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"
38 #define VG_NOACCESS_READ(__ptr) ({ \
39 VALGRIND_MAKE_MEM_DEFINED((__ptr), sizeof(*(__ptr))); \
40 __typeof(*(__ptr)) __val = *(__ptr); \
41 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr)));\
44 #define VG_NOACCESS_WRITE(__ptr, __val) ({ \
45 VALGRIND_MAKE_MEM_UNDEFINED((__ptr), sizeof(*(__ptr))); \
47 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr))); \
50 #define VG_NOACCESS_READ(__ptr) (*(__ptr))
51 #define VG_NOACCESS_WRITE(__ptr, __val) (*(__ptr) = (__val))
56 * - Lock free (except when resizing underlying bos)
58 * - Constant time allocation with typically only one atomic
60 * - Multiple allocation sizes without fragmentation
62 * - Can grow while keeping addresses and offset of contents stable
64 * - All allocations within one bo so we can point one of the
65 * STATE_BASE_ADDRESS pointers at it.
67 * The overall design is a two-level allocator: top level is a fixed size, big
68 * block (8k) allocator, which operates out of a bo. Allocation is done by
69 * either pulling a block from the free list or growing the used range of the
70 * bo. Growing the range may run out of space in the bo which we then need to
71 * grow. Growing the bo is tricky in a multi-threaded, lockless environment:
72 * we need to keep all pointers and contents in the old map valid. GEM bos in
73 * general can't grow, but we use a trick: we create a memfd and use ftruncate
74 * to grow it as necessary. We mmap the new size and then create a gem bo for
75 * it using the new gem userptr ioctl. Without heavy-handed locking around
76 * our allocation fast-path, there isn't really a way to munmap the old mmap,
77 * so we just keep it around until garbage collection time. While the block
78 * allocator is lockless for normal operations, we block other threads trying
79 * to allocate while we're growing the map. It sholdn't happen often, and
80 * growing is fast anyway.
82 * At the next level we can use various sub-allocators. The state pool is a
83 * pool of smaller, fixed size objects, which operates much like the block
84 * pool. It uses a free list for freeing objects, but when it runs out of
85 * space it just allocates a new block from the block pool. This allocator is
86 * intended for longer lived state objects such as SURFACE_STATE and most
87 * other persistent state objects in the API. We may need to track more info
88 * with these object and a pointer back to the CPU object (eg VkImage). In
89 * those cases we just allocate a slightly bigger object and put the extra
90 * state after the GPU state object.
92 * The state stream allocator works similar to how the i965 DRI driver streams
93 * all its state. Even with Vulkan, we need to emit transient state (whether
94 * surface state base or dynamic state base), and for that we can just get a
95 * block and fill it up. These cases are local to a command buffer and the
96 * sub-allocator need not be thread safe. The streaming allocator gets a new
97 * block when it runs out of space and chains them together so they can be
101 /* Allocations are always at least 64 byte aligned, so 1 is an invalid value.
102 * We use it to indicate the free list is empty. */
105 struct anv_mmap_cleanup
{
111 #define ANV_MMAP_CLEANUP_INIT ((struct anv_mmap_cleanup){0})
114 sys_futex(void *addr1
, int op
, int val1
,
115 struct timespec
*timeout
, void *addr2
, int val3
)
117 return syscall(SYS_futex
, addr1
, op
, val1
, timeout
, addr2
, val3
);
121 futex_wake(uint32_t *addr
, int count
)
123 return sys_futex(addr
, FUTEX_WAKE
, count
, NULL
, NULL
, 0);
127 futex_wait(uint32_t *addr
, int32_t value
)
129 return sys_futex(addr
, FUTEX_WAIT
, value
, NULL
, NULL
, 0);
133 memfd_create(const char *name
, unsigned int flags
)
135 return syscall(SYS_memfd_create
, name
, flags
);
138 static inline uint32_t
139 ilog2_round_up(uint32_t value
)
142 return 32 - __builtin_clz(value
- 1);
145 static inline uint32_t
146 round_to_power_of_two(uint32_t value
)
148 return 1 << ilog2_round_up(value
);
152 anv_free_list_pop(union anv_free_list
*list
, void **map
, int32_t *offset
)
154 union anv_free_list current
, new, old
;
156 current
.u64
= list
->u64
;
157 while (current
.offset
!= EMPTY
) {
158 /* We have to add a memory barrier here so that the list head (and
159 * offset) gets read before we read the map pointer. This way we
160 * know that the map pointer is valid for the given offset at the
161 * point where we read it.
163 __sync_synchronize();
165 int32_t *next_ptr
= *map
+ current
.offset
;
166 new.offset
= VG_NOACCESS_READ(next_ptr
);
167 new.count
= current
.count
+ 1;
168 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
169 if (old
.u64
== current
.u64
) {
170 *offset
= current
.offset
;
180 anv_free_list_push(union anv_free_list
*list
, void *map
, int32_t offset
)
182 union anv_free_list current
, old
, new;
183 int32_t *next_ptr
= map
+ offset
;
188 VG_NOACCESS_WRITE(next_ptr
, current
.offset
);
190 new.count
= current
.count
+ 1;
191 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
192 } while (old
.u64
!= current
.u64
);
195 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
196 * means that the bottom 12 bits should all be zero.
198 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
199 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
200 #define PFL_PACK(ptr, count) ({ \
201 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
205 anv_ptr_free_list_pop(void **list
, void **elem
)
207 void *current
= *list
;
208 while (PFL_PTR(current
) != NULL
) {
209 void **next_ptr
= PFL_PTR(current
);
210 void *new_ptr
= VG_NOACCESS_READ(next_ptr
);
211 unsigned new_count
= PFL_COUNT(current
) + 1;
212 void *new = PFL_PACK(new_ptr
, new_count
);
213 void *old
= __sync_val_compare_and_swap(list
, current
, new);
214 if (old
== current
) {
215 *elem
= PFL_PTR(current
);
225 anv_ptr_free_list_push(void **list
, void *elem
)
228 void **next_ptr
= elem
;
230 /* The pointer-based free list requires that the pointer be
231 * page-aligned. This is because we use the bottom 12 bits of the
232 * pointer to store a counter to solve the ABA concurrency problem.
234 assert(((uintptr_t)elem
& 0xfff) == 0);
239 VG_NOACCESS_WRITE(next_ptr
, PFL_PTR(current
));
240 unsigned new_count
= PFL_COUNT(current
) + 1;
241 void *new = PFL_PACK(elem
, new_count
);
242 old
= __sync_val_compare_and_swap(list
, current
, new);
243 } while (old
!= current
);
247 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
);
250 anv_block_pool_init(struct anv_block_pool
*pool
,
251 struct anv_device
*device
, uint32_t block_size
)
253 assert(util_is_power_of_two(block_size
));
255 pool
->device
= device
;
256 pool
->bo
.gem_handle
= 0;
259 pool
->bo
.is_winsys_bo
= false;
260 pool
->block_size
= block_size
;
261 pool
->free_list
= ANV_FREE_LIST_EMPTY
;
262 pool
->back_free_list
= ANV_FREE_LIST_EMPTY
;
264 pool
->fd
= memfd_create("block pool", MFD_CLOEXEC
);
268 /* Just make it 2GB up-front. The Linux kernel won't actually back it
269 * with pages until we either map and fault on one of them or we use
270 * userptr and send a chunk of it off to the GPU.
272 if (ftruncate(pool
->fd
, BLOCK_POOL_MEMFD_SIZE
) == -1)
275 u_vector_init(&pool
->mmap_cleanups
,
276 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)), 128);
278 pool
->state
.next
= 0;
280 pool
->back_state
.next
= 0;
281 pool
->back_state
.end
= 0;
283 /* Immediately grow the pool so we'll have a backing bo. */
284 pool
->state
.end
= anv_block_pool_grow(pool
, &pool
->state
);
288 anv_block_pool_finish(struct anv_block_pool
*pool
)
290 struct anv_mmap_cleanup
*cleanup
;
292 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
) {
294 munmap(cleanup
->map
, cleanup
->size
);
295 if (cleanup
->gem_handle
)
296 anv_gem_close(pool
->device
, cleanup
->gem_handle
);
299 u_vector_finish(&pool
->mmap_cleanups
);
304 #define PAGE_SIZE 4096
306 /** Grows and re-centers the block pool.
308 * We grow the block pool in one or both directions in such a way that the
309 * following conditions are met:
311 * 1) The size of the entire pool is always a power of two.
313 * 2) The pool only grows on both ends. Neither end can get
316 * 3) At the end of the allocation, we have about twice as much space
317 * allocated for each end as we have used. This way the pool doesn't
318 * grow too far in one direction or the other.
320 * 4) If the _alloc_back() has never been called, then the back portion of
321 * the pool retains a size of zero. (This makes it easier for users of
322 * the block pool that only want a one-sided pool.)
324 * 5) We have enough space allocated for at least one more block in
325 * whichever side `state` points to.
327 * 6) The center of the pool is always aligned to both the block_size of
328 * the pool and a 4K CPU page.
331 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
336 struct anv_mmap_cleanup
*cleanup
;
338 pthread_mutex_lock(&pool
->device
->mutex
);
340 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
342 /* Gather a little usage information on the pool. Since we may have
343 * threadsd waiting in queue to get some storage while we resize, it's
344 * actually possible that total_used will be larger than old_size. In
345 * particular, block_pool_alloc() increments state->next prior to
346 * calling block_pool_grow, so this ensures that we get enough space for
347 * which ever side tries to grow the pool.
349 * We align to a page size because it makes it easier to do our
350 * calculations later in such a way that we state page-aigned.
352 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
353 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
354 uint32_t total_used
= front_used
+ back_used
;
356 assert(state
== &pool
->state
|| back_used
> 0);
358 size_t old_size
= pool
->bo
.size
;
361 back_used
* 2 <= pool
->center_bo_offset
&&
362 front_used
* 2 <= (old_size
- pool
->center_bo_offset
)) {
363 /* If we're in this case then this isn't the firsta allocation and we
364 * already have enough space on both sides to hold double what we
365 * have allocated. There's nothing for us to do.
371 /* This is the first allocation */
372 size
= MAX2(32 * pool
->block_size
, PAGE_SIZE
);
377 /* We can't have a block pool bigger than 1GB because we use signed
378 * 32-bit offsets in the free list and we don't want overflow. We
379 * should never need a block pool bigger than 1GB anyway.
381 assert(size
<= (1u << 31));
383 /* We compute a new center_bo_offset such that, when we double the size
384 * of the pool, we maintain the ratio of how much is used by each side.
385 * This way things should remain more-or-less balanced.
387 uint32_t center_bo_offset
;
388 if (back_used
== 0) {
389 /* If we're in this case then we have never called alloc_back(). In
390 * this case, we want keep the offset at 0 to make things as simple
391 * as possible for users that don't care about back allocations.
393 center_bo_offset
= 0;
395 /* Try to "center" the allocation based on how much is currently in
396 * use on each side of the center line.
398 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
400 /* Align down to a multiple of both the block size and page size */
401 uint32_t granularity
= MAX2(pool
->block_size
, PAGE_SIZE
);
402 assert(util_is_power_of_two(granularity
));
403 center_bo_offset
&= ~(granularity
- 1);
405 assert(center_bo_offset
>= back_used
);
407 /* Make sure we don't shrink the back end of the pool */
408 if (center_bo_offset
< pool
->back_state
.end
)
409 center_bo_offset
= pool
->back_state
.end
;
411 /* Make sure that we don't shrink the front end of the pool */
412 if (size
- center_bo_offset
< pool
->state
.end
)
413 center_bo_offset
= size
- pool
->state
.end
;
416 assert(center_bo_offset
% pool
->block_size
== 0);
417 assert(center_bo_offset
% PAGE_SIZE
== 0);
419 /* Assert that we only ever grow the pool */
420 assert(center_bo_offset
>= pool
->back_state
.end
);
421 assert(size
- center_bo_offset
>= pool
->state
.end
);
423 cleanup
= u_vector_add(&pool
->mmap_cleanups
);
426 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
428 /* Just leak the old map until we destroy the pool. We can't munmap it
429 * without races or imposing locking on the block allocate fast path. On
430 * the whole the leaked maps adds up to less than the size of the
431 * current map. MAP_POPULATE seems like the right thing to do, but we
432 * should try to get some numbers.
434 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
435 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
436 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
438 cleanup
->size
= size
;
440 if (map
== MAP_FAILED
)
443 gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
446 cleanup
->gem_handle
= gem_handle
;
449 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
450 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
451 * always created as I915_CACHING_CACHED, which on non-LLC means
452 * snooped. That can be useful but comes with a bit of overheard. Since
453 * we're eplicitly clflushing and don't want the overhead we need to turn
455 if (!pool
->device
->info
.has_llc
) {
456 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_NONE
);
457 anv_gem_set_domain(pool
->device
, gem_handle
,
458 I915_GEM_DOMAIN_GTT
, I915_GEM_DOMAIN_GTT
);
462 /* Now that we successfull allocated everything, we can write the new
463 * values back into pool. */
464 pool
->map
= map
+ center_bo_offset
;
465 pool
->center_bo_offset
= center_bo_offset
;
466 pool
->bo
.gem_handle
= gem_handle
;
467 pool
->bo
.size
= size
;
472 pthread_mutex_unlock(&pool
->device
->mutex
);
474 /* Return the appropreate new size. This function never actually
475 * updates state->next. Instead, we let the caller do that because it
476 * needs to do so in order to maintain its concurrency model.
478 if (state
== &pool
->state
) {
479 return pool
->bo
.size
- pool
->center_bo_offset
;
481 assert(pool
->center_bo_offset
> 0);
482 return pool
->center_bo_offset
;
486 pthread_mutex_unlock(&pool
->device
->mutex
);
492 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
493 struct anv_block_state
*pool_state
)
495 struct anv_block_state state
, old
, new;
498 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, pool
->block_size
);
499 if (state
.next
< state
.end
) {
502 } else if (state
.next
== state
.end
) {
503 /* We allocated the first block outside the pool, we have to grow it.
504 * pool_state->next acts a mutex: threads who try to allocate now will
505 * get block indexes above the current limit and hit futex_wait
507 new.next
= state
.next
+ pool
->block_size
;
508 new.end
= anv_block_pool_grow(pool
, pool_state
);
509 assert(new.end
>= new.next
&& new.end
% pool
->block_size
== 0);
510 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
511 if (old
.next
!= state
.next
)
512 futex_wake(&pool_state
->end
, INT_MAX
);
515 futex_wait(&pool_state
->end
, state
.end
);
522 anv_block_pool_alloc(struct anv_block_pool
*pool
)
526 /* Try free list first. */
527 if (anv_free_list_pop(&pool
->free_list
, &pool
->map
, &offset
)) {
533 return anv_block_pool_alloc_new(pool
, &pool
->state
);
536 /* Allocates a block out of the back of the block pool.
538 * This will allocated a block earlier than the "start" of the block pool.
539 * The offsets returned from this function will be negative but will still
540 * be correct relative to the block pool's map pointer.
542 * If you ever use anv_block_pool_alloc_back, then you will have to do
543 * gymnastics with the block pool's BO when doing relocations.
546 anv_block_pool_alloc_back(struct anv_block_pool
*pool
)
550 /* Try free list first. */
551 if (anv_free_list_pop(&pool
->back_free_list
, &pool
->map
, &offset
)) {
557 offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
);
559 /* The offset we get out of anv_block_pool_alloc_new() is actually the
560 * number of bytes downwards from the middle to the end of the block.
561 * We need to turn it into a (negative) offset from the middle to the
562 * start of the block.
565 return -(offset
+ pool
->block_size
);
569 anv_block_pool_free(struct anv_block_pool
*pool
, int32_t offset
)
572 anv_free_list_push(&pool
->back_free_list
, pool
->map
, offset
);
574 anv_free_list_push(&pool
->free_list
, pool
->map
, offset
);
579 anv_fixed_size_state_pool_init(struct anv_fixed_size_state_pool
*pool
,
582 /* At least a cache line and must divide the block size. */
583 assert(state_size
>= 64 && util_is_power_of_two(state_size
));
585 pool
->state_size
= state_size
;
586 pool
->free_list
= ANV_FREE_LIST_EMPTY
;
587 pool
->block
.next
= 0;
592 anv_fixed_size_state_pool_alloc(struct anv_fixed_size_state_pool
*pool
,
593 struct anv_block_pool
*block_pool
)
596 struct anv_block_state block
, old
, new;
598 /* Try free list first. */
599 if (anv_free_list_pop(&pool
->free_list
, &block_pool
->map
, &offset
)) {
604 /* If free list was empty (or somebody raced us and took the items) we
605 * allocate a new item from the end of the block */
607 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, pool
->state_size
);
609 if (block
.next
< block
.end
) {
611 } else if (block
.next
== block
.end
) {
612 offset
= anv_block_pool_alloc(block_pool
);
613 new.next
= offset
+ pool
->state_size
;
614 new.end
= offset
+ block_pool
->block_size
;
615 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
616 if (old
.next
!= block
.next
)
617 futex_wake(&pool
->block
.end
, INT_MAX
);
620 futex_wait(&pool
->block
.end
, block
.end
);
626 anv_fixed_size_state_pool_free(struct anv_fixed_size_state_pool
*pool
,
627 struct anv_block_pool
*block_pool
,
630 anv_free_list_push(&pool
->free_list
, block_pool
->map
, offset
);
634 anv_state_pool_init(struct anv_state_pool
*pool
,
635 struct anv_block_pool
*block_pool
)
637 pool
->block_pool
= block_pool
;
638 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
639 size_t size
= 1 << (ANV_MIN_STATE_SIZE_LOG2
+ i
);
640 anv_fixed_size_state_pool_init(&pool
->buckets
[i
], size
);
642 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
646 anv_state_pool_finish(struct anv_state_pool
*pool
)
648 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
652 anv_state_pool_alloc(struct anv_state_pool
*pool
, size_t size
, size_t align
)
654 unsigned size_log2
= ilog2_round_up(size
< align
? align
: size
);
655 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
656 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
657 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
658 unsigned bucket
= size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
660 struct anv_state state
;
661 state
.alloc_size
= 1 << size_log2
;
662 state
.offset
= anv_fixed_size_state_pool_alloc(&pool
->buckets
[bucket
],
664 state
.map
= pool
->block_pool
->map
+ state
.offset
;
665 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
670 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
672 assert(util_is_power_of_two(state
.alloc_size
));
673 unsigned size_log2
= ilog2_round_up(state
.alloc_size
);
674 assert(size_log2
>= ANV_MIN_STATE_SIZE_LOG2
&&
675 size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
676 unsigned bucket
= size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
678 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
679 anv_fixed_size_state_pool_free(&pool
->buckets
[bucket
],
680 pool
->block_pool
, state
.offset
);
684 struct anv_state_stream_block
{
686 struct anv_state_stream_block
*next
;
688 /* The offset into the block pool at which this block starts */
692 /* A pointer to the first user-allocated thing in this block. This is
693 * what valgrind sees as the start of the block.
699 /* The state stream allocator is a one-shot, single threaded allocator for
700 * variable sized blocks. We use it for allocating dynamic state.
703 anv_state_stream_init(struct anv_state_stream
*stream
,
704 struct anv_block_pool
*block_pool
)
706 stream
->block_pool
= block_pool
;
707 stream
->block
= NULL
;
709 /* Ensure that next + whatever > end. This way the first call to
710 * state_stream_alloc fetches a new block.
715 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
719 anv_state_stream_finish(struct anv_state_stream
*stream
)
721 VG(const uint32_t block_size
= stream
->block_pool
->block_size
);
723 struct anv_state_stream_block
*next
= stream
->block
;
724 while (next
!= NULL
) {
725 VG(VALGRIND_MAKE_MEM_DEFINED(next
, sizeof(*next
)));
726 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
727 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
728 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, block_size
));
729 anv_block_pool_free(stream
->block_pool
, sb
.offset
);
733 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
737 anv_state_stream_alloc(struct anv_state_stream
*stream
,
738 uint32_t size
, uint32_t alignment
)
740 struct anv_state_stream_block
*sb
= stream
->block
;
742 struct anv_state state
;
744 state
.offset
= align_u32(stream
->next
, alignment
);
745 if (state
.offset
+ size
> stream
->end
) {
746 uint32_t block
= anv_block_pool_alloc(stream
->block_pool
);
747 sb
= stream
->block_pool
->map
+ block
;
749 VG(VALGRIND_MAKE_MEM_UNDEFINED(sb
, sizeof(*sb
)));
750 sb
->next
= stream
->block
;
752 VG(sb
->_vg_ptr
= NULL
);
753 VG(VALGRIND_MAKE_MEM_NOACCESS(sb
, stream
->block_pool
->block_size
));
756 stream
->start
= block
;
757 stream
->next
= block
+ sizeof(*sb
);
758 stream
->end
= block
+ stream
->block_pool
->block_size
;
760 state
.offset
= align_u32(stream
->next
, alignment
);
761 assert(state
.offset
+ size
<= stream
->end
);
764 assert(state
.offset
> stream
->start
);
765 state
.map
= (void *)sb
+ (state
.offset
- stream
->start
);
766 state
.alloc_size
= size
;
769 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
770 if (vg_ptr
== NULL
) {
772 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
773 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
775 void *state_end
= state
.map
+ state
.alloc_size
;
776 /* This only updates the mempool. The newly allocated chunk is still
777 * marked as NOACCESS. */
778 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
779 /* Mark the newly allocated chunk as undefined */
780 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
784 stream
->next
= state
.offset
+ size
;
789 struct bo_pool_bo_link
{
790 struct bo_pool_bo_link
*next
;
795 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
)
797 pool
->device
= device
;
798 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
800 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
804 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
806 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
807 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
808 while (link
!= NULL
) {
809 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
811 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
812 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
813 link
= link_copy
.next
;
817 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
821 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
825 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
826 const unsigned pow2_size
= 1 << size_log2
;
827 const unsigned bucket
= size_log2
- 12;
828 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
830 void *next_free_void
;
831 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
832 struct bo_pool_bo_link
*next_free
= next_free_void
;
833 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
834 assert(bo
->map
== next_free
);
835 assert(size
<= bo
->size
);
837 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
842 struct anv_bo new_bo
;
844 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
845 if (result
!= VK_SUCCESS
)
848 assert(new_bo
.size
== pow2_size
);
850 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
851 if (new_bo
.map
== NULL
) {
852 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
853 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
858 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
864 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
866 /* Make a copy in case the anv_bo happens to be storred in the BO */
867 struct anv_bo bo
= *bo_in
;
869 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
871 struct bo_pool_bo_link
*link
= bo
.map
;
872 VG_NOACCESS_WRITE(&link
->bo
, bo
);
874 assert(util_is_power_of_two(bo
.size
));
875 const unsigned size_log2
= ilog2_round_up(bo
.size
);
876 const unsigned bucket
= size_log2
- 12;
877 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
879 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
885 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
887 memset(pool
, 0, sizeof(*pool
));
891 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
893 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
894 for (unsigned i
= 0; i
< 16; i
++) {
895 struct anv_bo
*bo
= &pool
->bos
[i
][s
];
897 anv_gem_close(device
, bo
->gem_handle
);
903 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
904 gl_shader_stage stage
, unsigned per_thread_scratch
)
906 if (per_thread_scratch
== 0)
909 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
910 assert(scratch_size_log2
< 16);
912 struct anv_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
914 /* From now on, we go into a critical section. In order to remain
915 * thread-safe, we use the bo size as a lock. A value of 0 means we don't
916 * have a valid BO yet. A value of 1 means locked. A value greater than 1
917 * means we have a bo of the given size.
923 uint64_t size
= __sync_val_compare_and_swap(&bo
->size
, 0, 1);
925 /* We own the lock. Allocate a buffer */
927 const struct anv_physical_device
*physical_device
=
928 &device
->instance
->physicalDevice
;
929 const struct gen_device_info
*devinfo
= &physical_device
->info
;
931 /* WaCSScratchSize:hsw
933 * Haswell's scratch space address calculation appears to be sparse
934 * rather than tightly packed. The Thread ID has bits indicating which
935 * subslice, EU within a subslice, and thread within an EU it is.
936 * There's a maximum of two slices and two subslices, so these can be
937 * stored with a single bit. Even though there are only 10 EUs per
938 * subslice, this is stored in 4 bits, so there's an effective maximum
939 * value of 16 EUs. Similarly, although there are only 7 threads per EU,
940 * this is stored in a 3 bit number, giving an effective maximum value
941 * of 8 threads per EU.
943 * This means that we need to use 16 * 8 instead of 10 * 7 for the
944 * number of threads per subslice.
946 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
947 const unsigned scratch_ids_per_subslice
=
948 device
->info
.is_haswell
? 16 * 8 : devinfo
->max_cs_threads
;
950 uint32_t max_threads
[] = {
951 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
952 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
953 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
954 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
955 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
956 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
959 size
= per_thread_scratch
* max_threads
[stage
];
961 struct anv_bo new_bo
;
962 anv_bo_init_new(&new_bo
, device
, size
);
964 bo
->gem_handle
= new_bo
.gem_handle
;
966 /* Set the size last because we use it as a lock */
967 __sync_synchronize();
970 futex_wake((uint32_t *)&bo
->size
, INT_MAX
);
972 /* Someone else got here first */
973 while (bo
->size
== 1)
974 futex_wait((uint32_t *)&bo
->size
, 1);