Merge branch 'v3.10/topic/misc' into linux-linaro-lsk
[firefly-linux-kernel-4.4.55.git] / drivers / gpu / drm / ttm / ttm_bo.c
1 /**************************************************************************
2  *
3  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 /*
28  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29  */
30
31 #define pr_fmt(fmt) "[TTM] " fmt
32
33 #include <drm/ttm/ttm_module.h>
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <linux/jiffies.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 #include <linux/mm.h>
40 #include <linux/file.h>
41 #include <linux/module.h>
42 #include <linux/atomic.h>
43
44 #define TTM_ASSERT_LOCKED(param)
45 #define TTM_DEBUG(fmt, arg...)
46 #define TTM_BO_HASH_ORDER 13
47
48 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo);
49 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50 static void ttm_bo_global_kobj_release(struct kobject *kobj);
51
52 static struct attribute ttm_bo_count = {
53         .name = "bo_count",
54         .mode = S_IRUGO
55 };
56
57 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
58 {
59         int i;
60
61         for (i = 0; i <= TTM_PL_PRIV5; i++)
62                 if (flags & (1 << i)) {
63                         *mem_type = i;
64                         return 0;
65                 }
66         return -EINVAL;
67 }
68
69 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
70 {
71         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
72
73         pr_err("    has_type: %d\n", man->has_type);
74         pr_err("    use_type: %d\n", man->use_type);
75         pr_err("    flags: 0x%08X\n", man->flags);
76         pr_err("    gpu_offset: 0x%08lX\n", man->gpu_offset);
77         pr_err("    size: %llu\n", man->size);
78         pr_err("    available_caching: 0x%08X\n", man->available_caching);
79         pr_err("    default_caching: 0x%08X\n", man->default_caching);
80         if (mem_type != TTM_PL_SYSTEM)
81                 (*man->func->debug)(man, TTM_PFX);
82 }
83
84 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
85                                         struct ttm_placement *placement)
86 {
87         int i, ret, mem_type;
88
89         pr_err("No space for %p (%lu pages, %luK, %luM)\n",
90                bo, bo->mem.num_pages, bo->mem.size >> 10,
91                bo->mem.size >> 20);
92         for (i = 0; i < placement->num_placement; i++) {
93                 ret = ttm_mem_type_from_flags(placement->placement[i],
94                                                 &mem_type);
95                 if (ret)
96                         return;
97                 pr_err("  placement[%d]=0x%08X (%d)\n",
98                        i, placement->placement[i], mem_type);
99                 ttm_mem_type_debug(bo->bdev, mem_type);
100         }
101 }
102
103 static ssize_t ttm_bo_global_show(struct kobject *kobj,
104                                   struct attribute *attr,
105                                   char *buffer)
106 {
107         struct ttm_bo_global *glob =
108                 container_of(kobj, struct ttm_bo_global, kobj);
109
110         return snprintf(buffer, PAGE_SIZE, "%lu\n",
111                         (unsigned long) atomic_read(&glob->bo_count));
112 }
113
114 static struct attribute *ttm_bo_global_attrs[] = {
115         &ttm_bo_count,
116         NULL
117 };
118
119 static const struct sysfs_ops ttm_bo_global_ops = {
120         .show = &ttm_bo_global_show
121 };
122
123 static struct kobj_type ttm_bo_glob_kobj_type  = {
124         .release = &ttm_bo_global_kobj_release,
125         .sysfs_ops = &ttm_bo_global_ops,
126         .default_attrs = ttm_bo_global_attrs
127 };
128
129
130 static inline uint32_t ttm_bo_type_flags(unsigned type)
131 {
132         return 1 << (type);
133 }
134
135 static void ttm_bo_release_list(struct kref *list_kref)
136 {
137         struct ttm_buffer_object *bo =
138             container_of(list_kref, struct ttm_buffer_object, list_kref);
139         struct ttm_bo_device *bdev = bo->bdev;
140         size_t acc_size = bo->acc_size;
141
142         BUG_ON(atomic_read(&bo->list_kref.refcount));
143         BUG_ON(atomic_read(&bo->kref.refcount));
144         BUG_ON(atomic_read(&bo->cpu_writers));
145         BUG_ON(bo->sync_obj != NULL);
146         BUG_ON(bo->mem.mm_node != NULL);
147         BUG_ON(!list_empty(&bo->lru));
148         BUG_ON(!list_empty(&bo->ddestroy));
149
150         if (bo->ttm)
151                 ttm_tt_destroy(bo->ttm);
152         atomic_dec(&bo->glob->bo_count);
153         if (bo->destroy)
154                 bo->destroy(bo);
155         else {
156                 kfree(bo);
157         }
158         ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
159 }
160
161 static int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo,
162                                   bool interruptible)
163 {
164         if (interruptible) {
165                 return wait_event_interruptible(bo->event_queue,
166                                                !ttm_bo_is_reserved(bo));
167         } else {
168                 wait_event(bo->event_queue, !ttm_bo_is_reserved(bo));
169                 return 0;
170         }
171 }
172
173 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
174 {
175         struct ttm_bo_device *bdev = bo->bdev;
176         struct ttm_mem_type_manager *man;
177
178         BUG_ON(!ttm_bo_is_reserved(bo));
179
180         if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
181
182                 BUG_ON(!list_empty(&bo->lru));
183
184                 man = &bdev->man[bo->mem.mem_type];
185                 list_add_tail(&bo->lru, &man->lru);
186                 kref_get(&bo->list_kref);
187
188                 if (bo->ttm != NULL) {
189                         list_add_tail(&bo->swap, &bo->glob->swap_lru);
190                         kref_get(&bo->list_kref);
191                 }
192         }
193 }
194
195 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
196 {
197         int put_count = 0;
198
199         if (!list_empty(&bo->swap)) {
200                 list_del_init(&bo->swap);
201                 ++put_count;
202         }
203         if (!list_empty(&bo->lru)) {
204                 list_del_init(&bo->lru);
205                 ++put_count;
206         }
207
208         /*
209          * TODO: Add a driver hook to delete from
210          * driver-specific LRU's here.
211          */
212
213         return put_count;
214 }
215
216 int ttm_bo_reserve_nolru(struct ttm_buffer_object *bo,
217                           bool interruptible,
218                           bool no_wait, bool use_sequence, uint32_t sequence)
219 {
220         int ret;
221
222         while (unlikely(atomic_xchg(&bo->reserved, 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                 ret = ttm_bo_wait_unreserved(bo, interruptible);
244
245                 if (unlikely(ret))
246                         return ret;
247         }
248
249         if (use_sequence) {
250                 bool wake_up = false;
251                 /**
252                  * Wake up waiters that may need to recheck for deadlock,
253                  * if we decreased the sequence number.
254                  */
255                 if (unlikely((bo->val_seq - sequence < (1 << 31))
256                              || !bo->seq_valid))
257                         wake_up = true;
258
259                 /*
260                  * In the worst case with memory ordering these values can be
261                  * seen in the wrong order. However since we call wake_up_all
262                  * in that case, this will hopefully not pose a problem,
263                  * and the worst case would only cause someone to accidentally
264                  * hit -EAGAIN in ttm_bo_reserve when they see old value of
265                  * val_seq. However this would only happen if seq_valid was
266                  * written before val_seq was, and just means some slightly
267                  * increased cpu usage
268                  */
269                 bo->val_seq = sequence;
270                 bo->seq_valid = true;
271                 if (wake_up)
272                         wake_up_all(&bo->event_queue);
273         } else {
274                 bo->seq_valid = false;
275         }
276
277         return 0;
278 }
279 EXPORT_SYMBOL(ttm_bo_reserve);
280
281 static void ttm_bo_ref_bug(struct kref *list_kref)
282 {
283         BUG();
284 }
285
286 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
287                          bool never_free)
288 {
289         kref_sub(&bo->list_kref, count,
290                  (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
291 }
292
293 int ttm_bo_reserve(struct ttm_buffer_object *bo,
294                    bool interruptible,
295                    bool no_wait, bool use_sequence, uint32_t sequence)
296 {
297         struct ttm_bo_global *glob = bo->glob;
298         int put_count = 0;
299         int ret;
300
301         ret = ttm_bo_reserve_nolru(bo, interruptible, no_wait, use_sequence,
302                                    sequence);
303         if (likely(ret == 0)) {
304                 spin_lock(&glob->lru_lock);
305                 put_count = ttm_bo_del_from_lru(bo);
306                 spin_unlock(&glob->lru_lock);
307                 ttm_bo_list_ref_sub(bo, put_count, true);
308         }
309
310         return ret;
311 }
312
313 int ttm_bo_reserve_slowpath_nolru(struct ttm_buffer_object *bo,
314                                   bool interruptible, uint32_t sequence)
315 {
316         bool wake_up = false;
317         int ret;
318
319         while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) {
320                 WARN_ON(bo->seq_valid && sequence == bo->val_seq);
321
322                 ret = ttm_bo_wait_unreserved(bo, interruptible);
323
324                 if (unlikely(ret))
325                         return ret;
326         }
327
328         if ((bo->val_seq - sequence < (1 << 31)) || !bo->seq_valid)
329                 wake_up = true;
330
331         /**
332          * Wake up waiters that may need to recheck for deadlock,
333          * if we decreased the sequence number.
334          */
335         bo->val_seq = sequence;
336         bo->seq_valid = true;
337         if (wake_up)
338                 wake_up_all(&bo->event_queue);
339
340         return 0;
341 }
342
343 int ttm_bo_reserve_slowpath(struct ttm_buffer_object *bo,
344                             bool interruptible, uint32_t sequence)
345 {
346         struct ttm_bo_global *glob = bo->glob;
347         int put_count, ret;
348
349         ret = ttm_bo_reserve_slowpath_nolru(bo, interruptible, sequence);
350         if (likely(!ret)) {
351                 spin_lock(&glob->lru_lock);
352                 put_count = ttm_bo_del_from_lru(bo);
353                 spin_unlock(&glob->lru_lock);
354                 ttm_bo_list_ref_sub(bo, put_count, true);
355         }
356         return ret;
357 }
358 EXPORT_SYMBOL(ttm_bo_reserve_slowpath);
359
360 void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo)
361 {
362         ttm_bo_add_to_lru(bo);
363         atomic_set(&bo->reserved, 0);
364         wake_up_all(&bo->event_queue);
365 }
366
367 void ttm_bo_unreserve(struct ttm_buffer_object *bo)
368 {
369         struct ttm_bo_global *glob = bo->glob;
370
371         spin_lock(&glob->lru_lock);
372         ttm_bo_unreserve_locked(bo);
373         spin_unlock(&glob->lru_lock);
374 }
375 EXPORT_SYMBOL(ttm_bo_unreserve);
376
377 /*
378  * Call bo->mutex locked.
379  */
380 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
381 {
382         struct ttm_bo_device *bdev = bo->bdev;
383         struct ttm_bo_global *glob = bo->glob;
384         int ret = 0;
385         uint32_t page_flags = 0;
386
387         TTM_ASSERT_LOCKED(&bo->mutex);
388         bo->ttm = NULL;
389
390         if (bdev->need_dma32)
391                 page_flags |= TTM_PAGE_FLAG_DMA32;
392
393         switch (bo->type) {
394         case ttm_bo_type_device:
395                 if (zero_alloc)
396                         page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
397         case ttm_bo_type_kernel:
398                 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
399                                                       page_flags, glob->dummy_read_page);
400                 if (unlikely(bo->ttm == NULL))
401                         ret = -ENOMEM;
402                 break;
403         case ttm_bo_type_sg:
404                 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
405                                                       page_flags | TTM_PAGE_FLAG_SG,
406                                                       glob->dummy_read_page);
407                 if (unlikely(bo->ttm == NULL)) {
408                         ret = -ENOMEM;
409                         break;
410                 }
411                 bo->ttm->sg = bo->sg;
412                 break;
413         default:
414                 pr_err("Illegal buffer object type\n");
415                 ret = -EINVAL;
416                 break;
417         }
418
419         return ret;
420 }
421
422 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
423                                   struct ttm_mem_reg *mem,
424                                   bool evict, bool interruptible,
425                                   bool no_wait_gpu)
426 {
427         struct ttm_bo_device *bdev = bo->bdev;
428         bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
429         bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
430         struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
431         struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
432         int ret = 0;
433
434         if (old_is_pci || new_is_pci ||
435             ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
436                 ret = ttm_mem_io_lock(old_man, true);
437                 if (unlikely(ret != 0))
438                         goto out_err;
439                 ttm_bo_unmap_virtual_locked(bo);
440                 ttm_mem_io_unlock(old_man);
441         }
442
443         /*
444          * Create and bind a ttm if required.
445          */
446
447         if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
448                 if (bo->ttm == NULL) {
449                         bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
450                         ret = ttm_bo_add_ttm(bo, zero);
451                         if (ret)
452                                 goto out_err;
453                 }
454
455                 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
456                 if (ret)
457                         goto out_err;
458
459                 if (mem->mem_type != TTM_PL_SYSTEM) {
460                         ret = ttm_tt_bind(bo->ttm, mem);
461                         if (ret)
462                                 goto out_err;
463                 }
464
465                 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
466                         if (bdev->driver->move_notify)
467                                 bdev->driver->move_notify(bo, mem);
468                         bo->mem = *mem;
469                         mem->mm_node = NULL;
470                         goto moved;
471                 }
472         }
473
474         if (bdev->driver->move_notify)
475                 bdev->driver->move_notify(bo, mem);
476
477         if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
478             !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
479                 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
480         else if (bdev->driver->move)
481                 ret = bdev->driver->move(bo, evict, interruptible,
482                                          no_wait_gpu, mem);
483         else
484                 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
485
486         if (ret) {
487                 if (bdev->driver->move_notify) {
488                         struct ttm_mem_reg tmp_mem = *mem;
489                         *mem = bo->mem;
490                         bo->mem = tmp_mem;
491                         bdev->driver->move_notify(bo, mem);
492                         bo->mem = *mem;
493                         *mem = tmp_mem;
494                 }
495
496                 goto out_err;
497         }
498
499 moved:
500         if (bo->evicted) {
501                 if (bdev->driver->invalidate_caches) {
502                         ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
503                         if (ret)
504                                 pr_err("Can not flush read caches\n");
505                 }
506                 bo->evicted = false;
507         }
508
509         if (bo->mem.mm_node) {
510                 bo->offset = (bo->mem.start << PAGE_SHIFT) +
511                     bdev->man[bo->mem.mem_type].gpu_offset;
512                 bo->cur_placement = bo->mem.placement;
513         } else
514                 bo->offset = 0;
515
516         return 0;
517
518 out_err:
519         new_man = &bdev->man[bo->mem.mem_type];
520         if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
521                 ttm_tt_unbind(bo->ttm);
522                 ttm_tt_destroy(bo->ttm);
523                 bo->ttm = NULL;
524         }
525
526         return ret;
527 }
528
529 /**
530  * Call bo::reserved.
531  * Will release GPU memory type usage on destruction.
532  * This is the place to put in driver specific hooks to release
533  * driver private resources.
534  * Will release the bo::reserved lock.
535  */
536
537 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
538 {
539         if (bo->bdev->driver->move_notify)
540                 bo->bdev->driver->move_notify(bo, NULL);
541
542         if (bo->ttm) {
543                 ttm_tt_unbind(bo->ttm);
544                 ttm_tt_destroy(bo->ttm);
545                 bo->ttm = NULL;
546         }
547         ttm_bo_mem_put(bo, &bo->mem);
548
549         atomic_set(&bo->reserved, 0);
550         wake_up_all(&bo->event_queue);
551
552         /*
553          * Since the final reference to this bo may not be dropped by
554          * the current task we have to put a memory barrier here to make
555          * sure the changes done in this function are always visible.
556          *
557          * This function only needs protection against the final kref_put.
558          */
559         smp_mb__before_atomic_dec();
560 }
561
562 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
563 {
564         struct ttm_bo_device *bdev = bo->bdev;
565         struct ttm_bo_global *glob = bo->glob;
566         struct ttm_bo_driver *driver = bdev->driver;
567         void *sync_obj = NULL;
568         int put_count;
569         int ret;
570
571         spin_lock(&glob->lru_lock);
572         ret = ttm_bo_reserve_nolru(bo, false, true, false, 0);
573
574         spin_lock(&bdev->fence_lock);
575         (void) ttm_bo_wait(bo, false, false, true);
576         if (!ret && !bo->sync_obj) {
577                 spin_unlock(&bdev->fence_lock);
578                 put_count = ttm_bo_del_from_lru(bo);
579
580                 spin_unlock(&glob->lru_lock);
581                 ttm_bo_cleanup_memtype_use(bo);
582
583                 ttm_bo_list_ref_sub(bo, put_count, true);
584
585                 return;
586         }
587         if (bo->sync_obj)
588                 sync_obj = driver->sync_obj_ref(bo->sync_obj);
589         spin_unlock(&bdev->fence_lock);
590
591         if (!ret) {
592                 atomic_set(&bo->reserved, 0);
593                 wake_up_all(&bo->event_queue);
594         }
595
596         kref_get(&bo->list_kref);
597         list_add_tail(&bo->ddestroy, &bdev->ddestroy);
598         spin_unlock(&glob->lru_lock);
599
600         if (sync_obj) {
601                 driver->sync_obj_flush(sync_obj);
602                 driver->sync_obj_unref(&sync_obj);
603         }
604         schedule_delayed_work(&bdev->wq,
605                               ((HZ / 100) < 1) ? 1 : HZ / 100);
606 }
607
608 /**
609  * function ttm_bo_cleanup_refs_and_unlock
610  * If bo idle, remove from delayed- and lru lists, and unref.
611  * If not idle, do nothing.
612  *
613  * Must be called with lru_lock and reservation held, this function
614  * will drop both before returning.
615  *
616  * @interruptible         Any sleeps should occur interruptibly.
617  * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
618  */
619
620 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
621                                           bool interruptible,
622                                           bool no_wait_gpu)
623 {
624         struct ttm_bo_device *bdev = bo->bdev;
625         struct ttm_bo_driver *driver = bdev->driver;
626         struct ttm_bo_global *glob = bo->glob;
627         int put_count;
628         int ret;
629
630         spin_lock(&bdev->fence_lock);
631         ret = ttm_bo_wait(bo, false, false, true);
632
633         if (ret && !no_wait_gpu) {
634                 void *sync_obj;
635
636                 /*
637                  * Take a reference to the fence and unreserve,
638                  * at this point the buffer should be dead, so
639                  * no new sync objects can be attached.
640                  */
641                 sync_obj = driver->sync_obj_ref(bo->sync_obj);
642                 spin_unlock(&bdev->fence_lock);
643
644                 atomic_set(&bo->reserved, 0);
645                 wake_up_all(&bo->event_queue);
646                 spin_unlock(&glob->lru_lock);
647
648                 ret = driver->sync_obj_wait(sync_obj, false, interruptible);
649                 driver->sync_obj_unref(&sync_obj);
650                 if (ret)
651                         return ret;
652
653                 /*
654                  * remove sync_obj with ttm_bo_wait, the wait should be
655                  * finished, and no new wait object should have been added.
656                  */
657                 spin_lock(&bdev->fence_lock);
658                 ret = ttm_bo_wait(bo, false, false, true);
659                 WARN_ON(ret);
660                 spin_unlock(&bdev->fence_lock);
661                 if (ret)
662                         return ret;
663
664                 spin_lock(&glob->lru_lock);
665                 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0);
666
667                 /*
668                  * We raced, and lost, someone else holds the reservation now,
669                  * and is probably busy in ttm_bo_cleanup_memtype_use.
670                  *
671                  * Even if it's not the case, because we finished waiting any
672                  * delayed destruction would succeed, so just return success
673                  * here.
674                  */
675                 if (ret) {
676                         spin_unlock(&glob->lru_lock);
677                         return 0;
678                 }
679         } else
680                 spin_unlock(&bdev->fence_lock);
681
682         if (ret || unlikely(list_empty(&bo->ddestroy))) {
683                 atomic_set(&bo->reserved, 0);
684                 wake_up_all(&bo->event_queue);
685                 spin_unlock(&glob->lru_lock);
686                 return ret;
687         }
688
689         put_count = ttm_bo_del_from_lru(bo);
690         list_del_init(&bo->ddestroy);
691         ++put_count;
692
693         spin_unlock(&glob->lru_lock);
694         ttm_bo_cleanup_memtype_use(bo);
695
696         ttm_bo_list_ref_sub(bo, put_count, true);
697
698         return 0;
699 }
700
701 /**
702  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
703  * encountered buffers.
704  */
705
706 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
707 {
708         struct ttm_bo_global *glob = bdev->glob;
709         struct ttm_buffer_object *entry = NULL;
710         int ret = 0;
711
712         spin_lock(&glob->lru_lock);
713         if (list_empty(&bdev->ddestroy))
714                 goto out_unlock;
715
716         entry = list_first_entry(&bdev->ddestroy,
717                 struct ttm_buffer_object, ddestroy);
718         kref_get(&entry->list_kref);
719
720         for (;;) {
721                 struct ttm_buffer_object *nentry = NULL;
722
723                 if (entry->ddestroy.next != &bdev->ddestroy) {
724                         nentry = list_first_entry(&entry->ddestroy,
725                                 struct ttm_buffer_object, ddestroy);
726                         kref_get(&nentry->list_kref);
727                 }
728
729                 ret = ttm_bo_reserve_nolru(entry, false, true, false, 0);
730                 if (remove_all && ret) {
731                         spin_unlock(&glob->lru_lock);
732                         ret = ttm_bo_reserve_nolru(entry, false, false,
733                                                    false, 0);
734                         spin_lock(&glob->lru_lock);
735                 }
736
737                 if (!ret)
738                         ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
739                                                              !remove_all);
740                 else
741                         spin_unlock(&glob->lru_lock);
742
743                 kref_put(&entry->list_kref, ttm_bo_release_list);
744                 entry = nentry;
745
746                 if (ret || !entry)
747                         goto out;
748
749                 spin_lock(&glob->lru_lock);
750                 if (list_empty(&entry->ddestroy))
751                         break;
752         }
753
754 out_unlock:
755         spin_unlock(&glob->lru_lock);
756 out:
757         if (entry)
758                 kref_put(&entry->list_kref, ttm_bo_release_list);
759         return ret;
760 }
761
762 static void ttm_bo_delayed_workqueue(struct work_struct *work)
763 {
764         struct ttm_bo_device *bdev =
765             container_of(work, struct ttm_bo_device, wq.work);
766
767         if (ttm_bo_delayed_delete(bdev, false)) {
768                 schedule_delayed_work(&bdev->wq,
769                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
770         }
771 }
772
773 static void ttm_bo_release(struct kref *kref)
774 {
775         struct ttm_buffer_object *bo =
776             container_of(kref, struct ttm_buffer_object, kref);
777         struct ttm_bo_device *bdev = bo->bdev;
778         struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
779
780         write_lock(&bdev->vm_lock);
781         if (likely(bo->vm_node != NULL)) {
782                 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
783                 drm_mm_put_block(bo->vm_node);
784                 bo->vm_node = NULL;
785         }
786         write_unlock(&bdev->vm_lock);
787         ttm_mem_io_lock(man, false);
788         ttm_mem_io_free_vm(bo);
789         ttm_mem_io_unlock(man);
790         ttm_bo_cleanup_refs_or_queue(bo);
791         kref_put(&bo->list_kref, ttm_bo_release_list);
792 }
793
794 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
795 {
796         struct ttm_buffer_object *bo = *p_bo;
797
798         *p_bo = NULL;
799         kref_put(&bo->kref, ttm_bo_release);
800 }
801 EXPORT_SYMBOL(ttm_bo_unref);
802
803 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
804 {
805         return cancel_delayed_work_sync(&bdev->wq);
806 }
807 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
808
809 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
810 {
811         if (resched)
812                 schedule_delayed_work(&bdev->wq,
813                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
814 }
815 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
816
817 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
818                         bool no_wait_gpu)
819 {
820         struct ttm_bo_device *bdev = bo->bdev;
821         struct ttm_mem_reg evict_mem;
822         struct ttm_placement placement;
823         int ret = 0;
824
825         spin_lock(&bdev->fence_lock);
826         ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
827         spin_unlock(&bdev->fence_lock);
828
829         if (unlikely(ret != 0)) {
830                 if (ret != -ERESTARTSYS) {
831                         pr_err("Failed to expire sync object before buffer eviction\n");
832                 }
833                 goto out;
834         }
835
836         BUG_ON(!ttm_bo_is_reserved(bo));
837
838         evict_mem = bo->mem;
839         evict_mem.mm_node = NULL;
840         evict_mem.bus.io_reserved_vm = false;
841         evict_mem.bus.io_reserved_count = 0;
842
843         placement.fpfn = 0;
844         placement.lpfn = 0;
845         placement.num_placement = 0;
846         placement.num_busy_placement = 0;
847         bdev->driver->evict_flags(bo, &placement);
848         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
849                                 no_wait_gpu);
850         if (ret) {
851                 if (ret != -ERESTARTSYS) {
852                         pr_err("Failed to find memory space for buffer 0x%p eviction\n",
853                                bo);
854                         ttm_bo_mem_space_debug(bo, &placement);
855                 }
856                 goto out;
857         }
858
859         ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
860                                      no_wait_gpu);
861         if (ret) {
862                 if (ret != -ERESTARTSYS)
863                         pr_err("Buffer eviction failed\n");
864                 ttm_bo_mem_put(bo, &evict_mem);
865                 goto out;
866         }
867         bo->evicted = true;
868 out:
869         return ret;
870 }
871
872 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
873                                 uint32_t mem_type,
874                                 bool interruptible,
875                                 bool no_wait_gpu)
876 {
877         struct ttm_bo_global *glob = bdev->glob;
878         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
879         struct ttm_buffer_object *bo;
880         int ret = -EBUSY, put_count;
881
882         spin_lock(&glob->lru_lock);
883         list_for_each_entry(bo, &man->lru, lru) {
884                 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0);
885                 if (!ret)
886                         break;
887         }
888
889         if (ret) {
890                 spin_unlock(&glob->lru_lock);
891                 return ret;
892         }
893
894         kref_get(&bo->list_kref);
895
896         if (!list_empty(&bo->ddestroy)) {
897                 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
898                                                      no_wait_gpu);
899                 kref_put(&bo->list_kref, ttm_bo_release_list);
900                 return ret;
901         }
902
903         put_count = ttm_bo_del_from_lru(bo);
904         spin_unlock(&glob->lru_lock);
905
906         BUG_ON(ret != 0);
907
908         ttm_bo_list_ref_sub(bo, put_count, true);
909
910         ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
911         ttm_bo_unreserve(bo);
912
913         kref_put(&bo->list_kref, ttm_bo_release_list);
914         return ret;
915 }
916
917 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
918 {
919         struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
920
921         if (mem->mm_node)
922                 (*man->func->put_node)(man, mem);
923 }
924 EXPORT_SYMBOL(ttm_bo_mem_put);
925
926 /**
927  * Repeatedly evict memory from the LRU for @mem_type until we create enough
928  * space, or we've evicted everything and there isn't enough space.
929  */
930 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
931                                         uint32_t mem_type,
932                                         struct ttm_placement *placement,
933                                         struct ttm_mem_reg *mem,
934                                         bool interruptible,
935                                         bool no_wait_gpu)
936 {
937         struct ttm_bo_device *bdev = bo->bdev;
938         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
939         int ret;
940
941         do {
942                 ret = (*man->func->get_node)(man, bo, placement, mem);
943                 if (unlikely(ret != 0))
944                         return ret;
945                 if (mem->mm_node)
946                         break;
947                 ret = ttm_mem_evict_first(bdev, mem_type,
948                                           interruptible, no_wait_gpu);
949                 if (unlikely(ret != 0))
950                         return ret;
951         } while (1);
952         if (mem->mm_node == NULL)
953                 return -ENOMEM;
954         mem->mem_type = mem_type;
955         return 0;
956 }
957
958 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
959                                       uint32_t cur_placement,
960                                       uint32_t proposed_placement)
961 {
962         uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
963         uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
964
965         /**
966          * Keep current caching if possible.
967          */
968
969         if ((cur_placement & caching) != 0)
970                 result |= (cur_placement & caching);
971         else if ((man->default_caching & caching) != 0)
972                 result |= man->default_caching;
973         else if ((TTM_PL_FLAG_CACHED & caching) != 0)
974                 result |= TTM_PL_FLAG_CACHED;
975         else if ((TTM_PL_FLAG_WC & caching) != 0)
976                 result |= TTM_PL_FLAG_WC;
977         else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
978                 result |= TTM_PL_FLAG_UNCACHED;
979
980         return result;
981 }
982
983 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
984                                  uint32_t mem_type,
985                                  uint32_t proposed_placement,
986                                  uint32_t *masked_placement)
987 {
988         uint32_t cur_flags = ttm_bo_type_flags(mem_type);
989
990         if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
991                 return false;
992
993         if ((proposed_placement & man->available_caching) == 0)
994                 return false;
995
996         cur_flags |= (proposed_placement & man->available_caching);
997
998         *masked_placement = cur_flags;
999         return true;
1000 }
1001
1002 /**
1003  * Creates space for memory region @mem according to its type.
1004  *
1005  * This function first searches for free space in compatible memory types in
1006  * the priority order defined by the driver.  If free space isn't found, then
1007  * ttm_bo_mem_force_space is attempted in priority order to evict and find
1008  * space.
1009  */
1010 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
1011                         struct ttm_placement *placement,
1012                         struct ttm_mem_reg *mem,
1013                         bool interruptible,
1014                         bool no_wait_gpu)
1015 {
1016         struct ttm_bo_device *bdev = bo->bdev;
1017         struct ttm_mem_type_manager *man;
1018         uint32_t mem_type = TTM_PL_SYSTEM;
1019         uint32_t cur_flags = 0;
1020         bool type_found = false;
1021         bool type_ok = false;
1022         bool has_erestartsys = false;
1023         int i, ret;
1024
1025         mem->mm_node = NULL;
1026         for (i = 0; i < placement->num_placement; ++i) {
1027                 ret = ttm_mem_type_from_flags(placement->placement[i],
1028                                                 &mem_type);
1029                 if (ret)
1030                         return ret;
1031                 man = &bdev->man[mem_type];
1032
1033                 type_ok = ttm_bo_mt_compatible(man,
1034                                                 mem_type,
1035                                                 placement->placement[i],
1036                                                 &cur_flags);
1037
1038                 if (!type_ok)
1039                         continue;
1040
1041                 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1042                                                   cur_flags);
1043                 /*
1044                  * Use the access and other non-mapping-related flag bits from
1045                  * the memory placement flags to the current flags
1046                  */
1047                 ttm_flag_masked(&cur_flags, placement->placement[i],
1048                                 ~TTM_PL_MASK_MEMTYPE);
1049
1050                 if (mem_type == TTM_PL_SYSTEM)
1051                         break;
1052
1053                 if (man->has_type && man->use_type) {
1054                         type_found = true;
1055                         ret = (*man->func->get_node)(man, bo, placement, mem);
1056                         if (unlikely(ret))
1057                                 return ret;
1058                 }
1059                 if (mem->mm_node)
1060                         break;
1061         }
1062
1063         if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
1064                 mem->mem_type = mem_type;
1065                 mem->placement = cur_flags;
1066                 return 0;
1067         }
1068
1069         if (!type_found)
1070                 return -EINVAL;
1071
1072         for (i = 0; i < placement->num_busy_placement; ++i) {
1073                 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
1074                                                 &mem_type);
1075                 if (ret)
1076                         return ret;
1077                 man = &bdev->man[mem_type];
1078                 if (!man->has_type)
1079                         continue;
1080                 if (!ttm_bo_mt_compatible(man,
1081                                                 mem_type,
1082                                                 placement->busy_placement[i],
1083                                                 &cur_flags))
1084                         continue;
1085
1086                 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1087                                                   cur_flags);
1088                 /*
1089                  * Use the access and other non-mapping-related flag bits from
1090                  * the memory placement flags to the current flags
1091                  */
1092                 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1093                                 ~TTM_PL_MASK_MEMTYPE);
1094
1095
1096                 if (mem_type == TTM_PL_SYSTEM) {
1097                         mem->mem_type = mem_type;
1098                         mem->placement = cur_flags;
1099                         mem->mm_node = NULL;
1100                         return 0;
1101                 }
1102
1103                 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1104                                                 interruptible, no_wait_gpu);
1105                 if (ret == 0 && mem->mm_node) {
1106                         mem->placement = cur_flags;
1107                         return 0;
1108                 }
1109                 if (ret == -ERESTARTSYS)
1110                         has_erestartsys = true;
1111         }
1112         ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1113         return ret;
1114 }
1115 EXPORT_SYMBOL(ttm_bo_mem_space);
1116
1117 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1118                         struct ttm_placement *placement,
1119                         bool interruptible,
1120                         bool no_wait_gpu)
1121 {
1122         int ret = 0;
1123         struct ttm_mem_reg mem;
1124         struct ttm_bo_device *bdev = bo->bdev;
1125
1126         BUG_ON(!ttm_bo_is_reserved(bo));
1127
1128         /*
1129          * FIXME: It's possible to pipeline buffer moves.
1130          * Have the driver move function wait for idle when necessary,
1131          * instead of doing it here.
1132          */
1133         spin_lock(&bdev->fence_lock);
1134         ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1135         spin_unlock(&bdev->fence_lock);
1136         if (ret)
1137                 return ret;
1138         mem.num_pages = bo->num_pages;
1139         mem.size = mem.num_pages << PAGE_SHIFT;
1140         mem.page_alignment = bo->mem.page_alignment;
1141         mem.bus.io_reserved_vm = false;
1142         mem.bus.io_reserved_count = 0;
1143         /*
1144          * Determine where to move the buffer.
1145          */
1146         ret = ttm_bo_mem_space(bo, placement, &mem,
1147                                interruptible, no_wait_gpu);
1148         if (ret)
1149                 goto out_unlock;
1150         ret = ttm_bo_handle_move_mem(bo, &mem, false,
1151                                      interruptible, no_wait_gpu);
1152 out_unlock:
1153         if (ret && mem.mm_node)
1154                 ttm_bo_mem_put(bo, &mem);
1155         return ret;
1156 }
1157
1158 static bool ttm_bo_mem_compat(struct ttm_placement *placement,
1159                               struct ttm_mem_reg *mem,
1160                               uint32_t *new_flags)
1161 {
1162         int i;
1163
1164         if (mem->mm_node && placement->lpfn != 0 &&
1165             (mem->start < placement->fpfn ||
1166              mem->start + mem->num_pages > placement->lpfn))
1167                 return false;
1168
1169         for (i = 0; i < placement->num_placement; i++) {
1170                 *new_flags = placement->placement[i];
1171                 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1172                     (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1173                         return true;
1174         }
1175
1176         for (i = 0; i < placement->num_busy_placement; i++) {
1177                 *new_flags = placement->busy_placement[i];
1178                 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1179                     (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1180                         return true;
1181         }
1182
1183         return false;
1184 }
1185
1186 int ttm_bo_validate(struct ttm_buffer_object *bo,
1187                         struct ttm_placement *placement,
1188                         bool interruptible,
1189                         bool no_wait_gpu)
1190 {
1191         int ret;
1192         uint32_t new_flags;
1193
1194         BUG_ON(!ttm_bo_is_reserved(bo));
1195         /* Check that range is valid */
1196         if (placement->lpfn || placement->fpfn)
1197                 if (placement->fpfn > placement->lpfn ||
1198                         (placement->lpfn - placement->fpfn) < bo->num_pages)
1199                         return -EINVAL;
1200         /*
1201          * Check whether we need to move buffer.
1202          */
1203         if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1204                 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1205                                          no_wait_gpu);
1206                 if (ret)
1207                         return ret;
1208         } else {
1209                 /*
1210                  * Use the access and other non-mapping-related flag bits from
1211                  * the compatible memory placement flags to the active flags
1212                  */
1213                 ttm_flag_masked(&bo->mem.placement, new_flags,
1214                                 ~TTM_PL_MASK_MEMTYPE);
1215         }
1216         /*
1217          * We might need to add a TTM.
1218          */
1219         if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1220                 ret = ttm_bo_add_ttm(bo, true);
1221                 if (ret)
1222                         return ret;
1223         }
1224         return 0;
1225 }
1226 EXPORT_SYMBOL(ttm_bo_validate);
1227
1228 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1229                                 struct ttm_placement *placement)
1230 {
1231         BUG_ON((placement->fpfn || placement->lpfn) &&
1232                (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1233
1234         return 0;
1235 }
1236
1237 int ttm_bo_init(struct ttm_bo_device *bdev,
1238                 struct ttm_buffer_object *bo,
1239                 unsigned long size,
1240                 enum ttm_bo_type type,
1241                 struct ttm_placement *placement,
1242                 uint32_t page_alignment,
1243                 bool interruptible,
1244                 struct file *persistent_swap_storage,
1245                 size_t acc_size,
1246                 struct sg_table *sg,
1247                 void (*destroy) (struct ttm_buffer_object *))
1248 {
1249         int ret = 0;
1250         unsigned long num_pages;
1251         struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1252
1253         ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1254         if (ret) {
1255                 pr_err("Out of kernel memory\n");
1256                 if (destroy)
1257                         (*destroy)(bo);
1258                 else
1259                         kfree(bo);
1260                 return -ENOMEM;
1261         }
1262
1263         num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1264         if (num_pages == 0) {
1265                 pr_err("Illegal buffer object size\n");
1266                 if (destroy)
1267                         (*destroy)(bo);
1268                 else
1269                         kfree(bo);
1270                 ttm_mem_global_free(mem_glob, acc_size);
1271                 return -EINVAL;
1272         }
1273         bo->destroy = destroy;
1274
1275         kref_init(&bo->kref);
1276         kref_init(&bo->list_kref);
1277         atomic_set(&bo->cpu_writers, 0);
1278         atomic_set(&bo->reserved, 1);
1279         init_waitqueue_head(&bo->event_queue);
1280         INIT_LIST_HEAD(&bo->lru);
1281         INIT_LIST_HEAD(&bo->ddestroy);
1282         INIT_LIST_HEAD(&bo->swap);
1283         INIT_LIST_HEAD(&bo->io_reserve_lru);
1284         bo->bdev = bdev;
1285         bo->glob = bdev->glob;
1286         bo->type = type;
1287         bo->num_pages = num_pages;
1288         bo->mem.size = num_pages << PAGE_SHIFT;
1289         bo->mem.mem_type = TTM_PL_SYSTEM;
1290         bo->mem.num_pages = bo->num_pages;
1291         bo->mem.mm_node = NULL;
1292         bo->mem.page_alignment = page_alignment;
1293         bo->mem.bus.io_reserved_vm = false;
1294         bo->mem.bus.io_reserved_count = 0;
1295         bo->priv_flags = 0;
1296         bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1297         bo->seq_valid = false;
1298         bo->persistent_swap_storage = persistent_swap_storage;
1299         bo->acc_size = acc_size;
1300         bo->sg = sg;
1301         atomic_inc(&bo->glob->bo_count);
1302
1303         ret = ttm_bo_check_placement(bo, placement);
1304         if (unlikely(ret != 0))
1305                 goto out_err;
1306
1307         /*
1308          * For ttm_bo_type_device buffers, allocate
1309          * address space from the device.
1310          */
1311         if (bo->type == ttm_bo_type_device ||
1312             bo->type == ttm_bo_type_sg) {
1313                 ret = ttm_bo_setup_vm(bo);
1314                 if (ret)
1315                         goto out_err;
1316         }
1317
1318         ret = ttm_bo_validate(bo, placement, interruptible, false);
1319         if (ret)
1320                 goto out_err;
1321
1322         ttm_bo_unreserve(bo);
1323         return 0;
1324
1325 out_err:
1326         ttm_bo_unreserve(bo);
1327         ttm_bo_unref(&bo);
1328
1329         return ret;
1330 }
1331 EXPORT_SYMBOL(ttm_bo_init);
1332
1333 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1334                        unsigned long bo_size,
1335                        unsigned struct_size)
1336 {
1337         unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1338         size_t size = 0;
1339
1340         size += ttm_round_pot(struct_size);
1341         size += PAGE_ALIGN(npages * sizeof(void *));
1342         size += ttm_round_pot(sizeof(struct ttm_tt));
1343         return size;
1344 }
1345 EXPORT_SYMBOL(ttm_bo_acc_size);
1346
1347 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1348                            unsigned long bo_size,
1349                            unsigned struct_size)
1350 {
1351         unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1352         size_t size = 0;
1353
1354         size += ttm_round_pot(struct_size);
1355         size += PAGE_ALIGN(npages * sizeof(void *));
1356         size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1357         size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1358         return size;
1359 }
1360 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1361
1362 int ttm_bo_create(struct ttm_bo_device *bdev,
1363                         unsigned long size,
1364                         enum ttm_bo_type type,
1365                         struct ttm_placement *placement,
1366                         uint32_t page_alignment,
1367                         bool interruptible,
1368                         struct file *persistent_swap_storage,
1369                         struct ttm_buffer_object **p_bo)
1370 {
1371         struct ttm_buffer_object *bo;
1372         size_t acc_size;
1373         int ret;
1374
1375         bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1376         if (unlikely(bo == NULL))
1377                 return -ENOMEM;
1378
1379         acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1380         ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1381                           interruptible, persistent_swap_storage, acc_size,
1382                           NULL, NULL);
1383         if (likely(ret == 0))
1384                 *p_bo = bo;
1385
1386         return ret;
1387 }
1388 EXPORT_SYMBOL(ttm_bo_create);
1389
1390 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1391                                         unsigned mem_type, bool allow_errors)
1392 {
1393         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1394         struct ttm_bo_global *glob = bdev->glob;
1395         int ret;
1396
1397         /*
1398          * Can't use standard list traversal since we're unlocking.
1399          */
1400
1401         spin_lock(&glob->lru_lock);
1402         while (!list_empty(&man->lru)) {
1403                 spin_unlock(&glob->lru_lock);
1404                 ret = ttm_mem_evict_first(bdev, mem_type, false, false);
1405                 if (ret) {
1406                         if (allow_errors) {
1407                                 return ret;
1408                         } else {
1409                                 pr_err("Cleanup eviction failed\n");
1410                         }
1411                 }
1412                 spin_lock(&glob->lru_lock);
1413         }
1414         spin_unlock(&glob->lru_lock);
1415         return 0;
1416 }
1417
1418 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1419 {
1420         struct ttm_mem_type_manager *man;
1421         int ret = -EINVAL;
1422
1423         if (mem_type >= TTM_NUM_MEM_TYPES) {
1424                 pr_err("Illegal memory type %d\n", mem_type);
1425                 return ret;
1426         }
1427         man = &bdev->man[mem_type];
1428
1429         if (!man->has_type) {
1430                 pr_err("Trying to take down uninitialized memory manager type %u\n",
1431                        mem_type);
1432                 return ret;
1433         }
1434
1435         man->use_type = false;
1436         man->has_type = false;
1437
1438         ret = 0;
1439         if (mem_type > 0) {
1440                 ttm_bo_force_list_clean(bdev, mem_type, false);
1441
1442                 ret = (*man->func->takedown)(man);
1443         }
1444
1445         return ret;
1446 }
1447 EXPORT_SYMBOL(ttm_bo_clean_mm);
1448
1449 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1450 {
1451         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1452
1453         if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1454                 pr_err("Illegal memory manager memory type %u\n", mem_type);
1455                 return -EINVAL;
1456         }
1457
1458         if (!man->has_type) {
1459                 pr_err("Memory type %u has not been initialized\n", mem_type);
1460                 return 0;
1461         }
1462
1463         return ttm_bo_force_list_clean(bdev, mem_type, true);
1464 }
1465 EXPORT_SYMBOL(ttm_bo_evict_mm);
1466
1467 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1468                         unsigned long p_size)
1469 {
1470         int ret = -EINVAL;
1471         struct ttm_mem_type_manager *man;
1472
1473         BUG_ON(type >= TTM_NUM_MEM_TYPES);
1474         man = &bdev->man[type];
1475         BUG_ON(man->has_type);
1476         man->io_reserve_fastpath = true;
1477         man->use_io_reserve_lru = false;
1478         mutex_init(&man->io_reserve_mutex);
1479         INIT_LIST_HEAD(&man->io_reserve_lru);
1480
1481         ret = bdev->driver->init_mem_type(bdev, type, man);
1482         if (ret)
1483                 return ret;
1484         man->bdev = bdev;
1485
1486         ret = 0;
1487         if (type != TTM_PL_SYSTEM) {
1488                 ret = (*man->func->init)(man, p_size);
1489                 if (ret)
1490                         return ret;
1491         }
1492         man->has_type = true;
1493         man->use_type = true;
1494         man->size = p_size;
1495
1496         INIT_LIST_HEAD(&man->lru);
1497
1498         return 0;
1499 }
1500 EXPORT_SYMBOL(ttm_bo_init_mm);
1501
1502 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1503 {
1504         struct ttm_bo_global *glob =
1505                 container_of(kobj, struct ttm_bo_global, kobj);
1506
1507         ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1508         __free_page(glob->dummy_read_page);
1509         kfree(glob);
1510 }
1511
1512 void ttm_bo_global_release(struct drm_global_reference *ref)
1513 {
1514         struct ttm_bo_global *glob = ref->object;
1515
1516         kobject_del(&glob->kobj);
1517         kobject_put(&glob->kobj);
1518 }
1519 EXPORT_SYMBOL(ttm_bo_global_release);
1520
1521 int ttm_bo_global_init(struct drm_global_reference *ref)
1522 {
1523         struct ttm_bo_global_ref *bo_ref =
1524                 container_of(ref, struct ttm_bo_global_ref, ref);
1525         struct ttm_bo_global *glob = ref->object;
1526         int ret;
1527
1528         mutex_init(&glob->device_list_mutex);
1529         spin_lock_init(&glob->lru_lock);
1530         glob->mem_glob = bo_ref->mem_glob;
1531         glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1532
1533         if (unlikely(glob->dummy_read_page == NULL)) {
1534                 ret = -ENOMEM;
1535                 goto out_no_drp;
1536         }
1537
1538         INIT_LIST_HEAD(&glob->swap_lru);
1539         INIT_LIST_HEAD(&glob->device_list);
1540
1541         ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1542         ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1543         if (unlikely(ret != 0)) {
1544                 pr_err("Could not register buffer object swapout\n");
1545                 goto out_no_shrink;
1546         }
1547
1548         atomic_set(&glob->bo_count, 0);
1549
1550         ret = kobject_init_and_add(
1551                 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1552         if (unlikely(ret != 0))
1553                 kobject_put(&glob->kobj);
1554         return ret;
1555 out_no_shrink:
1556         __free_page(glob->dummy_read_page);
1557 out_no_drp:
1558         kfree(glob);
1559         return ret;
1560 }
1561 EXPORT_SYMBOL(ttm_bo_global_init);
1562
1563
1564 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1565 {
1566         int ret = 0;
1567         unsigned i = TTM_NUM_MEM_TYPES;
1568         struct ttm_mem_type_manager *man;
1569         struct ttm_bo_global *glob = bdev->glob;
1570
1571         while (i--) {
1572                 man = &bdev->man[i];
1573                 if (man->has_type) {
1574                         man->use_type = false;
1575                         if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1576                                 ret = -EBUSY;
1577                                 pr_err("DRM memory manager type %d is not clean\n",
1578                                        i);
1579                         }
1580                         man->has_type = false;
1581                 }
1582         }
1583
1584         mutex_lock(&glob->device_list_mutex);
1585         list_del(&bdev->device_list);
1586         mutex_unlock(&glob->device_list_mutex);
1587
1588         cancel_delayed_work_sync(&bdev->wq);
1589
1590         while (ttm_bo_delayed_delete(bdev, true))
1591                 ;
1592
1593         spin_lock(&glob->lru_lock);
1594         if (list_empty(&bdev->ddestroy))
1595                 TTM_DEBUG("Delayed destroy list was clean\n");
1596
1597         if (list_empty(&bdev->man[0].lru))
1598                 TTM_DEBUG("Swap list was clean\n");
1599         spin_unlock(&glob->lru_lock);
1600
1601         BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1602         write_lock(&bdev->vm_lock);
1603         drm_mm_takedown(&bdev->addr_space_mm);
1604         write_unlock(&bdev->vm_lock);
1605
1606         return ret;
1607 }
1608 EXPORT_SYMBOL(ttm_bo_device_release);
1609
1610 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1611                        struct ttm_bo_global *glob,
1612                        struct ttm_bo_driver *driver,
1613                        uint64_t file_page_offset,
1614                        bool need_dma32)
1615 {
1616         int ret = -EINVAL;
1617
1618         rwlock_init(&bdev->vm_lock);
1619         bdev->driver = driver;
1620
1621         memset(bdev->man, 0, sizeof(bdev->man));
1622
1623         /*
1624          * Initialize the system memory buffer type.
1625          * Other types need to be driver / IOCTL initialized.
1626          */
1627         ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1628         if (unlikely(ret != 0))
1629                 goto out_no_sys;
1630
1631         bdev->addr_space_rb = RB_ROOT;
1632         ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1633         if (unlikely(ret != 0))
1634                 goto out_no_addr_mm;
1635
1636         INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1637         INIT_LIST_HEAD(&bdev->ddestroy);
1638         bdev->dev_mapping = NULL;
1639         bdev->glob = glob;
1640         bdev->need_dma32 = need_dma32;
1641         bdev->val_seq = 0;
1642         spin_lock_init(&bdev->fence_lock);
1643         mutex_lock(&glob->device_list_mutex);
1644         list_add_tail(&bdev->device_list, &glob->device_list);
1645         mutex_unlock(&glob->device_list_mutex);
1646
1647         return 0;
1648 out_no_addr_mm:
1649         ttm_bo_clean_mm(bdev, 0);
1650 out_no_sys:
1651         return ret;
1652 }
1653 EXPORT_SYMBOL(ttm_bo_device_init);
1654
1655 /*
1656  * buffer object vm functions.
1657  */
1658
1659 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1660 {
1661         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1662
1663         if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1664                 if (mem->mem_type == TTM_PL_SYSTEM)
1665                         return false;
1666
1667                 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1668                         return false;
1669
1670                 if (mem->placement & TTM_PL_FLAG_CACHED)
1671                         return false;
1672         }
1673         return true;
1674 }
1675
1676 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1677 {
1678         struct ttm_bo_device *bdev = bo->bdev;
1679         loff_t offset = (loff_t) bo->addr_space_offset;
1680         loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1681
1682         if (!bdev->dev_mapping)
1683                 return;
1684         unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1685         ttm_mem_io_free_vm(bo);
1686 }
1687
1688 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1689 {
1690         struct ttm_bo_device *bdev = bo->bdev;
1691         struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1692
1693         ttm_mem_io_lock(man, false);
1694         ttm_bo_unmap_virtual_locked(bo);
1695         ttm_mem_io_unlock(man);
1696 }
1697
1698
1699 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1700
1701 static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1702 {
1703         struct ttm_bo_device *bdev = bo->bdev;
1704         struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1705         struct rb_node *parent = NULL;
1706         struct ttm_buffer_object *cur_bo;
1707         unsigned long offset = bo->vm_node->start;
1708         unsigned long cur_offset;
1709
1710         while (*cur) {
1711                 parent = *cur;
1712                 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1713                 cur_offset = cur_bo->vm_node->start;
1714                 if (offset < cur_offset)
1715                         cur = &parent->rb_left;
1716                 else if (offset > cur_offset)
1717                         cur = &parent->rb_right;
1718                 else
1719                         BUG();
1720         }
1721
1722         rb_link_node(&bo->vm_rb, parent, cur);
1723         rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1724 }
1725
1726 /**
1727  * ttm_bo_setup_vm:
1728  *
1729  * @bo: the buffer to allocate address space for
1730  *
1731  * Allocate address space in the drm device so that applications
1732  * can mmap the buffer and access the contents. This only
1733  * applies to ttm_bo_type_device objects as others are not
1734  * placed in the drm device address space.
1735  */
1736
1737 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1738 {
1739         struct ttm_bo_device *bdev = bo->bdev;
1740         int ret;
1741
1742 retry_pre_get:
1743         ret = drm_mm_pre_get(&bdev->addr_space_mm);
1744         if (unlikely(ret != 0))
1745                 return ret;
1746
1747         write_lock(&bdev->vm_lock);
1748         bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1749                                          bo->mem.num_pages, 0, 0);
1750
1751         if (unlikely(bo->vm_node == NULL)) {
1752                 ret = -ENOMEM;
1753                 goto out_unlock;
1754         }
1755
1756         bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1757                                               bo->mem.num_pages, 0);
1758
1759         if (unlikely(bo->vm_node == NULL)) {
1760                 write_unlock(&bdev->vm_lock);
1761                 goto retry_pre_get;
1762         }
1763
1764         ttm_bo_vm_insert_rb(bo);
1765         write_unlock(&bdev->vm_lock);
1766         bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1767
1768         return 0;
1769 out_unlock:
1770         write_unlock(&bdev->vm_lock);
1771         return ret;
1772 }
1773
1774 int ttm_bo_wait(struct ttm_buffer_object *bo,
1775                 bool lazy, bool interruptible, bool no_wait)
1776 {
1777         struct ttm_bo_driver *driver = bo->bdev->driver;
1778         struct ttm_bo_device *bdev = bo->bdev;
1779         void *sync_obj;
1780         int ret = 0;
1781
1782         if (likely(bo->sync_obj == NULL))
1783                 return 0;
1784
1785         while (bo->sync_obj) {
1786
1787                 if (driver->sync_obj_signaled(bo->sync_obj)) {
1788                         void *tmp_obj = bo->sync_obj;
1789                         bo->sync_obj = NULL;
1790                         clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1791                         spin_unlock(&bdev->fence_lock);
1792                         driver->sync_obj_unref(&tmp_obj);
1793                         spin_lock(&bdev->fence_lock);
1794                         continue;
1795                 }
1796
1797                 if (no_wait)
1798                         return -EBUSY;
1799
1800                 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1801                 spin_unlock(&bdev->fence_lock);
1802                 ret = driver->sync_obj_wait(sync_obj,
1803                                             lazy, interruptible);
1804                 if (unlikely(ret != 0)) {
1805                         driver->sync_obj_unref(&sync_obj);
1806                         spin_lock(&bdev->fence_lock);
1807                         return ret;
1808                 }
1809                 spin_lock(&bdev->fence_lock);
1810                 if (likely(bo->sync_obj == sync_obj)) {
1811                         void *tmp_obj = bo->sync_obj;
1812                         bo->sync_obj = NULL;
1813                         clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1814                                   &bo->priv_flags);
1815                         spin_unlock(&bdev->fence_lock);
1816                         driver->sync_obj_unref(&sync_obj);
1817                         driver->sync_obj_unref(&tmp_obj);
1818                         spin_lock(&bdev->fence_lock);
1819                 } else {
1820                         spin_unlock(&bdev->fence_lock);
1821                         driver->sync_obj_unref(&sync_obj);
1822                         spin_lock(&bdev->fence_lock);
1823                 }
1824         }
1825         return 0;
1826 }
1827 EXPORT_SYMBOL(ttm_bo_wait);
1828
1829 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1830 {
1831         struct ttm_bo_device *bdev = bo->bdev;
1832         int ret = 0;
1833
1834         /*
1835          * Using ttm_bo_reserve makes sure the lru lists are updated.
1836          */
1837
1838         ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1839         if (unlikely(ret != 0))
1840                 return ret;
1841         spin_lock(&bdev->fence_lock);
1842         ret = ttm_bo_wait(bo, false, true, no_wait);
1843         spin_unlock(&bdev->fence_lock);
1844         if (likely(ret == 0))
1845                 atomic_inc(&bo->cpu_writers);
1846         ttm_bo_unreserve(bo);
1847         return ret;
1848 }
1849 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1850
1851 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1852 {
1853         atomic_dec(&bo->cpu_writers);
1854 }
1855 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1856
1857 /**
1858  * A buffer object shrink method that tries to swap out the first
1859  * buffer object on the bo_global::swap_lru list.
1860  */
1861
1862 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1863 {
1864         struct ttm_bo_global *glob =
1865             container_of(shrink, struct ttm_bo_global, shrink);
1866         struct ttm_buffer_object *bo;
1867         int ret = -EBUSY;
1868         int put_count;
1869         uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1870
1871         spin_lock(&glob->lru_lock);
1872         list_for_each_entry(bo, &glob->swap_lru, swap) {
1873                 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0);
1874                 if (!ret)
1875                         break;
1876         }
1877
1878         if (ret) {
1879                 spin_unlock(&glob->lru_lock);
1880                 return ret;
1881         }
1882
1883         kref_get(&bo->list_kref);
1884
1885         if (!list_empty(&bo->ddestroy)) {
1886                 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1887                 kref_put(&bo->list_kref, ttm_bo_release_list);
1888                 return ret;
1889         }
1890
1891         put_count = ttm_bo_del_from_lru(bo);
1892         spin_unlock(&glob->lru_lock);
1893
1894         ttm_bo_list_ref_sub(bo, put_count, true);
1895
1896         /**
1897          * Wait for GPU, then move to system cached.
1898          */
1899
1900         spin_lock(&bo->bdev->fence_lock);
1901         ret = ttm_bo_wait(bo, false, false, false);
1902         spin_unlock(&bo->bdev->fence_lock);
1903
1904         if (unlikely(ret != 0))
1905                 goto out;
1906
1907         if ((bo->mem.placement & swap_placement) != swap_placement) {
1908                 struct ttm_mem_reg evict_mem;
1909
1910                 evict_mem = bo->mem;
1911                 evict_mem.mm_node = NULL;
1912                 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1913                 evict_mem.mem_type = TTM_PL_SYSTEM;
1914
1915                 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1916                                              false, false);
1917                 if (unlikely(ret != 0))
1918                         goto out;
1919         }
1920
1921         ttm_bo_unmap_virtual(bo);
1922
1923         /**
1924          * Swap out. Buffer will be swapped in again as soon as
1925          * anyone tries to access a ttm page.
1926          */
1927
1928         if (bo->bdev->driver->swap_notify)
1929                 bo->bdev->driver->swap_notify(bo);
1930
1931         ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1932 out:
1933
1934         /**
1935          *
1936          * Unreserve without putting on LRU to avoid swapping out an
1937          * already swapped buffer.
1938          */
1939
1940         atomic_set(&bo->reserved, 0);
1941         wake_up_all(&bo->event_queue);
1942         kref_put(&bo->list_kref, ttm_bo_release_list);
1943         return ret;
1944 }
1945
1946 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1947 {
1948         while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1949                 ;
1950 }
1951 EXPORT_SYMBOL(ttm_bo_swapout_all);