/drivers/gpu/drm/ttm/ttm_bo.c
C | 1891 lines | 1411 code | 300 blank | 180 comment | 293 complexity | 86bbe99c1cc53be018d80499d0ffc8f6 MD5 | raw file
Possible License(s): GPL-2.0, LGPL-2.0, AGPL-1.0
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 bo->ttm = NULL;
359 }
360 break;
361 default:
362 printk(KERN_ERR TTM_PFX "Illegal buffer object type\n");
363 ret = -EINVAL;
364 break;
365 }
366
367 return ret;
368}
369
370static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
371 struct ttm_mem_reg *mem,
372 bool evict, bool interruptible,
373 bool no_wait_reserve, bool no_wait_gpu)
374{
375 struct ttm_bo_device *bdev = bo->bdev;
376 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
377 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
378 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
379 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
380 int ret = 0;
381
382 if (old_is_pci || new_is_pci ||
383 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
384 ret = ttm_mem_io_lock(old_man, true);
385 if (unlikely(ret != 0))
386 goto out_err;
387 ttm_bo_unmap_virtual_locked(bo);
388 ttm_mem_io_unlock(old_man);
389 }
390
391 /*
392 * Create and bind a ttm if required.
393 */
394
395 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
396 if (bo->ttm == NULL) {
397 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
398 ret = ttm_bo_add_ttm(bo, zero);
399 if (ret)
400 goto out_err;
401 }
402
403 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
404 if (ret)
405 goto out_err;
406
407 if (mem->mem_type != TTM_PL_SYSTEM) {
408 ret = ttm_tt_bind(bo->ttm, mem);
409 if (ret)
410 goto out_err;
411 }
412
413 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
414 if (bdev->driver->move_notify)
415 bdev->driver->move_notify(bo, mem);
416 bo->mem = *mem;
417 mem->mm_node = NULL;
418 goto moved;
419 }
420 }
421
422 if (bdev->driver->move_notify)
423 bdev->driver->move_notify(bo, mem);
424
425 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
426 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
427 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
428 else if (bdev->driver->move)
429 ret = bdev->driver->move(bo, evict, interruptible,
430 no_wait_reserve, no_wait_gpu, mem);
431 else
432 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
433
434 if (ret)
435 goto out_err;
436
437moved:
438 if (bo->evicted) {
439 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
440 if (ret)
441 printk(KERN_ERR TTM_PFX "Can not flush read caches\n");
442 bo->evicted = false;
443 }
444
445 if (bo->mem.mm_node) {
446 bo->offset = (bo->mem.start << PAGE_SHIFT) +
447 bdev->man[bo->mem.mem_type].gpu_offset;
448 bo->cur_placement = bo->mem.placement;
449 } else
450 bo->offset = 0;
451
452 return 0;
453
454out_err:
455 new_man = &bdev->man[bo->mem.mem_type];
456 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
457 ttm_tt_unbind(bo->ttm);
458 ttm_tt_destroy(bo->ttm);
459 bo->ttm = NULL;
460 }
461
462 return ret;
463}
464
465/**
466 * Call bo::reserved.
467 * Will release GPU memory type usage on destruction.
468 * This is the place to put in driver specific hooks to release
469 * driver private resources.
470 * Will release the bo::reserved lock.
471 */
472
473static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
474{
475 if (bo->ttm) {
476 ttm_tt_unbind(bo->ttm);
477 ttm_tt_destroy(bo->ttm);
478 bo->ttm = NULL;
479 }
480 ttm_bo_mem_put(bo, &bo->mem);
481
482 atomic_set(&bo->reserved, 0);
483
484 /*
485 * Make processes trying to reserve really pick it up.
486 */
487 smp_mb__after_atomic_dec();
488 wake_up_all(&bo->event_queue);
489}
490
491static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
492{
493 struct ttm_bo_device *bdev = bo->bdev;
494 struct ttm_bo_global *glob = bo->glob;
495 struct ttm_bo_driver *driver;
496 void *sync_obj = NULL;
497 void *sync_obj_arg;
498 int put_count;
499 int ret;
500
501 spin_lock(&bdev->fence_lock);
502 (void) ttm_bo_wait(bo, false, false, true);
503 if (!bo->sync_obj) {
504
505 spin_lock(&glob->lru_lock);
506
507 /**
508 * Lock inversion between bo:reserve and bdev::fence_lock here,
509 * but that's OK, since we're only trylocking.
510 */
511
512 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
513
514 if (unlikely(ret == -EBUSY))
515 goto queue;
516
517 spin_unlock(&bdev->fence_lock);
518 put_count = ttm_bo_del_from_lru(bo);
519
520 spin_unlock(&glob->lru_lock);
521 ttm_bo_cleanup_memtype_use(bo);
522
523 ttm_bo_list_ref_sub(bo, put_count, true);
524
525 return;
526 } else {
527 spin_lock(&glob->lru_lock);
528 }
529queue:
530 driver = bdev->driver;
531 if (bo->sync_obj)
532 sync_obj = driver->sync_obj_ref(bo->sync_obj);
533 sync_obj_arg = bo->sync_obj_arg;
534
535 kref_get(&bo->list_kref);
536 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
537 spin_unlock(&glob->lru_lock);
538 spin_unlock(&bdev->fence_lock);
539
540 if (sync_obj) {
541 driver->sync_obj_flush(sync_obj, sync_obj_arg);
542 driver->sync_obj_unref(&sync_obj);
543 }
544 schedule_delayed_work(&bdev->wq,
545 ((HZ / 100) < 1) ? 1 : HZ / 100);
546}
547
548/**
549 * function ttm_bo_cleanup_refs
550 * If bo idle, remove from delayed- and lru lists, and unref.
551 * If not idle, do nothing.
552 *
553 * @interruptible Any sleeps should occur interruptibly.
554 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
555 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
556 */
557
558static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
559 bool interruptible,
560 bool no_wait_reserve,
561 bool no_wait_gpu)
562{
563 struct ttm_bo_device *bdev = bo->bdev;
564 struct ttm_bo_global *glob = bo->glob;
565 int put_count;
566 int ret = 0;
567
568retry:
569 spin_lock(&bdev->fence_lock);
570 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
571 spin_unlock(&bdev->fence_lock);
572
573 if (unlikely(ret != 0))
574 return ret;
575
576 spin_lock(&glob->lru_lock);
577 ret = ttm_bo_reserve_locked(bo, interruptible,
578 no_wait_reserve, false, 0);
579
580 if (unlikely(ret != 0) || list_empty(&bo->ddestroy)) {
581 spin_unlock(&glob->lru_lock);
582 return ret;
583 }
584
585 /**
586 * We can re-check for sync object without taking
587 * the bo::lock since setting the sync object requires
588 * also bo::reserved. A busy object at this point may
589 * be caused by another thread recently starting an accelerated
590 * eviction.
591 */
592
593 if (unlikely(bo->sync_obj)) {
594 atomic_set(&bo->reserved, 0);
595 wake_up_all(&bo->event_queue);
596 spin_unlock(&glob->lru_lock);
597 goto retry;
598 }
599
600 put_count = ttm_bo_del_from_lru(bo);
601 list_del_init(&bo->ddestroy);
602 ++put_count;
603
604 spin_unlock(&glob->lru_lock);
605 ttm_bo_cleanup_memtype_use(bo);
606
607 ttm_bo_list_ref_sub(bo, put_count, true);
608
609 return 0;
610}
611
612/**
613 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
614 * encountered buffers.
615 */
616
617static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
618{
619 struct ttm_bo_global *glob = bdev->glob;
620 struct ttm_buffer_object *entry = NULL;
621 int ret = 0;
622
623 spin_lock(&glob->lru_lock);
624 if (list_empty(&bdev->ddestroy))
625 goto out_unlock;
626
627 entry = list_first_entry(&bdev->ddestroy,
628 struct ttm_buffer_object, ddestroy);
629 kref_get(&entry->list_kref);
630
631 for (;;) {
632 struct ttm_buffer_object *nentry = NULL;
633
634 if (entry->ddestroy.next != &bdev->ddestroy) {
635 nentry = list_first_entry(&entry->ddestroy,
636 struct ttm_buffer_object, ddestroy);
637 kref_get(&nentry->list_kref);
638 }
639
640 spin_unlock(&glob->lru_lock);
641 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
642 !remove_all);
643 kref_put(&entry->list_kref, ttm_bo_release_list);
644 entry = nentry;
645
646 if (ret || !entry)
647 goto out;
648
649 spin_lock(&glob->lru_lock);
650 if (list_empty(&entry->ddestroy))
651 break;
652 }
653
654out_unlock:
655 spin_unlock(&glob->lru_lock);
656out:
657 if (entry)
658 kref_put(&entry->list_kref, ttm_bo_release_list);
659 return ret;
660}
661
662static void ttm_bo_delayed_workqueue(struct work_struct *work)
663{
664 struct ttm_bo_device *bdev =
665 container_of(work, struct ttm_bo_device, wq.work);
666
667 if (ttm_bo_delayed_delete(bdev, false)) {
668 schedule_delayed_work(&bdev->wq,
669 ((HZ / 100) < 1) ? 1 : HZ / 100);
670 }
671}
672
673static void ttm_bo_release(struct kref *kref)
674{
675 struct ttm_buffer_object *bo =
676 container_of(kref, struct ttm_buffer_object, kref);
677 struct ttm_bo_device *bdev = bo->bdev;
678 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
679
680 if (likely(bo->vm_node != NULL)) {
681 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
682 drm_mm_put_block(bo->vm_node);
683 bo->vm_node = NULL;
684 }
685 write_unlock(&bdev->vm_lock);
686 ttm_mem_io_lock(man, false);
687 ttm_mem_io_free_vm(bo);
688 ttm_mem_io_unlock(man);
689 ttm_bo_cleanup_refs_or_queue(bo);
690 kref_put(&bo->list_kref, ttm_bo_release_list);
691 write_lock(&bdev->vm_lock);
692}
693
694void ttm_bo_unref(struct ttm_buffer_object **p_bo)
695{
696 struct ttm_buffer_object *bo = *p_bo;
697 struct ttm_bo_device *bdev = bo->bdev;
698
699 *p_bo = NULL;
700 write_lock(&bdev->vm_lock);
701 kref_put(&bo->kref, ttm_bo_release);
702 write_unlock(&bdev->vm_lock);
703}
704EXPORT_SYMBOL(ttm_bo_unref);
705
706int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
707{
708 return cancel_delayed_work_sync(&bdev->wq);
709}
710EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
711
712void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
713{
714 if (resched)
715 schedule_delayed_work(&bdev->wq,
716 ((HZ / 100) < 1) ? 1 : HZ / 100);
717}
718EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
719
720static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
721 bool no_wait_reserve, bool no_wait_gpu)
722{
723 struct ttm_bo_device *bdev = bo->bdev;
724 struct ttm_mem_reg evict_mem;
725 struct ttm_placement placement;
726 int ret = 0;
727
728 spin_lock(&bdev->fence_lock);
729 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
730 spin_unlock(&bdev->fence_lock);
731
732 if (unlikely(ret != 0)) {
733 if (ret != -ERESTARTSYS) {
734 printk(KERN_ERR TTM_PFX
735 "Failed to expire sync object before "
736 "buffer eviction.\n");
737 }
738 goto out;
739 }
740
741 BUG_ON(!atomic_read(&bo->reserved));
742
743 evict_mem = bo->mem;
744 evict_mem.mm_node = NULL;
745 evict_mem.bus.io_reserved_vm = false;
746 evict_mem.bus.io_reserved_count = 0;
747
748 placement.fpfn = 0;
749 placement.lpfn = 0;
750 placement.num_placement = 0;
751 placement.num_busy_placement = 0;
752 bdev->driver->evict_flags(bo, &placement);
753 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
754 no_wait_reserve, no_wait_gpu);
755 if (ret) {
756 if (ret != -ERESTARTSYS) {
757 printk(KERN_ERR TTM_PFX
758 "Failed to find memory space for "
759 "buffer 0x%p eviction.\n", bo);
760 ttm_bo_mem_space_debug(bo, &placement);
761 }
762 goto out;
763 }
764
765 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
766 no_wait_reserve, no_wait_gpu);
767 if (ret) {
768 if (ret != -ERESTARTSYS)
769 printk(KERN_ERR TTM_PFX "Buffer eviction failed\n");
770 ttm_bo_mem_put(bo, &evict_mem);
771 goto out;
772 }
773 bo->evicted = true;
774out:
775 return ret;
776}
777
778static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
779 uint32_t mem_type,
780 bool interruptible, bool no_wait_reserve,
781 bool no_wait_gpu)
782{
783 struct ttm_bo_global *glob = bdev->glob;
784 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
785 struct ttm_buffer_object *bo;
786 int ret, put_count = 0;
787
788retry:
789 spin_lock(&glob->lru_lock);
790 if (list_empty(&man->lru)) {
791 spin_unlock(&glob->lru_lock);
792 return -EBUSY;
793 }
794
795 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
796 kref_get(&bo->list_kref);
797
798 if (!list_empty(&bo->ddestroy)) {
799 spin_unlock(&glob->lru_lock);
800 ret = ttm_bo_cleanup_refs(bo, interruptible,
801 no_wait_reserve, no_wait_gpu);
802 kref_put(&bo->list_kref, ttm_bo_release_list);
803
804 if (likely(ret == 0 || ret == -ERESTARTSYS))
805 return ret;
806
807 goto retry;
808 }
809
810 ret = ttm_bo_reserve_locked(bo, false, no_wait_reserve, false, 0);
811
812 if (unlikely(ret == -EBUSY)) {
813 spin_unlock(&glob->lru_lock);
814 if (likely(!no_wait_gpu))
815 ret = ttm_bo_wait_unreserved(bo, interruptible);
816
817 kref_put(&bo->list_kref, ttm_bo_release_list);
818
819 /**
820 * We *need* to retry after releasing the lru lock.
821 */
822
823 if (unlikely(ret != 0))
824 return ret;
825 goto retry;
826 }
827
828 put_count = ttm_bo_del_from_lru(bo);
829 spin_unlock(&glob->lru_lock);
830
831 BUG_ON(ret != 0);
832
833 ttm_bo_list_ref_sub(bo, put_count, true);
834
835 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
836 ttm_bo_unreserve(bo);
837
838 kref_put(&bo->list_kref, ttm_bo_release_list);
839 return ret;
840}
841
842void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
843{
844 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
845
846 if (mem->mm_node)
847 (*man->func->put_node)(man, mem);
848}
849EXPORT_SYMBOL(ttm_bo_mem_put);
850
851/**
852 * Repeatedly evict memory from the LRU for @mem_type until we create enough
853 * space, or we've evicted everything and there isn't enough space.
854 */
855static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
856 uint32_t mem_type,
857 struct ttm_placement *placement,
858 struct ttm_mem_reg *mem,
859 bool interruptible,
860 bool no_wait_reserve,
861 bool no_wait_gpu)
862{
863 struct ttm_bo_device *bdev = bo->bdev;
864 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
865 int ret;
866
867 do {
868 ret = (*man->func->get_node)(man, bo, placement, mem);
869 if (unlikely(ret != 0))
870 return ret;
871 if (mem->mm_node)
872 break;
873 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
874 no_wait_reserve, no_wait_gpu);
875 if (unlikely(ret != 0))
876 return ret;
877 } while (1);
878 if (mem->mm_node == NULL)
879 return -ENOMEM;
880 mem->mem_type = mem_type;
881 return 0;
882}
883
884static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
885 uint32_t cur_placement,
886 uint32_t proposed_placement)
887{
888 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
889 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
890
891 /**
892 * Keep current caching if possible.
893 */
894
895 if ((cur_placement & caching) != 0)
896 result |= (cur_placement & caching);
897 else if ((man->default_caching & caching) != 0)
898 result |= man->default_caching;
899 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
900 result |= TTM_PL_FLAG_CACHED;
901 else if ((TTM_PL_FLAG_WC & caching) != 0)
902 result |= TTM_PL_FLAG_WC;
903 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
904 result |= TTM_PL_FLAG_UNCACHED;
905
906 return result;
907}
908
909static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
910 bool disallow_fixed,
911 uint32_t mem_type,
912 uint32_t proposed_placement,
913 uint32_t *masked_placement)
914{
915 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
916
917 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && disallow_fixed)
918 return false;
919
920 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
921 return false;
922
923 if ((proposed_placement & man->available_caching) == 0)
924 return false;
925
926 cur_flags |= (proposed_placement & man->available_caching);
927
928 *masked_placement = cur_flags;
929 return true;
930}
931
932/**
933 * Creates space for memory region @mem according to its type.
934 *
935 * This function first searches for free space in compatible memory types in
936 * the priority order defined by the driver. If free space isn't found, then
937 * ttm_bo_mem_force_space is attempted in priority order to evict and find
938 * space.
939 */
940int ttm_bo_mem_space(struct ttm_buffer_object *bo,
941 struct ttm_placement *placement,
942 struct ttm_mem_reg *mem,
943 bool interruptible, bool no_wait_reserve,
944 bool no_wait_gpu)
945{
946 struct ttm_bo_device *bdev = bo->bdev;
947 struct ttm_mem_type_manager *man;
948 uint32_t mem_type = TTM_PL_SYSTEM;
949 uint32_t cur_flags = 0;
950 bool type_found = false;
951 bool type_ok = false;
952 bool has_erestartsys = false;
953 int i, ret;
954
955 mem->mm_node = NULL;
956 for (i = 0; i < placement->num_placement; ++i) {
957 ret = ttm_mem_type_from_flags(placement->placement[i],
958 &mem_type);
959 if (ret)
960 return ret;
961 man = &bdev->man[mem_type];
962
963 type_ok = ttm_bo_mt_compatible(man,
964 bo->type == ttm_bo_type_user,
965 mem_type,
966 placement->placement[i],
967 &cur_flags);
968
969 if (!type_ok)
970 continue;
971
972 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
973 cur_flags);
974 /*
975 * Use the access and other non-mapping-related flag bits from
976 * the memory placement flags to the current flags
977 */
978 ttm_flag_masked(&cur_flags, placement->placement[i],
979 ~TTM_PL_MASK_MEMTYPE);
980
981 if (mem_type == TTM_PL_SYSTEM)
982 break;
983
984 if (man->has_type && man->use_type) {
985 type_found = true;
986 ret = (*man->func->get_node)(man, bo, placement, mem);
987 if (unlikely(ret))
988 return ret;
989 }
990 if (mem->mm_node)
991 break;
992 }
993
994 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
995 mem->mem_type = mem_type;
996 mem->placement = cur_flags;
997 return 0;
998 }
999
1000 if (!type_found)
1001 return -EINVAL;
1002
1003 for (i = 0; i < placement->num_busy_placement; ++i) {
1004 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
1005 &mem_type);
1006 if (ret)
1007 return ret;
1008 man = &bdev->man[mem_type];
1009 if (!man->has_type)
1010 continue;
1011 if (!ttm_bo_mt_compatible(man,
1012 bo->type == ttm_bo_type_user,
1013 mem_type,
1014 placement->busy_placement[i],
1015 &cur_flags))
1016 continue;
1017
1018 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1019 cur_flags);
1020 /*
1021 * Use the access and other non-mapping-related flag bits from
1022 * the memory placement flags to the current flags
1023 */
1024 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1025 ~TTM_PL_MASK_MEMTYPE);
1026
1027
1028 if (mem_type == TTM_PL_SYSTEM) {
1029 mem->mem_type = mem_type;
1030 mem->placement = cur_flags;
1031 mem->mm_node = NULL;
1032 return 0;
1033 }
1034
1035 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1036 interruptible, no_wait_reserve, no_wait_gpu);
1037 if (ret == 0 && mem->mm_node) {
1038 mem->placement = cur_flags;
1039 return 0;
1040 }
1041 if (ret == -ERESTARTSYS)
1042 has_erestartsys = true;
1043 }
1044 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1045 return ret;
1046}
1047EXPORT_SYMBOL(ttm_bo_mem_space);
1048
1049int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1050{
1051 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1052 return -EBUSY;
1053
1054 return wait_event_interruptible(bo->event_queue,
1055 atomic_read(&bo->cpu_writers) == 0);
1056}
1057EXPORT_SYMBOL(ttm_bo_wait_cpu);
1058
1059int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1060 struct ttm_placement *placement,
1061 bool interruptible, bool no_wait_reserve,
1062 bool no_wait_gpu)
1063{
1064 int ret = 0;
1065 struct ttm_mem_reg mem;
1066 struct ttm_bo_device *bdev = bo->bdev;
1067
1068 BUG_ON(!atomic_read(&bo->reserved));
1069
1070 /*
1071 * FIXME: It's possible to pipeline buffer moves.
1072 * Have the driver move function wait for idle when necessary,
1073 * instead of doing it here.
1074 */
1075 spin_lock(&bdev->fence_lock);
1076 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1077 spin_unlock(&bdev->fence_lock);
1078 if (ret)
1079 return ret;
1080 mem.num_pages = bo->num_pages;
1081 mem.size = mem.num_pages << PAGE_SHIFT;
1082 mem.page_alignment = bo->mem.page_alignment;
1083 mem.bus.io_reserved_vm = false;
1084 mem.bus.io_reserved_count = 0;
1085 /*
1086 * Determine where to move the buffer.
1087 */
1088 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1089 if (ret)
1090 goto out_unlock;
1091 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1092out_unlock:
1093 if (ret && mem.mm_node)
1094 ttm_bo_mem_put(bo, &mem);
1095 return ret;
1096}
1097
1098static int ttm_bo_mem_compat(struct ttm_placement *placement,
1099 struct ttm_mem_reg *mem)
1100{
1101 int i;
1102
1103 if (mem->mm_node && placement->lpfn != 0 &&
1104 (mem->start < placement->fpfn ||
1105 mem->start + mem->num_pages > placement->lpfn))
1106 return -1;
1107
1108 for (i = 0; i < placement->num_placement; i++) {
1109 if ((placement->placement[i] & mem->placement &
1110 TTM_PL_MASK_CACHING) &&
1111 (placement->placement[i] & mem->placement &
1112 TTM_PL_MASK_MEM))
1113 return i;
1114 }
1115 return -1;
1116}
1117
1118int ttm_bo_validate(struct ttm_buffer_object *bo,
1119 struct ttm_placement *placement,
1120 bool interruptible, bool no_wait_reserve,
1121 bool no_wait_gpu)
1122{
1123 int ret;
1124
1125 BUG_ON(!atomic_read(&bo->reserved));
1126 /* Check that range is valid */
1127 if (placement->lpfn || placement->fpfn)
1128 if (placement->fpfn > placement->lpfn ||
1129 (placement->lpfn - placement->fpfn) < bo->num_pages)
1130 return -EINVAL;
1131 /*
1132 * Check whether we need to move buffer.
1133 */
1134 ret = ttm_bo_mem_compat(placement, &bo->mem);
1135 if (ret < 0) {
1136 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1137 if (ret)
1138 return ret;
1139 } else {
1140 /*
1141 * Use the access and other non-mapping-related flag bits from
1142 * the compatible memory placement flags to the active flags
1143 */
1144 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1145 ~TTM_PL_MASK_MEMTYPE);
1146 }
1147 /*
1148 * We might need to add a TTM.
1149 */
1150 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1151 ret = ttm_bo_add_ttm(bo, true);
1152 if (ret)
1153 return ret;
1154 }
1155 return 0;
1156}
1157EXPORT_SYMBOL(ttm_bo_validate);
1158
1159int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1160 struct ttm_placement *placement)
1161{
1162 BUG_ON((placement->fpfn || placement->lpfn) &&
1163 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1164
1165 return 0;
1166}
1167
1168int ttm_bo_init(struct ttm_bo_device *bdev,
1169 struct ttm_buffer_object *bo,
1170 unsigned long size,
1171 enum ttm_bo_type type,
1172 struct ttm_placement *placement,
1173 uint32_t page_alignment,
1174 unsigned long buffer_start,
1175 bool interruptible,
1176 struct file *persistent_swap_storage,
1177 size_t acc_size,
1178 void (*destroy) (struct ttm_buffer_object *))
1179{
1180 int ret = 0;
1181 unsigned long num_pages;
1182
1183 size += buffer_start & ~PAGE_MASK;
1184 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1185 if (num_pages == 0) {
1186 printk(KERN_ERR TTM_PFX "Illegal buffer object size.\n");
1187 if (destroy)
1188 (*destroy)(bo);
1189 else
1190 kfree(bo);
1191 return -EINVAL;
1192 }
1193 bo->destroy = destroy;
1194
1195 kref_init(&bo->kref);
1196 kref_init(&bo->list_kref);
1197 atomic_set(&bo->cpu_writers, 0);
1198 atomic_set(&bo->reserved, 1);
1199 init_waitqueue_head(&bo->event_queue);
1200 INIT_LIST_HEAD(&bo->lru);
1201 INIT_LIST_HEAD(&bo->ddestroy);
1202 INIT_LIST_HEAD(&bo->swap);
1203 INIT_LIST_HEAD(&bo->io_reserve_lru);
1204 bo->bdev = bdev;
1205 bo->glob = bdev->glob;
1206 bo->type = type;
1207 bo->num_pages = num_pages;
1208 bo->mem.size = num_pages << PAGE_SHIFT;
1209 bo->mem.mem_type = TTM_PL_SYSTEM;
1210 bo->mem.num_pages = bo->num_pages;
1211 bo->mem.mm_node = NULL;
1212 bo->mem.page_alignment = page_alignment;
1213 bo->mem.bus.io_reserved_vm = false;
1214 bo->mem.bus.io_reserved_count = 0;
1215 bo->buffer_start = buffer_start & PAGE_MASK;
1216 bo->priv_flags = 0;
1217 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1218 bo->seq_valid = false;
1219 bo->persistent_swap_storage = persistent_swap_storage;
1220 bo->acc_size = acc_size;
1221 atomic_inc(&bo->glob->bo_count);
1222
1223 ret = ttm_bo_check_placement(bo, placement);
1224 if (unlikely(ret != 0))
1225 goto out_err;
1226
1227 /*
1228 * For ttm_bo_type_device buffers, allocate
1229 * address space from the device.
1230 */
1231 if (bo->type == ttm_bo_type_device) {
1232 ret = ttm_bo_setup_vm(bo);
1233 if (ret)
1234 goto out_err;
1235 }
1236
1237 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1238 if (ret)
1239 goto out_err;
1240
1241 ttm_bo_unreserve(bo);
1242 return 0;
1243
1244out_err:
1245 ttm_bo_unreserve(bo);
1246 ttm_bo_unref(&bo);
1247
1248 return ret;
1249}
1250EXPORT_SYMBOL(ttm_bo_init);
1251
1252static inline size_t ttm_bo_size(struct ttm_bo_global *glob,
1253 unsigned long num_pages)
1254{
1255 size_t page_array_size = (num_pages * sizeof(void *) + PAGE_SIZE - 1) &
1256 PAGE_MASK;
1257
1258 return glob->ttm_bo_size + 2 * page_array_size;
1259}
1260
1261int ttm_bo_create(struct ttm_bo_device *bdev,
1262 unsigned long size,
1263 enum ttm_bo_type type,
1264 struct ttm_placement *placement,
1265 uint32_t page_alignment,
1266 unsigned long buffer_start,
1267 bool interruptible,
1268 struct file *persistent_swap_storage,
1269 struct ttm_buffer_object **p_bo)
1270{
1271 struct ttm_buffer_object *bo;
1272 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1273 int ret;
1274
1275 size_t acc_size =
1276 ttm_bo_size(bdev->glob, (size + PAGE_SIZE - 1) >> PAGE_SHIFT);
1277 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1278 if (unlikely(ret != 0))
1279 return ret;
1280
1281 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1282
1283 if (unlikely(bo == NULL)) {
1284 ttm_mem_global_free(mem_glob, acc_size);
1285 return -ENOMEM;
1286 }
1287
1288 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1289 buffer_start, interruptible,
1290 persistent_swap_storage, acc_size, NULL);
1291 if (likely(ret == 0))
1292 *p_bo = bo;
1293
1294 return ret;
1295}
1296
1297static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1298 unsigned mem_type, bool allow_errors)
1299{
1300 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1301 struct ttm_bo_global *glob = bdev->glob;
1302 int ret;
1303
1304 /*
1305 * Can't use standard list traversal since we're unlocking.
1306 */
1307
1308 spin_lock(&glob->lru_lock);
1309 while (!list_empty(&man->lru)) {
1310 spin_unlock(&glob->lru_lock);
1311 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1312 if (ret) {
1313 if (allow_errors) {
1314 return ret;
1315 } else {
1316 printk(KERN_ERR TTM_PFX
1317 "Cleanup eviction failed\n");
1318 }
1319 }
1320 spin_lock(&glob->lru_lock);
1321 }
1322 spin_unlock(&glob->lru_lock);
1323 return 0;
1324}
1325
1326int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1327{
1328 struct ttm_mem_type_manager *man;
1329 int ret = -EINVAL;
1330
1331 if (mem_type >= TTM_NUM_MEM_TYPES) {
1332 printk(KERN_ERR TTM_PFX "Illegal memory type %d\n", mem_type);
1333 return ret;
1334 }
1335 man = &bdev->man[mem_type];
1336
1337 if (!man->has_type) {
1338 printk(KERN_ERR TTM_PFX "Trying to take down uninitialized "
1339 "memory manager type %u\n", mem_type);
1340 return ret;
1341 }
1342
1343 man->use_type = false;
1344 man->has_type = false;
1345
1346 ret = 0;
1347 if (mem_type > 0) {
1348 ttm_bo_force_list_clean(bdev, mem_type, false);
1349
1350 ret = (*man->func->takedown)(man);
1351 }
1352
1353 return ret;
1354}
1355EXPORT_SYMBOL(ttm_bo_clean_mm);
1356
1357int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1358{
1359 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1360
1361 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1362 printk(KERN_ERR TTM_PFX
1363 "Illegal memory manager memory type %u.\n",
1364 mem_type);
1365 return -EINVAL;
1366 }
1367
1368 if (!man->has_type) {
1369 printk(KERN_ERR TTM_PFX
1370 "Memory type %u has not been initialized.\n",
1371 mem_type);
1372 return 0;
1373 }
1374
1375 return ttm_bo_force_list_clean(bdev, mem_type, true);
1376}
1377EXPORT_SYMBOL(ttm_bo_evict_mm);
1378
1379int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1380 unsigned long p_size)
1381{
1382 int ret = -EINVAL;
1383 struct ttm_mem_type_manager *man;
1384
1385 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1386 man = &bdev->man[type];
1387 BUG_ON(man->has_type);
1388 man->io_reserve_fastpath = true;
1389 man->use_io_reserve_lru = false;
1390 mutex_init(&man->io_reserve_mutex);
1391 INIT_LIST_HEAD(&man->io_reserve_lru);
1392
1393 ret = bdev->driver->init_mem_type(bdev, type, man);
1394 if (ret)
1395 return ret;
1396 man->bdev = bdev;
1397
1398 ret = 0;
1399 if (type != TTM_PL_SYSTEM) {
1400 ret = (*man->func->init)(man, p_size);
1401 if (ret)
1402 return ret;
1403 }
1404 man->has_type = true;
1405 man->use_type = true;
1406 man->size = p_size;
1407
1408 INIT_LIST_HEAD(&man->lru);
1409
1410 return 0;
1411}
1412EXPORT_SYMBOL(ttm_bo_init_mm);
1413
1414static void ttm_bo_global_kobj_release(struct kobject *kobj)
1415{
1416 struct ttm_bo_global *glob =
1417 container_of(kobj, struct ttm_bo_global, kobj);
1418
1419 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1420 __free_page(glob->dummy_read_page);
1421 kfree(glob);
1422}
1423
1424void ttm_bo_global_release(struct drm_global_reference *ref)
1425{
1426 struct ttm_bo_global *glob = ref->object;
1427
1428 kobject_del(&glob->kobj);
1429 kobject_put(&glob->kobj);
1430}
1431EXPORT_SYMBOL(ttm_bo_global_release);
1432
1433int ttm_bo_global_init(struct drm_global_reference *ref)
1434{
1435 struct ttm_bo_global_ref *bo_ref =
1436 container_of(ref, struct ttm_bo_global_ref, ref);
1437 struct ttm_bo_global *glob = ref->object;
1438 int ret;
1439
1440 mutex_init(&glob->device_list_mutex);
1441 spin_lock_init(&glob->lru_lock);
1442 glob->mem_glob = bo_ref->mem_glob;
1443 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1444
1445 if (unlikely(glob->dummy_read_page == NULL)) {
1446 ret = -ENOMEM;
1447 goto out_no_drp;
1448 }
1449
1450 INIT_LIST_HEAD(&glob->swap_lru);
1451 INIT_LIST_HEAD(&glob->device_list);
1452
1453 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1454 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1455 if (unlikely(ret != 0)) {
1456 printk(KERN_ERR TTM_PFX
1457 "Could not register buffer object swapout.\n");
1458 goto out_no_shrink;
1459 }
1460
1461 glob->ttm_bo_extra_size =
1462 ttm_round_pot(sizeof(struct ttm_tt)) +
1463 ttm_round_pot(sizeof(struct ttm_backend));
1464
1465 glob->ttm_bo_size = glob->ttm_bo_extra_size +
1466 ttm_round_pot(sizeof(struct ttm_buffer_object));
1467
1468 atomic_set(&glob->bo_count, 0);
1469
1470 ret = kobject_init_and_add(
1471 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1472 if (unlikely(ret != 0))
1473 kobject_put(&glob->kobj);
1474 return ret;
1475out_no_shrink:
1476 __free_page(glob->dummy_read_page);
1477out_no_drp:
1478 kfree(glob);
1479 return ret;
1480}
1481EXPORT_SYMBOL(ttm_bo_global_init);
1482
1483
1484int ttm_bo_device_release(struct ttm_bo_device *bdev)
1485{
1486 int ret = 0;
1487 unsigned i = TTM_NUM_MEM_TYPES;
1488 struct ttm_mem_type_manager *man;
1489 struct ttm_bo_global *glob = bdev->glob;
1490
1491 while (i--) {
1492 man = &bdev->man[i];
1493 if (man->has_type) {
1494 man->use_type = false;
1495 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1496 ret = -EBUSY;
1497 printk(KERN_ERR TTM_PFX
1498 "DRM memory manager type %d "
1499 "is not clean.\n", i);
1500 }
1501 man->has_type = false;
1502 }
1503 }
1504
1505 mutex_lock(&glob->device_list_mutex);
1506 list_del(&bdev->device_list);
1507 mutex_unlock(&glob->device_list_mutex);
1508
1509 cancel_delayed_work_sync(&bdev->wq);
1510
1511 while (ttm_bo_delayed_delete(bdev, true))
1512 ;
1513
1514 spin_lock(&glob->lru_lock);
1515 if (list_empty(&bdev->ddestroy))
1516 TTM_DEBUG("Delayed destroy list was clean\n");
1517
1518 if (list_empty(&bdev->man[0].lru))
1519 TTM_DEBUG("Swap list was clean\n");
1520 spin_unlock(&glob->lru_lock);
1521
1522 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1523 write_lock(&bdev->vm_lock);
1524 drm_mm_takedown(&bdev->addr_space_mm);
1525 write_unlock(&bdev->vm_lock);
1526
1527 return ret;
1528}
1529EXPORT_SYMBOL(ttm_bo_device_release);
1530
1531int ttm_bo_device_init(struct ttm_bo_device *bdev,
1532 struct ttm_bo_global *glob,
1533 struct ttm_bo_driver *driver,
1534 uint64_t file_page_offset,
1535 bool need_dma32)
1536{
1537 int ret = -EINVAL;
1538
1539 rwlock_init(&bdev->vm_lock);
1540 bdev->driver = driver;
1541
1542 memset(bdev->man, 0, sizeof(bdev->man));
1543
1544 /*
1545 * Initialize the system memory buffer type.
1546 * Other types need to be driver / IOCTL initialized.
1547 */
1548 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1549 if (unlikely(ret != 0))
1550 goto out_no_sys;
1551
1552 bdev->addr_space_rb = RB_ROOT;
1553 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1554 if (unlikely(ret != 0))
1555 goto out_no_addr_mm;
1556
1557 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1558 bdev->nice_mode = true;
1559 INIT_LIST_HEAD(&bdev->ddestroy);
1560 bdev->dev_mapping = NULL;
1561 bdev->glob = glob;
1562 bdev->need_dma32 = need_dma32;
1563 bdev->val_seq = 0;
1564 spin_lock_init(&bdev->fence_lock);
1565 mutex_lock(&glob->device_list_mutex);
1566 list_add_tail(&bdev->device_list, &glob->device_list);
1567 mutex_unlock(&glob->device_list_mutex);
1568
1569 return 0;
1570out_no_addr_mm:
1571 ttm_bo_clean_mm(bdev, 0);
1572out_no_sys:
1573 return ret;
1574}
1575EXPORT_SYMBOL(ttm_bo_device_init);
1576
1577/*
1578 * buffer object vm functions.
1579 */
1580
1581bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1582{
1583 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1584
1585 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1586 if (mem->mem_type == TTM_PL_SYSTEM)
1587 return false;
1588
1589 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1590 return false;
1591
1592 if (mem->placement & TTM_PL_FLAG_CACHED)
1593 return false;
1594 }
1595 return true;
1596}
1597
1598void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1599{
1600 struct ttm_bo_device *bdev = bo->bdev;
1601 loff_t offset = (loff_t) bo->addr_space_offset;
1602 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1603
1604 if (!bdev->dev_mapping)
1605 return;
1606 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1607 ttm_mem_io_free_vm(bo);
1608}
1609
1610void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1611{
1612 struct ttm_bo_device *bdev = bo->bdev;
1613 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1614
1615 ttm_mem_io_lock(man, false);
1616 ttm_bo_unmap_virtual_locked(bo);
1617 ttm_mem_io_unlock(man);
1618}
1619
1620
1621EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1622
1623static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1624{
1625 struct ttm_bo_device *bdev = bo->bdev;
1626 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1627 struct rb_node *parent = NULL;
1628 struct ttm_buffer_object *cur_bo;
1629 unsigned long offset = bo->vm_node->start;
1630 unsigned long cur_offset;
1631
1632 while (*cur) {
1633 parent = *cur;
1634 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1635 cur_offset = cur_bo->vm_node->start;
1636 if (offset < cur_offset)
1637 cur = &parent->rb_left;
1638 else if (offset > cur_offset)
1639 cur = &parent->rb_right;
1640 else
1641 BUG();
1642 }
1643
1644 rb_link_node(&bo->vm_rb, parent, cur);
1645 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1646}
1647
1648/**
1649 * ttm_bo_setup_vm:
1650 *
1651 * @bo: the buffer to allocate address space for
1652 *
1653 * Allocate address space in the drm device so that applications
1654 * can mmap the buffer and access the contents. This only
1655 * applies to ttm_bo_type_device objects as others are not
1656 * placed in the drm device address space.
1657 */
1658
1659static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1660{
1661 struct ttm_bo_device *bdev = bo->bdev;
1662 int ret;
1663
1664retry_pre_get:
1665 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1666 if (unlikely(ret != 0))
1667 return ret;
1668
1669 write_lock(&bdev->vm_lock);
1670 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1671 bo->mem.num_pages, 0, 0);
1672
1673 if (unlikely(bo->vm_node == NULL)) {
1674 ret = -ENOMEM;
1675 goto out_unlock;
1676 }
1677
1678 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1679 bo->mem.num_pages, 0);
1680
1681 if (unlikely(bo->vm_node == NULL)) {
1682 write_unlock(&bdev->vm_lock);
1683 goto retry_pre_get;
1684 }
1685
1686 ttm_bo_vm_insert_rb(bo);
1687 write_unlock(&bdev->vm_lock);
1688 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1689
1690 return 0;
1691out_unlock:
1692 write_unlock(&bdev->vm_lock);
1693 return ret;
1694}
1695
1696int ttm_bo_wait(struct ttm_buffer_object *bo,
1697 bool lazy, bool interruptible, bool no_wait)
1698{
1699 struct ttm_bo_driver *driver = bo->bdev->driver;
1700 struct ttm_bo_device *bdev = bo->bdev;
1701 void *sync_obj;
1702 void *sync_obj_arg;
1703 int ret = 0;
1704
1705 if (likely(bo->sync_obj == NULL))
1706 return 0;
1707
1708 while (bo->sync_obj) {
1709
1710 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1711 void *tmp_obj = bo->sync_obj;
1712 bo->sync_obj = NULL;
1713 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1714 spin_unlock(&bdev->fence_lock);
1715 driver->sync_obj_unref(&tmp_obj);
1716 spin_lock(&bdev->fence_lock);
1717 continue;
1718 }
1719
1720 if (no_wait)
1721 return -EBUSY;
1722
1723 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1724 sync_obj_arg = bo->sync_obj_arg;
1725 spin_unlock(&bdev->fence_lock);
1726 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1727 lazy, interruptible);
1728 if (unlikely(ret != 0)) {
1729 driver->sync_obj_unref(&sync_obj);
1730 spin_lock(&bdev->fence_lock);
1731 return ret;
1732 }
1733 spin_lock(&bdev->fence_lock);
1734 if (likely(bo->sync_obj == sync_obj &&
1735 bo->sync_obj_arg == sync_obj_arg)) {
1736 void *tmp_obj = bo->sync_obj;
1737 bo->sync_obj = NULL;
1738 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1739 &bo->priv_flags);
1740 spin_unlock(&bdev->fence_lock);
1741 driver->sync_obj_unref(&sync_obj);
1742 driver->sync_obj_unref(&tmp_obj);
1743 spin_lock(&bdev->fence_lock);
1744 } else {
1745 spin_unlock(&bdev->fence_lock);
1746 driver->sync_obj_unref(&sync_obj);
1747 spin_lock(&bdev->fence_lock);
1748 }
1749 }
1750 return 0;
1751}
1752EXPORT_SYMBOL(ttm_bo_wait);
1753
1754int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1755{
1756 struct ttm_bo_device *bdev = bo->bdev;
1757 int ret = 0;
1758
1759 /*
1760 * Using ttm_bo_reserve makes sure the lru lists are updated.
1761 */
1762
1763 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1764 if (unlikely(ret != 0))
1765 return ret;
1766 spin_lock(&bdev->fence_lock);
1767 ret = ttm_bo_wait(bo, false, true, no_wait);
1768 spin_unlock(&bdev->fence_lock);
1769 if (likely(ret == 0))
1770 atomic_inc(&bo->cpu_writers);
1771 ttm_bo_unreserve(bo);
1772 return ret;
1773}
1774EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1775
1776void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1777{
1778 if (atomic_dec_and_test(&bo->cpu_writers))
1779 wake_up_all(&bo->event_queue);
1780}
1781EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1782
1783/**
1784 * A buffer object shrink method that tries to swap out the first
1785 * buffer object on the bo_global::swap_lru list.
1786 */
1787
1788static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1789{
1790 struct ttm_bo_global *glob =
1791 container_of(shrink, struct ttm_bo_global, shrink);
1792 struct ttm_buffer_object *bo;
1793 int ret = -EBUSY;
1794 int put_count;
1795 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1796
1797 spin_lock(&glob->lru_lock);
1798 while (ret == -EBUSY) {
1799 if (unlikely(list_empty(&glob->swap_lru))) {
1800 spin_unlock(&glob->lru_lock);
1801 return -EBUSY;
1802 }
1803
1804 bo = list_first_entry(&glob->swap_lru,
1805 struct ttm_buffer_object, swap);
1806 kref_get(&bo->list_kref);
1807
1808 if (!list_empty(&bo->ddestroy)) {
1809 spin_unlock(&glob->lru_lock);
1810 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1811 kref_put(&bo->list_kref, ttm_bo_release_list);
1812 continue;
1813 }
1814
1815 /**
1816 * Reserve buffer. Since we unlock while sleeping, we need
1817 * to re-check that nobody removed us from the swap-list while
1818 * we slept.
1819 */
1820
1821 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1822 if (unlikely(ret == -EBUSY)) {
1823 spin_unlock(&glob->lru_lock);
1824 ttm_bo_wait_unreserved(bo, false);
1825 kref_put(&bo->list_kref, ttm_bo_release_list);
1826 spin_lock(&glob->lru_lock);
1827 }
1828 }
1829
1830 BUG_ON(ret != 0);
1831 put_count = ttm_bo_del_from_lru(bo);
1832 spin_unlock(&glob->lru_lock);
1833
1834 ttm_bo_list_ref_sub(bo, put_count, true);
1835
1836 /**
1837 * Wait for GPU, then move to system cached.
1838 */
1839
1840 spin_lock(&bo->bdev->fence_lock);
1841 ret = ttm_bo_wait(bo, false, false, false);
1842 spin_unlock(&bo->bdev->fence_lock);
1843
1844 if (unlikely(ret != 0))
1845 goto out;
1846
1847 if ((bo->mem.placement & swap_placement) != swap_placement) {
1848 struct ttm_mem_reg evict_mem;
1849
1850 evict_mem = bo->mem;
1851 evict_mem.mm_node = NULL;
1852 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1853 evict_mem.mem_type = TTM_PL_SYSTEM;
1854
1855 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1856 false, false, false);
1857 if (unlikely(ret != 0))
1858 goto out;
1859 }
1860
1861 ttm_bo_unmap_virtual(bo);
1862
1863 /**
1864 * Swap out. Buffer will be swapped in again as soon as
1865 * anyone tries to access a ttm page.
1866 */
1867
1868 if (bo->bdev->driver->swap_notify)
1869 bo->bdev->driver->swap_notify(bo);
1870
1871 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1872out:
1873
1874 /**
1875 *
1876 * Unreserve without putting on LRU to avoid swapping out an
1877 * already swapped buffer.
1878 */
1879
1880 atomic_set(&bo->reserved, 0);
1881 wake_up_all(&bo->event_queue);
1882 kref_put(&bo->list_kref, ttm_bo_release_list);
1883 return ret;
1884}
1885
1886void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1887{
1888 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1889 ;
1890}
1891EXPORT_SYMBOL(ttm_bo_swapout_all);