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
,
183 uint32_t size
, uint32_t count
)
185 union anv_free_list current
, old
, new;
186 int32_t *next_ptr
= map
+ offset
;
188 /* If we're returning more than one chunk, we need to build a chain to add
189 * to the list. Fortunately, we can do this without any atomics since we
190 * own everything in the chain right now. `offset` is left pointing to the
191 * head of our chain list while `next_ptr` points to the tail.
193 for (uint32_t i
= 1; i
< count
; i
++) {
194 VG_NOACCESS_WRITE(next_ptr
, offset
+ i
* size
);
195 next_ptr
= map
+ offset
+ i
* size
;
201 VG_NOACCESS_WRITE(next_ptr
, current
.offset
);
203 new.count
= current
.count
+ 1;
204 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
205 } while (old
.u64
!= current
.u64
);
208 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
209 * means that the bottom 12 bits should all be zero.
211 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
212 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
213 #define PFL_PACK(ptr, count) ({ \
214 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
218 anv_ptr_free_list_pop(void **list
, void **elem
)
220 void *current
= *list
;
221 while (PFL_PTR(current
) != NULL
) {
222 void **next_ptr
= PFL_PTR(current
);
223 void *new_ptr
= VG_NOACCESS_READ(next_ptr
);
224 unsigned new_count
= PFL_COUNT(current
) + 1;
225 void *new = PFL_PACK(new_ptr
, new_count
);
226 void *old
= __sync_val_compare_and_swap(list
, current
, new);
227 if (old
== current
) {
228 *elem
= PFL_PTR(current
);
238 anv_ptr_free_list_push(void **list
, void *elem
)
241 void **next_ptr
= elem
;
243 /* The pointer-based free list requires that the pointer be
244 * page-aligned. This is because we use the bottom 12 bits of the
245 * pointer to store a counter to solve the ABA concurrency problem.
247 assert(((uintptr_t)elem
& 0xfff) == 0);
252 VG_NOACCESS_WRITE(next_ptr
, PFL_PTR(current
));
253 unsigned new_count
= PFL_COUNT(current
) + 1;
254 void *new = PFL_PACK(elem
, new_count
);
255 old
= __sync_val_compare_and_swap(list
, current
, new);
256 } while (old
!= current
);
260 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
261 uint32_t center_bo_offset
, uint32_t size
);
264 anv_block_pool_init(struct anv_block_pool
*pool
,
265 struct anv_device
*device
,
266 uint32_t initial_size
)
270 pool
->device
= device
;
271 anv_bo_init(&pool
->bo
, 0, 0);
273 pool
->fd
= memfd_create("block pool", MFD_CLOEXEC
);
275 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
277 /* Just make it 2GB up-front. The Linux kernel won't actually back it
278 * with pages until we either map and fault on one of them or we use
279 * userptr and send a chunk of it off to the GPU.
281 if (ftruncate(pool
->fd
, BLOCK_POOL_MEMFD_SIZE
) == -1) {
282 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
286 if (!u_vector_init(&pool
->mmap_cleanups
,
287 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
289 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
293 pool
->state
.next
= 0;
295 pool
->back_state
.next
= 0;
296 pool
->back_state
.end
= 0;
298 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
299 if (result
!= VK_SUCCESS
)
300 goto fail_mmap_cleanups
;
305 u_vector_finish(&pool
->mmap_cleanups
);
313 anv_block_pool_finish(struct anv_block_pool
*pool
)
315 struct anv_mmap_cleanup
*cleanup
;
317 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
) {
319 munmap(cleanup
->map
, cleanup
->size
);
320 if (cleanup
->gem_handle
)
321 anv_gem_close(pool
->device
, cleanup
->gem_handle
);
324 u_vector_finish(&pool
->mmap_cleanups
);
329 #define PAGE_SIZE 4096
332 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
333 uint32_t center_bo_offset
, uint32_t size
)
337 struct anv_mmap_cleanup
*cleanup
;
339 /* Assert that we only ever grow the pool */
340 assert(center_bo_offset
>= pool
->back_state
.end
);
341 assert(size
- center_bo_offset
>= pool
->state
.end
);
343 /* Assert that we don't go outside the bounds of the memfd */
344 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
345 assert(size
- center_bo_offset
<=
346 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
348 cleanup
= u_vector_add(&pool
->mmap_cleanups
);
350 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
352 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
354 /* Just leak the old map until we destroy the pool. We can't munmap it
355 * without races or imposing locking on the block allocate fast path. On
356 * the whole the leaked maps adds up to less than the size of the
357 * current map. MAP_POPULATE seems like the right thing to do, but we
358 * should try to get some numbers.
360 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
361 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
362 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
363 if (map
== MAP_FAILED
)
364 return vk_errorf(pool
->device
->instance
, pool
->device
,
365 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
367 gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
368 if (gem_handle
== 0) {
370 return vk_errorf(pool
->device
->instance
, pool
->device
,
371 VK_ERROR_TOO_MANY_OBJECTS
, "userptr failed: %m");
375 cleanup
->size
= size
;
376 cleanup
->gem_handle
= gem_handle
;
379 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
380 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
381 * always created as I915_CACHING_CACHED, which on non-LLC means
382 * snooped. That can be useful but comes with a bit of overheard. Since
383 * we're eplicitly clflushing and don't want the overhead we need to turn
385 if (!pool
->device
->info
.has_llc
) {
386 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_NONE
);
387 anv_gem_set_domain(pool
->device
, gem_handle
,
388 I915_GEM_DOMAIN_GTT
, I915_GEM_DOMAIN_GTT
);
392 /* Now that we successfull allocated everything, we can write the new
393 * values back into pool. */
394 pool
->map
= map
+ center_bo_offset
;
395 pool
->center_bo_offset
= center_bo_offset
;
397 /* For block pool BOs we have to be a bit careful about where we place them
398 * in the GTT. There are two documented workarounds for state base address
399 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
400 * which state that those two base addresses do not support 48-bit
401 * addresses and need to be placed in the bottom 32-bit range.
402 * Unfortunately, this is not quite accurate.
404 * The real problem is that we always set the size of our state pools in
405 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
406 * likely significantly smaller. We do this because we do not no at the
407 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
408 * the pool during command buffer building so we don't actually have a
409 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
410 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
411 * as being out of bounds and returns zero. For dynamic state, this
412 * usually just leads to rendering corruptions, but shaders that are all
413 * zero hang the GPU immediately.
415 * The easiest solution to do is exactly what the bogus workarounds say to
416 * do: restrict these buffers to 32-bit addresses. We could also pin the
417 * BO to some particular location of our choosing, but that's significantly
418 * more work than just not setting a flag. So, we explicitly DO NOT set
419 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
422 anv_bo_init(&pool
->bo
, gem_handle
, size
);
428 /** Grows and re-centers the block pool.
430 * We grow the block pool in one or both directions in such a way that the
431 * following conditions are met:
433 * 1) The size of the entire pool is always a power of two.
435 * 2) The pool only grows on both ends. Neither end can get
438 * 3) At the end of the allocation, we have about twice as much space
439 * allocated for each end as we have used. This way the pool doesn't
440 * grow too far in one direction or the other.
442 * 4) If the _alloc_back() has never been called, then the back portion of
443 * the pool retains a size of zero. (This makes it easier for users of
444 * the block pool that only want a one-sided pool.)
446 * 5) We have enough space allocated for at least one more block in
447 * whichever side `state` points to.
449 * 6) The center of the pool is always aligned to both the block_size of
450 * the pool and a 4K CPU page.
453 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
455 VkResult result
= VK_SUCCESS
;
457 pthread_mutex_lock(&pool
->device
->mutex
);
459 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
461 /* Gather a little usage information on the pool. Since we may have
462 * threadsd waiting in queue to get some storage while we resize, it's
463 * actually possible that total_used will be larger than old_size. In
464 * particular, block_pool_alloc() increments state->next prior to
465 * calling block_pool_grow, so this ensures that we get enough space for
466 * which ever side tries to grow the pool.
468 * We align to a page size because it makes it easier to do our
469 * calculations later in such a way that we state page-aigned.
471 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
472 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
473 uint32_t total_used
= front_used
+ back_used
;
475 assert(state
== &pool
->state
|| back_used
> 0);
477 uint32_t old_size
= pool
->bo
.size
;
479 /* The block pool is always initialized to a nonzero size and this function
480 * is always called after initialization.
482 assert(old_size
> 0);
484 /* The back_used and front_used may actually be smaller than the actual
485 * requirement because they are based on the next pointers which are
486 * updated prior to calling this function.
488 uint32_t back_required
= MAX2(back_used
, pool
->center_bo_offset
);
489 uint32_t front_required
= MAX2(front_used
, old_size
- pool
->center_bo_offset
);
491 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
492 /* If we're in this case then this isn't the firsta allocation and we
493 * already have enough space on both sides to hold double what we
494 * have allocated. There's nothing for us to do.
499 uint32_t size
= old_size
* 2;
500 while (size
< back_required
+ front_required
)
503 assert(size
> pool
->bo
.size
);
505 /* We compute a new center_bo_offset such that, when we double the size
506 * of the pool, we maintain the ratio of how much is used by each side.
507 * This way things should remain more-or-less balanced.
509 uint32_t center_bo_offset
;
510 if (back_used
== 0) {
511 /* If we're in this case then we have never called alloc_back(). In
512 * this case, we want keep the offset at 0 to make things as simple
513 * as possible for users that don't care about back allocations.
515 center_bo_offset
= 0;
517 /* Try to "center" the allocation based on how much is currently in
518 * use on each side of the center line.
520 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
522 /* Align down to a multiple of the page size */
523 center_bo_offset
&= ~(PAGE_SIZE
- 1);
525 assert(center_bo_offset
>= back_used
);
527 /* Make sure we don't shrink the back end of the pool */
528 if (center_bo_offset
< pool
->back_state
.end
)
529 center_bo_offset
= pool
->back_state
.end
;
531 /* Make sure that we don't shrink the front end of the pool */
532 if (size
- center_bo_offset
< pool
->state
.end
)
533 center_bo_offset
= size
- pool
->state
.end
;
536 assert(center_bo_offset
% PAGE_SIZE
== 0);
538 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
540 if (pool
->device
->instance
->physicalDevice
.has_exec_async
)
541 pool
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
544 pthread_mutex_unlock(&pool
->device
->mutex
);
546 if (result
== VK_SUCCESS
) {
547 /* Return the appropriate new size. This function never actually
548 * updates state->next. Instead, we let the caller do that because it
549 * needs to do so in order to maintain its concurrency model.
551 if (state
== &pool
->state
) {
552 return pool
->bo
.size
- pool
->center_bo_offset
;
554 assert(pool
->center_bo_offset
> 0);
555 return pool
->center_bo_offset
;
563 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
564 struct anv_block_state
*pool_state
,
567 struct anv_block_state state
, old
, new;
570 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
571 if (state
.next
+ block_size
<= state
.end
) {
574 } else if (state
.next
<= state
.end
) {
575 /* We allocated the first block outside the pool so we have to grow
576 * the pool. pool_state->next acts a mutex: threads who try to
577 * allocate now will get block indexes above the current limit and
578 * hit futex_wait below.
580 new.next
= state
.next
+ block_size
;
582 new.end
= anv_block_pool_grow(pool
, pool_state
);
583 } while (new.end
< new.next
);
585 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
586 if (old
.next
!= state
.next
)
587 futex_wake(&pool_state
->end
, INT_MAX
);
590 futex_wait(&pool_state
->end
, state
.end
);
597 anv_block_pool_alloc(struct anv_block_pool
*pool
,
600 return anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
);
603 /* Allocates a block out of the back of the block pool.
605 * This will allocated a block earlier than the "start" of the block pool.
606 * The offsets returned from this function will be negative but will still
607 * be correct relative to the block pool's map pointer.
609 * If you ever use anv_block_pool_alloc_back, then you will have to do
610 * gymnastics with the block pool's BO when doing relocations.
613 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
616 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
619 /* The offset we get out of anv_block_pool_alloc_new() is actually the
620 * number of bytes downwards from the middle to the end of the block.
621 * We need to turn it into a (negative) offset from the middle to the
622 * start of the block.
625 return -(offset
+ block_size
);
629 anv_state_pool_init(struct anv_state_pool
*pool
,
630 struct anv_device
*device
,
633 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
635 if (result
!= VK_SUCCESS
)
638 assert(util_is_power_of_two(block_size
));
639 pool
->block_size
= block_size
;
640 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
641 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
642 pool
->buckets
[i
].free_list
= ANV_FREE_LIST_EMPTY
;
643 pool
->buckets
[i
].block
.next
= 0;
644 pool
->buckets
[i
].block
.end
= 0;
646 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
652 anv_state_pool_finish(struct anv_state_pool
*pool
)
654 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
655 anv_block_pool_finish(&pool
->block_pool
);
659 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
660 struct anv_block_pool
*block_pool
,
664 struct anv_block_state block
, old
, new;
667 /* If our state is large, we don't need any sub-allocation from a block.
668 * Instead, we just grab whole (potentially large) blocks.
670 if (state_size
>= block_size
)
671 return anv_block_pool_alloc(block_pool
, state_size
);
674 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
676 if (block
.next
< block
.end
) {
678 } else if (block
.next
== block
.end
) {
679 offset
= anv_block_pool_alloc(block_pool
, block_size
);
680 new.next
= offset
+ state_size
;
681 new.end
= offset
+ block_size
;
682 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
683 if (old
.next
!= block
.next
)
684 futex_wake(&pool
->block
.end
, INT_MAX
);
687 futex_wait(&pool
->block
.end
, block
.end
);
693 anv_state_pool_get_bucket(uint32_t size
)
695 unsigned size_log2
= ilog2_round_up(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 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
703 anv_state_pool_get_bucket_size(uint32_t bucket
)
705 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
706 return 1 << size_log2
;
709 static struct anv_state
710 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
711 uint32_t size
, uint32_t align
)
713 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
715 struct anv_state state
;
716 state
.alloc_size
= anv_state_pool_get_bucket_size(bucket
);
718 /* Try free list first. */
719 if (anv_free_list_pop(&pool
->buckets
[bucket
].free_list
,
720 &pool
->block_pool
.map
, &state
.offset
)) {
721 assert(state
.offset
>= 0);
725 /* Try to grab a chunk from some larger bucket and split it up */
726 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
727 int32_t chunk_offset
;
728 if (anv_free_list_pop(&pool
->buckets
[b
].free_list
,
729 &pool
->block_pool
.map
, &chunk_offset
)) {
730 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
732 /* We've found a chunk that's larger than the requested state size.
733 * There are a couple of options as to what we do with it:
735 * 1) We could fully split the chunk into state.alloc_size sized
736 * pieces. However, this would mean that allocating a 16B
737 * state could potentially split a 2MB chunk into 512K smaller
738 * chunks. This would lead to unnecessary fragmentation.
740 * 2) The classic "buddy allocator" method would have us split the
741 * chunk in half and return one half. Then we would split the
742 * remaining half in half and return one half, and repeat as
743 * needed until we get down to the size we want. However, if
744 * you are allocating a bunch of the same size state (which is
745 * the common case), this means that every other allocation has
746 * to go up a level and every fourth goes up two levels, etc.
747 * This is not nearly as efficient as it could be if we did a
748 * little more work up-front.
750 * 3) Split the difference between (1) and (2) by doing a
751 * two-level split. If it's bigger than some fixed block_size,
752 * we split it into block_size sized chunks and return all but
753 * one of them. Then we split what remains into
754 * state.alloc_size sized chunks and return all but one.
756 * We choose option (3).
758 if (chunk_size
> pool
->block_size
&&
759 state
.alloc_size
< pool
->block_size
) {
760 assert(chunk_size
% pool
->block_size
== 0);
761 /* We don't want to split giant chunks into tiny chunks. Instead,
762 * break anything bigger than a block into block-sized chunks and
763 * then break it down into bucket-sized chunks from there. Return
764 * all but the first block of the chunk to the block bucket.
766 const uint32_t block_bucket
=
767 anv_state_pool_get_bucket(pool
->block_size
);
768 anv_free_list_push(&pool
->buckets
[block_bucket
].free_list
,
769 pool
->block_pool
.map
,
770 chunk_offset
+ pool
->block_size
,
772 (chunk_size
/ pool
->block_size
) - 1);
773 chunk_size
= pool
->block_size
;
776 assert(chunk_size
% state
.alloc_size
== 0);
777 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
778 pool
->block_pool
.map
,
779 chunk_offset
+ state
.alloc_size
,
781 (chunk_size
/ state
.alloc_size
) - 1);
783 state
.offset
= chunk_offset
;
788 state
.offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
794 state
.map
= pool
->block_pool
.map
+ state
.offset
;
799 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
802 return ANV_STATE_NULL
;
804 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
805 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
810 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
812 struct anv_state state
;
813 state
.alloc_size
= pool
->block_size
;
815 if (anv_free_list_pop(&pool
->back_alloc_free_list
,
816 &pool
->block_pool
.map
, &state
.offset
)) {
817 assert(state
.offset
< 0);
821 state
.offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
825 state
.map
= pool
->block_pool
.map
+ state
.offset
;
826 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, state
.alloc_size
));
831 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
833 assert(util_is_power_of_two(state
.alloc_size
));
834 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
836 if (state
.offset
< 0) {
837 assert(state
.alloc_size
== pool
->block_size
);
838 anv_free_list_push(&pool
->back_alloc_free_list
,
839 pool
->block_pool
.map
, state
.offset
,
840 state
.alloc_size
, 1);
842 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
843 pool
->block_pool
.map
, state
.offset
,
844 state
.alloc_size
, 1);
849 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
851 if (state
.alloc_size
== 0)
854 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
855 anv_state_pool_free_no_vg(pool
, state
);
858 struct anv_state_stream_block
{
859 struct anv_state block
;
862 struct anv_state_stream_block
*next
;
865 /* A pointer to the first user-allocated thing in this block. This is
866 * what valgrind sees as the start of the block.
872 /* The state stream allocator is a one-shot, single threaded allocator for
873 * variable sized blocks. We use it for allocating dynamic state.
876 anv_state_stream_init(struct anv_state_stream
*stream
,
877 struct anv_state_pool
*state_pool
,
880 stream
->state_pool
= state_pool
;
881 stream
->block_size
= block_size
;
883 stream
->block
= ANV_STATE_NULL
;
885 stream
->block_list
= NULL
;
887 /* Ensure that next + whatever > block_size. This way the first call to
888 * state_stream_alloc fetches a new block.
890 stream
->next
= block_size
;
892 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
896 anv_state_stream_finish(struct anv_state_stream
*stream
)
898 struct anv_state_stream_block
*next
= stream
->block_list
;
899 while (next
!= NULL
) {
900 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
901 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
902 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
903 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
907 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
911 anv_state_stream_alloc(struct anv_state_stream
*stream
,
912 uint32_t size
, uint32_t alignment
)
915 return ANV_STATE_NULL
;
917 assert(alignment
<= PAGE_SIZE
);
919 uint32_t offset
= align_u32(stream
->next
, alignment
);
920 if (offset
+ size
> stream
->block
.alloc_size
) {
921 uint32_t block_size
= stream
->block_size
;
922 if (block_size
< size
)
923 block_size
= round_to_power_of_two(size
);
925 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
926 block_size
, PAGE_SIZE
);
928 struct anv_state_stream_block
*sb
= stream
->block
.map
;
929 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
930 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
931 stream
->block_list
= sb
;
932 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
934 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
936 /* Reset back to the start plus space for the header */
937 stream
->next
= sizeof(*sb
);
939 offset
= align_u32(stream
->next
, alignment
);
940 assert(offset
+ size
<= stream
->block
.alloc_size
);
943 struct anv_state state
= stream
->block
;
944 state
.offset
+= offset
;
945 state
.alloc_size
= size
;
948 stream
->next
= offset
+ size
;
951 struct anv_state_stream_block
*sb
= stream
->block_list
;
952 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
953 if (vg_ptr
== NULL
) {
955 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
956 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
958 void *state_end
= state
.map
+ state
.alloc_size
;
959 /* This only updates the mempool. The newly allocated chunk is still
960 * marked as NOACCESS. */
961 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
962 /* Mark the newly allocated chunk as undefined */
963 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
970 struct bo_pool_bo_link
{
971 struct bo_pool_bo_link
*next
;
976 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
)
978 pool
->device
= device
;
979 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
981 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
985 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
987 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
988 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
989 while (link
!= NULL
) {
990 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
992 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
993 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
994 link
= link_copy
.next
;
998 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1002 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
1006 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1007 const unsigned pow2_size
= 1 << size_log2
;
1008 const unsigned bucket
= size_log2
- 12;
1009 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1011 void *next_free_void
;
1012 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
1013 struct bo_pool_bo_link
*next_free
= next_free_void
;
1014 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
1015 assert(bo
->gem_handle
);
1016 assert(bo
->map
== next_free
);
1017 assert(size
<= bo
->size
);
1019 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1024 struct anv_bo new_bo
;
1026 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
1027 if (result
!= VK_SUCCESS
)
1030 if (pool
->device
->instance
->physicalDevice
.supports_48bit_addresses
)
1031 new_bo
.flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1033 if (pool
->device
->instance
->physicalDevice
.has_exec_async
)
1034 new_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1036 assert(new_bo
.size
== pow2_size
);
1038 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
1039 if (new_bo
.map
== MAP_FAILED
) {
1040 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
1041 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1046 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1052 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
1054 /* Make a copy in case the anv_bo happens to be storred in the BO */
1055 struct anv_bo bo
= *bo_in
;
1057 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
1059 struct bo_pool_bo_link
*link
= bo
.map
;
1060 VG_NOACCESS_WRITE(&link
->bo
, bo
);
1062 assert(util_is_power_of_two(bo
.size
));
1063 const unsigned size_log2
= ilog2_round_up(bo
.size
);
1064 const unsigned bucket
= size_log2
- 12;
1065 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1067 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
1073 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1075 memset(pool
, 0, sizeof(*pool
));
1079 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1081 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1082 for (unsigned i
= 0; i
< 16; i
++) {
1083 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
1085 anv_gem_close(device
, bo
->bo
.gem_handle
);
1091 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1092 gl_shader_stage stage
, unsigned per_thread_scratch
)
1094 if (per_thread_scratch
== 0)
1097 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1098 assert(scratch_size_log2
< 16);
1100 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
1102 /* We can use "exists" to shortcut and ignore the critical section */
1106 pthread_mutex_lock(&device
->mutex
);
1108 __sync_synchronize();
1112 const struct anv_physical_device
*physical_device
=
1113 &device
->instance
->physicalDevice
;
1114 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1116 /* WaCSScratchSize:hsw
1118 * Haswell's scratch space address calculation appears to be sparse
1119 * rather than tightly packed. The Thread ID has bits indicating which
1120 * subslice, EU within a subslice, and thread within an EU it is.
1121 * There's a maximum of two slices and two subslices, so these can be
1122 * stored with a single bit. Even though there are only 10 EUs per
1123 * subslice, this is stored in 4 bits, so there's an effective maximum
1124 * value of 16 EUs. Similarly, although there are only 7 threads per EU,
1125 * this is stored in a 3 bit number, giving an effective maximum value
1126 * of 8 threads per EU.
1128 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1129 * number of threads per subslice.
1131 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
1132 const unsigned scratch_ids_per_subslice
=
1133 device
->info
.is_haswell
? 16 * 8 : devinfo
->max_cs_threads
;
1135 uint32_t max_threads
[] = {
1136 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1137 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1138 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1139 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1140 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1141 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1144 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1146 anv_bo_init_new(&bo
->bo
, device
, size
);
1148 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1149 * are still relative to the general state base address. When we emit
1150 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1151 * to the maximum (1 page under 4GB). This allows us to just place the
1152 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1153 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1154 * However, in order to do so, we need to ensure that the kernel does not
1155 * place the scratch BO above the 32-bit boundary.
1157 * NOTE: Technically, it can't go "anywhere" because the top page is off
1158 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1159 * kernel allocates space using
1161 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1163 * so nothing will ever touch the top page.
1165 assert(!(bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
));
1167 if (device
->instance
->physicalDevice
.has_exec_async
)
1168 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1170 /* Set the exists last because it may be read by other threads */
1171 __sync_synchronize();
1174 pthread_mutex_unlock(&device
->mutex
);
1179 struct anv_cached_bo
{
1186 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1188 cache
->bo_map
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
1189 _mesa_key_pointer_equal
);
1191 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1193 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1194 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1195 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1196 "pthread_mutex_init failed: %m");
1203 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1205 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1206 pthread_mutex_destroy(&cache
->mutex
);
1209 static struct anv_cached_bo
*
1210 anv_bo_cache_lookup_locked(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1212 struct hash_entry
*entry
=
1213 _mesa_hash_table_search(cache
->bo_map
,
1214 (const void *)(uintptr_t)gem_handle
);
1218 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)entry
->data
;
1219 assert(bo
->bo
.gem_handle
== gem_handle
);
1224 UNUSED
static struct anv_bo
*
1225 anv_bo_cache_lookup(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1227 pthread_mutex_lock(&cache
->mutex
);
1229 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1231 pthread_mutex_unlock(&cache
->mutex
);
1233 return bo
? &bo
->bo
: NULL
;
1237 anv_bo_cache_alloc(struct anv_device
*device
,
1238 struct anv_bo_cache
*cache
,
1239 uint64_t size
, struct anv_bo
**bo_out
)
1241 struct anv_cached_bo
*bo
=
1242 vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1243 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1245 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1249 /* The kernel is going to give us whole pages anyway */
1250 size
= align_u64(size
, 4096);
1252 VkResult result
= anv_bo_init_new(&bo
->bo
, device
, size
);
1253 if (result
!= VK_SUCCESS
) {
1254 vk_free(&device
->alloc
, bo
);
1258 assert(bo
->bo
.gem_handle
);
1260 pthread_mutex_lock(&cache
->mutex
);
1262 _mesa_hash_table_insert(cache
->bo_map
,
1263 (void *)(uintptr_t)bo
->bo
.gem_handle
, bo
);
1265 pthread_mutex_unlock(&cache
->mutex
);
1273 anv_bo_cache_import(struct anv_device
*device
,
1274 struct anv_bo_cache
*cache
,
1275 int fd
, struct anv_bo
**bo_out
)
1277 pthread_mutex_lock(&cache
->mutex
);
1279 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1281 pthread_mutex_unlock(&cache
->mutex
);
1282 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
1285 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1287 __sync_fetch_and_add(&bo
->refcount
, 1);
1289 off_t size
= lseek(fd
, 0, SEEK_END
);
1290 if (size
== (off_t
)-1) {
1291 anv_gem_close(device
, gem_handle
);
1292 pthread_mutex_unlock(&cache
->mutex
);
1293 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
1296 bo
= vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1297 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1299 anv_gem_close(device
, gem_handle
);
1300 pthread_mutex_unlock(&cache
->mutex
);
1301 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1306 anv_bo_init(&bo
->bo
, gem_handle
, size
);
1308 _mesa_hash_table_insert(cache
->bo_map
, (void *)(uintptr_t)gem_handle
, bo
);
1311 pthread_mutex_unlock(&cache
->mutex
);
1318 anv_bo_cache_export(struct anv_device
*device
,
1319 struct anv_bo_cache
*cache
,
1320 struct anv_bo
*bo_in
, int *fd_out
)
1322 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1323 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1325 int fd
= anv_gem_handle_to_fd(device
, bo
->bo
.gem_handle
);
1327 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1335 atomic_dec_not_one(uint32_t *counter
)
1344 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1353 anv_bo_cache_release(struct anv_device
*device
,
1354 struct anv_bo_cache
*cache
,
1355 struct anv_bo
*bo_in
)
1357 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1358 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1360 /* Try to decrement the counter but don't go below one. If this succeeds
1361 * then the refcount has been decremented and we are not the last
1364 if (atomic_dec_not_one(&bo
->refcount
))
1367 pthread_mutex_lock(&cache
->mutex
);
1369 /* We are probably the last reference since our attempt to decrement above
1370 * failed. However, we can't actually know until we are inside the mutex.
1371 * Otherwise, someone could import the BO between the decrement and our
1374 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1375 /* Turns out we're not the last reference. Unlock and bail. */
1376 pthread_mutex_unlock(&cache
->mutex
);
1380 struct hash_entry
*entry
=
1381 _mesa_hash_table_search(cache
->bo_map
,
1382 (const void *)(uintptr_t)bo
->bo
.gem_handle
);
1384 _mesa_hash_table_remove(cache
->bo_map
, entry
);
1387 anv_gem_munmap(bo
->bo
.map
, bo
->bo
.size
);
1389 anv_gem_close(device
, bo
->bo
.gem_handle
);
1391 /* Don't unlock until we've actually closed the BO. The whole point of
1392 * the BO cache is to ensure that we correctly handle races with creating
1393 * and releasing GEM handles and we don't want to let someone import the BO
1394 * again between mutex unlock and closing the GEM handle.
1396 pthread_mutex_unlock(&cache
->mutex
);
1398 vk_free(&device
->alloc
, bo
);