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
29 #include <linux/futex.h>
30 #include <linux/memfd.h>
33 #include <sys/syscall.h>
35 #include "anv_private.h"
37 #include "util/hash_table.h"
40 #define VG_NOACCESS_READ(__ptr) ({ \
41 VALGRIND_MAKE_MEM_DEFINED((__ptr), sizeof(*(__ptr))); \
42 __typeof(*(__ptr)) __val = *(__ptr); \
43 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr)));\
46 #define VG_NOACCESS_WRITE(__ptr, __val) ({ \
47 VALGRIND_MAKE_MEM_UNDEFINED((__ptr), sizeof(*(__ptr))); \
49 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr))); \
52 #define VG_NOACCESS_READ(__ptr) (*(__ptr))
53 #define VG_NOACCESS_WRITE(__ptr, __val) (*(__ptr) = (__val))
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. */
107 struct anv_mmap_cleanup
{
113 #define ANV_MMAP_CLEANUP_INIT ((struct anv_mmap_cleanup){0})
116 sys_futex(void *addr1
, int op
, int val1
,
117 struct timespec
*timeout
, void *addr2
, int val3
)
119 return syscall(SYS_futex
, addr1
, op
, val1
, timeout
, addr2
, val3
);
123 futex_wake(uint32_t *addr
, int count
)
125 return sys_futex(addr
, FUTEX_WAKE
, count
, NULL
, NULL
, 0);
129 futex_wait(uint32_t *addr
, int32_t value
)
131 return sys_futex(addr
, FUTEX_WAIT
, value
, NULL
, NULL
, 0);
135 memfd_create(const char *name
, unsigned int flags
)
137 return syscall(SYS_memfd_create
, name
, flags
);
140 static inline uint32_t
141 ilog2_round_up(uint32_t value
)
144 return 32 - __builtin_clz(value
- 1);
147 static inline uint32_t
148 round_to_power_of_two(uint32_t value
)
150 return 1 << ilog2_round_up(value
);
154 anv_free_list_pop(union anv_free_list
*list
, void **map
, int32_t *offset
)
156 union anv_free_list current
, new, old
;
158 current
.u64
= list
->u64
;
159 while (current
.offset
!= EMPTY
) {
160 /* We have to add a memory barrier here so that the list head (and
161 * offset) gets read before we read the map pointer. This way we
162 * know that the map pointer is valid for the given offset at the
163 * point where we read it.
165 __sync_synchronize();
167 int32_t *next_ptr
= *map
+ current
.offset
;
168 new.offset
= VG_NOACCESS_READ(next_ptr
);
169 new.count
= current
.count
+ 1;
170 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
171 if (old
.u64
== current
.u64
) {
172 *offset
= current
.offset
;
182 anv_free_list_push(union anv_free_list
*list
, void *map
, int32_t offset
)
184 union anv_free_list current
, old
, new;
185 int32_t *next_ptr
= map
+ offset
;
190 VG_NOACCESS_WRITE(next_ptr
, current
.offset
);
192 new.count
= current
.count
+ 1;
193 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
194 } while (old
.u64
!= current
.u64
);
197 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
198 * means that the bottom 12 bits should all be zero.
200 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
201 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
202 #define PFL_PACK(ptr, count) ({ \
203 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
207 anv_ptr_free_list_pop(void **list
, void **elem
)
209 void *current
= *list
;
210 while (PFL_PTR(current
) != NULL
) {
211 void **next_ptr
= PFL_PTR(current
);
212 void *new_ptr
= VG_NOACCESS_READ(next_ptr
);
213 unsigned new_count
= PFL_COUNT(current
) + 1;
214 void *new = PFL_PACK(new_ptr
, new_count
);
215 void *old
= __sync_val_compare_and_swap(list
, current
, new);
216 if (old
== current
) {
217 *elem
= PFL_PTR(current
);
227 anv_ptr_free_list_push(void **list
, void *elem
)
230 void **next_ptr
= elem
;
232 /* The pointer-based free list requires that the pointer be
233 * page-aligned. This is because we use the bottom 12 bits of the
234 * pointer to store a counter to solve the ABA concurrency problem.
236 assert(((uintptr_t)elem
& 0xfff) == 0);
241 VG_NOACCESS_WRITE(next_ptr
, PFL_PTR(current
));
242 unsigned new_count
= PFL_COUNT(current
) + 1;
243 void *new = PFL_PACK(elem
, new_count
);
244 old
= __sync_val_compare_and_swap(list
, current
, new);
245 } while (old
!= current
);
249 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
);
252 anv_block_pool_init(struct anv_block_pool
*pool
,
253 struct anv_device
*device
, uint32_t block_size
)
257 assert(util_is_power_of_two(block_size
));
259 pool
->device
= device
;
260 anv_bo_init(&pool
->bo
, 0, 0);
261 pool
->block_size
= block_size
;
262 pool
->free_list
= ANV_FREE_LIST_EMPTY
;
263 pool
->back_free_list
= ANV_FREE_LIST_EMPTY
;
265 pool
->fd
= memfd_create("block pool", MFD_CLOEXEC
);
267 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
269 /* Just make it 2GB up-front. The Linux kernel won't actually back it
270 * with pages until we either map and fault on one of them or we use
271 * userptr and send a chunk of it off to the GPU.
273 if (ftruncate(pool
->fd
, BLOCK_POOL_MEMFD_SIZE
) == -1) {
274 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
278 if (!u_vector_init(&pool
->mmap_cleanups
,
279 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
281 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
285 pool
->state
.next
= 0;
287 pool
->back_state
.next
= 0;
288 pool
->back_state
.end
= 0;
290 /* Immediately grow the pool so we'll have a backing bo. */
291 pool
->state
.end
= anv_block_pool_grow(pool
, &pool
->state
);
302 anv_block_pool_finish(struct anv_block_pool
*pool
)
304 struct anv_mmap_cleanup
*cleanup
;
306 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
) {
308 munmap(cleanup
->map
, cleanup
->size
);
309 if (cleanup
->gem_handle
)
310 anv_gem_close(pool
->device
, cleanup
->gem_handle
);
313 u_vector_finish(&pool
->mmap_cleanups
);
318 #define PAGE_SIZE 4096
320 /** Grows and re-centers the block pool.
322 * We grow the block pool in one or both directions in such a way that the
323 * following conditions are met:
325 * 1) The size of the entire pool is always a power of two.
327 * 2) The pool only grows on both ends. Neither end can get
330 * 3) At the end of the allocation, we have about twice as much space
331 * allocated for each end as we have used. This way the pool doesn't
332 * grow too far in one direction or the other.
334 * 4) If the _alloc_back() has never been called, then the back portion of
335 * the pool retains a size of zero. (This makes it easier for users of
336 * the block pool that only want a one-sided pool.)
338 * 5) We have enough space allocated for at least one more block in
339 * whichever side `state` points to.
341 * 6) The center of the pool is always aligned to both the block_size of
342 * the pool and a 4K CPU page.
345 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
350 struct anv_mmap_cleanup
*cleanup
;
352 pthread_mutex_lock(&pool
->device
->mutex
);
354 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
356 /* Gather a little usage information on the pool. Since we may have
357 * threadsd waiting in queue to get some storage while we resize, it's
358 * actually possible that total_used will be larger than old_size. In
359 * particular, block_pool_alloc() increments state->next prior to
360 * calling block_pool_grow, so this ensures that we get enough space for
361 * which ever side tries to grow the pool.
363 * We align to a page size because it makes it easier to do our
364 * calculations later in such a way that we state page-aigned.
366 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
367 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
368 uint32_t total_used
= front_used
+ back_used
;
370 assert(state
== &pool
->state
|| back_used
> 0);
372 size_t old_size
= pool
->bo
.size
;
375 back_used
* 2 <= pool
->center_bo_offset
&&
376 front_used
* 2 <= (old_size
- pool
->center_bo_offset
)) {
377 /* If we're in this case then this isn't the firsta allocation and we
378 * already have enough space on both sides to hold double what we
379 * have allocated. There's nothing for us to do.
385 /* This is the first allocation */
386 size
= MAX2(32 * pool
->block_size
, PAGE_SIZE
);
391 /* We can't have a block pool bigger than 1GB because we use signed
392 * 32-bit offsets in the free list and we don't want overflow. We
393 * should never need a block pool bigger than 1GB anyway.
395 assert(size
<= (1u << 31));
397 /* We compute a new center_bo_offset such that, when we double the size
398 * of the pool, we maintain the ratio of how much is used by each side.
399 * This way things should remain more-or-less balanced.
401 uint32_t center_bo_offset
;
402 if (back_used
== 0) {
403 /* If we're in this case then we have never called alloc_back(). In
404 * this case, we want keep the offset at 0 to make things as simple
405 * as possible for users that don't care about back allocations.
407 center_bo_offset
= 0;
409 /* Try to "center" the allocation based on how much is currently in
410 * use on each side of the center line.
412 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
414 /* Align down to a multiple of both the block size and page size */
415 uint32_t granularity
= MAX2(pool
->block_size
, PAGE_SIZE
);
416 assert(util_is_power_of_two(granularity
));
417 center_bo_offset
&= ~(granularity
- 1);
419 assert(center_bo_offset
>= back_used
);
421 /* Make sure we don't shrink the back end of the pool */
422 if (center_bo_offset
< pool
->back_state
.end
)
423 center_bo_offset
= pool
->back_state
.end
;
425 /* Make sure that we don't shrink the front end of the pool */
426 if (size
- center_bo_offset
< pool
->state
.end
)
427 center_bo_offset
= size
- pool
->state
.end
;
430 assert(center_bo_offset
% pool
->block_size
== 0);
431 assert(center_bo_offset
% PAGE_SIZE
== 0);
433 /* Assert that we only ever grow the pool */
434 assert(center_bo_offset
>= pool
->back_state
.end
);
435 assert(size
- center_bo_offset
>= pool
->state
.end
);
437 cleanup
= u_vector_add(&pool
->mmap_cleanups
);
440 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
442 /* Just leak the old map until we destroy the pool. We can't munmap it
443 * without races or imposing locking on the block allocate fast path. On
444 * the whole the leaked maps adds up to less than the size of the
445 * current map. MAP_POPULATE seems like the right thing to do, but we
446 * should try to get some numbers.
448 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
449 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
450 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
452 cleanup
->size
= size
;
454 if (map
== MAP_FAILED
)
457 gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
460 cleanup
->gem_handle
= gem_handle
;
463 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
464 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
465 * always created as I915_CACHING_CACHED, which on non-LLC means
466 * snooped. That can be useful but comes with a bit of overheard. Since
467 * we're eplicitly clflushing and don't want the overhead we need to turn
469 if (!pool
->device
->info
.has_llc
) {
470 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_NONE
);
471 anv_gem_set_domain(pool
->device
, gem_handle
,
472 I915_GEM_DOMAIN_GTT
, I915_GEM_DOMAIN_GTT
);
476 /* Now that we successfull allocated everything, we can write the new
477 * values back into pool. */
478 pool
->map
= map
+ center_bo_offset
;
479 pool
->center_bo_offset
= center_bo_offset
;
481 /* For block pool BOs we have to be a bit careful about where we place them
482 * in the GTT. There are two documented workarounds for state base address
483 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
484 * which state that those two base addresses do not support 48-bit
485 * addresses and need to be placed in the bottom 32-bit range.
486 * Unfortunately, this is not quite accurate.
488 * The real problem is that we always set the size of our state pools in
489 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
490 * likely significantly smaller. We do this because we do not no at the
491 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
492 * the pool during command buffer building so we don't actually have a
493 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
494 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
495 * as being out of bounds and returns zero. For dynamic state, this
496 * usually just leads to rendering corruptions, but shaders that are all
497 * zero hang the GPU immediately.
499 * The easiest solution to do is exactly what the bogus workarounds say to
500 * do: restrict these buffers to 32-bit addresses. We could also pin the
501 * BO to some particular location of our choosing, but that's significantly
502 * more work than just not setting a flag. So, we explicitly DO NOT set
503 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
506 anv_bo_init(&pool
->bo
, gem_handle
, size
);
509 if (pool
->device
->instance
->physicalDevice
.has_exec_async
)
510 pool
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
513 pthread_mutex_unlock(&pool
->device
->mutex
);
515 /* Return the appropreate new size. This function never actually
516 * updates state->next. Instead, we let the caller do that because it
517 * needs to do so in order to maintain its concurrency model.
519 if (state
== &pool
->state
) {
520 return pool
->bo
.size
- pool
->center_bo_offset
;
522 assert(pool
->center_bo_offset
> 0);
523 return pool
->center_bo_offset
;
527 pthread_mutex_unlock(&pool
->device
->mutex
);
533 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
534 struct anv_block_state
*pool_state
)
536 struct anv_block_state state
, old
, new;
539 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, pool
->block_size
);
540 if (state
.next
< state
.end
) {
543 } else if (state
.next
== state
.end
) {
544 /* We allocated the first block outside the pool, we have to grow it.
545 * pool_state->next acts a mutex: threads who try to allocate now will
546 * get block indexes above the current limit and hit futex_wait
548 new.next
= state
.next
+ pool
->block_size
;
549 new.end
= anv_block_pool_grow(pool
, pool_state
);
550 assert(new.end
>= new.next
&& new.end
% pool
->block_size
== 0);
551 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
552 if (old
.next
!= state
.next
)
553 futex_wake(&pool_state
->end
, INT_MAX
);
556 futex_wait(&pool_state
->end
, state
.end
);
563 anv_block_pool_alloc(struct anv_block_pool
*pool
)
567 /* Try free list first. */
568 if (anv_free_list_pop(&pool
->free_list
, &pool
->map
, &offset
)) {
574 return anv_block_pool_alloc_new(pool
, &pool
->state
);
577 /* Allocates a block out of the back of the block pool.
579 * This will allocated a block earlier than the "start" of the block pool.
580 * The offsets returned from this function will be negative but will still
581 * be correct relative to the block pool's map pointer.
583 * If you ever use anv_block_pool_alloc_back, then you will have to do
584 * gymnastics with the block pool's BO when doing relocations.
587 anv_block_pool_alloc_back(struct anv_block_pool
*pool
)
591 /* Try free list first. */
592 if (anv_free_list_pop(&pool
->back_free_list
, &pool
->map
, &offset
)) {
598 offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
);
600 /* The offset we get out of anv_block_pool_alloc_new() is actually the
601 * number of bytes downwards from the middle to the end of the block.
602 * We need to turn it into a (negative) offset from the middle to the
603 * start of the block.
606 return -(offset
+ pool
->block_size
);
610 anv_block_pool_free(struct anv_block_pool
*pool
, int32_t offset
)
613 anv_free_list_push(&pool
->back_free_list
, pool
->map
, offset
);
615 anv_free_list_push(&pool
->free_list
, pool
->map
, offset
);
620 anv_fixed_size_state_pool_init(struct anv_fixed_size_state_pool
*pool
,
623 /* At least a cache line and must divide the block size. */
624 assert(state_size
>= 64 && util_is_power_of_two(state_size
));
626 pool
->state_size
= state_size
;
627 pool
->free_list
= ANV_FREE_LIST_EMPTY
;
628 pool
->block
.next
= 0;
633 anv_fixed_size_state_pool_alloc(struct anv_fixed_size_state_pool
*pool
,
634 struct anv_block_pool
*block_pool
)
637 struct anv_block_state block
, old
, new;
639 /* Try free list first. */
640 if (anv_free_list_pop(&pool
->free_list
, &block_pool
->map
, &offset
)) {
645 /* If free list was empty (or somebody raced us and took the items) we
646 * allocate a new item from the end of the block */
648 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, pool
->state_size
);
650 if (block
.next
< block
.end
) {
652 } else if (block
.next
== block
.end
) {
653 offset
= anv_block_pool_alloc(block_pool
);
654 new.next
= offset
+ pool
->state_size
;
655 new.end
= offset
+ block_pool
->block_size
;
656 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
657 if (old
.next
!= block
.next
)
658 futex_wake(&pool
->block
.end
, INT_MAX
);
661 futex_wait(&pool
->block
.end
, block
.end
);
667 anv_fixed_size_state_pool_free(struct anv_fixed_size_state_pool
*pool
,
668 struct anv_block_pool
*block_pool
,
671 anv_free_list_push(&pool
->free_list
, block_pool
->map
, offset
);
675 anv_state_pool_init(struct anv_state_pool
*pool
,
676 struct anv_block_pool
*block_pool
)
678 pool
->block_pool
= block_pool
;
679 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
680 size_t size
= 1 << (ANV_MIN_STATE_SIZE_LOG2
+ i
);
681 anv_fixed_size_state_pool_init(&pool
->buckets
[i
], size
);
683 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
687 anv_state_pool_finish(struct anv_state_pool
*pool
)
689 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
693 anv_state_pool_alloc(struct anv_state_pool
*pool
, size_t size
, size_t align
)
695 unsigned size_log2
= ilog2_round_up(size
< align
? align
: size
);
696 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
697 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
698 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
699 unsigned bucket
= size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
701 struct anv_state state
;
702 state
.alloc_size
= 1 << size_log2
;
703 state
.offset
= anv_fixed_size_state_pool_alloc(&pool
->buckets
[bucket
],
705 state
.map
= pool
->block_pool
->map
+ state
.offset
;
706 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
711 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
713 assert(util_is_power_of_two(state
.alloc_size
));
714 unsigned size_log2
= ilog2_round_up(state
.alloc_size
);
715 assert(size_log2
>= ANV_MIN_STATE_SIZE_LOG2
&&
716 size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
717 unsigned bucket
= size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
719 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
720 anv_fixed_size_state_pool_free(&pool
->buckets
[bucket
],
721 pool
->block_pool
, state
.offset
);
725 struct anv_state_stream_block
{
727 struct anv_state_stream_block
*next
;
729 /* The offset into the block pool at which this block starts */
733 /* A pointer to the first user-allocated thing in this block. This is
734 * what valgrind sees as the start of the block.
740 /* The state stream allocator is a one-shot, single threaded allocator for
741 * variable sized blocks. We use it for allocating dynamic state.
744 anv_state_stream_init(struct anv_state_stream
*stream
,
745 struct anv_block_pool
*block_pool
)
747 stream
->block_pool
= block_pool
;
748 stream
->block
= NULL
;
750 /* Ensure that next + whatever > end. This way the first call to
751 * state_stream_alloc fetches a new block.
756 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
760 anv_state_stream_finish(struct anv_state_stream
*stream
)
762 VG(const uint32_t block_size
= stream
->block_pool
->block_size
);
764 struct anv_state_stream_block
*next
= stream
->block
;
765 while (next
!= NULL
) {
766 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
767 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
768 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, block_size
));
769 anv_block_pool_free(stream
->block_pool
, sb
.offset
);
773 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
777 anv_state_stream_alloc(struct anv_state_stream
*stream
,
778 uint32_t size
, uint32_t alignment
)
780 struct anv_state_stream_block
*sb
= stream
->block
;
782 struct anv_state state
;
784 state
.offset
= align_u32(stream
->next
, alignment
);
785 if (state
.offset
+ size
> stream
->end
) {
786 uint32_t block
= anv_block_pool_alloc(stream
->block_pool
);
787 sb
= stream
->block_pool
->map
+ block
;
789 VG(VALGRIND_MAKE_MEM_UNDEFINED(sb
, sizeof(*sb
)));
790 sb
->next
= stream
->block
;
792 VG(sb
->_vg_ptr
= NULL
);
793 VG(VALGRIND_MAKE_MEM_NOACCESS(sb
, stream
->block_pool
->block_size
));
796 stream
->start
= block
;
797 stream
->next
= block
+ sizeof(*sb
);
798 stream
->end
= block
+ stream
->block_pool
->block_size
;
800 state
.offset
= align_u32(stream
->next
, alignment
);
801 assert(state
.offset
+ size
<= stream
->end
);
804 assert(state
.offset
> stream
->start
);
805 state
.map
= (void *)sb
+ (state
.offset
- stream
->start
);
806 state
.alloc_size
= size
;
809 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
810 if (vg_ptr
== NULL
) {
812 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
813 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
815 void *state_end
= state
.map
+ state
.alloc_size
;
816 /* This only updates the mempool. The newly allocated chunk is still
817 * marked as NOACCESS. */
818 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
819 /* Mark the newly allocated chunk as undefined */
820 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
824 stream
->next
= state
.offset
+ size
;
829 struct bo_pool_bo_link
{
830 struct bo_pool_bo_link
*next
;
835 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
)
837 pool
->device
= device
;
838 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
840 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
844 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
846 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
847 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
848 while (link
!= NULL
) {
849 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
851 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
852 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
853 link
= link_copy
.next
;
857 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
861 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
865 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
866 const unsigned pow2_size
= 1 << size_log2
;
867 const unsigned bucket
= size_log2
- 12;
868 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
870 void *next_free_void
;
871 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
872 struct bo_pool_bo_link
*next_free
= next_free_void
;
873 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
874 assert(bo
->gem_handle
);
875 assert(bo
->map
== next_free
);
876 assert(size
<= bo
->size
);
878 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
883 struct anv_bo new_bo
;
885 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
886 if (result
!= VK_SUCCESS
)
889 assert(new_bo
.size
== pow2_size
);
891 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
892 if (new_bo
.map
== NULL
) {
893 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
894 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
899 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
905 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
907 /* Make a copy in case the anv_bo happens to be storred in the BO */
908 struct anv_bo bo
= *bo_in
;
910 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
912 struct bo_pool_bo_link
*link
= bo
.map
;
913 VG_NOACCESS_WRITE(&link
->bo
, bo
);
915 assert(util_is_power_of_two(bo
.size
));
916 const unsigned size_log2
= ilog2_round_up(bo
.size
);
917 const unsigned bucket
= size_log2
- 12;
918 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
920 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
926 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
928 memset(pool
, 0, sizeof(*pool
));
932 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
934 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
935 for (unsigned i
= 0; i
< 16; i
++) {
936 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
938 anv_gem_close(device
, bo
->bo
.gem_handle
);
944 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
945 gl_shader_stage stage
, unsigned per_thread_scratch
)
947 if (per_thread_scratch
== 0)
950 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
951 assert(scratch_size_log2
< 16);
953 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
955 /* We can use "exists" to shortcut and ignore the critical section */
959 pthread_mutex_lock(&device
->mutex
);
961 __sync_synchronize();
965 const struct anv_physical_device
*physical_device
=
966 &device
->instance
->physicalDevice
;
967 const struct gen_device_info
*devinfo
= &physical_device
->info
;
969 /* WaCSScratchSize:hsw
971 * Haswell's scratch space address calculation appears to be sparse
972 * rather than tightly packed. The Thread ID has bits indicating which
973 * subslice, EU within a subslice, and thread within an EU it is.
974 * There's a maximum of two slices and two subslices, so these can be
975 * stored with a single bit. Even though there are only 10 EUs per
976 * subslice, this is stored in 4 bits, so there's an effective maximum
977 * value of 16 EUs. Similarly, although there are only 7 threads per EU,
978 * this is stored in a 3 bit number, giving an effective maximum value
979 * of 8 threads per EU.
981 * This means that we need to use 16 * 8 instead of 10 * 7 for the
982 * number of threads per subslice.
984 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
985 const unsigned scratch_ids_per_subslice
=
986 device
->info
.is_haswell
? 16 * 8 : devinfo
->max_cs_threads
;
988 uint32_t max_threads
[] = {
989 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
990 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
991 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
992 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
993 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
994 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
997 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
999 anv_bo_init_new(&bo
->bo
, device
, size
);
1001 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1002 * are still relative to the general state base address. When we emit
1003 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1004 * to the maximum (1 page under 4GB). This allows us to just place the
1005 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1006 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1007 * However, in order to do so, we need to ensure that the kernel does not
1008 * place the scratch BO above the 32-bit boundary.
1010 * NOTE: Technically, it can't go "anywhere" because the top page is off
1011 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1012 * kernel allocates space using
1014 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1016 * so nothing will ever touch the top page.
1018 bo
->bo
.flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1020 /* Set the exists last because it may be read by other threads */
1021 __sync_synchronize();
1024 pthread_mutex_unlock(&device
->mutex
);
1029 struct anv_cached_bo
{
1036 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1038 cache
->bo_map
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
1039 _mesa_key_pointer_equal
);
1041 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1043 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1044 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1045 return vk_errorf(VK_ERROR_OUT_OF_HOST_MEMORY
,
1046 "pthread_mutex_init failed: %m");
1053 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1055 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1056 pthread_mutex_destroy(&cache
->mutex
);
1059 static struct anv_cached_bo
*
1060 anv_bo_cache_lookup_locked(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1062 struct hash_entry
*entry
=
1063 _mesa_hash_table_search(cache
->bo_map
,
1064 (const void *)(uintptr_t)gem_handle
);
1068 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)entry
->data
;
1069 assert(bo
->bo
.gem_handle
== gem_handle
);
1074 static struct anv_bo
*
1075 anv_bo_cache_lookup(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1077 pthread_mutex_lock(&cache
->mutex
);
1079 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1081 pthread_mutex_unlock(&cache
->mutex
);
1083 return bo
? &bo
->bo
: NULL
;
1087 anv_bo_cache_alloc(struct anv_device
*device
,
1088 struct anv_bo_cache
*cache
,
1089 uint64_t size
, struct anv_bo
**bo_out
)
1091 struct anv_cached_bo
*bo
=
1092 vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1093 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1095 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1099 /* The kernel is going to give us whole pages anyway */
1100 size
= align_u64(size
, 4096);
1102 VkResult result
= anv_bo_init_new(&bo
->bo
, device
, size
);
1103 if (result
!= VK_SUCCESS
) {
1104 vk_free(&device
->alloc
, bo
);
1108 assert(bo
->bo
.gem_handle
);
1110 pthread_mutex_lock(&cache
->mutex
);
1112 _mesa_hash_table_insert(cache
->bo_map
,
1113 (void *)(uintptr_t)bo
->bo
.gem_handle
, bo
);
1115 pthread_mutex_unlock(&cache
->mutex
);
1123 anv_bo_cache_import(struct anv_device
*device
,
1124 struct anv_bo_cache
*cache
,
1125 int fd
, uint64_t size
, struct anv_bo
**bo_out
)
1127 pthread_mutex_lock(&cache
->mutex
);
1129 /* The kernel is going to give us whole pages anyway */
1130 size
= align_u64(size
, 4096);
1132 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1134 pthread_mutex_unlock(&cache
->mutex
);
1135 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHX
);
1138 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1140 if (bo
->bo
.size
!= size
) {
1141 pthread_mutex_unlock(&cache
->mutex
);
1142 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHX
);
1144 __sync_fetch_and_add(&bo
->refcount
, 1);
1146 /* For security purposes, we reject BO imports where the size does not
1147 * match exactly. This prevents a malicious client from passing a
1148 * buffer to a trusted client, lying about the size, and telling the
1149 * trusted client to try and texture from an image that goes
1150 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1151 * in the trusted client. The trusted client can protect itself against
1152 * this sort of attack but only if it can trust the buffer size.
1154 off_t import_size
= lseek(fd
, 0, SEEK_END
);
1155 if (import_size
== (off_t
)-1 || import_size
!= size
) {
1156 anv_gem_close(device
, gem_handle
);
1157 pthread_mutex_unlock(&cache
->mutex
);
1158 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHX
);
1161 bo
= vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1162 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1164 anv_gem_close(device
, gem_handle
);
1165 pthread_mutex_unlock(&cache
->mutex
);
1166 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1171 anv_bo_init(&bo
->bo
, gem_handle
, size
);
1173 if (device
->instance
->physicalDevice
.supports_48bit_addresses
)
1174 bo
->bo
.flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1176 if (device
->instance
->physicalDevice
.has_exec_async
)
1177 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1179 _mesa_hash_table_insert(cache
->bo_map
, (void *)(uintptr_t)gem_handle
, bo
);
1182 pthread_mutex_unlock(&cache
->mutex
);
1184 /* From the Vulkan spec:
1186 * "Importing memory from a file descriptor transfers ownership of
1187 * the file descriptor from the application to the Vulkan
1188 * implementation. The application must not perform any operations on
1189 * the file descriptor after a successful import."
1191 * If the import fails, we leave the file descriptor open.
1201 anv_bo_cache_export(struct anv_device
*device
,
1202 struct anv_bo_cache
*cache
,
1203 struct anv_bo
*bo_in
, int *fd_out
)
1205 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1206 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1208 int fd
= anv_gem_handle_to_fd(device
, bo
->bo
.gem_handle
);
1210 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1218 atomic_dec_not_one(uint32_t *counter
)
1227 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1236 anv_bo_cache_release(struct anv_device
*device
,
1237 struct anv_bo_cache
*cache
,
1238 struct anv_bo
*bo_in
)
1240 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1241 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1243 /* Try to decrement the counter but don't go below one. If this succeeds
1244 * then the refcount has been decremented and we are not the last
1247 if (atomic_dec_not_one(&bo
->refcount
))
1250 pthread_mutex_lock(&cache
->mutex
);
1252 /* We are probably the last reference since our attempt to decrement above
1253 * failed. However, we can't actually know until we are inside the mutex.
1254 * Otherwise, someone could import the BO between the decrement and our
1257 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1258 /* Turns out we're not the last reference. Unlock and bail. */
1259 pthread_mutex_unlock(&cache
->mutex
);
1263 struct hash_entry
*entry
=
1264 _mesa_hash_table_search(cache
->bo_map
,
1265 (const void *)(uintptr_t)bo
->bo
.gem_handle
);
1267 _mesa_hash_table_remove(cache
->bo_map
, entry
);
1270 anv_gem_munmap(bo
->bo
.map
, bo
->bo
.size
);
1272 anv_gem_close(device
, bo
->bo
.gem_handle
);
1274 /* Don't unlock until we've actually closed the BO. The whole point of
1275 * the BO cache is to ensure that we correctly handle races with creating
1276 * and releasing GEM handles and we don't want to let someone import the BO
1277 * again between mutex unlock and closing the GEM handle.
1279 pthread_mutex_unlock(&cache
->mutex
);
1281 vk_free(&device
->alloc
, bo
);