]> git.ipfire.org Git - thirdparty/linux.git/blob - drivers/gpu/drm/ttm/ttm_bo.c
projects / thirdparty / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branches 'clk-doc', 'clk-more-critical', 'clk-meson' and 'clk-basic-be' into...
[thirdparty/linux.git] / drivers / gpu / drm / ttm / ttm_bo.c
1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
3 *
4 * Copyright (c) 2006-2009 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
32 #define pr_fmt(fmt) "[TTM] " fmt
33
34 #include <drm/ttm/ttm_module.h>
35 #include <drm/ttm/ttm_bo_driver.h>
36 #include <drm/ttm/ttm_placement.h>
37 #include <linux/jiffies.h>
38 #include <linux/slab.h>
39 #include <linux/sched.h>
40 #include <linux/mm.h>
41 #include <linux/file.h>
42 #include <linux/module.h>
43 #include <linux/atomic.h>
44 #include <linux/reservation.h>
45
46 static void ttm_bo_global_kobj_release(struct kobject *kobj);
47
48 /**
49 * ttm_global_mutex - protecting the global BO state
50 */
51 DEFINE_MUTEX(ttm_global_mutex);
52 unsigned ttm_bo_glob_use_count;
53 struct ttm_bo_global ttm_bo_glob;
54
55 static struct attribute ttm_bo_count = {
56 .name = "bo_count",
57 .mode = S_IRUGO
58 };
59
60 /* default destructor */
61 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
62 {
63 kfree(bo);
64 }
65
66 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
67 uint32_t *mem_type)
68 {
69 int pos;
70
71 pos = ffs(place->flags & TTM_PL_MASK_MEM);
72 if (unlikely(!pos))
73 return -EINVAL;
74
75 *mem_type = pos - 1;
76 return 0;
77 }
78
79 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, struct drm_printer *p,
80 int mem_type)
81 {
82 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
83
84 drm_printf(p, " has_type: %d\n", man->has_type);
85 drm_printf(p, " use_type: %d\n", man->use_type);
86 drm_printf(p, " flags: 0x%08X\n", man->flags);
87 drm_printf(p, " gpu_offset: 0x%08llX\n", man->gpu_offset);
88 drm_printf(p, " size: %llu\n", man->size);
89 drm_printf(p, " available_caching: 0x%08X\n", man->available_caching);
90 drm_printf(p, " default_caching: 0x%08X\n", man->default_caching);
91 if (mem_type != TTM_PL_SYSTEM)
92 (*man->func->debug)(man, p);
93 }
94
95 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
96 struct ttm_placement *placement)
97 {
98 struct drm_printer p = drm_debug_printer(TTM_PFX);
99 int i, ret, mem_type;
100
101 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n",
102 bo, bo->mem.num_pages, bo->mem.size >> 10,
103 bo->mem.size >> 20);
104 for (i = 0; i < placement->num_placement; i++) {
105 ret = ttm_mem_type_from_place(&placement->placement[i],
106 &mem_type);
107 if (ret)
108 return;
109 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
110 i, placement->placement[i].flags, mem_type);
111 ttm_mem_type_debug(bo->bdev, &p, mem_type);
112 }
113 }
114
115 static ssize_t ttm_bo_global_show(struct kobject *kobj,
116 struct attribute *attr,
117 char *buffer)
118 {
119 struct ttm_bo_global *glob =
120 container_of(kobj, struct ttm_bo_global, kobj);
121
122 return snprintf(buffer, PAGE_SIZE, "%d\n",
123 atomic_read(&glob->bo_count));
124 }
125
126 static struct attribute *ttm_bo_global_attrs[] = {
127 &ttm_bo_count,
128 NULL
129 };
130
131 static const struct sysfs_ops ttm_bo_global_ops = {
132 .show = &ttm_bo_global_show
133 };
134
135 static struct kobj_type ttm_bo_glob_kobj_type = {
136 .release = &ttm_bo_global_kobj_release,
137 .sysfs_ops = &ttm_bo_global_ops,
138 .default_attrs = ttm_bo_global_attrs
139 };
140
141
142 static inline uint32_t ttm_bo_type_flags(unsigned type)
143 {
144 return 1 << (type);
145 }
146
147 static void ttm_bo_release_list(struct kref *list_kref)
148 {
149 struct ttm_buffer_object *bo =
150 container_of(list_kref, struct ttm_buffer_object, list_kref);
151 struct ttm_bo_device *bdev = bo->bdev;
152 size_t acc_size = bo->acc_size;
153
154 BUG_ON(kref_read(&bo->list_kref));
155 BUG_ON(kref_read(&bo->kref));
156 BUG_ON(atomic_read(&bo->cpu_writers));
157 BUG_ON(bo->mem.mm_node != NULL);
158 BUG_ON(!list_empty(&bo->lru));
159 BUG_ON(!list_empty(&bo->ddestroy));
160 ttm_tt_destroy(bo->ttm);
161 atomic_dec(&bo->bdev->glob->bo_count);
162 dma_fence_put(bo->moving);
163 reservation_object_fini(&bo->ttm_resv);
164 mutex_destroy(&bo->wu_mutex);
165 bo->destroy(bo);
166 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
167 }
168
169 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
170 {
171 struct ttm_bo_device *bdev = bo->bdev;
172 struct ttm_mem_type_manager *man;
173
174 reservation_object_assert_held(bo->resv);
175
176 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
177 BUG_ON(!list_empty(&bo->lru));
178
179 man = &bdev->man[bo->mem.mem_type];
180 list_add_tail(&bo->lru, &man->lru[bo->priority]);
181 kref_get(&bo->list_kref);
182
183 if (bo->ttm && !(bo->ttm->page_flags &
184 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) {
185 list_add_tail(&bo->swap,
186 &bdev->glob->swap_lru[bo->priority]);
187 kref_get(&bo->list_kref);
188 }
189 }
190 }
191 EXPORT_SYMBOL(ttm_bo_add_to_lru);
192
193 static void ttm_bo_ref_bug(struct kref *list_kref)
194 {
195 BUG();
196 }
197
198 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
199 {
200 struct ttm_bo_device *bdev = bo->bdev;
201 bool notify = false;
202
203 if (!list_empty(&bo->swap)) {
204 list_del_init(&bo->swap);
205 kref_put(&bo->list_kref, ttm_bo_ref_bug);
206 notify = true;
207 }
208 if (!list_empty(&bo->lru)) {
209 list_del_init(&bo->lru);
210 kref_put(&bo->list_kref, ttm_bo_ref_bug);
211 notify = true;
212 }
213
214 if (notify && bdev->driver->del_from_lru_notify)
215 bdev->driver->del_from_lru_notify(bo);
216 }
217
218 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
219 {
220 struct ttm_bo_global *glob = bo->bdev->glob;
221
222 spin_lock(&glob->lru_lock);
223 ttm_bo_del_from_lru(bo);
224 spin_unlock(&glob->lru_lock);
225 }
226 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
227
228 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
229 struct ttm_buffer_object *bo)
230 {
231 if (!pos->first)
232 pos->first = bo;
233 pos->last = bo;
234 }
235
236 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
237 struct ttm_lru_bulk_move *bulk)
238 {
239 reservation_object_assert_held(bo->resv);
240
241 ttm_bo_del_from_lru(bo);
242 ttm_bo_add_to_lru(bo);
243
244 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
245 switch (bo->mem.mem_type) {
246 case TTM_PL_TT:
247 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
248 break;
249
250 case TTM_PL_VRAM:
251 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
252 break;
253 }
254 if (bo->ttm && !(bo->ttm->page_flags &
255 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
256 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
257 }
258 }
259 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
260
261 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
262 {
263 unsigned i;
264
265 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
266 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
267 struct ttm_mem_type_manager *man;
268
269 if (!pos->first)
270 continue;
271
272 reservation_object_assert_held(pos->first->resv);
273 reservation_object_assert_held(pos->last->resv);
274
275 man = &pos->first->bdev->man[TTM_PL_TT];
276 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
277 &pos->last->lru);
278 }
279
280 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
281 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
282 struct ttm_mem_type_manager *man;
283
284 if (!pos->first)
285 continue;
286
287 reservation_object_assert_held(pos->first->resv);
288 reservation_object_assert_held(pos->last->resv);
289
290 man = &pos->first->bdev->man[TTM_PL_VRAM];
291 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
292 &pos->last->lru);
293 }
294
295 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
296 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
297 struct list_head *lru;
298
299 if (!pos->first)
300 continue;
301
302 reservation_object_assert_held(pos->first->resv);
303 reservation_object_assert_held(pos->last->resv);
304
305 lru = &pos->first->bdev->glob->swap_lru[i];
306 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
307 }
308 }
309 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
310
311 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
312 struct ttm_mem_reg *mem, bool evict,
313 struct ttm_operation_ctx *ctx)
314 {
315 struct ttm_bo_device *bdev = bo->bdev;
316 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
317 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
318 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
319 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
320 int ret = 0;
321
322 if (old_is_pci || new_is_pci ||
323 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
324 ret = ttm_mem_io_lock(old_man, true);
325 if (unlikely(ret != 0))
326 goto out_err;
327 ttm_bo_unmap_virtual_locked(bo);
328 ttm_mem_io_unlock(old_man);
329 }
330
331 /*
332 * Create and bind a ttm if required.
333 */
334
335 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
336 if (bo->ttm == NULL) {
337 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
338 ret = ttm_tt_create(bo, zero);
339 if (ret)
340 goto out_err;
341 }
342
343 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
344 if (ret)
345 goto out_err;
346
347 if (mem->mem_type != TTM_PL_SYSTEM) {
348 ret = ttm_tt_bind(bo->ttm, mem, ctx);
349 if (ret)
350 goto out_err;
351 }
352
353 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
354 if (bdev->driver->move_notify)
355 bdev->driver->move_notify(bo, evict, mem);
356 bo->mem = *mem;
357 mem->mm_node = NULL;
358 goto moved;
359 }
360 }
361
362 if (bdev->driver->move_notify)
363 bdev->driver->move_notify(bo, evict, mem);
364
365 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
366 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
367 ret = ttm_bo_move_ttm(bo, ctx, mem);
368 else if (bdev->driver->move)
369 ret = bdev->driver->move(bo, evict, ctx, mem);
370 else
371 ret = ttm_bo_move_memcpy(bo, ctx, mem);
372
373 if (ret) {
374 if (bdev->driver->move_notify) {
375 swap(*mem, bo->mem);
376 bdev->driver->move_notify(bo, false, mem);
377 swap(*mem, bo->mem);
378 }
379
380 goto out_err;
381 }
382
383 moved:
384 if (bo->evicted) {
385 if (bdev->driver->invalidate_caches) {
386 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
387 if (ret)
388 pr_err("Can not flush read caches\n");
389 }
390 bo->evicted = false;
391 }
392
393 if (bo->mem.mm_node)
394 bo->offset = (bo->mem.start << PAGE_SHIFT) +
395 bdev->man[bo->mem.mem_type].gpu_offset;
396 else
397 bo->offset = 0;
398
399 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
400 return 0;
401
402 out_err:
403 new_man = &bdev->man[bo->mem.mem_type];
404 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
405 ttm_tt_destroy(bo->ttm);
406 bo->ttm = NULL;
407 }
408
409 return ret;
410 }
411
412 /**
413 * Call bo::reserved.
414 * Will release GPU memory type usage on destruction.
415 * This is the place to put in driver specific hooks to release
416 * driver private resources.
417 * Will release the bo::reserved lock.
418 */
419
420 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
421 {
422 if (bo->bdev->driver->move_notify)
423 bo->bdev->driver->move_notify(bo, false, NULL);
424
425 ttm_tt_destroy(bo->ttm);
426 bo->ttm = NULL;
427 ttm_bo_mem_put(bo, &bo->mem);
428 }
429
430 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
431 {
432 int r;
433
434 if (bo->resv == &bo->ttm_resv)
435 return 0;
436
437 BUG_ON(!reservation_object_trylock(&bo->ttm_resv));
438
439 r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv);
440 if (r)
441 reservation_object_unlock(&bo->ttm_resv);
442
443 return r;
444 }
445
446 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
447 {
448 struct reservation_object_list *fobj;
449 struct dma_fence *fence;
450 int i;
451
452 fobj = reservation_object_get_list(&bo->ttm_resv);
453 fence = reservation_object_get_excl(&bo->ttm_resv);
454 if (fence && !fence->ops->signaled)
455 dma_fence_enable_sw_signaling(fence);
456
457 for (i = 0; fobj && i < fobj->shared_count; ++i) {
458 fence = rcu_dereference_protected(fobj->shared[i],
459 reservation_object_held(bo->resv));
460
461 if (!fence->ops->signaled)
462 dma_fence_enable_sw_signaling(fence);
463 }
464 }
465
466 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
467 {
468 struct ttm_bo_device *bdev = bo->bdev;
469 struct ttm_bo_global *glob = bdev->glob;
470 int ret;
471
472 ret = ttm_bo_individualize_resv(bo);
473 if (ret) {
474 /* Last resort, if we fail to allocate memory for the
475 * fences block for the BO to become idle
476 */
477 reservation_object_wait_timeout_rcu(bo->resv, true, false,
478 30 * HZ);
479 spin_lock(&glob->lru_lock);
480 goto error;
481 }
482
483 spin_lock(&glob->lru_lock);
484 ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY;
485 if (!ret) {
486 if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) {
487 ttm_bo_del_from_lru(bo);
488 spin_unlock(&glob->lru_lock);
489 if (bo->resv != &bo->ttm_resv)
490 reservation_object_unlock(&bo->ttm_resv);
491
492 ttm_bo_cleanup_memtype_use(bo);
493 reservation_object_unlock(bo->resv);
494 return;
495 }
496
497 ttm_bo_flush_all_fences(bo);
498
499 /*
500 * Make NO_EVICT bos immediately available to
501 * shrinkers, now that they are queued for
502 * destruction.
503 */
504 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
505 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
506 ttm_bo_add_to_lru(bo);
507 }
508
509 reservation_object_unlock(bo->resv);
510 }
511 if (bo->resv != &bo->ttm_resv)
512 reservation_object_unlock(&bo->ttm_resv);
513
514 error:
515 kref_get(&bo->list_kref);
516 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
517 spin_unlock(&glob->lru_lock);
518
519 schedule_delayed_work(&bdev->wq,
520 ((HZ / 100) < 1) ? 1 : HZ / 100);
521 }
522
523 /**
524 * function ttm_bo_cleanup_refs
525 * If bo idle, remove from delayed- and lru lists, and unref.
526 * If not idle, do nothing.
527 *
528 * Must be called with lru_lock and reservation held, this function
529 * will drop the lru lock and optionally the reservation lock before returning.
530 *
531 * @interruptible Any sleeps should occur interruptibly.
532 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
533 * @unlock_resv Unlock the reservation lock as well.
534 */
535
536 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
537 bool interruptible, bool no_wait_gpu,
538 bool unlock_resv)
539 {
540 struct ttm_bo_global *glob = bo->bdev->glob;
541 struct reservation_object *resv;
542 int ret;
543
544 if (unlikely(list_empty(&bo->ddestroy)))
545 resv = bo->resv;
546 else
547 resv = &bo->ttm_resv;
548
549 if (reservation_object_test_signaled_rcu(resv, true))
550 ret = 0;
551 else
552 ret = -EBUSY;
553
554 if (ret && !no_wait_gpu) {
555 long lret;
556
557 if (unlock_resv)
558 reservation_object_unlock(bo->resv);
559 spin_unlock(&glob->lru_lock);
560
561 lret = reservation_object_wait_timeout_rcu(resv, true,
562 interruptible,
563 30 * HZ);
564
565 if (lret < 0)
566 return lret;
567 else if (lret == 0)
568 return -EBUSY;
569
570 spin_lock(&glob->lru_lock);
571 if (unlock_resv && !reservation_object_trylock(bo->resv)) {
572 /*
573 * We raced, and lost, someone else holds the reservation now,
574 * and is probably busy in ttm_bo_cleanup_memtype_use.
575 *
576 * Even if it's not the case, because we finished waiting any
577 * delayed destruction would succeed, so just return success
578 * here.
579 */
580 spin_unlock(&glob->lru_lock);
581 return 0;
582 }
583 ret = 0;
584 }
585
586 if (ret || unlikely(list_empty(&bo->ddestroy))) {
587 if (unlock_resv)
588 reservation_object_unlock(bo->resv);
589 spin_unlock(&glob->lru_lock);
590 return ret;
591 }
592
593 ttm_bo_del_from_lru(bo);
594 list_del_init(&bo->ddestroy);
595 kref_put(&bo->list_kref, ttm_bo_ref_bug);
596
597 spin_unlock(&glob->lru_lock);
598 ttm_bo_cleanup_memtype_use(bo);
599
600 if (unlock_resv)
601 reservation_object_unlock(bo->resv);
602
603 return 0;
604 }
605
606 /**
607 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
608 * encountered buffers.
609 */
610 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
611 {
612 struct ttm_bo_global *glob = bdev->glob;
613 struct list_head removed;
614 bool empty;
615
616 INIT_LIST_HEAD(&removed);
617
618 spin_lock(&glob->lru_lock);
619 while (!list_empty(&bdev->ddestroy)) {
620 struct ttm_buffer_object *bo;
621
622 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
623 ddestroy);
624 kref_get(&bo->list_kref);
625 list_move_tail(&bo->ddestroy, &removed);
626
627 if (remove_all || bo->resv != &bo->ttm_resv) {
628 spin_unlock(&glob->lru_lock);
629 reservation_object_lock(bo->resv, NULL);
630
631 spin_lock(&glob->lru_lock);
632 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
633
634 } else if (reservation_object_trylock(bo->resv)) {
635 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
636 } else {
637 spin_unlock(&glob->lru_lock);
638 }
639
640 kref_put(&bo->list_kref, ttm_bo_release_list);
641 spin_lock(&glob->lru_lock);
642 }
643 list_splice_tail(&removed, &bdev->ddestroy);
644 empty = list_empty(&bdev->ddestroy);
645 spin_unlock(&glob->lru_lock);
646
647 return empty;
648 }
649
650 static void ttm_bo_delayed_workqueue(struct work_struct *work)
651 {
652 struct ttm_bo_device *bdev =
653 container_of(work, struct ttm_bo_device, wq.work);
654
655 if (!ttm_bo_delayed_delete(bdev, false))
656 schedule_delayed_work(&bdev->wq,
657 ((HZ / 100) < 1) ? 1 : HZ / 100);
658 }
659
660 static void ttm_bo_release(struct kref *kref)
661 {
662 struct ttm_buffer_object *bo =
663 container_of(kref, struct ttm_buffer_object, kref);
664 struct ttm_bo_device *bdev = bo->bdev;
665 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
666
667 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
668 ttm_mem_io_lock(man, false);
669 ttm_mem_io_free_vm(bo);
670 ttm_mem_io_unlock(man);
671 ttm_bo_cleanup_refs_or_queue(bo);
672 kref_put(&bo->list_kref, ttm_bo_release_list);
673 }
674
675 void ttm_bo_put(struct ttm_buffer_object *bo)
676 {
677 kref_put(&bo->kref, ttm_bo_release);
678 }
679 EXPORT_SYMBOL(ttm_bo_put);
680
681 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
682 {
683 return cancel_delayed_work_sync(&bdev->wq);
684 }
685 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
686
687 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
688 {
689 if (resched)
690 schedule_delayed_work(&bdev->wq,
691 ((HZ / 100) < 1) ? 1 : HZ / 100);
692 }
693 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
694
695 static int ttm_bo_evict(struct ttm_buffer_object *bo,
696 struct ttm_operation_ctx *ctx)
697 {
698 struct ttm_bo_device *bdev = bo->bdev;
699 struct ttm_mem_reg evict_mem;
700 struct ttm_placement placement;
701 int ret = 0;
702
703 reservation_object_assert_held(bo->resv);
704
705 placement.num_placement = 0;
706 placement.num_busy_placement = 0;
707 bdev->driver->evict_flags(bo, &placement);
708
709 if (!placement.num_placement && !placement.num_busy_placement) {
710 ret = ttm_bo_pipeline_gutting(bo);
711 if (ret)
712 return ret;
713
714 return ttm_tt_create(bo, false);
715 }
716
717 evict_mem = bo->mem;
718 evict_mem.mm_node = NULL;
719 evict_mem.bus.io_reserved_vm = false;
720 evict_mem.bus.io_reserved_count = 0;
721
722 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
723 if (ret) {
724 if (ret != -ERESTARTSYS) {
725 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
726 bo);
727 ttm_bo_mem_space_debug(bo, &placement);
728 }
729 goto out;
730 }
731
732 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
733 if (unlikely(ret)) {
734 if (ret != -ERESTARTSYS)
735 pr_err("Buffer eviction failed\n");
736 ttm_bo_mem_put(bo, &evict_mem);
737 goto out;
738 }
739 bo->evicted = true;
740 out:
741 return ret;
742 }
743
744 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
745 const struct ttm_place *place)
746 {
747 /* Don't evict this BO if it's outside of the
748 * requested placement range
749 */
750 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
751 (place->lpfn && place->lpfn <= bo->mem.start))
752 return false;
753
754 return true;
755 }
756 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
757
758 /**
759 * Check the target bo is allowable to be evicted or swapout, including cases:
760 *
761 * a. if share same reservation object with ctx->resv, have assumption
762 * reservation objects should already be locked, so not lock again and
763 * return true directly when either the opreation allow_reserved_eviction
764 * or the target bo already is in delayed free list;
765 *
766 * b. Otherwise, trylock it.
767 */
768 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
769 struct ttm_operation_ctx *ctx, bool *locked)
770 {
771 bool ret = false;
772
773 *locked = false;
774 if (bo->resv == ctx->resv) {
775 reservation_object_assert_held(bo->resv);
776 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT
777 || !list_empty(&bo->ddestroy))
778 ret = true;
779 } else {
780 *locked = reservation_object_trylock(bo->resv);
781 ret = *locked;
782 }
783
784 return ret;
785 }
786
787 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
788 uint32_t mem_type,
789 const struct ttm_place *place,
790 struct ttm_operation_ctx *ctx)
791 {
792 struct ttm_bo_global *glob = bdev->glob;
793 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
794 struct ttm_buffer_object *bo = NULL;
795 bool locked = false;
796 unsigned i;
797 int ret;
798
799 spin_lock(&glob->lru_lock);
800 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
801 list_for_each_entry(bo, &man->lru[i], lru) {
802 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked))
803 continue;
804
805 if (place && !bdev->driver->eviction_valuable(bo,
806 place)) {
807 if (locked)
808 reservation_object_unlock(bo->resv);
809 continue;
810 }
811 break;
812 }
813
814 /* If the inner loop terminated early, we have our candidate */
815 if (&bo->lru != &man->lru[i])
816 break;
817
818 bo = NULL;
819 }
820
821 if (!bo) {
822 spin_unlock(&glob->lru_lock);
823 return -EBUSY;
824 }
825
826 kref_get(&bo->list_kref);
827
828 if (!list_empty(&bo->ddestroy)) {
829 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
830 ctx->no_wait_gpu, locked);
831 kref_put(&bo->list_kref, ttm_bo_release_list);
832 return ret;
833 }
834
835 ttm_bo_del_from_lru(bo);
836 spin_unlock(&glob->lru_lock);
837
838 ret = ttm_bo_evict(bo, ctx);
839 if (locked) {
840 ttm_bo_unreserve(bo);
841 } else {
842 spin_lock(&glob->lru_lock);
843 ttm_bo_add_to_lru(bo);
844 spin_unlock(&glob->lru_lock);
845 }
846
847 kref_put(&bo->list_kref, ttm_bo_release_list);
848 return ret;
849 }
850
851 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
852 {
853 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
854
855 if (mem->mm_node)
856 (*man->func->put_node)(man, mem);
857 }
858 EXPORT_SYMBOL(ttm_bo_mem_put);
859
860 /**
861 * Add the last move fence to the BO and reserve a new shared slot.
862 */
863 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
864 struct ttm_mem_type_manager *man,
865 struct ttm_mem_reg *mem)
866 {
867 struct dma_fence *fence;
868 int ret;
869
870 spin_lock(&man->move_lock);
871 fence = dma_fence_get(man->move);
872 spin_unlock(&man->move_lock);
873
874 if (fence) {
875 reservation_object_add_shared_fence(bo->resv, fence);
876
877 ret = reservation_object_reserve_shared(bo->resv, 1);
878 if (unlikely(ret)) {
879 dma_fence_put(fence);
880 return ret;
881 }
882
883 dma_fence_put(bo->moving);
884 bo->moving = fence;
885 }
886
887 return 0;
888 }
889
890 /**
891 * Repeatedly evict memory from the LRU for @mem_type until we create enough
892 * space, or we've evicted everything and there isn't enough space.
893 */
894 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
895 uint32_t mem_type,
896 const struct ttm_place *place,
897 struct ttm_mem_reg *mem,
898 struct ttm_operation_ctx *ctx)
899 {
900 struct ttm_bo_device *bdev = bo->bdev;
901 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
902 int ret;
903
904 do {
905 ret = (*man->func->get_node)(man, bo, place, mem);
906 if (unlikely(ret != 0))
907 return ret;
908 if (mem->mm_node)
909 break;
910 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
911 if (unlikely(ret != 0))
912 return ret;
913 } while (1);
914 mem->mem_type = mem_type;
915 return ttm_bo_add_move_fence(bo, man, mem);
916 }
917
918 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
919 uint32_t cur_placement,
920 uint32_t proposed_placement)
921 {
922 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
923 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
924
925 /**
926 * Keep current caching if possible.
927 */
928
929 if ((cur_placement & caching) != 0)
930 result |= (cur_placement & caching);
931 else if ((man->default_caching & caching) != 0)
932 result |= man->default_caching;
933 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
934 result |= TTM_PL_FLAG_CACHED;
935 else if ((TTM_PL_FLAG_WC & caching) != 0)
936 result |= TTM_PL_FLAG_WC;
937 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
938 result |= TTM_PL_FLAG_UNCACHED;
939
940 return result;
941 }
942
943 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
944 uint32_t mem_type,
945 const struct ttm_place *place,
946 uint32_t *masked_placement)
947 {
948 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
949
950 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
951 return false;
952
953 if ((place->flags & man->available_caching) == 0)
954 return false;
955
956 cur_flags |= (place->flags & man->available_caching);
957
958 *masked_placement = cur_flags;
959 return true;
960 }
961
962 /**
963 * Creates space for memory region @mem according to its type.
964 *
965 * This function first searches for free space in compatible memory types in
966 * the priority order defined by the driver. If free space isn't found, then
967 * ttm_bo_mem_force_space is attempted in priority order to evict and find
968 * space.
969 */
970 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
971 struct ttm_placement *placement,
972 struct ttm_mem_reg *mem,
973 struct ttm_operation_ctx *ctx)
974 {
975 struct ttm_bo_device *bdev = bo->bdev;
976 struct ttm_mem_type_manager *man;
977 uint32_t mem_type = TTM_PL_SYSTEM;
978 uint32_t cur_flags = 0;
979 bool type_found = false;
980 bool type_ok = false;
981 bool has_erestartsys = false;
982 int i, ret;
983
984 ret = reservation_object_reserve_shared(bo->resv, 1);
985 if (unlikely(ret))
986 return ret;
987
988 mem->mm_node = NULL;
989 for (i = 0; i < placement->num_placement; ++i) {
990 const struct ttm_place *place = &placement->placement[i];
991
992 ret = ttm_mem_type_from_place(place, &mem_type);
993 if (ret)
994 return ret;
995 man = &bdev->man[mem_type];
996 if (!man->has_type || !man->use_type)
997 continue;
998
999 type_ok = ttm_bo_mt_compatible(man, mem_type, place,
1000 &cur_flags);
1001
1002 if (!type_ok)
1003 continue;
1004
1005 type_found = true;
1006 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1007 cur_flags);
1008 /*
1009 * Use the access and other non-mapping-related flag bits from
1010 * the memory placement flags to the current flags
1011 */
1012 ttm_flag_masked(&cur_flags, place->flags,
1013 ~TTM_PL_MASK_MEMTYPE);
1014
1015 if (mem_type == TTM_PL_SYSTEM)
1016 break;
1017
1018 ret = (*man->func->get_node)(man, bo, place, mem);
1019 if (unlikely(ret))
1020 return ret;
1021
1022 if (mem->mm_node) {
1023 ret = ttm_bo_add_move_fence(bo, man, mem);
1024 if (unlikely(ret)) {
1025 (*man->func->put_node)(man, mem);
1026 return ret;
1027 }
1028 break;
1029 }
1030 }
1031
1032 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
1033 mem->mem_type = mem_type;
1034 mem->placement = cur_flags;
1035 return 0;
1036 }
1037
1038 for (i = 0; i < placement->num_busy_placement; ++i) {
1039 const struct ttm_place *place = &placement->busy_placement[i];
1040
1041 ret = ttm_mem_type_from_place(place, &mem_type);
1042 if (ret)
1043 return ret;
1044 man = &bdev->man[mem_type];
1045 if (!man->has_type || !man->use_type)
1046 continue;
1047 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
1048 continue;
1049
1050 type_found = true;
1051 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1052 cur_flags);
1053 /*
1054 * Use the access and other non-mapping-related flag bits from
1055 * the memory placement flags to the current flags
1056 */
1057 ttm_flag_masked(&cur_flags, place->flags,
1058 ~TTM_PL_MASK_MEMTYPE);
1059
1060 if (mem_type == TTM_PL_SYSTEM) {
1061 mem->mem_type = mem_type;
1062 mem->placement = cur_flags;
1063 mem->mm_node = NULL;
1064 return 0;
1065 }
1066
1067 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx);
1068 if (ret == 0 && mem->mm_node) {
1069 mem->placement = cur_flags;
1070 return 0;
1071 }
1072 if (ret == -ERESTARTSYS)
1073 has_erestartsys = true;
1074 }
1075
1076 if (!type_found) {
1077 pr_err(TTM_PFX "No compatible memory type found\n");
1078 return -EINVAL;
1079 }
1080
1081 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1082 }
1083 EXPORT_SYMBOL(ttm_bo_mem_space);
1084
1085 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1086 struct ttm_placement *placement,
1087 struct ttm_operation_ctx *ctx)
1088 {
1089 int ret = 0;
1090 struct ttm_mem_reg mem;
1091
1092 reservation_object_assert_held(bo->resv);
1093
1094 mem.num_pages = bo->num_pages;
1095 mem.size = mem.num_pages << PAGE_SHIFT;
1096 mem.page_alignment = bo->mem.page_alignment;
1097 mem.bus.io_reserved_vm = false;
1098 mem.bus.io_reserved_count = 0;
1099 /*
1100 * Determine where to move the buffer.
1101 */
1102 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1103 if (ret)
1104 goto out_unlock;
1105 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1106 out_unlock:
1107 if (ret && mem.mm_node)
1108 ttm_bo_mem_put(bo, &mem);
1109 return ret;
1110 }
1111
1112 static bool ttm_bo_places_compat(const struct ttm_place *places,
1113 unsigned num_placement,
1114 struct ttm_mem_reg *mem,
1115 uint32_t *new_flags)
1116 {
1117 unsigned i;
1118
1119 for (i = 0; i < num_placement; i++) {
1120 const struct ttm_place *heap = &places[i];
1121
1122 if (mem->mm_node && (mem->start < heap->fpfn ||
1123 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1124 continue;
1125
1126 *new_flags = heap->flags;
1127 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1128 (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
1129 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1130 (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1131 return true;
1132 }
1133 return false;
1134 }
1135
1136 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1137 struct ttm_mem_reg *mem,
1138 uint32_t *new_flags)
1139 {
1140 if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1141 mem, new_flags))
1142 return true;
1143
1144 if ((placement->busy_placement != placement->placement ||
1145 placement->num_busy_placement > placement->num_placement) &&
1146 ttm_bo_places_compat(placement->busy_placement,
1147 placement->num_busy_placement,
1148 mem, new_flags))
1149 return true;
1150
1151 return false;
1152 }
1153 EXPORT_SYMBOL(ttm_bo_mem_compat);
1154
1155 int ttm_bo_validate(struct ttm_buffer_object *bo,
1156 struct ttm_placement *placement,
1157 struct ttm_operation_ctx *ctx)
1158 {
1159 int ret;
1160 uint32_t new_flags;
1161
1162 reservation_object_assert_held(bo->resv);
1163 /*
1164 * Check whether we need to move buffer.
1165 */
1166 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1167 ret = ttm_bo_move_buffer(bo, placement, ctx);
1168 if (ret)
1169 return ret;
1170 } else {
1171 /*
1172 * Use the access and other non-mapping-related flag bits from
1173 * the compatible memory placement flags to the active flags
1174 */
1175 ttm_flag_masked(&bo->mem.placement, new_flags,
1176 ~TTM_PL_MASK_MEMTYPE);
1177 }
1178 /*
1179 * We might need to add a TTM.
1180 */
1181 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1182 ret = ttm_tt_create(bo, true);
1183 if (ret)
1184 return ret;
1185 }
1186 return 0;
1187 }
1188 EXPORT_SYMBOL(ttm_bo_validate);
1189
1190 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1191 struct ttm_buffer_object *bo,
1192 unsigned long size,
1193 enum ttm_bo_type type,
1194 struct ttm_placement *placement,
1195 uint32_t page_alignment,
1196 struct ttm_operation_ctx *ctx,
1197 size_t acc_size,
1198 struct sg_table *sg,
1199 struct reservation_object *resv,
1200 void (*destroy) (struct ttm_buffer_object *))
1201 {
1202 int ret = 0;
1203 unsigned long num_pages;
1204 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1205 bool locked;
1206
1207 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1208 if (ret) {
1209 pr_err("Out of kernel memory\n");
1210 if (destroy)
1211 (*destroy)(bo);
1212 else
1213 kfree(bo);
1214 return -ENOMEM;
1215 }
1216
1217 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1218 if (num_pages == 0) {
1219 pr_err("Illegal buffer object size\n");
1220 if (destroy)
1221 (*destroy)(bo);
1222 else
1223 kfree(bo);
1224 ttm_mem_global_free(mem_glob, acc_size);
1225 return -EINVAL;
1226 }
1227 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1228
1229 kref_init(&bo->kref);
1230 kref_init(&bo->list_kref);
1231 atomic_set(&bo->cpu_writers, 0);
1232 INIT_LIST_HEAD(&bo->lru);
1233 INIT_LIST_HEAD(&bo->ddestroy);
1234 INIT_LIST_HEAD(&bo->swap);
1235 INIT_LIST_HEAD(&bo->io_reserve_lru);
1236 mutex_init(&bo->wu_mutex);
1237 bo->bdev = bdev;
1238 bo->type = type;
1239 bo->num_pages = num_pages;
1240 bo->mem.size = num_pages << PAGE_SHIFT;
1241 bo->mem.mem_type = TTM_PL_SYSTEM;
1242 bo->mem.num_pages = bo->num_pages;
1243 bo->mem.mm_node = NULL;
1244 bo->mem.page_alignment = page_alignment;
1245 bo->mem.bus.io_reserved_vm = false;
1246 bo->mem.bus.io_reserved_count = 0;
1247 bo->moving = NULL;
1248 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1249 bo->acc_size = acc_size;
1250 bo->sg = sg;
1251 if (resv) {
1252 bo->resv = resv;
1253 reservation_object_assert_held(bo->resv);
1254 } else {
1255 bo->resv = &bo->ttm_resv;
1256 }
1257 reservation_object_init(&bo->ttm_resv);
1258 atomic_inc(&bo->bdev->glob->bo_count);
1259 drm_vma_node_reset(&bo->vma_node);
1260
1261 /*
1262 * For ttm_bo_type_device buffers, allocate
1263 * address space from the device.
1264 */
1265 if (bo->type == ttm_bo_type_device ||
1266 bo->type == ttm_bo_type_sg)
1267 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1268 bo->mem.num_pages);
1269
1270 /* passed reservation objects should already be locked,
1271 * since otherwise lockdep will be angered in radeon.
1272 */
1273 if (!resv) {
1274 locked = reservation_object_trylock(bo->resv);
1275 WARN_ON(!locked);
1276 }
1277
1278 if (likely(!ret))
1279 ret = ttm_bo_validate(bo, placement, ctx);
1280
1281 if (unlikely(ret)) {
1282 if (!resv)
1283 ttm_bo_unreserve(bo);
1284
1285 ttm_bo_put(bo);
1286 return ret;
1287 }
1288
1289 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1290 spin_lock(&bdev->glob->lru_lock);
1291 ttm_bo_add_to_lru(bo);
1292 spin_unlock(&bdev->glob->lru_lock);
1293 }
1294
1295 return ret;
1296 }
1297 EXPORT_SYMBOL(ttm_bo_init_reserved);
1298
1299 int ttm_bo_init(struct ttm_bo_device *bdev,
1300 struct ttm_buffer_object *bo,
1301 unsigned long size,
1302 enum ttm_bo_type type,
1303 struct ttm_placement *placement,
1304 uint32_t page_alignment,
1305 bool interruptible,
1306 size_t acc_size,
1307 struct sg_table *sg,
1308 struct reservation_object *resv,
1309 void (*destroy) (struct ttm_buffer_object *))
1310 {
1311 struct ttm_operation_ctx ctx = { interruptible, false };
1312 int ret;
1313
1314 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1315 page_alignment, &ctx, acc_size,
1316 sg, resv, destroy);
1317 if (ret)
1318 return ret;
1319
1320 if (!resv)
1321 ttm_bo_unreserve(bo);
1322
1323 return 0;
1324 }
1325 EXPORT_SYMBOL(ttm_bo_init);
1326
1327 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1328 unsigned long bo_size,
1329 unsigned struct_size)
1330 {
1331 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1332 size_t size = 0;
1333
1334 size += ttm_round_pot(struct_size);
1335 size += ttm_round_pot(npages * sizeof(void *));
1336 size += ttm_round_pot(sizeof(struct ttm_tt));
1337 return size;
1338 }
1339 EXPORT_SYMBOL(ttm_bo_acc_size);
1340
1341 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1342 unsigned long bo_size,
1343 unsigned struct_size)
1344 {
1345 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1346 size_t size = 0;
1347
1348 size += ttm_round_pot(struct_size);
1349 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1350 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1351 return size;
1352 }
1353 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1354
1355 int ttm_bo_create(struct ttm_bo_device *bdev,
1356 unsigned long size,
1357 enum ttm_bo_type type,
1358 struct ttm_placement *placement,
1359 uint32_t page_alignment,
1360 bool interruptible,
1361 struct ttm_buffer_object **p_bo)
1362 {
1363 struct ttm_buffer_object *bo;
1364 size_t acc_size;
1365 int ret;
1366
1367 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1368 if (unlikely(bo == NULL))
1369 return -ENOMEM;
1370
1371 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1372 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1373 interruptible, acc_size,
1374 NULL, NULL, NULL);
1375 if (likely(ret == 0))
1376 *p_bo = bo;
1377
1378 return ret;
1379 }
1380 EXPORT_SYMBOL(ttm_bo_create);
1381
1382 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1383 unsigned mem_type)
1384 {
1385 struct ttm_operation_ctx ctx = {
1386 .interruptible = false,
1387 .no_wait_gpu = false,
1388 .flags = TTM_OPT_FLAG_FORCE_ALLOC
1389 };
1390 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1391 struct ttm_bo_global *glob = bdev->glob;
1392 struct dma_fence *fence;
1393 int ret;
1394 unsigned i;
1395
1396 /*
1397 * Can't use standard list traversal since we're unlocking.
1398 */
1399
1400 spin_lock(&glob->lru_lock);
1401 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1402 while (!list_empty(&man->lru[i])) {
1403 spin_unlock(&glob->lru_lock);
1404 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx);
1405 if (ret)
1406 return ret;
1407 spin_lock(&glob->lru_lock);
1408 }
1409 }
1410 spin_unlock(&glob->lru_lock);
1411
1412 spin_lock(&man->move_lock);
1413 fence = dma_fence_get(man->move);
1414 spin_unlock(&man->move_lock);
1415
1416 if (fence) {
1417 ret = dma_fence_wait(fence, false);
1418 dma_fence_put(fence);
1419 if (ret)
1420 return ret;
1421 }
1422
1423 return 0;
1424 }
1425
1426 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1427 {
1428 struct ttm_mem_type_manager *man;
1429 int ret = -EINVAL;
1430
1431 if (mem_type >= TTM_NUM_MEM_TYPES) {
1432 pr_err("Illegal memory type %d\n", mem_type);
1433 return ret;
1434 }
1435 man = &bdev->man[mem_type];
1436
1437 if (!man->has_type) {
1438 pr_err("Trying to take down uninitialized memory manager type %u\n",
1439 mem_type);
1440 return ret;
1441 }
1442
1443 man->use_type = false;
1444 man->has_type = false;
1445
1446 ret = 0;
1447 if (mem_type > 0) {
1448 ret = ttm_bo_force_list_clean(bdev, mem_type);
1449 if (ret) {
1450 pr_err("Cleanup eviction failed\n");
1451 return ret;
1452 }
1453
1454 ret = (*man->func->takedown)(man);
1455 }
1456
1457 dma_fence_put(man->move);
1458 man->move = NULL;
1459
1460 return ret;
1461 }
1462 EXPORT_SYMBOL(ttm_bo_clean_mm);
1463
1464 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1465 {
1466 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1467
1468 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1469 pr_err("Illegal memory manager memory type %u\n", mem_type);
1470 return -EINVAL;
1471 }
1472
1473 if (!man->has_type) {
1474 pr_err("Memory type %u has not been initialized\n", mem_type);
1475 return 0;
1476 }
1477
1478 return ttm_bo_force_list_clean(bdev, mem_type);
1479 }
1480 EXPORT_SYMBOL(ttm_bo_evict_mm);
1481
1482 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1483 unsigned long p_size)
1484 {
1485 int ret;
1486 struct ttm_mem_type_manager *man;
1487 unsigned i;
1488
1489 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1490 man = &bdev->man[type];
1491 BUG_ON(man->has_type);
1492 man->io_reserve_fastpath = true;
1493 man->use_io_reserve_lru = false;
1494 mutex_init(&man->io_reserve_mutex);
1495 spin_lock_init(&man->move_lock);
1496 INIT_LIST_HEAD(&man->io_reserve_lru);
1497
1498 ret = bdev->driver->init_mem_type(bdev, type, man);
1499 if (ret)
1500 return ret;
1501 man->bdev = bdev;
1502
1503 if (type != TTM_PL_SYSTEM) {
1504 ret = (*man->func->init)(man, p_size);
1505 if (ret)
1506 return ret;
1507 }
1508 man->has_type = true;
1509 man->use_type = true;
1510 man->size = p_size;
1511
1512 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1513 INIT_LIST_HEAD(&man->lru[i]);
1514 man->move = NULL;
1515
1516 return 0;
1517 }
1518 EXPORT_SYMBOL(ttm_bo_init_mm);
1519
1520 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1521 {
1522 struct ttm_bo_global *glob =
1523 container_of(kobj, struct ttm_bo_global, kobj);
1524
1525 __free_page(glob->dummy_read_page);
1526 }
1527
1528 static void ttm_bo_global_release(void)
1529 {
1530 struct ttm_bo_global *glob = &ttm_bo_glob;
1531
1532 mutex_lock(&ttm_global_mutex);
1533 if (--ttm_bo_glob_use_count > 0)
1534 goto out;
1535
1536 kobject_del(&glob->kobj);
1537 kobject_put(&glob->kobj);
1538 ttm_mem_global_release(&ttm_mem_glob);
1539 memset(glob, 0, sizeof(*glob));
1540 out:
1541 mutex_unlock(&ttm_global_mutex);
1542 }
1543
1544 static int ttm_bo_global_init(void)
1545 {
1546 struct ttm_bo_global *glob = &ttm_bo_glob;
1547 int ret = 0;
1548 unsigned i;
1549
1550 mutex_lock(&ttm_global_mutex);
1551 if (++ttm_bo_glob_use_count > 1)
1552 goto out;
1553
1554 ret = ttm_mem_global_init(&ttm_mem_glob);
1555 if (ret)
1556 goto out;
1557
1558 spin_lock_init(&glob->lru_lock);
1559 glob->mem_glob = &ttm_mem_glob;
1560 glob->mem_glob->bo_glob = glob;
1561 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1562
1563 if (unlikely(glob->dummy_read_page == NULL)) {
1564 ret = -ENOMEM;
1565 goto out;
1566 }
1567
1568 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1569 INIT_LIST_HEAD(&glob->swap_lru[i]);
1570 INIT_LIST_HEAD(&glob->device_list);
1571 atomic_set(&glob->bo_count, 0);
1572
1573 ret = kobject_init_and_add(
1574 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1575 if (unlikely(ret != 0))
1576 kobject_put(&glob->kobj);
1577 out:
1578 mutex_unlock(&ttm_global_mutex);
1579 return ret;
1580 }
1581
1582 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1583 {
1584 int ret = 0;
1585 unsigned i = TTM_NUM_MEM_TYPES;
1586 struct ttm_mem_type_manager *man;
1587 struct ttm_bo_global *glob = bdev->glob;
1588
1589 while (i--) {
1590 man = &bdev->man[i];
1591 if (man->has_type) {
1592 man->use_type = false;
1593 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1594 ret = -EBUSY;
1595 pr_err("DRM memory manager type %d is not clean\n",
1596 i);
1597 }
1598 man->has_type = false;
1599 }
1600 }
1601
1602 mutex_lock(&ttm_global_mutex);
1603 list_del(&bdev->device_list);
1604 mutex_unlock(&ttm_global_mutex);
1605
1606 cancel_delayed_work_sync(&bdev->wq);
1607
1608 if (ttm_bo_delayed_delete(bdev, true))
1609 pr_debug("Delayed destroy list was clean\n");
1610
1611 spin_lock(&glob->lru_lock);
1612 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1613 if (list_empty(&bdev->man[0].lru[0]))
1614 pr_debug("Swap list %d was clean\n", i);
1615 spin_unlock(&glob->lru_lock);
1616
1617 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1618
1619 if (!ret)
1620 ttm_bo_global_release();
1621
1622 return ret;
1623 }
1624 EXPORT_SYMBOL(ttm_bo_device_release);
1625
1626 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1627 struct ttm_bo_driver *driver,
1628 struct address_space *mapping,
1629 uint64_t file_page_offset,
1630 bool need_dma32)
1631 {
1632 struct ttm_bo_global *glob = &ttm_bo_glob;
1633 int ret;
1634
1635 ret = ttm_bo_global_init();
1636 if (ret)
1637 return ret;
1638
1639 bdev->driver = driver;
1640
1641 memset(bdev->man, 0, sizeof(bdev->man));
1642
1643 /*
1644 * Initialize the system memory buffer type.
1645 * Other types need to be driver / IOCTL initialized.
1646 */
1647 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1648 if (unlikely(ret != 0))
1649 goto out_no_sys;
1650
1651 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1652 0x10000000);
1653 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1654 INIT_LIST_HEAD(&bdev->ddestroy);
1655 bdev->dev_mapping = mapping;
1656 bdev->glob = glob;
1657 bdev->need_dma32 = need_dma32;
1658 mutex_lock(&ttm_global_mutex);
1659 list_add_tail(&bdev->device_list, &glob->device_list);
1660 mutex_unlock(&ttm_global_mutex);
1661
1662 return 0;
1663 out_no_sys:
1664 ttm_bo_global_release();
1665 return ret;
1666 }
1667 EXPORT_SYMBOL(ttm_bo_device_init);
1668
1669 /*
1670 * buffer object vm functions.
1671 */
1672
1673 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1674 {
1675 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1676
1677 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1678 if (mem->mem_type == TTM_PL_SYSTEM)
1679 return false;
1680
1681 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1682 return false;
1683
1684 if (mem->placement & TTM_PL_FLAG_CACHED)
1685 return false;
1686 }
1687 return true;
1688 }
1689
1690 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1691 {
1692 struct ttm_bo_device *bdev = bo->bdev;
1693
1694 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1695 ttm_mem_io_free_vm(bo);
1696 }
1697
1698 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1699 {
1700 struct ttm_bo_device *bdev = bo->bdev;
1701 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1702
1703 ttm_mem_io_lock(man, false);
1704 ttm_bo_unmap_virtual_locked(bo);
1705 ttm_mem_io_unlock(man);
1706 }
1707
1708
1709 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1710
1711 int ttm_bo_wait(struct ttm_buffer_object *bo,
1712 bool interruptible, bool no_wait)
1713 {
1714 long timeout = 15 * HZ;
1715
1716 if (no_wait) {
1717 if (reservation_object_test_signaled_rcu(bo->resv, true))
1718 return 0;
1719 else
1720 return -EBUSY;
1721 }
1722
1723 timeout = reservation_object_wait_timeout_rcu(bo->resv, true,
1724 interruptible, timeout);
1725 if (timeout < 0)
1726 return timeout;
1727
1728 if (timeout == 0)
1729 return -EBUSY;
1730
1731 reservation_object_add_excl_fence(bo->resv, NULL);
1732 return 0;
1733 }
1734 EXPORT_SYMBOL(ttm_bo_wait);
1735
1736 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1737 {
1738 int ret = 0;
1739
1740 /*
1741 * Using ttm_bo_reserve makes sure the lru lists are updated.
1742 */
1743
1744 ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1745 if (unlikely(ret != 0))
1746 return ret;
1747 ret = ttm_bo_wait(bo, true, no_wait);
1748 if (likely(ret == 0))
1749 atomic_inc(&bo->cpu_writers);
1750 ttm_bo_unreserve(bo);
1751 return ret;
1752 }
1753 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1754
1755 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1756 {
1757 atomic_dec(&bo->cpu_writers);
1758 }
1759 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1760
1761 /**
1762 * A buffer object shrink method that tries to swap out the first
1763 * buffer object on the bo_global::swap_lru list.
1764 */
1765 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1766 {
1767 struct ttm_buffer_object *bo;
1768 int ret = -EBUSY;
1769 bool locked;
1770 unsigned i;
1771
1772 spin_lock(&glob->lru_lock);
1773 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1774 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1775 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) {
1776 ret = 0;
1777 break;
1778 }
1779 }
1780 if (!ret)
1781 break;
1782 }
1783
1784 if (ret) {
1785 spin_unlock(&glob->lru_lock);
1786 return ret;
1787 }
1788
1789 kref_get(&bo->list_kref);
1790
1791 if (!list_empty(&bo->ddestroy)) {
1792 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1793 kref_put(&bo->list_kref, ttm_bo_release_list);
1794 return ret;
1795 }
1796
1797 ttm_bo_del_from_lru(bo);
1798 spin_unlock(&glob->lru_lock);
1799
1800 /**
1801 * Move to system cached
1802 */
1803
1804 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1805 bo->ttm->caching_state != tt_cached) {
1806 struct ttm_operation_ctx ctx = { false, false };
1807 struct ttm_mem_reg evict_mem;
1808
1809 evict_mem = bo->mem;
1810 evict_mem.mm_node = NULL;
1811 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1812 evict_mem.mem_type = TTM_PL_SYSTEM;
1813
1814 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1815 if (unlikely(ret != 0))
1816 goto out;
1817 }
1818
1819 /**
1820 * Make sure BO is idle.
1821 */
1822
1823 ret = ttm_bo_wait(bo, false, false);
1824 if (unlikely(ret != 0))
1825 goto out;
1826
1827 ttm_bo_unmap_virtual(bo);
1828
1829 /**
1830 * Swap out. Buffer will be swapped in again as soon as
1831 * anyone tries to access a ttm page.
1832 */
1833
1834 if (bo->bdev->driver->swap_notify)
1835 bo->bdev->driver->swap_notify(bo);
1836
1837 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1838 out:
1839
1840 /**
1841 *
1842 * Unreserve without putting on LRU to avoid swapping out an
1843 * already swapped buffer.
1844 */
1845 if (locked)
1846 reservation_object_unlock(bo->resv);
1847 kref_put(&bo->list_kref, ttm_bo_release_list);
1848 return ret;
1849 }
1850 EXPORT_SYMBOL(ttm_bo_swapout);
1851
1852 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1853 {
1854 struct ttm_operation_ctx ctx = {
1855 .interruptible = false,
1856 .no_wait_gpu = false
1857 };
1858
1859 while (ttm_bo_swapout(bdev->glob, &ctx) == 0)
1860 ;
1861 }
1862 EXPORT_SYMBOL(ttm_bo_swapout_all);
1863
1864 /**
1865 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1866 * unreserved
1867 *
1868 * @bo: Pointer to buffer
1869 */
1870 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1871 {
1872 int ret;
1873
1874 /*
1875 * In the absense of a wait_unlocked API,
1876 * Use the bo::wu_mutex to avoid triggering livelocks due to
1877 * concurrent use of this function. Note that this use of
1878 * bo::wu_mutex can go away if we change locking order to
1879 * mmap_sem -> bo::reserve.
1880 */
1881 ret = mutex_lock_interruptible(&bo->wu_mutex);
1882 if (unlikely(ret != 0))
1883 return -ERESTARTSYS;
1884 if (!ww_mutex_is_locked(&bo->resv->lock))
1885 goto out_unlock;
1886 ret = reservation_object_lock_interruptible(bo->resv, NULL);
1887 if (ret == -EINTR)
1888 ret = -ERESTARTSYS;
1889 if (unlikely(ret != 0))
1890 goto out_unlock;
1891 reservation_object_unlock(bo->resv);
1892
1893 out_unlock:
1894 mutex_unlock(&bo->wu_mutex);
1895 return ret;
1896 }
A mirror of Linus' kernel repository