?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
drm/ttm/vmwgfx: Have TTM manage the validation sequence.
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / drivers / gpu / drm / ttm / ttm_bo.c
blob25e4c2a1d1d8c8ba9be32edab75c12ea7c61e7d3
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 #include "ttm/ttm_module.h"
32 #include "ttm/ttm_bo_driver.h"
33 #include "ttm/ttm_placement.h"
34 #include <linux/jiffies.h>
35 #include <linux/slab.h>
36 #include <linux/sched.h>
37 #include <linux/mm.h>
38 #include <linux/file.h>
39 #include <linux/module.h>
40 #include <asm/atomic.h>
41
42 #define TTM_ASSERT_LOCKED(param)
43 #define TTM_DEBUG(fmt, arg...)
44 #define TTM_BO_HASH_ORDER 13
45
46 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo);
47 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
48 static void ttm_bo_global_kobj_release(struct kobject *kobj);
49
50 static struct attribute ttm_bo_count = {
51 .name = "bo_count",
52 .mode = S_IRUGO
53 };
54
55 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
56 {
57 int i;
58
59 for (i = 0; i <= TTM_PL_PRIV5; i++)
60 if (flags & (1 << i)) {
61 *mem_type = i;
62 return 0;
63 }
64 return -EINVAL;
65 }
66
67 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
68 {
69 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
70
71 printk(KERN_ERR TTM_PFX " has_type: %d\n", man->has_type);
72 printk(KERN_ERR TTM_PFX " use_type: %d\n", man->use_type);
73 printk(KERN_ERR TTM_PFX " flags: 0x%08X\n", man->flags);
74 printk(KERN_ERR TTM_PFX " gpu_offset: 0x%08lX\n", man->gpu_offset);
75 printk(KERN_ERR TTM_PFX " size: %llu\n", man->size);
76 printk(KERN_ERR TTM_PFX " available_caching: 0x%08X\n",
77 man->available_caching);
78 printk(KERN_ERR TTM_PFX " default_caching: 0x%08X\n",
79 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 printk(KERN_ERR TTM_PFX "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 printk(KERN_ERR TTM_PFX " 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
141 BUG_ON(atomic_read(&bo->list_kref.refcount));
142 BUG_ON(atomic_read(&bo->kref.refcount));
143 BUG_ON(atomic_read(&bo->cpu_writers));
144 BUG_ON(bo->sync_obj != NULL);
145 BUG_ON(bo->mem.mm_node != NULL);
146 BUG_ON(!list_empty(&bo->lru));
147 BUG_ON(!list_empty(&bo->ddestroy));
148
149 if (bo->ttm)
150 ttm_tt_destroy(bo->ttm);
151 atomic_dec(&bo->glob->bo_count);
152 if (bo->destroy)
153 bo->destroy(bo);
154 else {
155 ttm_mem_global_free(bdev->glob->mem_glob, bo->acc_size);
156 kfree(bo);
157 }
158 }
159
160 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, bool interruptible)
161 {
162 if (interruptible) {
163 return wait_event_interruptible(bo->event_queue,
164 atomic_read(&bo->reserved) == 0);
165 } else {
166 wait_event(bo->event_queue, atomic_read(&bo->reserved) == 0);
167 return 0;
168 }
169 }
170 EXPORT_SYMBOL(ttm_bo_wait_unreserved);
171
172 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
173 {
174 struct ttm_bo_device *bdev = bo->bdev;
175 struct ttm_mem_type_manager *man;
176
177 BUG_ON(!atomic_read(&bo->reserved));
178
179 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
180
181 BUG_ON(!list_empty(&bo->lru));
182
183 man = &bdev->man[bo->mem.mem_type];
184 list_add_tail(&bo->lru, &man->lru);
185 kref_get(&bo->list_kref);
186
187 if (bo->ttm != NULL) {
188 list_add_tail(&bo->swap, &bo->glob->swap_lru);
189 kref_get(&bo->list_kref);
190 }
191 }
192 }
193
194 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
195 {
196 int put_count = 0;
197
198 if (!list_empty(&bo->swap)) {
199 list_del_init(&bo->swap);
200 ++put_count;
201 }
202 if (!list_empty(&bo->lru)) {
203 list_del_init(&bo->lru);
204 ++put_count;
205 }
206
207 /*
208 * TODO: Add a driver hook to delete from
209 * driver-specific LRU's here.
210 */
211
212 return put_count;
213 }
214
215 int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
216 bool interruptible,
217 bool no_wait, bool use_sequence, uint32_t sequence)
218 {
219 struct ttm_bo_global *glob = bo->glob;
220 int ret;
221
222 while (unlikely(atomic_cmpxchg(&bo->reserved, 0, 1) != 0)) {
223 /**
224 * Deadlock avoidance for multi-bo reserving.
225 */
226 if (use_sequence && bo->seq_valid) {
227 /**
228 * We've already reserved this one.
229 */
230 if (unlikely(sequence == bo->val_seq))
231 return -EDEADLK;
232 /**
233 * Already reserved by a thread that will not back
234 * off for us. We need to back off.
235 */
236 if (unlikely(sequence - bo->val_seq < (1 << 31)))
237 return -EAGAIN;
238 }
239
240 if (no_wait)
241 return -EBUSY;
242
243 spin_unlock(&glob->lru_lock);
244 ret = ttm_bo_wait_unreserved(bo, interruptible);
245 spin_lock(&glob->lru_lock);
246
247 if (unlikely(ret))
248 return ret;
249 }
250
251 if (use_sequence) {
252 /**
253 * Wake up waiters that may need to recheck for deadlock,
254 * if we decreased the sequence number.
255 */
256 if (unlikely((bo->val_seq - sequence < (1 << 31))
257 || !bo->seq_valid))
258 wake_up_all(&bo->event_queue);
259
260 bo->val_seq = sequence;
261 bo->seq_valid = true;
262 } else {
263 bo->seq_valid = false;
264 }
265
266 return 0;
267 }
268 EXPORT_SYMBOL(ttm_bo_reserve);
269
270 static void ttm_bo_ref_bug(struct kref *list_kref)
271 {
272 BUG();
273 }
274
275 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
276 bool never_free)
277 {
278 kref_sub(&bo->list_kref, count,
279 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
280 }
281
282 int ttm_bo_reserve(struct ttm_buffer_object *bo,
283 bool interruptible,
284 bool no_wait, bool use_sequence, uint32_t sequence)
285 {
286 struct ttm_bo_global *glob = bo->glob;
287 int put_count = 0;
288 int ret;
289
290 spin_lock(&glob->lru_lock);
291 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
292 sequence);
293 if (likely(ret == 0))
294 put_count = ttm_bo_del_from_lru(bo);
295 spin_unlock(&glob->lru_lock);
296
297 ttm_bo_list_ref_sub(bo, put_count, true);
298
299 return ret;
300 }
301
302 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo)
303 {
304 ttm_bo_add_to_lru(bo);
305 atomic_set(&bo->reserved, 0);
306 wake_up_all(&bo->event_queue);
307 }
308
309 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
310 {
311 struct ttm_bo_global *glob = bo->glob;
312
313 spin_lock(&glob->lru_lock);
314 ttm_bo_unreserve_locked(bo);
315 spin_unlock(&glob->lru_lock);
316 }
317 EXPORT_SYMBOL(ttm_bo_unreserve);
318
319 /*
320 * Call bo->mutex locked.
321 */
322 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
323 {
324 struct ttm_bo_device *bdev = bo->bdev;
325 struct ttm_bo_global *glob = bo->glob;
326 int ret = 0;
327 uint32_t page_flags = 0;
328
329 TTM_ASSERT_LOCKED(&bo->mutex);
330 bo->ttm = NULL;
331
332 if (bdev->need_dma32)
333 page_flags |= TTM_PAGE_FLAG_DMA32;
334
335 switch (bo->type) {
336 case ttm_bo_type_device:
337 if (zero_alloc)
338 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
339 case ttm_bo_type_kernel:
340 bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
341 page_flags, glob->dummy_read_page);
342 if (unlikely(bo->ttm == NULL))
343 ret = -ENOMEM;
344 break;
345 case ttm_bo_type_user:
346 bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
347 page_flags | TTM_PAGE_FLAG_USER,
348 glob->dummy_read_page);
349 if (unlikely(bo->ttm == NULL)) {
350 ret = -ENOMEM;
351 break;
352 }
353
354 ret = ttm_tt_set_user(bo->ttm, current,
355 bo->buffer_start, bo->num_pages);
356 if (unlikely(ret != 0))
357 ttm_tt_destroy(bo->ttm);
358 break;
359 default:
360 printk(KERN_ERR TTM_PFX "Illegal buffer object type\n");
361 ret = -EINVAL;
362 break;
363 }
364
365 return ret;
366 }
367
368 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
369 struct ttm_mem_reg *mem,
370 bool evict, bool interruptible,
371 bool no_wait_reserve, bool no_wait_gpu)
372 {
373 struct ttm_bo_device *bdev = bo->bdev;
374 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
375 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
376 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
377 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
378 int ret = 0;
379
380 if (old_is_pci || new_is_pci ||
381 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0))
382 ttm_bo_unmap_virtual(bo);
383
384 /*
385 * Create and bind a ttm if required.
386 */
387
388 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && (bo->ttm == NULL)) {
389 ret = ttm_bo_add_ttm(bo, false);
390 if (ret)
391 goto out_err;
392
393 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
394 if (ret)
395 goto out_err;
396
397 if (mem->mem_type != TTM_PL_SYSTEM) {
398 ret = ttm_tt_bind(bo->ttm, mem);
399 if (ret)
400 goto out_err;
401 }
402
403 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
404 bo->mem = *mem;
405 mem->mm_node = NULL;
406 goto moved;
407 }
408
409 }
410
411 if (bdev->driver->move_notify)
412 bdev->driver->move_notify(bo, mem);
413
414 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
415 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
416 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
417 else if (bdev->driver->move)
418 ret = bdev->driver->move(bo, evict, interruptible,
419 no_wait_reserve, no_wait_gpu, mem);
420 else
421 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
422
423 if (ret)
424 goto out_err;
425
426 moved:
427 if (bo->evicted) {
428 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
429 if (ret)
430 printk(KERN_ERR TTM_PFX "Can not flush read caches\n");
431 bo->evicted = false;
432 }
433
434 if (bo->mem.mm_node) {
435 bo->offset = (bo->mem.start << PAGE_SHIFT) +
436 bdev->man[bo->mem.mem_type].gpu_offset;
437 bo->cur_placement = bo->mem.placement;
438 } else
439 bo->offset = 0;
440
441 return 0;
442
443 out_err:
444 new_man = &bdev->man[bo->mem.mem_type];
445 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
446 ttm_tt_unbind(bo->ttm);
447 ttm_tt_destroy(bo->ttm);
448 bo->ttm = NULL;
449 }
450
451 return ret;
452 }
453
454 /**
455 * Call bo::reserved.
456 * Will release GPU memory type usage on destruction.
457 * This is the place to put in driver specific hooks to release
458 * driver private resources.
459 * Will release the bo::reserved lock.
460 */
461
462 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
463 {
464 if (bo->ttm) {
465 ttm_tt_unbind(bo->ttm);
466 ttm_tt_destroy(bo->ttm);
467 bo->ttm = NULL;
468 }
469
470 ttm_bo_mem_put(bo, &bo->mem);
471
472 atomic_set(&bo->reserved, 0);
473
474 /*
475 * Make processes trying to reserve really pick it up.
476 */
477 smp_mb__after_atomic_dec();
478 wake_up_all(&bo->event_queue);
479 }
480
481 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
482 {
483 struct ttm_bo_device *bdev = bo->bdev;
484 struct ttm_bo_global *glob = bo->glob;
485 struct ttm_bo_driver *driver;
486 void *sync_obj = NULL;
487 void *sync_obj_arg;
488 int put_count;
489 int ret;
490
491 spin_lock(&bdev->fence_lock);
492 (void) ttm_bo_wait(bo, false, false, true);
493 if (!bo->sync_obj) {
494
495 spin_lock(&glob->lru_lock);
496
497 /**
498 * Lock inversion between bo:reserve and bdev::fence_lock here,
499 * but that's OK, since we're only trylocking.
500 */
501
502 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
503
504 if (unlikely(ret == -EBUSY))
505 goto queue;
506
507 spin_unlock(&bdev->fence_lock);
508 put_count = ttm_bo_del_from_lru(bo);
509
510 spin_unlock(&glob->lru_lock);
511 ttm_bo_cleanup_memtype_use(bo);
512
513 ttm_bo_list_ref_sub(bo, put_count, true);
514
515 return;
516 } else {
517 spin_lock(&glob->lru_lock);
518 }
519 queue:
520 driver = bdev->driver;
521 if (bo->sync_obj)
522 sync_obj = driver->sync_obj_ref(bo->sync_obj);
523 sync_obj_arg = bo->sync_obj_arg;
524
525 kref_get(&bo->list_kref);
526 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
527 spin_unlock(&glob->lru_lock);
528 spin_unlock(&bdev->fence_lock);
529
530 if (sync_obj) {
531 driver->sync_obj_flush(sync_obj, sync_obj_arg);
532 driver->sync_obj_unref(&sync_obj);
533 }
534 schedule_delayed_work(&bdev->wq,
535 ((HZ / 100) < 1) ? 1 : HZ / 100);
536 }
537
538 /**
539 * function ttm_bo_cleanup_refs
540 * If bo idle, remove from delayed- and lru lists, and unref.
541 * If not idle, do nothing.
542 *
543 * @interruptible Any sleeps should occur interruptibly.
544 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
545 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
546 */
547
548 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
549 bool interruptible,
550 bool no_wait_reserve,
551 bool no_wait_gpu)
552 {
553 struct ttm_bo_device *bdev = bo->bdev;
554 struct ttm_bo_global *glob = bo->glob;
555 int put_count;
556 int ret = 0;
557
558 retry:
559 spin_lock(&bdev->fence_lock);
560 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
561 spin_unlock(&bdev->fence_lock);
562
563 if (unlikely(ret != 0))
564 return ret;
565
566 spin_lock(&glob->lru_lock);
567 ret = ttm_bo_reserve_locked(bo, interruptible,
568 no_wait_reserve, false, 0);
569
570 if (unlikely(ret != 0) || list_empty(&bo->ddestroy)) {
571 spin_unlock(&glob->lru_lock);
572 return ret;
573 }
574
575 /**
576 * We can re-check for sync object without taking
577 * the bo::lock since setting the sync object requires
578 * also bo::reserved. A busy object at this point may
579 * be caused by another thread recently starting an accelerated
580 * eviction.
581 */
582
583 if (unlikely(bo->sync_obj)) {
584 atomic_set(&bo->reserved, 0);
585 wake_up_all(&bo->event_queue);
586 spin_unlock(&glob->lru_lock);
587 goto retry;
588 }
589
590 put_count = ttm_bo_del_from_lru(bo);
591 list_del_init(&bo->ddestroy);
592 ++put_count;
593
594 spin_unlock(&glob->lru_lock);
595 ttm_bo_cleanup_memtype_use(bo);
596
597 ttm_bo_list_ref_sub(bo, put_count, true);
598
599 return 0;
600 }
601
602 /**
603 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
604 * encountered buffers.
605 */
606
607 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
608 {
609 struct ttm_bo_global *glob = bdev->glob;
610 struct ttm_buffer_object *entry = NULL;
611 int ret = 0;
612
613 spin_lock(&glob->lru_lock);
614 if (list_empty(&bdev->ddestroy))
615 goto out_unlock;
616
617 entry = list_first_entry(&bdev->ddestroy,
618 struct ttm_buffer_object, ddestroy);
619 kref_get(&entry->list_kref);
620
621 for (;;) {
622 struct ttm_buffer_object *nentry = NULL;
623
624 if (entry->ddestroy.next != &bdev->ddestroy) {
625 nentry = list_first_entry(&entry->ddestroy,
626 struct ttm_buffer_object, ddestroy);
627 kref_get(&nentry->list_kref);
628 }
629
630 spin_unlock(&glob->lru_lock);
631 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
632 !remove_all);
633 kref_put(&entry->list_kref, ttm_bo_release_list);
634 entry = nentry;
635
636 if (ret || !entry)
637 goto out;
638
639 spin_lock(&glob->lru_lock);
640 if (list_empty(&entry->ddestroy))
641 break;
642 }
643
644 out_unlock:
645 spin_unlock(&glob->lru_lock);
646 out:
647 if (entry)
648 kref_put(&entry->list_kref, ttm_bo_release_list);
649 return ret;
650 }
651
652 static void ttm_bo_delayed_workqueue(struct work_struct *work)
653 {
654 struct ttm_bo_device *bdev =
655 container_of(work, struct ttm_bo_device, wq.work);
656
657 if (ttm_bo_delayed_delete(bdev, false)) {
658 schedule_delayed_work(&bdev->wq,
659 ((HZ / 100) < 1) ? 1 : HZ / 100);
660 }
661 }
662
663 static void ttm_bo_release(struct kref *kref)
664 {
665 struct ttm_buffer_object *bo =
666 container_of(kref, struct ttm_buffer_object, kref);
667 struct ttm_bo_device *bdev = bo->bdev;
668
669 if (likely(bo->vm_node != NULL)) {
670 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
671 drm_mm_put_block(bo->vm_node);
672 bo->vm_node = NULL;
673 }
674 write_unlock(&bdev->vm_lock);
675 ttm_bo_cleanup_refs_or_queue(bo);
676 kref_put(&bo->list_kref, ttm_bo_release_list);
677 write_lock(&bdev->vm_lock);
678 }
679
680 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
681 {
682 struct ttm_buffer_object *bo = *p_bo;
683 struct ttm_bo_device *bdev = bo->bdev;
684
685 *p_bo = NULL;
686 write_lock(&bdev->vm_lock);
687 kref_put(&bo->kref, ttm_bo_release);
688 write_unlock(&bdev->vm_lock);
689 }
690 EXPORT_SYMBOL(ttm_bo_unref);
691
692 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
693 {
694 return cancel_delayed_work_sync(&bdev->wq);
695 }
696 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
697
698 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
699 {
700 if (resched)
701 schedule_delayed_work(&bdev->wq,
702 ((HZ / 100) < 1) ? 1 : HZ / 100);
703 }
704 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
705
706 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
707 bool no_wait_reserve, bool no_wait_gpu)
708 {
709 struct ttm_bo_device *bdev = bo->bdev;
710 struct ttm_mem_reg evict_mem;
711 struct ttm_placement placement;
712 int ret = 0;
713
714 spin_lock(&bdev->fence_lock);
715 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
716 spin_unlock(&bdev->fence_lock);
717
718 if (unlikely(ret != 0)) {
719 if (ret != -ERESTARTSYS) {
720 printk(KERN_ERR TTM_PFX
721 "Failed to expire sync object before "
722 "buffer eviction.\n");
723 }
724 goto out;
725 }
726
727 BUG_ON(!atomic_read(&bo->reserved));
728
729 evict_mem = bo->mem;
730 evict_mem.mm_node = NULL;
731 evict_mem.bus.io_reserved = false;
732
733 placement.fpfn = 0;
734 placement.lpfn = 0;
735 placement.num_placement = 0;
736 placement.num_busy_placement = 0;
737 bdev->driver->evict_flags(bo, &placement);
738 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
739 no_wait_reserve, no_wait_gpu);
740 if (ret) {
741 if (ret != -ERESTARTSYS) {
742 printk(KERN_ERR TTM_PFX
743 "Failed to find memory space for "
744 "buffer 0x%p eviction.\n", bo);
745 ttm_bo_mem_space_debug(bo, &placement);
746 }
747 goto out;
748 }
749
750 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
751 no_wait_reserve, no_wait_gpu);
752 if (ret) {
753 if (ret != -ERESTARTSYS)
754 printk(KERN_ERR TTM_PFX "Buffer eviction failed\n");
755 ttm_bo_mem_put(bo, &evict_mem);
756 goto out;
757 }
758 bo->evicted = true;
759 out:
760 return ret;
761 }
762
763 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
764 uint32_t mem_type,
765 bool interruptible, bool no_wait_reserve,
766 bool no_wait_gpu)
767 {
768 struct ttm_bo_global *glob = bdev->glob;
769 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
770 struct ttm_buffer_object *bo;
771 int ret, put_count = 0;
772
773 retry:
774 spin_lock(&glob->lru_lock);
775 if (list_empty(&man->lru)) {
776 spin_unlock(&glob->lru_lock);
777 return -EBUSY;
778 }
779
780 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
781 kref_get(&bo->list_kref);
782
783 if (!list_empty(&bo->ddestroy)) {
784 spin_unlock(&glob->lru_lock);
785 ret = ttm_bo_cleanup_refs(bo, interruptible,
786 no_wait_reserve, no_wait_gpu);
787 kref_put(&bo->list_kref, ttm_bo_release_list);
788
789 if (likely(ret == 0 || ret == -ERESTARTSYS))
790 return ret;
791
792 goto retry;
793 }
794
795 ret = ttm_bo_reserve_locked(bo, false, no_wait_reserve, false, 0);
796
797 if (unlikely(ret == -EBUSY)) {
798 spin_unlock(&glob->lru_lock);
799 if (likely(!no_wait_gpu))
800 ret = ttm_bo_wait_unreserved(bo, interruptible);
801
802 kref_put(&bo->list_kref, ttm_bo_release_list);
803
804 /**
805 * We *need* to retry after releasing the lru lock.
806 */
807
808 if (unlikely(ret != 0))
809 return ret;
810 goto retry;
811 }
812
813 put_count = ttm_bo_del_from_lru(bo);
814 spin_unlock(&glob->lru_lock);
815
816 BUG_ON(ret != 0);
817
818 ttm_bo_list_ref_sub(bo, put_count, true);
819
820 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
821 ttm_bo_unreserve(bo);
822
823 kref_put(&bo->list_kref, ttm_bo_release_list);
824 return ret;
825 }
826
827 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
828 {
829 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
830
831 if (mem->mm_node)
832 (*man->func->put_node)(man, mem);
833 }
834 EXPORT_SYMBOL(ttm_bo_mem_put);
835
836 /**
837 * Repeatedly evict memory from the LRU for @mem_type until we create enough
838 * space, or we've evicted everything and there isn't enough space.
839 */
840 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
841 uint32_t mem_type,
842 struct ttm_placement *placement,
843 struct ttm_mem_reg *mem,
844 bool interruptible,
845 bool no_wait_reserve,
846 bool no_wait_gpu)
847 {
848 struct ttm_bo_device *bdev = bo->bdev;
849 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
850 int ret;
851
852 do {
853 ret = (*man->func->get_node)(man, bo, placement, mem);
854 if (unlikely(ret != 0))
855 return ret;
856 if (mem->mm_node)
857 break;
858 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
859 no_wait_reserve, no_wait_gpu);
860 if (unlikely(ret != 0))
861 return ret;
862 } while (1);
863 if (mem->mm_node == NULL)
864 return -ENOMEM;
865 mem->mem_type = mem_type;
866 return 0;
867 }
868
869 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
870 uint32_t cur_placement,
871 uint32_t proposed_placement)
872 {
873 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
874 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
875
876 /**
877 * Keep current caching if possible.
878 */
879
880 if ((cur_placement & caching) != 0)
881 result |= (cur_placement & caching);
882 else if ((man->default_caching & caching) != 0)
883 result |= man->default_caching;
884 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
885 result |= TTM_PL_FLAG_CACHED;
886 else if ((TTM_PL_FLAG_WC & caching) != 0)
887 result |= TTM_PL_FLAG_WC;
888 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
889 result |= TTM_PL_FLAG_UNCACHED;
890
891 return result;
892 }
893
894 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
895 bool disallow_fixed,
896 uint32_t mem_type,
897 uint32_t proposed_placement,
898 uint32_t *masked_placement)
899 {
900 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
901
902 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && disallow_fixed)
903 return false;
904
905 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
906 return false;
907
908 if ((proposed_placement & man->available_caching) == 0)
909 return false;
910
911 cur_flags |= (proposed_placement & man->available_caching);
912
913 *masked_placement = cur_flags;
914 return true;
915 }
916
917 /**
918 * Creates space for memory region @mem according to its type.
919 *
920 * This function first searches for free space in compatible memory types in
921 * the priority order defined by the driver. If free space isn't found, then
922 * ttm_bo_mem_force_space is attempted in priority order to evict and find
923 * space.
924 */
925 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
926 struct ttm_placement *placement,
927 struct ttm_mem_reg *mem,
928 bool interruptible, bool no_wait_reserve,
929 bool no_wait_gpu)
930 {
931 struct ttm_bo_device *bdev = bo->bdev;
932 struct ttm_mem_type_manager *man;
933 uint32_t mem_type = TTM_PL_SYSTEM;
934 uint32_t cur_flags = 0;
935 bool type_found = false;
936 bool type_ok = false;
937 bool has_erestartsys = false;
938 int i, ret;
939
940 mem->mm_node = NULL;
941 for (i = 0; i < placement->num_placement; ++i) {
942 ret = ttm_mem_type_from_flags(placement->placement[i],
943 &mem_type);
944 if (ret)
945 return ret;
946 man = &bdev->man[mem_type];
947
948 type_ok = ttm_bo_mt_compatible(man,
949 bo->type == ttm_bo_type_user,
950 mem_type,
951 placement->placement[i],
952 &cur_flags);
953
954 if (!type_ok)
955 continue;
956
957 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
958 cur_flags);
959 /*
960 * Use the access and other non-mapping-related flag bits from
961 * the memory placement flags to the current flags
962 */
963 ttm_flag_masked(&cur_flags, placement->placement[i],
964 ~TTM_PL_MASK_MEMTYPE);
965
966 if (mem_type == TTM_PL_SYSTEM)
967 break;
968
969 if (man->has_type && man->use_type) {
970 type_found = true;
971 ret = (*man->func->get_node)(man, bo, placement, mem);
972 if (unlikely(ret))
973 return ret;
974 }
975 if (mem->mm_node)
976 break;
977 }
978
979 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
980 mem->mem_type = mem_type;
981 mem->placement = cur_flags;
982 return 0;
983 }
984
985 if (!type_found)
986 return -EINVAL;
987
988 for (i = 0; i < placement->num_busy_placement; ++i) {
989 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
990 &mem_type);
991 if (ret)
992 return ret;
993 man = &bdev->man[mem_type];
994 if (!man->has_type)
995 continue;
996 if (!ttm_bo_mt_compatible(man,
997 bo->type == ttm_bo_type_user,
998 mem_type,
999 placement->busy_placement[i],
1000 &cur_flags))
1001 continue;
1002
1003 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1004 cur_flags);
1005 /*
1006 * Use the access and other non-mapping-related flag bits from
1007 * the memory placement flags to the current flags
1008 */
1009 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1010 ~TTM_PL_MASK_MEMTYPE);
1011
1012
1013 if (mem_type == TTM_PL_SYSTEM) {
1014 mem->mem_type = mem_type;
1015 mem->placement = cur_flags;
1016 mem->mm_node = NULL;
1017 return 0;
1018 }
1019
1020 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1021 interruptible, no_wait_reserve, no_wait_gpu);
1022 if (ret == 0 && mem->mm_node) {
1023 mem->placement = cur_flags;
1024 return 0;
1025 }
1026 if (ret == -ERESTARTSYS)
1027 has_erestartsys = true;
1028 }
1029 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1030 return ret;
1031 }
1032 EXPORT_SYMBOL(ttm_bo_mem_space);
1033
1034 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1035 {
1036 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1037 return -EBUSY;
1038
1039 return wait_event_interruptible(bo->event_queue,
1040 atomic_read(&bo->cpu_writers) == 0);
1041 }
1042 EXPORT_SYMBOL(ttm_bo_wait_cpu);
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(!atomic_read(&bo->reserved));
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 = false;
1069 /*
1070 * Determine where to move the buffer.
1071 */
1072 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1073 if (ret)
1074 goto out_unlock;
1075 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1076 out_unlock:
1077 if (ret && mem.mm_node)
1078 ttm_bo_mem_put(bo, &mem);
1079 return ret;
1080 }
1081
1082 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1083 struct ttm_mem_reg *mem)
1084 {
1085 int i;
1086
1087 if (mem->mm_node && placement->lpfn != 0 &&
1088 (mem->start < placement->fpfn ||
1089 mem->start + mem->num_pages > placement->lpfn))
1090 return -1;
1091
1092 for (i = 0; i < placement->num_placement; i++) {
1093 if ((placement->placement[i] & mem->placement &
1094 TTM_PL_MASK_CACHING) &&
1095 (placement->placement[i] & mem->placement &
1096 TTM_PL_MASK_MEM))
1097 return i;
1098 }
1099 return -1;
1100 }
1101
1102 int ttm_bo_validate(struct ttm_buffer_object *bo,
1103 struct ttm_placement *placement,
1104 bool interruptible, bool no_wait_reserve,
1105 bool no_wait_gpu)
1106 {
1107 int ret;
1108
1109 BUG_ON(!atomic_read(&bo->reserved));
1110 /* Check that range is valid */
1111 if (placement->lpfn || placement->fpfn)
1112 if (placement->fpfn > placement->lpfn ||
1113 (placement->lpfn - placement->fpfn) < bo->num_pages)
1114 return -EINVAL;
1115 /*
1116 * Check whether we need to move buffer.
1117 */
1118 ret = ttm_bo_mem_compat(placement, &bo->mem);
1119 if (ret < 0) {
1120 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1121 if (ret)
1122 return ret;
1123 } else {
1124 /*
1125 * Use the access and other non-mapping-related flag bits from
1126 * the compatible memory placement flags to the active flags
1127 */
1128 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1129 ~TTM_PL_MASK_MEMTYPE);
1130 }
1131 /*
1132 * We might need to add a TTM.
1133 */
1134 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1135 ret = ttm_bo_add_ttm(bo, true);
1136 if (ret)
1137 return ret;
1138 }
1139 return 0;
1140 }
1141 EXPORT_SYMBOL(ttm_bo_validate);
1142
1143 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1144 struct ttm_placement *placement)
1145 {
1146 BUG_ON((placement->fpfn || placement->lpfn) &&
1147 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1148
1149 return 0;
1150 }
1151
1152 int ttm_bo_init(struct ttm_bo_device *bdev,
1153 struct ttm_buffer_object *bo,
1154 unsigned long size,
1155 enum ttm_bo_type type,
1156 struct ttm_placement *placement,
1157 uint32_t page_alignment,
1158 unsigned long buffer_start,
1159 bool interruptible,
1160 struct file *persistant_swap_storage,
1161 size_t acc_size,
1162 void (*destroy) (struct ttm_buffer_object *))
1163 {
1164 int ret = 0;
1165 unsigned long num_pages;
1166
1167 size += buffer_start & ~PAGE_MASK;
1168 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1169 if (num_pages == 0) {
1170 printk(KERN_ERR TTM_PFX "Illegal buffer object size.\n");
1171 if (destroy)
1172 (*destroy)(bo);
1173 else
1174 kfree(bo);
1175 return -EINVAL;
1176 }
1177 bo->destroy = destroy;
1178
1179 kref_init(&bo->kref);
1180 kref_init(&bo->list_kref);
1181 atomic_set(&bo->cpu_writers, 0);
1182 atomic_set(&bo->reserved, 1);
1183 init_waitqueue_head(&bo->event_queue);
1184 INIT_LIST_HEAD(&bo->lru);
1185 INIT_LIST_HEAD(&bo->ddestroy);
1186 INIT_LIST_HEAD(&bo->swap);
1187 bo->bdev = bdev;
1188 bo->glob = bdev->glob;
1189 bo->type = type;
1190 bo->num_pages = num_pages;
1191 bo->mem.size = num_pages << PAGE_SHIFT;
1192 bo->mem.mem_type = TTM_PL_SYSTEM;
1193 bo->mem.num_pages = bo->num_pages;
1194 bo->mem.mm_node = NULL;
1195 bo->mem.page_alignment = page_alignment;
1196 bo->mem.bus.io_reserved = false;
1197 bo->buffer_start = buffer_start & PAGE_MASK;
1198 bo->priv_flags = 0;
1199 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1200 bo->seq_valid = false;
1201 bo->persistant_swap_storage = persistant_swap_storage;
1202 bo->acc_size = acc_size;
1203 atomic_inc(&bo->glob->bo_count);
1204
1205 ret = ttm_bo_check_placement(bo, placement);
1206 if (unlikely(ret != 0))
1207 goto out_err;
1208
1209 /*
1210 * For ttm_bo_type_device buffers, allocate
1211 * address space from the device.
1212 */
1213 if (bo->type == ttm_bo_type_device) {
1214 ret = ttm_bo_setup_vm(bo);
1215 if (ret)
1216 goto out_err;
1217 }
1218
1219 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1220 if (ret)
1221 goto out_err;
1222
1223 ttm_bo_unreserve(bo);
1224 return 0;
1225
1226 out_err:
1227 ttm_bo_unreserve(bo);
1228 ttm_bo_unref(&bo);
1229
1230 return ret;
1231 }
1232 EXPORT_SYMBOL(ttm_bo_init);
1233
1234 static inline size_t ttm_bo_size(struct ttm_bo_global *glob,
1235 unsigned long num_pages)
1236 {
1237 size_t page_array_size = (num_pages * sizeof(void *) + PAGE_SIZE - 1) &
1238 PAGE_MASK;
1239
1240 return glob->ttm_bo_size + 2 * page_array_size;
1241 }
1242
1243 int ttm_bo_create(struct ttm_bo_device *bdev,
1244 unsigned long size,
1245 enum ttm_bo_type type,
1246 struct ttm_placement *placement,
1247 uint32_t page_alignment,
1248 unsigned long buffer_start,
1249 bool interruptible,
1250 struct file *persistant_swap_storage,
1251 struct ttm_buffer_object **p_bo)
1252 {
1253 struct ttm_buffer_object *bo;
1254 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1255 int ret;
1256
1257 size_t acc_size =
1258 ttm_bo_size(bdev->glob, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
1259 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1260 if (unlikely(ret != 0))
1261 return ret;
1262
1263 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1264
1265 if (unlikely(bo == NULL)) {
1266 ttm_mem_global_free(mem_glob, acc_size);
1267 return -ENOMEM;
1268 }
1269
1270 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1271 buffer_start, interruptible,
1272 persistant_swap_storage, acc_size, NULL);
1273 if (likely(ret == 0))
1274 *p_bo = bo;
1275
1276 return ret;
1277 }
1278
1279 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1280 unsigned mem_type, bool allow_errors)
1281 {
1282 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1283 struct ttm_bo_global *glob = bdev->glob;
1284 int ret;
1285
1286 /*
1287 * Can't use standard list traversal since we're unlocking.
1288 */
1289
1290 spin_lock(&glob->lru_lock);
1291 while (!list_empty(&man->lru)) {
1292 spin_unlock(&glob->lru_lock);
1293 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1294 if (ret) {
1295 if (allow_errors) {
1296 return ret;
1297 } else {
1298 printk(KERN_ERR TTM_PFX
1299 "Cleanup eviction failed\n");
1300 }
1301 }
1302 spin_lock(&glob->lru_lock);
1303 }
1304 spin_unlock(&glob->lru_lock);
1305 return 0;
1306 }
1307
1308 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1309 {
1310 struct ttm_mem_type_manager *man;
1311 int ret = -EINVAL;
1312
1313 if (mem_type >= TTM_NUM_MEM_TYPES) {
1314 printk(KERN_ERR TTM_PFX "Illegal memory type %d\n", mem_type);
1315 return ret;
1316 }
1317 man = &bdev->man[mem_type];
1318
1319 if (!man->has_type) {
1320 printk(KERN_ERR TTM_PFX "Trying to take down uninitialized "
1321 "memory manager type %u\n", mem_type);
1322 return ret;
1323 }
1324
1325 man->use_type = false;
1326 man->has_type = false;
1327
1328 ret = 0;
1329 if (mem_type > 0) {
1330 ttm_bo_force_list_clean(bdev, mem_type, false);
1331
1332 ret = (*man->func->takedown)(man);
1333 }
1334
1335 return ret;
1336 }
1337 EXPORT_SYMBOL(ttm_bo_clean_mm);
1338
1339 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1340 {
1341 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1342
1343 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1344 printk(KERN_ERR TTM_PFX
1345 "Illegal memory manager memory type %u.\n",
1346 mem_type);
1347 return -EINVAL;
1348 }
1349
1350 if (!man->has_type) {
1351 printk(KERN_ERR TTM_PFX
1352 "Memory type %u has not been initialized.\n",
1353 mem_type);
1354 return 0;
1355 }
1356
1357 return ttm_bo_force_list_clean(bdev, mem_type, true);
1358 }
1359 EXPORT_SYMBOL(ttm_bo_evict_mm);
1360
1361 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1362 unsigned long p_size)
1363 {
1364 int ret = -EINVAL;
1365 struct ttm_mem_type_manager *man;
1366
1367 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1368 man = &bdev->man[type];
1369 BUG_ON(man->has_type);
1370
1371 ret = bdev->driver->init_mem_type(bdev, type, man);
1372 if (ret)
1373 return ret;
1374 man->bdev = bdev;
1375
1376 ret = 0;
1377 if (type != TTM_PL_SYSTEM) {
1378 ret = (*man->func->init)(man, p_size);
1379 if (ret)
1380 return ret;
1381 }
1382 man->has_type = true;
1383 man->use_type = true;
1384 man->size = p_size;
1385
1386 INIT_LIST_HEAD(&man->lru);
1387
1388 return 0;
1389 }
1390 EXPORT_SYMBOL(ttm_bo_init_mm);
1391
1392 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1393 {
1394 struct ttm_bo_global *glob =
1395 container_of(kobj, struct ttm_bo_global, kobj);
1396
1397 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1398 __free_page(glob->dummy_read_page);
1399 kfree(glob);
1400 }
1401
1402 void ttm_bo_global_release(struct drm_global_reference *ref)
1403 {
1404 struct ttm_bo_global *glob = ref->object;
1405
1406 kobject_del(&glob->kobj);
1407 kobject_put(&glob->kobj);
1408 }
1409 EXPORT_SYMBOL(ttm_bo_global_release);
1410
1411 int ttm_bo_global_init(struct drm_global_reference *ref)
1412 {
1413 struct ttm_bo_global_ref *bo_ref =
1414 container_of(ref, struct ttm_bo_global_ref, ref);
1415 struct ttm_bo_global *glob = ref->object;
1416 int ret;
1417
1418 mutex_init(&glob->device_list_mutex);
1419 spin_lock_init(&glob->lru_lock);
1420 glob->mem_glob = bo_ref->mem_glob;
1421 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1422
1423 if (unlikely(glob->dummy_read_page == NULL)) {
1424 ret = -ENOMEM;
1425 goto out_no_drp;
1426 }
1427
1428 INIT_LIST_HEAD(&glob->swap_lru);
1429 INIT_LIST_HEAD(&glob->device_list);
1430
1431 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1432 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1433 if (unlikely(ret != 0)) {
1434 printk(KERN_ERR TTM_PFX
1435 "Could not register buffer object swapout.\n");
1436 goto out_no_shrink;
1437 }
1438
1439 glob->ttm_bo_extra_size =
1440 ttm_round_pot(sizeof(struct ttm_tt)) +
1441 ttm_round_pot(sizeof(struct ttm_backend));
1442
1443 glob->ttm_bo_size = glob->ttm_bo_extra_size +
1444 ttm_round_pot(sizeof(struct ttm_buffer_object));
1445
1446 atomic_set(&glob->bo_count, 0);
1447
1448 ret = kobject_init_and_add(
1449 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1450 if (unlikely(ret != 0))
1451 kobject_put(&glob->kobj);
1452 return ret;
1453 out_no_shrink:
1454 __free_page(glob->dummy_read_page);
1455 out_no_drp:
1456 kfree(glob);
1457 return ret;
1458 }
1459 EXPORT_SYMBOL(ttm_bo_global_init);
1460
1461
1462 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1463 {
1464 int ret = 0;
1465 unsigned i = TTM_NUM_MEM_TYPES;
1466 struct ttm_mem_type_manager *man;
1467 struct ttm_bo_global *glob = bdev->glob;
1468
1469 while (i--) {
1470 man = &bdev->man[i];
1471 if (man->has_type) {
1472 man->use_type = false;
1473 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1474 ret = -EBUSY;
1475 printk(KERN_ERR TTM_PFX
1476 "DRM memory manager type %d "
1477 "is not clean.\n", i);
1478 }
1479 man->has_type = false;
1480 }
1481 }
1482
1483 mutex_lock(&glob->device_list_mutex);
1484 list_del(&bdev->device_list);
1485 mutex_unlock(&glob->device_list_mutex);
1486
1487 if (!cancel_delayed_work(&bdev->wq))
1488 flush_scheduled_work();
1489
1490 while (ttm_bo_delayed_delete(bdev, true))
1491 ;
1492
1493 spin_lock(&glob->lru_lock);
1494 if (list_empty(&bdev->ddestroy))
1495 TTM_DEBUG("Delayed destroy list was clean\n");
1496
1497 if (list_empty(&bdev->man[0].lru))
1498 TTM_DEBUG("Swap list was clean\n");
1499 spin_unlock(&glob->lru_lock);
1500
1501 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1502 write_lock(&bdev->vm_lock);
1503 drm_mm_takedown(&bdev->addr_space_mm);
1504 write_unlock(&bdev->vm_lock);
1505
1506 return ret;
1507 }
1508 EXPORT_SYMBOL(ttm_bo_device_release);
1509
1510 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1511 struct ttm_bo_global *glob,
1512 struct ttm_bo_driver *driver,
1513 uint64_t file_page_offset,
1514 bool need_dma32)
1515 {
1516 int ret = -EINVAL;
1517
1518 rwlock_init(&bdev->vm_lock);
1519 bdev->driver = driver;
1520
1521 memset(bdev->man, 0, sizeof(bdev->man));
1522
1523 /*
1524 * Initialize the system memory buffer type.
1525 * Other types need to be driver / IOCTL initialized.
1526 */
1527 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1528 if (unlikely(ret != 0))
1529 goto out_no_sys;
1530
1531 bdev->addr_space_rb = RB_ROOT;
1532 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1533 if (unlikely(ret != 0))
1534 goto out_no_addr_mm;
1535
1536 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1537 bdev->nice_mode = true;
1538 INIT_LIST_HEAD(&bdev->ddestroy);
1539 bdev->dev_mapping = NULL;
1540 bdev->glob = glob;
1541 bdev->need_dma32 = need_dma32;
1542 bdev->val_seq = 0;
1543 spin_lock_init(&bdev->fence_lock);
1544 mutex_lock(&glob->device_list_mutex);
1545 list_add_tail(&bdev->device_list, &glob->device_list);
1546 mutex_unlock(&glob->device_list_mutex);
1547
1548 return 0;
1549 out_no_addr_mm:
1550 ttm_bo_clean_mm(bdev, 0);
1551 out_no_sys:
1552 return ret;
1553 }
1554 EXPORT_SYMBOL(ttm_bo_device_init);
1555
1556 /*
1557 * buffer object vm functions.
1558 */
1559
1560 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1561 {
1562 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1563
1564 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1565 if (mem->mem_type == TTM_PL_SYSTEM)
1566 return false;
1567
1568 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1569 return false;
1570
1571 if (mem->placement & TTM_PL_FLAG_CACHED)
1572 return false;
1573 }
1574 return true;
1575 }
1576
1577 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1578 {
1579 struct ttm_bo_device *bdev = bo->bdev;
1580 loff_t offset = (loff_t) bo->addr_space_offset;
1581 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1582
1583 if (!bdev->dev_mapping)
1584 return;
1585 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1586 ttm_mem_io_free(bdev, &bo->mem);
1587 }
1588 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1589
1590 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1591 {
1592 struct ttm_bo_device *bdev = bo->bdev;
1593 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1594 struct rb_node *parent = NULL;
1595 struct ttm_buffer_object *cur_bo;
1596 unsigned long offset = bo->vm_node->start;
1597 unsigned long cur_offset;
1598
1599 while (*cur) {
1600 parent = *cur;
1601 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1602 cur_offset = cur_bo->vm_node->start;
1603 if (offset < cur_offset)
1604 cur = &parent->rb_left;
1605 else if (offset > cur_offset)
1606 cur = &parent->rb_right;
1607 else
1608 BUG();
1609 }
1610
1611 rb_link_node(&bo->vm_rb, parent, cur);
1612 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1613 }
1614
1615 /**
1616 * ttm_bo_setup_vm:
1617 *
1618 * @bo: the buffer to allocate address space for
1619 *
1620 * Allocate address space in the drm device so that applications
1621 * can mmap the buffer and access the contents. This only
1622 * applies to ttm_bo_type_device objects as others are not
1623 * placed in the drm device address space.
1624 */
1625
1626 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1627 {
1628 struct ttm_bo_device *bdev = bo->bdev;
1629 int ret;
1630
1631 retry_pre_get:
1632 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1633 if (unlikely(ret != 0))
1634 return ret;
1635
1636 write_lock(&bdev->vm_lock);
1637 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1638 bo->mem.num_pages, 0, 0);
1639
1640 if (unlikely(bo->vm_node == NULL)) {
1641 ret = -ENOMEM;
1642 goto out_unlock;
1643 }
1644
1645 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1646 bo->mem.num_pages, 0);
1647
1648 if (unlikely(bo->vm_node == NULL)) {
1649 write_unlock(&bdev->vm_lock);
1650 goto retry_pre_get;
1651 }
1652
1653 ttm_bo_vm_insert_rb(bo);
1654 write_unlock(&bdev->vm_lock);
1655 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1656
1657 return 0;
1658 out_unlock:
1659 write_unlock(&bdev->vm_lock);
1660 return ret;
1661 }
1662
1663 int ttm_bo_wait(struct ttm_buffer_object *bo,
1664 bool lazy, bool interruptible, bool no_wait)
1665 {
1666 struct ttm_bo_driver *driver = bo->bdev->driver;
1667 struct ttm_bo_device *bdev = bo->bdev;
1668 void *sync_obj;
1669 void *sync_obj_arg;
1670 int ret = 0;
1671
1672 if (likely(bo->sync_obj == NULL))
1673 return 0;
1674
1675 while (bo->sync_obj) {
1676
1677 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1678 void *tmp_obj = bo->sync_obj;
1679 bo->sync_obj = NULL;
1680 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1681 spin_unlock(&bdev->fence_lock);
1682 driver->sync_obj_unref(&tmp_obj);
1683 spin_lock(&bdev->fence_lock);
1684 continue;
1685 }
1686
1687 if (no_wait)
1688 return -EBUSY;
1689
1690 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1691 sync_obj_arg = bo->sync_obj_arg;
1692 spin_unlock(&bdev->fence_lock);
1693 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1694 lazy, interruptible);
1695 if (unlikely(ret != 0)) {
1696 driver->sync_obj_unref(&sync_obj);
1697 spin_lock(&bdev->fence_lock);
1698 return ret;
1699 }
1700 spin_lock(&bdev->fence_lock);
1701 if (likely(bo->sync_obj == sync_obj &&
1702 bo->sync_obj_arg == sync_obj_arg)) {
1703 void *tmp_obj = bo->sync_obj;
1704 bo->sync_obj = NULL;
1705 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1706 &bo->priv_flags);
1707 spin_unlock(&bdev->fence_lock);
1708 driver->sync_obj_unref(&sync_obj);
1709 driver->sync_obj_unref(&tmp_obj);
1710 spin_lock(&bdev->fence_lock);
1711 } else {
1712 spin_unlock(&bdev->fence_lock);
1713 driver->sync_obj_unref(&sync_obj);
1714 spin_lock(&bdev->fence_lock);
1715 }
1716 }
1717 return 0;
1718 }
1719 EXPORT_SYMBOL(ttm_bo_wait);
1720
1721 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1722 {
1723 struct ttm_bo_device *bdev = bo->bdev;
1724 int ret = 0;
1725
1726 /*
1727 * Using ttm_bo_reserve makes sure the lru lists are updated.
1728 */
1729
1730 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1731 if (unlikely(ret != 0))
1732 return ret;
1733 spin_lock(&bdev->fence_lock);
1734 ret = ttm_bo_wait(bo, false, true, no_wait);
1735 spin_unlock(&bdev->fence_lock);
1736 if (likely(ret == 0))
1737 atomic_inc(&bo->cpu_writers);
1738 ttm_bo_unreserve(bo);
1739 return ret;
1740 }
1741 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1742
1743 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1744 {
1745 if (atomic_dec_and_test(&bo->cpu_writers))
1746 wake_up_all(&bo->event_queue);
1747 }
1748 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1749
1750 /**
1751 * A buffer object shrink method that tries to swap out the first
1752 * buffer object on the bo_global::swap_lru list.
1753 */
1754
1755 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1756 {
1757 struct ttm_bo_global *glob =
1758 container_of(shrink, struct ttm_bo_global, shrink);
1759 struct ttm_buffer_object *bo;
1760 int ret = -EBUSY;
1761 int put_count;
1762 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1763
1764 spin_lock(&glob->lru_lock);
1765 while (ret == -EBUSY) {
1766 if (unlikely(list_empty(&glob->swap_lru))) {
1767 spin_unlock(&glob->lru_lock);
1768 return -EBUSY;
1769 }
1770
1771 bo = list_first_entry(&glob->swap_lru,
1772 struct ttm_buffer_object, swap);
1773 kref_get(&bo->list_kref);
1774
1775 if (!list_empty(&bo->ddestroy)) {
1776 spin_unlock(&glob->lru_lock);
1777 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1778 kref_put(&bo->list_kref, ttm_bo_release_list);
1779 continue;
1780 }
1781
1782 /**
1783 * Reserve buffer. Since we unlock while sleeping, we need
1784 * to re-check that nobody removed us from the swap-list while
1785 * we slept.
1786 */
1787
1788 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1789 if (unlikely(ret == -EBUSY)) {
1790 spin_unlock(&glob->lru_lock);
1791 ttm_bo_wait_unreserved(bo, false);
1792 kref_put(&bo->list_kref, ttm_bo_release_list);
1793 spin_lock(&glob->lru_lock);
1794 }
1795 }
1796
1797 BUG_ON(ret != 0);
1798 put_count = ttm_bo_del_from_lru(bo);
1799 spin_unlock(&glob->lru_lock);
1800
1801 ttm_bo_list_ref_sub(bo, put_count, true);
1802
1803 /**
1804 * Wait for GPU, then move to system cached.
1805 */
1806
1807 spin_lock(&bo->bdev->fence_lock);
1808 ret = ttm_bo_wait(bo, false, false, false);
1809 spin_unlock(&bo->bdev->fence_lock);
1810
1811 if (unlikely(ret != 0))
1812 goto out;
1813
1814 if ((bo->mem.placement & swap_placement) != swap_placement) {
1815 struct ttm_mem_reg evict_mem;
1816
1817 evict_mem = bo->mem;
1818 evict_mem.mm_node = NULL;
1819 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1820 evict_mem.mem_type = TTM_PL_SYSTEM;
1821
1822 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1823 false, false, false);
1824 if (unlikely(ret != 0))
1825 goto out;
1826 }
1827
1828 ttm_bo_unmap_virtual(bo);
1829
1830 /**
1831 * Swap out. Buffer will be swapped in again as soon as
1832 * anyone tries to access a ttm page.
1833 */
1834
1835 if (bo->bdev->driver->swap_notify)
1836 bo->bdev->driver->swap_notify(bo);
1837
1838 ret = ttm_tt_swapout(bo->ttm, bo->persistant_swap_storage);
1839 out:
1840
1841 /**
1842 *
1843 * Unreserve without putting on LRU to avoid swapping out an
1844 * already swapped buffer.
1845 */
1846
1847 atomic_set(&bo->reserved, 0);
1848 wake_up_all(&bo->event_queue);
1849 kref_put(&bo->list_kref, ttm_bo_release_list);
1850 return ret;
1851 }
1852
1853 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1854 {
1855 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1856 ;
1857 }
1858 EXPORT_SYMBOL(ttm_bo_swapout_all);
Linux 2.6 ibm-acpi/thinkpad-acpi driver git tree