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


b7781453bfd17728a3522520c46706b86f53e372
[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_read(&bo->reserved) != 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 atomic_set(&bo->reserved, 1);
253 if (use_sequence) {
254 /**
255 * Wake up waiters that may need to recheck for deadlock,
256 * if we decreased the sequence number.
257 */
258 if (unlikely((bo->val_seq - sequence < (1 << 31))
259 || !bo->seq_valid))
260 wake_up_all(&bo->event_queue);
261
262 bo->val_seq = sequence;
263 bo->seq_valid = true;
264 } else {
265 bo->seq_valid = false;
266 }
267
268 return 0;
269 }
270 EXPORT_SYMBOL(ttm_bo_reserve);
271
272 static void ttm_bo_ref_bug(struct kref *list_kref)
273 {
274 BUG();
275 }
276
277 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
278 bool never_free)
279 {
280 kref_sub(&bo->list_kref, count,
281 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
282 }
283
284 int ttm_bo_reserve(struct ttm_buffer_object *bo,
285 bool interruptible,
286 bool no_wait, bool use_sequence, uint32_t sequence)
287 {
288 struct ttm_bo_global *glob = bo->glob;
289 int put_count = 0;
290 int ret;
291
292 spin_lock(&glob->lru_lock);
293 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
294 sequence);
295 if (likely(ret == 0))
296 put_count = ttm_bo_del_from_lru(bo);
297 spin_unlock(&glob->lru_lock);
298
299 ttm_bo_list_ref_sub(bo, put_count, true);
300
301 return ret;
302 }
303
304 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo)
305 {
306 ttm_bo_add_to_lru(bo);
307 atomic_set(&bo->reserved, 0);
308 wake_up_all(&bo->event_queue);
309 }
310
311 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
312 {
313 struct ttm_bo_global *glob = bo->glob;
314
315 spin_lock(&glob->lru_lock);
316 ttm_bo_unreserve_locked(bo);
317 spin_unlock(&glob->lru_lock);
318 }
319 EXPORT_SYMBOL(ttm_bo_unreserve);
320
321 /*
322 * Call bo->mutex locked.
323 */
324 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
325 {
326 struct ttm_bo_device *bdev = bo->bdev;
327 struct ttm_bo_global *glob = bo->glob;
328 int ret = 0;
329 uint32_t page_flags = 0;
330
331 TTM_ASSERT_LOCKED(&bo->mutex);
332 bo->ttm = NULL;
333
334 if (bdev->need_dma32)
335 page_flags |= TTM_PAGE_FLAG_DMA32;
336
337 switch (bo->type) {
338 case ttm_bo_type_device:
339 if (zero_alloc)
340 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
341 case ttm_bo_type_kernel:
342 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
343 page_flags, glob->dummy_read_page);
344 if (unlikely(bo->ttm == NULL))
345 ret = -ENOMEM;
346 break;
347 case ttm_bo_type_sg:
348 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
349 page_flags | TTM_PAGE_FLAG_SG,
350 glob->dummy_read_page);
351 if (unlikely(bo->ttm == NULL)) {
352 ret = -ENOMEM;
353 break;
354 }
355 bo->ttm->sg = bo->sg;
356 break;
357 default:
358 pr_err("Illegal buffer object type\n");
359 ret = -EINVAL;
360 break;
361 }
362
363 return ret;
364 }
365
366 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
367 struct ttm_mem_reg *mem,
368 bool evict, bool interruptible,
369 bool no_wait_reserve, bool no_wait_gpu)
370 {
371 struct ttm_bo_device *bdev = bo->bdev;
372 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
373 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
374 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
375 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
376 int ret = 0;
377
378 if (old_is_pci || new_is_pci ||
379 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
380 ret = ttm_mem_io_lock(old_man, true);
381 if (unlikely(ret != 0))
382 goto out_err;
383 ttm_bo_unmap_virtual_locked(bo);
384 ttm_mem_io_unlock(old_man);
385 }
386
387 /*
388 * Create and bind a ttm if required.
389 */
390
391 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
392 if (bo->ttm == NULL) {
393 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
394 ret = ttm_bo_add_ttm(bo, zero);
395 if (ret)
396 goto out_err;
397 }
398
399 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
400 if (ret)
401 goto out_err;
402
403 if (mem->mem_type != TTM_PL_SYSTEM) {
404 ret = ttm_tt_bind(bo->ttm, mem);
405 if (ret)
406 goto out_err;
407 }
408
409 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
410 if (bdev->driver->move_notify)
411 bdev->driver->move_notify(bo, mem);
412 bo->mem = *mem;
413 mem->mm_node = NULL;
414 goto moved;
415 }
416 }
417
418 if (bdev->driver->move_notify)
419 bdev->driver->move_notify(bo, mem);
420
421 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
422 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
423 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
424 else if (bdev->driver->move)
425 ret = bdev->driver->move(bo, evict, interruptible,
426 no_wait_reserve, no_wait_gpu, mem);
427 else
428 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
429
430 if (ret) {
431 if (bdev->driver->move_notify) {
432 struct ttm_mem_reg tmp_mem = *mem;
433 *mem = bo->mem;
434 bo->mem = tmp_mem;
435 bdev->driver->move_notify(bo, mem);
436 bo->mem = *mem;
437 }
438
439 goto out_err;
440 }
441
442 moved:
443 if (bo->evicted) {
444 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
445 if (ret)
446 pr_err("Can not flush read caches\n");
447 bo->evicted = false;
448 }
449
450 if (bo->mem.mm_node) {
451 bo->offset = (bo->mem.start << PAGE_SHIFT) +
452 bdev->man[bo->mem.mem_type].gpu_offset;
453 bo->cur_placement = bo->mem.placement;
454 } else
455 bo->offset = 0;
456
457 return 0;
458
459 out_err:
460 new_man = &bdev->man[bo->mem.mem_type];
461 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
462 ttm_tt_unbind(bo->ttm);
463 ttm_tt_destroy(bo->ttm);
464 bo->ttm = NULL;
465 }
466
467 return ret;
468 }
469
470 /**
471 * Call bo::reserved.
472 * Will release GPU memory type usage on destruction.
473 * This is the place to put in driver specific hooks to release
474 * driver private resources.
475 * Will release the bo::reserved lock.
476 */
477
478 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
479 {
480 if (bo->bdev->driver->move_notify)
481 bo->bdev->driver->move_notify(bo, NULL);
482
483 if (bo->ttm) {
484 ttm_tt_unbind(bo->ttm);
485 ttm_tt_destroy(bo->ttm);
486 bo->ttm = NULL;
487 }
488 ttm_bo_mem_put(bo, &bo->mem);
489
490 atomic_set(&bo->reserved, 0);
491
492 /*
493 * Make processes trying to reserve really pick it up.
494 */
495 smp_mb__after_atomic_dec();
496 wake_up_all(&bo->event_queue);
497 }
498
499 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
500 {
501 struct ttm_bo_device *bdev = bo->bdev;
502 struct ttm_bo_global *glob = bo->glob;
503 struct ttm_bo_driver *driver = bdev->driver;
504 void *sync_obj = NULL;
505 int put_count;
506 int ret;
507
508 spin_lock(&glob->lru_lock);
509 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
510
511 spin_lock(&bdev->fence_lock);
512 (void) ttm_bo_wait(bo, false, false, true);
513 if (!ret && !bo->sync_obj) {
514 spin_unlock(&bdev->fence_lock);
515 put_count = ttm_bo_del_from_lru(bo);
516
517 spin_unlock(&glob->lru_lock);
518 ttm_bo_cleanup_memtype_use(bo);
519
520 ttm_bo_list_ref_sub(bo, put_count, true);
521
522 return;
523 }
524 if (bo->sync_obj)
525 sync_obj = driver->sync_obj_ref(bo->sync_obj);
526 spin_unlock(&bdev->fence_lock);
527
528 if (!ret) {
529 atomic_set(&bo->reserved, 0);
530 wake_up_all(&bo->event_queue);
531 }
532
533 kref_get(&bo->list_kref);
534 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
535 spin_unlock(&glob->lru_lock);
536
537 if (sync_obj) {
538 driver->sync_obj_flush(sync_obj);
539 driver->sync_obj_unref(&sync_obj);
540 }
541 schedule_delayed_work(&bdev->wq,
542 ((HZ / 100) < 1) ? 1 : HZ / 100);
543 }
544
545 /**
546 * function ttm_bo_cleanup_refs
547 * If bo idle, remove from delayed- and lru lists, and unref.
548 * If not idle, do nothing.
549 *
550 * @interruptible Any sleeps should occur interruptibly.
551 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
552 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
553 */
554
555 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
556 bool interruptible,
557 bool no_wait_reserve,
558 bool no_wait_gpu)
559 {
560 struct ttm_bo_device *bdev = bo->bdev;
561 struct ttm_bo_global *glob = bo->glob;
562 int put_count;
563 int ret = 0;
564
565 retry:
566 spin_lock(&bdev->fence_lock);
567 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
568 spin_unlock(&bdev->fence_lock);
569
570 if (unlikely(ret != 0))
571 return ret;
572
573 retry_reserve:
574 spin_lock(&glob->lru_lock);
575
576 if (unlikely(list_empty(&bo->ddestroy))) {
577 spin_unlock(&glob->lru_lock);
578 return 0;
579 }
580
581 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
582
583 if (unlikely(ret == -EBUSY)) {
584 spin_unlock(&glob->lru_lock);
585 if (likely(!no_wait_reserve))
586 ret = ttm_bo_wait_unreserved(bo, interruptible);
587 if (unlikely(ret != 0))
588 return ret;
589
590 goto retry_reserve;
591 }
592
593 BUG_ON(ret != 0);
594
595 /**
596 * We can re-check for sync object without taking
597 * the bo::lock since setting the sync object requires
598 * also bo::reserved. A busy object at this point may
599 * be caused by another thread recently starting an accelerated
600 * eviction.
601 */
602
603 if (unlikely(bo->sync_obj)) {
604 atomic_set(&bo->reserved, 0);
605 wake_up_all(&bo->event_queue);
606 spin_unlock(&glob->lru_lock);
607 goto retry;
608 }
609
610 put_count = ttm_bo_del_from_lru(bo);
611 list_del_init(&bo->ddestroy);
612 ++put_count;
613
614 spin_unlock(&glob->lru_lock);
615 ttm_bo_cleanup_memtype_use(bo);
616
617 ttm_bo_list_ref_sub(bo, put_count, true);
618
619 return 0;
620 }
621
622 /**
623 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
624 * encountered buffers.
625 */
626
627 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
628 {
629 struct ttm_bo_global *glob = bdev->glob;
630 struct ttm_buffer_object *entry = NULL;
631 int ret = 0;
632
633 spin_lock(&glob->lru_lock);
634 if (list_empty(&bdev->ddestroy))
635 goto out_unlock;
636
637 entry = list_first_entry(&bdev->ddestroy,
638 struct ttm_buffer_object, ddestroy);
639 kref_get(&entry->list_kref);
640
641 for (;;) {
642 struct ttm_buffer_object *nentry = NULL;
643
644 if (entry->ddestroy.next != &bdev->ddestroy) {
645 nentry = list_first_entry(&entry->ddestroy,
646 struct ttm_buffer_object, ddestroy);
647 kref_get(&nentry->list_kref);
648 }
649
650 spin_unlock(&glob->lru_lock);
651 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
652 !remove_all);
653 kref_put(&entry->list_kref, ttm_bo_release_list);
654 entry = nentry;
655
656 if (ret || !entry)
657 goto out;
658
659 spin_lock(&glob->lru_lock);
660 if (list_empty(&entry->ddestroy))
661 break;
662 }
663
664 out_unlock:
665 spin_unlock(&glob->lru_lock);
666 out:
667 if (entry)
668 kref_put(&entry->list_kref, ttm_bo_release_list);
669 return ret;
670 }
671
672 static void ttm_bo_delayed_workqueue(struct work_struct *work)
673 {
674 struct ttm_bo_device *bdev =
675 container_of(work, struct ttm_bo_device, wq.work);
676
677 if (ttm_bo_delayed_delete(bdev, false)) {
678 schedule_delayed_work(&bdev->wq,
679 ((HZ / 100) < 1) ? 1 : HZ / 100);
680 }
681 }
682
683 static void ttm_bo_release(struct kref *kref)
684 {
685 struct ttm_buffer_object *bo =
686 container_of(kref, struct ttm_buffer_object, kref);
687 struct ttm_bo_device *bdev = bo->bdev;
688 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
689
690 write_lock(&bdev->vm_lock);
691 if (likely(bo->vm_node != NULL)) {
692 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
693 drm_mm_put_block(bo->vm_node);
694 bo->vm_node = NULL;
695 }
696 write_unlock(&bdev->vm_lock);
697 ttm_mem_io_lock(man, false);
698 ttm_mem_io_free_vm(bo);
699 ttm_mem_io_unlock(man);
700 ttm_bo_cleanup_refs_or_queue(bo);
701 kref_put(&bo->list_kref, ttm_bo_release_list);
702 }
703
704 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
705 {
706 struct ttm_buffer_object *bo = *p_bo;
707
708 *p_bo = NULL;
709 kref_put(&bo->kref, ttm_bo_release);
710 }
711 EXPORT_SYMBOL(ttm_bo_unref);
712
713 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
714 {
715 return cancel_delayed_work_sync(&bdev->wq);
716 }
717 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
718
719 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
720 {
721 if (resched)
722 schedule_delayed_work(&bdev->wq,
723 ((HZ / 100) < 1) ? 1 : HZ / 100);
724 }
725 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
726
727 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
728 bool no_wait_reserve, bool no_wait_gpu)
729 {
730 struct ttm_bo_device *bdev = bo->bdev;
731 struct ttm_mem_reg evict_mem;
732 struct ttm_placement placement;
733 int ret = 0;
734
735 spin_lock(&bdev->fence_lock);
736 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
737 spin_unlock(&bdev->fence_lock);
738
739 if (unlikely(ret != 0)) {
740 if (ret != -ERESTARTSYS) {
741 pr_err("Failed to expire sync object before buffer eviction\n");
742 }
743 goto out;
744 }
745
746 BUG_ON(!ttm_bo_is_reserved(bo));
747
748 evict_mem = bo->mem;
749 evict_mem.mm_node = NULL;
750 evict_mem.bus.io_reserved_vm = false;
751 evict_mem.bus.io_reserved_count = 0;
752
753 placement.fpfn = 0;
754 placement.lpfn = 0;
755 placement.num_placement = 0;
756 placement.num_busy_placement = 0;
757 bdev->driver->evict_flags(bo, &placement);
758 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
759 no_wait_reserve, no_wait_gpu);
760 if (ret) {
761 if (ret != -ERESTARTSYS) {
762 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
763 bo);
764 ttm_bo_mem_space_debug(bo, &placement);
765 }
766 goto out;
767 }
768
769 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
770 no_wait_reserve, no_wait_gpu);
771 if (ret) {
772 if (ret != -ERESTARTSYS)
773 pr_err("Buffer eviction failed\n");
774 ttm_bo_mem_put(bo, &evict_mem);
775 goto out;
776 }
777 bo->evicted = true;
778 out:
779 return ret;
780 }
781
782 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
783 uint32_t mem_type,
784 bool interruptible, bool no_wait_reserve,
785 bool no_wait_gpu)
786 {
787 struct ttm_bo_global *glob = bdev->glob;
788 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
789 struct ttm_buffer_object *bo;
790 int ret, put_count = 0;
791
792 retry:
793 spin_lock(&glob->lru_lock);
794 if (list_empty(&man->lru)) {
795 spin_unlock(&glob->lru_lock);
796 return -EBUSY;
797 }
798
799 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
800 kref_get(&bo->list_kref);
801
802 if (!list_empty(&bo->ddestroy)) {
803 spin_unlock(&glob->lru_lock);
804 ret = ttm_bo_cleanup_refs(bo, interruptible,
805 no_wait_reserve, no_wait_gpu);
806 kref_put(&bo->list_kref, ttm_bo_release_list);
807
808 return ret;
809 }
810
811 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
812
813 if (unlikely(ret == -EBUSY)) {
814 spin_unlock(&glob->lru_lock);
815 if (likely(!no_wait_reserve))
816 ret = ttm_bo_wait_unreserved(bo, interruptible);
817
818 kref_put(&bo->list_kref, ttm_bo_release_list);
819
820 /**
821 * We *need* to retry after releasing the lru lock.
822 */
823
824 if (unlikely(ret != 0))
825 return ret;
826 goto retry;
827 }
828
829 put_count = ttm_bo_del_from_lru(bo);
830 spin_unlock(&glob->lru_lock);
831
832 BUG_ON(ret != 0);
833
834 ttm_bo_list_ref_sub(bo, put_count, true);
835
836 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
837 ttm_bo_unreserve(bo);
838
839 kref_put(&bo->list_kref, ttm_bo_release_list);
840 return ret;
841 }
842
843 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
844 {
845 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
846
847 if (mem->mm_node)
848 (*man->func->put_node)(man, mem);
849 }
850 EXPORT_SYMBOL(ttm_bo_mem_put);
851
852 /**
853 * Repeatedly evict memory from the LRU for @mem_type until we create enough
854 * space, or we've evicted everything and there isn't enough space.
855 */
856 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
857 uint32_t mem_type,
858 struct ttm_placement *placement,
859 struct ttm_mem_reg *mem,
860 bool interruptible,
861 bool no_wait_reserve,
862 bool no_wait_gpu)
863 {
864 struct ttm_bo_device *bdev = bo->bdev;
865 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
866 int ret;
867
868 do {
869 ret = (*man->func->get_node)(man, bo, placement, mem);
870 if (unlikely(ret != 0))
871 return ret;
872 if (mem->mm_node)
873 break;
874 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
875 no_wait_reserve, no_wait_gpu);
876 if (unlikely(ret != 0))
877 return ret;
878 } while (1);
879 if (mem->mm_node == NULL)
880 return -ENOMEM;
881 mem->mem_type = mem_type;
882 return 0;
883 }
884
885 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
886 uint32_t cur_placement,
887 uint32_t proposed_placement)
888 {
889 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
890 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
891
892 /**
893 * Keep current caching if possible.
894 */
895
896 if ((cur_placement & caching) != 0)
897 result |= (cur_placement & caching);
898 else if ((man->default_caching & caching) != 0)
899 result |= man->default_caching;
900 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
901 result |= TTM_PL_FLAG_CACHED;
902 else if ((TTM_PL_FLAG_WC & caching) != 0)
903 result |= TTM_PL_FLAG_WC;
904 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
905 result |= TTM_PL_FLAG_UNCACHED;
906
907 return result;
908 }
909
910 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
911 uint32_t mem_type,
912 uint32_t proposed_placement,
913 uint32_t *masked_placement)
914 {
915 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
916
917 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
918 return false;
919
920 if ((proposed_placement & man->available_caching) == 0)
921 return false;
922
923 cur_flags |= (proposed_placement & man->available_caching);
924
925 *masked_placement = cur_flags;
926 return true;
927 }
928
929 /**
930 * Creates space for memory region @mem according to its type.
931 *
932 * This function first searches for free space in compatible memory types in
933 * the priority order defined by the driver. If free space isn't found, then
934 * ttm_bo_mem_force_space is attempted in priority order to evict and find
935 * space.
936 */
937 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
938 struct ttm_placement *placement,
939 struct ttm_mem_reg *mem,
940 bool interruptible, bool no_wait_reserve,
941 bool no_wait_gpu)
942 {
943 struct ttm_bo_device *bdev = bo->bdev;
944 struct ttm_mem_type_manager *man;
945 uint32_t mem_type = TTM_PL_SYSTEM;
946 uint32_t cur_flags = 0;
947 bool type_found = false;
948 bool type_ok = false;
949 bool has_erestartsys = false;
950 int i, ret;
951
952 mem->mm_node = NULL;
953 for (i = 0; i < placement->num_placement; ++i) {
954 ret = ttm_mem_type_from_flags(placement->placement[i],
955 &mem_type);
956 if (ret)
957 return ret;
958 man = &bdev->man[mem_type];
959
960 type_ok = ttm_bo_mt_compatible(man,
961 mem_type,
962 placement->placement[i],
963 &cur_flags);
964
965 if (!type_ok)
966 continue;
967
968 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
969 cur_flags);
970 /*
971 * Use the access and other non-mapping-related flag bits from
972 * the memory placement flags to the current flags
973 */
974 ttm_flag_masked(&cur_flags, placement->placement[i],
975 ~TTM_PL_MASK_MEMTYPE);
976
977 if (mem_type == TTM_PL_SYSTEM)
978 break;
979
980 if (man->has_type && man->use_type) {
981 type_found = true;
982 ret = (*man->func->get_node)(man, bo, placement, mem);
983 if (unlikely(ret))
984 return ret;
985 }
986 if (mem->mm_node)
987 break;
988 }
989
990 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
991 mem->mem_type = mem_type;
992 mem->placement = cur_flags;
993 return 0;
994 }
995
996 if (!type_found)
997 return -EINVAL;
998
999 for (i = 0; i < placement->num_busy_placement; ++i) {
1000 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
1001 &mem_type);
1002 if (ret)
1003 return ret;
1004 man = &bdev->man[mem_type];
1005 if (!man->has_type)
1006 continue;
1007 if (!ttm_bo_mt_compatible(man,
1008 mem_type,
1009 placement->busy_placement[i],
1010 &cur_flags))
1011 continue;
1012
1013 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1014 cur_flags);
1015 /*
1016 * Use the access and other non-mapping-related flag bits from
1017 * the memory placement flags to the current flags
1018 */
1019 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1020 ~TTM_PL_MASK_MEMTYPE);
1021
1022
1023 if (mem_type == TTM_PL_SYSTEM) {
1024 mem->mem_type = mem_type;
1025 mem->placement = cur_flags;
1026 mem->mm_node = NULL;
1027 return 0;
1028 }
1029
1030 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1031 interruptible, no_wait_reserve, no_wait_gpu);
1032 if (ret == 0 && mem->mm_node) {
1033 mem->placement = cur_flags;
1034 return 0;
1035 }
1036 if (ret == -ERESTARTSYS)
1037 has_erestartsys = true;
1038 }
1039 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1040 return ret;
1041 }
1042 EXPORT_SYMBOL(ttm_bo_mem_space);
1043
1044 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1045 struct ttm_placement *placement,
1046 bool interruptible, bool no_wait_reserve,
1047 bool no_wait_gpu)
1048 {
1049 int ret = 0;
1050 struct ttm_mem_reg mem;
1051 struct ttm_bo_device *bdev = bo->bdev;
1052
1053 BUG_ON(!ttm_bo_is_reserved(bo));
1054
1055 /*
1056 * FIXME: It's possible to pipeline buffer moves.
1057 * Have the driver move function wait for idle when necessary,
1058 * instead of doing it here.
1059 */
1060 spin_lock(&bdev->fence_lock);
1061 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1062 spin_unlock(&bdev->fence_lock);
1063 if (ret)
1064 return ret;
1065 mem.num_pages = bo->num_pages;
1066 mem.size = mem.num_pages << PAGE_SHIFT;
1067 mem.page_alignment = bo->mem.page_alignment;
1068 mem.bus.io_reserved_vm = false;
1069 mem.bus.io_reserved_count = 0;
1070 /*
1071 * Determine where to move the buffer.
1072 */
1073 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1074 if (ret)
1075 goto out_unlock;
1076 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1077 out_unlock:
1078 if (ret && mem.mm_node)
1079 ttm_bo_mem_put(bo, &mem);
1080 return ret;
1081 }
1082
1083 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1084 struct ttm_mem_reg *mem)
1085 {
1086 int i;
1087
1088 if (mem->mm_node && placement->lpfn != 0 &&
1089 (mem->start < placement->fpfn ||
1090 mem->start + mem->num_pages > placement->lpfn))
1091 return -1;
1092
1093 for (i = 0; i < placement->num_placement; i++) {
1094 if ((placement->placement[i] & mem->placement &
1095 TTM_PL_MASK_CACHING) &&
1096 (placement->placement[i] & mem->placement &
1097 TTM_PL_MASK_MEM))
1098 return i;
1099 }
1100 return -1;
1101 }
1102
1103 int ttm_bo_validate(struct ttm_buffer_object *bo,
1104 struct ttm_placement *placement,
1105 bool interruptible, bool no_wait_reserve,
1106 bool no_wait_gpu)
1107 {
1108 int ret;
1109
1110 BUG_ON(!ttm_bo_is_reserved(bo));
1111 /* Check that range is valid */
1112 if (placement->lpfn || placement->fpfn)
1113 if (placement->fpfn > placement->lpfn ||
1114 (placement->lpfn - placement->fpfn) < bo->num_pages)
1115 return -EINVAL;
1116 /*
1117 * Check whether we need to move buffer.
1118 */
1119 ret = ttm_bo_mem_compat(placement, &bo->mem);
1120 if (ret < 0) {
1121 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1122 if (ret)
1123 return ret;
1124 } else {
1125 /*
1126 * Use the access and other non-mapping-related flag bits from
1127 * the compatible memory placement flags to the active flags
1128 */
1129 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1130 ~TTM_PL_MASK_MEMTYPE);
1131 }
1132 /*
1133 * We might need to add a TTM.
1134 */
1135 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1136 ret = ttm_bo_add_ttm(bo, true);
1137 if (ret)
1138 return ret;
1139 }
1140 return 0;
1141 }
1142 EXPORT_SYMBOL(ttm_bo_validate);
1143
1144 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1145 struct ttm_placement *placement)
1146 {
1147 BUG_ON((placement->fpfn || placement->lpfn) &&
1148 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1149
1150 return 0;
1151 }
1152
1153 int ttm_bo_init(struct ttm_bo_device *bdev,
1154 struct ttm_buffer_object *bo,
1155 unsigned long size,
1156 enum ttm_bo_type type,
1157 struct ttm_placement *placement,
1158 uint32_t page_alignment,
1159 bool interruptible,
1160 struct file *persistent_swap_storage,
1161 size_t acc_size,
1162 struct sg_table *sg,
1163 void (*destroy) (struct ttm_buffer_object *))
1164 {
1165 int ret = 0;
1166 unsigned long num_pages;
1167 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1168
1169 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1170 if (ret) {
1171 pr_err("Out of kernel memory\n");
1172 if (destroy)
1173 (*destroy)(bo);
1174 else
1175 kfree(bo);
1176 return -ENOMEM;
1177 }
1178
1179 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1180 if (num_pages == 0) {
1181 pr_err("Illegal buffer object size\n");
1182 if (destroy)
1183 (*destroy)(bo);
1184 else
1185 kfree(bo);
1186 ttm_mem_global_free(mem_glob, acc_size);
1187 return -EINVAL;
1188 }
1189 bo->destroy = destroy;
1190
1191 kref_init(&bo->kref);
1192 kref_init(&bo->list_kref);
1193 atomic_set(&bo->cpu_writers, 0);
1194 atomic_set(&bo->reserved, 1);
1195 init_waitqueue_head(&bo->event_queue);
1196 INIT_LIST_HEAD(&bo->lru);
1197 INIT_LIST_HEAD(&bo->ddestroy);
1198 INIT_LIST_HEAD(&bo->swap);
1199 INIT_LIST_HEAD(&bo->io_reserve_lru);
1200 bo->bdev = bdev;
1201 bo->glob = bdev->glob;
1202 bo->type = type;
1203 bo->num_pages = num_pages;
1204 bo->mem.size = num_pages << PAGE_SHIFT;
1205 bo->mem.mem_type = TTM_PL_SYSTEM;
1206 bo->mem.num_pages = bo->num_pages;
1207 bo->mem.mm_node = NULL;
1208 bo->mem.page_alignment = page_alignment;
1209 bo->mem.bus.io_reserved_vm = false;
1210 bo->mem.bus.io_reserved_count = 0;
1211 bo->priv_flags = 0;
1212 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1213 bo->seq_valid = false;
1214 bo->persistent_swap_storage = persistent_swap_storage;
1215 bo->acc_size = acc_size;
1216 bo->sg = sg;
1217 atomic_inc(&bo->glob->bo_count);
1218
1219 ret = ttm_bo_check_placement(bo, placement);
1220 if (unlikely(ret != 0))
1221 goto out_err;
1222
1223 /*
1224 * For ttm_bo_type_device buffers, allocate
1225 * address space from the device.
1226 */
1227 if (bo->type == ttm_bo_type_device ||
1228 bo->type == ttm_bo_type_sg) {
1229 ret = ttm_bo_setup_vm(bo);
1230 if (ret)
1231 goto out_err;
1232 }
1233
1234 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1235 if (ret)
1236 goto out_err;
1237
1238 ttm_bo_unreserve(bo);
1239 return 0;
1240
1241 out_err:
1242 ttm_bo_unreserve(bo);
1243 ttm_bo_unref(&bo);
1244
1245 return ret;
1246 }
1247 EXPORT_SYMBOL(ttm_bo_init);
1248
1249 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1250 unsigned long bo_size,
1251 unsigned struct_size)
1252 {
1253 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1254 size_t size = 0;
1255
1256 size += ttm_round_pot(struct_size);
1257 size += PAGE_ALIGN(npages * sizeof(void *));
1258 size += ttm_round_pot(sizeof(struct ttm_tt));
1259 return size;
1260 }
1261 EXPORT_SYMBOL(ttm_bo_acc_size);
1262
1263 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1264 unsigned long bo_size,
1265 unsigned struct_size)
1266 {
1267 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1268 size_t size = 0;
1269
1270 size += ttm_round_pot(struct_size);
1271 size += PAGE_ALIGN(npages * sizeof(void *));
1272 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1273 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1274 return size;
1275 }
1276 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1277
1278 int ttm_bo_create(struct ttm_bo_device *bdev,
1279 unsigned long size,
1280 enum ttm_bo_type type,
1281 struct ttm_placement *placement,
1282 uint32_t page_alignment,
1283 bool interruptible,
1284 struct file *persistent_swap_storage,
1285 struct ttm_buffer_object **p_bo)
1286 {
1287 struct ttm_buffer_object *bo;
1288 size_t acc_size;
1289 int ret;
1290
1291 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1292 if (unlikely(bo == NULL))
1293 return -ENOMEM;
1294
1295 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1296 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1297 interruptible, persistent_swap_storage, acc_size,
1298 NULL, NULL);
1299 if (likely(ret == 0))
1300 *p_bo = bo;
1301
1302 return ret;
1303 }
1304 EXPORT_SYMBOL(ttm_bo_create);
1305
1306 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1307 unsigned mem_type, bool allow_errors)
1308 {
1309 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1310 struct ttm_bo_global *glob = bdev->glob;
1311 int ret;
1312
1313 /*
1314 * Can't use standard list traversal since we're unlocking.
1315 */
1316
1317 spin_lock(&glob->lru_lock);
1318 while (!list_empty(&man->lru)) {
1319 spin_unlock(&glob->lru_lock);
1320 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1321 if (ret) {
1322 if (allow_errors) {
1323 return ret;
1324 } else {
1325 pr_err("Cleanup eviction failed\n");
1326 }
1327 }
1328 spin_lock(&glob->lru_lock);
1329 }
1330 spin_unlock(&glob->lru_lock);
1331 return 0;
1332 }
1333
1334 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1335 {
1336 struct ttm_mem_type_manager *man;
1337 int ret = -EINVAL;
1338
1339 if (mem_type >= TTM_NUM_MEM_TYPES) {
1340 pr_err("Illegal memory type %d\n", mem_type);
1341 return ret;
1342 }
1343 man = &bdev->man[mem_type];
1344
1345 if (!man->has_type) {
1346 pr_err("Trying to take down uninitialized memory manager type %u\n",
1347 mem_type);
1348 return ret;
1349 }
1350
1351 man->use_type = false;
1352 man->has_type = false;
1353
1354 ret = 0;
1355 if (mem_type > 0) {
1356 ttm_bo_force_list_clean(bdev, mem_type, false);
1357
1358 ret = (*man->func->takedown)(man);
1359 }
1360
1361 return ret;
1362 }
1363 EXPORT_SYMBOL(ttm_bo_clean_mm);
1364
1365 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1366 {
1367 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1368
1369 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1370 pr_err("Illegal memory manager memory type %u\n", mem_type);
1371 return -EINVAL;
1372 }
1373
1374 if (!man->has_type) {
1375 pr_err("Memory type %u has not been initialized\n", mem_type);
1376 return 0;
1377 }
1378
1379 return ttm_bo_force_list_clean(bdev, mem_type, true);
1380 }
1381 EXPORT_SYMBOL(ttm_bo_evict_mm);
1382
1383 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1384 unsigned long p_size)
1385 {
1386 int ret = -EINVAL;
1387 struct ttm_mem_type_manager *man;
1388
1389 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1390 man = &bdev->man[type];
1391 BUG_ON(man->has_type);
1392 man->io_reserve_fastpath = true;
1393 man->use_io_reserve_lru = false;
1394 mutex_init(&man->io_reserve_mutex);
1395 INIT_LIST_HEAD(&man->io_reserve_lru);
1396
1397 ret = bdev->driver->init_mem_type(bdev, type, man);
1398 if (ret)
1399 return ret;
1400 man->bdev = bdev;
1401
1402 ret = 0;
1403 if (type != TTM_PL_SYSTEM) {
1404 ret = (*man->func->init)(man, p_size);
1405 if (ret)
1406 return ret;
1407 }
1408 man->has_type = true;
1409 man->use_type = true;
1410 man->size = p_size;
1411
1412 INIT_LIST_HEAD(&man->lru);
1413
1414 return 0;
1415 }
1416 EXPORT_SYMBOL(ttm_bo_init_mm);
1417
1418 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1419 {
1420 struct ttm_bo_global *glob =
1421 container_of(kobj, struct ttm_bo_global, kobj);
1422
1423 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1424 __free_page(glob->dummy_read_page);
1425 kfree(glob);
1426 }
1427
1428 void ttm_bo_global_release(struct drm_global_reference *ref)
1429 {
1430 struct ttm_bo_global *glob = ref->object;
1431
1432 kobject_del(&glob->kobj);
1433 kobject_put(&glob->kobj);
1434 }
1435 EXPORT_SYMBOL(ttm_bo_global_release);
1436
1437 int ttm_bo_global_init(struct drm_global_reference *ref)
1438 {
1439 struct ttm_bo_global_ref *bo_ref =
1440 container_of(ref, struct ttm_bo_global_ref, ref);
1441 struct ttm_bo_global *glob = ref->object;
1442 int ret;
1443
1444 mutex_init(&glob->device_list_mutex);
1445 spin_lock_init(&glob->lru_lock);
1446 glob->mem_glob = bo_ref->mem_glob;
1447 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1448
1449 if (unlikely(glob->dummy_read_page == NULL)) {
1450 ret = -ENOMEM;
1451 goto out_no_drp;
1452 }
1453
1454 INIT_LIST_HEAD(&glob->swap_lru);
1455 INIT_LIST_HEAD(&glob->device_list);
1456
1457 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1458 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1459 if (unlikely(ret != 0)) {
1460 pr_err("Could not register buffer object swapout\n");
1461 goto out_no_shrink;
1462 }
1463
1464 atomic_set(&glob->bo_count, 0);
1465
1466 ret = kobject_init_and_add(
1467 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1468 if (unlikely(ret != 0))
1469 kobject_put(&glob->kobj);
1470 return ret;
1471 out_no_shrink:
1472 __free_page(glob->dummy_read_page);
1473 out_no_drp:
1474 kfree(glob);
1475 return ret;
1476 }
1477 EXPORT_SYMBOL(ttm_bo_global_init);
1478
1479
1480 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1481 {
1482 int ret = 0;
1483 unsigned i = TTM_NUM_MEM_TYPES;
1484 struct ttm_mem_type_manager *man;
1485 struct ttm_bo_global *glob = bdev->glob;
1486
1487 while (i--) {
1488 man = &bdev->man[i];
1489 if (man->has_type) {
1490 man->use_type = false;
1491 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1492 ret = -EBUSY;
1493 pr_err("DRM memory manager type %d is not clean\n",
1494 i);
1495 }
1496 man->has_type = false;
1497 }
1498 }
1499
1500 mutex_lock(&glob->device_list_mutex);
1501 list_del(&bdev->device_list);
1502 mutex_unlock(&glob->device_list_mutex);
1503
1504 cancel_delayed_work_sync(&bdev->wq);
1505
1506 while (ttm_bo_delayed_delete(bdev, true))
1507 ;
1508
1509 spin_lock(&glob->lru_lock);
1510 if (list_empty(&bdev->ddestroy))
1511 TTM_DEBUG("Delayed destroy list was clean\n");
1512
1513 if (list_empty(&bdev->man[0].lru))
1514 TTM_DEBUG("Swap list was clean\n");
1515 spin_unlock(&glob->lru_lock);
1516
1517 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1518 write_lock(&bdev->vm_lock);
1519 drm_mm_takedown(&bdev->addr_space_mm);
1520 write_unlock(&bdev->vm_lock);
1521
1522 return ret;
1523 }
1524 EXPORT_SYMBOL(ttm_bo_device_release);
1525
1526 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1527 struct ttm_bo_global *glob,
1528 struct ttm_bo_driver *driver,
1529 uint64_t file_page_offset,
1530 bool need_dma32)
1531 {
1532 int ret = -EINVAL;
1533
1534 rwlock_init(&bdev->vm_lock);
1535 bdev->driver = driver;
1536
1537 memset(bdev->man, 0, sizeof(bdev->man));
1538
1539 /*
1540 * Initialize the system memory buffer type.
1541 * Other types need to be driver / IOCTL initialized.
1542 */
1543 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1544 if (unlikely(ret != 0))
1545 goto out_no_sys;
1546
1547 bdev->addr_space_rb = RB_ROOT;
1548 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1549 if (unlikely(ret != 0))
1550 goto out_no_addr_mm;
1551
1552 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1553 INIT_LIST_HEAD(&bdev->ddestroy);
1554 bdev->dev_mapping = NULL;
1555 bdev->glob = glob;
1556 bdev->need_dma32 = need_dma32;
1557 bdev->val_seq = 0;
1558 spin_lock_init(&bdev->fence_lock);
1559 mutex_lock(&glob->device_list_mutex);
1560 list_add_tail(&bdev->device_list, &glob->device_list);
1561 mutex_unlock(&glob->device_list_mutex);
1562
1563 return 0;
1564 out_no_addr_mm:
1565 ttm_bo_clean_mm(bdev, 0);
1566 out_no_sys:
1567 return ret;
1568 }
1569 EXPORT_SYMBOL(ttm_bo_device_init);
1570
1571 /*
1572 * buffer object vm functions.
1573 */
1574
1575 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1576 {
1577 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1578
1579 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1580 if (mem->mem_type == TTM_PL_SYSTEM)
1581 return false;
1582
1583 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1584 return false;
1585
1586 if (mem->placement & TTM_PL_FLAG_CACHED)
1587 return false;
1588 }
1589 return true;
1590 }
1591
1592 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1593 {
1594 struct ttm_bo_device *bdev = bo->bdev;
1595 loff_t offset = (loff_t) bo->addr_space_offset;
1596 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1597
1598 if (!bdev->dev_mapping)
1599 return;
1600 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1601 ttm_mem_io_free_vm(bo);
1602 }
1603
1604 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1605 {
1606 struct ttm_bo_device *bdev = bo->bdev;
1607 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1608
1609 ttm_mem_io_lock(man, false);
1610 ttm_bo_unmap_virtual_locked(bo);
1611 ttm_mem_io_unlock(man);
1612 }
1613
1614
1615 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1616
1617 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1618 {
1619 struct ttm_bo_device *bdev = bo->bdev;
1620 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1621 struct rb_node *parent = NULL;
1622 struct ttm_buffer_object *cur_bo;
1623 unsigned long offset = bo->vm_node->start;
1624 unsigned long cur_offset;
1625
1626 while (*cur) {
1627 parent = *cur;
1628 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1629 cur_offset = cur_bo->vm_node->start;
1630 if (offset < cur_offset)
1631 cur = &parent->rb_left;
1632 else if (offset > cur_offset)
1633 cur = &parent->rb_right;
1634 else
1635 BUG();
1636 }
1637
1638 rb_link_node(&bo->vm_rb, parent, cur);
1639 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1640 }
1641
1642 /**
1643 * ttm_bo_setup_vm:
1644 *
1645 * @bo: the buffer to allocate address space for
1646 *
1647 * Allocate address space in the drm device so that applications
1648 * can mmap the buffer and access the contents. This only
1649 * applies to ttm_bo_type_device objects as others are not
1650 * placed in the drm device address space.
1651 */
1652
1653 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1654 {
1655 struct ttm_bo_device *bdev = bo->bdev;
1656 int ret;
1657
1658 retry_pre_get:
1659 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1660 if (unlikely(ret != 0))
1661 return ret;
1662
1663 write_lock(&bdev->vm_lock);
1664 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1665 bo->mem.num_pages, 0, 0);
1666
1667 if (unlikely(bo->vm_node == NULL)) {
1668 ret = -ENOMEM;
1669 goto out_unlock;
1670 }
1671
1672 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1673 bo->mem.num_pages, 0);
1674
1675 if (unlikely(bo->vm_node == NULL)) {
1676 write_unlock(&bdev->vm_lock);
1677 goto retry_pre_get;
1678 }
1679
1680 ttm_bo_vm_insert_rb(bo);
1681 write_unlock(&bdev->vm_lock);
1682 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1683
1684 return 0;
1685 out_unlock:
1686 write_unlock(&bdev->vm_lock);
1687 return ret;
1688 }
1689
1690 int ttm_bo_wait(struct ttm_buffer_object *bo,
1691 bool lazy, bool interruptible, bool no_wait)
1692 {
1693 struct ttm_bo_driver *driver = bo->bdev->driver;
1694 struct ttm_bo_device *bdev = bo->bdev;
1695 void *sync_obj;
1696 int ret = 0;
1697
1698 if (likely(bo->sync_obj == NULL))
1699 return 0;
1700
1701 while (bo->sync_obj) {
1702
1703 if (driver->sync_obj_signaled(bo->sync_obj)) {
1704 void *tmp_obj = bo->sync_obj;
1705 bo->sync_obj = NULL;
1706 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1707 spin_unlock(&bdev->fence_lock);
1708 driver->sync_obj_unref(&tmp_obj);
1709 spin_lock(&bdev->fence_lock);
1710 continue;
1711 }
1712
1713 if (no_wait)
1714 return -EBUSY;
1715
1716 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1717 spin_unlock(&bdev->fence_lock);
1718 ret = driver->sync_obj_wait(sync_obj,
1719 lazy, interruptible);
1720 if (unlikely(ret != 0)) {
1721 driver->sync_obj_unref(&sync_obj);
1722 spin_lock(&bdev->fence_lock);
1723 return ret;
1724 }
1725 spin_lock(&bdev->fence_lock);
1726 if (likely(bo->sync_obj == sync_obj)) {
1727 void *tmp_obj = bo->sync_obj;
1728 bo->sync_obj = NULL;
1729 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1730 &bo->priv_flags);
1731 spin_unlock(&bdev->fence_lock);
1732 driver->sync_obj_unref(&sync_obj);
1733 driver->sync_obj_unref(&tmp_obj);
1734 spin_lock(&bdev->fence_lock);
1735 } else {
1736 spin_unlock(&bdev->fence_lock);
1737 driver->sync_obj_unref(&sync_obj);
1738 spin_lock(&bdev->fence_lock);
1739 }
1740 }
1741 return 0;
1742 }
1743 EXPORT_SYMBOL(ttm_bo_wait);
1744
1745 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1746 {
1747 struct ttm_bo_device *bdev = bo->bdev;
1748 int ret = 0;
1749
1750 /*
1751 * Using ttm_bo_reserve makes sure the lru lists are updated.
1752 */
1753
1754 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1755 if (unlikely(ret != 0))
1756 return ret;
1757 spin_lock(&bdev->fence_lock);
1758 ret = ttm_bo_wait(bo, false, true, no_wait);
1759 spin_unlock(&bdev->fence_lock);
1760 if (likely(ret == 0))
1761 atomic_inc(&bo->cpu_writers);
1762 ttm_bo_unreserve(bo);
1763 return ret;
1764 }
1765 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1766
1767 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1768 {
1769 atomic_dec(&bo->cpu_writers);
1770 }
1771 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1772
1773 /**
1774 * A buffer object shrink method that tries to swap out the first
1775 * buffer object on the bo_global::swap_lru list.
1776 */
1777
1778 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1779 {
1780 struct ttm_bo_global *glob =
1781 container_of(shrink, struct ttm_bo_global, shrink);
1782 struct ttm_buffer_object *bo;
1783 int ret = -EBUSY;
1784 int put_count;
1785 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1786
1787 spin_lock(&glob->lru_lock);
1788 while (ret == -EBUSY) {
1789 if (unlikely(list_empty(&glob->swap_lru))) {
1790 spin_unlock(&glob->lru_lock);
1791 return -EBUSY;
1792 }
1793
1794 bo = list_first_entry(&glob->swap_lru,
1795 struct ttm_buffer_object, swap);
1796 kref_get(&bo->list_kref);
1797
1798 if (!list_empty(&bo->ddestroy)) {
1799 spin_unlock(&glob->lru_lock);
1800 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1801 kref_put(&bo->list_kref, ttm_bo_release_list);
1802 spin_lock(&glob->lru_lock);
1803 continue;
1804 }
1805
1806 /**
1807 * Reserve buffer. Since we unlock while sleeping, we need
1808 * to re-check that nobody removed us from the swap-list while
1809 * we slept.
1810 */
1811
1812 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1813 if (unlikely(ret == -EBUSY)) {
1814 spin_unlock(&glob->lru_lock);
1815 ttm_bo_wait_unreserved(bo, false);
1816 kref_put(&bo->list_kref, ttm_bo_release_list);
1817 spin_lock(&glob->lru_lock);
1818 }
1819 }
1820
1821 BUG_ON(ret != 0);
1822 put_count = ttm_bo_del_from_lru(bo);
1823 spin_unlock(&glob->lru_lock);
1824
1825 ttm_bo_list_ref_sub(bo, put_count, true);
1826
1827 /**
1828 * Wait for GPU, then move to system cached.
1829 */
1830
1831 spin_lock(&bo->bdev->fence_lock);
1832 ret = ttm_bo_wait(bo, false, false, false);
1833 spin_unlock(&bo->bdev->fence_lock);
1834
1835 if (unlikely(ret != 0))
1836 goto out;
1837
1838 if ((bo->mem.placement & swap_placement) != swap_placement) {
1839 struct ttm_mem_reg evict_mem;
1840
1841 evict_mem = bo->mem;
1842 evict_mem.mm_node = NULL;
1843 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1844 evict_mem.mem_type = TTM_PL_SYSTEM;
1845
1846 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1847 false, false, false);
1848 if (unlikely(ret != 0))
1849 goto out;
1850 }
1851
1852 ttm_bo_unmap_virtual(bo);
1853
1854 /**
1855 * Swap out. Buffer will be swapped in again as soon as
1856 * anyone tries to access a ttm page.
1857 */
1858
1859 if (bo->bdev->driver->swap_notify)
1860 bo->bdev->driver->swap_notify(bo);
1861
1862 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1863 out:
1864
1865 /**
1866 *
1867 * Unreserve without putting on LRU to avoid swapping out an
1868 * already swapped buffer.
1869 */
1870
1871 atomic_set(&bo->reserved, 0);
1872 wake_up_all(&bo->event_queue);
1873 kref_put(&bo->list_kref, ttm_bo_release_list);
1874 return ret;
1875 }
1876
1877 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1878 {
1879 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1880 ;
1881 }
1882 EXPORT_SYMBOL(ttm_bo_swapout_all);
Linux kernel - Deliverables
This page took 0.099999 seconds and 5 git commands to generate.