11875cb639962e9bb177eeacf38e1684601dff02
[mesa.git] / src / intel / vulkan / anv_allocator.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 #include <stdint.h>
25 #include <stdlib.h>
26 #include <unistd.h>
27 #include <limits.h>
28 #include <assert.h>
29 #include <linux/futex.h>
30 #include <linux/memfd.h>
31 #include <sys/time.h>
32 #include <sys/mman.h>
33 #include <sys/syscall.h>
34
35 #include "anv_private.h"
36
37 #ifdef HAVE_VALGRIND
38 #define VG_NOACCESS_READ(__ptr) ({ \
39 VALGRIND_MAKE_MEM_DEFINED((__ptr), sizeof(*(__ptr))); \
40 __typeof(*(__ptr)) __val = *(__ptr); \
41 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr)));\
42 __val; \
43 })
44 #define VG_NOACCESS_WRITE(__ptr, __val) ({ \
45 VALGRIND_MAKE_MEM_UNDEFINED((__ptr), sizeof(*(__ptr))); \
46 *(__ptr) = (__val); \
47 VALGRIND_MAKE_MEM_NOACCESS((__ptr), sizeof(*(__ptr))); \
48 })
49 #else
50 #define VG_NOACCESS_READ(__ptr) (*(__ptr))
51 #define VG_NOACCESS_WRITE(__ptr, __val) (*(__ptr) = (__val))
52 #endif
53
54 /* Design goals:
55 *
56 * - Lock free (except when resizing underlying bos)
57 *
58 * - Constant time allocation with typically only one atomic
59 *
60 * - Multiple allocation sizes without fragmentation
61 *
62 * - Can grow while keeping addresses and offset of contents stable
63 *
64 * - All allocations within one bo so we can point one of the
65 * STATE_BASE_ADDRESS pointers at it.
66 *
67 * The overall design is a two-level allocator: top level is a fixed size, big
68 * block (8k) allocator, which operates out of a bo. Allocation is done by
69 * either pulling a block from the free list or growing the used range of the
70 * bo. Growing the range may run out of space in the bo which we then need to
71 * grow. Growing the bo is tricky in a multi-threaded, lockless environment:
72 * we need to keep all pointers and contents in the old map valid. GEM bos in
73 * general can't grow, but we use a trick: we create a memfd and use ftruncate
74 * to grow it as necessary. We mmap the new size and then create a gem bo for
75 * it using the new gem userptr ioctl. Without heavy-handed locking around
76 * our allocation fast-path, there isn't really a way to munmap the old mmap,
77 * so we just keep it around until garbage collection time. While the block
78 * allocator is lockless for normal operations, we block other threads trying
79 * to allocate while we're growing the map. It sholdn't happen often, and
80 * growing is fast anyway.
81 *
82 * At the next level we can use various sub-allocators. The state pool is a
83 * pool of smaller, fixed size objects, which operates much like the block
84 * pool. It uses a free list for freeing objects, but when it runs out of
85 * space it just allocates a new block from the block pool. This allocator is
86 * intended for longer lived state objects such as SURFACE_STATE and most
87 * other persistent state objects in the API. We may need to track more info
88 * with these object and a pointer back to the CPU object (eg VkImage). In
89 * those cases we just allocate a slightly bigger object and put the extra
90 * state after the GPU state object.
91 *
92 * The state stream allocator works similar to how the i965 DRI driver streams
93 * all its state. Even with Vulkan, we need to emit transient state (whether
94 * surface state base or dynamic state base), and for that we can just get a
95 * block and fill it up. These cases are local to a command buffer and the
96 * sub-allocator need not be thread safe. The streaming allocator gets a new
97 * block when it runs out of space and chains them together so they can be
98 * easily freed.
99 */
100
101 /* Allocations are always at least 64 byte aligned, so 1 is an invalid value.
102 * We use it to indicate the free list is empty. */
103 #define EMPTY 1
104
105 struct anv_mmap_cleanup {
106 void *map;
107 size_t size;
108 uint32_t gem_handle;
109 };
110
111 #define ANV_MMAP_CLEANUP_INIT ((struct anv_mmap_cleanup){0})
112
113 static inline long
114 sys_futex(void *addr1, int op, int val1,
115 struct timespec *timeout, void *addr2, int val3)
116 {
117 return syscall(SYS_futex, addr1, op, val1, timeout, addr2, val3);
118 }
119
120 static inline int
121 futex_wake(uint32_t *addr, int count)
122 {
123 return sys_futex(addr, FUTEX_WAKE, count, NULL, NULL, 0);
124 }
125
126 static inline int
127 futex_wait(uint32_t *addr, int32_t value)
128 {
129 return sys_futex(addr, FUTEX_WAIT, value, NULL, NULL, 0);
130 }
131
132 static inline int
133 memfd_create(const char *name, unsigned int flags)
134 {
135 return syscall(SYS_memfd_create, name, flags);
136 }
137
138 static inline uint32_t
139 ilog2_round_up(uint32_t value)
140 {
141 assert(value != 0);
142 return 32 - __builtin_clz(value - 1);
143 }
144
145 static inline uint32_t
146 round_to_power_of_two(uint32_t value)
147 {
148 return 1 << ilog2_round_up(value);
149 }
150
151 static bool
152 anv_free_list_pop(union anv_free_list *list, void **map, int32_t *offset)
153 {
154 union anv_free_list current, new, old;
155
156 current.u64 = list->u64;
157 while (current.offset != EMPTY) {
158 /* We have to add a memory barrier here so that the list head (and
159 * offset) gets read before we read the map pointer. This way we
160 * know that the map pointer is valid for the given offset at the
161 * point where we read it.
162 */
163 __sync_synchronize();
164
165 int32_t *next_ptr = *map + current.offset;
166 new.offset = VG_NOACCESS_READ(next_ptr);
167 new.count = current.count + 1;
168 old.u64 = __sync_val_compare_and_swap(&list->u64, current.u64, new.u64);
169 if (old.u64 == current.u64) {
170 *offset = current.offset;
171 return true;
172 }
173 current = old;
174 }
175
176 return false;
177 }
178
179 static void
180 anv_free_list_push(union anv_free_list *list, void *map, int32_t offset)
181 {
182 union anv_free_list current, old, new;
183 int32_t *next_ptr = map + offset;
184
185 old = *list;
186 do {
187 current = old;
188 VG_NOACCESS_WRITE(next_ptr, current.offset);
189 new.offset = offset;
190 new.count = current.count + 1;
191 old.u64 = __sync_val_compare_and_swap(&list->u64, current.u64, new.u64);
192 } while (old.u64 != current.u64);
193 }
194
195 /* All pointers in the ptr_free_list are assumed to be page-aligned. This
196 * means that the bottom 12 bits should all be zero.
197 */
198 #define PFL_COUNT(x) ((uintptr_t)(x) & 0xfff)
199 #define PFL_PTR(x) ((void *)((uintptr_t)(x) & ~(uintptr_t)0xfff))
200 #define PFL_PACK(ptr, count) ({ \
201 (void *)(((uintptr_t)(ptr) & ~(uintptr_t)0xfff) | ((count) & 0xfff)); \
202 })
203
204 static bool
205 anv_ptr_free_list_pop(void **list, void **elem)
206 {
207 void *current = *list;
208 while (PFL_PTR(current) != NULL) {
209 void **next_ptr = PFL_PTR(current);
210 void *new_ptr = VG_NOACCESS_READ(next_ptr);
211 unsigned new_count = PFL_COUNT(current) + 1;
212 void *new = PFL_PACK(new_ptr, new_count);
213 void *old = __sync_val_compare_and_swap(list, current, new);
214 if (old == current) {
215 *elem = PFL_PTR(current);
216 return true;
217 }
218 current = old;
219 }
220
221 return false;
222 }
223
224 static void
225 anv_ptr_free_list_push(void **list, void *elem)
226 {
227 void *old, *current;
228 void **next_ptr = elem;
229
230 /* The pointer-based free list requires that the pointer be
231 * page-aligned. This is because we use the bottom 12 bits of the
232 * pointer to store a counter to solve the ABA concurrency problem.
233 */
234 assert(((uintptr_t)elem & 0xfff) == 0);
235
236 old = *list;
237 do {
238 current = old;
239 VG_NOACCESS_WRITE(next_ptr, PFL_PTR(current));
240 unsigned new_count = PFL_COUNT(current) + 1;
241 void *new = PFL_PACK(elem, new_count);
242 old = __sync_val_compare_and_swap(list, current, new);
243 } while (old != current);
244 }
245
246 static uint32_t
247 anv_block_pool_grow(struct anv_block_pool *pool, struct anv_block_state *state);
248
249 void
250 anv_block_pool_init(struct anv_block_pool *pool,
251 struct anv_device *device, uint32_t block_size)
252 {
253 assert(util_is_power_of_two(block_size));
254
255 pool->device = device;
256 anv_bo_init(&pool->bo, 0, 0);
257 pool->block_size = block_size;
258 pool->free_list = ANV_FREE_LIST_EMPTY;
259 pool->back_free_list = ANV_FREE_LIST_EMPTY;
260
261 pool->fd = memfd_create("block pool", MFD_CLOEXEC);
262 if (pool->fd == -1)
263 return;
264
265 /* Just make it 2GB up-front. The Linux kernel won't actually back it
266 * with pages until we either map and fault on one of them or we use
267 * userptr and send a chunk of it off to the GPU.
268 */
269 if (ftruncate(pool->fd, BLOCK_POOL_MEMFD_SIZE) == -1)
270 return;
271
272 u_vector_init(&pool->mmap_cleanups,
273 round_to_power_of_two(sizeof(struct anv_mmap_cleanup)), 128);
274
275 pool->state.next = 0;
276 pool->state.end = 0;
277 pool->back_state.next = 0;
278 pool->back_state.end = 0;
279
280 /* Immediately grow the pool so we'll have a backing bo. */
281 pool->state.end = anv_block_pool_grow(pool, &pool->state);
282 }
283
284 void
285 anv_block_pool_finish(struct anv_block_pool *pool)
286 {
287 struct anv_mmap_cleanup *cleanup;
288
289 u_vector_foreach(cleanup, &pool->mmap_cleanups) {
290 if (cleanup->map)
291 munmap(cleanup->map, cleanup->size);
292 if (cleanup->gem_handle)
293 anv_gem_close(pool->device, cleanup->gem_handle);
294 }
295
296 u_vector_finish(&pool->mmap_cleanups);
297
298 close(pool->fd);
299 }
300
301 #define PAGE_SIZE 4096
302
303 /** Grows and re-centers the block pool.
304 *
305 * We grow the block pool in one or both directions in such a way that the
306 * following conditions are met:
307 *
308 * 1) The size of the entire pool is always a power of two.
309 *
310 * 2) The pool only grows on both ends. Neither end can get
311 * shortened.
312 *
313 * 3) At the end of the allocation, we have about twice as much space
314 * allocated for each end as we have used. This way the pool doesn't
315 * grow too far in one direction or the other.
316 *
317 * 4) If the _alloc_back() has never been called, then the back portion of
318 * the pool retains a size of zero. (This makes it easier for users of
319 * the block pool that only want a one-sided pool.)
320 *
321 * 5) We have enough space allocated for at least one more block in
322 * whichever side `state` points to.
323 *
324 * 6) The center of the pool is always aligned to both the block_size of
325 * the pool and a 4K CPU page.
326 */
327 static uint32_t
328 anv_block_pool_grow(struct anv_block_pool *pool, struct anv_block_state *state)
329 {
330 size_t size;
331 void *map;
332 uint32_t gem_handle;
333 struct anv_mmap_cleanup *cleanup;
334
335 pthread_mutex_lock(&pool->device->mutex);
336
337 assert(state == &pool->state || state == &pool->back_state);
338
339 /* Gather a little usage information on the pool. Since we may have
340 * threadsd waiting in queue to get some storage while we resize, it's
341 * actually possible that total_used will be larger than old_size. In
342 * particular, block_pool_alloc() increments state->next prior to
343 * calling block_pool_grow, so this ensures that we get enough space for
344 * which ever side tries to grow the pool.
345 *
346 * We align to a page size because it makes it easier to do our
347 * calculations later in such a way that we state page-aigned.
348 */
349 uint32_t back_used = align_u32(pool->back_state.next, PAGE_SIZE);
350 uint32_t front_used = align_u32(pool->state.next, PAGE_SIZE);
351 uint32_t total_used = front_used + back_used;
352
353 assert(state == &pool->state || back_used > 0);
354
355 size_t old_size = pool->bo.size;
356
357 if (old_size != 0 &&
358 back_used * 2 <= pool->center_bo_offset &&
359 front_used * 2 <= (old_size - pool->center_bo_offset)) {
360 /* If we're in this case then this isn't the firsta allocation and we
361 * already have enough space on both sides to hold double what we
362 * have allocated. There's nothing for us to do.
363 */
364 goto done;
365 }
366
367 if (old_size == 0) {
368 /* This is the first allocation */
369 size = MAX2(32 * pool->block_size, PAGE_SIZE);
370 } else {
371 size = old_size * 2;
372 }
373
374 /* We can't have a block pool bigger than 1GB because we use signed
375 * 32-bit offsets in the free list and we don't want overflow. We
376 * should never need a block pool bigger than 1GB anyway.
377 */
378 assert(size <= (1u << 31));
379
380 /* We compute a new center_bo_offset such that, when we double the size
381 * of the pool, we maintain the ratio of how much is used by each side.
382 * This way things should remain more-or-less balanced.
383 */
384 uint32_t center_bo_offset;
385 if (back_used == 0) {
386 /* If we're in this case then we have never called alloc_back(). In
387 * this case, we want keep the offset at 0 to make things as simple
388 * as possible for users that don't care about back allocations.
389 */
390 center_bo_offset = 0;
391 } else {
392 /* Try to "center" the allocation based on how much is currently in
393 * use on each side of the center line.
394 */
395 center_bo_offset = ((uint64_t)size * back_used) / total_used;
396
397 /* Align down to a multiple of both the block size and page size */
398 uint32_t granularity = MAX2(pool->block_size, PAGE_SIZE);
399 assert(util_is_power_of_two(granularity));
400 center_bo_offset &= ~(granularity - 1);
401
402 assert(center_bo_offset >= back_used);
403
404 /* Make sure we don't shrink the back end of the pool */
405 if (center_bo_offset < pool->back_state.end)
406 center_bo_offset = pool->back_state.end;
407
408 /* Make sure that we don't shrink the front end of the pool */
409 if (size - center_bo_offset < pool->state.end)
410 center_bo_offset = size - pool->state.end;
411 }
412
413 assert(center_bo_offset % pool->block_size == 0);
414 assert(center_bo_offset % PAGE_SIZE == 0);
415
416 /* Assert that we only ever grow the pool */
417 assert(center_bo_offset >= pool->back_state.end);
418 assert(size - center_bo_offset >= pool->state.end);
419
420 cleanup = u_vector_add(&pool->mmap_cleanups);
421 if (!cleanup)
422 goto fail;
423 *cleanup = ANV_MMAP_CLEANUP_INIT;
424
425 /* Just leak the old map until we destroy the pool. We can't munmap it
426 * without races or imposing locking on the block allocate fast path. On
427 * the whole the leaked maps adds up to less than the size of the
428 * current map. MAP_POPULATE seems like the right thing to do, but we
429 * should try to get some numbers.
430 */
431 map = mmap(NULL, size, PROT_READ | PROT_WRITE,
432 MAP_SHARED | MAP_POPULATE, pool->fd,
433 BLOCK_POOL_MEMFD_CENTER - center_bo_offset);
434 cleanup->map = map;
435 cleanup->size = size;
436
437 if (map == MAP_FAILED)
438 goto fail;
439
440 gem_handle = anv_gem_userptr(pool->device, map, size);
441 if (gem_handle == 0)
442 goto fail;
443 cleanup->gem_handle = gem_handle;
444
445 #if 0
446 /* Regular objects are created I915_CACHING_CACHED on LLC platforms and
447 * I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
448 * always created as I915_CACHING_CACHED, which on non-LLC means
449 * snooped. That can be useful but comes with a bit of overheard. Since
450 * we're eplicitly clflushing and don't want the overhead we need to turn
451 * it off. */
452 if (!pool->device->info.has_llc) {
453 anv_gem_set_caching(pool->device, gem_handle, I915_CACHING_NONE);
454 anv_gem_set_domain(pool->device, gem_handle,
455 I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
456 }
457 #endif
458
459 /* Now that we successfull allocated everything, we can write the new
460 * values back into pool. */
461 pool->map = map + center_bo_offset;
462 pool->center_bo_offset = center_bo_offset;
463 anv_bo_init(&pool->bo, gem_handle, size);
464 pool->bo.map = map;
465
466 done:
467 pthread_mutex_unlock(&pool->device->mutex);
468
469 /* Return the appropreate new size. This function never actually
470 * updates state->next. Instead, we let the caller do that because it
471 * needs to do so in order to maintain its concurrency model.
472 */
473 if (state == &pool->state) {
474 return pool->bo.size - pool->center_bo_offset;
475 } else {
476 assert(pool->center_bo_offset > 0);
477 return pool->center_bo_offset;
478 }
479
480 fail:
481 pthread_mutex_unlock(&pool->device->mutex);
482
483 return 0;
484 }
485
486 static uint32_t
487 anv_block_pool_alloc_new(struct anv_block_pool *pool,
488 struct anv_block_state *pool_state)
489 {
490 struct anv_block_state state, old, new;
491
492 while (1) {
493 state.u64 = __sync_fetch_and_add(&pool_state->u64, pool->block_size);
494 if (state.next < state.end) {
495 assert(pool->map);
496 return state.next;
497 } else if (state.next == state.end) {
498 /* We allocated the first block outside the pool, we have to grow it.
499 * pool_state->next acts a mutex: threads who try to allocate now will
500 * get block indexes above the current limit and hit futex_wait
501 * below. */
502 new.next = state.next + pool->block_size;
503 new.end = anv_block_pool_grow(pool, pool_state);
504 assert(new.end >= new.next && new.end % pool->block_size == 0);
505 old.u64 = __sync_lock_test_and_set(&pool_state->u64, new.u64);
506 if (old.next != state.next)
507 futex_wake(&pool_state->end, INT_MAX);
508 return state.next;
509 } else {
510 futex_wait(&pool_state->end, state.end);
511 continue;
512 }
513 }
514 }
515
516 int32_t
517 anv_block_pool_alloc(struct anv_block_pool *pool)
518 {
519 int32_t offset;
520
521 /* Try free list first. */
522 if (anv_free_list_pop(&pool->free_list, &pool->map, &offset)) {
523 assert(offset >= 0);
524 assert(pool->map);
525 return offset;
526 }
527
528 return anv_block_pool_alloc_new(pool, &pool->state);
529 }
530
531 /* Allocates a block out of the back of the block pool.
532 *
533 * This will allocated a block earlier than the "start" of the block pool.
534 * The offsets returned from this function will be negative but will still
535 * be correct relative to the block pool's map pointer.
536 *
537 * If you ever use anv_block_pool_alloc_back, then you will have to do
538 * gymnastics with the block pool's BO when doing relocations.
539 */
540 int32_t
541 anv_block_pool_alloc_back(struct anv_block_pool *pool)
542 {
543 int32_t offset;
544
545 /* Try free list first. */
546 if (anv_free_list_pop(&pool->back_free_list, &pool->map, &offset)) {
547 assert(offset < 0);
548 assert(pool->map);
549 return offset;
550 }
551
552 offset = anv_block_pool_alloc_new(pool, &pool->back_state);
553
554 /* The offset we get out of anv_block_pool_alloc_new() is actually the
555 * number of bytes downwards from the middle to the end of the block.
556 * We need to turn it into a (negative) offset from the middle to the
557 * start of the block.
558 */
559 assert(offset >= 0);
560 return -(offset + pool->block_size);
561 }
562
563 void
564 anv_block_pool_free(struct anv_block_pool *pool, int32_t offset)
565 {
566 if (offset < 0) {
567 anv_free_list_push(&pool->back_free_list, pool->map, offset);
568 } else {
569 anv_free_list_push(&pool->free_list, pool->map, offset);
570 }
571 }
572
573 static void
574 anv_fixed_size_state_pool_init(struct anv_fixed_size_state_pool *pool,
575 size_t state_size)
576 {
577 /* At least a cache line and must divide the block size. */
578 assert(state_size >= 64 && util_is_power_of_two(state_size));
579
580 pool->state_size = state_size;
581 pool->free_list = ANV_FREE_LIST_EMPTY;
582 pool->block.next = 0;
583 pool->block.end = 0;
584 }
585
586 static uint32_t
587 anv_fixed_size_state_pool_alloc(struct anv_fixed_size_state_pool *pool,
588 struct anv_block_pool *block_pool)
589 {
590 int32_t offset;
591 struct anv_block_state block, old, new;
592
593 /* Try free list first. */
594 if (anv_free_list_pop(&pool->free_list, &block_pool->map, &offset)) {
595 assert(offset >= 0);
596 return offset;
597 }
598
599 /* If free list was empty (or somebody raced us and took the items) we
600 * allocate a new item from the end of the block */
601 restart:
602 block.u64 = __sync_fetch_and_add(&pool->block.u64, pool->state_size);
603
604 if (block.next < block.end) {
605 return block.next;
606 } else if (block.next == block.end) {
607 offset = anv_block_pool_alloc(block_pool);
608 new.next = offset + pool->state_size;
609 new.end = offset + block_pool->block_size;
610 old.u64 = __sync_lock_test_and_set(&pool->block.u64, new.u64);
611 if (old.next != block.next)
612 futex_wake(&pool->block.end, INT_MAX);
613 return offset;
614 } else {
615 futex_wait(&pool->block.end, block.end);
616 goto restart;
617 }
618 }
619
620 static void
621 anv_fixed_size_state_pool_free(struct anv_fixed_size_state_pool *pool,
622 struct anv_block_pool *block_pool,
623 uint32_t offset)
624 {
625 anv_free_list_push(&pool->free_list, block_pool->map, offset);
626 }
627
628 void
629 anv_state_pool_init(struct anv_state_pool *pool,
630 struct anv_block_pool *block_pool)
631 {
632 pool->block_pool = block_pool;
633 for (unsigned i = 0; i < ANV_STATE_BUCKETS; i++) {
634 size_t size = 1 << (ANV_MIN_STATE_SIZE_LOG2 + i);
635 anv_fixed_size_state_pool_init(&pool->buckets[i], size);
636 }
637 VG(VALGRIND_CREATE_MEMPOOL(pool, 0, false));
638 }
639
640 void
641 anv_state_pool_finish(struct anv_state_pool *pool)
642 {
643 VG(VALGRIND_DESTROY_MEMPOOL(pool));
644 }
645
646 struct anv_state
647 anv_state_pool_alloc(struct anv_state_pool *pool, size_t size, size_t align)
648 {
649 unsigned size_log2 = ilog2_round_up(size < align ? align : size);
650 assert(size_log2 <= ANV_MAX_STATE_SIZE_LOG2);
651 if (size_log2 < ANV_MIN_STATE_SIZE_LOG2)
652 size_log2 = ANV_MIN_STATE_SIZE_LOG2;
653 unsigned bucket = size_log2 - ANV_MIN_STATE_SIZE_LOG2;
654
655 struct anv_state state;
656 state.alloc_size = 1 << size_log2;
657 state.offset = anv_fixed_size_state_pool_alloc(&pool->buckets[bucket],
658 pool->block_pool);
659 state.map = pool->block_pool->map + state.offset;
660 VG(VALGRIND_MEMPOOL_ALLOC(pool, state.map, size));
661 return state;
662 }
663
664 void
665 anv_state_pool_free(struct anv_state_pool *pool, struct anv_state state)
666 {
667 assert(util_is_power_of_two(state.alloc_size));
668 unsigned size_log2 = ilog2_round_up(state.alloc_size);
669 assert(size_log2 >= ANV_MIN_STATE_SIZE_LOG2 &&
670 size_log2 <= ANV_MAX_STATE_SIZE_LOG2);
671 unsigned bucket = size_log2 - ANV_MIN_STATE_SIZE_LOG2;
672
673 VG(VALGRIND_MEMPOOL_FREE(pool, state.map));
674 anv_fixed_size_state_pool_free(&pool->buckets[bucket],
675 pool->block_pool, state.offset);
676 }
677
678 #define NULL_BLOCK 1
679 struct anv_state_stream_block {
680 /* The next block */
681 struct anv_state_stream_block *next;
682
683 /* The offset into the block pool at which this block starts */
684 uint32_t offset;
685
686 #ifdef HAVE_VALGRIND
687 /* A pointer to the first user-allocated thing in this block. This is
688 * what valgrind sees as the start of the block.
689 */
690 void *_vg_ptr;
691 #endif
692 };
693
694 /* The state stream allocator is a one-shot, single threaded allocator for
695 * variable sized blocks. We use it for allocating dynamic state.
696 */
697 void
698 anv_state_stream_init(struct anv_state_stream *stream,
699 struct anv_block_pool *block_pool)
700 {
701 stream->block_pool = block_pool;
702 stream->block = NULL;
703
704 /* Ensure that next + whatever > end. This way the first call to
705 * state_stream_alloc fetches a new block.
706 */
707 stream->next = 1;
708 stream->end = 0;
709
710 VG(VALGRIND_CREATE_MEMPOOL(stream, 0, false));
711 }
712
713 void
714 anv_state_stream_finish(struct anv_state_stream *stream)
715 {
716 VG(const uint32_t block_size = stream->block_pool->block_size);
717
718 struct anv_state_stream_block *next = stream->block;
719 while (next != NULL) {
720 VG(VALGRIND_MAKE_MEM_DEFINED(next, sizeof(*next)));
721 struct anv_state_stream_block sb = VG_NOACCESS_READ(next);
722 VG(VALGRIND_MEMPOOL_FREE(stream, sb._vg_ptr));
723 VG(VALGRIND_MAKE_MEM_UNDEFINED(next, block_size));
724 anv_block_pool_free(stream->block_pool, sb.offset);
725 next = sb.next;
726 }
727
728 VG(VALGRIND_DESTROY_MEMPOOL(stream));
729 }
730
731 struct anv_state
732 anv_state_stream_alloc(struct anv_state_stream *stream,
733 uint32_t size, uint32_t alignment)
734 {
735 struct anv_state_stream_block *sb = stream->block;
736
737 struct anv_state state;
738
739 state.offset = align_u32(stream->next, alignment);
740 if (state.offset + size > stream->end) {
741 uint32_t block = anv_block_pool_alloc(stream->block_pool);
742 sb = stream->block_pool->map + block;
743
744 VG(VALGRIND_MAKE_MEM_UNDEFINED(sb, sizeof(*sb)));
745 sb->next = stream->block;
746 sb->offset = block;
747 VG(sb->_vg_ptr = NULL);
748 VG(VALGRIND_MAKE_MEM_NOACCESS(sb, stream->block_pool->block_size));
749
750 stream->block = sb;
751 stream->start = block;
752 stream->next = block + sizeof(*sb);
753 stream->end = block + stream->block_pool->block_size;
754
755 state.offset = align_u32(stream->next, alignment);
756 assert(state.offset + size <= stream->end);
757 }
758
759 assert(state.offset > stream->start);
760 state.map = (void *)sb + (state.offset - stream->start);
761 state.alloc_size = size;
762
763 #ifdef HAVE_VALGRIND
764 void *vg_ptr = VG_NOACCESS_READ(&sb->_vg_ptr);
765 if (vg_ptr == NULL) {
766 vg_ptr = state.map;
767 VG_NOACCESS_WRITE(&sb->_vg_ptr, vg_ptr);
768 VALGRIND_MEMPOOL_ALLOC(stream, vg_ptr, size);
769 } else {
770 void *state_end = state.map + state.alloc_size;
771 /* This only updates the mempool. The newly allocated chunk is still
772 * marked as NOACCESS. */
773 VALGRIND_MEMPOOL_CHANGE(stream, vg_ptr, vg_ptr, state_end - vg_ptr);
774 /* Mark the newly allocated chunk as undefined */
775 VALGRIND_MAKE_MEM_UNDEFINED(state.map, state.alloc_size);
776 }
777 #endif
778
779 stream->next = state.offset + size;
780
781 return state;
782 }
783
784 struct bo_pool_bo_link {
785 struct bo_pool_bo_link *next;
786 struct anv_bo bo;
787 };
788
789 void
790 anv_bo_pool_init(struct anv_bo_pool *pool, struct anv_device *device)
791 {
792 pool->device = device;
793 memset(pool->free_list, 0, sizeof(pool->free_list));
794
795 VG(VALGRIND_CREATE_MEMPOOL(pool, 0, false));
796 }
797
798 void
799 anv_bo_pool_finish(struct anv_bo_pool *pool)
800 {
801 for (unsigned i = 0; i < ARRAY_SIZE(pool->free_list); i++) {
802 struct bo_pool_bo_link *link = PFL_PTR(pool->free_list[i]);
803 while (link != NULL) {
804 struct bo_pool_bo_link link_copy = VG_NOACCESS_READ(link);
805
806 anv_gem_munmap(link_copy.bo.map, link_copy.bo.size);
807 anv_gem_close(pool->device, link_copy.bo.gem_handle);
808 link = link_copy.next;
809 }
810 }
811
812 VG(VALGRIND_DESTROY_MEMPOOL(pool));
813 }
814
815 VkResult
816 anv_bo_pool_alloc(struct anv_bo_pool *pool, struct anv_bo *bo, uint32_t size)
817 {
818 VkResult result;
819
820 const unsigned size_log2 = size < 4096 ? 12 : ilog2_round_up(size);
821 const unsigned pow2_size = 1 << size_log2;
822 const unsigned bucket = size_log2 - 12;
823 assert(bucket < ARRAY_SIZE(pool->free_list));
824
825 void *next_free_void;
826 if (anv_ptr_free_list_pop(&pool->free_list[bucket], &next_free_void)) {
827 struct bo_pool_bo_link *next_free = next_free_void;
828 *bo = VG_NOACCESS_READ(&next_free->bo);
829 assert(bo->gem_handle);
830 assert(bo->map == next_free);
831 assert(size <= bo->size);
832
833 VG(VALGRIND_MEMPOOL_ALLOC(pool, bo->map, size));
834
835 return VK_SUCCESS;
836 }
837
838 struct anv_bo new_bo;
839
840 result = anv_bo_init_new(&new_bo, pool->device, pow2_size);
841 if (result != VK_SUCCESS)
842 return result;
843
844 assert(new_bo.size == pow2_size);
845
846 new_bo.map = anv_gem_mmap(pool->device, new_bo.gem_handle, 0, pow2_size, 0);
847 if (new_bo.map == NULL) {
848 anv_gem_close(pool->device, new_bo.gem_handle);
849 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
850 }
851
852 *bo = new_bo;
853
854 VG(VALGRIND_MEMPOOL_ALLOC(pool, bo->map, size));
855
856 return VK_SUCCESS;
857 }
858
859 void
860 anv_bo_pool_free(struct anv_bo_pool *pool, const struct anv_bo *bo_in)
861 {
862 /* Make a copy in case the anv_bo happens to be storred in the BO */
863 struct anv_bo bo = *bo_in;
864
865 VG(VALGRIND_MEMPOOL_FREE(pool, bo.map));
866
867 struct bo_pool_bo_link *link = bo.map;
868 VG_NOACCESS_WRITE(&link->bo, bo);
869
870 assert(util_is_power_of_two(bo.size));
871 const unsigned size_log2 = ilog2_round_up(bo.size);
872 const unsigned bucket = size_log2 - 12;
873 assert(bucket < ARRAY_SIZE(pool->free_list));
874
875 anv_ptr_free_list_push(&pool->free_list[bucket], link);
876 }
877
878 // Scratch pool
879
880 void
881 anv_scratch_pool_init(struct anv_device *device, struct anv_scratch_pool *pool)
882 {
883 memset(pool, 0, sizeof(*pool));
884 }
885
886 void
887 anv_scratch_pool_finish(struct anv_device *device, struct anv_scratch_pool *pool)
888 {
889 for (unsigned s = 0; s < MESA_SHADER_STAGES; s++) {
890 for (unsigned i = 0; i < 16; i++) {
891 struct anv_bo *bo = &pool->bos[i][s];
892 if (bo->size > 0)
893 anv_gem_close(device, bo->gem_handle);
894 }
895 }
896 }
897
898 struct anv_bo *
899 anv_scratch_pool_alloc(struct anv_device *device, struct anv_scratch_pool *pool,
900 gl_shader_stage stage, unsigned per_thread_scratch)
901 {
902 if (per_thread_scratch == 0)
903 return NULL;
904
905 unsigned scratch_size_log2 = ffs(per_thread_scratch / 2048);
906 assert(scratch_size_log2 < 16);
907
908 struct anv_bo *bo = &pool->bos[scratch_size_log2][stage];
909
910 /* From now on, we go into a critical section. In order to remain
911 * thread-safe, we use the bo size as a lock. A value of 0 means we don't
912 * have a valid BO yet. A value of 1 means locked. A value greater than 1
913 * means we have a bo of the given size.
914 */
915
916 if (bo->size > 1)
917 return bo;
918
919 uint64_t size = __sync_val_compare_and_swap(&bo->size, 0, 1);
920 if (size == 0) {
921 /* We own the lock. Allocate a buffer */
922
923 const struct anv_physical_device *physical_device =
924 &device->instance->physicalDevice;
925 const struct gen_device_info *devinfo = &physical_device->info;
926
927 /* WaCSScratchSize:hsw
928 *
929 * Haswell's scratch space address calculation appears to be sparse
930 * rather than tightly packed. The Thread ID has bits indicating which
931 * subslice, EU within a subslice, and thread within an EU it is.
932 * There's a maximum of two slices and two subslices, so these can be
933 * stored with a single bit. Even though there are only 10 EUs per
934 * subslice, this is stored in 4 bits, so there's an effective maximum
935 * value of 16 EUs. Similarly, although there are only 7 threads per EU,
936 * this is stored in a 3 bit number, giving an effective maximum value
937 * of 8 threads per EU.
938 *
939 * This means that we need to use 16 * 8 instead of 10 * 7 for the
940 * number of threads per subslice.
941 */
942 const unsigned subslices = MAX2(physical_device->subslice_total, 1);
943 const unsigned scratch_ids_per_subslice =
944 device->info.is_haswell ? 16 * 8 : devinfo->max_cs_threads;
945
946 uint32_t max_threads[] = {
947 [MESA_SHADER_VERTEX] = devinfo->max_vs_threads,
948 [MESA_SHADER_TESS_CTRL] = devinfo->max_tcs_threads,
949 [MESA_SHADER_TESS_EVAL] = devinfo->max_tes_threads,
950 [MESA_SHADER_GEOMETRY] = devinfo->max_gs_threads,
951 [MESA_SHADER_FRAGMENT] = devinfo->max_wm_threads,
952 [MESA_SHADER_COMPUTE] = scratch_ids_per_subslice * subslices,
953 };
954
955 size = per_thread_scratch * max_threads[stage];
956
957 struct anv_bo new_bo;
958 anv_bo_init_new(&new_bo, device, size);
959
960 bo->gem_handle = new_bo.gem_handle;
961
962 /* Set the size last because we use it as a lock */
963 __sync_synchronize();
964 bo->size = size;
965
966 futex_wake((uint32_t *)&bo->size, INT_MAX);
967 } else {
968 /* Someone else got here first */
969 while (bo->size == 1)
970 futex_wait((uint32_t *)&bo->size, 1);
971 }
972
973 return bo;
974 }