?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
drm/ttm: Add a bo list reserve fastpath (v2)
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / drivers / gpu / drm / ttm / ttm_bo.c
bloba586378b1b2b420110bebfc3f74c13d278d07e70
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 (sequence - bo->val_seq < (1 << 31))) {
228 return -EAGAIN;
229 }
230
231 if (no_wait)
232 return -EBUSY;
233
234 spin_unlock(&glob->lru_lock);
235 ret = ttm_bo_wait_unreserved(bo, interruptible);
236 spin_lock(&glob->lru_lock);
237
238 if (unlikely(ret))
239 return ret;
240 }
241
242 if (use_sequence) {
243 /**
244 * Wake up waiters that may need to recheck for deadlock,
245 * if we decreased the sequence number.
246 */
247 if (unlikely((bo->val_seq - sequence < (1 << 31))
248 || !bo->seq_valid))
249 wake_up_all(&bo->event_queue);
250
251 bo->val_seq = sequence;
252 bo->seq_valid = true;
253 } else {
254 bo->seq_valid = false;
255 }
256
257 return 0;
258 }
259 EXPORT_SYMBOL(ttm_bo_reserve);
260
261 static void ttm_bo_ref_bug(struct kref *list_kref)
262 {
263 BUG();
264 }
265
266 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
267 bool never_free)
268 {
269 while (count--)
270 kref_put(&bo->list_kref,
271 (never_free || (count >= 0)) ? ttm_bo_ref_bug :
272 ttm_bo_release_list);
273 }
274
275 int ttm_bo_reserve(struct ttm_buffer_object *bo,
276 bool interruptible,
277 bool no_wait, bool use_sequence, uint32_t sequence)
278 {
279 struct ttm_bo_global *glob = bo->glob;
280 int put_count = 0;
281 int ret;
282
283 spin_lock(&glob->lru_lock);
284 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
285 sequence);
286 if (likely(ret == 0))
287 put_count = ttm_bo_del_from_lru(bo);
288 spin_unlock(&glob->lru_lock);
289
290 ttm_bo_list_ref_sub(bo, put_count, true);
291
292 return ret;
293 }
294
295 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
296 {
297 struct ttm_bo_global *glob = bo->glob;
298
299 spin_lock(&glob->lru_lock);
300 ttm_bo_add_to_lru(bo);
301 atomic_set(&bo->reserved, 0);
302 wake_up_all(&bo->event_queue);
303 spin_unlock(&glob->lru_lock);
304 }
305 EXPORT_SYMBOL(ttm_bo_unreserve);
306
307 /*
308 * Call bo->mutex locked.
309 */
310 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
311 {
312 struct ttm_bo_device *bdev = bo->bdev;
313 struct ttm_bo_global *glob = bo->glob;
314 int ret = 0;
315 uint32_t page_flags = 0;
316
317 TTM_ASSERT_LOCKED(&bo->mutex);
318 bo->ttm = NULL;
319
320 if (bdev->need_dma32)
321 page_flags |= TTM_PAGE_FLAG_DMA32;
322
323 switch (bo->type) {
324 case ttm_bo_type_device:
325 if (zero_alloc)
326 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
327 case ttm_bo_type_kernel:
328 bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
329 page_flags, glob->dummy_read_page);
330 if (unlikely(bo->ttm == NULL))
331 ret = -ENOMEM;
332 break;
333 case ttm_bo_type_user:
334 bo->ttm = ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
335 page_flags | TTM_PAGE_FLAG_USER,
336 glob->dummy_read_page);
337 if (unlikely(bo->ttm == NULL)) {
338 ret = -ENOMEM;
339 break;
340 }
341
342 ret = ttm_tt_set_user(bo->ttm, current,
343 bo->buffer_start, bo->num_pages);
344 if (unlikely(ret != 0))
345 ttm_tt_destroy(bo->ttm);
346 break;
347 default:
348 printk(KERN_ERR TTM_PFX "Illegal buffer object type\n");
349 ret = -EINVAL;
350 break;
351 }
352
353 return ret;
354 }
355
356 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
357 struct ttm_mem_reg *mem,
358 bool evict, bool interruptible,
359 bool no_wait_reserve, bool no_wait_gpu)
360 {
361 struct ttm_bo_device *bdev = bo->bdev;
362 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
363 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
364 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
365 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
366 int ret = 0;
367
368 if (old_is_pci || new_is_pci ||
369 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0))
370 ttm_bo_unmap_virtual(bo);
371
372 /*
373 * Create and bind a ttm if required.
374 */
375
376 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && (bo->ttm == NULL)) {
377 ret = ttm_bo_add_ttm(bo, false);
378 if (ret)
379 goto out_err;
380
381 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
382 if (ret)
383 goto out_err;
384
385 if (mem->mem_type != TTM_PL_SYSTEM) {
386 ret = ttm_tt_bind(bo->ttm, mem);
387 if (ret)
388 goto out_err;
389 }
390
391 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
392 bo->mem = *mem;
393 mem->mm_node = NULL;
394 goto moved;
395 }
396
397 }
398
399 if (bdev->driver->move_notify)
400 bdev->driver->move_notify(bo, mem);
401
402 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
403 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
404 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
405 else if (bdev->driver->move)
406 ret = bdev->driver->move(bo, evict, interruptible,
407 no_wait_reserve, no_wait_gpu, mem);
408 else
409 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
410
411 if (ret)
412 goto out_err;
413
414 moved:
415 if (bo->evicted) {
416 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
417 if (ret)
418 printk(KERN_ERR TTM_PFX "Can not flush read caches\n");
419 bo->evicted = false;
420 }
421
422 if (bo->mem.mm_node) {
423 spin_lock(&bo->lock);
424 bo->offset = (bo->mem.start << PAGE_SHIFT) +
425 bdev->man[bo->mem.mem_type].gpu_offset;
426 bo->cur_placement = bo->mem.placement;
427 spin_unlock(&bo->lock);
428 } else
429 bo->offset = 0;
430
431 return 0;
432
433 out_err:
434 new_man = &bdev->man[bo->mem.mem_type];
435 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
436 ttm_tt_unbind(bo->ttm);
437 ttm_tt_destroy(bo->ttm);
438 bo->ttm = NULL;
439 }
440
441 return ret;
442 }
443
444 /**
445 * Call bo::reserved.
446 * Will release GPU memory type usage on destruction.
447 * This is the place to put in driver specific hooks to release
448 * driver private resources.
449 * Will release the bo::reserved lock.
450 */
451
452 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
453 {
454 if (bo->ttm) {
455 ttm_tt_unbind(bo->ttm);
456 ttm_tt_destroy(bo->ttm);
457 bo->ttm = NULL;
458 }
459
460 ttm_bo_mem_put(bo, &bo->mem);
461
462 atomic_set(&bo->reserved, 0);
463
464 /*
465 * Make processes trying to reserve really pick it up.
466 */
467 smp_mb__after_atomic_dec();
468 wake_up_all(&bo->event_queue);
469 }
470
471 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
472 {
473 struct ttm_bo_device *bdev = bo->bdev;
474 struct ttm_bo_global *glob = bo->glob;
475 struct ttm_bo_driver *driver;
476 void *sync_obj = NULL;
477 void *sync_obj_arg;
478 int put_count;
479 int ret;
480
481 spin_lock(&bo->lock);
482 (void) ttm_bo_wait(bo, false, false, true);
483 if (!bo->sync_obj) {
484
485 spin_lock(&glob->lru_lock);
486
487 /**
488 * Lock inversion between bo::reserve and bo::lock here,
489 * but that's OK, since we're only trylocking.
490 */
491
492 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
493
494 if (unlikely(ret == -EBUSY))
495 goto queue;
496
497 spin_unlock(&bo->lock);
498 put_count = ttm_bo_del_from_lru(bo);
499
500 spin_unlock(&glob->lru_lock);
501 ttm_bo_cleanup_memtype_use(bo);
502
503 ttm_bo_list_ref_sub(bo, put_count, true);
504
505 return;
506 } else {
507 spin_lock(&glob->lru_lock);
508 }
509 queue:
510 driver = bdev->driver;
511 if (bo->sync_obj)
512 sync_obj = driver->sync_obj_ref(bo->sync_obj);
513 sync_obj_arg = bo->sync_obj_arg;
514
515 kref_get(&bo->list_kref);
516 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
517 spin_unlock(&glob->lru_lock);
518 spin_unlock(&bo->lock);
519
520 if (sync_obj) {
521 driver->sync_obj_flush(sync_obj, sync_obj_arg);
522 driver->sync_obj_unref(&sync_obj);
523 }
524 schedule_delayed_work(&bdev->wq,
525 ((HZ / 100) < 1) ? 1 : HZ / 100);
526 }
527
528 /**
529 * function ttm_bo_cleanup_refs
530 * If bo idle, remove from delayed- and lru lists, and unref.
531 * If not idle, do nothing.
532 *
533 * @interruptible Any sleeps should occur interruptibly.
534 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
535 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
536 */
537
538 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
539 bool interruptible,
540 bool no_wait_reserve,
541 bool no_wait_gpu)
542 {
543 struct ttm_bo_global *glob = bo->glob;
544 int put_count;
545 int ret = 0;
546
547 retry:
548 spin_lock(&bo->lock);
549 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
550 spin_unlock(&bo->lock);
551
552 if (unlikely(ret != 0))
553 return ret;
554
555 spin_lock(&glob->lru_lock);
556 ret = ttm_bo_reserve_locked(bo, interruptible,
557 no_wait_reserve, false, 0);
558
559 if (unlikely(ret != 0) || list_empty(&bo->ddestroy)) {
560 spin_unlock(&glob->lru_lock);
561 return ret;
562 }
563
564 /**
565 * We can re-check for sync object without taking
566 * the bo::lock since setting the sync object requires
567 * also bo::reserved. A busy object at this point may
568 * be caused by another thread recently starting an accelerated
569 * eviction.
570 */
571
572 if (unlikely(bo->sync_obj)) {
573 atomic_set(&bo->reserved, 0);
574 wake_up_all(&bo->event_queue);
575 spin_unlock(&glob->lru_lock);
576 goto retry;
577 }
578
579 put_count = ttm_bo_del_from_lru(bo);
580 list_del_init(&bo->ddestroy);
581 ++put_count;
582
583 spin_unlock(&glob->lru_lock);
584 ttm_bo_cleanup_memtype_use(bo);
585
586 ttm_bo_list_ref_sub(bo, put_count, true);
587
588 return 0;
589 }
590
591 /**
592 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
593 * encountered buffers.
594 */
595
596 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
597 {
598 struct ttm_bo_global *glob = bdev->glob;
599 struct ttm_buffer_object *entry = NULL;
600 int ret = 0;
601
602 spin_lock(&glob->lru_lock);
603 if (list_empty(&bdev->ddestroy))
604 goto out_unlock;
605
606 entry = list_first_entry(&bdev->ddestroy,
607 struct ttm_buffer_object, ddestroy);
608 kref_get(&entry->list_kref);
609
610 for (;;) {
611 struct ttm_buffer_object *nentry = NULL;
612
613 if (entry->ddestroy.next != &bdev->ddestroy) {
614 nentry = list_first_entry(&entry->ddestroy,
615 struct ttm_buffer_object, ddestroy);
616 kref_get(&nentry->list_kref);
617 }
618
619 spin_unlock(&glob->lru_lock);
620 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
621 !remove_all);
622 kref_put(&entry->list_kref, ttm_bo_release_list);
623 entry = nentry;
624
625 if (ret || !entry)
626 goto out;
627
628 spin_lock(&glob->lru_lock);
629 if (list_empty(&entry->ddestroy))
630 break;
631 }
632
633 out_unlock:
634 spin_unlock(&glob->lru_lock);
635 out:
636 if (entry)
637 kref_put(&entry->list_kref, ttm_bo_release_list);
638 return ret;
639 }
640
641 static void ttm_bo_delayed_workqueue(struct work_struct *work)
642 {
643 struct ttm_bo_device *bdev =
644 container_of(work, struct ttm_bo_device, wq.work);
645
646 if (ttm_bo_delayed_delete(bdev, false)) {
647 schedule_delayed_work(&bdev->wq,
648 ((HZ / 100) < 1) ? 1 : HZ / 100);
649 }
650 }
651
652 static void ttm_bo_release(struct kref *kref)
653 {
654 struct ttm_buffer_object *bo =
655 container_of(kref, struct ttm_buffer_object, kref);
656 struct ttm_bo_device *bdev = bo->bdev;
657
658 if (likely(bo->vm_node != NULL)) {
659 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
660 drm_mm_put_block(bo->vm_node);
661 bo->vm_node = NULL;
662 }
663 write_unlock(&bdev->vm_lock);
664 ttm_bo_cleanup_refs_or_queue(bo);
665 kref_put(&bo->list_kref, ttm_bo_release_list);
666 write_lock(&bdev->vm_lock);
667 }
668
669 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
670 {
671 struct ttm_buffer_object *bo = *p_bo;
672 struct ttm_bo_device *bdev = bo->bdev;
673
674 *p_bo = NULL;
675 write_lock(&bdev->vm_lock);
676 kref_put(&bo->kref, ttm_bo_release);
677 write_unlock(&bdev->vm_lock);
678 }
679 EXPORT_SYMBOL(ttm_bo_unref);
680
681 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
682 {
683 return cancel_delayed_work_sync(&bdev->wq);
684 }
685 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
686
687 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
688 {
689 if (resched)
690 schedule_delayed_work(&bdev->wq,
691 ((HZ / 100) < 1) ? 1 : HZ / 100);
692 }
693 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
694
695 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
696 bool no_wait_reserve, bool no_wait_gpu)
697 {
698 struct ttm_bo_device *bdev = bo->bdev;
699 struct ttm_mem_reg evict_mem;
700 struct ttm_placement placement;
701 int ret = 0;
702
703 spin_lock(&bo->lock);
704 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
705 spin_unlock(&bo->lock);
706
707 if (unlikely(ret != 0)) {
708 if (ret != -ERESTARTSYS) {
709 printk(KERN_ERR TTM_PFX
710 "Failed to expire sync object before "
711 "buffer eviction.\n");
712 }
713 goto out;
714 }
715
716 BUG_ON(!atomic_read(&bo->reserved));
717
718 evict_mem = bo->mem;
719 evict_mem.mm_node = NULL;
720 evict_mem.bus.io_reserved = false;
721
722 placement.fpfn = 0;
723 placement.lpfn = 0;
724 placement.num_placement = 0;
725 placement.num_busy_placement = 0;
726 bdev->driver->evict_flags(bo, &placement);
727 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
728 no_wait_reserve, no_wait_gpu);
729 if (ret) {
730 if (ret != -ERESTARTSYS) {
731 printk(KERN_ERR TTM_PFX
732 "Failed to find memory space for "
733 "buffer 0x%p eviction.\n", bo);
734 ttm_bo_mem_space_debug(bo, &placement);
735 }
736 goto out;
737 }
738
739 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
740 no_wait_reserve, no_wait_gpu);
741 if (ret) {
742 if (ret != -ERESTARTSYS)
743 printk(KERN_ERR TTM_PFX "Buffer eviction failed\n");
744 ttm_bo_mem_put(bo, &evict_mem);
745 goto out;
746 }
747 bo->evicted = true;
748 out:
749 return ret;
750 }
751
752 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
753 uint32_t mem_type,
754 bool interruptible, bool no_wait_reserve,
755 bool no_wait_gpu)
756 {
757 struct ttm_bo_global *glob = bdev->glob;
758 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
759 struct ttm_buffer_object *bo;
760 int ret, put_count = 0;
761
762 retry:
763 spin_lock(&glob->lru_lock);
764 if (list_empty(&man->lru)) {
765 spin_unlock(&glob->lru_lock);
766 return -EBUSY;
767 }
768
769 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
770 kref_get(&bo->list_kref);
771
772 if (!list_empty(&bo->ddestroy)) {
773 spin_unlock(&glob->lru_lock);
774 ret = ttm_bo_cleanup_refs(bo, interruptible,
775 no_wait_reserve, no_wait_gpu);
776 kref_put(&bo->list_kref, ttm_bo_release_list);
777
778 if (likely(ret == 0 || ret == -ERESTARTSYS))
779 return ret;
780
781 goto retry;
782 }
783
784 ret = ttm_bo_reserve_locked(bo, false, no_wait_reserve, false, 0);
785
786 if (unlikely(ret == -EBUSY)) {
787 spin_unlock(&glob->lru_lock);
788 if (likely(!no_wait_gpu))
789 ret = ttm_bo_wait_unreserved(bo, interruptible);
790
791 kref_put(&bo->list_kref, ttm_bo_release_list);
792
793 /**
794 * We *need* to retry after releasing the lru lock.
795 */
796
797 if (unlikely(ret != 0))
798 return ret;
799 goto retry;
800 }
801
802 put_count = ttm_bo_del_from_lru(bo);
803 spin_unlock(&glob->lru_lock);
804
805 BUG_ON(ret != 0);
806
807 ttm_bo_list_ref_sub(bo, put_count, true);
808
809 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
810 ttm_bo_unreserve(bo);
811
812 kref_put(&bo->list_kref, ttm_bo_release_list);
813 return ret;
814 }
815
816 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
817 {
818 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
819
820 if (mem->mm_node)
821 (*man->func->put_node)(man, mem);
822 }
823 EXPORT_SYMBOL(ttm_bo_mem_put);
824
825 /**
826 * Repeatedly evict memory from the LRU for @mem_type until we create enough
827 * space, or we've evicted everything and there isn't enough space.
828 */
829 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
830 uint32_t mem_type,
831 struct ttm_placement *placement,
832 struct ttm_mem_reg *mem,
833 bool interruptible,
834 bool no_wait_reserve,
835 bool no_wait_gpu)
836 {
837 struct ttm_bo_device *bdev = bo->bdev;
838 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
839 int ret;
840
841 do {
842 ret = (*man->func->get_node)(man, bo, placement, mem);
843 if (unlikely(ret != 0))
844 return ret;
845 if (mem->mm_node)
846 break;
847 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
848 no_wait_reserve, no_wait_gpu);
849 if (unlikely(ret != 0))
850 return ret;
851 } while (1);
852 if (mem->mm_node == NULL)
853 return -ENOMEM;
854 mem->mem_type = mem_type;
855 return 0;
856 }
857
858 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
859 uint32_t cur_placement,
860 uint32_t proposed_placement)
861 {
862 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
863 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
864
865 /**
866 * Keep current caching if possible.
867 */
868
869 if ((cur_placement & caching) != 0)
870 result |= (cur_placement & caching);
871 else if ((man->default_caching & caching) != 0)
872 result |= man->default_caching;
873 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
874 result |= TTM_PL_FLAG_CACHED;
875 else if ((TTM_PL_FLAG_WC & caching) != 0)
876 result |= TTM_PL_FLAG_WC;
877 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
878 result |= TTM_PL_FLAG_UNCACHED;
879
880 return result;
881 }
882
883 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
884 bool disallow_fixed,
885 uint32_t mem_type,
886 uint32_t proposed_placement,
887 uint32_t *masked_placement)
888 {
889 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
890
891 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && disallow_fixed)
892 return false;
893
894 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
895 return false;
896
897 if ((proposed_placement & man->available_caching) == 0)
898 return false;
899
900 cur_flags |= (proposed_placement & man->available_caching);
901
902 *masked_placement = cur_flags;
903 return true;
904 }
905
906 /**
907 * Creates space for memory region @mem according to its type.
908 *
909 * This function first searches for free space in compatible memory types in
910 * the priority order defined by the driver. If free space isn't found, then
911 * ttm_bo_mem_force_space is attempted in priority order to evict and find
912 * space.
913 */
914 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
915 struct ttm_placement *placement,
916 struct ttm_mem_reg *mem,
917 bool interruptible, bool no_wait_reserve,
918 bool no_wait_gpu)
919 {
920 struct ttm_bo_device *bdev = bo->bdev;
921 struct ttm_mem_type_manager *man;
922 uint32_t mem_type = TTM_PL_SYSTEM;
923 uint32_t cur_flags = 0;
924 bool type_found = false;
925 bool type_ok = false;
926 bool has_erestartsys = false;
927 int i, ret;
928
929 mem->mm_node = NULL;
930 for (i = 0; i < placement->num_placement; ++i) {
931 ret = ttm_mem_type_from_flags(placement->placement[i],
932 &mem_type);
933 if (ret)
934 return ret;
935 man = &bdev->man[mem_type];
936
937 type_ok = ttm_bo_mt_compatible(man,
938 bo->type == ttm_bo_type_user,
939 mem_type,
940 placement->placement[i],
941 &cur_flags);
942
943 if (!type_ok)
944 continue;
945
946 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
947 cur_flags);
948 /*
949 * Use the access and other non-mapping-related flag bits from
950 * the memory placement flags to the current flags
951 */
952 ttm_flag_masked(&cur_flags, placement->placement[i],
953 ~TTM_PL_MASK_MEMTYPE);
954
955 if (mem_type == TTM_PL_SYSTEM)
956 break;
957
958 if (man->has_type && man->use_type) {
959 type_found = true;
960 ret = (*man->func->get_node)(man, bo, placement, mem);
961 if (unlikely(ret))
962 return ret;
963 }
964 if (mem->mm_node)
965 break;
966 }
967
968 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
969 mem->mem_type = mem_type;
970 mem->placement = cur_flags;
971 return 0;
972 }
973
974 if (!type_found)
975 return -EINVAL;
976
977 for (i = 0; i < placement->num_busy_placement; ++i) {
978 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
979 &mem_type);
980 if (ret)
981 return ret;
982 man = &bdev->man[mem_type];
983 if (!man->has_type)
984 continue;
985 if (!ttm_bo_mt_compatible(man,
986 bo->type == ttm_bo_type_user,
987 mem_type,
988 placement->busy_placement[i],
989 &cur_flags))
990 continue;
991
992 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
993 cur_flags);
994 /*
995 * Use the access and other non-mapping-related flag bits from
996 * the memory placement flags to the current flags
997 */
998 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
999 ~TTM_PL_MASK_MEMTYPE);
1000
1001
1002 if (mem_type == TTM_PL_SYSTEM) {
1003 mem->mem_type = mem_type;
1004 mem->placement = cur_flags;
1005 mem->mm_node = NULL;
1006 return 0;
1007 }
1008
1009 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1010 interruptible, no_wait_reserve, no_wait_gpu);
1011 if (ret == 0 && mem->mm_node) {
1012 mem->placement = cur_flags;
1013 return 0;
1014 }
1015 if (ret == -ERESTARTSYS)
1016 has_erestartsys = true;
1017 }
1018 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1019 return ret;
1020 }
1021 EXPORT_SYMBOL(ttm_bo_mem_space);
1022
1023 int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1024 {
1025 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1026 return -EBUSY;
1027
1028 return wait_event_interruptible(bo->event_queue,
1029 atomic_read(&bo->cpu_writers) == 0);
1030 }
1031 EXPORT_SYMBOL(ttm_bo_wait_cpu);
1032
1033 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1034 struct ttm_placement *placement,
1035 bool interruptible, bool no_wait_reserve,
1036 bool no_wait_gpu)
1037 {
1038 int ret = 0;
1039 struct ttm_mem_reg mem;
1040
1041 BUG_ON(!atomic_read(&bo->reserved));
1042
1043 /*
1044 * FIXME: It's possible to pipeline buffer moves.
1045 * Have the driver move function wait for idle when necessary,
1046 * instead of doing it here.
1047 */
1048 spin_lock(&bo->lock);
1049 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1050 spin_unlock(&bo->lock);
1051 if (ret)
1052 return ret;
1053 mem.num_pages = bo->num_pages;
1054 mem.size = mem.num_pages << PAGE_SHIFT;
1055 mem.page_alignment = bo->mem.page_alignment;
1056 mem.bus.io_reserved = false;
1057 /*
1058 * Determine where to move the buffer.
1059 */
1060 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1061 if (ret)
1062 goto out_unlock;
1063 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1064 out_unlock:
1065 if (ret && mem.mm_node)
1066 ttm_bo_mem_put(bo, &mem);
1067 return ret;
1068 }
1069
1070 static int ttm_bo_mem_compat(struct ttm_placement *placement,
1071 struct ttm_mem_reg *mem)
1072 {
1073 int i;
1074
1075 if (mem->mm_node && placement->lpfn != 0 &&
1076 (mem->start < placement->fpfn ||
1077 mem->start + mem->num_pages > placement->lpfn))
1078 return -1;
1079
1080 for (i = 0; i < placement->num_placement; i++) {
1081 if ((placement->placement[i] & mem->placement &
1082 TTM_PL_MASK_CACHING) &&
1083 (placement->placement[i] & mem->placement &
1084 TTM_PL_MASK_MEM))
1085 return i;
1086 }
1087 return -1;
1088 }
1089
1090 int ttm_bo_validate(struct ttm_buffer_object *bo,
1091 struct ttm_placement *placement,
1092 bool interruptible, bool no_wait_reserve,
1093 bool no_wait_gpu)
1094 {
1095 int ret;
1096
1097 BUG_ON(!atomic_read(&bo->reserved));
1098 /* Check that range is valid */
1099 if (placement->lpfn || placement->fpfn)
1100 if (placement->fpfn > placement->lpfn ||
1101 (placement->lpfn - placement->fpfn) < bo->num_pages)
1102 return -EINVAL;
1103 /*
1104 * Check whether we need to move buffer.
1105 */
1106 ret = ttm_bo_mem_compat(placement, &bo->mem);
1107 if (ret < 0) {
1108 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1109 if (ret)
1110 return ret;
1111 } else {
1112 /*
1113 * Use the access and other non-mapping-related flag bits from
1114 * the compatible memory placement flags to the active flags
1115 */
1116 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1117 ~TTM_PL_MASK_MEMTYPE);
1118 }
1119 /*
1120 * We might need to add a TTM.
1121 */
1122 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1123 ret = ttm_bo_add_ttm(bo, true);
1124 if (ret)
1125 return ret;
1126 }
1127 return 0;
1128 }
1129 EXPORT_SYMBOL(ttm_bo_validate);
1130
1131 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1132 struct ttm_placement *placement)
1133 {
1134 BUG_ON((placement->fpfn || placement->lpfn) &&
1135 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1136
1137 return 0;
1138 }
1139
1140 int ttm_bo_init(struct ttm_bo_device *bdev,
1141 struct ttm_buffer_object *bo,
1142 unsigned long size,
1143 enum ttm_bo_type type,
1144 struct ttm_placement *placement,
1145 uint32_t page_alignment,
1146 unsigned long buffer_start,
1147 bool interruptible,
1148 struct file *persistant_swap_storage,
1149 size_t acc_size,
1150 void (*destroy) (struct ttm_buffer_object *))
1151 {
1152 int ret = 0;
1153 unsigned long num_pages;
1154
1155 size += buffer_start & ~PAGE_MASK;
1156 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1157 if (num_pages == 0) {
1158 printk(KERN_ERR TTM_PFX "Illegal buffer object size.\n");
1159 if (destroy)
1160 (*destroy)(bo);
1161 else
1162 kfree(bo);
1163 return -EINVAL;
1164 }
1165 bo->destroy = destroy;
1166
1167 spin_lock_init(&bo->lock);
1168 kref_init(&bo->kref);
1169 kref_init(&bo->list_kref);
1170 atomic_set(&bo->cpu_writers, 0);
1171 atomic_set(&bo->reserved, 1);
1172 init_waitqueue_head(&bo->event_queue);
1173 INIT_LIST_HEAD(&bo->lru);
1174 INIT_LIST_HEAD(&bo->ddestroy);
1175 INIT_LIST_HEAD(&bo->swap);
1176 bo->bdev = bdev;
1177 bo->glob = bdev->glob;
1178 bo->type = type;
1179 bo->num_pages = num_pages;
1180 bo->mem.size = num_pages << PAGE_SHIFT;
1181 bo->mem.mem_type = TTM_PL_SYSTEM;
1182 bo->mem.num_pages = bo->num_pages;
1183 bo->mem.mm_node = NULL;
1184 bo->mem.page_alignment = page_alignment;
1185 bo->mem.bus.io_reserved = false;
1186 bo->buffer_start = buffer_start & PAGE_MASK;
1187 bo->priv_flags = 0;
1188 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1189 bo->seq_valid = false;
1190 bo->persistant_swap_storage = persistant_swap_storage;
1191 bo->acc_size = acc_size;
1192 atomic_inc(&bo->glob->bo_count);
1193
1194 ret = ttm_bo_check_placement(bo, placement);
1195 if (unlikely(ret != 0))
1196 goto out_err;
1197
1198 /*
1199 * For ttm_bo_type_device buffers, allocate
1200 * address space from the device.
1201 */
1202 if (bo->type == ttm_bo_type_device) {
1203 ret = ttm_bo_setup_vm(bo);
1204 if (ret)
1205 goto out_err;
1206 }
1207
1208 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1209 if (ret)
1210 goto out_err;
1211
1212 ttm_bo_unreserve(bo);
1213 return 0;
1214
1215 out_err:
1216 ttm_bo_unreserve(bo);
1217 ttm_bo_unref(&bo);
1218
1219 return ret;
1220 }
1221 EXPORT_SYMBOL(ttm_bo_init);
1222
1223 static inline size_t ttm_bo_size(struct ttm_bo_global *glob,
1224 unsigned long num_pages)
1225 {
1226 size_t page_array_size = (num_pages * sizeof(void *) + PAGE_SIZE - 1) &
1227 PAGE_MASK;
1228
1229 return glob->ttm_bo_size + 2 * page_array_size;
1230 }
1231
1232 int ttm_bo_create(struct ttm_bo_device *bdev,
1233 unsigned long size,
1234 enum ttm_bo_type type,
1235 struct ttm_placement *placement,
1236 uint32_t page_alignment,
1237 unsigned long buffer_start,
1238 bool interruptible,
1239 struct file *persistant_swap_storage,
1240 struct ttm_buffer_object **p_bo)
1241 {
1242 struct ttm_buffer_object *bo;
1243 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1244 int ret;
1245
1246 size_t acc_size =
1247 ttm_bo_size(bdev->glob, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
1248 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1249 if (unlikely(ret != 0))
1250 return ret;
1251
1252 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1253
1254 if (unlikely(bo == NULL)) {
1255 ttm_mem_global_free(mem_glob, acc_size);
1256 return -ENOMEM;
1257 }
1258
1259 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1260 buffer_start, interruptible,
1261 persistant_swap_storage, acc_size, NULL);
1262 if (likely(ret == 0))
1263 *p_bo = bo;
1264
1265 return ret;
1266 }
1267
1268 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1269 unsigned mem_type, bool allow_errors)
1270 {
1271 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1272 struct ttm_bo_global *glob = bdev->glob;
1273 int ret;
1274
1275 /*
1276 * Can't use standard list traversal since we're unlocking.
1277 */
1278
1279 spin_lock(&glob->lru_lock);
1280 while (!list_empty(&man->lru)) {
1281 spin_unlock(&glob->lru_lock);
1282 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1283 if (ret) {
1284 if (allow_errors) {
1285 return ret;
1286 } else {
1287 printk(KERN_ERR TTM_PFX
1288 "Cleanup eviction failed\n");
1289 }
1290 }
1291 spin_lock(&glob->lru_lock);
1292 }
1293 spin_unlock(&glob->lru_lock);
1294 return 0;
1295 }
1296
1297 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1298 {
1299 struct ttm_mem_type_manager *man;
1300 int ret = -EINVAL;
1301
1302 if (mem_type >= TTM_NUM_MEM_TYPES) {
1303 printk(KERN_ERR TTM_PFX "Illegal memory type %d\n", mem_type);
1304 return ret;
1305 }
1306 man = &bdev->man[mem_type];
1307
1308 if (!man->has_type) {
1309 printk(KERN_ERR TTM_PFX "Trying to take down uninitialized "
1310 "memory manager type %u\n", mem_type);
1311 return ret;
1312 }
1313
1314 man->use_type = false;
1315 man->has_type = false;
1316
1317 ret = 0;
1318 if (mem_type > 0) {
1319 ttm_bo_force_list_clean(bdev, mem_type, false);
1320
1321 ret = (*man->func->takedown)(man);
1322 }
1323
1324 return ret;
1325 }
1326 EXPORT_SYMBOL(ttm_bo_clean_mm);
1327
1328 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1329 {
1330 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1331
1332 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1333 printk(KERN_ERR TTM_PFX
1334 "Illegal memory manager memory type %u.\n",
1335 mem_type);
1336 return -EINVAL;
1337 }
1338
1339 if (!man->has_type) {
1340 printk(KERN_ERR TTM_PFX
1341 "Memory type %u has not been initialized.\n",
1342 mem_type);
1343 return 0;
1344 }
1345
1346 return ttm_bo_force_list_clean(bdev, mem_type, true);
1347 }
1348 EXPORT_SYMBOL(ttm_bo_evict_mm);
1349
1350 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1351 unsigned long p_size)
1352 {
1353 int ret = -EINVAL;
1354 struct ttm_mem_type_manager *man;
1355
1356 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1357 man = &bdev->man[type];
1358 BUG_ON(man->has_type);
1359
1360 ret = bdev->driver->init_mem_type(bdev, type, man);
1361 if (ret)
1362 return ret;
1363 man->bdev = bdev;
1364
1365 ret = 0;
1366 if (type != TTM_PL_SYSTEM) {
1367 ret = (*man->func->init)(man, p_size);
1368 if (ret)
1369 return ret;
1370 }
1371 man->has_type = true;
1372 man->use_type = true;
1373 man->size = p_size;
1374
1375 INIT_LIST_HEAD(&man->lru);
1376
1377 return 0;
1378 }
1379 EXPORT_SYMBOL(ttm_bo_init_mm);
1380
1381 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1382 {
1383 struct ttm_bo_global *glob =
1384 container_of(kobj, struct ttm_bo_global, kobj);
1385
1386 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1387 __free_page(glob->dummy_read_page);
1388 kfree(glob);
1389 }
1390
1391 void ttm_bo_global_release(struct drm_global_reference *ref)
1392 {
1393 struct ttm_bo_global *glob = ref->object;
1394
1395 kobject_del(&glob->kobj);
1396 kobject_put(&glob->kobj);
1397 }
1398 EXPORT_SYMBOL(ttm_bo_global_release);
1399
1400 int ttm_bo_global_init(struct drm_global_reference *ref)
1401 {
1402 struct ttm_bo_global_ref *bo_ref =
1403 container_of(ref, struct ttm_bo_global_ref, ref);
1404 struct ttm_bo_global *glob = ref->object;
1405 int ret;
1406
1407 mutex_init(&glob->device_list_mutex);
1408 spin_lock_init(&glob->lru_lock);
1409 glob->mem_glob = bo_ref->mem_glob;
1410 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1411
1412 if (unlikely(glob->dummy_read_page == NULL)) {
1413 ret = -ENOMEM;
1414 goto out_no_drp;
1415 }
1416
1417 INIT_LIST_HEAD(&glob->swap_lru);
1418 INIT_LIST_HEAD(&glob->device_list);
1419
1420 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1421 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1422 if (unlikely(ret != 0)) {
1423 printk(KERN_ERR TTM_PFX
1424 "Could not register buffer object swapout.\n");
1425 goto out_no_shrink;
1426 }
1427
1428 glob->ttm_bo_extra_size =
1429 ttm_round_pot(sizeof(struct ttm_tt)) +
1430 ttm_round_pot(sizeof(struct ttm_backend));
1431
1432 glob->ttm_bo_size = glob->ttm_bo_extra_size +
1433 ttm_round_pot(sizeof(struct ttm_buffer_object));
1434
1435 atomic_set(&glob->bo_count, 0);
1436
1437 ret = kobject_init_and_add(
1438 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1439 if (unlikely(ret != 0))
1440 kobject_put(&glob->kobj);
1441 return ret;
1442 out_no_shrink:
1443 __free_page(glob->dummy_read_page);
1444 out_no_drp:
1445 kfree(glob);
1446 return ret;
1447 }
1448 EXPORT_SYMBOL(ttm_bo_global_init);
1449
1450
1451 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1452 {
1453 int ret = 0;
1454 unsigned i = TTM_NUM_MEM_TYPES;
1455 struct ttm_mem_type_manager *man;
1456 struct ttm_bo_global *glob = bdev->glob;
1457
1458 while (i--) {
1459 man = &bdev->man[i];
1460 if (man->has_type) {
1461 man->use_type = false;
1462 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1463 ret = -EBUSY;
1464 printk(KERN_ERR TTM_PFX
1465 "DRM memory manager type %d "
1466 "is not clean.\n", i);
1467 }
1468 man->has_type = false;
1469 }
1470 }
1471
1472 mutex_lock(&glob->device_list_mutex);
1473 list_del(&bdev->device_list);
1474 mutex_unlock(&glob->device_list_mutex);
1475
1476 if (!cancel_delayed_work(&bdev->wq))
1477 flush_scheduled_work();
1478
1479 while (ttm_bo_delayed_delete(bdev, true))
1480 ;
1481
1482 spin_lock(&glob->lru_lock);
1483 if (list_empty(&bdev->ddestroy))
1484 TTM_DEBUG("Delayed destroy list was clean\n");
1485
1486 if (list_empty(&bdev->man[0].lru))
1487 TTM_DEBUG("Swap list was clean\n");
1488 spin_unlock(&glob->lru_lock);
1489
1490 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1491 write_lock(&bdev->vm_lock);
1492 drm_mm_takedown(&bdev->addr_space_mm);
1493 write_unlock(&bdev->vm_lock);
1494
1495 return ret;
1496 }
1497 EXPORT_SYMBOL(ttm_bo_device_release);
1498
1499 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1500 struct ttm_bo_global *glob,
1501 struct ttm_bo_driver *driver,
1502 uint64_t file_page_offset,
1503 bool need_dma32)
1504 {
1505 int ret = -EINVAL;
1506
1507 rwlock_init(&bdev->vm_lock);
1508 bdev->driver = driver;
1509
1510 memset(bdev->man, 0, sizeof(bdev->man));
1511
1512 /*
1513 * Initialize the system memory buffer type.
1514 * Other types need to be driver / IOCTL initialized.
1515 */
1516 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1517 if (unlikely(ret != 0))
1518 goto out_no_sys;
1519
1520 bdev->addr_space_rb = RB_ROOT;
1521 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1522 if (unlikely(ret != 0))
1523 goto out_no_addr_mm;
1524
1525 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1526 bdev->nice_mode = true;
1527 INIT_LIST_HEAD(&bdev->ddestroy);
1528 bdev->dev_mapping = NULL;
1529 bdev->glob = glob;
1530 bdev->need_dma32 = need_dma32;
1531
1532 mutex_lock(&glob->device_list_mutex);
1533 list_add_tail(&bdev->device_list, &glob->device_list);
1534 mutex_unlock(&glob->device_list_mutex);
1535
1536 return 0;
1537 out_no_addr_mm:
1538 ttm_bo_clean_mm(bdev, 0);
1539 out_no_sys:
1540 return ret;
1541 }
1542 EXPORT_SYMBOL(ttm_bo_device_init);
1543
1544 /*
1545 * buffer object vm functions.
1546 */
1547
1548 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1549 {
1550 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1551
1552 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1553 if (mem->mem_type == TTM_PL_SYSTEM)
1554 return false;
1555
1556 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1557 return false;
1558
1559 if (mem->placement & TTM_PL_FLAG_CACHED)
1560 return false;
1561 }
1562 return true;
1563 }
1564
1565 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1566 {
1567 struct ttm_bo_device *bdev = bo->bdev;
1568 loff_t offset = (loff_t) bo->addr_space_offset;
1569 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1570
1571 if (!bdev->dev_mapping)
1572 return;
1573 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1574 ttm_mem_io_free(bdev, &bo->mem);
1575 }
1576 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1577
1578 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1579 {
1580 struct ttm_bo_device *bdev = bo->bdev;
1581 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1582 struct rb_node *parent = NULL;
1583 struct ttm_buffer_object *cur_bo;
1584 unsigned long offset = bo->vm_node->start;
1585 unsigned long cur_offset;
1586
1587 while (*cur) {
1588 parent = *cur;
1589 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1590 cur_offset = cur_bo->vm_node->start;
1591 if (offset < cur_offset)
1592 cur = &parent->rb_left;
1593 else if (offset > cur_offset)
1594 cur = &parent->rb_right;
1595 else
1596 BUG();
1597 }
1598
1599 rb_link_node(&bo->vm_rb, parent, cur);
1600 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1601 }
1602
1603 /**
1604 * ttm_bo_setup_vm:
1605 *
1606 * @bo: the buffer to allocate address space for
1607 *
1608 * Allocate address space in the drm device so that applications
1609 * can mmap the buffer and access the contents. This only
1610 * applies to ttm_bo_type_device objects as others are not
1611 * placed in the drm device address space.
1612 */
1613
1614 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1615 {
1616 struct ttm_bo_device *bdev = bo->bdev;
1617 int ret;
1618
1619 retry_pre_get:
1620 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1621 if (unlikely(ret != 0))
1622 return ret;
1623
1624 write_lock(&bdev->vm_lock);
1625 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1626 bo->mem.num_pages, 0, 0);
1627
1628 if (unlikely(bo->vm_node == NULL)) {
1629 ret = -ENOMEM;
1630 goto out_unlock;
1631 }
1632
1633 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1634 bo->mem.num_pages, 0);
1635
1636 if (unlikely(bo->vm_node == NULL)) {
1637 write_unlock(&bdev->vm_lock);
1638 goto retry_pre_get;
1639 }
1640
1641 ttm_bo_vm_insert_rb(bo);
1642 write_unlock(&bdev->vm_lock);
1643 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1644
1645 return 0;
1646 out_unlock:
1647 write_unlock(&bdev->vm_lock);
1648 return ret;
1649 }
1650
1651 int ttm_bo_wait(struct ttm_buffer_object *bo,
1652 bool lazy, bool interruptible, bool no_wait)
1653 {
1654 struct ttm_bo_driver *driver = bo->bdev->driver;
1655 void *sync_obj;
1656 void *sync_obj_arg;
1657 int ret = 0;
1658
1659 if (likely(bo->sync_obj == NULL))
1660 return 0;
1661
1662 while (bo->sync_obj) {
1663
1664 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1665 void *tmp_obj = bo->sync_obj;
1666 bo->sync_obj = NULL;
1667 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1668 spin_unlock(&bo->lock);
1669 driver->sync_obj_unref(&tmp_obj);
1670 spin_lock(&bo->lock);
1671 continue;
1672 }
1673
1674 if (no_wait)
1675 return -EBUSY;
1676
1677 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1678 sync_obj_arg = bo->sync_obj_arg;
1679 spin_unlock(&bo->lock);
1680 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1681 lazy, interruptible);
1682 if (unlikely(ret != 0)) {
1683 driver->sync_obj_unref(&sync_obj);
1684 spin_lock(&bo->lock);
1685 return ret;
1686 }
1687 spin_lock(&bo->lock);
1688 if (likely(bo->sync_obj == sync_obj &&
1689 bo->sync_obj_arg == sync_obj_arg)) {
1690 void *tmp_obj = bo->sync_obj;
1691 bo->sync_obj = NULL;
1692 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1693 &bo->priv_flags);
1694 spin_unlock(&bo->lock);
1695 driver->sync_obj_unref(&sync_obj);
1696 driver->sync_obj_unref(&tmp_obj);
1697 spin_lock(&bo->lock);
1698 } else {
1699 spin_unlock(&bo->lock);
1700 driver->sync_obj_unref(&sync_obj);
1701 spin_lock(&bo->lock);
1702 }
1703 }
1704 return 0;
1705 }
1706 EXPORT_SYMBOL(ttm_bo_wait);
1707
1708 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1709 {
1710 int ret = 0;
1711
1712 /*
1713 * Using ttm_bo_reserve makes sure the lru lists are updated.
1714 */
1715
1716 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1717 if (unlikely(ret != 0))
1718 return ret;
1719 spin_lock(&bo->lock);
1720 ret = ttm_bo_wait(bo, false, true, no_wait);
1721 spin_unlock(&bo->lock);
1722 if (likely(ret == 0))
1723 atomic_inc(&bo->cpu_writers);
1724 ttm_bo_unreserve(bo);
1725 return ret;
1726 }
1727 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1728
1729 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1730 {
1731 if (atomic_dec_and_test(&bo->cpu_writers))
1732 wake_up_all(&bo->event_queue);
1733 }
1734 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1735
1736 /**
1737 * A buffer object shrink method that tries to swap out the first
1738 * buffer object on the bo_global::swap_lru list.
1739 */
1740
1741 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1742 {
1743 struct ttm_bo_global *glob =
1744 container_of(shrink, struct ttm_bo_global, shrink);
1745 struct ttm_buffer_object *bo;
1746 int ret = -EBUSY;
1747 int put_count;
1748 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1749
1750 spin_lock(&glob->lru_lock);
1751 while (ret == -EBUSY) {
1752 if (unlikely(list_empty(&glob->swap_lru))) {
1753 spin_unlock(&glob->lru_lock);
1754 return -EBUSY;
1755 }
1756
1757 bo = list_first_entry(&glob->swap_lru,
1758 struct ttm_buffer_object, swap);
1759 kref_get(&bo->list_kref);
1760
1761 if (!list_empty(&bo->ddestroy)) {
1762 spin_unlock(&glob->lru_lock);
1763 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1764 kref_put(&bo->list_kref, ttm_bo_release_list);
1765 continue;
1766 }
1767
1768 /**
1769 * Reserve buffer. Since we unlock while sleeping, we need
1770 * to re-check that nobody removed us from the swap-list while
1771 * we slept.
1772 */
1773
1774 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1775 if (unlikely(ret == -EBUSY)) {
1776 spin_unlock(&glob->lru_lock);
1777 ttm_bo_wait_unreserved(bo, false);
1778 kref_put(&bo->list_kref, ttm_bo_release_list);
1779 spin_lock(&glob->lru_lock);
1780 }
1781 }
1782
1783 BUG_ON(ret != 0);
1784 put_count = ttm_bo_del_from_lru(bo);
1785 spin_unlock(&glob->lru_lock);
1786
1787 ttm_bo_list_ref_sub(bo, put_count, true);
1788
1789 /**
1790 * Wait for GPU, then move to system cached.
1791 */
1792
1793 spin_lock(&bo->lock);
1794 ret = ttm_bo_wait(bo, false, false, false);
1795 spin_unlock(&bo->lock);
1796
1797 if (unlikely(ret != 0))
1798 goto out;
1799
1800 if ((bo->mem.placement & swap_placement) != swap_placement) {
1801 struct ttm_mem_reg evict_mem;
1802
1803 evict_mem = bo->mem;
1804 evict_mem.mm_node = NULL;
1805 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1806 evict_mem.mem_type = TTM_PL_SYSTEM;
1807
1808 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1809 false, false, false);
1810 if (unlikely(ret != 0))
1811 goto out;
1812 }
1813
1814 ttm_bo_unmap_virtual(bo);
1815
1816 /**
1817 * Swap out. Buffer will be swapped in again as soon as
1818 * anyone tries to access a ttm page.
1819 */
1820
1821 if (bo->bdev->driver->swap_notify)
1822 bo->bdev->driver->swap_notify(bo);
1823
1824 ret = ttm_tt_swapout(bo->ttm, bo->persistant_swap_storage);
1825 out:
1826
1827 /**
1828 *
1829 * Unreserve without putting on LRU to avoid swapping out an
1830 * already swapped buffer.
1831 */
1832
1833 atomic_set(&bo->reserved, 0);
1834 wake_up_all(&bo->event_queue);
1835 kref_put(&bo->list_kref, ttm_bo_release_list);
1836 return ret;
1837 }
1838
1839 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1840 {
1841 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1842 ;
1843 }
1844 EXPORT_SYMBOL(ttm_bo_swapout_all);
Linux 2.6 ibm-acpi/thinkpad-acpi driver git tree