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
;
439 pool
->center_bo_offset
= 0;
440 pool
->start_address
= gen_canonical_address(start_address
);
443 /* This pointer will always point to the first BO in the list */
444 pool
->bo
= &pool
->bos
[0];
446 anv_bo_init(pool
->bo
, 0, 0);
448 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
)) {
449 pool
->fd
= memfd_create("block pool", MFD_CLOEXEC
);
451 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
453 /* Just make it 2GB up-front. The Linux kernel won't actually back it
454 * with pages until we either map and fault on one of them or we use
455 * userptr and send a chunk of it off to the GPU.
457 if (ftruncate(pool
->fd
, BLOCK_POOL_MEMFD_SIZE
) == -1) {
458 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
465 if (!u_vector_init(&pool
->mmap_cleanups
,
466 round_to_power_of_two(sizeof(struct anv_mmap_cleanup
)),
468 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
472 pool
->state
.next
= 0;
474 pool
->back_state
.next
= 0;
475 pool
->back_state
.end
= 0;
477 result
= anv_block_pool_expand_range(pool
, 0, initial_size
);
478 if (result
!= VK_SUCCESS
)
479 goto fail_mmap_cleanups
;
484 u_vector_finish(&pool
->mmap_cleanups
);
486 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
493 anv_block_pool_finish(struct anv_block_pool
*pool
)
495 struct anv_mmap_cleanup
*cleanup
;
496 const bool use_softpin
= !!(pool
->bo_flags
& EXEC_OBJECT_PINNED
);
498 u_vector_foreach(cleanup
, &pool
->mmap_cleanups
) {
500 anv_gem_munmap(cleanup
->map
, cleanup
->size
);
502 munmap(cleanup
->map
, cleanup
->size
);
504 if (cleanup
->gem_handle
)
505 anv_gem_close(pool
->device
, cleanup
->gem_handle
);
508 u_vector_finish(&pool
->mmap_cleanups
);
509 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
514 anv_block_pool_expand_range(struct anv_block_pool
*pool
,
515 uint32_t center_bo_offset
, uint32_t size
)
519 struct anv_mmap_cleanup
*cleanup
;
520 const bool use_softpin
= !!(pool
->bo_flags
& EXEC_OBJECT_PINNED
);
522 /* Assert that we only ever grow the pool */
523 assert(center_bo_offset
>= pool
->back_state
.end
);
524 assert(size
- center_bo_offset
>= pool
->state
.end
);
526 /* Assert that we don't go outside the bounds of the memfd */
527 assert(center_bo_offset
<= BLOCK_POOL_MEMFD_CENTER
);
528 assert(use_softpin
||
529 size
- center_bo_offset
<=
530 BLOCK_POOL_MEMFD_SIZE
- BLOCK_POOL_MEMFD_CENTER
);
532 cleanup
= u_vector_add(&pool
->mmap_cleanups
);
534 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
536 *cleanup
= ANV_MMAP_CLEANUP_INIT
;
538 uint32_t newbo_size
= size
- pool
->size
;
540 gem_handle
= anv_gem_create(pool
->device
, newbo_size
);
541 map
= anv_gem_mmap(pool
->device
, gem_handle
, 0, newbo_size
, 0);
542 if (map
== MAP_FAILED
)
543 return vk_errorf(pool
->device
->instance
, pool
->device
,
544 VK_ERROR_MEMORY_MAP_FAILED
, "gem mmap failed: %m");
545 assert(center_bo_offset
== 0);
547 /* Just leak the old map until we destroy the pool. We can't munmap it
548 * without races or imposing locking on the block allocate fast path. On
549 * the whole the leaked maps adds up to less than the size of the
550 * current map. MAP_POPULATE seems like the right thing to do, but we
551 * should try to get some numbers.
553 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
554 MAP_SHARED
| MAP_POPULATE
, pool
->fd
,
555 BLOCK_POOL_MEMFD_CENTER
- center_bo_offset
);
556 if (map
== MAP_FAILED
)
557 return vk_errorf(pool
->device
->instance
, pool
->device
,
558 VK_ERROR_MEMORY_MAP_FAILED
, "mmap failed: %m");
560 /* Now that we mapped the new memory, we can write the new
561 * center_bo_offset back into pool and update pool->map. */
562 pool
->center_bo_offset
= center_bo_offset
;
563 pool
->map
= map
+ center_bo_offset
;
564 gem_handle
= anv_gem_userptr(pool
->device
, map
, size
);
565 if (gem_handle
== 0) {
567 return vk_errorf(pool
->device
->instance
, pool
->device
,
568 VK_ERROR_TOO_MANY_OBJECTS
, "userptr failed: %m");
573 cleanup
->size
= use_softpin
? newbo_size
: size
;
574 cleanup
->gem_handle
= gem_handle
;
576 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
577 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
578 * always created as I915_CACHING_CACHED, which on non-LLC means
581 * On platforms that support softpin, we are not going to use userptr
582 * anymore, but we still want to rely on the snooped states. So make sure
583 * everything is set to I915_CACHING_CACHED.
585 if (!pool
->device
->info
.has_llc
)
586 anv_gem_set_caching(pool
->device
, gem_handle
, I915_CACHING_CACHED
);
588 /* For block pool BOs we have to be a bit careful about where we place them
589 * in the GTT. There are two documented workarounds for state base address
590 * placement : Wa32bitGeneralStateOffset and Wa32bitInstructionBaseOffset
591 * which state that those two base addresses do not support 48-bit
592 * addresses and need to be placed in the bottom 32-bit range.
593 * Unfortunately, this is not quite accurate.
595 * The real problem is that we always set the size of our state pools in
596 * STATE_BASE_ADDRESS to 0xfffff (the maximum) even though the BO is most
597 * likely significantly smaller. We do this because we do not no at the
598 * time we emit STATE_BASE_ADDRESS whether or not we will need to expand
599 * the pool during command buffer building so we don't actually have a
600 * valid final size. If the address + size, as seen by STATE_BASE_ADDRESS
601 * overflows 48 bits, the GPU appears to treat all accesses to the buffer
602 * as being out of bounds and returns zero. For dynamic state, this
603 * usually just leads to rendering corruptions, but shaders that are all
604 * zero hang the GPU immediately.
606 * The easiest solution to do is exactly what the bogus workarounds say to
607 * do: restrict these buffers to 32-bit addresses. We could also pin the
608 * BO to some particular location of our choosing, but that's significantly
609 * more work than just not setting a flag. So, we explicitly DO NOT set
610 * the EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and the kernel does all of the
617 assert(pool
->nbos
< ANV_MAX_BLOCK_POOL_BOS
);
620 /* With softpin, we add a new BO to the pool, and set its offset to right
621 * where the previous BO ends (the end of the pool).
623 bo
= &pool
->bos
[pool
->nbos
++];
624 bo_size
= newbo_size
;
625 bo_offset
= pool
->start_address
+ pool
->size
;
627 /* Without softpin, we just need one BO, and we already have a pointer to
628 * it. Simply "allocate" it from our array if we didn't do it before.
629 * The offset doesn't matter since we are not pinning the BO anyway.
638 anv_bo_init(bo
, gem_handle
, bo_size
);
639 bo
->offset
= bo_offset
;
640 bo
->flags
= pool
->bo_flags
;
647 static struct anv_bo
*
648 anv_block_pool_get_bo(struct anv_block_pool
*pool
, int32_t *offset
)
650 struct anv_bo
*bo
, *bo_found
= NULL
;
651 int32_t cur_offset
= 0;
655 if (!(pool
->bo_flags
& EXEC_OBJECT_PINNED
))
658 anv_block_pool_foreach_bo(bo
, pool
) {
659 if (*offset
< cur_offset
+ bo
->size
) {
663 cur_offset
+= bo
->size
;
666 assert(bo_found
!= NULL
);
667 *offset
-= cur_offset
;
672 /** Returns current memory map of the block pool.
674 * The returned pointer points to the map for the memory at the specified
675 * offset. The offset parameter is relative to the "center" of the block pool
676 * rather than the start of the block pool BO map.
679 anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
)
681 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
) {
682 struct anv_bo
*bo
= anv_block_pool_get_bo(pool
, &offset
);
683 return bo
->map
+ offset
;
685 return pool
->map
+ offset
;
689 /** Grows and re-centers the block pool.
691 * We grow the block pool in one or both directions in such a way that the
692 * following conditions are met:
694 * 1) The size of the entire pool is always a power of two.
696 * 2) The pool only grows on both ends. Neither end can get
699 * 3) At the end of the allocation, we have about twice as much space
700 * allocated for each end as we have used. This way the pool doesn't
701 * grow too far in one direction or the other.
703 * 4) If the _alloc_back() has never been called, then the back portion of
704 * the pool retains a size of zero. (This makes it easier for users of
705 * the block pool that only want a one-sided pool.)
707 * 5) We have enough space allocated for at least one more block in
708 * whichever side `state` points to.
710 * 6) The center of the pool is always aligned to both the block_size of
711 * the pool and a 4K CPU page.
714 anv_block_pool_grow(struct anv_block_pool
*pool
, struct anv_block_state
*state
)
716 VkResult result
= VK_SUCCESS
;
718 pthread_mutex_lock(&pool
->device
->mutex
);
720 assert(state
== &pool
->state
|| state
== &pool
->back_state
);
722 /* Gather a little usage information on the pool. Since we may have
723 * threadsd waiting in queue to get some storage while we resize, it's
724 * actually possible that total_used will be larger than old_size. In
725 * particular, block_pool_alloc() increments state->next prior to
726 * calling block_pool_grow, so this ensures that we get enough space for
727 * which ever side tries to grow the pool.
729 * We align to a page size because it makes it easier to do our
730 * calculations later in such a way that we state page-aigned.
732 uint32_t back_used
= align_u32(pool
->back_state
.next
, PAGE_SIZE
);
733 uint32_t front_used
= align_u32(pool
->state
.next
, PAGE_SIZE
);
734 uint32_t total_used
= front_used
+ back_used
;
736 assert(state
== &pool
->state
|| back_used
> 0);
738 uint32_t old_size
= pool
->size
;
740 /* The block pool is always initialized to a nonzero size and this function
741 * is always called after initialization.
743 assert(old_size
> 0);
745 /* The back_used and front_used may actually be smaller than the actual
746 * requirement because they are based on the next pointers which are
747 * updated prior to calling this function.
749 uint32_t back_required
= MAX2(back_used
, pool
->center_bo_offset
);
750 uint32_t front_required
= MAX2(front_used
, old_size
- pool
->center_bo_offset
);
752 if (back_used
* 2 <= back_required
&& front_used
* 2 <= front_required
) {
753 /* If we're in this case then this isn't the firsta allocation and we
754 * already have enough space on both sides to hold double what we
755 * have allocated. There's nothing for us to do.
760 uint32_t size
= old_size
* 2;
761 while (size
< back_required
+ front_required
)
764 assert(size
> pool
->size
);
766 /* We compute a new center_bo_offset such that, when we double the size
767 * of the pool, we maintain the ratio of how much is used by each side.
768 * This way things should remain more-or-less balanced.
770 uint32_t center_bo_offset
;
771 if (back_used
== 0) {
772 /* If we're in this case then we have never called alloc_back(). In
773 * this case, we want keep the offset at 0 to make things as simple
774 * as possible for users that don't care about back allocations.
776 center_bo_offset
= 0;
778 /* Try to "center" the allocation based on how much is currently in
779 * use on each side of the center line.
781 center_bo_offset
= ((uint64_t)size
* back_used
) / total_used
;
783 /* Align down to a multiple of the page size */
784 center_bo_offset
&= ~(PAGE_SIZE
- 1);
786 assert(center_bo_offset
>= back_used
);
788 /* Make sure we don't shrink the back end of the pool */
789 if (center_bo_offset
< back_required
)
790 center_bo_offset
= back_required
;
792 /* Make sure that we don't shrink the front end of the pool */
793 if (size
- center_bo_offset
< front_required
)
794 center_bo_offset
= size
- front_required
;
797 assert(center_bo_offset
% PAGE_SIZE
== 0);
799 result
= anv_block_pool_expand_range(pool
, center_bo_offset
, size
);
801 pool
->bo
->flags
= pool
->bo_flags
;
804 pthread_mutex_unlock(&pool
->device
->mutex
);
806 if (result
== VK_SUCCESS
) {
807 /* Return the appropriate new size. This function never actually
808 * updates state->next. Instead, we let the caller do that because it
809 * needs to do so in order to maintain its concurrency model.
811 if (state
== &pool
->state
) {
812 return pool
->size
- pool
->center_bo_offset
;
814 assert(pool
->center_bo_offset
> 0);
815 return pool
->center_bo_offset
;
823 anv_block_pool_alloc_new(struct anv_block_pool
*pool
,
824 struct anv_block_state
*pool_state
,
825 uint32_t block_size
, uint32_t *padding
)
827 struct anv_block_state state
, old
, new;
829 /* Most allocations won't generate any padding */
834 state
.u64
= __sync_fetch_and_add(&pool_state
->u64
, block_size
);
835 if (state
.next
+ block_size
<= state
.end
) {
837 } else if (state
.next
<= state
.end
) {
838 if (pool
->bo_flags
& EXEC_OBJECT_PINNED
&& state
.next
< state
.end
) {
839 /* We need to grow the block pool, but still have some leftover
840 * space that can't be used by that particular allocation. So we
841 * add that as a "padding", and return it.
843 uint32_t leftover
= state
.end
- state
.next
;
845 /* If there is some leftover space in the pool, the caller must
848 assert(leftover
== 0 || padding
);
851 state
.next
+= leftover
;
854 /* We allocated the first block outside the pool so we have to grow
855 * the pool. pool_state->next acts a mutex: threads who try to
856 * allocate now will get block indexes above the current limit and
857 * hit futex_wait below.
859 new.next
= state
.next
+ block_size
;
861 new.end
= anv_block_pool_grow(pool
, pool_state
);
862 } while (new.end
< new.next
);
864 old
.u64
= __sync_lock_test_and_set(&pool_state
->u64
, new.u64
);
865 if (old
.next
!= state
.next
)
866 futex_wake(&pool_state
->end
, INT_MAX
);
869 futex_wait(&pool_state
->end
, state
.end
, NULL
);
876 anv_block_pool_alloc(struct anv_block_pool
*pool
,
877 uint32_t block_size
, uint32_t *padding
)
881 offset
= anv_block_pool_alloc_new(pool
, &pool
->state
, block_size
, padding
);
886 /* Allocates a block out of the back of the block pool.
888 * This will allocated a block earlier than the "start" of the block pool.
889 * The offsets returned from this function will be negative but will still
890 * be correct relative to the block pool's map pointer.
892 * If you ever use anv_block_pool_alloc_back, then you will have to do
893 * gymnastics with the block pool's BO when doing relocations.
896 anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
899 int32_t offset
= anv_block_pool_alloc_new(pool
, &pool
->back_state
,
902 /* The offset we get out of anv_block_pool_alloc_new() is actually the
903 * number of bytes downwards from the middle to the end of the block.
904 * We need to turn it into a (negative) offset from the middle to the
905 * start of the block.
908 return -(offset
+ block_size
);
912 anv_state_pool_init(struct anv_state_pool
*pool
,
913 struct anv_device
*device
,
914 uint64_t start_address
,
918 VkResult result
= anv_block_pool_init(&pool
->block_pool
, device
,
922 if (result
!= VK_SUCCESS
)
925 result
= anv_state_table_init(&pool
->table
, device
, 64);
926 if (result
!= VK_SUCCESS
) {
927 anv_block_pool_finish(&pool
->block_pool
);
931 assert(util_is_power_of_two_or_zero(block_size
));
932 pool
->block_size
= block_size
;
933 pool
->back_alloc_free_list
= ANV_FREE_LIST_EMPTY
;
934 for (unsigned i
= 0; i
< ANV_STATE_BUCKETS
; i
++) {
935 pool
->buckets
[i
].free_list
= ANV_FREE_LIST_EMPTY
;
936 pool
->buckets
[i
].block
.next
= 0;
937 pool
->buckets
[i
].block
.end
= 0;
939 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
945 anv_state_pool_finish(struct anv_state_pool
*pool
)
947 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
948 anv_state_table_finish(&pool
->table
);
949 anv_block_pool_finish(&pool
->block_pool
);
953 anv_fixed_size_state_pool_alloc_new(struct anv_fixed_size_state_pool
*pool
,
954 struct anv_block_pool
*block_pool
,
959 struct anv_block_state block
, old
, new;
962 /* We don't always use anv_block_pool_alloc(), which would set *padding to
963 * zero for us. So if we have a pointer to padding, we must zero it out
964 * ourselves here, to make sure we always return some sensible value.
969 /* If our state is large, we don't need any sub-allocation from a block.
970 * Instead, we just grab whole (potentially large) blocks.
972 if (state_size
>= block_size
)
973 return anv_block_pool_alloc(block_pool
, state_size
, padding
);
976 block
.u64
= __sync_fetch_and_add(&pool
->block
.u64
, state_size
);
978 if (block
.next
< block
.end
) {
980 } else if (block
.next
== block
.end
) {
981 offset
= anv_block_pool_alloc(block_pool
, block_size
, padding
);
982 new.next
= offset
+ state_size
;
983 new.end
= offset
+ block_size
;
984 old
.u64
= __sync_lock_test_and_set(&pool
->block
.u64
, new.u64
);
985 if (old
.next
!= block
.next
)
986 futex_wake(&pool
->block
.end
, INT_MAX
);
989 futex_wait(&pool
->block
.end
, block
.end
, NULL
);
995 anv_state_pool_get_bucket(uint32_t size
)
997 unsigned size_log2
= ilog2_round_up(size
);
998 assert(size_log2
<= ANV_MAX_STATE_SIZE_LOG2
);
999 if (size_log2
< ANV_MIN_STATE_SIZE_LOG2
)
1000 size_log2
= ANV_MIN_STATE_SIZE_LOG2
;
1001 return size_log2
- ANV_MIN_STATE_SIZE_LOG2
;
1005 anv_state_pool_get_bucket_size(uint32_t bucket
)
1007 uint32_t size_log2
= bucket
+ ANV_MIN_STATE_SIZE_LOG2
;
1008 return 1 << size_log2
;
1011 /** Helper to push a chunk into the state table.
1013 * It creates 'count' entries into the state table and update their sizes,
1014 * offsets and maps, also pushing them as "free" states.
1017 anv_state_pool_return_blocks(struct anv_state_pool
*pool
,
1018 uint32_t chunk_offset
, uint32_t count
,
1019 uint32_t block_size
)
1021 /* Disallow returning 0 chunks */
1024 /* Make sure we always return chunks aligned to the block_size */
1025 assert(chunk_offset
% block_size
== 0);
1028 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &st_idx
, count
);
1029 assert(result
== VK_SUCCESS
);
1030 for (int i
= 0; i
< count
; i
++) {
1031 /* update states that were added back to the state table */
1032 struct anv_state
*state_i
= anv_state_table_get(&pool
->table
,
1034 state_i
->alloc_size
= block_size
;
1035 state_i
->offset
= chunk_offset
+ block_size
* i
;
1036 state_i
->map
= anv_block_pool_map(&pool
->block_pool
, state_i
->offset
);
1039 uint32_t block_bucket
= anv_state_pool_get_bucket(block_size
);
1040 anv_free_list_push(&pool
->buckets
[block_bucket
].free_list
,
1041 &pool
->table
, st_idx
, count
);
1044 /** Returns a chunk of memory back to the state pool.
1046 * Do a two-level split. If chunk_size is bigger than divisor
1047 * (pool->block_size), we return as many divisor sized blocks as we can, from
1048 * the end of the chunk.
1050 * The remaining is then split into smaller blocks (starting at small_size if
1051 * it is non-zero), with larger blocks always being taken from the end of the
1055 anv_state_pool_return_chunk(struct anv_state_pool
*pool
,
1056 uint32_t chunk_offset
, uint32_t chunk_size
,
1057 uint32_t small_size
)
1059 uint32_t divisor
= pool
->block_size
;
1060 uint32_t nblocks
= chunk_size
/ divisor
;
1061 uint32_t rest
= chunk_size
- nblocks
* divisor
;
1064 /* First return divisor aligned and sized chunks. We start returning
1065 * larger blocks from the end fo the chunk, since they should already be
1066 * aligned to divisor. Also anv_state_pool_return_blocks() only accepts
1069 uint32_t offset
= chunk_offset
+ rest
;
1070 anv_state_pool_return_blocks(pool
, offset
, nblocks
, divisor
);
1076 if (small_size
> 0 && small_size
< divisor
)
1077 divisor
= small_size
;
1079 uint32_t min_size
= 1 << ANV_MIN_STATE_SIZE_LOG2
;
1081 /* Just as before, return larger divisor aligned blocks from the end of the
1084 while (chunk_size
> 0 && divisor
>= min_size
) {
1085 nblocks
= chunk_size
/ divisor
;
1086 rest
= chunk_size
- nblocks
* divisor
;
1088 anv_state_pool_return_blocks(pool
, chunk_offset
+ rest
,
1096 static struct anv_state
1097 anv_state_pool_alloc_no_vg(struct anv_state_pool
*pool
,
1098 uint32_t size
, uint32_t align
)
1100 uint32_t bucket
= anv_state_pool_get_bucket(MAX2(size
, align
));
1102 struct anv_state
*state
;
1103 uint32_t alloc_size
= anv_state_pool_get_bucket_size(bucket
);
1106 /* Try free list first. */
1107 state
= anv_free_list_pop(&pool
->buckets
[bucket
].free_list
,
1110 assert(state
->offset
>= 0);
1114 /* Try to grab a chunk from some larger bucket and split it up */
1115 for (unsigned b
= bucket
+ 1; b
< ANV_STATE_BUCKETS
; b
++) {
1116 state
= anv_free_list_pop(&pool
->buckets
[b
].free_list
, &pool
->table
);
1118 unsigned chunk_size
= anv_state_pool_get_bucket_size(b
);
1119 int32_t chunk_offset
= state
->offset
;
1121 /* First lets update the state we got to its new size. offset and map
1124 state
->alloc_size
= alloc_size
;
1126 /* Now return the unused part of the chunk back to the pool as free
1129 * There are a couple of options as to what we do with it:
1131 * 1) We could fully split the chunk into state.alloc_size sized
1132 * pieces. However, this would mean that allocating a 16B
1133 * state could potentially split a 2MB chunk into 512K smaller
1134 * chunks. This would lead to unnecessary fragmentation.
1136 * 2) The classic "buddy allocator" method would have us split the
1137 * chunk in half and return one half. Then we would split the
1138 * remaining half in half and return one half, and repeat as
1139 * needed until we get down to the size we want. However, if
1140 * you are allocating a bunch of the same size state (which is
1141 * the common case), this means that every other allocation has
1142 * to go up a level and every fourth goes up two levels, etc.
1143 * This is not nearly as efficient as it could be if we did a
1144 * little more work up-front.
1146 * 3) Split the difference between (1) and (2) by doing a
1147 * two-level split. If it's bigger than some fixed block_size,
1148 * we split it into block_size sized chunks and return all but
1149 * one of them. Then we split what remains into
1150 * state.alloc_size sized chunks and return them.
1152 * We choose something close to option (3), which is implemented with
1153 * anv_state_pool_return_chunk(). That is done by returning the
1154 * remaining of the chunk, with alloc_size as a hint of the size that
1155 * we want the smaller chunk split into.
1157 anv_state_pool_return_chunk(pool
, chunk_offset
+ alloc_size
,
1158 chunk_size
- alloc_size
, alloc_size
);
1164 offset
= anv_fixed_size_state_pool_alloc_new(&pool
->buckets
[bucket
],
1169 /* Everytime we allocate a new state, add it to the state pool */
1171 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1172 assert(result
== VK_SUCCESS
);
1174 state
= anv_state_table_get(&pool
->table
, idx
);
1175 state
->offset
= offset
;
1176 state
->alloc_size
= alloc_size
;
1177 state
->map
= anv_block_pool_map(&pool
->block_pool
, offset
);
1180 uint32_t return_offset
= offset
- padding
;
1181 anv_state_pool_return_chunk(pool
, return_offset
, padding
, 0);
1189 anv_state_pool_alloc(struct anv_state_pool
*pool
, uint32_t size
, uint32_t align
)
1192 return ANV_STATE_NULL
;
1194 struct anv_state state
= anv_state_pool_alloc_no_vg(pool
, size
, align
);
1195 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
.map
, size
));
1200 anv_state_pool_alloc_back(struct anv_state_pool
*pool
)
1202 struct anv_state
*state
;
1203 uint32_t alloc_size
= pool
->block_size
;
1205 state
= anv_free_list_pop(&pool
->back_alloc_free_list
, &pool
->table
);
1207 assert(state
->offset
< 0);
1212 offset
= anv_block_pool_alloc_back(&pool
->block_pool
,
1215 UNUSED VkResult result
= anv_state_table_add(&pool
->table
, &idx
, 1);
1216 assert(result
== VK_SUCCESS
);
1218 state
= anv_state_table_get(&pool
->table
, idx
);
1219 state
->offset
= offset
;
1220 state
->alloc_size
= alloc_size
;
1221 state
->map
= anv_block_pool_map(&pool
->block_pool
, state
->offset
);
1224 VG(VALGRIND_MEMPOOL_ALLOC(pool
, state
->map
, state
->alloc_size
));
1229 anv_state_pool_free_no_vg(struct anv_state_pool
*pool
, struct anv_state state
)
1231 assert(util_is_power_of_two_or_zero(state
.alloc_size
));
1232 unsigned bucket
= anv_state_pool_get_bucket(state
.alloc_size
);
1234 if (state
.offset
< 0) {
1235 assert(state
.alloc_size
== pool
->block_size
);
1236 anv_free_list_push(&pool
->back_alloc_free_list
,
1237 &pool
->table
, state
.idx
, 1);
1239 anv_free_list_push(&pool
->buckets
[bucket
].free_list
,
1240 &pool
->table
, state
.idx
, 1);
1245 anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
)
1247 if (state
.alloc_size
== 0)
1250 VG(VALGRIND_MEMPOOL_FREE(pool
, state
.map
));
1251 anv_state_pool_free_no_vg(pool
, state
);
1254 struct anv_state_stream_block
{
1255 struct anv_state block
;
1257 /* The next block */
1258 struct anv_state_stream_block
*next
;
1260 #ifdef HAVE_VALGRIND
1261 /* A pointer to the first user-allocated thing in this block. This is
1262 * what valgrind sees as the start of the block.
1268 /* The state stream allocator is a one-shot, single threaded allocator for
1269 * variable sized blocks. We use it for allocating dynamic state.
1272 anv_state_stream_init(struct anv_state_stream
*stream
,
1273 struct anv_state_pool
*state_pool
,
1274 uint32_t block_size
)
1276 stream
->state_pool
= state_pool
;
1277 stream
->block_size
= block_size
;
1279 stream
->block
= ANV_STATE_NULL
;
1281 stream
->block_list
= NULL
;
1283 /* Ensure that next + whatever > block_size. This way the first call to
1284 * state_stream_alloc fetches a new block.
1286 stream
->next
= block_size
;
1288 VG(VALGRIND_CREATE_MEMPOOL(stream
, 0, false));
1292 anv_state_stream_finish(struct anv_state_stream
*stream
)
1294 struct anv_state_stream_block
*next
= stream
->block_list
;
1295 while (next
!= NULL
) {
1296 struct anv_state_stream_block sb
= VG_NOACCESS_READ(next
);
1297 VG(VALGRIND_MEMPOOL_FREE(stream
, sb
._vg_ptr
));
1298 VG(VALGRIND_MAKE_MEM_UNDEFINED(next
, stream
->block_size
));
1299 anv_state_pool_free_no_vg(stream
->state_pool
, sb
.block
);
1303 VG(VALGRIND_DESTROY_MEMPOOL(stream
));
1307 anv_state_stream_alloc(struct anv_state_stream
*stream
,
1308 uint32_t size
, uint32_t alignment
)
1311 return ANV_STATE_NULL
;
1313 assert(alignment
<= PAGE_SIZE
);
1315 uint32_t offset
= align_u32(stream
->next
, alignment
);
1316 if (offset
+ size
> stream
->block
.alloc_size
) {
1317 uint32_t block_size
= stream
->block_size
;
1318 if (block_size
< size
)
1319 block_size
= round_to_power_of_two(size
);
1321 stream
->block
= anv_state_pool_alloc_no_vg(stream
->state_pool
,
1322 block_size
, PAGE_SIZE
);
1324 struct anv_state_stream_block
*sb
= stream
->block
.map
;
1325 VG_NOACCESS_WRITE(&sb
->block
, stream
->block
);
1326 VG_NOACCESS_WRITE(&sb
->next
, stream
->block_list
);
1327 stream
->block_list
= sb
;
1328 VG(VG_NOACCESS_WRITE(&sb
->_vg_ptr
, NULL
));
1330 VG(VALGRIND_MAKE_MEM_NOACCESS(stream
->block
.map
, stream
->block_size
));
1332 /* Reset back to the start plus space for the header */
1333 stream
->next
= sizeof(*sb
);
1335 offset
= align_u32(stream
->next
, alignment
);
1336 assert(offset
+ size
<= stream
->block
.alloc_size
);
1339 struct anv_state state
= stream
->block
;
1340 state
.offset
+= offset
;
1341 state
.alloc_size
= size
;
1342 state
.map
+= offset
;
1344 stream
->next
= offset
+ size
;
1346 #ifdef HAVE_VALGRIND
1347 struct anv_state_stream_block
*sb
= stream
->block_list
;
1348 void *vg_ptr
= VG_NOACCESS_READ(&sb
->_vg_ptr
);
1349 if (vg_ptr
== NULL
) {
1351 VG_NOACCESS_WRITE(&sb
->_vg_ptr
, vg_ptr
);
1352 VALGRIND_MEMPOOL_ALLOC(stream
, vg_ptr
, size
);
1354 void *state_end
= state
.map
+ state
.alloc_size
;
1355 /* This only updates the mempool. The newly allocated chunk is still
1356 * marked as NOACCESS. */
1357 VALGRIND_MEMPOOL_CHANGE(stream
, vg_ptr
, vg_ptr
, state_end
- vg_ptr
);
1358 /* Mark the newly allocated chunk as undefined */
1359 VALGRIND_MAKE_MEM_UNDEFINED(state
.map
, state
.alloc_size
);
1366 struct bo_pool_bo_link
{
1367 struct bo_pool_bo_link
*next
;
1372 anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
,
1375 pool
->device
= device
;
1376 pool
->bo_flags
= bo_flags
;
1377 memset(pool
->free_list
, 0, sizeof(pool
->free_list
));
1379 VG(VALGRIND_CREATE_MEMPOOL(pool
, 0, false));
1383 anv_bo_pool_finish(struct anv_bo_pool
*pool
)
1385 for (unsigned i
= 0; i
< ARRAY_SIZE(pool
->free_list
); i
++) {
1386 struct bo_pool_bo_link
*link
= PFL_PTR(pool
->free_list
[i
]);
1387 while (link
!= NULL
) {
1388 struct bo_pool_bo_link link_copy
= VG_NOACCESS_READ(link
);
1390 anv_gem_munmap(link_copy
.bo
.map
, link_copy
.bo
.size
);
1391 anv_vma_free(pool
->device
, &link_copy
.bo
);
1392 anv_gem_close(pool
->device
, link_copy
.bo
.gem_handle
);
1393 link
= link_copy
.next
;
1397 VG(VALGRIND_DESTROY_MEMPOOL(pool
));
1401 anv_bo_pool_alloc(struct anv_bo_pool
*pool
, struct anv_bo
*bo
, uint32_t size
)
1405 const unsigned size_log2
= size
< 4096 ? 12 : ilog2_round_up(size
);
1406 const unsigned pow2_size
= 1 << size_log2
;
1407 const unsigned bucket
= size_log2
- 12;
1408 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1410 void *next_free_void
;
1411 if (anv_ptr_free_list_pop(&pool
->free_list
[bucket
], &next_free_void
)) {
1412 struct bo_pool_bo_link
*next_free
= next_free_void
;
1413 *bo
= VG_NOACCESS_READ(&next_free
->bo
);
1414 assert(bo
->gem_handle
);
1415 assert(bo
->map
== next_free
);
1416 assert(size
<= bo
->size
);
1418 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1423 struct anv_bo new_bo
;
1425 result
= anv_bo_init_new(&new_bo
, pool
->device
, pow2_size
);
1426 if (result
!= VK_SUCCESS
)
1429 new_bo
.flags
= pool
->bo_flags
;
1431 if (!anv_vma_alloc(pool
->device
, &new_bo
))
1432 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1434 assert(new_bo
.size
== pow2_size
);
1436 new_bo
.map
= anv_gem_mmap(pool
->device
, new_bo
.gem_handle
, 0, pow2_size
, 0);
1437 if (new_bo
.map
== MAP_FAILED
) {
1438 anv_gem_close(pool
->device
, new_bo
.gem_handle
);
1439 anv_vma_free(pool
->device
, &new_bo
);
1440 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1443 /* We are removing the state flushes, so lets make sure that these buffers
1444 * are cached/snooped.
1446 if (!pool
->device
->info
.has_llc
) {
1447 anv_gem_set_caching(pool
->device
, new_bo
.gem_handle
,
1448 I915_CACHING_CACHED
);
1453 VG(VALGRIND_MEMPOOL_ALLOC(pool
, bo
->map
, size
));
1459 anv_bo_pool_free(struct anv_bo_pool
*pool
, const struct anv_bo
*bo_in
)
1461 /* Make a copy in case the anv_bo happens to be storred in the BO */
1462 struct anv_bo bo
= *bo_in
;
1464 VG(VALGRIND_MEMPOOL_FREE(pool
, bo
.map
));
1466 struct bo_pool_bo_link
*link
= bo
.map
;
1467 VG_NOACCESS_WRITE(&link
->bo
, bo
);
1469 assert(util_is_power_of_two_or_zero(bo
.size
));
1470 const unsigned size_log2
= ilog2_round_up(bo
.size
);
1471 const unsigned bucket
= size_log2
- 12;
1472 assert(bucket
< ARRAY_SIZE(pool
->free_list
));
1474 anv_ptr_free_list_push(&pool
->free_list
[bucket
], link
);
1480 anv_scratch_pool_init(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1482 memset(pool
, 0, sizeof(*pool
));
1486 anv_scratch_pool_finish(struct anv_device
*device
, struct anv_scratch_pool
*pool
)
1488 for (unsigned s
= 0; s
< MESA_SHADER_STAGES
; s
++) {
1489 for (unsigned i
= 0; i
< 16; i
++) {
1490 struct anv_scratch_bo
*bo
= &pool
->bos
[i
][s
];
1491 if (bo
->exists
> 0) {
1492 anv_vma_free(device
, &bo
->bo
);
1493 anv_gem_close(device
, bo
->bo
.gem_handle
);
1500 anv_scratch_pool_alloc(struct anv_device
*device
, struct anv_scratch_pool
*pool
,
1501 gl_shader_stage stage
, unsigned per_thread_scratch
)
1503 if (per_thread_scratch
== 0)
1506 unsigned scratch_size_log2
= ffs(per_thread_scratch
/ 2048);
1507 assert(scratch_size_log2
< 16);
1509 struct anv_scratch_bo
*bo
= &pool
->bos
[scratch_size_log2
][stage
];
1511 /* We can use "exists" to shortcut and ignore the critical section */
1515 pthread_mutex_lock(&device
->mutex
);
1517 __sync_synchronize();
1519 pthread_mutex_unlock(&device
->mutex
);
1523 const struct anv_physical_device
*physical_device
=
1524 &device
->instance
->physicalDevice
;
1525 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1527 const unsigned subslices
= MAX2(physical_device
->subslice_total
, 1);
1529 unsigned scratch_ids_per_subslice
;
1530 if (devinfo
->is_haswell
) {
1531 /* WaCSScratchSize:hsw
1533 * Haswell's scratch space address calculation appears to be sparse
1534 * rather than tightly packed. The Thread ID has bits indicating
1535 * which subslice, EU within a subslice, and thread within an EU it
1536 * is. There's a maximum of two slices and two subslices, so these
1537 * can be stored with a single bit. Even though there are only 10 EUs
1538 * per subslice, this is stored in 4 bits, so there's an effective
1539 * maximum value of 16 EUs. Similarly, although there are only 7
1540 * threads per EU, this is stored in a 3 bit number, giving an
1541 * effective maximum value of 8 threads per EU.
1543 * This means that we need to use 16 * 8 instead of 10 * 7 for the
1544 * number of threads per subslice.
1546 scratch_ids_per_subslice
= 16 * 8;
1547 } else if (devinfo
->is_cherryview
) {
1548 /* Cherryview devices have either 6 or 8 EUs per subslice, and each EU
1549 * has 7 threads. The 6 EU devices appear to calculate thread IDs as if
1552 scratch_ids_per_subslice
= 8 * 7;
1554 scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
1557 uint32_t max_threads
[] = {
1558 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
1559 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
1560 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
1561 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
1562 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
1563 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslices
,
1566 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
1568 anv_bo_init_new(&bo
->bo
, device
, size
);
1570 /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
1571 * are still relative to the general state base address. When we emit
1572 * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
1573 * to the maximum (1 page under 4GB). This allows us to just place the
1574 * scratch buffers anywhere we wish in the bottom 32 bits of address space
1575 * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
1576 * However, in order to do so, we need to ensure that the kernel does not
1577 * place the scratch BO above the 32-bit boundary.
1579 * NOTE: Technically, it can't go "anywhere" because the top page is off
1580 * limits. However, when EXEC_OBJECT_SUPPORTS_48B_ADDRESS is set, the
1581 * kernel allocates space using
1583 * end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
1585 * so nothing will ever touch the top page.
1587 assert(!(bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
));
1589 if (device
->instance
->physicalDevice
.has_exec_async
)
1590 bo
->bo
.flags
|= EXEC_OBJECT_ASYNC
;
1592 if (device
->instance
->physicalDevice
.use_softpin
)
1593 bo
->bo
.flags
|= EXEC_OBJECT_PINNED
;
1595 anv_vma_alloc(device
, &bo
->bo
);
1597 /* Set the exists last because it may be read by other threads */
1598 __sync_synchronize();
1601 pthread_mutex_unlock(&device
->mutex
);
1606 struct anv_cached_bo
{
1613 anv_bo_cache_init(struct anv_bo_cache
*cache
)
1615 cache
->bo_map
= _mesa_pointer_hash_table_create(NULL
);
1617 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1619 if (pthread_mutex_init(&cache
->mutex
, NULL
)) {
1620 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1621 return vk_errorf(NULL
, NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
,
1622 "pthread_mutex_init failed: %m");
1629 anv_bo_cache_finish(struct anv_bo_cache
*cache
)
1631 _mesa_hash_table_destroy(cache
->bo_map
, NULL
);
1632 pthread_mutex_destroy(&cache
->mutex
);
1635 static struct anv_cached_bo
*
1636 anv_bo_cache_lookup_locked(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1638 struct hash_entry
*entry
=
1639 _mesa_hash_table_search(cache
->bo_map
,
1640 (const void *)(uintptr_t)gem_handle
);
1644 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)entry
->data
;
1645 assert(bo
->bo
.gem_handle
== gem_handle
);
1650 UNUSED
static struct anv_bo
*
1651 anv_bo_cache_lookup(struct anv_bo_cache
*cache
, uint32_t gem_handle
)
1653 pthread_mutex_lock(&cache
->mutex
);
1655 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1657 pthread_mutex_unlock(&cache
->mutex
);
1659 return bo
? &bo
->bo
: NULL
;
1662 #define ANV_BO_CACHE_SUPPORTED_FLAGS \
1663 (EXEC_OBJECT_WRITE | \
1664 EXEC_OBJECT_ASYNC | \
1665 EXEC_OBJECT_SUPPORTS_48B_ADDRESS | \
1666 EXEC_OBJECT_PINNED | \
1670 anv_bo_cache_alloc(struct anv_device
*device
,
1671 struct anv_bo_cache
*cache
,
1672 uint64_t size
, uint64_t bo_flags
,
1673 struct anv_bo
**bo_out
)
1675 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1677 struct anv_cached_bo
*bo
=
1678 vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1679 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1681 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1685 /* The kernel is going to give us whole pages anyway */
1686 size
= align_u64(size
, 4096);
1688 VkResult result
= anv_bo_init_new(&bo
->bo
, device
, size
);
1689 if (result
!= VK_SUCCESS
) {
1690 vk_free(&device
->alloc
, bo
);
1694 bo
->bo
.flags
= bo_flags
;
1696 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1697 anv_gem_close(device
, bo
->bo
.gem_handle
);
1698 vk_free(&device
->alloc
, bo
);
1699 return vk_errorf(device
->instance
, NULL
,
1700 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1701 "failed to allocate virtual address for BO");
1704 assert(bo
->bo
.gem_handle
);
1706 pthread_mutex_lock(&cache
->mutex
);
1708 _mesa_hash_table_insert(cache
->bo_map
,
1709 (void *)(uintptr_t)bo
->bo
.gem_handle
, bo
);
1711 pthread_mutex_unlock(&cache
->mutex
);
1719 anv_bo_cache_import_host_ptr(struct anv_device
*device
,
1720 struct anv_bo_cache
*cache
,
1721 void *host_ptr
, uint32_t size
,
1722 uint64_t bo_flags
, struct anv_bo
**bo_out
)
1724 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1725 assert((bo_flags
& ANV_BO_EXTERNAL
) == 0);
1727 uint32_t gem_handle
= anv_gem_userptr(device
, host_ptr
, size
);
1729 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1731 pthread_mutex_lock(&cache
->mutex
);
1733 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1735 /* VK_EXT_external_memory_host doesn't require handling importing the
1736 * same pointer twice at the same time, but we don't get in the way. If
1737 * kernel gives us the same gem_handle, only succeed if the flags match.
1739 if (bo_flags
!= bo
->bo
.flags
) {
1740 pthread_mutex_unlock(&cache
->mutex
);
1741 return vk_errorf(device
->instance
, NULL
,
1742 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1743 "same host pointer imported two different ways");
1745 __sync_fetch_and_add(&bo
->refcount
, 1);
1747 bo
= vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1748 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1750 anv_gem_close(device
, gem_handle
);
1751 pthread_mutex_unlock(&cache
->mutex
);
1752 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1757 anv_bo_init(&bo
->bo
, gem_handle
, size
);
1758 bo
->bo
.flags
= bo_flags
;
1760 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1761 anv_gem_close(device
, bo
->bo
.gem_handle
);
1762 pthread_mutex_unlock(&cache
->mutex
);
1763 vk_free(&device
->alloc
, bo
);
1764 return vk_errorf(device
->instance
, NULL
,
1765 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1766 "failed to allocate virtual address for BO");
1769 _mesa_hash_table_insert(cache
->bo_map
, (void *)(uintptr_t)gem_handle
, bo
);
1772 pthread_mutex_unlock(&cache
->mutex
);
1779 anv_bo_cache_import(struct anv_device
*device
,
1780 struct anv_bo_cache
*cache
,
1781 int fd
, uint64_t bo_flags
,
1782 struct anv_bo
**bo_out
)
1784 assert(bo_flags
== (bo_flags
& ANV_BO_CACHE_SUPPORTED_FLAGS
));
1785 assert(bo_flags
& ANV_BO_EXTERNAL
);
1787 pthread_mutex_lock(&cache
->mutex
);
1789 uint32_t gem_handle
= anv_gem_fd_to_handle(device
, fd
);
1791 pthread_mutex_unlock(&cache
->mutex
);
1792 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1795 struct anv_cached_bo
*bo
= anv_bo_cache_lookup_locked(cache
, gem_handle
);
1797 /* We have to be careful how we combine flags so that it makes sense.
1798 * Really, though, if we get to this case and it actually matters, the
1799 * client has imported a BO twice in different ways and they get what
1802 uint64_t new_flags
= ANV_BO_EXTERNAL
;
1803 new_flags
|= (bo
->bo
.flags
| bo_flags
) & EXEC_OBJECT_WRITE
;
1804 new_flags
|= (bo
->bo
.flags
& bo_flags
) & EXEC_OBJECT_ASYNC
;
1805 new_flags
|= (bo
->bo
.flags
& bo_flags
) & EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1806 new_flags
|= (bo
->bo
.flags
| bo_flags
) & EXEC_OBJECT_PINNED
;
1808 /* It's theoretically possible for a BO to get imported such that it's
1809 * both pinned and not pinned. The only way this can happen is if it
1810 * gets imported as both a semaphore and a memory object and that would
1811 * be an application error. Just fail out in that case.
1813 if ((bo
->bo
.flags
& EXEC_OBJECT_PINNED
) !=
1814 (bo_flags
& EXEC_OBJECT_PINNED
)) {
1815 pthread_mutex_unlock(&cache
->mutex
);
1816 return vk_errorf(device
->instance
, NULL
,
1817 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1818 "The same BO was imported two different ways");
1821 /* It's also theoretically possible that someone could export a BO from
1822 * one heap and import it into another or to import the same BO into two
1823 * different heaps. If this happens, we could potentially end up both
1824 * allowing and disallowing 48-bit addresses. There's not much we can
1825 * do about it if we're pinning so we just throw an error and hope no
1826 * app is actually that stupid.
1828 if ((new_flags
& EXEC_OBJECT_PINNED
) &&
1829 (bo
->bo
.flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) !=
1830 (bo_flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
)) {
1831 pthread_mutex_unlock(&cache
->mutex
);
1832 return vk_errorf(device
->instance
, NULL
,
1833 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
1834 "The same BO was imported on two different heaps");
1837 bo
->bo
.flags
= new_flags
;
1839 __sync_fetch_and_add(&bo
->refcount
, 1);
1841 off_t size
= lseek(fd
, 0, SEEK_END
);
1842 if (size
== (off_t
)-1) {
1843 anv_gem_close(device
, gem_handle
);
1844 pthread_mutex_unlock(&cache
->mutex
);
1845 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
1848 bo
= vk_alloc(&device
->alloc
, sizeof(struct anv_cached_bo
), 8,
1849 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1851 anv_gem_close(device
, gem_handle
);
1852 pthread_mutex_unlock(&cache
->mutex
);
1853 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1858 anv_bo_init(&bo
->bo
, gem_handle
, size
);
1859 bo
->bo
.flags
= bo_flags
;
1861 if (!anv_vma_alloc(device
, &bo
->bo
)) {
1862 anv_gem_close(device
, bo
->bo
.gem_handle
);
1863 pthread_mutex_unlock(&cache
->mutex
);
1864 vk_free(&device
->alloc
, bo
);
1865 return vk_errorf(device
->instance
, NULL
,
1866 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1867 "failed to allocate virtual address for BO");
1870 _mesa_hash_table_insert(cache
->bo_map
, (void *)(uintptr_t)gem_handle
, bo
);
1873 pthread_mutex_unlock(&cache
->mutex
);
1880 anv_bo_cache_export(struct anv_device
*device
,
1881 struct anv_bo_cache
*cache
,
1882 struct anv_bo
*bo_in
, int *fd_out
)
1884 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1885 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1887 /* This BO must have been flagged external in order for us to be able
1888 * to export it. This is done based on external options passed into
1889 * anv_AllocateMemory.
1891 assert(bo
->bo
.flags
& ANV_BO_EXTERNAL
);
1893 int fd
= anv_gem_handle_to_fd(device
, bo
->bo
.gem_handle
);
1895 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1903 atomic_dec_not_one(uint32_t *counter
)
1912 old
= __sync_val_compare_and_swap(counter
, val
, val
- 1);
1921 anv_bo_cache_release(struct anv_device
*device
,
1922 struct anv_bo_cache
*cache
,
1923 struct anv_bo
*bo_in
)
1925 assert(anv_bo_cache_lookup(cache
, bo_in
->gem_handle
) == bo_in
);
1926 struct anv_cached_bo
*bo
= (struct anv_cached_bo
*)bo_in
;
1928 /* Try to decrement the counter but don't go below one. If this succeeds
1929 * then the refcount has been decremented and we are not the last
1932 if (atomic_dec_not_one(&bo
->refcount
))
1935 pthread_mutex_lock(&cache
->mutex
);
1937 /* We are probably the last reference since our attempt to decrement above
1938 * failed. However, we can't actually know until we are inside the mutex.
1939 * Otherwise, someone could import the BO between the decrement and our
1942 if (unlikely(__sync_sub_and_fetch(&bo
->refcount
, 1) > 0)) {
1943 /* Turns out we're not the last reference. Unlock and bail. */
1944 pthread_mutex_unlock(&cache
->mutex
);
1948 struct hash_entry
*entry
=
1949 _mesa_hash_table_search(cache
->bo_map
,
1950 (const void *)(uintptr_t)bo
->bo
.gem_handle
);
1952 _mesa_hash_table_remove(cache
->bo_map
, entry
);
1955 anv_gem_munmap(bo
->bo
.map
, bo
->bo
.size
);
1957 anv_vma_free(device
, &bo
->bo
);
1959 anv_gem_close(device
, bo
->bo
.gem_handle
);
1961 /* Don't unlock until we've actually closed the BO. The whole point of
1962 * the BO cache is to ensure that we correctly handle races with creating
1963 * and releasing GEM handles and we don't want to let someone import the BO
1964 * again between mutex unlock and closing the GEM handle.
1966 pthread_mutex_unlock(&cache
->mutex
);
1968 vk_free(&device
->alloc
, bo
);