projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
drm/ttm: convert to unified vma offset manager
[deliverable/linux.git] / drivers / gpu / drm / ttm / ttm_bo.c
1 /**************************************************************************
2 *
3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27 /*
28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29 */
30
31 #define pr_fmt(fmt) "[TTM] " fmt
32
33 #include <drm/ttm/ttm_module.h>
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <linux/jiffies.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 #include <linux/mm.h>
40 #include <linux/file.h>
41 #include <linux/module.h>
42 #include <linux/atomic.h>
43
44 #define TTM_ASSERT_LOCKED(param)
45 #define TTM_DEBUG(fmt, arg...)
46 #define TTM_BO_HASH_ORDER 13
47
48 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo);
49 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50 static void ttm_bo_global_kobj_release(struct kobject *kobj);
51
52 static struct attribute ttm_bo_count = {
53 .name = "bo_count",
54 .mode = S_IRUGO
55 };
56
57 static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
58 {
59 int i;
60
61 for (i = 0; i <= TTM_PL_PRIV5; i++)
62 if (flags & (1 << i)) {
63 *mem_type = i;
64 return 0;
65 }
66 return -EINVAL;
67 }
68
69 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
70 {
71 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
72
73 pr_err(" has_type: %d\n", man->has_type);
74 pr_err(" use_type: %d\n", man->use_type);
75 pr_err(" flags: 0x%08X\n", man->flags);
76 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset);
77 pr_err(" size: %llu\n", man->size);
78 pr_err(" available_caching: 0x%08X\n", man->available_caching);
79 pr_err(" default_caching: 0x%08X\n", man->default_caching);
80 if (mem_type != TTM_PL_SYSTEM)
81 (*man->func->debug)(man, TTM_PFX);
82 }
83
84 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
85 struct ttm_placement *placement)
86 {
87 int i, ret, mem_type;
88
89 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
90 bo, bo->mem.num_pages, bo->mem.size >> 10,
91 bo->mem.size >> 20);
92 for (i = 0; i < placement->num_placement; i++) {
93 ret = ttm_mem_type_from_flags(placement->placement[i],
94 &mem_type);
95 if (ret)
96 return;
97 pr_err(" placement[%d]=0x%08X (%d)\n",
98 i, placement->placement[i], mem_type);
99 ttm_mem_type_debug(bo->bdev, mem_type);
100 }
101 }
102
103 static ssize_t ttm_bo_global_show(struct kobject *kobj,
104 struct attribute *attr,
105 char *buffer)
106 {
107 struct ttm_bo_global *glob =
108 container_of(kobj, struct ttm_bo_global, kobj);
109
110 return snprintf(buffer, PAGE_SIZE, "%lu\n",
111 (unsigned long) atomic_read(&glob->bo_count));
112 }
113
114 static struct attribute *ttm_bo_global_attrs[] = {
115 &ttm_bo_count,
116 NULL
117 };
118
119 static const struct sysfs_ops ttm_bo_global_ops = {
120 .show = &ttm_bo_global_show
121 };
122
123 static struct kobj_type ttm_bo_glob_kobj_type = {
124 .release = &ttm_bo_global_kobj_release,
125 .sysfs_ops = &ttm_bo_global_ops,
126 .default_attrs = ttm_bo_global_attrs
127 };
128
129
130 static inline uint32_t ttm_bo_type_flags(unsigned type)
131 {
132 return 1 << (type);
133 }
134
135 static void ttm_bo_release_list(struct kref *list_kref)
136 {
137 struct ttm_buffer_object *bo =
138 container_of(list_kref, struct ttm_buffer_object, list_kref);
139 struct ttm_bo_device *bdev = bo->bdev;
140 size_t acc_size = bo->acc_size;
141
142 BUG_ON(atomic_read(&bo->list_kref.refcount));
143 BUG_ON(atomic_read(&bo->kref.refcount));
144 BUG_ON(atomic_read(&bo->cpu_writers));
145 BUG_ON(bo->sync_obj != NULL);
146 BUG_ON(bo->mem.mm_node != NULL);
147 BUG_ON(!list_empty(&bo->lru));
148 BUG_ON(!list_empty(&bo->ddestroy));
149
150 if (bo->ttm)
151 ttm_tt_destroy(bo->ttm);
152 atomic_dec(&bo->glob->bo_count);
153 if (bo->resv == &bo->ttm_resv)
154 reservation_object_fini(&bo->ttm_resv);
155
156 if (bo->destroy)
157 bo->destroy(bo);
158 else {
159 kfree(bo);
160 }
161 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
162 }
163
164 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
165 {
166 struct ttm_bo_device *bdev = bo->bdev;
167 struct ttm_mem_type_manager *man;
168
169 lockdep_assert_held(&bo->resv->lock.base);
170
171 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
172
173 BUG_ON(!list_empty(&bo->lru));
174
175 man = &bdev->man[bo->mem.mem_type];
176 list_add_tail(&bo->lru, &man->lru);
177 kref_get(&bo->list_kref);
178
179 if (bo->ttm != NULL) {
180 list_add_tail(&bo->swap, &bo->glob->swap_lru);
181 kref_get(&bo->list_kref);
182 }
183 }
184 }
185 EXPORT_SYMBOL(ttm_bo_add_to_lru);
186
187 int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
188 {
189 int put_count = 0;
190
191 if (!list_empty(&bo->swap)) {
192 list_del_init(&bo->swap);
193 ++put_count;
194 }
195 if (!list_empty(&bo->lru)) {
196 list_del_init(&bo->lru);
197 ++put_count;
198 }
199
200 /*
201 * TODO: Add a driver hook to delete from
202 * driver-specific LRU's here.
203 */
204
205 return put_count;
206 }
207
208 static void ttm_bo_ref_bug(struct kref *list_kref)
209 {
210 BUG();
211 }
212
213 void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
214 bool never_free)
215 {
216 kref_sub(&bo->list_kref, count,
217 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
218 }
219
220 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
221 {
222 int put_count;
223
224 spin_lock(&bo->glob->lru_lock);
225 put_count = ttm_bo_del_from_lru(bo);
226 spin_unlock(&bo->glob->lru_lock);
227 ttm_bo_list_ref_sub(bo, put_count, true);
228 }
229 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
230
231 /*
232 * Call bo->mutex locked.
233 */
234 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
235 {
236 struct ttm_bo_device *bdev = bo->bdev;
237 struct ttm_bo_global *glob = bo->glob;
238 int ret = 0;
239 uint32_t page_flags = 0;
240
241 TTM_ASSERT_LOCKED(&bo->mutex);
242 bo->ttm = NULL;
243
244 if (bdev->need_dma32)
245 page_flags |= TTM_PAGE_FLAG_DMA32;
246
247 switch (bo->type) {
248 case ttm_bo_type_device:
249 if (zero_alloc)
250 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
251 case ttm_bo_type_kernel:
252 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
253 page_flags, glob->dummy_read_page);
254 if (unlikely(bo->ttm == NULL))
255 ret = -ENOMEM;
256 break;
257 case ttm_bo_type_sg:
258 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
259 page_flags | TTM_PAGE_FLAG_SG,
260 glob->dummy_read_page);
261 if (unlikely(bo->ttm == NULL)) {
262 ret = -ENOMEM;
263 break;
264 }
265 bo->ttm->sg = bo->sg;
266 break;
267 default:
268 pr_err("Illegal buffer object type\n");
269 ret = -EINVAL;
270 break;
271 }
272
273 return ret;
274 }
275
276 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
277 struct ttm_mem_reg *mem,
278 bool evict, bool interruptible,
279 bool no_wait_gpu)
280 {
281 struct ttm_bo_device *bdev = bo->bdev;
282 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
283 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
284 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
285 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
286 int ret = 0;
287
288 if (old_is_pci || new_is_pci ||
289 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
290 ret = ttm_mem_io_lock(old_man, true);
291 if (unlikely(ret != 0))
292 goto out_err;
293 ttm_bo_unmap_virtual_locked(bo);
294 ttm_mem_io_unlock(old_man);
295 }
296
297 /*
298 * Create and bind a ttm if required.
299 */
300
301 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
302 if (bo->ttm == NULL) {
303 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
304 ret = ttm_bo_add_ttm(bo, zero);
305 if (ret)
306 goto out_err;
307 }
308
309 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
310 if (ret)
311 goto out_err;
312
313 if (mem->mem_type != TTM_PL_SYSTEM) {
314 ret = ttm_tt_bind(bo->ttm, mem);
315 if (ret)
316 goto out_err;
317 }
318
319 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
320 if (bdev->driver->move_notify)
321 bdev->driver->move_notify(bo, mem);
322 bo->mem = *mem;
323 mem->mm_node = NULL;
324 goto moved;
325 }
326 }
327
328 if (bdev->driver->move_notify)
329 bdev->driver->move_notify(bo, mem);
330
331 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
332 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
333 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
334 else if (bdev->driver->move)
335 ret = bdev->driver->move(bo, evict, interruptible,
336 no_wait_gpu, mem);
337 else
338 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
339
340 if (ret) {
341 if (bdev->driver->move_notify) {
342 struct ttm_mem_reg tmp_mem = *mem;
343 *mem = bo->mem;
344 bo->mem = tmp_mem;
345 bdev->driver->move_notify(bo, mem);
346 bo->mem = *mem;
347 *mem = tmp_mem;
348 }
349
350 goto out_err;
351 }
352
353 moved:
354 if (bo->evicted) {
355 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
356 if (ret)
357 pr_err("Can not flush read caches\n");
358 bo->evicted = false;
359 }
360
361 if (bo->mem.mm_node) {
362 bo->offset = (bo->mem.start << PAGE_SHIFT) +
363 bdev->man[bo->mem.mem_type].gpu_offset;
364 bo->cur_placement = bo->mem.placement;
365 } else
366 bo->offset = 0;
367
368 return 0;
369
370 out_err:
371 new_man = &bdev->man[bo->mem.mem_type];
372 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
373 ttm_tt_unbind(bo->ttm);
374 ttm_tt_destroy(bo->ttm);
375 bo->ttm = NULL;
376 }
377
378 return ret;
379 }
380
381 /**
382 * Call bo::reserved.
383 * Will release GPU memory type usage on destruction.
384 * This is the place to put in driver specific hooks to release
385 * driver private resources.
386 * Will release the bo::reserved lock.
387 */
388
389 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
390 {
391 if (bo->bdev->driver->move_notify)
392 bo->bdev->driver->move_notify(bo, NULL);
393
394 if (bo->ttm) {
395 ttm_tt_unbind(bo->ttm);
396 ttm_tt_destroy(bo->ttm);
397 bo->ttm = NULL;
398 }
399 ttm_bo_mem_put(bo, &bo->mem);
400
401 ww_mutex_unlock (&bo->resv->lock);
402 }
403
404 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
405 {
406 struct ttm_bo_device *bdev = bo->bdev;
407 struct ttm_bo_global *glob = bo->glob;
408 struct ttm_bo_driver *driver = bdev->driver;
409 void *sync_obj = NULL;
410 int put_count;
411 int ret;
412
413 spin_lock(&glob->lru_lock);
414 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0);
415
416 spin_lock(&bdev->fence_lock);
417 (void) ttm_bo_wait(bo, false, false, true);
418 if (!ret && !bo->sync_obj) {
419 spin_unlock(&bdev->fence_lock);
420 put_count = ttm_bo_del_from_lru(bo);
421
422 spin_unlock(&glob->lru_lock);
423 ttm_bo_cleanup_memtype_use(bo);
424
425 ttm_bo_list_ref_sub(bo, put_count, true);
426
427 return;
428 }
429 if (bo->sync_obj)
430 sync_obj = driver->sync_obj_ref(bo->sync_obj);
431 spin_unlock(&bdev->fence_lock);
432
433 if (!ret)
434 ww_mutex_unlock(&bo->resv->lock);
435
436 kref_get(&bo->list_kref);
437 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
438 spin_unlock(&glob->lru_lock);
439
440 if (sync_obj) {
441 driver->sync_obj_flush(sync_obj);
442 driver->sync_obj_unref(&sync_obj);
443 }
444 schedule_delayed_work(&bdev->wq,
445 ((HZ / 100) < 1) ? 1 : HZ / 100);
446 }
447
448 /**
449 * function ttm_bo_cleanup_refs_and_unlock
450 * If bo idle, remove from delayed- and lru lists, and unref.
451 * If not idle, do nothing.
452 *
453 * Must be called with lru_lock and reservation held, this function
454 * will drop both before returning.
455 *
456 * @interruptible Any sleeps should occur interruptibly.
457 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
458 */
459
460 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
461 bool interruptible,
462 bool no_wait_gpu)
463 {
464 struct ttm_bo_device *bdev = bo->bdev;
465 struct ttm_bo_driver *driver = bdev->driver;
466 struct ttm_bo_global *glob = bo->glob;
467 int put_count;
468 int ret;
469
470 spin_lock(&bdev->fence_lock);
471 ret = ttm_bo_wait(bo, false, false, true);
472
473 if (ret && !no_wait_gpu) {
474 void *sync_obj;
475
476 /*
477 * Take a reference to the fence and unreserve,
478 * at this point the buffer should be dead, so
479 * no new sync objects can be attached.
480 */
481 sync_obj = driver->sync_obj_ref(bo->sync_obj);
482 spin_unlock(&bdev->fence_lock);
483
484 ww_mutex_unlock(&bo->resv->lock);
485 spin_unlock(&glob->lru_lock);
486
487 ret = driver->sync_obj_wait(sync_obj, false, interruptible);
488 driver->sync_obj_unref(&sync_obj);
489 if (ret)
490 return ret;
491
492 /*
493 * remove sync_obj with ttm_bo_wait, the wait should be
494 * finished, and no new wait object should have been added.
495 */
496 spin_lock(&bdev->fence_lock);
497 ret = ttm_bo_wait(bo, false, false, true);
498 WARN_ON(ret);
499 spin_unlock(&bdev->fence_lock);
500 if (ret)
501 return ret;
502
503 spin_lock(&glob->lru_lock);
504 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0);
505
506 /*
507 * We raced, and lost, someone else holds the reservation now,
508 * and is probably busy in ttm_bo_cleanup_memtype_use.
509 *
510 * Even if it's not the case, because we finished waiting any
511 * delayed destruction would succeed, so just return success
512 * here.
513 */
514 if (ret) {
515 spin_unlock(&glob->lru_lock);
516 return 0;
517 }
518 } else
519 spin_unlock(&bdev->fence_lock);
520
521 if (ret || unlikely(list_empty(&bo->ddestroy))) {
522 ww_mutex_unlock(&bo->resv->lock);
523 spin_unlock(&glob->lru_lock);
524 return ret;
525 }
526
527 put_count = ttm_bo_del_from_lru(bo);
528 list_del_init(&bo->ddestroy);
529 ++put_count;
530
531 spin_unlock(&glob->lru_lock);
532 ttm_bo_cleanup_memtype_use(bo);
533
534 ttm_bo_list_ref_sub(bo, put_count, true);
535
536 return 0;
537 }
538
539 /**
540 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
541 * encountered buffers.
542 */
543
544 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
545 {
546 struct ttm_bo_global *glob = bdev->glob;
547 struct ttm_buffer_object *entry = NULL;
548 int ret = 0;
549
550 spin_lock(&glob->lru_lock);
551 if (list_empty(&bdev->ddestroy))
552 goto out_unlock;
553
554 entry = list_first_entry(&bdev->ddestroy,
555 struct ttm_buffer_object, ddestroy);
556 kref_get(&entry->list_kref);
557
558 for (;;) {
559 struct ttm_buffer_object *nentry = NULL;
560
561 if (entry->ddestroy.next != &bdev->ddestroy) {
562 nentry = list_first_entry(&entry->ddestroy,
563 struct ttm_buffer_object, ddestroy);
564 kref_get(&nentry->list_kref);
565 }
566
567 ret = ttm_bo_reserve_nolru(entry, false, true, false, 0);
568 if (remove_all && ret) {
569 spin_unlock(&glob->lru_lock);
570 ret = ttm_bo_reserve_nolru(entry, false, false,
571 false, 0);
572 spin_lock(&glob->lru_lock);
573 }
574
575 if (!ret)
576 ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
577 !remove_all);
578 else
579 spin_unlock(&glob->lru_lock);
580
581 kref_put(&entry->list_kref, ttm_bo_release_list);
582 entry = nentry;
583
584 if (ret || !entry)
585 goto out;
586
587 spin_lock(&glob->lru_lock);
588 if (list_empty(&entry->ddestroy))
589 break;
590 }
591
592 out_unlock:
593 spin_unlock(&glob->lru_lock);
594 out:
595 if (entry)
596 kref_put(&entry->list_kref, ttm_bo_release_list);
597 return ret;
598 }
599
600 static void ttm_bo_delayed_workqueue(struct work_struct *work)
601 {
602 struct ttm_bo_device *bdev =
603 container_of(work, struct ttm_bo_device, wq.work);
604
605 if (ttm_bo_delayed_delete(bdev, false)) {
606 schedule_delayed_work(&bdev->wq,
607 ((HZ / 100) < 1) ? 1 : HZ / 100);
608 }
609 }
610
611 static void ttm_bo_release(struct kref *kref)
612 {
613 struct ttm_buffer_object *bo =
614 container_of(kref, struct ttm_buffer_object, kref);
615 struct ttm_bo_device *bdev = bo->bdev;
616 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
617
618 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
619 ttm_mem_io_lock(man, false);
620 ttm_mem_io_free_vm(bo);
621 ttm_mem_io_unlock(man);
622 ttm_bo_cleanup_refs_or_queue(bo);
623 kref_put(&bo->list_kref, ttm_bo_release_list);
624 }
625
626 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
627 {
628 struct ttm_buffer_object *bo = *p_bo;
629
630 *p_bo = NULL;
631 kref_put(&bo->kref, ttm_bo_release);
632 }
633 EXPORT_SYMBOL(ttm_bo_unref);
634
635 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
636 {
637 return cancel_delayed_work_sync(&bdev->wq);
638 }
639 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
640
641 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
642 {
643 if (resched)
644 schedule_delayed_work(&bdev->wq,
645 ((HZ / 100) < 1) ? 1 : HZ / 100);
646 }
647 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
648
649 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
650 bool no_wait_gpu)
651 {
652 struct ttm_bo_device *bdev = bo->bdev;
653 struct ttm_mem_reg evict_mem;
654 struct ttm_placement placement;
655 int ret = 0;
656
657 spin_lock(&bdev->fence_lock);
658 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
659 spin_unlock(&bdev->fence_lock);
660
661 if (unlikely(ret != 0)) {
662 if (ret != -ERESTARTSYS) {
663 pr_err("Failed to expire sync object before buffer eviction\n");
664 }
665 goto out;
666 }
667
668 lockdep_assert_held(&bo->resv->lock.base);
669
670 evict_mem = bo->mem;
671 evict_mem.mm_node = NULL;
672 evict_mem.bus.io_reserved_vm = false;
673 evict_mem.bus.io_reserved_count = 0;
674
675 placement.fpfn = 0;
676 placement.lpfn = 0;
677 placement.num_placement = 0;
678 placement.num_busy_placement = 0;
679 bdev->driver->evict_flags(bo, &placement);
680 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
681 no_wait_gpu);
682 if (ret) {
683 if (ret != -ERESTARTSYS) {
684 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
685 bo);
686 ttm_bo_mem_space_debug(bo, &placement);
687 }
688 goto out;
689 }
690
691 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
692 no_wait_gpu);
693 if (ret) {
694 if (ret != -ERESTARTSYS)
695 pr_err("Buffer eviction failed\n");
696 ttm_bo_mem_put(bo, &evict_mem);
697 goto out;
698 }
699 bo->evicted = true;
700 out:
701 return ret;
702 }
703
704 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
705 uint32_t mem_type,
706 bool interruptible,
707 bool no_wait_gpu)
708 {
709 struct ttm_bo_global *glob = bdev->glob;
710 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
711 struct ttm_buffer_object *bo;
712 int ret = -EBUSY, put_count;
713
714 spin_lock(&glob->lru_lock);
715 list_for_each_entry(bo, &man->lru, lru) {
716 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0);
717 if (!ret)
718 break;
719 }
720
721 if (ret) {
722 spin_unlock(&glob->lru_lock);
723 return ret;
724 }
725
726 kref_get(&bo->list_kref);
727
728 if (!list_empty(&bo->ddestroy)) {
729 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
730 no_wait_gpu);
731 kref_put(&bo->list_kref, ttm_bo_release_list);
732 return ret;
733 }
734
735 put_count = ttm_bo_del_from_lru(bo);
736 spin_unlock(&glob->lru_lock);
737
738 BUG_ON(ret != 0);
739
740 ttm_bo_list_ref_sub(bo, put_count, true);
741
742 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
743 ttm_bo_unreserve(bo);
744
745 kref_put(&bo->list_kref, ttm_bo_release_list);
746 return ret;
747 }
748
749 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
750 {
751 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
752
753 if (mem->mm_node)
754 (*man->func->put_node)(man, mem);
755 }
756 EXPORT_SYMBOL(ttm_bo_mem_put);
757
758 /**
759 * Repeatedly evict memory from the LRU for @mem_type until we create enough
760 * space, or we've evicted everything and there isn't enough space.
761 */
762 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
763 uint32_t mem_type,
764 struct ttm_placement *placement,
765 struct ttm_mem_reg *mem,
766 bool interruptible,
767 bool no_wait_gpu)
768 {
769 struct ttm_bo_device *bdev = bo->bdev;
770 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
771 int ret;
772
773 do {
774 ret = (*man->func->get_node)(man, bo, placement, mem);
775 if (unlikely(ret != 0))
776 return ret;
777 if (mem->mm_node)
778 break;
779 ret = ttm_mem_evict_first(bdev, mem_type,
780 interruptible, no_wait_gpu);
781 if (unlikely(ret != 0))
782 return ret;
783 } while (1);
784 if (mem->mm_node == NULL)
785 return -ENOMEM;
786 mem->mem_type = mem_type;
787 return 0;
788 }
789
790 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
791 uint32_t cur_placement,
792 uint32_t proposed_placement)
793 {
794 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
795 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
796
797 /**
798 * Keep current caching if possible.
799 */
800
801 if ((cur_placement & caching) != 0)
802 result |= (cur_placement & caching);
803 else if ((man->default_caching & caching) != 0)
804 result |= man->default_caching;
805 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
806 result |= TTM_PL_FLAG_CACHED;
807 else if ((TTM_PL_FLAG_WC & caching) != 0)
808 result |= TTM_PL_FLAG_WC;
809 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
810 result |= TTM_PL_FLAG_UNCACHED;
811
812 return result;
813 }
814
815 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
816 uint32_t mem_type,
817 uint32_t proposed_placement,
818 uint32_t *masked_placement)
819 {
820 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
821
822 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
823 return false;
824
825 if ((proposed_placement & man->available_caching) == 0)
826 return false;
827
828 cur_flags |= (proposed_placement & man->available_caching);
829
830 *masked_placement = cur_flags;
831 return true;
832 }
833
834 /**
835 * Creates space for memory region @mem according to its type.
836 *
837 * This function first searches for free space in compatible memory types in
838 * the priority order defined by the driver. If free space isn't found, then
839 * ttm_bo_mem_force_space is attempted in priority order to evict and find
840 * space.
841 */
842 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
843 struct ttm_placement *placement,
844 struct ttm_mem_reg *mem,
845 bool interruptible,
846 bool no_wait_gpu)
847 {
848 struct ttm_bo_device *bdev = bo->bdev;
849 struct ttm_mem_type_manager *man;
850 uint32_t mem_type = TTM_PL_SYSTEM;
851 uint32_t cur_flags = 0;
852 bool type_found = false;
853 bool type_ok = false;
854 bool has_erestartsys = false;
855 int i, ret;
856
857 mem->mm_node = NULL;
858 for (i = 0; i < placement->num_placement; ++i) {
859 ret = ttm_mem_type_from_flags(placement->placement[i],
860 &mem_type);
861 if (ret)
862 return ret;
863 man = &bdev->man[mem_type];
864
865 type_ok = ttm_bo_mt_compatible(man,
866 mem_type,
867 placement->placement[i],
868 &cur_flags);
869
870 if (!type_ok)
871 continue;
872
873 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
874 cur_flags);
875 /*
876 * Use the access and other non-mapping-related flag bits from
877 * the memory placement flags to the current flags
878 */
879 ttm_flag_masked(&cur_flags, placement->placement[i],
880 ~TTM_PL_MASK_MEMTYPE);
881
882 if (mem_type == TTM_PL_SYSTEM)
883 break;
884
885 if (man->has_type && man->use_type) {
886 type_found = true;
887 ret = (*man->func->get_node)(man, bo, placement, mem);
888 if (unlikely(ret))
889 return ret;
890 }
891 if (mem->mm_node)
892 break;
893 }
894
895 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
896 mem->mem_type = mem_type;
897 mem->placement = cur_flags;
898 return 0;
899 }
900
901 if (!type_found)
902 return -EINVAL;
903
904 for (i = 0; i < placement->num_busy_placement; ++i) {
905 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
906 &mem_type);
907 if (ret)
908 return ret;
909 man = &bdev->man[mem_type];
910 if (!man->has_type)
911 continue;
912 if (!ttm_bo_mt_compatible(man,
913 mem_type,
914 placement->busy_placement[i],
915 &cur_flags))
916 continue;
917
918 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
919 cur_flags);
920 /*
921 * Use the access and other non-mapping-related flag bits from
922 * the memory placement flags to the current flags
923 */
924 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
925 ~TTM_PL_MASK_MEMTYPE);
926
927
928 if (mem_type == TTM_PL_SYSTEM) {
929 mem->mem_type = mem_type;
930 mem->placement = cur_flags;
931 mem->mm_node = NULL;
932 return 0;
933 }
934
935 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
936 interruptible, no_wait_gpu);
937 if (ret == 0 && mem->mm_node) {
938 mem->placement = cur_flags;
939 return 0;
940 }
941 if (ret == -ERESTARTSYS)
942 has_erestartsys = true;
943 }
944 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
945 return ret;
946 }
947 EXPORT_SYMBOL(ttm_bo_mem_space);
948
949 int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
950 struct ttm_placement *placement,
951 bool interruptible,
952 bool no_wait_gpu)
953 {
954 int ret = 0;
955 struct ttm_mem_reg mem;
956 struct ttm_bo_device *bdev = bo->bdev;
957
958 lockdep_assert_held(&bo->resv->lock.base);
959
960 /*
961 * FIXME: It's possible to pipeline buffer moves.
962 * Have the driver move function wait for idle when necessary,
963 * instead of doing it here.
964 */
965 spin_lock(&bdev->fence_lock);
966 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
967 spin_unlock(&bdev->fence_lock);
968 if (ret)
969 return ret;
970 mem.num_pages = bo->num_pages;
971 mem.size = mem.num_pages << PAGE_SHIFT;
972 mem.page_alignment = bo->mem.page_alignment;
973 mem.bus.io_reserved_vm = false;
974 mem.bus.io_reserved_count = 0;
975 /*
976 * Determine where to move the buffer.
977 */
978 ret = ttm_bo_mem_space(bo, placement, &mem,
979 interruptible, no_wait_gpu);
980 if (ret)
981 goto out_unlock;
982 ret = ttm_bo_handle_move_mem(bo, &mem, false,
983 interruptible, no_wait_gpu);
984 out_unlock:
985 if (ret && mem.mm_node)
986 ttm_bo_mem_put(bo, &mem);
987 return ret;
988 }
989
990 static int ttm_bo_mem_compat(struct ttm_placement *placement,
991 struct ttm_mem_reg *mem)
992 {
993 int i;
994
995 if (mem->mm_node && placement->lpfn != 0 &&
996 (mem->start < placement->fpfn ||
997 mem->start + mem->num_pages > placement->lpfn))
998 return -1;
999
1000 for (i = 0; i < placement->num_placement; i++) {
1001 if ((placement->placement[i] & mem->placement &
1002 TTM_PL_MASK_CACHING) &&
1003 (placement->placement[i] & mem->placement &
1004 TTM_PL_MASK_MEM))
1005 return i;
1006 }
1007 return -1;
1008 }
1009
1010 int ttm_bo_validate(struct ttm_buffer_object *bo,
1011 struct ttm_placement *placement,
1012 bool interruptible,
1013 bool no_wait_gpu)
1014 {
1015 int ret;
1016
1017 lockdep_assert_held(&bo->resv->lock.base);
1018 /* Check that range is valid */
1019 if (placement->lpfn || placement->fpfn)
1020 if (placement->fpfn > placement->lpfn ||
1021 (placement->lpfn - placement->fpfn) < bo->num_pages)
1022 return -EINVAL;
1023 /*
1024 * Check whether we need to move buffer.
1025 */
1026 ret = ttm_bo_mem_compat(placement, &bo->mem);
1027 if (ret < 0) {
1028 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1029 no_wait_gpu);
1030 if (ret)
1031 return ret;
1032 } else {
1033 /*
1034 * Use the access and other non-mapping-related flag bits from
1035 * the compatible memory placement flags to the active flags
1036 */
1037 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1038 ~TTM_PL_MASK_MEMTYPE);
1039 }
1040 /*
1041 * We might need to add a TTM.
1042 */
1043 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1044 ret = ttm_bo_add_ttm(bo, true);
1045 if (ret)
1046 return ret;
1047 }
1048 return 0;
1049 }
1050 EXPORT_SYMBOL(ttm_bo_validate);
1051
1052 int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1053 struct ttm_placement *placement)
1054 {
1055 BUG_ON((placement->fpfn || placement->lpfn) &&
1056 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1057
1058 return 0;
1059 }
1060
1061 int ttm_bo_init(struct ttm_bo_device *bdev,
1062 struct ttm_buffer_object *bo,
1063 unsigned long size,
1064 enum ttm_bo_type type,
1065 struct ttm_placement *placement,
1066 uint32_t page_alignment,
1067 bool interruptible,
1068 struct file *persistent_swap_storage,
1069 size_t acc_size,
1070 struct sg_table *sg,
1071 void (*destroy) (struct ttm_buffer_object *))
1072 {
1073 int ret = 0;
1074 unsigned long num_pages;
1075 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1076 bool locked;
1077
1078 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1079 if (ret) {
1080 pr_err("Out of kernel memory\n");
1081 if (destroy)
1082 (*destroy)(bo);
1083 else
1084 kfree(bo);
1085 return -ENOMEM;
1086 }
1087
1088 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1089 if (num_pages == 0) {
1090 pr_err("Illegal buffer object size\n");
1091 if (destroy)
1092 (*destroy)(bo);
1093 else
1094 kfree(bo);
1095 ttm_mem_global_free(mem_glob, acc_size);
1096 return -EINVAL;
1097 }
1098 bo->destroy = destroy;
1099
1100 kref_init(&bo->kref);
1101 kref_init(&bo->list_kref);
1102 atomic_set(&bo->cpu_writers, 0);
1103 INIT_LIST_HEAD(&bo->lru);
1104 INIT_LIST_HEAD(&bo->ddestroy);
1105 INIT_LIST_HEAD(&bo->swap);
1106 INIT_LIST_HEAD(&bo->io_reserve_lru);
1107 bo->bdev = bdev;
1108 bo->glob = bdev->glob;
1109 bo->type = type;
1110 bo->num_pages = num_pages;
1111 bo->mem.size = num_pages << PAGE_SHIFT;
1112 bo->mem.mem_type = TTM_PL_SYSTEM;
1113 bo->mem.num_pages = bo->num_pages;
1114 bo->mem.mm_node = NULL;
1115 bo->mem.page_alignment = page_alignment;
1116 bo->mem.bus.io_reserved_vm = false;
1117 bo->mem.bus.io_reserved_count = 0;
1118 bo->priv_flags = 0;
1119 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1120 bo->persistent_swap_storage = persistent_swap_storage;
1121 bo->acc_size = acc_size;
1122 bo->sg = sg;
1123 bo->resv = &bo->ttm_resv;
1124 reservation_object_init(bo->resv);
1125 atomic_inc(&bo->glob->bo_count);
1126 drm_vma_node_reset(&bo->vma_node);
1127
1128 ret = ttm_bo_check_placement(bo, placement);
1129
1130 /*
1131 * For ttm_bo_type_device buffers, allocate
1132 * address space from the device.
1133 */
1134 if (likely(!ret) &&
1135 (bo->type == ttm_bo_type_device ||
1136 bo->type == ttm_bo_type_sg))
1137 ret = ttm_bo_setup_vm(bo);
1138
1139 locked = ww_mutex_trylock(&bo->resv->lock);
1140 WARN_ON(!locked);
1141
1142 if (likely(!ret))
1143 ret = ttm_bo_validate(bo, placement, interruptible, false);
1144
1145 ttm_bo_unreserve(bo);
1146
1147 if (unlikely(ret))
1148 ttm_bo_unref(&bo);
1149
1150 return ret;
1151 }
1152 EXPORT_SYMBOL(ttm_bo_init);
1153
1154 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1155 unsigned long bo_size,
1156 unsigned struct_size)
1157 {
1158 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1159 size_t size = 0;
1160
1161 size += ttm_round_pot(struct_size);
1162 size += PAGE_ALIGN(npages * sizeof(void *));
1163 size += ttm_round_pot(sizeof(struct ttm_tt));
1164 return size;
1165 }
1166 EXPORT_SYMBOL(ttm_bo_acc_size);
1167
1168 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1169 unsigned long bo_size,
1170 unsigned struct_size)
1171 {
1172 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1173 size_t size = 0;
1174
1175 size += ttm_round_pot(struct_size);
1176 size += PAGE_ALIGN(npages * sizeof(void *));
1177 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1178 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1179 return size;
1180 }
1181 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1182
1183 int ttm_bo_create(struct ttm_bo_device *bdev,
1184 unsigned long size,
1185 enum ttm_bo_type type,
1186 struct ttm_placement *placement,
1187 uint32_t page_alignment,
1188 bool interruptible,
1189 struct file *persistent_swap_storage,
1190 struct ttm_buffer_object **p_bo)
1191 {
1192 struct ttm_buffer_object *bo;
1193 size_t acc_size;
1194 int ret;
1195
1196 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1197 if (unlikely(bo == NULL))
1198 return -ENOMEM;
1199
1200 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1201 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1202 interruptible, persistent_swap_storage, acc_size,
1203 NULL, NULL);
1204 if (likely(ret == 0))
1205 *p_bo = bo;
1206
1207 return ret;
1208 }
1209 EXPORT_SYMBOL(ttm_bo_create);
1210
1211 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1212 unsigned mem_type, bool allow_errors)
1213 {
1214 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1215 struct ttm_bo_global *glob = bdev->glob;
1216 int ret;
1217
1218 /*
1219 * Can't use standard list traversal since we're unlocking.
1220 */
1221
1222 spin_lock(&glob->lru_lock);
1223 while (!list_empty(&man->lru)) {
1224 spin_unlock(&glob->lru_lock);
1225 ret = ttm_mem_evict_first(bdev, mem_type, false, false);
1226 if (ret) {
1227 if (allow_errors) {
1228 return ret;
1229 } else {
1230 pr_err("Cleanup eviction failed\n");
1231 }
1232 }
1233 spin_lock(&glob->lru_lock);
1234 }
1235 spin_unlock(&glob->lru_lock);
1236 return 0;
1237 }
1238
1239 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1240 {
1241 struct ttm_mem_type_manager *man;
1242 int ret = -EINVAL;
1243
1244 if (mem_type >= TTM_NUM_MEM_TYPES) {
1245 pr_err("Illegal memory type %d\n", mem_type);
1246 return ret;
1247 }
1248 man = &bdev->man[mem_type];
1249
1250 if (!man->has_type) {
1251 pr_err("Trying to take down uninitialized memory manager type %u\n",
1252 mem_type);
1253 return ret;
1254 }
1255
1256 man->use_type = false;
1257 man->has_type = false;
1258
1259 ret = 0;
1260 if (mem_type > 0) {
1261 ttm_bo_force_list_clean(bdev, mem_type, false);
1262
1263 ret = (*man->func->takedown)(man);
1264 }
1265
1266 return ret;
1267 }
1268 EXPORT_SYMBOL(ttm_bo_clean_mm);
1269
1270 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1271 {
1272 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1273
1274 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1275 pr_err("Illegal memory manager memory type %u\n", mem_type);
1276 return -EINVAL;
1277 }
1278
1279 if (!man->has_type) {
1280 pr_err("Memory type %u has not been initialized\n", mem_type);
1281 return 0;
1282 }
1283
1284 return ttm_bo_force_list_clean(bdev, mem_type, true);
1285 }
1286 EXPORT_SYMBOL(ttm_bo_evict_mm);
1287
1288 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1289 unsigned long p_size)
1290 {
1291 int ret = -EINVAL;
1292 struct ttm_mem_type_manager *man;
1293
1294 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1295 man = &bdev->man[type];
1296 BUG_ON(man->has_type);
1297 man->io_reserve_fastpath = true;
1298 man->use_io_reserve_lru = false;
1299 mutex_init(&man->io_reserve_mutex);
1300 INIT_LIST_HEAD(&man->io_reserve_lru);
1301
1302 ret = bdev->driver->init_mem_type(bdev, type, man);
1303 if (ret)
1304 return ret;
1305 man->bdev = bdev;
1306
1307 ret = 0;
1308 if (type != TTM_PL_SYSTEM) {
1309 ret = (*man->func->init)(man, p_size);
1310 if (ret)
1311 return ret;
1312 }
1313 man->has_type = true;
1314 man->use_type = true;
1315 man->size = p_size;
1316
1317 INIT_LIST_HEAD(&man->lru);
1318
1319 return 0;
1320 }
1321 EXPORT_SYMBOL(ttm_bo_init_mm);
1322
1323 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1324 {
1325 struct ttm_bo_global *glob =
1326 container_of(kobj, struct ttm_bo_global, kobj);
1327
1328 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1329 __free_page(glob->dummy_read_page);
1330 kfree(glob);
1331 }
1332
1333 void ttm_bo_global_release(struct drm_global_reference *ref)
1334 {
1335 struct ttm_bo_global *glob = ref->object;
1336
1337 kobject_del(&glob->kobj);
1338 kobject_put(&glob->kobj);
1339 }
1340 EXPORT_SYMBOL(ttm_bo_global_release);
1341
1342 int ttm_bo_global_init(struct drm_global_reference *ref)
1343 {
1344 struct ttm_bo_global_ref *bo_ref =
1345 container_of(ref, struct ttm_bo_global_ref, ref);
1346 struct ttm_bo_global *glob = ref->object;
1347 int ret;
1348
1349 mutex_init(&glob->device_list_mutex);
1350 spin_lock_init(&glob->lru_lock);
1351 glob->mem_glob = bo_ref->mem_glob;
1352 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1353
1354 if (unlikely(glob->dummy_read_page == NULL)) {
1355 ret = -ENOMEM;
1356 goto out_no_drp;
1357 }
1358
1359 INIT_LIST_HEAD(&glob->swap_lru);
1360 INIT_LIST_HEAD(&glob->device_list);
1361
1362 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1363 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1364 if (unlikely(ret != 0)) {
1365 pr_err("Could not register buffer object swapout\n");
1366 goto out_no_shrink;
1367 }
1368
1369 atomic_set(&glob->bo_count, 0);
1370
1371 ret = kobject_init_and_add(
1372 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1373 if (unlikely(ret != 0))
1374 kobject_put(&glob->kobj);
1375 return ret;
1376 out_no_shrink:
1377 __free_page(glob->dummy_read_page);
1378 out_no_drp:
1379 kfree(glob);
1380 return ret;
1381 }
1382 EXPORT_SYMBOL(ttm_bo_global_init);
1383
1384
1385 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1386 {
1387 int ret = 0;
1388 unsigned i = TTM_NUM_MEM_TYPES;
1389 struct ttm_mem_type_manager *man;
1390 struct ttm_bo_global *glob = bdev->glob;
1391
1392 while (i--) {
1393 man = &bdev->man[i];
1394 if (man->has_type) {
1395 man->use_type = false;
1396 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1397 ret = -EBUSY;
1398 pr_err("DRM memory manager type %d is not clean\n",
1399 i);
1400 }
1401 man->has_type = false;
1402 }
1403 }
1404
1405 mutex_lock(&glob->device_list_mutex);
1406 list_del(&bdev->device_list);
1407 mutex_unlock(&glob->device_list_mutex);
1408
1409 cancel_delayed_work_sync(&bdev->wq);
1410
1411 while (ttm_bo_delayed_delete(bdev, true))
1412 ;
1413
1414 spin_lock(&glob->lru_lock);
1415 if (list_empty(&bdev->ddestroy))
1416 TTM_DEBUG("Delayed destroy list was clean\n");
1417
1418 if (list_empty(&bdev->man[0].lru))
1419 TTM_DEBUG("Swap list was clean\n");
1420 spin_unlock(&glob->lru_lock);
1421
1422 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1423
1424 return ret;
1425 }
1426 EXPORT_SYMBOL(ttm_bo_device_release);
1427
1428 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1429 struct ttm_bo_global *glob,
1430 struct ttm_bo_driver *driver,
1431 uint64_t file_page_offset,
1432 bool need_dma32)
1433 {
1434 int ret = -EINVAL;
1435
1436 bdev->driver = driver;
1437
1438 memset(bdev->man, 0, sizeof(bdev->man));
1439
1440 /*
1441 * Initialize the system memory buffer type.
1442 * Other types need to be driver / IOCTL initialized.
1443 */
1444 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1445 if (unlikely(ret != 0))
1446 goto out_no_sys;
1447
1448 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1449 0x10000000);
1450 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1451 INIT_LIST_HEAD(&bdev->ddestroy);
1452 bdev->dev_mapping = NULL;
1453 bdev->glob = glob;
1454 bdev->need_dma32 = need_dma32;
1455 bdev->val_seq = 0;
1456 spin_lock_init(&bdev->fence_lock);
1457 mutex_lock(&glob->device_list_mutex);
1458 list_add_tail(&bdev->device_list, &glob->device_list);
1459 mutex_unlock(&glob->device_list_mutex);
1460
1461 return 0;
1462 out_no_sys:
1463 return ret;
1464 }
1465 EXPORT_SYMBOL(ttm_bo_device_init);
1466
1467 /*
1468 * buffer object vm functions.
1469 */
1470
1471 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1472 {
1473 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1474
1475 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1476 if (mem->mem_type == TTM_PL_SYSTEM)
1477 return false;
1478
1479 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1480 return false;
1481
1482 if (mem->placement & TTM_PL_FLAG_CACHED)
1483 return false;
1484 }
1485 return true;
1486 }
1487
1488 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1489 {
1490 struct ttm_bo_device *bdev = bo->bdev;
1491 loff_t offset, holelen;
1492
1493 if (!bdev->dev_mapping)
1494 return;
1495
1496 if (drm_vma_node_has_offset(&bo->vma_node)) {
1497 offset = (loff_t) drm_vma_node_offset_addr(&bo->vma_node);
1498 holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1499
1500 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1501 }
1502 ttm_mem_io_free_vm(bo);
1503 }
1504
1505 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1506 {
1507 struct ttm_bo_device *bdev = bo->bdev;
1508 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1509
1510 ttm_mem_io_lock(man, false);
1511 ttm_bo_unmap_virtual_locked(bo);
1512 ttm_mem_io_unlock(man);
1513 }
1514
1515
1516 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1517
1518 /**
1519 * ttm_bo_setup_vm:
1520 *
1521 * @bo: the buffer to allocate address space for
1522 *
1523 * Allocate address space in the drm device so that applications
1524 * can mmap the buffer and access the contents. This only
1525 * applies to ttm_bo_type_device objects as others are not
1526 * placed in the drm device address space.
1527 */
1528
1529 static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1530 {
1531 struct ttm_bo_device *bdev = bo->bdev;
1532
1533 return drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1534 bo->mem.num_pages);
1535 }
1536
1537 int ttm_bo_wait(struct ttm_buffer_object *bo,
1538 bool lazy, bool interruptible, bool no_wait)
1539 {
1540 struct ttm_bo_driver *driver = bo->bdev->driver;
1541 struct ttm_bo_device *bdev = bo->bdev;
1542 void *sync_obj;
1543 int ret = 0;
1544
1545 if (likely(bo->sync_obj == NULL))
1546 return 0;
1547
1548 while (bo->sync_obj) {
1549
1550 if (driver->sync_obj_signaled(bo->sync_obj)) {
1551 void *tmp_obj = bo->sync_obj;
1552 bo->sync_obj = NULL;
1553 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1554 spin_unlock(&bdev->fence_lock);
1555 driver->sync_obj_unref(&tmp_obj);
1556 spin_lock(&bdev->fence_lock);
1557 continue;
1558 }
1559
1560 if (no_wait)
1561 return -EBUSY;
1562
1563 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1564 spin_unlock(&bdev->fence_lock);
1565 ret = driver->sync_obj_wait(sync_obj,
1566 lazy, interruptible);
1567 if (unlikely(ret != 0)) {
1568 driver->sync_obj_unref(&sync_obj);
1569 spin_lock(&bdev->fence_lock);
1570 return ret;
1571 }
1572 spin_lock(&bdev->fence_lock);
1573 if (likely(bo->sync_obj == sync_obj)) {
1574 void *tmp_obj = bo->sync_obj;
1575 bo->sync_obj = NULL;
1576 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1577 &bo->priv_flags);
1578 spin_unlock(&bdev->fence_lock);
1579 driver->sync_obj_unref(&sync_obj);
1580 driver->sync_obj_unref(&tmp_obj);
1581 spin_lock(&bdev->fence_lock);
1582 } else {
1583 spin_unlock(&bdev->fence_lock);
1584 driver->sync_obj_unref(&sync_obj);
1585 spin_lock(&bdev->fence_lock);
1586 }
1587 }
1588 return 0;
1589 }
1590 EXPORT_SYMBOL(ttm_bo_wait);
1591
1592 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1593 {
1594 struct ttm_bo_device *bdev = bo->bdev;
1595 int ret = 0;
1596
1597 /*
1598 * Using ttm_bo_reserve makes sure the lru lists are updated.
1599 */
1600
1601 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1602 if (unlikely(ret != 0))
1603 return ret;
1604 spin_lock(&bdev->fence_lock);
1605 ret = ttm_bo_wait(bo, false, true, no_wait);
1606 spin_unlock(&bdev->fence_lock);
1607 if (likely(ret == 0))
1608 atomic_inc(&bo->cpu_writers);
1609 ttm_bo_unreserve(bo);
1610 return ret;
1611 }
1612 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1613
1614 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1615 {
1616 atomic_dec(&bo->cpu_writers);
1617 }
1618 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1619
1620 /**
1621 * A buffer object shrink method that tries to swap out the first
1622 * buffer object on the bo_global::swap_lru list.
1623 */
1624
1625 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1626 {
1627 struct ttm_bo_global *glob =
1628 container_of(shrink, struct ttm_bo_global, shrink);
1629 struct ttm_buffer_object *bo;
1630 int ret = -EBUSY;
1631 int put_count;
1632 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1633
1634 spin_lock(&glob->lru_lock);
1635 list_for_each_entry(bo, &glob->swap_lru, swap) {
1636 ret = ttm_bo_reserve_nolru(bo, false, true, false, 0);
1637 if (!ret)
1638 break;
1639 }
1640
1641 if (ret) {
1642 spin_unlock(&glob->lru_lock);
1643 return ret;
1644 }
1645
1646 kref_get(&bo->list_kref);
1647
1648 if (!list_empty(&bo->ddestroy)) {
1649 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1650 kref_put(&bo->list_kref, ttm_bo_release_list);
1651 return ret;
1652 }
1653
1654 put_count = ttm_bo_del_from_lru(bo);
1655 spin_unlock(&glob->lru_lock);
1656
1657 ttm_bo_list_ref_sub(bo, put_count, true);
1658
1659 /**
1660 * Wait for GPU, then move to system cached.
1661 */
1662
1663 spin_lock(&bo->bdev->fence_lock);
1664 ret = ttm_bo_wait(bo, false, false, false);
1665 spin_unlock(&bo->bdev->fence_lock);
1666
1667 if (unlikely(ret != 0))
1668 goto out;
1669
1670 if ((bo->mem.placement & swap_placement) != swap_placement) {
1671 struct ttm_mem_reg evict_mem;
1672
1673 evict_mem = bo->mem;
1674 evict_mem.mm_node = NULL;
1675 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1676 evict_mem.mem_type = TTM_PL_SYSTEM;
1677
1678 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1679 false, false);
1680 if (unlikely(ret != 0))
1681 goto out;
1682 }
1683
1684 ttm_bo_unmap_virtual(bo);
1685
1686 /**
1687 * Swap out. Buffer will be swapped in again as soon as
1688 * anyone tries to access a ttm page.
1689 */
1690
1691 if (bo->bdev->driver->swap_notify)
1692 bo->bdev->driver->swap_notify(bo);
1693
1694 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1695 out:
1696
1697 /**
1698 *
1699 * Unreserve without putting on LRU to avoid swapping out an
1700 * already swapped buffer.
1701 */
1702
1703 ww_mutex_unlock(&bo->resv->lock);
1704 kref_put(&bo->list_kref, ttm_bo_release_list);
1705 return ret;
1706 }
1707
1708 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1709 {
1710 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1711 ;
1712 }
1713 EXPORT_SYMBOL(ttm_bo_swapout_all);
Linux kernel - Deliverables
This page took 0.114904 seconds and 5 git commands to generate.