creates 'count' entries into the state table and update their sizes,
945 * offsets and maps, also pushing them as "free" states.
948 anv_state_pool_return_blocks(struct anv_state_pool
*pool
,
949 uint32_t chunk_offset
, uint32_t count
,
952 /* Disallow returning 0 chunks */
955 /* Make sure we always return chunks aligned to the block_size */
956 assert(chunk_offset
% block_size
== 0);
959 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &st_idx
, count
);
960 assert(result
== VK_SUCCESS
);
961 for (int i
= 0; i
< count
; i
++) {
962 /* update states that were added back to the state table */
963 struct anv_state
*state_i
= anv_state_table_get(&pool
->table
,
965 state_i
->alloc_size
= block_size
;
966 state_i
->offset
= chunk_offset
+ block_size
* i
;
967 state_i
->map
= anv_block_pool_map(&pool
->block_pool
, state_i
->offset
);
970 uint32_t block_bucket
= anv_state_pool_get_bucket(block_size
);
971 anv_free_list_push(&pool
->buckets
[block_bucket
].free_list
,
972 &pool
->table
, st_idx
, count
);
975 /** Returns a chunk of memory back to the state pool.
977 * Do a two-level split. If chunk_size is bigger than divisor
978 * (pool->block_size), we return as many divisor sized blocks as we can, from
979 * the end of the chunk.
981 * The remaining is then split into smaller blocks (starting at small_size if
982 * it is non-zero), with larger blocks always being taken from the end of the
986 anv_state_pool_return_chunk(struct anv_state_pool
*pool
,
987 uint32_t chunk_offset
, uint32_t chunk_size
,
990 uint32_t divisor
= pool
->block_size
;
991 uint32_t nblocks
= chunk_size
/ divisor
;
992 uint32_t rest
= chunk_size
- nblocks
* divisor
;
995 /* First return divisor aligned and sized chunks. We start returning
996 * larger blocks from the end fo the chunk, since they should already be
997 * aligned to divisor. Also anv_state_pool_return_blocks() only accepts
1000 uint32_t offset
= chunk_offset
+ rest
;
1001 anv_state_pool_return_blocks(pool
, offset
, nblocks
, divisor
);
1007 if (small_size
> 0 && small_size
< divisor
)
1008 divisor
= small_size
;
1010 uint32_t min_size
= 1 << ANV_MIN_STATE_SIZE_LOG2
;
1012 /* Just as before, return larger divisor aligned blocks from the end of the
1015 while (chunk_size
> 0 && divisor
>= min_size
) {
1016 nblocks
= chunk_size
/ divisor
;
1017 rest
= chunk_size
- nblocks
* divisor
;
1019 anv_state_pool_return_blocks(pool
, chunk_offset
+ rest
,
1027 static struct anv_state
1028 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
1029 uint32_t size
, uint32_t align
)
1031 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
1033 struct anv_state
*state
;
1034 uint32_t alloc_size
= anv_state_pool_get_bucket_size(bucket
);
1037 /* Try free list first. */
1038 state
= anv_free_list_pop(&pool
->buckets
[bucket
].free_list
,
1041 assert(state
->offset
>= 0);
1045 /* Try to grab a chunk from some larger bucket and split it up */
1046 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
1047 state
= anv_free_list_pop(&pool
->buckets
[b
].free_list
, &pool
->table
);
1049 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
1050 int32_t chunk_offset
= state
->offset
;
1052 /* First lets update the state we got to its new size. offset and map
1055 state
->alloc_size
= alloc_size
;
1057 /* Now return the unused part of the chunk back to the pool as free
1060 * There are a couple of options as to what we do with it:
1062 * 1) We could fully split the chunk into state.alloc_size sized
1063 * pieces. However, this would mean that allocating a 16B
1064 * state could potentially split a 2MB chunk into 512K smaller
1065 * chunks. This would lead to unnecessary fragmentation.
1067 * 2) The classic "buddy allocator" method would have us split the
1068 * chunk in half and return one half. Then we would split the
1069 * remaining half in half and return one half, and repeat as
1070 * needed until we get down to the size we want. However, if
1071 * you are allocating a bunch of the same size state (which is
1072 * the common case), this means that every other allocation has
1073 * to go up a level and every fourth goes up two levels, etc.
1074 * This is not nearly as efficient as it could be if we did a
1075 * little more work up-front.
1077 * 3) Split the difference between (1) and (2) by doing a
1078 * two-level split. If it's bigger than some fixed block_size,
1079 * we split it into block_size sized chunks and return all but
1080 * one of them. Then we split what remains into
1081 * state.alloc_size sized chunks and return them.
1083 * We choose something close to option (3), which is implemented with
1084 * anv_state_pool_return_chunk(). That is done by returning the
1085 * remaining of the chunk, with alloc_size as a hint of the size that
1086 * we want the smaller chunk split into.
1088 anv_state_pool_return_chunk(pool
, chunk_offset
+ alloc_size
,
1089 chunk_size
- alloc_size
, alloc_size
);
1095 offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
1100 /* Everytime we allocate a new state, add it to the state pool */
1102 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1103 assert(result
== VK_SUCCESS
);
1105 state
= anv_state_table_get(&pool
->table
, idx
);
1106 state
->offset
= offset
;
1107 state
->alloc_size
= alloc_size
;
1108 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
);
1111 uint32_t return_offset
= offset
- padding
;
1112 anv_state_pool_return_chunk(pool
, return_offset
, padding
, 0);
1120 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
1123 return ANV_STATE_NULL
;
1125 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
1126 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
1131 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
1133 struct anv_state
*state
;
1134 uint32_t alloc_size
= pool
->block_size
;
1136 state
= anv_free_list_pop(&pool
->back_alloc_free_list
, &pool
->table
);
1138 assert(state
->offset
< 0);
1143 offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
1146 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1147 assert(result
== VK_SUCCESS
);
1149 state
= anv_state_table_get(&pool
->table
, idx
);
1150 state
->offset
= offset
;
1151 state
->alloc_size
= alloc_size
;
1152 state
->map
= anv_block_pool_map(&pool
->block_pool
, state
->offset
);
1155 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
->map
, state
->alloc_size
));
1160 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
1162 assert(util_is_power_of_two_or_zero(state
.alloc_size
));
1163 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
1165 if (state
.offset
< 0) {
1166 assert(state
.alloc_size
== pool
->block_size
);
1167 anv_free_list_push(&pool
->back_alloc_free_list
,
1168 &pool
->table
, state
.idx
, 1);
1170 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
1171 &pool
->table
, state
.idx
, 1);
1176 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
1178 if (state
.alloc_size
== 0)
1181 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
1182 anv_state_pool_free_no_vg(pool
, state
);
1185 struct anv_state_stream_block
{
1186 struct anv_state block
;
1188 /* The next block */
1189 struct anv_state_stream_block
*next
;
1191 #ifdef HAVE_VALGRIND
1192 /* A pointer to the first user-allocated thing in this block. This is
1193 * what valgrind sees as the start of the block.
1199 /* The state stream allocator is a one-shot, single threaded allocator for
1200 * variable sized blocks. We use it for allocating dynamic state.
1203 anv_state_stream_init(struct anv_state_stream
*stream
,
1204 struct anv_state_pool
*state_pool
,
1205 uint32_t block_size
)
1207 stream
->state_pool
= state_pool
;
1208 stream
->block_size
= block_size
;
1210 stream
->block
= ANV_STATE_NULL
;
1212 stream
->block_list
= NULL
;
1214 /* Ensure that next + whatever > block_size. This way the first call to
1215 * state_stream_alloc fetches a new block.
1217 stream
->next
= block_size
;
1219 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
1223 anv_state_stream_finish(struct anv_state_stream
*stream
)
1225 struct anv_state_stream_block
*next
= stream
->block_list
;
1226 while (next
!= NULL
) {
1227 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
1228 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
1229 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
1230 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
1234 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
1238 anv_state_stream_alloc(struct anv_state_stream
*stream
,
1239 uint32_t size
, uint32_t alignment
)
1242 return ANV_STATE_NULL
;
1244 assert(alignment
<= PAGE_SIZE
);
1246 uint32_t offset
= align_u32(stream
->next
, alignment
);
1247 if (offset
+ size
> stream
->block
.alloc_size
) {
1248 uint32_t block_size
= stream
->block_size
;
1249 if (block_size
< size
)
1250 block_size
= round_to_power_of_two(size
);
1252 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
1253 block_size
, PAGE_SIZE
);
1255 struct anv_state_stream_block
*sb
= stream
->block
.map
;
1256 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
1257 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
1258 stream
->block_list
= sb
;
1259 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
1261 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
1263 /* Reset back to the start plus space for the header */
1264 stream
->next
= sizeof(*sb
);
1266 offset
= align_u32(stream
->next
, alignment
);
1267 assert(offset
+ size
<= stream
->block
.alloc_size
);
1270 struct anv_state state
= stream
->block
;
1271 state
.offset
+= offset
;
1272 state
.alloc_size
= size
;
1273 state
.map
+= offset
;
1275 stream
->next
= offset
+ size
;
1277 #ifdef HAVE_VALGRIND
1278 struct anv_state_stream_block
*sb
= stream
->block_list
;
1279 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
1280 if (vg_ptr
== NULL
) {
1282 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
1283 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
1285 void *state_end
= state
.map
+ state
.alloc_size
;
1286 /* This only updates the mempool. The newly allocated chunk is still
1287 * marked as NOACCESS. */
1288 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
1289 /* Mark the newly allocated chunk as undefined */
1290 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
1297 struct bo_pool_bo_link
{
1298 struct bo_pool_bo_link
*next
;
1303 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
,
1306 pool
->device
= device
;
1307 pool
->bo_flags
= bo_flags
;
1308 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
1310 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
1314 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
1316 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1317 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
1318 while (link
!= NULL
) {
1319 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
1321 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
1322 anv_vma_free(pool
->device
, &link_copy
.bo
);
1323 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
1324 link
= link_copy
.next
;
1328 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1332 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
1336 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1337 const unsigned pow2_size
= 1 << size_log2
;
1338 const unsigned bucket
= size_log2
- 12;
1339 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1341 void *next_free_void
;
1342 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
1343 struct bo_pool_bo_link
*next_free
= next_free_void
;
1344 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
1345 assert(bo
->gem_handle
);
1346 assert(bo
->map
== next_free
);
1347 assert(size
<= bo
->size
);
1349 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1354 struct anv_bo new_bo
;
1356 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
1357 if (result
!= VK_SUCCESS
)
1360 new_bo
.flags
= pool
->bo_flags
;
1362 if (!anv_vma_alloc(pool
->device
, &new_bo
))
1363 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1365 assert(new_bo
.size
== pow2_size
);
1367 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
1368 if (new_bo
.map
== MAP_FAILED
) {
1369 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
1370 anv_vma_free(pool
->device
, &new_bo
);
1371 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1374 /* We are removing the state flushes, so lets make sure that these buffers
1375 * are cached/snooped.
1377 if (!pool
->device
->info
.has_llc
) {
1378 anv_gem_set_caching(pool
->device
, new_bo
.gem_handle
,
1379 I915_CACHING_CACHED
);
1384 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1390 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
1392 /* Make a copy in case the anv_bo happens to be storred in the BO */
1393 struct anv_bo bo
= *bo_in
;
1395 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
1397 struct bo_pool_bo_link
*link
= bo
.map
;
1398 VG_NOACCESS_WRITE(&link
->bo
, bo
);
1400 assert(util_is_power_of_two_or_zero(bo
.size
));
1401 const unsigned size_log2
= ilog2_round_up(bo
.size
);
1402 const unsigned bucket
= size_log2
- 12;
1403 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1405 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
1411 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1413 memset(pool
, 0, sizeof(*pool
));
1417 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1419 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1420 for (unsigned i
= 0; i
< 16; i
++) {
1421 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
1422 if (bo
->exists
> 0) {
1423 anv_vma_free(device
, &bo
->bo
);
1424 anv_gem_close(device
, bo
->bo
.gem_handle
);
1431 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1432 gl_shader_stage stage
, unsigned per_thread_scratch
)
1434 if (per_thread_scratch
== 0)
1437 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1438 assert(scratch_size_log2
< 16);
1440 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
1442 /* We can use "exists" to shortcut and ignore the critical section */
1446 pthread_mutex_lock(&device
->mutex
);
1448 __sync_synchronize();
1450 pthread_mutex_unlock(&device
->mutex
);
1454 const struct anv_physical_device
*physical_device
=
1455 &device
->instance
->physicalDevice
;
1456 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1458 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
1460 unsigned scratch_ids_per_subslice
;
1461 if (devinfo
->gen
>= 11) {
1462 /* The MEDIA_VFE_STATE docs say:
1464 * "Starting with this configuration, the Maximum Number of
1465 * Threads must be set to (#EU * 8) for GPGPU dispatches.
1467 * Although there are only 7 threads per EU in the configuration,
1468 * the FFTID is calculated as if there are 8 threads per EU,
1469 * which in turn requires a larger amount of Scratch Space to be
1470 * allocated by the driver."
1472 scratch_ids_per_subslice
= 8 * 8;
1473 } else if (devinfo
->is_haswell
) {
1474 /* WaCSScratchSize:hsw
1476 * Haswell's scratch space address calculation appears to be sparse
1477 * rather than tightly packed. The Thread ID has bits indicating
1478 * which subslice, EU within a subslice, and thread within an EU it
1479 * is. There's a maximum of two slices and two subslices, so these
1480 * can be stored with a single bit. Even though there are only 10 EUs
1481 * per subslice, this is stored in 4 bits, so there's an effective
1482 * maximum value of 16 EUs. Similarly, although there are only 7
1483 * threads per EU, this is stored in a 3 bit number, giving an
1484 * effective maximum value of 8 threads per EU.
1486 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1487 * number of threads per subslice.
1489 scratch_ids_per_subslice
= 16 * 8;
1490 } else if (devinfo
->is_cherryview
) {
1491 /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
1492 * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
1495 scratch_ids_per_subslice
= 8 * 7;
1497 scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
1500 uint32_t max_threads
[] = {
1501 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1502 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1503 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1504 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1505 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1506 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1509 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1511 anv_bo_init_new(&bo
->bo
, device
, size
);
1513 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1514 * are still relative to the general state base address. When we emit
1515 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1516 * to the maximum (1 page under 4GB). This allows us to just place the
1517 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1518 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1519 * However, in order to do so, we need to ensure that the kernel does not
1520 * place the scratch BO above the 32-bit boundary.
1522 * NOTE: Technically, it can't go "anywhere" because the top page is off
1523 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1524 * kernel allocates space using
1526 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1528 * so nothing will ever touch the top page.
1530 assert(!(bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
));
1532 if (device
->instance
->physicalDevice
.has_exec_async
)
1533 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1535 if (device
->instance
->physicalDevice
.use_softpin
)
1536 bo
->bo
.flags
|= EXEC_OBJECT_PINNED
;
1538 anv_vma_alloc(device
, &bo
->bo
);
1540 /* Set the exists last because it may be read by other threads */
1541 __sync_synchronize();
1544 pthread_mutex_unlock(&device
->mutex
);
1550 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1552 util_sparse_array_init(&cache
->bo_map
, sizeof(struct anv_bo
), 1024);
1554 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1555 util_sparse_array_finish(&cache
->bo_map
);
1556 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1557 "pthread_mutex_init failed: %m");
1564 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1566 util_sparse_array_finish(&cache
->bo_map
);
1567 pthread_mutex_destroy(&cache
->mutex
);
1570 static struct anv_bo
*
1571 anv_bo_cache_lookup(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1573 return util_sparse_array_get(&cache
->bo_map
, gem_handle
);
1576 #define ANV_BO_CACHE_SUPPORTED_FLAGS \
1577 (EXEC_OBJECT_WRITE | \
1578 EXEC_OBJECT_ASYNC | \
1579 EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
1583 anv_bo_cache_alloc(struct anv_device
*device
,
1584 struct anv_bo_cache
*cache
,
1585 uint64_t size
, uint64_t bo_flags
,
1587 struct anv_bo
**bo_out
)
1589 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1591 /* The kernel is going to give us whole pages anyway */
1592 size
= align_u64(size
, 4096);
1594 struct anv_bo new_bo
;
1595 VkResult result
= anv_bo_init_new(&new_bo
, device
, size
);
1596 if (result
!= VK_SUCCESS
)
1599 new_bo
.flags
= bo_flags
;
1600 new_bo
.is_external
= is_external
;
1602 if (!anv_vma_alloc(device
, &new_bo
)) {
1603 anv_gem_close(device
, new_bo
.gem_handle
);
1604 return vk_errorf(device
->instance
, NULL
,
1605 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1606 "failed to allocate virtual address for BO");
1609 assert(new_bo
.gem_handle
);
1611 /* If we just got this gem_handle from anv_bo_init_new then we know no one
1612 * else is touching this BO at the moment so we don't need to lock here.
1614 struct anv_bo
*bo
= anv_bo_cache_lookup(cache
, new_bo
.gem_handle
);
1623 anv_bo_cache_import_host_ptr(struct anv_device
*device
,
1624 struct anv_bo_cache
*cache
,
1625 void *host_ptr
, uint32_t size
,
1626 uint64_t bo_flags
, struct anv_bo
**bo_out
)
1628 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1630 uint32_t gem_handle
= anv_gem_userptr(device
, host_ptr
, size
);
1632 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1634 pthread_mutex_lock(&cache
->mutex
);
1636 struct anv_bo
*bo
= anv_bo_cache_lookup(cache
, gem_handle
);
1637 if (bo
->refcount
> 0) {
1638 /* VK_EXT_external_memory_host doesn't require handling importing the
1639 * same pointer twice at the same time, but we don't get in the way. If
1640 * kernel gives us the same gem_handle, only succeed if the flags match.
1642 assert(bo
->gem_handle
== gem_handle
);
1643 if (bo_flags
!= bo
->flags
) {
1644 pthread_mutex_unlock(&cache
->mutex
);
1645 return vk_errorf(device
->instance
, NULL
,
1646 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1647 "same host pointer imported two different ways");
1649 __sync_fetch_and_add(&bo
->refcount
, 1);
1651 struct anv_bo new_bo
;
1652 anv_bo_init(&new_bo
, gem_handle
, size
);
1653 new_bo
.flags
= bo_flags
;
1654 new_bo
.is_external
= true;
1656 if (!anv_vma_alloc(device
, &new_bo
)) {
1657 anv_gem_close(device
, new_bo
.gem_handle
);
1658 pthread_mutex_unlock(&cache
->mutex
);
1659 return vk_errorf(device
->instance
, NULL
,
1660 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1661 "failed to allocate virtual address for BO");
1667 pthread_mutex_unlock(&cache
->mutex
);
1674 anv_bo_cache_import(struct anv_device
*device
,
1675 struct anv_bo_cache
*cache
,
1676 int fd
, uint64_t bo_flags
,
1677 struct anv_bo
**bo_out
)
1679 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1681 pthread_mutex_lock(&cache
->mutex
);
1683 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1685 pthread_mutex_unlock(&cache
->mutex
);
1686 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1689 struct anv_bo
*bo
= anv_bo_cache_lookup(cache
, gem_handle
);
1690 if (bo
->refcount
> 0) {
1691 /* We have to be careful how we combine flags so that it makes sense.
1692 * Really, though, if we get to this case and it actually matters, the
1693 * client has imported a BO twice in different ways and they get what
1696 uint64_t new_flags
= 0;
1697 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_WRITE
;
1698 new_flags
|= (bo
->flags
& bo_flags
) & EXEC_OBJECT_ASYNC
;
1699 new_flags
|= (bo
->flags
& bo_flags
) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1700 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_PINNED
;
1702 /* It's theoretically possible for a BO to get imported such that it's
1703 * both pinned and not pinned. The only way this can happen is if it
1704 * gets imported as both a semaphore and a memory object and that would
1705 * be an application error. Just fail out in that case.
1707 if ((bo
->flags
& EXEC_OBJECT_PINNED
) !=
1708 (bo_flags
& EXEC_OBJECT_PINNED
)) {
1709 pthread_mutex_unlock(&cache
->mutex
);
1710 return vk_errorf(device
->instance
, NULL
,
1711 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1712 "The same BO was imported two different ways");
1715 /* It's also theoretically possible that someone could export a BO from
1716 * one heap and import it into another or to import the same BO into two
1717 * different heaps. If this happens, we could potentially end up both
1718 * allowing and disallowing 48-bit addresses. There's not much we can
1719 * do about it if we're pinning so we just throw an error and hope no
1720 * app is actually that stupid.
1722 if ((new_flags
& EXEC_OBJECT_PINNED
) &&
1723 (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) !=
1724 (bo_flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
)) {
1725 pthread_mutex_unlock(&cache
->mutex
);
1726 return vk_errorf(device
->instance
, NULL
,
1727 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1728 "The same BO was imported on two different heaps");
1731 bo
->flags
= new_flags
;
1733 __sync_fetch_and_add(&bo
->refcount
, 1);
1735 off_t size
= lseek(fd
, 0, SEEK_END
);
1736 if (size
== (off_t
)-1) {
1737 anv_gem_close(device
, gem_handle
);
1738 pthread_mutex_unlock(&cache
->mutex
);
1739 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1742 struct anv_bo new_bo
;
1743 anv_bo_init(&new_bo
, gem_handle
, size
);
1744 new_bo
.flags
= bo_flags
;
1745 new_bo
.is_external
= true;
1747 if (!anv_vma_alloc(device
, &new_bo
)) {
1748 anv_gem_close(device
, new_bo
.gem_handle
);
1749 pthread_mutex_unlock(&cache
->mutex
);
1750 return vk_errorf(device
->instance
, NULL
,
1751 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1752 "failed to allocate virtual address for BO");
1758 pthread_mutex_unlock(&cache
->mutex
);
1765 anv_bo_cache_export(struct anv_device
*device
,
1766 struct anv_bo_cache
*cache
,
1767 struct anv_bo
*bo
, int *fd_out
)
1769 assert(anv_bo_cache_lookup(cache
, bo
->gem_handle
) == bo
);
1771 /* This BO must have been flagged external in order for us to be able
1772 * to export it. This is done based on external options passed into
1773 * anv_AllocateMemory.
1775 assert(bo
->is_external
);
1777 int fd
= anv_gem_handle_to_fd(device
, bo
->gem_handle
);
1779 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1787 atomic_dec_not_one(uint32_t *counter
)
1796 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1805 anv_bo_cache_release(struct anv_device
*device
,
1806 struct anv_bo_cache
*cache
,
1809 assert(anv_bo_cache_lookup(cache
, bo
->gem_handle
) == bo
);
1811 /* Try to decrement the counter but don't go below one. If this succeeds
1812 * then the refcount has been decremented and we are not the last
1815 if (atomic_dec_not_one(&bo
->refcount
))
1818 pthread_mutex_lock(&cache
->mutex
);
1820 /* We are probably the last reference since our attempt to decrement above
1821 * failed. However, we can't actually know until we are inside the mutex.
1822 * Otherwise, someone could import the BO between the decrement and our
1825 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1826 /* Turns out we're not the last reference. Unlock and bail. */
1827 pthread_mutex_unlock(&cache
->mutex
);
1830 assert(bo
->refcount
== 0);
1833 anv_gem_munmap(bo
->map
, bo
->size
);
1835 anv_vma_free(device
, bo
);
1837 anv_gem_close(device
, bo
->gem_handle
);
1839 /* Don't unlock until we've actually closed the BO. The whole point of
1840 * the BO cache is to ensure that we correctly handle races with creating
1841 * and releasing GEM handles and we don't want to let someone import the BO
1842 * again between mutex unlock and closing the GEM handle.
1844 pthread_mutex_unlock(&cache
->mutex
);