1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
3 *
4 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * 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 above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28 /*
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30 */
31 #include <linux/swap.h>
32 #include <linux/vmalloc.h>
33
34 #include <drm/ttm/ttm_bo.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <drm/ttm/ttm_tt.h>
37
38 #include <drm/drm_cache.h>
39
40 struct ttm_transfer_obj {
41 struct ttm_buffer_object base;
42 struct ttm_buffer_object *bo;
43 };
44
ttm_mem_io_reserve(struct ttm_device * bdev,struct ttm_resource * mem)45 int ttm_mem_io_reserve(struct ttm_device *bdev,
46 struct ttm_resource *mem)
47 {
48 if (mem->bus.offset || mem->bus.addr)
49 return 0;
50
51 mem->bus.is_iomem = false;
52 if (!bdev->funcs->io_mem_reserve)
53 return 0;
54
55 return bdev->funcs->io_mem_reserve(bdev, mem);
56 }
57
ttm_mem_io_free(struct ttm_device * bdev,struct ttm_resource * mem)58 void ttm_mem_io_free(struct ttm_device *bdev,
59 struct ttm_resource *mem)
60 {
61 if (!mem)
62 return;
63
64 if (!mem->bus.offset && !mem->bus.addr)
65 return;
66
67 if (bdev->funcs->io_mem_free)
68 bdev->funcs->io_mem_free(bdev, mem);
69
70 mem->bus.offset = 0;
71 mem->bus.addr = NULL;
72 }
73
74 /**
75 * ttm_move_memcpy - Helper to perform a memcpy ttm move operation.
76 * @clear: Whether to clear rather than copy.
77 * @num_pages: Number of pages of the operation.
78 * @dst_iter: A struct ttm_kmap_iter representing the destination resource.
79 * @src_iter: A struct ttm_kmap_iter representing the source resource.
80 *
81 * This function is intended to be able to move out async under a
82 * dma-fence if desired.
83 */
ttm_move_memcpy(bool clear,u32 num_pages,struct ttm_kmap_iter * dst_iter,struct ttm_kmap_iter * src_iter)84 void ttm_move_memcpy(bool clear,
85 u32 num_pages,
86 struct ttm_kmap_iter *dst_iter,
87 struct ttm_kmap_iter *src_iter)
88 {
89 const struct ttm_kmap_iter_ops *dst_ops = dst_iter->ops;
90 const struct ttm_kmap_iter_ops *src_ops = src_iter->ops;
91 struct iosys_map src_map, dst_map;
92 pgoff_t i;
93
94 /* Single TTM move. NOP */
95 if (dst_ops->maps_tt && src_ops->maps_tt)
96 return;
97
98 /* Don't move nonexistent data. Clear destination instead. */
99 if (clear) {
100 for (i = 0; i < num_pages; ++i) {
101 dst_ops->map_local(dst_iter, &dst_map, i);
102 if (dst_map.is_iomem)
103 memset_io(dst_map.vaddr_iomem, 0, PAGE_SIZE);
104 else
105 memset(dst_map.vaddr, 0, PAGE_SIZE);
106 if (dst_ops->unmap_local)
107 dst_ops->unmap_local(dst_iter, &dst_map);
108 }
109 return;
110 }
111
112 for (i = 0; i < num_pages; ++i) {
113 dst_ops->map_local(dst_iter, &dst_map, i);
114 src_ops->map_local(src_iter, &src_map, i);
115
116 drm_memcpy_from_wc(&dst_map, &src_map, PAGE_SIZE);
117
118 if (src_ops->unmap_local)
119 src_ops->unmap_local(src_iter, &src_map);
120 if (dst_ops->unmap_local)
121 dst_ops->unmap_local(dst_iter, &dst_map);
122 }
123 }
124 EXPORT_SYMBOL(ttm_move_memcpy);
125
126 /**
127 * ttm_bo_move_memcpy
128 *
129 * @bo: A pointer to a struct ttm_buffer_object.
130 * @ctx: operation context
131 * @dst_mem: struct ttm_resource indicating where to move.
132 *
133 * Fallback move function for a mappable buffer object in mappable memory.
134 * The function will, if successful,
135 * free any old aperture space, and set (@new_mem)->mm_node to NULL,
136 * and update the (@bo)->mem placement flags. If unsuccessful, the old
137 * data remains untouched, and it's up to the caller to free the
138 * memory space indicated by @new_mem.
139 * Returns:
140 * !0: Failure.
141 */
ttm_bo_move_memcpy(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,struct ttm_resource * dst_mem)142 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo,
143 struct ttm_operation_ctx *ctx,
144 struct ttm_resource *dst_mem)
145 {
146 struct ttm_device *bdev = bo->bdev;
147 struct ttm_resource_manager *dst_man =
148 ttm_manager_type(bo->bdev, dst_mem->mem_type);
149 struct ttm_tt *ttm = bo->ttm;
150 struct ttm_resource *src_mem = bo->resource;
151 struct ttm_resource_manager *src_man;
152 union {
153 struct ttm_kmap_iter_tt tt;
154 struct ttm_kmap_iter_linear_io io;
155 } _dst_iter, _src_iter;
156 struct ttm_kmap_iter *dst_iter, *src_iter;
157 bool clear;
158 int ret = 0;
159
160 if (WARN_ON(!src_mem))
161 return -EINVAL;
162
163 src_man = ttm_manager_type(bdev, src_mem->mem_type);
164 if (ttm && ((ttm->page_flags & TTM_TT_FLAG_SWAPPED) ||
165 dst_man->use_tt)) {
166 ret = ttm_bo_populate(bo, ctx);
167 if (ret)
168 return ret;
169 }
170
171 dst_iter = ttm_kmap_iter_linear_io_init(&_dst_iter.io, bdev, dst_mem);
172 if (PTR_ERR(dst_iter) == -EINVAL && dst_man->use_tt)
173 dst_iter = ttm_kmap_iter_tt_init(&_dst_iter.tt, bo->ttm);
174 if (IS_ERR(dst_iter))
175 return PTR_ERR(dst_iter);
176
177 src_iter = ttm_kmap_iter_linear_io_init(&_src_iter.io, bdev, src_mem);
178 if (PTR_ERR(src_iter) == -EINVAL && src_man->use_tt)
179 src_iter = ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm);
180 if (IS_ERR(src_iter)) {
181 ret = PTR_ERR(src_iter);
182 goto out_src_iter;
183 }
184
185 clear = src_iter->ops->maps_tt && (!ttm || !ttm_tt_is_populated(ttm));
186 if (!(clear && ttm && !(ttm->page_flags & TTM_TT_FLAG_ZERO_ALLOC)))
187 ttm_move_memcpy(clear, PFN_UP(dst_mem->size), dst_iter, src_iter);
188
189 if (!src_iter->ops->maps_tt)
190 ttm_kmap_iter_linear_io_fini(&_src_iter.io, bdev, src_mem);
191 ttm_bo_move_sync_cleanup(bo, dst_mem);
192
193 out_src_iter:
194 if (!dst_iter->ops->maps_tt)
195 ttm_kmap_iter_linear_io_fini(&_dst_iter.io, bdev, dst_mem);
196
197 return ret;
198 }
199 EXPORT_SYMBOL(ttm_bo_move_memcpy);
200
ttm_transfered_destroy(struct ttm_buffer_object * bo)201 static void ttm_transfered_destroy(struct ttm_buffer_object *bo)
202 {
203 struct ttm_transfer_obj *fbo;
204
205 fbo = container_of(bo, struct ttm_transfer_obj, base);
206 dma_resv_fini(&fbo->base.base._resv);
207 ttm_bo_put(fbo->bo);
208 kfree(fbo);
209 }
210
211 /**
212 * ttm_buffer_object_transfer
213 *
214 * @bo: A pointer to a struct ttm_buffer_object.
215 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object,
216 * holding the data of @bo with the old placement.
217 *
218 * This is a utility function that may be called after an accelerated move
219 * has been scheduled. A new buffer object is created as a placeholder for
220 * the old data while it's being copied. When that buffer object is idle,
221 * it can be destroyed, releasing the space of the old placement.
222 * Returns:
223 * !0: Failure.
224 */
225
ttm_buffer_object_transfer(struct ttm_buffer_object * bo,struct ttm_buffer_object ** new_obj)226 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
227 struct ttm_buffer_object **new_obj)
228 {
229 struct ttm_transfer_obj *fbo;
230 int ret;
231
232 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
233 if (!fbo)
234 return -ENOMEM;
235
236 fbo->base = *bo;
237
238 /**
239 * Fix up members that we shouldn't copy directly:
240 * TODO: Explicit member copy would probably be better here.
241 */
242
243 atomic_inc(&ttm_glob.bo_count);
244 drm_vma_node_reset(&fbo->base.base.vma_node);
245
246 kref_init(&fbo->base.kref);
247 fbo->base.destroy = &ttm_transfered_destroy;
248 fbo->base.pin_count = 0;
249 if (bo->type != ttm_bo_type_sg)
250 fbo->base.base.resv = &fbo->base.base._resv;
251
252 dma_resv_init(&fbo->base.base._resv);
253 fbo->base.base.dev = NULL;
254 ret = dma_resv_trylock(&fbo->base.base._resv);
255 WARN_ON(!ret);
256
257 if (fbo->base.resource) {
258 ttm_resource_set_bo(fbo->base.resource, &fbo->base);
259 bo->resource = NULL;
260 ttm_bo_set_bulk_move(&fbo->base, NULL);
261 } else {
262 fbo->base.bulk_move = NULL;
263 }
264
265 ret = dma_resv_reserve_fences(&fbo->base.base._resv, 1);
266 if (ret) {
267 kfree(fbo);
268 return ret;
269 }
270
271 ttm_bo_get(bo);
272 fbo->bo = bo;
273
274 ttm_bo_move_to_lru_tail_unlocked(&fbo->base);
275
276 *new_obj = &fbo->base;
277 return 0;
278 }
279
280 /**
281 * ttm_io_prot
282 *
283 * @bo: ttm buffer object
284 * @res: ttm resource object
285 * @tmp: Page protection flag for a normal, cached mapping.
286 *
287 * Utility function that returns the pgprot_t that should be used for
288 * setting up a PTE with the caching model indicated by @c_state.
289 */
ttm_io_prot(struct ttm_buffer_object * bo,struct ttm_resource * res,pgprot_t tmp)290 pgprot_t ttm_io_prot(struct ttm_buffer_object *bo, struct ttm_resource *res,
291 pgprot_t tmp)
292 {
293 struct ttm_resource_manager *man;
294 enum ttm_caching caching;
295
296 man = ttm_manager_type(bo->bdev, res->mem_type);
297 if (man->use_tt) {
298 caching = bo->ttm->caching;
299 if (bo->ttm->page_flags & TTM_TT_FLAG_DECRYPTED)
300 tmp = pgprot_decrypted(tmp);
301 } else {
302 caching = res->bus.caching;
303 }
304
305 return ttm_prot_from_caching(caching, tmp);
306 }
307 EXPORT_SYMBOL(ttm_io_prot);
308
ttm_bo_ioremap(struct ttm_buffer_object * bo,unsigned long offset,unsigned long size,struct ttm_bo_kmap_obj * map)309 static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
310 unsigned long offset,
311 unsigned long size,
312 struct ttm_bo_kmap_obj *map)
313 {
314 struct ttm_resource *mem = bo->resource;
315
316 if (bo->resource->bus.addr) {
317 map->bo_kmap_type = ttm_bo_map_premapped;
318 map->virtual = ((u8 *)bo->resource->bus.addr) + offset;
319 } else {
320 resource_size_t res = bo->resource->bus.offset + offset;
321
322 map->bo_kmap_type = ttm_bo_map_iomap;
323 if (mem->bus.caching == ttm_write_combined)
324 map->virtual = ioremap_wc(res, size);
325 #ifdef CONFIG_X86
326 else if (mem->bus.caching == ttm_cached)
327 map->virtual = ioremap_cache(res, size);
328 #endif
329 else
330 map->virtual = ioremap(res, size);
331 }
332 return (!map->virtual) ? -ENOMEM : 0;
333 }
334
ttm_bo_kmap_ttm(struct ttm_buffer_object * bo,unsigned long start_page,unsigned long num_pages,struct ttm_bo_kmap_obj * map)335 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo,
336 unsigned long start_page,
337 unsigned long num_pages,
338 struct ttm_bo_kmap_obj *map)
339 {
340 struct ttm_resource *mem = bo->resource;
341 struct ttm_operation_ctx ctx = {
342 .interruptible = false,
343 .no_wait_gpu = false
344 };
345 struct ttm_tt *ttm = bo->ttm;
346 struct ttm_resource_manager *man =
347 ttm_manager_type(bo->bdev, bo->resource->mem_type);
348 pgprot_t prot;
349 int ret;
350
351 BUG_ON(!ttm);
352
353 ret = ttm_bo_populate(bo, &ctx);
354 if (ret)
355 return ret;
356
357 if (num_pages == 1 && ttm->caching == ttm_cached &&
358 !(man->use_tt && (ttm->page_flags & TTM_TT_FLAG_DECRYPTED))) {
359 /*
360 * We're mapping a single page, and the desired
361 * page protection is consistent with the bo.
362 */
363
364 map->bo_kmap_type = ttm_bo_map_kmap;
365 map->page = ttm->pages[start_page];
366 map->virtual = kmap(map->page);
367 } else {
368 /*
369 * We need to use vmap to get the desired page protection
370 * or to make the buffer object look contiguous.
371 */
372 prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
373 map->bo_kmap_type = ttm_bo_map_vmap;
374 map->virtual = vmap(ttm->pages + start_page, num_pages,
375 0, prot);
376 }
377 return (!map->virtual) ? -ENOMEM : 0;
378 }
379
380 /**
381 * ttm_bo_kmap
382 *
383 * @bo: The buffer object.
384 * @start_page: The first page to map.
385 * @num_pages: Number of pages to map.
386 * @map: pointer to a struct ttm_bo_kmap_obj representing the map.
387 *
388 * Sets up a kernel virtual mapping, using ioremap, vmap or kmap to the
389 * data in the buffer object. The ttm_kmap_obj_virtual function can then be
390 * used to obtain a virtual address to the data.
391 *
392 * Returns
393 * -ENOMEM: Out of memory.
394 * -EINVAL: Invalid range.
395 */
ttm_bo_kmap(struct ttm_buffer_object * bo,unsigned long start_page,unsigned long num_pages,struct ttm_bo_kmap_obj * map)396 int ttm_bo_kmap(struct ttm_buffer_object *bo,
397 unsigned long start_page, unsigned long num_pages,
398 struct ttm_bo_kmap_obj *map)
399 {
400 unsigned long offset, size;
401 int ret;
402
403 map->virtual = NULL;
404 map->bo = bo;
405 if (num_pages > PFN_UP(bo->resource->size))
406 return -EINVAL;
407 if ((start_page + num_pages) > PFN_UP(bo->resource->size))
408 return -EINVAL;
409
410 ret = ttm_mem_io_reserve(bo->bdev, bo->resource);
411 if (ret)
412 return ret;
413 if (!bo->resource->bus.is_iomem) {
414 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map);
415 } else {
416 offset = start_page << PAGE_SHIFT;
417 size = num_pages << PAGE_SHIFT;
418 return ttm_bo_ioremap(bo, offset, size, map);
419 }
420 }
421 EXPORT_SYMBOL(ttm_bo_kmap);
422
423 /**
424 * ttm_bo_kunmap
425 *
426 * @map: Object describing the map to unmap.
427 *
428 * Unmaps a kernel map set up by ttm_bo_kmap.
429 */
ttm_bo_kunmap(struct ttm_bo_kmap_obj * map)430 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
431 {
432 if (!map->virtual)
433 return;
434 switch (map->bo_kmap_type) {
435 case ttm_bo_map_iomap:
436 iounmap(map->virtual);
437 break;
438 case ttm_bo_map_vmap:
439 vunmap(map->virtual);
440 break;
441 case ttm_bo_map_kmap:
442 kunmap(map->page);
443 break;
444 case ttm_bo_map_premapped:
445 break;
446 default:
447 BUG();
448 }
449 ttm_mem_io_free(map->bo->bdev, map->bo->resource);
450 map->virtual = NULL;
451 map->page = NULL;
452 }
453 EXPORT_SYMBOL(ttm_bo_kunmap);
454
455 /**
456 * ttm_bo_vmap
457 *
458 * @bo: The buffer object.
459 * @map: pointer to a struct iosys_map representing the map.
460 *
461 * Sets up a kernel virtual mapping, using ioremap or vmap to the
462 * data in the buffer object. The parameter @map returns the virtual
463 * address as struct iosys_map. Unmap the buffer with ttm_bo_vunmap().
464 *
465 * Returns
466 * -ENOMEM: Out of memory.
467 * -EINVAL: Invalid range.
468 */
ttm_bo_vmap(struct ttm_buffer_object * bo,struct iosys_map * map)469 int ttm_bo_vmap(struct ttm_buffer_object *bo, struct iosys_map *map)
470 {
471 struct ttm_resource *mem = bo->resource;
472 int ret;
473
474 dma_resv_assert_held(bo->base.resv);
475
476 ret = ttm_mem_io_reserve(bo->bdev, mem);
477 if (ret)
478 return ret;
479
480 if (mem->bus.is_iomem) {
481 void __iomem *vaddr_iomem;
482
483 if (mem->bus.addr)
484 vaddr_iomem = (void __iomem *)mem->bus.addr;
485 else if (mem->bus.caching == ttm_write_combined)
486 vaddr_iomem = ioremap_wc(mem->bus.offset,
487 bo->base.size);
488 #ifdef CONFIG_X86
489 else if (mem->bus.caching == ttm_cached)
490 vaddr_iomem = ioremap_cache(mem->bus.offset,
491 bo->base.size);
492 #endif
493 else
494 vaddr_iomem = ioremap(mem->bus.offset, bo->base.size);
495
496 if (!vaddr_iomem)
497 return -ENOMEM;
498
499 iosys_map_set_vaddr_iomem(map, vaddr_iomem);
500
501 } else {
502 struct ttm_operation_ctx ctx = {
503 .interruptible = false,
504 .no_wait_gpu = false
505 };
506 struct ttm_tt *ttm = bo->ttm;
507 pgprot_t prot;
508 void *vaddr;
509
510 ret = ttm_bo_populate(bo, &ctx);
511 if (ret)
512 return ret;
513
514 /*
515 * We need to use vmap to get the desired page protection
516 * or to make the buffer object look contiguous.
517 */
518 prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
519 vaddr = vmap(ttm->pages, ttm->num_pages, 0, prot);
520 if (!vaddr)
521 return -ENOMEM;
522
523 iosys_map_set_vaddr(map, vaddr);
524 }
525
526 return 0;
527 }
528 EXPORT_SYMBOL(ttm_bo_vmap);
529
530 /**
531 * ttm_bo_vunmap
532 *
533 * @bo: The buffer object.
534 * @map: Object describing the map to unmap.
535 *
536 * Unmaps a kernel map set up by ttm_bo_vmap().
537 */
ttm_bo_vunmap(struct ttm_buffer_object * bo,struct iosys_map * map)538 void ttm_bo_vunmap(struct ttm_buffer_object *bo, struct iosys_map *map)
539 {
540 struct ttm_resource *mem = bo->resource;
541
542 dma_resv_assert_held(bo->base.resv);
543
544 if (iosys_map_is_null(map))
545 return;
546
547 if (!map->is_iomem)
548 vunmap(map->vaddr);
549 else if (!mem->bus.addr)
550 iounmap(map->vaddr_iomem);
551 iosys_map_clear(map);
552
553 ttm_mem_io_free(bo->bdev, bo->resource);
554 }
555 EXPORT_SYMBOL(ttm_bo_vunmap);
556
ttm_bo_wait_free_node(struct ttm_buffer_object * bo,bool dst_use_tt)557 static int ttm_bo_wait_free_node(struct ttm_buffer_object *bo,
558 bool dst_use_tt)
559 {
560 long ret;
561
562 ret = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
563 false, 15 * HZ);
564 if (ret == 0)
565 return -EBUSY;
566 if (ret < 0)
567 return ret;
568
569 if (!dst_use_tt)
570 ttm_bo_tt_destroy(bo);
571 ttm_resource_free(bo, &bo->resource);
572 return 0;
573 }
574
ttm_bo_move_to_ghost(struct ttm_buffer_object * bo,struct dma_fence * fence,bool dst_use_tt)575 static int ttm_bo_move_to_ghost(struct ttm_buffer_object *bo,
576 struct dma_fence *fence,
577 bool dst_use_tt)
578 {
579 struct ttm_buffer_object *ghost_obj;
580 int ret;
581
582 /**
583 * This should help pipeline ordinary buffer moves.
584 *
585 * Hang old buffer memory on a new buffer object,
586 * and leave it to be released when the GPU
587 * operation has completed.
588 */
589
590 ret = ttm_buffer_object_transfer(bo, &ghost_obj);
591 if (ret)
592 return ret;
593
594 dma_resv_add_fence(&ghost_obj->base._resv, fence,
595 DMA_RESV_USAGE_KERNEL);
596
597 /**
598 * If we're not moving to fixed memory, the TTM object
599 * needs to stay alive. Otherwhise hang it on the ghost
600 * bo to be unbound and destroyed.
601 */
602
603 if (dst_use_tt)
604 ghost_obj->ttm = NULL;
605 else
606 bo->ttm = NULL;
607
608 dma_resv_unlock(&ghost_obj->base._resv);
609 ttm_bo_put(ghost_obj);
610 return 0;
611 }
612
ttm_bo_move_pipeline_evict(struct ttm_buffer_object * bo,struct dma_fence * fence)613 static void ttm_bo_move_pipeline_evict(struct ttm_buffer_object *bo,
614 struct dma_fence *fence)
615 {
616 struct ttm_device *bdev = bo->bdev;
617 struct ttm_resource_manager *from;
618
619 from = ttm_manager_type(bdev, bo->resource->mem_type);
620
621 /**
622 * BO doesn't have a TTM we need to bind/unbind. Just remember
623 * this eviction and free up the allocation
624 */
625 spin_lock(&from->move_lock);
626 if (!from->move || dma_fence_is_later(fence, from->move)) {
627 dma_fence_put(from->move);
628 from->move = dma_fence_get(fence);
629 }
630 spin_unlock(&from->move_lock);
631
632 ttm_resource_free(bo, &bo->resource);
633 }
634
635 /**
636 * ttm_bo_move_accel_cleanup - cleanup helper for hw copies
637 *
638 * @bo: A pointer to a struct ttm_buffer_object.
639 * @fence: A fence object that signals when moving is complete.
640 * @evict: This is an evict move. Don't return until the buffer is idle.
641 * @pipeline: evictions are to be pipelined.
642 * @new_mem: struct ttm_resource indicating where to move.
643 *
644 * Accelerated move function to be called when an accelerated move
645 * has been scheduled. The function will create a new temporary buffer object
646 * representing the old placement, and put the sync object on both buffer
647 * objects. After that the newly created buffer object is unref'd to be
648 * destroyed when the move is complete. This will help pipeline
649 * buffer moves.
650 */
ttm_bo_move_accel_cleanup(struct ttm_buffer_object * bo,struct dma_fence * fence,bool evict,bool pipeline,struct ttm_resource * new_mem)651 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
652 struct dma_fence *fence,
653 bool evict,
654 bool pipeline,
655 struct ttm_resource *new_mem)
656 {
657 struct ttm_device *bdev = bo->bdev;
658 struct ttm_resource_manager *from = ttm_manager_type(bdev, bo->resource->mem_type);
659 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
660 int ret = 0;
661
662 dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
663 if (!evict)
664 ret = ttm_bo_move_to_ghost(bo, fence, man->use_tt);
665 else if (!from->use_tt && pipeline)
666 ttm_bo_move_pipeline_evict(bo, fence);
667 else
668 ret = ttm_bo_wait_free_node(bo, man->use_tt);
669
670 if (ret)
671 return ret;
672
673 ttm_bo_assign_mem(bo, new_mem);
674
675 return 0;
676 }
677 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
678
679 /**
680 * ttm_bo_move_sync_cleanup - cleanup by waiting for the move to finish
681 *
682 * @bo: A pointer to a struct ttm_buffer_object.
683 * @new_mem: struct ttm_resource indicating where to move.
684 *
685 * Special case of ttm_bo_move_accel_cleanup where the bo is guaranteed
686 * by the caller to be idle. Typically used after memcpy buffer moves.
687 */
ttm_bo_move_sync_cleanup(struct ttm_buffer_object * bo,struct ttm_resource * new_mem)688 void ttm_bo_move_sync_cleanup(struct ttm_buffer_object *bo,
689 struct ttm_resource *new_mem)
690 {
691 struct ttm_device *bdev = bo->bdev;
692 struct ttm_resource_manager *man = ttm_manager_type(bdev, new_mem->mem_type);
693 int ret;
694
695 ret = ttm_bo_wait_free_node(bo, man->use_tt);
696 if (WARN_ON(ret))
697 return;
698
699 ttm_bo_assign_mem(bo, new_mem);
700 }
701 EXPORT_SYMBOL(ttm_bo_move_sync_cleanup);
702
703 /**
704 * ttm_bo_pipeline_gutting - purge the contents of a bo
705 * @bo: The buffer object
706 *
707 * Purge the contents of a bo, async if the bo is not idle.
708 * After a successful call, the bo is left unpopulated in
709 * system placement. The function may wait uninterruptible
710 * for idle on OOM.
711 *
712 * Return: 0 if successful, negative error code on failure.
713 */
ttm_bo_pipeline_gutting(struct ttm_buffer_object * bo)714 int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
715 {
716 struct ttm_buffer_object *ghost;
717 struct ttm_tt *ttm;
718 int ret;
719
720 /* If already idle, no need for ghost object dance. */
721 if (dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP)) {
722 if (!bo->ttm) {
723 /* See comment below about clearing. */
724 ret = ttm_tt_create(bo, true);
725 if (ret)
726 return ret;
727 } else {
728 ttm_tt_unpopulate(bo->bdev, bo->ttm);
729 if (bo->type == ttm_bo_type_device)
730 ttm_tt_mark_for_clear(bo->ttm);
731 }
732 ttm_resource_free(bo, &bo->resource);
733 return 0;
734 }
735
736 /*
737 * We need an unpopulated ttm_tt after giving our current one,
738 * if any, to the ghost object. And we can't afford to fail
739 * creating one *after* the operation. If the bo subsequently gets
740 * resurrected, make sure it's cleared (if ttm_bo_type_device)
741 * to avoid leaking sensitive information to user-space.
742 */
743
744 ttm = bo->ttm;
745 bo->ttm = NULL;
746 ret = ttm_tt_create(bo, true);
747 swap(bo->ttm, ttm);
748 if (ret)
749 return ret;
750
751 ret = ttm_buffer_object_transfer(bo, &ghost);
752 if (ret)
753 goto error_destroy_tt;
754
755 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
756 /* Last resort, wait for the BO to be idle when we are OOM */
757 if (ret) {
758 dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
759 false, MAX_SCHEDULE_TIMEOUT);
760 }
761
762 dma_resv_unlock(&ghost->base._resv);
763 ttm_bo_put(ghost);
764 bo->ttm = ttm;
765 return 0;
766
767 error_destroy_tt:
768 ttm_tt_destroy(bo->bdev, ttm);
769 return ret;
770 }
771
ttm_lru_walk_trylock(struct ttm_operation_ctx * ctx,struct ttm_buffer_object * bo,bool * needs_unlock)772 static bool ttm_lru_walk_trylock(struct ttm_operation_ctx *ctx,
773 struct ttm_buffer_object *bo,
774 bool *needs_unlock)
775 {
776 *needs_unlock = false;
777
778 if (dma_resv_trylock(bo->base.resv)) {
779 *needs_unlock = true;
780 return true;
781 }
782
783 if (bo->base.resv == ctx->resv && ctx->allow_res_evict) {
784 dma_resv_assert_held(bo->base.resv);
785 return true;
786 }
787
788 return false;
789 }
790
ttm_lru_walk_ticketlock(struct ttm_lru_walk * walk,struct ttm_buffer_object * bo,bool * needs_unlock)791 static int ttm_lru_walk_ticketlock(struct ttm_lru_walk *walk,
792 struct ttm_buffer_object *bo,
793 bool *needs_unlock)
794 {
795 struct dma_resv *resv = bo->base.resv;
796 int ret;
797
798 if (walk->ctx->interruptible)
799 ret = dma_resv_lock_interruptible(resv, walk->ticket);
800 else
801 ret = dma_resv_lock(resv, walk->ticket);
802
803 if (!ret) {
804 *needs_unlock = true;
805 /*
806 * Only a single ticketlock per loop. Ticketlocks are prone
807 * to return -EDEADLK causing the eviction to fail, so
808 * after waiting for the ticketlock, revert back to
809 * trylocking for this walk.
810 */
811 walk->ticket = NULL;
812 } else if (ret == -EDEADLK) {
813 /* Caller needs to exit the ww transaction. */
814 ret = -ENOSPC;
815 }
816
817 return ret;
818 }
819
ttm_lru_walk_unlock(struct ttm_buffer_object * bo,bool locked)820 static void ttm_lru_walk_unlock(struct ttm_buffer_object *bo, bool locked)
821 {
822 if (locked)
823 dma_resv_unlock(bo->base.resv);
824 }
825
826 /**
827 * ttm_lru_walk_for_evict() - Perform a LRU list walk, with actions taken on
828 * valid items.
829 * @walk: describe the walks and actions taken
830 * @bdev: The TTM device.
831 * @man: The struct ttm_resource manager whose LRU lists we're walking.
832 * @target: The end condition for the walk.
833 *
834 * The LRU lists of @man are walk, and for each struct ttm_resource encountered,
835 * the corresponding ttm_buffer_object is locked and taken a reference on, and
836 * the LRU lock is dropped. the LRU lock may be dropped before locking and, in
837 * that case, it's verified that the item actually remains on the LRU list after
838 * the lock, and that the buffer object didn't switch resource in between.
839 *
840 * With a locked object, the actions indicated by @walk->process_bo are
841 * performed, and after that, the bo is unlocked, the refcount dropped and the
842 * next struct ttm_resource is processed. Here, the walker relies on
843 * TTM's restartable LRU list implementation.
844 *
845 * Typically @walk->process_bo() would return the number of pages evicted,
846 * swapped or shrunken, so that when the total exceeds @target, or when the
847 * LRU list has been walked in full, iteration is terminated. It's also terminated
848 * on error. Note that the definition of @target is done by the caller, it
849 * could have a different meaning than the number of pages.
850 *
851 * Note that the way dma_resv individualization is done, locking needs to be done
852 * either with the LRU lock held (trylocking only) or with a reference on the
853 * object.
854 *
855 * Return: The progress made towards target or negative error code on error.
856 */
ttm_lru_walk_for_evict(struct ttm_lru_walk * walk,struct ttm_device * bdev,struct ttm_resource_manager * man,s64 target)857 s64 ttm_lru_walk_for_evict(struct ttm_lru_walk *walk, struct ttm_device *bdev,
858 struct ttm_resource_manager *man, s64 target)
859 {
860 struct ttm_resource_cursor cursor;
861 struct ttm_resource *res;
862 s64 progress = 0;
863 s64 lret;
864
865 spin_lock(&bdev->lru_lock);
866 ttm_resource_cursor_init(&cursor, man);
867 ttm_resource_manager_for_each_res(&cursor, res) {
868 struct ttm_buffer_object *bo = res->bo;
869 bool bo_needs_unlock = false;
870 bool bo_locked = false;
871 int mem_type;
872
873 /*
874 * Attempt a trylock before taking a reference on the bo,
875 * since if we do it the other way around, and the trylock fails,
876 * we need to drop the lru lock to put the bo.
877 */
878 if (ttm_lru_walk_trylock(walk->ctx, bo, &bo_needs_unlock))
879 bo_locked = true;
880 else if (!walk->ticket || walk->ctx->no_wait_gpu ||
881 walk->trylock_only)
882 continue;
883
884 if (!ttm_bo_get_unless_zero(bo)) {
885 ttm_lru_walk_unlock(bo, bo_needs_unlock);
886 continue;
887 }
888
889 mem_type = res->mem_type;
890 spin_unlock(&bdev->lru_lock);
891
892 lret = 0;
893 if (!bo_locked)
894 lret = ttm_lru_walk_ticketlock(walk, bo, &bo_needs_unlock);
895
896 /*
897 * Note that in between the release of the lru lock and the
898 * ticketlock, the bo may have switched resource,
899 * and also memory type, since the resource may have been
900 * freed and allocated again with a different memory type.
901 * In that case, just skip it.
902 */
903 if (!lret && bo->resource && bo->resource->mem_type == mem_type)
904 lret = walk->ops->process_bo(walk, bo);
905
906 ttm_lru_walk_unlock(bo, bo_needs_unlock);
907 ttm_bo_put(bo);
908 if (lret == -EBUSY || lret == -EALREADY)
909 lret = 0;
910 progress = (lret < 0) ? lret : progress + lret;
911
912 spin_lock(&bdev->lru_lock);
913 if (progress < 0 || progress >= target)
914 break;
915 }
916 ttm_resource_cursor_fini(&cursor);
917 spin_unlock(&bdev->lru_lock);
918
919 return progress;
920 }
921 EXPORT_SYMBOL(ttm_lru_walk_for_evict);
922
ttm_bo_lru_cursor_cleanup_bo(struct ttm_bo_lru_cursor * curs)923 static void ttm_bo_lru_cursor_cleanup_bo(struct ttm_bo_lru_cursor *curs)
924 {
925 struct ttm_buffer_object *bo = curs->bo;
926
927 if (bo) {
928 if (curs->needs_unlock)
929 dma_resv_unlock(bo->base.resv);
930 ttm_bo_put(bo);
931 curs->bo = NULL;
932 }
933 }
934
935 /**
936 * ttm_bo_lru_cursor_fini() - Stop using a struct ttm_bo_lru_cursor
937 * and clean up any iteration it was used for.
938 * @curs: The cursor.
939 */
ttm_bo_lru_cursor_fini(struct ttm_bo_lru_cursor * curs)940 void ttm_bo_lru_cursor_fini(struct ttm_bo_lru_cursor *curs)
941 {
942 spinlock_t *lru_lock = &curs->res_curs.man->bdev->lru_lock;
943
944 ttm_bo_lru_cursor_cleanup_bo(curs);
945 spin_lock(lru_lock);
946 ttm_resource_cursor_fini(&curs->res_curs);
947 spin_unlock(lru_lock);
948 }
949 EXPORT_SYMBOL(ttm_bo_lru_cursor_fini);
950
951 /**
952 * ttm_bo_lru_cursor_init() - Initialize a struct ttm_bo_lru_cursor
953 * @curs: The ttm_bo_lru_cursor to initialize.
954 * @man: The ttm resource_manager whose LRU lists to iterate over.
955 * @ctx: The ttm_operation_ctx to govern the locking.
956 *
957 * Initialize a struct ttm_bo_lru_cursor. Currently only trylocking
958 * or prelocked buffer objects are available as detailed by
959 * @ctx::resv and @ctx::allow_res_evict. Ticketlocking is not
960 * supported.
961 *
962 * Return: Pointer to @curs. The function does not fail.
963 */
964 struct ttm_bo_lru_cursor *
ttm_bo_lru_cursor_init(struct ttm_bo_lru_cursor * curs,struct ttm_resource_manager * man,struct ttm_operation_ctx * ctx)965 ttm_bo_lru_cursor_init(struct ttm_bo_lru_cursor *curs,
966 struct ttm_resource_manager *man,
967 struct ttm_operation_ctx *ctx)
968 {
969 memset(curs, 0, sizeof(*curs));
970 ttm_resource_cursor_init(&curs->res_curs, man);
971 curs->ctx = ctx;
972
973 return curs;
974 }
975 EXPORT_SYMBOL(ttm_bo_lru_cursor_init);
976
977 static struct ttm_buffer_object *
ttm_bo_from_res_reserved(struct ttm_resource * res,struct ttm_bo_lru_cursor * curs)978 ttm_bo_from_res_reserved(struct ttm_resource *res, struct ttm_bo_lru_cursor *curs)
979 {
980 struct ttm_buffer_object *bo = res->bo;
981
982 if (!ttm_lru_walk_trylock(curs->ctx, bo, &curs->needs_unlock))
983 return NULL;
984
985 if (!ttm_bo_get_unless_zero(bo)) {
986 if (curs->needs_unlock)
987 dma_resv_unlock(bo->base.resv);
988 return NULL;
989 }
990
991 curs->bo = bo;
992 return bo;
993 }
994
995 /**
996 * ttm_bo_lru_cursor_next() - Continue iterating a manager's LRU lists
997 * to find and lock buffer object.
998 * @curs: The cursor initialized using ttm_bo_lru_cursor_init() and
999 * ttm_bo_lru_cursor_first().
1000 *
1001 * Return: A pointer to a locked and reference-counted buffer object,
1002 * or NULL if none could be found and looping should be terminated.
1003 */
ttm_bo_lru_cursor_next(struct ttm_bo_lru_cursor * curs)1004 struct ttm_buffer_object *ttm_bo_lru_cursor_next(struct ttm_bo_lru_cursor *curs)
1005 {
1006 spinlock_t *lru_lock = &curs->res_curs.man->bdev->lru_lock;
1007 struct ttm_resource *res = NULL;
1008 struct ttm_buffer_object *bo;
1009
1010 ttm_bo_lru_cursor_cleanup_bo(curs);
1011
1012 spin_lock(lru_lock);
1013 for (;;) {
1014 res = ttm_resource_manager_next(&curs->res_curs);
1015 if (!res)
1016 break;
1017
1018 bo = ttm_bo_from_res_reserved(res, curs);
1019 if (bo)
1020 break;
1021 }
1022
1023 spin_unlock(lru_lock);
1024 return res ? bo : NULL;
1025 }
1026 EXPORT_SYMBOL(ttm_bo_lru_cursor_next);
1027
1028 /**
1029 * ttm_bo_lru_cursor_first() - Start iterating a manager's LRU lists
1030 * to find and lock buffer object.
1031 * @curs: The cursor initialized using ttm_bo_lru_cursor_init().
1032 *
1033 * Return: A pointer to a locked and reference-counted buffer object,
1034 * or NULL if none could be found and looping should be terminated.
1035 */
ttm_bo_lru_cursor_first(struct ttm_bo_lru_cursor * curs)1036 struct ttm_buffer_object *ttm_bo_lru_cursor_first(struct ttm_bo_lru_cursor *curs)
1037 {
1038 spinlock_t *lru_lock = &curs->res_curs.man->bdev->lru_lock;
1039 struct ttm_buffer_object *bo;
1040 struct ttm_resource *res;
1041
1042 spin_lock(lru_lock);
1043 res = ttm_resource_manager_first(&curs->res_curs);
1044 if (!res) {
1045 spin_unlock(lru_lock);
1046 return NULL;
1047 }
1048
1049 bo = ttm_bo_from_res_reserved(res, curs);
1050 spin_unlock(lru_lock);
1051
1052 return bo ? bo : ttm_bo_lru_cursor_next(curs);
1053 }
1054 EXPORT_SYMBOL(ttm_bo_lru_cursor_first);
1055
1056 /**
1057 * ttm_bo_shrink() - Helper to shrink a ttm buffer object.
1058 * @ctx: The struct ttm_operation_ctx used for the shrinking operation.
1059 * @bo: The buffer object.
1060 * @flags: Flags governing the shrinking behaviour.
1061 *
1062 * The function uses the ttm_tt_back_up functionality to back up or
1063 * purge a struct ttm_tt. If the bo is not in system, it's first
1064 * moved there.
1065 *
1066 * Return: The number of pages shrunken or purged, or
1067 * negative error code on failure.
1068 */
ttm_bo_shrink(struct ttm_operation_ctx * ctx,struct ttm_buffer_object * bo,const struct ttm_bo_shrink_flags flags)1069 long ttm_bo_shrink(struct ttm_operation_ctx *ctx, struct ttm_buffer_object *bo,
1070 const struct ttm_bo_shrink_flags flags)
1071 {
1072 static const struct ttm_place sys_placement_flags = {
1073 .fpfn = 0,
1074 .lpfn = 0,
1075 .mem_type = TTM_PL_SYSTEM,
1076 .flags = 0,
1077 };
1078 static struct ttm_placement sys_placement = {
1079 .num_placement = 1,
1080 .placement = &sys_placement_flags,
1081 };
1082 struct ttm_tt *tt = bo->ttm;
1083 long lret;
1084
1085 dma_resv_assert_held(bo->base.resv);
1086
1087 if (flags.allow_move && bo->resource->mem_type != TTM_PL_SYSTEM) {
1088 int ret = ttm_bo_validate(bo, &sys_placement, ctx);
1089
1090 /* Consider -ENOMEM and -ENOSPC non-fatal. */
1091 if (ret) {
1092 if (ret == -ENOMEM || ret == -ENOSPC)
1093 ret = -EBUSY;
1094 return ret;
1095 }
1096 }
1097
1098 ttm_bo_unmap_virtual(bo);
1099 lret = ttm_bo_wait_ctx(bo, ctx);
1100 if (lret < 0)
1101 return lret;
1102
1103 if (bo->bulk_move) {
1104 spin_lock(&bo->bdev->lru_lock);
1105 ttm_resource_del_bulk_move(bo->resource, bo);
1106 spin_unlock(&bo->bdev->lru_lock);
1107 }
1108
1109 lret = ttm_tt_backup(bo->bdev, tt, (struct ttm_backup_flags)
1110 {.purge = flags.purge,
1111 .writeback = flags.writeback});
1112
1113 if (lret <= 0 && bo->bulk_move) {
1114 spin_lock(&bo->bdev->lru_lock);
1115 ttm_resource_add_bulk_move(bo->resource, bo);
1116 spin_unlock(&bo->bdev->lru_lock);
1117 }
1118
1119 if (lret < 0 && lret != -EINTR)
1120 return -EBUSY;
1121
1122 return lret;
1123 }
1124 EXPORT_SYMBOL(ttm_bo_shrink);
1125
1126 /**
1127 * ttm_bo_shrink_suitable() - Whether a bo is suitable for shinking
1128 * @ctx: The struct ttm_operation_ctx governing the shrinking.
1129 * @bo: The candidate for shrinking.
1130 *
1131 * Check whether the object, given the information available to TTM,
1132 * is suitable for shinking, This function can and should be used
1133 * before attempting to shrink an object.
1134 *
1135 * Return: true if suitable. false if not.
1136 */
ttm_bo_shrink_suitable(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)1137 bool ttm_bo_shrink_suitable(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx)
1138 {
1139 return bo->ttm && ttm_tt_is_populated(bo->ttm) && !bo->pin_count &&
1140 (!ctx->no_wait_gpu ||
1141 dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP));
1142 }
1143 EXPORT_SYMBOL(ttm_bo_shrink_suitable);
1144
1145 /**
1146 * ttm_bo_shrink_avoid_wait() - Whether to avoid waiting for GPU
1147 * during shrinking
1148 *
1149 * In some situations, like direct reclaim, waiting (in particular gpu waiting)
1150 * should be avoided since it may stall a system that could otherwise make progress
1151 * shrinking something else less time consuming.
1152 *
1153 * Return: true if gpu waiting should be avoided, false if not.
1154 */
ttm_bo_shrink_avoid_wait(void)1155 bool ttm_bo_shrink_avoid_wait(void)
1156 {
1157 return !current_is_kswapd();
1158 }
1159 EXPORT_SYMBOL(ttm_bo_shrink_avoid_wait);
1160