?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
drm/ttm: Fix spinlock imbalance
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / drivers / gpu / drm / ttm / ttm_bo.c
blobb67cfcaa661f87bffbb84d3f0ed9786efb591e5a
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 "ttm/ttm_module.h"
34 #include "ttm/ttm_bo_driver.h"
35 #include "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 atomic_read(&bo->reserved) == 0);
166 } else {
167 wait_event(bo->event_queue, atomic_read(&bo->reserved) == 0);
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(!atomic_read(&bo->reserved));
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 spin_lock(&glob->lru_lock);
584
585 if (unlikely(list_empty(&bo->ddestroy))) {
586 spin_unlock(&glob->lru_lock);
587 return 0;
588 }
589
590 ret = ttm_bo_reserve_locked(bo, interruptible,
591 no_wait_reserve, false, 0);
592
593 if (unlikely(ret != 0)) {
594 spin_unlock(&glob->lru_lock);
595 return ret;
596 }
597
598 /**
599 * We can re-check for sync object without taking
600 * the bo::lock since setting the sync object requires
601 * also bo::reserved. A busy object at this point may
602 * be caused by another thread recently starting an accelerated
603 * eviction.
604 */
605
606 if (unlikely(bo->sync_obj)) {
607 atomic_set(&bo->reserved, 0);
608 wake_up_all(&bo->event_queue);
609 spin_unlock(&glob->lru_lock);
610 goto retry;
611 }
612
613 put_count = ttm_bo_del_from_lru(bo);
614 list_del_init(&bo->ddestroy);
615 ++put_count;
616
617 spin_unlock(&glob->lru_lock);
618 ttm_bo_cleanup_memtype_use(bo);
619
620 ttm_bo_list_ref_sub(bo, put_count, true);
621
622 return 0;
623 }
624
625 /**
626 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
627 * encountered buffers.
628 */
629
630 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
631 {
632 struct ttm_bo_global *glob = bdev->glob;
633 struct ttm_buffer_object *entry = NULL;
634 int ret = 0;
635
636 spin_lock(&glob->lru_lock);
637 if (list_empty(&bdev->ddestroy))
638 goto out_unlock;
639
640 entry = list_first_entry(&bdev->ddestroy,
641 struct ttm_buffer_object, ddestroy);
642 kref_get(&entry->list_kref);
643
644 for (;;) {
645 struct ttm_buffer_object *nentry = NULL;
646
647 if (entry->ddestroy.next != &bdev->ddestroy) {
648 nentry = list_first_entry(&entry->ddestroy,
649 struct ttm_buffer_object, ddestroy);
650 kref_get(&nentry->list_kref);
651 }
652
653 spin_unlock(&glob->lru_lock);
654 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
655 !remove_all);
656 kref_put(&entry->list_kref, ttm_bo_release_list);
657 entry = nentry;
658
659 if (ret || !entry)
660 goto out;
661
662 spin_lock(&glob->lru_lock);
663 if (list_empty(&entry->ddestroy))
664 break;
665 }
666
667 out_unlock:
668 spin_unlock(&glob->lru_lock);
669 out:
670 if (entry)
671 kref_put(&entry->list_kref, ttm_bo_release_list);
672 return ret;
673 }
674
675 static void ttm_bo_delayed_workqueue(struct work_struct *work)
676 {
677 struct ttm_bo_device *bdev =
678 container_of(work, struct ttm_bo_device, wq.work);
679
680 if (ttm_bo_delayed_delete(bdev, false)) {
681 schedule_delayed_work(&bdev->wq,
682 ((HZ / 100) < 1) ? 1 : HZ / 100);
683 }
684 }
685
686 static void ttm_bo_release(struct kref *kref)
687 {
688 struct ttm_buffer_object *bo =
689 container_of(kref, struct ttm_buffer_object, kref);
690 struct ttm_bo_device *bdev = bo->bdev;
691 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
692
693 if (likely(bo->vm_node != NULL)) {
694 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
695 drm_mm_put_block(bo->vm_node);
696 bo->vm_node = NULL;
697 }
698 write_unlock(&bdev->vm_lock);
699 ttm_mem_io_lock(man, false);
700 ttm_mem_io_free_vm(bo);
701 ttm_mem_io_unlock(man);
702 ttm_bo_cleanup_refs_or_queue(bo);
703 kref_put(&bo->list_kref, ttm_bo_release_list);
704 write_lock(&bdev->vm_lock);
705 }
706
707 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
708 {
709 struct ttm_buffer_object *bo = *p_bo;
710 struct ttm_bo_device *bdev = bo->bdev;
711
712 *p_bo = NULL;
713 write_lock(&bdev->vm_lock);
714 kref_put(&bo->kref, ttm_bo_release);
715 write_unlock(&bdev->vm_lock);
716 }
717 EXPORT_SYMBOL(ttm_bo_unref);
718
719 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
720 {
721 return cancel_delayed_work_sync(&bdev->wq);
722 }
723 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
724
725 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
726 {
727 if (resched)
728 schedule_delayed_work(&bdev->wq,
729 ((HZ / 100) < 1) ? 1 : HZ / 100);
730 }
731 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
732
733 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
734 bool no_wait_reserve, bool no_wait_gpu)
735 {
736 struct ttm_bo_device *bdev = bo->bdev;
737 struct ttm_mem_reg evict_mem;
738 struct ttm_placement placement;
739 int ret = 0;
740
741 spin_lock(&bdev->fence_lock);
742 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
743 spin_unlock(&bdev->fence_lock);
744
745 if (unlikely(ret != 0)) {
746 if (ret != -ERESTARTSYS) {
747 pr_err("Failed to expire sync object before buffer eviction\n");
748 }
749 goto out;
750 }
751
752 BUG_ON(!atomic_read(&bo->reserved));
753
754 evict_mem = bo->mem;
755 evict_mem.mm_node = NULL;
756 evict_mem.bus.io_reserved_vm = false;
757 evict_mem.bus.io_reserved_count = 0;
758
759 placement.fpfn = 0;
760 placement.lpfn = 0;
761 placement.num_placement = 0;
762 placement.num_busy_placement = 0;
763 bdev->driver->evict_flags(bo, &placement);
764 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
765 no_wait_reserve, no_wait_gpu);
766 if (ret) {
767 if (ret != -ERESTARTSYS) {
768 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
769 bo);
770 ttm_bo_mem_space_debug(bo, &placement);
771 }
772 goto out;
773 }
774
775 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
776 no_wait_reserve, no_wait_gpu);
777 if (ret) {
778 if (ret != -ERESTARTSYS)
779 pr_err("Buffer eviction failed\n");
780 ttm_bo_mem_put(bo, &evict_mem);
781 goto out;
782 }
783 bo->evicted = true;
784 out:
785 return ret;
786 }
787
788 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
789 uint32_t mem_type,
790 bool interruptible, bool no_wait_reserve,
791 bool no_wait_gpu)
792 {
793 struct ttm_bo_global *glob = bdev->glob;
794 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
795 struct ttm_buffer_object *bo;
796 int ret, put_count = 0;
797
798 retry:
799 spin_lock(&glob->lru_lock);
800 if (list_empty(&man->lru)) {
801 spin_unlock(&glob->lru_lock);
802 return -EBUSY;
803 }
804
805 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
806 kref_get(&bo->list_kref);
807
808 if (!list_empty(&bo->ddestroy)) {
809 spin_unlock(&glob->lru_lock);
810 ret = ttm_bo_cleanup_refs(bo, interruptible,
811 no_wait_reserve, no_wait_gpu);
812 kref_put(&bo->list_kref, ttm_bo_release_list);
813
814 if (likely(ret == 0 || ret == -ERESTARTSYS))
815 return ret;
816
817 goto retry;
818 }
819
820 ret = ttm_bo_reserve_locked(bo, false, no_wait_reserve, false, 0);
821
822 if (unlikely(ret == -EBUSY)) {
823 spin_unlock(&glob->lru_lock);
824 if (likely(!no_wait_gpu))
825 ret = ttm_bo_wait_unreserved(bo, interruptible);
826
827 kref_put(&bo->list_kref, ttm_bo_release_list);
828
829 /**
830 * We *need* to retry after releasing the lru lock.
831 */
832
833 if (unlikely(ret != 0))
834 return ret;
835 goto retry;
836 }
837
838 put_count = ttm_bo_del_from_lru(bo);
839 spin_unlock(&glob->lru_lock);
840
841 BUG_ON(ret != 0);
842
843 ttm_bo_list_ref_sub(bo, put_count, true);
844
845 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
846 ttm_bo_unreserve(bo);
847
848 kref_put(&bo->list_kref, ttm_bo_release_list);
849 return ret;
850 }
851
852 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
853 {
854 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
855
856 if (mem->mm_node)
857 (*man->func->put_node)(man, mem);
858 }
859 EXPORT_SYMBOL(ttm_bo_mem_put);
860
861 /**
862 * Repeatedly evict memory from the LRU for @mem_type until we create enough
863 * space, or we've evicted everything and there isn't enough space.
864 */
865 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
866 uint32_t mem_type,
867 struct ttm_placement *placement,
868 struct ttm_mem_reg *mem,
869 bool interruptible,
870 bool no_wait_reserve,
871 bool no_wait_gpu)
872 {
873 struct ttm_bo_device *bdev = bo->bdev;
874 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
875 int ret;
876
877 do {
878 ret = (*man->func->get_node)(man, bo, placement, mem);
879 if (unlikely(ret != 0))
880 return ret;
881 if (mem->mm_node)
882 break;
883 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
884 no_wait_reserve, no_wait_gpu);
885 if (unlikely(ret != 0))
886 return ret;
887 } while (1);
888 if (mem->mm_node == NULL)
889 return -ENOMEM;
890 mem->mem_type = mem_type;
891 return 0;
892 }
893
894 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
895 uint32_t cur_placement,
896 uint32_t proposed_placement)
897 {
898 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
899 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
900
901 /**
902 * Keep current caching if possible.
903 */
904
905 if ((cur_placement & caching) != 0)
906 result |= (cur_placement & caching);
907 else if ((man->default_caching & caching) != 0)
908 result |= man->default_caching;
909 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
910 result |= TTM_PL_FLAG_CACHED;
911 else if ((TTM_PL_FLAG_WC & caching) != 0)
912 result |= TTM_PL_FLAG_WC;
913 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
914 result |= TTM_PL_FLAG_UNCACHED;
915
916 return result;
917 }
918
919 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
920 uint32_t mem_type,
921 uint32_t proposed_placement,
922 uint32_t *masked_placement)
923 {
924 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
925
926 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
927 return false;
928
929 if ((proposed_placement & man->available_caching) == 0)
930 return false;
931
932 cur_flags |= (proposed_placement & man->available_caching);
933
934 *masked_placement = cur_flags;
935 return true;
936 }
937
938 /**
939 * Creates space for memory region @mem according to its type.
940 *
941 * This function first searches for free space in compatible memory types in
942 * the priority order defined by the driver. If free space isn't found, then
943 * ttm_bo_mem_force_space is attempted in priority order to evict and find
944 * space.
945 */
946 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
947 struct ttm_placement *placement,
948 struct ttm_mem_reg *mem,
949 bool interruptible, bool no_wait_reserve,
950 bool no_wait_gpu)
951 {
952 struct ttm_bo_device *bdev = bo->bdev;
953 struct ttm_mem_type_manager *man;
954 uint32_t mem_type = TTM_PL_SYSTEM;
955 uint32_t cur_flags = 0;
956 bool type_found = false;
957 bool type_ok = false;
958 bool has_erestartsys = false;
959 int i, ret;
960
961 mem->mm_node = NULL;
962 for (i = 0; i < placement->num_placement; ++i) {
963 ret = ttm_mem_type_from_flags(placement->placement[i],
964 &mem_type);
965 if (ret)
966 return ret;
967 man = &bdev->man[mem_type];
968
969 type_ok = ttm_bo_mt_compatible(man,
970 mem_type,
971 placement->placement[i],
972 &cur_flags);
973
974 if (!type_ok)
975 continue;
976
977 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
978 cur_flags);
979 /*
980 * Use the access and other non-mapping-related flag bits from
981 * the memory placement flags to the current flags
982 */
983 ttm_flag_masked(&cur_flags, placement->placement[i],
984 ~TTM_PL_MASK_MEMTYPE);
985
986 if (mem_type == TTM_PL_SYSTEM)
987 break;
988
989 if (man->has_type && man->use_type) {
990 type_found = true;
991 ret = (*man->func->get_node)(man, bo, placement, mem);
992 if (unlikely(ret))
993 return ret;
994 }
995 if (mem->mm_node)
996 break;
997 }
998
999 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
1000 mem->mem_type = mem_type;
1001 mem->placement = cur_flags;
1002 return 0;
1003 }
1004
1005 if (!type_found)
1006 return -EINVAL;
1007
1008 for (i = 0; i < placement->num_busy_placement; ++i) {
1009 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
1010 &mem_type);
1011 if (ret)
1012 return ret;
1013 man = &bdev->man[mem_type];
1014 if (!man->has_type)
1015 continue;
1016 if (!ttm_bo_mt_compatible(man,
1017 mem_type,
1018 placement->busy_placement[i],
1019 &cur_flags))
1020 continue;
1021
1022 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1023 cur_flags);
1024 /*
1025 * Use the access and other non-mapping-related flag bits from
1026 * the memory placement flags to the current flags
1027 */
1028 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1029 ~TTM_PL_MASK_MEMTYPE);
1030
1031
1032 if (mem_type == TTM_PL_SYSTEM) {
1033 mem->mem_type = mem_type;
1034 mem->placement = cur_flags;
1035 mem->mm_node = NULL;
1036 return 0;
1037 }
1038
1039 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1040 interruptible, no_wait_reserve, no_wait_gpu);
1041 if (ret == 0 && mem->mm_node) {
1042 mem->placement = cur_flags;
1043 return 0;
1044 }
1045 if (ret == -ERESTARTSYS)
1046 has_erestartsys = true;
1047 }
1048 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1049 return ret;
1050 }
1051 EXPORT_SYMBOL(ttm_bo_mem_space);
1052
1053 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1054 {
1055 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1056 return -EBUSY;
1057
1058 return wait_event_interruptible(bo->event_queue,
1059 atomic_read(&bo->cpu_writers) == 0);
1060 }
1061 EXPORT_SYMBOL(ttm_bo_wait_cpu);
1062
1063 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1064 struct ttm_placement *placement,
1065 bool interruptible, bool no_wait_reserve,
1066 bool no_wait_gpu)
1067 {
1068 int ret = 0;
1069 struct ttm_mem_reg mem;
1070 struct ttm_bo_device *bdev = bo->bdev;
1071
1072 BUG_ON(!atomic_read(&bo->reserved));
1073
1074 /*
1075 * FIXME: It's possible to pipeline buffer moves.
1076 * Have the driver move function wait for idle when necessary,
1077 * instead of doing it here.
1078 */
1079 spin_lock(&bdev->fence_lock);
1080 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1081 spin_unlock(&bdev->fence_lock);
1082 if (ret)
1083 return ret;
1084 mem.num_pages = bo->num_pages;
1085 mem.size = mem.num_pages << PAGE_SHIFT;
1086 mem.page_alignment = bo->mem.page_alignment;
1087 mem.bus.io_reserved_vm = false;
1088 mem.bus.io_reserved_count = 0;
1089 /*
1090 * Determine where to move the buffer.
1091 */
1092 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1093 if (ret)
1094 goto out_unlock;
1095 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1096 out_unlock:
1097 if (ret && mem.mm_node)
1098 ttm_bo_mem_put(bo, &mem);
1099 return ret;
1100 }
1101
1102 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1103 struct ttm_mem_reg *mem)
1104 {
1105 int i;
1106
1107 if (mem->mm_node && placement->lpfn != 0 &&
1108 (mem->start < placement->fpfn ||
1109 mem->start + mem->num_pages > placement->lpfn))
1110 return -1;
1111
1112 for (i = 0; i < placement->num_placement; i++) {
1113 if ((placement->placement[i] & mem->placement &
1114 TTM_PL_MASK_CACHING) &&
1115 (placement->placement[i] & mem->placement &
1116 TTM_PL_MASK_MEM))
1117 return i;
1118 }
1119 return -1;
1120 }
1121
1122 int ttm_bo_validate(struct ttm_buffer_object *bo,
1123 struct ttm_placement *placement,
1124 bool interruptible, bool no_wait_reserve,
1125 bool no_wait_gpu)
1126 {
1127 int ret;
1128
1129 BUG_ON(!atomic_read(&bo->reserved));
1130 /* Check that range is valid */
1131 if (placement->lpfn || placement->fpfn)
1132 if (placement->fpfn > placement->lpfn ||
1133 (placement->lpfn - placement->fpfn) < bo->num_pages)
1134 return -EINVAL;
1135 /*
1136 * Check whether we need to move buffer.
1137 */
1138 ret = ttm_bo_mem_compat(placement, &bo->mem);
1139 if (ret < 0) {
1140 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1141 if (ret)
1142 return ret;
1143 } else {
1144 /*
1145 * Use the access and other non-mapping-related flag bits from
1146 * the compatible memory placement flags to the active flags
1147 */
1148 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1149 ~TTM_PL_MASK_MEMTYPE);
1150 }
1151 /*
1152 * We might need to add a TTM.
1153 */
1154 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1155 ret = ttm_bo_add_ttm(bo, true);
1156 if (ret)
1157 return ret;
1158 }
1159 return 0;
1160 }
1161 EXPORT_SYMBOL(ttm_bo_validate);
1162
1163 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1164 struct ttm_placement *placement)
1165 {
1166 BUG_ON((placement->fpfn || placement->lpfn) &&
1167 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1168
1169 return 0;
1170 }
1171
1172 int ttm_bo_init(struct ttm_bo_device *bdev,
1173 struct ttm_buffer_object *bo,
1174 unsigned long size,
1175 enum ttm_bo_type type,
1176 struct ttm_placement *placement,
1177 uint32_t page_alignment,
1178 unsigned long buffer_start,
1179 bool interruptible,
1180 struct file *persistent_swap_storage,
1181 size_t acc_size,
1182 struct sg_table *sg,
1183 void (*destroy) (struct ttm_buffer_object *))
1184 {
1185 int ret = 0;
1186 unsigned long num_pages;
1187 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1188
1189 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1190 if (ret) {
1191 pr_err("Out of kernel memory\n");
1192 if (destroy)
1193 (*destroy)(bo);
1194 else
1195 kfree(bo);
1196 return -ENOMEM;
1197 }
1198
1199 size += buffer_start & ~PAGE_MASK;
1200 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1201 if (num_pages == 0) {
1202 pr_err("Illegal buffer object size\n");
1203 if (destroy)
1204 (*destroy)(bo);
1205 else
1206 kfree(bo);
1207 return -EINVAL;
1208 }
1209 bo->destroy = destroy;
1210
1211 kref_init(&bo->kref);
1212 kref_init(&bo->list_kref);
1213 atomic_set(&bo->cpu_writers, 0);
1214 atomic_set(&bo->reserved, 1);
1215 init_waitqueue_head(&bo->event_queue);
1216 INIT_LIST_HEAD(&bo->lru);
1217 INIT_LIST_HEAD(&bo->ddestroy);
1218 INIT_LIST_HEAD(&bo->swap);
1219 INIT_LIST_HEAD(&bo->io_reserve_lru);
1220 bo->bdev = bdev;
1221 bo->glob = bdev->glob;
1222 bo->type = type;
1223 bo->num_pages = num_pages;
1224 bo->mem.size = num_pages << PAGE_SHIFT;
1225 bo->mem.mem_type = TTM_PL_SYSTEM;
1226 bo->mem.num_pages = bo->num_pages;
1227 bo->mem.mm_node = NULL;
1228 bo->mem.page_alignment = page_alignment;
1229 bo->mem.bus.io_reserved_vm = false;
1230 bo->mem.bus.io_reserved_count = 0;
1231 bo->buffer_start = buffer_start & PAGE_MASK;
1232 bo->priv_flags = 0;
1233 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1234 bo->seq_valid = false;
1235 bo->persistent_swap_storage = persistent_swap_storage;
1236 bo->acc_size = acc_size;
1237 bo->sg = sg;
1238 atomic_inc(&bo->glob->bo_count);
1239
1240 ret = ttm_bo_check_placement(bo, placement);
1241 if (unlikely(ret != 0))
1242 goto out_err;
1243
1244 /*
1245 * For ttm_bo_type_device buffers, allocate
1246 * address space from the device.
1247 */
1248 if (bo->type == ttm_bo_type_device ||
1249 bo->type == ttm_bo_type_sg) {
1250 ret = ttm_bo_setup_vm(bo);
1251 if (ret)
1252 goto out_err;
1253 }
1254
1255 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1256 if (ret)
1257 goto out_err;
1258
1259 ttm_bo_unreserve(bo);
1260 return 0;
1261
1262 out_err:
1263 ttm_bo_unreserve(bo);
1264 ttm_bo_unref(&bo);
1265
1266 return ret;
1267 }
1268 EXPORT_SYMBOL(ttm_bo_init);
1269
1270 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1271 unsigned long bo_size,
1272 unsigned struct_size)
1273 {
1274 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1275 size_t size = 0;
1276
1277 size += ttm_round_pot(struct_size);
1278 size += PAGE_ALIGN(npages * sizeof(void *));
1279 size += ttm_round_pot(sizeof(struct ttm_tt));
1280 return size;
1281 }
1282 EXPORT_SYMBOL(ttm_bo_acc_size);
1283
1284 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1285 unsigned long bo_size,
1286 unsigned struct_size)
1287 {
1288 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1289 size_t size = 0;
1290
1291 size += ttm_round_pot(struct_size);
1292 size += PAGE_ALIGN(npages * sizeof(void *));
1293 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1294 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1295 return size;
1296 }
1297 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1298
1299 int ttm_bo_create(struct ttm_bo_device *bdev,
1300 unsigned long size,
1301 enum ttm_bo_type type,
1302 struct ttm_placement *placement,
1303 uint32_t page_alignment,
1304 unsigned long buffer_start,
1305 bool interruptible,
1306 struct file *persistent_swap_storage,
1307 struct ttm_buffer_object **p_bo)
1308 {
1309 struct ttm_buffer_object *bo;
1310 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1311 size_t acc_size;
1312 int ret;
1313
1314 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1315 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1316 if (unlikely(ret != 0))
1317 return ret;
1318
1319 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1320
1321 if (unlikely(bo == NULL)) {
1322 ttm_mem_global_free(mem_glob, acc_size);
1323 return -ENOMEM;
1324 }
1325
1326 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1327 buffer_start, interruptible,
1328 persistent_swap_storage, acc_size, NULL, NULL);
1329 if (likely(ret == 0))
1330 *p_bo = bo;
1331
1332 return ret;
1333 }
1334 EXPORT_SYMBOL(ttm_bo_create);
1335
1336 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1337 unsigned mem_type, bool allow_errors)
1338 {
1339 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1340 struct ttm_bo_global *glob = bdev->glob;
1341 int ret;
1342
1343 /*
1344 * Can't use standard list traversal since we're unlocking.
1345 */
1346
1347 spin_lock(&glob->lru_lock);
1348 while (!list_empty(&man->lru)) {
1349 spin_unlock(&glob->lru_lock);
1350 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1351 if (ret) {
1352 if (allow_errors) {
1353 return ret;
1354 } else {
1355 pr_err("Cleanup eviction failed\n");
1356 }
1357 }
1358 spin_lock(&glob->lru_lock);
1359 }
1360 spin_unlock(&glob->lru_lock);
1361 return 0;
1362 }
1363
1364 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1365 {
1366 struct ttm_mem_type_manager *man;
1367 int ret = -EINVAL;
1368
1369 if (mem_type >= TTM_NUM_MEM_TYPES) {
1370 pr_err("Illegal memory type %d\n", mem_type);
1371 return ret;
1372 }
1373 man = &bdev->man[mem_type];
1374
1375 if (!man->has_type) {
1376 pr_err("Trying to take down uninitialized memory manager type %u\n",
1377 mem_type);
1378 return ret;
1379 }
1380
1381 man->use_type = false;
1382 man->has_type = false;
1383
1384 ret = 0;
1385 if (mem_type > 0) {
1386 ttm_bo_force_list_clean(bdev, mem_type, false);
1387
1388 ret = (*man->func->takedown)(man);
1389 }
1390
1391 return ret;
1392 }
1393 EXPORT_SYMBOL(ttm_bo_clean_mm);
1394
1395 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1396 {
1397 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1398
1399 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1400 pr_err("Illegal memory manager memory type %u\n", mem_type);
1401 return -EINVAL;
1402 }
1403
1404 if (!man->has_type) {
1405 pr_err("Memory type %u has not been initialized\n", mem_type);
1406 return 0;
1407 }
1408
1409 return ttm_bo_force_list_clean(bdev, mem_type, true);
1410 }
1411 EXPORT_SYMBOL(ttm_bo_evict_mm);
1412
1413 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1414 unsigned long p_size)
1415 {
1416 int ret = -EINVAL;
1417 struct ttm_mem_type_manager *man;
1418
1419 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1420 man = &bdev->man[type];
1421 BUG_ON(man->has_type);
1422 man->io_reserve_fastpath = true;
1423 man->use_io_reserve_lru = false;
1424 mutex_init(&man->io_reserve_mutex);
1425 INIT_LIST_HEAD(&man->io_reserve_lru);
1426
1427 ret = bdev->driver->init_mem_type(bdev, type, man);
1428 if (ret)
1429 return ret;
1430 man->bdev = bdev;
1431
1432 ret = 0;
1433 if (type != TTM_PL_SYSTEM) {
1434 ret = (*man->func->init)(man, p_size);
1435 if (ret)
1436 return ret;
1437 }
1438 man->has_type = true;
1439 man->use_type = true;
1440 man->size = p_size;
1441
1442 INIT_LIST_HEAD(&man->lru);
1443
1444 return 0;
1445 }
1446 EXPORT_SYMBOL(ttm_bo_init_mm);
1447
1448 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1449 {
1450 struct ttm_bo_global *glob =
1451 container_of(kobj, struct ttm_bo_global, kobj);
1452
1453 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1454 __free_page(glob->dummy_read_page);
1455 kfree(glob);
1456 }
1457
1458 void ttm_bo_global_release(struct drm_global_reference *ref)
1459 {
1460 struct ttm_bo_global *glob = ref->object;
1461
1462 kobject_del(&glob->kobj);
1463 kobject_put(&glob->kobj);
1464 }
1465 EXPORT_SYMBOL(ttm_bo_global_release);
1466
1467 int ttm_bo_global_init(struct drm_global_reference *ref)
1468 {
1469 struct ttm_bo_global_ref *bo_ref =
1470 container_of(ref, struct ttm_bo_global_ref, ref);
1471 struct ttm_bo_global *glob = ref->object;
1472 int ret;
1473
1474 mutex_init(&glob->device_list_mutex);
1475 spin_lock_init(&glob->lru_lock);
1476 glob->mem_glob = bo_ref->mem_glob;
1477 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1478
1479 if (unlikely(glob->dummy_read_page == NULL)) {
1480 ret = -ENOMEM;
1481 goto out_no_drp;
1482 }
1483
1484 INIT_LIST_HEAD(&glob->swap_lru);
1485 INIT_LIST_HEAD(&glob->device_list);
1486
1487 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1488 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1489 if (unlikely(ret != 0)) {
1490 pr_err("Could not register buffer object swapout\n");
1491 goto out_no_shrink;
1492 }
1493
1494 atomic_set(&glob->bo_count, 0);
1495
1496 ret = kobject_init_and_add(
1497 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1498 if (unlikely(ret != 0))
1499 kobject_put(&glob->kobj);
1500 return ret;
1501 out_no_shrink:
1502 __free_page(glob->dummy_read_page);
1503 out_no_drp:
1504 kfree(glob);
1505 return ret;
1506 }
1507 EXPORT_SYMBOL(ttm_bo_global_init);
1508
1509
1510 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1511 {
1512 int ret = 0;
1513 unsigned i = TTM_NUM_MEM_TYPES;
1514 struct ttm_mem_type_manager *man;
1515 struct ttm_bo_global *glob = bdev->glob;
1516
1517 while (i--) {
1518 man = &bdev->man[i];
1519 if (man->has_type) {
1520 man->use_type = false;
1521 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1522 ret = -EBUSY;
1523 pr_err("DRM memory manager type %d is not clean\n",
1524 i);
1525 }
1526 man->has_type = false;
1527 }
1528 }
1529
1530 mutex_lock(&glob->device_list_mutex);
1531 list_del(&bdev->device_list);
1532 mutex_unlock(&glob->device_list_mutex);
1533
1534 cancel_delayed_work_sync(&bdev->wq);
1535
1536 while (ttm_bo_delayed_delete(bdev, true))
1537 ;
1538
1539 spin_lock(&glob->lru_lock);
1540 if (list_empty(&bdev->ddestroy))
1541 TTM_DEBUG("Delayed destroy list was clean\n");
1542
1543 if (list_empty(&bdev->man[0].lru))
1544 TTM_DEBUG("Swap list was clean\n");
1545 spin_unlock(&glob->lru_lock);
1546
1547 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1548 write_lock(&bdev->vm_lock);
1549 drm_mm_takedown(&bdev->addr_space_mm);
1550 write_unlock(&bdev->vm_lock);
1551
1552 return ret;
1553 }
1554 EXPORT_SYMBOL(ttm_bo_device_release);
1555
1556 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1557 struct ttm_bo_global *glob,
1558 struct ttm_bo_driver *driver,
1559 uint64_t file_page_offset,
1560 bool need_dma32)
1561 {
1562 int ret = -EINVAL;
1563
1564 rwlock_init(&bdev->vm_lock);
1565 bdev->driver = driver;
1566
1567 memset(bdev->man, 0, sizeof(bdev->man));
1568
1569 /*
1570 * Initialize the system memory buffer type.
1571 * Other types need to be driver / IOCTL initialized.
1572 */
1573 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1574 if (unlikely(ret != 0))
1575 goto out_no_sys;
1576
1577 bdev->addr_space_rb = RB_ROOT;
1578 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1579 if (unlikely(ret != 0))
1580 goto out_no_addr_mm;
1581
1582 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1583 bdev->nice_mode = true;
1584 INIT_LIST_HEAD(&bdev->ddestroy);
1585 bdev->dev_mapping = NULL;
1586 bdev->glob = glob;
1587 bdev->need_dma32 = need_dma32;
1588 bdev->val_seq = 0;
1589 spin_lock_init(&bdev->fence_lock);
1590 mutex_lock(&glob->device_list_mutex);
1591 list_add_tail(&bdev->device_list, &glob->device_list);
1592 mutex_unlock(&glob->device_list_mutex);
1593
1594 return 0;
1595 out_no_addr_mm:
1596 ttm_bo_clean_mm(bdev, 0);
1597 out_no_sys:
1598 return ret;
1599 }
1600 EXPORT_SYMBOL(ttm_bo_device_init);
1601
1602 /*
1603 * buffer object vm functions.
1604 */
1605
1606 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1607 {
1608 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1609
1610 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1611 if (mem->mem_type == TTM_PL_SYSTEM)
1612 return false;
1613
1614 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1615 return false;
1616
1617 if (mem->placement & TTM_PL_FLAG_CACHED)
1618 return false;
1619 }
1620 return true;
1621 }
1622
1623 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1624 {
1625 struct ttm_bo_device *bdev = bo->bdev;
1626 loff_t offset = (loff_t) bo->addr_space_offset;
1627 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1628
1629 if (!bdev->dev_mapping)
1630 return;
1631 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1632 ttm_mem_io_free_vm(bo);
1633 }
1634
1635 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1636 {
1637 struct ttm_bo_device *bdev = bo->bdev;
1638 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1639
1640 ttm_mem_io_lock(man, false);
1641 ttm_bo_unmap_virtual_locked(bo);
1642 ttm_mem_io_unlock(man);
1643 }
1644
1645
1646 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1647
1648 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1649 {
1650 struct ttm_bo_device *bdev = bo->bdev;
1651 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1652 struct rb_node *parent = NULL;
1653 struct ttm_buffer_object *cur_bo;
1654 unsigned long offset = bo->vm_node->start;
1655 unsigned long cur_offset;
1656
1657 while (*cur) {
1658 parent = *cur;
1659 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1660 cur_offset = cur_bo->vm_node->start;
1661 if (offset < cur_offset)
1662 cur = &parent->rb_left;
1663 else if (offset > cur_offset)
1664 cur = &parent->rb_right;
1665 else
1666 BUG();
1667 }
1668
1669 rb_link_node(&bo->vm_rb, parent, cur);
1670 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1671 }
1672
1673 /**
1674 * ttm_bo_setup_vm:
1675 *
1676 * @bo: the buffer to allocate address space for
1677 *
1678 * Allocate address space in the drm device so that applications
1679 * can mmap the buffer and access the contents. This only
1680 * applies to ttm_bo_type_device objects as others are not
1681 * placed in the drm device address space.
1682 */
1683
1684 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1685 {
1686 struct ttm_bo_device *bdev = bo->bdev;
1687 int ret;
1688
1689 retry_pre_get:
1690 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1691 if (unlikely(ret != 0))
1692 return ret;
1693
1694 write_lock(&bdev->vm_lock);
1695 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1696 bo->mem.num_pages, 0, 0);
1697
1698 if (unlikely(bo->vm_node == NULL)) {
1699 ret = -ENOMEM;
1700 goto out_unlock;
1701 }
1702
1703 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1704 bo->mem.num_pages, 0);
1705
1706 if (unlikely(bo->vm_node == NULL)) {
1707 write_unlock(&bdev->vm_lock);
1708 goto retry_pre_get;
1709 }
1710
1711 ttm_bo_vm_insert_rb(bo);
1712 write_unlock(&bdev->vm_lock);
1713 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1714
1715 return 0;
1716 out_unlock:
1717 write_unlock(&bdev->vm_lock);
1718 return ret;
1719 }
1720
1721 int ttm_bo_wait(struct ttm_buffer_object *bo,
1722 bool lazy, bool interruptible, bool no_wait)
1723 {
1724 struct ttm_bo_driver *driver = bo->bdev->driver;
1725 struct ttm_bo_device *bdev = bo->bdev;
1726 void *sync_obj;
1727 void *sync_obj_arg;
1728 int ret = 0;
1729
1730 if (likely(bo->sync_obj == NULL))
1731 return 0;
1732
1733 while (bo->sync_obj) {
1734
1735 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1736 void *tmp_obj = bo->sync_obj;
1737 bo->sync_obj = NULL;
1738 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1739 spin_unlock(&bdev->fence_lock);
1740 driver->sync_obj_unref(&tmp_obj);
1741 spin_lock(&bdev->fence_lock);
1742 continue;
1743 }
1744
1745 if (no_wait)
1746 return -EBUSY;
1747
1748 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1749 sync_obj_arg = bo->sync_obj_arg;
1750 spin_unlock(&bdev->fence_lock);
1751 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1752 lazy, interruptible);
1753 if (unlikely(ret != 0)) {
1754 driver->sync_obj_unref(&sync_obj);
1755 spin_lock(&bdev->fence_lock);
1756 return ret;
1757 }
1758 spin_lock(&bdev->fence_lock);
1759 if (likely(bo->sync_obj == sync_obj &&
1760 bo->sync_obj_arg == sync_obj_arg)) {
1761 void *tmp_obj = bo->sync_obj;
1762 bo->sync_obj = NULL;
1763 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1764 &bo->priv_flags);
1765 spin_unlock(&bdev->fence_lock);
1766 driver->sync_obj_unref(&sync_obj);
1767 driver->sync_obj_unref(&tmp_obj);
1768 spin_lock(&bdev->fence_lock);
1769 } else {
1770 spin_unlock(&bdev->fence_lock);
1771 driver->sync_obj_unref(&sync_obj);
1772 spin_lock(&bdev->fence_lock);
1773 }
1774 }
1775 return 0;
1776 }
1777 EXPORT_SYMBOL(ttm_bo_wait);
1778
1779 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1780 {
1781 struct ttm_bo_device *bdev = bo->bdev;
1782 int ret = 0;
1783
1784 /*
1785 * Using ttm_bo_reserve makes sure the lru lists are updated.
1786 */
1787
1788 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1789 if (unlikely(ret != 0))
1790 return ret;
1791 spin_lock(&bdev->fence_lock);
1792 ret = ttm_bo_wait(bo, false, true, no_wait);
1793 spin_unlock(&bdev->fence_lock);
1794 if (likely(ret == 0))
1795 atomic_inc(&bo->cpu_writers);
1796 ttm_bo_unreserve(bo);
1797 return ret;
1798 }
1799 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1800
1801 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1802 {
1803 if (atomic_dec_and_test(&bo->cpu_writers))
1804 wake_up_all(&bo->event_queue);
1805 }
1806 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1807
1808 /**
1809 * A buffer object shrink method that tries to swap out the first
1810 * buffer object on the bo_global::swap_lru list.
1811 */
1812
1813 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1814 {
1815 struct ttm_bo_global *glob =
1816 container_of(shrink, struct ttm_bo_global, shrink);
1817 struct ttm_buffer_object *bo;
1818 int ret = -EBUSY;
1819 int put_count;
1820 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1821
1822 spin_lock(&glob->lru_lock);
1823 while (ret == -EBUSY) {
1824 if (unlikely(list_empty(&glob->swap_lru))) {
1825 spin_unlock(&glob->lru_lock);
1826 return -EBUSY;
1827 }
1828
1829 bo = list_first_entry(&glob->swap_lru,
1830 struct ttm_buffer_object, swap);
1831 kref_get(&bo->list_kref);
1832
1833 if (!list_empty(&bo->ddestroy)) {
1834 spin_unlock(&glob->lru_lock);
1835 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1836 kref_put(&bo->list_kref, ttm_bo_release_list);
1837 spin_lock(&glob->lru_lock);
1838 continue;
1839 }
1840
1841 /**
1842 * Reserve buffer. Since we unlock while sleeping, we need
1843 * to re-check that nobody removed us from the swap-list while
1844 * we slept.
1845 */
1846
1847 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1848 if (unlikely(ret == -EBUSY)) {
1849 spin_unlock(&glob->lru_lock);
1850 ttm_bo_wait_unreserved(bo, false);
1851 kref_put(&bo->list_kref, ttm_bo_release_list);
1852 spin_lock(&glob->lru_lock);
1853 }
1854 }
1855
1856 BUG_ON(ret != 0);
1857 put_count = ttm_bo_del_from_lru(bo);
1858 spin_unlock(&glob->lru_lock);
1859
1860 ttm_bo_list_ref_sub(bo, put_count, true);
1861
1862 /**
1863 * Wait for GPU, then move to system cached.
1864 */
1865
1866 spin_lock(&bo->bdev->fence_lock);
1867 ret = ttm_bo_wait(bo, false, false, false);
1868 spin_unlock(&bo->bdev->fence_lock);
1869
1870 if (unlikely(ret != 0))
1871 goto out;
1872
1873 if ((bo->mem.placement & swap_placement) != swap_placement) {
1874 struct ttm_mem_reg evict_mem;
1875
1876 evict_mem = bo->mem;
1877 evict_mem.mm_node = NULL;
1878 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1879 evict_mem.mem_type = TTM_PL_SYSTEM;
1880
1881 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1882 false, false, false);
1883 if (unlikely(ret != 0))
1884 goto out;
1885 }
1886
1887 ttm_bo_unmap_virtual(bo);
1888
1889 /**
1890 * Swap out. Buffer will be swapped in again as soon as
1891 * anyone tries to access a ttm page.
1892 */
1893
1894 if (bo->bdev->driver->swap_notify)
1895 bo->bdev->driver->swap_notify(bo);
1896
1897 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1898 out:
1899
1900 /**
1901 *
1902 * Unreserve without putting on LRU to avoid swapping out an
1903 * already swapped buffer.
1904 */
1905
1906 atomic_set(&bo->reserved, 0);
1907 wake_up_all(&bo->event_queue);
1908 kref_put(&bo->list_kref, ttm_bo_release_list);
1909 return ret;
1910 }
1911
1912 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1913 {
1914 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1915 ;
1916 }
1917 EXPORT_SYMBOL(ttm_bo_swapout_all);
Linux 2.6 ibm-acpi/thinkpad-acpi driver git tree