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
28 #include <linux/memfd.h>
31 #include "anv_private.h"
33 #include "util/hash_table.h"
34 #include "util/simple_mtx.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 #ifndef HAVE_MEMFD_CREATE
116 memfd_create(const char *name
, unsigned int flags
)
118 return syscall(SYS_memfd_create
, name
, flags
);
122 static inline uint32_t
123 ilog2_round_up(uint32_t value
)
126 return 32 - __builtin_clz(value
- 1);
129 static inline uint32_t
130 round_to_power_of_two(uint32_t value
)
132 return 1 << ilog2_round_up(value
);
135 struct anv_state_table_cleanup
{
140 #define ANV_STATE_TABLE_CLEANUP_INIT ((struct anv_state_table_cleanup){0})
141 #define ANV_STATE_ENTRY_SIZE (sizeof(struct anv_free_entry))
144 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
);
147 anv_state_table_init(struct anv_state_table
*table
,
148 struct anv_device
*device
,
149 uint32_t initial_entries
)
153 table
->device
= device
;
155 table
->fd
= memfd_create("state table", MFD_CLOEXEC
);
157 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
159 /* Just make it 2GB up-front. The Linux kernel won't actually back it
160 * with pages until we either map and fault on one of them or we use
161 * userptr and send a chunk of it off to the GPU.
163 if (ftruncate(table
->fd
, BLOCK_POOL_MEMFD_SIZE
) == -1) {
164 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
168 if (!u_vector_init(&table
->mmap_cleanups
,
169 round_to_power_of_two(sizeof(struct anv_state_table_cleanup
)),
171 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
175 table
->state
.next
= 0;
176 table
->state
.end
= 0;
179 uint32_t initial_size
= initial_entries
* ANV_STATE_ENTRY_SIZE
;
180 result
= anv_state_table_expand_range(table
, initial_size
);
181 if (result
!= VK_SUCCESS
)
182 goto fail_mmap_cleanups
;
187 u_vector_finish(&table
->mmap_cleanups
);
195 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
)
198 struct anv_mmap_cleanup
*cleanup
;
200 /* Assert that we only ever grow the pool */
201 assert(size
>= table
->state
.end
);
203 /* Make sure that we don't go outside the bounds of the memfd */
204 if (size
> BLOCK_POOL_MEMFD_SIZE
)
205 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
207 cleanup
= u_vector_add(&table
->mmap_cleanups
);
209 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
211 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
213 /* Just leak the old map until we destroy the pool. We can't munmap it
214 * without races or imposing locking on the block allocate fast path. On
215 * the whole the leaked maps adds up to less than the size of the
216 * current map. MAP_POPULATE seems like the right thing to do, but we
217 * should try to get some numbers.
219 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
220 MAP_SHARED
| MAP_POPULATE
, table
->fd
, 0);
221 if (map
== MAP_FAILED
) {
222 return vk_errorf(table
->device
->instance
, table
->device
,
223 VK_ERROR_OUT_OF_HOST_MEMORY
, "mmap failed: %m");
227 cleanup
->size
= size
;
236 anv_state_table_grow(struct anv_state_table
*table
)
238 VkResult result
= VK_SUCCESS
;
240 uint32_t used
= align_u32(table
->state
.next
* ANV_STATE_ENTRY_SIZE
,
242 uint32_t old_size
= table
->size
;
244 /* The block pool is always initialized to a nonzero size and this function
245 * is always called after initialization.
247 assert(old_size
> 0);
249 uint32_t required
= MAX2(used
, old_size
);
250 if (used
* 2 <= required
) {
251 /* If we're in this case then this isn't the firsta allocation and we
252 * already have enough space on both sides to hold double what we
253 * have allocated. There's nothing for us to do.
258 uint32_t size
= old_size
* 2;
259 while (size
< required
)
262 assert(size
> table
->size
);
264 result
= anv_state_table_expand_range(table
, size
);
271 anv_state_table_finish(struct anv_state_table
*table
)
273 struct anv_state_table_cleanup
*cleanup
;
275 u_vector_foreach(cleanup
, &table
->mmap_cleanups
) {
277 munmap(cleanup
->map
, cleanup
->size
);
280 u_vector_finish(&table
->mmap_cleanups
);
286 anv_state_table_add(struct anv_state_table
*table
, uint32_t *idx
,
289 struct anv_block_state state
, old
, new;
295 state
.u64
= __sync_fetch_and_add(&table
->state
.u64
, count
);
296 if (state
.next
+ count
<= state
.end
) {
298 struct anv_free_entry
*entry
= &table
->map
[state
.next
];
299 for (int i
= 0; i
< count
; i
++) {
300 entry
[i
].state
.idx
= state
.next
+ i
;
304 } else if (state
.next
<= state
.end
) {
305 /* We allocated the first block outside the pool so we have to grow
306 * the pool. pool_state->next acts a mutex: threads who try to
307 * allocate now will get block indexes above the current limit and
308 * hit futex_wait below.
310 new.next
= state
.next
+ count
;
312 result
= anv_state_table_grow(table
);
313 if (result
!= VK_SUCCESS
)
315 new.end
= table
->size
/ ANV_STATE_ENTRY_SIZE
;
316 } while (new.end
< new.next
);
318 old
.u64
= __sync_lock_test_and_set(&table
->state
.u64
, new.u64
);
319 if (old
.next
!= state
.next
)
320 futex_wake(&table
->state
.end
, INT_MAX
);
322 futex_wait(&table
->state
.end
, state
.end
, NULL
);
329 anv_free_list_push(union anv_free_list
*list
,
330 struct anv_state_table
*table
,
331 uint32_t first
, uint32_t count
)
333 union anv_free_list current
, old
, new;
334 uint32_t last
= first
;
336 for (uint32_t i
= 1; i
< count
; i
++, last
++)
337 table
->map
[last
].next
= last
+ 1;
342 table
->map
[last
].next
= current
.offset
;
344 new.count
= current
.count
+ 1;
345 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
346 } while (old
.u64
!= current
.u64
);
350 anv_free_list_pop(union anv_free_list
*list
,
351 struct anv_state_table
*table
)
353 union anv_free_list current
, new, old
;
355 current
.u64
= list
->u64
;
356 while (current
.offset
!= EMPTY
) {
357 __sync_synchronize();
358 new.offset
= table
->map
[current
.offset
].next
;
359 new.count
= current
.count
+ 1;
360 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
361 if (old
.u64
== current
.u64
) {
362 struct anv_free_entry
*entry
= &table
->map
[current
.offset
];
363 return &entry
->state
;
371 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
372 * means that the bottom 12 bits should all be zero.
374 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
375 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
376 #define PFL_PACK(ptr, count) ({ \
377 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
381 anv_ptr_free_list_pop(void **list
, void **elem
)
383 void *current
= *list
;
384 while (PFL_PTR(current
) != NULL
) {
385 void **next_ptr
= PFL_PTR(current
);
386 void *new_ptr
= VG_NOACCESS_READ(next_ptr
);
387 unsigned new_count
= PFL_COUNT(current
) + 1;
388 void *new = PFL_PACK(new_ptr
, new_count
);
389 void *old
= __sync_val_compare_and_swap(list
, current
, new);
390 if (old
== current
) {
391 *elem
= PFL_PTR(current
);
401 anv_ptr_free_list_push(void **list
, void *elem
)
404 void **next_ptr
= elem
;
406 /* The pointer-based free list requires that the pointer be
407 * page-aligned. This is because we use the bottom 12 bits of the
408 * pointer to store a counter to solve the ABA concurrency problem.
410 assert(((uintptr_t)elem
& 0xfff) == 0);
415 VG_NOACCESS_WRITE(next_ptr
, PFL_PTR(current
));
416 unsigned new_count
= PFL_COUNT(current
) + 1;
417 void *new = PFL_PACK(elem
, new_count
);
418 old
= __sync_val_compare_and_swap(list
, current
, new);
419 } while (old
!= current
);
423 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
424 uint32_t center_bo_offset
, uint32_t size
);
427 anv_block_pool_init(struct anv_block_pool
*pool
,
428 struct anv_device
*device
,
429 uint64_t start_address
,
430 uint32_t initial_size
,
435 pool
->device
= device
;
436 pool
->bo_flags
= bo_flags
;
437 pool
->start_address
= gen_canonical_address(start_address
);
439 anv_bo_init(&pool
->bo
, 0, 0);
441 pool
->fd
= memfd_create("block pool", MFD_CLOEXEC
);
443 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
445 /* Just make it 2GB up-front. The Linux kernel won't actually back it
446 * with pages until we either map and fault on one of them or we use
447 * userptr and send a chunk of it off to the GPU.
449 if (ftruncate(pool
->fd
, BLOCK_POOL_MEMFD_SIZE
) == -1) {
450 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
454 if (!u_vector_init(&pool
->mmap_cleanups
,
455 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
457 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
461 pool
->state
.next
= 0;
463 pool
->back_state
.next
= 0;
464 pool
->back_state
.end
= 0;
466 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
467 if (result
!= VK_SUCCESS
)
468 goto fail_mmap_cleanups
;
473 u_vector_finish(&pool
->mmap_cleanups
);
481 anv_block_pool_finish(struct anv_block_pool
*pool
)
483 struct anv_mmap_cleanup
*cleanup
;
485 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
) {
487 munmap(cleanup
->map
, cleanup
->size
);
488 if (cleanup
->gem_handle
)
489 anv_gem_close(pool
->device
, cleanup
->gem_handle
);
492 u_vector_finish(&pool
->mmap_cleanups
);
498 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
499 uint32_t center_bo_offset
, uint32_t size
)
503 struct anv_mmap_cleanup
*cleanup
;
505 /* Assert that we only ever grow the pool */
506 assert(center_bo_offset
>= pool
->back_state
.end
);
507 assert(size
- center_bo_offset
>= pool
->state
.end
);
509 /* Assert that we don't go outside the bounds of the memfd */
510 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
511 assert(size
- center_bo_offset
<=
512 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
514 cleanup
= u_vector_add(&pool
->mmap_cleanups
);
516 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
518 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
520 /* Just leak the old map until we destroy the pool. We can't munmap it
521 * without races or imposing locking on the block allocate fast path. On
522 * the whole the leaked maps adds up to less than the size of the
523 * current map. MAP_POPULATE seems like the right thing to do, but we
524 * should try to get some numbers.
526 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
527 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
528 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
529 if (map
== MAP_FAILED
)
530 return vk_errorf(pool
->device
->instance
, pool
->device
,
531 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
533 gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
534 if (gem_handle
== 0) {
536 return vk_errorf(pool
->device
->instance
, pool
->device
,
537 VK_ERROR_TOO_MANY_OBJECTS
, "userptr failed: %m");
541 cleanup
->size
= size
;
542 cleanup
->gem_handle
= gem_handle
;
545 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
546 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
547 * always created as I915_CACHING_CACHED, which on non-LLC means
548 * snooped. That can be useful but comes with a bit of overheard. Since
549 * we're eplicitly clflushing and don't want the overhead we need to turn
551 if (!pool
->device
->info
.has_llc
) {
552 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_NONE
);
553 anv_gem_set_domain(pool
->device
, gem_handle
,
554 I915_GEM_DOMAIN_GTT
, I915_GEM_DOMAIN_GTT
);
558 /* Now that we successfull allocated everything, we can write the new
559 * values back into pool. */
560 pool
->map
= map
+ center_bo_offset
;
561 pool
->center_bo_offset
= center_bo_offset
;
563 /* For block pool BOs we have to be a bit careful about where we place them
564 * in the GTT. There are two documented workarounds for state base address
565 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
566 * which state that those two base addresses do not support 48-bit
567 * addresses and need to be placed in the bottom 32-bit range.
568 * Unfortunately, this is not quite accurate.
570 * The real problem is that we always set the size of our state pools in
571 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
572 * likely significantly smaller. We do this because we do not no at the
573 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
574 * the pool during command buffer building so we don't actually have a
575 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
576 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
577 * as being out of bounds and returns zero. For dynamic state, this
578 * usually just leads to rendering corruptions, but shaders that are all
579 * zero hang the GPU immediately.
581 * The easiest solution to do is exactly what the bogus workarounds say to
582 * do: restrict these buffers to 32-bit addresses. We could also pin the
583 * BO to some particular location of our choosing, but that's significantly
584 * more work than just not setting a flag. So, we explicitly DO NOT set
585 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
588 anv_bo_init(&pool
->bo
, gem_handle
, size
);
589 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
) {
590 pool
->bo
.offset
= pool
->start_address
+ BLOCK_POOL_MEMFD_CENTER
-
593 pool
->bo
.flags
= pool
->bo_flags
;
599 /** Returns current memory map of the block pool.
601 * The returned pointer points to the map for the memory at the specified
602 * offset. The offset parameter is relative to the "center" of the block pool
603 * rather than the start of the block pool BO map.
606 anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
)
608 return pool
->map
+ offset
;
611 /** Grows and re-centers the block pool.
613 * We grow the block pool in one or both directions in such a way that the
614 * following conditions are met:
616 * 1) The size of the entire pool is always a power of two.
618 * 2) The pool only grows on both ends. Neither end can get
621 * 3) At the end of the allocation, we have about twice as much space
622 * allocated for each end as we have used. This way the pool doesn't
623 * grow too far in one direction or the other.
625 * 4) If the _alloc_back() has never been called, then the back portion of
626 * the pool retains a size of zero. (This makes it easier for users of
627 * the block pool that only want a one-sided pool.)
629 * 5) We have enough space allocated for at least one more block in
630 * whichever side `state` points to.
632 * 6) The center of the pool is always aligned to both the block_size of
633 * the pool and a 4K CPU page.
636 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
638 VkResult result
= VK_SUCCESS
;
640 pthread_mutex_lock(&pool
->device
->mutex
);
642 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
644 /* Gather a little usage information on the pool. Since we may have
645 * threadsd waiting in queue to get some storage while we resize, it's
646 * actually possible that total_used will be larger than old_size. In
647 * particular, block_pool_alloc() increments state->next prior to
648 * calling block_pool_grow, so this ensures that we get enough space for
649 * which ever side tries to grow the pool.
651 * We align to a page size because it makes it easier to do our
652 * calculations later in such a way that we state page-aigned.
654 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
655 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
656 uint32_t total_used
= front_used
+ back_used
;
658 assert(state
== &pool
->state
|| back_used
> 0);
660 uint32_t old_size
= pool
->bo
.size
;
662 /* The block pool is always initialized to a nonzero size and this function
663 * is always called after initialization.
665 assert(old_size
> 0);
667 /* The back_used and front_used may actually be smaller than the actual
668 * requirement because they are based on the next pointers which are
669 * updated prior to calling this function.
671 uint32_t back_required
= MAX2(back_used
, pool
->center_bo_offset
);
672 uint32_t front_required
= MAX2(front_used
, old_size
- pool
->center_bo_offset
);
674 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
675 /* If we're in this case then this isn't the firsta allocation and we
676 * already have enough space on both sides to hold double what we
677 * have allocated. There's nothing for us to do.
682 uint32_t size
= old_size
* 2;
683 while (size
< back_required
+ front_required
)
686 assert(size
> pool
->bo
.size
);
688 /* We compute a new center_bo_offset such that, when we double the size
689 * of the pool, we maintain the ratio of how much is used by each side.
690 * This way things should remain more-or-less balanced.
692 uint32_t center_bo_offset
;
693 if (back_used
== 0) {
694 /* If we're in this case then we have never called alloc_back(). In
695 * this case, we want keep the offset at 0 to make things as simple
696 * as possible for users that don't care about back allocations.
698 center_bo_offset
= 0;
700 /* Try to "center" the allocation based on how much is currently in
701 * use on each side of the center line.
703 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
705 /* Align down to a multiple of the page size */
706 center_bo_offset
&= ~(PAGE_SIZE
- 1);
708 assert(center_bo_offset
>= back_used
);
710 /* Make sure we don't shrink the back end of the pool */
711 if (center_bo_offset
< back_required
)
712 center_bo_offset
= back_required
;
714 /* Make sure that we don't shrink the front end of the pool */
715 if (size
- center_bo_offset
< front_required
)
716 center_bo_offset
= size
- front_required
;
719 assert(center_bo_offset
% PAGE_SIZE
== 0);
721 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
723 pool
->bo
.flags
= pool
->bo_flags
;
726 pthread_mutex_unlock(&pool
->device
->mutex
);
728 if (result
== VK_SUCCESS
) {
729 /* Return the appropriate new size. This function never actually
730 * updates state->next. Instead, we let the caller do that because it
731 * needs to do so in order to maintain its concurrency model.
733 if (state
== &pool
->state
) {
734 return pool
->bo
.size
- pool
->center_bo_offset
;
736 assert(pool
->center_bo_offset
> 0);
737 return pool
->center_bo_offset
;
745 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
746 struct anv_block_state
*pool_state
,
749 struct anv_block_state state
, old
, new;
752 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
753 if (state
.next
+ block_size
<= state
.end
) {
756 } else if (state
.next
<= state
.end
) {
757 /* We allocated the first block outside the pool so we have to grow
758 * the pool. pool_state->next acts a mutex: threads who try to
759 * allocate now will get block indexes above the current limit and
760 * hit futex_wait below.
762 new.next
= state
.next
+ block_size
;
764 new.end
= anv_block_pool_grow(pool
, pool_state
);
765 } while (new.end
< new.next
);
767 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
768 if (old
.next
!= state
.next
)
769 futex_wake(&pool_state
->end
, INT_MAX
);
772 futex_wait(&pool_state
->end
, state
.end
, NULL
);
779 anv_block_pool_alloc(struct anv_block_pool
*pool
,
782 return anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
);
785 /* Allocates a block out of the back of the block pool.
787 * This will allocated a block earlier than the "start" of the block pool.
788 * The offsets returned from this function will be negative but will still
789 * be correct relative to the block pool's map pointer.
791 * If you ever use anv_block_pool_alloc_back, then you will have to do
792 * gymnastics with the block pool's BO when doing relocations.
795 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
798 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
801 /* The offset we get out of anv_block_pool_alloc_new() is actually the
802 * number of bytes downwards from the middle to the end of the block.
803 * We need to turn it into a (negative) offset from the middle to the
804 * start of the block.
807 return -(offset
+ block_size
);
811 anv_state_pool_init(struct anv_state_pool
*pool
,
812 struct anv_device
*device
,
813 uint64_t start_address
,
817 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
821 if (result
!= VK_SUCCESS
)
824 result
= anv_state_table_init(&pool
->table
, device
, 64);
825 if (result
!= VK_SUCCESS
) {
826 anv_block_pool_finish(&pool
->block_pool
);
830 assert(util_is_power_of_two_or_zero(block_size
));
831 pool
->block_size
= block_size
;
832 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
833 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
834 pool
->buckets
[i
].free_list
= ANV_FREE_LIST_EMPTY
;
835 pool
->buckets
[i
].block
.next
= 0;
836 pool
->buckets
[i
].block
.end
= 0;
838 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
844 anv_state_pool_finish(struct anv_state_pool
*pool
)
846 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
847 anv_state_table_finish(&pool
->table
);
848 anv_block_pool_finish(&pool
->block_pool
);
852 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
853 struct anv_block_pool
*block_pool
,
857 struct anv_block_state block
, old
, new;
860 /* If our state is large, we don't need any sub-allocation from a block.
861 * Instead, we just grab whole (potentially large) blocks.
863 if (state_size
>= block_size
)
864 return anv_block_pool_alloc(block_pool
, state_size
);
867 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
869 if (block
.next
< block
.end
) {
871 } else if (block
.next
== block
.end
) {
872 offset
= anv_block_pool_alloc(block_pool
, block_size
);
873 new.next
= offset
+ state_size
;
874 new.end
= offset
+ block_size
;
875 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
876 if (old
.next
!= block
.next
)
877 futex_wake(&pool
->block
.end
, INT_MAX
);
880 futex_wait(&pool
->block
.end
, block
.end
, NULL
);
886 anv_state_pool_get_bucket(uint32_t size
)
888 unsigned size_log2
= ilog2_round_up(size
);
889 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
890 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
891 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
892 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
896 anv_state_pool_get_bucket_size(uint32_t bucket
)
898 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
899 return 1 << size_log2
;
902 /** Helper to push a chunk into the state table.
904 * It creates 'count' entries into the state table and update their sizes,
905 * offsets and maps, also pushing them as "free" states.
908 anv_state_pool_return_blocks(struct anv_state_pool
*pool
,
909 uint32_t chunk_offset
, uint32_t count
,
915 /* Make sure we always return chunks aligned to the block_size */
916 assert(chunk_offset
% block_size
== 0);
919 VkResult result
= anv_state_table_add(&pool
->table
, &st_idx
, count
);
920 assert(result
== VK_SUCCESS
);
921 for (int i
= 0; i
< count
; i
++) {
922 /* update states that were added back to the state table */
923 struct anv_state
*state_i
= anv_state_table_get(&pool
->table
,
925 state_i
->alloc_size
= block_size
;
926 state_i
->offset
= chunk_offset
+ block_size
* i
;
927 state_i
->map
= anv_block_pool_map(&pool
->block_pool
, state_i
->offset
);
930 uint32_t block_bucket
= anv_state_pool_get_bucket(block_size
);
931 anv_free_list_push(&pool
->buckets
[block_bucket
].free_list
,
932 &pool
->table
, st_idx
, count
);
935 static struct anv_state
936 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
937 uint32_t size
, uint32_t align
)
939 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
941 struct anv_state
*state
;
942 uint32_t alloc_size
= anv_state_pool_get_bucket_size(bucket
);
945 /* Try free list first. */
946 state
= anv_free_list_pop(&pool
->buckets
[bucket
].free_list
,
949 assert(state
->offset
>= 0);
953 /* Try to grab a chunk from some larger bucket and split it up */
954 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
955 state
= anv_free_list_pop(&pool
->buckets
[b
].free_list
, &pool
->table
);
957 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
958 int32_t chunk_offset
= state
->offset
;
960 /* First lets update the state we got to its new size. offset and map
963 state
->alloc_size
= alloc_size
;
965 /* We've found a chunk that's larger than the requested state size.
966 * There are a couple of options as to what we do with it:
968 * 1) We could fully split the chunk into state.alloc_size sized
969 * pieces. However, this would mean that allocating a 16B
970 * state could potentially split a 2MB chunk into 512K smaller
971 * chunks. This would lead to unnecessary fragmentation.
973 * 2) The classic "buddy allocator" method would have us split the
974 * chunk in half and return one half. Then we would split the
975 * remaining half in half and return one half, and repeat as
976 * needed until we get down to the size we want. However, if
977 * you are allocating a bunch of the same size state (which is
978 * the common case), this means that every other allocation has
979 * to go up a level and every fourth goes up two levels, etc.
980 * This is not nearly as efficient as it could be if we did a
981 * little more work up-front.
983 * 3) Split the difference between (1) and (2) by doing a
984 * two-level split. If it's bigger than some fixed block_size,
985 * we split it into block_size sized chunks and return all but
986 * one of them. Then we split what remains into
987 * state.alloc_size sized chunks and return all but one.
989 * We choose option (3).
991 if (chunk_size
> pool
->block_size
&&
992 alloc_size
< pool
->block_size
) {
993 assert(chunk_size
% pool
->block_size
== 0);
994 /* We don't want to split giant chunks into tiny chunks. Instead,
995 * break anything bigger than a block into block-sized chunks and
996 * then break it down into bucket-sized chunks from there. Return
997 * all but the first block of the chunk to the block bucket.
999 uint32_t push_back
= (chunk_size
/ pool
->block_size
) - 1;
1000 anv_state_pool_return_blocks(pool
, chunk_offset
+ pool
->block_size
,
1001 push_back
, pool
->block_size
);
1002 chunk_size
= pool
->block_size
;
1005 assert(chunk_size
% alloc_size
== 0);
1006 uint32_t push_back
= (chunk_size
/ alloc_size
) - 1;
1007 anv_state_pool_return_blocks(pool
, chunk_offset
+ alloc_size
,
1008 push_back
, alloc_size
);
1013 offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
1017 /* Everytime we allocate a new state, add it to the state pool */
1019 VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1020 assert(result
== VK_SUCCESS
);
1022 state
= anv_state_table_get(&pool
->table
, idx
);
1023 state
->offset
= offset
;
1024 state
->alloc_size
= alloc_size
;
1025 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
);
1032 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
1035 return ANV_STATE_NULL
;
1037 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
1038 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
1043 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
1045 struct anv_state
*state
;
1046 uint32_t alloc_size
= pool
->block_size
;
1048 state
= anv_free_list_pop(&pool
->back_alloc_free_list
, &pool
->table
);
1050 assert(state
->offset
< 0);
1055 offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
1058 VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1059 assert(result
== VK_SUCCESS
);
1061 state
= anv_state_table_get(&pool
->table
, idx
);
1062 state
->offset
= offset
;
1063 state
->alloc_size
= alloc_size
;
1064 state
->map
= pool
->block_pool
.map
+ state
->offset
;
1067 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
->map
, state
->alloc_size
));
1072 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
1074 assert(util_is_power_of_two_or_zero(state
.alloc_size
));
1075 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
1077 if (state
.offset
< 0) {
1078 assert(state
.alloc_size
== pool
->block_size
);
1079 anv_free_list_push(&pool
->back_alloc_free_list
,
1080 &pool
->table
, state
.idx
, 1);
1082 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
1083 &pool
->table
, state
.idx
, 1);
1088 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
1090 if (state
.alloc_size
== 0)
1093 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
1094 anv_state_pool_free_no_vg(pool
, state
);
1097 struct anv_state_stream_block
{
1098 struct anv_state block
;
1100 /* The next block */
1101 struct anv_state_stream_block
*next
;
1103 #ifdef HAVE_VALGRIND
1104 /* A pointer to the first user-allocated thing in this block. This is
1105 * what valgrind sees as the start of the block.
1111 /* The state stream allocator is a one-shot, single threaded allocator for
1112 * variable sized blocks. We use it for allocating dynamic state.
1115 anv_state_stream_init(struct anv_state_stream
*stream
,
1116 struct anv_state_pool
*state_pool
,
1117 uint32_t block_size
)
1119 stream
->state_pool
= state_pool
;
1120 stream
->block_size
= block_size
;
1122 stream
->block
= ANV_STATE_NULL
;
1124 stream
->block_list
= NULL
;
1126 /* Ensure that next + whatever > block_size. This way the first call to
1127 * state_stream_alloc fetches a new block.
1129 stream
->next
= block_size
;
1131 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
1135 anv_state_stream_finish(struct anv_state_stream
*stream
)
1137 struct anv_state_stream_block
*next
= stream
->block_list
;
1138 while (next
!= NULL
) {
1139 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
1140 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
1141 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
1142 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
1146 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
1150 anv_state_stream_alloc(struct anv_state_stream
*stream
,
1151 uint32_t size
, uint32_t alignment
)
1154 return ANV_STATE_NULL
;
1156 assert(alignment
<= PAGE_SIZE
);
1158 uint32_t offset
= align_u32(stream
->next
, alignment
);
1159 if (offset
+ size
> stream
->block
.alloc_size
) {
1160 uint32_t block_size
= stream
->block_size
;
1161 if (block_size
< size
)
1162 block_size
= round_to_power_of_two(size
);
1164 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
1165 block_size
, PAGE_SIZE
);
1167 struct anv_state_stream_block
*sb
= stream
->block
.map
;
1168 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
1169 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
1170 stream
->block_list
= sb
;
1171 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
1173 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
1175 /* Reset back to the start plus space for the header */
1176 stream
->next
= sizeof(*sb
);
1178 offset
= align_u32(stream
->next
, alignment
);
1179 assert(offset
+ size
<= stream
->block
.alloc_size
);
1182 struct anv_state state
= stream
->block
;
1183 state
.offset
+= offset
;
1184 state
.alloc_size
= size
;
1185 state
.map
+= offset
;
1187 stream
->next
= offset
+ size
;
1189 #ifdef HAVE_VALGRIND
1190 struct anv_state_stream_block
*sb
= stream
->block_list
;
1191 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
1192 if (vg_ptr
== NULL
) {
1194 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
1195 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
1197 void *state_end
= state
.map
+ state
.alloc_size
;
1198 /* This only updates the mempool. The newly allocated chunk is still
1199 * marked as NOACCESS. */
1200 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
1201 /* Mark the newly allocated chunk as undefined */
1202 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
1209 struct bo_pool_bo_link
{
1210 struct bo_pool_bo_link
*next
;
1215 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
,
1218 pool
->device
= device
;
1219 pool
->bo_flags
= bo_flags
;
1220 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
1222 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
1226 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
1228 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1229 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
1230 while (link
!= NULL
) {
1231 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
1233 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
1234 anv_vma_free(pool
->device
, &link_copy
.bo
);
1235 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
1236 link
= link_copy
.next
;
1240 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1244 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
1248 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1249 const unsigned pow2_size
= 1 << size_log2
;
1250 const unsigned bucket
= size_log2
- 12;
1251 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1253 void *next_free_void
;
1254 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
1255 struct bo_pool_bo_link
*next_free
= next_free_void
;
1256 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
1257 assert(bo
->gem_handle
);
1258 assert(bo
->map
== next_free
);
1259 assert(size
<= bo
->size
);
1261 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1266 struct anv_bo new_bo
;
1268 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
1269 if (result
!= VK_SUCCESS
)
1272 new_bo
.flags
= pool
->bo_flags
;
1274 if (!anv_vma_alloc(pool
->device
, &new_bo
))
1275 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1277 assert(new_bo
.size
== pow2_size
);
1279 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
1280 if (new_bo
.map
== MAP_FAILED
) {
1281 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
1282 anv_vma_free(pool
->device
, &new_bo
);
1283 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1288 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1294 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
1296 /* Make a copy in case the anv_bo happens to be storred in the BO */
1297 struct anv_bo bo
= *bo_in
;
1299 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
1301 struct bo_pool_bo_link
*link
= bo
.map
;
1302 VG_NOACCESS_WRITE(&link
->bo
, bo
);
1304 assert(util_is_power_of_two_or_zero(bo
.size
));
1305 const unsigned size_log2
= ilog2_round_up(bo
.size
);
1306 const unsigned bucket
= size_log2
- 12;
1307 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1309 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
1315 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1317 memset(pool
, 0, sizeof(*pool
));
1321 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1323 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1324 for (unsigned i
= 0; i
< 16; i
++) {
1325 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
1326 if (bo
->exists
> 0) {
1327 anv_vma_free(device
, &bo
->bo
);
1328 anv_gem_close(device
, bo
->bo
.gem_handle
);
1335 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1336 gl_shader_stage stage
, unsigned per_thread_scratch
)
1338 if (per_thread_scratch
== 0)
1341 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1342 assert(scratch_size_log2
< 16);
1344 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
1346 /* We can use "exists" to shortcut and ignore the critical section */
1350 pthread_mutex_lock(&device
->mutex
);
1352 __sync_synchronize();
1354 pthread_mutex_unlock(&device
->mutex
);
1358 const struct anv_physical_device
*physical_device
=
1359 &device
->instance
->physicalDevice
;
1360 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1362 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
1364 unsigned scratch_ids_per_subslice
;
1365 if (devinfo
->is_haswell
) {
1366 /* WaCSScratchSize:hsw
1368 * Haswell's scratch space address calculation appears to be sparse
1369 * rather than tightly packed. The Thread ID has bits indicating
1370 * which subslice, EU within a subslice, and thread within an EU it
1371 * is. There's a maximum of two slices and two subslices, so these
1372 * can be stored with a single bit. Even though there are only 10 EUs
1373 * per subslice, this is stored in 4 bits, so there's an effective
1374 * maximum value of 16 EUs. Similarly, although there are only 7
1375 * threads per EU, this is stored in a 3 bit number, giving an
1376 * effective maximum value of 8 threads per EU.
1378 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1379 * number of threads per subslice.
1381 scratch_ids_per_subslice
= 16 * 8;
1382 } else if (devinfo
->is_cherryview
) {
1383 /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
1384 * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
1387 scratch_ids_per_subslice
= 8 * 7;
1389 scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
1392 uint32_t max_threads
[] = {
1393 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1394 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1395 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1396 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1397 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1398 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1401 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1403 anv_bo_init_new(&bo
->bo
, device
, size
);
1405 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1406 * are still relative to the general state base address. When we emit
1407 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1408 * to the maximum (1 page under 4GB). This allows us to just place the
1409 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1410 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1411 * However, in order to do so, we need to ensure that the kernel does not
1412 * place the scratch BO above the 32-bit boundary.
1414 * NOTE: Technically, it can't go "anywhere" because the top page is off
1415 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1416 * kernel allocates space using
1418 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1420 * so nothing will ever touch the top page.
1422 assert(!(bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
));
1424 if (device
->instance
->physicalDevice
.has_exec_async
)
1425 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1427 if (device
->instance
->physicalDevice
.use_softpin
)
1428 bo
->bo
.flags
|= EXEC_OBJECT_PINNED
;
1430 anv_vma_alloc(device
, &bo
->bo
);
1432 /* Set the exists last because it may be read by other threads */
1433 __sync_synchronize();
1436 pthread_mutex_unlock(&device
->mutex
);
1441 struct anv_cached_bo
{
1448 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1450 cache
->bo_map
= _mesa_pointer_hash_table_create(NULL
);
1452 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1454 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1455 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1456 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1457 "pthread_mutex_init failed: %m");
1464 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1466 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1467 pthread_mutex_destroy(&cache
->mutex
);
1470 static struct anv_cached_bo
*
1471 anv_bo_cache_lookup_locked(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1473 struct hash_entry
*entry
=
1474 _mesa_hash_table_search(cache
->bo_map
,
1475 (const void *)(uintptr_t)gem_handle
);
1479 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)entry
->data
;
1480 assert(bo
->bo
.gem_handle
== gem_handle
);
1485 UNUSED
static struct anv_bo
*
1486 anv_bo_cache_lookup(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1488 pthread_mutex_lock(&cache
->mutex
);
1490 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1492 pthread_mutex_unlock(&cache
->mutex
);
1494 return bo
? &bo
->bo
: NULL
;
1497 #define ANV_BO_CACHE_SUPPORTED_FLAGS \
1498 (EXEC_OBJECT_WRITE | \
1499 EXEC_OBJECT_ASYNC | \
1500 EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
1501 EXEC_OBJECT_PINNED | \
1505 anv_bo_cache_alloc(struct anv_device
*device
,
1506 struct anv_bo_cache
*cache
,
1507 uint64_t size
, uint64_t bo_flags
,
1508 struct anv_bo
**bo_out
)
1510 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1512 struct anv_cached_bo
*bo
=
1513 vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1514 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1516 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1520 /* The kernel is going to give us whole pages anyway */
1521 size
= align_u64(size
, 4096);
1523 VkResult result
= anv_bo_init_new(&bo
->bo
, device
, size
);
1524 if (result
!= VK_SUCCESS
) {
1525 vk_free(&device
->alloc
, bo
);
1529 bo
->bo
.flags
= bo_flags
;
1531 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1532 anv_gem_close(device
, bo
->bo
.gem_handle
);
1533 vk_free(&device
->alloc
, bo
);
1534 return vk_errorf(device
->instance
, NULL
,
1535 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1536 "failed to allocate virtual address for BO");
1539 assert(bo
->bo
.gem_handle
);
1541 pthread_mutex_lock(&cache
->mutex
);
1543 _mesa_hash_table_insert(cache
->bo_map
,
1544 (void *)(uintptr_t)bo
->bo
.gem_handle
, bo
);
1546 pthread_mutex_unlock(&cache
->mutex
);
1554 anv_bo_cache_import(struct anv_device
*device
,
1555 struct anv_bo_cache
*cache
,
1556 int fd
, uint64_t bo_flags
,
1557 struct anv_bo
**bo_out
)
1559 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1560 assert(bo_flags
& ANV_BO_EXTERNAL
);
1562 pthread_mutex_lock(&cache
->mutex
);
1564 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1566 pthread_mutex_unlock(&cache
->mutex
);
1567 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1570 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1572 /* We have to be careful how we combine flags so that it makes sense.
1573 * Really, though, if we get to this case and it actually matters, the
1574 * client has imported a BO twice in different ways and they get what
1577 uint64_t new_flags
= ANV_BO_EXTERNAL
;
1578 new_flags
|= (bo
->bo
.flags
| bo_flags
) & EXEC_OBJECT_WRITE
;
1579 new_flags
|= (bo
->bo
.flags
& bo_flags
) & EXEC_OBJECT_ASYNC
;
1580 new_flags
|= (bo
->bo
.flags
& bo_flags
) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1581 new_flags
|= (bo
->bo
.flags
| bo_flags
) & EXEC_OBJECT_PINNED
;
1583 /* It's theoretically possible for a BO to get imported such that it's
1584 * both pinned and not pinned. The only way this can happen is if it
1585 * gets imported as both a semaphore and a memory object and that would
1586 * be an application error. Just fail out in that case.
1588 if ((bo
->bo
.flags
& EXEC_OBJECT_PINNED
) !=
1589 (bo_flags
& EXEC_OBJECT_PINNED
)) {
1590 pthread_mutex_unlock(&cache
->mutex
);
1591 return vk_errorf(device
->instance
, NULL
,
1592 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1593 "The same BO was imported two different ways");
1596 /* It's also theoretically possible that someone could export a BO from
1597 * one heap and import it into another or to import the same BO into two
1598 * different heaps. If this happens, we could potentially end up both
1599 * allowing and disallowing 48-bit addresses. There's not much we can
1600 * do about it if we're pinning so we just throw an error and hope no
1601 * app is actually that stupid.
1603 if ((new_flags
& EXEC_OBJECT_PINNED
) &&
1604 (bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) !=
1605 (bo_flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
)) {
1606 pthread_mutex_unlock(&cache
->mutex
);
1607 return vk_errorf(device
->instance
, NULL
,
1608 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1609 "The same BO was imported on two different heaps");
1612 bo
->bo
.flags
= new_flags
;
1614 __sync_fetch_and_add(&bo
->refcount
, 1);
1616 off_t size
= lseek(fd
, 0, SEEK_END
);
1617 if (size
== (off_t
)-1) {
1618 anv_gem_close(device
, gem_handle
);
1619 pthread_mutex_unlock(&cache
->mutex
);
1620 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1623 bo
= vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1624 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1626 anv_gem_close(device
, gem_handle
);
1627 pthread_mutex_unlock(&cache
->mutex
);
1628 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1633 anv_bo_init(&bo
->bo
, gem_handle
, size
);
1634 bo
->bo
.flags
= bo_flags
;
1636 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1637 anv_gem_close(device
, bo
->bo
.gem_handle
);
1638 pthread_mutex_unlock(&cache
->mutex
);
1639 vk_free(&device
->alloc
, bo
);
1640 return vk_errorf(device
->instance
, NULL
,
1641 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1642 "failed to allocate virtual address for BO");
1645 _mesa_hash_table_insert(cache
->bo_map
, (void *)(uintptr_t)gem_handle
, bo
);
1648 pthread_mutex_unlock(&cache
->mutex
);
1655 anv_bo_cache_export(struct anv_device
*device
,
1656 struct anv_bo_cache
*cache
,
1657 struct anv_bo
*bo_in
, int *fd_out
)
1659 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1660 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1662 /* This BO must have been flagged external in order for us to be able
1663 * to export it. This is done based on external options passed into
1664 * anv_AllocateMemory.
1666 assert(bo
->bo
.flags
& ANV_BO_EXTERNAL
);
1668 int fd
= anv_gem_handle_to_fd(device
, bo
->bo
.gem_handle
);
1670 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1678 atomic_dec_not_one(uint32_t *counter
)
1687 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1696 anv_bo_cache_release(struct anv_device
*device
,
1697 struct anv_bo_cache
*cache
,
1698 struct anv_bo
*bo_in
)
1700 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1701 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1703 /* Try to decrement the counter but don't go below one. If this succeeds
1704 * then the refcount has been decremented and we are not the last
1707 if (atomic_dec_not_one(&bo
->refcount
))
1710 pthread_mutex_lock(&cache
->mutex
);
1712 /* We are probably the last reference since our attempt to decrement above
1713 * failed. However, we can't actually know until we are inside the mutex.
1714 * Otherwise, someone could import the BO between the decrement and our
1717 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1718 /* Turns out we're not the last reference. Unlock and bail. */
1719 pthread_mutex_unlock(&cache
->mutex
);
1723 struct hash_entry
*entry
=
1724 _mesa_hash_table_search(cache
->bo_map
,
1725 (const void *)(uintptr_t)bo
->bo
.gem_handle
);
1727 _mesa_hash_table_remove(cache
->bo_map
, entry
);
1730 anv_gem_munmap(bo
->bo
.map
, bo
->bo
.size
);
1732 anv_vma_free(device
, &bo
->bo
);
1734 anv_gem_close(device
, bo
->bo
.gem_handle
);
1736 /* Don't unlock until we've actually closed the BO. The whole point of
1737 * the BO cache is to ensure that we correctly handle races with creating
1738 * and releasing GEM handles and we don't want to let someone import the BO
1739 * again between mutex unlock and closing the GEM handle.
1741 pthread_mutex_unlock(&cache
->mutex
);
1743 vk_free(&device
->alloc
, bo
);