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
{
115 #define ANV_MMAP_CLEANUP_INIT ((struct anv_mmap_cleanup){0})
117 static inline uint32_t
118 ilog2_round_up(uint32_t value
)
121 return 32 - __builtin_clz(value
- 1);
124 static inline uint32_t
125 round_to_power_of_two(uint32_t value
)
127 return 1 << ilog2_round_up(value
);
130 struct anv_state_table_cleanup
{
135 #define ANV_STATE_TABLE_CLEANUP_INIT ((struct anv_state_table_cleanup){0})
136 #define ANV_STATE_ENTRY_SIZE (sizeof(struct anv_free_entry))
139 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
);
142 anv_state_table_init(struct anv_state_table
*table
,
143 struct anv_device
*device
,
144 uint32_t initial_entries
)
148 table
->device
= device
;
150 /* Just make it 2GB up-front. The Linux kernel won't actually back it
151 * with pages until we either map and fault on one of them or we use
152 * userptr and send a chunk of it off to the GPU.
154 table
->fd
= os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE
, "state table");
155 if (table
->fd
== -1) {
156 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
160 if (!u_vector_init(&table
->cleanups
,
161 round_to_power_of_two(sizeof(struct anv_state_table_cleanup
)),
163 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
167 table
->state
.next
= 0;
168 table
->state
.end
= 0;
171 uint32_t initial_size
= initial_entries
* ANV_STATE_ENTRY_SIZE
;
172 result
= anv_state_table_expand_range(table
, initial_size
);
173 if (result
!= VK_SUCCESS
)
179 u_vector_finish(&table
->cleanups
);
187 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
)
190 struct anv_state_table_cleanup
*cleanup
;
192 /* Assert that we only ever grow the pool */
193 assert(size
>= table
->state
.end
);
195 /* Make sure that we don't go outside the bounds of the memfd */
196 if (size
> BLOCK_POOL_MEMFD_SIZE
)
197 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
199 cleanup
= u_vector_add(&table
->cleanups
);
201 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
203 *cleanup
= ANV_STATE_TABLE_CLEANUP_INIT
;
205 /* Just leak the old map until we destroy the pool. We can't munmap it
206 * without races or imposing locking on the block allocate fast path. On
207 * the whole the leaked maps adds up to less than the size of the
208 * current map. MAP_POPULATE seems like the right thing to do, but we
209 * should try to get some numbers.
211 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
212 MAP_SHARED
| MAP_POPULATE
, table
->fd
, 0);
213 if (map
== MAP_FAILED
) {
214 return vk_errorf(table
->device
->instance
, table
->device
,
215 VK_ERROR_OUT_OF_HOST_MEMORY
, "mmap failed: %m");
219 cleanup
->size
= size
;
228 anv_state_table_grow(struct anv_state_table
*table
)
230 VkResult result
= VK_SUCCESS
;
232 uint32_t used
= align_u32(table
->state
.next
* ANV_STATE_ENTRY_SIZE
,
234 uint32_t old_size
= table
->size
;
236 /* The block pool is always initialized to a nonzero size and this function
237 * is always called after initialization.
239 assert(old_size
> 0);
241 uint32_t required
= MAX2(used
, old_size
);
242 if (used
* 2 <= required
) {
243 /* If we're in this case then this isn't the firsta allocation and we
244 * already have enough space on both sides to hold double what we
245 * have allocated. There's nothing for us to do.
250 uint32_t size
= old_size
* 2;
251 while (size
< required
)
254 assert(size
> table
->size
);
256 result
= anv_state_table_expand_range(table
, size
);
263 anv_state_table_finish(struct anv_state_table
*table
)
265 struct anv_state_table_cleanup
*cleanup
;
267 u_vector_foreach(cleanup
, &table
->cleanups
) {
269 munmap(cleanup
->map
, cleanup
->size
);
272 u_vector_finish(&table
->cleanups
);
278 anv_state_table_add(struct anv_state_table
*table
, uint32_t *idx
,
281 struct anv_block_state state
, old
, new;
287 state
.u64
= __sync_fetch_and_add(&table
->state
.u64
, count
);
288 if (state
.next
+ count
<= state
.end
) {
290 struct anv_free_entry
*entry
= &table
->map
[state
.next
];
291 for (int i
= 0; i
< count
; i
++) {
292 entry
[i
].state
.idx
= state
.next
+ i
;
296 } else if (state
.next
<= state
.end
) {
297 /* We allocated the first block outside the pool so we have to grow
298 * the pool. pool_state->next acts a mutex: threads who try to
299 * allocate now will get block indexes above the current limit and
300 * hit futex_wait below.
302 new.next
= state
.next
+ count
;
304 result
= anv_state_table_grow(table
);
305 if (result
!= VK_SUCCESS
)
307 new.end
= table
->size
/ ANV_STATE_ENTRY_SIZE
;
308 } while (new.end
< new.next
);
310 old
.u64
= __sync_lock_test_and_set(&table
->state
.u64
, new.u64
);
311 if (old
.next
!= state
.next
)
312 futex_wake(&table
->state
.end
, INT_MAX
);
314 futex_wait(&table
->state
.end
, state
.end
, NULL
);
321 anv_free_list_push(union anv_free_list
*list
,
322 struct anv_state_table
*table
,
323 uint32_t first
, uint32_t count
)
325 union anv_free_list current
, old
, new;
326 uint32_t last
= first
;
328 for (uint32_t i
= 1; i
< count
; i
++, last
++)
329 table
->map
[last
].next
= last
+ 1;
334 table
->map
[last
].next
= current
.offset
;
336 new.count
= current
.count
+ 1;
337 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
338 } while (old
.u64
!= current
.u64
);
342 anv_free_list_pop(union anv_free_list
*list
,
343 struct anv_state_table
*table
)
345 union anv_free_list current
, new, old
;
347 current
.u64
= list
->u64
;
348 while (current
.offset
!= EMPTY
) {
349 __sync_synchronize();
350 new.offset
= table
->map
[current
.offset
].next
;
351 new.count
= current
.count
+ 1;
352 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
353 if (old
.u64
== current
.u64
) {
354 struct anv_free_entry
*entry
= &table
->map
[current
.offset
];
355 return &entry
->state
;
363 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
364 * means that the bottom 12 bits should all be zero.
366 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
367 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
368 #define PFL_PACK(ptr, count) ({ \
369 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
373 anv_ptr_free_list_pop(void **list
, void **elem
)
375 void *current
= *list
;
376 while (PFL_PTR(current
) != NULL
) {
377 void **next_ptr
= PFL_PTR(current
);
378 void *new_ptr
= VG_NOACCESS_READ(next_ptr
);
379 unsigned new_count
= PFL_COUNT(current
) + 1;
380 void *new = PFL_PACK(new_ptr
, new_count
);
381 void *old
= __sync_val_compare_and_swap(list
, current
, new);
382 if (old
== current
) {
383 *elem
= PFL_PTR(current
);
393 anv_ptr_free_list_push(void **list
, void *elem
)
396 void **next_ptr
= elem
;
398 /* The pointer-based free list requires that the pointer be
399 * page-aligned. This is because we use the bottom 12 bits of the
400 * pointer to store a counter to solve the ABA concurrency problem.
402 assert(((uintptr_t)elem
& 0xfff) == 0);
407 VG_NOACCESS_WRITE(next_ptr
, PFL_PTR(current
));
408 unsigned new_count
= PFL_COUNT(current
) + 1;
409 void *new = PFL_PACK(elem
, new_count
);
410 old
= __sync_val_compare_and_swap(list
, current
, new);
411 } while (old
!= current
);
415 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
416 uint32_t center_bo_offset
, uint32_t size
);
419 anv_block_pool_init(struct anv_block_pool
*pool
,
420 struct anv_device
*device
,
421 uint64_t start_address
,
422 uint32_t initial_size
,
427 pool
->device
= device
;
428 pool
->bo_flags
= bo_flags
;
431 pool
->center_bo_offset
= 0;
432 pool
->start_address
= gen_canonical_address(start_address
);
435 /* This pointer will always point to the first BO in the list */
436 pool
->bo
= &pool
->bos
[0];
438 anv_bo_init(pool
->bo
, 0, 0);
440 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
)) {
441 /* Just make it 2GB up-front. The Linux kernel won't actually back it
442 * with pages until we either map and fault on one of them or we use
443 * userptr and send a chunk of it off to the GPU.
445 pool
->fd
= os_create_anonymous_file(BLOCK_POOL_MEMFD_SIZE
, "block pool");
447 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
452 if (!u_vector_init(&pool
->mmap_cleanups
,
453 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
455 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
459 pool
->state
.next
= 0;
461 pool
->back_state
.next
= 0;
462 pool
->back_state
.end
= 0;
464 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
465 if (result
!= VK_SUCCESS
)
466 goto fail_mmap_cleanups
;
468 /* Make the entire pool available in the front of the pool. If back
469 * allocation needs to use this space, the "ends" will be re-arranged.
471 pool
->state
.end
= pool
->size
;
476 u_vector_finish(&pool
->mmap_cleanups
);
478 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
485 anv_block_pool_finish(struct anv_block_pool
*pool
)
487 struct anv_mmap_cleanup
*cleanup
;
488 const bool use_softpin
= !!(pool
->bo_flags
& EXEC_OBJECT_PINNED
);
490 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
) {
492 anv_gem_munmap(cleanup
->map
, cleanup
->size
);
494 munmap(cleanup
->map
, cleanup
->size
);
496 if (cleanup
->gem_handle
)
497 anv_gem_close(pool
->device
, cleanup
->gem_handle
);
500 u_vector_finish(&pool
->mmap_cleanups
);
501 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
506 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
507 uint32_t center_bo_offset
, uint32_t size
)
511 struct anv_mmap_cleanup
*cleanup
;
512 const bool use_softpin
= !!(pool
->bo_flags
& EXEC_OBJECT_PINNED
);
514 /* Assert that we only ever grow the pool */
515 assert(center_bo_offset
>= pool
->back_state
.end
);
516 assert(size
- center_bo_offset
>= pool
->state
.end
);
518 /* Assert that we don't go outside the bounds of the memfd */
519 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
520 assert(use_softpin
||
521 size
- center_bo_offset
<=
522 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
524 cleanup
= u_vector_add(&pool
->mmap_cleanups
);
526 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
528 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
530 uint32_t newbo_size
= size
- pool
->size
;
532 gem_handle
= anv_gem_create(pool
->device
, newbo_size
);
533 map
= anv_gem_mmap(pool
->device
, gem_handle
, 0, newbo_size
, 0);
534 if (map
== MAP_FAILED
)
535 return vk_errorf(pool
->device
->instance
, pool
->device
,
536 VK_ERROR_MEMORY_MAP_FAILED
, "gem mmap failed: %m");
537 assert(center_bo_offset
== 0);
539 /* Just leak the old map until we destroy the pool. We can't munmap it
540 * without races or imposing locking on the block allocate fast path. On
541 * the whole the leaked maps adds up to less than the size of the
542 * current map. MAP_POPULATE seems like the right thing to do, but we
543 * should try to get some numbers.
545 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
546 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
547 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
548 if (map
== MAP_FAILED
)
549 return vk_errorf(pool
->device
->instance
, pool
->device
,
550 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
552 /* Now that we mapped the new memory, we can write the new
553 * center_bo_offset back into pool and update pool->map. */
554 pool
->center_bo_offset
= center_bo_offset
;
555 pool
->map
= map
+ center_bo_offset
;
556 gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
557 if (gem_handle
== 0) {
559 return vk_errorf(pool
->device
->instance
, pool
->device
,
560 VK_ERROR_TOO_MANY_OBJECTS
, "userptr failed: %m");
565 cleanup
->size
= use_softpin
? newbo_size
: size
;
566 cleanup
->gem_handle
= gem_handle
;
568 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
569 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
570 * always created as I915_CACHING_CACHED, which on non-LLC means
573 * On platforms that support softpin, we are not going to use userptr
574 * anymore, but we still want to rely on the snooped states. So make sure
575 * everything is set to I915_CACHING_CACHED.
577 if (!pool
->device
->info
.has_llc
)
578 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_CACHED
);
580 /* For block pool BOs we have to be a bit careful about where we place them
581 * in the GTT. There are two documented workarounds for state base address
582 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
583 * which state that those two base addresses do not support 48-bit
584 * addresses and need to be placed in the bottom 32-bit range.
585 * Unfortunately, this is not quite accurate.
587 * The real problem is that we always set the size of our state pools in
588 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
589 * likely significantly smaller. We do this because we do not no at the
590 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
591 * the pool during command buffer building so we don't actually have a
592 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
593 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
594 * as being out of bounds and returns zero. For dynamic state, this
595 * usually just leads to rendering corruptions, but shaders that are all
596 * zero hang the GPU immediately.
598 * The easiest solution to do is exactly what the bogus workarounds say to
599 * do: restrict these buffers to 32-bit addresses. We could also pin the
600 * BO to some particular location of our choosing, but that's significantly
601 * more work than just not setting a flag. So, we explicitly DO NOT set
602 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
609 assert(pool
->nbos
< ANV_MAX_BLOCK_POOL_BOS
);
612 /* With softpin, we add a new BO to the pool, and set its offset to right
613 * where the previous BO ends (the end of the pool).
615 bo
= &pool
->bos
[pool
->nbos
++];
616 bo_size
= newbo_size
;
617 bo_offset
= pool
->start_address
+ pool
->size
;
619 /* Without softpin, we just need one BO, and we already have a pointer to
620 * it. Simply "allocate" it from our array if we didn't do it before.
621 * The offset doesn't matter since we are not pinning the BO anyway.
630 anv_bo_init(bo
, gem_handle
, bo_size
);
631 bo
->offset
= bo_offset
;
632 bo
->flags
= pool
->bo_flags
;
639 /** Returns current memory map of the block pool.
641 * The returned pointer points to the map for the memory at the specified
642 * offset. The offset parameter is relative to the "center" of the block pool
643 * rather than the start of the block pool BO map.
646 anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
)
648 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
) {
649 struct anv_bo
*bo
= NULL
;
650 int32_t bo_offset
= 0;
651 anv_block_pool_foreach_bo(iter_bo
, pool
) {
652 if (offset
< bo_offset
+ iter_bo
->size
) {
656 bo_offset
+= iter_bo
->size
;
659 assert(offset
>= bo_offset
);
661 return bo
->map
+ (offset
- bo_offset
);
663 return pool
->map
+ offset
;
667 /** Grows and re-centers the block pool.
669 * We grow the block pool in one or both directions in such a way that the
670 * following conditions are met:
672 * 1) The size of the entire pool is always a power of two.
674 * 2) The pool only grows on both ends. Neither end can get
677 * 3) At the end of the allocation, we have about twice as much space
678 * allocated for each end as we have used. This way the pool doesn't
679 * grow too far in one direction or the other.
681 * 4) If the _alloc_back() has never been called, then the back portion of
682 * the pool retains a size of zero. (This makes it easier for users of
683 * the block pool that only want a one-sided pool.)
685 * 5) We have enough space allocated for at least one more block in
686 * whichever side `state` points to.
688 * 6) The center of the pool is always aligned to both the block_size of
689 * the pool and a 4K CPU page.
692 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
694 VkResult result
= VK_SUCCESS
;
696 pthread_mutex_lock(&pool
->device
->mutex
);
698 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
700 /* Gather a little usage information on the pool. Since we may have
701 * threadsd waiting in queue to get some storage while we resize, it's
702 * actually possible that total_used will be larger than old_size. In
703 * particular, block_pool_alloc() increments state->next prior to
704 * calling block_pool_grow, so this ensures that we get enough space for
705 * which ever side tries to grow the pool.
707 * We align to a page size because it makes it easier to do our
708 * calculations later in such a way that we state page-aigned.
710 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
711 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
712 uint32_t total_used
= front_used
+ back_used
;
714 assert(state
== &pool
->state
|| back_used
> 0);
716 uint32_t old_size
= pool
->size
;
718 /* The block pool is always initialized to a nonzero size and this function
719 * is always called after initialization.
721 assert(old_size
> 0);
723 /* The back_used and front_used may actually be smaller than the actual
724 * requirement because they are based on the next pointers which are
725 * updated prior to calling this function.
727 uint32_t back_required
= MAX2(back_used
, pool
->center_bo_offset
);
728 uint32_t front_required
= MAX2(front_used
, old_size
- pool
->center_bo_offset
);
730 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
731 /* If we're in this case then this isn't the firsta allocation and we
732 * already have enough space on both sides to hold double what we
733 * have allocated. There's nothing for us to do.
738 uint32_t size
= old_size
* 2;
739 while (size
< back_required
+ front_required
)
742 assert(size
> pool
->size
);
744 /* We compute a new center_bo_offset such that, when we double the size
745 * of the pool, we maintain the ratio of how much is used by each side.
746 * This way things should remain more-or-less balanced.
748 uint32_t center_bo_offset
;
749 if (back_used
== 0) {
750 /* If we're in this case then we have never called alloc_back(). In
751 * this case, we want keep the offset at 0 to make things as simple
752 * as possible for users that don't care about back allocations.
754 center_bo_offset
= 0;
756 /* Try to "center" the allocation based on how much is currently in
757 * use on each side of the center line.
759 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
761 /* Align down to a multiple of the page size */
762 center_bo_offset
&= ~(PAGE_SIZE
- 1);
764 assert(center_bo_offset
>= back_used
);
766 /* Make sure we don't shrink the back end of the pool */
767 if (center_bo_offset
< back_required
)
768 center_bo_offset
= back_required
;
770 /* Make sure that we don't shrink the front end of the pool */
771 if (size
- center_bo_offset
< front_required
)
772 center_bo_offset
= size
- front_required
;
775 assert(center_bo_offset
% PAGE_SIZE
== 0);
777 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
779 pool
->bo
->flags
= pool
->bo_flags
;
782 pthread_mutex_unlock(&pool
->device
->mutex
);
784 if (result
== VK_SUCCESS
) {
785 /* Return the appropriate new size. This function never actually
786 * updates state->next. Instead, we let the caller do that because it
787 * needs to do so in order to maintain its concurrency model.
789 if (state
== &pool
->state
) {
790 return pool
->size
- pool
->center_bo_offset
;
792 assert(pool
->center_bo_offset
> 0);
793 return pool
->center_bo_offset
;
801 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
802 struct anv_block_state
*pool_state
,
803 uint32_t block_size
, uint32_t *padding
)
805 struct anv_block_state state
, old
, new;
807 /* Most allocations won't generate any padding */
812 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
813 if (state
.next
+ block_size
<= state
.end
) {
815 } else if (state
.next
<= state
.end
) {
816 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
&& state
.next
< state
.end
) {
817 /* We need to grow the block pool, but still have some leftover
818 * space that can't be used by that particular allocation. So we
819 * add that as a "padding", and return it.
821 uint32_t leftover
= state
.end
- state
.next
;
823 /* If there is some leftover space in the pool, the caller must
826 assert(leftover
== 0 || padding
);
829 state
.next
+= leftover
;
832 /* We allocated the first block outside the pool so we have to grow
833 * the pool. pool_state->next acts a mutex: threads who try to
834 * allocate now will get block indexes above the current limit and
835 * hit futex_wait below.
837 new.next
= state
.next
+ block_size
;
839 new.end
= anv_block_pool_grow(pool
, pool_state
);
840 } while (new.end
< new.next
);
842 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
843 if (old
.next
!= state
.next
)
844 futex_wake(&pool_state
->end
, INT_MAX
);
847 futex_wait(&pool_state
->end
, state
.end
, NULL
);
854 anv_block_pool_alloc(struct anv_block_pool
*pool
,
855 uint32_t block_size
, uint32_t *padding
)
859 offset
= anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
, padding
);
864 /* Allocates a block out of the back of the block pool.
866 * This will allocated a block earlier than the "start" of the block pool.
867 * The offsets returned from this function will be negative but will still
868 * be correct relative to the block pool's map pointer.
870 * If you ever use anv_block_pool_alloc_back, then you will have to do
871 * gymnastics with the block pool's BO when doing relocations.
874 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
877 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
880 /* The offset we get out of anv_block_pool_alloc_new() is actually the
881 * number of bytes downwards from the middle to the end of the block.
882 * We need to turn it into a (negative) offset from the middle to the
883 * start of the block.
886 return -(offset
+ block_size
);
890 anv_state_pool_init(struct anv_state_pool
*pool
,
891 struct anv_device
*device
,
892 uint64_t start_address
,
896 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
900 if (result
!= VK_SUCCESS
)
903 result
= anv_state_table_init(&pool
->table
, device
, 64);
904 if (result
!= VK_SUCCESS
) {
905 anv_block_pool_finish(&pool
->block_pool
);
909 assert(util_is_power_of_two_or_zero(block_size
));
910 pool
->block_size
= block_size
;
911 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
912 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
913 pool
->buckets
[i
].free_list
= ANV_FREE_LIST_EMPTY
;
914 pool
->buckets
[i
].block
.next
= 0;
915 pool
->buckets
[i
].block
.end
= 0;
917 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
923 anv_state_pool_finish(struct anv_state_pool
*pool
)
925 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
926 anv_state_table_finish(&pool
->table
);
927 anv_block_pool_finish(&pool
->block_pool
);
931 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
932 struct anv_block_pool
*block_pool
,
937 struct anv_block_state block
, old
, new;
940 /* We don't always use anv_block_pool_alloc(), which would set *padding to
941 * zero for us. So if we have a pointer to padding, we must zero it out
942 * ourselves here, to make sure we always return some sensible value.
947 /* If our state is large, we don't need any sub-allocation from a block.
948 * Instead, we just grab whole (potentially large) blocks.
950 if (state_size
>= block_size
)
951 return anv_block_pool_alloc(block_pool
, state_size
, padding
);
954 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
956 if (block
.next
< block
.end
) {
958 } else if (block
.next
== block
.end
) {
959 offset
= anv_block_pool_alloc(block_pool
, block_size
, padding
);
960 new.next
= offset
+ state_size
;
961 new.end
= offset
+ block_size
;
962 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
963 if (old
.next
!= block
.next
)
964 futex_wake(&pool
->block
.end
, INT_MAX
);
967 futex_wait(&pool
->block
.end
, block
.end
, NULL
);
973 anv_state_pool_get_bucket(uint32_t size
)
975 unsigned size_log2
= ilog2_round_up(size
);
976 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
977 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
978 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
979 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
983 anv_state_pool_get_bucket_size(uint32_t bucket
)
985 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
986 return 1 << size_log2
;
989 /** Helper to push a chunk into the state table.
991 * It creates 'count' entries into the state table and update their sizes,
992 * offsets and maps, also pushing them as "free" states.
995 anv_state_pool_return_blocks(struct anv_state_pool
*pool
,
996 uint32_t chunk_offset
, uint32_t count
,
999 /* Disallow returning 0 chunks */
1002 /* Make sure we always return chunks aligned to the block_size */
1003 assert(chunk_offset
% block_size
== 0);
1006 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &st_idx
, count
);
1007 assert(result
== VK_SUCCESS
);
1008 for (int i
= 0; i
< count
; i
++) {
1009 /* update states that were added back to the state table */
1010 struct anv_state
*state_i
= anv_state_table_get(&pool
->table
,
1012 state_i
->alloc_size
= block_size
;
1013 state_i
->offset
= chunk_offset
+ block_size
* i
;
1014 state_i
->map
= anv_block_pool_map(&pool
->block_pool
, state_i
->offset
);
1017 uint32_t block_bucket
= anv_state_pool_get_bucket(block_size
);
1018 anv_free_list_push(&pool
->buckets
[block_bucket
].free_list
,
1019 &pool
->table
, st_idx
, count
);
1022 /** Returns a chunk of memory back to the state pool.
1024 * Do a two-level split. If chunk_size is bigger than divisor
1025 * (pool->block_size), we return as many divisor sized blocks as we can, from
1026 * the end of the chunk.
1028 * The remaining is then split into smaller blocks (starting at small_size if
1029 * it is non-zero), with larger blocks always being taken from the end of the
1033 anv_state_pool_return_chunk(struct anv_state_pool
*pool
,
1034 uint32_t chunk_offset
, uint32_t chunk_size
,
1035 uint32_t small_size
)
1037 uint32_t divisor
= pool
->block_size
;
1038 uint32_t nblocks
= chunk_size
/ divisor
;
1039 uint32_t rest
= chunk_size
- nblocks
* divisor
;
1042 /* First return divisor aligned and sized chunks. We start returning
1043 * larger blocks from the end fo the chunk, since they should already be
1044 * aligned to divisor. Also anv_state_pool_return_blocks() only accepts
1047 uint32_t offset
= chunk_offset
+ rest
;
1048 anv_state_pool_return_blocks(pool
, offset
, nblocks
, divisor
);
1054 if (small_size
> 0 && small_size
< divisor
)
1055 divisor
= small_size
;
1057 uint32_t min_size
= 1 << ANV_MIN_STATE_SIZE_LOG2
;
1059 /* Just as before, return larger divisor aligned blocks from the end of the
1062 while (chunk_size
> 0 && divisor
>= min_size
) {
1063 nblocks
= chunk_size
/ divisor
;
1064 rest
= chunk_size
- nblocks
* divisor
;
1066 anv_state_pool_return_blocks(pool
, chunk_offset
+ rest
,
1074 static struct anv_state
1075 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
1076 uint32_t size
, uint32_t align
)
1078 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
1080 struct anv_state
*state
;
1081 uint32_t alloc_size
= anv_state_pool_get_bucket_size(bucket
);
1084 /* Try free list first. */
1085 state
= anv_free_list_pop(&pool
->buckets
[bucket
].free_list
,
1088 assert(state
->offset
>= 0);
1092 /* Try to grab a chunk from some larger bucket and split it up */
1093 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
1094 state
= anv_free_list_pop(&pool
->buckets
[b
].free_list
, &pool
->table
);
1096 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
1097 int32_t chunk_offset
= state
->offset
;
1099 /* First lets update the state we got to its new size. offset and map
1102 state
->alloc_size
= alloc_size
;
1104 /* Now return the unused part of the chunk back to the pool as free
1107 * There are a couple of options as to what we do with it:
1109 * 1) We could fully split the chunk into state.alloc_size sized
1110 * pieces. However, this would mean that allocating a 16B
1111 * state could potentially split a 2MB chunk into 512K smaller
1112 * chunks. This would lead to unnecessary fragmentation.
1114 * 2) The classic "buddy allocator" method would have us split the
1115 * chunk in half and return one half. Then we would split the
1116 * remaining half in half and return one half, and repeat as
1117 * needed until we get down to the size we want. However, if
1118 * you are allocating a bunch of the same size state (which is
1119 * the common case), this means that every other allocation has
1120 * to go up a level and every fourth goes up two levels, etc.
1121 * This is not nearly as efficient as it could be if we did a
1122 * little more work up-front.
1124 * 3) Split the difference between (1) and (2) by doing a
1125 * two-level split. If it's bigger than some fixed block_size,
1126 * we split it into block_size sized chunks and return all but
1127 * one of them. Then we split what remains into
1128 * state.alloc_size sized chunks and return them.
1130 * We choose something close to option (3), which is implemented with
1131 * anv_state_pool_return_chunk(). That is done by returning the
1132 * remaining of the chunk, with alloc_size as a hint of the size that
1133 * we want the smaller chunk split into.
1135 anv_state_pool_return_chunk(pool
, chunk_offset
+ alloc_size
,
1136 chunk_size
- alloc_size
, alloc_size
);
1142 offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
1147 /* Everytime we allocate a new state, add it to the state pool */
1149 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1150 assert(result
== VK_SUCCESS
);
1152 state
= anv_state_table_get(&pool
->table
, idx
);
1153 state
->offset
= offset
;
1154 state
->alloc_size
= alloc_size
;
1155 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
);
1158 uint32_t return_offset
= offset
- padding
;
1159 anv_state_pool_return_chunk(pool
, return_offset
, padding
, 0);
1167 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
1170 return ANV_STATE_NULL
;
1172 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
1173 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
1178 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
1180 struct anv_state
*state
;
1181 uint32_t alloc_size
= pool
->block_size
;
1183 state
= anv_free_list_pop(&pool
->back_alloc_free_list
, &pool
->table
);
1185 assert(state
->offset
< 0);
1190 offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
1193 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1194 assert(result
== VK_SUCCESS
);
1196 state
= anv_state_table_get(&pool
->table
, idx
);
1197 state
->offset
= offset
;
1198 state
->alloc_size
= alloc_size
;
1199 state
->map
= anv_block_pool_map(&pool
->block_pool
, state
->offset
);
1202 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
->map
, state
->alloc_size
));
1207 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
1209 assert(util_is_power_of_two_or_zero(state
.alloc_size
));
1210 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
1212 if (state
.offset
< 0) {
1213 assert(state
.alloc_size
== pool
->block_size
);
1214 anv_free_list_push(&pool
->back_alloc_free_list
,
1215 &pool
->table
, state
.idx
, 1);
1217 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
1218 &pool
->table
, state
.idx
, 1);
1223 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
1225 if (state
.alloc_size
== 0)
1228 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
1229 anv_state_pool_free_no_vg(pool
, state
);
1232 struct anv_state_stream_block
{
1233 struct anv_state block
;
1235 /* The next block */
1236 struct anv_state_stream_block
*next
;
1238 #ifdef HAVE_VALGRIND
1239 /* A pointer to the first user-allocated thing in this block. This is
1240 * what valgrind sees as the start of the block.
1246 /* The state stream allocator is a one-shot, single threaded allocator for
1247 * variable sized blocks. We use it for allocating dynamic state.
1250 anv_state_stream_init(struct anv_state_stream
*stream
,
1251 struct anv_state_pool
*state_pool
,
1252 uint32_t block_size
)
1254 stream
->state_pool
= state_pool
;
1255 stream
->block_size
= block_size
;
1257 stream
->block
= ANV_STATE_NULL
;
1259 stream
->block_list
= NULL
;
1261 /* Ensure that next + whatever > block_size. This way the first call to
1262 * state_stream_alloc fetches a new block.
1264 stream
->next
= block_size
;
1266 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
1270 anv_state_stream_finish(struct anv_state_stream
*stream
)
1272 struct anv_state_stream_block
*next
= stream
->block_list
;
1273 while (next
!= NULL
) {
1274 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
1275 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
1276 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
1277 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
1281 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
1285 anv_state_stream_alloc(struct anv_state_stream
*stream
,
1286 uint32_t size
, uint32_t alignment
)
1289 return ANV_STATE_NULL
;
1291 assert(alignment
<= PAGE_SIZE
);
1293 uint32_t offset
= align_u32(stream
->next
, alignment
);
1294 if (offset
+ size
> stream
->block
.alloc_size
) {
1295 uint32_t block_size
= stream
->block_size
;
1296 if (block_size
< size
)
1297 block_size
= round_to_power_of_two(size
);
1299 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
1300 block_size
, PAGE_SIZE
);
1302 struct anv_state_stream_block
*sb
= stream
->block
.map
;
1303 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
1304 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
1305 stream
->block_list
= sb
;
1306 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
1308 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
1310 /* Reset back to the start plus space for the header */
1311 stream
->next
= sizeof(*sb
);
1313 offset
= align_u32(stream
->next
, alignment
);
1314 assert(offset
+ size
<= stream
->block
.alloc_size
);
1317 struct anv_state state
= stream
->block
;
1318 state
.offset
+= offset
;
1319 state
.alloc_size
= size
;
1320 state
.map
+= offset
;
1322 stream
->next
= offset
+ size
;
1324 #ifdef HAVE_VALGRIND
1325 struct anv_state_stream_block
*sb
= stream
->block_list
;
1326 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
1327 if (vg_ptr
== NULL
) {
1329 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
1330 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
1332 void *state_end
= state
.map
+ state
.alloc_size
;
1333 /* This only updates the mempool. The newly allocated chunk is still
1334 * marked as NOACCESS. */
1335 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
1336 /* Mark the newly allocated chunk as undefined */
1337 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
1344 struct bo_pool_bo_link
{
1345 struct bo_pool_bo_link
*next
;
1350 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
,
1353 pool
->device
= device
;
1354 pool
->bo_flags
= bo_flags
;
1355 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
1357 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
1361 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
1363 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1364 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
1365 while (link
!= NULL
) {
1366 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
1368 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
1369 anv_vma_free(pool
->device
, &link_copy
.bo
);
1370 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
1371 link
= link_copy
.next
;
1375 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1379 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
1383 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1384 const unsigned pow2_size
= 1 << size_log2
;
1385 const unsigned bucket
= size_log2
- 12;
1386 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1388 void *next_free_void
;
1389 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
1390 struct bo_pool_bo_link
*next_free
= next_free_void
;
1391 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
1392 assert(bo
->gem_handle
);
1393 assert(bo
->map
== next_free
);
1394 assert(size
<= bo
->size
);
1396 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1401 struct anv_bo new_bo
;
1403 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
1404 if (result
!= VK_SUCCESS
)
1407 new_bo
.flags
= pool
->bo_flags
;
1409 if (!anv_vma_alloc(pool
->device
, &new_bo
))
1410 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1412 assert(new_bo
.size
== pow2_size
);
1414 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
1415 if (new_bo
.map
== MAP_FAILED
) {
1416 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
1417 anv_vma_free(pool
->device
, &new_bo
);
1418 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1421 /* We are removing the state flushes, so lets make sure that these buffers
1422 * are cached/snooped.
1424 if (!pool
->device
->info
.has_llc
) {
1425 anv_gem_set_caching(pool
->device
, new_bo
.gem_handle
,
1426 I915_CACHING_CACHED
);
1431 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1437 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
1439 /* Make a copy in case the anv_bo happens to be storred in the BO */
1440 struct anv_bo bo
= *bo_in
;
1442 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
1444 struct bo_pool_bo_link
*link
= bo
.map
;
1445 VG_NOACCESS_WRITE(&link
->bo
, bo
);
1447 assert(util_is_power_of_two_or_zero(bo
.size
));
1448 const unsigned size_log2
= ilog2_round_up(bo
.size
);
1449 const unsigned bucket
= size_log2
- 12;
1450 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1452 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
1458 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1460 memset(pool
, 0, sizeof(*pool
));
1464 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1466 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1467 for (unsigned i
= 0; i
< 16; i
++) {
1468 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
1469 if (bo
->exists
> 0) {
1470 anv_vma_free(device
, &bo
->bo
);
1471 anv_gem_close(device
, bo
->bo
.gem_handle
);
1478 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1479 gl_shader_stage stage
, unsigned per_thread_scratch
)
1481 if (per_thread_scratch
== 0)
1484 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1485 assert(scratch_size_log2
< 16);
1487 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
1489 /* We can use "exists" to shortcut and ignore the critical section */
1493 pthread_mutex_lock(&device
->mutex
);
1495 __sync_synchronize();
1497 pthread_mutex_unlock(&device
->mutex
);
1501 const struct anv_physical_device
*physical_device
=
1502 &device
->instance
->physicalDevice
;
1503 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1505 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
1507 unsigned scratch_ids_per_subslice
;
1508 if (devinfo
->gen
>= 11) {
1509 /* The MEDIA_VFE_STATE docs say:
1511 * "Starting with this configuration, the Maximum Number of
1512 * Threads must be set to (#EU * 8) for GPGPU dispatches.
1514 * Although there are only 7 threads per EU in the configuration,
1515 * the FFTID is calculated as if there are 8 threads per EU,
1516 * which in turn requires a larger amount of Scratch Space to be
1517 * allocated by the driver."
1519 scratch_ids_per_subslice
= 8 * 8;
1520 } else if (devinfo
->is_haswell
) {
1521 /* WaCSScratchSize:hsw
1523 * Haswell's scratch space address calculation appears to be sparse
1524 * rather than tightly packed. The Thread ID has bits indicating
1525 * which subslice, EU within a subslice, and thread within an EU it
1526 * is. There's a maximum of two slices and two subslices, so these
1527 * can be stored with a single bit. Even though there are only 10 EUs
1528 * per subslice, this is stored in 4 bits, so there's an effective
1529 * maximum value of 16 EUs. Similarly, although there are only 7
1530 * threads per EU, this is stored in a 3 bit number, giving an
1531 * effective maximum value of 8 threads per EU.
1533 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1534 * number of threads per subslice.
1536 scratch_ids_per_subslice
= 16 * 8;
1537 } else if (devinfo
->is_cherryview
) {
1538 /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
1539 * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
1542 scratch_ids_per_subslice
= 8 * 7;
1544 scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
1547 uint32_t max_threads
[] = {
1548 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1549 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1550 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1551 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1552 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1553 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1556 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1558 anv_bo_init_new(&bo
->bo
, device
, size
);
1560 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1561 * are still relative to the general state base address. When we emit
1562 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1563 * to the maximum (1 page under 4GB). This allows us to just place the
1564 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1565 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1566 * However, in order to do so, we need to ensure that the kernel does not
1567 * place the scratch BO above the 32-bit boundary.
1569 * NOTE: Technically, it can't go "anywhere" because the top page is off
1570 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1571 * kernel allocates space using
1573 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1575 * so nothing will ever touch the top page.
1577 assert(!(bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
));
1579 if (device
->instance
->physicalDevice
.has_exec_async
)
1580 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1582 if (device
->instance
->physicalDevice
.use_softpin
)
1583 bo
->bo
.flags
|= EXEC_OBJECT_PINNED
;
1585 anv_vma_alloc(device
, &bo
->bo
);
1587 /* Set the exists last because it may be read by other threads */
1588 __sync_synchronize();
1591 pthread_mutex_unlock(&device
->mutex
);
1597 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1599 util_sparse_array_init(&cache
->bo_map
, sizeof(struct anv_bo
), 1024);
1601 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1602 util_sparse_array_finish(&cache
->bo_map
);
1603 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1604 "pthread_mutex_init failed: %m");
1611 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1613 util_sparse_array_finish(&cache
->bo_map
);
1614 pthread_mutex_destroy(&cache
->mutex
);
1617 static struct anv_bo
*
1618 anv_bo_cache_lookup(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1620 return util_sparse_array_get(&cache
->bo_map
, gem_handle
);
1623 #define ANV_BO_CACHE_SUPPORTED_FLAGS \
1624 (EXEC_OBJECT_WRITE | \
1625 EXEC_OBJECT_ASYNC | \
1626 EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
1630 anv_bo_cache_alloc(struct anv_device
*device
,
1631 struct anv_bo_cache
*cache
,
1632 uint64_t size
, uint64_t bo_flags
,
1634 struct anv_bo
**bo_out
)
1636 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1638 /* The kernel is going to give us whole pages anyway */
1639 size
= align_u64(size
, 4096);
1641 struct anv_bo new_bo
;
1642 VkResult result
= anv_bo_init_new(&new_bo
, device
, size
);
1643 if (result
!= VK_SUCCESS
)
1646 new_bo
.flags
= bo_flags
;
1647 new_bo
.is_external
= is_external
;
1649 if (!anv_vma_alloc(device
, &new_bo
)) {
1650 anv_gem_close(device
, new_bo
.gem_handle
);
1651 return vk_errorf(device
->instance
, NULL
,
1652 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1653 "failed to allocate virtual address for BO");
1656 assert(new_bo
.gem_handle
);
1658 /* If we just got this gem_handle from anv_bo_init_new then we know no one
1659 * else is touching this BO at the moment so we don't need to lock here.
1661 struct anv_bo
*bo
= anv_bo_cache_lookup(cache
, new_bo
.gem_handle
);
1670 anv_bo_cache_import_host_ptr(struct anv_device
*device
,
1671 struct anv_bo_cache
*cache
,
1672 void *host_ptr
, uint32_t size
,
1673 uint64_t bo_flags
, struct anv_bo
**bo_out
)
1675 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1677 uint32_t gem_handle
= anv_gem_userptr(device
, host_ptr
, size
);
1679 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1681 pthread_mutex_lock(&cache
->mutex
);
1683 struct anv_bo
*bo
= anv_bo_cache_lookup(cache
, gem_handle
);
1684 if (bo
->refcount
> 0) {
1685 /* VK_EXT_external_memory_host doesn't require handling importing the
1686 * same pointer twice at the same time, but we don't get in the way. If
1687 * kernel gives us the same gem_handle, only succeed if the flags match.
1689 assert(bo
->gem_handle
== gem_handle
);
1690 if (bo_flags
!= bo
->flags
) {
1691 pthread_mutex_unlock(&cache
->mutex
);
1692 return vk_errorf(device
->instance
, NULL
,
1693 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1694 "same host pointer imported two different ways");
1696 __sync_fetch_and_add(&bo
->refcount
, 1);
1698 struct anv_bo new_bo
;
1699 anv_bo_init(&new_bo
, gem_handle
, size
);
1700 new_bo
.flags
= bo_flags
;
1701 new_bo
.is_external
= true;
1703 if (!anv_vma_alloc(device
, &new_bo
)) {
1704 anv_gem_close(device
, new_bo
.gem_handle
);
1705 pthread_mutex_unlock(&cache
->mutex
);
1706 return vk_errorf(device
->instance
, NULL
,
1707 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1708 "failed to allocate virtual address for BO");
1714 pthread_mutex_unlock(&cache
->mutex
);
1721 anv_bo_cache_import(struct anv_device
*device
,
1722 struct anv_bo_cache
*cache
,
1723 int fd
, uint64_t bo_flags
,
1724 struct anv_bo
**bo_out
)
1726 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1728 pthread_mutex_lock(&cache
->mutex
);
1730 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1732 pthread_mutex_unlock(&cache
->mutex
);
1733 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1736 struct anv_bo
*bo
= anv_bo_cache_lookup(cache
, gem_handle
);
1737 if (bo
->refcount
> 0) {
1738 /* We have to be careful how we combine flags so that it makes sense.
1739 * Really, though, if we get to this case and it actually matters, the
1740 * client has imported a BO twice in different ways and they get what
1743 uint64_t new_flags
= 0;
1744 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_WRITE
;
1745 new_flags
|= (bo
->flags
& bo_flags
) & EXEC_OBJECT_ASYNC
;
1746 new_flags
|= (bo
->flags
& bo_flags
) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1747 new_flags
|= (bo
->flags
| bo_flags
) & EXEC_OBJECT_PINNED
;
1749 /* It's theoretically possible for a BO to get imported such that it's
1750 * both pinned and not pinned. The only way this can happen is if it
1751 * gets imported as both a semaphore and a memory object and that would
1752 * be an application error. Just fail out in that case.
1754 if ((bo
->flags
& EXEC_OBJECT_PINNED
) !=
1755 (bo_flags
& EXEC_OBJECT_PINNED
)) {
1756 pthread_mutex_unlock(&cache
->mutex
);
1757 return vk_errorf(device
->instance
, NULL
,
1758 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1759 "The same BO was imported two different ways");
1762 /* It's also theoretically possible that someone could export a BO from
1763 * one heap and import it into another or to import the same BO into two
1764 * different heaps. If this happens, we could potentially end up both
1765 * allowing and disallowing 48-bit addresses. There's not much we can
1766 * do about it if we're pinning so we just throw an error and hope no
1767 * app is actually that stupid.
1769 if ((new_flags
& EXEC_OBJECT_PINNED
) &&
1770 (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) !=
1771 (bo_flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
)) {
1772 pthread_mutex_unlock(&cache
->mutex
);
1773 return vk_errorf(device
->instance
, NULL
,
1774 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1775 "The same BO was imported on two different heaps");
1778 bo
->flags
= new_flags
;
1780 __sync_fetch_and_add(&bo
->refcount
, 1);
1782 off_t size
= lseek(fd
, 0, SEEK_END
);
1783 if (size
== (off_t
)-1) {
1784 anv_gem_close(device
, gem_handle
);
1785 pthread_mutex_unlock(&cache
->mutex
);
1786 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1789 struct anv_bo new_bo
;
1790 anv_bo_init(&new_bo
, gem_handle
, size
);
1791 new_bo
.flags
= bo_flags
;
1792 new_bo
.is_external
= true;
1794 if (!anv_vma_alloc(device
, &new_bo
)) {
1795 anv_gem_close(device
, new_bo
.gem_handle
);
1796 pthread_mutex_unlock(&cache
->mutex
);
1797 return vk_errorf(device
->instance
, NULL
,
1798 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1799 "failed to allocate virtual address for BO");
1805 pthread_mutex_unlock(&cache
->mutex
);
1812 anv_bo_cache_export(struct anv_device
*device
,
1813 struct anv_bo_cache
*cache
,
1814 struct anv_bo
*bo
, int *fd_out
)
1816 assert(anv_bo_cache_lookup(cache
, bo
->gem_handle
) == bo
);
1818 /* This BO must have been flagged external in order for us to be able
1819 * to export it. This is done based on external options passed into
1820 * anv_AllocateMemory.
1822 assert(bo
->is_external
);
1824 int fd
= anv_gem_handle_to_fd(device
, bo
->gem_handle
);
1826 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1834 atomic_dec_not_one(uint32_t *counter
)
1843 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1852 anv_bo_cache_release(struct anv_device
*device
,
1853 struct anv_bo_cache
*cache
,
1856 assert(anv_bo_cache_lookup(cache
, bo
->gem_handle
) == bo
);
1858 /* Try to decrement the counter but don't go below one. If this succeeds
1859 * then the refcount has been decremented and we are not the last
1862 if (atomic_dec_not_one(&bo
->refcount
))
1865 pthread_mutex_lock(&cache
->mutex
);
1867 /* We are probably the last reference since our attempt to decrement above
1868 * failed. However, we can't actually know until we are inside the mutex.
1869 * Otherwise, someone could import the BO between the decrement and our
1872 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1873 /* Turns out we're not the last reference. Unlock and bail. */
1874 pthread_mutex_unlock(&cache
->mutex
);
1877 assert(bo
->refcount
== 0);
1880 anv_gem_munmap(bo
->map
, bo
->size
);
1882 anv_vma_free(device
, bo
);
1884 anv_gem_close(device
, bo
->gem_handle
);
1886 /* Don't unlock until we've actually closed the BO. The whole point of
1887 * the BO cache is to ensure that we correctly handle races with creating
1888 * and releasing GEM handles and we don't want to let someone import the BO
1889 * again between mutex unlock and closing the GEM handle.
1891 pthread_mutex_unlock(&cache
->mutex
);