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 #include <linux/reservation.h>
44
45 #define TTM_ASSERT_LOCKED(param)
46 #define TTM_DEBUG(fmt, arg...)
47 #define TTM_BO_HASH_ORDER 13
48
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
ttm_mem_type_from_place(const struct ttm_place * place,uint32_t * mem_type)57 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
58 uint32_t *mem_type)
59 {
60 int pos;
61
62 pos = ffs(place->flags & TTM_PL_MASK_MEM);
63 if (unlikely(!pos))
64 return -EINVAL;
65
66 *mem_type = pos - 1;
67 return 0;
68 }
69
ttm_mem_type_debug(struct ttm_bo_device * bdev,int mem_type)70 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
71 {
72 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
73
74 pr_err(" has_type: %d\n", man->has_type);
75 pr_err(" use_type: %d\n", man->use_type);
76 pr_err(" flags: 0x%08X\n", man->flags);
77 pr_err(" gpu_offset: 0x%08llX\n", man->gpu_offset);
78 pr_err(" size: %llu\n", man->size);
79 pr_err(" available_caching: 0x%08X\n", man->available_caching);
80 pr_err(" default_caching: 0x%08X\n", man->default_caching);
81 if (mem_type != TTM_PL_SYSTEM)
82 (*man->func->debug)(man, TTM_PFX);
83 }
84
ttm_bo_mem_space_debug(struct ttm_buffer_object * bo,struct ttm_placement * placement)85 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
86 struct ttm_placement *placement)
87 {
88 int i, ret, mem_type;
89
90 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
91 bo, bo->mem.num_pages, bo->mem.size >> 10,
92 bo->mem.size >> 20);
93 for (i = 0; i < placement->num_placement; i++) {
94 ret = ttm_mem_type_from_place(&placement->placement[i],
95 &mem_type);
96 if (ret)
97 return;
98 pr_err(" placement[%d]=0x%08X (%d)\n",
99 i, placement->placement[i].flags, mem_type);
100 ttm_mem_type_debug(bo->bdev, mem_type);
101 }
102 }
103
ttm_bo_global_show(struct kobject * kobj,struct attribute * attr,char * buffer)104 static ssize_t ttm_bo_global_show(struct kobject *kobj,
105 struct attribute *attr,
106 char *buffer)
107 {
108 struct ttm_bo_global *glob =
109 container_of(kobj, struct ttm_bo_global, kobj);
110
111 return snprintf(buffer, PAGE_SIZE, "%lu\n",
112 (unsigned long) atomic_read(&glob->bo_count));
113 }
114
115 static struct attribute *ttm_bo_global_attrs[] = {
116 &ttm_bo_count,
117 NULL
118 };
119
120 static const struct sysfs_ops ttm_bo_global_ops = {
121 .show = &ttm_bo_global_show
122 };
123
124 static struct kobj_type ttm_bo_glob_kobj_type = {
125 .release = &ttm_bo_global_kobj_release,
126 .sysfs_ops = &ttm_bo_global_ops,
127 .default_attrs = ttm_bo_global_attrs
128 };
129
130
ttm_bo_type_flags(unsigned type)131 static inline uint32_t ttm_bo_type_flags(unsigned type)
132 {
133 return 1 << (type);
134 }
135
ttm_bo_release_list(struct kref * list_kref)136 static void ttm_bo_release_list(struct kref *list_kref)
137 {
138 struct ttm_buffer_object *bo =
139 container_of(list_kref, struct ttm_buffer_object, list_kref);
140 struct ttm_bo_device *bdev = bo->bdev;
141 size_t acc_size = bo->acc_size;
142
143 BUG_ON(atomic_read(&bo->list_kref.refcount));
144 BUG_ON(atomic_read(&bo->kref.refcount));
145 BUG_ON(atomic_read(&bo->cpu_writers));
146 BUG_ON(bo->mem.mm_node != NULL);
147 BUG_ON(!list_empty(&bo->lru));
148 BUG_ON(!list_empty(&bo->ddestroy));
149 ttm_tt_destroy(bo->ttm);
150 atomic_dec(&bo->glob->bo_count);
151 fence_put(bo->moving);
152 if (bo->resv == &bo->ttm_resv)
153 reservation_object_fini(&bo->ttm_resv);
154 mutex_destroy(&bo->wu_mutex);
155 if (bo->destroy)
156 bo->destroy(bo);
157 else {
158 kfree(bo);
159 }
160 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
161 }
162
ttm_bo_add_to_lru(struct ttm_buffer_object * bo)163 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
164 {
165 struct ttm_bo_device *bdev = bo->bdev;
166
167 lockdep_assert_held(&bo->resv->lock.base);
168
169 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
170
171 BUG_ON(!list_empty(&bo->lru));
172
173 list_add(&bo->lru, bdev->driver->lru_tail(bo));
174 kref_get(&bo->list_kref);
175
176 if (bo->ttm && !(bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
177 list_add(&bo->swap, bdev->driver->swap_lru_tail(bo));
178 kref_get(&bo->list_kref);
179 }
180 }
181 }
182 EXPORT_SYMBOL(ttm_bo_add_to_lru);
183
ttm_bo_del_from_lru(struct ttm_buffer_object * bo)184 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
185 {
186 struct ttm_bo_device *bdev = bo->bdev;
187 int put_count = 0;
188
189 if (bdev->driver->lru_removal)
190 bdev->driver->lru_removal(bo);
191
192 if (!list_empty(&bo->swap)) {
193 list_del_init(&bo->swap);
194 ++put_count;
195 }
196 if (!list_empty(&bo->lru)) {
197 list_del_init(&bo->lru);
198 ++put_count;
199 }
200
201 return put_count;
202 }
203
ttm_bo_ref_bug(struct kref * list_kref)204 static void ttm_bo_ref_bug(struct kref *list_kref)
205 {
206 BUG();
207 }
208
ttm_bo_list_ref_sub(struct ttm_buffer_object * bo,int count,bool never_free)209 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
210 bool never_free)
211 {
212 kref_sub(&bo->list_kref, count,
213 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
214 }
215
ttm_bo_del_sub_from_lru(struct ttm_buffer_object * bo)216 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
217 {
218 int put_count;
219
220 spin_lock(&bo->glob->lru_lock);
221 put_count = ttm_bo_del_from_lru(bo);
222 spin_unlock(&bo->glob->lru_lock);
223 ttm_bo_list_ref_sub(bo, put_count, true);
224 }
225 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
226
ttm_bo_move_to_lru_tail(struct ttm_buffer_object * bo)227 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
228 {
229 struct ttm_bo_device *bdev = bo->bdev;
230 int put_count = 0;
231
232 lockdep_assert_held(&bo->resv->lock.base);
233
234 if (bdev->driver->lru_removal)
235 bdev->driver->lru_removal(bo);
236
237 put_count = ttm_bo_del_from_lru(bo);
238 ttm_bo_list_ref_sub(bo, put_count, true);
239 ttm_bo_add_to_lru(bo);
240 }
241 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
242
ttm_bo_default_lru_tail(struct ttm_buffer_object * bo)243 struct list_head *ttm_bo_default_lru_tail(struct ttm_buffer_object *bo)
244 {
245 return bo->bdev->man[bo->mem.mem_type].lru.prev;
246 }
247 EXPORT_SYMBOL(ttm_bo_default_lru_tail);
248
ttm_bo_default_swap_lru_tail(struct ttm_buffer_object * bo)249 struct list_head *ttm_bo_default_swap_lru_tail(struct ttm_buffer_object *bo)
250 {
251 return bo->glob->swap_lru.prev;
252 }
253 EXPORT_SYMBOL(ttm_bo_default_swap_lru_tail);
254
255 /*
256 * Call bo->mutex locked.
257 */
ttm_bo_add_ttm(struct ttm_buffer_object * bo,bool zero_alloc)258 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
259 {
260 struct ttm_bo_device *bdev = bo->bdev;
261 struct ttm_bo_global *glob = bo->glob;
262 int ret = 0;
263 uint32_t page_flags = 0;
264
265 TTM_ASSERT_LOCKED(&bo->mutex);
266 bo->ttm = NULL;
267
268 if (bdev->need_dma32)
269 page_flags |= TTM_PAGE_FLAG_DMA32;
270
271 switch (bo->type) {
272 case ttm_bo_type_device:
273 if (zero_alloc)
274 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
275 case ttm_bo_type_kernel:
276 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
277 page_flags, glob->dummy_read_page);
278 if (unlikely(bo->ttm == NULL))
279 ret = -ENOMEM;
280 break;
281 case ttm_bo_type_sg:
282 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
283 page_flags | TTM_PAGE_FLAG_SG,
284 glob->dummy_read_page);
285 if (unlikely(bo->ttm == NULL)) {
286 ret = -ENOMEM;
287 break;
288 }
289 bo->ttm->sg = bo->sg;
290 break;
291 default:
292 pr_err("Illegal buffer object type\n");
293 ret = -EINVAL;
294 break;
295 }
296
297 return ret;
298 }
299
ttm_bo_handle_move_mem(struct ttm_buffer_object * bo,struct ttm_mem_reg * mem,bool evict,bool interruptible,bool no_wait_gpu)300 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
301 struct ttm_mem_reg *mem,
302 bool evict, bool interruptible,
303 bool no_wait_gpu)
304 {
305 struct ttm_bo_device *bdev = bo->bdev;
306 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
307 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
308 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
309 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
310 int ret = 0;
311
312 if (old_is_pci || new_is_pci ||
313 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
314 ret = ttm_mem_io_lock(old_man, true);
315 if (unlikely(ret != 0))
316 goto out_err;
317 ttm_bo_unmap_virtual_locked(bo);
318 ttm_mem_io_unlock(old_man);
319 }
320
321 /*
322 * Create and bind a ttm if required.
323 */
324
325 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
326 if (bo->ttm == NULL) {
327 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
328 ret = ttm_bo_add_ttm(bo, zero);
329 if (ret)
330 goto out_err;
331 }
332
333 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
334 if (ret)
335 goto out_err;
336
337 if (mem->mem_type != TTM_PL_SYSTEM) {
338 ret = ttm_tt_bind(bo->ttm, mem);
339 if (ret)
340 goto out_err;
341 }
342
343 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
344 if (bdev->driver->move_notify)
345 bdev->driver->move_notify(bo, mem);
346 bo->mem = *mem;
347 mem->mm_node = NULL;
348 goto moved;
349 }
350 }
351
352 if (bdev->driver->move_notify)
353 bdev->driver->move_notify(bo, mem);
354
355 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
356 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
357 ret = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, mem);
358 else if (bdev->driver->move)
359 ret = bdev->driver->move(bo, evict, interruptible,
360 no_wait_gpu, mem);
361 else
362 ret = ttm_bo_move_memcpy(bo, interruptible, no_wait_gpu, mem);
363
364 if (ret) {
365 if (bdev->driver->move_notify) {
366 struct ttm_mem_reg tmp_mem = *mem;
367 *mem = bo->mem;
368 bo->mem = tmp_mem;
369 bdev->driver->move_notify(bo, mem);
370 bo->mem = *mem;
371 *mem = tmp_mem;
372 }
373
374 goto out_err;
375 }
376
377 moved:
378 if (bo->evicted) {
379 if (bdev->driver->invalidate_caches) {
380 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
381 if (ret)
382 pr_err("Can not flush read caches\n");
383 }
384 bo->evicted = false;
385 }
386
387 if (bo->mem.mm_node) {
388 bo->offset = (bo->mem.start << PAGE_SHIFT) +
389 bdev->man[bo->mem.mem_type].gpu_offset;
390 bo->cur_placement = bo->mem.placement;
391 } else
392 bo->offset = 0;
393
394 return 0;
395
396 out_err:
397 new_man = &bdev->man[bo->mem.mem_type];
398 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
399 ttm_tt_destroy(bo->ttm);
400 bo->ttm = NULL;
401 }
402
403 return ret;
404 }
405
406 /**
407 * Call bo::reserved.
408 * Will release GPU memory type usage on destruction.
409 * This is the place to put in driver specific hooks to release
410 * driver private resources.
411 * Will release the bo::reserved lock.
412 */
413
ttm_bo_cleanup_memtype_use(struct ttm_buffer_object * bo)414 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
415 {
416 if (bo->bdev->driver->move_notify)
417 bo->bdev->driver->move_notify(bo, NULL);
418
419 ttm_tt_destroy(bo->ttm);
420 bo->ttm = NULL;
421 ttm_bo_mem_put(bo, &bo->mem);
422
423 ww_mutex_unlock (&bo->resv->lock);
424 }
425
ttm_bo_flush_all_fences(struct ttm_buffer_object * bo)426 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
427 {
428 struct reservation_object_list *fobj;
429 struct fence *fence;
430 int i;
431
432 fobj = reservation_object_get_list(bo->resv);
433 fence = reservation_object_get_excl(bo->resv);
434 if (fence && !fence->ops->signaled)
435 fence_enable_sw_signaling(fence);
436
437 for (i = 0; fobj && i < fobj->shared_count; ++i) {
438 fence = rcu_dereference_protected(fobj->shared[i],
439 reservation_object_held(bo->resv));
440
441 if (!fence->ops->signaled)
442 fence_enable_sw_signaling(fence);
443 }
444 }
445
ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object * bo)446 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
447 {
448 struct ttm_bo_device *bdev = bo->bdev;
449 struct ttm_bo_global *glob = bo->glob;
450 int put_count;
451 int ret;
452
453 spin_lock(&glob->lru_lock);
454 ret = __ttm_bo_reserve(bo, false, true, NULL);
455
456 if (!ret) {
457 if (!ttm_bo_wait(bo, false, true)) {
458 put_count = ttm_bo_del_from_lru(bo);
459
460 spin_unlock(&glob->lru_lock);
461 ttm_bo_cleanup_memtype_use(bo);
462
463 ttm_bo_list_ref_sub(bo, put_count, true);
464
465 return;
466 } else
467 ttm_bo_flush_all_fences(bo);
468
469 /*
470 * Make NO_EVICT bos immediately available to
471 * shrinkers, now that they are queued for
472 * destruction.
473 */
474 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
475 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
476 ttm_bo_add_to_lru(bo);
477 }
478
479 __ttm_bo_unreserve(bo);
480 }
481
482 kref_get(&bo->list_kref);
483 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
484 spin_unlock(&glob->lru_lock);
485
486 schedule_delayed_work(&bdev->wq,
487 ((HZ / 100) < 1) ? 1 : HZ / 100);
488 }
489
490 /**
491 * function ttm_bo_cleanup_refs_and_unlock
492 * If bo idle, remove from delayed- and lru lists, and unref.
493 * If not idle, do nothing.
494 *
495 * Must be called with lru_lock and reservation held, this function
496 * will drop both before returning.
497 *
498 * @interruptible Any sleeps should occur interruptibly.
499 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
500 */
501
ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object * bo,bool interruptible,bool no_wait_gpu)502 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
503 bool interruptible,
504 bool no_wait_gpu)
505 {
506 struct ttm_bo_global *glob = bo->glob;
507 int put_count;
508 int ret;
509
510 ret = ttm_bo_wait(bo, false, true);
511
512 if (ret && !no_wait_gpu) {
513 long lret;
514 ww_mutex_unlock(&bo->resv->lock);
515 spin_unlock(&glob->lru_lock);
516
517 lret = reservation_object_wait_timeout_rcu(bo->resv,
518 true,
519 interruptible,
520 30 * HZ);
521
522 if (lret < 0)
523 return lret;
524 else if (lret == 0)
525 return -EBUSY;
526
527 spin_lock(&glob->lru_lock);
528 ret = __ttm_bo_reserve(bo, false, true, NULL);
529
530 /*
531 * We raced, and lost, someone else holds the reservation now,
532 * and is probably busy in ttm_bo_cleanup_memtype_use.
533 *
534 * Even if it's not the case, because we finished waiting any
535 * delayed destruction would succeed, so just return success
536 * here.
537 */
538 if (ret) {
539 spin_unlock(&glob->lru_lock);
540 return 0;
541 }
542
543 /*
544 * remove sync_obj with ttm_bo_wait, the wait should be
545 * finished, and no new wait object should have been added.
546 */
547 ret = ttm_bo_wait(bo, false, true);
548 WARN_ON(ret);
549 }
550
551 if (ret || unlikely(list_empty(&bo->ddestroy))) {
552 __ttm_bo_unreserve(bo);
553 spin_unlock(&glob->lru_lock);
554 return ret;
555 }
556
557 put_count = ttm_bo_del_from_lru(bo);
558 list_del_init(&bo->ddestroy);
559 ++put_count;
560
561 spin_unlock(&glob->lru_lock);
562 ttm_bo_cleanup_memtype_use(bo);
563
564 ttm_bo_list_ref_sub(bo, put_count, true);
565
566 return 0;
567 }
568
569 /**
570 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
571 * encountered buffers.
572 */
573
ttm_bo_delayed_delete(struct ttm_bo_device * bdev,bool remove_all)574 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
575 {
576 struct ttm_bo_global *glob = bdev->glob;
577 struct ttm_buffer_object *entry = NULL;
578 int ret = 0;
579
580 spin_lock(&glob->lru_lock);
581 if (list_empty(&bdev->ddestroy))
582 goto out_unlock;
583
584 entry = list_first_entry(&bdev->ddestroy,
585 struct ttm_buffer_object, ddestroy);
586 kref_get(&entry->list_kref);
587
588 for (;;) {
589 struct ttm_buffer_object *nentry = NULL;
590
591 if (entry->ddestroy.next != &bdev->ddestroy) {
592 nentry = list_first_entry(&entry->ddestroy,
593 struct ttm_buffer_object, ddestroy);
594 kref_get(&nentry->list_kref);
595 }
596
597 ret = __ttm_bo_reserve(entry, false, true, NULL);
598 if (remove_all && ret) {
599 spin_unlock(&glob->lru_lock);
600 ret = __ttm_bo_reserve(entry, false, false, NULL);
601 spin_lock(&glob->lru_lock);
602 }
603
604 if (!ret)
605 ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
606 !remove_all);
607 else
608 spin_unlock(&glob->lru_lock);
609
610 kref_put(&entry->list_kref, ttm_bo_release_list);
611 entry = nentry;
612
613 if (ret || !entry)
614 goto out;
615
616 spin_lock(&glob->lru_lock);
617 if (list_empty(&entry->ddestroy))
618 break;
619 }
620
621 out_unlock:
622 spin_unlock(&glob->lru_lock);
623 out:
624 if (entry)
625 kref_put(&entry->list_kref, ttm_bo_release_list);
626 return ret;
627 }
628
ttm_bo_delayed_workqueue(struct work_struct * work)629 static void ttm_bo_delayed_workqueue(struct work_struct *work)
630 {
631 struct ttm_bo_device *bdev =
632 container_of(work, struct ttm_bo_device, wq.work);
633
634 if (ttm_bo_delayed_delete(bdev, false)) {
635 schedule_delayed_work(&bdev->wq,
636 ((HZ / 100) < 1) ? 1 : HZ / 100);
637 }
638 }
639
ttm_bo_release(struct kref * kref)640 static void ttm_bo_release(struct kref *kref)
641 {
642 struct ttm_buffer_object *bo =
643 container_of(kref, struct ttm_buffer_object, kref);
644 struct ttm_bo_device *bdev = bo->bdev;
645 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
646
647 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
648 ttm_mem_io_lock(man, false);
649 ttm_mem_io_free_vm(bo);
650 ttm_mem_io_unlock(man);
651 ttm_bo_cleanup_refs_or_queue(bo);
652 kref_put(&bo->list_kref, ttm_bo_release_list);
653 }
654
ttm_bo_unref(struct ttm_buffer_object ** p_bo)655 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
656 {
657 struct ttm_buffer_object *bo = *p_bo;
658
659 *p_bo = NULL;
660 kref_put(&bo->kref, ttm_bo_release);
661 }
662 EXPORT_SYMBOL(ttm_bo_unref);
663
ttm_bo_lock_delayed_workqueue(struct ttm_bo_device * bdev)664 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
665 {
666 return cancel_delayed_work_sync(&bdev->wq);
667 }
668 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
669
ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device * bdev,int resched)670 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
671 {
672 if (resched)
673 schedule_delayed_work(&bdev->wq,
674 ((HZ / 100) < 1) ? 1 : HZ / 100);
675 }
676 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
677
ttm_bo_evict(struct ttm_buffer_object * bo,bool interruptible,bool no_wait_gpu)678 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
679 bool no_wait_gpu)
680 {
681 struct ttm_bo_device *bdev = bo->bdev;
682 struct ttm_mem_reg evict_mem;
683 struct ttm_placement placement;
684 int ret = 0;
685
686 lockdep_assert_held(&bo->resv->lock.base);
687
688 evict_mem = bo->mem;
689 evict_mem.mm_node = NULL;
690 evict_mem.bus.io_reserved_vm = false;
691 evict_mem.bus.io_reserved_count = 0;
692
693 placement.num_placement = 0;
694 placement.num_busy_placement = 0;
695 bdev->driver->evict_flags(bo, &placement);
696 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
697 no_wait_gpu);
698 if (ret) {
699 if (ret != -ERESTARTSYS) {
700 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
701 bo);
702 ttm_bo_mem_space_debug(bo, &placement);
703 }
704 goto out;
705 }
706
707 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
708 no_wait_gpu);
709 if (unlikely(ret)) {
710 if (ret != -ERESTARTSYS)
711 pr_err("Buffer eviction failed\n");
712 ttm_bo_mem_put(bo, &evict_mem);
713 goto out;
714 }
715 bo->evicted = true;
716 out:
717 return ret;
718 }
719
ttm_mem_evict_first(struct ttm_bo_device * bdev,uint32_t mem_type,const struct ttm_place * place,bool interruptible,bool no_wait_gpu)720 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
721 uint32_t mem_type,
722 const struct ttm_place *place,
723 bool interruptible,
724 bool no_wait_gpu)
725 {
726 struct ttm_bo_global *glob = bdev->glob;
727 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
728 struct ttm_buffer_object *bo;
729 int ret = -EBUSY, put_count;
730
731 spin_lock(&glob->lru_lock);
732 list_for_each_entry(bo, &man->lru, lru) {
733 ret = __ttm_bo_reserve(bo, false, true, NULL);
734 if (!ret) {
735 if (place && (place->fpfn || place->lpfn)) {
736 /* Don't evict this BO if it's outside of the
737 * requested placement range
738 */
739 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
740 (place->lpfn && place->lpfn <= bo->mem.start)) {
741 __ttm_bo_unreserve(bo);
742 ret = -EBUSY;
743 continue;
744 }
745 }
746
747 break;
748 }
749 }
750
751 if (ret) {
752 spin_unlock(&glob->lru_lock);
753 return ret;
754 }
755
756 kref_get(&bo->list_kref);
757
758 if (!list_empty(&bo->ddestroy)) {
759 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
760 no_wait_gpu);
761 kref_put(&bo->list_kref, ttm_bo_release_list);
762 return ret;
763 }
764
765 put_count = ttm_bo_del_from_lru(bo);
766 spin_unlock(&glob->lru_lock);
767
768 BUG_ON(ret != 0);
769
770 ttm_bo_list_ref_sub(bo, put_count, true);
771
772 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
773 ttm_bo_unreserve(bo);
774
775 kref_put(&bo->list_kref, ttm_bo_release_list);
776 return ret;
777 }
778
ttm_bo_mem_put(struct ttm_buffer_object * bo,struct ttm_mem_reg * mem)779 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
780 {
781 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
782
783 if (mem->mm_node)
784 (*man->func->put_node)(man, mem);
785 }
786 EXPORT_SYMBOL(ttm_bo_mem_put);
787
788 /**
789 * Add the last move fence to the BO and reserve a new shared slot.
790 */
ttm_bo_add_move_fence(struct ttm_buffer_object * bo,struct ttm_mem_type_manager * man,struct ttm_mem_reg * mem)791 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
792 struct ttm_mem_type_manager *man,
793 struct ttm_mem_reg *mem)
794 {
795 struct fence *fence;
796 int ret;
797
798 spin_lock(&man->move_lock);
799 fence = fence_get(man->move);
800 spin_unlock(&man->move_lock);
801
802 if (fence) {
803 reservation_object_add_shared_fence(bo->resv, fence);
804
805 ret = reservation_object_reserve_shared(bo->resv);
806 if (unlikely(ret))
807 return ret;
808
809 fence_put(bo->moving);
810 bo->moving = fence;
811 }
812
813 return 0;
814 }
815
816 /**
817 * Repeatedly evict memory from the LRU for @mem_type until we create enough
818 * space, or we've evicted everything and there isn't enough space.
819 */
ttm_bo_mem_force_space(struct ttm_buffer_object * bo,uint32_t mem_type,const struct ttm_place * place,struct ttm_mem_reg * mem,bool interruptible,bool no_wait_gpu)820 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
821 uint32_t mem_type,
822 const struct ttm_place *place,
823 struct ttm_mem_reg *mem,
824 bool interruptible,
825 bool no_wait_gpu)
826 {
827 struct ttm_bo_device *bdev = bo->bdev;
828 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
829 int ret;
830
831 do {
832 ret = (*man->func->get_node)(man, bo, place, mem);
833 if (unlikely(ret != 0))
834 return ret;
835 if (mem->mm_node)
836 break;
837 ret = ttm_mem_evict_first(bdev, mem_type, place,
838 interruptible, no_wait_gpu);
839 if (unlikely(ret != 0))
840 return ret;
841 } while (1);
842 mem->mem_type = mem_type;
843 return ttm_bo_add_move_fence(bo, man, mem);
844 }
845
ttm_bo_select_caching(struct ttm_mem_type_manager * man,uint32_t cur_placement,uint32_t proposed_placement)846 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
847 uint32_t cur_placement,
848 uint32_t proposed_placement)
849 {
850 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
851 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
852
853 /**
854 * Keep current caching if possible.
855 */
856
857 if ((cur_placement & caching) != 0)
858 result |= (cur_placement & caching);
859 else if ((man->default_caching & caching) != 0)
860 result |= man->default_caching;
861 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
862 result |= TTM_PL_FLAG_CACHED;
863 else if ((TTM_PL_FLAG_WC & caching) != 0)
864 result |= TTM_PL_FLAG_WC;
865 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
866 result |= TTM_PL_FLAG_UNCACHED;
867
868 return result;
869 }
870
ttm_bo_mt_compatible(struct ttm_mem_type_manager * man,uint32_t mem_type,const struct ttm_place * place,uint32_t * masked_placement)871 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
872 uint32_t mem_type,
873 const struct ttm_place *place,
874 uint32_t *masked_placement)
875 {
876 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
877
878 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
879 return false;
880
881 if ((place->flags & man->available_caching) == 0)
882 return false;
883
884 cur_flags |= (place->flags & man->available_caching);
885
886 *masked_placement = cur_flags;
887 return true;
888 }
889
890 /**
891 * Creates space for memory region @mem according to its type.
892 *
893 * This function first searches for free space in compatible memory types in
894 * the priority order defined by the driver. If free space isn't found, then
895 * ttm_bo_mem_force_space is attempted in priority order to evict and find
896 * space.
897 */
ttm_bo_mem_space(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_mem_reg * mem,bool interruptible,bool no_wait_gpu)898 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
899 struct ttm_placement *placement,
900 struct ttm_mem_reg *mem,
901 bool interruptible,
902 bool no_wait_gpu)
903 {
904 struct ttm_bo_device *bdev = bo->bdev;
905 struct ttm_mem_type_manager *man;
906 uint32_t mem_type = TTM_PL_SYSTEM;
907 uint32_t cur_flags = 0;
908 bool type_found = false;
909 bool type_ok = false;
910 bool has_erestartsys = false;
911 int i, ret;
912
913 ret = reservation_object_reserve_shared(bo->resv);
914 if (unlikely(ret))
915 return ret;
916
917 mem->mm_node = NULL;
918 for (i = 0; i < placement->num_placement; ++i) {
919 const struct ttm_place *place = &placement->placement[i];
920
921 ret = ttm_mem_type_from_place(place, &mem_type);
922 if (ret)
923 return ret;
924 man = &bdev->man[mem_type];
925 if (!man->has_type || !man->use_type)
926 continue;
927
928 type_ok = ttm_bo_mt_compatible(man, mem_type, place,
929 &cur_flags);
930
931 if (!type_ok)
932 continue;
933
934 type_found = true;
935 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
936 cur_flags);
937 /*
938 * Use the access and other non-mapping-related flag bits from
939 * the memory placement flags to the current flags
940 */
941 ttm_flag_masked(&cur_flags, place->flags,
942 ~TTM_PL_MASK_MEMTYPE);
943
944 if (mem_type == TTM_PL_SYSTEM)
945 break;
946
947 ret = (*man->func->get_node)(man, bo, place, mem);
948 if (unlikely(ret))
949 return ret;
950
951 if (mem->mm_node) {
952 ret = ttm_bo_add_move_fence(bo, man, mem);
953 if (unlikely(ret)) {
954 (*man->func->put_node)(man, mem);
955 return ret;
956 }
957 break;
958 }
959 }
960
961 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
962 mem->mem_type = mem_type;
963 mem->placement = cur_flags;
964 return 0;
965 }
966
967 for (i = 0; i < placement->num_busy_placement; ++i) {
968 const struct ttm_place *place = &placement->busy_placement[i];
969
970 ret = ttm_mem_type_from_place(place, &mem_type);
971 if (ret)
972 return ret;
973 man = &bdev->man[mem_type];
974 if (!man->has_type || !man->use_type)
975 continue;
976 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
977 continue;
978
979 type_found = true;
980 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
981 cur_flags);
982 /*
983 * Use the access and other non-mapping-related flag bits from
984 * the memory placement flags to the current flags
985 */
986 ttm_flag_masked(&cur_flags, place->flags,
987 ~TTM_PL_MASK_MEMTYPE);
988
989 if (mem_type == TTM_PL_SYSTEM) {
990 mem->mem_type = mem_type;
991 mem->placement = cur_flags;
992 mem->mm_node = NULL;
993 return 0;
994 }
995
996 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
997 interruptible, no_wait_gpu);
998 if (ret == 0 && mem->mm_node) {
999 mem->placement = cur_flags;
1000 return 0;
1001 }
1002 if (ret == -ERESTARTSYS)
1003 has_erestartsys = true;
1004 }
1005
1006 if (!type_found) {
1007 printk(KERN_ERR TTM_PFX "No compatible memory type found.\n");
1008 return -EINVAL;
1009 }
1010
1011 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1012 }
1013 EXPORT_SYMBOL(ttm_bo_mem_space);
1014
ttm_bo_move_buffer(struct ttm_buffer_object * bo,struct ttm_placement * placement,bool interruptible,bool no_wait_gpu)1015 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1016 struct ttm_placement *placement,
1017 bool interruptible,
1018 bool no_wait_gpu)
1019 {
1020 int ret = 0;
1021 struct ttm_mem_reg mem;
1022
1023 lockdep_assert_held(&bo->resv->lock.base);
1024
1025 mem.num_pages = bo->num_pages;
1026 mem.size = mem.num_pages << PAGE_SHIFT;
1027 mem.page_alignment = bo->mem.page_alignment;
1028 mem.bus.io_reserved_vm = false;
1029 mem.bus.io_reserved_count = 0;
1030 /*
1031 * Determine where to move the buffer.
1032 */
1033 ret = ttm_bo_mem_space(bo, placement, &mem,
1034 interruptible, no_wait_gpu);
1035 if (ret)
1036 goto out_unlock;
1037 ret = ttm_bo_handle_move_mem(bo, &mem, false,
1038 interruptible, no_wait_gpu);
1039 out_unlock:
1040 if (ret && mem.mm_node)
1041 ttm_bo_mem_put(bo, &mem);
1042 return ret;
1043 }
1044
ttm_bo_mem_compat(struct ttm_placement * placement,struct ttm_mem_reg * mem,uint32_t * new_flags)1045 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1046 struct ttm_mem_reg *mem,
1047 uint32_t *new_flags)
1048 {
1049 int i;
1050
1051 for (i = 0; i < placement->num_placement; i++) {
1052 const struct ttm_place *heap = &placement->placement[i];
1053 if (mem->mm_node &&
1054 (mem->start < heap->fpfn ||
1055 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1056 continue;
1057
1058 *new_flags = heap->flags;
1059 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1060 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1061 return true;
1062 }
1063
1064 for (i = 0; i < placement->num_busy_placement; i++) {
1065 const struct ttm_place *heap = &placement->busy_placement[i];
1066 if (mem->mm_node &&
1067 (mem->start < heap->fpfn ||
1068 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1069 continue;
1070
1071 *new_flags = heap->flags;
1072 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1073 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1074 return true;
1075 }
1076
1077 return false;
1078 }
1079 EXPORT_SYMBOL(ttm_bo_mem_compat);
1080
ttm_bo_validate(struct ttm_buffer_object * bo,struct ttm_placement * placement,bool interruptible,bool no_wait_gpu)1081 int ttm_bo_validate(struct ttm_buffer_object *bo,
1082 struct ttm_placement *placement,
1083 bool interruptible,
1084 bool no_wait_gpu)
1085 {
1086 int ret;
1087 uint32_t new_flags;
1088
1089 lockdep_assert_held(&bo->resv->lock.base);
1090 /*
1091 * Check whether we need to move buffer.
1092 */
1093 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1094 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1095 no_wait_gpu);
1096 if (ret)
1097 return ret;
1098 } else {
1099 /*
1100 * Use the access and other non-mapping-related flag bits from
1101 * the compatible memory placement flags to the active flags
1102 */
1103 ttm_flag_masked(&bo->mem.placement, new_flags,
1104 ~TTM_PL_MASK_MEMTYPE);
1105 }
1106 /*
1107 * We might need to add a TTM.
1108 */
1109 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1110 ret = ttm_bo_add_ttm(bo, true);
1111 if (ret)
1112 return ret;
1113 }
1114 return 0;
1115 }
1116 EXPORT_SYMBOL(ttm_bo_validate);
1117
ttm_bo_init(struct ttm_bo_device * bdev,struct ttm_buffer_object * bo,unsigned long size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,bool interruptible,struct file * persistent_swap_storage,size_t acc_size,struct sg_table * sg,struct reservation_object * resv,void (* destroy)(struct ttm_buffer_object *))1118 int ttm_bo_init(struct ttm_bo_device *bdev,
1119 struct ttm_buffer_object *bo,
1120 unsigned long size,
1121 enum ttm_bo_type type,
1122 struct ttm_placement *placement,
1123 uint32_t page_alignment,
1124 bool interruptible,
1125 struct file *persistent_swap_storage,
1126 size_t acc_size,
1127 struct sg_table *sg,
1128 struct reservation_object *resv,
1129 void (*destroy) (struct ttm_buffer_object *))
1130 {
1131 int ret = 0;
1132 unsigned long num_pages;
1133 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1134 bool locked;
1135
1136 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1137 if (ret) {
1138 pr_err("Out of kernel memory\n");
1139 if (destroy)
1140 (*destroy)(bo);
1141 else
1142 kfree(bo);
1143 return -ENOMEM;
1144 }
1145
1146 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1147 if (num_pages == 0) {
1148 pr_err("Illegal buffer object size\n");
1149 if (destroy)
1150 (*destroy)(bo);
1151 else
1152 kfree(bo);
1153 ttm_mem_global_free(mem_glob, acc_size);
1154 return -EINVAL;
1155 }
1156 bo->destroy = destroy;
1157
1158 kref_init(&bo->kref);
1159 kref_init(&bo->list_kref);
1160 atomic_set(&bo->cpu_writers, 0);
1161 INIT_LIST_HEAD(&bo->lru);
1162 INIT_LIST_HEAD(&bo->ddestroy);
1163 INIT_LIST_HEAD(&bo->swap);
1164 INIT_LIST_HEAD(&bo->io_reserve_lru);
1165 mutex_init(&bo->wu_mutex);
1166 bo->bdev = bdev;
1167 bo->glob = bdev->glob;
1168 bo->type = type;
1169 bo->num_pages = num_pages;
1170 bo->mem.size = num_pages << PAGE_SHIFT;
1171 bo->mem.mem_type = TTM_PL_SYSTEM;
1172 bo->mem.num_pages = bo->num_pages;
1173 bo->mem.mm_node = NULL;
1174 bo->mem.page_alignment = page_alignment;
1175 bo->mem.bus.io_reserved_vm = false;
1176 bo->mem.bus.io_reserved_count = 0;
1177 bo->moving = NULL;
1178 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1179 bo->persistent_swap_storage = persistent_swap_storage;
1180 bo->acc_size = acc_size;
1181 bo->sg = sg;
1182 if (resv) {
1183 bo->resv = resv;
1184 lockdep_assert_held(&bo->resv->lock.base);
1185 } else {
1186 bo->resv = &bo->ttm_resv;
1187 reservation_object_init(&bo->ttm_resv);
1188 }
1189 atomic_inc(&bo->glob->bo_count);
1190 drm_vma_node_reset(&bo->vma_node);
1191
1192 /*
1193 * For ttm_bo_type_device buffers, allocate
1194 * address space from the device.
1195 */
1196 if (bo->type == ttm_bo_type_device ||
1197 bo->type == ttm_bo_type_sg)
1198 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1199 bo->mem.num_pages);
1200
1201 /* passed reservation objects should already be locked,
1202 * since otherwise lockdep will be angered in radeon.
1203 */
1204 if (!resv) {
1205 locked = ww_mutex_trylock(&bo->resv->lock);
1206 WARN_ON(!locked);
1207 }
1208
1209 if (likely(!ret))
1210 ret = ttm_bo_validate(bo, placement, interruptible, false);
1211
1212 if (!resv)
1213 ttm_bo_unreserve(bo);
1214
1215 if (unlikely(ret)) {
1216 ttm_bo_unref(&bo);
1217 return ret;
1218 }
1219
1220 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1221 spin_lock(&bo->glob->lru_lock);
1222 ttm_bo_add_to_lru(bo);
1223 spin_unlock(&bo->glob->lru_lock);
1224 }
1225
1226 return ret;
1227 }
1228 EXPORT_SYMBOL(ttm_bo_init);
1229
ttm_bo_acc_size(struct ttm_bo_device * bdev,unsigned long bo_size,unsigned struct_size)1230 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1231 unsigned long bo_size,
1232 unsigned struct_size)
1233 {
1234 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1235 size_t size = 0;
1236
1237 size += ttm_round_pot(struct_size);
1238 size += ttm_round_pot(npages * sizeof(void *));
1239 size += ttm_round_pot(sizeof(struct ttm_tt));
1240 return size;
1241 }
1242 EXPORT_SYMBOL(ttm_bo_acc_size);
1243
ttm_bo_dma_acc_size(struct ttm_bo_device * bdev,unsigned long bo_size,unsigned struct_size)1244 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1245 unsigned long bo_size,
1246 unsigned struct_size)
1247 {
1248 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1249 size_t size = 0;
1250
1251 size += ttm_round_pot(struct_size);
1252 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1253 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1254 return size;
1255 }
1256 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1257
ttm_bo_create(struct ttm_bo_device * bdev,unsigned long size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,bool interruptible,struct file * persistent_swap_storage,struct ttm_buffer_object ** p_bo)1258 int ttm_bo_create(struct ttm_bo_device *bdev,
1259 unsigned long size,
1260 enum ttm_bo_type type,
1261 struct ttm_placement *placement,
1262 uint32_t page_alignment,
1263 bool interruptible,
1264 struct file *persistent_swap_storage,
1265 struct ttm_buffer_object **p_bo)
1266 {
1267 struct ttm_buffer_object *bo;
1268 size_t acc_size;
1269 int ret;
1270
1271 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1272 if (unlikely(bo == NULL))
1273 return -ENOMEM;
1274
1275 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1276 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1277 interruptible, persistent_swap_storage, acc_size,
1278 NULL, NULL, NULL);
1279 if (likely(ret == 0))
1280 *p_bo = bo;
1281
1282 return ret;
1283 }
1284 EXPORT_SYMBOL(ttm_bo_create);
1285
ttm_bo_force_list_clean(struct ttm_bo_device * bdev,unsigned mem_type,bool allow_errors)1286 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1287 unsigned mem_type, bool allow_errors)
1288 {
1289 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1290 struct ttm_bo_global *glob = bdev->glob;
1291 struct fence *fence;
1292 int ret;
1293
1294 /*
1295 * Can't use standard list traversal since we're unlocking.
1296 */
1297
1298 spin_lock(&glob->lru_lock);
1299 while (!list_empty(&man->lru)) {
1300 spin_unlock(&glob->lru_lock);
1301 ret = ttm_mem_evict_first(bdev, mem_type, NULL, false, false);
1302 if (ret) {
1303 if (allow_errors) {
1304 return ret;
1305 } else {
1306 pr_err("Cleanup eviction failed\n");
1307 }
1308 }
1309 spin_lock(&glob->lru_lock);
1310 }
1311 spin_unlock(&glob->lru_lock);
1312
1313 spin_lock(&man->move_lock);
1314 fence = fence_get(man->move);
1315 spin_unlock(&man->move_lock);
1316
1317 if (fence) {
1318 ret = fence_wait(fence, false);
1319 fence_put(fence);
1320 if (ret) {
1321 if (allow_errors) {
1322 return ret;
1323 } else {
1324 pr_err("Cleanup eviction failed\n");
1325 }
1326 }
1327 }
1328
1329 return 0;
1330 }
1331
ttm_bo_clean_mm(struct ttm_bo_device * bdev,unsigned mem_type)1332 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1333 {
1334 struct ttm_mem_type_manager *man;
1335 int ret = -EINVAL;
1336
1337 if (mem_type >= TTM_NUM_MEM_TYPES) {
1338 pr_err("Illegal memory type %d\n", mem_type);
1339 return ret;
1340 }
1341 man = &bdev->man[mem_type];
1342
1343 if (!man->has_type) {
1344 pr_err("Trying to take down uninitialized memory manager type %u\n",
1345 mem_type);
1346 return ret;
1347 }
1348
1349 man->use_type = false;
1350 man->has_type = false;
1351
1352 ret = 0;
1353 if (mem_type > 0) {
1354 ttm_bo_force_list_clean(bdev, mem_type, false);
1355
1356 ret = (*man->func->takedown)(man);
1357 }
1358
1359 fence_put(man->move);
1360 man->move = NULL;
1361
1362 return ret;
1363 }
1364 EXPORT_SYMBOL(ttm_bo_clean_mm);
1365
ttm_bo_evict_mm(struct ttm_bo_device * bdev,unsigned mem_type)1366 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1367 {
1368 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1369
1370 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1371 pr_err("Illegal memory manager memory type %u\n", mem_type);
1372 return -EINVAL;
1373 }
1374
1375 if (!man->has_type) {
1376 pr_err("Memory type %u has not been initialized\n", mem_type);
1377 return 0;
1378 }
1379
1380 return ttm_bo_force_list_clean(bdev, mem_type, true);
1381 }
1382 EXPORT_SYMBOL(ttm_bo_evict_mm);
1383
ttm_bo_init_mm(struct ttm_bo_device * bdev,unsigned type,unsigned long p_size)1384 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1385 unsigned long p_size)
1386 {
1387 int ret = -EINVAL;
1388 struct ttm_mem_type_manager *man;
1389
1390 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1391 man = &bdev->man[type];
1392 BUG_ON(man->has_type);
1393 man->io_reserve_fastpath = true;
1394 man->use_io_reserve_lru = false;
1395 mutex_init(&man->io_reserve_mutex);
1396 spin_lock_init(&man->move_lock);
1397 INIT_LIST_HEAD(&man->io_reserve_lru);
1398
1399 ret = bdev->driver->init_mem_type(bdev, type, man);
1400 if (ret)
1401 return ret;
1402 man->bdev = bdev;
1403
1404 ret = 0;
1405 if (type != TTM_PL_SYSTEM) {
1406 ret = (*man->func->init)(man, p_size);
1407 if (ret)
1408 return ret;
1409 }
1410 man->has_type = true;
1411 man->use_type = true;
1412 man->size = p_size;
1413
1414 INIT_LIST_HEAD(&man->lru);
1415 man->move = NULL;
1416
1417 return 0;
1418 }
1419 EXPORT_SYMBOL(ttm_bo_init_mm);
1420
ttm_bo_global_kobj_release(struct kobject * kobj)1421 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1422 {
1423 struct ttm_bo_global *glob =
1424 container_of(kobj, struct ttm_bo_global, kobj);
1425
1426 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1427 __free_page(glob->dummy_read_page);
1428 kfree(glob);
1429 }
1430
ttm_bo_global_release(struct drm_global_reference * ref)1431 void ttm_bo_global_release(struct drm_global_reference *ref)
1432 {
1433 struct ttm_bo_global *glob = ref->object;
1434
1435 kobject_del(&glob->kobj);
1436 kobject_put(&glob->kobj);
1437 }
1438 EXPORT_SYMBOL(ttm_bo_global_release);
1439
ttm_bo_global_init(struct drm_global_reference * ref)1440 int ttm_bo_global_init(struct drm_global_reference *ref)
1441 {
1442 struct ttm_bo_global_ref *bo_ref =
1443 container_of(ref, struct ttm_bo_global_ref, ref);
1444 struct ttm_bo_global *glob = ref->object;
1445 int ret;
1446
1447 mutex_init(&glob->device_list_mutex);
1448 spin_lock_init(&glob->lru_lock);
1449 glob->mem_glob = bo_ref->mem_glob;
1450 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1451
1452 if (unlikely(glob->dummy_read_page == NULL)) {
1453 ret = -ENOMEM;
1454 goto out_no_drp;
1455 }
1456
1457 INIT_LIST_HEAD(&glob->swap_lru);
1458 INIT_LIST_HEAD(&glob->device_list);
1459
1460 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1461 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1462 if (unlikely(ret != 0)) {
1463 pr_err("Could not register buffer object swapout\n");
1464 goto out_no_shrink;
1465 }
1466
1467 atomic_set(&glob->bo_count, 0);
1468
1469 ret = kobject_init_and_add(
1470 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1471 if (unlikely(ret != 0))
1472 kobject_put(&glob->kobj);
1473 return ret;
1474 out_no_shrink:
1475 __free_page(glob->dummy_read_page);
1476 out_no_drp:
1477 kfree(glob);
1478 return ret;
1479 }
1480 EXPORT_SYMBOL(ttm_bo_global_init);
1481
1482
ttm_bo_device_release(struct ttm_bo_device * bdev)1483 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1484 {
1485 int ret = 0;
1486 unsigned i = TTM_NUM_MEM_TYPES;
1487 struct ttm_mem_type_manager *man;
1488 struct ttm_bo_global *glob = bdev->glob;
1489
1490 while (i--) {
1491 man = &bdev->man[i];
1492 if (man->has_type) {
1493 man->use_type = false;
1494 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1495 ret = -EBUSY;
1496 pr_err("DRM memory manager type %d is not clean\n",
1497 i);
1498 }
1499 man->has_type = false;
1500 }
1501 }
1502
1503 mutex_lock(&glob->device_list_mutex);
1504 list_del(&bdev->device_list);
1505 mutex_unlock(&glob->device_list_mutex);
1506
1507 cancel_delayed_work_sync(&bdev->wq);
1508
1509 while (ttm_bo_delayed_delete(bdev, true))
1510 ;
1511
1512 spin_lock(&glob->lru_lock);
1513 if (list_empty(&bdev->ddestroy))
1514 TTM_DEBUG("Delayed destroy list was clean\n");
1515
1516 if (list_empty(&bdev->man[0].lru))
1517 TTM_DEBUG("Swap list was clean\n");
1518 spin_unlock(&glob->lru_lock);
1519
1520 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1521
1522 return ret;
1523 }
1524 EXPORT_SYMBOL(ttm_bo_device_release);
1525
ttm_bo_device_init(struct ttm_bo_device * bdev,struct ttm_bo_global * glob,struct ttm_bo_driver * driver,struct address_space * mapping,uint64_t file_page_offset,bool need_dma32)1526 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1527 struct ttm_bo_global *glob,
1528 struct ttm_bo_driver *driver,
1529 struct address_space *mapping,
1530 uint64_t file_page_offset,
1531 bool need_dma32)
1532 {
1533 int ret = -EINVAL;
1534
1535 bdev->driver = driver;
1536
1537 memset(bdev->man, 0, sizeof(bdev->man));
1538
1539 /*
1540 * Initialize the system memory buffer type.
1541 * Other types need to be driver / IOCTL initialized.
1542 */
1543 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1544 if (unlikely(ret != 0))
1545 goto out_no_sys;
1546
1547 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1548 0x10000000);
1549 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1550 INIT_LIST_HEAD(&bdev->ddestroy);
1551 bdev->dev_mapping = mapping;
1552 bdev->glob = glob;
1553 bdev->need_dma32 = need_dma32;
1554 mutex_lock(&glob->device_list_mutex);
1555 list_add_tail(&bdev->device_list, &glob->device_list);
1556 mutex_unlock(&glob->device_list_mutex);
1557
1558 return 0;
1559 out_no_sys:
1560 return ret;
1561 }
1562 EXPORT_SYMBOL(ttm_bo_device_init);
1563
1564 /*
1565 * buffer object vm functions.
1566 */
1567
ttm_mem_reg_is_pci(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)1568 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1569 {
1570 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1571
1572 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1573 if (mem->mem_type == TTM_PL_SYSTEM)
1574 return false;
1575
1576 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1577 return false;
1578
1579 if (mem->placement & TTM_PL_FLAG_CACHED)
1580 return false;
1581 }
1582 return true;
1583 }
1584
ttm_bo_unmap_virtual_locked(struct ttm_buffer_object * bo)1585 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1586 {
1587 struct ttm_bo_device *bdev = bo->bdev;
1588
1589 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1590 ttm_mem_io_free_vm(bo);
1591 }
1592
ttm_bo_unmap_virtual(struct ttm_buffer_object * bo)1593 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1594 {
1595 struct ttm_bo_device *bdev = bo->bdev;
1596 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1597
1598 ttm_mem_io_lock(man, false);
1599 ttm_bo_unmap_virtual_locked(bo);
1600 ttm_mem_io_unlock(man);
1601 }
1602
1603
1604 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1605
ttm_bo_wait(struct ttm_buffer_object * bo,bool interruptible,bool no_wait)1606 int ttm_bo_wait(struct ttm_buffer_object *bo,
1607 bool interruptible, bool no_wait)
1608 {
1609 long timeout = 15 * HZ;
1610
1611 if (no_wait) {
1612 if (reservation_object_test_signaled_rcu(bo->resv, true))
1613 return 0;
1614 else
1615 return -EBUSY;
1616 }
1617
1618 timeout = reservation_object_wait_timeout_rcu(bo->resv, true,
1619 interruptible, timeout);
1620 if (timeout < 0)
1621 return timeout;
1622
1623 if (timeout == 0)
1624 return -EBUSY;
1625
1626 reservation_object_add_excl_fence(bo->resv, NULL);
1627 return 0;
1628 }
1629 EXPORT_SYMBOL(ttm_bo_wait);
1630
ttm_bo_synccpu_write_grab(struct ttm_buffer_object * bo,bool no_wait)1631 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1632 {
1633 int ret = 0;
1634
1635 /*
1636 * Using ttm_bo_reserve makes sure the lru lists are updated.
1637 */
1638
1639 ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1640 if (unlikely(ret != 0))
1641 return ret;
1642 ret = ttm_bo_wait(bo, true, no_wait);
1643 if (likely(ret == 0))
1644 atomic_inc(&bo->cpu_writers);
1645 ttm_bo_unreserve(bo);
1646 return ret;
1647 }
1648 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1649
ttm_bo_synccpu_write_release(struct ttm_buffer_object * bo)1650 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1651 {
1652 atomic_dec(&bo->cpu_writers);
1653 }
1654 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1655
1656 /**
1657 * A buffer object shrink method that tries to swap out the first
1658 * buffer object on the bo_global::swap_lru list.
1659 */
1660
ttm_bo_swapout(struct ttm_mem_shrink * shrink)1661 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1662 {
1663 struct ttm_bo_global *glob =
1664 container_of(shrink, struct ttm_bo_global, shrink);
1665 struct ttm_buffer_object *bo;
1666 int ret = -EBUSY;
1667 int put_count;
1668
1669 spin_lock(&glob->lru_lock);
1670 list_for_each_entry(bo, &glob->swap_lru, swap) {
1671 ret = __ttm_bo_reserve(bo, false, true, NULL);
1672 if (!ret)
1673 break;
1674 }
1675
1676 if (ret) {
1677 spin_unlock(&glob->lru_lock);
1678 return ret;
1679 }
1680
1681 kref_get(&bo->list_kref);
1682
1683 if (!list_empty(&bo->ddestroy)) {
1684 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1685 kref_put(&bo->list_kref, ttm_bo_release_list);
1686 return ret;
1687 }
1688
1689 put_count = ttm_bo_del_from_lru(bo);
1690 spin_unlock(&glob->lru_lock);
1691
1692 ttm_bo_list_ref_sub(bo, put_count, true);
1693
1694 /**
1695 * Move to system cached
1696 */
1697
1698 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1699 bo->ttm->caching_state != tt_cached) {
1700 struct ttm_mem_reg evict_mem;
1701
1702 evict_mem = bo->mem;
1703 evict_mem.mm_node = NULL;
1704 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1705 evict_mem.mem_type = TTM_PL_SYSTEM;
1706
1707 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1708 false, false);
1709 if (unlikely(ret != 0))
1710 goto out;
1711 }
1712
1713 /**
1714 * Make sure BO is idle.
1715 */
1716
1717 ret = ttm_bo_wait(bo, false, false);
1718 if (unlikely(ret != 0))
1719 goto out;
1720
1721 ttm_bo_unmap_virtual(bo);
1722
1723 /**
1724 * Swap out. Buffer will be swapped in again as soon as
1725 * anyone tries to access a ttm page.
1726 */
1727
1728 if (bo->bdev->driver->swap_notify)
1729 bo->bdev->driver->swap_notify(bo);
1730
1731 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1732 out:
1733
1734 /**
1735 *
1736 * Unreserve without putting on LRU to avoid swapping out an
1737 * already swapped buffer.
1738 */
1739
1740 __ttm_bo_unreserve(bo);
1741 kref_put(&bo->list_kref, ttm_bo_release_list);
1742 return ret;
1743 }
1744
ttm_bo_swapout_all(struct ttm_bo_device * bdev)1745 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1746 {
1747 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1748 ;
1749 }
1750 EXPORT_SYMBOL(ttm_bo_swapout_all);
1751
1752 /**
1753 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1754 * unreserved
1755 *
1756 * @bo: Pointer to buffer
1757 */
ttm_bo_wait_unreserved(struct ttm_buffer_object * bo)1758 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1759 {
1760 int ret;
1761
1762 /*
1763 * In the absense of a wait_unlocked API,
1764 * Use the bo::wu_mutex to avoid triggering livelocks due to
1765 * concurrent use of this function. Note that this use of
1766 * bo::wu_mutex can go away if we change locking order to
1767 * mmap_sem -> bo::reserve.
1768 */
1769 ret = mutex_lock_interruptible(&bo->wu_mutex);
1770 if (unlikely(ret != 0))
1771 return -ERESTARTSYS;
1772 if (!ww_mutex_is_locked(&bo->resv->lock))
1773 goto out_unlock;
1774 ret = __ttm_bo_reserve(bo, true, false, NULL);
1775 if (unlikely(ret != 0))
1776 goto out_unlock;
1777 __ttm_bo_unreserve(bo);
1778
1779 out_unlock:
1780 mutex_unlock(&bo->wu_mutex);
1781 return ret;
1782 }
1783