projects / deliverable / linux.git / blob
? search:


f65182667825787979f155dc41c6b71e37a27265
[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_setup_vm(struct ttm_buffer_object *bo);
49 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50 static void ttm_bo_global_kobj_release(struct kobject *kobj);
51
52 static struct attribute ttm_bo_count = {
53 .name = "bo_count",
54 .mode = S_IRUGO
55 };
56
57 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
58 {
59 int i;
60
61 for (i = 0; i <= TTM_PL_PRIV5; i++)
62 if (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_flags(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], 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->destroy)
154 bo->destroy(bo);
155 else {
156 kfree(bo);
157 }
158 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
159 }
160
161 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, bool interruptible)
162 {
163 if (interruptible) {
164 return wait_event_interruptible(bo->event_queue,
165 !ttm_bo_is_reserved(bo));
166 } else {
167 wait_event(bo->event_queue, !ttm_bo_is_reserved(bo));
168 return 0;
169 }
170 }
171 EXPORT_SYMBOL(ttm_bo_wait_unreserved);
172
173 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
174 {
175 struct ttm_bo_device *bdev = bo->bdev;
176 struct ttm_mem_type_manager *man;
177
178 BUG_ON(!ttm_bo_is_reserved(bo));
179
180 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
181
182 BUG_ON(!list_empty(&bo->lru));
183
184 man = &bdev->man[bo->mem.mem_type];
185 list_add_tail(&bo->lru, &man->lru);
186 kref_get(&bo->list_kref);
187
188 if (bo->ttm != NULL) {
189 list_add_tail(&bo->swap, &bo->glob->swap_lru);
190 kref_get(&bo->list_kref);
191 }
192 }
193 }
194
195 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
196 {
197 int put_count = 0;
198
199 if (!list_empty(&bo->swap)) {
200 list_del_init(&bo->swap);
201 ++put_count;
202 }
203 if (!list_empty(&bo->lru)) {
204 list_del_init(&bo->lru);
205 ++put_count;
206 }
207
208 /*
209 * TODO: Add a driver hook to delete from
210 * driver-specific LRU's here.
211 */
212
213 return put_count;
214 }
215
216 int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
217 bool interruptible,
218 bool no_wait, bool use_sequence, uint32_t sequence)
219 {
220 struct ttm_bo_global *glob = bo->glob;
221 int ret;
222
223 while (unlikely(atomic_cmpxchg(&bo->reserved, 0, 1) != 0)) {
224 /**
225 * Deadlock avoidance for multi-bo reserving.
226 */
227 if (use_sequence && bo->seq_valid) {
228 /**
229 * We've already reserved this one.
230 */
231 if (unlikely(sequence == bo->val_seq))
232 return -EDEADLK;
233 /**
234 * Already reserved by a thread that will not back
235 * off for us. We need to back off.
236 */
237 if (unlikely(sequence - bo->val_seq < (1 << 31)))
238 return -EAGAIN;
239 }
240
241 if (no_wait)
242 return -EBUSY;
243
244 spin_unlock(&glob->lru_lock);
245 ret = ttm_bo_wait_unreserved(bo, interruptible);
246 spin_lock(&glob->lru_lock);
247
248 if (unlikely(ret))
249 return ret;
250 }
251
252 if (use_sequence) {
253 /**
254 * Wake up waiters that may need to recheck for deadlock,
255 * if we decreased the sequence number.
256 */
257 if (unlikely((bo->val_seq - sequence < (1 << 31))
258 || !bo->seq_valid))
259 wake_up_all(&bo->event_queue);
260
261 bo->val_seq = sequence;
262 bo->seq_valid = true;
263 } else {
264 bo->seq_valid = false;
265 }
266
267 return 0;
268 }
269 EXPORT_SYMBOL(ttm_bo_reserve);
270
271 static void ttm_bo_ref_bug(struct kref *list_kref)
272 {
273 BUG();
274 }
275
276 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
277 bool never_free)
278 {
279 kref_sub(&bo->list_kref, count,
280 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
281 }
282
283 int ttm_bo_reserve(struct ttm_buffer_object *bo,
284 bool interruptible,
285 bool no_wait, bool use_sequence, uint32_t sequence)
286 {
287 struct ttm_bo_global *glob = bo->glob;
288 int put_count = 0;
289 int ret;
290
291 spin_lock(&glob->lru_lock);
292 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
293 sequence);
294 if (likely(ret == 0))
295 put_count = ttm_bo_del_from_lru(bo);
296 spin_unlock(&glob->lru_lock);
297
298 ttm_bo_list_ref_sub(bo, put_count, true);
299
300 return ret;
301 }
302
303 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo)
304 {
305 ttm_bo_add_to_lru(bo);
306 atomic_set(&bo->reserved, 0);
307 wake_up_all(&bo->event_queue);
308 }
309
310 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
311 {
312 struct ttm_bo_global *glob = bo->glob;
313
314 spin_lock(&glob->lru_lock);
315 ttm_bo_unreserve_locked(bo);
316 spin_unlock(&glob->lru_lock);
317 }
318 EXPORT_SYMBOL(ttm_bo_unreserve);
319
320 /*
321 * Call bo->mutex locked.
322 */
323 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
324 {
325 struct ttm_bo_device *bdev = bo->bdev;
326 struct ttm_bo_global *glob = bo->glob;
327 int ret = 0;
328 uint32_t page_flags = 0;
329
330 TTM_ASSERT_LOCKED(&bo->mutex);
331 bo->ttm = NULL;
332
333 if (bdev->need_dma32)
334 page_flags |= TTM_PAGE_FLAG_DMA32;
335
336 switch (bo->type) {
337 case ttm_bo_type_device:
338 if (zero_alloc)
339 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
340 case ttm_bo_type_kernel:
341 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
342 page_flags, glob->dummy_read_page);
343 if (unlikely(bo->ttm == NULL))
344 ret = -ENOMEM;
345 break;
346 case ttm_bo_type_sg:
347 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
348 page_flags | TTM_PAGE_FLAG_SG,
349 glob->dummy_read_page);
350 if (unlikely(bo->ttm == NULL)) {
351 ret = -ENOMEM;
352 break;
353 }
354 bo->ttm->sg = bo->sg;
355 break;
356 default:
357 pr_err("Illegal buffer object type\n");
358 ret = -EINVAL;
359 break;
360 }
361
362 return ret;
363 }
364
365 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
366 struct ttm_mem_reg *mem,
367 bool evict, bool interruptible,
368 bool no_wait_reserve, bool no_wait_gpu)
369 {
370 struct ttm_bo_device *bdev = bo->bdev;
371 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
372 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
373 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
374 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
375 int ret = 0;
376
377 if (old_is_pci || new_is_pci ||
378 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
379 ret = ttm_mem_io_lock(old_man, true);
380 if (unlikely(ret != 0))
381 goto out_err;
382 ttm_bo_unmap_virtual_locked(bo);
383 ttm_mem_io_unlock(old_man);
384 }
385
386 /*
387 * Create and bind a ttm if required.
388 */
389
390 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
391 if (bo->ttm == NULL) {
392 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
393 ret = ttm_bo_add_ttm(bo, zero);
394 if (ret)
395 goto out_err;
396 }
397
398 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
399 if (ret)
400 goto out_err;
401
402 if (mem->mem_type != TTM_PL_SYSTEM) {
403 ret = ttm_tt_bind(bo->ttm, mem);
404 if (ret)
405 goto out_err;
406 }
407
408 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
409 if (bdev->driver->move_notify)
410 bdev->driver->move_notify(bo, mem);
411 bo->mem = *mem;
412 mem->mm_node = NULL;
413 goto moved;
414 }
415 }
416
417 if (bdev->driver->move_notify)
418 bdev->driver->move_notify(bo, mem);
419
420 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
421 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
422 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
423 else if (bdev->driver->move)
424 ret = bdev->driver->move(bo, evict, interruptible,
425 no_wait_reserve, no_wait_gpu, mem);
426 else
427 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
428
429 if (ret) {
430 if (bdev->driver->move_notify) {
431 struct ttm_mem_reg tmp_mem = *mem;
432 *mem = bo->mem;
433 bo->mem = tmp_mem;
434 bdev->driver->move_notify(bo, mem);
435 bo->mem = *mem;
436 }
437
438 goto out_err;
439 }
440
441 moved:
442 if (bo->evicted) {
443 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
444 if (ret)
445 pr_err("Can not flush read caches\n");
446 bo->evicted = false;
447 }
448
449 if (bo->mem.mm_node) {
450 bo->offset = (bo->mem.start << PAGE_SHIFT) +
451 bdev->man[bo->mem.mem_type].gpu_offset;
452 bo->cur_placement = bo->mem.placement;
453 } else
454 bo->offset = 0;
455
456 return 0;
457
458 out_err:
459 new_man = &bdev->man[bo->mem.mem_type];
460 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
461 ttm_tt_unbind(bo->ttm);
462 ttm_tt_destroy(bo->ttm);
463 bo->ttm = NULL;
464 }
465
466 return ret;
467 }
468
469 /**
470 * Call bo::reserved.
471 * Will release GPU memory type usage on destruction.
472 * This is the place to put in driver specific hooks to release
473 * driver private resources.
474 * Will release the bo::reserved lock.
475 */
476
477 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
478 {
479 if (bo->bdev->driver->move_notify)
480 bo->bdev->driver->move_notify(bo, NULL);
481
482 if (bo->ttm) {
483 ttm_tt_unbind(bo->ttm);
484 ttm_tt_destroy(bo->ttm);
485 bo->ttm = NULL;
486 }
487 ttm_bo_mem_put(bo, &bo->mem);
488
489 atomic_set(&bo->reserved, 0);
490
491 /*
492 * Make processes trying to reserve really pick it up.
493 */
494 smp_mb__after_atomic_dec();
495 wake_up_all(&bo->event_queue);
496 }
497
498 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
499 {
500 struct ttm_bo_device *bdev = bo->bdev;
501 struct ttm_bo_global *glob = bo->glob;
502 struct ttm_bo_driver *driver;
503 void *sync_obj = NULL;
504 void *sync_obj_arg;
505 int put_count;
506 int ret;
507
508 spin_lock(&bdev->fence_lock);
509 (void) ttm_bo_wait(bo, false, false, true);
510 if (!bo->sync_obj) {
511
512 spin_lock(&glob->lru_lock);
513
514 /**
515 * Lock inversion between bo:reserve and bdev::fence_lock here,
516 * but that's OK, since we're only trylocking.
517 */
518
519 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
520
521 if (unlikely(ret == -EBUSY))
522 goto queue;
523
524 spin_unlock(&bdev->fence_lock);
525 put_count = ttm_bo_del_from_lru(bo);
526
527 spin_unlock(&glob->lru_lock);
528 ttm_bo_cleanup_memtype_use(bo);
529
530 ttm_bo_list_ref_sub(bo, put_count, true);
531
532 return;
533 } else {
534 spin_lock(&glob->lru_lock);
535 }
536 queue:
537 driver = bdev->driver;
538 if (bo->sync_obj)
539 sync_obj = driver->sync_obj_ref(bo->sync_obj);
540 sync_obj_arg = bo->sync_obj_arg;
541
542 kref_get(&bo->list_kref);
543 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
544 spin_unlock(&glob->lru_lock);
545 spin_unlock(&bdev->fence_lock);
546
547 if (sync_obj) {
548 driver->sync_obj_flush(sync_obj, sync_obj_arg);
549 driver->sync_obj_unref(&sync_obj);
550 }
551 schedule_delayed_work(&bdev->wq,
552 ((HZ / 100) < 1) ? 1 : HZ / 100);
553 }
554
555 /**
556 * function ttm_bo_cleanup_refs
557 * If bo idle, remove from delayed- and lru lists, and unref.
558 * If not idle, do nothing.
559 *
560 * @interruptible Any sleeps should occur interruptibly.
561 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
562 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
563 */
564
565 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
566 bool interruptible,
567 bool no_wait_reserve,
568 bool no_wait_gpu)
569 {
570 struct ttm_bo_device *bdev = bo->bdev;
571 struct ttm_bo_global *glob = bo->glob;
572 int put_count;
573 int ret = 0;
574
575 retry:
576 spin_lock(&bdev->fence_lock);
577 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
578 spin_unlock(&bdev->fence_lock);
579
580 if (unlikely(ret != 0))
581 return ret;
582
583 retry_reserve:
584 spin_lock(&glob->lru_lock);
585
586 if (unlikely(list_empty(&bo->ddestroy))) {
587 spin_unlock(&glob->lru_lock);
588 return 0;
589 }
590
591 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
592
593 if (unlikely(ret == -EBUSY)) {
594 spin_unlock(&glob->lru_lock);
595 if (likely(!no_wait_reserve))
596 ret = ttm_bo_wait_unreserved(bo, interruptible);
597 if (unlikely(ret != 0))
598 return ret;
599
600 goto retry_reserve;
601 }
602
603 BUG_ON(ret != 0);
604
605 /**
606 * We can re-check for sync object without taking
607 * the bo::lock since setting the sync object requires
608 * also bo::reserved. A busy object at this point may
609 * be caused by another thread recently starting an accelerated
610 * eviction.
611 */
612
613 if (unlikely(bo->sync_obj)) {
614 atomic_set(&bo->reserved, 0);
615 wake_up_all(&bo->event_queue);
616 spin_unlock(&glob->lru_lock);
617 goto retry;
618 }
619
620 put_count = ttm_bo_del_from_lru(bo);
621 list_del_init(&bo->ddestroy);
622 ++put_count;
623
624 spin_unlock(&glob->lru_lock);
625 ttm_bo_cleanup_memtype_use(bo);
626
627 ttm_bo_list_ref_sub(bo, put_count, true);
628
629 return 0;
630 }
631
632 /**
633 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
634 * encountered buffers.
635 */
636
637 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
638 {
639 struct ttm_bo_global *glob = bdev->glob;
640 struct ttm_buffer_object *entry = NULL;
641 int ret = 0;
642
643 spin_lock(&glob->lru_lock);
644 if (list_empty(&bdev->ddestroy))
645 goto out_unlock;
646
647 entry = list_first_entry(&bdev->ddestroy,
648 struct ttm_buffer_object, ddestroy);
649 kref_get(&entry->list_kref);
650
651 for (;;) {
652 struct ttm_buffer_object *nentry = NULL;
653
654 if (entry->ddestroy.next != &bdev->ddestroy) {
655 nentry = list_first_entry(&entry->ddestroy,
656 struct ttm_buffer_object, ddestroy);
657 kref_get(&nentry->list_kref);
658 }
659
660 spin_unlock(&glob->lru_lock);
661 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
662 !remove_all);
663 kref_put(&entry->list_kref, ttm_bo_release_list);
664 entry = nentry;
665
666 if (ret || !entry)
667 goto out;
668
669 spin_lock(&glob->lru_lock);
670 if (list_empty(&entry->ddestroy))
671 break;
672 }
673
674 out_unlock:
675 spin_unlock(&glob->lru_lock);
676 out:
677 if (entry)
678 kref_put(&entry->list_kref, ttm_bo_release_list);
679 return ret;
680 }
681
682 static void ttm_bo_delayed_workqueue(struct work_struct *work)
683 {
684 struct ttm_bo_device *bdev =
685 container_of(work, struct ttm_bo_device, wq.work);
686
687 if (ttm_bo_delayed_delete(bdev, false)) {
688 schedule_delayed_work(&bdev->wq,
689 ((HZ / 100) < 1) ? 1 : HZ / 100);
690 }
691 }
692
693 static void ttm_bo_release(struct kref *kref)
694 {
695 struct ttm_buffer_object *bo =
696 container_of(kref, struct ttm_buffer_object, kref);
697 struct ttm_bo_device *bdev = bo->bdev;
698 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
699
700 if (likely(bo->vm_node != NULL)) {
701 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
702 drm_mm_put_block(bo->vm_node);
703 bo->vm_node = NULL;
704 }
705 write_unlock(&bdev->vm_lock);
706 ttm_mem_io_lock(man, false);
707 ttm_mem_io_free_vm(bo);
708 ttm_mem_io_unlock(man);
709 ttm_bo_cleanup_refs_or_queue(bo);
710 kref_put(&bo->list_kref, ttm_bo_release_list);
711 write_lock(&bdev->vm_lock);
712 }
713
714 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
715 {
716 struct ttm_buffer_object *bo = *p_bo;
717 struct ttm_bo_device *bdev = bo->bdev;
718
719 *p_bo = NULL;
720 write_lock(&bdev->vm_lock);
721 kref_put(&bo->kref, ttm_bo_release);
722 write_unlock(&bdev->vm_lock);
723 }
724 EXPORT_SYMBOL(ttm_bo_unref);
725
726 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
727 {
728 return cancel_delayed_work_sync(&bdev->wq);
729 }
730 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
731
732 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
733 {
734 if (resched)
735 schedule_delayed_work(&bdev->wq,
736 ((HZ / 100) < 1) ? 1 : HZ / 100);
737 }
738 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
739
740 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
741 bool no_wait_reserve, bool no_wait_gpu)
742 {
743 struct ttm_bo_device *bdev = bo->bdev;
744 struct ttm_mem_reg evict_mem;
745 struct ttm_placement placement;
746 int ret = 0;
747
748 spin_lock(&bdev->fence_lock);
749 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
750 spin_unlock(&bdev->fence_lock);
751
752 if (unlikely(ret != 0)) {
753 if (ret != -ERESTARTSYS) {
754 pr_err("Failed to expire sync object before buffer eviction\n");
755 }
756 goto out;
757 }
758
759 BUG_ON(!ttm_bo_is_reserved(bo));
760
761 evict_mem = bo->mem;
762 evict_mem.mm_node = NULL;
763 evict_mem.bus.io_reserved_vm = false;
764 evict_mem.bus.io_reserved_count = 0;
765
766 placement.fpfn = 0;
767 placement.lpfn = 0;
768 placement.num_placement = 0;
769 placement.num_busy_placement = 0;
770 bdev->driver->evict_flags(bo, &placement);
771 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
772 no_wait_reserve, no_wait_gpu);
773 if (ret) {
774 if (ret != -ERESTARTSYS) {
775 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
776 bo);
777 ttm_bo_mem_space_debug(bo, &placement);
778 }
779 goto out;
780 }
781
782 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
783 no_wait_reserve, no_wait_gpu);
784 if (ret) {
785 if (ret != -ERESTARTSYS)
786 pr_err("Buffer eviction failed\n");
787 ttm_bo_mem_put(bo, &evict_mem);
788 goto out;
789 }
790 bo->evicted = true;
791 out:
792 return ret;
793 }
794
795 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
796 uint32_t mem_type,
797 bool interruptible, bool no_wait_reserve,
798 bool no_wait_gpu)
799 {
800 struct ttm_bo_global *glob = bdev->glob;
801 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
802 struct ttm_buffer_object *bo;
803 int ret, put_count = 0;
804
805 retry:
806 spin_lock(&glob->lru_lock);
807 if (list_empty(&man->lru)) {
808 spin_unlock(&glob->lru_lock);
809 return -EBUSY;
810 }
811
812 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
813 kref_get(&bo->list_kref);
814
815 if (!list_empty(&bo->ddestroy)) {
816 spin_unlock(&glob->lru_lock);
817 ret = ttm_bo_cleanup_refs(bo, interruptible,
818 no_wait_reserve, no_wait_gpu);
819 kref_put(&bo->list_kref, ttm_bo_release_list);
820
821 return ret;
822 }
823
824 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
825
826 if (unlikely(ret == -EBUSY)) {
827 spin_unlock(&glob->lru_lock);
828 if (likely(!no_wait_reserve))
829 ret = ttm_bo_wait_unreserved(bo, interruptible);
830
831 kref_put(&bo->list_kref, ttm_bo_release_list);
832
833 /**
834 * We *need* to retry after releasing the lru lock.
835 */
836
837 if (unlikely(ret != 0))
838 return ret;
839 goto retry;
840 }
841
842 put_count = ttm_bo_del_from_lru(bo);
843 spin_unlock(&glob->lru_lock);
844
845 BUG_ON(ret != 0);
846
847 ttm_bo_list_ref_sub(bo, put_count, true);
848
849 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
850 ttm_bo_unreserve(bo);
851
852 kref_put(&bo->list_kref, ttm_bo_release_list);
853 return ret;
854 }
855
856 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
857 {
858 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
859
860 if (mem->mm_node)
861 (*man->func->put_node)(man, mem);
862 }
863 EXPORT_SYMBOL(ttm_bo_mem_put);
864
865 /**
866 * Repeatedly evict memory from the LRU for @mem_type until we create enough
867 * space, or we've evicted everything and there isn't enough space.
868 */
869 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
870 uint32_t mem_type,
871 struct ttm_placement *placement,
872 struct ttm_mem_reg *mem,
873 bool interruptible,
874 bool no_wait_reserve,
875 bool no_wait_gpu)
876 {
877 struct ttm_bo_device *bdev = bo->bdev;
878 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
879 int ret;
880
881 do {
882 ret = (*man->func->get_node)(man, bo, placement, mem);
883 if (unlikely(ret != 0))
884 return ret;
885 if (mem->mm_node)
886 break;
887 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
888 no_wait_reserve, no_wait_gpu);
889 if (unlikely(ret != 0))
890 return ret;
891 } while (1);
892 if (mem->mm_node == NULL)
893 return -ENOMEM;
894 mem->mem_type = mem_type;
895 return 0;
896 }
897
898 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
899 uint32_t cur_placement,
900 uint32_t proposed_placement)
901 {
902 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
903 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
904
905 /**
906 * Keep current caching if possible.
907 */
908
909 if ((cur_placement & caching) != 0)
910 result |= (cur_placement & caching);
911 else if ((man->default_caching & caching) != 0)
912 result |= man->default_caching;
913 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
914 result |= TTM_PL_FLAG_CACHED;
915 else if ((TTM_PL_FLAG_WC & caching) != 0)
916 result |= TTM_PL_FLAG_WC;
917 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
918 result |= TTM_PL_FLAG_UNCACHED;
919
920 return result;
921 }
922
923 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
924 uint32_t mem_type,
925 uint32_t proposed_placement,
926 uint32_t *masked_placement)
927 {
928 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
929
930 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
931 return false;
932
933 if ((proposed_placement & man->available_caching) == 0)
934 return false;
935
936 cur_flags |= (proposed_placement & man->available_caching);
937
938 *masked_placement = cur_flags;
939 return true;
940 }
941
942 /**
943 * Creates space for memory region @mem according to its type.
944 *
945 * This function first searches for free space in compatible memory types in
946 * the priority order defined by the driver. If free space isn't found, then
947 * ttm_bo_mem_force_space is attempted in priority order to evict and find
948 * space.
949 */
950 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
951 struct ttm_placement *placement,
952 struct ttm_mem_reg *mem,
953 bool interruptible, bool no_wait_reserve,
954 bool no_wait_gpu)
955 {
956 struct ttm_bo_device *bdev = bo->bdev;
957 struct ttm_mem_type_manager *man;
958 uint32_t mem_type = TTM_PL_SYSTEM;
959 uint32_t cur_flags = 0;
960 bool type_found = false;
961 bool type_ok = false;
962 bool has_erestartsys = false;
963 int i, ret;
964
965 mem->mm_node = NULL;
966 for (i = 0; i < placement->num_placement; ++i) {
967 ret = ttm_mem_type_from_flags(placement->placement[i],
968 &mem_type);
969 if (ret)
970 return ret;
971 man = &bdev->man[mem_type];
972
973 type_ok = ttm_bo_mt_compatible(man,
974 mem_type,
975 placement->placement[i],
976 &cur_flags);
977
978 if (!type_ok)
979 continue;
980
981 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
982 cur_flags);
983 /*
984 * Use the access and other non-mapping-related flag bits from
985 * the memory placement flags to the current flags
986 */
987 ttm_flag_masked(&cur_flags, placement->placement[i],
988 ~TTM_PL_MASK_MEMTYPE);
989
990 if (mem_type == TTM_PL_SYSTEM)
991 break;
992
993 if (man->has_type && man->use_type) {
994 type_found = true;
995 ret = (*man->func->get_node)(man, bo, placement, mem);
996 if (unlikely(ret))
997 return ret;
998 }
999 if (mem->mm_node)
1000 break;
1001 }
1002
1003 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
1004 mem->mem_type = mem_type;
1005 mem->placement = cur_flags;
1006 return 0;
1007 }
1008
1009 if (!type_found)
1010 return -EINVAL;
1011
1012 for (i = 0; i < placement->num_busy_placement; ++i) {
1013 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
1014 &mem_type);
1015 if (ret)
1016 return ret;
1017 man = &bdev->man[mem_type];
1018 if (!man->has_type)
1019 continue;
1020 if (!ttm_bo_mt_compatible(man,
1021 mem_type,
1022 placement->busy_placement[i],
1023 &cur_flags))
1024 continue;
1025
1026 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1027 cur_flags);
1028 /*
1029 * Use the access and other non-mapping-related flag bits from
1030 * the memory placement flags to the current flags
1031 */
1032 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1033 ~TTM_PL_MASK_MEMTYPE);
1034
1035
1036 if (mem_type == TTM_PL_SYSTEM) {
1037 mem->mem_type = mem_type;
1038 mem->placement = cur_flags;
1039 mem->mm_node = NULL;
1040 return 0;
1041 }
1042
1043 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1044 interruptible, no_wait_reserve, no_wait_gpu);
1045 if (ret == 0 && mem->mm_node) {
1046 mem->placement = cur_flags;
1047 return 0;
1048 }
1049 if (ret == -ERESTARTSYS)
1050 has_erestartsys = true;
1051 }
1052 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1053 return ret;
1054 }
1055 EXPORT_SYMBOL(ttm_bo_mem_space);
1056
1057 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1058 {
1059 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1060 return -EBUSY;
1061
1062 return wait_event_interruptible(bo->event_queue,
1063 atomic_read(&bo->cpu_writers) == 0);
1064 }
1065 EXPORT_SYMBOL(ttm_bo_wait_cpu);
1066
1067 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1068 struct ttm_placement *placement,
1069 bool interruptible, bool no_wait_reserve,
1070 bool no_wait_gpu)
1071 {
1072 int ret = 0;
1073 struct ttm_mem_reg mem;
1074 struct ttm_bo_device *bdev = bo->bdev;
1075
1076 BUG_ON(!ttm_bo_is_reserved(bo));
1077
1078 /*
1079 * FIXME: It's possible to pipeline buffer moves.
1080 * Have the driver move function wait for idle when necessary,
1081 * instead of doing it here.
1082 */
1083 spin_lock(&bdev->fence_lock);
1084 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1085 spin_unlock(&bdev->fence_lock);
1086 if (ret)
1087 return ret;
1088 mem.num_pages = bo->num_pages;
1089 mem.size = mem.num_pages << PAGE_SHIFT;
1090 mem.page_alignment = bo->mem.page_alignment;
1091 mem.bus.io_reserved_vm = false;
1092 mem.bus.io_reserved_count = 0;
1093 /*
1094 * Determine where to move the buffer.
1095 */
1096 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1097 if (ret)
1098 goto out_unlock;
1099 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1100 out_unlock:
1101 if (ret && mem.mm_node)
1102 ttm_bo_mem_put(bo, &mem);
1103 return ret;
1104 }
1105
1106 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1107 struct ttm_mem_reg *mem)
1108 {
1109 int i;
1110
1111 if (mem->mm_node && placement->lpfn != 0 &&
1112 (mem->start < placement->fpfn ||
1113 mem->start + mem->num_pages > placement->lpfn))
1114 return -1;
1115
1116 for (i = 0; i < placement->num_placement; i++) {
1117 if ((placement->placement[i] & mem->placement &
1118 TTM_PL_MASK_CACHING) &&
1119 (placement->placement[i] & mem->placement &
1120 TTM_PL_MASK_MEM))
1121 return i;
1122 }
1123 return -1;
1124 }
1125
1126 int ttm_bo_validate(struct ttm_buffer_object *bo,
1127 struct ttm_placement *placement,
1128 bool interruptible, bool no_wait_reserve,
1129 bool no_wait_gpu)
1130 {
1131 int ret;
1132
1133 BUG_ON(!ttm_bo_is_reserved(bo));
1134 /* Check that range is valid */
1135 if (placement->lpfn || placement->fpfn)
1136 if (placement->fpfn > placement->lpfn ||
1137 (placement->lpfn - placement->fpfn) < bo->num_pages)
1138 return -EINVAL;
1139 /*
1140 * Check whether we need to move buffer.
1141 */
1142 ret = ttm_bo_mem_compat(placement, &bo->mem);
1143 if (ret < 0) {
1144 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1145 if (ret)
1146 return ret;
1147 } else {
1148 /*
1149 * Use the access and other non-mapping-related flag bits from
1150 * the compatible memory placement flags to the active flags
1151 */
1152 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1153 ~TTM_PL_MASK_MEMTYPE);
1154 }
1155 /*
1156 * We might need to add a TTM.
1157 */
1158 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1159 ret = ttm_bo_add_ttm(bo, true);
1160 if (ret)
1161 return ret;
1162 }
1163 return 0;
1164 }
1165 EXPORT_SYMBOL(ttm_bo_validate);
1166
1167 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1168 struct ttm_placement *placement)
1169 {
1170 BUG_ON((placement->fpfn || placement->lpfn) &&
1171 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1172
1173 return 0;
1174 }
1175
1176 int ttm_bo_init(struct ttm_bo_device *bdev,
1177 struct ttm_buffer_object *bo,
1178 unsigned long size,
1179 enum ttm_bo_type type,
1180 struct ttm_placement *placement,
1181 uint32_t page_alignment,
1182 bool interruptible,
1183 struct file *persistent_swap_storage,
1184 size_t acc_size,
1185 struct sg_table *sg,
1186 void (*destroy) (struct ttm_buffer_object *))
1187 {
1188 int ret = 0;
1189 unsigned long num_pages;
1190 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1191
1192 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1193 if (ret) {
1194 pr_err("Out of kernel memory\n");
1195 if (destroy)
1196 (*destroy)(bo);
1197 else
1198 kfree(bo);
1199 return -ENOMEM;
1200 }
1201
1202 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1203 if (num_pages == 0) {
1204 pr_err("Illegal buffer object size\n");
1205 if (destroy)
1206 (*destroy)(bo);
1207 else
1208 kfree(bo);
1209 ttm_mem_global_free(mem_glob, acc_size);
1210 return -EINVAL;
1211 }
1212 bo->destroy = destroy;
1213
1214 kref_init(&bo->kref);
1215 kref_init(&bo->list_kref);
1216 atomic_set(&bo->cpu_writers, 0);
1217 atomic_set(&bo->reserved, 1);
1218 init_waitqueue_head(&bo->event_queue);
1219 INIT_LIST_HEAD(&bo->lru);
1220 INIT_LIST_HEAD(&bo->ddestroy);
1221 INIT_LIST_HEAD(&bo->swap);
1222 INIT_LIST_HEAD(&bo->io_reserve_lru);
1223 bo->bdev = bdev;
1224 bo->glob = bdev->glob;
1225 bo->type = type;
1226 bo->num_pages = num_pages;
1227 bo->mem.size = num_pages << PAGE_SHIFT;
1228 bo->mem.mem_type = TTM_PL_SYSTEM;
1229 bo->mem.num_pages = bo->num_pages;
1230 bo->mem.mm_node = NULL;
1231 bo->mem.page_alignment = page_alignment;
1232 bo->mem.bus.io_reserved_vm = false;
1233 bo->mem.bus.io_reserved_count = 0;
1234 bo->priv_flags = 0;
1235 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1236 bo->seq_valid = false;
1237 bo->persistent_swap_storage = persistent_swap_storage;
1238 bo->acc_size = acc_size;
1239 bo->sg = sg;
1240 atomic_inc(&bo->glob->bo_count);
1241
1242 ret = ttm_bo_check_placement(bo, placement);
1243 if (unlikely(ret != 0))
1244 goto out_err;
1245
1246 /*
1247 * For ttm_bo_type_device buffers, allocate
1248 * address space from the device.
1249 */
1250 if (bo->type == ttm_bo_type_device ||
1251 bo->type == ttm_bo_type_sg) {
1252 ret = ttm_bo_setup_vm(bo);
1253 if (ret)
1254 goto out_err;
1255 }
1256
1257 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1258 if (ret)
1259 goto out_err;
1260
1261 ttm_bo_unreserve(bo);
1262 return 0;
1263
1264 out_err:
1265 ttm_bo_unreserve(bo);
1266 ttm_bo_unref(&bo);
1267
1268 return ret;
1269 }
1270 EXPORT_SYMBOL(ttm_bo_init);
1271
1272 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1273 unsigned long bo_size,
1274 unsigned struct_size)
1275 {
1276 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1277 size_t size = 0;
1278
1279 size += ttm_round_pot(struct_size);
1280 size += PAGE_ALIGN(npages * sizeof(void *));
1281 size += ttm_round_pot(sizeof(struct ttm_tt));
1282 return size;
1283 }
1284 EXPORT_SYMBOL(ttm_bo_acc_size);
1285
1286 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1287 unsigned long bo_size,
1288 unsigned struct_size)
1289 {
1290 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1291 size_t size = 0;
1292
1293 size += ttm_round_pot(struct_size);
1294 size += PAGE_ALIGN(npages * sizeof(void *));
1295 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1296 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1297 return size;
1298 }
1299 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1300
1301 int ttm_bo_create(struct ttm_bo_device *bdev,
1302 unsigned long size,
1303 enum ttm_bo_type type,
1304 struct ttm_placement *placement,
1305 uint32_t page_alignment,
1306 bool interruptible,
1307 struct file *persistent_swap_storage,
1308 struct ttm_buffer_object **p_bo)
1309 {
1310 struct ttm_buffer_object *bo;
1311 size_t acc_size;
1312 int ret;
1313
1314 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1315 if (unlikely(bo == NULL))
1316 return -ENOMEM;
1317
1318 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1319 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1320 interruptible, persistent_swap_storage, acc_size,
1321 NULL, NULL);
1322 if (likely(ret == 0))
1323 *p_bo = bo;
1324
1325 return ret;
1326 }
1327 EXPORT_SYMBOL(ttm_bo_create);
1328
1329 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1330 unsigned mem_type, bool allow_errors)
1331 {
1332 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1333 struct ttm_bo_global *glob = bdev->glob;
1334 int ret;
1335
1336 /*
1337 * Can't use standard list traversal since we're unlocking.
1338 */
1339
1340 spin_lock(&glob->lru_lock);
1341 while (!list_empty(&man->lru)) {
1342 spin_unlock(&glob->lru_lock);
1343 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1344 if (ret) {
1345 if (allow_errors) {
1346 return ret;
1347 } else {
1348 pr_err("Cleanup eviction failed\n");
1349 }
1350 }
1351 spin_lock(&glob->lru_lock);
1352 }
1353 spin_unlock(&glob->lru_lock);
1354 return 0;
1355 }
1356
1357 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1358 {
1359 struct ttm_mem_type_manager *man;
1360 int ret = -EINVAL;
1361
1362 if (mem_type >= TTM_NUM_MEM_TYPES) {
1363 pr_err("Illegal memory type %d\n", mem_type);
1364 return ret;
1365 }
1366 man = &bdev->man[mem_type];
1367
1368 if (!man->has_type) {
1369 pr_err("Trying to take down uninitialized memory manager type %u\n",
1370 mem_type);
1371 return ret;
1372 }
1373
1374 man->use_type = false;
1375 man->has_type = false;
1376
1377 ret = 0;
1378 if (mem_type > 0) {
1379 ttm_bo_force_list_clean(bdev, mem_type, false);
1380
1381 ret = (*man->func->takedown)(man);
1382 }
1383
1384 return ret;
1385 }
1386 EXPORT_SYMBOL(ttm_bo_clean_mm);
1387
1388 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1389 {
1390 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1391
1392 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1393 pr_err("Illegal memory manager memory type %u\n", mem_type);
1394 return -EINVAL;
1395 }
1396
1397 if (!man->has_type) {
1398 pr_err("Memory type %u has not been initialized\n", mem_type);
1399 return 0;
1400 }
1401
1402 return ttm_bo_force_list_clean(bdev, mem_type, true);
1403 }
1404 EXPORT_SYMBOL(ttm_bo_evict_mm);
1405
1406 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1407 unsigned long p_size)
1408 {
1409 int ret = -EINVAL;
1410 struct ttm_mem_type_manager *man;
1411
1412 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1413 man = &bdev->man[type];
1414 BUG_ON(man->has_type);
1415 man->io_reserve_fastpath = true;
1416 man->use_io_reserve_lru = false;
1417 mutex_init(&man->io_reserve_mutex);
1418 INIT_LIST_HEAD(&man->io_reserve_lru);
1419
1420 ret = bdev->driver->init_mem_type(bdev, type, man);
1421 if (ret)
1422 return ret;
1423 man->bdev = bdev;
1424
1425 ret = 0;
1426 if (type != TTM_PL_SYSTEM) {
1427 ret = (*man->func->init)(man, p_size);
1428 if (ret)
1429 return ret;
1430 }
1431 man->has_type = true;
1432 man->use_type = true;
1433 man->size = p_size;
1434
1435 INIT_LIST_HEAD(&man->lru);
1436
1437 return 0;
1438 }
1439 EXPORT_SYMBOL(ttm_bo_init_mm);
1440
1441 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1442 {
1443 struct ttm_bo_global *glob =
1444 container_of(kobj, struct ttm_bo_global, kobj);
1445
1446 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1447 __free_page(glob->dummy_read_page);
1448 kfree(glob);
1449 }
1450
1451 void ttm_bo_global_release(struct drm_global_reference *ref)
1452 {
1453 struct ttm_bo_global *glob = ref->object;
1454
1455 kobject_del(&glob->kobj);
1456 kobject_put(&glob->kobj);
1457 }
1458 EXPORT_SYMBOL(ttm_bo_global_release);
1459
1460 int ttm_bo_global_init(struct drm_global_reference *ref)
1461 {
1462 struct ttm_bo_global_ref *bo_ref =
1463 container_of(ref, struct ttm_bo_global_ref, ref);
1464 struct ttm_bo_global *glob = ref->object;
1465 int ret;
1466
1467 mutex_init(&glob->device_list_mutex);
1468 spin_lock_init(&glob->lru_lock);
1469 glob->mem_glob = bo_ref->mem_glob;
1470 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1471
1472 if (unlikely(glob->dummy_read_page == NULL)) {
1473 ret = -ENOMEM;
1474 goto out_no_drp;
1475 }
1476
1477 INIT_LIST_HEAD(&glob->swap_lru);
1478 INIT_LIST_HEAD(&glob->device_list);
1479
1480 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1481 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1482 if (unlikely(ret != 0)) {
1483 pr_err("Could not register buffer object swapout\n");
1484 goto out_no_shrink;
1485 }
1486
1487 atomic_set(&glob->bo_count, 0);
1488
1489 ret = kobject_init_and_add(
1490 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1491 if (unlikely(ret != 0))
1492 kobject_put(&glob->kobj);
1493 return ret;
1494 out_no_shrink:
1495 __free_page(glob->dummy_read_page);
1496 out_no_drp:
1497 kfree(glob);
1498 return ret;
1499 }
1500 EXPORT_SYMBOL(ttm_bo_global_init);
1501
1502
1503 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1504 {
1505 int ret = 0;
1506 unsigned i = TTM_NUM_MEM_TYPES;
1507 struct ttm_mem_type_manager *man;
1508 struct ttm_bo_global *glob = bdev->glob;
1509
1510 while (i--) {
1511 man = &bdev->man[i];
1512 if (man->has_type) {
1513 man->use_type = false;
1514 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1515 ret = -EBUSY;
1516 pr_err("DRM memory manager type %d is not clean\n",
1517 i);
1518 }
1519 man->has_type = false;
1520 }
1521 }
1522
1523 mutex_lock(&glob->device_list_mutex);
1524 list_del(&bdev->device_list);
1525 mutex_unlock(&glob->device_list_mutex);
1526
1527 cancel_delayed_work_sync(&bdev->wq);
1528
1529 while (ttm_bo_delayed_delete(bdev, true))
1530 ;
1531
1532 spin_lock(&glob->lru_lock);
1533 if (list_empty(&bdev->ddestroy))
1534 TTM_DEBUG("Delayed destroy list was clean\n");
1535
1536 if (list_empty(&bdev->man[0].lru))
1537 TTM_DEBUG("Swap list was clean\n");
1538 spin_unlock(&glob->lru_lock);
1539
1540 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1541 write_lock(&bdev->vm_lock);
1542 drm_mm_takedown(&bdev->addr_space_mm);
1543 write_unlock(&bdev->vm_lock);
1544
1545 return ret;
1546 }
1547 EXPORT_SYMBOL(ttm_bo_device_release);
1548
1549 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1550 struct ttm_bo_global *glob,
1551 struct ttm_bo_driver *driver,
1552 uint64_t file_page_offset,
1553 bool need_dma32)
1554 {
1555 int ret = -EINVAL;
1556
1557 rwlock_init(&bdev->vm_lock);
1558 bdev->driver = driver;
1559
1560 memset(bdev->man, 0, sizeof(bdev->man));
1561
1562 /*
1563 * Initialize the system memory buffer type.
1564 * Other types need to be driver / IOCTL initialized.
1565 */
1566 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1567 if (unlikely(ret != 0))
1568 goto out_no_sys;
1569
1570 bdev->addr_space_rb = RB_ROOT;
1571 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1572 if (unlikely(ret != 0))
1573 goto out_no_addr_mm;
1574
1575 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1576 bdev->nice_mode = true;
1577 INIT_LIST_HEAD(&bdev->ddestroy);
1578 bdev->dev_mapping = NULL;
1579 bdev->glob = glob;
1580 bdev->need_dma32 = need_dma32;
1581 bdev->val_seq = 0;
1582 spin_lock_init(&bdev->fence_lock);
1583 mutex_lock(&glob->device_list_mutex);
1584 list_add_tail(&bdev->device_list, &glob->device_list);
1585 mutex_unlock(&glob->device_list_mutex);
1586
1587 return 0;
1588 out_no_addr_mm:
1589 ttm_bo_clean_mm(bdev, 0);
1590 out_no_sys:
1591 return ret;
1592 }
1593 EXPORT_SYMBOL(ttm_bo_device_init);
1594
1595 /*
1596 * buffer object vm functions.
1597 */
1598
1599 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1600 {
1601 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1602
1603 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1604 if (mem->mem_type == TTM_PL_SYSTEM)
1605 return false;
1606
1607 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1608 return false;
1609
1610 if (mem->placement & TTM_PL_FLAG_CACHED)
1611 return false;
1612 }
1613 return true;
1614 }
1615
1616 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1617 {
1618 struct ttm_bo_device *bdev = bo->bdev;
1619 loff_t offset = (loff_t) bo->addr_space_offset;
1620 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1621
1622 if (!bdev->dev_mapping)
1623 return;
1624 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1625 ttm_mem_io_free_vm(bo);
1626 }
1627
1628 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1629 {
1630 struct ttm_bo_device *bdev = bo->bdev;
1631 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1632
1633 ttm_mem_io_lock(man, false);
1634 ttm_bo_unmap_virtual_locked(bo);
1635 ttm_mem_io_unlock(man);
1636 }
1637
1638
1639 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1640
1641 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1642 {
1643 struct ttm_bo_device *bdev = bo->bdev;
1644 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1645 struct rb_node *parent = NULL;
1646 struct ttm_buffer_object *cur_bo;
1647 unsigned long offset = bo->vm_node->start;
1648 unsigned long cur_offset;
1649
1650 while (*cur) {
1651 parent = *cur;
1652 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1653 cur_offset = cur_bo->vm_node->start;
1654 if (offset < cur_offset)
1655 cur = &parent->rb_left;
1656 else if (offset > cur_offset)
1657 cur = &parent->rb_right;
1658 else
1659 BUG();
1660 }
1661
1662 rb_link_node(&bo->vm_rb, parent, cur);
1663 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1664 }
1665
1666 /**
1667 * ttm_bo_setup_vm:
1668 *
1669 * @bo: the buffer to allocate address space for
1670 *
1671 * Allocate address space in the drm device so that applications
1672 * can mmap the buffer and access the contents. This only
1673 * applies to ttm_bo_type_device objects as others are not
1674 * placed in the drm device address space.
1675 */
1676
1677 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1678 {
1679 struct ttm_bo_device *bdev = bo->bdev;
1680 int ret;
1681
1682 retry_pre_get:
1683 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1684 if (unlikely(ret != 0))
1685 return ret;
1686
1687 write_lock(&bdev->vm_lock);
1688 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1689 bo->mem.num_pages, 0, 0);
1690
1691 if (unlikely(bo->vm_node == NULL)) {
1692 ret = -ENOMEM;
1693 goto out_unlock;
1694 }
1695
1696 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1697 bo->mem.num_pages, 0);
1698
1699 if (unlikely(bo->vm_node == NULL)) {
1700 write_unlock(&bdev->vm_lock);
1701 goto retry_pre_get;
1702 }
1703
1704 ttm_bo_vm_insert_rb(bo);
1705 write_unlock(&bdev->vm_lock);
1706 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1707
1708 return 0;
1709 out_unlock:
1710 write_unlock(&bdev->vm_lock);
1711 return ret;
1712 }
1713
1714 int ttm_bo_wait(struct ttm_buffer_object *bo,
1715 bool lazy, bool interruptible, bool no_wait)
1716 {
1717 struct ttm_bo_driver *driver = bo->bdev->driver;
1718 struct ttm_bo_device *bdev = bo->bdev;
1719 void *sync_obj;
1720 void *sync_obj_arg;
1721 int ret = 0;
1722
1723 if (likely(bo->sync_obj == NULL))
1724 return 0;
1725
1726 while (bo->sync_obj) {
1727
1728 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1729 void *tmp_obj = bo->sync_obj;
1730 bo->sync_obj = NULL;
1731 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1732 spin_unlock(&bdev->fence_lock);
1733 driver->sync_obj_unref(&tmp_obj);
1734 spin_lock(&bdev->fence_lock);
1735 continue;
1736 }
1737
1738 if (no_wait)
1739 return -EBUSY;
1740
1741 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1742 sync_obj_arg = bo->sync_obj_arg;
1743 spin_unlock(&bdev->fence_lock);
1744 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1745 lazy, interruptible);
1746 if (unlikely(ret != 0)) {
1747 driver->sync_obj_unref(&sync_obj);
1748 spin_lock(&bdev->fence_lock);
1749 return ret;
1750 }
1751 spin_lock(&bdev->fence_lock);
1752 if (likely(bo->sync_obj == sync_obj &&
1753 bo->sync_obj_arg == sync_obj_arg)) {
1754 void *tmp_obj = bo->sync_obj;
1755 bo->sync_obj = NULL;
1756 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1757 &bo->priv_flags);
1758 spin_unlock(&bdev->fence_lock);
1759 driver->sync_obj_unref(&sync_obj);
1760 driver->sync_obj_unref(&tmp_obj);
1761 spin_lock(&bdev->fence_lock);
1762 } else {
1763 spin_unlock(&bdev->fence_lock);
1764 driver->sync_obj_unref(&sync_obj);
1765 spin_lock(&bdev->fence_lock);
1766 }
1767 }
1768 return 0;
1769 }
1770 EXPORT_SYMBOL(ttm_bo_wait);
1771
1772 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1773 {
1774 struct ttm_bo_device *bdev = bo->bdev;
1775 int ret = 0;
1776
1777 /*
1778 * Using ttm_bo_reserve makes sure the lru lists are updated.
1779 */
1780
1781 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1782 if (unlikely(ret != 0))
1783 return ret;
1784 spin_lock(&bdev->fence_lock);
1785 ret = ttm_bo_wait(bo, false, true, no_wait);
1786 spin_unlock(&bdev->fence_lock);
1787 if (likely(ret == 0))
1788 atomic_inc(&bo->cpu_writers);
1789 ttm_bo_unreserve(bo);
1790 return ret;
1791 }
1792 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1793
1794 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1795 {
1796 if (atomic_dec_and_test(&bo->cpu_writers))
1797 wake_up_all(&bo->event_queue);
1798 }
1799 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1800
1801 /**
1802 * A buffer object shrink method that tries to swap out the first
1803 * buffer object on the bo_global::swap_lru list.
1804 */
1805
1806 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1807 {
1808 struct ttm_bo_global *glob =
1809 container_of(shrink, struct ttm_bo_global, shrink);
1810 struct ttm_buffer_object *bo;
1811 int ret = -EBUSY;
1812 int put_count;
1813 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1814
1815 spin_lock(&glob->lru_lock);
1816 while (ret == -EBUSY) {
1817 if (unlikely(list_empty(&glob->swap_lru))) {
1818 spin_unlock(&glob->lru_lock);
1819 return -EBUSY;
1820 }
1821
1822 bo = list_first_entry(&glob->swap_lru,
1823 struct ttm_buffer_object, swap);
1824 kref_get(&bo->list_kref);
1825
1826 if (!list_empty(&bo->ddestroy)) {
1827 spin_unlock(&glob->lru_lock);
1828 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1829 kref_put(&bo->list_kref, ttm_bo_release_list);
1830 spin_lock(&glob->lru_lock);
1831 continue;
1832 }
1833
1834 /**
1835 * Reserve buffer. Since we unlock while sleeping, we need
1836 * to re-check that nobody removed us from the swap-list while
1837 * we slept.
1838 */
1839
1840 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1841 if (unlikely(ret == -EBUSY)) {
1842 spin_unlock(&glob->lru_lock);
1843 ttm_bo_wait_unreserved(bo, false);
1844 kref_put(&bo->list_kref, ttm_bo_release_list);
1845 spin_lock(&glob->lru_lock);
1846 }
1847 }
1848
1849 BUG_ON(ret != 0);
1850 put_count = ttm_bo_del_from_lru(bo);
1851 spin_unlock(&glob->lru_lock);
1852
1853 ttm_bo_list_ref_sub(bo, put_count, true);
1854
1855 /**
1856 * Wait for GPU, then move to system cached.
1857 */
1858
1859 spin_lock(&bo->bdev->fence_lock);
1860 ret = ttm_bo_wait(bo, false, false, false);
1861 spin_unlock(&bo->bdev->fence_lock);
1862
1863 if (unlikely(ret != 0))
1864 goto out;
1865
1866 if ((bo->mem.placement & swap_placement) != swap_placement) {
1867 struct ttm_mem_reg evict_mem;
1868
1869 evict_mem = bo->mem;
1870 evict_mem.mm_node = NULL;
1871 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1872 evict_mem.mem_type = TTM_PL_SYSTEM;
1873
1874 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1875 false, false, false);
1876 if (unlikely(ret != 0))
1877 goto out;
1878 }
1879
1880 ttm_bo_unmap_virtual(bo);
1881
1882 /**
1883 * Swap out. Buffer will be swapped in again as soon as
1884 * anyone tries to access a ttm page.
1885 */
1886
1887 if (bo->bdev->driver->swap_notify)
1888 bo->bdev->driver->swap_notify(bo);
1889
1890 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1891 out:
1892
1893 /**
1894 *
1895 * Unreserve without putting on LRU to avoid swapping out an
1896 * already swapped buffer.
1897 */
1898
1899 atomic_set(&bo->reserved, 0);
1900 wake_up_all(&bo->event_queue);
1901 kref_put(&bo->list_kref, ttm_bo_release_list);
1902 return ret;
1903 }
1904
1905 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1906 {
1907 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1908 ;
1909 }
1910 EXPORT_SYMBOL(ttm_bo_swapout_all);
Linux kernel - Deliverables
This page took 0.133465 seconds and 5 git commands to generate.