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/hash_table.h"
33 #include "util/simple_mtx.h"
34 #include "util/anon_file.h"
37 #define VG_NOACCESS_READ(__ptr) ({ \
38 VALGRIND_MAKE_MEM_DEFINED((__ptr), sizeof(*(__ptr))); \
39 __typeof(*(__ptr)) __val = *(__ptr); \
40 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr)));\
43 #define VG_NOACCESS_WRITE(__ptr, __val) ({ \
44 VALGRIND_MAKE_MEM_UNDEFINED((__ptr), sizeof(*(__ptr))); \
46 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr))); \
49 #define VG_NOACCESS_READ(__ptr) (*(__ptr))
50 #define VG_NOACCESS_WRITE(__ptr, __val) (*(__ptr) = (__val))
55 * - Lock free (except when resizing underlying bos)
57 * - Constant time allocation with typically only one atomic
59 * - Multiple allocation sizes without fragmentation
61 * - Can grow while keeping addresses and offset of contents stable
63 * - All allocations within one bo so we can point one of the
64 * STATE_BASE_ADDRESS pointers at it.
66 * The overall design is a two-level allocator: top level is a fixed size, big
67 * block (8k) allocator, which operates out of a bo. Allocation is done by
68 * either pulling a block from the free list or growing the used range of the
69 * bo. Growing the range may run out of space in the bo which we then need to
70 * grow. Growing the bo is tricky in a multi-threaded, lockless environment:
71 * we need to keep all pointers and contents in the old map valid. GEM bos in
72 * general can't grow, but we use a trick: we create a memfd and use ftruncate
73 * to grow it as necessary. We mmap the new size and then create a gem bo for
74 * it using the new gem userptr ioctl. Without heavy-handed locking around
75 * our allocation fast-path, there isn't really a way to munmap the old mmap,
76 * so we just keep it around until garbage collection time. While the block
77 * allocator is lockless for normal operations, we block other threads trying
78 * to allocate while we're growing the map. It sholdn't happen often, and
79 * growing is fast anyway.
81 * At the next level we can use various sub-allocators. The state pool is a
82 * pool of smaller, fixed size objects, which operates much like the block
83 * pool. It uses a free list for freeing objects, but when it runs out of
84 * space it just allocates a new block from the block pool. This allocator is
85 * intended for longer lived state objects such as SURFACE_STATE and most
86 * other persistent state objects in the API. We may need to track more info
87 * with these object and a pointer back to the CPU object (eg VkImage). In
88 * those cases we just allocate a slightly bigger object and put the extra
89 * state after the GPU state object.
91 * The state stream allocator works similar to how the i965 DRI driver streams
92 * all its state. Even with Vulkan, we need to emit transient state (whether
93 * surface state base or dynamic state base), and for that we can just get a
94 * block and fill it up. These cases are local to a command buffer and the
95 * sub-allocator need not be thread safe. The streaming allocator gets a new
96 * block when it runs out of space and chains them together so they can be
100 /* Allocations are always at least 64 byte aligned, so 1 is an invalid value.
101 * We use it to indicate the free list is empty. */
102 #define EMPTY UINT32_MAX
104 #define PAGE_SIZE 4096
106 struct anv_mmap_cleanup
{
112 #define ANV_MMAP_CLEANUP_INIT ((struct anv_mmap_cleanup){0})
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
,
424 pool
->device
= device
;
425 pool
->bo_flags
= bo_flags
;
428 pool
->center_bo_offset
= 0;
429 pool
->start_address
= gen_canonical_address(start_address
);
432 /* This pointer will always point to the first BO in the list */
433 pool
->bo
= &pool
->bos
[0];
435 anv_bo_init(pool
->bo
, 0, 0);
437 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
)) {
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
);
449 if (!u_vector_init(&pool
->mmap_cleanups
,
450 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
452 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
456 pool
->state
.next
= 0;
458 pool
->back_state
.next
= 0;
459 pool
->back_state
.end
= 0;
461 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
462 if (result
!= VK_SUCCESS
)
463 goto fail_mmap_cleanups
;
465 /* Make the entire pool available in the front of the pool. If back
466 * allocation needs to use this space, the "ends" will be re-arranged.
468 pool
->state
.end
= pool
->size
;
473 u_vector_finish(&pool
->mmap_cleanups
);
475 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
482 anv_block_pool_finish(struct anv_block_pool
*pool
)
484 struct anv_mmap_cleanup
*cleanup
;
485 const bool use_softpin
= !!(pool
->bo_flags
& EXEC_OBJECT_PINNED
);
487 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
) {
489 anv_gem_munmap(cleanup
->map
, cleanup
->size
);
491 munmap(cleanup
->map
, cleanup
->size
);
493 if (cleanup
->gem_handle
)
494 anv_gem_close(pool
->device
, cleanup
->gem_handle
);
497 u_vector_finish(&pool
->mmap_cleanups
);
498 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
503 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
504 uint32_t center_bo_offset
, uint32_t size
)
508 struct anv_mmap_cleanup
*cleanup
;
509 const bool use_softpin
= !!(pool
->bo_flags
& EXEC_OBJECT_PINNED
);
511 /* Assert that we only ever grow the pool */
512 assert(center_bo_offset
>= pool
->back_state
.end
);
513 assert(size
- center_bo_offset
>= pool
->state
.end
);
515 /* Assert that we don't go outside the bounds of the memfd */
516 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
517 assert(use_softpin
||
518 size
- center_bo_offset
<=
519 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
521 cleanup
= u_vector_add(&pool
->mmap_cleanups
);
523 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
525 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
527 uint32_t newbo_size
= size
- pool
->size
;
529 gem_handle
= anv_gem_create(pool
->device
, newbo_size
);
530 map
= anv_gem_mmap(pool
->device
, gem_handle
, 0, newbo_size
, 0);
531 if (map
== MAP_FAILED
)
532 return vk_errorf(pool
->device
->instance
, pool
->device
,
533 VK_ERROR_MEMORY_MAP_FAILED
, "gem mmap failed: %m");
534 assert(center_bo_offset
== 0);
536 /* Just leak the old map until we destroy the pool. We can't munmap it
537 * without races or imposing locking on the block allocate fast path. On
538 * the whole the leaked maps adds up to less than the size of the
539 * current map. MAP_POPULATE seems like the right thing to do, but we
540 * should try to get some numbers.
542 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
543 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
544 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
545 if (map
== MAP_FAILED
)
546 return vk_errorf(pool
->device
->instance
, pool
->device
,
547 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
549 /* Now that we mapped the new memory, we can write the new
550 * center_bo_offset back into pool and update pool->map. */
551 pool
->center_bo_offset
= center_bo_offset
;
552 pool
->map
= map
+ center_bo_offset
;
553 gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
554 if (gem_handle
== 0) {
556 return vk_errorf(pool
->device
->instance
, pool
->device
,
557 VK_ERROR_TOO_MANY_OBJECTS
, "userptr failed: %m");
562 cleanup
->size
= use_softpin
? newbo_size
: size
;
563 cleanup
->gem_handle
= gem_handle
;
565 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
566 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
567 * always created as I915_CACHING_CACHED, which on non-LLC means
570 * On platforms that support softpin, we are not going to use userptr
571 * anymore, but we still want to rely on the snooped states. So make sure
572 * everything is set to I915_CACHING_CACHED.
574 if (!pool
->device
->info
.has_llc
)
575 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_CACHED
);
577 /* For block pool BOs we have to be a bit careful about where we place them
578 * in the GTT. There are two documented workarounds for state base address
579 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
580 * which state that those two base addresses do not support 48-bit
581 * addresses and need to be placed in the bottom 32-bit range.
582 * Unfortunately, this is not quite accurate.
584 * The real problem is that we always set the size of our state pools in
585 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
586 * likely significantly smaller. We do this because we do not no at the
587 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
588 * the pool during command buffer building so we don't actually have a
589 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
590 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
591 * as being out of bounds and returns zero. For dynamic state, this
592 * usually just leads to rendering corruptions, but shaders that are all
593 * zero hang the GPU immediately.
595 * The easiest solution to do is exactly what the bogus workarounds say to
596 * do: restrict these buffers to 32-bit addresses. We could also pin the
597 * BO to some particular location of our choosing, but that's significantly
598 * more work than just not setting a flag. So, we explicitly DO NOT set
599 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
606 assert(pool
->nbos
< ANV_MAX_BLOCK_POOL_BOS
);
609 /* With softpin, we add a new BO to the pool, and set its offset to right
610 * where the previous BO ends (the end of the pool).
612 bo
= &pool
->bos
[pool
->nbos
++];
613 bo_size
= newbo_size
;
614 bo_offset
= pool
->start_address
+ pool
->size
;
616 /* Without softpin, we just need one BO, and we already have a pointer to
617 * it. Simply "allocate" it from our array if we didn't do it before.
618 * The offset doesn't matter since we are not pinning the BO anyway.
627 anv_bo_init(bo
, gem_handle
, bo_size
);
628 bo
->offset
= bo_offset
;
629 bo
->flags
= pool
->bo_flags
;
636 static struct anv_bo
*
637 anv_block_pool_get_bo(struct anv_block_pool
*pool
, int32_t *offset
)
639 struct anv_bo
*bo
, *bo_found
= NULL
;
640 int32_t cur_offset
= 0;
644 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
647 anv_block_pool_foreach_bo(bo
, pool
) {
648 if (*offset
< cur_offset
+ bo
->size
) {
652 cur_offset
+= bo
->size
;
655 assert(bo_found
!= NULL
);
656 *offset
-= cur_offset
;
661 /** Returns current memory map of the block pool.
663 * The returned pointer points to the map for the memory at the specified
664 * offset. The offset parameter is relative to the "center" of the block pool
665 * rather than the start of the block pool BO map.
668 anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
)
670 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
) {
671 struct anv_bo
*bo
= anv_block_pool_get_bo(pool
, &offset
);
672 return bo
->map
+ offset
;
674 return pool
->map
+ offset
;
678 /** Grows and re-centers the block pool.
680 * We grow the block pool in one or both directions in such a way that the
681 * following conditions are met:
683 * 1) The size of the entire pool is always a power of two.
685 * 2) The pool only grows on both ends. Neither end can get
688 * 3) At the end of the allocation, we have about twice as much space
689 * allocated for each end as we have used. This way the pool doesn't
690 * grow too far in one direction or the other.
692 * 4) If the _alloc_back() has never been called, then the back portion of
693 * the pool retains a size of zero. (This makes it easier for users of
694 * the block pool that only want a one-sided pool.)
696 * 5) We have enough space allocated for at least one more block in
697 * whichever side `state` points to.
699 * 6) The center of the pool is always aligned to both the block_size of
700 * the pool and a 4K CPU page.
703 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
705 VkResult result
= VK_SUCCESS
;
707 pthread_mutex_lock(&pool
->device
->mutex
);
709 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
711 /* Gather a little usage information on the pool. Since we may have
712 * threadsd waiting in queue to get some storage while we resize, it's
713 * actually possible that total_used will be larger than old_size. In
714 * particular, block_pool_alloc() increments state->next prior to
715 * calling block_pool_grow, so this ensures that we get enough space for
716 * which ever side tries to grow the pool.
718 * We align to a page size because it makes it easier to do our
719 * calculations later in such a way that we state page-aigned.
721 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
722 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
723 uint32_t total_used
= front_used
+ back_used
;
725 assert(state
== &pool
->state
|| back_used
> 0);
727 uint32_t old_size
= pool
->size
;
729 /* The block pool is always initialized to a nonzero size and this function
730 * is always called after initialization.
732 assert(old_size
> 0);
734 /* The back_used and front_used may actually be smaller than the actual
735 * requirement because they are based on the next pointers which are
736 * updated prior to calling this function.
738 uint32_t back_required
= MAX2(back_used
, pool
->center_bo_offset
);
739 uint32_t front_required
= MAX2(front_used
, old_size
- pool
->center_bo_offset
);
741 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
742 /* If we're in this case then this isn't the firsta allocation and we
743 * already have enough space on both sides to hold double what we
744 * have allocated. There's nothing for us to do.
749 uint32_t size
= old_size
* 2;
750 while (size
< back_required
+ front_required
)
753 assert(size
> pool
->size
);
755 /* We compute a new center_bo_offset such that, when we double the size
756 * of the pool, we maintain the ratio of how much is used by each side.
757 * This way things should remain more-or-less balanced.
759 uint32_t center_bo_offset
;
760 if (back_used
== 0) {
761 /* If we're in this case then we have never called alloc_back(). In
762 * this case, we want keep the offset at 0 to make things as simple
763 * as possible for users that don't care about back allocations.
765 center_bo_offset
= 0;
767 /* Try to "center" the allocation based on how much is currently in
768 * use on each side of the center line.
770 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
772 /* Align down to a multiple of the page size */
773 center_bo_offset
&= ~(PAGE_SIZE
- 1);
775 assert(center_bo_offset
>= back_used
);
777 /* Make sure we don't shrink the back end of the pool */
778 if (center_bo_offset
< back_required
)
779 center_bo_offset
= back_required
;
781 /* Make sure that we don't shrink the front end of the pool */
782 if (size
- center_bo_offset
< front_required
)
783 center_bo_offset
= size
- front_required
;
786 assert(center_bo_offset
% PAGE_SIZE
== 0);
788 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
790 pool
->bo
->flags
= pool
->bo_flags
;
793 pthread_mutex_unlock(&pool
->device
->mutex
);
795 if (result
== VK_SUCCESS
) {
796 /* Return the appropriate new size. This function never actually
797 * updates state->next. Instead, we let the caller do that because it
798 * needs to do so in order to maintain its concurrency model.
800 if (state
== &pool
->state
) {
801 return pool
->size
- pool
->center_bo_offset
;
803 assert(pool
->center_bo_offset
> 0);
804 return pool
->center_bo_offset
;
812 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
813 struct anv_block_state
*pool_state
,
814 uint32_t block_size
, uint32_t *padding
)
816 struct anv_block_state state
, old
, new;
818 /* Most allocations won't generate any padding */
823 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
824 if (state
.next
+ block_size
<= state
.end
) {
826 } else if (state
.next
<= state
.end
) {
827 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
&& state
.next
< state
.end
) {
828 /* We need to grow the block pool, but still have some leftover
829 * space that can't be used by that particular allocation. So we
830 * add that as a "padding", and return it.
832 uint32_t leftover
= state
.end
- state
.next
;
834 /* If there is some leftover space in the pool, the caller must
837 assert(leftover
== 0 || padding
);
840 state
.next
+= leftover
;
843 /* We allocated the first block outside the pool so we have to grow
844 * the pool. pool_state->next acts a mutex: threads who try to
845 * allocate now will get block indexes above the current limit and
846 * hit futex_wait below.
848 new.next
= state
.next
+ block_size
;
850 new.end
= anv_block_pool_grow(pool
, pool_state
);
851 } while (new.end
< new.next
);
853 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
854 if (old
.next
!= state
.next
)
855 futex_wake(&pool_state
->end
, INT_MAX
);
858 futex_wait(&pool_state
->end
, state
.end
, NULL
);
865 anv_block_pool_alloc(struct anv_block_pool
*pool
,
866 uint32_t block_size
, uint32_t *padding
)
870 offset
= anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
, padding
);
875 /* Allocates a block out of the back of the block pool.
877 * This will allocated a block earlier than the "start" of the block pool.
878 * The offsets returned from this function will be negative but will still
879 * be correct relative to the block pool's map pointer.
881 * If you ever use anv_block_pool_alloc_back, then you will have to do
882 * gymnastics with the block pool's BO when doing relocations.
885 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
888 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
891 /* The offset we get out of anv_block_pool_alloc_new() is actually the
892 * number of bytes downwards from the middle to the end of the block.
893 * We need to turn it into a (negative) offset from the middle to the
894 * start of the block.
897 return -(offset
+ block_size
);
901 anv_state_pool_init(struct anv_state_pool
*pool
,
902 struct anv_device
*device
,
903 uint64_t start_address
,
907 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
911 if (result
!= VK_SUCCESS
)
914 result
= anv_state_table_init(&pool
->table
, device
, 64);
915 if (result
!= VK_SUCCESS
) {
916 anv_block_pool_finish(&pool
->block_pool
);
920 assert(util_is_power_of_two_or_zero(block_size
));
921 pool
->block_size
= block_size
;
922 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
923 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
924 pool
->buckets
[i
].free_list
= ANV_FREE_LIST_EMPTY
;
925 pool
->buckets
[i
].block
.next
= 0;
926 pool
->buckets
[i
].block
.end
= 0;
928 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
934 anv_state_pool_finish(struct anv_state_pool
*pool
)
936 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
937 anv_state_table_finish(&pool
->table
);
938 anv_block_pool_finish(&pool
->block_pool
);
942 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
943 struct anv_block_pool
*block_pool
,
948 struct anv_block_state block
, old
, new;
951 /* We don't always use anv_block_pool_alloc(), which would set *padding to
952 * zero for us. So if we have a pointer to padding, we must zero it out
953 * ourselves here, to make sure we always return some sensible value.
958 /* If our state is large, we don't need any sub-allocation from a block.
959 * Instead, we just grab whole (potentially large) blocks.
961 if (state_size
>= block_size
)
962 return anv_block_pool_alloc(block_pool
, state_size
, padding
);
965 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
967 if (block
.next
< block
.end
) {
969 } else if (block
.next
== block
.end
) {
970 offset
= anv_block_pool_alloc(block_pool
, block_size
, padding
);
971 new.next
= offset
+ state_size
;
972 new.end
= offset
+ block_size
;
973 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
974 if (old
.next
!= block
.next
)
975 futex_wake(&pool
->block
.end
, INT_MAX
);
978 futex_wait(&pool
->block
.end
, block
.end
, NULL
);
984 anv_state_pool_get_bucket(uint32_t size
)
986 unsigned size_log2
= ilog2_round_up(size
);
987 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
988 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
989 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
990 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
994 anv_state_pool_get_bucket_size(uint32_t bucket
)
996 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
997 return 1 << size_log2
;
1000 /** Helper to push a chunk into the state table.
1002 * It creates 'count' entries into the state table and update their sizes,
1003 * offsets and maps, also pushing them as "free" states.
1006 anv_state_pool_return_blocks(struct anv_state_pool
*pool
,
1007 uint32_t chunk_offset
, uint32_t count
,
1008 uint32_t block_size
)
1010 /* Disallow returning 0 chunks */
1013 /* Make sure we always return chunks aligned to the block_size */
1014 assert(chunk_offset
% block_size
== 0);
1017 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &st_idx
, count
);
1018 assert(result
== VK_SUCCESS
);
1019 for (int i
= 0; i
< count
; i
++) {
1020 /* update states that were added back to the state table */
1021 struct anv_state
*state_i
= anv_state_table_get(&pool
->table
,
1023 state_i
->alloc_size
= block_size
;
1024 state_i
->offset
= chunk_offset
+ block_size
* i
;
1025 state_i
->map
= anv_block_pool_map(&pool
->block_pool
, state_i
->offset
);
1028 uint32_t block_bucket
= anv_state_pool_get_bucket(block_size
);
1029 anv_free_list_push(&pool
->buckets
[block_bucket
].free_list
,
1030 &pool
->table
, st_idx
, count
);
1033 /** Returns a chunk of memory back to the state pool.
1035 * Do a two-level split. If chunk_size is bigger than divisor
1036 * (pool->block_size), we return as many divisor sized blocks as we can, from
1037 * the end of the chunk.
1039 * The remaining is then split into smaller blocks (starting at small_size if
1040 * it is non-zero), with larger blocks always being taken from the end of the
1044 anv_state_pool_return_chunk(struct anv_state_pool
*pool
,
1045 uint32_t chunk_offset
, uint32_t chunk_size
,
1046 uint32_t small_size
)
1048 uint32_t divisor
= pool
->block_size
;
1049 uint32_t nblocks
= chunk_size
/ divisor
;
1050 uint32_t rest
= chunk_size
- nblocks
* divisor
;
1053 /* First return divisor aligned and sized chunks. We start returning
1054 * larger blocks from the end fo the chunk, since they should already be
1055 * aligned to divisor. Also anv_state_pool_return_blocks() only accepts
1058 uint32_t offset
= chunk_offset
+ rest
;
1059 anv_state_pool_return_blocks(pool
, offset
, nblocks
, divisor
);
1065 if (small_size
> 0 && small_size
< divisor
)
1066 divisor
= small_size
;
1068 uint32_t min_size
= 1 << ANV_MIN_STATE_SIZE_LOG2
;
1070 /* Just as before, return larger divisor aligned blocks from the end of the
1073 while (chunk_size
> 0 && divisor
>= min_size
) {
1074 nblocks
= chunk_size
/ divisor
;
1075 rest
= chunk_size
- nblocks
* divisor
;
1077 anv_state_pool_return_blocks(pool
, chunk_offset
+ rest
,
1085 static struct anv_state
1086 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
1087 uint32_t size
, uint32_t align
)
1089 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
1091 struct anv_state
*state
;
1092 uint32_t alloc_size
= anv_state_pool_get_bucket_size(bucket
);
1095 /* Try free list first. */
1096 state
= anv_free_list_pop(&pool
->buckets
[bucket
].free_list
,
1099 assert(state
->offset
>= 0);
1103 /* Try to grab a chunk from some larger bucket and split it up */
1104 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
1105 state
= anv_free_list_pop(&pool
->buckets
[b
].free_list
, &pool
->table
);
1107 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
1108 int32_t chunk_offset
= state
->offset
;
1110 /* First lets update the state we got to its new size. offset and map
1113 state
->alloc_size
= alloc_size
;
1115 /* Now return the unused part of the chunk back to the pool as free
1118 * There are a couple of options as to what we do with it:
1120 * 1) We could fully split the chunk into state.alloc_size sized
1121 * pieces. However, this would mean that allocating a 16B
1122 * state could potentially split a 2MB chunk into 512K smaller
1123 * chunks. This would lead to unnecessary fragmentation.
1125 * 2) The classic "buddy allocator" method would have us split the
1126 * chunk in half and return one half. Then we would split the
1127 * remaining half in half and return one half, and repeat as
1128 * needed until we get down to the size we want. However, if
1129 * you are allocating a bunch of the same size state (which is
1130 * the common case), this means that every other allocation has
1131 * to go up a level and every fourth goes up two levels, etc.
1132 * This is not nearly as efficient as it could be if we did a
1133 * little more work up-front.
1135 * 3) Split the difference between (1) and (2) by doing a
1136 * two-level split. If it's bigger than some fixed block_size,
1137 * we split it into block_size sized chunks and return all but
1138 * one of them. Then we split what remains into
1139 * state.alloc_size sized chunks and return them.
1141 * We choose something close to option (3), which is implemented with
1142 * anv_state_pool_return_chunk(). That is done by returning the
1143 * remaining of the chunk, with alloc_size as a hint of the size that
1144 * we want the smaller chunk split into.
1146 anv_state_pool_return_chunk(pool
, chunk_offset
+ alloc_size
,
1147 chunk_size
- alloc_size
, alloc_size
);
1153 offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
1158 /* Everytime we allocate a new state, add it to the state pool */
1160 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1161 assert(result
== VK_SUCCESS
);
1163 state
= anv_state_table_get(&pool
->table
, idx
);
1164 state
->offset
= offset
;
1165 state
->alloc_size
= alloc_size
;
1166 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
);
1169 uint32_t return_offset
= offset
- padding
;
1170 anv_state_pool_return_chunk(pool
, return_offset
, padding
, 0);
1178 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
1181 return ANV_STATE_NULL
;
1183 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
1184 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
1189 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
1191 struct anv_state
*state
;
1192 uint32_t alloc_size
= pool
->block_size
;
1194 state
= anv_free_list_pop(&pool
->back_alloc_free_list
, &pool
->table
);
1196 assert(state
->offset
< 0);
1201 offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
1204 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1205 assert(result
== VK_SUCCESS
);
1207 state
= anv_state_table_get(&pool
->table
, idx
);
1208 state
->offset
= offset
;
1209 state
->alloc_size
= alloc_size
;
1210 state
->map
= anv_block_pool_map(&pool
->block_pool
, state
->offset
);
1213 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
->map
, state
->alloc_size
));
1218 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
1220 assert(util_is_power_of_two_or_zero(state
.alloc_size
));
1221 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
1223 if (state
.offset
< 0) {
1224 assert(state
.alloc_size
== pool
->block_size
);
1225 anv_free_list_push(&pool
->back_alloc_free_list
,
1226 &pool
->table
, state
.idx
, 1);
1228 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
1229 &pool
->table
, state
.idx
, 1);
1234 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
1236 if (state
.alloc_size
== 0)
1239 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
1240 anv_state_pool_free_no_vg(pool
, state
);
1243 struct anv_state_stream_block
{
1244 struct anv_state block
;
1246 /* The next block */
1247 struct anv_state_stream_block
*next
;
1249 #ifdef HAVE_VALGRIND
1250 /* A pointer to the first user-allocated thing in this block. This is
1251 * what valgrind sees as the start of the block.
1257 /* The state stream allocator is a one-shot, single threaded allocator for
1258 * variable sized blocks. We use it for allocating dynamic state.
1261 anv_state_stream_init(struct anv_state_stream
*stream
,
1262 struct anv_state_pool
*state_pool
,
1263 uint32_t block_size
)
1265 stream
->state_pool
= state_pool
;
1266 stream
->block_size
= block_size
;
1268 stream
->block
= ANV_STATE_NULL
;
1270 stream
->block_list
= NULL
;
1272 /* Ensure that next + whatever > block_size. This way the first call to
1273 * state_stream_alloc fetches a new block.
1275 stream
->next
= block_size
;
1277 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
1281 anv_state_stream_finish(struct anv_state_stream
*stream
)
1283 struct anv_state_stream_block
*next
= stream
->block_list
;
1284 while (next
!= NULL
) {
1285 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
1286 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
1287 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
1288 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
1292 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
1296 anv_state_stream_alloc(struct anv_state_stream
*stream
,
1297 uint32_t size
, uint32_t alignment
)
1300 return ANV_STATE_NULL
;
1302 assert(alignment
<= PAGE_SIZE
);
1304 uint32_t offset
= align_u32(stream
->next
, alignment
);
1305 if (offset
+ size
> stream
->block
.alloc_size
) {
1306 uint32_t block_size
= stream
->block_size
;
1307 if (block_size
< size
)
1308 block_size
= round_to_power_of_two(size
);
1310 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
1311 block_size
, PAGE_SIZE
);
1313 struct anv_state_stream_block
*sb
= stream
->block
.map
;
1314 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
1315 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
1316 stream
->block_list
= sb
;
1317 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
1319 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
1321 /* Reset back to the start plus space for the header */
1322 stream
->next
= sizeof(*sb
);
1324 offset
= align_u32(stream
->next
, alignment
);
1325 assert(offset
+ size
<= stream
->block
.alloc_size
);
1328 struct anv_state state
= stream
->block
;
1329 state
.offset
+= offset
;
1330 state
.alloc_size
= size
;
1331 state
.map
+= offset
;
1333 stream
->next
= offset
+ size
;
1335 #ifdef HAVE_VALGRIND
1336 struct anv_state_stream_block
*sb
= stream
->block_list
;
1337 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
1338 if (vg_ptr
== NULL
) {
1340 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
1341 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
1343 void *state_end
= state
.map
+ state
.alloc_size
;
1344 /* This only updates the mempool. The newly allocated chunk is still
1345 * marked as NOACCESS. */
1346 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
1347 /* Mark the newly allocated chunk as undefined */
1348 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
1355 struct bo_pool_bo_link
{
1356 struct bo_pool_bo_link
*next
;
1361 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
,
1364 pool
->device
= device
;
1365 pool
->bo_flags
= bo_flags
;
1366 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
1368 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
1372 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
1374 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1375 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
1376 while (link
!= NULL
) {
1377 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
1379 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
1380 anv_vma_free(pool
->device
, &link_copy
.bo
);
1381 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
1382 link
= link_copy
.next
;
1386 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1390 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
1394 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1395 const unsigned pow2_size
= 1 << size_log2
;
1396 const unsigned bucket
= size_log2
- 12;
1397 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1399 void *next_free_void
;
1400 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
1401 struct bo_pool_bo_link
*next_free
= next_free_void
;
1402 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
1403 assert(bo
->gem_handle
);
1404 assert(bo
->map
== next_free
);
1405 assert(size
<= bo
->size
);
1407 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1412 struct anv_bo new_bo
;
1414 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
1415 if (result
!= VK_SUCCESS
)
1418 new_bo
.flags
= pool
->bo_flags
;
1420 if (!anv_vma_alloc(pool
->device
, &new_bo
))
1421 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1423 assert(new_bo
.size
== pow2_size
);
1425 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
1426 if (new_bo
.map
== MAP_FAILED
) {
1427 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
1428 anv_vma_free(pool
->device
, &new_bo
);
1429 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1432 /* We are removing the state flushes, so lets make sure that these buffers
1433 * are cached/snooped.
1435 if (!pool
->device
->info
.has_llc
) {
1436 anv_gem_set_caching(pool
->device
, new_bo
.gem_handle
,
1437 I915_CACHING_CACHED
);
1442 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1448 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
1450 /* Make a copy in case the anv_bo happens to be storred in the BO */
1451 struct anv_bo bo
= *bo_in
;
1453 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
1455 struct bo_pool_bo_link
*link
= bo
.map
;
1456 VG_NOACCESS_WRITE(&link
->bo
, bo
);
1458 assert(util_is_power_of_two_or_zero(bo
.size
));
1459 const unsigned size_log2
= ilog2_round_up(bo
.size
);
1460 const unsigned bucket
= size_log2
- 12;
1461 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1463 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
1469 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1471 memset(pool
, 0, sizeof(*pool
));
1475 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1477 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1478 for (unsigned i
= 0; i
< 16; i
++) {
1479 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
1480 if (bo
->exists
> 0) {
1481 anv_vma_free(device
, &bo
->bo
);
1482 anv_gem_close(device
, bo
->bo
.gem_handle
);
1489 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1490 gl_shader_stage stage
, unsigned per_thread_scratch
)
1492 if (per_thread_scratch
== 0)
1495 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1496 assert(scratch_size_log2
< 16);
1498 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
1500 /* We can use "exists" to shortcut and ignore the critical section */
1504 pthread_mutex_lock(&device
->mutex
);
1506 __sync_synchronize();
1508 pthread_mutex_unlock(&device
->mutex
);
1512 const struct anv_physical_device
*physical_device
=
1513 &device
->instance
->physicalDevice
;
1514 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1516 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
1518 unsigned scratch_ids_per_subslice
;
1519 if (devinfo
->is_haswell
) {
1520 /* WaCSScratchSize:hsw
1522 * Haswell's scratch space address calculation appears to be sparse
1523 * rather than tightly packed. The Thread ID has bits indicating
1524 * which subslice, EU within a subslice, and thread within an EU it
1525 * is. There's a maximum of two slices and two subslices, so these
1526 * can be stored with a single bit. Even though there are only 10 EUs
1527 * per subslice, this is stored in 4 bits, so there's an effective
1528 * maximum value of 16 EUs. Similarly, although there are only 7
1529 * threads per EU, this is stored in a 3 bit number, giving an
1530 * effective maximum value of 8 threads per EU.
1532 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1533 * number of threads per subslice.
1535 scratch_ids_per_subslice
= 16 * 8;
1536 } else if (devinfo
->is_cherryview
) {
1537 /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
1538 * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
1541 scratch_ids_per_subslice
= 8 * 7;
1543 scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
1546 uint32_t max_threads
[] = {
1547 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1548 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1549 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1550 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1551 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1552 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1555 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1557 anv_bo_init_new(&bo
->bo
, device
, size
);
1559 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1560 * are still relative to the general state base address. When we emit
1561 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1562 * to the maximum (1 page under 4GB). This allows us to just place the
1563 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1564 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1565 * However, in order to do so, we need to ensure that the kernel does not
1566 * place the scratch BO above the 32-bit boundary.
1568 * NOTE: Technically, it can't go "anywhere" because the top page is off
1569 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1570 * kernel allocates space using
1572 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1574 * so nothing will ever touch the top page.
1576 assert(!(bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
));
1578 if (device
->instance
->physicalDevice
.has_exec_async
)
1579 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1581 if (device
->instance
->physicalDevice
.use_softpin
)
1582 bo
->bo
.flags
|= EXEC_OBJECT_PINNED
;
1584 anv_vma_alloc(device
, &bo
->bo
);
1586 /* Set the exists last because it may be read by other threads */
1587 __sync_synchronize();
1590 pthread_mutex_unlock(&device
->mutex
);
1595 struct anv_cached_bo
{
1602 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1604 cache
->bo_map
= _mesa_pointer_hash_table_create(NULL
);
1606 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1608 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1609 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1610 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1611 "pthread_mutex_init failed: %m");
1618 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1620 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1621 pthread_mutex_destroy(&cache
->mutex
);
1624 static struct anv_cached_bo
*
1625 anv_bo_cache_lookup_locked(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1627 struct hash_entry
*entry
=
1628 _mesa_hash_table_search(cache
->bo_map
,
1629 (const void *)(uintptr_t)gem_handle
);
1633 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)entry
->data
;
1634 assert(bo
->bo
.gem_handle
== gem_handle
);
1639 UNUSED
static struct anv_bo
*
1640 anv_bo_cache_lookup(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1642 pthread_mutex_lock(&cache
->mutex
);
1644 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1646 pthread_mutex_unlock(&cache
->mutex
);
1648 return bo
? &bo
->bo
: NULL
;
1651 #define ANV_BO_CACHE_SUPPORTED_FLAGS \
1652 (EXEC_OBJECT_WRITE | \
1653 EXEC_OBJECT_ASYNC | \
1654 EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
1655 EXEC_OBJECT_PINNED | \
1659 anv_bo_cache_alloc(struct anv_device
*device
,
1660 struct anv_bo_cache
*cache
,
1661 uint64_t size
, uint64_t bo_flags
,
1662 struct anv_bo
**bo_out
)
1664 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1666 struct anv_cached_bo
*bo
=
1667 vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1668 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1670 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1674 /* The kernel is going to give us whole pages anyway */
1675 size
= align_u64(size
, 4096);
1677 VkResult result
= anv_bo_init_new(&bo
->bo
, device
, size
);
1678 if (result
!= VK_SUCCESS
) {
1679 vk_free(&device
->alloc
, bo
);
1683 bo
->bo
.flags
= bo_flags
;
1685 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1686 anv_gem_close(device
, bo
->bo
.gem_handle
);
1687 vk_free(&device
->alloc
, bo
);
1688 return vk_errorf(device
->instance
, NULL
,
1689 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1690 "failed to allocate virtual address for BO");
1693 assert(bo
->bo
.gem_handle
);
1695 pthread_mutex_lock(&cache
->mutex
);
1697 _mesa_hash_table_insert(cache
->bo_map
,
1698 (void *)(uintptr_t)bo
->bo
.gem_handle
, bo
);
1700 pthread_mutex_unlock(&cache
->mutex
);
1708 anv_bo_cache_import_host_ptr(struct anv_device
*device
,
1709 struct anv_bo_cache
*cache
,
1710 void *host_ptr
, uint32_t size
,
1711 uint64_t bo_flags
, struct anv_bo
**bo_out
)
1713 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1714 assert((bo_flags
& ANV_BO_EXTERNAL
) == 0);
1716 uint32_t gem_handle
= anv_gem_userptr(device
, host_ptr
, size
);
1718 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1720 pthread_mutex_lock(&cache
->mutex
);
1722 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1724 /* VK_EXT_external_memory_host doesn't require handling importing the
1725 * same pointer twice at the same time, but we don't get in the way. If
1726 * kernel gives us the same gem_handle, only succeed if the flags match.
1728 if (bo_flags
!= bo
->bo
.flags
) {
1729 pthread_mutex_unlock(&cache
->mutex
);
1730 return vk_errorf(device
->instance
, NULL
,
1731 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1732 "same host pointer imported two different ways");
1734 __sync_fetch_and_add(&bo
->refcount
, 1);
1736 bo
= vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1737 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1739 anv_gem_close(device
, gem_handle
);
1740 pthread_mutex_unlock(&cache
->mutex
);
1741 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1746 anv_bo_init(&bo
->bo
, gem_handle
, size
);
1747 bo
->bo
.flags
= bo_flags
;
1749 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1750 anv_gem_close(device
, bo
->bo
.gem_handle
);
1751 pthread_mutex_unlock(&cache
->mutex
);
1752 vk_free(&device
->alloc
, bo
);
1753 return vk_errorf(device
->instance
, NULL
,
1754 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1755 "failed to allocate virtual address for BO");
1758 _mesa_hash_table_insert(cache
->bo_map
, (void *)(uintptr_t)gem_handle
, bo
);
1761 pthread_mutex_unlock(&cache
->mutex
);
1768 anv_bo_cache_import(struct anv_device
*device
,
1769 struct anv_bo_cache
*cache
,
1770 int fd
, uint64_t bo_flags
,
1771 struct anv_bo
**bo_out
)
1773 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1774 assert(bo_flags
& ANV_BO_EXTERNAL
);
1776 pthread_mutex_lock(&cache
->mutex
);
1778 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1780 pthread_mutex_unlock(&cache
->mutex
);
1781 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1784 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1786 /* We have to be careful how we combine flags so that it makes sense.
1787 * Really, though, if we get to this case and it actually matters, the
1788 * client has imported a BO twice in different ways and they get what
1791 uint64_t new_flags
= ANV_BO_EXTERNAL
;
1792 new_flags
|= (bo
->bo
.flags
| bo_flags
) & EXEC_OBJECT_WRITE
;
1793 new_flags
|= (bo
->bo
.flags
& bo_flags
) & EXEC_OBJECT_ASYNC
;
1794 new_flags
|= (bo
->bo
.flags
& bo_flags
) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1795 new_flags
|= (bo
->bo
.flags
| bo_flags
) & EXEC_OBJECT_PINNED
;
1797 /* It's theoretically possible for a BO to get imported such that it's
1798 * both pinned and not pinned. The only way this can happen is if it
1799 * gets imported as both a semaphore and a memory object and that would
1800 * be an application error. Just fail out in that case.
1802 if ((bo
->bo
.flags
& EXEC_OBJECT_PINNED
) !=
1803 (bo_flags
& EXEC_OBJECT_PINNED
)) {
1804 pthread_mutex_unlock(&cache
->mutex
);
1805 return vk_errorf(device
->instance
, NULL
,
1806 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1807 "The same BO was imported two different ways");
1810 /* It's also theoretically possible that someone could export a BO from
1811 * one heap and import it into another or to import the same BO into two
1812 * different heaps. If this happens, we could potentially end up both
1813 * allowing and disallowing 48-bit addresses. There's not much we can
1814 * do about it if we're pinning so we just throw an error and hope no
1815 * app is actually that stupid.
1817 if ((new_flags
& EXEC_OBJECT_PINNED
) &&
1818 (bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) !=
1819 (bo_flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
)) {
1820 pthread_mutex_unlock(&cache
->mutex
);
1821 return vk_errorf(device
->instance
, NULL
,
1822 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1823 "The same BO was imported on two different heaps");
1826 bo
->bo
.flags
= new_flags
;
1828 __sync_fetch_and_add(&bo
->refcount
, 1);
1830 off_t size
= lseek(fd
, 0, SEEK_END
);
1831 if (size
== (off_t
)-1) {
1832 anv_gem_close(device
, gem_handle
);
1833 pthread_mutex_unlock(&cache
->mutex
);
1834 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1837 bo
= vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1838 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1840 anv_gem_close(device
, gem_handle
);
1841 pthread_mutex_unlock(&cache
->mutex
);
1842 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1847 anv_bo_init(&bo
->bo
, gem_handle
, size
);
1848 bo
->bo
.flags
= bo_flags
;
1850 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1851 anv_gem_close(device
, bo
->bo
.gem_handle
);
1852 pthread_mutex_unlock(&cache
->mutex
);
1853 vk_free(&device
->alloc
, bo
);
1854 return vk_errorf(device
->instance
, NULL
,
1855 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1856 "failed to allocate virtual address for BO");
1859 _mesa_hash_table_insert(cache
->bo_map
, (void *)(uintptr_t)gem_handle
, bo
);
1862 pthread_mutex_unlock(&cache
->mutex
);
1869 anv_bo_cache_export(struct anv_device
*device
,
1870 struct anv_bo_cache
*cache
,
1871 struct anv_bo
*bo_in
, int *fd_out
)
1873 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1874 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1876 /* This BO must have been flagged external in order for us to be able
1877 * to export it. This is done based on external options passed into
1878 * anv_AllocateMemory.
1880 assert(bo
->bo
.flags
& ANV_BO_EXTERNAL
);
1882 int fd
= anv_gem_handle_to_fd(device
, bo
->bo
.gem_handle
);
1884 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1892 atomic_dec_not_one(uint32_t *counter
)
1901 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1910 anv_bo_cache_release(struct anv_device
*device
,
1911 struct anv_bo_cache
*cache
,
1912 struct anv_bo
*bo_in
)
1914 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1915 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1917 /* Try to decrement the counter but don't go below one. If this succeeds
1918 * then the refcount has been decremented and we are not the last
1921 if (atomic_dec_not_one(&bo
->refcount
))
1924 pthread_mutex_lock(&cache
->mutex
);
1926 /* We are probably the last reference since our attempt to decrement above
1927 * failed. However, we can't actually know until we are inside the mutex.
1928 * Otherwise, someone could import the BO between the decrement and our
1931 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1932 /* Turns out we're not the last reference. Unlock and bail. */
1933 pthread_mutex_unlock(&cache
->mutex
);
1937 struct hash_entry
*entry
=
1938 _mesa_hash_table_search(cache
->bo_map
,
1939 (const void *)(uintptr_t)bo
->bo
.gem_handle
);
1941 _mesa_hash_table_remove(cache
->bo_map
, entry
);
1944 anv_gem_munmap(bo
->bo
.map
, bo
->bo
.size
);
1946 anv_vma_free(device
, &bo
->bo
);
1948 anv_gem_close(device
, bo
->bo
.gem_handle
);
1950 /* Don't unlock until we've actually closed the BO. The whole point of
1951 * the BO cache is to ensure that we correctly handle races with creating
1952 * and releasing GEM handles and we don't want to let someone import the BO
1953 * again between mutex unlock and closing the GEM handle.
1955 pthread_mutex_unlock(&cache
->mutex
);
1957 vk_free(&device
->alloc
, bo
);