Merge v5.8-rc1 into drm-misc-fixes
[sfrench/cifs-2.6.git] / drivers / gpu / drm / ttm / ttm_bo.c
1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
3  *
4  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the
9  * "Software"), to deal in the Software without restriction, including
10  * without limitation the rights to use, copy, modify, merge, publish,
11  * distribute, sub license, and/or sell copies of the Software, and to
12  * permit persons to whom the Software is furnished to do so, subject to
13  * the following conditions:
14  *
15  * The above copyright notice and this permission notice (including the
16  * next paragraph) shall be included in all copies or substantial portions
17  * of the Software.
18  *
19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25  * USE OR OTHER DEALINGS IN THE SOFTWARE.
26  *
27  **************************************************************************/
28 /*
29  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30  */
31
32 #define pr_fmt(fmt) "[TTM] " fmt
33
34 #include <drm/ttm/ttm_module.h>
35 #include <drm/ttm/ttm_bo_driver.h>
36 #include <drm/ttm/ttm_placement.h>
37 #include <linux/jiffies.h>
38 #include <linux/slab.h>
39 #include <linux/sched.h>
40 #include <linux/mm.h>
41 #include <linux/file.h>
42 #include <linux/module.h>
43 #include <linux/atomic.h>
44 #include <linux/dma-resv.h>
45
46 static void ttm_bo_global_kobj_release(struct kobject *kobj);
47
48 /**
49  * ttm_global_mutex - protecting the global BO state
50  */
51 DEFINE_MUTEX(ttm_global_mutex);
52 unsigned ttm_bo_glob_use_count;
53 struct ttm_bo_global ttm_bo_glob;
54 EXPORT_SYMBOL(ttm_bo_glob);
55
56 static struct attribute ttm_bo_count = {
57         .name = "bo_count",
58         .mode = S_IRUGO
59 };
60
61 /* default destructor */
62 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
63 {
64         kfree(bo);
65 }
66
67 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
68                                           uint32_t *mem_type)
69 {
70         int pos;
71
72         pos = ffs(place->flags & TTM_PL_MASK_MEM);
73         if (unlikely(!pos))
74                 return -EINVAL;
75
76         *mem_type = pos - 1;
77         return 0;
78 }
79
80 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, struct drm_printer *p,
81                                int mem_type)
82 {
83         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
84
85         drm_printf(p, "    has_type: %d\n", man->has_type);
86         drm_printf(p, "    use_type: %d\n", man->use_type);
87         drm_printf(p, "    flags: 0x%08X\n", man->flags);
88         drm_printf(p, "    gpu_offset: 0x%08llX\n", man->gpu_offset);
89         drm_printf(p, "    size: %llu\n", man->size);
90         drm_printf(p, "    available_caching: 0x%08X\n", man->available_caching);
91         drm_printf(p, "    default_caching: 0x%08X\n", man->default_caching);
92         if (mem_type != TTM_PL_SYSTEM)
93                 (*man->func->debug)(man, p);
94 }
95
96 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
97                                         struct ttm_placement *placement)
98 {
99         struct drm_printer p = drm_debug_printer(TTM_PFX);
100         int i, ret, mem_type;
101
102         drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n",
103                    bo, bo->mem.num_pages, bo->mem.size >> 10,
104                    bo->mem.size >> 20);
105         for (i = 0; i < placement->num_placement; i++) {
106                 ret = ttm_mem_type_from_place(&placement->placement[i],
107                                                 &mem_type);
108                 if (ret)
109                         return;
110                 drm_printf(&p, "  placement[%d]=0x%08X (%d)\n",
111                            i, placement->placement[i].flags, mem_type);
112                 ttm_mem_type_debug(bo->bdev, &p, mem_type);
113         }
114 }
115
116 static ssize_t ttm_bo_global_show(struct kobject *kobj,
117                                   struct attribute *attr,
118                                   char *buffer)
119 {
120         struct ttm_bo_global *glob =
121                 container_of(kobj, struct ttm_bo_global, kobj);
122
123         return snprintf(buffer, PAGE_SIZE, "%d\n",
124                                 atomic_read(&glob->bo_count));
125 }
126
127 static struct attribute *ttm_bo_global_attrs[] = {
128         &ttm_bo_count,
129         NULL
130 };
131
132 static const struct sysfs_ops ttm_bo_global_ops = {
133         .show = &ttm_bo_global_show
134 };
135
136 static struct kobj_type ttm_bo_glob_kobj_type  = {
137         .release = &ttm_bo_global_kobj_release,
138         .sysfs_ops = &ttm_bo_global_ops,
139         .default_attrs = ttm_bo_global_attrs
140 };
141
142
143 static inline uint32_t ttm_bo_type_flags(unsigned type)
144 {
145         return 1 << (type);
146 }
147
148 static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo,
149                                   struct ttm_mem_reg *mem)
150 {
151         struct ttm_bo_device *bdev = bo->bdev;
152         struct ttm_mem_type_manager *man;
153
154         if (!list_empty(&bo->lru))
155                 return;
156
157         if (mem->placement & TTM_PL_FLAG_NO_EVICT)
158                 return;
159
160         man = &bdev->man[mem->mem_type];
161         list_add_tail(&bo->lru, &man->lru[bo->priority]);
162
163         if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm &&
164             !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG |
165                                      TTM_PAGE_FLAG_SWAPPED))) {
166                 list_add_tail(&bo->swap, &ttm_bo_glob.swap_lru[bo->priority]);
167         }
168 }
169
170 static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
171 {
172         struct ttm_bo_device *bdev = bo->bdev;
173         bool notify = false;
174
175         if (!list_empty(&bo->swap)) {
176                 list_del_init(&bo->swap);
177                 notify = true;
178         }
179         if (!list_empty(&bo->lru)) {
180                 list_del_init(&bo->lru);
181                 notify = true;
182         }
183
184         if (notify && bdev->driver->del_from_lru_notify)
185                 bdev->driver->del_from_lru_notify(bo);
186 }
187
188 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
189                                      struct ttm_buffer_object *bo)
190 {
191         if (!pos->first)
192                 pos->first = bo;
193         pos->last = bo;
194 }
195
196 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
197                              struct ttm_lru_bulk_move *bulk)
198 {
199         dma_resv_assert_held(bo->base.resv);
200
201         ttm_bo_del_from_lru(bo);
202         ttm_bo_add_mem_to_lru(bo, &bo->mem);
203
204         if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
205                 switch (bo->mem.mem_type) {
206                 case TTM_PL_TT:
207                         ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
208                         break;
209
210                 case TTM_PL_VRAM:
211                         ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
212                         break;
213                 }
214                 if (bo->ttm && !(bo->ttm->page_flags &
215                                  (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
216                         ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
217         }
218 }
219 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
220
221 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
222 {
223         unsigned i;
224
225         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
226                 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
227                 struct ttm_mem_type_manager *man;
228
229                 if (!pos->first)
230                         continue;
231
232                 dma_resv_assert_held(pos->first->base.resv);
233                 dma_resv_assert_held(pos->last->base.resv);
234
235                 man = &pos->first->bdev->man[TTM_PL_TT];
236                 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
237                                     &pos->last->lru);
238         }
239
240         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
241                 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
242                 struct ttm_mem_type_manager *man;
243
244                 if (!pos->first)
245                         continue;
246
247                 dma_resv_assert_held(pos->first->base.resv);
248                 dma_resv_assert_held(pos->last->base.resv);
249
250                 man = &pos->first->bdev->man[TTM_PL_VRAM];
251                 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
252                                     &pos->last->lru);
253         }
254
255         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
256                 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
257                 struct list_head *lru;
258
259                 if (!pos->first)
260                         continue;
261
262                 dma_resv_assert_held(pos->first->base.resv);
263                 dma_resv_assert_held(pos->last->base.resv);
264
265                 lru = &ttm_bo_glob.swap_lru[i];
266                 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
267         }
268 }
269 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
270
271 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
272                                   struct ttm_mem_reg *mem, bool evict,
273                                   struct ttm_operation_ctx *ctx)
274 {
275         struct ttm_bo_device *bdev = bo->bdev;
276         bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
277         bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
278         struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
279         struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
280         int ret = 0;
281
282         if (old_is_pci || new_is_pci ||
283             ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
284                 ret = ttm_mem_io_lock(old_man, true);
285                 if (unlikely(ret != 0))
286                         goto out_err;
287                 ttm_bo_unmap_virtual_locked(bo);
288                 ttm_mem_io_unlock(old_man);
289         }
290
291         /*
292          * Create and bind a ttm if required.
293          */
294
295         if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
296                 if (bo->ttm == NULL) {
297                         bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
298                         ret = ttm_tt_create(bo, zero);
299                         if (ret)
300                                 goto out_err;
301                 }
302
303                 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
304                 if (ret)
305                         goto out_err;
306
307                 if (mem->mem_type != TTM_PL_SYSTEM) {
308                         ret = ttm_tt_bind(bo->ttm, mem, ctx);
309                         if (ret)
310                                 goto out_err;
311                 }
312
313                 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
314                         if (bdev->driver->move_notify)
315                                 bdev->driver->move_notify(bo, evict, mem);
316                         bo->mem = *mem;
317                         mem->mm_node = NULL;
318                         goto moved;
319                 }
320         }
321
322         if (bdev->driver->move_notify)
323                 bdev->driver->move_notify(bo, evict, mem);
324
325         if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
326             !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
327                 ret = ttm_bo_move_ttm(bo, ctx, mem);
328         else if (bdev->driver->move)
329                 ret = bdev->driver->move(bo, evict, ctx, mem);
330         else
331                 ret = ttm_bo_move_memcpy(bo, ctx, mem);
332
333         if (ret) {
334                 if (bdev->driver->move_notify) {
335                         swap(*mem, bo->mem);
336                         bdev->driver->move_notify(bo, false, mem);
337                         swap(*mem, bo->mem);
338                 }
339
340                 goto out_err;
341         }
342
343 moved:
344         bo->evicted = false;
345
346         if (bo->mem.mm_node)
347                 bo->offset = (bo->mem.start << PAGE_SHIFT) +
348                     bdev->man[bo->mem.mem_type].gpu_offset;
349         else
350                 bo->offset = 0;
351
352         ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
353         return 0;
354
355 out_err:
356         new_man = &bdev->man[bo->mem.mem_type];
357         if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
358                 ttm_tt_destroy(bo->ttm);
359                 bo->ttm = NULL;
360         }
361
362         return ret;
363 }
364
365 /**
366  * Call bo::reserved.
367  * Will release GPU memory type usage on destruction.
368  * This is the place to put in driver specific hooks to release
369  * driver private resources.
370  * Will release the bo::reserved lock.
371  */
372
373 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
374 {
375         if (bo->bdev->driver->move_notify)
376                 bo->bdev->driver->move_notify(bo, false, NULL);
377
378         ttm_tt_destroy(bo->ttm);
379         bo->ttm = NULL;
380         ttm_bo_mem_put(bo, &bo->mem);
381 }
382
383 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
384 {
385         int r;
386
387         if (bo->base.resv == &bo->base._resv)
388                 return 0;
389
390         BUG_ON(!dma_resv_trylock(&bo->base._resv));
391
392         r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
393         dma_resv_unlock(&bo->base._resv);
394         if (r)
395                 return r;
396
397         if (bo->type != ttm_bo_type_sg) {
398                 /* This works because the BO is about to be destroyed and nobody
399                  * reference it any more. The only tricky case is the trylock on
400                  * the resv object while holding the lru_lock.
401                  */
402                 spin_lock(&ttm_bo_glob.lru_lock);
403                 bo->base.resv = &bo->base._resv;
404                 spin_unlock(&ttm_bo_glob.lru_lock);
405         }
406
407         return r;
408 }
409
410 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
411 {
412         struct dma_resv *resv = &bo->base._resv;
413         struct dma_resv_list *fobj;
414         struct dma_fence *fence;
415         int i;
416
417         rcu_read_lock();
418         fobj = rcu_dereference(resv->fence);
419         fence = rcu_dereference(resv->fence_excl);
420         if (fence && !fence->ops->signaled)
421                 dma_fence_enable_sw_signaling(fence);
422
423         for (i = 0; fobj && i < fobj->shared_count; ++i) {
424                 fence = rcu_dereference(fobj->shared[i]);
425
426                 if (!fence->ops->signaled)
427                         dma_fence_enable_sw_signaling(fence);
428         }
429         rcu_read_unlock();
430 }
431
432 /**
433  * function ttm_bo_cleanup_refs
434  * If bo idle, remove from lru lists, and unref.
435  * If not idle, block if possible.
436  *
437  * Must be called with lru_lock and reservation held, this function
438  * will drop the lru lock and optionally the reservation lock before returning.
439  *
440  * @interruptible         Any sleeps should occur interruptibly.
441  * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
442  * @unlock_resv           Unlock the reservation lock as well.
443  */
444
445 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
446                                bool interruptible, bool no_wait_gpu,
447                                bool unlock_resv)
448 {
449         struct dma_resv *resv = &bo->base._resv;
450         int ret;
451
452         if (dma_resv_test_signaled_rcu(resv, true))
453                 ret = 0;
454         else
455                 ret = -EBUSY;
456
457         if (ret && !no_wait_gpu) {
458                 long lret;
459
460                 if (unlock_resv)
461                         dma_resv_unlock(bo->base.resv);
462                 spin_unlock(&ttm_bo_glob.lru_lock);
463
464                 lret = dma_resv_wait_timeout_rcu(resv, true, interruptible,
465                                                  30 * HZ);
466
467                 if (lret < 0)
468                         return lret;
469                 else if (lret == 0)
470                         return -EBUSY;
471
472                 spin_lock(&ttm_bo_glob.lru_lock);
473                 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
474                         /*
475                          * We raced, and lost, someone else holds the reservation now,
476                          * and is probably busy in ttm_bo_cleanup_memtype_use.
477                          *
478                          * Even if it's not the case, because we finished waiting any
479                          * delayed destruction would succeed, so just return success
480                          * here.
481                          */
482                         spin_unlock(&ttm_bo_glob.lru_lock);
483                         return 0;
484                 }
485                 ret = 0;
486         }
487
488         if (ret || unlikely(list_empty(&bo->ddestroy))) {
489                 if (unlock_resv)
490                         dma_resv_unlock(bo->base.resv);
491                 spin_unlock(&ttm_bo_glob.lru_lock);
492                 return ret;
493         }
494
495         ttm_bo_del_from_lru(bo);
496         list_del_init(&bo->ddestroy);
497         spin_unlock(&ttm_bo_glob.lru_lock);
498         ttm_bo_cleanup_memtype_use(bo);
499
500         if (unlock_resv)
501                 dma_resv_unlock(bo->base.resv);
502
503         ttm_bo_put(bo);
504
505         return 0;
506 }
507
508 /**
509  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
510  * encountered buffers.
511  */
512 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
513 {
514         struct ttm_bo_global *glob = &ttm_bo_glob;
515         struct list_head removed;
516         bool empty;
517
518         INIT_LIST_HEAD(&removed);
519
520         spin_lock(&glob->lru_lock);
521         while (!list_empty(&bdev->ddestroy)) {
522                 struct ttm_buffer_object *bo;
523
524                 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
525                                       ddestroy);
526                 list_move_tail(&bo->ddestroy, &removed);
527                 if (!ttm_bo_get_unless_zero(bo))
528                         continue;
529
530                 if (remove_all || bo->base.resv != &bo->base._resv) {
531                         spin_unlock(&glob->lru_lock);
532                         dma_resv_lock(bo->base.resv, NULL);
533
534                         spin_lock(&glob->lru_lock);
535                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
536
537                 } else if (dma_resv_trylock(bo->base.resv)) {
538                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
539                 } else {
540                         spin_unlock(&glob->lru_lock);
541                 }
542
543                 ttm_bo_put(bo);
544                 spin_lock(&glob->lru_lock);
545         }
546         list_splice_tail(&removed, &bdev->ddestroy);
547         empty = list_empty(&bdev->ddestroy);
548         spin_unlock(&glob->lru_lock);
549
550         return empty;
551 }
552
553 static void ttm_bo_delayed_workqueue(struct work_struct *work)
554 {
555         struct ttm_bo_device *bdev =
556             container_of(work, struct ttm_bo_device, wq.work);
557
558         if (!ttm_bo_delayed_delete(bdev, false))
559                 schedule_delayed_work(&bdev->wq,
560                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
561 }
562
563 static void ttm_bo_release(struct kref *kref)
564 {
565         struct ttm_buffer_object *bo =
566             container_of(kref, struct ttm_buffer_object, kref);
567         struct ttm_bo_device *bdev = bo->bdev;
568         struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
569         size_t acc_size = bo->acc_size;
570         int ret;
571
572         if (!bo->deleted) {
573                 ret = ttm_bo_individualize_resv(bo);
574                 if (ret) {
575                         /* Last resort, if we fail to allocate memory for the
576                          * fences block for the BO to become idle
577                          */
578                         dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
579                                                   30 * HZ);
580                 }
581
582                 if (bo->bdev->driver->release_notify)
583                         bo->bdev->driver->release_notify(bo);
584
585                 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
586                 ttm_mem_io_lock(man, false);
587                 ttm_mem_io_free_vm(bo);
588                 ttm_mem_io_unlock(man);
589         }
590
591         if (!dma_resv_test_signaled_rcu(bo->base.resv, true) ||
592             !dma_resv_trylock(bo->base.resv)) {
593                 /* The BO is not idle, resurrect it for delayed destroy */
594                 ttm_bo_flush_all_fences(bo);
595                 bo->deleted = true;
596
597                 spin_lock(&ttm_bo_glob.lru_lock);
598
599                 /*
600                  * Make NO_EVICT bos immediately available to
601                  * shrinkers, now that they are queued for
602                  * destruction.
603                  */
604                 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
605                         bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
606                         ttm_bo_del_from_lru(bo);
607                         ttm_bo_add_mem_to_lru(bo, &bo->mem);
608                 }
609
610                 kref_init(&bo->kref);
611                 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
612                 spin_unlock(&ttm_bo_glob.lru_lock);
613
614                 schedule_delayed_work(&bdev->wq,
615                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
616                 return;
617         }
618
619         spin_lock(&ttm_bo_glob.lru_lock);
620         ttm_bo_del_from_lru(bo);
621         list_del(&bo->ddestroy);
622         spin_unlock(&ttm_bo_glob.lru_lock);
623
624         ttm_bo_cleanup_memtype_use(bo);
625         dma_resv_unlock(bo->base.resv);
626
627         BUG_ON(bo->mem.mm_node != NULL);
628         atomic_dec(&ttm_bo_glob.bo_count);
629         dma_fence_put(bo->moving);
630         if (!ttm_bo_uses_embedded_gem_object(bo))
631                 dma_resv_fini(&bo->base._resv);
632         bo->destroy(bo);
633         ttm_mem_global_free(&ttm_mem_glob, acc_size);
634 }
635
636 void ttm_bo_put(struct ttm_buffer_object *bo)
637 {
638         kref_put(&bo->kref, ttm_bo_release);
639 }
640 EXPORT_SYMBOL(ttm_bo_put);
641
642 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
643 {
644         return cancel_delayed_work_sync(&bdev->wq);
645 }
646 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
647
648 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
649 {
650         if (resched)
651                 schedule_delayed_work(&bdev->wq,
652                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
653 }
654 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
655
656 static int ttm_bo_evict(struct ttm_buffer_object *bo,
657                         struct ttm_operation_ctx *ctx)
658 {
659         struct ttm_bo_device *bdev = bo->bdev;
660         struct ttm_mem_reg evict_mem;
661         struct ttm_placement placement;
662         int ret = 0;
663
664         dma_resv_assert_held(bo->base.resv);
665
666         placement.num_placement = 0;
667         placement.num_busy_placement = 0;
668         bdev->driver->evict_flags(bo, &placement);
669
670         if (!placement.num_placement && !placement.num_busy_placement) {
671                 ret = ttm_bo_pipeline_gutting(bo);
672                 if (ret)
673                         return ret;
674
675                 return ttm_tt_create(bo, false);
676         }
677
678         evict_mem = bo->mem;
679         evict_mem.mm_node = NULL;
680         evict_mem.bus.io_reserved_vm = false;
681         evict_mem.bus.io_reserved_count = 0;
682
683         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
684         if (ret) {
685                 if (ret != -ERESTARTSYS) {
686                         pr_err("Failed to find memory space for buffer 0x%p eviction\n",
687                                bo);
688                         ttm_bo_mem_space_debug(bo, &placement);
689                 }
690                 goto out;
691         }
692
693         ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
694         if (unlikely(ret)) {
695                 if (ret != -ERESTARTSYS)
696                         pr_err("Buffer eviction failed\n");
697                 ttm_bo_mem_put(bo, &evict_mem);
698                 goto out;
699         }
700         bo->evicted = true;
701 out:
702         return ret;
703 }
704
705 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
706                               const struct ttm_place *place)
707 {
708         /* Don't evict this BO if it's outside of the
709          * requested placement range
710          */
711         if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
712             (place->lpfn && place->lpfn <= bo->mem.start))
713                 return false;
714
715         return true;
716 }
717 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
718
719 /**
720  * Check the target bo is allowable to be evicted or swapout, including cases:
721  *
722  * a. if share same reservation object with ctx->resv, have assumption
723  * reservation objects should already be locked, so not lock again and
724  * return true directly when either the opreation allow_reserved_eviction
725  * or the target bo already is in delayed free list;
726  *
727  * b. Otherwise, trylock it.
728  */
729 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
730                         struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
731 {
732         bool ret = false;
733
734         if (bo->base.resv == ctx->resv) {
735                 dma_resv_assert_held(bo->base.resv);
736                 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT)
737                         ret = true;
738                 *locked = false;
739                 if (busy)
740                         *busy = false;
741         } else {
742                 ret = dma_resv_trylock(bo->base.resv);
743                 *locked = ret;
744                 if (busy)
745                         *busy = !ret;
746         }
747
748         return ret;
749 }
750
751 /**
752  * ttm_mem_evict_wait_busy - wait for a busy BO to become available
753  *
754  * @busy_bo: BO which couldn't be locked with trylock
755  * @ctx: operation context
756  * @ticket: acquire ticket
757  *
758  * Try to lock a busy buffer object to avoid failing eviction.
759  */
760 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
761                                    struct ttm_operation_ctx *ctx,
762                                    struct ww_acquire_ctx *ticket)
763 {
764         int r;
765
766         if (!busy_bo || !ticket)
767                 return -EBUSY;
768
769         if (ctx->interruptible)
770                 r = dma_resv_lock_interruptible(busy_bo->base.resv,
771                                                           ticket);
772         else
773                 r = dma_resv_lock(busy_bo->base.resv, ticket);
774
775         /*
776          * TODO: It would be better to keep the BO locked until allocation is at
777          * least tried one more time, but that would mean a much larger rework
778          * of TTM.
779          */
780         if (!r)
781                 dma_resv_unlock(busy_bo->base.resv);
782
783         return r == -EDEADLK ? -EBUSY : r;
784 }
785
786 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
787                                uint32_t mem_type,
788                                const struct ttm_place *place,
789                                struct ttm_operation_ctx *ctx,
790                                struct ww_acquire_ctx *ticket)
791 {
792         struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
793         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
794         bool locked = false;
795         unsigned i;
796         int ret;
797
798         spin_lock(&ttm_bo_glob.lru_lock);
799         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
800                 list_for_each_entry(bo, &man->lru[i], lru) {
801                         bool busy;
802
803                         if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
804                                                             &busy)) {
805                                 if (busy && !busy_bo && ticket !=
806                                     dma_resv_locking_ctx(bo->base.resv))
807                                         busy_bo = bo;
808                                 continue;
809                         }
810
811                         if (place && !bdev->driver->eviction_valuable(bo,
812                                                                       place)) {
813                                 if (locked)
814                                         dma_resv_unlock(bo->base.resv);
815                                 continue;
816                         }
817                         if (!ttm_bo_get_unless_zero(bo)) {
818                                 if (locked)
819                                         dma_resv_unlock(bo->base.resv);
820                                 continue;
821                         }
822                         break;
823                 }
824
825                 /* If the inner loop terminated early, we have our candidate */
826                 if (&bo->lru != &man->lru[i])
827                         break;
828
829                 bo = NULL;
830         }
831
832         if (!bo) {
833                 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
834                         busy_bo = NULL;
835                 spin_unlock(&ttm_bo_glob.lru_lock);
836                 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
837                 if (busy_bo)
838                         ttm_bo_put(busy_bo);
839                 return ret;
840         }
841
842         if (bo->deleted) {
843                 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
844                                           ctx->no_wait_gpu, locked);
845                 ttm_bo_put(bo);
846                 return ret;
847         }
848
849         spin_unlock(&ttm_bo_glob.lru_lock);
850
851         ret = ttm_bo_evict(bo, ctx);
852         if (locked)
853                 ttm_bo_unreserve(bo);
854
855         ttm_bo_put(bo);
856         return ret;
857 }
858
859 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
860 {
861         struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
862
863         if (mem->mm_node)
864                 (*man->func->put_node)(man, mem);
865 }
866 EXPORT_SYMBOL(ttm_bo_mem_put);
867
868 /**
869  * Add the last move fence to the BO and reserve a new shared slot.
870  */
871 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
872                                  struct ttm_mem_type_manager *man,
873                                  struct ttm_mem_reg *mem,
874                                  bool no_wait_gpu)
875 {
876         struct dma_fence *fence;
877         int ret;
878
879         spin_lock(&man->move_lock);
880         fence = dma_fence_get(man->move);
881         spin_unlock(&man->move_lock);
882
883         if (!fence)
884                 return 0;
885
886         if (no_wait_gpu) {
887                 dma_fence_put(fence);
888                 return -EBUSY;
889         }
890
891         dma_resv_add_shared_fence(bo->base.resv, fence);
892
893         ret = dma_resv_reserve_shared(bo->base.resv, 1);
894         if (unlikely(ret)) {
895                 dma_fence_put(fence);
896                 return ret;
897         }
898
899         dma_fence_put(bo->moving);
900         bo->moving = fence;
901         return 0;
902 }
903
904 /**
905  * Repeatedly evict memory from the LRU for @mem_type until we create enough
906  * space, or we've evicted everything and there isn't enough space.
907  */
908 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
909                                   const struct ttm_place *place,
910                                   struct ttm_mem_reg *mem,
911                                   struct ttm_operation_ctx *ctx)
912 {
913         struct ttm_bo_device *bdev = bo->bdev;
914         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
915         struct ww_acquire_ctx *ticket;
916         int ret;
917
918         ticket = dma_resv_locking_ctx(bo->base.resv);
919         do {
920                 ret = (*man->func->get_node)(man, bo, place, mem);
921                 if (unlikely(ret != 0))
922                         return ret;
923                 if (mem->mm_node)
924                         break;
925                 ret = ttm_mem_evict_first(bdev, mem->mem_type, place, ctx,
926                                           ticket);
927                 if (unlikely(ret != 0))
928                         return ret;
929         } while (1);
930
931         return ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
932 }
933
934 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
935                                       uint32_t cur_placement,
936                                       uint32_t proposed_placement)
937 {
938         uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
939         uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
940
941         /**
942          * Keep current caching if possible.
943          */
944
945         if ((cur_placement & caching) != 0)
946                 result |= (cur_placement & caching);
947         else if ((man->default_caching & caching) != 0)
948                 result |= man->default_caching;
949         else if ((TTM_PL_FLAG_CACHED & caching) != 0)
950                 result |= TTM_PL_FLAG_CACHED;
951         else if ((TTM_PL_FLAG_WC & caching) != 0)
952                 result |= TTM_PL_FLAG_WC;
953         else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
954                 result |= TTM_PL_FLAG_UNCACHED;
955
956         return result;
957 }
958
959 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
960                                  uint32_t mem_type,
961                                  const struct ttm_place *place,
962                                  uint32_t *masked_placement)
963 {
964         uint32_t cur_flags = ttm_bo_type_flags(mem_type);
965
966         if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
967                 return false;
968
969         if ((place->flags & man->available_caching) == 0)
970                 return false;
971
972         cur_flags |= (place->flags & man->available_caching);
973
974         *masked_placement = cur_flags;
975         return true;
976 }
977
978 /**
979  * ttm_bo_mem_placement - check if placement is compatible
980  * @bo: BO to find memory for
981  * @place: where to search
982  * @mem: the memory object to fill in
983  * @ctx: operation context
984  *
985  * Check if placement is compatible and fill in mem structure.
986  * Returns -EBUSY if placement won't work or negative error code.
987  * 0 when placement can be used.
988  */
989 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo,
990                                 const struct ttm_place *place,
991                                 struct ttm_mem_reg *mem,
992                                 struct ttm_operation_ctx *ctx)
993 {
994         struct ttm_bo_device *bdev = bo->bdev;
995         uint32_t mem_type = TTM_PL_SYSTEM;
996         struct ttm_mem_type_manager *man;
997         uint32_t cur_flags = 0;
998         int ret;
999
1000         ret = ttm_mem_type_from_place(place, &mem_type);
1001         if (ret)
1002                 return ret;
1003
1004         man = &bdev->man[mem_type];
1005         if (!man->has_type || !man->use_type)
1006                 return -EBUSY;
1007
1008         if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
1009                 return -EBUSY;
1010
1011         cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags);
1012         /*
1013          * Use the access and other non-mapping-related flag bits from
1014          * the memory placement flags to the current flags
1015          */
1016         ttm_flag_masked(&cur_flags, place->flags, ~TTM_PL_MASK_MEMTYPE);
1017
1018         mem->mem_type = mem_type;
1019         mem->placement = cur_flags;
1020
1021         spin_lock(&ttm_bo_glob.lru_lock);
1022         ttm_bo_del_from_lru(bo);
1023         ttm_bo_add_mem_to_lru(bo, mem);
1024         spin_unlock(&ttm_bo_glob.lru_lock);
1025
1026         return 0;
1027 }
1028
1029 /**
1030  * Creates space for memory region @mem according to its type.
1031  *
1032  * This function first searches for free space in compatible memory types in
1033  * the priority order defined by the driver.  If free space isn't found, then
1034  * ttm_bo_mem_force_space is attempted in priority order to evict and find
1035  * space.
1036  */
1037 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
1038                         struct ttm_placement *placement,
1039                         struct ttm_mem_reg *mem,
1040                         struct ttm_operation_ctx *ctx)
1041 {
1042         struct ttm_bo_device *bdev = bo->bdev;
1043         bool type_found = false;
1044         int i, ret;
1045
1046         ret = dma_resv_reserve_shared(bo->base.resv, 1);
1047         if (unlikely(ret))
1048                 return ret;
1049
1050         mem->mm_node = NULL;
1051         for (i = 0; i < placement->num_placement; ++i) {
1052                 const struct ttm_place *place = &placement->placement[i];
1053                 struct ttm_mem_type_manager *man;
1054
1055                 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
1056                 if (ret == -EBUSY)
1057                         continue;
1058                 if (ret)
1059                         goto error;
1060
1061                 type_found = true;
1062                 mem->mm_node = NULL;
1063                 if (mem->mem_type == TTM_PL_SYSTEM)
1064                         return 0;
1065
1066                 man = &bdev->man[mem->mem_type];
1067                 ret = (*man->func->get_node)(man, bo, place, mem);
1068                 if (unlikely(ret))
1069                         goto error;
1070
1071                 if (!mem->mm_node)
1072                         continue;
1073
1074                 ret = ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
1075                 if (unlikely(ret)) {
1076                         (*man->func->put_node)(man, mem);
1077                         if (ret == -EBUSY)
1078                                 continue;
1079
1080                         goto error;
1081                 }
1082                 return 0;
1083         }
1084
1085         for (i = 0; i < placement->num_busy_placement; ++i) {
1086                 const struct ttm_place *place = &placement->busy_placement[i];
1087
1088                 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
1089                 if (ret == -EBUSY)
1090                         continue;
1091                 if (ret)
1092                         goto error;
1093
1094                 type_found = true;
1095                 mem->mm_node = NULL;
1096                 if (mem->mem_type == TTM_PL_SYSTEM)
1097                         return 0;
1098
1099                 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
1100                 if (ret == 0 && mem->mm_node)
1101                         return 0;
1102
1103                 if (ret && ret != -EBUSY)
1104                         goto error;
1105         }
1106
1107         ret = -ENOMEM;
1108         if (!type_found) {
1109                 pr_err(TTM_PFX "No compatible memory type found\n");
1110                 ret = -EINVAL;
1111         }
1112
1113 error:
1114         if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) {
1115                 spin_lock(&ttm_bo_glob.lru_lock);
1116                 ttm_bo_move_to_lru_tail(bo, NULL);
1117                 spin_unlock(&ttm_bo_glob.lru_lock);
1118         }
1119
1120         return ret;
1121 }
1122 EXPORT_SYMBOL(ttm_bo_mem_space);
1123
1124 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1125                               struct ttm_placement *placement,
1126                               struct ttm_operation_ctx *ctx)
1127 {
1128         int ret = 0;
1129         struct ttm_mem_reg mem;
1130
1131         dma_resv_assert_held(bo->base.resv);
1132
1133         mem.num_pages = bo->num_pages;
1134         mem.size = mem.num_pages << PAGE_SHIFT;
1135         mem.page_alignment = bo->mem.page_alignment;
1136         mem.bus.io_reserved_vm = false;
1137         mem.bus.io_reserved_count = 0;
1138         /*
1139          * Determine where to move the buffer.
1140          */
1141         ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1142         if (ret)
1143                 goto out_unlock;
1144         ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1145 out_unlock:
1146         if (ret && mem.mm_node)
1147                 ttm_bo_mem_put(bo, &mem);
1148         return ret;
1149 }
1150
1151 static bool ttm_bo_places_compat(const struct ttm_place *places,
1152                                  unsigned num_placement,
1153                                  struct ttm_mem_reg *mem,
1154                                  uint32_t *new_flags)
1155 {
1156         unsigned i;
1157
1158         for (i = 0; i < num_placement; i++) {
1159                 const struct ttm_place *heap = &places[i];
1160
1161                 if (mem->mm_node && (mem->start < heap->fpfn ||
1162                      (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1163                         continue;
1164
1165                 *new_flags = heap->flags;
1166                 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1167                     (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
1168                     (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1169                      (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1170                         return true;
1171         }
1172         return false;
1173 }
1174
1175 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1176                        struct ttm_mem_reg *mem,
1177                        uint32_t *new_flags)
1178 {
1179         if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1180                                  mem, new_flags))
1181                 return true;
1182
1183         if ((placement->busy_placement != placement->placement ||
1184              placement->num_busy_placement > placement->num_placement) &&
1185             ttm_bo_places_compat(placement->busy_placement,
1186                                  placement->num_busy_placement,
1187                                  mem, new_flags))
1188                 return true;
1189
1190         return false;
1191 }
1192 EXPORT_SYMBOL(ttm_bo_mem_compat);
1193
1194 int ttm_bo_validate(struct ttm_buffer_object *bo,
1195                     struct ttm_placement *placement,
1196                     struct ttm_operation_ctx *ctx)
1197 {
1198         int ret;
1199         uint32_t new_flags;
1200
1201         dma_resv_assert_held(bo->base.resv);
1202
1203         /*
1204          * Remove the backing store if no placement is given.
1205          */
1206         if (!placement->num_placement && !placement->num_busy_placement) {
1207                 ret = ttm_bo_pipeline_gutting(bo);
1208                 if (ret)
1209                         return ret;
1210
1211                 return ttm_tt_create(bo, false);
1212         }
1213
1214         /*
1215          * Check whether we need to move buffer.
1216          */
1217         if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1218                 ret = ttm_bo_move_buffer(bo, placement, ctx);
1219                 if (ret)
1220                         return ret;
1221         } else {
1222                 /*
1223                  * Use the access and other non-mapping-related flag bits from
1224                  * the compatible memory placement flags to the active flags
1225                  */
1226                 ttm_flag_masked(&bo->mem.placement, new_flags,
1227                                 ~TTM_PL_MASK_MEMTYPE);
1228         }
1229         /*
1230          * We might need to add a TTM.
1231          */
1232         if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1233                 ret = ttm_tt_create(bo, true);
1234                 if (ret)
1235                         return ret;
1236         }
1237         return 0;
1238 }
1239 EXPORT_SYMBOL(ttm_bo_validate);
1240
1241 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1242                          struct ttm_buffer_object *bo,
1243                          unsigned long size,
1244                          enum ttm_bo_type type,
1245                          struct ttm_placement *placement,
1246                          uint32_t page_alignment,
1247                          struct ttm_operation_ctx *ctx,
1248                          size_t acc_size,
1249                          struct sg_table *sg,
1250                          struct dma_resv *resv,
1251                          void (*destroy) (struct ttm_buffer_object *))
1252 {
1253         struct ttm_mem_global *mem_glob = &ttm_mem_glob;
1254         int ret = 0;
1255         unsigned long num_pages;
1256         bool locked;
1257
1258         ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1259         if (ret) {
1260                 pr_err("Out of kernel memory\n");
1261                 if (destroy)
1262                         (*destroy)(bo);
1263                 else
1264                         kfree(bo);
1265                 return -ENOMEM;
1266         }
1267
1268         num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1269         if (num_pages == 0) {
1270                 pr_err("Illegal buffer object size\n");
1271                 if (destroy)
1272                         (*destroy)(bo);
1273                 else
1274                         kfree(bo);
1275                 ttm_mem_global_free(mem_glob, acc_size);
1276                 return -EINVAL;
1277         }
1278         bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1279
1280         kref_init(&bo->kref);
1281         INIT_LIST_HEAD(&bo->lru);
1282         INIT_LIST_HEAD(&bo->ddestroy);
1283         INIT_LIST_HEAD(&bo->swap);
1284         INIT_LIST_HEAD(&bo->io_reserve_lru);
1285         bo->bdev = bdev;
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->moving = NULL;
1296         bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1297         bo->acc_size = acc_size;
1298         bo->sg = sg;
1299         if (resv) {
1300                 bo->base.resv = resv;
1301                 dma_resv_assert_held(bo->base.resv);
1302         } else {
1303                 bo->base.resv = &bo->base._resv;
1304         }
1305         if (!ttm_bo_uses_embedded_gem_object(bo)) {
1306                 /*
1307                  * bo.gem is not initialized, so we have to setup the
1308                  * struct elements we want use regardless.
1309                  */
1310                 dma_resv_init(&bo->base._resv);
1311                 drm_vma_node_reset(&bo->base.vma_node);
1312         }
1313         atomic_inc(&ttm_bo_glob.bo_count);
1314
1315         /*
1316          * For ttm_bo_type_device buffers, allocate
1317          * address space from the device.
1318          */
1319         if (bo->type == ttm_bo_type_device ||
1320             bo->type == ttm_bo_type_sg)
1321                 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1322                                          bo->mem.num_pages);
1323
1324         /* passed reservation objects should already be locked,
1325          * since otherwise lockdep will be angered in radeon.
1326          */
1327         if (!resv) {
1328                 locked = dma_resv_trylock(bo->base.resv);
1329                 WARN_ON(!locked);
1330         }
1331
1332         if (likely(!ret))
1333                 ret = ttm_bo_validate(bo, placement, ctx);
1334
1335         if (unlikely(ret)) {
1336                 if (!resv)
1337                         ttm_bo_unreserve(bo);
1338
1339                 ttm_bo_put(bo);
1340                 return ret;
1341         }
1342
1343         spin_lock(&ttm_bo_glob.lru_lock);
1344         ttm_bo_move_to_lru_tail(bo, NULL);
1345         spin_unlock(&ttm_bo_glob.lru_lock);
1346
1347         return ret;
1348 }
1349 EXPORT_SYMBOL(ttm_bo_init_reserved);
1350
1351 int ttm_bo_init(struct ttm_bo_device *bdev,
1352                 struct ttm_buffer_object *bo,
1353                 unsigned long size,
1354                 enum ttm_bo_type type,
1355                 struct ttm_placement *placement,
1356                 uint32_t page_alignment,
1357                 bool interruptible,
1358                 size_t acc_size,
1359                 struct sg_table *sg,
1360                 struct dma_resv *resv,
1361                 void (*destroy) (struct ttm_buffer_object *))
1362 {
1363         struct ttm_operation_ctx ctx = { interruptible, false };
1364         int ret;
1365
1366         ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1367                                    page_alignment, &ctx, acc_size,
1368                                    sg, resv, destroy);
1369         if (ret)
1370                 return ret;
1371
1372         if (!resv)
1373                 ttm_bo_unreserve(bo);
1374
1375         return 0;
1376 }
1377 EXPORT_SYMBOL(ttm_bo_init);
1378
1379 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1380                        unsigned long bo_size,
1381                        unsigned struct_size)
1382 {
1383         unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1384         size_t size = 0;
1385
1386         size += ttm_round_pot(struct_size);
1387         size += ttm_round_pot(npages * sizeof(void *));
1388         size += ttm_round_pot(sizeof(struct ttm_tt));
1389         return size;
1390 }
1391 EXPORT_SYMBOL(ttm_bo_acc_size);
1392
1393 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1394                            unsigned long bo_size,
1395                            unsigned struct_size)
1396 {
1397         unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1398         size_t size = 0;
1399
1400         size += ttm_round_pot(struct_size);
1401         size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1402         size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1403         return size;
1404 }
1405 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1406
1407 int ttm_bo_create(struct ttm_bo_device *bdev,
1408                         unsigned long size,
1409                         enum ttm_bo_type type,
1410                         struct ttm_placement *placement,
1411                         uint32_t page_alignment,
1412                         bool interruptible,
1413                         struct ttm_buffer_object **p_bo)
1414 {
1415         struct ttm_buffer_object *bo;
1416         size_t acc_size;
1417         int ret;
1418
1419         bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1420         if (unlikely(bo == NULL))
1421                 return -ENOMEM;
1422
1423         acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1424         ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1425                           interruptible, acc_size,
1426                           NULL, NULL, NULL);
1427         if (likely(ret == 0))
1428                 *p_bo = bo;
1429
1430         return ret;
1431 }
1432 EXPORT_SYMBOL(ttm_bo_create);
1433
1434 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1435                                    unsigned mem_type)
1436 {
1437         struct ttm_operation_ctx ctx = {
1438                 .interruptible = false,
1439                 .no_wait_gpu = false,
1440                 .flags = TTM_OPT_FLAG_FORCE_ALLOC
1441         };
1442         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1443         struct ttm_bo_global *glob = &ttm_bo_glob;
1444         struct dma_fence *fence;
1445         int ret;
1446         unsigned i;
1447
1448         /*
1449          * Can't use standard list traversal since we're unlocking.
1450          */
1451
1452         spin_lock(&glob->lru_lock);
1453         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1454                 while (!list_empty(&man->lru[i])) {
1455                         spin_unlock(&glob->lru_lock);
1456                         ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx,
1457                                                   NULL);
1458                         if (ret)
1459                                 return ret;
1460                         spin_lock(&glob->lru_lock);
1461                 }
1462         }
1463         spin_unlock(&glob->lru_lock);
1464
1465         spin_lock(&man->move_lock);
1466         fence = dma_fence_get(man->move);
1467         spin_unlock(&man->move_lock);
1468
1469         if (fence) {
1470                 ret = dma_fence_wait(fence, false);
1471                 dma_fence_put(fence);
1472                 if (ret)
1473                         return ret;
1474         }
1475
1476         return 0;
1477 }
1478
1479 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1480 {
1481         struct ttm_mem_type_manager *man;
1482         int ret = -EINVAL;
1483
1484         if (mem_type >= TTM_NUM_MEM_TYPES) {
1485                 pr_err("Illegal memory type %d\n", mem_type);
1486                 return ret;
1487         }
1488         man = &bdev->man[mem_type];
1489
1490         if (!man->has_type) {
1491                 pr_err("Trying to take down uninitialized memory manager type %u\n",
1492                        mem_type);
1493                 return ret;
1494         }
1495
1496         man->use_type = false;
1497         man->has_type = false;
1498
1499         ret = 0;
1500         if (mem_type > 0) {
1501                 ret = ttm_bo_force_list_clean(bdev, mem_type);
1502                 if (ret) {
1503                         pr_err("Cleanup eviction failed\n");
1504                         return ret;
1505                 }
1506
1507                 ret = (*man->func->takedown)(man);
1508         }
1509
1510         dma_fence_put(man->move);
1511         man->move = NULL;
1512
1513         return ret;
1514 }
1515 EXPORT_SYMBOL(ttm_bo_clean_mm);
1516
1517 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1518 {
1519         struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1520
1521         if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1522                 pr_err("Illegal memory manager memory type %u\n", mem_type);
1523                 return -EINVAL;
1524         }
1525
1526         if (!man->has_type) {
1527                 pr_err("Memory type %u has not been initialized\n", mem_type);
1528                 return 0;
1529         }
1530
1531         return ttm_bo_force_list_clean(bdev, mem_type);
1532 }
1533 EXPORT_SYMBOL(ttm_bo_evict_mm);
1534
1535 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1536                         unsigned long p_size)
1537 {
1538         int ret;
1539         struct ttm_mem_type_manager *man;
1540         unsigned i;
1541
1542         BUG_ON(type >= TTM_NUM_MEM_TYPES);
1543         man = &bdev->man[type];
1544         BUG_ON(man->has_type);
1545         man->io_reserve_fastpath = true;
1546         man->use_io_reserve_lru = false;
1547         mutex_init(&man->io_reserve_mutex);
1548         spin_lock_init(&man->move_lock);
1549         INIT_LIST_HEAD(&man->io_reserve_lru);
1550
1551         ret = bdev->driver->init_mem_type(bdev, type, man);
1552         if (ret)
1553                 return ret;
1554         man->bdev = bdev;
1555
1556         if (type != TTM_PL_SYSTEM) {
1557                 ret = (*man->func->init)(man, p_size);
1558                 if (ret)
1559                         return ret;
1560         }
1561         man->has_type = true;
1562         man->use_type = true;
1563         man->size = p_size;
1564
1565         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1566                 INIT_LIST_HEAD(&man->lru[i]);
1567         man->move = NULL;
1568
1569         return 0;
1570 }
1571 EXPORT_SYMBOL(ttm_bo_init_mm);
1572
1573 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1574 {
1575         struct ttm_bo_global *glob =
1576                 container_of(kobj, struct ttm_bo_global, kobj);
1577
1578         __free_page(glob->dummy_read_page);
1579 }
1580
1581 static void ttm_bo_global_release(void)
1582 {
1583         struct ttm_bo_global *glob = &ttm_bo_glob;
1584
1585         mutex_lock(&ttm_global_mutex);
1586         if (--ttm_bo_glob_use_count > 0)
1587                 goto out;
1588
1589         kobject_del(&glob->kobj);
1590         kobject_put(&glob->kobj);
1591         ttm_mem_global_release(&ttm_mem_glob);
1592         memset(glob, 0, sizeof(*glob));
1593 out:
1594         mutex_unlock(&ttm_global_mutex);
1595 }
1596
1597 static int ttm_bo_global_init(void)
1598 {
1599         struct ttm_bo_global *glob = &ttm_bo_glob;
1600         int ret = 0;
1601         unsigned i;
1602
1603         mutex_lock(&ttm_global_mutex);
1604         if (++ttm_bo_glob_use_count > 1)
1605                 goto out;
1606
1607         ret = ttm_mem_global_init(&ttm_mem_glob);
1608         if (ret)
1609                 goto out;
1610
1611         spin_lock_init(&glob->lru_lock);
1612         glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1613
1614         if (unlikely(glob->dummy_read_page == NULL)) {
1615                 ret = -ENOMEM;
1616                 goto out;
1617         }
1618
1619         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1620                 INIT_LIST_HEAD(&glob->swap_lru[i]);
1621         INIT_LIST_HEAD(&glob->device_list);
1622         atomic_set(&glob->bo_count, 0);
1623
1624         ret = kobject_init_and_add(
1625                 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1626         if (unlikely(ret != 0))
1627                 kobject_put(&glob->kobj);
1628 out:
1629         mutex_unlock(&ttm_global_mutex);
1630         return ret;
1631 }
1632
1633 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1634 {
1635         struct ttm_bo_global *glob = &ttm_bo_glob;
1636         int ret = 0;
1637         unsigned i = TTM_NUM_MEM_TYPES;
1638         struct ttm_mem_type_manager *man;
1639
1640         while (i--) {
1641                 man = &bdev->man[i];
1642                 if (man->has_type) {
1643                         man->use_type = false;
1644                         if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1645                                 ret = -EBUSY;
1646                                 pr_err("DRM memory manager type %d is not clean\n",
1647                                        i);
1648                         }
1649                         man->has_type = false;
1650                 }
1651         }
1652
1653         mutex_lock(&ttm_global_mutex);
1654         list_del(&bdev->device_list);
1655         mutex_unlock(&ttm_global_mutex);
1656
1657         cancel_delayed_work_sync(&bdev->wq);
1658
1659         if (ttm_bo_delayed_delete(bdev, true))
1660                 pr_debug("Delayed destroy list was clean\n");
1661
1662         spin_lock(&glob->lru_lock);
1663         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1664                 if (list_empty(&bdev->man[0].lru[0]))
1665                         pr_debug("Swap list %d was clean\n", i);
1666         spin_unlock(&glob->lru_lock);
1667
1668         if (!ret)
1669                 ttm_bo_global_release();
1670
1671         return ret;
1672 }
1673 EXPORT_SYMBOL(ttm_bo_device_release);
1674
1675 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1676                        struct ttm_bo_driver *driver,
1677                        struct address_space *mapping,
1678                        struct drm_vma_offset_manager *vma_manager,
1679                        bool need_dma32)
1680 {
1681         struct ttm_bo_global *glob = &ttm_bo_glob;
1682         int ret;
1683
1684         if (WARN_ON(vma_manager == NULL))
1685                 return -EINVAL;
1686
1687         ret = ttm_bo_global_init();
1688         if (ret)
1689                 return ret;
1690
1691         bdev->driver = driver;
1692
1693         memset(bdev->man, 0, sizeof(bdev->man));
1694
1695         /*
1696          * Initialize the system memory buffer type.
1697          * Other types need to be driver / IOCTL initialized.
1698          */
1699         ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1700         if (unlikely(ret != 0))
1701                 goto out_no_sys;
1702
1703         bdev->vma_manager = vma_manager;
1704         INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1705         INIT_LIST_HEAD(&bdev->ddestroy);
1706         bdev->dev_mapping = mapping;
1707         bdev->need_dma32 = need_dma32;
1708         mutex_lock(&ttm_global_mutex);
1709         list_add_tail(&bdev->device_list, &glob->device_list);
1710         mutex_unlock(&ttm_global_mutex);
1711
1712         return 0;
1713 out_no_sys:
1714         ttm_bo_global_release();
1715         return ret;
1716 }
1717 EXPORT_SYMBOL(ttm_bo_device_init);
1718
1719 /*
1720  * buffer object vm functions.
1721  */
1722
1723 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1724 {
1725         struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1726
1727         if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1728                 if (mem->mem_type == TTM_PL_SYSTEM)
1729                         return false;
1730
1731                 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1732                         return false;
1733
1734                 if (mem->placement & TTM_PL_FLAG_CACHED)
1735                         return false;
1736         }
1737         return true;
1738 }
1739
1740 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1741 {
1742         struct ttm_bo_device *bdev = bo->bdev;
1743
1744         drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1745         ttm_mem_io_free_vm(bo);
1746 }
1747
1748 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1749 {
1750         struct ttm_bo_device *bdev = bo->bdev;
1751         struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1752
1753         ttm_mem_io_lock(man, false);
1754         ttm_bo_unmap_virtual_locked(bo);
1755         ttm_mem_io_unlock(man);
1756 }
1757
1758
1759 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1760
1761 int ttm_bo_wait(struct ttm_buffer_object *bo,
1762                 bool interruptible, bool no_wait)
1763 {
1764         long timeout = 15 * HZ;
1765
1766         if (no_wait) {
1767                 if (dma_resv_test_signaled_rcu(bo->base.resv, true))
1768                         return 0;
1769                 else
1770                         return -EBUSY;
1771         }
1772
1773         timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true,
1774                                                       interruptible, timeout);
1775         if (timeout < 0)
1776                 return timeout;
1777
1778         if (timeout == 0)
1779                 return -EBUSY;
1780
1781         dma_resv_add_excl_fence(bo->base.resv, NULL);
1782         return 0;
1783 }
1784 EXPORT_SYMBOL(ttm_bo_wait);
1785
1786 /**
1787  * A buffer object shrink method that tries to swap out the first
1788  * buffer object on the bo_global::swap_lru list.
1789  */
1790 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1791 {
1792         struct ttm_buffer_object *bo;
1793         int ret = -EBUSY;
1794         bool locked;
1795         unsigned i;
1796
1797         spin_lock(&glob->lru_lock);
1798         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1799                 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1800                         if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
1801                                                             NULL))
1802                                 continue;
1803
1804                         if (!ttm_bo_get_unless_zero(bo)) {
1805                                 if (locked)
1806                                         dma_resv_unlock(bo->base.resv);
1807                                 continue;
1808                         }
1809
1810                         ret = 0;
1811                         break;
1812                 }
1813                 if (!ret)
1814                         break;
1815         }
1816
1817         if (ret) {
1818                 spin_unlock(&glob->lru_lock);
1819                 return ret;
1820         }
1821
1822         if (bo->deleted) {
1823                 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1824                 ttm_bo_put(bo);
1825                 return ret;
1826         }
1827
1828         ttm_bo_del_from_lru(bo);
1829         spin_unlock(&glob->lru_lock);
1830
1831         /**
1832          * Move to system cached
1833          */
1834
1835         if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1836             bo->ttm->caching_state != tt_cached) {
1837                 struct ttm_operation_ctx ctx = { false, false };
1838                 struct ttm_mem_reg evict_mem;
1839
1840                 evict_mem = bo->mem;
1841                 evict_mem.mm_node = NULL;
1842                 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1843                 evict_mem.mem_type = TTM_PL_SYSTEM;
1844
1845                 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1846                 if (unlikely(ret != 0))
1847                         goto out;
1848         }
1849
1850         /**
1851          * Make sure BO is idle.
1852          */
1853
1854         ret = ttm_bo_wait(bo, false, false);
1855         if (unlikely(ret != 0))
1856                 goto out;
1857
1858         ttm_bo_unmap_virtual(bo);
1859
1860         /**
1861          * Swap out. Buffer will be swapped in again as soon as
1862          * anyone tries to access a ttm page.
1863          */
1864
1865         if (bo->bdev->driver->swap_notify)
1866                 bo->bdev->driver->swap_notify(bo);
1867
1868         ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1869 out:
1870
1871         /**
1872          *
1873          * Unreserve without putting on LRU to avoid swapping out an
1874          * already swapped buffer.
1875          */
1876         if (locked)
1877                 dma_resv_unlock(bo->base.resv);
1878         ttm_bo_put(bo);
1879         return ret;
1880 }
1881 EXPORT_SYMBOL(ttm_bo_swapout);
1882
1883 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1884 {
1885         struct ttm_operation_ctx ctx = {
1886                 .interruptible = false,
1887                 .no_wait_gpu = false
1888         };
1889
1890         while (ttm_bo_swapout(&ttm_bo_glob, &ctx) == 0);
1891 }
1892 EXPORT_SYMBOL(ttm_bo_swapout_all);