2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
30 #include "anv_private.h"
32 #include "util/simple_mtx.h"
33 #include "util/anon_file.h"
36 #define VG_NOACCESS_READ(__ptr) ({ \
37 VALGRIND_MAKE_MEM_DEFINED((__ptr), sizeof(*(__ptr))); \
38 __typeof(*(__ptr)) __val = *(__ptr); \
39 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr)));\
42 #define VG_NOACCESS_WRITE(__ptr, __val) ({ \
43 VALGRIND_MAKE_MEM_UNDEFINED((__ptr), sizeof(*(__ptr))); \
45 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr))); \
48 #define VG_NOACCESS_READ(__ptr) (*(__ptr))
49 #define VG_NOACCESS_WRITE(__ptr, __val) (*(__ptr) = (__val))
53 #define MAP_POPULATE 0
58 * - Lock free (except when resizing underlying bos)
60 * - Constant time allocation with typically only one atomic
62 * - Multiple allocation sizes without fragmentation
64 * - Can grow while keeping addresses and offset of contents stable
66 * - All allocations within one bo so we can point one of the
67 * STATE_BASE_ADDRESS pointers at it.
69 * The overall design is a two-level allocator: top level is a fixed size, big
70 * block (8k) allocator, which operates out of a bo. Allocation is done by
71 * either pulling a block from the free list or growing the used range of the
72 * bo. Growing the range may run out of space in the bo which we then need to
73 * grow. Growing the bo is tricky in a multi-threaded, lockless environment:
74 * we need to keep all pointers and contents in the old map valid. GEM bos in
75 * general can't grow, but we use a trick: we create a memfd and use ftruncate
76 * to grow it as necessary. We mmap the new size and then create a gem bo for
77 * it using the new gem userptr ioctl. Without heavy-handed locking around
78 * our allocation fast-path, there isn't really a way to munmap the old mmap,
79 * so we just keep it around until garbage collection time. While the block
80 * allocator is lockless for normal operations, we block other threads trying
81 * to allocate while we're growing the map. It sholdn't happen often, and
82 * growing is fast anyway.
84 * At the next level we can use various sub-allocators. The state pool is a
85 * pool of smaller, fixed size objects, which operates much like the block
86 * pool. It uses a free list for freeing objects, but when it runs out of
87 * space it just allocates a new block from the block pool. This allocator is
88 * intended for longer lived state objects such as SURFACE_STATE and most
89 * other persistent state objects in the API. We may need to track more info
90 * with these object and a pointer back to the CPU object (eg VkImage). In
91 * those cases we just allocate a slightly bigger object and put the extra
92 * state after the GPU state object.
94 * The state stream allocator works similar to how the i965 DRI driver streams
95 * all its state. Even with Vulkan, we need to emit transient state (whether
96 * surface state base or dynamic state base), and for that we can just get a
97 * block and fill it up. These cases are local to a command buffer and the
98 * sub-allocator need not be thread safe. The streaming allocator gets a new
99 * block when it runs out of space and chains them together so they can be
103 /* Allocations are always at least 64 byte aligned, so 1 is an invalid value.
104 * We use it to indicate the free list is empty. */
105 #define EMPTY UINT32_MAX
107 #define PAGE_SIZE 4096
109 struct anv_mmap_cleanup
{
114 static inline uint32_t
115 ilog2_round_up(uint32_t value
)
118 return 32 - __builtin_clz(value
- 1);
121 static inline uint32_t
122 round_to_power_of_two(uint32_t value
)
124 return 1 << ilog2_round_up(value
);
127 struct anv_state_table_cleanup
{
132 #define ANV_STATE_TABLE_CLEANUP_INIT ((struct anv_state_table_cleanup){0})
133 #define ANV_STATE_ENTRY_SIZE (sizeof(struct anv_free_entry))
136 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
);
139 anv_state_table_init(struct anv_state_table
*table
,
140 struct anv_device
*device
,
141 uint32_t initial_entries
)
145 table
->device
= device
;
147 /* Just make it 2GB up-front. The Linux kernel won't actually back it
148 * with pages until we either map and fault on one of them or we use
149 * userptr and send a chunk of it off to the GPU.
151 table
->fd
= os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE
, "state table");
152 if (table
->fd
== -1) {
153 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
157 if (!u_vector_init(&table
->cleanups
,
158 round_to_power_of_two(sizeof(struct anv_state_table_cleanup
)),
160 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
164 table
->state
.next
= 0;
165 table
->state
.end
= 0;
168 uint32_t initial_size
= initial_entries
* ANV_STATE_ENTRY_SIZE
;
169 result
= anv_state_table_expand_range(table
, initial_size
);
170 if (result
!= VK_SUCCESS
)
176 u_vector_finish(&table
->cleanups
);
184 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
)
187 struct anv_state_table_cleanup
*cleanup
;
189 /* Assert that we only ever grow the pool */
190 assert(size
>= table
->state
.end
);
192 /* Make sure that we don't go outside the bounds of the memfd */
193 if (size
> BLOCK_POOL_MEMFD_SIZE
)
194 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
196 cleanup
= u_vector_add(&table
->cleanups
);
198 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
200 *cleanup
= ANV_STATE_TABLE_CLEANUP_INIT
;
202 /* Just leak the old map until we destroy the pool. We can't munmap it
203 * without races or imposing locking on the block allocate fast path. On
204 * the whole the leaked maps adds up to less than the size of the
205 * current map. MAP_POPULATE seems like the right thing to do, but we
206 * should try to get some numbers.
208 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
209 MAP_SHARED
| MAP_POPULATE
, table
->fd
, 0);
210 if (map
== MAP_FAILED
) {
211 return vk_errorf(table
->device
->instance
, table
->device
,
212 VK_ERROR_OUT_OF_HOST_MEMORY
, "mmap failed: %m");
216 cleanup
->size
= size
;
225 anv_state_table_grow(struct anv_state_table
*table
)
227 VkResult result
= VK_SUCCESS
;
229 uint32_t used
= align_u32(table
->state
.next
* ANV_STATE_ENTRY_SIZE
,
231 uint32_t old_size
= table
->size
;
233 /* The block pool is always initialized to a nonzero size and this function
234 * is always called after initialization.
236 assert(old_size
> 0);
238 uint32_t required
= MAX2(used
, old_size
);
239 if (used
* 2 <= required
) {
240 /* If we're in this case then this isn't the firsta allocation and we
241 * already have enough space on both sides to hold double what we
242 * have allocated. There's nothing for us to do.
247 uint32_t size
= old_size
* 2;
248 while (size
< required
)
251 assert(size
> table
->size
);
253 result
= anv_state_table_expand_range(table
, size
);
260 anv_state_table_finish(struct anv_state_table
*table
)
262 struct anv_state_table_cleanup
*cleanup
;
264 u_vector_foreach(cleanup
, &table
->cleanups
) {
266 munmap(cleanup
->map
, cleanup
->size
);
269 u_vector_finish(&table
->cleanups
);
275 anv_state_table_add(struct anv_state_table
*table
, uint32_t *idx
,
278 struct anv_block_state state
, old
, new;
284 state
.u64
= __sync_fetch_and_add(&table
->state
.u64
, count
);
285 if (state
.next
+ count
<= state
.end
) {
287 struct anv_free_entry
*entry
= &table
->map
[state
.next
];
288 for (int i
= 0; i
< count
; i
++) {
289 entry
[i
].state
.idx
= state
.next
+ i
;
293 } else if (state
.next
<= state
.end
) {
294 /* We allocated the first block outside the pool so we have to grow
295 * the pool. pool_state->next acts a mutex: threads who try to
296 * allocate now will get block indexes above the current limit and
297 * hit futex_wait below.
299 new.next
= state
.next
+ count
;
301 result
= anv_state_table_grow(table
);
302 if (result
!= VK_SUCCESS
)
304 new.end
= table
->size
/ ANV_STATE_ENTRY_SIZE
;
305 } while (new.end
< new.next
);
307 old
.u64
= __sync_lock_test_and_set(&table
->state
.u64
, new.u64
);
308 if (old
.next
!= state
.next
)
309 futex_wake(&table
->state
.end
, INT_MAX
);
311 futex_wait(&table
->state
.end
, state
.end
, NULL
);
318 anv_free_list_push(union anv_free_list
*list
,
319 struct anv_state_table
*table
,
320 uint32_t first
, uint32_t count
)
322 union anv_free_list current
, old
, new;
323 uint32_t last
= first
;
325 for (uint32_t i
= 1; i
< count
; i
++, last
++)
326 table
->map
[last
].next
= last
+ 1;
331 table
->map
[last
].next
= current
.offset
;
333 new.count
= current
.count
+ 1;
334 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
335 } while (old
.u64
!= current
.u64
);
339 anv_free_list_pop(union anv_free_list
*list
,
340 struct anv_state_table
*table
)
342 union anv_free_list current
, new, old
;
344 current
.u64
= list
->u64
;
345 while (current
.offset
!= EMPTY
) {
346 __sync_synchronize();
347 new.offset
= table
->map
[current
.offset
].next
;
348 new.count
= current
.count
+ 1;
349 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
350 if (old
.u64
== current
.u64
) {
351 struct anv_free_entry
*entry
= &table
->map
[current
.offset
];
352 return &entry
->state
;
360 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
361 * means that the bottom 12 bits should all be zero.
363 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
364 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
365 #define PFL_PACK(ptr, count) ({ \
366 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
370 anv_ptr_free_list_pop(void **list
, void **elem
)
372 void *current
= *list
;
373 while (PFL_PTR(current
) != NULL
) {
374 void **next_ptr
= PFL_PTR(current
);
375 void *new_ptr
= VG_NOACCESS_READ(next_ptr
);
376 unsigned new_count
= PFL_COUNT(current
) + 1;
377 void *new = PFL_PACK(new_ptr
, new_count
);
378 void *old
= __sync_val_compare_and_swap(list
, current
, new);
379 if (old
== current
) {
380 *elem
= PFL_PTR(current
);
390 anv_ptr_free_list_push(void **list
, void *elem
)
393 void **next_ptr
= elem
;
395 /* The pointer-based free list requires that the pointer be
396 * page-aligned. This is because we use the bottom 12 bits of the
397 * pointer to store a counter to solve the ABA concurrency problem.
399 assert(((uintptr_t)elem
& 0xfff) == 0);
404 VG_NOACCESS_WRITE(next_ptr
, PFL_PTR(current
));
405 unsigned new_count
= PFL_COUNT(current
) + 1;
406 void *new = PFL_PACK(elem
, new_count
);
407 old
= __sync_val_compare_and_swap(list
, current
, new);
408 } while (old
!= current
);
412 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
413 uint32_t center_bo_offset
, uint32_t size
);
416 anv_block_pool_init(struct anv_block_pool
*pool
,
417 struct anv_device
*device
,
418 uint64_t start_address
,
419 uint32_t initial_size
)
423 pool
->device
= device
;
424 pool
->use_softpin
= device
->instance
->physicalDevice
.use_softpin
;
427 pool
->center_bo_offset
= 0;
428 pool
->start_address
= gen_canonical_address(start_address
);
431 if (pool
->use_softpin
) {
432 /* This pointer will always point to the first BO in the list */
433 anv_bo_init(&pool
->bos
[0], 0, 0);
434 pool
->bo
= &pool
->bos
[0];
438 /* Just make it 2GB up-front. The Linux kernel won't actually back it
439 * with pages until we either map and fault on one of them or we use
440 * userptr and send a chunk of it off to the GPU.
442 pool
->fd
= os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE
, "block pool");
444 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
446 anv_bo_init(&pool
->wrapper_bo
, 0, 0);
447 pool
->wrapper_bo
.is_wrapper
= true;
448 pool
->bo
= &pool
->wrapper_bo
;
451 if (!u_vector_init(&pool
->mmap_cleanups
,
452 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
454 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
458 pool
->state
.next
= 0;
460 pool
->back_state
.next
= 0;
461 pool
->back_state
.end
= 0;
463 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
464 if (result
!= VK_SUCCESS
)
465 goto fail_mmap_cleanups
;
467 /* Make the entire pool available in the front of the pool. If back
468 * allocation needs to use this space, the "ends" will be re-arranged.
470 pool
->state
.end
= pool
->size
;
475 u_vector_finish(&pool
->mmap_cleanups
);
484 anv_block_pool_finish(struct anv_block_pool
*pool
)
486 anv_block_pool_foreach_bo(bo
, pool
) {
488 anv_gem_munmap(bo
->map
, bo
->size
);
489 anv_gem_close(pool
->device
, bo
->gem_handle
);
492 struct anv_mmap_cleanup
*cleanup
;
493 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
)
494 munmap(cleanup
->map
, cleanup
->size
);
495 u_vector_finish(&pool
->mmap_cleanups
);
502 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
503 uint32_t center_bo_offset
, uint32_t size
)
505 /* Assert that we only ever grow the pool */
506 assert(center_bo_offset
>= pool
->back_state
.end
);
507 assert(size
- center_bo_offset
>= pool
->state
.end
);
509 /* Assert that we don't go outside the bounds of the memfd */
510 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
511 assert(pool
->use_softpin
||
512 size
- center_bo_offset
<=
513 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
515 /* For state pool BOs we have to be a bit careful about where we place them
516 * in the GTT. There are two documented workarounds for state base address
517 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
518 * which state that those two base addresses do not support 48-bit
519 * addresses and need to be placed in the bottom 32-bit range.
520 * Unfortunately, this is not quite accurate.
522 * The real problem is that we always set the size of our state pools in
523 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
524 * likely significantly smaller. We do this because we do not no at the
525 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
526 * the pool during command buffer building so we don't actually have a
527 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
528 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
529 * as being out of bounds and returns zero. For dynamic state, this
530 * usually just leads to rendering corruptions, but shaders that are all
531 * zero hang the GPU immediately.
533 * The easiest solution to do is exactly what the bogus workarounds say to
534 * do: restrict these buffers to 32-bit addresses. We could also pin the
535 * BO to some particular location of our choosing, but that's significantly
536 * more work than just not setting a flag. So, we explicitly DO NOT set
537 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
538 * hard work for us. When using softpin, we're in control and the fixed
539 * addresses we choose are fine for base addresses.
541 uint64_t bo_flags
= 0;
542 if (pool
->use_softpin
) {
543 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
|
547 if (pool
->device
->instance
->physicalDevice
.has_exec_async
)
548 bo_flags
|= EXEC_OBJECT_ASYNC
;
550 if (pool
->device
->instance
->physicalDevice
.has_exec_capture
)
551 bo_flags
|= EXEC_OBJECT_CAPTURE
;
553 if (pool
->use_softpin
) {
554 uint32_t newbo_size
= size
- pool
->size
;
555 uint32_t gem_handle
= anv_gem_create(pool
->device
, newbo_size
);
556 void *map
= anv_gem_mmap(pool
->device
, gem_handle
, 0, newbo_size
, 0);
557 if (map
== MAP_FAILED
) {
558 anv_gem_close(pool
->device
, gem_handle
);
559 return vk_errorf(pool
->device
->instance
, pool
->device
,
560 VK_ERROR_MEMORY_MAP_FAILED
, "gem mmap failed: %m");
563 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
564 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
565 * always created as I915_CACHING_CACHED, which on non-LLC means
568 * On platforms that support softpin, we are not going to use userptr
569 * anymore, but we still want to rely on the snooped states. So make
570 * sure everything is set to I915_CACHING_CACHED.
572 if (!pool
->device
->info
.has_llc
)
573 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_CACHED
);
575 assert(center_bo_offset
== 0);
577 struct anv_bo
*bo
= &pool
->bos
[pool
->nbos
++];
578 anv_bo_init(bo
, gem_handle
, newbo_size
);
579 bo
->offset
= pool
->start_address
+ pool
->size
;
580 bo
->flags
= bo_flags
;
583 /* Just leak the old map until we destroy the pool. We can't munmap it
584 * without races or imposing locking on the block allocate fast path. On
585 * the whole the leaked maps adds up to less than the size of the
586 * current map. MAP_POPULATE seems like the right thing to do, but we
587 * should try to get some numbers.
589 void *map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
590 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
591 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
592 if (map
== MAP_FAILED
)
593 return vk_errorf(pool
->device
->instance
, pool
->device
,
594 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
596 uint32_t gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
597 if (gem_handle
== 0) {
599 return vk_errorf(pool
->device
->instance
, pool
->device
,
600 VK_ERROR_TOO_MANY_OBJECTS
, "userptr failed: %m");
603 struct anv_mmap_cleanup
*cleanup
= u_vector_add(&pool
->mmap_cleanups
);
606 anv_gem_close(pool
->device
, gem_handle
);
607 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
610 cleanup
->size
= size
;
612 /* Now that we mapped the new memory, we can write the new
613 * center_bo_offset back into pool and update pool->map. */
614 pool
->center_bo_offset
= center_bo_offset
;
615 pool
->map
= map
+ center_bo_offset
;
617 struct anv_bo
*bo
= &pool
->bos
[pool
->nbos
++];
618 anv_bo_init(bo
, gem_handle
, size
);
619 bo
->flags
= bo_flags
;
620 pool
->wrapper_bo
.map
= bo
;
623 assert(pool
->nbos
< ANV_MAX_BLOCK_POOL_BOS
);
629 /** Returns current memory map of the block pool.
631 * The returned pointer points to the map for the memory at the specified
632 * offset. The offset parameter is relative to the "center" of the block pool
633 * rather than the start of the block pool BO map.
636 anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
)
638 if (pool
->use_softpin
) {
639 struct anv_bo
*bo
= NULL
;
640 int32_t bo_offset
= 0;
641 anv_block_pool_foreach_bo(iter_bo
, pool
) {
642 if (offset
< bo_offset
+ iter_bo
->size
) {
646 bo_offset
+= iter_bo
->size
;
649 assert(offset
>= bo_offset
);
651 return bo
->map
+ (offset
- bo_offset
);
653 return pool
->map
+ offset
;
657 /** Grows and re-centers the block pool.
659 * We grow the block pool in one or both directions in such a way that the
660 * following conditions are met:
662 * 1) The size of the entire pool is always a power of two.
664 * 2) The pool only grows on both ends. Neither end can get
667 * 3) At the end of the allocation, we have about twice as much space
668 * allocated for each end as we have used. This way the pool doesn't
669 * grow too far in one direction or the other.
671 * 4) If the _alloc_back() has never been called, then the back portion of
672 * the pool retains a size of zero. (This makes it easier for users of
673 * the block pool that only want a one-sided pool.)
675 * 5) We have enough space allocated for at least one more block in
676 * whichever side `state` points to.
678 * 6) The center of the pool is always aligned to both the block_size of
679 * the pool and a 4K CPU page.
682 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
684 VkResult result
= VK_SUCCESS
;
686 pthread_mutex_lock(&pool
->device
->mutex
);
688 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
690 /* Gather a little usage information on the pool. Since we may have
691 * threadsd waiting in queue to get some storage while we resize, it's
692 * actually possible that total_used will be larger than old_size. In
693 * particular, block_pool_alloc() increments state->next prior to
694 * calling block_pool_grow, so this ensures that we get enough space for
695 * which ever side tries to grow the pool.
697 * We align to a page size because it makes it easier to do our
698 * calculations later in such a way that we state page-aigned.
700 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
701 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
702 uint32_t total_used
= front_used
+ back_used
;
704 assert(state
== &pool
->state
|| back_used
> 0);
706 uint32_t old_size
= pool
->size
;
708 /* The block pool is always initialized to a nonzero size and this function
709 * is always called after initialization.
711 assert(old_size
> 0);
713 /* The back_used and front_used may actually be smaller than the actual
714 * requirement because they are based on the next pointers which are
715 * updated prior to calling this function.
717 uint32_t back_required
= MAX2(back_used
, pool
->center_bo_offset
);
718 uint32_t front_required
= MAX2(front_used
, old_size
- pool
->center_bo_offset
);
720 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
721 /* If we're in this case then this isn't the firsta allocation and we
722 * already have enough space on both sides to hold double what we
723 * have allocated. There's nothing for us to do.
728 uint32_t size
= old_size
* 2;
729 while (size
< back_required
+ front_required
)
732 assert(size
> pool
->size
);
734 /* We compute a new center_bo_offset such that, when we double the size
735 * of the pool, we maintain the ratio of how much is used by each side.
736 * This way things should remain more-or-less balanced.
738 uint32_t center_bo_offset
;
739 if (back_used
== 0) {
740 /* If we're in this case then we have never called alloc_back(). In
741 * this case, we want keep the offset at 0 to make things as simple
742 * as possible for users that don't care about back allocations.
744 center_bo_offset
= 0;
746 /* Try to "center" the allocation based on how much is currently in
747 * use on each side of the center line.
749 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
751 /* Align down to a multiple of the page size */
752 center_bo_offset
&= ~(PAGE_SIZE
- 1);
754 assert(center_bo_offset
>= back_used
);
756 /* Make sure we don't shrink the back end of the pool */
757 if (center_bo_offset
< back_required
)
758 center_bo_offset
= back_required
;
760 /* Make sure that we don't shrink the front end of the pool */
761 if (size
- center_bo_offset
< front_required
)
762 center_bo_offset
= size
- front_required
;
765 assert(center_bo_offset
% PAGE_SIZE
== 0);
767 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
770 pthread_mutex_unlock(&pool
->device
->mutex
);
772 if (result
== VK_SUCCESS
) {
773 /* Return the appropriate new size. This function never actually
774 * updates state->next. Instead, we let the caller do that because it
775 * needs to do so in order to maintain its concurrency model.
777 if (state
== &pool
->state
) {
778 return pool
->size
- pool
->center_bo_offset
;
780 assert(pool
->center_bo_offset
> 0);
781 return pool
->center_bo_offset
;
789 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
790 struct anv_block_state
*pool_state
,
791 uint32_t block_size
, uint32_t *padding
)
793 struct anv_block_state state
, old
, new;
795 /* Most allocations won't generate any padding */
800 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
801 if (state
.next
+ block_size
<= state
.end
) {
803 } else if (state
.next
<= state
.end
) {
804 if (pool
->use_softpin
&& state
.next
< state
.end
) {
805 /* We need to grow the block pool, but still have some leftover
806 * space that can't be used by that particular allocation. So we
807 * add that as a "padding", and return it.
809 uint32_t leftover
= state
.end
- state
.next
;
811 /* If there is some leftover space in the pool, the caller must
814 assert(leftover
== 0 || padding
);
817 state
.next
+= leftover
;
820 /* We allocated the first block outside the pool so we have to grow
821 * the pool. pool_state->next acts a mutex: threads who try to
822 * allocate now will get block indexes above the current limit and
823 * hit futex_wait below.
825 new.next
= state
.next
+ block_size
;
827 new.end
= anv_block_pool_grow(pool
, pool_state
);
828 } while (new.end
< new.next
);
830 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
831 if (old
.next
!= state
.next
)
832 futex_wake(&pool_state
->end
, INT_MAX
);
835 futex_wait(&pool_state
->end
, state
.end
, NULL
);
842 anv_block_pool_alloc(struct anv_block_pool
*pool
,
843 uint32_t block_size
, uint32_t *padding
)
847 offset
= anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
, padding
);
852 /* Allocates a block out of the back of the block pool.
854 * This will allocated a block earlier than the "start" of the block pool.
855 * The offsets returned from this function will be negative but will still
856 * be correct relative to the block pool's map pointer.
858 * If you ever use anv_block_pool_alloc_back, then you will have to do
859 * gymnastics with the block pool's BO when doing relocations.
862 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
865 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
868 /* The offset we get out of anv_block_pool_alloc_new() is actually the
869 * number of bytes downwards from the middle to the end of the block.
870 * We need to turn it into a (negative) offset from the middle to the
871 * start of the block.
874 return -(offset
+ block_size
);
878 anv_state_pool_init(struct anv_state_pool
*pool
,
879 struct anv_device
*device
,
880 uint64_t start_address
,
883 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
886 if (result
!= VK_SUCCESS
)
889 result
= anv_state_table_init(&pool
->table
, device
, 64);
890 if (result
!= VK_SUCCESS
) {
891 anv_block_pool_finish(&pool
->block_pool
);
895 assert(util_is_power_of_two_or_zero(block_size
));
896 pool
->block_size
= block_size
;
897 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
898 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
899 pool
->buckets
[i
].free_list
= ANV_FREE_LIST_EMPTY
;
900 pool
->buckets
[i
].block
.next
= 0;
901 pool
->buckets
[i
].block
.end
= 0;
903 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
909 anv_state_pool_finish(struct anv_state_pool
*pool
)
911 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
912 anv_state_table_finish(&pool
->table
);
913 anv_block_pool_finish(&pool
->block_pool
);
917 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
918 struct anv_block_pool
*block_pool
,
923 struct anv_block_state block
, old
, new;
926 /* We don't always use anv_block_pool_alloc(), which would set *padding to
927 * zero for us. So if we have a pointer to padding, we must zero it out
928 * ourselves here, to make sure we always return some sensible value.
933 /* If our state is large, we don't need any sub-allocation from a block.
934 * Instead, we just grab whole (potentially large) blocks.
936 if (state_size
>= block_size
)
937 return anv_block_pool_alloc(block_pool
, state_size
, padding
);
940 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
942 if (block
.next
< block
.end
) {
944 } else if (block
.next
== block
.end
) {
945 offset
= anv_block_pool_alloc(block_pool
, block_size
, padding
);
946 new.next
= offset
+ state_size
;
947 new.end
= offset
+ block_size
;
948 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
949 if (old
.next
!= block
.next
)
950 futex_wake(&pool
->block
.end
, INT_MAX
);
953 futex_wait(&pool
->block
.end
, block
.end
, NULL
);
959 anv_state_pool_get_bucket(uint32_t size
)
961 unsigned size_log2
= ilog2_round_up(size
);
962 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
963 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
964 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
965 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
969 anv_state_pool_get_bucket_size(uint32_t bucket
)
971 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
972 return 1 << size_log2
;
975 /** Helper to push a chunk into the state table.
977 * It creates 'count' entries into the state table and update their sizes,
978 * offsets and maps, also pushing them as "free" states.
981 anv_state_pool_return_blocks(struct anv_state_pool
*pool
,
982 uint32_t chunk_offset
, uint32_t count
,
985 /* Disallow returning 0 chunks */
988 /* Make sure we always return chunks aligned to the block_size */
989 assert(chunk_offset
% block_size
== 0);
992 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &st_idx
, count
);
993 assert(result
== VK_SUCCESS
);
994 for (int i
= 0; i
< count
; i
++) {
995 /* update states that were added back to the state table */
996 struct anv_state
*state_i
= anv_state_table_get(&pool
->table
,
998 state_i
->alloc_size
= block_size
;
999 state_i
->offset
= chunk_offset
+ block_size
* i
;
1000 state_i
->map
= anv_block_pool_map(&pool
->block_pool
, state_i
->offset
);
1003 uint32_t block_bucket
= anv_state_pool_get_bucket(block_size
);
1004 anv_free_list_push(&pool
->buckets
[block_bucket
].free_list
,
1005 &pool
->table
, st_idx
, count
);
1008 /** Returns a chunk of memory back to the state pool.
1010 * Do a two-level split. If chunk_size is bigger than divisor
1011 * (pool->block_size), we return as many divisor sized blocks as we can, from
1012 * the end of the chunk.
1014 * The remaining is then split into smaller blocks (starting at small_size if
1015 * it is non-zero), with larger blocks always being taken from the end of the
1019 anv_state_pool_return_chunk(struct anv_state_pool
*pool
,
1020 uint32_t chunk_offset
, uint32_t chunk_size
,
1021 uint32_t small_size
)
1023 uint32_t divisor
= pool
->block_size
;
1024 uint32_t nblocks
= chunk_size
/ divisor
;
1025 uint32_t rest
= chunk_size
- nblocks
* divisor
;
1028 /* First return divisor aligned and sized chunks. We start returning
1029 * larger blocks from the end fo the chunk, since they should already be
1030 * aligned to divisor. Also anv_state_pool_return_blocks() only accepts
1033 uint32_t offset
= chunk_offset
+ rest
;
1034 anv_state_pool_return_blocks(pool
, offset
, nblocks
, divisor
);
1040 if (small_size
> 0 && small_size
< divisor
)
1041 divisor
= small_size
;
1043 uint32_t min_size
= 1 << ANV_MIN_STATE_SIZE_LOG2
;
1045 /* Just as before, return larger divisor aligned blocks from the end of the
1048 while (chunk_size
> 0 && divisor
>= min_size
) {
1049 nblocks
= chunk_size
/ divisor
;
1050 rest
= chunk_size
- nblocks
* divisor
;
1052 anv_state_pool_return_blocks(pool
, chunk_offset
+ rest
,
1060 static struct anv_state
1061 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
1062 uint32_t size
, uint32_t align
)
1064 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
1066 struct anv_state
*state
;
1067 uint32_t alloc_size
= anv_state_pool_get_bucket_size(bucket
);
1070 /* Try free list first. */
1071 state
= anv_free_list_pop(&pool
->buckets
[bucket
].free_list
,
1074 assert(state
->offset
>= 0);
1078 /* Try to grab a chunk from some larger bucket and split it up */
1079 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
1080 state
= anv_free_list_pop(&pool
->buckets
[b
].free_list
, &pool
->table
);
1082 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
1083 int32_t chunk_offset
= state
->offset
;
1085 /* First lets update the state we got to its new size. offset and map
1088 state
->alloc_size
= alloc_size
;
1090 /* Now return the unused part of the chunk back to the pool as free
1093 * There are a couple of options as to what we do with it:
1095 * 1) We could fully split the chunk into state.alloc_size sized
1096 * pieces. However, this would mean that allocating a 16B
1097 * state could potentially split a 2MB chunk into 512K smaller
1098 * chunks. This would lead to unnecessary fragmentation.
1100 * 2) The classic "buddy allocator" method would have us split the
1101 * chunk in half and return one half. Then we would split the
1102 * remaining half in half and return one half, and repeat as
1103 * needed until we get down to the size we want. However, if
1104 * you are allocating a bunch of the same size state (which is
1105 * the common case), this means that every other allocation has
1106 * to go up a level and every fourth goes up two levels, etc.
1107 * This is not nearly as efficient as it could be if we did a
1108 * little more work up-front.
1110 * 3) Split the difference between (1) and (2) by doing a
1111 * two-level split. If it's bigger than some fixed block_size,
1112 * we split it into block_size sized chunks and return all but
1113 * one of them. Then we split what remains into
1114 * state.alloc_size sized chunks and return them.
1116 * We choose something close to option (3), which is implemented with
1117 * anv_state_pool_return_chunk(). That is done by returning the
1118 * remaining of the chunk, with alloc_size as a hint of the size that
1119 * we want the smaller chunk split into.
1121 anv_state_pool_return_chunk(pool
, chunk_offset
+ alloc_size
,
1122 chunk_size
- alloc_size
, alloc_size
);
1128 offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
1133 /* Everytime we allocate a new state, add it to the state pool */
1135 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1136 assert(result
== VK_SUCCESS
);
1138 state
= anv_state_table_get(&pool
->table
, idx
);
1139 state
->offset
= offset
;
1140 state
->alloc_size
= alloc_size
;
1141 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
);
1144 uint32_t return_offset
= offset
- padding
;
1145 anv_state_pool_return_chunk(pool
, return_offset
, padding
, 0);
1153 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
1156 return ANV_STATE_NULL
;
1158 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
1159 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
1164 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
1166 struct anv_state
*state
;
1167 uint32_t alloc_size
= pool
->block_size
;
1169 state
= anv_free_list_pop(&pool
->back_alloc_free_list
, &pool
->table
);
1171 assert(state
->offset
< 0);
1176 offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
1179 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1180 assert(result
== VK_SUCCESS
);
1182 state
= anv_state_table_get(&pool
->table
, idx
);
1183 state
->offset
= offset
;
1184 state
->alloc_size
= alloc_size
;
1185 state
->map
= anv_block_pool_map(&pool
->block_pool
, state
->offset
);
1188 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
->map
, state
->alloc_size
));
1193 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
1195 assert(util_is_power_of_two_or_zero(state
.alloc_size
));
1196 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
1198 if (state
.offset
< 0) {
1199 assert(state
.alloc_size
== pool
->block_size
);
1200 anv_free_list_push(&pool
->back_alloc_free_list
,
1201 &pool
->table
, state
.idx
, 1);
1203 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
1204 &pool
->table
, state
.idx
, 1);
1209 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
1211 if (state
.alloc_size
== 0)
1214 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
1215 anv_state_pool_free_no_vg(pool
, state
);
1218 struct anv_state_stream_block
{
1219 struct anv_state block
;
1221 /* The next block */
1222 struct anv_state_stream_block
*next
;
1224 #ifdef HAVE_VALGRIND
1225 /* A pointer to the first user-allocated thing in this block. This is
1226 * what valgrind sees as the start of the block.
1232 /* The state stream allocator is a one-shot, single threaded allocator for
1233 * variable sized blocks. We use it for allocating dynamic state.
1236 anv_state_stream_init(struct anv_state_stream
*stream
,
1237 struct anv_state_pool
*state_pool
,
1238 uint32_t block_size
)
1240 stream
->state_pool
= state_pool
;
1241 stream
->block_size
= block_size
;
1243 stream
->block
= ANV_STATE_NULL
;
1245 stream
->block_list
= NULL
;
1247 /* Ensure that next + whatever > block_size. This way the first call to
1248 * state_stream_alloc fetches a new block.
1250 stream
->next
= block_size
;
1252 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
1256 anv_state_stream_finish(struct anv_state_stream
*stream
)
1258 struct anv_state_stream_block
*next
= stream
->block_list
;
1259 while (next
!= NULL
) {
1260 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
1261 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
1262 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
1263 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
1267 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
1271 anv_state_stream_alloc(struct anv_state_stream
*stream
,
1272 uint32_t size
, uint32_t alignment
)
1275 return ANV_STATE_NULL
;
1277 assert(alignment
<= PAGE_SIZE
);
1279 uint32_t offset
= align_u32(stream
->next
, alignment
);
1280 if (offset
+ size
> stream
->block
.alloc_size
) {
1281 uint32_t block_size
= stream
->block_size
;
1282 if (block_size
< size
)
1283 block_size
= round_to_power_of_two(size
);
1285 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
1286 block_size
, PAGE_SIZE
);
1288 struct anv_state_stream_block
*sb
= stream
->block
.map
;
1289 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
1290 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
1291 stream
->block_list
= sb
;
1292 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
1294 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
1296 /* Reset back to the start plus space for the header */
1297 stream
->next
= sizeof(*sb
);
1299 offset
= align_u32(stream
->next
, alignment
);
1300 assert(offset
+ size
<= stream
->block
.alloc_size
);
1303 struct anv_state state
= stream
->block
;
1304 state
.offset
+= offset
;
1305 state
.alloc_size
= size
;
1306 state
.map
+= offset
;
1308 stream
->next
= offset
+ size
;
1310 #ifdef HAVE_VALGRIND
1311 struct anv_state_stream_block
*sb
= stream
->block_list
;
1312 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
1313 if (vg_ptr
== NULL
) {
1315 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
1316 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
1318 void *state_end
= state
.map
+ state
.alloc_size
;
1319 /* This only updates the mempool. The newly allocated chunk is still
1320 * marked as NOACCESS. */
1321 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
1322 /* Mark the newly allocated chunk as undefined */
1323 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
1330 struct bo_pool_bo_link
{
1331 struct bo_pool_bo_link
*next
;
1336 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
,
1339 pool
->device
= device
;
1340 pool
->bo_flags
= bo_flags
;
1341 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
1343 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
1347 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
1349 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1350 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
1351 while (link
!= NULL
) {
1352 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
1354 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
1355 anv_vma_free(pool
->device
, &link_copy
.bo
);
1356 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
1357 link
= link_copy
.next
;
1361 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1365 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
1369 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1370 const unsigned pow2_size
= 1 << size_log2
;
1371 const unsigned bucket
= size_log2
- 12;
1372 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1374 void *next_free_void
;
1375 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
1376 struct bo_pool_bo_link
*next_free
= next_free_void
;
1377 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
1378 assert(bo
->gem_handle
);
1379 assert(bo
->map
== next_free
);
1380 assert(size
<= bo
->size
);
1382 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1387 struct anv_bo new_bo
;
1389 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
1390 if (result
!= VK_SUCCESS
)
1393 new_bo
.flags
= pool
->bo_flags
;
1395 if (!anv_vma_alloc(pool
->device
, &new_bo
))
1396 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1398 assert(new_bo
.size
== pow2_size
);
1400 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
1401 if (new_bo
.map
== MAP_FAILED
) {
1402 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
1403 anv_vma_free(pool
->device
, &new_bo
);
1404 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1407 /* We are removing the state flushes, so lets make sure that these buffers
1408 * are cached/snooped.
1410 if (!pool
->device
->info
.has_llc
) {
1411 anv_gem_set_caching(pool
->device
, new_bo
.gem_handle
,
1412 I915_CACHING_CACHED
);
1417 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1423 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
1425 /* Make a copy in case the anv_bo happens to be storred in the BO */
1426 struct anv_bo bo
= *bo_in
;
1428 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
1430 struct bo_pool_bo_link
*link
= bo
.map
;
1431 VG_NOACCESS_WRITE(&link
->bo
, bo
);
1433 assert(util_is_power_of_two_or_zero(bo
.size
));
1434 const unsigned size_log2
= ilog2_round_up(bo
.size
);
1435 const unsigned bucket
= size_log2
- 12;
1436 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1438 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
1444 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1446 memset(pool
, 0, sizeof(*pool
));
1450 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1452 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1453 for (unsigned i
= 0; i
< 16; i
++) {
1454 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
1455 if (bo
->exists
> 0) {
1456 anv_vma_free(device
, &bo
->bo
);
1457 anv_gem_close(device
, bo
->bo
.gem_handle
);
1464 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1465 gl_shader_stage stage
, unsigned per_thread_scratch
)
1467 if (per_thread_scratch
== 0)
1470 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1471 assert(scratch_size_log2
< 16);
1473 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
1475 /* We can use "exists" to shortcut and ignore the critical section */
1479 pthread_mutex_lock(&device
->mutex
);
1481 __sync_synchronize();
1483 pthread_mutex_unlock(&device
->mutex
);
1487 const struct anv_physical_device
*physical_device
=
1488 &device
->instance
->physicalDevice
;
1489 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1491 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
1493 unsigned scratch_ids_per_subslice
;
1494 if (devinfo
->gen
>= 11) {
1495 /* The MEDIA_VFE_STATE docs say:
1497 * "Starting with this configuration, the Maximum Number of
1498 * Threads must be set to (#EU * 8) for GPGPU dispatches.
1500 * Although there are only 7 threads per EU in the configuration,
1501 * the FFTID is calculated as if there are 8 threads per EU,
1502 * which in turn requires a larger amount of Scratch Space to be
1503 * allocated by the driver."
1505 scratch_ids_per_subslice
= 8 * 8;
1506 } else if (devinfo
->is_haswell
) {
1507 /* WaCSScratchSize:hsw
1509 * Haswell's scratch space address calculation appears to be sparse
1510 * rather than tightly packed. The Thread ID has bits indicating
1511 * which subslice, EU within a subslice, and thread within an EU it
1512 * is. There's a maximum of two slices and two subslices, so these
1513 * can be stored with a single bit. Even though there are only 10 EUs
1514 * per subslice, this is stored in 4 bits, so there's an effective
1515 * maximum value of 16 EUs. Similarly, although there are only 7
1516 * threads per EU, this is stored in a 3 bit number, giving an
1517 * effective maximum value of 8 threads per EU.
1519 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1520 * number of threads per subslice.
1522 scratch_ids_per_subslice
= 16 * 8;
1523 } else if (devinfo
->is_cherryview
) {
1524 /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
1525 * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
1528 scratch_ids_per_subslice
= 8 * 7;
1530 scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
1533 uint32_t max_threads
[] = {
1534 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1535 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1536 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1537 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1538 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1539 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1542 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1544 anv_bo_init_new(&bo
->bo
, device
, size
);
1546 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1547 * are still relative to the general state base address. When we emit
1548 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1549 * to the maximum (1 page under 4GB). This allows us to just place the
1550 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1551 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1552 * However, in order to do so, we need to ensure that the kernel does not
1553 * place the scratch BO above the 32-bit boundary.
1555 * NOTE: Technically, it can't go "anywhere" because the top page is off
1556 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1557 * kernel allocates space using
1559 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1561 * so nothing will ever touch the top page.
1563 assert(!(bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
));
1565 if (device
->instance
->physicalDevice
.has_exec_async
)
1566 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1568 if (device
->instance
->physicalDevice
.use_softpin
)
1569 bo
->bo
.flags
|= EXEC_OBJECT_PINNED
;
1571 anv_vma_alloc(device
, &bo
->bo
);
1573 /* Set the exists last because it may be read by other threads */
1574 __sync_synchronize();
1577 pthread_mutex_unlock(&device
->mutex
);
1583 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1585 util_sparse_array_init(&cache
->bo_map
, sizeof(struct anv_bo
), 1024);
1587 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1588 util_sparse_array_finish(&cache
->bo_map
);
1589 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1590 "pthread_mutex_init failed: %m");
1597 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1599 util_sparse_array_finish(&cache
->bo_map
);
1600 pthread_mutex_destroy(&cache
->mutex
);
1603 #define ANV_BO_CACHE_SUPPORTED_FLAGS \
1604 (EXEC_OBJECT_WRITE | \
1605 EXEC_OBJECT_ASYNC | \
1606 EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
1607 EXEC_OBJECT_PINNED | \
1608 EXEC_OBJECT_CAPTURE)
1611 anv_bo_alloc_flags_to_bo_flags(struct anv_device
*device
,
1612 enum anv_bo_alloc_flags alloc_flags
)
1614 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1616 uint64_t bo_flags
= 0;
1617 if (!(alloc_flags
& ANV_BO_ALLOC_32BIT_ADDRESS
) &&
1618 pdevice
->supports_48bit_addresses
)
1619 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1621 if ((alloc_flags
& ANV_BO_ALLOC_CAPTURE
) && pdevice
->has_exec_capture
)
1622 bo_flags
|= EXEC_OBJECT_CAPTURE
;
1624 if (alloc_flags
& ANV_BO_ALLOC_IMPLICIT_WRITE
) {
1625 assert(alloc_flags
& ANV_BO_ALLOC_IMPLICIT_SYNC
);
1626 bo_flags
|= EXEC_OBJECT_WRITE
;
1629 if (!(alloc_flags
& ANV_BO_ALLOC_IMPLICIT_SYNC
) && pdevice
->has_exec_async
)
1630 bo_flags
|= EXEC_OBJECT_ASYNC
;
1632 if (pdevice
->use_softpin
)
1633 bo_flags
|= EXEC_OBJECT_PINNED
;
1639 anv_device_alloc_bo(struct anv_device
*device
,
1641 enum anv_bo_alloc_flags alloc_flags
,
1642 struct anv_bo
**bo_out
)
1644 const uint32_t bo_flags
=
1645 anv_bo_alloc_flags_to_bo_flags(device
, alloc_flags
);
1646 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1648 /* The kernel is going to give us whole pages anyway */
1649 size
= align_u64(size
, 4096);
1651 struct anv_bo new_bo
;
1652 VkResult result
= anv_bo_init_new(&new_bo
, device
, size
);
1653 if (result
!= VK_SUCCESS
)
1656 new_bo
.flags
= bo_flags
;
1657 new_bo
.is_external
= (alloc_flags
& ANV_BO_ALLOC_EXTERNAL
);
1659 if (alloc_flags
& ANV_BO_ALLOC_MAPPED
) {
1660 new_bo
.map
= anv_gem_mmap(device
, new_bo
.gem_handle
, 0, size
, 0);
1661 if (new_bo
.map
== MAP_FAILED
) {
1662 anv_gem_close(device
, new_bo
.gem_handle
);
1663 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1667 if (alloc_flags
& ANV_BO_ALLOC_SNOOPED
) {
1668 assert(alloc_flags
& ANV_BO_ALLOC_MAPPED
);
1669 /* We don't want to change these defaults if it's going to be shared
1670 * with another process.
1672 assert(!(alloc_flags
& ANV_BO_ALLOC_EXTERNAL
));
1674 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
1675 * I915_CACHING_NONE on non-LLC platforms. For many internal state
1676 * objects, we'd rather take the snooping overhead than risk forgetting
1677 * a CLFLUSH somewhere. Userptr objects are always created as
1678 * I915_CACHING_CACHED, which on non-LLC means snooped so there's no
1679 * need to do this there.
1681 if (!device
->info
.has_llc
) {
1682 anv_gem_set_caching(device
, new_bo
.gem_handle
,
1683 I915_CACHING_CACHED
);
1687 if (alloc_flags
& ANV_BO_ALLOC_FIXED_ADDRESS
) {
1688 new_bo
.has_fixed_address
= true;
1690 if (!anv_vma_alloc(device
, &new_bo
)) {
1692 anv_gem_munmap(new_bo
.map
, size
);
1693 anv_gem_close(device
, new_bo
.gem_handle
);
1694 return vk_errorf(device
->instance
, NULL
,
1695 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1696 "failed to allocate virtual address for BO");
1700 assert(new_bo
.gem_handle
);
1702 /* If we just got this gem_handle from anv_bo_init_new then we know no one
1703 * else is touching this BO at the moment so we don't need to lock here.
1705 struct anv_bo
*bo
= anv_device_lookup_bo(device
, new_bo
.gem_handle
);
1714 anv_device_import_bo_from_host_ptr(struct anv_device
*device
,
1715 void *host_ptr
, uint32_t size
,
1716 enum anv_bo_alloc_flags alloc_flags
,
1717 struct anv_bo
**bo_out
)
1719 assert(!(alloc_flags
& (ANV_BO_ALLOC_MAPPED
|
1720 ANV_BO_ALLOC_SNOOPED
|
1721 ANV_BO_ALLOC_FIXED_ADDRESS
)));
1723 struct anv_bo_cache
*cache
= &device
->bo_cache
;
1724 const uint32_t bo_flags
=
1725 anv_bo_alloc_flags_to_bo_flags(device
, alloc_flags
);
1726 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1728 uint32_t gem_handle
= anv_gem_userptr(device
, host_ptr
, size
);
1730 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1732 pthread_mutex_lock(&cache
->mutex
);
1734 struct anv_bo
*bo
= anv_device_lookup_bo(device
, gem_handle
);
1735 if (bo
->refcount
> 0) {
1736 /* VK_EXT_external_memory_host doesn't require handling importing the
1737 * same pointer twice at the same time, but we don't get in the way. If
1738 * kernel gives us the same gem_handle, only succeed if the flags match.
1740 assert(bo
->gem_handle
== gem_handle
);
1741 if (bo_flags
!= bo
->flags
) {
1742 pthread_mutex_unlock(&cache
->mutex
);
1743 return vk_errorf(device
->instance
, NULL
,
1744 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1745 "same host pointer imported two different ways");
1747 __sync_fetch_and_add(&bo
->refcount
, 1);
1749 struct anv_bo new_bo
;
1750 anv_bo_init(&new_bo
, gem_handle
, size
);
1751 new_bo
.map
= host_ptr
;
1752 new_bo
.flags
= bo_flags
;
1753 new_bo
.is_external
= true;
1754 new_bo
.from_host_ptr
= true;
1756 if (!anv_vma_alloc(device
, &new_bo
)) {
1757 anv_gem_close(device
, new_bo
.gem_handle
);
1758 pthread_mutex_unlock(&cache
->mutex
);
1759 return vk_errorf(device
->instance
, NULL
,
1760 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1761 "failed to allocate virtual address for BO");
1767 pthread_mutex_unlock(&cache
->mutex
);
1774 anv_device_import_bo(struct anv_device
*device
,
1776 enum anv_bo_alloc_flags alloc_flags
,
1777 struct anv_bo
**bo_out
)
1779 assert(!(alloc_flags
& (ANV_BO_ALLOC_MAPPED
|
1780 ANV_BO_ALLOC_SNOOPED
|
1781 ANV_BO_ALLOC_FIXED_ADDRESS
)));
1783 struct anv_bo_cache
*cache
= &device
->bo_cache
;
1784 const uint32_t bo_flags
=
1785 anv_bo_alloc_flags_to_bo_flags(device
, alloc_flags
);
1786 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1788 pthread_mutex_lock(&cache
->mutex
);
1790 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1792 pthread_mutex_unlock(&cache
->mutex
);
1793 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1796 struct anv_bo
*bo
= anv_device_lookup_bo(device
, gem_handle
);
1797 if (bo
->refcount
> 0) {
1798 /* We have to be careful how we combine flags so that it makes sense.
1799 * Really, though, if we get to this case and it actually matters, the
1800 * client has imported a BO twice in different ways and they get what
1803 uint64_t new_flags
= 0;
1804 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_WRITE
;
1805 new_flags
|= (bo
->flags
& bo_flags
) & EXEC_OBJECT_ASYNC
;
1806 new_flags
|= (bo
->flags
& bo_flags
) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1807 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_PINNED
;
1808 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_CAPTURE
;
1810 /* It's theoretically possible for a BO to get imported such that it's
1811 * both pinned and not pinned. The only way this can happen is if it
1812 * gets imported as both a semaphore and a memory object and that would
1813 * be an application error. Just fail out in that case.
1815 if ((bo
->flags
& EXEC_OBJECT_PINNED
) !=
1816 (bo_flags
& EXEC_OBJECT_PINNED
)) {
1817 pthread_mutex_unlock(&cache
->mutex
);
1818 return vk_errorf(device
->instance
, NULL
,
1819 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1820 "The same BO was imported two different ways");
1823 /* It's also theoretically possible that someone could export a BO from
1824 * one heap and import it into another or to import the same BO into two
1825 * different heaps. If this happens, we could potentially end up both
1826 * allowing and disallowing 48-bit addresses. There's not much we can
1827 * do about it if we're pinning so we just throw an error and hope no
1828 * app is actually that stupid.
1830 if ((new_flags
& EXEC_OBJECT_PINNED
) &&
1831 (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) !=
1832 (bo_flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
)) {
1833 pthread_mutex_unlock(&cache
->mutex
);
1834 return vk_errorf(device
->instance
, NULL
,
1835 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1836 "The same BO was imported on two different heaps");
1839 bo
->flags
= new_flags
;
1841 __sync_fetch_and_add(&bo
->refcount
, 1);
1843 off_t size
= lseek(fd
, 0, SEEK_END
);
1844 if (size
== (off_t
)-1) {
1845 anv_gem_close(device
, gem_handle
);
1846 pthread_mutex_unlock(&cache
->mutex
);
1847 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1850 struct anv_bo new_bo
;
1851 anv_bo_init(&new_bo
, gem_handle
, size
);
1852 new_bo
.flags
= bo_flags
;
1853 new_bo
.is_external
= true;
1855 if (!anv_vma_alloc(device
, &new_bo
)) {
1856 anv_gem_close(device
, new_bo
.gem_handle
);
1857 pthread_mutex_unlock(&cache
->mutex
);
1858 return vk_errorf(device
->instance
, NULL
,
1859 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1860 "failed to allocate virtual address for BO");
1866 pthread_mutex_unlock(&cache
->mutex
);
1873 anv_device_export_bo(struct anv_device
*device
,
1874 struct anv_bo
*bo
, int *fd_out
)
1876 assert(anv_device_lookup_bo(device
, bo
->gem_handle
) == bo
);
1878 /* This BO must have been flagged external in order for us to be able
1879 * to export it. This is done based on external options passed into
1880 * anv_AllocateMemory.
1882 assert(bo
->is_external
);
1884 int fd
= anv_gem_handle_to_fd(device
, bo
->gem_handle
);
1886 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1894 atomic_dec_not_one(uint32_t *counter
)
1903 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1912 anv_device_release_bo(struct anv_device
*device
,
1915 struct anv_bo_cache
*cache
= &device
->bo_cache
;
1916 assert(anv_device_lookup_bo(device
, bo
->gem_handle
) == bo
);
1918 /* Try to decrement the counter but don't go below one. If this succeeds
1919 * then the refcount has been decremented and we are not the last
1922 if (atomic_dec_not_one(&bo
->refcount
))
1925 pthread_mutex_lock(&cache
->mutex
);
1927 /* We are probably the last reference since our attempt to decrement above
1928 * failed. However, we can't actually know until we are inside the mutex.
1929 * Otherwise, someone could import the BO between the decrement and our
1932 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1933 /* Turns out we're not the last reference. Unlock and bail. */
1934 pthread_mutex_unlock(&cache
->mutex
);
1937 assert(bo
->refcount
== 0);
1939 if (bo
->map
&& !bo
->from_host_ptr
)
1940 anv_gem_munmap(bo
->map
, bo
->size
);
1942 if (!bo
->has_fixed_address
)
1943 anv_vma_free(device
, bo
);
1945 anv_gem_close(device
, bo
->gem_handle
);
1947 /* Don't unlock until we've actually closed the BO. The whole point of
1948 * the BO cache is to ensure that we correctly handle races with creating
1949 * and releasing GEM handles and we don't want to let someone import the BO
1950 * again between mutex unlock and closing the GEM handle.
1952 pthread_mutex_unlock(&cache
->mutex
);