projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
drm/ttm: call ttm_bo_wait while inside a reservation
[deliverable/linux.git] / drivers / gpu / drm / ttm / ttm_bo.c
1 /**************************************************************************
2 *
3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27 /*
28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29 */
30
31 #define pr_fmt(fmt) "[TTM] " fmt
32
33 #include <drm/ttm/ttm_module.h>
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <linux/jiffies.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 #include <linux/mm.h>
40 #include <linux/file.h>
41 #include <linux/module.h>
42 #include <linux/atomic.h>
43
44 #define TTM_ASSERT_LOCKED(param)
45 #define TTM_DEBUG(fmt, arg...)
46 #define TTM_BO_HASH_ORDER 13
47
48 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
49 static void ttm_bo_global_kobj_release(struct kobject *kobj);
50
51 static struct attribute ttm_bo_count = {
52 .name = "bo_count",
53 .mode = S_IRUGO
54 };
55
56 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
57 uint32_t *mem_type)
58 {
59 int i;
60
61 for (i = 0; i <= TTM_PL_PRIV5; i++)
62 if (place->flags & (1 << i)) {
63 *mem_type = i;
64 return 0;
65 }
66 return -EINVAL;
67 }
68
69 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
70 {
71 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
72
73 pr_err(" has_type: %d\n", man->has_type);
74 pr_err(" use_type: %d\n", man->use_type);
75 pr_err(" flags: 0x%08X\n", man->flags);
76 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset);
77 pr_err(" size: %llu\n", man->size);
78 pr_err(" available_caching: 0x%08X\n", man->available_caching);
79 pr_err(" default_caching: 0x%08X\n", man->default_caching);
80 if (mem_type != TTM_PL_SYSTEM)
81 (*man->func->debug)(man, TTM_PFX);
82 }
83
84 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
85 struct ttm_placement *placement)
86 {
87 int i, ret, mem_type;
88
89 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
90 bo, bo->mem.num_pages, bo->mem.size >> 10,
91 bo->mem.size >> 20);
92 for (i = 0; i < placement->num_placement; i++) {
93 ret = ttm_mem_type_from_place(&placement->placement[i],
94 &mem_type);
95 if (ret)
96 return;
97 pr_err(" placement[%d]=0x%08X (%d)\n",
98 i, placement->placement[i].flags, mem_type);
99 ttm_mem_type_debug(bo->bdev, mem_type);
100 }
101 }
102
103 static ssize_t ttm_bo_global_show(struct kobject *kobj,
104 struct attribute *attr,
105 char *buffer)
106 {
107 struct ttm_bo_global *glob =
108 container_of(kobj, struct ttm_bo_global, kobj);
109
110 return snprintf(buffer, PAGE_SIZE, "%lu\n",
111 (unsigned long) atomic_read(&glob->bo_count));
112 }
113
114 static struct attribute *ttm_bo_global_attrs[] = {
115 &ttm_bo_count,
116 NULL
117 };
118
119 static const struct sysfs_ops ttm_bo_global_ops = {
120 .show = &ttm_bo_global_show
121 };
122
123 static struct kobj_type ttm_bo_glob_kobj_type = {
124 .release = &ttm_bo_global_kobj_release,
125 .sysfs_ops = &ttm_bo_global_ops,
126 .default_attrs = ttm_bo_global_attrs
127 };
128
129
130 static inline uint32_t ttm_bo_type_flags(unsigned type)
131 {
132 return 1 << (type);
133 }
134
135 static void ttm_bo_release_list(struct kref *list_kref)
136 {
137 struct ttm_buffer_object *bo =
138 container_of(list_kref, struct ttm_buffer_object, list_kref);
139 struct ttm_bo_device *bdev = bo->bdev;
140 size_t acc_size = bo->acc_size;
141
142 BUG_ON(atomic_read(&bo->list_kref.refcount));
143 BUG_ON(atomic_read(&bo->kref.refcount));
144 BUG_ON(atomic_read(&bo->cpu_writers));
145 BUG_ON(bo->sync_obj != NULL);
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
164 void 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 != NULL) {
180 list_add_tail(&bo->swap, &bo->glob->swap_lru);
181 kref_get(&bo->list_kref);
182 }
183 }
184 }
185 EXPORT_SYMBOL(ttm_bo_add_to_lru);
186
187 int 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
208 static void ttm_bo_ref_bug(struct kref *list_kref)
209 {
210 BUG();
211 }
212
213 void 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
220 void 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 }
229 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
230
231 /*
232 * Call bo->mutex locked.
233 */
234 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
235 {
236 struct ttm_bo_device *bdev = bo->bdev;
237 struct ttm_bo_global *glob = bo->glob;
238 int ret = 0;
239 uint32_t page_flags = 0;
240
241 TTM_ASSERT_LOCKED(&bo->mutex);
242 bo->ttm = NULL;
243
244 if (bdev->need_dma32)
245 page_flags |= TTM_PAGE_FLAG_DMA32;
246
247 switch (bo->type) {
248 case ttm_bo_type_device:
249 if (zero_alloc)
250 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
251 case ttm_bo_type_kernel:
252 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
253 page_flags, glob->dummy_read_page);
254 if (unlikely(bo->ttm == NULL))
255 ret = -ENOMEM;
256 break;
257 case ttm_bo_type_sg:
258 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
259 page_flags | TTM_PAGE_FLAG_SG,
260 glob->dummy_read_page);
261 if (unlikely(bo->ttm == NULL)) {
262 ret = -ENOMEM;
263 break;
264 }
265 bo->ttm->sg = bo->sg;
266 break;
267 default:
268 pr_err("Illegal buffer object type\n");
269 ret = -EINVAL;
270 break;
271 }
272
273 return ret;
274 }
275
276 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
277 struct ttm_mem_reg *mem,
278 bool evict, bool interruptible,
279 bool no_wait_gpu)
280 {
281 struct ttm_bo_device *bdev = bo->bdev;
282 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
283 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
284 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
285 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
286 int ret = 0;
287
288 if (old_is_pci || new_is_pci ||
289 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
290 ret = ttm_mem_io_lock(old_man, true);
291 if (unlikely(ret != 0))
292 goto out_err;
293 ttm_bo_unmap_virtual_locked(bo);
294 ttm_mem_io_unlock(old_man);
295 }
296
297 /*
298 * Create and bind a ttm if required.
299 */
300
301 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
302 if (bo->ttm == NULL) {
303 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
304 ret = ttm_bo_add_ttm(bo, zero);
305 if (ret)
306 goto out_err;
307 }
308
309 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
310 if (ret)
311 goto out_err;
312
313 if (mem->mem_type != TTM_PL_SYSTEM) {
314 ret = ttm_tt_bind(bo->ttm, mem);
315 if (ret)
316 goto out_err;
317 }
318
319 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
320 if (bdev->driver->move_notify)
321 bdev->driver->move_notify(bo, mem);
322 bo->mem = *mem;
323 mem->mm_node = NULL;
324 goto moved;
325 }
326 }
327
328 if (bdev->driver->move_notify)
329 bdev->driver->move_notify(bo, mem);
330
331 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
332 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
333 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
334 else if (bdev->driver->move)
335 ret = bdev->driver->move(bo, evict, interruptible,
336 no_wait_gpu, mem);
337 else
338 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
339
340 if (ret) {
341 if (bdev->driver->move_notify) {
342 struct ttm_mem_reg tmp_mem = *mem;
343 *mem = bo->mem;
344 bo->mem = tmp_mem;
345 bdev->driver->move_notify(bo, mem);
346 bo->mem = *mem;
347 *mem = tmp_mem;
348 }
349
350 goto out_err;
351 }
352
353 moved:
354 if (bo->evicted) {
355 if (bdev->driver->invalidate_caches) {
356 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
357 if (ret)
358 pr_err("Can not flush read caches\n");
359 }
360 bo->evicted = false;
361 }
362
363 if (bo->mem.mm_node) {
364 bo->offset = (bo->mem.start << PAGE_SHIFT) +
365 bdev->man[bo->mem.mem_type].gpu_offset;
366 bo->cur_placement = bo->mem.placement;
367 } else
368 bo->offset = 0;
369
370 return 0;
371
372 out_err:
373 new_man = &bdev->man[bo->mem.mem_type];
374 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
375 ttm_tt_unbind(bo->ttm);
376 ttm_tt_destroy(bo->ttm);
377 bo->ttm = NULL;
378 }
379
380 return ret;
381 }
382
383 /**
384 * Call bo::reserved.
385 * Will release GPU memory type usage on destruction.
386 * This is the place to put in driver specific hooks to release
387 * driver private resources.
388 * Will release the bo::reserved lock.
389 */
390
391 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
392 {
393 if (bo->bdev->driver->move_notify)
394 bo->bdev->driver->move_notify(bo, NULL);
395
396 if (bo->ttm) {
397 ttm_tt_unbind(bo->ttm);
398 ttm_tt_destroy(bo->ttm);
399 bo->ttm = NULL;
400 }
401 ttm_bo_mem_put(bo, &bo->mem);
402
403 ww_mutex_unlock (&bo->resv->lock);
404 }
405
406 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
407 {
408 struct ttm_bo_device *bdev = bo->bdev;
409 struct ttm_bo_global *glob = bo->glob;
410 struct ttm_bo_driver *driver = bdev->driver;
411 void *sync_obj = NULL;
412 int put_count;
413 int ret;
414
415 spin_lock(&glob->lru_lock);
416 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
417
418 spin_lock(&bdev->fence_lock);
419 (void) ttm_bo_wait(bo, false, false, true);
420 if (!ret && !bo->sync_obj) {
421 spin_unlock(&bdev->fence_lock);
422 put_count = ttm_bo_del_from_lru(bo);
423
424 spin_unlock(&glob->lru_lock);
425 ttm_bo_cleanup_memtype_use(bo);
426
427 ttm_bo_list_ref_sub(bo, put_count, true);
428
429 return;
430 }
431 if (bo->sync_obj)
432 sync_obj = driver->sync_obj_ref(bo->sync_obj);
433 spin_unlock(&bdev->fence_lock);
434
435 if (!ret) {
436
437 /*
438 * Make NO_EVICT bos immediately available to
439 * shrinkers, now that they are queued for
440 * destruction.
441 */
442 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
443 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
444 ttm_bo_add_to_lru(bo);
445 }
446
447 __ttm_bo_unreserve(bo);
448 }
449
450 kref_get(&bo->list_kref);
451 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
452 spin_unlock(&glob->lru_lock);
453
454 if (sync_obj) {
455 driver->sync_obj_flush(sync_obj);
456 driver->sync_obj_unref(&sync_obj);
457 }
458 schedule_delayed_work(&bdev->wq,
459 ((HZ / 100) < 1) ? 1 : HZ / 100);
460 }
461
462 /**
463 * function ttm_bo_cleanup_refs_and_unlock
464 * If bo idle, remove from delayed- and lru lists, and unref.
465 * If not idle, do nothing.
466 *
467 * Must be called with lru_lock and reservation held, this function
468 * will drop both before returning.
469 *
470 * @interruptible Any sleeps should occur interruptibly.
471 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
472 */
473
474 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
475 bool interruptible,
476 bool no_wait_gpu)
477 {
478 struct ttm_bo_device *bdev = bo->bdev;
479 struct ttm_bo_driver *driver = bdev->driver;
480 struct ttm_bo_global *glob = bo->glob;
481 int put_count;
482 int ret;
483
484 spin_lock(&bdev->fence_lock);
485 ret = ttm_bo_wait(bo, false, false, true);
486
487 if (ret && !no_wait_gpu) {
488 void *sync_obj;
489
490 /*
491 * Take a reference to the fence and unreserve,
492 * at this point the buffer should be dead, so
493 * no new sync objects can be attached.
494 */
495 sync_obj = driver->sync_obj_ref(bo->sync_obj);
496 spin_unlock(&bdev->fence_lock);
497
498 __ttm_bo_unreserve(bo);
499 spin_unlock(&glob->lru_lock);
500
501 ret = driver->sync_obj_wait(sync_obj, false, interruptible);
502 driver->sync_obj_unref(&sync_obj);
503 if (ret)
504 return ret;
505
506 spin_lock(&glob->lru_lock);
507 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
508
509 /*
510 * We raced, and lost, someone else holds the reservation now,
511 * and is probably busy in ttm_bo_cleanup_memtype_use.
512 *
513 * Even if it's not the case, because we finished waiting any
514 * delayed destruction would succeed, so just return success
515 * here.
516 */
517 if (ret) {
518 spin_unlock(&glob->lru_lock);
519 return 0;
520 }
521
522 /*
523 * remove sync_obj with ttm_bo_wait, the wait should be
524 * finished, and no new wait object should have been added.
525 */
526 spin_lock(&bdev->fence_lock);
527 ret = ttm_bo_wait(bo, false, false, true);
528 WARN_ON(ret);
529 }
530 spin_unlock(&bdev->fence_lock);
531
532 if (ret || unlikely(list_empty(&bo->ddestroy))) {
533 __ttm_bo_unreserve(bo);
534 spin_unlock(&glob->lru_lock);
535 return ret;
536 }
537
538 put_count = ttm_bo_del_from_lru(bo);
539 list_del_init(&bo->ddestroy);
540 ++put_count;
541
542 spin_unlock(&glob->lru_lock);
543 ttm_bo_cleanup_memtype_use(bo);
544
545 ttm_bo_list_ref_sub(bo, put_count, true);
546
547 return 0;
548 }
549
550 /**
551 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
552 * encountered buffers.
553 */
554
555 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
556 {
557 struct ttm_bo_global *glob = bdev->glob;
558 struct ttm_buffer_object *entry = NULL;
559 int ret = 0;
560
561 spin_lock(&glob->lru_lock);
562 if (list_empty(&bdev->ddestroy))
563 goto out_unlock;
564
565 entry = list_first_entry(&bdev->ddestroy,
566 struct ttm_buffer_object, ddestroy);
567 kref_get(&entry->list_kref);
568
569 for (;;) {
570 struct ttm_buffer_object *nentry = NULL;
571
572 if (entry->ddestroy.next != &bdev->ddestroy) {
573 nentry = list_first_entry(&entry->ddestroy,
574 struct ttm_buffer_object, ddestroy);
575 kref_get(&nentry->list_kref);
576 }
577
578 ret = __ttm_bo_reserve(entry, false, true, false, NULL);
579 if (remove_all && ret) {
580 spin_unlock(&glob->lru_lock);
581 ret = __ttm_bo_reserve(entry, false, false,
582 false, NULL);
583 spin_lock(&glob->lru_lock);
584 }
585
586 if (!ret)
587 ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
588 !remove_all);
589 else
590 spin_unlock(&glob->lru_lock);
591
592 kref_put(&entry->list_kref, ttm_bo_release_list);
593 entry = nentry;
594
595 if (ret || !entry)
596 goto out;
597
598 spin_lock(&glob->lru_lock);
599 if (list_empty(&entry->ddestroy))
600 break;
601 }
602
603 out_unlock:
604 spin_unlock(&glob->lru_lock);
605 out:
606 if (entry)
607 kref_put(&entry->list_kref, ttm_bo_release_list);
608 return ret;
609 }
610
611 static void ttm_bo_delayed_workqueue(struct work_struct *work)
612 {
613 struct ttm_bo_device *bdev =
614 container_of(work, struct ttm_bo_device, wq.work);
615
616 if (ttm_bo_delayed_delete(bdev, false)) {
617 schedule_delayed_work(&bdev->wq,
618 ((HZ / 100) < 1) ? 1 : HZ / 100);
619 }
620 }
621
622 static void ttm_bo_release(struct kref *kref)
623 {
624 struct ttm_buffer_object *bo =
625 container_of(kref, struct ttm_buffer_object, kref);
626 struct ttm_bo_device *bdev = bo->bdev;
627 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
628
629 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
630 ttm_mem_io_lock(man, false);
631 ttm_mem_io_free_vm(bo);
632 ttm_mem_io_unlock(man);
633 ttm_bo_cleanup_refs_or_queue(bo);
634 kref_put(&bo->list_kref, ttm_bo_release_list);
635 }
636
637 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
638 {
639 struct ttm_buffer_object *bo = *p_bo;
640
641 *p_bo = NULL;
642 kref_put(&bo->kref, ttm_bo_release);
643 }
644 EXPORT_SYMBOL(ttm_bo_unref);
645
646 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
647 {
648 return cancel_delayed_work_sync(&bdev->wq);
649 }
650 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
651
652 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
653 {
654 if (resched)
655 schedule_delayed_work(&bdev->wq,
656 ((HZ / 100) < 1) ? 1 : HZ / 100);
657 }
658 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
659
660 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
661 bool no_wait_gpu)
662 {
663 struct ttm_bo_device *bdev = bo->bdev;
664 struct ttm_mem_reg evict_mem;
665 struct ttm_placement placement;
666 int ret = 0;
667
668 spin_lock(&bdev->fence_lock);
669 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
670 spin_unlock(&bdev->fence_lock);
671
672 if (unlikely(ret != 0)) {
673 if (ret != -ERESTARTSYS) {
674 pr_err("Failed to expire sync object before buffer eviction\n");
675 }
676 goto out;
677 }
678
679 lockdep_assert_held(&bo->resv->lock.base);
680
681 evict_mem = bo->mem;
682 evict_mem.mm_node = NULL;
683 evict_mem.bus.io_reserved_vm = false;
684 evict_mem.bus.io_reserved_count = 0;
685
686 placement.num_placement = 0;
687 placement.num_busy_placement = 0;
688 bdev->driver->evict_flags(bo, &placement);
689 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
690 no_wait_gpu);
691 if (ret) {
692 if (ret != -ERESTARTSYS) {
693 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
694 bo);
695 ttm_bo_mem_space_debug(bo, &placement);
696 }
697 goto out;
698 }
699
700 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
701 no_wait_gpu);
702 if (ret) {
703 if (ret != -ERESTARTSYS)
704 pr_err("Buffer eviction failed\n");
705 ttm_bo_mem_put(bo, &evict_mem);
706 goto out;
707 }
708 bo->evicted = true;
709 out:
710 return ret;
711 }
712
713 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
714 uint32_t mem_type,
715 bool interruptible,
716 bool no_wait_gpu)
717 {
718 struct ttm_bo_global *glob = bdev->glob;
719 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
720 struct ttm_buffer_object *bo;
721 int ret = -EBUSY, put_count;
722
723 spin_lock(&glob->lru_lock);
724 list_for_each_entry(bo, &man->lru, lru) {
725 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
726 if (!ret)
727 break;
728 }
729
730 if (ret) {
731 spin_unlock(&glob->lru_lock);
732 return ret;
733 }
734
735 kref_get(&bo->list_kref);
736
737 if (!list_empty(&bo->ddestroy)) {
738 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
739 no_wait_gpu);
740 kref_put(&bo->list_kref, ttm_bo_release_list);
741 return ret;
742 }
743
744 put_count = ttm_bo_del_from_lru(bo);
745 spin_unlock(&glob->lru_lock);
746
747 BUG_ON(ret != 0);
748
749 ttm_bo_list_ref_sub(bo, put_count, true);
750
751 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
752 ttm_bo_unreserve(bo);
753
754 kref_put(&bo->list_kref, ttm_bo_release_list);
755 return ret;
756 }
757
758 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
759 {
760 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
761
762 if (mem->mm_node)
763 (*man->func->put_node)(man, mem);
764 }
765 EXPORT_SYMBOL(ttm_bo_mem_put);
766
767 /**
768 * Repeatedly evict memory from the LRU for @mem_type until we create enough
769 * space, or we've evicted everything and there isn't enough space.
770 */
771 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
772 uint32_t mem_type,
773 const struct ttm_place *place,
774 struct ttm_mem_reg *mem,
775 bool interruptible,
776 bool no_wait_gpu)
777 {
778 struct ttm_bo_device *bdev = bo->bdev;
779 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
780 int ret;
781
782 do {
783 ret = (*man->func->get_node)(man, bo, place, mem);
784 if (unlikely(ret != 0))
785 return ret;
786 if (mem->mm_node)
787 break;
788 ret = ttm_mem_evict_first(bdev, mem_type,
789 interruptible, no_wait_gpu);
790 if (unlikely(ret != 0))
791 return ret;
792 } while (1);
793 if (mem->mm_node == NULL)
794 return -ENOMEM;
795 mem->mem_type = mem_type;
796 return 0;
797 }
798
799 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
800 uint32_t cur_placement,
801 uint32_t proposed_placement)
802 {
803 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
804 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
805
806 /**
807 * Keep current caching if possible.
808 */
809
810 if ((cur_placement & caching) != 0)
811 result |= (cur_placement & caching);
812 else if ((man->default_caching & caching) != 0)
813 result |= man->default_caching;
814 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
815 result |= TTM_PL_FLAG_CACHED;
816 else if ((TTM_PL_FLAG_WC & caching) != 0)
817 result |= TTM_PL_FLAG_WC;
818 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
819 result |= TTM_PL_FLAG_UNCACHED;
820
821 return result;
822 }
823
824 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
825 uint32_t mem_type,
826 const struct ttm_place *place,
827 uint32_t *masked_placement)
828 {
829 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
830
831 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
832 return false;
833
834 if ((place->flags & man->available_caching) == 0)
835 return false;
836
837 cur_flags |= (place->flags & man->available_caching);
838
839 *masked_placement = cur_flags;
840 return true;
841 }
842
843 /**
844 * Creates space for memory region @mem according to its type.
845 *
846 * This function first searches for free space in compatible memory types in
847 * the priority order defined by the driver. If free space isn't found, then
848 * ttm_bo_mem_force_space is attempted in priority order to evict and find
849 * space.
850 */
851 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
852 struct ttm_placement *placement,
853 struct ttm_mem_reg *mem,
854 bool interruptible,
855 bool no_wait_gpu)
856 {
857 struct ttm_bo_device *bdev = bo->bdev;
858 struct ttm_mem_type_manager *man;
859 uint32_t mem_type = TTM_PL_SYSTEM;
860 uint32_t cur_flags = 0;
861 bool type_found = false;
862 bool type_ok = false;
863 bool has_erestartsys = false;
864 int i, ret;
865
866 mem->mm_node = NULL;
867 for (i = 0; i < placement->num_placement; ++i) {
868 const struct ttm_place *place = &placement->placement[i];
869
870 ret = ttm_mem_type_from_place(place, &mem_type);
871 if (ret)
872 return ret;
873 man = &bdev->man[mem_type];
874
875 type_ok = ttm_bo_mt_compatible(man, mem_type, place,
876 &cur_flags);
877
878 if (!type_ok)
879 continue;
880
881 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
882 cur_flags);
883 /*
884 * Use the access and other non-mapping-related flag bits from
885 * the memory placement flags to the current flags
886 */
887 ttm_flag_masked(&cur_flags, place->flags,
888 ~TTM_PL_MASK_MEMTYPE);
889
890 if (mem_type == TTM_PL_SYSTEM)
891 break;
892
893 if (man->has_type && man->use_type) {
894 type_found = true;
895 ret = (*man->func->get_node)(man, bo, place, mem);
896 if (unlikely(ret))
897 return ret;
898 }
899 if (mem->mm_node)
900 break;
901 }
902
903 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
904 mem->mem_type = mem_type;
905 mem->placement = cur_flags;
906 return 0;
907 }
908
909 if (!type_found)
910 return -EINVAL;
911
912 for (i = 0; i < placement->num_busy_placement; ++i) {
913 const struct ttm_place *place = &placement->busy_placement[i];
914
915 ret = ttm_mem_type_from_place(place, &mem_type);
916 if (ret)
917 return ret;
918 man = &bdev->man[mem_type];
919 if (!man->has_type)
920 continue;
921 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
922 continue;
923
924 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
925 cur_flags);
926 /*
927 * Use the access and other non-mapping-related flag bits from
928 * the memory placement flags to the current flags
929 */
930 ttm_flag_masked(&cur_flags, place->flags,
931 ~TTM_PL_MASK_MEMTYPE);
932
933 if (mem_type == TTM_PL_SYSTEM) {
934 mem->mem_type = mem_type;
935 mem->placement = cur_flags;
936 mem->mm_node = NULL;
937 return 0;
938 }
939
940 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
941 interruptible, no_wait_gpu);
942 if (ret == 0 && mem->mm_node) {
943 mem->placement = cur_flags;
944 return 0;
945 }
946 if (ret == -ERESTARTSYS)
947 has_erestartsys = true;
948 }
949 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
950 return ret;
951 }
952 EXPORT_SYMBOL(ttm_bo_mem_space);
953
954 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
955 struct ttm_placement *placement,
956 bool interruptible,
957 bool no_wait_gpu)
958 {
959 int ret = 0;
960 struct ttm_mem_reg mem;
961 struct ttm_bo_device *bdev = bo->bdev;
962
963 lockdep_assert_held(&bo->resv->lock.base);
964
965 /*
966 * FIXME: It's possible to pipeline buffer moves.
967 * Have the driver move function wait for idle when necessary,
968 * instead of doing it here.
969 */
970 spin_lock(&bdev->fence_lock);
971 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
972 spin_unlock(&bdev->fence_lock);
973 if (ret)
974 return ret;
975 mem.num_pages = bo->num_pages;
976 mem.size = mem.num_pages << PAGE_SHIFT;
977 mem.page_alignment = bo->mem.page_alignment;
978 mem.bus.io_reserved_vm = false;
979 mem.bus.io_reserved_count = 0;
980 /*
981 * Determine where to move the buffer.
982 */
983 ret = ttm_bo_mem_space(bo, placement, &mem,
984 interruptible, no_wait_gpu);
985 if (ret)
986 goto out_unlock;
987 ret = ttm_bo_handle_move_mem(bo, &mem, false,
988 interruptible, no_wait_gpu);
989 out_unlock:
990 if (ret && mem.mm_node)
991 ttm_bo_mem_put(bo, &mem);
992 return ret;
993 }
994
995 static bool ttm_bo_mem_compat(struct ttm_placement *placement,
996 struct ttm_mem_reg *mem,
997 uint32_t *new_flags)
998 {
999 int i;
1000
1001 for (i = 0; i < placement->num_placement; i++) {
1002 const struct ttm_place *heap = &placement->placement[i];
1003 if (mem->mm_node && heap->lpfn != 0 &&
1004 (mem->start < heap->fpfn ||
1005 mem->start + mem->num_pages > heap->lpfn))
1006 continue;
1007
1008 *new_flags = heap->flags;
1009 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1010 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1011 return true;
1012 }
1013
1014 for (i = 0; i < placement->num_busy_placement; i++) {
1015 const struct ttm_place *heap = &placement->busy_placement[i];
1016 if (mem->mm_node && heap->lpfn != 0 &&
1017 (mem->start < heap->fpfn ||
1018 mem->start + mem->num_pages > heap->lpfn))
1019 continue;
1020
1021 *new_flags = heap->flags;
1022 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1023 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1024 return true;
1025 }
1026
1027 return false;
1028 }
1029
1030 int ttm_bo_validate(struct ttm_buffer_object *bo,
1031 struct ttm_placement *placement,
1032 bool interruptible,
1033 bool no_wait_gpu)
1034 {
1035 int ret;
1036 uint32_t new_flags;
1037
1038 lockdep_assert_held(&bo->resv->lock.base);
1039 /*
1040 * Check whether we need to move buffer.
1041 */
1042 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1043 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1044 no_wait_gpu);
1045 if (ret)
1046 return ret;
1047 } else {
1048 /*
1049 * Use the access and other non-mapping-related flag bits from
1050 * the compatible memory placement flags to the active flags
1051 */
1052 ttm_flag_masked(&bo->mem.placement, new_flags,
1053 ~TTM_PL_MASK_MEMTYPE);
1054 }
1055 /*
1056 * We might need to add a TTM.
1057 */
1058 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1059 ret = ttm_bo_add_ttm(bo, true);
1060 if (ret)
1061 return ret;
1062 }
1063 return 0;
1064 }
1065 EXPORT_SYMBOL(ttm_bo_validate);
1066
1067 int ttm_bo_init(struct ttm_bo_device *bdev,
1068 struct ttm_buffer_object *bo,
1069 unsigned long size,
1070 enum ttm_bo_type type,
1071 struct ttm_placement *placement,
1072 uint32_t page_alignment,
1073 bool interruptible,
1074 struct file *persistent_swap_storage,
1075 size_t acc_size,
1076 struct sg_table *sg,
1077 void (*destroy) (struct ttm_buffer_object *))
1078 {
1079 int ret = 0;
1080 unsigned long num_pages;
1081 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1082 bool locked;
1083
1084 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1085 if (ret) {
1086 pr_err("Out of kernel memory\n");
1087 if (destroy)
1088 (*destroy)(bo);
1089 else
1090 kfree(bo);
1091 return -ENOMEM;
1092 }
1093
1094 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1095 if (num_pages == 0) {
1096 pr_err("Illegal buffer object size\n");
1097 if (destroy)
1098 (*destroy)(bo);
1099 else
1100 kfree(bo);
1101 ttm_mem_global_free(mem_glob, acc_size);
1102 return -EINVAL;
1103 }
1104 bo->destroy = destroy;
1105
1106 kref_init(&bo->kref);
1107 kref_init(&bo->list_kref);
1108 atomic_set(&bo->cpu_writers, 0);
1109 INIT_LIST_HEAD(&bo->lru);
1110 INIT_LIST_HEAD(&bo->ddestroy);
1111 INIT_LIST_HEAD(&bo->swap);
1112 INIT_LIST_HEAD(&bo->io_reserve_lru);
1113 mutex_init(&bo->wu_mutex);
1114 bo->bdev = bdev;
1115 bo->glob = bdev->glob;
1116 bo->type = type;
1117 bo->num_pages = num_pages;
1118 bo->mem.size = num_pages << PAGE_SHIFT;
1119 bo->mem.mem_type = TTM_PL_SYSTEM;
1120 bo->mem.num_pages = bo->num_pages;
1121 bo->mem.mm_node = NULL;
1122 bo->mem.page_alignment = page_alignment;
1123 bo->mem.bus.io_reserved_vm = false;
1124 bo->mem.bus.io_reserved_count = 0;
1125 bo->priv_flags = 0;
1126 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1127 bo->persistent_swap_storage = persistent_swap_storage;
1128 bo->acc_size = acc_size;
1129 bo->sg = sg;
1130 bo->resv = &bo->ttm_resv;
1131 reservation_object_init(bo->resv);
1132 atomic_inc(&bo->glob->bo_count);
1133 drm_vma_node_reset(&bo->vma_node);
1134
1135 /*
1136 * For ttm_bo_type_device buffers, allocate
1137 * address space from the device.
1138 */
1139 if (bo->type == ttm_bo_type_device ||
1140 bo->type == ttm_bo_type_sg)
1141 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1142 bo->mem.num_pages);
1143
1144 locked = ww_mutex_trylock(&bo->resv->lock);
1145 WARN_ON(!locked);
1146
1147 if (likely(!ret))
1148 ret = ttm_bo_validate(bo, placement, interruptible, false);
1149
1150 ttm_bo_unreserve(bo);
1151
1152 if (unlikely(ret))
1153 ttm_bo_unref(&bo);
1154
1155 return ret;
1156 }
1157 EXPORT_SYMBOL(ttm_bo_init);
1158
1159 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1160 unsigned long bo_size,
1161 unsigned struct_size)
1162 {
1163 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1164 size_t size = 0;
1165
1166 size += ttm_round_pot(struct_size);
1167 size += PAGE_ALIGN(npages * sizeof(void *));
1168 size += ttm_round_pot(sizeof(struct ttm_tt));
1169 return size;
1170 }
1171 EXPORT_SYMBOL(ttm_bo_acc_size);
1172
1173 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1174 unsigned long bo_size,
1175 unsigned struct_size)
1176 {
1177 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1178 size_t size = 0;
1179
1180 size += ttm_round_pot(struct_size);
1181 size += PAGE_ALIGN(npages * sizeof(void *));
1182 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1183 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1184 return size;
1185 }
1186 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1187
1188 int ttm_bo_create(struct ttm_bo_device *bdev,
1189 unsigned long size,
1190 enum ttm_bo_type type,
1191 struct ttm_placement *placement,
1192 uint32_t page_alignment,
1193 bool interruptible,
1194 struct file *persistent_swap_storage,
1195 struct ttm_buffer_object **p_bo)
1196 {
1197 struct ttm_buffer_object *bo;
1198 size_t acc_size;
1199 int ret;
1200
1201 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1202 if (unlikely(bo == NULL))
1203 return -ENOMEM;
1204
1205 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1206 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1207 interruptible, persistent_swap_storage, acc_size,
1208 NULL, NULL);
1209 if (likely(ret == 0))
1210 *p_bo = bo;
1211
1212 return ret;
1213 }
1214 EXPORT_SYMBOL(ttm_bo_create);
1215
1216 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1217 unsigned mem_type, bool allow_errors)
1218 {
1219 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1220 struct ttm_bo_global *glob = bdev->glob;
1221 int ret;
1222
1223 /*
1224 * Can't use standard list traversal since we're unlocking.
1225 */
1226
1227 spin_lock(&glob->lru_lock);
1228 while (!list_empty(&man->lru)) {
1229 spin_unlock(&glob->lru_lock);
1230 ret = ttm_mem_evict_first(bdev, mem_type, false, false);
1231 if (ret) {
1232 if (allow_errors) {
1233 return ret;
1234 } else {
1235 pr_err("Cleanup eviction failed\n");
1236 }
1237 }
1238 spin_lock(&glob->lru_lock);
1239 }
1240 spin_unlock(&glob->lru_lock);
1241 return 0;
1242 }
1243
1244 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1245 {
1246 struct ttm_mem_type_manager *man;
1247 int ret = -EINVAL;
1248
1249 if (mem_type >= TTM_NUM_MEM_TYPES) {
1250 pr_err("Illegal memory type %d\n", mem_type);
1251 return ret;
1252 }
1253 man = &bdev->man[mem_type];
1254
1255 if (!man->has_type) {
1256 pr_err("Trying to take down uninitialized memory manager type %u\n",
1257 mem_type);
1258 return ret;
1259 }
1260
1261 man->use_type = false;
1262 man->has_type = false;
1263
1264 ret = 0;
1265 if (mem_type > 0) {
1266 ttm_bo_force_list_clean(bdev, mem_type, false);
1267
1268 ret = (*man->func->takedown)(man);
1269 }
1270
1271 return ret;
1272 }
1273 EXPORT_SYMBOL(ttm_bo_clean_mm);
1274
1275 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1276 {
1277 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1278
1279 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1280 pr_err("Illegal memory manager memory type %u\n", mem_type);
1281 return -EINVAL;
1282 }
1283
1284 if (!man->has_type) {
1285 pr_err("Memory type %u has not been initialized\n", mem_type);
1286 return 0;
1287 }
1288
1289 return ttm_bo_force_list_clean(bdev, mem_type, true);
1290 }
1291 EXPORT_SYMBOL(ttm_bo_evict_mm);
1292
1293 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1294 unsigned long p_size)
1295 {
1296 int ret = -EINVAL;
1297 struct ttm_mem_type_manager *man;
1298
1299 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1300 man = &bdev->man[type];
1301 BUG_ON(man->has_type);
1302 man->io_reserve_fastpath = true;
1303 man->use_io_reserve_lru = false;
1304 mutex_init(&man->io_reserve_mutex);
1305 INIT_LIST_HEAD(&man->io_reserve_lru);
1306
1307 ret = bdev->driver->init_mem_type(bdev, type, man);
1308 if (ret)
1309 return ret;
1310 man->bdev = bdev;
1311
1312 ret = 0;
1313 if (type != TTM_PL_SYSTEM) {
1314 ret = (*man->func->init)(man, p_size);
1315 if (ret)
1316 return ret;
1317 }
1318 man->has_type = true;
1319 man->use_type = true;
1320 man->size = p_size;
1321
1322 INIT_LIST_HEAD(&man->lru);
1323
1324 return 0;
1325 }
1326 EXPORT_SYMBOL(ttm_bo_init_mm);
1327
1328 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1329 {
1330 struct ttm_bo_global *glob =
1331 container_of(kobj, struct ttm_bo_global, kobj);
1332
1333 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1334 __free_page(glob->dummy_read_page);
1335 kfree(glob);
1336 }
1337
1338 void ttm_bo_global_release(struct drm_global_reference *ref)
1339 {
1340 struct ttm_bo_global *glob = ref->object;
1341
1342 kobject_del(&glob->kobj);
1343 kobject_put(&glob->kobj);
1344 }
1345 EXPORT_SYMBOL(ttm_bo_global_release);
1346
1347 int ttm_bo_global_init(struct drm_global_reference *ref)
1348 {
1349 struct ttm_bo_global_ref *bo_ref =
1350 container_of(ref, struct ttm_bo_global_ref, ref);
1351 struct ttm_bo_global *glob = ref->object;
1352 int ret;
1353
1354 mutex_init(&glob->device_list_mutex);
1355 spin_lock_init(&glob->lru_lock);
1356 glob->mem_glob = bo_ref->mem_glob;
1357 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1358
1359 if (unlikely(glob->dummy_read_page == NULL)) {
1360 ret = -ENOMEM;
1361 goto out_no_drp;
1362 }
1363
1364 INIT_LIST_HEAD(&glob->swap_lru);
1365 INIT_LIST_HEAD(&glob->device_list);
1366
1367 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1368 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1369 if (unlikely(ret != 0)) {
1370 pr_err("Could not register buffer object swapout\n");
1371 goto out_no_shrink;
1372 }
1373
1374 atomic_set(&glob->bo_count, 0);
1375
1376 ret = kobject_init_and_add(
1377 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1378 if (unlikely(ret != 0))
1379 kobject_put(&glob->kobj);
1380 return ret;
1381 out_no_shrink:
1382 __free_page(glob->dummy_read_page);
1383 out_no_drp:
1384 kfree(glob);
1385 return ret;
1386 }
1387 EXPORT_SYMBOL(ttm_bo_global_init);
1388
1389
1390 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1391 {
1392 int ret = 0;
1393 unsigned i = TTM_NUM_MEM_TYPES;
1394 struct ttm_mem_type_manager *man;
1395 struct ttm_bo_global *glob = bdev->glob;
1396
1397 while (i--) {
1398 man = &bdev->man[i];
1399 if (man->has_type) {
1400 man->use_type = false;
1401 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1402 ret = -EBUSY;
1403 pr_err("DRM memory manager type %d is not clean\n",
1404 i);
1405 }
1406 man->has_type = false;
1407 }
1408 }
1409
1410 mutex_lock(&glob->device_list_mutex);
1411 list_del(&bdev->device_list);
1412 mutex_unlock(&glob->device_list_mutex);
1413
1414 cancel_delayed_work_sync(&bdev->wq);
1415
1416 while (ttm_bo_delayed_delete(bdev, true))
1417 ;
1418
1419 spin_lock(&glob->lru_lock);
1420 if (list_empty(&bdev->ddestroy))
1421 TTM_DEBUG("Delayed destroy list was clean\n");
1422
1423 if (list_empty(&bdev->man[0].lru))
1424 TTM_DEBUG("Swap list was clean\n");
1425 spin_unlock(&glob->lru_lock);
1426
1427 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1428
1429 return ret;
1430 }
1431 EXPORT_SYMBOL(ttm_bo_device_release);
1432
1433 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1434 struct ttm_bo_global *glob,
1435 struct ttm_bo_driver *driver,
1436 struct address_space *mapping,
1437 uint64_t file_page_offset,
1438 bool need_dma32)
1439 {
1440 int ret = -EINVAL;
1441
1442 bdev->driver = driver;
1443
1444 memset(bdev->man, 0, sizeof(bdev->man));
1445
1446 /*
1447 * Initialize the system memory buffer type.
1448 * Other types need to be driver / IOCTL initialized.
1449 */
1450 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1451 if (unlikely(ret != 0))
1452 goto out_no_sys;
1453
1454 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1455 0x10000000);
1456 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1457 INIT_LIST_HEAD(&bdev->ddestroy);
1458 bdev->dev_mapping = mapping;
1459 bdev->glob = glob;
1460 bdev->need_dma32 = need_dma32;
1461 bdev->val_seq = 0;
1462 spin_lock_init(&bdev->fence_lock);
1463 mutex_lock(&glob->device_list_mutex);
1464 list_add_tail(&bdev->device_list, &glob->device_list);
1465 mutex_unlock(&glob->device_list_mutex);
1466
1467 return 0;
1468 out_no_sys:
1469 return ret;
1470 }
1471 EXPORT_SYMBOL(ttm_bo_device_init);
1472
1473 /*
1474 * buffer object vm functions.
1475 */
1476
1477 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1478 {
1479 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1480
1481 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1482 if (mem->mem_type == TTM_PL_SYSTEM)
1483 return false;
1484
1485 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1486 return false;
1487
1488 if (mem->placement & TTM_PL_FLAG_CACHED)
1489 return false;
1490 }
1491 return true;
1492 }
1493
1494 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1495 {
1496 struct ttm_bo_device *bdev = bo->bdev;
1497
1498 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1499 ttm_mem_io_free_vm(bo);
1500 }
1501
1502 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1503 {
1504 struct ttm_bo_device *bdev = bo->bdev;
1505 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1506
1507 ttm_mem_io_lock(man, false);
1508 ttm_bo_unmap_virtual_locked(bo);
1509 ttm_mem_io_unlock(man);
1510 }
1511
1512
1513 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1514
1515
1516 int ttm_bo_wait(struct ttm_buffer_object *bo,
1517 bool lazy, bool interruptible, bool no_wait)
1518 {
1519 struct ttm_bo_driver *driver = bo->bdev->driver;
1520 struct ttm_bo_device *bdev = bo->bdev;
1521 void *sync_obj;
1522 int ret = 0;
1523
1524 lockdep_assert_held(&bo->resv->lock.base);
1525
1526 if (likely(bo->sync_obj == NULL))
1527 return 0;
1528
1529 while (bo->sync_obj) {
1530
1531 if (driver->sync_obj_signaled(bo->sync_obj)) {
1532 void *tmp_obj = bo->sync_obj;
1533 bo->sync_obj = NULL;
1534 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1535 spin_unlock(&bdev->fence_lock);
1536 driver->sync_obj_unref(&tmp_obj);
1537 spin_lock(&bdev->fence_lock);
1538 continue;
1539 }
1540
1541 if (no_wait)
1542 return -EBUSY;
1543
1544 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1545 spin_unlock(&bdev->fence_lock);
1546 ret = driver->sync_obj_wait(sync_obj,
1547 lazy, interruptible);
1548 if (unlikely(ret != 0)) {
1549 driver->sync_obj_unref(&sync_obj);
1550 spin_lock(&bdev->fence_lock);
1551 return ret;
1552 }
1553 spin_lock(&bdev->fence_lock);
1554 if (likely(bo->sync_obj == sync_obj)) {
1555 void *tmp_obj = bo->sync_obj;
1556 bo->sync_obj = NULL;
1557 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1558 &bo->priv_flags);
1559 spin_unlock(&bdev->fence_lock);
1560 driver->sync_obj_unref(&sync_obj);
1561 driver->sync_obj_unref(&tmp_obj);
1562 spin_lock(&bdev->fence_lock);
1563 } else {
1564 spin_unlock(&bdev->fence_lock);
1565 driver->sync_obj_unref(&sync_obj);
1566 spin_lock(&bdev->fence_lock);
1567 }
1568 }
1569 return 0;
1570 }
1571 EXPORT_SYMBOL(ttm_bo_wait);
1572
1573 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1574 {
1575 struct ttm_bo_device *bdev = bo->bdev;
1576 int ret = 0;
1577
1578 /*
1579 * Using ttm_bo_reserve makes sure the lru lists are updated.
1580 */
1581
1582 ret = ttm_bo_reserve(bo, true, no_wait, false, NULL);
1583 if (unlikely(ret != 0))
1584 return ret;
1585 spin_lock(&bdev->fence_lock);
1586 ret = ttm_bo_wait(bo, false, true, no_wait);
1587 spin_unlock(&bdev->fence_lock);
1588 if (likely(ret == 0))
1589 atomic_inc(&bo->cpu_writers);
1590 ttm_bo_unreserve(bo);
1591 return ret;
1592 }
1593 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1594
1595 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1596 {
1597 atomic_dec(&bo->cpu_writers);
1598 }
1599 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1600
1601 /**
1602 * A buffer object shrink method that tries to swap out the first
1603 * buffer object on the bo_global::swap_lru list.
1604 */
1605
1606 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1607 {
1608 struct ttm_bo_global *glob =
1609 container_of(shrink, struct ttm_bo_global, shrink);
1610 struct ttm_buffer_object *bo;
1611 int ret = -EBUSY;
1612 int put_count;
1613 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1614
1615 spin_lock(&glob->lru_lock);
1616 list_for_each_entry(bo, &glob->swap_lru, swap) {
1617 ret = __ttm_bo_reserve(bo, false, true, false, NULL);
1618 if (!ret)
1619 break;
1620 }
1621
1622 if (ret) {
1623 spin_unlock(&glob->lru_lock);
1624 return ret;
1625 }
1626
1627 kref_get(&bo->list_kref);
1628
1629 if (!list_empty(&bo->ddestroy)) {
1630 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1631 kref_put(&bo->list_kref, ttm_bo_release_list);
1632 return ret;
1633 }
1634
1635 put_count = ttm_bo_del_from_lru(bo);
1636 spin_unlock(&glob->lru_lock);
1637
1638 ttm_bo_list_ref_sub(bo, put_count, true);
1639
1640 /**
1641 * Wait for GPU, then move to system cached.
1642 */
1643
1644 spin_lock(&bo->bdev->fence_lock);
1645 ret = ttm_bo_wait(bo, false, false, false);
1646 spin_unlock(&bo->bdev->fence_lock);
1647
1648 if (unlikely(ret != 0))
1649 goto out;
1650
1651 if ((bo->mem.placement & swap_placement) != swap_placement) {
1652 struct ttm_mem_reg evict_mem;
1653
1654 evict_mem = bo->mem;
1655 evict_mem.mm_node = NULL;
1656 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1657 evict_mem.mem_type = TTM_PL_SYSTEM;
1658
1659 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1660 false, false);
1661 if (unlikely(ret != 0))
1662 goto out;
1663 }
1664
1665 ttm_bo_unmap_virtual(bo);
1666
1667 /**
1668 * Swap out. Buffer will be swapped in again as soon as
1669 * anyone tries to access a ttm page.
1670 */
1671
1672 if (bo->bdev->driver->swap_notify)
1673 bo->bdev->driver->swap_notify(bo);
1674
1675 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1676 out:
1677
1678 /**
1679 *
1680 * Unreserve without putting on LRU to avoid swapping out an
1681 * already swapped buffer.
1682 */
1683
1684 __ttm_bo_unreserve(bo);
1685 kref_put(&bo->list_kref, ttm_bo_release_list);
1686 return ret;
1687 }
1688
1689 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1690 {
1691 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1692 ;
1693 }
1694 EXPORT_SYMBOL(ttm_bo_swapout_all);
1695
1696 /**
1697 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1698 * unreserved
1699 *
1700 * @bo: Pointer to buffer
1701 */
1702 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1703 {
1704 int ret;
1705
1706 /*
1707 * In the absense of a wait_unlocked API,
1708 * Use the bo::wu_mutex to avoid triggering livelocks due to
1709 * concurrent use of this function. Note that this use of
1710 * bo::wu_mutex can go away if we change locking order to
1711 * mmap_sem -> bo::reserve.
1712 */
1713 ret = mutex_lock_interruptible(&bo->wu_mutex);
1714 if (unlikely(ret != 0))
1715 return -ERESTARTSYS;
1716 if (!ww_mutex_is_locked(&bo->resv->lock))
1717 goto out_unlock;
1718 ret = __ttm_bo_reserve(bo, true, false, false, NULL);
1719 if (unlikely(ret != 0))
1720 goto out_unlock;
1721 __ttm_bo_unreserve(bo);
1722
1723 out_unlock:
1724 mutex_unlock(&bo->wu_mutex);
1725 return ret;
1726 }
Linux kernel - Deliverables
This page took 0.092118 seconds and 5 git commands to generate.