Loading...
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
49static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50static void ttm_bo_global_kobj_release(struct kobject *kobj);
51
52static struct attribute ttm_bo_count = {
53 .name = "bo_count",
54 .mode = S_IRUGO
55};
56
57static 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
70static 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
85static 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
104static 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
115static struct attribute *ttm_bo_global_attrs[] = {
116 &ttm_bo_count,
117 NULL
118};
119
120static const struct sysfs_ops ttm_bo_global_ops = {
121 .show = &ttm_bo_global_show
122};
123
124static 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
131static inline uint32_t ttm_bo_type_flags(unsigned type)
132{
133 return 1 << (type);
134}
135
136static 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 if (bo->ttm)
151 ttm_tt_destroy(bo->ttm);
152 atomic_dec(&bo->glob->bo_count);
153 if (bo->resv == &bo->ttm_resv)
154 reservation_object_fini(&bo->ttm_resv);
155 mutex_destroy(&bo->wu_mutex);
156 if (bo->destroy)
157 bo->destroy(bo);
158 else {
159 kfree(bo);
160 }
161 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
162}
163
164void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
165{
166 struct ttm_bo_device *bdev = bo->bdev;
167 struct ttm_mem_type_manager *man;
168
169 lockdep_assert_held(&bo->resv->lock.base);
170
171 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
172
173 BUG_ON(!list_empty(&bo->lru));
174
175 man = &bdev->man[bo->mem.mem_type];
176 list_add_tail(&bo->lru, &man->lru);
177 kref_get(&bo->list_kref);
178
179 if (bo->ttm && !(bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
180 list_add_tail(&bo->swap, &bo->glob->swap_lru);
181 kref_get(&bo->list_kref);
182 }
183 }
184}
185EXPORT_SYMBOL(ttm_bo_add_to_lru);
186
187int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
188{
189 int put_count = 0;
190
191 if (!list_empty(&bo->swap)) {
192 list_del_init(&bo->swap);
193 ++put_count;
194 }
195 if (!list_empty(&bo->lru)) {
196 list_del_init(&bo->lru);
197 ++put_count;
198 }
199
200 /*
201 * TODO: Add a driver hook to delete from
202 * driver-specific LRU's here.
203 */
204
205 return put_count;
206}
207
208static void ttm_bo_ref_bug(struct kref *list_kref)
209{
210 BUG();
211}
212
213void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
214 bool never_free)
215{
216 kref_sub(&bo->list_kref, count,
217 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
218}
219
220void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
221{
222 int put_count;
223
224 spin_lock(&bo->glob->lru_lock);
225 put_count = ttm_bo_del_from_lru(bo);
226 spin_unlock(&bo->glob->lru_lock);
227 ttm_bo_list_ref_sub(bo, put_count, true);
228}
229EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
230
231void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
232{
233 int put_count = 0;
234
235 lockdep_assert_held(&bo->resv->lock.base);
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}
241EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
242
243/*
244 * Call bo->mutex locked.
245 */
246static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
247{
248 struct ttm_bo_device *bdev = bo->bdev;
249 struct ttm_bo_global *glob = bo->glob;
250 int ret = 0;
251 uint32_t page_flags = 0;
252
253 TTM_ASSERT_LOCKED(&bo->mutex);
254 bo->ttm = NULL;
255
256 if (bdev->need_dma32)
257 page_flags |= TTM_PAGE_FLAG_DMA32;
258
259 switch (bo->type) {
260 case ttm_bo_type_device:
261 if (zero_alloc)
262 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
263 case ttm_bo_type_kernel:
264 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
265 page_flags, glob->dummy_read_page);
266 if (unlikely(bo->ttm == NULL))
267 ret = -ENOMEM;
268 break;
269 case ttm_bo_type_sg:
270 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
271 page_flags | TTM_PAGE_FLAG_SG,
272 glob->dummy_read_page);
273 if (unlikely(bo->ttm == NULL)) {
274 ret = -ENOMEM;
275 break;
276 }
277 bo->ttm->sg = bo->sg;
278 break;
279 default:
280 pr_err("Illegal buffer object type\n");
281 ret = -EINVAL;
282 break;
283 }
284
285 return ret;
286}
287
288static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
289 struct ttm_mem_reg *mem,
290 bool evict, bool interruptible,
291 bool no_wait_gpu)
292{
293 struct ttm_bo_device *bdev = bo->bdev;
294 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
295 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
296 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
297 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
298 int ret = 0;
299
300 if (old_is_pci || new_is_pci ||
301 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
302 ret = ttm_mem_io_lock(old_man, true);
303 if (unlikely(ret != 0))
304 goto out_err;
305 ttm_bo_unmap_virtual_locked(bo);
306 ttm_mem_io_unlock(old_man);
307 }
308
309 /*
310 * Create and bind a ttm if required.
311 */
312
313 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
314 if (bo->ttm == NULL) {
315 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
316 ret = ttm_bo_add_ttm(bo, zero);
317 if (ret)
318 goto out_err;
319 }
320
321 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
322 if (ret)
323 goto out_err;
324
325 if (mem->mem_type != TTM_PL_SYSTEM) {
326 ret = ttm_tt_bind(bo->ttm, mem);
327 if (ret)
328 goto out_err;
329 }
330
331 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
332 if (bdev->driver->move_notify)
333 bdev->driver->move_notify(bo, mem);
334 bo->mem = *mem;
335 mem->mm_node = NULL;
336 goto moved;
337 }
338 }
339
340 if (bdev->driver->move_notify)
341 bdev->driver->move_notify(bo, mem);
342
343 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
344 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
345 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
346 else if (bdev->driver->move)
347 ret = bdev->driver->move(bo, evict, interruptible,
348 no_wait_gpu, mem);
349 else
350 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
351
352 if (ret) {
353 if (bdev->driver->move_notify) {
354 struct ttm_mem_reg tmp_mem = *mem;
355 *mem = bo->mem;
356 bo->mem = tmp_mem;
357 bdev->driver->move_notify(bo, mem);
358 bo->mem = *mem;
359 *mem = tmp_mem;
360 }
361
362 goto out_err;
363 }
364
365moved:
366 if (bo->evicted) {
367 if (bdev->driver->invalidate_caches) {
368 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
369 if (ret)
370 pr_err("Can not flush read caches\n");
371 }
372 bo->evicted = false;
373 }
374
375 if (bo->mem.mm_node) {
376 bo->offset = (bo->mem.start << PAGE_SHIFT) +
377 bdev->man[bo->mem.mem_type].gpu_offset;
378 bo->cur_placement = bo->mem.placement;
379 } else
380 bo->offset = 0;
381
382 return 0;
383
384out_err:
385 new_man = &bdev->man[bo->mem.mem_type];
386 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
387 ttm_tt_unbind(bo->ttm);
388 ttm_tt_destroy(bo->ttm);
389 bo->ttm = NULL;
390 }
391
392 return ret;
393}
394
395/**
396 * Call bo::reserved.
397 * Will release GPU memory type usage on destruction.
398 * This is the place to put in driver specific hooks to release
399 * driver private resources.
400 * Will release the bo::reserved lock.
401 */
402
403static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
404{
405 if (bo->bdev->driver->move_notify)
406 bo->bdev->driver->move_notify(bo, NULL);
407
408 if (bo->ttm) {
409 ttm_tt_unbind(bo->ttm);
410 ttm_tt_destroy(bo->ttm);
411 bo->ttm = NULL;
412 }
413 ttm_bo_mem_put(bo, &bo->mem);
414
415 ww_mutex_unlock (&bo->resv->lock);
416}
417
418static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
419{
420 struct reservation_object_list *fobj;
421 struct fence *fence;
422 int i;
423
424 fobj = reservation_object_get_list(bo->resv);
425 fence = reservation_object_get_excl(bo->resv);
426 if (fence && !fence->ops->signaled)
427 fence_enable_sw_signaling(fence);
428
429 for (i = 0; fobj && i < fobj->shared_count; ++i) {
430 fence = rcu_dereference_protected(fobj->shared[i],
431 reservation_object_held(bo->resv));
432
433 if (!fence->ops->signaled)
434 fence_enable_sw_signaling(fence);
435 }
436}
437
438static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
439{
440 struct ttm_bo_device *bdev = bo->bdev;
441 struct ttm_bo_global *glob = bo->glob;
442 int put_count;
443 int ret;
444
445 spin_lock(&glob->lru_lock);
446 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
447
448 if (!ret) {
449 if (!ttm_bo_wait(bo, false, false, true)) {
450 put_count = ttm_bo_del_from_lru(bo);
451
452 spin_unlock(&glob->lru_lock);
453 ttm_bo_cleanup_memtype_use(bo);
454
455 ttm_bo_list_ref_sub(bo, put_count, true);
456
457 return;
458 } else
459 ttm_bo_flush_all_fences(bo);
460
461 /*
462 * Make NO_EVICT bos immediately available to
463 * shrinkers, now that they are queued for
464 * destruction.
465 */
466 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
467 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
468 ttm_bo_add_to_lru(bo);
469 }
470
471 __ttm_bo_unreserve(bo);
472 }
473
474 kref_get(&bo->list_kref);
475 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
476 spin_unlock(&glob->lru_lock);
477
478 schedule_delayed_work(&bdev->wq,
479 ((HZ / 100) < 1) ? 1 : HZ / 100);
480}
481
482/**
483 * function ttm_bo_cleanup_refs_and_unlock
484 * If bo idle, remove from delayed- and lru lists, and unref.
485 * If not idle, do nothing.
486 *
487 * Must be called with lru_lock and reservation held, this function
488 * will drop both before returning.
489 *
490 * @interruptible Any sleeps should occur interruptibly.
491 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
492 */
493
494static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
495 bool interruptible,
496 bool no_wait_gpu)
497{
498 struct ttm_bo_global *glob = bo->glob;
499 int put_count;
500 int ret;
501
502 ret = ttm_bo_wait(bo, false, false, true);
503
504 if (ret && !no_wait_gpu) {
505 long lret;
506 ww_mutex_unlock(&bo->resv->lock);
507 spin_unlock(&glob->lru_lock);
508
509 lret = reservation_object_wait_timeout_rcu(bo->resv,
510 true,
511 interruptible,
512 30 * HZ);
513
514 if (lret < 0)
515 return lret;
516 else if (lret == 0)
517 return -EBUSY;
518
519 spin_lock(&glob->lru_lock);
520 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
521
522 /*
523 * We raced, and lost, someone else holds the reservation now,
524 * and is probably busy in ttm_bo_cleanup_memtype_use.
525 *
526 * Even if it's not the case, because we finished waiting any
527 * delayed destruction would succeed, so just return success
528 * here.
529 */
530 if (ret) {
531 spin_unlock(&glob->lru_lock);
532 return 0;
533 }
534
535 /*
536 * remove sync_obj with ttm_bo_wait, the wait should be
537 * finished, and no new wait object should have been added.
538 */
539 ret = ttm_bo_wait(bo, false, false, true);
540 WARN_ON(ret);
541 }
542
543 if (ret || unlikely(list_empty(&bo->ddestroy))) {
544 __ttm_bo_unreserve(bo);
545 spin_unlock(&glob->lru_lock);
546 return ret;
547 }
548
549 put_count = ttm_bo_del_from_lru(bo);
550 list_del_init(&bo->ddestroy);
551 ++put_count;
552
553 spin_unlock(&glob->lru_lock);
554 ttm_bo_cleanup_memtype_use(bo);
555
556 ttm_bo_list_ref_sub(bo, put_count, true);
557
558 return 0;
559}
560
561/**
562 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
563 * encountered buffers.
564 */
565
566static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
567{
568 struct ttm_bo_global *glob = bdev->glob;
569 struct ttm_buffer_object *entry = NULL;
570 int ret = 0;
571
572 spin_lock(&glob->lru_lock);
573 if (list_empty(&bdev->ddestroy))
574 goto out_unlock;
575
576 entry = list_first_entry(&bdev->ddestroy,
577 struct ttm_buffer_object, ddestroy);
578 kref_get(&entry->list_kref);
579
580 for (;;) {
581 struct ttm_buffer_object *nentry = NULL;
582
583 if (entry->ddestroy.next != &bdev->ddestroy) {
584 nentry = list_first_entry(&entry->ddestroy,
585 struct ttm_buffer_object, ddestroy);
586 kref_get(&nentry->list_kref);
587 }
588
589 ret = __ttm_bo_reserve(entry, false, true, false, NULL);
590 if (remove_all && ret) {
591 spin_unlock(&glob->lru_lock);
592 ret = __ttm_bo_reserve(entry, false, false,
593 false, NULL);
594 spin_lock(&glob->lru_lock);
595 }
596
597 if (!ret)
598 ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
599 !remove_all);
600 else
601 spin_unlock(&glob->lru_lock);
602
603 kref_put(&entry->list_kref, ttm_bo_release_list);
604 entry = nentry;
605
606 if (ret || !entry)
607 goto out;
608
609 spin_lock(&glob->lru_lock);
610 if (list_empty(&entry->ddestroy))
611 break;
612 }
613
614out_unlock:
615 spin_unlock(&glob->lru_lock);
616out:
617 if (entry)
618 kref_put(&entry->list_kref, ttm_bo_release_list);
619 return ret;
620}
621
622static void ttm_bo_delayed_workqueue(struct work_struct *work)
623{
624 struct ttm_bo_device *bdev =
625 container_of(work, struct ttm_bo_device, wq.work);
626
627 if (ttm_bo_delayed_delete(bdev, false)) {
628 schedule_delayed_work(&bdev->wq,
629 ((HZ / 100) < 1) ? 1 : HZ / 100);
630 }
631}
632
633static void ttm_bo_release(struct kref *kref)
634{
635 struct ttm_buffer_object *bo =
636 container_of(kref, struct ttm_buffer_object, kref);
637 struct ttm_bo_device *bdev = bo->bdev;
638 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
639
640 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
641 ttm_mem_io_lock(man, false);
642 ttm_mem_io_free_vm(bo);
643 ttm_mem_io_unlock(man);
644 ttm_bo_cleanup_refs_or_queue(bo);
645 kref_put(&bo->list_kref, ttm_bo_release_list);
646}
647
648void ttm_bo_unref(struct ttm_buffer_object **p_bo)
649{
650 struct ttm_buffer_object *bo = *p_bo;
651
652 *p_bo = NULL;
653 kref_put(&bo->kref, ttm_bo_release);
654}
655EXPORT_SYMBOL(ttm_bo_unref);
656
657int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
658{
659 return cancel_delayed_work_sync(&bdev->wq);
660}
661EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
662
663void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
664{
665 if (resched)
666 schedule_delayed_work(&bdev->wq,
667 ((HZ / 100) < 1) ? 1 : HZ / 100);
668}
669EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
670
671static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
672 bool no_wait_gpu)
673{
674 struct ttm_bo_device *bdev = bo->bdev;
675 struct ttm_mem_reg evict_mem;
676 struct ttm_placement placement;
677 int ret = 0;
678
679 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
680
681 if (unlikely(ret != 0)) {
682 if (ret != -ERESTARTSYS) {
683 pr_err("Failed to expire sync object before buffer eviction\n");
684 }
685 goto out;
686 }
687
688 lockdep_assert_held(&bo->resv->lock.base);
689
690 evict_mem = bo->mem;
691 evict_mem.mm_node = NULL;
692 evict_mem.bus.io_reserved_vm = false;
693 evict_mem.bus.io_reserved_count = 0;
694
695 placement.num_placement = 0;
696 placement.num_busy_placement = 0;
697 bdev->driver->evict_flags(bo, &placement);
698 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
699 no_wait_gpu);
700 if (ret) {
701 if (ret != -ERESTARTSYS) {
702 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
703 bo);
704 ttm_bo_mem_space_debug(bo, &placement);
705 }
706 goto out;
707 }
708
709 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
710 no_wait_gpu);
711 if (ret) {
712 if (ret != -ERESTARTSYS)
713 pr_err("Buffer eviction failed\n");
714 ttm_bo_mem_put(bo, &evict_mem);
715 goto out;
716 }
717 bo->evicted = true;
718out:
719 return ret;
720}
721
722static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
723 uint32_t mem_type,
724 const struct ttm_place *place,
725 bool interruptible,
726 bool no_wait_gpu)
727{
728 struct ttm_bo_global *glob = bdev->glob;
729 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
730 struct ttm_buffer_object *bo;
731 int ret = -EBUSY, put_count;
732
733 spin_lock(&glob->lru_lock);
734 list_for_each_entry(bo, &man->lru, lru) {
735 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
736 if (!ret) {
737 if (place && (place->fpfn || place->lpfn)) {
738 /* Don't evict this BO if it's outside of the
739 * requested placement range
740 */
741 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
742 (place->lpfn && place->lpfn <= bo->mem.start)) {
743 __ttm_bo_unreserve(bo);
744 ret = -EBUSY;
745 continue;
746 }
747 }
748
749 break;
750 }
751 }
752
753 if (ret) {
754 spin_unlock(&glob->lru_lock);
755 return ret;
756 }
757
758 kref_get(&bo->list_kref);
759
760 if (!list_empty(&bo->ddestroy)) {
761 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
762 no_wait_gpu);
763 kref_put(&bo->list_kref, ttm_bo_release_list);
764 return ret;
765 }
766
767 put_count = ttm_bo_del_from_lru(bo);
768 spin_unlock(&glob->lru_lock);
769
770 BUG_ON(ret != 0);
771
772 ttm_bo_list_ref_sub(bo, put_count, true);
773
774 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
775 ttm_bo_unreserve(bo);
776
777 kref_put(&bo->list_kref, ttm_bo_release_list);
778 return ret;
779}
780
781void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
782{
783 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
784
785 if (mem->mm_node)
786 (*man->func->put_node)(man, mem);
787}
788EXPORT_SYMBOL(ttm_bo_mem_put);
789
790/**
791 * Repeatedly evict memory from the LRU for @mem_type until we create enough
792 * space, or we've evicted everything and there isn't enough space.
793 */
794static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
795 uint32_t mem_type,
796 const struct ttm_place *place,
797 struct ttm_mem_reg *mem,
798 bool interruptible,
799 bool no_wait_gpu)
800{
801 struct ttm_bo_device *bdev = bo->bdev;
802 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
803 int ret;
804
805 do {
806 ret = (*man->func->get_node)(man, bo, place, mem);
807 if (unlikely(ret != 0))
808 return ret;
809 if (mem->mm_node)
810 break;
811 ret = ttm_mem_evict_first(bdev, mem_type, place,
812 interruptible, no_wait_gpu);
813 if (unlikely(ret != 0))
814 return ret;
815 } while (1);
816 if (mem->mm_node == NULL)
817 return -ENOMEM;
818 mem->mem_type = mem_type;
819 return 0;
820}
821
822static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
823 uint32_t cur_placement,
824 uint32_t proposed_placement)
825{
826 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
827 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
828
829 /**
830 * Keep current caching if possible.
831 */
832
833 if ((cur_placement & caching) != 0)
834 result |= (cur_placement & caching);
835 else if ((man->default_caching & caching) != 0)
836 result |= man->default_caching;
837 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
838 result |= TTM_PL_FLAG_CACHED;
839 else if ((TTM_PL_FLAG_WC & caching) != 0)
840 result |= TTM_PL_FLAG_WC;
841 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
842 result |= TTM_PL_FLAG_UNCACHED;
843
844 return result;
845}
846
847static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
848 uint32_t mem_type,
849 const struct ttm_place *place,
850 uint32_t *masked_placement)
851{
852 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
853
854 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
855 return false;
856
857 if ((place->flags & man->available_caching) == 0)
858 return false;
859
860 cur_flags |= (place->flags & man->available_caching);
861
862 *masked_placement = cur_flags;
863 return true;
864}
865
866/**
867 * Creates space for memory region @mem according to its type.
868 *
869 * This function first searches for free space in compatible memory types in
870 * the priority order defined by the driver. If free space isn't found, then
871 * ttm_bo_mem_force_space is attempted in priority order to evict and find
872 * space.
873 */
874int ttm_bo_mem_space(struct ttm_buffer_object *bo,
875 struct ttm_placement *placement,
876 struct ttm_mem_reg *mem,
877 bool interruptible,
878 bool no_wait_gpu)
879{
880 struct ttm_bo_device *bdev = bo->bdev;
881 struct ttm_mem_type_manager *man;
882 uint32_t mem_type = TTM_PL_SYSTEM;
883 uint32_t cur_flags = 0;
884 bool type_found = false;
885 bool type_ok = false;
886 bool has_erestartsys = false;
887 int i, ret;
888
889 mem->mm_node = NULL;
890 for (i = 0; i < placement->num_placement; ++i) {
891 const struct ttm_place *place = &placement->placement[i];
892
893 ret = ttm_mem_type_from_place(place, &mem_type);
894 if (ret)
895 return ret;
896 man = &bdev->man[mem_type];
897 if (!man->has_type || !man->use_type)
898 continue;
899
900 type_ok = ttm_bo_mt_compatible(man, mem_type, place,
901 &cur_flags);
902
903 if (!type_ok)
904 continue;
905
906 type_found = true;
907 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
908 cur_flags);
909 /*
910 * Use the access and other non-mapping-related flag bits from
911 * the memory placement flags to the current flags
912 */
913 ttm_flag_masked(&cur_flags, place->flags,
914 ~TTM_PL_MASK_MEMTYPE);
915
916 if (mem_type == TTM_PL_SYSTEM)
917 break;
918
919 ret = (*man->func->get_node)(man, bo, place, mem);
920 if (unlikely(ret))
921 return ret;
922
923 if (mem->mm_node)
924 break;
925 }
926
927 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
928 mem->mem_type = mem_type;
929 mem->placement = cur_flags;
930 return 0;
931 }
932
933 for (i = 0; i < placement->num_busy_placement; ++i) {
934 const struct ttm_place *place = &placement->busy_placement[i];
935
936 ret = ttm_mem_type_from_place(place, &mem_type);
937 if (ret)
938 return ret;
939 man = &bdev->man[mem_type];
940 if (!man->has_type || !man->use_type)
941 continue;
942 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
943 continue;
944
945 type_found = true;
946 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
947 cur_flags);
948 /*
949 * Use the access and other non-mapping-related flag bits from
950 * the memory placement flags to the current flags
951 */
952 ttm_flag_masked(&cur_flags, place->flags,
953 ~TTM_PL_MASK_MEMTYPE);
954
955 if (mem_type == TTM_PL_SYSTEM) {
956 mem->mem_type = mem_type;
957 mem->placement = cur_flags;
958 mem->mm_node = NULL;
959 return 0;
960 }
961
962 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
963 interruptible, no_wait_gpu);
964 if (ret == 0 && mem->mm_node) {
965 mem->placement = cur_flags;
966 return 0;
967 }
968 if (ret == -ERESTARTSYS)
969 has_erestartsys = true;
970 }
971
972 if (!type_found) {
973 printk(KERN_ERR TTM_PFX "No compatible memory type found.\n");
974 return -EINVAL;
975 }
976
977 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
978}
979EXPORT_SYMBOL(ttm_bo_mem_space);
980
981static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
982 struct ttm_placement *placement,
983 bool interruptible,
984 bool no_wait_gpu)
985{
986 int ret = 0;
987 struct ttm_mem_reg mem;
988
989 lockdep_assert_held(&bo->resv->lock.base);
990
991 /*
992 * FIXME: It's possible to pipeline buffer moves.
993 * Have the driver move function wait for idle when necessary,
994 * instead of doing it here.
995 */
996 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
997 if (ret)
998 return ret;
999 mem.num_pages = bo->num_pages;
1000 mem.size = mem.num_pages << PAGE_SHIFT;
1001 mem.page_alignment = bo->mem.page_alignment;
1002 mem.bus.io_reserved_vm = false;
1003 mem.bus.io_reserved_count = 0;
1004 /*
1005 * Determine where to move the buffer.
1006 */
1007 ret = ttm_bo_mem_space(bo, placement, &mem,
1008 interruptible, no_wait_gpu);
1009 if (ret)
1010 goto out_unlock;
1011 ret = ttm_bo_handle_move_mem(bo, &mem, false,
1012 interruptible, no_wait_gpu);
1013out_unlock:
1014 if (ret && mem.mm_node)
1015 ttm_bo_mem_put(bo, &mem);
1016 return ret;
1017}
1018
1019static bool ttm_bo_mem_compat(struct ttm_placement *placement,
1020 struct ttm_mem_reg *mem,
1021 uint32_t *new_flags)
1022{
1023 int i;
1024
1025 for (i = 0; i < placement->num_placement; i++) {
1026 const struct ttm_place *heap = &placement->placement[i];
1027 if (mem->mm_node &&
1028 (mem->start < heap->fpfn ||
1029 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1030 continue;
1031
1032 *new_flags = heap->flags;
1033 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1034 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1035 return true;
1036 }
1037
1038 for (i = 0; i < placement->num_busy_placement; i++) {
1039 const struct ttm_place *heap = &placement->busy_placement[i];
1040 if (mem->mm_node &&
1041 (mem->start < heap->fpfn ||
1042 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1043 continue;
1044
1045 *new_flags = heap->flags;
1046 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1047 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1048 return true;
1049 }
1050
1051 return false;
1052}
1053
1054int ttm_bo_validate(struct ttm_buffer_object *bo,
1055 struct ttm_placement *placement,
1056 bool interruptible,
1057 bool no_wait_gpu)
1058{
1059 int ret;
1060 uint32_t new_flags;
1061
1062 lockdep_assert_held(&bo->resv->lock.base);
1063 /*
1064 * Check whether we need to move buffer.
1065 */
1066 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1067 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1068 no_wait_gpu);
1069 if (ret)
1070 return ret;
1071 } else {
1072 /*
1073 * Use the access and other non-mapping-related flag bits from
1074 * the compatible memory placement flags to the active flags
1075 */
1076 ttm_flag_masked(&bo->mem.placement, new_flags,
1077 ~TTM_PL_MASK_MEMTYPE);
1078 }
1079 /*
1080 * We might need to add a TTM.
1081 */
1082 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1083 ret = ttm_bo_add_ttm(bo, true);
1084 if (ret)
1085 return ret;
1086 }
1087 return 0;
1088}
1089EXPORT_SYMBOL(ttm_bo_validate);
1090
1091int ttm_bo_init(struct ttm_bo_device *bdev,
1092 struct ttm_buffer_object *bo,
1093 unsigned long size,
1094 enum ttm_bo_type type,
1095 struct ttm_placement *placement,
1096 uint32_t page_alignment,
1097 bool interruptible,
1098 struct file *persistent_swap_storage,
1099 size_t acc_size,
1100 struct sg_table *sg,
1101 struct reservation_object *resv,
1102 void (*destroy) (struct ttm_buffer_object *))
1103{
1104 int ret = 0;
1105 unsigned long num_pages;
1106 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1107 bool locked;
1108
1109 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1110 if (ret) {
1111 pr_err("Out of kernel memory\n");
1112 if (destroy)
1113 (*destroy)(bo);
1114 else
1115 kfree(bo);
1116 return -ENOMEM;
1117 }
1118
1119 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1120 if (num_pages == 0) {
1121 pr_err("Illegal buffer object size\n");
1122 if (destroy)
1123 (*destroy)(bo);
1124 else
1125 kfree(bo);
1126 ttm_mem_global_free(mem_glob, acc_size);
1127 return -EINVAL;
1128 }
1129 bo->destroy = destroy;
1130
1131 kref_init(&bo->kref);
1132 kref_init(&bo->list_kref);
1133 atomic_set(&bo->cpu_writers, 0);
1134 INIT_LIST_HEAD(&bo->lru);
1135 INIT_LIST_HEAD(&bo->ddestroy);
1136 INIT_LIST_HEAD(&bo->swap);
1137 INIT_LIST_HEAD(&bo->io_reserve_lru);
1138 mutex_init(&bo->wu_mutex);
1139 bo->bdev = bdev;
1140 bo->glob = bdev->glob;
1141 bo->type = type;
1142 bo->num_pages = num_pages;
1143 bo->mem.size = num_pages << PAGE_SHIFT;
1144 bo->mem.mem_type = TTM_PL_SYSTEM;
1145 bo->mem.num_pages = bo->num_pages;
1146 bo->mem.mm_node = NULL;
1147 bo->mem.page_alignment = page_alignment;
1148 bo->mem.bus.io_reserved_vm = false;
1149 bo->mem.bus.io_reserved_count = 0;
1150 bo->priv_flags = 0;
1151 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1152 bo->persistent_swap_storage = persistent_swap_storage;
1153 bo->acc_size = acc_size;
1154 bo->sg = sg;
1155 if (resv) {
1156 bo->resv = resv;
1157 lockdep_assert_held(&bo->resv->lock.base);
1158 } else {
1159 bo->resv = &bo->ttm_resv;
1160 reservation_object_init(&bo->ttm_resv);
1161 }
1162 atomic_inc(&bo->glob->bo_count);
1163 drm_vma_node_reset(&bo->vma_node);
1164
1165 /*
1166 * For ttm_bo_type_device buffers, allocate
1167 * address space from the device.
1168 */
1169 if (bo->type == ttm_bo_type_device ||
1170 bo->type == ttm_bo_type_sg)
1171 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1172 bo->mem.num_pages);
1173
1174 /* passed reservation objects should already be locked,
1175 * since otherwise lockdep will be angered in radeon.
1176 */
1177 if (!resv) {
1178 locked = ww_mutex_trylock(&bo->resv->lock);
1179 WARN_ON(!locked);
1180 }
1181
1182 if (likely(!ret))
1183 ret = ttm_bo_validate(bo, placement, interruptible, false);
1184
1185 if (!resv) {
1186 ttm_bo_unreserve(bo);
1187
1188 } else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1189 spin_lock(&bo->glob->lru_lock);
1190 ttm_bo_add_to_lru(bo);
1191 spin_unlock(&bo->glob->lru_lock);
1192 }
1193
1194 if (unlikely(ret))
1195 ttm_bo_unref(&bo);
1196
1197 return ret;
1198}
1199EXPORT_SYMBOL(ttm_bo_init);
1200
1201size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1202 unsigned long bo_size,
1203 unsigned struct_size)
1204{
1205 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1206 size_t size = 0;
1207
1208 size += ttm_round_pot(struct_size);
1209 size += PAGE_ALIGN(npages * sizeof(void *));
1210 size += ttm_round_pot(sizeof(struct ttm_tt));
1211 return size;
1212}
1213EXPORT_SYMBOL(ttm_bo_acc_size);
1214
1215size_t ttm_bo_dma_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 += PAGE_ALIGN(npages * sizeof(void *));
1224 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1225 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1226 return size;
1227}
1228EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1229
1230int ttm_bo_create(struct ttm_bo_device *bdev,
1231 unsigned long size,
1232 enum ttm_bo_type type,
1233 struct ttm_placement *placement,
1234 uint32_t page_alignment,
1235 bool interruptible,
1236 struct file *persistent_swap_storage,
1237 struct ttm_buffer_object **p_bo)
1238{
1239 struct ttm_buffer_object *bo;
1240 size_t acc_size;
1241 int ret;
1242
1243 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1244 if (unlikely(bo == NULL))
1245 return -ENOMEM;
1246
1247 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1248 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1249 interruptible, persistent_swap_storage, acc_size,
1250 NULL, NULL, NULL);
1251 if (likely(ret == 0))
1252 *p_bo = bo;
1253
1254 return ret;
1255}
1256EXPORT_SYMBOL(ttm_bo_create);
1257
1258static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1259 unsigned mem_type, bool allow_errors)
1260{
1261 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1262 struct ttm_bo_global *glob = bdev->glob;
1263 int ret;
1264
1265 /*
1266 * Can't use standard list traversal since we're unlocking.
1267 */
1268
1269 spin_lock(&glob->lru_lock);
1270 while (!list_empty(&man->lru)) {
1271 spin_unlock(&glob->lru_lock);
1272 ret = ttm_mem_evict_first(bdev, mem_type, NULL, false, false);
1273 if (ret) {
1274 if (allow_errors) {
1275 return ret;
1276 } else {
1277 pr_err("Cleanup eviction failed\n");
1278 }
1279 }
1280 spin_lock(&glob->lru_lock);
1281 }
1282 spin_unlock(&glob->lru_lock);
1283 return 0;
1284}
1285
1286int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1287{
1288 struct ttm_mem_type_manager *man;
1289 int ret = -EINVAL;
1290
1291 if (mem_type >= TTM_NUM_MEM_TYPES) {
1292 pr_err("Illegal memory type %d\n", mem_type);
1293 return ret;
1294 }
1295 man = &bdev->man[mem_type];
1296
1297 if (!man->has_type) {
1298 pr_err("Trying to take down uninitialized memory manager type %u\n",
1299 mem_type);
1300 return ret;
1301 }
1302
1303 man->use_type = false;
1304 man->has_type = false;
1305
1306 ret = 0;
1307 if (mem_type > 0) {
1308 ttm_bo_force_list_clean(bdev, mem_type, false);
1309
1310 ret = (*man->func->takedown)(man);
1311 }
1312
1313 return ret;
1314}
1315EXPORT_SYMBOL(ttm_bo_clean_mm);
1316
1317int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1318{
1319 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1320
1321 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1322 pr_err("Illegal memory manager memory type %u\n", mem_type);
1323 return -EINVAL;
1324 }
1325
1326 if (!man->has_type) {
1327 pr_err("Memory type %u has not been initialized\n", mem_type);
1328 return 0;
1329 }
1330
1331 return ttm_bo_force_list_clean(bdev, mem_type, true);
1332}
1333EXPORT_SYMBOL(ttm_bo_evict_mm);
1334
1335int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1336 unsigned long p_size)
1337{
1338 int ret = -EINVAL;
1339 struct ttm_mem_type_manager *man;
1340
1341 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1342 man = &bdev->man[type];
1343 BUG_ON(man->has_type);
1344 man->io_reserve_fastpath = true;
1345 man->use_io_reserve_lru = false;
1346 mutex_init(&man->io_reserve_mutex);
1347 INIT_LIST_HEAD(&man->io_reserve_lru);
1348
1349 ret = bdev->driver->init_mem_type(bdev, type, man);
1350 if (ret)
1351 return ret;
1352 man->bdev = bdev;
1353
1354 ret = 0;
1355 if (type != TTM_PL_SYSTEM) {
1356 ret = (*man->func->init)(man, p_size);
1357 if (ret)
1358 return ret;
1359 }
1360 man->has_type = true;
1361 man->use_type = true;
1362 man->size = p_size;
1363
1364 INIT_LIST_HEAD(&man->lru);
1365
1366 return 0;
1367}
1368EXPORT_SYMBOL(ttm_bo_init_mm);
1369
1370static void ttm_bo_global_kobj_release(struct kobject *kobj)
1371{
1372 struct ttm_bo_global *glob =
1373 container_of(kobj, struct ttm_bo_global, kobj);
1374
1375 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1376 __free_page(glob->dummy_read_page);
1377 kfree(glob);
1378}
1379
1380void ttm_bo_global_release(struct drm_global_reference *ref)
1381{
1382 struct ttm_bo_global *glob = ref->object;
1383
1384 kobject_del(&glob->kobj);
1385 kobject_put(&glob->kobj);
1386}
1387EXPORT_SYMBOL(ttm_bo_global_release);
1388
1389int ttm_bo_global_init(struct drm_global_reference *ref)
1390{
1391 struct ttm_bo_global_ref *bo_ref =
1392 container_of(ref, struct ttm_bo_global_ref, ref);
1393 struct ttm_bo_global *glob = ref->object;
1394 int ret;
1395
1396 mutex_init(&glob->device_list_mutex);
1397 spin_lock_init(&glob->lru_lock);
1398 glob->mem_glob = bo_ref->mem_glob;
1399 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1400
1401 if (unlikely(glob->dummy_read_page == NULL)) {
1402 ret = -ENOMEM;
1403 goto out_no_drp;
1404 }
1405
1406 INIT_LIST_HEAD(&glob->swap_lru);
1407 INIT_LIST_HEAD(&glob->device_list);
1408
1409 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1410 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1411 if (unlikely(ret != 0)) {
1412 pr_err("Could not register buffer object swapout\n");
1413 goto out_no_shrink;
1414 }
1415
1416 atomic_set(&glob->bo_count, 0);
1417
1418 ret = kobject_init_and_add(
1419 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1420 if (unlikely(ret != 0))
1421 kobject_put(&glob->kobj);
1422 return ret;
1423out_no_shrink:
1424 __free_page(glob->dummy_read_page);
1425out_no_drp:
1426 kfree(glob);
1427 return ret;
1428}
1429EXPORT_SYMBOL(ttm_bo_global_init);
1430
1431
1432int ttm_bo_device_release(struct ttm_bo_device *bdev)
1433{
1434 int ret = 0;
1435 unsigned i = TTM_NUM_MEM_TYPES;
1436 struct ttm_mem_type_manager *man;
1437 struct ttm_bo_global *glob = bdev->glob;
1438
1439 while (i--) {
1440 man = &bdev->man[i];
1441 if (man->has_type) {
1442 man->use_type = false;
1443 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1444 ret = -EBUSY;
1445 pr_err("DRM memory manager type %d is not clean\n",
1446 i);
1447 }
1448 man->has_type = false;
1449 }
1450 }
1451
1452 mutex_lock(&glob->device_list_mutex);
1453 list_del(&bdev->device_list);
1454 mutex_unlock(&glob->device_list_mutex);
1455
1456 cancel_delayed_work_sync(&bdev->wq);
1457
1458 while (ttm_bo_delayed_delete(bdev, true))
1459 ;
1460
1461 spin_lock(&glob->lru_lock);
1462 if (list_empty(&bdev->ddestroy))
1463 TTM_DEBUG("Delayed destroy list was clean\n");
1464
1465 if (list_empty(&bdev->man[0].lru))
1466 TTM_DEBUG("Swap list was clean\n");
1467 spin_unlock(&glob->lru_lock);
1468
1469 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1470
1471 return ret;
1472}
1473EXPORT_SYMBOL(ttm_bo_device_release);
1474
1475int ttm_bo_device_init(struct ttm_bo_device *bdev,
1476 struct ttm_bo_global *glob,
1477 struct ttm_bo_driver *driver,
1478 struct address_space *mapping,
1479 uint64_t file_page_offset,
1480 bool need_dma32)
1481{
1482 int ret = -EINVAL;
1483
1484 bdev->driver = driver;
1485
1486 memset(bdev->man, 0, sizeof(bdev->man));
1487
1488 /*
1489 * Initialize the system memory buffer type.
1490 * Other types need to be driver / IOCTL initialized.
1491 */
1492 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1493 if (unlikely(ret != 0))
1494 goto out_no_sys;
1495
1496 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1497 0x10000000);
1498 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1499 INIT_LIST_HEAD(&bdev->ddestroy);
1500 bdev->dev_mapping = mapping;
1501 bdev->glob = glob;
1502 bdev->need_dma32 = need_dma32;
1503 bdev->val_seq = 0;
1504 mutex_lock(&glob->device_list_mutex);
1505 list_add_tail(&bdev->device_list, &glob->device_list);
1506 mutex_unlock(&glob->device_list_mutex);
1507
1508 return 0;
1509out_no_sys:
1510 return ret;
1511}
1512EXPORT_SYMBOL(ttm_bo_device_init);
1513
1514/*
1515 * buffer object vm functions.
1516 */
1517
1518bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1519{
1520 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1521
1522 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1523 if (mem->mem_type == TTM_PL_SYSTEM)
1524 return false;
1525
1526 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1527 return false;
1528
1529 if (mem->placement & TTM_PL_FLAG_CACHED)
1530 return false;
1531 }
1532 return true;
1533}
1534
1535void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1536{
1537 struct ttm_bo_device *bdev = bo->bdev;
1538
1539 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1540 ttm_mem_io_free_vm(bo);
1541}
1542
1543void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1544{
1545 struct ttm_bo_device *bdev = bo->bdev;
1546 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1547
1548 ttm_mem_io_lock(man, false);
1549 ttm_bo_unmap_virtual_locked(bo);
1550 ttm_mem_io_unlock(man);
1551}
1552
1553
1554EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1555
1556int ttm_bo_wait(struct ttm_buffer_object *bo,
1557 bool lazy, bool interruptible, bool no_wait)
1558{
1559 struct reservation_object_list *fobj;
1560 struct reservation_object *resv;
1561 struct fence *excl;
1562 long timeout = 15 * HZ;
1563 int i;
1564
1565 resv = bo->resv;
1566 fobj = reservation_object_get_list(resv);
1567 excl = reservation_object_get_excl(resv);
1568 if (excl) {
1569 if (!fence_is_signaled(excl)) {
1570 if (no_wait)
1571 return -EBUSY;
1572
1573 timeout = fence_wait_timeout(excl,
1574 interruptible, timeout);
1575 }
1576 }
1577
1578 for (i = 0; fobj && timeout > 0 && i < fobj->shared_count; ++i) {
1579 struct fence *fence;
1580 fence = rcu_dereference_protected(fobj->shared[i],
1581 reservation_object_held(resv));
1582
1583 if (!fence_is_signaled(fence)) {
1584 if (no_wait)
1585 return -EBUSY;
1586
1587 timeout = fence_wait_timeout(fence,
1588 interruptible, timeout);
1589 }
1590 }
1591
1592 if (timeout < 0)
1593 return timeout;
1594
1595 if (timeout == 0)
1596 return -EBUSY;
1597
1598 reservation_object_add_excl_fence(resv, NULL);
1599 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1600 return 0;
1601}
1602EXPORT_SYMBOL(ttm_bo_wait);
1603
1604int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1605{
1606 int ret = 0;
1607
1608 /*
1609 * Using ttm_bo_reserve makes sure the lru lists are updated.
1610 */
1611
1612 ret = ttm_bo_reserve(bo, true, no_wait, false, NULL);
1613 if (unlikely(ret != 0))
1614 return ret;
1615 ret = ttm_bo_wait(bo, false, true, no_wait);
1616 if (likely(ret == 0))
1617 atomic_inc(&bo->cpu_writers);
1618 ttm_bo_unreserve(bo);
1619 return ret;
1620}
1621EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1622
1623void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1624{
1625 atomic_dec(&bo->cpu_writers);
1626}
1627EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1628
1629/**
1630 * A buffer object shrink method that tries to swap out the first
1631 * buffer object on the bo_global::swap_lru list.
1632 */
1633
1634static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1635{
1636 struct ttm_bo_global *glob =
1637 container_of(shrink, struct ttm_bo_global, shrink);
1638 struct ttm_buffer_object *bo;
1639 int ret = -EBUSY;
1640 int put_count;
1641 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1642
1643 spin_lock(&glob->lru_lock);
1644 list_for_each_entry(bo, &glob->swap_lru, swap) {
1645 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
1646 if (!ret)
1647 break;
1648 }
1649
1650 if (ret) {
1651 spin_unlock(&glob->lru_lock);
1652 return ret;
1653 }
1654
1655 kref_get(&bo->list_kref);
1656
1657 if (!list_empty(&bo->ddestroy)) {
1658 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1659 kref_put(&bo->list_kref, ttm_bo_release_list);
1660 return ret;
1661 }
1662
1663 put_count = ttm_bo_del_from_lru(bo);
1664 spin_unlock(&glob->lru_lock);
1665
1666 ttm_bo_list_ref_sub(bo, put_count, true);
1667
1668 /**
1669 * Wait for GPU, then move to system cached.
1670 */
1671
1672 ret = ttm_bo_wait(bo, false, false, false);
1673
1674 if (unlikely(ret != 0))
1675 goto out;
1676
1677 if ((bo->mem.placement & swap_placement) != swap_placement) {
1678 struct ttm_mem_reg evict_mem;
1679
1680 evict_mem = bo->mem;
1681 evict_mem.mm_node = NULL;
1682 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1683 evict_mem.mem_type = TTM_PL_SYSTEM;
1684
1685 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1686 false, false);
1687 if (unlikely(ret != 0))
1688 goto out;
1689 }
1690
1691 ttm_bo_unmap_virtual(bo);
1692
1693 /**
1694 * Swap out. Buffer will be swapped in again as soon as
1695 * anyone tries to access a ttm page.
1696 */
1697
1698 if (bo->bdev->driver->swap_notify)
1699 bo->bdev->driver->swap_notify(bo);
1700
1701 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1702out:
1703
1704 /**
1705 *
1706 * Unreserve without putting on LRU to avoid swapping out an
1707 * already swapped buffer.
1708 */
1709
1710 __ttm_bo_unreserve(bo);
1711 kref_put(&bo->list_kref, ttm_bo_release_list);
1712 return ret;
1713}
1714
1715void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1716{
1717 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1718 ;
1719}
1720EXPORT_SYMBOL(ttm_bo_swapout_all);
1721
1722/**
1723 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1724 * unreserved
1725 *
1726 * @bo: Pointer to buffer
1727 */
1728int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1729{
1730 int ret;
1731
1732 /*
1733 * In the absense of a wait_unlocked API,
1734 * Use the bo::wu_mutex to avoid triggering livelocks due to
1735 * concurrent use of this function. Note that this use of
1736 * bo::wu_mutex can go away if we change locking order to
1737 * mmap_sem -> bo::reserve.
1738 */
1739 ret = mutex_lock_interruptible(&bo->wu_mutex);
1740 if (unlikely(ret != 0))
1741 return -ERESTARTSYS;
1742 if (!ww_mutex_is_locked(&bo->resv->lock))
1743 goto out_unlock;
1744 ret = __ttm_bo_reserve(bo, true, false, false, NULL);
1745 if (unlikely(ret != 0))
1746 goto out_unlock;
1747 __ttm_bo_unreserve(bo);
1748
1749out_unlock:
1750 mutex_unlock(&bo->wu_mutex);
1751 return ret;
1752}
1/* SPDX-License-Identifier: GPL-2.0 OR MIT */
2/**************************************************************************
3 *
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28/*
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30 */
31
32#define pr_fmt(fmt) "[TTM] " fmt
33
34#include <drm/ttm/ttm_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/dma-resv.h>
44
45#include "ttm_module.h"
46
47/* default destructor */
48static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
49{
50 kfree(bo);
51}
52
53static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
54 struct ttm_placement *placement)
55{
56 struct drm_printer p = drm_debug_printer(TTM_PFX);
57 struct ttm_resource_manager *man;
58 int i, mem_type;
59
60 drm_printf(&p, "No space for %p (%lu pages, %zuK, %zuM)\n",
61 bo, bo->resource->num_pages, bo->base.size >> 10,
62 bo->base.size >> 20);
63 for (i = 0; i < placement->num_placement; i++) {
64 mem_type = placement->placement[i].mem_type;
65 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
66 i, placement->placement[i].flags, mem_type);
67 man = ttm_manager_type(bo->bdev, mem_type);
68 ttm_resource_manager_debug(man, &p);
69 }
70}
71
72static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
73{
74 struct ttm_device *bdev = bo->bdev;
75
76 list_del_init(&bo->lru);
77
78 if (bdev->funcs->del_from_lru_notify)
79 bdev->funcs->del_from_lru_notify(bo);
80}
81
82static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
83 struct ttm_buffer_object *bo)
84{
85 if (!pos->first)
86 pos->first = bo;
87 pos->last = bo;
88}
89
90void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
91 struct ttm_resource *mem,
92 struct ttm_lru_bulk_move *bulk)
93{
94 struct ttm_device *bdev = bo->bdev;
95 struct ttm_resource_manager *man;
96
97 if (!bo->deleted)
98 dma_resv_assert_held(bo->base.resv);
99
100 if (bo->pin_count) {
101 ttm_bo_del_from_lru(bo);
102 return;
103 }
104
105 if (!mem)
106 return;
107
108 man = ttm_manager_type(bdev, mem->mem_type);
109 list_move_tail(&bo->lru, &man->lru[bo->priority]);
110
111 if (bdev->funcs->del_from_lru_notify)
112 bdev->funcs->del_from_lru_notify(bo);
113
114 if (bulk && !bo->pin_count) {
115 switch (bo->resource->mem_type) {
116 case TTM_PL_TT:
117 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
118 break;
119
120 case TTM_PL_VRAM:
121 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
122 break;
123 }
124 }
125}
126EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
127
128void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
129{
130 unsigned i;
131
132 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
133 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
134 struct ttm_resource_manager *man;
135
136 if (!pos->first)
137 continue;
138
139 dma_resv_assert_held(pos->first->base.resv);
140 dma_resv_assert_held(pos->last->base.resv);
141
142 man = ttm_manager_type(pos->first->bdev, TTM_PL_TT);
143 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
144 &pos->last->lru);
145 }
146
147 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
148 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
149 struct ttm_resource_manager *man;
150
151 if (!pos->first)
152 continue;
153
154 dma_resv_assert_held(pos->first->base.resv);
155 dma_resv_assert_held(pos->last->base.resv);
156
157 man = ttm_manager_type(pos->first->bdev, TTM_PL_VRAM);
158 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
159 &pos->last->lru);
160 }
161}
162EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
163
164static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
165 struct ttm_resource *mem, bool evict,
166 struct ttm_operation_ctx *ctx,
167 struct ttm_place *hop)
168{
169 struct ttm_resource_manager *old_man, *new_man;
170 struct ttm_device *bdev = bo->bdev;
171 int ret;
172
173 old_man = ttm_manager_type(bdev, bo->resource->mem_type);
174 new_man = ttm_manager_type(bdev, mem->mem_type);
175
176 ttm_bo_unmap_virtual(bo);
177
178 /*
179 * Create and bind a ttm if required.
180 */
181
182 if (new_man->use_tt) {
183 /* Zero init the new TTM structure if the old location should
184 * have used one as well.
185 */
186 ret = ttm_tt_create(bo, old_man->use_tt);
187 if (ret)
188 goto out_err;
189
190 if (mem->mem_type != TTM_PL_SYSTEM) {
191 ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
192 if (ret)
193 goto out_err;
194 }
195 }
196
197 ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
198 if (ret) {
199 if (ret == -EMULTIHOP)
200 return ret;
201 goto out_err;
202 }
203
204 ctx->bytes_moved += bo->base.size;
205 return 0;
206
207out_err:
208 new_man = ttm_manager_type(bdev, bo->resource->mem_type);
209 if (!new_man->use_tt)
210 ttm_bo_tt_destroy(bo);
211
212 return ret;
213}
214
215/*
216 * Call bo::reserved.
217 * Will release GPU memory type usage on destruction.
218 * This is the place to put in driver specific hooks to release
219 * driver private resources.
220 * Will release the bo::reserved lock.
221 */
222
223static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
224{
225 if (bo->bdev->funcs->delete_mem_notify)
226 bo->bdev->funcs->delete_mem_notify(bo);
227
228 ttm_bo_tt_destroy(bo);
229 ttm_resource_free(bo, &bo->resource);
230}
231
232static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
233{
234 int r;
235
236 if (bo->base.resv == &bo->base._resv)
237 return 0;
238
239 BUG_ON(!dma_resv_trylock(&bo->base._resv));
240
241 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
242 dma_resv_unlock(&bo->base._resv);
243 if (r)
244 return r;
245
246 if (bo->type != ttm_bo_type_sg) {
247 /* This works because the BO is about to be destroyed and nobody
248 * reference it any more. The only tricky case is the trylock on
249 * the resv object while holding the lru_lock.
250 */
251 spin_lock(&bo->bdev->lru_lock);
252 bo->base.resv = &bo->base._resv;
253 spin_unlock(&bo->bdev->lru_lock);
254 }
255
256 return r;
257}
258
259static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
260{
261 struct dma_resv *resv = &bo->base._resv;
262 struct dma_resv_list *fobj;
263 struct dma_fence *fence;
264 int i;
265
266 rcu_read_lock();
267 fobj = dma_resv_shared_list(resv);
268 fence = dma_resv_excl_fence(resv);
269 if (fence && !fence->ops->signaled)
270 dma_fence_enable_sw_signaling(fence);
271
272 for (i = 0; fobj && i < fobj->shared_count; ++i) {
273 fence = rcu_dereference(fobj->shared[i]);
274
275 if (!fence->ops->signaled)
276 dma_fence_enable_sw_signaling(fence);
277 }
278 rcu_read_unlock();
279}
280
281/**
282 * ttm_bo_cleanup_refs
283 * If bo idle, remove from lru lists, and unref.
284 * If not idle, block if possible.
285 *
286 * Must be called with lru_lock and reservation held, this function
287 * will drop the lru lock and optionally the reservation lock before returning.
288 *
289 * @bo: The buffer object to clean-up
290 * @interruptible: Any sleeps should occur interruptibly.
291 * @no_wait_gpu: Never wait for gpu. Return -EBUSY instead.
292 * @unlock_resv: Unlock the reservation lock as well.
293 */
294
295static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
296 bool interruptible, bool no_wait_gpu,
297 bool unlock_resv)
298{
299 struct dma_resv *resv = &bo->base._resv;
300 int ret;
301
302 if (dma_resv_test_signaled(resv, true))
303 ret = 0;
304 else
305 ret = -EBUSY;
306
307 if (ret && !no_wait_gpu) {
308 long lret;
309
310 if (unlock_resv)
311 dma_resv_unlock(bo->base.resv);
312 spin_unlock(&bo->bdev->lru_lock);
313
314 lret = dma_resv_wait_timeout(resv, true, interruptible,
315 30 * HZ);
316
317 if (lret < 0)
318 return lret;
319 else if (lret == 0)
320 return -EBUSY;
321
322 spin_lock(&bo->bdev->lru_lock);
323 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
324 /*
325 * We raced, and lost, someone else holds the reservation now,
326 * and is probably busy in ttm_bo_cleanup_memtype_use.
327 *
328 * Even if it's not the case, because we finished waiting any
329 * delayed destruction would succeed, so just return success
330 * here.
331 */
332 spin_unlock(&bo->bdev->lru_lock);
333 return 0;
334 }
335 ret = 0;
336 }
337
338 if (ret || unlikely(list_empty(&bo->ddestroy))) {
339 if (unlock_resv)
340 dma_resv_unlock(bo->base.resv);
341 spin_unlock(&bo->bdev->lru_lock);
342 return ret;
343 }
344
345 ttm_bo_del_from_lru(bo);
346 list_del_init(&bo->ddestroy);
347 spin_unlock(&bo->bdev->lru_lock);
348 ttm_bo_cleanup_memtype_use(bo);
349
350 if (unlock_resv)
351 dma_resv_unlock(bo->base.resv);
352
353 ttm_bo_put(bo);
354
355 return 0;
356}
357
358/*
359 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
360 * encountered buffers.
361 */
362bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all)
363{
364 struct list_head removed;
365 bool empty;
366
367 INIT_LIST_HEAD(&removed);
368
369 spin_lock(&bdev->lru_lock);
370 while (!list_empty(&bdev->ddestroy)) {
371 struct ttm_buffer_object *bo;
372
373 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
374 ddestroy);
375 list_move_tail(&bo->ddestroy, &removed);
376 if (!ttm_bo_get_unless_zero(bo))
377 continue;
378
379 if (remove_all || bo->base.resv != &bo->base._resv) {
380 spin_unlock(&bdev->lru_lock);
381 dma_resv_lock(bo->base.resv, NULL);
382
383 spin_lock(&bdev->lru_lock);
384 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
385
386 } else if (dma_resv_trylock(bo->base.resv)) {
387 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
388 } else {
389 spin_unlock(&bdev->lru_lock);
390 }
391
392 ttm_bo_put(bo);
393 spin_lock(&bdev->lru_lock);
394 }
395 list_splice_tail(&removed, &bdev->ddestroy);
396 empty = list_empty(&bdev->ddestroy);
397 spin_unlock(&bdev->lru_lock);
398
399 return empty;
400}
401
402static void ttm_bo_release(struct kref *kref)
403{
404 struct ttm_buffer_object *bo =
405 container_of(kref, struct ttm_buffer_object, kref);
406 struct ttm_device *bdev = bo->bdev;
407 int ret;
408
409 WARN_ON_ONCE(bo->pin_count);
410
411 if (!bo->deleted) {
412 ret = ttm_bo_individualize_resv(bo);
413 if (ret) {
414 /* Last resort, if we fail to allocate memory for the
415 * fences block for the BO to become idle
416 */
417 dma_resv_wait_timeout(bo->base.resv, true, false,
418 30 * HZ);
419 }
420
421 if (bo->bdev->funcs->release_notify)
422 bo->bdev->funcs->release_notify(bo);
423
424 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
425 ttm_mem_io_free(bdev, bo->resource);
426 }
427
428 if (!dma_resv_test_signaled(bo->base.resv, true) ||
429 !dma_resv_trylock(bo->base.resv)) {
430 /* The BO is not idle, resurrect it for delayed destroy */
431 ttm_bo_flush_all_fences(bo);
432 bo->deleted = true;
433
434 spin_lock(&bo->bdev->lru_lock);
435
436 /*
437 * Make pinned bos immediately available to
438 * shrinkers, now that they are queued for
439 * destruction.
440 *
441 * FIXME: QXL is triggering this. Can be removed when the
442 * driver is fixed.
443 */
444 if (bo->pin_count) {
445 bo->pin_count = 0;
446 ttm_bo_move_to_lru_tail(bo, bo->resource, NULL);
447 }
448
449 kref_init(&bo->kref);
450 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
451 spin_unlock(&bo->bdev->lru_lock);
452
453 schedule_delayed_work(&bdev->wq,
454 ((HZ / 100) < 1) ? 1 : HZ / 100);
455 return;
456 }
457
458 spin_lock(&bo->bdev->lru_lock);
459 ttm_bo_del_from_lru(bo);
460 list_del(&bo->ddestroy);
461 spin_unlock(&bo->bdev->lru_lock);
462
463 ttm_bo_cleanup_memtype_use(bo);
464 dma_resv_unlock(bo->base.resv);
465
466 atomic_dec(&ttm_glob.bo_count);
467 dma_fence_put(bo->moving);
468 bo->destroy(bo);
469}
470
471void ttm_bo_put(struct ttm_buffer_object *bo)
472{
473 kref_put(&bo->kref, ttm_bo_release);
474}
475EXPORT_SYMBOL(ttm_bo_put);
476
477int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev)
478{
479 return cancel_delayed_work_sync(&bdev->wq);
480}
481EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
482
483void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched)
484{
485 if (resched)
486 schedule_delayed_work(&bdev->wq,
487 ((HZ / 100) < 1) ? 1 : HZ / 100);
488}
489EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
490
491static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
492 struct ttm_resource **mem,
493 struct ttm_operation_ctx *ctx,
494 struct ttm_place *hop)
495{
496 struct ttm_placement hop_placement;
497 struct ttm_resource *hop_mem;
498 int ret;
499
500 hop_placement.num_placement = hop_placement.num_busy_placement = 1;
501 hop_placement.placement = hop_placement.busy_placement = hop;
502
503 /* find space in the bounce domain */
504 ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
505 if (ret)
506 return ret;
507 /* move to the bounce domain */
508 ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
509 if (ret) {
510 ttm_resource_free(bo, &hop_mem);
511 return ret;
512 }
513 return 0;
514}
515
516static int ttm_bo_evict(struct ttm_buffer_object *bo,
517 struct ttm_operation_ctx *ctx)
518{
519 struct ttm_device *bdev = bo->bdev;
520 struct ttm_resource *evict_mem;
521 struct ttm_placement placement;
522 struct ttm_place hop;
523 int ret = 0;
524
525 memset(&hop, 0, sizeof(hop));
526
527 dma_resv_assert_held(bo->base.resv);
528
529 placement.num_placement = 0;
530 placement.num_busy_placement = 0;
531 bdev->funcs->evict_flags(bo, &placement);
532
533 if (!placement.num_placement && !placement.num_busy_placement) {
534 ret = ttm_bo_wait(bo, true, false);
535 if (ret)
536 return ret;
537
538 /*
539 * Since we've already synced, this frees backing store
540 * immediately.
541 */
542 return ttm_bo_pipeline_gutting(bo);
543 }
544
545 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
546 if (ret) {
547 if (ret != -ERESTARTSYS) {
548 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
549 bo);
550 ttm_bo_mem_space_debug(bo, &placement);
551 }
552 goto out;
553 }
554
555bounce:
556 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
557 if (ret == -EMULTIHOP) {
558 ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
559 if (ret) {
560 pr_err("Buffer eviction failed\n");
561 ttm_resource_free(bo, &evict_mem);
562 goto out;
563 }
564 /* try and move to final place now. */
565 goto bounce;
566 }
567out:
568 return ret;
569}
570
571bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
572 const struct ttm_place *place)
573{
574 dma_resv_assert_held(bo->base.resv);
575 if (bo->resource->mem_type == TTM_PL_SYSTEM)
576 return true;
577
578 /* Don't evict this BO if it's outside of the
579 * requested placement range
580 */
581 if (place->fpfn >= (bo->resource->start + bo->resource->num_pages) ||
582 (place->lpfn && place->lpfn <= bo->resource->start))
583 return false;
584
585 return true;
586}
587EXPORT_SYMBOL(ttm_bo_eviction_valuable);
588
589/*
590 * Check the target bo is allowable to be evicted or swapout, including cases:
591 *
592 * a. if share same reservation object with ctx->resv, have assumption
593 * reservation objects should already be locked, so not lock again and
594 * return true directly when either the opreation allow_reserved_eviction
595 * or the target bo already is in delayed free list;
596 *
597 * b. Otherwise, trylock it.
598 */
599static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
600 struct ttm_operation_ctx *ctx,
601 const struct ttm_place *place,
602 bool *locked, bool *busy)
603{
604 bool ret = false;
605
606 if (bo->base.resv == ctx->resv) {
607 dma_resv_assert_held(bo->base.resv);
608 if (ctx->allow_res_evict)
609 ret = true;
610 *locked = false;
611 if (busy)
612 *busy = false;
613 } else {
614 ret = dma_resv_trylock(bo->base.resv);
615 *locked = ret;
616 if (busy)
617 *busy = !ret;
618 }
619
620 if (ret && place && !bo->bdev->funcs->eviction_valuable(bo, place)) {
621 ret = false;
622 if (*locked) {
623 dma_resv_unlock(bo->base.resv);
624 *locked = false;
625 }
626 }
627
628 return ret;
629}
630
631/**
632 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
633 *
634 * @busy_bo: BO which couldn't be locked with trylock
635 * @ctx: operation context
636 * @ticket: acquire ticket
637 *
638 * Try to lock a busy buffer object to avoid failing eviction.
639 */
640static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
641 struct ttm_operation_ctx *ctx,
642 struct ww_acquire_ctx *ticket)
643{
644 int r;
645
646 if (!busy_bo || !ticket)
647 return -EBUSY;
648
649 if (ctx->interruptible)
650 r = dma_resv_lock_interruptible(busy_bo->base.resv,
651 ticket);
652 else
653 r = dma_resv_lock(busy_bo->base.resv, ticket);
654
655 /*
656 * TODO: It would be better to keep the BO locked until allocation is at
657 * least tried one more time, but that would mean a much larger rework
658 * of TTM.
659 */
660 if (!r)
661 dma_resv_unlock(busy_bo->base.resv);
662
663 return r == -EDEADLK ? -EBUSY : r;
664}
665
666int ttm_mem_evict_first(struct ttm_device *bdev,
667 struct ttm_resource_manager *man,
668 const struct ttm_place *place,
669 struct ttm_operation_ctx *ctx,
670 struct ww_acquire_ctx *ticket)
671{
672 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
673 bool locked = false;
674 unsigned i;
675 int ret;
676
677 spin_lock(&bdev->lru_lock);
678 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
679 list_for_each_entry(bo, &man->lru[i], lru) {
680 bool busy;
681
682 if (!ttm_bo_evict_swapout_allowable(bo, ctx, place,
683 &locked, &busy)) {
684 if (busy && !busy_bo && ticket !=
685 dma_resv_locking_ctx(bo->base.resv))
686 busy_bo = bo;
687 continue;
688 }
689
690 if (!ttm_bo_get_unless_zero(bo)) {
691 if (locked)
692 dma_resv_unlock(bo->base.resv);
693 continue;
694 }
695 break;
696 }
697
698 /* If the inner loop terminated early, we have our candidate */
699 if (&bo->lru != &man->lru[i])
700 break;
701
702 bo = NULL;
703 }
704
705 if (!bo) {
706 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
707 busy_bo = NULL;
708 spin_unlock(&bdev->lru_lock);
709 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
710 if (busy_bo)
711 ttm_bo_put(busy_bo);
712 return ret;
713 }
714
715 if (bo->deleted) {
716 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
717 ctx->no_wait_gpu, locked);
718 ttm_bo_put(bo);
719 return ret;
720 }
721
722 spin_unlock(&bdev->lru_lock);
723
724 ret = ttm_bo_evict(bo, ctx);
725 if (locked)
726 ttm_bo_unreserve(bo);
727
728 ttm_bo_put(bo);
729 return ret;
730}
731
732/*
733 * Add the last move fence to the BO and reserve a new shared slot. We only use
734 * a shared slot to avoid unecessary sync and rely on the subsequent bo move to
735 * either stall or use an exclusive fence respectively set bo->moving.
736 */
737static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
738 struct ttm_resource_manager *man,
739 struct ttm_resource *mem,
740 bool no_wait_gpu)
741{
742 struct dma_fence *fence;
743 int ret;
744
745 spin_lock(&man->move_lock);
746 fence = dma_fence_get(man->move);
747 spin_unlock(&man->move_lock);
748
749 if (!fence)
750 return 0;
751
752 if (no_wait_gpu) {
753 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
754 dma_fence_put(fence);
755 return ret;
756 }
757
758 dma_resv_add_shared_fence(bo->base.resv, fence);
759
760 ret = dma_resv_reserve_shared(bo->base.resv, 1);
761 if (unlikely(ret)) {
762 dma_fence_put(fence);
763 return ret;
764 }
765
766 dma_fence_put(bo->moving);
767 bo->moving = fence;
768 return 0;
769}
770
771/*
772 * Repeatedly evict memory from the LRU for @mem_type until we create enough
773 * space, or we've evicted everything and there isn't enough space.
774 */
775static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
776 const struct ttm_place *place,
777 struct ttm_resource **mem,
778 struct ttm_operation_ctx *ctx)
779{
780 struct ttm_device *bdev = bo->bdev;
781 struct ttm_resource_manager *man;
782 struct ww_acquire_ctx *ticket;
783 int ret;
784
785 man = ttm_manager_type(bdev, place->mem_type);
786 ticket = dma_resv_locking_ctx(bo->base.resv);
787 do {
788 ret = ttm_resource_alloc(bo, place, mem);
789 if (likely(!ret))
790 break;
791 if (unlikely(ret != -ENOSPC))
792 return ret;
793 ret = ttm_mem_evict_first(bdev, man, place, ctx,
794 ticket);
795 if (unlikely(ret != 0))
796 return ret;
797 } while (1);
798
799 return ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
800}
801
802/*
803 * Creates space for memory region @mem according to its type.
804 *
805 * This function first searches for free space in compatible memory types in
806 * the priority order defined by the driver. If free space isn't found, then
807 * ttm_bo_mem_force_space is attempted in priority order to evict and find
808 * space.
809 */
810int ttm_bo_mem_space(struct ttm_buffer_object *bo,
811 struct ttm_placement *placement,
812 struct ttm_resource **mem,
813 struct ttm_operation_ctx *ctx)
814{
815 struct ttm_device *bdev = bo->bdev;
816 bool type_found = false;
817 int i, ret;
818
819 ret = dma_resv_reserve_shared(bo->base.resv, 1);
820 if (unlikely(ret))
821 return ret;
822
823 for (i = 0; i < placement->num_placement; ++i) {
824 const struct ttm_place *place = &placement->placement[i];
825 struct ttm_resource_manager *man;
826
827 man = ttm_manager_type(bdev, place->mem_type);
828 if (!man || !ttm_resource_manager_used(man))
829 continue;
830
831 type_found = true;
832 ret = ttm_resource_alloc(bo, place, mem);
833 if (ret == -ENOSPC)
834 continue;
835 if (unlikely(ret))
836 goto error;
837
838 ret = ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
839 if (unlikely(ret)) {
840 ttm_resource_free(bo, mem);
841 if (ret == -EBUSY)
842 continue;
843
844 goto error;
845 }
846 return 0;
847 }
848
849 for (i = 0; i < placement->num_busy_placement; ++i) {
850 const struct ttm_place *place = &placement->busy_placement[i];
851 struct ttm_resource_manager *man;
852
853 man = ttm_manager_type(bdev, place->mem_type);
854 if (!man || !ttm_resource_manager_used(man))
855 continue;
856
857 type_found = true;
858 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
859 if (likely(!ret))
860 return 0;
861
862 if (ret && ret != -EBUSY)
863 goto error;
864 }
865
866 ret = -ENOMEM;
867 if (!type_found) {
868 pr_err(TTM_PFX "No compatible memory type found\n");
869 ret = -EINVAL;
870 }
871
872error:
873 if (bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
874 ttm_bo_move_to_lru_tail_unlocked(bo);
875
876 return ret;
877}
878EXPORT_SYMBOL(ttm_bo_mem_space);
879
880static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
881 struct ttm_placement *placement,
882 struct ttm_operation_ctx *ctx)
883{
884 struct ttm_resource *mem;
885 struct ttm_place hop;
886 int ret;
887
888 dma_resv_assert_held(bo->base.resv);
889
890 /*
891 * Determine where to move the buffer.
892 *
893 * If driver determines move is going to need
894 * an extra step then it will return -EMULTIHOP
895 * and the buffer will be moved to the temporary
896 * stop and the driver will be called to make
897 * the second hop.
898 */
899 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
900 if (ret)
901 return ret;
902bounce:
903 ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
904 if (ret == -EMULTIHOP) {
905 ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
906 if (ret)
907 goto out;
908 /* try and move to final place now. */
909 goto bounce;
910 }
911out:
912 if (ret)
913 ttm_resource_free(bo, &mem);
914 return ret;
915}
916
917static bool ttm_bo_places_compat(const struct ttm_place *places,
918 unsigned num_placement,
919 struct ttm_resource *mem,
920 uint32_t *new_flags)
921{
922 unsigned i;
923
924 for (i = 0; i < num_placement; i++) {
925 const struct ttm_place *heap = &places[i];
926
927 if ((mem->start < heap->fpfn ||
928 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
929 continue;
930
931 *new_flags = heap->flags;
932 if ((mem->mem_type == heap->mem_type) &&
933 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
934 (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
935 return true;
936 }
937 return false;
938}
939
940bool ttm_bo_mem_compat(struct ttm_placement *placement,
941 struct ttm_resource *mem,
942 uint32_t *new_flags)
943{
944 if (ttm_bo_places_compat(placement->placement, placement->num_placement,
945 mem, new_flags))
946 return true;
947
948 if ((placement->busy_placement != placement->placement ||
949 placement->num_busy_placement > placement->num_placement) &&
950 ttm_bo_places_compat(placement->busy_placement,
951 placement->num_busy_placement,
952 mem, new_flags))
953 return true;
954
955 return false;
956}
957EXPORT_SYMBOL(ttm_bo_mem_compat);
958
959int ttm_bo_validate(struct ttm_buffer_object *bo,
960 struct ttm_placement *placement,
961 struct ttm_operation_ctx *ctx)
962{
963 int ret;
964 uint32_t new_flags;
965
966 dma_resv_assert_held(bo->base.resv);
967
968 /*
969 * Remove the backing store if no placement is given.
970 */
971 if (!placement->num_placement && !placement->num_busy_placement)
972 return ttm_bo_pipeline_gutting(bo);
973
974 /*
975 * Check whether we need to move buffer.
976 */
977 if (!ttm_bo_mem_compat(placement, bo->resource, &new_flags)) {
978 ret = ttm_bo_move_buffer(bo, placement, ctx);
979 if (ret)
980 return ret;
981 }
982 /*
983 * We might need to add a TTM.
984 */
985 if (bo->resource->mem_type == TTM_PL_SYSTEM) {
986 ret = ttm_tt_create(bo, true);
987 if (ret)
988 return ret;
989 }
990 return 0;
991}
992EXPORT_SYMBOL(ttm_bo_validate);
993
994int ttm_bo_init_reserved(struct ttm_device *bdev,
995 struct ttm_buffer_object *bo,
996 size_t size,
997 enum ttm_bo_type type,
998 struct ttm_placement *placement,
999 uint32_t page_alignment,
1000 struct ttm_operation_ctx *ctx,
1001 struct sg_table *sg,
1002 struct dma_resv *resv,
1003 void (*destroy) (struct ttm_buffer_object *))
1004{
1005 static const struct ttm_place sys_mem = { .mem_type = TTM_PL_SYSTEM };
1006 bool locked;
1007 int ret;
1008
1009 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1010
1011 kref_init(&bo->kref);
1012 INIT_LIST_HEAD(&bo->lru);
1013 INIT_LIST_HEAD(&bo->ddestroy);
1014 bo->bdev = bdev;
1015 bo->type = type;
1016 bo->page_alignment = page_alignment;
1017 bo->moving = NULL;
1018 bo->pin_count = 0;
1019 bo->sg = sg;
1020 if (resv) {
1021 bo->base.resv = resv;
1022 dma_resv_assert_held(bo->base.resv);
1023 } else {
1024 bo->base.resv = &bo->base._resv;
1025 }
1026 atomic_inc(&ttm_glob.bo_count);
1027
1028 ret = ttm_resource_alloc(bo, &sys_mem, &bo->resource);
1029 if (unlikely(ret)) {
1030 ttm_bo_put(bo);
1031 return ret;
1032 }
1033
1034 /*
1035 * For ttm_bo_type_device buffers, allocate
1036 * address space from the device.
1037 */
1038 if (bo->type == ttm_bo_type_device ||
1039 bo->type == ttm_bo_type_sg)
1040 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1041 bo->resource->num_pages);
1042
1043 /* passed reservation objects should already be locked,
1044 * since otherwise lockdep will be angered in radeon.
1045 */
1046 if (!resv) {
1047 locked = dma_resv_trylock(bo->base.resv);
1048 WARN_ON(!locked);
1049 }
1050
1051 if (likely(!ret))
1052 ret = ttm_bo_validate(bo, placement, ctx);
1053
1054 if (unlikely(ret)) {
1055 if (!resv)
1056 ttm_bo_unreserve(bo);
1057
1058 ttm_bo_put(bo);
1059 return ret;
1060 }
1061
1062 ttm_bo_move_to_lru_tail_unlocked(bo);
1063
1064 return ret;
1065}
1066EXPORT_SYMBOL(ttm_bo_init_reserved);
1067
1068int ttm_bo_init(struct ttm_device *bdev,
1069 struct ttm_buffer_object *bo,
1070 size_t size,
1071 enum ttm_bo_type type,
1072 struct ttm_placement *placement,
1073 uint32_t page_alignment,
1074 bool interruptible,
1075 struct sg_table *sg,
1076 struct dma_resv *resv,
1077 void (*destroy) (struct ttm_buffer_object *))
1078{
1079 struct ttm_operation_ctx ctx = { interruptible, false };
1080 int ret;
1081
1082 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1083 page_alignment, &ctx, sg, resv, destroy);
1084 if (ret)
1085 return ret;
1086
1087 if (!resv)
1088 ttm_bo_unreserve(bo);
1089
1090 return 0;
1091}
1092EXPORT_SYMBOL(ttm_bo_init);
1093
1094/*
1095 * buffer object vm functions.
1096 */
1097
1098void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1099{
1100 struct ttm_device *bdev = bo->bdev;
1101
1102 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1103 ttm_mem_io_free(bdev, bo->resource);
1104}
1105EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1106
1107int ttm_bo_wait(struct ttm_buffer_object *bo,
1108 bool interruptible, bool no_wait)
1109{
1110 long timeout = 15 * HZ;
1111
1112 if (no_wait) {
1113 if (dma_resv_test_signaled(bo->base.resv, true))
1114 return 0;
1115 else
1116 return -EBUSY;
1117 }
1118
1119 timeout = dma_resv_wait_timeout(bo->base.resv, true, interruptible,
1120 timeout);
1121 if (timeout < 0)
1122 return timeout;
1123
1124 if (timeout == 0)
1125 return -EBUSY;
1126
1127 dma_resv_add_excl_fence(bo->base.resv, NULL);
1128 return 0;
1129}
1130EXPORT_SYMBOL(ttm_bo_wait);
1131
1132int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1133 gfp_t gfp_flags)
1134{
1135 struct ttm_place place;
1136 bool locked;
1137 int ret;
1138
1139 /*
1140 * While the bo may already reside in SYSTEM placement, set
1141 * SYSTEM as new placement to cover also the move further below.
1142 * The driver may use the fact that we're moving from SYSTEM
1143 * as an indication that we're about to swap out.
1144 */
1145 memset(&place, 0, sizeof(place));
1146 place.mem_type = TTM_PL_SYSTEM;
1147 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1148 return -EBUSY;
1149
1150 if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1151 bo->ttm->page_flags & TTM_PAGE_FLAG_SG ||
1152 bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED ||
1153 !ttm_bo_get_unless_zero(bo)) {
1154 if (locked)
1155 dma_resv_unlock(bo->base.resv);
1156 return -EBUSY;
1157 }
1158
1159 if (bo->deleted) {
1160 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1161 ttm_bo_put(bo);
1162 return ret == -EBUSY ? -ENOSPC : ret;
1163 }
1164
1165 ttm_bo_del_from_lru(bo);
1166 /* TODO: Cleanup the locking */
1167 spin_unlock(&bo->bdev->lru_lock);
1168
1169 /*
1170 * Move to system cached
1171 */
1172 if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1173 struct ttm_operation_ctx ctx = { false, false };
1174 struct ttm_resource *evict_mem;
1175 struct ttm_place hop;
1176
1177 memset(&hop, 0, sizeof(hop));
1178 ret = ttm_resource_alloc(bo, &place, &evict_mem);
1179 if (unlikely(ret))
1180 goto out;
1181
1182 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, &ctx, &hop);
1183 if (unlikely(ret != 0)) {
1184 WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1185 goto out;
1186 }
1187 }
1188
1189 /*
1190 * Make sure BO is idle.
1191 */
1192 ret = ttm_bo_wait(bo, false, false);
1193 if (unlikely(ret != 0))
1194 goto out;
1195
1196 ttm_bo_unmap_virtual(bo);
1197
1198 /*
1199 * Swap out. Buffer will be swapped in again as soon as
1200 * anyone tries to access a ttm page.
1201 */
1202 if (bo->bdev->funcs->swap_notify)
1203 bo->bdev->funcs->swap_notify(bo);
1204
1205 if (ttm_tt_is_populated(bo->ttm))
1206 ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1207out:
1208
1209 /*
1210 * Unreserve without putting on LRU to avoid swapping out an
1211 * already swapped buffer.
1212 */
1213 if (locked)
1214 dma_resv_unlock(bo->base.resv);
1215 ttm_bo_put(bo);
1216 return ret == -EBUSY ? -ENOSPC : ret;
1217}
1218
1219void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1220{
1221 if (bo->ttm == NULL)
1222 return;
1223
1224 ttm_tt_destroy(bo->bdev, bo->ttm);
1225 bo->ttm = NULL;
1226}