]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - drivers/gpu/drm/ttm/ttm_bo.c
projects / mirror_ubuntu-artful-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
drm/ttm: wait for BO idle in ttm_bo_move_memcpy
[mirror_ubuntu-artful-kernel.git] / drivers / gpu / drm / ttm / ttm_bo.c
1 /**************************************************************************
2 *
3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27 /*
28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29 */
30
31 #define pr_fmt(fmt) "[TTM] " fmt
32
33 #include <drm/ttm/ttm_module.h>
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <linux/jiffies.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 #include <linux/mm.h>
40 #include <linux/file.h>
41 #include <linux/module.h>
42 #include <linux/atomic.h>
43 #include <linux/reservation.h>
44
45 #define TTM_ASSERT_LOCKED(param)
46 #define TTM_DEBUG(fmt, arg...)
47 #define TTM_BO_HASH_ORDER 13
48
49 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50 static void ttm_bo_global_kobj_release(struct kobject *kobj);
51
52 static struct attribute ttm_bo_count = {
53 .name = "bo_count",
54 .mode = S_IRUGO
55 };
56
57 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
58 uint32_t *mem_type)
59 {
60 int i;
61
62 for (i = 0; i <= TTM_PL_PRIV5; i++)
63 if (place->flags & (1 << i)) {
64 *mem_type = i;
65 return 0;
66 }
67 return -EINVAL;
68 }
69
70 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
71 {
72 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
73
74 pr_err(" has_type: %d\n", man->has_type);
75 pr_err(" use_type: %d\n", man->use_type);
76 pr_err(" flags: 0x%08X\n", man->flags);
77 pr_err(" gpu_offset: 0x%08llX\n", man->gpu_offset);
78 pr_err(" size: %llu\n", man->size);
79 pr_err(" available_caching: 0x%08X\n", man->available_caching);
80 pr_err(" default_caching: 0x%08X\n", man->default_caching);
81 if (mem_type != TTM_PL_SYSTEM)
82 (*man->func->debug)(man, TTM_PFX);
83 }
84
85 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
86 struct ttm_placement *placement)
87 {
88 int i, ret, mem_type;
89
90 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
91 bo, bo->mem.num_pages, bo->mem.size >> 10,
92 bo->mem.size >> 20);
93 for (i = 0; i < placement->num_placement; i++) {
94 ret = ttm_mem_type_from_place(&placement->placement[i],
95 &mem_type);
96 if (ret)
97 return;
98 pr_err(" placement[%d]=0x%08X (%d)\n",
99 i, placement->placement[i].flags, mem_type);
100 ttm_mem_type_debug(bo->bdev, mem_type);
101 }
102 }
103
104 static ssize_t ttm_bo_global_show(struct kobject *kobj,
105 struct attribute *attr,
106 char *buffer)
107 {
108 struct ttm_bo_global *glob =
109 container_of(kobj, struct ttm_bo_global, kobj);
110
111 return snprintf(buffer, PAGE_SIZE, "%lu\n",
112 (unsigned long) atomic_read(&glob->bo_count));
113 }
114
115 static struct attribute *ttm_bo_global_attrs[] = {
116 &ttm_bo_count,
117 NULL
118 };
119
120 static const struct sysfs_ops ttm_bo_global_ops = {
121 .show = &ttm_bo_global_show
122 };
123
124 static struct kobj_type ttm_bo_glob_kobj_type = {
125 .release = &ttm_bo_global_kobj_release,
126 .sysfs_ops = &ttm_bo_global_ops,
127 .default_attrs = ttm_bo_global_attrs
128 };
129
130
131 static inline uint32_t ttm_bo_type_flags(unsigned type)
132 {
133 return 1 << (type);
134 }
135
136 static void ttm_bo_release_list(struct kref *list_kref)
137 {
138 struct ttm_buffer_object *bo =
139 container_of(list_kref, struct ttm_buffer_object, list_kref);
140 struct ttm_bo_device *bdev = bo->bdev;
141 size_t acc_size = bo->acc_size;
142
143 BUG_ON(atomic_read(&bo->list_kref.refcount));
144 BUG_ON(atomic_read(&bo->kref.refcount));
145 BUG_ON(atomic_read(&bo->cpu_writers));
146 BUG_ON(bo->mem.mm_node != NULL);
147 BUG_ON(!list_empty(&bo->lru));
148 BUG_ON(!list_empty(&bo->ddestroy));
149
150 ttm_tt_destroy(bo->ttm);
151 atomic_dec(&bo->glob->bo_count);
152 if (bo->resv == &bo->ttm_resv)
153 reservation_object_fini(&bo->ttm_resv);
154 mutex_destroy(&bo->wu_mutex);
155 if (bo->destroy)
156 bo->destroy(bo);
157 else {
158 kfree(bo);
159 }
160 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
161 }
162
163 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
164 {
165 struct ttm_bo_device *bdev = bo->bdev;
166
167 lockdep_assert_held(&bo->resv->lock.base);
168
169 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
170
171 BUG_ON(!list_empty(&bo->lru));
172
173 list_add(&bo->lru, bdev->driver->lru_tail(bo));
174 kref_get(&bo->list_kref);
175
176 if (bo->ttm && !(bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
177 list_add(&bo->swap, bdev->driver->swap_lru_tail(bo));
178 kref_get(&bo->list_kref);
179 }
180 }
181 }
182 EXPORT_SYMBOL(ttm_bo_add_to_lru);
183
184 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
185 {
186 struct ttm_bo_device *bdev = bo->bdev;
187 int put_count = 0;
188
189 if (bdev->driver->lru_removal)
190 bdev->driver->lru_removal(bo);
191
192 if (!list_empty(&bo->swap)) {
193 list_del_init(&bo->swap);
194 ++put_count;
195 }
196 if (!list_empty(&bo->lru)) {
197 list_del_init(&bo->lru);
198 ++put_count;
199 }
200
201 return put_count;
202 }
203
204 static void ttm_bo_ref_bug(struct kref *list_kref)
205 {
206 BUG();
207 }
208
209 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
210 bool never_free)
211 {
212 kref_sub(&bo->list_kref, count,
213 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
214 }
215
216 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
217 {
218 int put_count;
219
220 spin_lock(&bo->glob->lru_lock);
221 put_count = ttm_bo_del_from_lru(bo);
222 spin_unlock(&bo->glob->lru_lock);
223 ttm_bo_list_ref_sub(bo, put_count, true);
224 }
225 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
226
227 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
228 {
229 struct ttm_bo_device *bdev = bo->bdev;
230 int put_count = 0;
231
232 lockdep_assert_held(&bo->resv->lock.base);
233
234 if (bdev->driver->lru_removal)
235 bdev->driver->lru_removal(bo);
236
237 put_count = ttm_bo_del_from_lru(bo);
238 ttm_bo_list_ref_sub(bo, put_count, true);
239 ttm_bo_add_to_lru(bo);
240 }
241 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
242
243 struct list_head *ttm_bo_default_lru_tail(struct ttm_buffer_object *bo)
244 {
245 return bo->bdev->man[bo->mem.mem_type].lru.prev;
246 }
247 EXPORT_SYMBOL(ttm_bo_default_lru_tail);
248
249 struct list_head *ttm_bo_default_swap_lru_tail(struct ttm_buffer_object *bo)
250 {
251 return bo->glob->swap_lru.prev;
252 }
253 EXPORT_SYMBOL(ttm_bo_default_swap_lru_tail);
254
255 /*
256 * Call bo->mutex locked.
257 */
258 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
259 {
260 struct ttm_bo_device *bdev = bo->bdev;
261 struct ttm_bo_global *glob = bo->glob;
262 int ret = 0;
263 uint32_t page_flags = 0;
264
265 TTM_ASSERT_LOCKED(&bo->mutex);
266 bo->ttm = NULL;
267
268 if (bdev->need_dma32)
269 page_flags |= TTM_PAGE_FLAG_DMA32;
270
271 switch (bo->type) {
272 case ttm_bo_type_device:
273 if (zero_alloc)
274 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
275 case ttm_bo_type_kernel:
276 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
277 page_flags, glob->dummy_read_page);
278 if (unlikely(bo->ttm == NULL))
279 ret = -ENOMEM;
280 break;
281 case ttm_bo_type_sg:
282 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
283 page_flags | TTM_PAGE_FLAG_SG,
284 glob->dummy_read_page);
285 if (unlikely(bo->ttm == NULL)) {
286 ret = -ENOMEM;
287 break;
288 }
289 bo->ttm->sg = bo->sg;
290 break;
291 default:
292 pr_err("Illegal buffer object type\n");
293 ret = -EINVAL;
294 break;
295 }
296
297 return ret;
298 }
299
300 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
301 struct ttm_mem_reg *mem,
302 bool evict, bool interruptible,
303 bool no_wait_gpu)
304 {
305 struct ttm_bo_device *bdev = bo->bdev;
306 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
307 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
308 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
309 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
310 int ret = 0;
311
312 if (old_is_pci || new_is_pci ||
313 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
314 ret = ttm_mem_io_lock(old_man, true);
315 if (unlikely(ret != 0))
316 goto out_err;
317 ttm_bo_unmap_virtual_locked(bo);
318 ttm_mem_io_unlock(old_man);
319 }
320
321 /*
322 * Create and bind a ttm if required.
323 */
324
325 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
326 if (bo->ttm == NULL) {
327 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
328 ret = ttm_bo_add_ttm(bo, zero);
329 if (ret)
330 goto out_err;
331 }
332
333 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
334 if (ret)
335 goto out_err;
336
337 if (mem->mem_type != TTM_PL_SYSTEM) {
338 ret = ttm_tt_bind(bo->ttm, mem);
339 if (ret)
340 goto out_err;
341 }
342
343 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
344 if (bdev->driver->move_notify)
345 bdev->driver->move_notify(bo, mem);
346 bo->mem = *mem;
347 mem->mm_node = NULL;
348 goto moved;
349 }
350 }
351
352 if (bdev->driver->move_notify)
353 bdev->driver->move_notify(bo, mem);
354
355 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
356 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
357 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
358 else if (bdev->driver->move)
359 ret = bdev->driver->move(bo, evict, interruptible,
360 no_wait_gpu, mem);
361 else
362 ret = ttm_bo_move_memcpy(bo, evict, interruptible,
363 no_wait_gpu, mem);
364
365 if (ret) {
366 if (bdev->driver->move_notify) {
367 struct ttm_mem_reg tmp_mem = *mem;
368 *mem = bo->mem;
369 bo->mem = tmp_mem;
370 bdev->driver->move_notify(bo, mem);
371 bo->mem = *mem;
372 *mem = tmp_mem;
373 }
374
375 goto out_err;
376 }
377
378 moved:
379 if (bo->evicted) {
380 if (bdev->driver->invalidate_caches) {
381 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
382 if (ret)
383 pr_err("Can not flush read caches\n");
384 }
385 bo->evicted = false;
386 }
387
388 if (bo->mem.mm_node) {
389 bo->offset = (bo->mem.start << PAGE_SHIFT) +
390 bdev->man[bo->mem.mem_type].gpu_offset;
391 bo->cur_placement = bo->mem.placement;
392 } else
393 bo->offset = 0;
394
395 return 0;
396
397 out_err:
398 new_man = &bdev->man[bo->mem.mem_type];
399 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
400 ttm_tt_destroy(bo->ttm);
401 bo->ttm = NULL;
402 }
403
404 return ret;
405 }
406
407 /**
408 * Call bo::reserved.
409 * Will release GPU memory type usage on destruction.
410 * This is the place to put in driver specific hooks to release
411 * driver private resources.
412 * Will release the bo::reserved lock.
413 */
414
415 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
416 {
417 if (bo->bdev->driver->move_notify)
418 bo->bdev->driver->move_notify(bo, NULL);
419
420 ttm_tt_destroy(bo->ttm);
421 bo->ttm = NULL;
422 ttm_bo_mem_put(bo, &bo->mem);
423
424 ww_mutex_unlock (&bo->resv->lock);
425 }
426
427 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
428 {
429 struct reservation_object_list *fobj;
430 struct fence *fence;
431 int i;
432
433 fobj = reservation_object_get_list(bo->resv);
434 fence = reservation_object_get_excl(bo->resv);
435 if (fence && !fence->ops->signaled)
436 fence_enable_sw_signaling(fence);
437
438 for (i = 0; fobj && i < fobj->shared_count; ++i) {
439 fence = rcu_dereference_protected(fobj->shared[i],
440 reservation_object_held(bo->resv));
441
442 if (!fence->ops->signaled)
443 fence_enable_sw_signaling(fence);
444 }
445 }
446
447 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
448 {
449 struct ttm_bo_device *bdev = bo->bdev;
450 struct ttm_bo_global *glob = bo->glob;
451 int put_count;
452 int ret;
453
454 spin_lock(&glob->lru_lock);
455 ret = __ttm_bo_reserve(bo, false, true, NULL);
456
457 if (!ret) {
458 if (!ttm_bo_wait(bo, false, true)) {
459 put_count = ttm_bo_del_from_lru(bo);
460
461 spin_unlock(&glob->lru_lock);
462 ttm_bo_cleanup_memtype_use(bo);
463
464 ttm_bo_list_ref_sub(bo, put_count, true);
465
466 return;
467 } else
468 ttm_bo_flush_all_fences(bo);
469
470 /*
471 * Make NO_EVICT bos immediately available to
472 * shrinkers, now that they are queued for
473 * destruction.
474 */
475 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
476 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
477 ttm_bo_add_to_lru(bo);
478 }
479
480 __ttm_bo_unreserve(bo);
481 }
482
483 kref_get(&bo->list_kref);
484 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
485 spin_unlock(&glob->lru_lock);
486
487 schedule_delayed_work(&bdev->wq,
488 ((HZ / 100) < 1) ? 1 : HZ / 100);
489 }
490
491 /**
492 * function ttm_bo_cleanup_refs_and_unlock
493 * If bo idle, remove from delayed- and lru lists, and unref.
494 * If not idle, do nothing.
495 *
496 * Must be called with lru_lock and reservation held, this function
497 * will drop both before returning.
498 *
499 * @interruptible Any sleeps should occur interruptibly.
500 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
501 */
502
503 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
504 bool interruptible,
505 bool no_wait_gpu)
506 {
507 struct ttm_bo_global *glob = bo->glob;
508 int put_count;
509 int ret;
510
511 ret = ttm_bo_wait(bo, false, true);
512
513 if (ret && !no_wait_gpu) {
514 long lret;
515 ww_mutex_unlock(&bo->resv->lock);
516 spin_unlock(&glob->lru_lock);
517
518 lret = reservation_object_wait_timeout_rcu(bo->resv,
519 true,
520 interruptible,
521 30 * HZ);
522
523 if (lret < 0)
524 return lret;
525 else if (lret == 0)
526 return -EBUSY;
527
528 spin_lock(&glob->lru_lock);
529 ret = __ttm_bo_reserve(bo, false, true, NULL);
530
531 /*
532 * We raced, and lost, someone else holds the reservation now,
533 * and is probably busy in ttm_bo_cleanup_memtype_use.
534 *
535 * Even if it's not the case, because we finished waiting any
536 * delayed destruction would succeed, so just return success
537 * here.
538 */
539 if (ret) {
540 spin_unlock(&glob->lru_lock);
541 return 0;
542 }
543
544 /*
545 * remove sync_obj with ttm_bo_wait, the wait should be
546 * finished, and no new wait object should have been added.
547 */
548 ret = ttm_bo_wait(bo, false, true);
549 WARN_ON(ret);
550 }
551
552 if (ret || unlikely(list_empty(&bo->ddestroy))) {
553 __ttm_bo_unreserve(bo);
554 spin_unlock(&glob->lru_lock);
555 return ret;
556 }
557
558 put_count = ttm_bo_del_from_lru(bo);
559 list_del_init(&bo->ddestroy);
560 ++put_count;
561
562 spin_unlock(&glob->lru_lock);
563 ttm_bo_cleanup_memtype_use(bo);
564
565 ttm_bo_list_ref_sub(bo, put_count, true);
566
567 return 0;
568 }
569
570 /**
571 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
572 * encountered buffers.
573 */
574
575 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
576 {
577 struct ttm_bo_global *glob = bdev->glob;
578 struct ttm_buffer_object *entry = NULL;
579 int ret = 0;
580
581 spin_lock(&glob->lru_lock);
582 if (list_empty(&bdev->ddestroy))
583 goto out_unlock;
584
585 entry = list_first_entry(&bdev->ddestroy,
586 struct ttm_buffer_object, ddestroy);
587 kref_get(&entry->list_kref);
588
589 for (;;) {
590 struct ttm_buffer_object *nentry = NULL;
591
592 if (entry->ddestroy.next != &bdev->ddestroy) {
593 nentry = list_first_entry(&entry->ddestroy,
594 struct ttm_buffer_object, ddestroy);
595 kref_get(&nentry->list_kref);
596 }
597
598 ret = __ttm_bo_reserve(entry, false, true, NULL);
599 if (remove_all && ret) {
600 spin_unlock(&glob->lru_lock);
601 ret = __ttm_bo_reserve(entry, false, false, NULL);
602 spin_lock(&glob->lru_lock);
603 }
604
605 if (!ret)
606 ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
607 !remove_all);
608 else
609 spin_unlock(&glob->lru_lock);
610
611 kref_put(&entry->list_kref, ttm_bo_release_list);
612 entry = nentry;
613
614 if (ret || !entry)
615 goto out;
616
617 spin_lock(&glob->lru_lock);
618 if (list_empty(&entry->ddestroy))
619 break;
620 }
621
622 out_unlock:
623 spin_unlock(&glob->lru_lock);
624 out:
625 if (entry)
626 kref_put(&entry->list_kref, ttm_bo_release_list);
627 return ret;
628 }
629
630 static void ttm_bo_delayed_workqueue(struct work_struct *work)
631 {
632 struct ttm_bo_device *bdev =
633 container_of(work, struct ttm_bo_device, wq.work);
634
635 if (ttm_bo_delayed_delete(bdev, false)) {
636 schedule_delayed_work(&bdev->wq,
637 ((HZ / 100) < 1) ? 1 : HZ / 100);
638 }
639 }
640
641 static void ttm_bo_release(struct kref *kref)
642 {
643 struct ttm_buffer_object *bo =
644 container_of(kref, struct ttm_buffer_object, kref);
645 struct ttm_bo_device *bdev = bo->bdev;
646 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
647
648 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
649 ttm_mem_io_lock(man, false);
650 ttm_mem_io_free_vm(bo);
651 ttm_mem_io_unlock(man);
652 ttm_bo_cleanup_refs_or_queue(bo);
653 kref_put(&bo->list_kref, ttm_bo_release_list);
654 }
655
656 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
657 {
658 struct ttm_buffer_object *bo = *p_bo;
659
660 *p_bo = NULL;
661 kref_put(&bo->kref, ttm_bo_release);
662 }
663 EXPORT_SYMBOL(ttm_bo_unref);
664
665 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
666 {
667 return cancel_delayed_work_sync(&bdev->wq);
668 }
669 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
670
671 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
672 {
673 if (resched)
674 schedule_delayed_work(&bdev->wq,
675 ((HZ / 100) < 1) ? 1 : HZ / 100);
676 }
677 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
678
679 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
680 bool no_wait_gpu)
681 {
682 struct ttm_bo_device *bdev = bo->bdev;
683 struct ttm_mem_reg evict_mem;
684 struct ttm_placement placement;
685 int ret = 0;
686
687 ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
688
689 if (unlikely(ret != 0)) {
690 if (ret != -ERESTARTSYS) {
691 pr_err("Failed to expire sync object before buffer eviction\n");
692 }
693 goto out;
694 }
695
696 lockdep_assert_held(&bo->resv->lock.base);
697
698 evict_mem = bo->mem;
699 evict_mem.mm_node = NULL;
700 evict_mem.bus.io_reserved_vm = false;
701 evict_mem.bus.io_reserved_count = 0;
702
703 placement.num_placement = 0;
704 placement.num_busy_placement = 0;
705 bdev->driver->evict_flags(bo, &placement);
706 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
707 no_wait_gpu);
708 if (ret) {
709 if (ret != -ERESTARTSYS) {
710 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
711 bo);
712 ttm_bo_mem_space_debug(bo, &placement);
713 }
714 goto out;
715 }
716
717 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
718 no_wait_gpu);
719 if (ret) {
720 if (ret != -ERESTARTSYS)
721 pr_err("Buffer eviction failed\n");
722 ttm_bo_mem_put(bo, &evict_mem);
723 goto out;
724 }
725 bo->evicted = true;
726 out:
727 return ret;
728 }
729
730 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
731 uint32_t mem_type,
732 const struct ttm_place *place,
733 bool interruptible,
734 bool no_wait_gpu)
735 {
736 struct ttm_bo_global *glob = bdev->glob;
737 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
738 struct ttm_buffer_object *bo;
739 int ret = -EBUSY, put_count;
740
741 spin_lock(&glob->lru_lock);
742 list_for_each_entry(bo, &man->lru, lru) {
743 ret = __ttm_bo_reserve(bo, false, true, NULL);
744 if (!ret) {
745 if (place && (place->fpfn || place->lpfn)) {
746 /* Don't evict this BO if it's outside of the
747 * requested placement range
748 */
749 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
750 (place->lpfn && place->lpfn <= bo->mem.start)) {
751 __ttm_bo_unreserve(bo);
752 ret = -EBUSY;
753 continue;
754 }
755 }
756
757 break;
758 }
759 }
760
761 if (ret) {
762 spin_unlock(&glob->lru_lock);
763 return ret;
764 }
765
766 kref_get(&bo->list_kref);
767
768 if (!list_empty(&bo->ddestroy)) {
769 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
770 no_wait_gpu);
771 kref_put(&bo->list_kref, ttm_bo_release_list);
772 return ret;
773 }
774
775 put_count = ttm_bo_del_from_lru(bo);
776 spin_unlock(&glob->lru_lock);
777
778 BUG_ON(ret != 0);
779
780 ttm_bo_list_ref_sub(bo, put_count, true);
781
782 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
783 ttm_bo_unreserve(bo);
784
785 kref_put(&bo->list_kref, ttm_bo_release_list);
786 return ret;
787 }
788
789 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
790 {
791 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
792
793 if (mem->mm_node)
794 (*man->func->put_node)(man, mem);
795 }
796 EXPORT_SYMBOL(ttm_bo_mem_put);
797
798 /**
799 * Repeatedly evict memory from the LRU for @mem_type until we create enough
800 * space, or we've evicted everything and there isn't enough space.
801 */
802 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
803 uint32_t mem_type,
804 const struct ttm_place *place,
805 struct ttm_mem_reg *mem,
806 bool interruptible,
807 bool no_wait_gpu)
808 {
809 struct ttm_bo_device *bdev = bo->bdev;
810 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
811 int ret;
812
813 do {
814 ret = (*man->func->get_node)(man, bo, place, mem);
815 if (unlikely(ret != 0))
816 return ret;
817 if (mem->mm_node)
818 break;
819 ret = ttm_mem_evict_first(bdev, mem_type, place,
820 interruptible, no_wait_gpu);
821 if (unlikely(ret != 0))
822 return ret;
823 } while (1);
824 if (mem->mm_node == NULL)
825 return -ENOMEM;
826 mem->mem_type = mem_type;
827 return 0;
828 }
829
830 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
831 uint32_t cur_placement,
832 uint32_t proposed_placement)
833 {
834 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
835 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
836
837 /**
838 * Keep current caching if possible.
839 */
840
841 if ((cur_placement & caching) != 0)
842 result |= (cur_placement & caching);
843 else if ((man->default_caching & caching) != 0)
844 result |= man->default_caching;
845 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
846 result |= TTM_PL_FLAG_CACHED;
847 else if ((TTM_PL_FLAG_WC & caching) != 0)
848 result |= TTM_PL_FLAG_WC;
849 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
850 result |= TTM_PL_FLAG_UNCACHED;
851
852 return result;
853 }
854
855 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
856 uint32_t mem_type,
857 const struct ttm_place *place,
858 uint32_t *masked_placement)
859 {
860 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
861
862 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
863 return false;
864
865 if ((place->flags & man->available_caching) == 0)
866 return false;
867
868 cur_flags |= (place->flags & man->available_caching);
869
870 *masked_placement = cur_flags;
871 return true;
872 }
873
874 /**
875 * Creates space for memory region @mem according to its type.
876 *
877 * This function first searches for free space in compatible memory types in
878 * the priority order defined by the driver. If free space isn't found, then
879 * ttm_bo_mem_force_space is attempted in priority order to evict and find
880 * space.
881 */
882 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
883 struct ttm_placement *placement,
884 struct ttm_mem_reg *mem,
885 bool interruptible,
886 bool no_wait_gpu)
887 {
888 struct ttm_bo_device *bdev = bo->bdev;
889 struct ttm_mem_type_manager *man;
890 uint32_t mem_type = TTM_PL_SYSTEM;
891 uint32_t cur_flags = 0;
892 bool type_found = false;
893 bool type_ok = false;
894 bool has_erestartsys = false;
895 int i, ret;
896
897 mem->mm_node = NULL;
898 for (i = 0; i < placement->num_placement; ++i) {
899 const struct ttm_place *place = &placement->placement[i];
900
901 ret = ttm_mem_type_from_place(place, &mem_type);
902 if (ret)
903 return ret;
904 man = &bdev->man[mem_type];
905 if (!man->has_type || !man->use_type)
906 continue;
907
908 type_ok = ttm_bo_mt_compatible(man, mem_type, place,
909 &cur_flags);
910
911 if (!type_ok)
912 continue;
913
914 type_found = true;
915 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
916 cur_flags);
917 /*
918 * Use the access and other non-mapping-related flag bits from
919 * the memory placement flags to the current flags
920 */
921 ttm_flag_masked(&cur_flags, place->flags,
922 ~TTM_PL_MASK_MEMTYPE);
923
924 if (mem_type == TTM_PL_SYSTEM)
925 break;
926
927 ret = (*man->func->get_node)(man, bo, place, mem);
928 if (unlikely(ret))
929 return ret;
930
931 if (mem->mm_node)
932 break;
933 }
934
935 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
936 mem->mem_type = mem_type;
937 mem->placement = cur_flags;
938 return 0;
939 }
940
941 for (i = 0; i < placement->num_busy_placement; ++i) {
942 const struct ttm_place *place = &placement->busy_placement[i];
943
944 ret = ttm_mem_type_from_place(place, &mem_type);
945 if (ret)
946 return ret;
947 man = &bdev->man[mem_type];
948 if (!man->has_type || !man->use_type)
949 continue;
950 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
951 continue;
952
953 type_found = true;
954 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
955 cur_flags);
956 /*
957 * Use the access and other non-mapping-related flag bits from
958 * the memory placement flags to the current flags
959 */
960 ttm_flag_masked(&cur_flags, place->flags,
961 ~TTM_PL_MASK_MEMTYPE);
962
963 if (mem_type == TTM_PL_SYSTEM) {
964 mem->mem_type = mem_type;
965 mem->placement = cur_flags;
966 mem->mm_node = NULL;
967 return 0;
968 }
969
970 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
971 interruptible, no_wait_gpu);
972 if (ret == 0 && mem->mm_node) {
973 mem->placement = cur_flags;
974 return 0;
975 }
976 if (ret == -ERESTARTSYS)
977 has_erestartsys = true;
978 }
979
980 if (!type_found) {
981 printk(KERN_ERR TTM_PFX "No compatible memory type found.\n");
982 return -EINVAL;
983 }
984
985 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
986 }
987 EXPORT_SYMBOL(ttm_bo_mem_space);
988
989 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
990 struct ttm_placement *placement,
991 bool interruptible,
992 bool no_wait_gpu)
993 {
994 int ret = 0;
995 struct ttm_mem_reg mem;
996
997 lockdep_assert_held(&bo->resv->lock.base);
998
999 /*
1000 * Don't wait for the BO on initial allocation. This is important when
1001 * the BO has an imported reservation object.
1002 */
1003 if (bo->mem.mem_type != TTM_PL_SYSTEM || bo->ttm != NULL) {
1004 /*
1005 * FIXME: It's possible to pipeline buffer moves.
1006 * Have the driver move function wait for idle when necessary,
1007 * instead of doing it here.
1008 */
1009 ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
1010 if (ret)
1011 return ret;
1012 }
1013 mem.num_pages = bo->num_pages;
1014 mem.size = mem.num_pages << PAGE_SHIFT;
1015 mem.page_alignment = bo->mem.page_alignment;
1016 mem.bus.io_reserved_vm = false;
1017 mem.bus.io_reserved_count = 0;
1018 /*
1019 * Determine where to move the buffer.
1020 */
1021 ret = ttm_bo_mem_space(bo, placement, &mem,
1022 interruptible, no_wait_gpu);
1023 if (ret)
1024 goto out_unlock;
1025 ret = ttm_bo_handle_move_mem(bo, &mem, false,
1026 interruptible, no_wait_gpu);
1027 out_unlock:
1028 if (ret && mem.mm_node)
1029 ttm_bo_mem_put(bo, &mem);
1030 return ret;
1031 }
1032
1033 static bool ttm_bo_mem_compat(struct ttm_placement *placement,
1034 struct ttm_mem_reg *mem,
1035 uint32_t *new_flags)
1036 {
1037 int i;
1038
1039 for (i = 0; i < placement->num_placement; i++) {
1040 const struct ttm_place *heap = &placement->placement[i];
1041 if (mem->mm_node &&
1042 (mem->start < heap->fpfn ||
1043 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1044 continue;
1045
1046 *new_flags = heap->flags;
1047 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1048 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1049 return true;
1050 }
1051
1052 for (i = 0; i < placement->num_busy_placement; i++) {
1053 const struct ttm_place *heap = &placement->busy_placement[i];
1054 if (mem->mm_node &&
1055 (mem->start < heap->fpfn ||
1056 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1057 continue;
1058
1059 *new_flags = heap->flags;
1060 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1061 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1062 return true;
1063 }
1064
1065 return false;
1066 }
1067
1068 int ttm_bo_validate(struct ttm_buffer_object *bo,
1069 struct ttm_placement *placement,
1070 bool interruptible,
1071 bool no_wait_gpu)
1072 {
1073 int ret;
1074 uint32_t new_flags;
1075
1076 lockdep_assert_held(&bo->resv->lock.base);
1077 /*
1078 * Check whether we need to move buffer.
1079 */
1080 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1081 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1082 no_wait_gpu);
1083 if (ret)
1084 return ret;
1085 } else {
1086 /*
1087 * Use the access and other non-mapping-related flag bits from
1088 * the compatible memory placement flags to the active flags
1089 */
1090 ttm_flag_masked(&bo->mem.placement, new_flags,
1091 ~TTM_PL_MASK_MEMTYPE);
1092 }
1093 /*
1094 * We might need to add a TTM.
1095 */
1096 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1097 ret = ttm_bo_add_ttm(bo, true);
1098 if (ret)
1099 return ret;
1100 }
1101 return 0;
1102 }
1103 EXPORT_SYMBOL(ttm_bo_validate);
1104
1105 int ttm_bo_init(struct ttm_bo_device *bdev,
1106 struct ttm_buffer_object *bo,
1107 unsigned long size,
1108 enum ttm_bo_type type,
1109 struct ttm_placement *placement,
1110 uint32_t page_alignment,
1111 bool interruptible,
1112 struct file *persistent_swap_storage,
1113 size_t acc_size,
1114 struct sg_table *sg,
1115 struct reservation_object *resv,
1116 void (*destroy) (struct ttm_buffer_object *))
1117 {
1118 int ret = 0;
1119 unsigned long num_pages;
1120 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1121 bool locked;
1122
1123 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1124 if (ret) {
1125 pr_err("Out of kernel memory\n");
1126 if (destroy)
1127 (*destroy)(bo);
1128 else
1129 kfree(bo);
1130 return -ENOMEM;
1131 }
1132
1133 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1134 if (num_pages == 0) {
1135 pr_err("Illegal buffer object size\n");
1136 if (destroy)
1137 (*destroy)(bo);
1138 else
1139 kfree(bo);
1140 ttm_mem_global_free(mem_glob, acc_size);
1141 return -EINVAL;
1142 }
1143 bo->destroy = destroy;
1144
1145 kref_init(&bo->kref);
1146 kref_init(&bo->list_kref);
1147 atomic_set(&bo->cpu_writers, 0);
1148 INIT_LIST_HEAD(&bo->lru);
1149 INIT_LIST_HEAD(&bo->ddestroy);
1150 INIT_LIST_HEAD(&bo->swap);
1151 INIT_LIST_HEAD(&bo->io_reserve_lru);
1152 mutex_init(&bo->wu_mutex);
1153 bo->bdev = bdev;
1154 bo->glob = bdev->glob;
1155 bo->type = type;
1156 bo->num_pages = num_pages;
1157 bo->mem.size = num_pages << PAGE_SHIFT;
1158 bo->mem.mem_type = TTM_PL_SYSTEM;
1159 bo->mem.num_pages = bo->num_pages;
1160 bo->mem.mm_node = NULL;
1161 bo->mem.page_alignment = page_alignment;
1162 bo->mem.bus.io_reserved_vm = false;
1163 bo->mem.bus.io_reserved_count = 0;
1164 bo->priv_flags = 0;
1165 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1166 bo->persistent_swap_storage = persistent_swap_storage;
1167 bo->acc_size = acc_size;
1168 bo->sg = sg;
1169 if (resv) {
1170 bo->resv = resv;
1171 lockdep_assert_held(&bo->resv->lock.base);
1172 } else {
1173 bo->resv = &bo->ttm_resv;
1174 reservation_object_init(&bo->ttm_resv);
1175 }
1176 atomic_inc(&bo->glob->bo_count);
1177 drm_vma_node_reset(&bo->vma_node);
1178
1179 /*
1180 * For ttm_bo_type_device buffers, allocate
1181 * address space from the device.
1182 */
1183 if (bo->type == ttm_bo_type_device ||
1184 bo->type == ttm_bo_type_sg)
1185 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1186 bo->mem.num_pages);
1187
1188 /* passed reservation objects should already be locked,
1189 * since otherwise lockdep will be angered in radeon.
1190 */
1191 if (!resv) {
1192 locked = ww_mutex_trylock(&bo->resv->lock);
1193 WARN_ON(!locked);
1194 }
1195
1196 if (likely(!ret))
1197 ret = ttm_bo_validate(bo, placement, interruptible, false);
1198
1199 if (!resv) {
1200 ttm_bo_unreserve(bo);
1201
1202 } else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1203 spin_lock(&bo->glob->lru_lock);
1204 ttm_bo_add_to_lru(bo);
1205 spin_unlock(&bo->glob->lru_lock);
1206 }
1207
1208 if (unlikely(ret))
1209 ttm_bo_unref(&bo);
1210
1211 return ret;
1212 }
1213 EXPORT_SYMBOL(ttm_bo_init);
1214
1215 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1216 unsigned long bo_size,
1217 unsigned struct_size)
1218 {
1219 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1220 size_t size = 0;
1221
1222 size += ttm_round_pot(struct_size);
1223 size += ttm_round_pot(npages * sizeof(void *));
1224 size += ttm_round_pot(sizeof(struct ttm_tt));
1225 return size;
1226 }
1227 EXPORT_SYMBOL(ttm_bo_acc_size);
1228
1229 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1230 unsigned long bo_size,
1231 unsigned struct_size)
1232 {
1233 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1234 size_t size = 0;
1235
1236 size += ttm_round_pot(struct_size);
1237 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1238 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1239 return size;
1240 }
1241 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1242
1243 int ttm_bo_create(struct ttm_bo_device *bdev,
1244 unsigned long size,
1245 enum ttm_bo_type type,
1246 struct ttm_placement *placement,
1247 uint32_t page_alignment,
1248 bool interruptible,
1249 struct file *persistent_swap_storage,
1250 struct ttm_buffer_object **p_bo)
1251 {
1252 struct ttm_buffer_object *bo;
1253 size_t acc_size;
1254 int ret;
1255
1256 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1257 if (unlikely(bo == NULL))
1258 return -ENOMEM;
1259
1260 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1261 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1262 interruptible, persistent_swap_storage, acc_size,
1263 NULL, NULL, NULL);
1264 if (likely(ret == 0))
1265 *p_bo = bo;
1266
1267 return ret;
1268 }
1269 EXPORT_SYMBOL(ttm_bo_create);
1270
1271 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1272 unsigned mem_type, bool allow_errors)
1273 {
1274 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1275 struct ttm_bo_global *glob = bdev->glob;
1276 int ret;
1277
1278 /*
1279 * Can't use standard list traversal since we're unlocking.
1280 */
1281
1282 spin_lock(&glob->lru_lock);
1283 while (!list_empty(&man->lru)) {
1284 spin_unlock(&glob->lru_lock);
1285 ret = ttm_mem_evict_first(bdev, mem_type, NULL, false, false);
1286 if (ret) {
1287 if (allow_errors) {
1288 return ret;
1289 } else {
1290 pr_err("Cleanup eviction failed\n");
1291 }
1292 }
1293 spin_lock(&glob->lru_lock);
1294 }
1295 spin_unlock(&glob->lru_lock);
1296 return 0;
1297 }
1298
1299 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1300 {
1301 struct ttm_mem_type_manager *man;
1302 int ret = -EINVAL;
1303
1304 if (mem_type >= TTM_NUM_MEM_TYPES) {
1305 pr_err("Illegal memory type %d\n", mem_type);
1306 return ret;
1307 }
1308 man = &bdev->man[mem_type];
1309
1310 if (!man->has_type) {
1311 pr_err("Trying to take down uninitialized memory manager type %u\n",
1312 mem_type);
1313 return ret;
1314 }
1315
1316 man->use_type = false;
1317 man->has_type = false;
1318
1319 ret = 0;
1320 if (mem_type > 0) {
1321 ttm_bo_force_list_clean(bdev, mem_type, false);
1322
1323 ret = (*man->func->takedown)(man);
1324 }
1325
1326 return ret;
1327 }
1328 EXPORT_SYMBOL(ttm_bo_clean_mm);
1329
1330 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1331 {
1332 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1333
1334 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1335 pr_err("Illegal memory manager memory type %u\n", mem_type);
1336 return -EINVAL;
1337 }
1338
1339 if (!man->has_type) {
1340 pr_err("Memory type %u has not been initialized\n", mem_type);
1341 return 0;
1342 }
1343
1344 return ttm_bo_force_list_clean(bdev, mem_type, true);
1345 }
1346 EXPORT_SYMBOL(ttm_bo_evict_mm);
1347
1348 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1349 unsigned long p_size)
1350 {
1351 int ret = -EINVAL;
1352 struct ttm_mem_type_manager *man;
1353
1354 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1355 man = &bdev->man[type];
1356 BUG_ON(man->has_type);
1357 man->io_reserve_fastpath = true;
1358 man->use_io_reserve_lru = false;
1359 mutex_init(&man->io_reserve_mutex);
1360 INIT_LIST_HEAD(&man->io_reserve_lru);
1361
1362 ret = bdev->driver->init_mem_type(bdev, type, man);
1363 if (ret)
1364 return ret;
1365 man->bdev = bdev;
1366
1367 ret = 0;
1368 if (type != TTM_PL_SYSTEM) {
1369 ret = (*man->func->init)(man, p_size);
1370 if (ret)
1371 return ret;
1372 }
1373 man->has_type = true;
1374 man->use_type = true;
1375 man->size = p_size;
1376
1377 INIT_LIST_HEAD(&man->lru);
1378
1379 return 0;
1380 }
1381 EXPORT_SYMBOL(ttm_bo_init_mm);
1382
1383 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1384 {
1385 struct ttm_bo_global *glob =
1386 container_of(kobj, struct ttm_bo_global, kobj);
1387
1388 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1389 __free_page(glob->dummy_read_page);
1390 kfree(glob);
1391 }
1392
1393 void ttm_bo_global_release(struct drm_global_reference *ref)
1394 {
1395 struct ttm_bo_global *glob = ref->object;
1396
1397 kobject_del(&glob->kobj);
1398 kobject_put(&glob->kobj);
1399 }
1400 EXPORT_SYMBOL(ttm_bo_global_release);
1401
1402 int ttm_bo_global_init(struct drm_global_reference *ref)
1403 {
1404 struct ttm_bo_global_ref *bo_ref =
1405 container_of(ref, struct ttm_bo_global_ref, ref);
1406 struct ttm_bo_global *glob = ref->object;
1407 int ret;
1408
1409 mutex_init(&glob->device_list_mutex);
1410 spin_lock_init(&glob->lru_lock);
1411 glob->mem_glob = bo_ref->mem_glob;
1412 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1413
1414 if (unlikely(glob->dummy_read_page == NULL)) {
1415 ret = -ENOMEM;
1416 goto out_no_drp;
1417 }
1418
1419 INIT_LIST_HEAD(&glob->swap_lru);
1420 INIT_LIST_HEAD(&glob->device_list);
1421
1422 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1423 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1424 if (unlikely(ret != 0)) {
1425 pr_err("Could not register buffer object swapout\n");
1426 goto out_no_shrink;
1427 }
1428
1429 atomic_set(&glob->bo_count, 0);
1430
1431 ret = kobject_init_and_add(
1432 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1433 if (unlikely(ret != 0))
1434 kobject_put(&glob->kobj);
1435 return ret;
1436 out_no_shrink:
1437 __free_page(glob->dummy_read_page);
1438 out_no_drp:
1439 kfree(glob);
1440 return ret;
1441 }
1442 EXPORT_SYMBOL(ttm_bo_global_init);
1443
1444
1445 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1446 {
1447 int ret = 0;
1448 unsigned i = TTM_NUM_MEM_TYPES;
1449 struct ttm_mem_type_manager *man;
1450 struct ttm_bo_global *glob = bdev->glob;
1451
1452 while (i--) {
1453 man = &bdev->man[i];
1454 if (man->has_type) {
1455 man->use_type = false;
1456 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1457 ret = -EBUSY;
1458 pr_err("DRM memory manager type %d is not clean\n",
1459 i);
1460 }
1461 man->has_type = false;
1462 }
1463 }
1464
1465 mutex_lock(&glob->device_list_mutex);
1466 list_del(&bdev->device_list);
1467 mutex_unlock(&glob->device_list_mutex);
1468
1469 cancel_delayed_work_sync(&bdev->wq);
1470
1471 while (ttm_bo_delayed_delete(bdev, true))
1472 ;
1473
1474 spin_lock(&glob->lru_lock);
1475 if (list_empty(&bdev->ddestroy))
1476 TTM_DEBUG("Delayed destroy list was clean\n");
1477
1478 if (list_empty(&bdev->man[0].lru))
1479 TTM_DEBUG("Swap list was clean\n");
1480 spin_unlock(&glob->lru_lock);
1481
1482 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1483
1484 return ret;
1485 }
1486 EXPORT_SYMBOL(ttm_bo_device_release);
1487
1488 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1489 struct ttm_bo_global *glob,
1490 struct ttm_bo_driver *driver,
1491 struct address_space *mapping,
1492 uint64_t file_page_offset,
1493 bool need_dma32)
1494 {
1495 int ret = -EINVAL;
1496
1497 bdev->driver = driver;
1498
1499 memset(bdev->man, 0, sizeof(bdev->man));
1500
1501 /*
1502 * Initialize the system memory buffer type.
1503 * Other types need to be driver / IOCTL initialized.
1504 */
1505 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1506 if (unlikely(ret != 0))
1507 goto out_no_sys;
1508
1509 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1510 0x10000000);
1511 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1512 INIT_LIST_HEAD(&bdev->ddestroy);
1513 bdev->dev_mapping = mapping;
1514 bdev->glob = glob;
1515 bdev->need_dma32 = need_dma32;
1516 mutex_lock(&glob->device_list_mutex);
1517 list_add_tail(&bdev->device_list, &glob->device_list);
1518 mutex_unlock(&glob->device_list_mutex);
1519
1520 return 0;
1521 out_no_sys:
1522 return ret;
1523 }
1524 EXPORT_SYMBOL(ttm_bo_device_init);
1525
1526 /*
1527 * buffer object vm functions.
1528 */
1529
1530 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1531 {
1532 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1533
1534 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1535 if (mem->mem_type == TTM_PL_SYSTEM)
1536 return false;
1537
1538 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1539 return false;
1540
1541 if (mem->placement & TTM_PL_FLAG_CACHED)
1542 return false;
1543 }
1544 return true;
1545 }
1546
1547 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1548 {
1549 struct ttm_bo_device *bdev = bo->bdev;
1550
1551 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1552 ttm_mem_io_free_vm(bo);
1553 }
1554
1555 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1556 {
1557 struct ttm_bo_device *bdev = bo->bdev;
1558 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1559
1560 ttm_mem_io_lock(man, false);
1561 ttm_bo_unmap_virtual_locked(bo);
1562 ttm_mem_io_unlock(man);
1563 }
1564
1565
1566 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1567
1568 int ttm_bo_wait(struct ttm_buffer_object *bo,
1569 bool interruptible, bool no_wait)
1570 {
1571 struct reservation_object_list *fobj;
1572 struct reservation_object *resv;
1573 struct fence *excl;
1574 long timeout = 15 * HZ;
1575 int i;
1576
1577 resv = bo->resv;
1578 fobj = reservation_object_get_list(resv);
1579 excl = reservation_object_get_excl(resv);
1580 if (excl) {
1581 if (!fence_is_signaled(excl)) {
1582 if (no_wait)
1583 return -EBUSY;
1584
1585 timeout = fence_wait_timeout(excl,
1586 interruptible, timeout);
1587 }
1588 }
1589
1590 for (i = 0; fobj && timeout > 0 && i < fobj->shared_count; ++i) {
1591 struct fence *fence;
1592 fence = rcu_dereference_protected(fobj->shared[i],
1593 reservation_object_held(resv));
1594
1595 if (!fence_is_signaled(fence)) {
1596 if (no_wait)
1597 return -EBUSY;
1598
1599 timeout = fence_wait_timeout(fence,
1600 interruptible, timeout);
1601 }
1602 }
1603
1604 if (timeout < 0)
1605 return timeout;
1606
1607 if (timeout == 0)
1608 return -EBUSY;
1609
1610 reservation_object_add_excl_fence(resv, NULL);
1611 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1612 return 0;
1613 }
1614 EXPORT_SYMBOL(ttm_bo_wait);
1615
1616 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1617 {
1618 int ret = 0;
1619
1620 /*
1621 * Using ttm_bo_reserve makes sure the lru lists are updated.
1622 */
1623
1624 ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1625 if (unlikely(ret != 0))
1626 return ret;
1627 ret = ttm_bo_wait(bo, true, no_wait);
1628 if (likely(ret == 0))
1629 atomic_inc(&bo->cpu_writers);
1630 ttm_bo_unreserve(bo);
1631 return ret;
1632 }
1633 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1634
1635 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1636 {
1637 atomic_dec(&bo->cpu_writers);
1638 }
1639 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1640
1641 /**
1642 * A buffer object shrink method that tries to swap out the first
1643 * buffer object on the bo_global::swap_lru list.
1644 */
1645
1646 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1647 {
1648 struct ttm_bo_global *glob =
1649 container_of(shrink, struct ttm_bo_global, shrink);
1650 struct ttm_buffer_object *bo;
1651 int ret = -EBUSY;
1652 int put_count;
1653 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1654
1655 spin_lock(&glob->lru_lock);
1656 list_for_each_entry(bo, &glob->swap_lru, swap) {
1657 ret = __ttm_bo_reserve(bo, false, true, NULL);
1658 if (!ret)
1659 break;
1660 }
1661
1662 if (ret) {
1663 spin_unlock(&glob->lru_lock);
1664 return ret;
1665 }
1666
1667 kref_get(&bo->list_kref);
1668
1669 if (!list_empty(&bo->ddestroy)) {
1670 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1671 kref_put(&bo->list_kref, ttm_bo_release_list);
1672 return ret;
1673 }
1674
1675 put_count = ttm_bo_del_from_lru(bo);
1676 spin_unlock(&glob->lru_lock);
1677
1678 ttm_bo_list_ref_sub(bo, put_count, true);
1679
1680 /**
1681 * Wait for GPU, then move to system cached.
1682 */
1683
1684 ret = ttm_bo_wait(bo, false, false);
1685
1686 if (unlikely(ret != 0))
1687 goto out;
1688
1689 if ((bo->mem.placement & swap_placement) != swap_placement) {
1690 struct ttm_mem_reg evict_mem;
1691
1692 evict_mem = bo->mem;
1693 evict_mem.mm_node = NULL;
1694 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1695 evict_mem.mem_type = TTM_PL_SYSTEM;
1696
1697 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1698 false, false);
1699 if (unlikely(ret != 0))
1700 goto out;
1701 }
1702
1703 ttm_bo_unmap_virtual(bo);
1704
1705 /**
1706 * Swap out. Buffer will be swapped in again as soon as
1707 * anyone tries to access a ttm page.
1708 */
1709
1710 if (bo->bdev->driver->swap_notify)
1711 bo->bdev->driver->swap_notify(bo);
1712
1713 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1714 out:
1715
1716 /**
1717 *
1718 * Unreserve without putting on LRU to avoid swapping out an
1719 * already swapped buffer.
1720 */
1721
1722 __ttm_bo_unreserve(bo);
1723 kref_put(&bo->list_kref, ttm_bo_release_list);
1724 return ret;
1725 }
1726
1727 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1728 {
1729 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1730 ;
1731 }
1732 EXPORT_SYMBOL(ttm_bo_swapout_all);
1733
1734 /**
1735 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1736 * unreserved
1737 *
1738 * @bo: Pointer to buffer
1739 */
1740 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1741 {
1742 int ret;
1743
1744 /*
1745 * In the absense of a wait_unlocked API,
1746 * Use the bo::wu_mutex to avoid triggering livelocks due to
1747 * concurrent use of this function. Note that this use of
1748 * bo::wu_mutex can go away if we change locking order to
1749 * mmap_sem -> bo::reserve.
1750 */
1751 ret = mutex_lock_interruptible(&bo->wu_mutex);
1752 if (unlikely(ret != 0))
1753 return -ERESTARTSYS;
1754 if (!ww_mutex_is_locked(&bo->resv->lock))
1755 goto out_unlock;
1756 ret = __ttm_bo_reserve(bo, true, false, NULL);
1757 if (unlikely(ret != 0))
1758 goto out_unlock;
1759 __ttm_bo_unreserve(bo);
1760
1761 out_unlock:
1762 mutex_unlock(&bo->wu_mutex);
1763 return ret;
1764 }
Ubuntu Artful kernel mirror