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 "common/gen_aux_map.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
, 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
;
361 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
362 uint32_t center_bo_offset
, uint32_t size
);
365 anv_block_pool_init(struct anv_block_pool
*pool
,
366 struct anv_device
*device
,
367 uint64_t start_address
,
368 uint32_t initial_size
)
372 pool
->device
= device
;
373 pool
->use_softpin
= device
->physical
->use_softpin
;
376 pool
->center_bo_offset
= 0;
377 pool
->start_address
= gen_canonical_address(start_address
);
380 if (pool
->use_softpin
) {
384 /* Just make it 2GB up-front. The Linux kernel won't actually back it
385 * with pages until we either map and fault on one of them or we use
386 * userptr and send a chunk of it off to the GPU.
388 pool
->fd
= os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE
, "block pool");
390 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
392 pool
->wrapper_bo
= (struct anv_bo
) {
397 pool
->bo
= &pool
->wrapper_bo
;
400 if (!u_vector_init(&pool
->mmap_cleanups
,
401 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
403 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
407 pool
->state
.next
= 0;
409 pool
->back_state
.next
= 0;
410 pool
->back_state
.end
= 0;
412 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
413 if (result
!= VK_SUCCESS
)
414 goto fail_mmap_cleanups
;
416 /* Make the entire pool available in the front of the pool. If back
417 * allocation needs to use this space, the "ends" will be re-arranged.
419 pool
->state
.end
= pool
->size
;
424 u_vector_finish(&pool
->mmap_cleanups
);
433 anv_block_pool_finish(struct anv_block_pool
*pool
)
435 anv_block_pool_foreach_bo(bo
, pool
) {
437 anv_gem_munmap(bo
->map
, bo
->size
);
438 anv_gem_close(pool
->device
, bo
->gem_handle
);
441 struct anv_mmap_cleanup
*cleanup
;
442 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
)
443 munmap(cleanup
->map
, cleanup
->size
);
444 u_vector_finish(&pool
->mmap_cleanups
);
451 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
452 uint32_t center_bo_offset
, uint32_t size
)
454 /* Assert that we only ever grow the pool */
455 assert(center_bo_offset
>= pool
->back_state
.end
);
456 assert(size
- center_bo_offset
>= pool
->state
.end
);
458 /* Assert that we don't go outside the bounds of the memfd */
459 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
460 assert(pool
->use_softpin
||
461 size
- center_bo_offset
<=
462 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
464 /* For state pool BOs we have to be a bit careful about where we place them
465 * in the GTT. There are two documented workarounds for state base address
466 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
467 * which state that those two base addresses do not support 48-bit
468 * addresses and need to be placed in the bottom 32-bit range.
469 * Unfortunately, this is not quite accurate.
471 * The real problem is that we always set the size of our state pools in
472 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
473 * likely significantly smaller. We do this because we do not no at the
474 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
475 * the pool during command buffer building so we don't actually have a
476 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
477 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
478 * as being out of bounds and returns zero. For dynamic state, this
479 * usually just leads to rendering corruptions, but shaders that are all
480 * zero hang the GPU immediately.
482 * The easiest solution to do is exactly what the bogus workarounds say to
483 * do: restrict these buffers to 32-bit addresses. We could also pin the
484 * BO to some particular location of our choosing, but that's significantly
485 * more work than just not setting a flag. So, we explicitly DO NOT set
486 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
487 * hard work for us. When using softpin, we're in control and the fixed
488 * addresses we choose are fine for base addresses.
490 enum anv_bo_alloc_flags bo_alloc_flags
= ANV_BO_ALLOC_CAPTURE
;
491 if (!pool
->use_softpin
)
492 bo_alloc_flags
|= ANV_BO_ALLOC_32BIT_ADDRESS
;
494 if (pool
->use_softpin
) {
495 uint32_t new_bo_size
= size
- pool
->size
;
496 struct anv_bo
*new_bo
;
497 assert(center_bo_offset
== 0);
498 VkResult result
= anv_device_alloc_bo(pool
->device
, new_bo_size
,
500 ANV_BO_ALLOC_FIXED_ADDRESS
|
501 ANV_BO_ALLOC_MAPPED
|
502 ANV_BO_ALLOC_SNOOPED
,
503 pool
->start_address
+ pool
->size
,
505 if (result
!= VK_SUCCESS
)
508 pool
->bos
[pool
->nbos
++] = new_bo
;
510 /* This pointer will always point to the first BO in the list */
511 pool
->bo
= pool
->bos
[0];
513 /* Just leak the old map until we destroy the pool. We can't munmap it
514 * without races or imposing locking on the block allocate fast path. On
515 * the whole the leaked maps adds up to less than the size of the
516 * current map. MAP_POPULATE seems like the right thing to do, but we
517 * should try to get some numbers.
519 void *map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
520 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
521 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
522 if (map
== MAP_FAILED
)
523 return vk_errorf(pool
->device
, pool
->device
,
524 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
526 struct anv_bo
*new_bo
;
527 VkResult result
= anv_device_import_bo_from_host_ptr(pool
->device
,
530 0 /* client_address */,
532 if (result
!= VK_SUCCESS
) {
537 struct anv_mmap_cleanup
*cleanup
= u_vector_add(&pool
->mmap_cleanups
);
540 anv_device_release_bo(pool
->device
, new_bo
);
541 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
544 cleanup
->size
= size
;
546 /* Now that we mapped the new memory, we can write the new
547 * center_bo_offset back into pool and update pool->map. */
548 pool
->center_bo_offset
= center_bo_offset
;
549 pool
->map
= map
+ center_bo_offset
;
551 pool
->bos
[pool
->nbos
++] = new_bo
;
552 pool
->wrapper_bo
.map
= new_bo
;
555 assert(pool
->nbos
< ANV_MAX_BLOCK_POOL_BOS
);
561 /** Returns current memory map of the block pool.
563 * The returned pointer points to the map for the memory at the specified
564 * offset. The offset parameter is relative to the "center" of the block pool
565 * rather than the start of the block pool BO map.
568 anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
, uint32_t size
)
570 if (pool
->use_softpin
) {
571 struct anv_bo
*bo
= NULL
;
572 int32_t bo_offset
= 0;
573 anv_block_pool_foreach_bo(iter_bo
, pool
) {
574 if (offset
< bo_offset
+ iter_bo
->size
) {
578 bo_offset
+= iter_bo
->size
;
581 assert(offset
>= bo_offset
);
582 assert((offset
- bo_offset
) + size
<= bo
->size
);
584 return bo
->map
+ (offset
- bo_offset
);
586 return pool
->map
+ offset
;
590 /** Grows and re-centers the block pool.
592 * We grow the block pool in one or both directions in such a way that the
593 * following conditions are met:
595 * 1) The size of the entire pool is always a power of two.
597 * 2) The pool only grows on both ends. Neither end can get
600 * 3) At the end of the allocation, we have about twice as much space
601 * allocated for each end as we have used. This way the pool doesn't
602 * grow too far in one direction or the other.
604 * 4) If the _alloc_back() has never been called, then the back portion of
605 * the pool retains a size of zero. (This makes it easier for users of
606 * the block pool that only want a one-sided pool.)
608 * 5) We have enough space allocated for at least one more block in
609 * whichever side `state` points to.
611 * 6) The center of the pool is always aligned to both the block_size of
612 * the pool and a 4K CPU page.
615 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
,
616 uint32_t contiguous_size
)
618 VkResult result
= VK_SUCCESS
;
620 pthread_mutex_lock(&pool
->device
->mutex
);
622 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
624 /* Gather a little usage information on the pool. Since we may have
625 * threadsd waiting in queue to get some storage while we resize, it's
626 * actually possible that total_used will be larger than old_size. In
627 * particular, block_pool_alloc() increments state->next prior to
628 * calling block_pool_grow, so this ensures that we get enough space for
629 * which ever side tries to grow the pool.
631 * We align to a page size because it makes it easier to do our
632 * calculations later in such a way that we state page-aigned.
634 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
635 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
636 uint32_t total_used
= front_used
+ back_used
;
638 assert(state
== &pool
->state
|| back_used
> 0);
640 uint32_t old_size
= pool
->size
;
642 /* The block pool is always initialized to a nonzero size and this function
643 * is always called after initialization.
645 assert(old_size
> 0);
647 const uint32_t old_back
= pool
->center_bo_offset
;
648 const uint32_t old_front
= old_size
- pool
->center_bo_offset
;
650 /* The back_used and front_used may actually be smaller than the actual
651 * requirement because they are based on the next pointers which are
652 * updated prior to calling this function.
654 uint32_t back_required
= MAX2(back_used
, old_back
);
655 uint32_t front_required
= MAX2(front_used
, old_front
);
657 if (pool
->use_softpin
) {
658 /* With softpin, the pool is made up of a bunch of buffers with separate
659 * maps. Make sure we have enough contiguous space that we can get a
660 * properly contiguous map for the next chunk.
662 assert(old_back
== 0);
663 front_required
= MAX2(front_required
, old_front
+ contiguous_size
);
666 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
667 /* If we're in this case then this isn't the firsta allocation and we
668 * already have enough space on both sides to hold double what we
669 * have allocated. There's nothing for us to do.
674 uint32_t size
= old_size
* 2;
675 while (size
< back_required
+ front_required
)
678 assert(size
> pool
->size
);
680 /* We compute a new center_bo_offset such that, when we double the size
681 * of the pool, we maintain the ratio of how much is used by each side.
682 * This way things should remain more-or-less balanced.
684 uint32_t center_bo_offset
;
685 if (back_used
== 0) {
686 /* If we're in this case then we have never called alloc_back(). In
687 * this case, we want keep the offset at 0 to make things as simple
688 * as possible for users that don't care about back allocations.
690 center_bo_offset
= 0;
692 /* Try to "center" the allocation based on how much is currently in
693 * use on each side of the center line.
695 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
697 /* Align down to a multiple of the page size */
698 center_bo_offset
&= ~(PAGE_SIZE
- 1);
700 assert(center_bo_offset
>= back_used
);
702 /* Make sure we don't shrink the back end of the pool */
703 if (center_bo_offset
< back_required
)
704 center_bo_offset
= back_required
;
706 /* Make sure that we don't shrink the front end of the pool */
707 if (size
- center_bo_offset
< front_required
)
708 center_bo_offset
= size
- front_required
;
711 assert(center_bo_offset
% PAGE_SIZE
== 0);
713 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
716 pthread_mutex_unlock(&pool
->device
->mutex
);
718 if (result
== VK_SUCCESS
) {
719 /* Return the appropriate new size. This function never actually
720 * updates state->next. Instead, we let the caller do that because it
721 * needs to do so in order to maintain its concurrency model.
723 if (state
== &pool
->state
) {
724 return pool
->size
- pool
->center_bo_offset
;
726 assert(pool
->center_bo_offset
> 0);
727 return pool
->center_bo_offset
;
735 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
736 struct anv_block_state
*pool_state
,
737 uint32_t block_size
, uint32_t *padding
)
739 struct anv_block_state state
, old
, new;
741 /* Most allocations won't generate any padding */
746 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
747 if (state
.next
+ block_size
<= state
.end
) {
749 } else if (state
.next
<= state
.end
) {
750 if (pool
->use_softpin
&& state
.next
< state
.end
) {
751 /* We need to grow the block pool, but still have some leftover
752 * space that can't be used by that particular allocation. So we
753 * add that as a "padding", and return it.
755 uint32_t leftover
= state
.end
- state
.next
;
757 /* If there is some leftover space in the pool, the caller must
760 assert(leftover
== 0 || padding
);
763 state
.next
+= leftover
;
766 /* We allocated the first block outside the pool so we have to grow
767 * the pool. pool_state->next acts a mutex: threads who try to
768 * allocate now will get block indexes above the current limit and
769 * hit futex_wait below.
771 new.next
= state
.next
+ block_size
;
773 new.end
= anv_block_pool_grow(pool
, pool_state
, block_size
);
774 } while (new.end
< new.next
);
776 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
777 if (old
.next
!= state
.next
)
778 futex_wake(&pool_state
->end
, INT_MAX
);
781 futex_wait(&pool_state
->end
, state
.end
, NULL
);
788 anv_block_pool_alloc(struct anv_block_pool
*pool
,
789 uint32_t block_size
, uint32_t *padding
)
793 offset
= anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
, padding
);
798 /* Allocates a block out of the back of the block pool.
800 * This will allocated a block earlier than the "start" of the block pool.
801 * The offsets returned from this function will be negative but will still
802 * be correct relative to the block pool's map pointer.
804 * If you ever use anv_block_pool_alloc_back, then you will have to do
805 * gymnastics with the block pool's BO when doing relocations.
808 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
811 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
814 /* The offset we get out of anv_block_pool_alloc_new() is actually the
815 * number of bytes downwards from the middle to the end of the block.
816 * We need to turn it into a (negative) offset from the middle to the
817 * start of the block.
820 return -(offset
+ block_size
);
824 anv_state_pool_init(struct anv_state_pool
*pool
,
825 struct anv_device
*device
,
826 uint64_t start_address
,
829 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
832 if (result
!= VK_SUCCESS
)
835 result
= anv_state_table_init(&pool
->table
, device
, 64);
836 if (result
!= VK_SUCCESS
) {
837 anv_block_pool_finish(&pool
->block_pool
);
841 assert(util_is_power_of_two_or_zero(block_size
));
842 pool
->block_size
= block_size
;
843 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
844 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
845 pool
->buckets
[i
].free_list
= ANV_FREE_LIST_EMPTY
;
846 pool
->buckets
[i
].block
.next
= 0;
847 pool
->buckets
[i
].block
.end
= 0;
849 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
855 anv_state_pool_finish(struct anv_state_pool
*pool
)
857 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
858 anv_state_table_finish(&pool
->table
);
859 anv_block_pool_finish(&pool
->block_pool
);
863 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
864 struct anv_block_pool
*block_pool
,
869 struct anv_block_state block
, old
, new;
872 /* We don't always use anv_block_pool_alloc(), which would set *padding to
873 * zero for us. So if we have a pointer to padding, we must zero it out
874 * ourselves here, to make sure we always return some sensible value.
879 /* If our state is large, we don't need any sub-allocation from a block.
880 * Instead, we just grab whole (potentially large) blocks.
882 if (state_size
>= block_size
)
883 return anv_block_pool_alloc(block_pool
, state_size
, padding
);
886 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
888 if (block
.next
< block
.end
) {
890 } else if (block
.next
== block
.end
) {
891 offset
= anv_block_pool_alloc(block_pool
, block_size
, padding
);
892 new.next
= offset
+ state_size
;
893 new.end
= offset
+ block_size
;
894 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
895 if (old
.next
!= block
.next
)
896 futex_wake(&pool
->block
.end
, INT_MAX
);
899 futex_wait(&pool
->block
.end
, block
.end
, NULL
);
905 anv_state_pool_get_bucket(uint32_t size
)
907 unsigned size_log2
= ilog2_round_up(size
);
908 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
909 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
910 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
911 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
915 anv_state_pool_get_bucket_size(uint32_t bucket
)
917 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
918 return 1 << size_log2
;
921 /** Helper to push a chunk into the state table.
923 * It creates 'count' entries into the state table and update their sizes,
924 * offsets and maps, also pushing them as "free" states.
927 anv_state_pool_return_blocks(struct anv_state_pool
*pool
,
928 uint32_t chunk_offset
, uint32_t count
,
931 /* Disallow returning 0 chunks */
934 /* Make sure we always return chunks aligned to the block_size */
935 assert(chunk_offset
% block_size
== 0);
938 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &st_idx
, count
);
939 assert(result
== VK_SUCCESS
);
940 for (int i
= 0; i
< count
; i
++) {
941 /* update states that were added back to the state table */
942 struct anv_state
*state_i
= anv_state_table_get(&pool
->table
,
944 state_i
->alloc_size
= block_size
;
945 state_i
->offset
= chunk_offset
+ block_size
* i
;
946 state_i
->map
= anv_block_pool_map(&pool
->block_pool
,
948 state_i
->alloc_size
);
951 uint32_t block_bucket
= anv_state_pool_get_bucket(block_size
);
952 anv_free_list_push(&pool
->buckets
[block_bucket
].free_list
,
953 &pool
->table
, st_idx
, count
);
956 /** Returns a chunk of memory back to the state pool.
958 * Do a two-level split. If chunk_size is bigger than divisor
959 * (pool->block_size), we return as many divisor sized blocks as we can, from
960 * the end of the chunk.
962 * The remaining is then split into smaller blocks (starting at small_size if
963 * it is non-zero), with larger blocks always being taken from the end of the
967 anv_state_pool_return_chunk(struct anv_state_pool
*pool
,
968 uint32_t chunk_offset
, uint32_t chunk_size
,
971 uint32_t divisor
= pool
->block_size
;
972 uint32_t nblocks
= chunk_size
/ divisor
;
973 uint32_t rest
= chunk_size
- nblocks
* divisor
;
976 /* First return divisor aligned and sized chunks. We start returning
977 * larger blocks from the end fo the chunk, since they should already be
978 * aligned to divisor. Also anv_state_pool_return_blocks() only accepts
981 uint32_t offset
= chunk_offset
+ rest
;
982 anv_state_pool_return_blocks(pool
, offset
, nblocks
, divisor
);
988 if (small_size
> 0 && small_size
< divisor
)
989 divisor
= small_size
;
991 uint32_t min_size
= 1 << ANV_MIN_STATE_SIZE_LOG2
;
993 /* Just as before, return larger divisor aligned blocks from the end of the
996 while (chunk_size
> 0 && divisor
>= min_size
) {
997 nblocks
= chunk_size
/ divisor
;
998 rest
= chunk_size
- nblocks
* divisor
;
1000 anv_state_pool_return_blocks(pool
, chunk_offset
+ rest
,
1008 static struct anv_state
1009 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
1010 uint32_t size
, uint32_t align
)
1012 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
1014 struct anv_state
*state
;
1015 uint32_t alloc_size
= anv_state_pool_get_bucket_size(bucket
);
1018 /* Try free list first. */
1019 state
= anv_free_list_pop(&pool
->buckets
[bucket
].free_list
,
1022 assert(state
->offset
>= 0);
1026 /* Try to grab a chunk from some larger bucket and split it up */
1027 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
1028 state
= anv_free_list_pop(&pool
->buckets
[b
].free_list
, &pool
->table
);
1030 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
1031 int32_t chunk_offset
= state
->offset
;
1033 /* First lets update the state we got to its new size. offset and map
1036 state
->alloc_size
= alloc_size
;
1038 /* Now return the unused part of the chunk back to the pool as free
1041 * There are a couple of options as to what we do with it:
1043 * 1) We could fully split the chunk into state.alloc_size sized
1044 * pieces. However, this would mean that allocating a 16B
1045 * state could potentially split a 2MB chunk into 512K smaller
1046 * chunks. This would lead to unnecessary fragmentation.
1048 * 2) The classic "buddy allocator" method would have us split the
1049 * chunk in half and return one half. Then we would split the
1050 * remaining half in half and return one half, and repeat as
1051 * needed until we get down to the size we want. However, if
1052 * you are allocating a bunch of the same size state (which is
1053 * the common case), this means that every other allocation has
1054 * to go up a level and every fourth goes up two levels, etc.
1055 * This is not nearly as efficient as it could be if we did a
1056 * little more work up-front.
1058 * 3) Split the difference between (1) and (2) by doing a
1059 * two-level split. If it's bigger than some fixed block_size,
1060 * we split it into block_size sized chunks and return all but
1061 * one of them. Then we split what remains into
1062 * state.alloc_size sized chunks and return them.
1064 * We choose something close to option (3), which is implemented with
1065 * anv_state_pool_return_chunk(). That is done by returning the
1066 * remaining of the chunk, with alloc_size as a hint of the size that
1067 * we want the smaller chunk split into.
1069 anv_state_pool_return_chunk(pool
, chunk_offset
+ alloc_size
,
1070 chunk_size
- alloc_size
, alloc_size
);
1076 offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
1081 /* Everytime we allocate a new state, add it to the state pool */
1083 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1084 assert(result
== VK_SUCCESS
);
1086 state
= anv_state_table_get(&pool
->table
, idx
);
1087 state
->offset
= offset
;
1088 state
->alloc_size
= alloc_size
;
1089 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
, alloc_size
);
1092 uint32_t return_offset
= offset
- padding
;
1093 anv_state_pool_return_chunk(pool
, return_offset
, padding
, 0);
1101 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
1104 return ANV_STATE_NULL
;
1106 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
1107 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
1112 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
1114 struct anv_state
*state
;
1115 uint32_t alloc_size
= pool
->block_size
;
1117 state
= anv_free_list_pop(&pool
->back_alloc_free_list
, &pool
->table
);
1119 assert(state
->offset
< 0);
1124 offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
1127 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1128 assert(result
== VK_SUCCESS
);
1130 state
= anv_state_table_get(&pool
->table
, idx
);
1131 state
->offset
= offset
;
1132 state
->alloc_size
= alloc_size
;
1133 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
, alloc_size
);
1136 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
->map
, state
->alloc_size
));
1141 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
1143 assert(util_is_power_of_two_or_zero(state
.alloc_size
));
1144 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
1146 if (state
.offset
< 0) {
1147 assert(state
.alloc_size
== pool
->block_size
);
1148 anv_free_list_push(&pool
->back_alloc_free_list
,
1149 &pool
->table
, state
.idx
, 1);
1151 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
1152 &pool
->table
, state
.idx
, 1);
1157 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
1159 if (state
.alloc_size
== 0)
1162 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
1163 anv_state_pool_free_no_vg(pool
, state
);
1166 struct anv_state_stream_block
{
1167 struct anv_state block
;
1169 /* The next block */
1170 struct anv_state_stream_block
*next
;
1172 #ifdef HAVE_VALGRIND
1173 /* A pointer to the first user-allocated thing in this block. This is
1174 * what valgrind sees as the start of the block.
1180 /* The state stream allocator is a one-shot, single threaded allocator for
1181 * variable sized blocks. We use it for allocating dynamic state.
1184 anv_state_stream_init(struct anv_state_stream
*stream
,
1185 struct anv_state_pool
*state_pool
,
1186 uint32_t block_size
)
1188 stream
->state_pool
= state_pool
;
1189 stream
->block_size
= block_size
;
1191 stream
->block
= ANV_STATE_NULL
;
1193 stream
->block_list
= NULL
;
1195 /* Ensure that next + whatever > block_size. This way the first call to
1196 * state_stream_alloc fetches a new block.
1198 stream
->next
= block_size
;
1200 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
1204 anv_state_stream_finish(struct anv_state_stream
*stream
)
1206 struct anv_state_stream_block
*next
= stream
->block_list
;
1207 while (next
!= NULL
) {
1208 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
1209 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
1210 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
1211 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
1215 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
1219 anv_state_stream_alloc(struct anv_state_stream
*stream
,
1220 uint32_t size
, uint32_t alignment
)
1223 return ANV_STATE_NULL
;
1225 assert(alignment
<= PAGE_SIZE
);
1227 uint32_t offset
= align_u32(stream
->next
, alignment
);
1228 if (offset
+ size
> stream
->block
.alloc_size
) {
1229 uint32_t min_block_size
= size
+ sizeof(struct anv_state_stream_block
);
1230 uint32_t block_size
= stream
->block_size
;
1231 if (block_size
< min_block_size
)
1232 block_size
= round_to_power_of_two(min_block_size
);
1234 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
1235 block_size
, PAGE_SIZE
);
1237 struct anv_state_stream_block
*sb
= stream
->block
.map
;
1238 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
1239 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
1240 stream
->block_list
= sb
;
1241 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
1243 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
1245 /* Reset back to the start plus space for the header */
1246 stream
->next
= sizeof(*sb
);
1248 offset
= align_u32(stream
->next
, alignment
);
1249 assert(offset
+ size
<= stream
->block
.alloc_size
);
1252 struct anv_state state
= stream
->block
;
1253 state
.offset
+= offset
;
1254 state
.alloc_size
= size
;
1255 state
.map
+= offset
;
1257 stream
->next
= offset
+ size
;
1259 #ifdef HAVE_VALGRIND
1260 struct anv_state_stream_block
*sb
= stream
->block_list
;
1261 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
1262 if (vg_ptr
== NULL
) {
1264 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
1265 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
1267 void *state_end
= state
.map
+ state
.alloc_size
;
1268 /* This only updates the mempool. The newly allocated chunk is still
1269 * marked as NOACCESS. */
1270 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
1271 /* Mark the newly allocated chunk as undefined */
1272 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
1280 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
)
1282 pool
->device
= device
;
1283 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1284 util_sparse_array_free_list_init(&pool
->free_list
[i
],
1285 &device
->bo_cache
.bo_map
, 0,
1286 offsetof(struct anv_bo
, free_index
));
1289 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
1293 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
1295 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1298 util_sparse_array_free_list_pop_elem(&pool
->free_list
[i
]);
1302 /* anv_device_release_bo is going to "free" it */
1303 VG(VALGRIND_MALLOCLIKE_BLOCK(bo
->map
, bo
->size
, 0, 1));
1304 anv_device_release_bo(pool
->device
, bo
);
1308 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1312 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, uint32_t size
,
1313 struct anv_bo
**bo_out
)
1315 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1316 const unsigned pow2_size
= 1 << size_log2
;
1317 const unsigned bucket
= size_log2
- 12;
1318 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1321 util_sparse_array_free_list_pop_elem(&pool
->free_list
[bucket
]);
1323 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1328 VkResult result
= anv_device_alloc_bo(pool
->device
,
1330 ANV_BO_ALLOC_MAPPED
|
1331 ANV_BO_ALLOC_SNOOPED
|
1332 ANV_BO_ALLOC_CAPTURE
,
1333 0 /* explicit_address */,
1335 if (result
!= VK_SUCCESS
)
1338 /* We want it to look like it came from this pool */
1339 VG(VALGRIND_FREELIKE_BLOCK(bo
->map
, 0));
1340 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1348 anv_bo_pool_free(struct anv_bo_pool
*pool
, struct anv_bo
*bo
)
1350 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
->map
));
1352 assert(util_is_power_of_two_or_zero(bo
->size
));
1353 const unsigned size_log2
= ilog2_round_up(bo
->size
);
1354 const unsigned bucket
= size_log2
- 12;
1355 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1357 assert(util_sparse_array_get(&pool
->device
->bo_cache
.bo_map
,
1358 bo
->gem_handle
) == bo
);
1359 util_sparse_array_free_list_push(&pool
->free_list
[bucket
],
1360 &bo
->gem_handle
, 1);
1366 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1368 memset(pool
, 0, sizeof(*pool
));
1372 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1374 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1375 for (unsigned i
= 0; i
< 16; i
++) {
1376 if (pool
->bos
[i
][s
] != NULL
)
1377 anv_device_release_bo(device
, pool
->bos
[i
][s
]);
1383 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1384 gl_shader_stage stage
, unsigned per_thread_scratch
)
1386 if (per_thread_scratch
== 0)
1389 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1390 assert(scratch_size_log2
< 16);
1392 struct anv_bo
*bo
= p_atomic_read(&pool
->bos
[scratch_size_log2
][stage
]);
1397 const struct gen_device_info
*devinfo
= &device
->info
;
1399 unsigned subslices
= MAX2(device
->physical
->subslice_total
, 1);
1401 /* The documentation for 3DSTATE_PS "Scratch Space Base Pointer" says:
1403 * "Scratch Space per slice is computed based on 4 sub-slices. SW
1404 * must allocate scratch space enough so that each slice has 4
1407 * According to the other driver team, this applies to compute shaders
1408 * as well. This is not currently documented at all.
1410 * This hack is no longer necessary on Gen11+.
1412 * For, Gen11+, scratch space allocation is based on the number of threads
1413 * in the base configuration.
1415 if (devinfo
->gen
>= 12)
1416 subslices
= devinfo
->num_subslices
[0];
1417 else if (devinfo
->gen
== 11)
1419 else if (devinfo
->gen
>= 9)
1420 subslices
= 4 * devinfo
->num_slices
;
1422 unsigned scratch_ids_per_subslice
;
1423 if (devinfo
->gen
>= 12) {
1424 /* Same as ICL below, but with 16 EUs. */
1425 scratch_ids_per_subslice
= 16 * 8;
1426 } else if (devinfo
->gen
== 11) {
1427 /* The MEDIA_VFE_STATE docs say:
1429 * "Starting with this configuration, the Maximum Number of
1430 * Threads must be set to (#EU * 8) for GPGPU dispatches.
1432 * Although there are only 7 threads per EU in the configuration,
1433 * the FFTID is calculated as if there are 8 threads per EU,
1434 * which in turn requires a larger amount of Scratch Space to be
1435 * allocated by the driver."
1437 scratch_ids_per_subslice
= 8 * 8;
1438 } else if (devinfo
->is_haswell
) {
1439 /* WaCSScratchSize:hsw
1441 * Haswell's scratch space address calculation appears to be sparse
1442 * rather than tightly packed. The Thread ID has bits indicating
1443 * which subslice, EU within a subslice, and thread within an EU it
1444 * is. There's a maximum of two slices and two subslices, so these
1445 * can be stored with a single bit. Even though there are only 10 EUs
1446 * per subslice, this is stored in 4 bits, so there's an effective
1447 * maximum value of 16 EUs. Similarly, although there are only 7
1448 * threads per EU, this is stored in a 3 bit number, giving an
1449 * effective maximum value of 8 threads per EU.
1451 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1452 * number of threads per subslice.
1454 scratch_ids_per_subslice
= 16 * 8;
1455 } else if (devinfo
->is_cherryview
) {
1456 /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
1457 * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
1460 scratch_ids_per_subslice
= 8 * 7;
1462 scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
1465 uint32_t max_threads
[] = {
1466 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1467 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1468 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1469 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1470 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1471 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1474 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1476 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1477 * are still relative to the general state base address. When we emit
1478 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1479 * to the maximum (1 page under 4GB). This allows us to just place the
1480 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1481 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1482 * However, in order to do so, we need to ensure that the kernel does not
1483 * place the scratch BO above the 32-bit boundary.
1485 * NOTE: Technically, it can't go "anywhere" because the top page is off
1486 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1487 * kernel allocates space using
1489 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1491 * so nothing will ever touch the top page.
1493 VkResult result
= anv_device_alloc_bo(device
, size
,
1494 ANV_BO_ALLOC_32BIT_ADDRESS
,
1495 0 /* explicit_address */,
1497 if (result
!= VK_SUCCESS
)
1498 return NULL
; /* TODO */
1500 struct anv_bo
*current_bo
=
1501 p_atomic_cmpxchg(&pool
->bos
[scratch_size_log2
][stage
], NULL
, bo
);
1503 anv_device_release_bo(device
, bo
);
1511 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1513 util_sparse_array_init(&cache
->bo_map
, sizeof(struct anv_bo
), 1024);
1515 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1516 util_sparse_array_finish(&cache
->bo_map
);
1517 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1518 "pthread_mutex_init failed: %m");
1525 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1527 util_sparse_array_finish(&cache
->bo_map
);
1528 pthread_mutex_destroy(&cache
->mutex
);
1531 #define ANV_BO_CACHE_SUPPORTED_FLAGS \
1532 (EXEC_OBJECT_WRITE | \
1533 EXEC_OBJECT_ASYNC | \
1534 EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
1535 EXEC_OBJECT_PINNED | \
1536 EXEC_OBJECT_CAPTURE)
1539 anv_bo_alloc_flags_to_bo_flags(struct anv_device
*device
,
1540 enum anv_bo_alloc_flags alloc_flags
)
1542 struct anv_physical_device
*pdevice
= device
->physical
;
1544 uint64_t bo_flags
= 0;
1545 if (!(alloc_flags
& ANV_BO_ALLOC_32BIT_ADDRESS
) &&
1546 pdevice
->supports_48bit_addresses
)
1547 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1549 if ((alloc_flags
& ANV_BO_ALLOC_CAPTURE
) && pdevice
->has_exec_capture
)
1550 bo_flags
|= EXEC_OBJECT_CAPTURE
;
1552 if (alloc_flags
& ANV_BO_ALLOC_IMPLICIT_WRITE
) {
1553 assert(alloc_flags
& ANV_BO_ALLOC_IMPLICIT_SYNC
);
1554 bo_flags
|= EXEC_OBJECT_WRITE
;
1557 if (!(alloc_flags
& ANV_BO_ALLOC_IMPLICIT_SYNC
) && pdevice
->has_exec_async
)
1558 bo_flags
|= EXEC_OBJECT_ASYNC
;
1560 if (pdevice
->use_softpin
)
1561 bo_flags
|= EXEC_OBJECT_PINNED
;
1567 anv_device_get_bo_align(struct anv_device
*device
,
1568 enum anv_bo_alloc_flags alloc_flags
)
1570 /* Gen12 CCS surface addresses need to be 64K aligned. */
1571 if (device
->info
.gen
>= 12 && (alloc_flags
& ANV_BO_ALLOC_IMPLICIT_CCS
))
1578 anv_device_alloc_bo(struct anv_device
*device
,
1580 enum anv_bo_alloc_flags alloc_flags
,
1581 uint64_t explicit_address
,
1582 struct anv_bo
**bo_out
)
1584 if (!device
->physical
->has_implicit_ccs
)
1585 assert(!(alloc_flags
& ANV_BO_ALLOC_IMPLICIT_CCS
));
1587 const uint32_t bo_flags
=
1588 anv_bo_alloc_flags_to_bo_flags(device
, alloc_flags
);
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 const uint32_t align
= anv_device_get_bo_align(device
, alloc_flags
);
1596 uint64_t ccs_size
= 0;
1597 if (device
->info
.has_aux_map
&& (alloc_flags
& ANV_BO_ALLOC_IMPLICIT_CCS
)) {
1598 /* Align the size up to the next multiple of 64K so we don't have any
1599 * AUX-TT entries pointing from a 64K page to itself.
1601 size
= align_u64(size
, 64 * 1024);
1603 /* See anv_bo::_ccs_size */
1604 ccs_size
= align_u64(DIV_ROUND_UP(size
, GEN_AUX_MAP_GEN12_CCS_SCALE
), 4096);
1607 uint32_t gem_handle
= anv_gem_create(device
, size
+ ccs_size
);
1608 if (gem_handle
== 0)
1609 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1611 struct anv_bo new_bo
= {
1612 .gem_handle
= gem_handle
,
1616 ._ccs_size
= ccs_size
,
1618 .is_external
= (alloc_flags
& ANV_BO_ALLOC_EXTERNAL
),
1619 .has_client_visible_address
=
1620 (alloc_flags
& ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
) != 0,
1621 .has_implicit_ccs
= ccs_size
> 0,
1624 if (alloc_flags
& ANV_BO_ALLOC_MAPPED
) {
1625 new_bo
.map
= anv_gem_mmap(device
, new_bo
.gem_handle
, 0, size
, 0);
1626 if (new_bo
.map
== MAP_FAILED
) {
1627 anv_gem_close(device
, new_bo
.gem_handle
);
1628 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1632 if (alloc_flags
& ANV_BO_ALLOC_SNOOPED
) {
1633 assert(alloc_flags
& ANV_BO_ALLOC_MAPPED
);
1634 /* We don't want to change these defaults if it's going to be shared
1635 * with another process.
1637 assert(!(alloc_flags
& ANV_BO_ALLOC_EXTERNAL
));
1639 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
1640 * I915_CACHING_NONE on non-LLC platforms. For many internal state
1641 * objects, we'd rather take the snooping overhead than risk forgetting
1642 * a CLFLUSH somewhere. Userptr objects are always created as
1643 * I915_CACHING_CACHED, which on non-LLC means snooped so there's no
1644 * need to do this there.
1646 if (!device
->info
.has_llc
) {
1647 anv_gem_set_caching(device
, new_bo
.gem_handle
,
1648 I915_CACHING_CACHED
);
1652 if (alloc_flags
& ANV_BO_ALLOC_FIXED_ADDRESS
) {
1653 new_bo
.has_fixed_address
= true;
1654 new_bo
.offset
= explicit_address
;
1655 } else if (new_bo
.flags
& EXEC_OBJECT_PINNED
) {
1656 new_bo
.offset
= anv_vma_alloc(device
, new_bo
.size
+ new_bo
._ccs_size
,
1657 align
, alloc_flags
, explicit_address
);
1658 if (new_bo
.offset
== 0) {
1660 anv_gem_munmap(new_bo
.map
, size
);
1661 anv_gem_close(device
, new_bo
.gem_handle
);
1662 return vk_errorf(device
, NULL
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1663 "failed to allocate virtual address for BO");
1666 assert(!new_bo
.has_client_visible_address
);
1669 if (new_bo
._ccs_size
> 0) {
1670 assert(device
->info
.has_aux_map
);
1671 gen_aux_map_add_mapping(device
->aux_map_ctx
,
1672 gen_canonical_address(new_bo
.offset
),
1673 gen_canonical_address(new_bo
.offset
+ new_bo
.size
),
1674 new_bo
.size
, 0 /* format_bits */);
1677 assert(new_bo
.gem_handle
);
1679 /* If we just got this gem_handle from anv_bo_init_new then we know no one
1680 * else is touching this BO at the moment so we don't need to lock here.
1682 struct anv_bo
*bo
= anv_device_lookup_bo(device
, new_bo
.gem_handle
);
1691 anv_device_import_bo_from_host_ptr(struct anv_device
*device
,
1692 void *host_ptr
, uint32_t size
,
1693 enum anv_bo_alloc_flags alloc_flags
,
1694 uint64_t client_address
,
1695 struct anv_bo
**bo_out
)
1697 assert(!(alloc_flags
& (ANV_BO_ALLOC_MAPPED
|
1698 ANV_BO_ALLOC_SNOOPED
|
1699 ANV_BO_ALLOC_FIXED_ADDRESS
)));
1701 /* We can't do implicit CCS with an aux table on shared memory */
1702 if (!device
->physical
->has_implicit_ccs
|| device
->info
.has_aux_map
)
1703 assert(!(alloc_flags
& ANV_BO_ALLOC_IMPLICIT_CCS
));
1705 struct anv_bo_cache
*cache
= &device
->bo_cache
;
1706 const uint32_t bo_flags
=
1707 anv_bo_alloc_flags_to_bo_flags(device
, alloc_flags
);
1708 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1710 uint32_t gem_handle
= anv_gem_userptr(device
, host_ptr
, size
);
1712 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1714 pthread_mutex_lock(&cache
->mutex
);
1716 struct anv_bo
*bo
= anv_device_lookup_bo(device
, gem_handle
);
1717 if (bo
->refcount
> 0) {
1718 /* VK_EXT_external_memory_host doesn't require handling importing the
1719 * same pointer twice at the same time, but we don't get in the way. If
1720 * kernel gives us the same gem_handle, only succeed if the flags match.
1722 assert(bo
->gem_handle
== gem_handle
);
1723 if (bo_flags
!= bo
->flags
) {
1724 pthread_mutex_unlock(&cache
->mutex
);
1725 return vk_errorf(device
, NULL
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1726 "same host pointer imported two different ways");
1729 if (bo
->has_client_visible_address
!=
1730 ((alloc_flags
& ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
) != 0)) {
1731 pthread_mutex_unlock(&cache
->mutex
);
1732 return vk_errorf(device
, NULL
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1733 "The same BO was imported with and without buffer "
1737 if (client_address
&& client_address
!= gen_48b_address(bo
->offset
)) {
1738 pthread_mutex_unlock(&cache
->mutex
);
1739 return vk_errorf(device
, NULL
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1740 "The same BO was imported at two different "
1744 __sync_fetch_and_add(&bo
->refcount
, 1);
1746 struct anv_bo new_bo
= {
1747 .gem_handle
= gem_handle
,
1753 .is_external
= true,
1754 .from_host_ptr
= true,
1755 .has_client_visible_address
=
1756 (alloc_flags
& ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
) != 0,
1759 assert(client_address
== gen_48b_address(client_address
));
1760 if (new_bo
.flags
& EXEC_OBJECT_PINNED
) {
1761 assert(new_bo
._ccs_size
== 0);
1762 new_bo
.offset
= anv_vma_alloc(device
, new_bo
.size
,
1763 anv_device_get_bo_align(device
,
1765 alloc_flags
, client_address
);
1766 if (new_bo
.offset
== 0) {
1767 anv_gem_close(device
, new_bo
.gem_handle
);
1768 pthread_mutex_unlock(&cache
->mutex
);
1769 return vk_errorf(device
, NULL
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1770 "failed to allocate virtual address for BO");
1773 assert(!new_bo
.has_client_visible_address
);
1779 pthread_mutex_unlock(&cache
->mutex
);
1786 anv_device_import_bo(struct anv_device
*device
,
1788 enum anv_bo_alloc_flags alloc_flags
,
1789 uint64_t client_address
,
1790 struct anv_bo
**bo_out
)
1792 assert(!(alloc_flags
& (ANV_BO_ALLOC_MAPPED
|
1793 ANV_BO_ALLOC_SNOOPED
|
1794 ANV_BO_ALLOC_FIXED_ADDRESS
)));
1796 /* We can't do implicit CCS with an aux table on shared memory */
1797 if (!device
->physical
->has_implicit_ccs
|| device
->info
.has_aux_map
)
1798 assert(!(alloc_flags
& ANV_BO_ALLOC_IMPLICIT_CCS
));
1800 struct anv_bo_cache
*cache
= &device
->bo_cache
;
1801 const uint32_t bo_flags
=
1802 anv_bo_alloc_flags_to_bo_flags(device
, alloc_flags
);
1803 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1805 pthread_mutex_lock(&cache
->mutex
);
1807 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1809 pthread_mutex_unlock(&cache
->mutex
);
1810 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1813 struct anv_bo
*bo
= anv_device_lookup_bo(device
, gem_handle
);
1814 if (bo
->refcount
> 0) {
1815 /* We have to be careful how we combine flags so that it makes sense.
1816 * Really, though, if we get to this case and it actually matters, the
1817 * client has imported a BO twice in different ways and they get what
1820 uint64_t new_flags
= 0;
1821 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_WRITE
;
1822 new_flags
|= (bo
->flags
& bo_flags
) & EXEC_OBJECT_ASYNC
;
1823 new_flags
|= (bo
->flags
& bo_flags
) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1824 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_PINNED
;
1825 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_CAPTURE
;
1827 /* It's theoretically possible for a BO to get imported such that it's
1828 * both pinned and not pinned. The only way this can happen is if it
1829 * gets imported as both a semaphore and a memory object and that would
1830 * be an application error. Just fail out in that case.
1832 if ((bo
->flags
& EXEC_OBJECT_PINNED
) !=
1833 (bo_flags
& EXEC_OBJECT_PINNED
)) {
1834 pthread_mutex_unlock(&cache
->mutex
);
1835 return vk_errorf(device
, NULL
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1836 "The same BO was imported two different ways");
1839 /* It's also theoretically possible that someone could export a BO from
1840 * one heap and import it into another or to import the same BO into two
1841 * different heaps. If this happens, we could potentially end up both
1842 * allowing and disallowing 48-bit addresses. There's not much we can
1843 * do about it if we're pinning so we just throw an error and hope no
1844 * app is actually that stupid.
1846 if ((new_flags
& EXEC_OBJECT_PINNED
) &&
1847 (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) !=
1848 (bo_flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
)) {
1849 pthread_mutex_unlock(&cache
->mutex
);
1850 return vk_errorf(device
, NULL
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1851 "The same BO was imported on two different heaps");
1854 if (bo
->has_client_visible_address
!=
1855 ((alloc_flags
& ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
) != 0)) {
1856 pthread_mutex_unlock(&cache
->mutex
);
1857 return vk_errorf(device
, NULL
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1858 "The same BO was imported with and without buffer "
1862 if (client_address
&& client_address
!= gen_48b_address(bo
->offset
)) {
1863 pthread_mutex_unlock(&cache
->mutex
);
1864 return vk_errorf(device
, NULL
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1865 "The same BO was imported at two different "
1869 bo
->flags
= new_flags
;
1871 __sync_fetch_and_add(&bo
->refcount
, 1);
1873 off_t size
= lseek(fd
, 0, SEEK_END
);
1874 if (size
== (off_t
)-1) {
1875 anv_gem_close(device
, gem_handle
);
1876 pthread_mutex_unlock(&cache
->mutex
);
1877 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1880 struct anv_bo new_bo
= {
1881 .gem_handle
= gem_handle
,
1886 .is_external
= true,
1887 .has_client_visible_address
=
1888 (alloc_flags
& ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
) != 0,
1891 assert(client_address
== gen_48b_address(client_address
));
1892 if (new_bo
.flags
& EXEC_OBJECT_PINNED
) {
1893 assert(new_bo
._ccs_size
== 0);
1894 new_bo
.offset
= anv_vma_alloc(device
, new_bo
.size
,
1895 anv_device_get_bo_align(device
,
1897 alloc_flags
, client_address
);
1898 if (new_bo
.offset
== 0) {
1899 anv_gem_close(device
, new_bo
.gem_handle
);
1900 pthread_mutex_unlock(&cache
->mutex
);
1901 return vk_errorf(device
, NULL
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1902 "failed to allocate virtual address for BO");
1905 assert(!new_bo
.has_client_visible_address
);
1911 pthread_mutex_unlock(&cache
->mutex
);
1918 anv_device_export_bo(struct anv_device
*device
,
1919 struct anv_bo
*bo
, int *fd_out
)
1921 assert(anv_device_lookup_bo(device
, bo
->gem_handle
) == bo
);
1923 /* This BO must have been flagged external in order for us to be able
1924 * to export it. This is done based on external options passed into
1925 * anv_AllocateMemory.
1927 assert(bo
->is_external
);
1929 int fd
= anv_gem_handle_to_fd(device
, bo
->gem_handle
);
1931 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1939 atomic_dec_not_one(uint32_t *counter
)
1948 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1957 anv_device_release_bo(struct anv_device
*device
,
1960 struct anv_bo_cache
*cache
= &device
->bo_cache
;
1961 assert(anv_device_lookup_bo(device
, bo
->gem_handle
) == bo
);
1963 /* Try to decrement the counter but don't go below one. If this succeeds
1964 * then the refcount has been decremented and we are not the last
1967 if (atomic_dec_not_one(&bo
->refcount
))
1970 pthread_mutex_lock(&cache
->mutex
);
1972 /* We are probably the last reference since our attempt to decrement above
1973 * failed. However, we can't actually know until we are inside the mutex.
1974 * Otherwise, someone could import the BO between the decrement and our
1977 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1978 /* Turns out we're not the last reference. Unlock and bail. */
1979 pthread_mutex_unlock(&cache
->mutex
);
1982 assert(bo
->refcount
== 0);
1984 if (bo
->map
&& !bo
->from_host_ptr
)
1985 anv_gem_munmap(bo
->map
, bo
->size
);
1987 if (bo
->_ccs_size
> 0) {
1988 assert(device
->physical
->has_implicit_ccs
);
1989 assert(device
->info
.has_aux_map
);
1990 assert(bo
->has_implicit_ccs
);
1991 gen_aux_map_unmap_range(device
->aux_map_ctx
,
1992 gen_canonical_address(bo
->offset
),
1996 if ((bo
->flags
& EXEC_OBJECT_PINNED
) && !bo
->has_fixed_address
)
1997 anv_vma_free(device
, bo
->offset
, bo
->size
+ bo
->_ccs_size
);
1999 uint32_t gem_handle
= bo
->gem_handle
;
2001 /* Memset the BO just in case. The refcount being zero should be enough to
2002 * prevent someone from assuming the data is valid but it's safer to just
2003 * stomp to zero just in case. We explicitly do this *before* we close the
2004 * GEM handle to ensure that if anyone allocates something and gets the
2005 * same GEM handle, the memset has already happen and won't stomp all over
2006 * any data they may write in this BO.
2008 memset(bo
, 0, sizeof(*bo
));
2010 anv_gem_close(device
, gem_handle
);
2012 /* Don't unlock until we've actually closed the BO. The whole point of
2013 * the BO cache is to ensure that we correctly handle races with creating
2014 * and releasing GEM handles and we don't want to let someone import the BO
2015 * again between mutex unlock and closing the GEM handle.
2017 pthread_mutex_unlock(&cache
->mutex
);