?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
treewide: remove redundant IS_ERR() before error code check
[linux/fpc-iii.git] / drivers / gpu / drm / ttm / ttm_bo.c
blob5df596fb0280c7f63745d277bdf8322dab99d0a0
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/dma-resv.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 EXPORT_SYMBOL(ttm_bo_glob);
55
56 static struct attribute ttm_bo_count = {
57 .name = "bo_count",
58 .mode = S_IRUGO
59 };
60
61 /* default destructor */
62 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
63 {
64 kfree(bo);
65 }
66
67 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
68 uint32_t *mem_type)
69 {
70 int pos;
71
72 pos = ffs(place->flags & TTM_PL_MASK_MEM);
73 if (unlikely(!pos))
74 return -EINVAL;
75
76 *mem_type = pos - 1;
77 return 0;
78 }
79
80 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, struct drm_printer *p,
81 int mem_type)
82 {
83 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
84
85 drm_printf(p, " has_type: %d\n", man->has_type);
86 drm_printf(p, " use_type: %d\n", man->use_type);
87 drm_printf(p, " flags: 0x%08X\n", man->flags);
88 drm_printf(p, " gpu_offset: 0x%08llX\n", man->gpu_offset);
89 drm_printf(p, " size: %llu\n", man->size);
90 drm_printf(p, " available_caching: 0x%08X\n", man->available_caching);
91 drm_printf(p, " default_caching: 0x%08X\n", man->default_caching);
92 if (mem_type != TTM_PL_SYSTEM)
93 (*man->func->debug)(man, p);
94 }
95
96 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
97 struct ttm_placement *placement)
98 {
99 struct drm_printer p = drm_debug_printer(TTM_PFX);
100 int i, ret, mem_type;
101
102 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n",
103 bo, bo->mem.num_pages, bo->mem.size >> 10,
104 bo->mem.size >> 20);
105 for (i = 0; i < placement->num_placement; i++) {
106 ret = ttm_mem_type_from_place(&placement->placement[i],
107 &mem_type);
108 if (ret)
109 return;
110 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
111 i, placement->placement[i].flags, mem_type);
112 ttm_mem_type_debug(bo->bdev, &p, mem_type);
113 }
114 }
115
116 static ssize_t ttm_bo_global_show(struct kobject *kobj,
117 struct attribute *attr,
118 char *buffer)
119 {
120 struct ttm_bo_global *glob =
121 container_of(kobj, struct ttm_bo_global, kobj);
122
123 return snprintf(buffer, PAGE_SIZE, "%d\n",
124 atomic_read(&glob->bo_count));
125 }
126
127 static struct attribute *ttm_bo_global_attrs[] = {
128 &ttm_bo_count,
129 NULL
130 };
131
132 static const struct sysfs_ops ttm_bo_global_ops = {
133 .show = &ttm_bo_global_show
134 };
135
136 static struct kobj_type ttm_bo_glob_kobj_type = {
137 .release = &ttm_bo_global_kobj_release,
138 .sysfs_ops = &ttm_bo_global_ops,
139 .default_attrs = ttm_bo_global_attrs
140 };
141
142
143 static inline uint32_t ttm_bo_type_flags(unsigned type)
144 {
145 return 1 << (type);
146 }
147
148 static void ttm_bo_release_list(struct kref *list_kref)
149 {
150 struct ttm_buffer_object *bo =
151 container_of(list_kref, struct ttm_buffer_object, list_kref);
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(bo->mem.mm_node != NULL);
157 BUG_ON(!list_empty(&bo->lru));
158 BUG_ON(!list_empty(&bo->ddestroy));
159 ttm_tt_destroy(bo->ttm);
160 atomic_dec(&ttm_bo_glob.bo_count);
161 dma_fence_put(bo->moving);
162 if (!ttm_bo_uses_embedded_gem_object(bo))
163 dma_resv_fini(&bo->base._resv);
164 bo->destroy(bo);
165 ttm_mem_global_free(&ttm_mem_glob, acc_size);
166 }
167
168 static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo,
169 struct ttm_mem_reg *mem)
170 {
171 struct ttm_bo_device *bdev = bo->bdev;
172 struct ttm_mem_type_manager *man;
173
174 dma_resv_assert_held(bo->base.resv);
175
176 if (!list_empty(&bo->lru))
177 return;
178
179 if (mem->placement & TTM_PL_FLAG_NO_EVICT)
180 return;
181
182 man = &bdev->man[mem->mem_type];
183 list_add_tail(&bo->lru, &man->lru[bo->priority]);
184 kref_get(&bo->list_kref);
185
186 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm &&
187 !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG |
188 TTM_PAGE_FLAG_SWAPPED))) {
189 list_add_tail(&bo->swap, &ttm_bo_glob.swap_lru[bo->priority]);
190 kref_get(&bo->list_kref);
191 }
192 }
193
194 static void ttm_bo_ref_bug(struct kref *list_kref)
195 {
196 BUG();
197 }
198
199 static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
200 {
201 struct ttm_bo_device *bdev = bo->bdev;
202 bool notify = false;
203
204 if (!list_empty(&bo->swap)) {
205 list_del_init(&bo->swap);
206 kref_put(&bo->list_kref, ttm_bo_ref_bug);
207 notify = true;
208 }
209 if (!list_empty(&bo->lru)) {
210 list_del_init(&bo->lru);
211 kref_put(&bo->list_kref, ttm_bo_ref_bug);
212 notify = true;
213 }
214
215 if (notify && bdev->driver->del_from_lru_notify)
216 bdev->driver->del_from_lru_notify(bo);
217 }
218
219 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
220 struct ttm_buffer_object *bo)
221 {
222 if (!pos->first)
223 pos->first = bo;
224 pos->last = bo;
225 }
226
227 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
228 struct ttm_lru_bulk_move *bulk)
229 {
230 dma_resv_assert_held(bo->base.resv);
231
232 ttm_bo_del_from_lru(bo);
233 ttm_bo_add_mem_to_lru(bo, &bo->mem);
234
235 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
236 switch (bo->mem.mem_type) {
237 case TTM_PL_TT:
238 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
239 break;
240
241 case TTM_PL_VRAM:
242 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
243 break;
244 }
245 if (bo->ttm && !(bo->ttm->page_flags &
246 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
247 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
248 }
249 }
250 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
251
252 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
253 {
254 unsigned i;
255
256 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
257 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
258 struct ttm_mem_type_manager *man;
259
260 if (!pos->first)
261 continue;
262
263 dma_resv_assert_held(pos->first->base.resv);
264 dma_resv_assert_held(pos->last->base.resv);
265
266 man = &pos->first->bdev->man[TTM_PL_TT];
267 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
268 &pos->last->lru);
269 }
270
271 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
272 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
273 struct ttm_mem_type_manager *man;
274
275 if (!pos->first)
276 continue;
277
278 dma_resv_assert_held(pos->first->base.resv);
279 dma_resv_assert_held(pos->last->base.resv);
280
281 man = &pos->first->bdev->man[TTM_PL_VRAM];
282 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
283 &pos->last->lru);
284 }
285
286 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
287 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
288 struct list_head *lru;
289
290 if (!pos->first)
291 continue;
292
293 dma_resv_assert_held(pos->first->base.resv);
294 dma_resv_assert_held(pos->last->base.resv);
295
296 lru = &ttm_bo_glob.swap_lru[i];
297 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
298 }
299 }
300 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
301
302 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
303 struct ttm_mem_reg *mem, bool evict,
304 struct ttm_operation_ctx *ctx)
305 {
306 struct ttm_bo_device *bdev = bo->bdev;
307 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
308 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
309 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
310 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
311 int ret = 0;
312
313 if (old_is_pci || new_is_pci ||
314 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
315 ret = ttm_mem_io_lock(old_man, true);
316 if (unlikely(ret != 0))
317 goto out_err;
318 ttm_bo_unmap_virtual_locked(bo);
319 ttm_mem_io_unlock(old_man);
320 }
321
322 /*
323 * Create and bind a ttm if required.
324 */
325
326 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
327 if (bo->ttm == NULL) {
328 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
329 ret = ttm_tt_create(bo, zero);
330 if (ret)
331 goto out_err;
332 }
333
334 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
335 if (ret)
336 goto out_err;
337
338 if (mem->mem_type != TTM_PL_SYSTEM) {
339 ret = ttm_tt_bind(bo->ttm, mem, ctx);
340 if (ret)
341 goto out_err;
342 }
343
344 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
345 if (bdev->driver->move_notify)
346 bdev->driver->move_notify(bo, evict, mem);
347 bo->mem = *mem;
348 mem->mm_node = NULL;
349 goto moved;
350 }
351 }
352
353 if (bdev->driver->move_notify)
354 bdev->driver->move_notify(bo, evict, mem);
355
356 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
357 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
358 ret = ttm_bo_move_ttm(bo, ctx, mem);
359 else if (bdev->driver->move)
360 ret = bdev->driver->move(bo, evict, ctx, mem);
361 else
362 ret = ttm_bo_move_memcpy(bo, ctx, mem);
363
364 if (ret) {
365 if (bdev->driver->move_notify) {
366 swap(*mem, bo->mem);
367 bdev->driver->move_notify(bo, false, mem);
368 swap(*mem, bo->mem);
369 }
370
371 goto out_err;
372 }
373
374 moved:
375 if (bo->evicted) {
376 if (bdev->driver->invalidate_caches) {
377 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
378 if (ret)
379 pr_err("Can not flush read caches\n");
380 }
381 bo->evicted = false;
382 }
383
384 if (bo->mem.mm_node)
385 bo->offset = (bo->mem.start << PAGE_SHIFT) +
386 bdev->man[bo->mem.mem_type].gpu_offset;
387 else
388 bo->offset = 0;
389
390 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
391 return 0;
392
393 out_err:
394 new_man = &bdev->man[bo->mem.mem_type];
395 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
396 ttm_tt_destroy(bo->ttm);
397 bo->ttm = NULL;
398 }
399
400 return ret;
401 }
402
403 /**
404 * Call bo::reserved.
405 * Will release GPU memory type usage on destruction.
406 * This is the place to put in driver specific hooks to release
407 * driver private resources.
408 * Will release the bo::reserved lock.
409 */
410
411 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
412 {
413 if (bo->bdev->driver->move_notify)
414 bo->bdev->driver->move_notify(bo, false, NULL);
415
416 ttm_tt_destroy(bo->ttm);
417 bo->ttm = NULL;
418 ttm_bo_mem_put(bo, &bo->mem);
419 }
420
421 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
422 {
423 int r;
424
425 if (bo->base.resv == &bo->base._resv)
426 return 0;
427
428 BUG_ON(!dma_resv_trylock(&bo->base._resv));
429
430 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
431 if (r)
432 dma_resv_unlock(&bo->base._resv);
433
434 return r;
435 }
436
437 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
438 {
439 struct dma_resv_list *fobj;
440 struct dma_fence *fence;
441 int i;
442
443 fobj = dma_resv_get_list(&bo->base._resv);
444 fence = dma_resv_get_excl(&bo->base._resv);
445 if (fence && !fence->ops->signaled)
446 dma_fence_enable_sw_signaling(fence);
447
448 for (i = 0; fobj && i < fobj->shared_count; ++i) {
449 fence = rcu_dereference_protected(fobj->shared[i],
450 dma_resv_held(bo->base.resv));
451
452 if (!fence->ops->signaled)
453 dma_fence_enable_sw_signaling(fence);
454 }
455 }
456
457 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
458 {
459 struct ttm_bo_device *bdev = bo->bdev;
460 int ret;
461
462 ret = ttm_bo_individualize_resv(bo);
463 if (ret) {
464 /* Last resort, if we fail to allocate memory for the
465 * fences block for the BO to become idle
466 */
467 dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
468 30 * HZ);
469 spin_lock(&ttm_bo_glob.lru_lock);
470 goto error;
471 }
472
473 spin_lock(&ttm_bo_glob.lru_lock);
474 ret = dma_resv_trylock(bo->base.resv) ? 0 : -EBUSY;
475 if (!ret) {
476 if (dma_resv_test_signaled_rcu(&bo->base._resv, true)) {
477 ttm_bo_del_from_lru(bo);
478 spin_unlock(&ttm_bo_glob.lru_lock);
479 if (bo->base.resv != &bo->base._resv)
480 dma_resv_unlock(&bo->base._resv);
481
482 ttm_bo_cleanup_memtype_use(bo);
483 dma_resv_unlock(bo->base.resv);
484 return;
485 }
486
487 ttm_bo_flush_all_fences(bo);
488
489 /*
490 * Make NO_EVICT bos immediately available to
491 * shrinkers, now that they are queued for
492 * destruction.
493 */
494 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
495 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
496 ttm_bo_move_to_lru_tail(bo, NULL);
497 }
498
499 dma_resv_unlock(bo->base.resv);
500 }
501 if (bo->base.resv != &bo->base._resv)
502 dma_resv_unlock(&bo->base._resv);
503
504 error:
505 kref_get(&bo->list_kref);
506 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
507 spin_unlock(&ttm_bo_glob.lru_lock);
508
509 schedule_delayed_work(&bdev->wq,
510 ((HZ / 100) < 1) ? 1 : HZ / 100);
511 }
512
513 /**
514 * function ttm_bo_cleanup_refs
515 * If bo idle, remove from delayed- and lru lists, and unref.
516 * If not idle, do nothing.
517 *
518 * Must be called with lru_lock and reservation held, this function
519 * will drop the lru lock and optionally the reservation lock before returning.
520 *
521 * @interruptible Any sleeps should occur interruptibly.
522 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
523 * @unlock_resv Unlock the reservation lock as well.
524 */
525
526 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
527 bool interruptible, bool no_wait_gpu,
528 bool unlock_resv)
529 {
530 struct dma_resv *resv;
531 int ret;
532
533 if (unlikely(list_empty(&bo->ddestroy)))
534 resv = bo->base.resv;
535 else
536 resv = &bo->base._resv;
537
538 if (dma_resv_test_signaled_rcu(resv, true))
539 ret = 0;
540 else
541 ret = -EBUSY;
542
543 if (ret && !no_wait_gpu) {
544 long lret;
545
546 if (unlock_resv)
547 dma_resv_unlock(bo->base.resv);
548 spin_unlock(&ttm_bo_glob.lru_lock);
549
550 lret = dma_resv_wait_timeout_rcu(resv, true,
551 interruptible,
552 30 * HZ);
553
554 if (lret < 0)
555 return lret;
556 else if (lret == 0)
557 return -EBUSY;
558
559 spin_lock(&ttm_bo_glob.lru_lock);
560 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
561 /*
562 * We raced, and lost, someone else holds the reservation now,
563 * and is probably busy in ttm_bo_cleanup_memtype_use.
564 *
565 * Even if it's not the case, because we finished waiting any
566 * delayed destruction would succeed, so just return success
567 * here.
568 */
569 spin_unlock(&ttm_bo_glob.lru_lock);
570 return 0;
571 }
572 ret = 0;
573 }
574
575 if (ret || unlikely(list_empty(&bo->ddestroy))) {
576 if (unlock_resv)
577 dma_resv_unlock(bo->base.resv);
578 spin_unlock(&ttm_bo_glob.lru_lock);
579 return ret;
580 }
581
582 ttm_bo_del_from_lru(bo);
583 list_del_init(&bo->ddestroy);
584 kref_put(&bo->list_kref, ttm_bo_ref_bug);
585
586 spin_unlock(&ttm_bo_glob.lru_lock);
587 ttm_bo_cleanup_memtype_use(bo);
588
589 if (unlock_resv)
590 dma_resv_unlock(bo->base.resv);
591
592 return 0;
593 }
594
595 /**
596 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
597 * encountered buffers.
598 */
599 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
600 {
601 struct ttm_bo_global *glob = &ttm_bo_glob;
602 struct list_head removed;
603 bool empty;
604
605 INIT_LIST_HEAD(&removed);
606
607 spin_lock(&glob->lru_lock);
608 while (!list_empty(&bdev->ddestroy)) {
609 struct ttm_buffer_object *bo;
610
611 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
612 ddestroy);
613 kref_get(&bo->list_kref);
614 list_move_tail(&bo->ddestroy, &removed);
615
616 if (remove_all || bo->base.resv != &bo->base._resv) {
617 spin_unlock(&glob->lru_lock);
618 dma_resv_lock(bo->base.resv, NULL);
619
620 spin_lock(&glob->lru_lock);
621 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
622
623 } else if (dma_resv_trylock(bo->base.resv)) {
624 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
625 } else {
626 spin_unlock(&glob->lru_lock);
627 }
628
629 kref_put(&bo->list_kref, ttm_bo_release_list);
630 spin_lock(&glob->lru_lock);
631 }
632 list_splice_tail(&removed, &bdev->ddestroy);
633 empty = list_empty(&bdev->ddestroy);
634 spin_unlock(&glob->lru_lock);
635
636 return empty;
637 }
638
639 static void ttm_bo_delayed_workqueue(struct work_struct *work)
640 {
641 struct ttm_bo_device *bdev =
642 container_of(work, struct ttm_bo_device, wq.work);
643
644 if (!ttm_bo_delayed_delete(bdev, false))
645 schedule_delayed_work(&bdev->wq,
646 ((HZ / 100) < 1) ? 1 : HZ / 100);
647 }
648
649 static void ttm_bo_release(struct kref *kref)
650 {
651 struct ttm_buffer_object *bo =
652 container_of(kref, struct ttm_buffer_object, kref);
653 struct ttm_bo_device *bdev = bo->bdev;
654 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
655
656 if (bo->bdev->driver->release_notify)
657 bo->bdev->driver->release_notify(bo);
658
659 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
660 ttm_mem_io_lock(man, false);
661 ttm_mem_io_free_vm(bo);
662 ttm_mem_io_unlock(man);
663 ttm_bo_cleanup_refs_or_queue(bo);
664 kref_put(&bo->list_kref, ttm_bo_release_list);
665 }
666
667 void ttm_bo_put(struct ttm_buffer_object *bo)
668 {
669 kref_put(&bo->kref, ttm_bo_release);
670 }
671 EXPORT_SYMBOL(ttm_bo_put);
672
673 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
674 {
675 return cancel_delayed_work_sync(&bdev->wq);
676 }
677 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
678
679 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
680 {
681 if (resched)
682 schedule_delayed_work(&bdev->wq,
683 ((HZ / 100) < 1) ? 1 : HZ / 100);
684 }
685 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
686
687 static int ttm_bo_evict(struct ttm_buffer_object *bo,
688 struct ttm_operation_ctx *ctx)
689 {
690 struct ttm_bo_device *bdev = bo->bdev;
691 struct ttm_mem_reg evict_mem;
692 struct ttm_placement placement;
693 int ret = 0;
694
695 dma_resv_assert_held(bo->base.resv);
696
697 placement.num_placement = 0;
698 placement.num_busy_placement = 0;
699 bdev->driver->evict_flags(bo, &placement);
700
701 if (!placement.num_placement && !placement.num_busy_placement) {
702 ret = ttm_bo_pipeline_gutting(bo);
703 if (ret)
704 return ret;
705
706 return ttm_tt_create(bo, false);
707 }
708
709 evict_mem = bo->mem;
710 evict_mem.mm_node = NULL;
711 evict_mem.bus.io_reserved_vm = false;
712 evict_mem.bus.io_reserved_count = 0;
713
714 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
715 if (ret) {
716 if (ret != -ERESTARTSYS) {
717 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
718 bo);
719 ttm_bo_mem_space_debug(bo, &placement);
720 }
721 goto out;
722 }
723
724 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
725 if (unlikely(ret)) {
726 if (ret != -ERESTARTSYS)
727 pr_err("Buffer eviction failed\n");
728 ttm_bo_mem_put(bo, &evict_mem);
729 goto out;
730 }
731 bo->evicted = true;
732 out:
733 return ret;
734 }
735
736 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
737 const struct ttm_place *place)
738 {
739 /* Don't evict this BO if it's outside of the
740 * requested placement range
741 */
742 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
743 (place->lpfn && place->lpfn <= bo->mem.start))
744 return false;
745
746 return true;
747 }
748 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
749
750 /**
751 * Check the target bo is allowable to be evicted or swapout, including cases:
752 *
753 * a. if share same reservation object with ctx->resv, have assumption
754 * reservation objects should already be locked, so not lock again and
755 * return true directly when either the opreation allow_reserved_eviction
756 * or the target bo already is in delayed free list;
757 *
758 * b. Otherwise, trylock it.
759 */
760 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
761 struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
762 {
763 bool ret = false;
764
765 if (bo->base.resv == ctx->resv) {
766 dma_resv_assert_held(bo->base.resv);
767 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT
768 || !list_empty(&bo->ddestroy))
769 ret = true;
770 *locked = false;
771 if (busy)
772 *busy = false;
773 } else {
774 ret = dma_resv_trylock(bo->base.resv);
775 *locked = ret;
776 if (busy)
777 *busy = !ret;
778 }
779
780 return ret;
781 }
782
783 /**
784 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
785 *
786 * @busy_bo: BO which couldn't be locked with trylock
787 * @ctx: operation context
788 * @ticket: acquire ticket
789 *
790 * Try to lock a busy buffer object to avoid failing eviction.
791 */
792 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
793 struct ttm_operation_ctx *ctx,
794 struct ww_acquire_ctx *ticket)
795 {
796 int r;
797
798 if (!busy_bo || !ticket)
799 return -EBUSY;
800
801 if (ctx->interruptible)
802 r = dma_resv_lock_interruptible(busy_bo->base.resv,
803 ticket);
804 else
805 r = dma_resv_lock(busy_bo->base.resv, ticket);
806
807 /*
808 * TODO: It would be better to keep the BO locked until allocation is at
809 * least tried one more time, but that would mean a much larger rework
810 * of TTM.
811 */
812 if (!r)
813 dma_resv_unlock(busy_bo->base.resv);
814
815 return r == -EDEADLK ? -EBUSY : r;
816 }
817
818 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
819 uint32_t mem_type,
820 const struct ttm_place *place,
821 struct ttm_operation_ctx *ctx,
822 struct ww_acquire_ctx *ticket)
823 {
824 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
825 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
826 bool locked = false;
827 unsigned i;
828 int ret;
829
830 spin_lock(&ttm_bo_glob.lru_lock);
831 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
832 list_for_each_entry(bo, &man->lru[i], lru) {
833 bool busy;
834
835 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
836 &busy)) {
837 if (busy && !busy_bo && ticket !=
838 dma_resv_locking_ctx(bo->base.resv))
839 busy_bo = bo;
840 continue;
841 }
842
843 if (place && !bdev->driver->eviction_valuable(bo,
844 place)) {
845 if (locked)
846 dma_resv_unlock(bo->base.resv);
847 continue;
848 }
849 break;
850 }
851
852 /* If the inner loop terminated early, we have our candidate */
853 if (&bo->lru != &man->lru[i])
854 break;
855
856 bo = NULL;
857 }
858
859 if (!bo) {
860 if (busy_bo)
861 kref_get(&busy_bo->list_kref);
862 spin_unlock(&ttm_bo_glob.lru_lock);
863 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
864 if (busy_bo)
865 kref_put(&busy_bo->list_kref, ttm_bo_release_list);
866 return ret;
867 }
868
869 kref_get(&bo->list_kref);
870
871 if (!list_empty(&bo->ddestroy)) {
872 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
873 ctx->no_wait_gpu, locked);
874 kref_put(&bo->list_kref, ttm_bo_release_list);
875 return ret;
876 }
877
878 spin_unlock(&ttm_bo_glob.lru_lock);
879
880 ret = ttm_bo_evict(bo, ctx);
881 if (locked)
882 ttm_bo_unreserve(bo);
883
884 kref_put(&bo->list_kref, ttm_bo_release_list);
885 return ret;
886 }
887
888 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
889 {
890 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
891
892 if (mem->mm_node)
893 (*man->func->put_node)(man, mem);
894 }
895 EXPORT_SYMBOL(ttm_bo_mem_put);
896
897 /**
898 * Add the last move fence to the BO and reserve a new shared slot.
899 */
900 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
901 struct ttm_mem_type_manager *man,
902 struct ttm_mem_reg *mem,
903 bool no_wait_gpu)
904 {
905 struct dma_fence *fence;
906 int ret;
907
908 spin_lock(&man->move_lock);
909 fence = dma_fence_get(man->move);
910 spin_unlock(&man->move_lock);
911
912 if (!fence)
913 return 0;
914
915 if (no_wait_gpu)
916 return -EBUSY;
917
918 dma_resv_add_shared_fence(bo->base.resv, fence);
919
920 ret = dma_resv_reserve_shared(bo->base.resv, 1);
921 if (unlikely(ret)) {
922 dma_fence_put(fence);
923 return ret;
924 }
925
926 dma_fence_put(bo->moving);
927 bo->moving = fence;
928 return 0;
929 }
930
931 /**
932 * Repeatedly evict memory from the LRU for @mem_type until we create enough
933 * space, or we've evicted everything and there isn't enough space.
934 */
935 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
936 const struct ttm_place *place,
937 struct ttm_mem_reg *mem,
938 struct ttm_operation_ctx *ctx)
939 {
940 struct ttm_bo_device *bdev = bo->bdev;
941 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
942 struct ww_acquire_ctx *ticket;
943 int ret;
944
945 ticket = dma_resv_locking_ctx(bo->base.resv);
946 do {
947 ret = (*man->func->get_node)(man, bo, place, mem);
948 if (unlikely(ret != 0))
949 return ret;
950 if (mem->mm_node)
951 break;
952 ret = ttm_mem_evict_first(bdev, mem->mem_type, place, ctx,
953 ticket);
954 if (unlikely(ret != 0))
955 return ret;
956 } while (1);
957
958 return ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
959 }
960
961 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
962 uint32_t cur_placement,
963 uint32_t proposed_placement)
964 {
965 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
966 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
967
968 /**
969 * Keep current caching if possible.
970 */
971
972 if ((cur_placement & caching) != 0)
973 result |= (cur_placement & caching);
974 else if ((man->default_caching & caching) != 0)
975 result |= man->default_caching;
976 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
977 result |= TTM_PL_FLAG_CACHED;
978 else if ((TTM_PL_FLAG_WC & caching) != 0)
979 result |= TTM_PL_FLAG_WC;
980 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
981 result |= TTM_PL_FLAG_UNCACHED;
982
983 return result;
984 }
985
986 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
987 uint32_t mem_type,
988 const struct ttm_place *place,
989 uint32_t *masked_placement)
990 {
991 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
992
993 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
994 return false;
995
996 if ((place->flags & man->available_caching) == 0)
997 return false;
998
999 cur_flags |= (place->flags & man->available_caching);
1000
1001 *masked_placement = cur_flags;
1002 return true;
1003 }
1004
1005 /**
1006 * ttm_bo_mem_placement - check if placement is compatible
1007 * @bo: BO to find memory for
1008 * @place: where to search
1009 * @mem: the memory object to fill in
1010 * @ctx: operation context
1011 *
1012 * Check if placement is compatible and fill in mem structure.
1013 * Returns -EBUSY if placement won't work or negative error code.
1014 * 0 when placement can be used.
1015 */
1016 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo,
1017 const struct ttm_place *place,
1018 struct ttm_mem_reg *mem,
1019 struct ttm_operation_ctx *ctx)
1020 {
1021 struct ttm_bo_device *bdev = bo->bdev;
1022 uint32_t mem_type = TTM_PL_SYSTEM;
1023 struct ttm_mem_type_manager *man;
1024 uint32_t cur_flags = 0;
1025 int ret;
1026
1027 ret = ttm_mem_type_from_place(place, &mem_type);
1028 if (ret)
1029 return ret;
1030
1031 man = &bdev->man[mem_type];
1032 if (!man->has_type || !man->use_type)
1033 return -EBUSY;
1034
1035 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
1036 return -EBUSY;
1037
1038 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags);
1039 /*
1040 * Use the access and other non-mapping-related flag bits from
1041 * the memory placement flags to the current flags
1042 */
1043 ttm_flag_masked(&cur_flags, place->flags, ~TTM_PL_MASK_MEMTYPE);
1044
1045 mem->mem_type = mem_type;
1046 mem->placement = cur_flags;
1047
1048 spin_lock(&ttm_bo_glob.lru_lock);
1049 ttm_bo_del_from_lru(bo);
1050 ttm_bo_add_mem_to_lru(bo, mem);
1051 spin_unlock(&ttm_bo_glob.lru_lock);
1052
1053 return 0;
1054 }
1055
1056 /**
1057 * Creates space for memory region @mem according to its type.
1058 *
1059 * This function first searches for free space in compatible memory types in
1060 * the priority order defined by the driver. If free space isn't found, then
1061 * ttm_bo_mem_force_space is attempted in priority order to evict and find
1062 * space.
1063 */
1064 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
1065 struct ttm_placement *placement,
1066 struct ttm_mem_reg *mem,
1067 struct ttm_operation_ctx *ctx)
1068 {
1069 struct ttm_bo_device *bdev = bo->bdev;
1070 bool type_found = false;
1071 int i, ret;
1072
1073 ret = dma_resv_reserve_shared(bo->base.resv, 1);
1074 if (unlikely(ret))
1075 return ret;
1076
1077 mem->mm_node = NULL;
1078 for (i = 0; i < placement->num_placement; ++i) {
1079 const struct ttm_place *place = &placement->placement[i];
1080 struct ttm_mem_type_manager *man;
1081
1082 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
1083 if (ret == -EBUSY)
1084 continue;
1085 if (ret)
1086 goto error;
1087
1088 type_found = true;
1089 mem->mm_node = NULL;
1090 if (mem->mem_type == TTM_PL_SYSTEM)
1091 return 0;
1092
1093 man = &bdev->man[mem->mem_type];
1094 ret = (*man->func->get_node)(man, bo, place, mem);
1095 if (unlikely(ret))
1096 goto error;
1097
1098 if (!mem->mm_node)
1099 continue;
1100
1101 ret = ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
1102 if (unlikely(ret)) {
1103 (*man->func->put_node)(man, mem);
1104 if (ret == -EBUSY)
1105 continue;
1106
1107 goto error;
1108 }
1109 return 0;
1110 }
1111
1112 for (i = 0; i < placement->num_busy_placement; ++i) {
1113 const struct ttm_place *place = &placement->busy_placement[i];
1114
1115 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
1116 if (ret == -EBUSY)
1117 continue;
1118 if (ret)
1119 goto error;
1120
1121 type_found = true;
1122 mem->mm_node = NULL;
1123 if (mem->mem_type == TTM_PL_SYSTEM)
1124 return 0;
1125
1126 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
1127 if (ret == 0 && mem->mm_node)
1128 return 0;
1129
1130 if (ret && ret != -EBUSY)
1131 goto error;
1132 }
1133
1134 ret = -ENOMEM;
1135 if (!type_found) {
1136 pr_err(TTM_PFX "No compatible memory type found\n");
1137 ret = -EINVAL;
1138 }
1139
1140 error:
1141 if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) {
1142 spin_lock(&ttm_bo_glob.lru_lock);
1143 ttm_bo_move_to_lru_tail(bo, NULL);
1144 spin_unlock(&ttm_bo_glob.lru_lock);
1145 }
1146
1147 return ret;
1148 }
1149 EXPORT_SYMBOL(ttm_bo_mem_space);
1150
1151 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1152 struct ttm_placement *placement,
1153 struct ttm_operation_ctx *ctx)
1154 {
1155 int ret = 0;
1156 struct ttm_mem_reg mem;
1157
1158 dma_resv_assert_held(bo->base.resv);
1159
1160 mem.num_pages = bo->num_pages;
1161 mem.size = mem.num_pages << PAGE_SHIFT;
1162 mem.page_alignment = bo->mem.page_alignment;
1163 mem.bus.io_reserved_vm = false;
1164 mem.bus.io_reserved_count = 0;
1165 /*
1166 * Determine where to move the buffer.
1167 */
1168 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1169 if (ret)
1170 goto out_unlock;
1171 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1172 out_unlock:
1173 if (ret && mem.mm_node)
1174 ttm_bo_mem_put(bo, &mem);
1175 return ret;
1176 }
1177
1178 static bool ttm_bo_places_compat(const struct ttm_place *places,
1179 unsigned num_placement,
1180 struct ttm_mem_reg *mem,
1181 uint32_t *new_flags)
1182 {
1183 unsigned i;
1184
1185 for (i = 0; i < num_placement; i++) {
1186 const struct ttm_place *heap = &places[i];
1187
1188 if (mem->mm_node && (mem->start < heap->fpfn ||
1189 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1190 continue;
1191
1192 *new_flags = heap->flags;
1193 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1194 (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
1195 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1196 (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1197 return true;
1198 }
1199 return false;
1200 }
1201
1202 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1203 struct ttm_mem_reg *mem,
1204 uint32_t *new_flags)
1205 {
1206 if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1207 mem, new_flags))
1208 return true;
1209
1210 if ((placement->busy_placement != placement->placement ||
1211 placement->num_busy_placement > placement->num_placement) &&
1212 ttm_bo_places_compat(placement->busy_placement,
1213 placement->num_busy_placement,
1214 mem, new_flags))
1215 return true;
1216
1217 return false;
1218 }
1219 EXPORT_SYMBOL(ttm_bo_mem_compat);
1220
1221 int ttm_bo_validate(struct ttm_buffer_object *bo,
1222 struct ttm_placement *placement,
1223 struct ttm_operation_ctx *ctx)
1224 {
1225 int ret;
1226 uint32_t new_flags;
1227
1228 dma_resv_assert_held(bo->base.resv);
1229 /*
1230 * Check whether we need to move buffer.
1231 */
1232 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1233 ret = ttm_bo_move_buffer(bo, placement, ctx);
1234 if (ret)
1235 return ret;
1236 } else {
1237 /*
1238 * Use the access and other non-mapping-related flag bits from
1239 * the compatible memory placement flags to the active flags
1240 */
1241 ttm_flag_masked(&bo->mem.placement, new_flags,
1242 ~TTM_PL_MASK_MEMTYPE);
1243 }
1244 /*
1245 * We might need to add a TTM.
1246 */
1247 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1248 ret = ttm_tt_create(bo, true);
1249 if (ret)
1250 return ret;
1251 }
1252 return 0;
1253 }
1254 EXPORT_SYMBOL(ttm_bo_validate);
1255
1256 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1257 struct ttm_buffer_object *bo,
1258 unsigned long size,
1259 enum ttm_bo_type type,
1260 struct ttm_placement *placement,
1261 uint32_t page_alignment,
1262 struct ttm_operation_ctx *ctx,
1263 size_t acc_size,
1264 struct sg_table *sg,
1265 struct dma_resv *resv,
1266 void (*destroy) (struct ttm_buffer_object *))
1267 {
1268 struct ttm_mem_global *mem_glob = &ttm_mem_glob;
1269 int ret = 0;
1270 unsigned long num_pages;
1271 bool locked;
1272
1273 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1274 if (ret) {
1275 pr_err("Out of kernel memory\n");
1276 if (destroy)
1277 (*destroy)(bo);
1278 else
1279 kfree(bo);
1280 return -ENOMEM;
1281 }
1282
1283 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1284 if (num_pages == 0) {
1285 pr_err("Illegal buffer object size\n");
1286 if (destroy)
1287 (*destroy)(bo);
1288 else
1289 kfree(bo);
1290 ttm_mem_global_free(mem_glob, acc_size);
1291 return -EINVAL;
1292 }
1293 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1294
1295 kref_init(&bo->kref);
1296 kref_init(&bo->list_kref);
1297 INIT_LIST_HEAD(&bo->lru);
1298 INIT_LIST_HEAD(&bo->ddestroy);
1299 INIT_LIST_HEAD(&bo->swap);
1300 INIT_LIST_HEAD(&bo->io_reserve_lru);
1301 bo->bdev = bdev;
1302 bo->type = type;
1303 bo->num_pages = num_pages;
1304 bo->mem.size = num_pages << PAGE_SHIFT;
1305 bo->mem.mem_type = TTM_PL_SYSTEM;
1306 bo->mem.num_pages = bo->num_pages;
1307 bo->mem.mm_node = NULL;
1308 bo->mem.page_alignment = page_alignment;
1309 bo->mem.bus.io_reserved_vm = false;
1310 bo->mem.bus.io_reserved_count = 0;
1311 bo->moving = NULL;
1312 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1313 bo->acc_size = acc_size;
1314 bo->sg = sg;
1315 if (resv) {
1316 bo->base.resv = resv;
1317 dma_resv_assert_held(bo->base.resv);
1318 } else {
1319 bo->base.resv = &bo->base._resv;
1320 }
1321 if (!ttm_bo_uses_embedded_gem_object(bo)) {
1322 /*
1323 * bo.gem is not initialized, so we have to setup the
1324 * struct elements we want use regardless.
1325 */
1326 dma_resv_init(&bo->base._resv);
1327 drm_vma_node_reset(&bo->base.vma_node);
1328 }
1329 atomic_inc(&ttm_bo_glob.bo_count);
1330
1331 /*
1332 * For ttm_bo_type_device buffers, allocate
1333 * address space from the device.
1334 */
1335 if (bo->type == ttm_bo_type_device ||
1336 bo->type == ttm_bo_type_sg)
1337 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1338 bo->mem.num_pages);
1339
1340 /* passed reservation objects should already be locked,
1341 * since otherwise lockdep will be angered in radeon.
1342 */
1343 if (!resv) {
1344 locked = dma_resv_trylock(bo->base.resv);
1345 WARN_ON(!locked);
1346 }
1347
1348 if (likely(!ret))
1349 ret = ttm_bo_validate(bo, placement, ctx);
1350
1351 if (unlikely(ret)) {
1352 if (!resv)
1353 ttm_bo_unreserve(bo);
1354
1355 ttm_bo_put(bo);
1356 return ret;
1357 }
1358
1359 spin_lock(&ttm_bo_glob.lru_lock);
1360 ttm_bo_move_to_lru_tail(bo, NULL);
1361 spin_unlock(&ttm_bo_glob.lru_lock);
1362
1363 return ret;
1364 }
1365 EXPORT_SYMBOL(ttm_bo_init_reserved);
1366
1367 int ttm_bo_init(struct ttm_bo_device *bdev,
1368 struct ttm_buffer_object *bo,
1369 unsigned long size,
1370 enum ttm_bo_type type,
1371 struct ttm_placement *placement,
1372 uint32_t page_alignment,
1373 bool interruptible,
1374 size_t acc_size,
1375 struct sg_table *sg,
1376 struct dma_resv *resv,
1377 void (*destroy) (struct ttm_buffer_object *))
1378 {
1379 struct ttm_operation_ctx ctx = { interruptible, false };
1380 int ret;
1381
1382 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1383 page_alignment, &ctx, acc_size,
1384 sg, resv, destroy);
1385 if (ret)
1386 return ret;
1387
1388 if (!resv)
1389 ttm_bo_unreserve(bo);
1390
1391 return 0;
1392 }
1393 EXPORT_SYMBOL(ttm_bo_init);
1394
1395 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1396 unsigned long bo_size,
1397 unsigned struct_size)
1398 {
1399 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1400 size_t size = 0;
1401
1402 size += ttm_round_pot(struct_size);
1403 size += ttm_round_pot(npages * sizeof(void *));
1404 size += ttm_round_pot(sizeof(struct ttm_tt));
1405 return size;
1406 }
1407 EXPORT_SYMBOL(ttm_bo_acc_size);
1408
1409 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1410 unsigned long bo_size,
1411 unsigned struct_size)
1412 {
1413 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1414 size_t size = 0;
1415
1416 size += ttm_round_pot(struct_size);
1417 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1418 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1419 return size;
1420 }
1421 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1422
1423 int ttm_bo_create(struct ttm_bo_device *bdev,
1424 unsigned long size,
1425 enum ttm_bo_type type,
1426 struct ttm_placement *placement,
1427 uint32_t page_alignment,
1428 bool interruptible,
1429 struct ttm_buffer_object **p_bo)
1430 {
1431 struct ttm_buffer_object *bo;
1432 size_t acc_size;
1433 int ret;
1434
1435 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1436 if (unlikely(bo == NULL))
1437 return -ENOMEM;
1438
1439 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1440 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1441 interruptible, acc_size,
1442 NULL, NULL, NULL);
1443 if (likely(ret == 0))
1444 *p_bo = bo;
1445
1446 return ret;
1447 }
1448 EXPORT_SYMBOL(ttm_bo_create);
1449
1450 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1451 unsigned mem_type)
1452 {
1453 struct ttm_operation_ctx ctx = {
1454 .interruptible = false,
1455 .no_wait_gpu = false,
1456 .flags = TTM_OPT_FLAG_FORCE_ALLOC
1457 };
1458 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1459 struct ttm_bo_global *glob = &ttm_bo_glob;
1460 struct dma_fence *fence;
1461 int ret;
1462 unsigned i;
1463
1464 /*
1465 * Can't use standard list traversal since we're unlocking.
1466 */
1467
1468 spin_lock(&glob->lru_lock);
1469 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1470 while (!list_empty(&man->lru[i])) {
1471 spin_unlock(&glob->lru_lock);
1472 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx,
1473 NULL);
1474 if (ret)
1475 return ret;
1476 spin_lock(&glob->lru_lock);
1477 }
1478 }
1479 spin_unlock(&glob->lru_lock);
1480
1481 spin_lock(&man->move_lock);
1482 fence = dma_fence_get(man->move);
1483 spin_unlock(&man->move_lock);
1484
1485 if (fence) {
1486 ret = dma_fence_wait(fence, false);
1487 dma_fence_put(fence);
1488 if (ret)
1489 return ret;
1490 }
1491
1492 return 0;
1493 }
1494
1495 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1496 {
1497 struct ttm_mem_type_manager *man;
1498 int ret = -EINVAL;
1499
1500 if (mem_type >= TTM_NUM_MEM_TYPES) {
1501 pr_err("Illegal memory type %d\n", mem_type);
1502 return ret;
1503 }
1504 man = &bdev->man[mem_type];
1505
1506 if (!man->has_type) {
1507 pr_err("Trying to take down uninitialized memory manager type %u\n",
1508 mem_type);
1509 return ret;
1510 }
1511
1512 man->use_type = false;
1513 man->has_type = false;
1514
1515 ret = 0;
1516 if (mem_type > 0) {
1517 ret = ttm_bo_force_list_clean(bdev, mem_type);
1518 if (ret) {
1519 pr_err("Cleanup eviction failed\n");
1520 return ret;
1521 }
1522
1523 ret = (*man->func->takedown)(man);
1524 }
1525
1526 dma_fence_put(man->move);
1527 man->move = NULL;
1528
1529 return ret;
1530 }
1531 EXPORT_SYMBOL(ttm_bo_clean_mm);
1532
1533 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1534 {
1535 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1536
1537 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1538 pr_err("Illegal memory manager memory type %u\n", mem_type);
1539 return -EINVAL;
1540 }
1541
1542 if (!man->has_type) {
1543 pr_err("Memory type %u has not been initialized\n", mem_type);
1544 return 0;
1545 }
1546
1547 return ttm_bo_force_list_clean(bdev, mem_type);
1548 }
1549 EXPORT_SYMBOL(ttm_bo_evict_mm);
1550
1551 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1552 unsigned long p_size)
1553 {
1554 int ret;
1555 struct ttm_mem_type_manager *man;
1556 unsigned i;
1557
1558 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1559 man = &bdev->man[type];
1560 BUG_ON(man->has_type);
1561 man->io_reserve_fastpath = true;
1562 man->use_io_reserve_lru = false;
1563 mutex_init(&man->io_reserve_mutex);
1564 spin_lock_init(&man->move_lock);
1565 INIT_LIST_HEAD(&man->io_reserve_lru);
1566
1567 ret = bdev->driver->init_mem_type(bdev, type, man);
1568 if (ret)
1569 return ret;
1570 man->bdev = bdev;
1571
1572 if (type != TTM_PL_SYSTEM) {
1573 ret = (*man->func->init)(man, p_size);
1574 if (ret)
1575 return ret;
1576 }
1577 man->has_type = true;
1578 man->use_type = true;
1579 man->size = p_size;
1580
1581 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1582 INIT_LIST_HEAD(&man->lru[i]);
1583 man->move = NULL;
1584
1585 return 0;
1586 }
1587 EXPORT_SYMBOL(ttm_bo_init_mm);
1588
1589 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1590 {
1591 struct ttm_bo_global *glob =
1592 container_of(kobj, struct ttm_bo_global, kobj);
1593
1594 __free_page(glob->dummy_read_page);
1595 }
1596
1597 static void ttm_bo_global_release(void)
1598 {
1599 struct ttm_bo_global *glob = &ttm_bo_glob;
1600
1601 mutex_lock(&ttm_global_mutex);
1602 if (--ttm_bo_glob_use_count > 0)
1603 goto out;
1604
1605 kobject_del(&glob->kobj);
1606 kobject_put(&glob->kobj);
1607 ttm_mem_global_release(&ttm_mem_glob);
1608 memset(glob, 0, sizeof(*glob));
1609 out:
1610 mutex_unlock(&ttm_global_mutex);
1611 }
1612
1613 static int ttm_bo_global_init(void)
1614 {
1615 struct ttm_bo_global *glob = &ttm_bo_glob;
1616 int ret = 0;
1617 unsigned i;
1618
1619 mutex_lock(&ttm_global_mutex);
1620 if (++ttm_bo_glob_use_count > 1)
1621 goto out;
1622
1623 ret = ttm_mem_global_init(&ttm_mem_glob);
1624 if (ret)
1625 goto out;
1626
1627 spin_lock_init(&glob->lru_lock);
1628 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1629
1630 if (unlikely(glob->dummy_read_page == NULL)) {
1631 ret = -ENOMEM;
1632 goto out;
1633 }
1634
1635 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1636 INIT_LIST_HEAD(&glob->swap_lru[i]);
1637 INIT_LIST_HEAD(&glob->device_list);
1638 atomic_set(&glob->bo_count, 0);
1639
1640 ret = kobject_init_and_add(
1641 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1642 if (unlikely(ret != 0))
1643 kobject_put(&glob->kobj);
1644 out:
1645 mutex_unlock(&ttm_global_mutex);
1646 return ret;
1647 }
1648
1649 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1650 {
1651 struct ttm_bo_global *glob = &ttm_bo_glob;
1652 int ret = 0;
1653 unsigned i = TTM_NUM_MEM_TYPES;
1654 struct ttm_mem_type_manager *man;
1655
1656 while (i--) {
1657 man = &bdev->man[i];
1658 if (man->has_type) {
1659 man->use_type = false;
1660 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1661 ret = -EBUSY;
1662 pr_err("DRM memory manager type %d is not clean\n",
1663 i);
1664 }
1665 man->has_type = false;
1666 }
1667 }
1668
1669 mutex_lock(&ttm_global_mutex);
1670 list_del(&bdev->device_list);
1671 mutex_unlock(&ttm_global_mutex);
1672
1673 cancel_delayed_work_sync(&bdev->wq);
1674
1675 if (ttm_bo_delayed_delete(bdev, true))
1676 pr_debug("Delayed destroy list was clean\n");
1677
1678 spin_lock(&glob->lru_lock);
1679 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1680 if (list_empty(&bdev->man[0].lru[0]))
1681 pr_debug("Swap list %d was clean\n", i);
1682 spin_unlock(&glob->lru_lock);
1683
1684 if (!ret)
1685 ttm_bo_global_release();
1686
1687 return ret;
1688 }
1689 EXPORT_SYMBOL(ttm_bo_device_release);
1690
1691 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1692 struct ttm_bo_driver *driver,
1693 struct address_space *mapping,
1694 struct drm_vma_offset_manager *vma_manager,
1695 bool need_dma32)
1696 {
1697 struct ttm_bo_global *glob = &ttm_bo_glob;
1698 int ret;
1699
1700 if (WARN_ON(vma_manager == NULL))
1701 return -EINVAL;
1702
1703 ret = ttm_bo_global_init();
1704 if (ret)
1705 return ret;
1706
1707 bdev->driver = driver;
1708
1709 memset(bdev->man, 0, sizeof(bdev->man));
1710
1711 /*
1712 * Initialize the system memory buffer type.
1713 * Other types need to be driver / IOCTL initialized.
1714 */
1715 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1716 if (unlikely(ret != 0))
1717 goto out_no_sys;
1718
1719 bdev->vma_manager = vma_manager;
1720 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1721 INIT_LIST_HEAD(&bdev->ddestroy);
1722 bdev->dev_mapping = mapping;
1723 bdev->need_dma32 = need_dma32;
1724 mutex_lock(&ttm_global_mutex);
1725 list_add_tail(&bdev->device_list, &glob->device_list);
1726 mutex_unlock(&ttm_global_mutex);
1727
1728 return 0;
1729 out_no_sys:
1730 ttm_bo_global_release();
1731 return ret;
1732 }
1733 EXPORT_SYMBOL(ttm_bo_device_init);
1734
1735 /*
1736 * buffer object vm functions.
1737 */
1738
1739 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1740 {
1741 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1742
1743 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1744 if (mem->mem_type == TTM_PL_SYSTEM)
1745 return false;
1746
1747 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1748 return false;
1749
1750 if (mem->placement & TTM_PL_FLAG_CACHED)
1751 return false;
1752 }
1753 return true;
1754 }
1755
1756 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1757 {
1758 struct ttm_bo_device *bdev = bo->bdev;
1759
1760 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1761 ttm_mem_io_free_vm(bo);
1762 }
1763
1764 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1765 {
1766 struct ttm_bo_device *bdev = bo->bdev;
1767 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1768
1769 ttm_mem_io_lock(man, false);
1770 ttm_bo_unmap_virtual_locked(bo);
1771 ttm_mem_io_unlock(man);
1772 }
1773
1774
1775 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1776
1777 int ttm_bo_wait(struct ttm_buffer_object *bo,
1778 bool interruptible, bool no_wait)
1779 {
1780 long timeout = 15 * HZ;
1781
1782 if (no_wait) {
1783 if (dma_resv_test_signaled_rcu(bo->base.resv, true))
1784 return 0;
1785 else
1786 return -EBUSY;
1787 }
1788
1789 timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true,
1790 interruptible, timeout);
1791 if (timeout < 0)
1792 return timeout;
1793
1794 if (timeout == 0)
1795 return -EBUSY;
1796
1797 dma_resv_add_excl_fence(bo->base.resv, NULL);
1798 return 0;
1799 }
1800 EXPORT_SYMBOL(ttm_bo_wait);
1801
1802 /**
1803 * A buffer object shrink method that tries to swap out the first
1804 * buffer object on the bo_global::swap_lru list.
1805 */
1806 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1807 {
1808 struct ttm_buffer_object *bo;
1809 int ret = -EBUSY;
1810 bool locked;
1811 unsigned i;
1812
1813 spin_lock(&glob->lru_lock);
1814 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1815 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1816 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
1817 NULL)) {
1818 ret = 0;
1819 break;
1820 }
1821 }
1822 if (!ret)
1823 break;
1824 }
1825
1826 if (ret) {
1827 spin_unlock(&glob->lru_lock);
1828 return ret;
1829 }
1830
1831 kref_get(&bo->list_kref);
1832
1833 if (!list_empty(&bo->ddestroy)) {
1834 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1835 kref_put(&bo->list_kref, ttm_bo_release_list);
1836 return ret;
1837 }
1838
1839 ttm_bo_del_from_lru(bo);
1840 spin_unlock(&glob->lru_lock);
1841
1842 /**
1843 * Move to system cached
1844 */
1845
1846 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1847 bo->ttm->caching_state != tt_cached) {
1848 struct ttm_operation_ctx ctx = { false, false };
1849 struct ttm_mem_reg evict_mem;
1850
1851 evict_mem = bo->mem;
1852 evict_mem.mm_node = NULL;
1853 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1854 evict_mem.mem_type = TTM_PL_SYSTEM;
1855
1856 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1857 if (unlikely(ret != 0))
1858 goto out;
1859 }
1860
1861 /**
1862 * Make sure BO is idle.
1863 */
1864
1865 ret = ttm_bo_wait(bo, false, false);
1866 if (unlikely(ret != 0))
1867 goto out;
1868
1869 ttm_bo_unmap_virtual(bo);
1870
1871 /**
1872 * Swap out. Buffer will be swapped in again as soon as
1873 * anyone tries to access a ttm page.
1874 */
1875
1876 if (bo->bdev->driver->swap_notify)
1877 bo->bdev->driver->swap_notify(bo);
1878
1879 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1880 out:
1881
1882 /**
1883 *
1884 * Unreserve without putting on LRU to avoid swapping out an
1885 * already swapped buffer.
1886 */
1887 if (locked)
1888 dma_resv_unlock(bo->base.resv);
1889 kref_put(&bo->list_kref, ttm_bo_release_list);
1890 return ret;
1891 }
1892 EXPORT_SYMBOL(ttm_bo_swapout);
1893
1894 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1895 {
1896 struct ttm_operation_ctx ctx = {
1897 .interruptible = false,
1898 .no_wait_gpu = false
1899 };
1900
1901 while (ttm_bo_swapout(&ttm_bo_glob, &ctx) == 0);
1902 }
1903 EXPORT_SYMBOL(ttm_bo_swapout_all);
Linux with added FPC-III support