Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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
46static int ttm_bo_setup_vm(struct ttm_buffer_object *bo);
47static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
48static void ttm_bo_global_kobj_release(struct kobject *kobj);
49
50static struct attribute ttm_bo_count = {
51 .name = "bo_count",
52 .mode = S_IRUGO
53};
54
55static 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
67static 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
84static 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
103static 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
114static struct attribute *ttm_bo_global_attrs[] = {
115 &ttm_bo_count,
116 NULL
117};
118
119static const struct sysfs_ops ttm_bo_global_ops = {
120 .show = &ttm_bo_global_show
121};
122
123static 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
130static inline uint32_t ttm_bo_type_flags(unsigned type)
131{
132 return 1 << (type);
133}
134
135static 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
160int 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}
170EXPORT_SYMBOL(ttm_bo_wait_unreserved);
171
172void 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
194int 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
215int 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}
268EXPORT_SYMBOL(ttm_bo_reserve);
269
270static void ttm_bo_ref_bug(struct kref *list_kref)
271{
272 BUG();
273}
274
275void 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
282int 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
302void 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
309void 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}
317EXPORT_SYMBOL(ttm_bo_unreserve);
318
319/*
320 * Call bo->mutex locked.
321 */
322static 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
368static 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 ret = ttm_mem_io_lock(old_man, true);
383 if (unlikely(ret != 0))
384 goto out_err;
385 ttm_bo_unmap_virtual_locked(bo);
386 ttm_mem_io_unlock(old_man);
387 }
388
389 /*
390 * Create and bind a ttm if required.
391 */
392
393 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && (bo->ttm == NULL)) {
394 ret = ttm_bo_add_ttm(bo, false);
395 if (ret)
396 goto out_err;
397
398 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
399 if (ret)
400 goto out_err;
401
402 if (mem->mem_type != TTM_PL_SYSTEM) {
403 ret = ttm_tt_bind(bo->ttm, mem);
404 if (ret)
405 goto out_err;
406 }
407
408 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
409 bo->mem = *mem;
410 mem->mm_node = NULL;
411 goto moved;
412 }
413
414 }
415
416 if (bdev->driver->move_notify)
417 bdev->driver->move_notify(bo, mem);
418
419 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
420 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
421 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
422 else if (bdev->driver->move)
423 ret = bdev->driver->move(bo, evict, interruptible,
424 no_wait_reserve, no_wait_gpu, mem);
425 else
426 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
427
428 if (ret)
429 goto out_err;
430
431moved:
432 if (bo->evicted) {
433 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
434 if (ret)
435 printk(KERN_ERR TTM_PFX "Can not flush read caches\n");
436 bo->evicted = false;
437 }
438
439 if (bo->mem.mm_node) {
440 bo->offset = (bo->mem.start << PAGE_SHIFT) +
441 bdev->man[bo->mem.mem_type].gpu_offset;
442 bo->cur_placement = bo->mem.placement;
443 } else
444 bo->offset = 0;
445
446 return 0;
447
448out_err:
449 new_man = &bdev->man[bo->mem.mem_type];
450 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
451 ttm_tt_unbind(bo->ttm);
452 ttm_tt_destroy(bo->ttm);
453 bo->ttm = NULL;
454 }
455
456 return ret;
457}
458
459/**
460 * Call bo::reserved.
461 * Will release GPU memory type usage on destruction.
462 * This is the place to put in driver specific hooks to release
463 * driver private resources.
464 * Will release the bo::reserved lock.
465 */
466
467static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
468{
469 if (bo->ttm) {
470 ttm_tt_unbind(bo->ttm);
471 ttm_tt_destroy(bo->ttm);
472 bo->ttm = NULL;
473 }
474 ttm_bo_mem_put(bo, &bo->mem);
475
476 atomic_set(&bo->reserved, 0);
477
478 /*
479 * Make processes trying to reserve really pick it up.
480 */
481 smp_mb__after_atomic_dec();
482 wake_up_all(&bo->event_queue);
483}
484
485static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
486{
487 struct ttm_bo_device *bdev = bo->bdev;
488 struct ttm_bo_global *glob = bo->glob;
489 struct ttm_bo_driver *driver;
490 void *sync_obj = NULL;
491 void *sync_obj_arg;
492 int put_count;
493 int ret;
494
495 spin_lock(&bdev->fence_lock);
496 (void) ttm_bo_wait(bo, false, false, true);
497 if (!bo->sync_obj) {
498
499 spin_lock(&glob->lru_lock);
500
501 /**
502 * Lock inversion between bo:reserve and bdev::fence_lock here,
503 * but that's OK, since we're only trylocking.
504 */
505
506 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
507
508 if (unlikely(ret == -EBUSY))
509 goto queue;
510
511 spin_unlock(&bdev->fence_lock);
512 put_count = ttm_bo_del_from_lru(bo);
513
514 spin_unlock(&glob->lru_lock);
515 ttm_bo_cleanup_memtype_use(bo);
516
517 ttm_bo_list_ref_sub(bo, put_count, true);
518
519 return;
520 } else {
521 spin_lock(&glob->lru_lock);
522 }
523queue:
524 driver = bdev->driver;
525 if (bo->sync_obj)
526 sync_obj = driver->sync_obj_ref(bo->sync_obj);
527 sync_obj_arg = bo->sync_obj_arg;
528
529 kref_get(&bo->list_kref);
530 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
531 spin_unlock(&glob->lru_lock);
532 spin_unlock(&bdev->fence_lock);
533
534 if (sync_obj) {
535 driver->sync_obj_flush(sync_obj, sync_obj_arg);
536 driver->sync_obj_unref(&sync_obj);
537 }
538 schedule_delayed_work(&bdev->wq,
539 ((HZ / 100) < 1) ? 1 : HZ / 100);
540}
541
542/**
543 * function ttm_bo_cleanup_refs
544 * If bo idle, remove from delayed- and lru lists, and unref.
545 * If not idle, do nothing.
546 *
547 * @interruptible Any sleeps should occur interruptibly.
548 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
549 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
550 */
551
552static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
553 bool interruptible,
554 bool no_wait_reserve,
555 bool no_wait_gpu)
556{
557 struct ttm_bo_device *bdev = bo->bdev;
558 struct ttm_bo_global *glob = bo->glob;
559 int put_count;
560 int ret = 0;
561
562retry:
563 spin_lock(&bdev->fence_lock);
564 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
565 spin_unlock(&bdev->fence_lock);
566
567 if (unlikely(ret != 0))
568 return ret;
569
570 spin_lock(&glob->lru_lock);
571 ret = ttm_bo_reserve_locked(bo, interruptible,
572 no_wait_reserve, false, 0);
573
574 if (unlikely(ret != 0) || list_empty(&bo->ddestroy)) {
575 spin_unlock(&glob->lru_lock);
576 return ret;
577 }
578
579 /**
580 * We can re-check for sync object without taking
581 * the bo::lock since setting the sync object requires
582 * also bo::reserved. A busy object at this point may
583 * be caused by another thread recently starting an accelerated
584 * eviction.
585 */
586
587 if (unlikely(bo->sync_obj)) {
588 atomic_set(&bo->reserved, 0);
589 wake_up_all(&bo->event_queue);
590 spin_unlock(&glob->lru_lock);
591 goto retry;
592 }
593
594 put_count = ttm_bo_del_from_lru(bo);
595 list_del_init(&bo->ddestroy);
596 ++put_count;
597
598 spin_unlock(&glob->lru_lock);
599 ttm_bo_cleanup_memtype_use(bo);
600
601 ttm_bo_list_ref_sub(bo, put_count, true);
602
603 return 0;
604}
605
606/**
607 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
608 * encountered buffers.
609 */
610
611static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
612{
613 struct ttm_bo_global *glob = bdev->glob;
614 struct ttm_buffer_object *entry = NULL;
615 int ret = 0;
616
617 spin_lock(&glob->lru_lock);
618 if (list_empty(&bdev->ddestroy))
619 goto out_unlock;
620
621 entry = list_first_entry(&bdev->ddestroy,
622 struct ttm_buffer_object, ddestroy);
623 kref_get(&entry->list_kref);
624
625 for (;;) {
626 struct ttm_buffer_object *nentry = NULL;
627
628 if (entry->ddestroy.next != &bdev->ddestroy) {
629 nentry = list_first_entry(&entry->ddestroy,
630 struct ttm_buffer_object, ddestroy);
631 kref_get(&nentry->list_kref);
632 }
633
634 spin_unlock(&glob->lru_lock);
635 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
636 !remove_all);
637 kref_put(&entry->list_kref, ttm_bo_release_list);
638 entry = nentry;
639
640 if (ret || !entry)
641 goto out;
642
643 spin_lock(&glob->lru_lock);
644 if (list_empty(&entry->ddestroy))
645 break;
646 }
647
648out_unlock:
649 spin_unlock(&glob->lru_lock);
650out:
651 if (entry)
652 kref_put(&entry->list_kref, ttm_bo_release_list);
653 return ret;
654}
655
656static void ttm_bo_delayed_workqueue(struct work_struct *work)
657{
658 struct ttm_bo_device *bdev =
659 container_of(work, struct ttm_bo_device, wq.work);
660
661 if (ttm_bo_delayed_delete(bdev, false)) {
662 schedule_delayed_work(&bdev->wq,
663 ((HZ / 100) < 1) ? 1 : HZ / 100);
664 }
665}
666
667static void ttm_bo_release(struct kref *kref)
668{
669 struct ttm_buffer_object *bo =
670 container_of(kref, struct ttm_buffer_object, kref);
671 struct ttm_bo_device *bdev = bo->bdev;
672 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
673
674 if (likely(bo->vm_node != NULL)) {
675 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
676 drm_mm_put_block(bo->vm_node);
677 bo->vm_node = NULL;
678 }
679 write_unlock(&bdev->vm_lock);
680 ttm_mem_io_lock(man, false);
681 ttm_mem_io_free_vm(bo);
682 ttm_mem_io_unlock(man);
683 ttm_bo_cleanup_refs_or_queue(bo);
684 kref_put(&bo->list_kref, ttm_bo_release_list);
685 write_lock(&bdev->vm_lock);
686}
687
688void ttm_bo_unref(struct ttm_buffer_object **p_bo)
689{
690 struct ttm_buffer_object *bo = *p_bo;
691 struct ttm_bo_device *bdev = bo->bdev;
692
693 *p_bo = NULL;
694 write_lock(&bdev->vm_lock);
695 kref_put(&bo->kref, ttm_bo_release);
696 write_unlock(&bdev->vm_lock);
697}
698EXPORT_SYMBOL(ttm_bo_unref);
699
700int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
701{
702 return cancel_delayed_work_sync(&bdev->wq);
703}
704EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
705
706void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
707{
708 if (resched)
709 schedule_delayed_work(&bdev->wq,
710 ((HZ / 100) < 1) ? 1 : HZ / 100);
711}
712EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
713
714static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
715 bool no_wait_reserve, bool no_wait_gpu)
716{
717 struct ttm_bo_device *bdev = bo->bdev;
718 struct ttm_mem_reg evict_mem;
719 struct ttm_placement placement;
720 int ret = 0;
721
722 spin_lock(&bdev->fence_lock);
723 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
724 spin_unlock(&bdev->fence_lock);
725
726 if (unlikely(ret != 0)) {
727 if (ret != -ERESTARTSYS) {
728 printk(KERN_ERR TTM_PFX
729 "Failed to expire sync object before "
730 "buffer eviction.\n");
731 }
732 goto out;
733 }
734
735 BUG_ON(!atomic_read(&bo->reserved));
736
737 evict_mem = bo->mem;
738 evict_mem.mm_node = NULL;
739 evict_mem.bus.io_reserved_vm = false;
740 evict_mem.bus.io_reserved_count = 0;
741
742 placement.fpfn = 0;
743 placement.lpfn = 0;
744 placement.num_placement = 0;
745 placement.num_busy_placement = 0;
746 bdev->driver->evict_flags(bo, &placement);
747 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
748 no_wait_reserve, no_wait_gpu);
749 if (ret) {
750 if (ret != -ERESTARTSYS) {
751 printk(KERN_ERR TTM_PFX
752 "Failed to find memory space for "
753 "buffer 0x%p eviction.\n", bo);
754 ttm_bo_mem_space_debug(bo, &placement);
755 }
756 goto out;
757 }
758
759 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
760 no_wait_reserve, no_wait_gpu);
761 if (ret) {
762 if (ret != -ERESTARTSYS)
763 printk(KERN_ERR TTM_PFX "Buffer eviction failed\n");
764 ttm_bo_mem_put(bo, &evict_mem);
765 goto out;
766 }
767 bo->evicted = true;
768out:
769 return ret;
770}
771
772static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
773 uint32_t mem_type,
774 bool interruptible, bool no_wait_reserve,
775 bool no_wait_gpu)
776{
777 struct ttm_bo_global *glob = bdev->glob;
778 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
779 struct ttm_buffer_object *bo;
780 int ret, put_count = 0;
781
782retry:
783 spin_lock(&glob->lru_lock);
784 if (list_empty(&man->lru)) {
785 spin_unlock(&glob->lru_lock);
786 return -EBUSY;
787 }
788
789 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
790 kref_get(&bo->list_kref);
791
792 if (!list_empty(&bo->ddestroy)) {
793 spin_unlock(&glob->lru_lock);
794 ret = ttm_bo_cleanup_refs(bo, interruptible,
795 no_wait_reserve, no_wait_gpu);
796 kref_put(&bo->list_kref, ttm_bo_release_list);
797
798 if (likely(ret == 0 || ret == -ERESTARTSYS))
799 return ret;
800
801 goto retry;
802 }
803
804 ret = ttm_bo_reserve_locked(bo, false, no_wait_reserve, false, 0);
805
806 if (unlikely(ret == -EBUSY)) {
807 spin_unlock(&glob->lru_lock);
808 if (likely(!no_wait_gpu))
809 ret = ttm_bo_wait_unreserved(bo, interruptible);
810
811 kref_put(&bo->list_kref, ttm_bo_release_list);
812
813 /**
814 * We *need* to retry after releasing the lru lock.
815 */
816
817 if (unlikely(ret != 0))
818 return ret;
819 goto retry;
820 }
821
822 put_count = ttm_bo_del_from_lru(bo);
823 spin_unlock(&glob->lru_lock);
824
825 BUG_ON(ret != 0);
826
827 ttm_bo_list_ref_sub(bo, put_count, true);
828
829 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
830 ttm_bo_unreserve(bo);
831
832 kref_put(&bo->list_kref, ttm_bo_release_list);
833 return ret;
834}
835
836void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
837{
838 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
839
840 if (mem->mm_node)
841 (*man->func->put_node)(man, mem);
842}
843EXPORT_SYMBOL(ttm_bo_mem_put);
844
845/**
846 * Repeatedly evict memory from the LRU for @mem_type until we create enough
847 * space, or we've evicted everything and there isn't enough space.
848 */
849static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
850 uint32_t mem_type,
851 struct ttm_placement *placement,
852 struct ttm_mem_reg *mem,
853 bool interruptible,
854 bool no_wait_reserve,
855 bool no_wait_gpu)
856{
857 struct ttm_bo_device *bdev = bo->bdev;
858 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
859 int ret;
860
861 do {
862 ret = (*man->func->get_node)(man, bo, placement, mem);
863 if (unlikely(ret != 0))
864 return ret;
865 if (mem->mm_node)
866 break;
867 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
868 no_wait_reserve, no_wait_gpu);
869 if (unlikely(ret != 0))
870 return ret;
871 } while (1);
872 if (mem->mm_node == NULL)
873 return -ENOMEM;
874 mem->mem_type = mem_type;
875 return 0;
876}
877
878static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
879 uint32_t cur_placement,
880 uint32_t proposed_placement)
881{
882 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
883 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
884
885 /**
886 * Keep current caching if possible.
887 */
888
889 if ((cur_placement & caching) != 0)
890 result |= (cur_placement & caching);
891 else if ((man->default_caching & caching) != 0)
892 result |= man->default_caching;
893 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
894 result |= TTM_PL_FLAG_CACHED;
895 else if ((TTM_PL_FLAG_WC & caching) != 0)
896 result |= TTM_PL_FLAG_WC;
897 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
898 result |= TTM_PL_FLAG_UNCACHED;
899
900 return result;
901}
902
903static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
904 bool disallow_fixed,
905 uint32_t mem_type,
906 uint32_t proposed_placement,
907 uint32_t *masked_placement)
908{
909 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
910
911 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && disallow_fixed)
912 return false;
913
914 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
915 return false;
916
917 if ((proposed_placement & man->available_caching) == 0)
918 return false;
919
920 cur_flags |= (proposed_placement & man->available_caching);
921
922 *masked_placement = cur_flags;
923 return true;
924}
925
926/**
927 * Creates space for memory region @mem according to its type.
928 *
929 * This function first searches for free space in compatible memory types in
930 * the priority order defined by the driver. If free space isn't found, then
931 * ttm_bo_mem_force_space is attempted in priority order to evict and find
932 * space.
933 */
934int ttm_bo_mem_space(struct ttm_buffer_object *bo,
935 struct ttm_placement *placement,
936 struct ttm_mem_reg *mem,
937 bool interruptible, bool no_wait_reserve,
938 bool no_wait_gpu)
939{
940 struct ttm_bo_device *bdev = bo->bdev;
941 struct ttm_mem_type_manager *man;
942 uint32_t mem_type = TTM_PL_SYSTEM;
943 uint32_t cur_flags = 0;
944 bool type_found = false;
945 bool type_ok = false;
946 bool has_erestartsys = false;
947 int i, ret;
948
949 mem->mm_node = NULL;
950 for (i = 0; i < placement->num_placement; ++i) {
951 ret = ttm_mem_type_from_flags(placement->placement[i],
952 &mem_type);
953 if (ret)
954 return ret;
955 man = &bdev->man[mem_type];
956
957 type_ok = ttm_bo_mt_compatible(man,
958 bo->type == ttm_bo_type_user,
959 mem_type,
960 placement->placement[i],
961 &cur_flags);
962
963 if (!type_ok)
964 continue;
965
966 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
967 cur_flags);
968 /*
969 * Use the access and other non-mapping-related flag bits from
970 * the memory placement flags to the current flags
971 */
972 ttm_flag_masked(&cur_flags, placement->placement[i],
973 ~TTM_PL_MASK_MEMTYPE);
974
975 if (mem_type == TTM_PL_SYSTEM)
976 break;
977
978 if (man->has_type && man->use_type) {
979 type_found = true;
980 ret = (*man->func->get_node)(man, bo, placement, mem);
981 if (unlikely(ret))
982 return ret;
983 }
984 if (mem->mm_node)
985 break;
986 }
987
988 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
989 mem->mem_type = mem_type;
990 mem->placement = cur_flags;
991 return 0;
992 }
993
994 if (!type_found)
995 return -EINVAL;
996
997 for (i = 0; i < placement->num_busy_placement; ++i) {
998 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
999 &mem_type);
1000 if (ret)
1001 return ret;
1002 man = &bdev->man[mem_type];
1003 if (!man->has_type)
1004 continue;
1005 if (!ttm_bo_mt_compatible(man,
1006 bo->type == ttm_bo_type_user,
1007 mem_type,
1008 placement->busy_placement[i],
1009 &cur_flags))
1010 continue;
1011
1012 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1013 cur_flags);
1014 /*
1015 * Use the access and other non-mapping-related flag bits from
1016 * the memory placement flags to the current flags
1017 */
1018 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1019 ~TTM_PL_MASK_MEMTYPE);
1020
1021
1022 if (mem_type == TTM_PL_SYSTEM) {
1023 mem->mem_type = mem_type;
1024 mem->placement = cur_flags;
1025 mem->mm_node = NULL;
1026 return 0;
1027 }
1028
1029 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1030 interruptible, no_wait_reserve, no_wait_gpu);
1031 if (ret == 0 && mem->mm_node) {
1032 mem->placement = cur_flags;
1033 return 0;
1034 }
1035 if (ret == -ERESTARTSYS)
1036 has_erestartsys = true;
1037 }
1038 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1039 return ret;
1040}
1041EXPORT_SYMBOL(ttm_bo_mem_space);
1042
1043int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1044{
1045 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1046 return -EBUSY;
1047
1048 return wait_event_interruptible(bo->event_queue,
1049 atomic_read(&bo->cpu_writers) == 0);
1050}
1051EXPORT_SYMBOL(ttm_bo_wait_cpu);
1052
1053int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1054 struct ttm_placement *placement,
1055 bool interruptible, bool no_wait_reserve,
1056 bool no_wait_gpu)
1057{
1058 int ret = 0;
1059 struct ttm_mem_reg mem;
1060 struct ttm_bo_device *bdev = bo->bdev;
1061
1062 BUG_ON(!atomic_read(&bo->reserved));
1063
1064 /*
1065 * FIXME: It's possible to pipeline buffer moves.
1066 * Have the driver move function wait for idle when necessary,
1067 * instead of doing it here.
1068 */
1069 spin_lock(&bdev->fence_lock);
1070 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1071 spin_unlock(&bdev->fence_lock);
1072 if (ret)
1073 return ret;
1074 mem.num_pages = bo->num_pages;
1075 mem.size = mem.num_pages << PAGE_SHIFT;
1076 mem.page_alignment = bo->mem.page_alignment;
1077 mem.bus.io_reserved_vm = false;
1078 mem.bus.io_reserved_count = 0;
1079 /*
1080 * Determine where to move the buffer.
1081 */
1082 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1083 if (ret)
1084 goto out_unlock;
1085 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1086out_unlock:
1087 if (ret && mem.mm_node)
1088 ttm_bo_mem_put(bo, &mem);
1089 return ret;
1090}
1091
1092static int ttm_bo_mem_compat(struct ttm_placement *placement,
1093 struct ttm_mem_reg *mem)
1094{
1095 int i;
1096
1097 if (mem->mm_node && placement->lpfn != 0 &&
1098 (mem->start < placement->fpfn ||
1099 mem->start + mem->num_pages > placement->lpfn))
1100 return -1;
1101
1102 for (i = 0; i < placement->num_placement; i++) {
1103 if ((placement->placement[i] & mem->placement &
1104 TTM_PL_MASK_CACHING) &&
1105 (placement->placement[i] & mem->placement &
1106 TTM_PL_MASK_MEM))
1107 return i;
1108 }
1109 return -1;
1110}
1111
1112int ttm_bo_validate(struct ttm_buffer_object *bo,
1113 struct ttm_placement *placement,
1114 bool interruptible, bool no_wait_reserve,
1115 bool no_wait_gpu)
1116{
1117 int ret;
1118
1119 BUG_ON(!atomic_read(&bo->reserved));
1120 /* Check that range is valid */
1121 if (placement->lpfn || placement->fpfn)
1122 if (placement->fpfn > placement->lpfn ||
1123 (placement->lpfn - placement->fpfn) < bo->num_pages)
1124 return -EINVAL;
1125 /*
1126 * Check whether we need to move buffer.
1127 */
1128 ret = ttm_bo_mem_compat(placement, &bo->mem);
1129 if (ret < 0) {
1130 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1131 if (ret)
1132 return ret;
1133 } else {
1134 /*
1135 * Use the access and other non-mapping-related flag bits from
1136 * the compatible memory placement flags to the active flags
1137 */
1138 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1139 ~TTM_PL_MASK_MEMTYPE);
1140 }
1141 /*
1142 * We might need to add a TTM.
1143 */
1144 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1145 ret = ttm_bo_add_ttm(bo, true);
1146 if (ret)
1147 return ret;
1148 }
1149 return 0;
1150}
1151EXPORT_SYMBOL(ttm_bo_validate);
1152
1153int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1154 struct ttm_placement *placement)
1155{
1156 BUG_ON((placement->fpfn || placement->lpfn) &&
1157 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1158
1159 return 0;
1160}
1161
1162int ttm_bo_init(struct ttm_bo_device *bdev,
1163 struct ttm_buffer_object *bo,
1164 unsigned long size,
1165 enum ttm_bo_type type,
1166 struct ttm_placement *placement,
1167 uint32_t page_alignment,
1168 unsigned long buffer_start,
1169 bool interruptible,
1170 struct file *persistant_swap_storage,
1171 size_t acc_size,
1172 void (*destroy) (struct ttm_buffer_object *))
1173{
1174 int ret = 0;
1175 unsigned long num_pages;
1176
1177 size += buffer_start & ~PAGE_MASK;
1178 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1179 if (num_pages == 0) {
1180 printk(KERN_ERR TTM_PFX "Illegal buffer object size.\n");
1181 if (destroy)
1182 (*destroy)(bo);
1183 else
1184 kfree(bo);
1185 return -EINVAL;
1186 }
1187 bo->destroy = destroy;
1188
1189 kref_init(&bo->kref);
1190 kref_init(&bo->list_kref);
1191 atomic_set(&bo->cpu_writers, 0);
1192 atomic_set(&bo->reserved, 1);
1193 init_waitqueue_head(&bo->event_queue);
1194 INIT_LIST_HEAD(&bo->lru);
1195 INIT_LIST_HEAD(&bo->ddestroy);
1196 INIT_LIST_HEAD(&bo->swap);
1197 INIT_LIST_HEAD(&bo->io_reserve_lru);
1198 bo->bdev = bdev;
1199 bo->glob = bdev->glob;
1200 bo->type = type;
1201 bo->num_pages = num_pages;
1202 bo->mem.size = num_pages << PAGE_SHIFT;
1203 bo->mem.mem_type = TTM_PL_SYSTEM;
1204 bo->mem.num_pages = bo->num_pages;
1205 bo->mem.mm_node = NULL;
1206 bo->mem.page_alignment = page_alignment;
1207 bo->mem.bus.io_reserved_vm = false;
1208 bo->mem.bus.io_reserved_count = 0;
1209 bo->buffer_start = buffer_start & PAGE_MASK;
1210 bo->priv_flags = 0;
1211 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1212 bo->seq_valid = false;
1213 bo->persistant_swap_storage = persistant_swap_storage;
1214 bo->acc_size = acc_size;
1215 atomic_inc(&bo->glob->bo_count);
1216
1217 ret = ttm_bo_check_placement(bo, placement);
1218 if (unlikely(ret != 0))
1219 goto out_err;
1220
1221 /*
1222 * For ttm_bo_type_device buffers, allocate
1223 * address space from the device.
1224 */
1225 if (bo->type == ttm_bo_type_device) {
1226 ret = ttm_bo_setup_vm(bo);
1227 if (ret)
1228 goto out_err;
1229 }
1230
1231 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1232 if (ret)
1233 goto out_err;
1234
1235 ttm_bo_unreserve(bo);
1236 return 0;
1237
1238out_err:
1239 ttm_bo_unreserve(bo);
1240 ttm_bo_unref(&bo);
1241
1242 return ret;
1243}
1244EXPORT_SYMBOL(ttm_bo_init);
1245
1246static inline size_t ttm_bo_size(struct ttm_bo_global *glob,
1247 unsigned long num_pages)
1248{
1249 size_t page_array_size = (num_pages * sizeof(void *) + PAGE_SIZE - 1) &
1250 PAGE_MASK;
1251
1252 return glob->ttm_bo_size + 2 * page_array_size;
1253}
1254
1255int ttm_bo_create(struct ttm_bo_device *bdev,
1256 unsigned long size,
1257 enum ttm_bo_type type,
1258 struct ttm_placement *placement,
1259 uint32_t page_alignment,
1260 unsigned long buffer_start,
1261 bool interruptible,
1262 struct file *persistant_swap_storage,
1263 struct ttm_buffer_object **p_bo)
1264{
1265 struct ttm_buffer_object *bo;
1266 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1267 int ret;
1268
1269 size_t acc_size =
1270 ttm_bo_size(bdev->glob, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
1271 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1272 if (unlikely(ret != 0))
1273 return ret;
1274
1275 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1276
1277 if (unlikely(bo == NULL)) {
1278 ttm_mem_global_free(mem_glob, acc_size);
1279 return -ENOMEM;
1280 }
1281
1282 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1283 buffer_start, interruptible,
1284 persistant_swap_storage, acc_size, NULL);
1285 if (likely(ret == 0))
1286 *p_bo = bo;
1287
1288 return ret;
1289}
1290
1291static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1292 unsigned mem_type, bool allow_errors)
1293{
1294 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1295 struct ttm_bo_global *glob = bdev->glob;
1296 int ret;
1297
1298 /*
1299 * Can't use standard list traversal since we're unlocking.
1300 */
1301
1302 spin_lock(&glob->lru_lock);
1303 while (!list_empty(&man->lru)) {
1304 spin_unlock(&glob->lru_lock);
1305 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1306 if (ret) {
1307 if (allow_errors) {
1308 return ret;
1309 } else {
1310 printk(KERN_ERR TTM_PFX
1311 "Cleanup eviction failed\n");
1312 }
1313 }
1314 spin_lock(&glob->lru_lock);
1315 }
1316 spin_unlock(&glob->lru_lock);
1317 return 0;
1318}
1319
1320int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1321{
1322 struct ttm_mem_type_manager *man;
1323 int ret = -EINVAL;
1324
1325 if (mem_type >= TTM_NUM_MEM_TYPES) {
1326 printk(KERN_ERR TTM_PFX "Illegal memory type %d\n", mem_type);
1327 return ret;
1328 }
1329 man = &bdev->man[mem_type];
1330
1331 if (!man->has_type) {
1332 printk(KERN_ERR TTM_PFX "Trying to take down uninitialized "
1333 "memory manager type %u\n", mem_type);
1334 return ret;
1335 }
1336
1337 man->use_type = false;
1338 man->has_type = false;
1339
1340 ret = 0;
1341 if (mem_type > 0) {
1342 ttm_bo_force_list_clean(bdev, mem_type, false);
1343
1344 ret = (*man->func->takedown)(man);
1345 }
1346
1347 return ret;
1348}
1349EXPORT_SYMBOL(ttm_bo_clean_mm);
1350
1351int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1352{
1353 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1354
1355 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1356 printk(KERN_ERR TTM_PFX
1357 "Illegal memory manager memory type %u.\n",
1358 mem_type);
1359 return -EINVAL;
1360 }
1361
1362 if (!man->has_type) {
1363 printk(KERN_ERR TTM_PFX
1364 "Memory type %u has not been initialized.\n",
1365 mem_type);
1366 return 0;
1367 }
1368
1369 return ttm_bo_force_list_clean(bdev, mem_type, true);
1370}
1371EXPORT_SYMBOL(ttm_bo_evict_mm);
1372
1373int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1374 unsigned long p_size)
1375{
1376 int ret = -EINVAL;
1377 struct ttm_mem_type_manager *man;
1378
1379 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1380 man = &bdev->man[type];
1381 BUG_ON(man->has_type);
1382 man->io_reserve_fastpath = true;
1383 man->use_io_reserve_lru = false;
1384 mutex_init(&man->io_reserve_mutex);
1385 INIT_LIST_HEAD(&man->io_reserve_lru);
1386
1387 ret = bdev->driver->init_mem_type(bdev, type, man);
1388 if (ret)
1389 return ret;
1390 man->bdev = bdev;
1391
1392 ret = 0;
1393 if (type != TTM_PL_SYSTEM) {
1394 ret = (*man->func->init)(man, p_size);
1395 if (ret)
1396 return ret;
1397 }
1398 man->has_type = true;
1399 man->use_type = true;
1400 man->size = p_size;
1401
1402 INIT_LIST_HEAD(&man->lru);
1403
1404 return 0;
1405}
1406EXPORT_SYMBOL(ttm_bo_init_mm);
1407
1408static void ttm_bo_global_kobj_release(struct kobject *kobj)
1409{
1410 struct ttm_bo_global *glob =
1411 container_of(kobj, struct ttm_bo_global, kobj);
1412
1413 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1414 __free_page(glob->dummy_read_page);
1415 kfree(glob);
1416}
1417
1418void ttm_bo_global_release(struct drm_global_reference *ref)
1419{
1420 struct ttm_bo_global *glob = ref->object;
1421
1422 kobject_del(&glob->kobj);
1423 kobject_put(&glob->kobj);
1424}
1425EXPORT_SYMBOL(ttm_bo_global_release);
1426
1427int ttm_bo_global_init(struct drm_global_reference *ref)
1428{
1429 struct ttm_bo_global_ref *bo_ref =
1430 container_of(ref, struct ttm_bo_global_ref, ref);
1431 struct ttm_bo_global *glob = ref->object;
1432 int ret;
1433
1434 mutex_init(&glob->device_list_mutex);
1435 spin_lock_init(&glob->lru_lock);
1436 glob->mem_glob = bo_ref->mem_glob;
1437 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1438
1439 if (unlikely(glob->dummy_read_page == NULL)) {
1440 ret = -ENOMEM;
1441 goto out_no_drp;
1442 }
1443
1444 INIT_LIST_HEAD(&glob->swap_lru);
1445 INIT_LIST_HEAD(&glob->device_list);
1446
1447 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1448 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1449 if (unlikely(ret != 0)) {
1450 printk(KERN_ERR TTM_PFX
1451 "Could not register buffer object swapout.\n");
1452 goto out_no_shrink;
1453 }
1454
1455 glob->ttm_bo_extra_size =
1456 ttm_round_pot(sizeof(struct ttm_tt)) +
1457 ttm_round_pot(sizeof(struct ttm_backend));
1458
1459 glob->ttm_bo_size = glob->ttm_bo_extra_size +
1460 ttm_round_pot(sizeof(struct ttm_buffer_object));
1461
1462 atomic_set(&glob->bo_count, 0);
1463
1464 ret = kobject_init_and_add(
1465 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1466 if (unlikely(ret != 0))
1467 kobject_put(&glob->kobj);
1468 return ret;
1469out_no_shrink:
1470 __free_page(glob->dummy_read_page);
1471out_no_drp:
1472 kfree(glob);
1473 return ret;
1474}
1475EXPORT_SYMBOL(ttm_bo_global_init);
1476
1477
1478int ttm_bo_device_release(struct ttm_bo_device *bdev)
1479{
1480 int ret = 0;
1481 unsigned i = TTM_NUM_MEM_TYPES;
1482 struct ttm_mem_type_manager *man;
1483 struct ttm_bo_global *glob = bdev->glob;
1484
1485 while (i--) {
1486 man = &bdev->man[i];
1487 if (man->has_type) {
1488 man->use_type = false;
1489 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1490 ret = -EBUSY;
1491 printk(KERN_ERR TTM_PFX
1492 "DRM memory manager type %d "
1493 "is not clean.\n", i);
1494 }
1495 man->has_type = false;
1496 }
1497 }
1498
1499 mutex_lock(&glob->device_list_mutex);
1500 list_del(&bdev->device_list);
1501 mutex_unlock(&glob->device_list_mutex);
1502
1503 cancel_delayed_work_sync(&bdev->wq);
1504
1505 while (ttm_bo_delayed_delete(bdev, true))
1506 ;
1507
1508 spin_lock(&glob->lru_lock);
1509 if (list_empty(&bdev->ddestroy))
1510 TTM_DEBUG("Delayed destroy list was clean\n");
1511
1512 if (list_empty(&bdev->man[0].lru))
1513 TTM_DEBUG("Swap list was clean\n");
1514 spin_unlock(&glob->lru_lock);
1515
1516 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1517 write_lock(&bdev->vm_lock);
1518 drm_mm_takedown(&bdev->addr_space_mm);
1519 write_unlock(&bdev->vm_lock);
1520
1521 return ret;
1522}
1523EXPORT_SYMBOL(ttm_bo_device_release);
1524
1525int ttm_bo_device_init(struct ttm_bo_device *bdev,
1526 struct ttm_bo_global *glob,
1527 struct ttm_bo_driver *driver,
1528 uint64_t file_page_offset,
1529 bool need_dma32)
1530{
1531 int ret = -EINVAL;
1532
1533 rwlock_init(&bdev->vm_lock);
1534 bdev->driver = driver;
1535
1536 memset(bdev->man, 0, sizeof(bdev->man));
1537
1538 /*
1539 * Initialize the system memory buffer type.
1540 * Other types need to be driver / IOCTL initialized.
1541 */
1542 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1543 if (unlikely(ret != 0))
1544 goto out_no_sys;
1545
1546 bdev->addr_space_rb = RB_ROOT;
1547 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1548 if (unlikely(ret != 0))
1549 goto out_no_addr_mm;
1550
1551 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1552 bdev->nice_mode = true;
1553 INIT_LIST_HEAD(&bdev->ddestroy);
1554 bdev->dev_mapping = NULL;
1555 bdev->glob = glob;
1556 bdev->need_dma32 = need_dma32;
1557 bdev->val_seq = 0;
1558 spin_lock_init(&bdev->fence_lock);
1559 mutex_lock(&glob->device_list_mutex);
1560 list_add_tail(&bdev->device_list, &glob->device_list);
1561 mutex_unlock(&glob->device_list_mutex);
1562
1563 return 0;
1564out_no_addr_mm:
1565 ttm_bo_clean_mm(bdev, 0);
1566out_no_sys:
1567 return ret;
1568}
1569EXPORT_SYMBOL(ttm_bo_device_init);
1570
1571/*
1572 * buffer object vm functions.
1573 */
1574
1575bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1576{
1577 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1578
1579 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1580 if (mem->mem_type == TTM_PL_SYSTEM)
1581 return false;
1582
1583 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1584 return false;
1585
1586 if (mem->placement & TTM_PL_FLAG_CACHED)
1587 return false;
1588 }
1589 return true;
1590}
1591
1592void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1593{
1594 struct ttm_bo_device *bdev = bo->bdev;
1595 loff_t offset = (loff_t) bo->addr_space_offset;
1596 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1597
1598 if (!bdev->dev_mapping)
1599 return;
1600 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1601 ttm_mem_io_free_vm(bo);
1602}
1603
1604void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1605{
1606 struct ttm_bo_device *bdev = bo->bdev;
1607 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1608
1609 ttm_mem_io_lock(man, false);
1610 ttm_bo_unmap_virtual_locked(bo);
1611 ttm_mem_io_unlock(man);
1612}
1613
1614
1615EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1616
1617static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1618{
1619 struct ttm_bo_device *bdev = bo->bdev;
1620 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1621 struct rb_node *parent = NULL;
1622 struct ttm_buffer_object *cur_bo;
1623 unsigned long offset = bo->vm_node->start;
1624 unsigned long cur_offset;
1625
1626 while (*cur) {
1627 parent = *cur;
1628 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1629 cur_offset = cur_bo->vm_node->start;
1630 if (offset < cur_offset)
1631 cur = &parent->rb_left;
1632 else if (offset > cur_offset)
1633 cur = &parent->rb_right;
1634 else
1635 BUG();
1636 }
1637
1638 rb_link_node(&bo->vm_rb, parent, cur);
1639 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1640}
1641
1642/**
1643 * ttm_bo_setup_vm:
1644 *
1645 * @bo: the buffer to allocate address space for
1646 *
1647 * Allocate address space in the drm device so that applications
1648 * can mmap the buffer and access the contents. This only
1649 * applies to ttm_bo_type_device objects as others are not
1650 * placed in the drm device address space.
1651 */
1652
1653static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1654{
1655 struct ttm_bo_device *bdev = bo->bdev;
1656 int ret;
1657
1658retry_pre_get:
1659 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1660 if (unlikely(ret != 0))
1661 return ret;
1662
1663 write_lock(&bdev->vm_lock);
1664 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1665 bo->mem.num_pages, 0, 0);
1666
1667 if (unlikely(bo->vm_node == NULL)) {
1668 ret = -ENOMEM;
1669 goto out_unlock;
1670 }
1671
1672 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1673 bo->mem.num_pages, 0);
1674
1675 if (unlikely(bo->vm_node == NULL)) {
1676 write_unlock(&bdev->vm_lock);
1677 goto retry_pre_get;
1678 }
1679
1680 ttm_bo_vm_insert_rb(bo);
1681 write_unlock(&bdev->vm_lock);
1682 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1683
1684 return 0;
1685out_unlock:
1686 write_unlock(&bdev->vm_lock);
1687 return ret;
1688}
1689
1690int ttm_bo_wait(struct ttm_buffer_object *bo,
1691 bool lazy, bool interruptible, bool no_wait)
1692{
1693 struct ttm_bo_driver *driver = bo->bdev->driver;
1694 struct ttm_bo_device *bdev = bo->bdev;
1695 void *sync_obj;
1696 void *sync_obj_arg;
1697 int ret = 0;
1698
1699 if (likely(bo->sync_obj == NULL))
1700 return 0;
1701
1702 while (bo->sync_obj) {
1703
1704 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1705 void *tmp_obj = bo->sync_obj;
1706 bo->sync_obj = NULL;
1707 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1708 spin_unlock(&bdev->fence_lock);
1709 driver->sync_obj_unref(&tmp_obj);
1710 spin_lock(&bdev->fence_lock);
1711 continue;
1712 }
1713
1714 if (no_wait)
1715 return -EBUSY;
1716
1717 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1718 sync_obj_arg = bo->sync_obj_arg;
1719 spin_unlock(&bdev->fence_lock);
1720 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1721 lazy, interruptible);
1722 if (unlikely(ret != 0)) {
1723 driver->sync_obj_unref(&sync_obj);
1724 spin_lock(&bdev->fence_lock);
1725 return ret;
1726 }
1727 spin_lock(&bdev->fence_lock);
1728 if (likely(bo->sync_obj == sync_obj &&
1729 bo->sync_obj_arg == sync_obj_arg)) {
1730 void *tmp_obj = bo->sync_obj;
1731 bo->sync_obj = NULL;
1732 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1733 &bo->priv_flags);
1734 spin_unlock(&bdev->fence_lock);
1735 driver->sync_obj_unref(&sync_obj);
1736 driver->sync_obj_unref(&tmp_obj);
1737 spin_lock(&bdev->fence_lock);
1738 } else {
1739 spin_unlock(&bdev->fence_lock);
1740 driver->sync_obj_unref(&sync_obj);
1741 spin_lock(&bdev->fence_lock);
1742 }
1743 }
1744 return 0;
1745}
1746EXPORT_SYMBOL(ttm_bo_wait);
1747
1748int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1749{
1750 struct ttm_bo_device *bdev = bo->bdev;
1751 int ret = 0;
1752
1753 /*
1754 * Using ttm_bo_reserve makes sure the lru lists are updated.
1755 */
1756
1757 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1758 if (unlikely(ret != 0))
1759 return ret;
1760 spin_lock(&bdev->fence_lock);
1761 ret = ttm_bo_wait(bo, false, true, no_wait);
1762 spin_unlock(&bdev->fence_lock);
1763 if (likely(ret == 0))
1764 atomic_inc(&bo->cpu_writers);
1765 ttm_bo_unreserve(bo);
1766 return ret;
1767}
1768EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1769
1770void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1771{
1772 if (atomic_dec_and_test(&bo->cpu_writers))
1773 wake_up_all(&bo->event_queue);
1774}
1775EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1776
1777/**
1778 * A buffer object shrink method that tries to swap out the first
1779 * buffer object on the bo_global::swap_lru list.
1780 */
1781
1782static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1783{
1784 struct ttm_bo_global *glob =
1785 container_of(shrink, struct ttm_bo_global, shrink);
1786 struct ttm_buffer_object *bo;
1787 int ret = -EBUSY;
1788 int put_count;
1789 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1790
1791 spin_lock(&glob->lru_lock);
1792 while (ret == -EBUSY) {
1793 if (unlikely(list_empty(&glob->swap_lru))) {
1794 spin_unlock(&glob->lru_lock);
1795 return -EBUSY;
1796 }
1797
1798 bo = list_first_entry(&glob->swap_lru,
1799 struct ttm_buffer_object, swap);
1800 kref_get(&bo->list_kref);
1801
1802 if (!list_empty(&bo->ddestroy)) {
1803 spin_unlock(&glob->lru_lock);
1804 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1805 kref_put(&bo->list_kref, ttm_bo_release_list);
1806 continue;
1807 }
1808
1809 /**
1810 * Reserve buffer. Since we unlock while sleeping, we need
1811 * to re-check that nobody removed us from the swap-list while
1812 * we slept.
1813 */
1814
1815 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1816 if (unlikely(ret == -EBUSY)) {
1817 spin_unlock(&glob->lru_lock);
1818 ttm_bo_wait_unreserved(bo, false);
1819 kref_put(&bo->list_kref, ttm_bo_release_list);
1820 spin_lock(&glob->lru_lock);
1821 }
1822 }
1823
1824 BUG_ON(ret != 0);
1825 put_count = ttm_bo_del_from_lru(bo);
1826 spin_unlock(&glob->lru_lock);
1827
1828 ttm_bo_list_ref_sub(bo, put_count, true);
1829
1830 /**
1831 * Wait for GPU, then move to system cached.
1832 */
1833
1834 spin_lock(&bo->bdev->fence_lock);
1835 ret = ttm_bo_wait(bo, false, false, false);
1836 spin_unlock(&bo->bdev->fence_lock);
1837
1838 if (unlikely(ret != 0))
1839 goto out;
1840
1841 if ((bo->mem.placement & swap_placement) != swap_placement) {
1842 struct ttm_mem_reg evict_mem;
1843
1844 evict_mem = bo->mem;
1845 evict_mem.mm_node = NULL;
1846 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1847 evict_mem.mem_type = TTM_PL_SYSTEM;
1848
1849 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1850 false, false, false);
1851 if (unlikely(ret != 0))
1852 goto out;
1853 }
1854
1855 ttm_bo_unmap_virtual(bo);
1856
1857 /**
1858 * Swap out. Buffer will be swapped in again as soon as
1859 * anyone tries to access a ttm page.
1860 */
1861
1862 if (bo->bdev->driver->swap_notify)
1863 bo->bdev->driver->swap_notify(bo);
1864
1865 ret = ttm_tt_swapout(bo->ttm, bo->persistant_swap_storage);
1866out:
1867
1868 /**
1869 *
1870 * Unreserve without putting on LRU to avoid swapping out an
1871 * already swapped buffer.
1872 */
1873
1874 atomic_set(&bo->reserved, 0);
1875 wake_up_all(&bo->event_queue);
1876 kref_put(&bo->list_kref, ttm_bo_release_list);
1877 return ret;
1878}
1879
1880void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1881{
1882 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1883 ;
1884}
1885EXPORT_SYMBOL(ttm_bo_swapout_all);