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. */
103 #define EMPTY2 UINT32_MAX
105 #define PAGE_SIZE 4096
107 struct anv_mmap_cleanup
{
113 #define ANV_MMAP_CLEANUP_INIT ((struct anv_mmap_cleanup){0})
115 #ifndef HAVE_MEMFD_CREATE
117 memfd_create(const char *name
, unsigned int flags
)
119 return syscall(SYS_memfd_create
, name
, flags
);
123 static inline uint32_t
124 ilog2_round_up(uint32_t value
)
127 return 32 - __builtin_clz(value
- 1);
130 static inline uint32_t
131 round_to_power_of_two(uint32_t value
)
133 return 1 << ilog2_round_up(value
);
136 struct anv_state_table_cleanup
{
141 #define ANV_STATE_TABLE_CLEANUP_INIT ((struct anv_state_table_cleanup){0})
142 #define ANV_STATE_ENTRY_SIZE (sizeof(struct anv_free_entry))
145 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
);
148 anv_state_table_init(struct anv_state_table
*table
,
149 struct anv_device
*device
,
150 uint32_t initial_entries
)
154 table
->device
= device
;
156 table
->fd
= memfd_create("state table", MFD_CLOEXEC
);
158 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
160 /* Just make it 2GB up-front. The Linux kernel won't actually back it
161 * with pages until we either map and fault on one of them or we use
162 * userptr and send a chunk of it off to the GPU.
164 if (ftruncate(table
->fd
, BLOCK_POOL_MEMFD_SIZE
) == -1) {
165 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
169 if (!u_vector_init(&table
->mmap_cleanups
,
170 round_to_power_of_two(sizeof(struct anv_state_table_cleanup
)),
172 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
176 table
->state
.next
= 0;
177 table
->state
.end
= 0;
180 uint32_t initial_size
= initial_entries
* ANV_STATE_ENTRY_SIZE
;
181 result
= anv_state_table_expand_range(table
, initial_size
);
182 if (result
!= VK_SUCCESS
)
183 goto fail_mmap_cleanups
;
188 u_vector_finish(&table
->mmap_cleanups
);
196 anv_state_table_expand_range(struct anv_state_table
*table
, uint32_t size
)
199 struct anv_mmap_cleanup
*cleanup
;
201 /* Assert that we only ever grow the pool */
202 assert(size
>= table
->state
.end
);
204 /* Make sure that we don't go outside the bounds of the memfd */
205 if (size
> BLOCK_POOL_MEMFD_SIZE
)
206 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
208 cleanup
= u_vector_add(&table
->mmap_cleanups
);
210 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
212 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
214 /* Just leak the old map until we destroy the pool. We can't munmap it
215 * without races or imposing locking on the block allocate fast path. On
216 * the whole the leaked maps adds up to less than the size of the
217 * current map. MAP_POPULATE seems like the right thing to do, but we
218 * should try to get some numbers.
220 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
221 MAP_SHARED
| MAP_POPULATE
, table
->fd
, 0);
222 if (map
== MAP_FAILED
) {
223 return vk_errorf(table
->device
->instance
, table
->device
,
224 VK_ERROR_OUT_OF_HOST_MEMORY
, "mmap failed: %m");
228 cleanup
->size
= size
;
237 anv_state_table_grow(struct anv_state_table
*table
)
239 VkResult result
= VK_SUCCESS
;
241 uint32_t used
= align_u32(table
->state
.next
* ANV_STATE_ENTRY_SIZE
,
243 uint32_t old_size
= table
->size
;
245 /* The block pool is always initialized to a nonzero size and this function
246 * is always called after initialization.
248 assert(old_size
> 0);
250 uint32_t required
= MAX2(used
, old_size
);
251 if (used
* 2 <= required
) {
252 /* If we're in this case then this isn't the firsta allocation and we
253 * already have enough space on both sides to hold double what we
254 * have allocated. There's nothing for us to do.
259 uint32_t size
= old_size
* 2;
260 while (size
< required
)
263 assert(size
> table
->size
);
265 result
= anv_state_table_expand_range(table
, size
);
272 anv_state_table_finish(struct anv_state_table
*table
)
274 struct anv_state_table_cleanup
*cleanup
;
276 u_vector_foreach(cleanup
, &table
->mmap_cleanups
) {
278 munmap(cleanup
->map
, cleanup
->size
);
281 u_vector_finish(&table
->mmap_cleanups
);
287 anv_state_table_add(struct anv_state_table
*table
, uint32_t *idx
,
290 struct anv_block_state state
, old
, new;
296 state
.u64
= __sync_fetch_and_add(&table
->state
.u64
, count
);
297 if (state
.next
+ count
<= state
.end
) {
299 struct anv_free_entry
*entry
= &table
->map
[state
.next
];
300 for (int i
= 0; i
< count
; i
++) {
301 entry
[i
].state
.idx
= state
.next
+ i
;
305 } else if (state
.next
<= state
.end
) {
306 /* We allocated the first block outside the pool so we have to grow
307 * the pool. pool_state->next acts a mutex: threads who try to
308 * allocate now will get block indexes above the current limit and
309 * hit futex_wait below.
311 new.next
= state
.next
+ count
;
313 result
= anv_state_table_grow(table
);
314 if (result
!= VK_SUCCESS
)
316 new.end
= table
->size
/ ANV_STATE_ENTRY_SIZE
;
317 } while (new.end
< new.next
);
319 old
.u64
= __sync_lock_test_and_set(&table
->state
.u64
, new.u64
);
320 if (old
.next
!= state
.next
)
321 futex_wake(&table
->state
.end
, INT_MAX
);
323 futex_wait(&table
->state
.end
, state
.end
, NULL
);
330 anv_free_list_push2(union anv_free_list2
*list
,
331 struct anv_state_table
*table
,
332 uint32_t first
, uint32_t count
)
334 union anv_free_list2 current
, old
, new;
335 uint32_t last
= first
;
337 for (uint32_t i
= 1; i
< count
; i
++, last
++)
338 table
->map
[last
].next
= last
+ 1;
343 table
->map
[last
].next
= current
.offset
;
345 new.count
= current
.count
+ 1;
346 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
347 } while (old
.u64
!= current
.u64
);
351 anv_free_list_pop2(union anv_free_list2
*list
,
352 struct anv_state_table
*table
)
354 union anv_free_list2 current
, new, old
;
356 current
.u64
= list
->u64
;
357 while (current
.offset
!= EMPTY2
) {
358 __sync_synchronize();
359 new.offset
= table
->map
[current
.offset
].next
;
360 new.count
= current
.count
+ 1;
361 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
362 if (old
.u64
== current
.u64
) {
363 struct anv_free_entry
*entry
= &table
->map
[current
.offset
];
364 return &entry
->state
;
373 anv_free_list_pop(union anv_free_list
*list
, void **map
, int32_t *offset
)
375 union anv_free_list current
, new, old
;
377 current
.u64
= list
->u64
;
378 while (current
.offset
!= EMPTY
) {
379 /* We have to add a memory barrier here so that the list head (and
380 * offset) gets read before we read the map pointer. This way we
381 * know that the map pointer is valid for the given offset at the
382 * point where we read it.
384 __sync_synchronize();
386 int32_t *next_ptr
= *map
+ current
.offset
;
387 new.offset
= VG_NOACCESS_READ(next_ptr
);
388 new.count
= current
.count
+ 1;
389 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
390 if (old
.u64
== current
.u64
) {
391 *offset
= current
.offset
;
401 anv_free_list_push(union anv_free_list
*list
, void *map
, int32_t offset
,
402 uint32_t size
, uint32_t count
)
404 union anv_free_list current
, old
, new;
405 int32_t *next_ptr
= map
+ offset
;
407 /* If we're returning more than one chunk, we need to build a chain to add
408 * to the list. Fortunately, we can do this without any atomics since we
409 * own everything in the chain right now. `offset` is left pointing to the
410 * head of our chain list while `next_ptr` points to the tail.
412 for (uint32_t i
= 1; i
< count
; i
++) {
413 VG_NOACCESS_WRITE(next_ptr
, offset
+ i
* size
);
414 next_ptr
= map
+ offset
+ i
* size
;
420 VG_NOACCESS_WRITE(next_ptr
, current
.offset
);
422 new.count
= current
.count
+ 1;
423 old
.u64
= __sync_val_compare_and_swap(&list
->u64
, current
.u64
, new.u64
);
424 } while (old
.u64
!= current
.u64
);
427 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
428 * means that the bottom 12 bits should all be zero.
430 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
431 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
432 #define PFL_PACK(ptr, count) ({ \
433 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
437 anv_ptr_free_list_pop(void **list
, void **elem
)
439 void *current
= *list
;
440 while (PFL_PTR(current
) != NULL
) {
441 void **next_ptr
= PFL_PTR(current
);
442 void *new_ptr
= VG_NOACCESS_READ(next_ptr
);
443 unsigned new_count
= PFL_COUNT(current
) + 1;
444 void *new = PFL_PACK(new_ptr
, new_count
);
445 void *old
= __sync_val_compare_and_swap(list
, current
, new);
446 if (old
== current
) {
447 *elem
= PFL_PTR(current
);
457 anv_ptr_free_list_push(void **list
, void *elem
)
460 void **next_ptr
= elem
;
462 /* The pointer-based free list requires that the pointer be
463 * page-aligned. This is because we use the bottom 12 bits of the
464 * pointer to store a counter to solve the ABA concurrency problem.
466 assert(((uintptr_t)elem
& 0xfff) == 0);
471 VG_NOACCESS_WRITE(next_ptr
, PFL_PTR(current
));
472 unsigned new_count
= PFL_COUNT(current
) + 1;
473 void *new = PFL_PACK(elem
, new_count
);
474 old
= __sync_val_compare_and_swap(list
, current
, new);
475 } while (old
!= current
);
479 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
480 uint32_t center_bo_offset
, uint32_t size
);
483 anv_block_pool_init(struct anv_block_pool
*pool
,
484 struct anv_device
*device
,
485 uint64_t start_address
,
486 uint32_t initial_size
,
491 pool
->device
= device
;
492 pool
->bo_flags
= bo_flags
;
493 pool
->start_address
= gen_canonical_address(start_address
);
495 anv_bo_init(&pool
->bo
, 0, 0);
497 pool
->fd
= memfd_create("block pool", MFD_CLOEXEC
);
499 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
501 /* Just make it 2GB up-front. The Linux kernel won't actually back it
502 * with pages until we either map and fault on one of them or we use
503 * userptr and send a chunk of it off to the GPU.
505 if (ftruncate(pool
->fd
, BLOCK_POOL_MEMFD_SIZE
) == -1) {
506 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
510 if (!u_vector_init(&pool
->mmap_cleanups
,
511 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
513 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
517 pool
->state
.next
= 0;
519 pool
->back_state
.next
= 0;
520 pool
->back_state
.end
= 0;
522 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
523 if (result
!= VK_SUCCESS
)
524 goto fail_mmap_cleanups
;
529 u_vector_finish(&pool
->mmap_cleanups
);
537 anv_block_pool_finish(struct anv_block_pool
*pool
)
539 struct anv_mmap_cleanup
*cleanup
;
541 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
) {
543 munmap(cleanup
->map
, cleanup
->size
);
544 if (cleanup
->gem_handle
)
545 anv_gem_close(pool
->device
, cleanup
->gem_handle
);
548 u_vector_finish(&pool
->mmap_cleanups
);
554 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
555 uint32_t center_bo_offset
, uint32_t size
)
559 struct anv_mmap_cleanup
*cleanup
;
561 /* Assert that we only ever grow the pool */
562 assert(center_bo_offset
>= pool
->back_state
.end
);
563 assert(size
- center_bo_offset
>= pool
->state
.end
);
565 /* Assert that we don't go outside the bounds of the memfd */
566 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
567 assert(size
- center_bo_offset
<=
568 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
570 cleanup
= u_vector_add(&pool
->mmap_cleanups
);
572 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
574 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
576 /* Just leak the old map until we destroy the pool. We can't munmap it
577 * without races or imposing locking on the block allocate fast path. On
578 * the whole the leaked maps adds up to less than the size of the
579 * current map. MAP_POPULATE seems like the right thing to do, but we
580 * should try to get some numbers.
582 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
583 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
584 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
585 if (map
== MAP_FAILED
)
586 return vk_errorf(pool
->device
->instance
, pool
->device
,
587 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
589 gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
590 if (gem_handle
== 0) {
592 return vk_errorf(pool
->device
->instance
, pool
->device
,
593 VK_ERROR_TOO_MANY_OBJECTS
, "userptr failed: %m");
597 cleanup
->size
= size
;
598 cleanup
->gem_handle
= gem_handle
;
601 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
602 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
603 * always created as I915_CACHING_CACHED, which on non-LLC means
604 * snooped. That can be useful but comes with a bit of overheard. Since
605 * we're eplicitly clflushing and don't want the overhead we need to turn
607 if (!pool
->device
->info
.has_llc
) {
608 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_NONE
);
609 anv_gem_set_domain(pool
->device
, gem_handle
,
610 I915_GEM_DOMAIN_GTT
, I915_GEM_DOMAIN_GTT
);
614 /* Now that we successfull allocated everything, we can write the new
615 * values back into pool. */
616 pool
->map
= map
+ center_bo_offset
;
617 pool
->center_bo_offset
= center_bo_offset
;
619 /* For block pool BOs we have to be a bit careful about where we place them
620 * in the GTT. There are two documented workarounds for state base address
621 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
622 * which state that those two base addresses do not support 48-bit
623 * addresses and need to be placed in the bottom 32-bit range.
624 * Unfortunately, this is not quite accurate.
626 * The real problem is that we always set the size of our state pools in
627 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
628 * likely significantly smaller. We do this because we do not no at the
629 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
630 * the pool during command buffer building so we don't actually have a
631 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
632 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
633 * as being out of bounds and returns zero. For dynamic state, this
634 * usually just leads to rendering corruptions, but shaders that are all
635 * zero hang the GPU immediately.
637 * The easiest solution to do is exactly what the bogus workarounds say to
638 * do: restrict these buffers to 32-bit addresses. We could also pin the
639 * BO to some particular location of our choosing, but that's significantly
640 * more work than just not setting a flag. So, we explicitly DO NOT set
641 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
644 anv_bo_init(&pool
->bo
, gem_handle
, size
);
645 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
) {
646 pool
->bo
.offset
= pool
->start_address
+ BLOCK_POOL_MEMFD_CENTER
-
649 pool
->bo
.flags
= pool
->bo_flags
;
655 /** Returns current memory map of the block pool.
657 * The returned pointer points to the map for the memory at the specified
658 * offset. The offset parameter is relative to the "center" of the block pool
659 * rather than the start of the block pool BO map.
662 anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
)
664 return pool
->map
+ offset
;
667 /** Grows and re-centers the block pool.
669 * We grow the block pool in one or both directions in such a way that the
670 * following conditions are met:
672 * 1) The size of the entire pool is always a power of two.
674 * 2) The pool only grows on both ends. Neither end can get
677 * 3) At the end of the allocation, we have about twice as much space
678 * allocated for each end as we have used. This way the pool doesn't
679 * grow too far in one direction or the other.
681 * 4) If the _alloc_back() has never been called, then the back portion of
682 * the pool retains a size of zero. (This makes it easier for users of
683 * the block pool that only want a one-sided pool.)
685 * 5) We have enough space allocated for at least one more block in
686 * whichever side `state` points to.
688 * 6) The center of the pool is always aligned to both the block_size of
689 * the pool and a 4K CPU page.
692 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
694 VkResult result
= VK_SUCCESS
;
696 pthread_mutex_lock(&pool
->device
->mutex
);
698 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
700 /* Gather a little usage information on the pool. Since we may have
701 * threadsd waiting in queue to get some storage while we resize, it's
702 * actually possible that total_used will be larger than old_size. In
703 * particular, block_pool_alloc() increments state->next prior to
704 * calling block_pool_grow, so this ensures that we get enough space for
705 * which ever side tries to grow the pool.
707 * We align to a page size because it makes it easier to do our
708 * calculations later in such a way that we state page-aigned.
710 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
711 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
712 uint32_t total_used
= front_used
+ back_used
;
714 assert(state
== &pool
->state
|| back_used
> 0);
716 uint32_t old_size
= pool
->bo
.size
;
718 /* The block pool is always initialized to a nonzero size and this function
719 * is always called after initialization.
721 assert(old_size
> 0);
723 /* The back_used and front_used may actually be smaller than the actual
724 * requirement because they are based on the next pointers which are
725 * updated prior to calling this function.
727 uint32_t back_required
= MAX2(back_used
, pool
->center_bo_offset
);
728 uint32_t front_required
= MAX2(front_used
, old_size
- pool
->center_bo_offset
);
730 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
731 /* If we're in this case then this isn't the firsta allocation and we
732 * already have enough space on both sides to hold double what we
733 * have allocated. There's nothing for us to do.
738 uint32_t size
= old_size
* 2;
739 while (size
< back_required
+ front_required
)
742 assert(size
> pool
->bo
.size
);
744 /* We compute a new center_bo_offset such that, when we double the size
745 * of the pool, we maintain the ratio of how much is used by each side.
746 * This way things should remain more-or-less balanced.
748 uint32_t center_bo_offset
;
749 if (back_used
== 0) {
750 /* If we're in this case then we have never called alloc_back(). In
751 * this case, we want keep the offset at 0 to make things as simple
752 * as possible for users that don't care about back allocations.
754 center_bo_offset
= 0;
756 /* Try to "center" the allocation based on how much is currently in
757 * use on each side of the center line.
759 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
761 /* Align down to a multiple of the page size */
762 center_bo_offset
&= ~(PAGE_SIZE
- 1);
764 assert(center_bo_offset
>= back_used
);
766 /* Make sure we don't shrink the back end of the pool */
767 if (center_bo_offset
< back_required
)
768 center_bo_offset
= back_required
;
770 /* Make sure that we don't shrink the front end of the pool */
771 if (size
- center_bo_offset
< front_required
)
772 center_bo_offset
= size
- front_required
;
775 assert(center_bo_offset
% PAGE_SIZE
== 0);
777 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
779 pool
->bo
.flags
= pool
->bo_flags
;
782 pthread_mutex_unlock(&pool
->device
->mutex
);
784 if (result
== VK_SUCCESS
) {
785 /* Return the appropriate new size. This function never actually
786 * updates state->next. Instead, we let the caller do that because it
787 * needs to do so in order to maintain its concurrency model.
789 if (state
== &pool
->state
) {
790 return pool
->bo
.size
- pool
->center_bo_offset
;
792 assert(pool
->center_bo_offset
> 0);
793 return pool
->center_bo_offset
;
801 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
802 struct anv_block_state
*pool_state
,
805 struct anv_block_state state
, old
, new;
808 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
809 if (state
.next
+ block_size
<= state
.end
) {
812 } else if (state
.next
<= state
.end
) {
813 /* We allocated the first block outside the pool so we have to grow
814 * the pool. pool_state->next acts a mutex: threads who try to
815 * allocate now will get block indexes above the current limit and
816 * hit futex_wait below.
818 new.next
= state
.next
+ block_size
;
820 new.end
= anv_block_pool_grow(pool
, pool_state
);
821 } while (new.end
< new.next
);
823 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
824 if (old
.next
!= state
.next
)
825 futex_wake(&pool_state
->end
, INT_MAX
);
828 futex_wait(&pool_state
->end
, state
.end
, NULL
);
835 anv_block_pool_alloc(struct anv_block_pool
*pool
,
838 return anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
);
841 /* Allocates a block out of the back of the block pool.
843 * This will allocated a block earlier than the "start" of the block pool.
844 * The offsets returned from this function will be negative but will still
845 * be correct relative to the block pool's map pointer.
847 * If you ever use anv_block_pool_alloc_back, then you will have to do
848 * gymnastics with the block pool's BO when doing relocations.
851 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
854 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
857 /* The offset we get out of anv_block_pool_alloc_new() is actually the
858 * number of bytes downwards from the middle to the end of the block.
859 * We need to turn it into a (negative) offset from the middle to the
860 * start of the block.
863 return -(offset
+ block_size
);
867 anv_state_pool_init(struct anv_state_pool
*pool
,
868 struct anv_device
*device
,
869 uint64_t start_address
,
873 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
877 if (result
!= VK_SUCCESS
)
880 result
= anv_state_table_init(&pool
->table
, device
, 64);
881 if (result
!= VK_SUCCESS
) {
882 anv_block_pool_finish(&pool
->block_pool
);
886 assert(util_is_power_of_two_or_zero(block_size
));
887 pool
->block_size
= block_size
;
888 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
889 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
890 pool
->buckets
[i
].free_list
= ANV_FREE_LIST2_EMPTY
;
891 pool
->buckets
[i
].block
.next
= 0;
892 pool
->buckets
[i
].block
.end
= 0;
894 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
900 anv_state_pool_finish(struct anv_state_pool
*pool
)
902 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
903 anv_state_table_finish(&pool
->table
);
904 anv_block_pool_finish(&pool
->block_pool
);
908 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
909 struct anv_block_pool
*block_pool
,
913 struct anv_block_state block
, old
, new;
916 /* If our state is large, we don't need any sub-allocation from a block.
917 * Instead, we just grab whole (potentially large) blocks.
919 if (state_size
>= block_size
)
920 return anv_block_pool_alloc(block_pool
, state_size
);
923 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
925 if (block
.next
< block
.end
) {
927 } else if (block
.next
== block
.end
) {
928 offset
= anv_block_pool_alloc(block_pool
, block_size
);
929 new.next
= offset
+ state_size
;
930 new.end
= offset
+ block_size
;
931 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
932 if (old
.next
!= block
.next
)
933 futex_wake(&pool
->block
.end
, INT_MAX
);
936 futex_wait(&pool
->block
.end
, block
.end
, NULL
);
942 anv_state_pool_get_bucket(uint32_t size
)
944 unsigned size_log2
= ilog2_round_up(size
);
945 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
946 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
947 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
948 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
952 anv_state_pool_get_bucket_size(uint32_t bucket
)
954 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
955 return 1 << size_log2
;
958 /** Helper to push a chunk into the state table.
960 * It creates 'count' entries into the state table and update their sizes,
961 * offsets and maps, also pushing them as "free" states.
964 anv_state_pool_return_blocks(struct anv_state_pool
*pool
,
965 uint32_t chunk_offset
, uint32_t count
,
971 /* Make sure we always return chunks aligned to the block_size */
972 assert(chunk_offset
% block_size
== 0);
975 VkResult result
= anv_state_table_add(&pool
->table
, &st_idx
, count
);
976 assert(result
== VK_SUCCESS
);
977 for (int i
= 0; i
< count
; i
++) {
978 /* update states that were added back to the state table */
979 struct anv_state
*state_i
= anv_state_table_get(&pool
->table
,
981 state_i
->alloc_size
= block_size
;
982 state_i
->offset
= chunk_offset
+ block_size
* i
;
983 state_i
->map
= anv_block_pool_map(&pool
->block_pool
, state_i
->offset
);
986 uint32_t block_bucket
= anv_state_pool_get_bucket(block_size
);
987 anv_free_list_push2(&pool
->buckets
[block_bucket
].free_list
,
988 &pool
->table
, st_idx
, count
);
991 static struct anv_state
992 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
993 uint32_t size
, uint32_t align
)
995 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
997 struct anv_state
*state
;
998 uint32_t alloc_size
= anv_state_pool_get_bucket_size(bucket
);
1001 /* Try free list first. */
1002 state
= anv_free_list_pop2(&pool
->buckets
[bucket
].free_list
,
1005 assert(state
->offset
>= 0);
1009 /* Try to grab a chunk from some larger bucket and split it up */
1010 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
1011 state
= anv_free_list_pop2(&pool
->buckets
[b
].free_list
, &pool
->table
);
1013 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
1014 int32_t chunk_offset
= state
->offset
;
1016 /* First lets update the state we got to its new size. offset and map
1019 state
->alloc_size
= alloc_size
;
1021 /* We've found a chunk that's larger than the requested state size.
1022 * There are a couple of options as to what we do with it:
1024 * 1) We could fully split the chunk into state.alloc_size sized
1025 * pieces. However, this would mean that allocating a 16B
1026 * state could potentially split a 2MB chunk into 512K smaller
1027 * chunks. This would lead to unnecessary fragmentation.
1029 * 2) The classic "buddy allocator" method would have us split the
1030 * chunk in half and return one half. Then we would split the
1031 * remaining half in half and return one half, and repeat as
1032 * needed until we get down to the size we want. However, if
1033 * you are allocating a bunch of the same size state (which is
1034 * the common case), this means that every other allocation has
1035 * to go up a level and every fourth goes up two levels, etc.
1036 * This is not nearly as efficient as it could be if we did a
1037 * little more work up-front.
1039 * 3) Split the difference between (1) and (2) by doing a
1040 * two-level split. If it's bigger than some fixed block_size,
1041 * we split it into block_size sized chunks and return all but
1042 * one of them. Then we split what remains into
1043 * state.alloc_size sized chunks and return all but one.
1045 * We choose option (3).
1047 if (chunk_size
> pool
->block_size
&&
1048 alloc_size
< pool
->block_size
) {
1049 assert(chunk_size
% pool
->block_size
== 0);
1050 /* We don't want to split giant chunks into tiny chunks. Instead,
1051 * break anything bigger than a block into block-sized chunks and
1052 * then break it down into bucket-sized chunks from there. Return
1053 * all but the first block of the chunk to the block bucket.
1055 uint32_t push_back
= (chunk_size
/ pool
->block_size
) - 1;
1056 anv_state_pool_return_blocks(pool
, chunk_offset
+ pool
->block_size
,
1057 push_back
, pool
->block_size
);
1058 chunk_size
= pool
->block_size
;
1061 assert(chunk_size
% alloc_size
== 0);
1062 uint32_t push_back
= (chunk_size
/ alloc_size
) - 1;
1063 anv_state_pool_return_blocks(pool
, chunk_offset
+ alloc_size
,
1064 push_back
, alloc_size
);
1069 offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
1073 /* Everytime we allocate a new state, add it to the state pool */
1075 VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1076 assert(result
== VK_SUCCESS
);
1078 state
= anv_state_table_get(&pool
->table
, idx
);
1079 state
->offset
= offset
;
1080 state
->alloc_size
= alloc_size
;
1081 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
);
1088 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
1091 return ANV_STATE_NULL
;
1093 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
1094 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
1099 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
1101 struct anv_state state
;
1102 state
.alloc_size
= pool
->block_size
;
1104 if (anv_free_list_pop(&pool
->back_alloc_free_list
,
1105 &pool
->block_pool
.map
, &state
.offset
)) {
1106 assert(state
.offset
< 0);
1110 state
.offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
1114 state
.map
= pool
->block_pool
.map
+ state
.offset
;
1115 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, state
.alloc_size
));
1120 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
1122 assert(util_is_power_of_two_or_zero(state
.alloc_size
));
1123 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
1125 if (state
.offset
< 0) {
1126 assert(state
.alloc_size
== pool
->block_size
);
1127 anv_free_list_push(&pool
->back_alloc_free_list
,
1128 pool
->block_pool
.map
, state
.offset
,
1129 state
.alloc_size
, 1);
1131 anv_free_list_push2(&pool
->buckets
[bucket
].free_list
,
1132 &pool
->table
, state
.idx
, 1);
1137 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
1139 if (state
.alloc_size
== 0)
1142 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
1143 anv_state_pool_free_no_vg(pool
, state
);
1146 struct anv_state_stream_block
{
1147 struct anv_state block
;
1149 /* The next block */
1150 struct anv_state_stream_block
*next
;
1152 #ifdef HAVE_VALGRIND
1153 /* A pointer to the first user-allocated thing in this block. This is
1154 * what valgrind sees as the start of the block.
1160 /* The state stream allocator is a one-shot, single threaded allocator for
1161 * variable sized blocks. We use it for allocating dynamic state.
1164 anv_state_stream_init(struct anv_state_stream
*stream
,
1165 struct anv_state_pool
*state_pool
,
1166 uint32_t block_size
)
1168 stream
->state_pool
= state_pool
;
1169 stream
->block_size
= block_size
;
1171 stream
->block
= ANV_STATE_NULL
;
1173 stream
->block_list
= NULL
;
1175 /* Ensure that next + whatever > block_size. This way the first call to
1176 * state_stream_alloc fetches a new block.
1178 stream
->next
= block_size
;
1180 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
1184 anv_state_stream_finish(struct anv_state_stream
*stream
)
1186 struct anv_state_stream_block
*next
= stream
->block_list
;
1187 while (next
!= NULL
) {
1188 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
1189 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
1190 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
1191 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
1195 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
1199 anv_state_stream_alloc(struct anv_state_stream
*stream
,
1200 uint32_t size
, uint32_t alignment
)
1203 return ANV_STATE_NULL
;
1205 assert(alignment
<= PAGE_SIZE
);
1207 uint32_t offset
= align_u32(stream
->next
, alignment
);
1208 if (offset
+ size
> stream
->block
.alloc_size
) {
1209 uint32_t block_size
= stream
->block_size
;
1210 if (block_size
< size
)
1211 block_size
= round_to_power_of_two(size
);
1213 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
1214 block_size
, PAGE_SIZE
);
1216 struct anv_state_stream_block
*sb
= stream
->block
.map
;
1217 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
1218 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
1219 stream
->block_list
= sb
;
1220 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
1222 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
1224 /* Reset back to the start plus space for the header */
1225 stream
->next
= sizeof(*sb
);
1227 offset
= align_u32(stream
->next
, alignment
);
1228 assert(offset
+ size
<= stream
->block
.alloc_size
);
1231 struct anv_state state
= stream
->block
;
1232 state
.offset
+= offset
;
1233 state
.alloc_size
= size
;
1234 state
.map
+= offset
;
1236 stream
->next
= offset
+ size
;
1238 #ifdef HAVE_VALGRIND
1239 struct anv_state_stream_block
*sb
= stream
->block_list
;
1240 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
1241 if (vg_ptr
== NULL
) {
1243 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
1244 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
1246 void *state_end
= state
.map
+ state
.alloc_size
;
1247 /* This only updates the mempool. The newly allocated chunk is still
1248 * marked as NOACCESS. */
1249 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
1250 /* Mark the newly allocated chunk as undefined */
1251 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
1258 struct bo_pool_bo_link
{
1259 struct bo_pool_bo_link
*next
;
1264 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
,
1267 pool
->device
= device
;
1268 pool
->bo_flags
= bo_flags
;
1269 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
1271 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
1275 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
1277 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1278 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
1279 while (link
!= NULL
) {
1280 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
1282 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
1283 anv_vma_free(pool
->device
, &link_copy
.bo
);
1284 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
1285 link
= link_copy
.next
;
1289 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1293 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
1297 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1298 const unsigned pow2_size
= 1 << size_log2
;
1299 const unsigned bucket
= size_log2
- 12;
1300 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1302 void *next_free_void
;
1303 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
1304 struct bo_pool_bo_link
*next_free
= next_free_void
;
1305 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
1306 assert(bo
->gem_handle
);
1307 assert(bo
->map
== next_free
);
1308 assert(size
<= bo
->size
);
1310 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1315 struct anv_bo new_bo
;
1317 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
1318 if (result
!= VK_SUCCESS
)
1321 new_bo
.flags
= pool
->bo_flags
;
1323 if (!anv_vma_alloc(pool
->device
, &new_bo
))
1324 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1326 assert(new_bo
.size
== pow2_size
);
1328 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
1329 if (new_bo
.map
== MAP_FAILED
) {
1330 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
1331 anv_vma_free(pool
->device
, &new_bo
);
1332 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1337 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1343 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
1345 /* Make a copy in case the anv_bo happens to be storred in the BO */
1346 struct anv_bo bo
= *bo_in
;
1348 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
1350 struct bo_pool_bo_link
*link
= bo
.map
;
1351 VG_NOACCESS_WRITE(&link
->bo
, bo
);
1353 assert(util_is_power_of_two_or_zero(bo
.size
));
1354 const unsigned size_log2
= ilog2_round_up(bo
.size
);
1355 const unsigned bucket
= size_log2
- 12;
1356 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1358 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
1364 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1366 memset(pool
, 0, sizeof(*pool
));
1370 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1372 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1373 for (unsigned i
= 0; i
< 16; i
++) {
1374 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
1375 if (bo
->exists
> 0) {
1376 anv_vma_free(device
, &bo
->bo
);
1377 anv_gem_close(device
, bo
->bo
.gem_handle
);
1384 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1385 gl_shader_stage stage
, unsigned per_thread_scratch
)
1387 if (per_thread_scratch
== 0)
1390 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1391 assert(scratch_size_log2
< 16);
1393 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
1395 /* We can use "exists" to shortcut and ignore the critical section */
1399 pthread_mutex_lock(&device
->mutex
);
1401 __sync_synchronize();
1403 pthread_mutex_unlock(&device
->mutex
);
1407 const struct anv_physical_device
*physical_device
=
1408 &device
->instance
->physicalDevice
;
1409 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1411 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
1413 unsigned scratch_ids_per_subslice
;
1414 if (devinfo
->is_haswell
) {
1415 /* WaCSScratchSize:hsw
1417 * Haswell's scratch space address calculation appears to be sparse
1418 * rather than tightly packed. The Thread ID has bits indicating
1419 * which subslice, EU within a subslice, and thread within an EU it
1420 * is. There's a maximum of two slices and two subslices, so these
1421 * can be stored with a single bit. Even though there are only 10 EUs
1422 * per subslice, this is stored in 4 bits, so there's an effective
1423 * maximum value of 16 EUs. Similarly, although there are only 7
1424 * threads per EU, this is stored in a 3 bit number, giving an
1425 * effective maximum value of 8 threads per EU.
1427 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1428 * number of threads per subslice.
1430 scratch_ids_per_subslice
= 16 * 8;
1431 } else if (devinfo
->is_cherryview
) {
1432 /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
1433 * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
1436 scratch_ids_per_subslice
= 8 * 7;
1438 scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
1441 uint32_t max_threads
[] = {
1442 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1443 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1444 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1445 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1446 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1447 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1450 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1452 anv_bo_init_new(&bo
->bo
, device
, size
);
1454 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1455 * are still relative to the general state base address. When we emit
1456 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1457 * to the maximum (1 page under 4GB). This allows us to just place the
1458 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1459 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1460 * However, in order to do so, we need to ensure that the kernel does not
1461 * place the scratch BO above the 32-bit boundary.
1463 * NOTE: Technically, it can't go "anywhere" because the top page is off
1464 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1465 * kernel allocates space using
1467 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1469 * so nothing will ever touch the top page.
1471 assert(!(bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
));
1473 if (device
->instance
->physicalDevice
.has_exec_async
)
1474 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1476 if (device
->instance
->physicalDevice
.use_softpin
)
1477 bo
->bo
.flags
|= EXEC_OBJECT_PINNED
;
1479 anv_vma_alloc(device
, &bo
->bo
);
1481 /* Set the exists last because it may be read by other threads */
1482 __sync_synchronize();
1485 pthread_mutex_unlock(&device
->mutex
);
1490 struct anv_cached_bo
{
1497 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1499 cache
->bo_map
= _mesa_pointer_hash_table_create(NULL
);
1501 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1503 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1504 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1505 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1506 "pthread_mutex_init failed: %m");
1513 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1515 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1516 pthread_mutex_destroy(&cache
->mutex
);
1519 static struct anv_cached_bo
*
1520 anv_bo_cache_lookup_locked(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1522 struct hash_entry
*entry
=
1523 _mesa_hash_table_search(cache
->bo_map
,
1524 (const void *)(uintptr_t)gem_handle
);
1528 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)entry
->data
;
1529 assert(bo
->bo
.gem_handle
== gem_handle
);
1534 UNUSED
static struct anv_bo
*
1535 anv_bo_cache_lookup(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1537 pthread_mutex_lock(&cache
->mutex
);
1539 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1541 pthread_mutex_unlock(&cache
->mutex
);
1543 return bo
? &bo
->bo
: NULL
;
1546 #define ANV_BO_CACHE_SUPPORTED_FLAGS \
1547 (EXEC_OBJECT_WRITE | \
1548 EXEC_OBJECT_ASYNC | \
1549 EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
1550 EXEC_OBJECT_PINNED | \
1554 anv_bo_cache_alloc(struct anv_device
*device
,
1555 struct anv_bo_cache
*cache
,
1556 uint64_t size
, uint64_t bo_flags
,
1557 struct anv_bo
**bo_out
)
1559 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1561 struct anv_cached_bo
*bo
=
1562 vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1563 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1565 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1569 /* The kernel is going to give us whole pages anyway */
1570 size
= align_u64(size
, 4096);
1572 VkResult result
= anv_bo_init_new(&bo
->bo
, device
, size
);
1573 if (result
!= VK_SUCCESS
) {
1574 vk_free(&device
->alloc
, bo
);
1578 bo
->bo
.flags
= bo_flags
;
1580 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1581 anv_gem_close(device
, bo
->bo
.gem_handle
);
1582 vk_free(&device
->alloc
, bo
);
1583 return vk_errorf(device
->instance
, NULL
,
1584 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1585 "failed to allocate virtual address for BO");
1588 assert(bo
->bo
.gem_handle
);
1590 pthread_mutex_lock(&cache
->mutex
);
1592 _mesa_hash_table_insert(cache
->bo_map
,
1593 (void *)(uintptr_t)bo
->bo
.gem_handle
, bo
);
1595 pthread_mutex_unlock(&cache
->mutex
);
1603 anv_bo_cache_import(struct anv_device
*device
,
1604 struct anv_bo_cache
*cache
,
1605 int fd
, uint64_t bo_flags
,
1606 struct anv_bo
**bo_out
)
1608 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1609 assert(bo_flags
& ANV_BO_EXTERNAL
);
1611 pthread_mutex_lock(&cache
->mutex
);
1613 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1615 pthread_mutex_unlock(&cache
->mutex
);
1616 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1619 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1621 /* We have to be careful how we combine flags so that it makes sense.
1622 * Really, though, if we get to this case and it actually matters, the
1623 * client has imported a BO twice in different ways and they get what
1626 uint64_t new_flags
= ANV_BO_EXTERNAL
;
1627 new_flags
|= (bo
->bo
.flags
| bo_flags
) & EXEC_OBJECT_WRITE
;
1628 new_flags
|= (bo
->bo
.flags
& bo_flags
) & EXEC_OBJECT_ASYNC
;
1629 new_flags
|= (bo
->bo
.flags
& bo_flags
) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1630 new_flags
|= (bo
->bo
.flags
| bo_flags
) & EXEC_OBJECT_PINNED
;
1632 /* It's theoretically possible for a BO to get imported such that it's
1633 * both pinned and not pinned. The only way this can happen is if it
1634 * gets imported as both a semaphore and a memory object and that would
1635 * be an application error. Just fail out in that case.
1637 if ((bo
->bo
.flags
& EXEC_OBJECT_PINNED
) !=
1638 (bo_flags
& EXEC_OBJECT_PINNED
)) {
1639 pthread_mutex_unlock(&cache
->mutex
);
1640 return vk_errorf(device
->instance
, NULL
,
1641 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1642 "The same BO was imported two different ways");
1645 /* It's also theoretically possible that someone could export a BO from
1646 * one heap and import it into another or to import the same BO into two
1647 * different heaps. If this happens, we could potentially end up both
1648 * allowing and disallowing 48-bit addresses. There's not much we can
1649 * do about it if we're pinning so we just throw an error and hope no
1650 * app is actually that stupid.
1652 if ((new_flags
& EXEC_OBJECT_PINNED
) &&
1653 (bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) !=
1654 (bo_flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
)) {
1655 pthread_mutex_unlock(&cache
->mutex
);
1656 return vk_errorf(device
->instance
, NULL
,
1657 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1658 "The same BO was imported on two different heaps");
1661 bo
->bo
.flags
= new_flags
;
1663 __sync_fetch_and_add(&bo
->refcount
, 1);
1665 off_t size
= lseek(fd
, 0, SEEK_END
);
1666 if (size
== (off_t
)-1) {
1667 anv_gem_close(device
, gem_handle
);
1668 pthread_mutex_unlock(&cache
->mutex
);
1669 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1672 bo
= vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1673 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1675 anv_gem_close(device
, gem_handle
);
1676 pthread_mutex_unlock(&cache
->mutex
);
1677 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1682 anv_bo_init(&bo
->bo
, gem_handle
, size
);
1683 bo
->bo
.flags
= bo_flags
;
1685 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1686 anv_gem_close(device
, bo
->bo
.gem_handle
);
1687 pthread_mutex_unlock(&cache
->mutex
);
1688 vk_free(&device
->alloc
, bo
);
1689 return vk_errorf(device
->instance
, NULL
,
1690 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1691 "failed to allocate virtual address for BO");
1694 _mesa_hash_table_insert(cache
->bo_map
, (void *)(uintptr_t)gem_handle
, bo
);
1697 pthread_mutex_unlock(&cache
->mutex
);
1704 anv_bo_cache_export(struct anv_device
*device
,
1705 struct anv_bo_cache
*cache
,
1706 struct anv_bo
*bo_in
, int *fd_out
)
1708 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1709 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1711 /* This BO must have been flagged external in order for us to be able
1712 * to export it. This is done based on external options passed into
1713 * anv_AllocateMemory.
1715 assert(bo
->bo
.flags
& ANV_BO_EXTERNAL
);
1717 int fd
= anv_gem_handle_to_fd(device
, bo
->bo
.gem_handle
);
1719 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1727 atomic_dec_not_one(uint32_t *counter
)
1736 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1745 anv_bo_cache_release(struct anv_device
*device
,
1746 struct anv_bo_cache
*cache
,
1747 struct anv_bo
*bo_in
)
1749 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1750 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1752 /* Try to decrement the counter but don't go below one. If this succeeds
1753 * then the refcount has been decremented and we are not the last
1756 if (atomic_dec_not_one(&bo
->refcount
))
1759 pthread_mutex_lock(&cache
->mutex
);
1761 /* We are probably the last reference since our attempt to decrement above
1762 * failed. However, we can't actually know until we are inside the mutex.
1763 * Otherwise, someone could import the BO between the decrement and our
1766 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1767 /* Turns out we're not the last reference. Unlock and bail. */
1768 pthread_mutex_unlock(&cache
->mutex
);
1772 struct hash_entry
*entry
=
1773 _mesa_hash_table_search(cache
->bo_map
,
1774 (const void *)(uintptr_t)bo
->bo
.gem_handle
);
1776 _mesa_hash_table_remove(cache
->bo_map
, entry
);
1779 anv_gem_munmap(bo
->bo
.map
, bo
->bo
.size
);
1781 anv_vma_free(device
, &bo
->bo
);
1783 anv_gem_close(device
, bo
->bo
.gem_handle
);
1785 /* Don't unlock until we've actually closed the BO. The whole point of
1786 * the BO cache is to ensure that we correctly handle races with creating
1787 * and releasing GEM handles and we don't want to let someone import the BO
1788 * again between mutex unlock and closing the GEM handle.
1790 pthread_mutex_unlock(&cache
->mutex
);
1792 vk_free(&device
->alloc
, bo
);