ttm_bo.c revision 1.1.1.3 1 /* $NetBSD: ttm_bo.c,v 1.1.1.3 2018/08/27 01:34:59 riastradh Exp $ */
2
3 /**************************************************************************
4 *
5 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
6 * All Rights Reserved.
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the
10 * "Software"), to deal in the Software without restriction, including
11 * without limitation the rights to use, copy, modify, merge, publish,
12 * distribute, sub license, and/or sell copies of the Software, and to
13 * permit persons to whom the Software is furnished to do so, subject to
14 * the following conditions:
15 *
16 * The above copyright notice and this permission notice (including the
17 * next paragraph) shall be included in all copies or substantial portions
18 * of the Software.
19 *
20 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
21 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
23 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
24 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
25 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
26 * USE OR OTHER DEALINGS IN THE SOFTWARE.
27 *
28 **************************************************************************/
29 /*
30 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
31 */
32
33 #include <sys/cdefs.h>
34 __KERNEL_RCSID(0, "$NetBSD: ttm_bo.c,v 1.1.1.3 2018/08/27 01:34:59 riastradh Exp $");
35
36 #define pr_fmt(fmt) "[TTM] " fmt
37
38 #include <drm/ttm/ttm_module.h>
39 #include <drm/ttm/ttm_bo_driver.h>
40 #include <drm/ttm/ttm_placement.h>
41 #include <linux/jiffies.h>
42 #include <linux/slab.h>
43 #include <linux/sched.h>
44 #include <linux/mm.h>
45 #include <linux/file.h>
46 #include <linux/module.h>
47 #include <linux/atomic.h>
48 #include <linux/reservation.h>
49
50 #define TTM_ASSERT_LOCKED(param)
51 #define TTM_DEBUG(fmt, arg...)
52 #define TTM_BO_HASH_ORDER 13
53
54 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
55 static void ttm_bo_global_kobj_release(struct kobject *kobj);
56
57 static struct attribute ttm_bo_count = {
58 .name = "bo_count",
59 .mode = S_IRUGO
60 };
61
62 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
63 uint32_t *mem_type)
64 {
65 int i;
66
67 for (i = 0; i <= TTM_PL_PRIV5; i++)
68 if (place->flags & (1 << i)) {
69 *mem_type = i;
70 return 0;
71 }
72 return -EINVAL;
73 }
74
75 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
76 {
77 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
78
79 pr_err(" has_type: %d\n", man->has_type);
80 pr_err(" use_type: %d\n", man->use_type);
81 pr_err(" flags: 0x%08X\n", man->flags);
82 pr_err(" gpu_offset: 0x%08llX\n", man->gpu_offset);
83 pr_err(" size: %llu\n", man->size);
84 pr_err(" available_caching: 0x%08X\n", man->available_caching);
85 pr_err(" default_caching: 0x%08X\n", man->default_caching);
86 if (mem_type != TTM_PL_SYSTEM)
87 (*man->func->debug)(man, TTM_PFX);
88 }
89
90 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
91 struct ttm_placement *placement)
92 {
93 int i, ret, mem_type;
94
95 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
96 bo, bo->mem.num_pages, bo->mem.size >> 10,
97 bo->mem.size >> 20);
98 for (i = 0; i < placement->num_placement; i++) {
99 ret = ttm_mem_type_from_place(&placement->placement[i],
100 &mem_type);
101 if (ret)
102 return;
103 pr_err(" placement[%d]=0x%08X (%d)\n",
104 i, placement->placement[i].flags, mem_type);
105 ttm_mem_type_debug(bo->bdev, mem_type);
106 }
107 }
108
109 static ssize_t ttm_bo_global_show(struct kobject *kobj,
110 struct attribute *attr,
111 char *buffer)
112 {
113 struct ttm_bo_global *glob =
114 container_of(kobj, struct ttm_bo_global, kobj);
115
116 return snprintf(buffer, PAGE_SIZE, "%lu\n",
117 (unsigned long) atomic_read(&glob->bo_count));
118 }
119
120 static struct attribute *ttm_bo_global_attrs[] = {
121 &ttm_bo_count,
122 NULL
123 };
124
125 static const struct sysfs_ops ttm_bo_global_ops = {
126 .show = &ttm_bo_global_show
127 };
128
129 static struct kobj_type ttm_bo_glob_kobj_type = {
130 .release = &ttm_bo_global_kobj_release,
131 .sysfs_ops = &ttm_bo_global_ops,
132 .default_attrs = ttm_bo_global_attrs
133 };
134
135
136 static inline uint32_t ttm_bo_type_flags(unsigned type)
137 {
138 return 1 << (type);
139 }
140
141 static void ttm_bo_release_list(struct kref *list_kref)
142 {
143 struct ttm_buffer_object *bo =
144 container_of(list_kref, struct ttm_buffer_object, list_kref);
145 struct ttm_bo_device *bdev = bo->bdev;
146 size_t acc_size = bo->acc_size;
147
148 BUG_ON(atomic_read(&bo->list_kref.refcount));
149 BUG_ON(atomic_read(&bo->kref.refcount));
150 BUG_ON(atomic_read(&bo->cpu_writers));
151 BUG_ON(bo->mem.mm_node != NULL);
152 BUG_ON(!list_empty(&bo->lru));
153 BUG_ON(!list_empty(&bo->ddestroy));
154
155 if (bo->ttm)
156 ttm_tt_destroy(bo->ttm);
157 atomic_dec(&bo->glob->bo_count);
158 if (bo->resv == &bo->ttm_resv)
159 reservation_object_fini(&bo->ttm_resv);
160 mutex_destroy(&bo->wu_mutex);
161 if (bo->destroy)
162 bo->destroy(bo);
163 else {
164 kfree(bo);
165 }
166 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
167 }
168
169 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
170 {
171 struct ttm_bo_device *bdev = bo->bdev;
172 struct ttm_mem_type_manager *man;
173
174 lockdep_assert_held(&bo->resv->lock.base);
175
176 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
177
178 BUG_ON(!list_empty(&bo->lru));
179
180 man = &bdev->man[bo->mem.mem_type];
181 list_add_tail(&bo->lru, &man->lru);
182 kref_get(&bo->list_kref);
183
184 if (bo->ttm != NULL) {
185 list_add_tail(&bo->swap, &bo->glob->swap_lru);
186 kref_get(&bo->list_kref);
187 }
188 }
189 }
190 EXPORT_SYMBOL(ttm_bo_add_to_lru);
191
192 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
193 {
194 int put_count = 0;
195
196 if (!list_empty(&bo->swap)) {
197 list_del_init(&bo->swap);
198 ++put_count;
199 }
200 if (!list_empty(&bo->lru)) {
201 list_del_init(&bo->lru);
202 ++put_count;
203 }
204
205 /*
206 * TODO: Add a driver hook to delete from
207 * driver-specific LRU's here.
208 */
209
210 return put_count;
211 }
212
213 static void ttm_bo_ref_bug(struct kref *list_kref)
214 {
215 BUG();
216 }
217
218 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
219 bool never_free)
220 {
221 kref_sub(&bo->list_kref, count,
222 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
223 }
224
225 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
226 {
227 int put_count;
228
229 spin_lock(&bo->glob->lru_lock);
230 put_count = ttm_bo_del_from_lru(bo);
231 spin_unlock(&bo->glob->lru_lock);
232 ttm_bo_list_ref_sub(bo, put_count, true);
233 }
234 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
235
236 /*
237 * Call bo->mutex locked.
238 */
239 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
240 {
241 struct ttm_bo_device *bdev = bo->bdev;
242 struct ttm_bo_global *glob = bo->glob;
243 int ret = 0;
244 uint32_t page_flags = 0;
245
246 TTM_ASSERT_LOCKED(&bo->mutex);
247 bo->ttm = NULL;
248
249 if (bdev->need_dma32)
250 page_flags |= TTM_PAGE_FLAG_DMA32;
251
252 switch (bo->type) {
253 case ttm_bo_type_device:
254 if (zero_alloc)
255 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
256 case ttm_bo_type_kernel:
257 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
258 page_flags, glob->dummy_read_page);
259 if (unlikely(bo->ttm == NULL))
260 ret = -ENOMEM;
261 break;
262 case ttm_bo_type_sg:
263 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
264 page_flags | TTM_PAGE_FLAG_SG,
265 glob->dummy_read_page);
266 if (unlikely(bo->ttm == NULL)) {
267 ret = -ENOMEM;
268 break;
269 }
270 bo->ttm->sg = bo->sg;
271 break;
272 default:
273 pr_err("Illegal buffer object type\n");
274 ret = -EINVAL;
275 break;
276 }
277
278 return ret;
279 }
280
281 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
282 struct ttm_mem_reg *mem,
283 bool evict, bool interruptible,
284 bool no_wait_gpu)
285 {
286 struct ttm_bo_device *bdev = bo->bdev;
287 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
288 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
289 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
290 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
291 int ret = 0;
292
293 if (old_is_pci || new_is_pci ||
294 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
295 ret = ttm_mem_io_lock(old_man, true);
296 if (unlikely(ret != 0))
297 goto out_err;
298 ttm_bo_unmap_virtual_locked(bo);
299 ttm_mem_io_unlock(old_man);
300 }
301
302 /*
303 * Create and bind a ttm if required.
304 */
305
306 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
307 if (bo->ttm == NULL) {
308 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
309 ret = ttm_bo_add_ttm(bo, zero);
310 if (ret)
311 goto out_err;
312 }
313
314 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
315 if (ret)
316 goto out_err;
317
318 if (mem->mem_type != TTM_PL_SYSTEM) {
319 ret = ttm_tt_bind(bo->ttm, mem);
320 if (ret)
321 goto out_err;
322 }
323
324 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
325 if (bdev->driver->move_notify)
326 bdev->driver->move_notify(bo, mem);
327 bo->mem = *mem;
328 mem->mm_node = NULL;
329 goto moved;
330 }
331 }
332
333 if (bdev->driver->move_notify)
334 bdev->driver->move_notify(bo, mem);
335
336 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
337 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
338 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
339 else if (bdev->driver->move)
340 ret = bdev->driver->move(bo, evict, interruptible,
341 no_wait_gpu, mem);
342 else
343 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
344
345 if (ret) {
346 if (bdev->driver->move_notify) {
347 struct ttm_mem_reg tmp_mem = *mem;
348 *mem = bo->mem;
349 bo->mem = tmp_mem;
350 bdev->driver->move_notify(bo, mem);
351 bo->mem = *mem;
352 *mem = tmp_mem;
353 }
354
355 goto out_err;
356 }
357
358 moved:
359 if (bo->evicted) {
360 if (bdev->driver->invalidate_caches) {
361 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
362 if (ret)
363 pr_err("Can not flush read caches\n");
364 }
365 bo->evicted = false;
366 }
367
368 if (bo->mem.mm_node) {
369 bo->offset = (bo->mem.start << PAGE_SHIFT) +
370 bdev->man[bo->mem.mem_type].gpu_offset;
371 bo->cur_placement = bo->mem.placement;
372 } else
373 bo->offset = 0;
374
375 return 0;
376
377 out_err:
378 new_man = &bdev->man[bo->mem.mem_type];
379 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
380 ttm_tt_unbind(bo->ttm);
381 ttm_tt_destroy(bo->ttm);
382 bo->ttm = NULL;
383 }
384
385 return ret;
386 }
387
388 /**
389 * Call bo::reserved.
390 * Will release GPU memory type usage on destruction.
391 * This is the place to put in driver specific hooks to release
392 * driver private resources.
393 * Will release the bo::reserved lock.
394 */
395
396 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
397 {
398 if (bo->bdev->driver->move_notify)
399 bo->bdev->driver->move_notify(bo, NULL);
400
401 if (bo->ttm) {
402 ttm_tt_unbind(bo->ttm);
403 ttm_tt_destroy(bo->ttm);
404 bo->ttm = NULL;
405 }
406 ttm_bo_mem_put(bo, &bo->mem);
407
408 ww_mutex_unlock (&bo->resv->lock);
409 }
410
411 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
412 {
413 struct reservation_object_list *fobj;
414 struct fence *fence;
415 int i;
416
417 fobj = reservation_object_get_list(bo->resv);
418 fence = reservation_object_get_excl(bo->resv);
419 if (fence && !fence->ops->signaled)
420 fence_enable_sw_signaling(fence);
421
422 for (i = 0; fobj && i < fobj->shared_count; ++i) {
423 fence = rcu_dereference_protected(fobj->shared[i],
424 reservation_object_held(bo->resv));
425
426 if (!fence->ops->signaled)
427 fence_enable_sw_signaling(fence);
428 }
429 }
430
431 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
432 {
433 struct ttm_bo_device *bdev = bo->bdev;
434 struct ttm_bo_global *glob = bo->glob;
435 int put_count;
436 int ret;
437
438 spin_lock(&glob->lru_lock);
439 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
440
441 if (!ret) {
442 if (!ttm_bo_wait(bo, false, false, true)) {
443 put_count = ttm_bo_del_from_lru(bo);
444
445 spin_unlock(&glob->lru_lock);
446 ttm_bo_cleanup_memtype_use(bo);
447
448 ttm_bo_list_ref_sub(bo, put_count, true);
449
450 return;
451 } else
452 ttm_bo_flush_all_fences(bo);
453
454 /*
455 * Make NO_EVICT bos immediately available to
456 * shrinkers, now that they are queued for
457 * destruction.
458 */
459 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
460 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
461 ttm_bo_add_to_lru(bo);
462 }
463
464 __ttm_bo_unreserve(bo);
465 }
466
467 kref_get(&bo->list_kref);
468 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
469 spin_unlock(&glob->lru_lock);
470
471 schedule_delayed_work(&bdev->wq,
472 ((HZ / 100) < 1) ? 1 : HZ / 100);
473 }
474
475 /**
476 * function ttm_bo_cleanup_refs_and_unlock
477 * If bo idle, remove from delayed- and lru lists, and unref.
478 * If not idle, do nothing.
479 *
480 * Must be called with lru_lock and reservation held, this function
481 * will drop both before returning.
482 *
483 * @interruptible Any sleeps should occur interruptibly.
484 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
485 */
486
487 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
488 bool interruptible,
489 bool no_wait_gpu)
490 {
491 struct ttm_bo_global *glob = bo->glob;
492 int put_count;
493 int ret;
494
495 ret = ttm_bo_wait(bo, false, false, true);
496
497 if (ret && !no_wait_gpu) {
498 long lret;
499 ww_mutex_unlock(&bo->resv->lock);
500 spin_unlock(&glob->lru_lock);
501
502 lret = reservation_object_wait_timeout_rcu(bo->resv,
503 true,
504 interruptible,
505 30 * HZ);
506
507 if (lret < 0)
508 return lret;
509 else if (lret == 0)
510 return -EBUSY;
511
512 spin_lock(&glob->lru_lock);
513 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
514
515 /*
516 * We raced, and lost, someone else holds the reservation now,
517 * and is probably busy in ttm_bo_cleanup_memtype_use.
518 *
519 * Even if it's not the case, because we finished waiting any
520 * delayed destruction would succeed, so just return success
521 * here.
522 */
523 if (ret) {
524 spin_unlock(&glob->lru_lock);
525 return 0;
526 }
527
528 /*
529 * remove sync_obj with ttm_bo_wait, the wait should be
530 * finished, and no new wait object should have been added.
531 */
532 ret = ttm_bo_wait(bo, false, false, true);
533 WARN_ON(ret);
534 }
535
536 if (ret || unlikely(list_empty(&bo->ddestroy))) {
537 __ttm_bo_unreserve(bo);
538 spin_unlock(&glob->lru_lock);
539 return ret;
540 }
541
542 put_count = ttm_bo_del_from_lru(bo);
543 list_del_init(&bo->ddestroy);
544 ++put_count;
545
546 spin_unlock(&glob->lru_lock);
547 ttm_bo_cleanup_memtype_use(bo);
548
549 ttm_bo_list_ref_sub(bo, put_count, true);
550
551 return 0;
552 }
553
554 /**
555 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
556 * encountered buffers.
557 */
558
559 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
560 {
561 struct ttm_bo_global *glob = bdev->glob;
562 struct ttm_buffer_object *entry = NULL;
563 int ret = 0;
564
565 spin_lock(&glob->lru_lock);
566 if (list_empty(&bdev->ddestroy))
567 goto out_unlock;
568
569 entry = list_first_entry(&bdev->ddestroy,
570 struct ttm_buffer_object, ddestroy);
571 kref_get(&entry->list_kref);
572
573 for (;;) {
574 struct ttm_buffer_object *nentry = NULL;
575
576 if (entry->ddestroy.next != &bdev->ddestroy) {
577 nentry = list_first_entry(&entry->ddestroy,
578 struct ttm_buffer_object, ddestroy);
579 kref_get(&nentry->list_kref);
580 }
581
582 ret = __ttm_bo_reserve(entry, false, true, false, NULL);
583 if (remove_all && ret) {
584 spin_unlock(&glob->lru_lock);
585 ret = __ttm_bo_reserve(entry, false, false,
586 false, NULL);
587 spin_lock(&glob->lru_lock);
588 }
589
590 if (!ret)
591 ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
592 !remove_all);
593 else
594 spin_unlock(&glob->lru_lock);
595
596 kref_put(&entry->list_kref, ttm_bo_release_list);
597 entry = nentry;
598
599 if (ret || !entry)
600 goto out;
601
602 spin_lock(&glob->lru_lock);
603 if (list_empty(&entry->ddestroy))
604 break;
605 }
606
607 out_unlock:
608 spin_unlock(&glob->lru_lock);
609 out:
610 if (entry)
611 kref_put(&entry->list_kref, ttm_bo_release_list);
612 return ret;
613 }
614
615 static void ttm_bo_delayed_workqueue(struct work_struct *work)
616 {
617 struct ttm_bo_device *bdev =
618 container_of(work, struct ttm_bo_device, wq.work);
619
620 if (ttm_bo_delayed_delete(bdev, false)) {
621 schedule_delayed_work(&bdev->wq,
622 ((HZ / 100) < 1) ? 1 : HZ / 100);
623 }
624 }
625
626 static void ttm_bo_release(struct kref *kref)
627 {
628 struct ttm_buffer_object *bo =
629 container_of(kref, struct ttm_buffer_object, kref);
630 struct ttm_bo_device *bdev = bo->bdev;
631 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
632
633 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
634 ttm_mem_io_lock(man, false);
635 ttm_mem_io_free_vm(bo);
636 ttm_mem_io_unlock(man);
637 ttm_bo_cleanup_refs_or_queue(bo);
638 kref_put(&bo->list_kref, ttm_bo_release_list);
639 }
640
641 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
642 {
643 struct ttm_buffer_object *bo = *p_bo;
644
645 *p_bo = NULL;
646 kref_put(&bo->kref, ttm_bo_release);
647 }
648 EXPORT_SYMBOL(ttm_bo_unref);
649
650 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
651 {
652 return cancel_delayed_work_sync(&bdev->wq);
653 }
654 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
655
656 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
657 {
658 if (resched)
659 schedule_delayed_work(&bdev->wq,
660 ((HZ / 100) < 1) ? 1 : HZ / 100);
661 }
662 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
663
664 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
665 bool no_wait_gpu)
666 {
667 struct ttm_bo_device *bdev = bo->bdev;
668 struct ttm_mem_reg evict_mem;
669 struct ttm_placement placement;
670 int ret = 0;
671
672 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
673
674 if (unlikely(ret != 0)) {
675 if (ret != -ERESTARTSYS) {
676 pr_err("Failed to expire sync object before buffer eviction\n");
677 }
678 goto out;
679 }
680
681 lockdep_assert_held(&bo->resv->lock.base);
682
683 evict_mem = bo->mem;
684 evict_mem.mm_node = NULL;
685 evict_mem.bus.io_reserved_vm = false;
686 evict_mem.bus.io_reserved_count = 0;
687
688 placement.num_placement = 0;
689 placement.num_busy_placement = 0;
690 bdev->driver->evict_flags(bo, &placement);
691 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
692 no_wait_gpu);
693 if (ret) {
694 if (ret != -ERESTARTSYS) {
695 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
696 bo);
697 ttm_bo_mem_space_debug(bo, &placement);
698 }
699 goto out;
700 }
701
702 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
703 no_wait_gpu);
704 if (ret) {
705 if (ret != -ERESTARTSYS)
706 pr_err("Buffer eviction failed\n");
707 ttm_bo_mem_put(bo, &evict_mem);
708 goto out;
709 }
710 bo->evicted = true;
711 out:
712 return ret;
713 }
714
715 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
716 uint32_t mem_type,
717 const struct ttm_place *place,
718 bool interruptible,
719 bool no_wait_gpu)
720 {
721 struct ttm_bo_global *glob = bdev->glob;
722 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
723 struct ttm_buffer_object *bo;
724 int ret = -EBUSY, put_count;
725
726 spin_lock(&glob->lru_lock);
727 list_for_each_entry(bo, &man->lru, lru) {
728 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
729 if (!ret) {
730 if (place && (place->fpfn || place->lpfn)) {
731 /* Don't evict this BO if it's outside of the
732 * requested placement range
733 */
734 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
735 (place->lpfn && place->lpfn <= bo->mem.start)) {
736 __ttm_bo_unreserve(bo);
737 ret = -EBUSY;
738 continue;
739 }
740 }
741
742 break;
743 }
744 }
745
746 if (ret) {
747 spin_unlock(&glob->lru_lock);
748 return ret;
749 }
750
751 kref_get(&bo->list_kref);
752
753 if (!list_empty(&bo->ddestroy)) {
754 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
755 no_wait_gpu);
756 kref_put(&bo->list_kref, ttm_bo_release_list);
757 return ret;
758 }
759
760 put_count = ttm_bo_del_from_lru(bo);
761 spin_unlock(&glob->lru_lock);
762
763 BUG_ON(ret != 0);
764
765 ttm_bo_list_ref_sub(bo, put_count, true);
766
767 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
768 ttm_bo_unreserve(bo);
769
770 kref_put(&bo->list_kref, ttm_bo_release_list);
771 return ret;
772 }
773
774 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
775 {
776 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
777
778 if (mem->mm_node)
779 (*man->func->put_node)(man, mem);
780 }
781 EXPORT_SYMBOL(ttm_bo_mem_put);
782
783 /**
784 * Repeatedly evict memory from the LRU for @mem_type until we create enough
785 * space, or we've evicted everything and there isn't enough space.
786 */
787 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
788 uint32_t mem_type,
789 const struct ttm_place *place,
790 struct ttm_mem_reg *mem,
791 bool interruptible,
792 bool no_wait_gpu)
793 {
794 struct ttm_bo_device *bdev = bo->bdev;
795 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
796 int ret;
797
798 do {
799 ret = (*man->func->get_node)(man, bo, place, mem);
800 if (unlikely(ret != 0))
801 return ret;
802 if (mem->mm_node)
803 break;
804 ret = ttm_mem_evict_first(bdev, mem_type, place,
805 interruptible, no_wait_gpu);
806 if (unlikely(ret != 0))
807 return ret;
808 } while (1);
809 if (mem->mm_node == NULL)
810 return -ENOMEM;
811 mem->mem_type = mem_type;
812 return 0;
813 }
814
815 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
816 uint32_t cur_placement,
817 uint32_t proposed_placement)
818 {
819 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
820 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
821
822 /**
823 * Keep current caching if possible.
824 */
825
826 if ((cur_placement & caching) != 0)
827 result |= (cur_placement & caching);
828 else if ((man->default_caching & caching) != 0)
829 result |= man->default_caching;
830 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
831 result |= TTM_PL_FLAG_CACHED;
832 else if ((TTM_PL_FLAG_WC & caching) != 0)
833 result |= TTM_PL_FLAG_WC;
834 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
835 result |= TTM_PL_FLAG_UNCACHED;
836
837 return result;
838 }
839
840 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
841 uint32_t mem_type,
842 const struct ttm_place *place,
843 uint32_t *masked_placement)
844 {
845 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
846
847 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
848 return false;
849
850 if ((place->flags & man->available_caching) == 0)
851 return false;
852
853 cur_flags |= (place->flags & man->available_caching);
854
855 *masked_placement = cur_flags;
856 return true;
857 }
858
859 /**
860 * Creates space for memory region @mem according to its type.
861 *
862 * This function first searches for free space in compatible memory types in
863 * the priority order defined by the driver. If free space isn't found, then
864 * ttm_bo_mem_force_space is attempted in priority order to evict and find
865 * space.
866 */
867 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
868 struct ttm_placement *placement,
869 struct ttm_mem_reg *mem,
870 bool interruptible,
871 bool no_wait_gpu)
872 {
873 struct ttm_bo_device *bdev = bo->bdev;
874 struct ttm_mem_type_manager *man;
875 uint32_t mem_type = TTM_PL_SYSTEM;
876 uint32_t cur_flags = 0;
877 bool type_found = false;
878 bool type_ok = false;
879 bool has_erestartsys = false;
880 int i, ret;
881
882 mem->mm_node = NULL;
883 for (i = 0; i < placement->num_placement; ++i) {
884 const struct ttm_place *place = &placement->placement[i];
885
886 ret = ttm_mem_type_from_place(place, &mem_type);
887 if (ret)
888 return ret;
889 man = &bdev->man[mem_type];
890 if (!man->has_type || !man->use_type)
891 continue;
892
893 type_ok = ttm_bo_mt_compatible(man, mem_type, place,
894 &cur_flags);
895
896 if (!type_ok)
897 continue;
898
899 type_found = true;
900 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
901 cur_flags);
902 /*
903 * Use the access and other non-mapping-related flag bits from
904 * the memory placement flags to the current flags
905 */
906 ttm_flag_masked(&cur_flags, place->flags,
907 ~TTM_PL_MASK_MEMTYPE);
908
909 if (mem_type == TTM_PL_SYSTEM)
910 break;
911
912 ret = (*man->func->get_node)(man, bo, place, mem);
913 if (unlikely(ret))
914 return ret;
915
916 if (mem->mm_node)
917 break;
918 }
919
920 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
921 mem->mem_type = mem_type;
922 mem->placement = cur_flags;
923 return 0;
924 }
925
926 for (i = 0; i < placement->num_busy_placement; ++i) {
927 const struct ttm_place *place = &placement->busy_placement[i];
928
929 ret = ttm_mem_type_from_place(place, &mem_type);
930 if (ret)
931 return ret;
932 man = &bdev->man[mem_type];
933 if (!man->has_type || !man->use_type)
934 continue;
935 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
936 continue;
937
938 type_found = true;
939 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
940 cur_flags);
941 /*
942 * Use the access and other non-mapping-related flag bits from
943 * the memory placement flags to the current flags
944 */
945 ttm_flag_masked(&cur_flags, place->flags,
946 ~TTM_PL_MASK_MEMTYPE);
947
948 if (mem_type == TTM_PL_SYSTEM) {
949 mem->mem_type = mem_type;
950 mem->placement = cur_flags;
951 mem->mm_node = NULL;
952 return 0;
953 }
954
955 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
956 interruptible, no_wait_gpu);
957 if (ret == 0 && mem->mm_node) {
958 mem->placement = cur_flags;
959 return 0;
960 }
961 if (ret == -ERESTARTSYS)
962 has_erestartsys = true;
963 }
964
965 if (!type_found) {
966 printk(KERN_ERR TTM_PFX "No compatible memory type found.\n");
967 return -EINVAL;
968 }
969
970 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
971 }
972 EXPORT_SYMBOL(ttm_bo_mem_space);
973
974 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
975 struct ttm_placement *placement,
976 bool interruptible,
977 bool no_wait_gpu)
978 {
979 int ret = 0;
980 struct ttm_mem_reg mem;
981
982 lockdep_assert_held(&bo->resv->lock.base);
983
984 /*
985 * FIXME: It's possible to pipeline buffer moves.
986 * Have the driver move function wait for idle when necessary,
987 * instead of doing it here.
988 */
989 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
990 if (ret)
991 return ret;
992 mem.num_pages = bo->num_pages;
993 mem.size = mem.num_pages << PAGE_SHIFT;
994 mem.page_alignment = bo->mem.page_alignment;
995 mem.bus.io_reserved_vm = false;
996 mem.bus.io_reserved_count = 0;
997 /*
998 * Determine where to move the buffer.
999 */
1000 ret = ttm_bo_mem_space(bo, placement, &mem,
1001 interruptible, no_wait_gpu);
1002 if (ret)
1003 goto out_unlock;
1004 ret = ttm_bo_handle_move_mem(bo, &mem, false,
1005 interruptible, no_wait_gpu);
1006 out_unlock:
1007 if (ret && mem.mm_node)
1008 ttm_bo_mem_put(bo, &mem);
1009 return ret;
1010 }
1011
1012 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1013 struct ttm_mem_reg *mem,
1014 uint32_t *new_flags)
1015 {
1016 int i;
1017
1018 for (i = 0; i < placement->num_placement; i++) {
1019 const struct ttm_place *heap = &placement->placement[i];
1020 if (mem->mm_node &&
1021 (mem->start < heap->fpfn ||
1022 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1023 continue;
1024
1025 *new_flags = heap->flags;
1026 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1027 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1028 return true;
1029 }
1030
1031 for (i = 0; i < placement->num_busy_placement; i++) {
1032 const struct ttm_place *heap = &placement->busy_placement[i];
1033 if (mem->mm_node &&
1034 (mem->start < heap->fpfn ||
1035 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1036 continue;
1037
1038 *new_flags = heap->flags;
1039 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1040 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1041 return true;
1042 }
1043
1044 return false;
1045 }
1046 EXPORT_SYMBOL(ttm_bo_mem_compat);
1047
1048 int ttm_bo_validate(struct ttm_buffer_object *bo,
1049 struct ttm_placement *placement,
1050 bool interruptible,
1051 bool no_wait_gpu)
1052 {
1053 int ret;
1054 uint32_t new_flags;
1055
1056 lockdep_assert_held(&bo->resv->lock.base);
1057 /*
1058 * Check whether we need to move buffer.
1059 */
1060 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1061 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1062 no_wait_gpu);
1063 if (ret)
1064 return ret;
1065 } else {
1066 /*
1067 * Use the access and other non-mapping-related flag bits from
1068 * the compatible memory placement flags to the active flags
1069 */
1070 ttm_flag_masked(&bo->mem.placement, new_flags,
1071 ~TTM_PL_MASK_MEMTYPE);
1072 }
1073 /*
1074 * We might need to add a TTM.
1075 */
1076 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1077 ret = ttm_bo_add_ttm(bo, true);
1078 if (ret)
1079 return ret;
1080 }
1081 return 0;
1082 }
1083 EXPORT_SYMBOL(ttm_bo_validate);
1084
1085 int ttm_bo_init(struct ttm_bo_device *bdev,
1086 struct ttm_buffer_object *bo,
1087 unsigned long size,
1088 enum ttm_bo_type type,
1089 struct ttm_placement *placement,
1090 uint32_t page_alignment,
1091 bool interruptible,
1092 struct file *persistent_swap_storage,
1093 size_t acc_size,
1094 struct sg_table *sg,
1095 struct reservation_object *resv,
1096 void (*destroy) (struct ttm_buffer_object *))
1097 {
1098 int ret = 0;
1099 unsigned long num_pages;
1100 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1101 bool locked;
1102
1103 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1104 if (ret) {
1105 pr_err("Out of kernel memory\n");
1106 if (destroy)
1107 (*destroy)(bo);
1108 else
1109 kfree(bo);
1110 return -ENOMEM;
1111 }
1112
1113 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1114 if (num_pages == 0) {
1115 pr_err("Illegal buffer object size\n");
1116 if (destroy)
1117 (*destroy)(bo);
1118 else
1119 kfree(bo);
1120 ttm_mem_global_free(mem_glob, acc_size);
1121 return -EINVAL;
1122 }
1123 bo->destroy = destroy;
1124
1125 kref_init(&bo->kref);
1126 kref_init(&bo->list_kref);
1127 atomic_set(&bo->cpu_writers, 0);
1128 INIT_LIST_HEAD(&bo->lru);
1129 INIT_LIST_HEAD(&bo->ddestroy);
1130 INIT_LIST_HEAD(&bo->swap);
1131 INIT_LIST_HEAD(&bo->io_reserve_lru);
1132 mutex_init(&bo->wu_mutex);
1133 bo->bdev = bdev;
1134 bo->glob = bdev->glob;
1135 bo->type = type;
1136 bo->num_pages = num_pages;
1137 bo->mem.size = num_pages << PAGE_SHIFT;
1138 bo->mem.mem_type = TTM_PL_SYSTEM;
1139 bo->mem.num_pages = bo->num_pages;
1140 bo->mem.mm_node = NULL;
1141 bo->mem.page_alignment = page_alignment;
1142 bo->mem.bus.io_reserved_vm = false;
1143 bo->mem.bus.io_reserved_count = 0;
1144 bo->priv_flags = 0;
1145 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1146 bo->persistent_swap_storage = persistent_swap_storage;
1147 bo->acc_size = acc_size;
1148 bo->sg = sg;
1149 if (resv) {
1150 bo->resv = resv;
1151 lockdep_assert_held(&bo->resv->lock.base);
1152 } else {
1153 bo->resv = &bo->ttm_resv;
1154 reservation_object_init(&bo->ttm_resv);
1155 }
1156 atomic_inc(&bo->glob->bo_count);
1157 drm_vma_node_reset(&bo->vma_node);
1158
1159 /*
1160 * For ttm_bo_type_device buffers, allocate
1161 * address space from the device.
1162 */
1163 if (bo->type == ttm_bo_type_device ||
1164 bo->type == ttm_bo_type_sg)
1165 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1166 bo->mem.num_pages);
1167
1168 /* passed reservation objects should already be locked,
1169 * since otherwise lockdep will be angered in radeon.
1170 */
1171 if (!resv) {
1172 locked = ww_mutex_trylock(&bo->resv->lock);
1173 WARN_ON(!locked);
1174 }
1175
1176 if (likely(!ret))
1177 ret = ttm_bo_validate(bo, placement, interruptible, false);
1178
1179 if (!resv)
1180 ttm_bo_unreserve(bo);
1181
1182 if (unlikely(ret))
1183 ttm_bo_unref(&bo);
1184
1185 return ret;
1186 }
1187 EXPORT_SYMBOL(ttm_bo_init);
1188
1189 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1190 unsigned long bo_size,
1191 unsigned struct_size)
1192 {
1193 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1194 size_t size = 0;
1195
1196 size += ttm_round_pot(struct_size);
1197 size += PAGE_ALIGN(npages * sizeof(void *));
1198 size += ttm_round_pot(sizeof(struct ttm_tt));
1199 return size;
1200 }
1201 EXPORT_SYMBOL(ttm_bo_acc_size);
1202
1203 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1204 unsigned long bo_size,
1205 unsigned struct_size)
1206 {
1207 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1208 size_t size = 0;
1209
1210 size += ttm_round_pot(struct_size);
1211 size += PAGE_ALIGN(npages * sizeof(void *));
1212 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1213 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1214 return size;
1215 }
1216 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1217
1218 int ttm_bo_create(struct ttm_bo_device *bdev,
1219 unsigned long size,
1220 enum ttm_bo_type type,
1221 struct ttm_placement *placement,
1222 uint32_t page_alignment,
1223 bool interruptible,
1224 struct file *persistent_swap_storage,
1225 struct ttm_buffer_object **p_bo)
1226 {
1227 struct ttm_buffer_object *bo;
1228 size_t acc_size;
1229 int ret;
1230
1231 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1232 if (unlikely(bo == NULL))
1233 return -ENOMEM;
1234
1235 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1236 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1237 interruptible, persistent_swap_storage, acc_size,
1238 NULL, NULL, NULL);
1239 if (likely(ret == 0))
1240 *p_bo = bo;
1241
1242 return ret;
1243 }
1244 EXPORT_SYMBOL(ttm_bo_create);
1245
1246 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1247 unsigned mem_type, bool allow_errors)
1248 {
1249 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1250 struct ttm_bo_global *glob = bdev->glob;
1251 int ret;
1252
1253 /*
1254 * Can't use standard list traversal since we're unlocking.
1255 */
1256
1257 spin_lock(&glob->lru_lock);
1258 while (!list_empty(&man->lru)) {
1259 spin_unlock(&glob->lru_lock);
1260 ret = ttm_mem_evict_first(bdev, mem_type, NULL, false, false);
1261 if (ret) {
1262 if (allow_errors) {
1263 return ret;
1264 } else {
1265 pr_err("Cleanup eviction failed\n");
1266 }
1267 }
1268 spin_lock(&glob->lru_lock);
1269 }
1270 spin_unlock(&glob->lru_lock);
1271 return 0;
1272 }
1273
1274 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1275 {
1276 struct ttm_mem_type_manager *man;
1277 int ret = -EINVAL;
1278
1279 if (mem_type >= TTM_NUM_MEM_TYPES) {
1280 pr_err("Illegal memory type %d\n", mem_type);
1281 return ret;
1282 }
1283 man = &bdev->man[mem_type];
1284
1285 if (!man->has_type) {
1286 pr_err("Trying to take down uninitialized memory manager type %u\n",
1287 mem_type);
1288 return ret;
1289 }
1290
1291 man->use_type = false;
1292 man->has_type = false;
1293
1294 ret = 0;
1295 if (mem_type > 0) {
1296 ttm_bo_force_list_clean(bdev, mem_type, false);
1297
1298 ret = (*man->func->takedown)(man);
1299 }
1300
1301 return ret;
1302 }
1303 EXPORT_SYMBOL(ttm_bo_clean_mm);
1304
1305 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1306 {
1307 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1308
1309 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1310 pr_err("Illegal memory manager memory type %u\n", mem_type);
1311 return -EINVAL;
1312 }
1313
1314 if (!man->has_type) {
1315 pr_err("Memory type %u has not been initialized\n", mem_type);
1316 return 0;
1317 }
1318
1319 return ttm_bo_force_list_clean(bdev, mem_type, true);
1320 }
1321 EXPORT_SYMBOL(ttm_bo_evict_mm);
1322
1323 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1324 unsigned long p_size)
1325 {
1326 int ret = -EINVAL;
1327 struct ttm_mem_type_manager *man;
1328
1329 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1330 man = &bdev->man[type];
1331 BUG_ON(man->has_type);
1332 man->io_reserve_fastpath = true;
1333 man->use_io_reserve_lru = false;
1334 mutex_init(&man->io_reserve_mutex);
1335 INIT_LIST_HEAD(&man->io_reserve_lru);
1336
1337 ret = bdev->driver->init_mem_type(bdev, type, man);
1338 if (ret)
1339 return ret;
1340 man->bdev = bdev;
1341
1342 ret = 0;
1343 if (type != TTM_PL_SYSTEM) {
1344 ret = (*man->func->init)(man, p_size);
1345 if (ret)
1346 return ret;
1347 }
1348 man->has_type = true;
1349 man->use_type = true;
1350 man->size = p_size;
1351
1352 INIT_LIST_HEAD(&man->lru);
1353
1354 return 0;
1355 }
1356 EXPORT_SYMBOL(ttm_bo_init_mm);
1357
1358 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1359 {
1360 struct ttm_bo_global *glob =
1361 container_of(kobj, struct ttm_bo_global, kobj);
1362
1363 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1364 __free_page(glob->dummy_read_page);
1365 kfree(glob);
1366 }
1367
1368 void ttm_bo_global_release(struct drm_global_reference *ref)
1369 {
1370 struct ttm_bo_global *glob = ref->object;
1371
1372 kobject_del(&glob->kobj);
1373 kobject_put(&glob->kobj);
1374 }
1375 EXPORT_SYMBOL(ttm_bo_global_release);
1376
1377 int ttm_bo_global_init(struct drm_global_reference *ref)
1378 {
1379 struct ttm_bo_global_ref *bo_ref =
1380 container_of(ref, struct ttm_bo_global_ref, ref);
1381 struct ttm_bo_global *glob = ref->object;
1382 int ret;
1383
1384 mutex_init(&glob->device_list_mutex);
1385 spin_lock_init(&glob->lru_lock);
1386 glob->mem_glob = bo_ref->mem_glob;
1387 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1388
1389 if (unlikely(glob->dummy_read_page == NULL)) {
1390 ret = -ENOMEM;
1391 goto out_no_drp;
1392 }
1393
1394 INIT_LIST_HEAD(&glob->swap_lru);
1395 INIT_LIST_HEAD(&glob->device_list);
1396
1397 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1398 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1399 if (unlikely(ret != 0)) {
1400 pr_err("Could not register buffer object swapout\n");
1401 goto out_no_shrink;
1402 }
1403
1404 atomic_set(&glob->bo_count, 0);
1405
1406 ret = kobject_init_and_add(
1407 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1408 if (unlikely(ret != 0))
1409 kobject_put(&glob->kobj);
1410 return ret;
1411 out_no_shrink:
1412 __free_page(glob->dummy_read_page);
1413 out_no_drp:
1414 kfree(glob);
1415 return ret;
1416 }
1417 EXPORT_SYMBOL(ttm_bo_global_init);
1418
1419
1420 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1421 {
1422 int ret = 0;
1423 unsigned i = TTM_NUM_MEM_TYPES;
1424 struct ttm_mem_type_manager *man;
1425 struct ttm_bo_global *glob = bdev->glob;
1426
1427 while (i--) {
1428 man = &bdev->man[i];
1429 if (man->has_type) {
1430 man->use_type = false;
1431 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1432 ret = -EBUSY;
1433 pr_err("DRM memory manager type %d is not clean\n",
1434 i);
1435 }
1436 man->has_type = false;
1437 }
1438 }
1439
1440 mutex_lock(&glob->device_list_mutex);
1441 list_del(&bdev->device_list);
1442 mutex_unlock(&glob->device_list_mutex);
1443
1444 cancel_delayed_work_sync(&bdev->wq);
1445
1446 while (ttm_bo_delayed_delete(bdev, true))
1447 ;
1448
1449 spin_lock(&glob->lru_lock);
1450 if (list_empty(&bdev->ddestroy))
1451 TTM_DEBUG("Delayed destroy list was clean\n");
1452
1453 if (list_empty(&bdev->man[0].lru))
1454 TTM_DEBUG("Swap list was clean\n");
1455 spin_unlock(&glob->lru_lock);
1456
1457 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1458
1459 return ret;
1460 }
1461 EXPORT_SYMBOL(ttm_bo_device_release);
1462
1463 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1464 struct ttm_bo_global *glob,
1465 struct ttm_bo_driver *driver,
1466 struct address_space *mapping,
1467 uint64_t file_page_offset,
1468 bool need_dma32)
1469 {
1470 int ret = -EINVAL;
1471
1472 bdev->driver = driver;
1473
1474 memset(bdev->man, 0, sizeof(bdev->man));
1475
1476 /*
1477 * Initialize the system memory buffer type.
1478 * Other types need to be driver / IOCTL initialized.
1479 */
1480 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1481 if (unlikely(ret != 0))
1482 goto out_no_sys;
1483
1484 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1485 0x10000000);
1486 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1487 INIT_LIST_HEAD(&bdev->ddestroy);
1488 bdev->dev_mapping = mapping;
1489 bdev->glob = glob;
1490 bdev->need_dma32 = need_dma32;
1491 bdev->val_seq = 0;
1492 mutex_lock(&glob->device_list_mutex);
1493 list_add_tail(&bdev->device_list, &glob->device_list);
1494 mutex_unlock(&glob->device_list_mutex);
1495
1496 return 0;
1497 out_no_sys:
1498 return ret;
1499 }
1500 EXPORT_SYMBOL(ttm_bo_device_init);
1501
1502 /*
1503 * buffer object vm functions.
1504 */
1505
1506 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1507 {
1508 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1509
1510 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1511 if (mem->mem_type == TTM_PL_SYSTEM)
1512 return false;
1513
1514 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1515 return false;
1516
1517 if (mem->placement & TTM_PL_FLAG_CACHED)
1518 return false;
1519 }
1520 return true;
1521 }
1522
1523 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1524 {
1525 struct ttm_bo_device *bdev = bo->bdev;
1526
1527 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1528 ttm_mem_io_free_vm(bo);
1529 }
1530
1531 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1532 {
1533 struct ttm_bo_device *bdev = bo->bdev;
1534 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1535
1536 ttm_mem_io_lock(man, false);
1537 ttm_bo_unmap_virtual_locked(bo);
1538 ttm_mem_io_unlock(man);
1539 }
1540
1541
1542 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1543
1544 int ttm_bo_wait(struct ttm_buffer_object *bo,
1545 bool lazy, bool interruptible, bool no_wait)
1546 {
1547 struct reservation_object_list *fobj;
1548 struct reservation_object *resv;
1549 struct fence *excl;
1550 long timeout = 15 * HZ;
1551 int i;
1552
1553 resv = bo->resv;
1554 fobj = reservation_object_get_list(resv);
1555 excl = reservation_object_get_excl(resv);
1556 if (excl) {
1557 if (!fence_is_signaled(excl)) {
1558 if (no_wait)
1559 return -EBUSY;
1560
1561 timeout = fence_wait_timeout(excl,
1562 interruptible, timeout);
1563 }
1564 }
1565
1566 for (i = 0; fobj && timeout > 0 && i < fobj->shared_count; ++i) {
1567 struct fence *fence;
1568 fence = rcu_dereference_protected(fobj->shared[i],
1569 reservation_object_held(resv));
1570
1571 if (!fence_is_signaled(fence)) {
1572 if (no_wait)
1573 return -EBUSY;
1574
1575 timeout = fence_wait_timeout(fence,
1576 interruptible, timeout);
1577 }
1578 }
1579
1580 if (timeout < 0)
1581 return timeout;
1582
1583 if (timeout == 0)
1584 return -EBUSY;
1585
1586 reservation_object_add_excl_fence(resv, NULL);
1587 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1588 return 0;
1589 }
1590 EXPORT_SYMBOL(ttm_bo_wait);
1591
1592 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1593 {
1594 int ret = 0;
1595
1596 /*
1597 * Using ttm_bo_reserve makes sure the lru lists are updated.
1598 */
1599
1600 ret = ttm_bo_reserve(bo, true, no_wait, false, NULL);
1601 if (unlikely(ret != 0))
1602 return ret;
1603 ret = ttm_bo_wait(bo, false, true, no_wait);
1604 if (likely(ret == 0))
1605 atomic_inc(&bo->cpu_writers);
1606 ttm_bo_unreserve(bo);
1607 return ret;
1608 }
1609 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1610
1611 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1612 {
1613 atomic_dec(&bo->cpu_writers);
1614 }
1615 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1616
1617 /**
1618 * A buffer object shrink method that tries to swap out the first
1619 * buffer object on the bo_global::swap_lru list.
1620 */
1621
1622 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1623 {
1624 struct ttm_bo_global *glob =
1625 container_of(shrink, struct ttm_bo_global, shrink);
1626 struct ttm_buffer_object *bo;
1627 int ret = -EBUSY;
1628 int put_count;
1629
1630 spin_lock(&glob->lru_lock);
1631 list_for_each_entry(bo, &glob->swap_lru, swap) {
1632 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
1633 if (!ret)
1634 break;
1635 }
1636
1637 if (ret) {
1638 spin_unlock(&glob->lru_lock);
1639 return ret;
1640 }
1641
1642 kref_get(&bo->list_kref);
1643
1644 if (!list_empty(&bo->ddestroy)) {
1645 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1646 kref_put(&bo->list_kref, ttm_bo_release_list);
1647 return ret;
1648 }
1649
1650 put_count = ttm_bo_del_from_lru(bo);
1651 spin_unlock(&glob->lru_lock);
1652
1653 ttm_bo_list_ref_sub(bo, put_count, true);
1654
1655 /**
1656 * Wait for GPU, then move to system cached.
1657 */
1658
1659 ret = ttm_bo_wait(bo, false, false, false);
1660
1661 if (unlikely(ret != 0))
1662 goto out;
1663
1664 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1665 bo->ttm->caching_state != tt_cached) {
1666 struct ttm_mem_reg evict_mem;
1667
1668 evict_mem = bo->mem;
1669 evict_mem.mm_node = NULL;
1670 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1671 evict_mem.mem_type = TTM_PL_SYSTEM;
1672
1673 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1674 false, false);
1675 if (unlikely(ret != 0))
1676 goto out;
1677 }
1678
1679 ttm_bo_unmap_virtual(bo);
1680
1681 /**
1682 * Swap out. Buffer will be swapped in again as soon as
1683 * anyone tries to access a ttm page.
1684 */
1685
1686 if (bo->bdev->driver->swap_notify)
1687 bo->bdev->driver->swap_notify(bo);
1688
1689 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1690 out:
1691
1692 /**
1693 *
1694 * Unreserve without putting on LRU to avoid swapping out an
1695 * already swapped buffer.
1696 */
1697
1698 __ttm_bo_unreserve(bo);
1699 kref_put(&bo->list_kref, ttm_bo_release_list);
1700 return ret;
1701 }
1702
1703 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1704 {
1705 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1706 ;
1707 }
1708 EXPORT_SYMBOL(ttm_bo_swapout_all);
1709
1710 /**
1711 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1712 * unreserved
1713 *
1714 * @bo: Pointer to buffer
1715 */
1716 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1717 {
1718 int ret;
1719
1720 /*
1721 * In the absense of a wait_unlocked API,
1722 * Use the bo::wu_mutex to avoid triggering livelocks due to
1723 * concurrent use of this function. Note that this use of
1724 * bo::wu_mutex can go away if we change locking order to
1725 * mmap_sem -> bo::reserve.
1726 */
1727 ret = mutex_lock_interruptible(&bo->wu_mutex);
1728 if (unlikely(ret != 0))
1729 return -ERESTARTSYS;
1730 if (!ww_mutex_is_locked(&bo->resv->lock))
1731 goto out_unlock;
1732 ret = __ttm_bo_reserve(bo, true, false, false, NULL);
1733 if (unlikely(ret != 0))
1734 goto out_unlock;
1735 __ttm_bo_unreserve(bo);
1736
1737 out_unlock:
1738 mutex_unlock(&bo->wu_mutex);
1739 return ret;
1740 }
1741