1 /*
2 * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
3 * Copyright © 2015 Advanced Micro Devices, Inc.
4 * Copyright © 2021 Valve Corporation
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining
8 * a copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
16 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
17 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
18 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
19 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 *
24 * The above copyright notice and this permission notice (including the
25 * next paragraph) shall be included in all copies or substantial portions
26 * of the Software.
27 *
28 * Authors:
29 * Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
30 */
31
32 #include "zink_context.h"
33 #include "zink_bo.h"
34 #include "zink_resource.h"
35 #include "zink_screen.h"
36 #include "util/u_hash_table.h"
37
38 #if !defined(__APPLE__) && !defined(_WIN32)
39 #define ZINK_USE_DMABUF
40 #include <xf86drm.h>
41 #endif
42
43 struct zink_bo;
44
45 struct zink_sparse_backing_chunk {
46 uint32_t begin, end;
47 };
48
49
50 /*
51 * Sub-allocation information for a real buffer used as backing memory of a
52 * sparse buffer.
53 */
54 struct zink_sparse_backing {
55 struct list_head list;
56
57 struct zink_bo *bo;
58
59 /* Sorted list of free chunks. */
60 struct zink_sparse_backing_chunk *chunks;
61 uint32_t max_chunks;
62 uint32_t num_chunks;
63 };
64
65 struct zink_sparse_commitment {
66 struct zink_sparse_backing *backing;
67 uint32_t page;
68 };
69
70 struct zink_slab {
71 struct pb_slab base;
72 unsigned entry_size;
73 struct zink_bo *buffer;
74 struct zink_bo *entries;
75 };
76
77
78 ALWAYS_INLINE static struct zink_slab *
zink_slab(struct pb_slab * pslab)79 zink_slab(struct pb_slab *pslab)
80 {
81 return (struct zink_slab*)pslab;
82 }
83
84 static struct pb_slabs *
get_slabs(struct zink_screen * screen,uint64_t size,enum zink_alloc_flag flags)85 get_slabs(struct zink_screen *screen, uint64_t size, enum zink_alloc_flag flags)
86 {
87 //struct pb_slabs *bo_slabs = ((flags & RADEON_FLAG_ENCRYPTED) && screen->info.has_tmz_support) ?
88 //screen->bo_slabs_encrypted : screen->bo_slabs;
89
90 struct pb_slabs *bo_slabs = screen->pb.bo_slabs;
91 /* Find the correct slab allocator for the given size. */
92 for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
93 struct pb_slabs *slabs = &bo_slabs[i];
94
95 if (size <= 1ULL << (slabs->min_order + slabs->num_orders - 1))
96 return slabs;
97 }
98
99 assert(0);
100 return NULL;
101 }
102
103 /* Return the power of two size of a slab entry matching the input size. */
104 static unsigned
get_slab_pot_entry_size(struct zink_screen * screen,unsigned size)105 get_slab_pot_entry_size(struct zink_screen *screen, unsigned size)
106 {
107 unsigned entry_size = util_next_power_of_two(size);
108 unsigned min_entry_size = 1 << screen->pb.bo_slabs[0].min_order;
109
110 return MAX2(entry_size, min_entry_size);
111 }
112
113 /* Return the slab entry alignment. */
get_slab_entry_alignment(struct zink_screen * screen,unsigned size)114 static unsigned get_slab_entry_alignment(struct zink_screen *screen, unsigned size)
115 {
116 unsigned entry_size = get_slab_pot_entry_size(screen, size);
117
118 if (size <= entry_size * 3 / 4)
119 return entry_size / 4;
120
121 return entry_size;
122 }
123
124 static void
bo_destroy(struct zink_screen * screen,struct pb_buffer * pbuf)125 bo_destroy(struct zink_screen *screen, struct pb_buffer *pbuf)
126 {
127 struct zink_bo *bo = zink_bo(pbuf);
128
129 #ifdef ZINK_USE_DMABUF
130 if (bo->mem && !bo->u.real.use_reusable_pool) {
131 simple_mtx_lock(&bo->u.real.export_lock);
132 list_for_each_entry_safe(struct bo_export, export, &bo->u.real.exports, link) {
133 struct drm_gem_close args = { .handle = export->gem_handle };
134 drmIoctl(export->drm_fd, DRM_IOCTL_GEM_CLOSE, &args);
135 list_del(&export->link);
136 free(export);
137 }
138 simple_mtx_unlock(&bo->u.real.export_lock);
139 simple_mtx_destroy(&bo->u.real.export_lock);
140 }
141 #endif
142
143 if (!bo->u.real.is_user_ptr && bo->u.real.cpu_ptr) {
144 bo->u.real.map_count = 1;
145 bo->u.real.cpu_ptr = NULL;
146 zink_bo_unmap(screen, bo);
147 }
148
149 VKSCR(FreeMemory)(screen->dev, bo->mem, NULL);
150
151 simple_mtx_destroy(&bo->lock);
152 FREE(bo);
153 }
154
155 static bool
bo_can_reclaim(struct zink_screen * screen,struct pb_buffer * pbuf)156 bo_can_reclaim(struct zink_screen *screen, struct pb_buffer *pbuf)
157 {
158 struct zink_bo *bo = zink_bo(pbuf);
159
160 return zink_screen_usage_check_completion(screen, bo->reads) && zink_screen_usage_check_completion(screen, bo->writes);
161 }
162
163 static bool
bo_can_reclaim_slab(void * priv,struct pb_slab_entry * entry)164 bo_can_reclaim_slab(void *priv, struct pb_slab_entry *entry)
165 {
166 struct zink_bo *bo = container_of(entry, struct zink_bo, u.slab.entry);
167
168 return bo_can_reclaim(priv, &bo->base);
169 }
170
171 static void
bo_slab_free(struct zink_screen * screen,struct pb_slab * pslab)172 bo_slab_free(struct zink_screen *screen, struct pb_slab *pslab)
173 {
174 struct zink_slab *slab = zink_slab(pslab);
175 ASSERTED unsigned slab_size = slab->buffer->base.size;
176
177 assert(slab->base.num_entries * slab->entry_size <= slab_size);
178 FREE(slab->entries);
179 zink_bo_unref(screen, slab->buffer);
180 FREE(slab);
181 }
182
183 static void
bo_slab_destroy(struct zink_screen * screen,struct pb_buffer * pbuf)184 bo_slab_destroy(struct zink_screen *screen, struct pb_buffer *pbuf)
185 {
186 struct zink_bo *bo = zink_bo(pbuf);
187
188 assert(!bo->mem);
189
190 //if (bo->base.usage & RADEON_FLAG_ENCRYPTED)
191 //pb_slab_free(get_slabs(screen, bo->base.size, RADEON_FLAG_ENCRYPTED), &bo->u.slab.entry);
192 //else
193 pb_slab_free(get_slabs(screen, bo->base.size, 0), &bo->u.slab.entry);
194 }
195
196 static void
clean_up_buffer_managers(struct zink_screen * screen)197 clean_up_buffer_managers(struct zink_screen *screen)
198 {
199 for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
200 pb_slabs_reclaim(&screen->pb.bo_slabs[i]);
201 //if (screen->info.has_tmz_support)
202 //pb_slabs_reclaim(&screen->bo_slabs_encrypted[i]);
203 }
204
205 pb_cache_release_all_buffers(&screen->pb.bo_cache);
206 }
207
208 static unsigned
get_optimal_alignment(struct zink_screen * screen,uint64_t size,unsigned alignment)209 get_optimal_alignment(struct zink_screen *screen, uint64_t size, unsigned alignment)
210 {
211 /* Increase the alignment for faster address translation and better memory
212 * access pattern.
213 */
214 if (size >= 4096) {
215 alignment = MAX2(alignment, 4096);
216 } else if (size) {
217 unsigned msb = util_last_bit(size);
218
219 alignment = MAX2(alignment, 1u << (msb - 1));
220 }
221 return alignment;
222 }
223
224 static void
bo_destroy_or_cache(struct zink_screen * screen,struct pb_buffer * pbuf)225 bo_destroy_or_cache(struct zink_screen *screen, struct pb_buffer *pbuf)
226 {
227 struct zink_bo *bo = zink_bo(pbuf);
228
229 assert(bo->mem); /* slab buffers have a separate vtbl */
230 bo->reads = NULL;
231 bo->writes = NULL;
232
233 if (bo->u.real.use_reusable_pool)
234 pb_cache_add_buffer(bo->cache_entry);
235 else
236 bo_destroy(screen, pbuf);
237 }
238
239 static const struct pb_vtbl bo_vtbl = {
240 /* Cast to void* because one of the function parameters is a struct pointer instead of void*. */
241 (void*)bo_destroy_or_cache
242 /* other functions are never called */
243 };
244
245 static struct zink_bo *
bo_create_internal(struct zink_screen * screen,uint64_t size,unsigned alignment,enum zink_heap heap,unsigned flags,const void * pNext)246 bo_create_internal(struct zink_screen *screen,
247 uint64_t size,
248 unsigned alignment,
249 enum zink_heap heap,
250 unsigned flags,
251 const void *pNext)
252 {
253 struct zink_bo *bo = NULL;
254 bool init_pb_cache;
255
256 /* too big for vk alloc */
257 if (size > UINT32_MAX)
258 return NULL;
259
260 alignment = get_optimal_alignment(screen, size, alignment);
261
262 VkMemoryAllocateInfo mai;
263 mai.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
264 mai.pNext = pNext;
265 mai.allocationSize = size;
266 mai.memoryTypeIndex = screen->heap_map[heap];
267 if (screen->info.mem_props.memoryTypes[mai.memoryTypeIndex].propertyFlags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
268 alignment = MAX2(alignment, screen->info.props.limits.minMemoryMapAlignment);
269 mai.allocationSize = align64(mai.allocationSize, screen->info.props.limits.minMemoryMapAlignment);
270 }
271 unsigned heap_idx = screen->info.mem_props.memoryTypes[screen->heap_map[heap]].heapIndex;
272 if (mai.allocationSize > screen->info.mem_props.memoryHeaps[heap_idx].size) {
273 mesa_loge("zink: can't allocate %"PRIu64" bytes from heap that's only %"PRIu64" bytes!\n", mai.allocationSize, screen->info.mem_props.memoryHeaps[heap_idx].size);
274 return NULL;
275 }
276
277 /* all non-suballocated bo can cache */
278 init_pb_cache = !pNext;
279
280 if (!bo)
281 bo = CALLOC(1, sizeof(struct zink_bo) + init_pb_cache * sizeof(struct pb_cache_entry));
282 if (!bo) {
283 return NULL;
284 }
285
286 VkResult ret = VKSCR(AllocateMemory)(screen->dev, &mai, NULL, &bo->mem);
287 if (!zink_screen_handle_vkresult(screen, ret)) {
288 mesa_loge("zink: couldn't allocate memory: heap=%u size=%" PRIu64, heap, size);
289 goto fail;
290 }
291
292 if (init_pb_cache) {
293 bo->u.real.use_reusable_pool = true;
294 pb_cache_init_entry(&screen->pb.bo_cache, bo->cache_entry, &bo->base, heap);
295 } else {
296 #ifdef ZINK_USE_DMABUF
297 list_inithead(&bo->u.real.exports);
298 simple_mtx_init(&bo->u.real.export_lock, mtx_plain);
299 #endif
300 }
301
302
303 simple_mtx_init(&bo->lock, mtx_plain);
304 pipe_reference_init(&bo->base.reference, 1);
305 bo->base.alignment_log2 = util_logbase2(alignment);
306 bo->base.size = mai.allocationSize;
307 bo->base.vtbl = &bo_vtbl;
308 bo->base.placement = screen->heap_flags[heap];
309 bo->base.usage = flags;
310 bo->unique_id = p_atomic_inc_return(&screen->pb.next_bo_unique_id);
311
312 return bo;
313
314 fail:
315 bo_destroy(screen, (void*)bo);
316 return NULL;
317 }
318
319 /*
320 * Attempt to allocate the given number of backing pages. Fewer pages may be
321 * allocated (depending on the fragmentation of existing backing buffers),
322 * which will be reflected by a change to *pnum_pages.
323 */
324 static struct zink_sparse_backing *
sparse_backing_alloc(struct zink_screen * screen,struct zink_bo * bo,uint32_t * pstart_page,uint32_t * pnum_pages)325 sparse_backing_alloc(struct zink_screen *screen, struct zink_bo *bo,
326 uint32_t *pstart_page, uint32_t *pnum_pages)
327 {
328 struct zink_sparse_backing *best_backing;
329 unsigned best_idx;
330 uint32_t best_num_pages;
331
332 best_backing = NULL;
333 best_idx = 0;
334 best_num_pages = 0;
335
336 /* This is a very simple and inefficient best-fit algorithm. */
337 list_for_each_entry(struct zink_sparse_backing, backing, &bo->u.sparse.backing, list) {
338 for (unsigned idx = 0; idx < backing->num_chunks; ++idx) {
339 uint32_t cur_num_pages = backing->chunks[idx].end - backing->chunks[idx].begin;
340 if ((best_num_pages < *pnum_pages && cur_num_pages > best_num_pages) ||
341 (best_num_pages > *pnum_pages && cur_num_pages < best_num_pages)) {
342 best_backing = backing;
343 best_idx = idx;
344 best_num_pages = cur_num_pages;
345 }
346 }
347 }
348
349 /* Allocate a new backing buffer if necessary. */
350 if (!best_backing) {
351 struct pb_buffer *buf;
352 uint64_t size;
353 uint32_t pages;
354
355 best_backing = CALLOC_STRUCT(zink_sparse_backing);
356 if (!best_backing)
357 return NULL;
358
359 best_backing->max_chunks = 4;
360 best_backing->chunks = CALLOC(best_backing->max_chunks,
361 sizeof(*best_backing->chunks));
362 if (!best_backing->chunks) {
363 FREE(best_backing);
364 return NULL;
365 }
366
367 assert(bo->u.sparse.num_backing_pages < DIV_ROUND_UP(bo->base.size, ZINK_SPARSE_BUFFER_PAGE_SIZE));
368
369 size = MIN3(bo->base.size / 16,
370 8 * 1024 * 1024,
371 bo->base.size - (uint64_t)bo->u.sparse.num_backing_pages * ZINK_SPARSE_BUFFER_PAGE_SIZE);
372 size = MAX2(size, ZINK_SPARSE_BUFFER_PAGE_SIZE);
373
374 buf = zink_bo_create(screen, size, ZINK_SPARSE_BUFFER_PAGE_SIZE,
375 ZINK_HEAP_DEVICE_LOCAL, 0, NULL);
376 if (!buf) {
377 FREE(best_backing->chunks);
378 FREE(best_backing);
379 return NULL;
380 }
381
382 /* We might have gotten a bigger buffer than requested via caching. */
383 pages = buf->size / ZINK_SPARSE_BUFFER_PAGE_SIZE;
384
385 best_backing->bo = zink_bo(buf);
386 best_backing->num_chunks = 1;
387 best_backing->chunks[0].begin = 0;
388 best_backing->chunks[0].end = pages;
389
390 list_add(&best_backing->list, &bo->u.sparse.backing);
391 bo->u.sparse.num_backing_pages += pages;
392
393 best_idx = 0;
394 best_num_pages = pages;
395 }
396
397 *pnum_pages = MIN2(*pnum_pages, best_num_pages);
398 *pstart_page = best_backing->chunks[best_idx].begin;
399 best_backing->chunks[best_idx].begin += *pnum_pages;
400
401 if (best_backing->chunks[best_idx].begin >= best_backing->chunks[best_idx].end) {
402 memmove(&best_backing->chunks[best_idx], &best_backing->chunks[best_idx + 1],
403 sizeof(*best_backing->chunks) * (best_backing->num_chunks - best_idx - 1));
404 best_backing->num_chunks--;
405 }
406
407 return best_backing;
408 }
409
410 static void
sparse_free_backing_buffer(struct zink_screen * screen,struct zink_bo * bo,struct zink_sparse_backing * backing)411 sparse_free_backing_buffer(struct zink_screen *screen, struct zink_bo *bo,
412 struct zink_sparse_backing *backing)
413 {
414 bo->u.sparse.num_backing_pages -= backing->bo->base.size / ZINK_SPARSE_BUFFER_PAGE_SIZE;
415
416 list_del(&backing->list);
417 zink_bo_unref(screen, backing->bo);
418 FREE(backing->chunks);
419 FREE(backing);
420 }
421
422 /*
423 * Return a range of pages from the given backing buffer back into the
424 * free structure.
425 */
426 static bool
sparse_backing_free(struct zink_screen * screen,struct zink_bo * bo,struct zink_sparse_backing * backing,uint32_t start_page,uint32_t num_pages)427 sparse_backing_free(struct zink_screen *screen, struct zink_bo *bo,
428 struct zink_sparse_backing *backing,
429 uint32_t start_page, uint32_t num_pages)
430 {
431 uint32_t end_page = start_page + num_pages;
432 unsigned low = 0;
433 unsigned high = backing->num_chunks;
434
435 /* Find the first chunk with begin >= start_page. */
436 while (low < high) {
437 unsigned mid = low + (high - low) / 2;
438
439 if (backing->chunks[mid].begin >= start_page)
440 high = mid;
441 else
442 low = mid + 1;
443 }
444
445 assert(low >= backing->num_chunks || end_page <= backing->chunks[low].begin);
446 assert(low == 0 || backing->chunks[low - 1].end <= start_page);
447
448 if (low > 0 && backing->chunks[low - 1].end == start_page) {
449 backing->chunks[low - 1].end = end_page;
450
451 if (low < backing->num_chunks && end_page == backing->chunks[low].begin) {
452 backing->chunks[low - 1].end = backing->chunks[low].end;
453 memmove(&backing->chunks[low], &backing->chunks[low + 1],
454 sizeof(*backing->chunks) * (backing->num_chunks - low - 1));
455 backing->num_chunks--;
456 }
457 } else if (low < backing->num_chunks && end_page == backing->chunks[low].begin) {
458 backing->chunks[low].begin = start_page;
459 } else {
460 if (backing->num_chunks >= backing->max_chunks) {
461 unsigned new_max_chunks = 2 * backing->max_chunks;
462 struct zink_sparse_backing_chunk *new_chunks =
463 REALLOC(backing->chunks,
464 sizeof(*backing->chunks) * backing->max_chunks,
465 sizeof(*backing->chunks) * new_max_chunks);
466 if (!new_chunks)
467 return false;
468
469 backing->max_chunks = new_max_chunks;
470 backing->chunks = new_chunks;
471 }
472
473 memmove(&backing->chunks[low + 1], &backing->chunks[low],
474 sizeof(*backing->chunks) * (backing->num_chunks - low));
475 backing->chunks[low].begin = start_page;
476 backing->chunks[low].end = end_page;
477 backing->num_chunks++;
478 }
479
480 if (backing->num_chunks == 1 && backing->chunks[0].begin == 0 &&
481 backing->chunks[0].end == backing->bo->base.size / ZINK_SPARSE_BUFFER_PAGE_SIZE)
482 sparse_free_backing_buffer(screen, bo, backing);
483
484 return true;
485 }
486
487 static void
bo_sparse_destroy(struct zink_screen * screen,struct pb_buffer * pbuf)488 bo_sparse_destroy(struct zink_screen *screen, struct pb_buffer *pbuf)
489 {
490 struct zink_bo *bo = zink_bo(pbuf);
491
492 assert(!bo->mem && bo->base.usage & ZINK_ALLOC_SPARSE);
493
494 while (!list_is_empty(&bo->u.sparse.backing)) {
495 sparse_free_backing_buffer(screen, bo,
496 container_of(bo->u.sparse.backing.next,
497 struct zink_sparse_backing, list));
498 }
499
500 FREE(bo->u.sparse.commitments);
501 simple_mtx_destroy(&bo->lock);
502 FREE(bo);
503 }
504
505 static const struct pb_vtbl bo_sparse_vtbl = {
506 /* Cast to void* because one of the function parameters is a struct pointer instead of void*. */
507 (void*)bo_sparse_destroy
508 /* other functions are never called */
509 };
510
511 static struct pb_buffer *
bo_sparse_create(struct zink_screen * screen,uint64_t size)512 bo_sparse_create(struct zink_screen *screen, uint64_t size)
513 {
514 struct zink_bo *bo;
515
516 /* We use 32-bit page numbers; refuse to attempt allocating sparse buffers
517 * that exceed this limit. This is not really a restriction: we don't have
518 * that much virtual address space anyway.
519 */
520 if (size > (uint64_t)INT32_MAX * ZINK_SPARSE_BUFFER_PAGE_SIZE)
521 return NULL;
522
523 bo = CALLOC_STRUCT(zink_bo);
524 if (!bo)
525 return NULL;
526
527 simple_mtx_init(&bo->lock, mtx_plain);
528 pipe_reference_init(&bo->base.reference, 1);
529 bo->base.alignment_log2 = util_logbase2(ZINK_SPARSE_BUFFER_PAGE_SIZE);
530 bo->base.size = size;
531 bo->base.vtbl = &bo_sparse_vtbl;
532 bo->base.placement = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
533 bo->unique_id = p_atomic_inc_return(&screen->pb.next_bo_unique_id);
534 bo->base.usage = ZINK_ALLOC_SPARSE;
535
536 bo->u.sparse.num_va_pages = DIV_ROUND_UP(size, ZINK_SPARSE_BUFFER_PAGE_SIZE);
537 bo->u.sparse.commitments = CALLOC(bo->u.sparse.num_va_pages,
538 sizeof(*bo->u.sparse.commitments));
539 if (!bo->u.sparse.commitments)
540 goto error_alloc_commitments;
541
542 list_inithead(&bo->u.sparse.backing);
543
544 return &bo->base;
545
546 error_alloc_commitments:
547 simple_mtx_destroy(&bo->lock);
548 FREE(bo);
549 return NULL;
550 }
551
552 struct pb_buffer *
zink_bo_create(struct zink_screen * screen,uint64_t size,unsigned alignment,enum zink_heap heap,enum zink_alloc_flag flags,const void * pNext)553 zink_bo_create(struct zink_screen *screen, uint64_t size, unsigned alignment, enum zink_heap heap, enum zink_alloc_flag flags, const void *pNext)
554 {
555 struct zink_bo *bo;
556 /* pull in sparse flag */
557 flags |= zink_alloc_flags_from_heap(heap);
558
559 //struct pb_slabs *slabs = ((flags & RADEON_FLAG_ENCRYPTED) && screen->info.has_tmz_support) ?
560 //screen->bo_slabs_encrypted : screen->bo_slabs;
561 struct pb_slabs *slabs = screen->pb.bo_slabs;
562
563 struct pb_slabs *last_slab = &slabs[NUM_SLAB_ALLOCATORS - 1];
564 unsigned max_slab_entry_size = 1 << (last_slab->min_order + last_slab->num_orders - 1);
565
566 /* Sub-allocate small buffers from slabs. */
567 if (!(flags & (ZINK_ALLOC_NO_SUBALLOC | ZINK_ALLOC_SPARSE)) &&
568 size <= max_slab_entry_size) {
569 struct pb_slab_entry *entry;
570
571 if (heap < 0 || heap >= ZINK_HEAP_MAX)
572 goto no_slab;
573
574 unsigned alloc_size = size;
575
576 /* Always use slabs for sizes less than 4 KB because the kernel aligns
577 * everything to 4 KB.
578 */
579 if (size < alignment && alignment <= 4 * 1024)
580 alloc_size = alignment;
581
582 if (alignment > get_slab_entry_alignment(screen, alloc_size)) {
583 /* 3/4 allocations can return too small alignment. Try again with a power of two
584 * allocation size.
585 */
586 unsigned pot_size = get_slab_pot_entry_size(screen, alloc_size);
587
588 if (alignment <= pot_size) {
589 /* This size works but wastes some memory to fulfil the alignment. */
590 alloc_size = pot_size;
591 } else {
592 goto no_slab; /* can't fulfil alignment requirements */
593 }
594 }
595
596 struct pb_slabs *slabs = get_slabs(screen, alloc_size, flags);
597 bool reclaim_all = false;
598 if (heap == ZINK_HEAP_DEVICE_LOCAL_VISIBLE && !screen->resizable_bar) {
599 unsigned low_bound = 128 * 1024 * 1024; //128MB is a very small BAR
600 if (screen->info.driver_props.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY)
601 low_bound *= 2; //nvidia has fat textures or something
602 unsigned heapidx = screen->info.mem_props.memoryTypes[screen->heap_map[heap]].heapIndex;
603 reclaim_all = screen->info.mem_props.memoryHeaps[heapidx].size <= low_bound;
604 }
605 entry = pb_slab_alloc_reclaimed(slabs, alloc_size, heap, reclaim_all);
606 if (!entry) {
607 /* Clean up buffer managers and try again. */
608 clean_up_buffer_managers(screen);
609
610 entry = pb_slab_alloc_reclaimed(slabs, alloc_size, heap, true);
611 }
612 if (!entry)
613 return NULL;
614
615 bo = container_of(entry, struct zink_bo, u.slab.entry);
616 pipe_reference_init(&bo->base.reference, 1);
617 bo->base.size = size;
618 assert(alignment <= 1 << bo->base.alignment_log2);
619
620 return &bo->base;
621 }
622 no_slab:
623
624 if (flags & ZINK_ALLOC_SPARSE) {
625 assert(ZINK_SPARSE_BUFFER_PAGE_SIZE % alignment == 0);
626
627 return bo_sparse_create(screen, size);
628 }
629
630 /* Align size to page size. This is the minimum alignment for normal
631 * BOs. Aligning this here helps the cached bufmgr. Especially small BOs,
632 * like constant/uniform buffers, can benefit from better and more reuse.
633 */
634 if (heap == ZINK_HEAP_DEVICE_LOCAL_VISIBLE) {
635 size = align64(size, screen->info.props.limits.minMemoryMapAlignment);
636 alignment = align(alignment, screen->info.props.limits.minMemoryMapAlignment);
637 }
638
639 bool use_reusable_pool = !(flags & ZINK_ALLOC_NO_SUBALLOC);
640
641 if (use_reusable_pool) {
642 /* Get a buffer from the cache. */
643 bo = (struct zink_bo*)
644 pb_cache_reclaim_buffer(&screen->pb.bo_cache, size, alignment, 0, heap);
645 if (bo)
646 return &bo->base;
647 }
648
649 /* Create a new one. */
650 bo = bo_create_internal(screen, size, alignment, heap, flags, pNext);
651 if (!bo) {
652 /* Clean up buffer managers and try again. */
653 clean_up_buffer_managers(screen);
654
655 bo = bo_create_internal(screen, size, alignment, heap, flags, pNext);
656 if (!bo)
657 return NULL;
658 }
659
660 return &bo->base;
661 }
662
663 void *
zink_bo_map(struct zink_screen * screen,struct zink_bo * bo)664 zink_bo_map(struct zink_screen *screen, struct zink_bo *bo)
665 {
666 void *cpu = NULL;
667 uint64_t offset = 0;
668 struct zink_bo *real;
669
670 if (bo->mem) {
671 real = bo;
672 } else {
673 real = bo->u.slab.real;
674 offset = bo->offset - real->offset;
675 }
676
677 cpu = p_atomic_read(&real->u.real.cpu_ptr);
678 if (!cpu) {
679 simple_mtx_lock(&real->lock);
680 /* Must re-check due to the possibility of a race. Re-check need not
681 * be atomic thanks to the lock. */
682 cpu = real->u.real.cpu_ptr;
683 if (!cpu) {
684 VkResult result = VKSCR(MapMemory)(screen->dev, real->mem, 0, real->base.size, 0, &cpu);
685 if (result != VK_SUCCESS) {
686 mesa_loge("ZINK: vkMapMemory failed (%s)", vk_Result_to_str(result));
687 simple_mtx_unlock(&real->lock);
688 return NULL;
689 }
690 p_atomic_set(&real->u.real.cpu_ptr, cpu);
691 }
692 simple_mtx_unlock(&real->lock);
693 }
694 p_atomic_inc(&real->u.real.map_count);
695
696 return (uint8_t*)cpu + offset;
697 }
698
699 void
zink_bo_unmap(struct zink_screen * screen,struct zink_bo * bo)700 zink_bo_unmap(struct zink_screen *screen, struct zink_bo *bo)
701 {
702 struct zink_bo *real = bo->mem ? bo : bo->u.slab.real;
703
704 assert(real->u.real.map_count != 0 && "too many unmaps");
705
706 if (p_atomic_dec_zero(&real->u.real.map_count)) {
707 p_atomic_set(&real->u.real.cpu_ptr, NULL);
708 VKSCR(UnmapMemory)(screen->dev, real->mem);
709 }
710 }
711
712 static VkSemaphore
get_semaphore(struct zink_screen * screen)713 get_semaphore(struct zink_screen *screen)
714 {
715 VkSemaphoreCreateInfo sci = {
716 VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
717 NULL,
718 0
719 };
720 VkSemaphore sem;
721 VkResult ret = VKSCR(CreateSemaphore)(screen->dev, &sci, NULL, &sem);
722 return ret == VK_SUCCESS ? sem : VK_NULL_HANDLE;
723 }
724
725 static VkSemaphore
buffer_commit_single(struct zink_screen * screen,struct zink_resource * res,struct zink_bo * bo,uint32_t bo_offset,uint32_t offset,uint32_t size,bool commit,VkSemaphore wait)726 buffer_commit_single(struct zink_screen *screen, struct zink_resource *res, struct zink_bo *bo, uint32_t bo_offset, uint32_t offset, uint32_t size, bool commit, VkSemaphore wait)
727 {
728 VkSemaphore sem = get_semaphore(screen);
729 VkBindSparseInfo sparse = {0};
730 sparse.sType = VK_STRUCTURE_TYPE_BIND_SPARSE_INFO;
731 sparse.bufferBindCount = res->obj->storage_buffer ? 2 : 1;
732 sparse.waitSemaphoreCount = !!wait;
733 sparse.pWaitSemaphores = &wait;
734 sparse.signalSemaphoreCount = 1;
735 sparse.pSignalSemaphores = &sem;
736
737 VkSparseBufferMemoryBindInfo sparse_bind[2];
738 sparse_bind[0].buffer = res->obj->buffer;
739 sparse_bind[1].buffer = res->obj->storage_buffer;
740 sparse_bind[0].bindCount = 1;
741 sparse_bind[1].bindCount = 1;
742 sparse.pBufferBinds = sparse_bind;
743
744 VkSparseMemoryBind mem_bind;
745 mem_bind.resourceOffset = offset;
746 mem_bind.size = MIN2(res->base.b.width0 - offset, size);
747 mem_bind.memory = commit ? (bo->mem ? bo->mem : bo->u.slab.real->mem) : VK_NULL_HANDLE;
748 mem_bind.memoryOffset = bo_offset * ZINK_SPARSE_BUFFER_PAGE_SIZE + (commit ? (bo->mem ? 0 : bo->offset) : 0);
749 mem_bind.flags = 0;
750 sparse_bind[0].pBinds = &mem_bind;
751 sparse_bind[1].pBinds = &mem_bind;
752
753 VkResult ret = VKSCR(QueueBindSparse)(screen->queue_sparse, 1, &sparse, VK_NULL_HANDLE);
754 if (zink_screen_handle_vkresult(screen, ret))
755 return sem;
756 VKSCR(DestroySemaphore)(screen->dev, sem, NULL);
757 return VK_NULL_HANDLE;
758 }
759
760 static bool
buffer_bo_commit(struct zink_screen * screen,struct zink_resource * res,uint32_t offset,uint32_t size,bool commit,VkSemaphore * sem)761 buffer_bo_commit(struct zink_screen *screen, struct zink_resource *res, uint32_t offset, uint32_t size, bool commit, VkSemaphore *sem)
762 {
763 bool ok = true;
764 struct zink_bo *bo = res->obj->bo;
765 assert(offset % ZINK_SPARSE_BUFFER_PAGE_SIZE == 0);
766 assert(offset <= bo->base.size);
767 assert(size <= bo->base.size - offset);
768 assert(size % ZINK_SPARSE_BUFFER_PAGE_SIZE == 0 || offset + size == bo->base.size);
769
770 struct zink_sparse_commitment *comm = bo->u.sparse.commitments;
771
772 uint32_t va_page = offset / ZINK_SPARSE_BUFFER_PAGE_SIZE;
773 uint32_t end_va_page = va_page + DIV_ROUND_UP(size, ZINK_SPARSE_BUFFER_PAGE_SIZE);
774 VkSemaphore cur_sem = VK_NULL_HANDLE;
775 if (commit) {
776 while (va_page < end_va_page) {
777 uint32_t span_va_page;
778
779 /* Skip pages that are already committed. */
780 if (comm[va_page].backing) {
781 va_page++;
782 continue;
783 }
784
785 /* Determine length of uncommitted span. */
786 span_va_page = va_page;
787 while (va_page < end_va_page && !comm[va_page].backing)
788 va_page++;
789
790 /* Fill the uncommitted span with chunks of backing memory. */
791 while (span_va_page < va_page) {
792 struct zink_sparse_backing *backing;
793 uint32_t backing_start, backing_size;
794
795 backing_size = va_page - span_va_page;
796 backing = sparse_backing_alloc(screen, bo, &backing_start, &backing_size);
797 if (!backing) {
798 ok = false;
799 goto out;
800 }
801 cur_sem = buffer_commit_single(screen, res, backing->bo, backing_start,
802 (uint64_t)span_va_page * ZINK_SPARSE_BUFFER_PAGE_SIZE,
803 (uint64_t)backing_size * ZINK_SPARSE_BUFFER_PAGE_SIZE, true, cur_sem);
804 if (!cur_sem) {
805 ok = sparse_backing_free(screen, bo, backing, backing_start, backing_size);
806 assert(ok && "sufficient memory should already be allocated");
807
808 ok = false;
809 goto out;
810 }
811
812 while (backing_size) {
813 comm[span_va_page].backing = backing;
814 comm[span_va_page].page = backing_start;
815 span_va_page++;
816 backing_start++;
817 backing_size--;
818 }
819 }
820 }
821 } else {
822 bool done = false;
823 uint32_t base_page = va_page;
824 while (va_page < end_va_page) {
825 struct zink_sparse_backing *backing;
826 uint32_t backing_start;
827 uint32_t span_pages;
828
829 /* Skip pages that are already uncommitted. */
830 if (!comm[va_page].backing) {
831 va_page++;
832 continue;
833 }
834
835 if (!done) {
836 cur_sem = buffer_commit_single(screen, res, NULL, 0,
837 (uint64_t)base_page * ZINK_SPARSE_BUFFER_PAGE_SIZE,
838 (uint64_t)(end_va_page - base_page) * ZINK_SPARSE_BUFFER_PAGE_SIZE, false, cur_sem);
839 if (!cur_sem) {
840 ok = false;
841 goto out;
842 }
843 }
844 done = true;
845
846 /* Group contiguous spans of pages. */
847 backing = comm[va_page].backing;
848 backing_start = comm[va_page].page;
849 comm[va_page].backing = NULL;
850
851 span_pages = 1;
852 va_page++;
853
854 while (va_page < end_va_page &&
855 comm[va_page].backing == backing &&
856 comm[va_page].page == backing_start + span_pages) {
857 comm[va_page].backing = NULL;
858 va_page++;
859 span_pages++;
860 }
861
862 if (!sparse_backing_free(screen, bo, backing, backing_start, span_pages)) {
863 /* Couldn't allocate tracking data structures, so we have to leak */
864 fprintf(stderr, "zink: leaking sparse backing memory\n");
865 ok = false;
866 }
867 }
868 }
869 out:
870 *sem = cur_sem;
871 return ok;
872 }
873
874 static VkSemaphore
texture_commit_single(struct zink_screen * screen,struct zink_resource * res,VkSparseImageMemoryBind * ibind,unsigned num_binds,bool commit,VkSemaphore wait)875 texture_commit_single(struct zink_screen *screen, struct zink_resource *res, VkSparseImageMemoryBind *ibind, unsigned num_binds, bool commit, VkSemaphore wait)
876 {
877 VkSemaphore sem = get_semaphore(screen);
878 VkBindSparseInfo sparse = {0};
879 sparse.sType = VK_STRUCTURE_TYPE_BIND_SPARSE_INFO;
880 sparse.imageBindCount = 1;
881 sparse.waitSemaphoreCount = !!wait;
882 sparse.pWaitSemaphores = &wait;
883 sparse.signalSemaphoreCount = 1;
884 sparse.pSignalSemaphores = &sem;
885
886 VkSparseImageMemoryBindInfo sparse_ibind;
887 sparse_ibind.image = res->obj->image;
888 sparse_ibind.bindCount = num_binds;
889 sparse_ibind.pBinds = ibind;
890 sparse.pImageBinds = &sparse_ibind;
891
892 VkResult ret = VKSCR(QueueBindSparse)(screen->queue_sparse, 1, &sparse, VK_NULL_HANDLE);
893 if (zink_screen_handle_vkresult(screen, ret))
894 return sem;
895 VKSCR(DestroySemaphore)(screen->dev, sem, NULL);
896 return VK_NULL_HANDLE;
897 }
898
899 static VkSemaphore
texture_commit_miptail(struct zink_screen * screen,struct zink_resource * res,struct zink_bo * bo,uint32_t bo_offset,uint32_t offset,bool commit,VkSemaphore wait)900 texture_commit_miptail(struct zink_screen *screen, struct zink_resource *res, struct zink_bo *bo, uint32_t bo_offset, uint32_t offset, bool commit, VkSemaphore wait)
901 {
902 VkSemaphore sem = get_semaphore(screen);
903 VkBindSparseInfo sparse = {0};
904 sparse.sType = VK_STRUCTURE_TYPE_BIND_SPARSE_INFO;
905 sparse.imageOpaqueBindCount = 1;
906 sparse.waitSemaphoreCount = !!wait;
907 sparse.pWaitSemaphores = &wait;
908 sparse.signalSemaphoreCount = 1;
909 sparse.pSignalSemaphores = &sem;
910
911 VkSparseImageOpaqueMemoryBindInfo sparse_bind;
912 sparse_bind.image = res->obj->image;
913 sparse_bind.bindCount = 1;
914 sparse.pImageOpaqueBinds = &sparse_bind;
915
916 VkSparseMemoryBind mem_bind;
917 mem_bind.resourceOffset = offset;
918 mem_bind.size = MIN2(ZINK_SPARSE_BUFFER_PAGE_SIZE, res->sparse.imageMipTailSize - offset);
919 mem_bind.memory = commit ? (bo->mem ? bo->mem : bo->u.slab.real->mem) : VK_NULL_HANDLE;
920 mem_bind.memoryOffset = bo_offset + (commit ? (bo->mem ? 0 : bo->offset) : 0);
921 mem_bind.flags = 0;
922 sparse_bind.pBinds = &mem_bind;
923
924 VkResult ret = VKSCR(QueueBindSparse)(screen->queue_sparse, 1, &sparse, VK_NULL_HANDLE);
925 if (zink_screen_handle_vkresult(screen, ret))
926 return sem;
927 VKSCR(DestroySemaphore)(screen->dev, sem, NULL);
928 return VK_NULL_HANDLE;
929 }
930
931 bool
zink_bo_commit(struct zink_screen * screen,struct zink_resource * res,unsigned level,struct pipe_box * box,bool commit,VkSemaphore * sem)932 zink_bo_commit(struct zink_screen *screen, struct zink_resource *res, unsigned level, struct pipe_box *box, bool commit, VkSemaphore *sem)
933 {
934 bool ok = true;
935 struct zink_bo *bo = res->obj->bo;
936 VkSemaphore cur_sem = VK_NULL_HANDLE;
937
938 if (screen->faked_e5sparse && res->base.b.format == PIPE_FORMAT_R9G9B9E5_FLOAT)
939 return true;
940
941 simple_mtx_lock(&screen->queue_lock);
942 simple_mtx_lock(&bo->lock);
943 if (res->base.b.target == PIPE_BUFFER) {
944 ok = buffer_bo_commit(screen, res, box->x, box->width, commit, sem);
945 goto out;
946 }
947
948 int gwidth, gheight, gdepth;
949 gwidth = res->sparse.formatProperties.imageGranularity.width;
950 gheight = res->sparse.formatProperties.imageGranularity.height;
951 gdepth = res->sparse.formatProperties.imageGranularity.depth;
952 assert(gwidth && gheight && gdepth);
953
954 struct zink_sparse_commitment *comm = bo->u.sparse.commitments;
955 VkImageSubresource subresource = { res->aspect, level, 0 };
956 unsigned nwidth = DIV_ROUND_UP(box->width, gwidth);
957 unsigned nheight = DIV_ROUND_UP(box->height, gheight);
958 unsigned ndepth = DIV_ROUND_UP(box->depth, gdepth);
959 VkExtent3D lastBlockExtent = {
960 (box->width % gwidth) ? box->width % gwidth : gwidth,
961 (box->height % gheight) ? box->height % gheight : gheight,
962 (box->depth % gdepth) ? box->depth % gdepth : gdepth
963 };
964 #define NUM_BATCHED_BINDS 50
965 VkSparseImageMemoryBind ibind[NUM_BATCHED_BINDS];
966 uint32_t backing_start[NUM_BATCHED_BINDS], backing_size[NUM_BATCHED_BINDS];
967 struct zink_sparse_backing *backing[NUM_BATCHED_BINDS];
968 unsigned i = 0;
969 bool commits_pending = false;
970 uint32_t va_page_offset = 0;
971 for (unsigned l = 0; l < level; l++) {
972 unsigned mipwidth = DIV_ROUND_UP(MAX2(res->base.b.width0 >> l, 1), gwidth);
973 unsigned mipheight = DIV_ROUND_UP(MAX2(res->base.b.height0 >> l, 1), gheight);
974 unsigned mipdepth = DIV_ROUND_UP(res->base.b.array_size > 1 ? res->base.b.array_size : MAX2(res->base.b.depth0 >> l, 1), gdepth);
975 va_page_offset += mipwidth * mipheight * mipdepth;
976 }
977 for (unsigned d = 0; d < ndepth; d++) {
978 for (unsigned h = 0; h < nheight; h++) {
979 for (unsigned w = 0; w < nwidth; w++) {
980 ibind[i].subresource = subresource;
981 ibind[i].flags = 0;
982 // Offset
983 ibind[i].offset.x = w * gwidth;
984 ibind[i].offset.y = h * gheight;
985 if (res->base.b.array_size > 1) {
986 ibind[i].subresource.arrayLayer = d * gdepth;
987 ibind[i].offset.z = 0;
988 } else {
989 ibind[i].offset.z = d * gdepth;
990 }
991 // Size of the page
992 ibind[i].extent.width = (w == nwidth - 1) ? lastBlockExtent.width : gwidth;
993 ibind[i].extent.height = (h == nheight - 1) ? lastBlockExtent.height : gheight;
994 ibind[i].extent.depth = (d == ndepth - 1 && res->base.b.target != PIPE_TEXTURE_CUBE) ? lastBlockExtent.depth : gdepth;
995 uint32_t va_page = va_page_offset +
996 (d + (box->z / gdepth)) * ((MAX2(res->base.b.width0 >> level, 1) / gwidth) * (MAX2(res->base.b.height0 >> level, 1) / gheight)) +
997 (h + (box->y / gheight)) * (MAX2(res->base.b.width0 >> level, 1) / gwidth) +
998 (w + (box->x / gwidth));
999
1000 uint32_t end_va_page = va_page + 1;
1001
1002 if (commit) {
1003 while (va_page < end_va_page) {
1004 uint32_t span_va_page;
1005
1006 /* Skip pages that are already committed. */
1007 if (comm[va_page].backing) {
1008 va_page++;
1009 continue;
1010 }
1011
1012 /* Determine length of uncommitted span. */
1013 span_va_page = va_page;
1014 while (va_page < end_va_page && !comm[va_page].backing)
1015 va_page++;
1016
1017 /* Fill the uncommitted span with chunks of backing memory. */
1018 while (span_va_page < va_page) {
1019 backing_size[i] = va_page - span_va_page;
1020 backing[i] = sparse_backing_alloc(screen, bo, &backing_start[i], &backing_size[i]);
1021 if (!backing[i]) {
1022 ok = false;
1023 goto out;
1024 }
1025 if (level >= res->sparse.imageMipTailFirstLod) {
1026 uint32_t offset = res->sparse.imageMipTailOffset + d * res->sparse.imageMipTailStride;
1027 cur_sem = texture_commit_miptail(screen, res, backing[i]->bo, backing_start[i], offset, commit, cur_sem);
1028 if (!cur_sem)
1029 goto out;
1030 } else {
1031 ibind[i].memory = backing[i]->bo->mem ? backing[i]->bo->mem : backing[i]->bo->u.slab.real->mem;
1032 ibind[i].memoryOffset = backing_start[i] * ZINK_SPARSE_BUFFER_PAGE_SIZE +
1033 (backing[i]->bo->mem ? 0 : backing[i]->bo->offset);
1034 commits_pending = true;
1035 }
1036
1037 while (backing_size[i]) {
1038 comm[span_va_page].backing = backing[i];
1039 comm[span_va_page].page = backing_start[i];
1040 span_va_page++;
1041 backing_start[i]++;
1042 backing_size[i]--;
1043 }
1044 i++;
1045 }
1046 }
1047 } else {
1048 ibind[i].memory = VK_NULL_HANDLE;
1049 ibind[i].memoryOffset = 0;
1050
1051 while (va_page < end_va_page) {
1052 /* Skip pages that are already uncommitted. */
1053 if (!comm[va_page].backing) {
1054 va_page++;
1055 continue;
1056 }
1057
1058 /* Group contiguous spans of pages. */
1059 backing[i] = comm[va_page].backing;
1060 backing_start[i] = comm[va_page].page;
1061 comm[va_page].backing = NULL;
1062
1063 backing_size[i] = 1;
1064 va_page++;
1065
1066 while (va_page < end_va_page &&
1067 comm[va_page].backing == backing[i] &&
1068 comm[va_page].page == backing_start[i] + backing_size[i]) {
1069 comm[va_page].backing = NULL;
1070 va_page++;
1071 backing_size[i]++;
1072 }
1073 if (level >= res->sparse.imageMipTailFirstLod) {
1074 uint32_t offset = res->sparse.imageMipTailOffset + d * res->sparse.imageMipTailStride;
1075 cur_sem = texture_commit_miptail(screen, res, NULL, 0, offset, commit, cur_sem);
1076 if (!cur_sem)
1077 goto out;
1078 } else {
1079 commits_pending = true;
1080 }
1081 i++;
1082 }
1083 }
1084 if (i == ARRAY_SIZE(ibind)) {
1085 cur_sem = texture_commit_single(screen, res, ibind, ARRAY_SIZE(ibind), commit, cur_sem);
1086 if (!cur_sem) {
1087 for (unsigned s = 0; s < i; s++) {
1088 ok = sparse_backing_free(screen, backing[s]->bo, backing[s], backing_start[s], backing_size[s]);
1089 if (!ok) {
1090 /* Couldn't allocate tracking data structures, so we have to leak */
1091 fprintf(stderr, "zink: leaking sparse backing memory\n");
1092 }
1093 }
1094 ok = false;
1095 goto out;
1096 }
1097 commits_pending = false;
1098 i = 0;
1099 }
1100 }
1101 }
1102 }
1103 if (commits_pending) {
1104 cur_sem = texture_commit_single(screen, res, ibind, i, commit, cur_sem);
1105 if (!cur_sem) {
1106 for (unsigned s = 0; s < i; s++) {
1107 ok = sparse_backing_free(screen, backing[s]->bo, backing[s], backing_start[s], backing_size[s]);
1108 if (!ok) {
1109 /* Couldn't allocate tracking data structures, so we have to leak */
1110 fprintf(stderr, "zink: leaking sparse backing memory\n");
1111 }
1112 }
1113 }
1114 ok = false;
1115 }
1116 out:
1117
1118 simple_mtx_unlock(&bo->lock);
1119 simple_mtx_unlock(&screen->queue_lock);
1120 *sem = cur_sem;
1121 return ok;
1122 }
1123
1124 bool
zink_bo_get_kms_handle(struct zink_screen * screen,struct zink_bo * bo,int fd,uint32_t * handle)1125 zink_bo_get_kms_handle(struct zink_screen *screen, struct zink_bo *bo, int fd, uint32_t *handle)
1126 {
1127 #ifdef ZINK_USE_DMABUF
1128 assert(bo->mem && !bo->u.real.use_reusable_pool);
1129 simple_mtx_lock(&bo->u.real.export_lock);
1130 list_for_each_entry(struct bo_export, export, &bo->u.real.exports, link) {
1131 if (export->drm_fd == fd) {
1132 simple_mtx_unlock(&bo->u.real.export_lock);
1133 *handle = export->gem_handle;
1134 return true;
1135 }
1136 }
1137 struct bo_export *export = CALLOC_STRUCT(bo_export);
1138 if (!export) {
1139 simple_mtx_unlock(&bo->u.real.export_lock);
1140 return false;
1141 }
1142 bool success = drmPrimeFDToHandle(screen->drm_fd, fd, handle) == 0;
1143 if (success) {
1144 list_addtail(&export->link, &bo->u.real.exports);
1145 export->gem_handle = *handle;
1146 export->drm_fd = screen->drm_fd;
1147 } else {
1148 mesa_loge("zink: failed drmPrimeFDToHandle %s", strerror(errno));
1149 FREE(export);
1150 }
1151 simple_mtx_unlock(&bo->u.real.export_lock);
1152 return success;
1153 #else
1154 return false;
1155 #endif
1156 }
1157
1158 static const struct pb_vtbl bo_slab_vtbl = {
1159 /* Cast to void* because one of the function parameters is a struct pointer instead of void*. */
1160 (void*)bo_slab_destroy
1161 /* other functions are never called */
1162 };
1163
1164 static struct pb_slab *
bo_slab_alloc(void * priv,unsigned heap,unsigned entry_size,unsigned group_index,bool encrypted)1165 bo_slab_alloc(void *priv, unsigned heap, unsigned entry_size, unsigned group_index, bool encrypted)
1166 {
1167 struct zink_screen *screen = priv;
1168 uint32_t base_id;
1169 unsigned slab_size = 0;
1170 struct zink_slab *slab = CALLOC_STRUCT(zink_slab);
1171
1172 if (!slab)
1173 return NULL;
1174
1175 //struct pb_slabs *slabs = ((flags & RADEON_FLAG_ENCRYPTED) && screen->info.has_tmz_support) ?
1176 //screen->bo_slabs_encrypted : screen->bo_slabs;
1177 struct pb_slabs *slabs = screen->pb.bo_slabs;
1178
1179 /* Determine the slab buffer size. */
1180 for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
1181 unsigned max_entry_size = 1 << (slabs[i].min_order + slabs[i].num_orders - 1);
1182
1183 if (entry_size <= max_entry_size) {
1184 /* The slab size is twice the size of the largest possible entry. */
1185 slab_size = max_entry_size * 2;
1186
1187 if (!util_is_power_of_two_nonzero(entry_size)) {
1188 assert(util_is_power_of_two_nonzero(entry_size * 4 / 3));
1189
1190 /* If the entry size is 3/4 of a power of two, we would waste space and not gain
1191 * anything if we allocated only twice the power of two for the backing buffer:
1192 * 2 * 3/4 = 1.5 usable with buffer size 2
1193 *
1194 * Allocating 5 times the entry size leads us to the next power of two and results
1195 * in a much better memory utilization:
1196 * 5 * 3/4 = 3.75 usable with buffer size 4
1197 */
1198 if (entry_size * 5 > slab_size)
1199 slab_size = util_next_power_of_two(entry_size * 5);
1200 }
1201
1202 break;
1203 }
1204 }
1205 assert(slab_size != 0);
1206
1207 slab->buffer = zink_bo(zink_bo_create(screen, slab_size, slab_size, heap, 0, NULL));
1208 if (!slab->buffer)
1209 goto fail;
1210
1211 slab_size = slab->buffer->base.size;
1212
1213 slab->base.num_entries = slab_size / entry_size;
1214 slab->base.num_free = slab->base.num_entries;
1215 slab->entry_size = entry_size;
1216 slab->entries = CALLOC(slab->base.num_entries, sizeof(*slab->entries));
1217 if (!slab->entries)
1218 goto fail_buffer;
1219
1220 list_inithead(&slab->base.free);
1221
1222 base_id = p_atomic_fetch_add(&screen->pb.next_bo_unique_id, slab->base.num_entries);
1223 for (unsigned i = 0; i < slab->base.num_entries; ++i) {
1224 struct zink_bo *bo = &slab->entries[i];
1225
1226 simple_mtx_init(&bo->lock, mtx_plain);
1227 bo->base.alignment_log2 = util_logbase2(get_slab_entry_alignment(screen, entry_size));
1228 bo->base.size = entry_size;
1229 bo->base.vtbl = &bo_slab_vtbl;
1230 bo->offset = slab->buffer->offset + i * entry_size;
1231 bo->unique_id = base_id + i;
1232 bo->u.slab.entry.slab = &slab->base;
1233 bo->u.slab.entry.group_index = group_index;
1234 bo->u.slab.entry.entry_size = entry_size;
1235
1236 if (slab->buffer->mem) {
1237 /* The slab is not suballocated. */
1238 bo->u.slab.real = slab->buffer;
1239 } else {
1240 /* The slab is allocated out of a bigger slab. */
1241 bo->u.slab.real = slab->buffer->u.slab.real;
1242 assert(bo->u.slab.real->mem);
1243 }
1244 bo->base.placement = bo->u.slab.real->base.placement;
1245
1246 list_addtail(&bo->u.slab.entry.head, &slab->base.free);
1247 }
1248
1249 /* Wasted alignment due to slabs with 3/4 allocations being aligned to a power of two. */
1250 assert(slab->base.num_entries * entry_size <= slab_size);
1251
1252 return &slab->base;
1253
1254 fail_buffer:
1255 zink_bo_unref(screen, slab->buffer);
1256 fail:
1257 FREE(slab);
1258 return NULL;
1259 }
1260
1261 static struct pb_slab *
bo_slab_alloc_normal(void * priv,unsigned heap,unsigned entry_size,unsigned group_index)1262 bo_slab_alloc_normal(void *priv, unsigned heap, unsigned entry_size, unsigned group_index)
1263 {
1264 return bo_slab_alloc(priv, heap, entry_size, group_index, false);
1265 }
1266
1267 bool
zink_bo_init(struct zink_screen * screen)1268 zink_bo_init(struct zink_screen *screen)
1269 {
1270 uint64_t total_mem = 0;
1271 for (uint32_t i = 0; i < screen->info.mem_props.memoryHeapCount; ++i)
1272 total_mem += screen->info.mem_props.memoryHeaps[i].size;
1273 /* Create managers. */
1274 pb_cache_init(&screen->pb.bo_cache, ZINK_HEAP_MAX,
1275 500000, 2.0f, 0,
1276 total_mem / 8, screen,
1277 (void*)bo_destroy, (void*)bo_can_reclaim);
1278
1279 unsigned min_slab_order = MIN_SLAB_ORDER; /* 256 bytes */
1280 unsigned max_slab_order = 20; /* 1 MB (slab size = 2 MB) */
1281 unsigned num_slab_orders_per_allocator = (max_slab_order - min_slab_order) /
1282 NUM_SLAB_ALLOCATORS;
1283
1284 /* Divide the size order range among slab managers. */
1285 for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
1286 unsigned min_order = min_slab_order;
1287 unsigned max_order = MIN2(min_order + num_slab_orders_per_allocator,
1288 max_slab_order);
1289
1290 if (!pb_slabs_init(&screen->pb.bo_slabs[i],
1291 min_order, max_order,
1292 ZINK_HEAP_MAX, true,
1293 screen,
1294 bo_can_reclaim_slab,
1295 bo_slab_alloc_normal,
1296 (void*)bo_slab_free)) {
1297 return false;
1298 }
1299 min_slab_order = max_order + 1;
1300 }
1301 screen->pb.min_alloc_size = 1 << screen->pb.bo_slabs[0].min_order;
1302 return true;
1303 }
1304
1305 void
zink_bo_deinit(struct zink_screen * screen)1306 zink_bo_deinit(struct zink_screen *screen)
1307 {
1308 for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
1309 if (screen->pb.bo_slabs[i].groups)
1310 pb_slabs_deinit(&screen->pb.bo_slabs[i]);
1311 }
1312 pb_cache_deinit(&screen->pb.bo_cache);
1313 }
1314