1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
3 *
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28 /*
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30 */
31
32 #define pr_fmt(fmt) "[TTM] " fmt
33
34 #include <drm/ttm/ttm_module.h>
35 #include <drm/ttm/ttm_bo_driver.h>
36 #include <drm/ttm/ttm_placement.h>
37 #include <linux/jiffies.h>
38 #include <linux/slab.h>
39 #include <linux/sched.h>
40 #include <linux/mm.h>
41 #include <linux/file.h>
42 #include <linux/module.h>
43 #include <linux/atomic.h>
44 #include <linux/dma-resv.h>
45
46 static void ttm_bo_global_kobj_release(struct kobject *kobj);
47
48 /**
49 * ttm_global_mutex - protecting the global BO state
50 */
51 DEFINE_MUTEX(ttm_global_mutex);
52 unsigned ttm_bo_glob_use_count;
53 struct ttm_bo_global ttm_bo_glob;
54 EXPORT_SYMBOL(ttm_bo_glob);
55
56 static struct attribute ttm_bo_count = {
57 .name = "bo_count",
58 .mode = S_IRUGO
59 };
60
61 /* default destructor */
ttm_bo_default_destroy(struct ttm_buffer_object * bo)62 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
63 {
64 kfree(bo);
65 }
66
ttm_bo_mem_space_debug(struct ttm_buffer_object * bo,struct ttm_placement * placement)67 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
68 struct ttm_placement *placement)
69 {
70 struct drm_printer p = drm_debug_printer(TTM_PFX);
71 struct ttm_resource_manager *man;
72 int i, mem_type;
73
74 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n",
75 bo, bo->mem.num_pages, bo->mem.size >> 10,
76 bo->mem.size >> 20);
77 for (i = 0; i < placement->num_placement; i++) {
78 mem_type = placement->placement[i].mem_type;
79 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
80 i, placement->placement[i].flags, mem_type);
81 man = ttm_manager_type(bo->bdev, mem_type);
82 ttm_resource_manager_debug(man, &p);
83 }
84 }
85
ttm_bo_global_show(struct kobject * kobj,struct attribute * attr,char * buffer)86 static ssize_t ttm_bo_global_show(struct kobject *kobj,
87 struct attribute *attr,
88 char *buffer)
89 {
90 struct ttm_bo_global *glob =
91 container_of(kobj, struct ttm_bo_global, kobj);
92
93 return snprintf(buffer, PAGE_SIZE, "%d\n",
94 atomic_read(&glob->bo_count));
95 }
96
97 static struct attribute *ttm_bo_global_attrs[] = {
98 &ttm_bo_count,
99 NULL
100 };
101
102 static const struct sysfs_ops ttm_bo_global_ops = {
103 .show = &ttm_bo_global_show
104 };
105
106 static struct kobj_type ttm_bo_glob_kobj_type = {
107 .release = &ttm_bo_global_kobj_release,
108 .sysfs_ops = &ttm_bo_global_ops,
109 .default_attrs = ttm_bo_global_attrs
110 };
111
ttm_bo_add_mem_to_lru(struct ttm_buffer_object * bo,struct ttm_resource * mem)112 static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo,
113 struct ttm_resource *mem)
114 {
115 struct ttm_bo_device *bdev = bo->bdev;
116 struct ttm_resource_manager *man;
117
118 if (!list_empty(&bo->lru) || bo->pin_count)
119 return;
120
121 if (mem->placement & TTM_PL_FLAG_NO_EVICT)
122 return;
123
124 man = ttm_manager_type(bdev, mem->mem_type);
125 list_add_tail(&bo->lru, &man->lru[bo->priority]);
126
127 if (man->use_tt && bo->ttm &&
128 !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG |
129 TTM_PAGE_FLAG_SWAPPED))) {
130 list_add_tail(&bo->swap, &ttm_bo_glob.swap_lru[bo->priority]);
131 }
132 }
133
ttm_bo_del_from_lru(struct ttm_buffer_object * bo)134 static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
135 {
136 struct ttm_bo_device *bdev = bo->bdev;
137 bool notify = false;
138
139 if (!list_empty(&bo->swap)) {
140 list_del_init(&bo->swap);
141 notify = true;
142 }
143 if (!list_empty(&bo->lru)) {
144 list_del_init(&bo->lru);
145 notify = true;
146 }
147
148 if (notify && bdev->driver->del_from_lru_notify)
149 bdev->driver->del_from_lru_notify(bo);
150 }
151
ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos * pos,struct ttm_buffer_object * bo)152 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
153 struct ttm_buffer_object *bo)
154 {
155 if (!pos->first)
156 pos->first = bo;
157 pos->last = bo;
158 }
159
ttm_bo_move_to_lru_tail(struct ttm_buffer_object * bo,struct ttm_lru_bulk_move * bulk)160 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
161 struct ttm_lru_bulk_move *bulk)
162 {
163 dma_resv_assert_held(bo->base.resv);
164
165 ttm_bo_del_from_lru(bo);
166 ttm_bo_add_mem_to_lru(bo, &bo->mem);
167
168 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT) &&
169 !bo->pin_count) {
170 switch (bo->mem.mem_type) {
171 case TTM_PL_TT:
172 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
173 break;
174
175 case TTM_PL_VRAM:
176 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
177 break;
178 }
179 if (bo->ttm && !(bo->ttm->page_flags &
180 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
181 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
182 }
183 }
184 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
185
ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move * bulk)186 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
187 {
188 unsigned i;
189
190 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
191 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
192 struct ttm_resource_manager *man;
193
194 if (!pos->first)
195 continue;
196
197 dma_resv_assert_held(pos->first->base.resv);
198 dma_resv_assert_held(pos->last->base.resv);
199
200 man = ttm_manager_type(pos->first->bdev, TTM_PL_TT);
201 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
202 &pos->last->lru);
203 }
204
205 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
206 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
207 struct ttm_resource_manager *man;
208
209 if (!pos->first)
210 continue;
211
212 dma_resv_assert_held(pos->first->base.resv);
213 dma_resv_assert_held(pos->last->base.resv);
214
215 man = ttm_manager_type(pos->first->bdev, TTM_PL_VRAM);
216 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
217 &pos->last->lru);
218 }
219
220 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
221 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
222 struct list_head *lru;
223
224 if (!pos->first)
225 continue;
226
227 dma_resv_assert_held(pos->first->base.resv);
228 dma_resv_assert_held(pos->last->base.resv);
229
230 lru = &ttm_bo_glob.swap_lru[i];
231 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
232 }
233 }
234 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
235
ttm_bo_handle_move_mem(struct ttm_buffer_object * bo,struct ttm_resource * mem,bool evict,struct ttm_operation_ctx * ctx)236 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
237 struct ttm_resource *mem, bool evict,
238 struct ttm_operation_ctx *ctx)
239 {
240 struct ttm_bo_device *bdev = bo->bdev;
241 struct ttm_resource_manager *old_man = ttm_manager_type(bdev, bo->mem.mem_type);
242 struct ttm_resource_manager *new_man = ttm_manager_type(bdev, mem->mem_type);
243 int ret;
244
245 ttm_bo_unmap_virtual(bo);
246
247 /*
248 * Create and bind a ttm if required.
249 */
250
251 if (new_man->use_tt) {
252 /* Zero init the new TTM structure if the old location should
253 * have used one as well.
254 */
255 ret = ttm_tt_create(bo, old_man->use_tt);
256 if (ret)
257 goto out_err;
258
259 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
260 if (ret)
261 goto out_err;
262
263 if (mem->mem_type != TTM_PL_SYSTEM) {
264 ret = ttm_tt_populate(bdev, bo->ttm, ctx);
265 if (ret)
266 goto out_err;
267
268 ret = ttm_bo_tt_bind(bo, mem);
269 if (ret)
270 goto out_err;
271 }
272
273 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
274 if (bdev->driver->move_notify)
275 bdev->driver->move_notify(bo, evict, mem);
276 bo->mem = *mem;
277 goto moved;
278 }
279 }
280
281 if (bdev->driver->move_notify)
282 bdev->driver->move_notify(bo, evict, mem);
283
284 if (old_man->use_tt && new_man->use_tt)
285 ret = ttm_bo_move_ttm(bo, ctx, mem);
286 else if (bdev->driver->move)
287 ret = bdev->driver->move(bo, evict, ctx, mem);
288 else
289 ret = ttm_bo_move_memcpy(bo, ctx, mem);
290
291 if (ret) {
292 if (bdev->driver->move_notify) {
293 swap(*mem, bo->mem);
294 bdev->driver->move_notify(bo, false, mem);
295 swap(*mem, bo->mem);
296 }
297
298 goto out_err;
299 }
300
301 moved:
302 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
303 return 0;
304
305 out_err:
306 new_man = ttm_manager_type(bdev, bo->mem.mem_type);
307 if (!new_man->use_tt)
308 ttm_bo_tt_destroy(bo);
309
310 return ret;
311 }
312
313 /**
314 * Call bo::reserved.
315 * Will release GPU memory type usage on destruction.
316 * This is the place to put in driver specific hooks to release
317 * driver private resources.
318 * Will release the bo::reserved lock.
319 */
320
ttm_bo_cleanup_memtype_use(struct ttm_buffer_object * bo)321 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
322 {
323 if (bo->bdev->driver->move_notify)
324 bo->bdev->driver->move_notify(bo, false, NULL);
325
326 ttm_bo_tt_destroy(bo);
327 ttm_resource_free(bo, &bo->mem);
328 }
329
ttm_bo_individualize_resv(struct ttm_buffer_object * bo)330 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
331 {
332 int r;
333
334 if (bo->base.resv == &bo->base._resv)
335 return 0;
336
337 BUG_ON(!dma_resv_trylock(&bo->base._resv));
338
339 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
340 dma_resv_unlock(&bo->base._resv);
341 if (r)
342 return r;
343
344 if (bo->type != ttm_bo_type_sg) {
345 /* This works because the BO is about to be destroyed and nobody
346 * reference it any more. The only tricky case is the trylock on
347 * the resv object while holding the lru_lock.
348 */
349 spin_lock(&ttm_bo_glob.lru_lock);
350 bo->base.resv = &bo->base._resv;
351 spin_unlock(&ttm_bo_glob.lru_lock);
352 }
353
354 return r;
355 }
356
ttm_bo_flush_all_fences(struct ttm_buffer_object * bo)357 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
358 {
359 struct dma_resv *resv = &bo->base._resv;
360 struct dma_resv_list *fobj;
361 struct dma_fence *fence;
362 int i;
363
364 rcu_read_lock();
365 fobj = rcu_dereference(resv->fence);
366 fence = rcu_dereference(resv->fence_excl);
367 if (fence && !fence->ops->signaled)
368 dma_fence_enable_sw_signaling(fence);
369
370 for (i = 0; fobj && i < fobj->shared_count; ++i) {
371 fence = rcu_dereference(fobj->shared[i]);
372
373 if (!fence->ops->signaled)
374 dma_fence_enable_sw_signaling(fence);
375 }
376 rcu_read_unlock();
377 }
378
379 /**
380 * function ttm_bo_cleanup_refs
381 * If bo idle, remove from lru lists, and unref.
382 * If not idle, block if possible.
383 *
384 * Must be called with lru_lock and reservation held, this function
385 * will drop the lru lock and optionally the reservation lock before returning.
386 *
387 * @interruptible Any sleeps should occur interruptibly.
388 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
389 * @unlock_resv Unlock the reservation lock as well.
390 */
391
ttm_bo_cleanup_refs(struct ttm_buffer_object * bo,bool interruptible,bool no_wait_gpu,bool unlock_resv)392 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
393 bool interruptible, bool no_wait_gpu,
394 bool unlock_resv)
395 {
396 struct dma_resv *resv = &bo->base._resv;
397 int ret;
398
399 if (dma_resv_test_signaled_rcu(resv, true))
400 ret = 0;
401 else
402 ret = -EBUSY;
403
404 if (ret && !no_wait_gpu) {
405 long lret;
406
407 if (unlock_resv)
408 dma_resv_unlock(bo->base.resv);
409 spin_unlock(&ttm_bo_glob.lru_lock);
410
411 lret = dma_resv_wait_timeout_rcu(resv, true, interruptible,
412 30 * HZ);
413
414 if (lret < 0)
415 return lret;
416 else if (lret == 0)
417 return -EBUSY;
418
419 spin_lock(&ttm_bo_glob.lru_lock);
420 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
421 /*
422 * We raced, and lost, someone else holds the reservation now,
423 * and is probably busy in ttm_bo_cleanup_memtype_use.
424 *
425 * Even if it's not the case, because we finished waiting any
426 * delayed destruction would succeed, so just return success
427 * here.
428 */
429 spin_unlock(&ttm_bo_glob.lru_lock);
430 return 0;
431 }
432 ret = 0;
433 }
434
435 if (ret || unlikely(list_empty(&bo->ddestroy))) {
436 if (unlock_resv)
437 dma_resv_unlock(bo->base.resv);
438 spin_unlock(&ttm_bo_glob.lru_lock);
439 return ret;
440 }
441
442 ttm_bo_del_from_lru(bo);
443 list_del_init(&bo->ddestroy);
444 spin_unlock(&ttm_bo_glob.lru_lock);
445 ttm_bo_cleanup_memtype_use(bo);
446
447 if (unlock_resv)
448 dma_resv_unlock(bo->base.resv);
449
450 ttm_bo_put(bo);
451
452 return 0;
453 }
454
455 /**
456 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
457 * encountered buffers.
458 */
ttm_bo_delayed_delete(struct ttm_bo_device * bdev,bool remove_all)459 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
460 {
461 struct ttm_bo_global *glob = &ttm_bo_glob;
462 struct list_head removed;
463 bool empty;
464
465 INIT_LIST_HEAD(&removed);
466
467 spin_lock(&glob->lru_lock);
468 while (!list_empty(&bdev->ddestroy)) {
469 struct ttm_buffer_object *bo;
470
471 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
472 ddestroy);
473 list_move_tail(&bo->ddestroy, &removed);
474 if (!ttm_bo_get_unless_zero(bo))
475 continue;
476
477 if (remove_all || bo->base.resv != &bo->base._resv) {
478 spin_unlock(&glob->lru_lock);
479 dma_resv_lock(bo->base.resv, NULL);
480
481 spin_lock(&glob->lru_lock);
482 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
483
484 } else if (dma_resv_trylock(bo->base.resv)) {
485 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
486 } else {
487 spin_unlock(&glob->lru_lock);
488 }
489
490 ttm_bo_put(bo);
491 spin_lock(&glob->lru_lock);
492 }
493 list_splice_tail(&removed, &bdev->ddestroy);
494 empty = list_empty(&bdev->ddestroy);
495 spin_unlock(&glob->lru_lock);
496
497 return empty;
498 }
499
ttm_bo_delayed_workqueue(struct work_struct * work)500 static void ttm_bo_delayed_workqueue(struct work_struct *work)
501 {
502 struct ttm_bo_device *bdev =
503 container_of(work, struct ttm_bo_device, wq.work);
504
505 if (!ttm_bo_delayed_delete(bdev, false))
506 schedule_delayed_work(&bdev->wq,
507 ((HZ / 100) < 1) ? 1 : HZ / 100);
508 }
509
ttm_bo_release(struct kref * kref)510 static void ttm_bo_release(struct kref *kref)
511 {
512 struct ttm_buffer_object *bo =
513 container_of(kref, struct ttm_buffer_object, kref);
514 struct ttm_bo_device *bdev = bo->bdev;
515 size_t acc_size = bo->acc_size;
516 int ret;
517
518 if (!bo->deleted) {
519 ret = ttm_bo_individualize_resv(bo);
520 if (ret) {
521 /* Last resort, if we fail to allocate memory for the
522 * fences block for the BO to become idle
523 */
524 dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
525 30 * HZ);
526 }
527
528 if (bo->bdev->driver->release_notify)
529 bo->bdev->driver->release_notify(bo);
530
531 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
532 ttm_mem_io_free(bdev, &bo->mem);
533 }
534
535 if (!dma_resv_test_signaled_rcu(bo->base.resv, true) ||
536 !dma_resv_trylock(bo->base.resv)) {
537 /* The BO is not idle, resurrect it for delayed destroy */
538 ttm_bo_flush_all_fences(bo);
539 bo->deleted = true;
540
541 spin_lock(&ttm_bo_glob.lru_lock);
542
543 /*
544 * Make NO_EVICT bos immediately available to
545 * shrinkers, now that they are queued for
546 * destruction.
547 */
548 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT || bo->pin_count) {
549 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
550 bo->pin_count = 0;
551 ttm_bo_del_from_lru(bo);
552 ttm_bo_add_mem_to_lru(bo, &bo->mem);
553 }
554
555 kref_init(&bo->kref);
556 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
557 spin_unlock(&ttm_bo_glob.lru_lock);
558
559 schedule_delayed_work(&bdev->wq,
560 ((HZ / 100) < 1) ? 1 : HZ / 100);
561 return;
562 }
563
564 spin_lock(&ttm_bo_glob.lru_lock);
565 ttm_bo_del_from_lru(bo);
566 list_del(&bo->ddestroy);
567 spin_unlock(&ttm_bo_glob.lru_lock);
568
569 ttm_bo_cleanup_memtype_use(bo);
570 dma_resv_unlock(bo->base.resv);
571
572 atomic_dec(&ttm_bo_glob.bo_count);
573 dma_fence_put(bo->moving);
574 if (!ttm_bo_uses_embedded_gem_object(bo))
575 dma_resv_fini(&bo->base._resv);
576 bo->destroy(bo);
577 ttm_mem_global_free(&ttm_mem_glob, acc_size);
578 }
579
ttm_bo_put(struct ttm_buffer_object * bo)580 void ttm_bo_put(struct ttm_buffer_object *bo)
581 {
582 kref_put(&bo->kref, ttm_bo_release);
583 }
584 EXPORT_SYMBOL(ttm_bo_put);
585
ttm_bo_lock_delayed_workqueue(struct ttm_bo_device * bdev)586 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
587 {
588 return cancel_delayed_work_sync(&bdev->wq);
589 }
590 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
591
ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device * bdev,int resched)592 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
593 {
594 if (resched)
595 schedule_delayed_work(&bdev->wq,
596 ((HZ / 100) < 1) ? 1 : HZ / 100);
597 }
598 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
599
ttm_bo_evict(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)600 static int ttm_bo_evict(struct ttm_buffer_object *bo,
601 struct ttm_operation_ctx *ctx)
602 {
603 struct ttm_bo_device *bdev = bo->bdev;
604 struct ttm_resource evict_mem;
605 struct ttm_placement placement;
606 int ret = 0;
607
608 dma_resv_assert_held(bo->base.resv);
609
610 placement.num_placement = 0;
611 placement.num_busy_placement = 0;
612 bdev->driver->evict_flags(bo, &placement);
613
614 if (!placement.num_placement && !placement.num_busy_placement) {
615 ttm_bo_wait(bo, false, false);
616
617 ttm_bo_cleanup_memtype_use(bo);
618 return ttm_tt_create(bo, false);
619 }
620
621 evict_mem = bo->mem;
622 evict_mem.mm_node = NULL;
623 evict_mem.bus.offset = 0;
624 evict_mem.bus.addr = NULL;
625
626 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
627 if (ret) {
628 if (ret != -ERESTARTSYS) {
629 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
630 bo);
631 ttm_bo_mem_space_debug(bo, &placement);
632 }
633 goto out;
634 }
635
636 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
637 if (unlikely(ret)) {
638 if (ret != -ERESTARTSYS)
639 pr_err("Buffer eviction failed\n");
640 ttm_resource_free(bo, &evict_mem);
641 }
642 out:
643 return ret;
644 }
645
ttm_bo_eviction_valuable(struct ttm_buffer_object * bo,const struct ttm_place * place)646 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
647 const struct ttm_place *place)
648 {
649 /* Don't evict this BO if it's outside of the
650 * requested placement range
651 */
652 if (place->fpfn >= (bo->mem.start + bo->mem.num_pages) ||
653 (place->lpfn && place->lpfn <= bo->mem.start))
654 return false;
655
656 return true;
657 }
658 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
659
660 /**
661 * Check the target bo is allowable to be evicted or swapout, including cases:
662 *
663 * a. if share same reservation object with ctx->resv, have assumption
664 * reservation objects should already be locked, so not lock again and
665 * return true directly when either the opreation allow_reserved_eviction
666 * or the target bo already is in delayed free list;
667 *
668 * b. Otherwise, trylock it.
669 */
ttm_bo_evict_swapout_allowable(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,bool * locked,bool * busy)670 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
671 struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
672 {
673 bool ret = false;
674
675 if (bo->pin_count) {
676 *locked = false;
677 if (busy)
678 *busy = false;
679 return false;
680 }
681
682 if (bo->base.resv == ctx->resv) {
683 dma_resv_assert_held(bo->base.resv);
684 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT)
685 ret = true;
686 *locked = false;
687 if (busy)
688 *busy = false;
689 } else {
690 ret = dma_resv_trylock(bo->base.resv);
691 *locked = ret;
692 if (busy)
693 *busy = !ret;
694 }
695
696 return ret;
697 }
698
699 /**
700 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
701 *
702 * @busy_bo: BO which couldn't be locked with trylock
703 * @ctx: operation context
704 * @ticket: acquire ticket
705 *
706 * Try to lock a busy buffer object to avoid failing eviction.
707 */
ttm_mem_evict_wait_busy(struct ttm_buffer_object * busy_bo,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)708 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
709 struct ttm_operation_ctx *ctx,
710 struct ww_acquire_ctx *ticket)
711 {
712 int r;
713
714 if (!busy_bo || !ticket)
715 return -EBUSY;
716
717 if (ctx->interruptible)
718 r = dma_resv_lock_interruptible(busy_bo->base.resv,
719 ticket);
720 else
721 r = dma_resv_lock(busy_bo->base.resv, ticket);
722
723 /*
724 * TODO: It would be better to keep the BO locked until allocation is at
725 * least tried one more time, but that would mean a much larger rework
726 * of TTM.
727 */
728 if (!r)
729 dma_resv_unlock(busy_bo->base.resv);
730
731 return r == -EDEADLK ? -EBUSY : r;
732 }
733
ttm_mem_evict_first(struct ttm_bo_device * bdev,struct ttm_resource_manager * man,const struct ttm_place * place,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)734 int ttm_mem_evict_first(struct ttm_bo_device *bdev,
735 struct ttm_resource_manager *man,
736 const struct ttm_place *place,
737 struct ttm_operation_ctx *ctx,
738 struct ww_acquire_ctx *ticket)
739 {
740 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
741 bool locked = false;
742 unsigned i;
743 int ret;
744
745 spin_lock(&ttm_bo_glob.lru_lock);
746 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
747 list_for_each_entry(bo, &man->lru[i], lru) {
748 bool busy;
749
750 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
751 &busy)) {
752 if (busy && !busy_bo && ticket !=
753 dma_resv_locking_ctx(bo->base.resv))
754 busy_bo = bo;
755 continue;
756 }
757
758 if (place && !bdev->driver->eviction_valuable(bo,
759 place)) {
760 if (locked)
761 dma_resv_unlock(bo->base.resv);
762 continue;
763 }
764 if (!ttm_bo_get_unless_zero(bo)) {
765 if (locked)
766 dma_resv_unlock(bo->base.resv);
767 continue;
768 }
769 break;
770 }
771
772 /* If the inner loop terminated early, we have our candidate */
773 if (&bo->lru != &man->lru[i])
774 break;
775
776 bo = NULL;
777 }
778
779 if (!bo) {
780 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
781 busy_bo = NULL;
782 spin_unlock(&ttm_bo_glob.lru_lock);
783 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
784 if (busy_bo)
785 ttm_bo_put(busy_bo);
786 return ret;
787 }
788
789 if (bo->deleted) {
790 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
791 ctx->no_wait_gpu, locked);
792 ttm_bo_put(bo);
793 return ret;
794 }
795
796 spin_unlock(&ttm_bo_glob.lru_lock);
797
798 ret = ttm_bo_evict(bo, ctx);
799 if (locked)
800 ttm_bo_unreserve(bo);
801 else
802 ttm_bo_move_to_lru_tail_unlocked(bo);
803
804 ttm_bo_put(bo);
805 return ret;
806 }
807
808 /**
809 * Add the last move fence to the BO and reserve a new shared slot.
810 */
ttm_bo_add_move_fence(struct ttm_buffer_object * bo,struct ttm_resource_manager * man,struct ttm_resource * mem,bool no_wait_gpu)811 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
812 struct ttm_resource_manager *man,
813 struct ttm_resource *mem,
814 bool no_wait_gpu)
815 {
816 struct dma_fence *fence;
817 int ret;
818
819 spin_lock(&man->move_lock);
820 fence = dma_fence_get(man->move);
821 spin_unlock(&man->move_lock);
822
823 if (!fence)
824 return 0;
825
826 if (no_wait_gpu) {
827 dma_fence_put(fence);
828 return -EBUSY;
829 }
830
831 dma_resv_add_shared_fence(bo->base.resv, fence);
832
833 ret = dma_resv_reserve_shared(bo->base.resv, 1);
834 if (unlikely(ret)) {
835 dma_fence_put(fence);
836 return ret;
837 }
838
839 dma_fence_put(bo->moving);
840 bo->moving = fence;
841 return 0;
842 }
843
844 /**
845 * Repeatedly evict memory from the LRU for @mem_type until we create enough
846 * space, or we've evicted everything and there isn't enough space.
847 */
ttm_bo_mem_force_space(struct ttm_buffer_object * bo,const struct ttm_place * place,struct ttm_resource * mem,struct ttm_operation_ctx * ctx)848 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
849 const struct ttm_place *place,
850 struct ttm_resource *mem,
851 struct ttm_operation_ctx *ctx)
852 {
853 struct ttm_bo_device *bdev = bo->bdev;
854 struct ttm_resource_manager *man = ttm_manager_type(bdev, mem->mem_type);
855 struct ww_acquire_ctx *ticket;
856 int ret;
857
858 ticket = dma_resv_locking_ctx(bo->base.resv);
859 do {
860 ret = ttm_resource_alloc(bo, place, mem);
861 if (likely(!ret))
862 break;
863 if (unlikely(ret != -ENOSPC))
864 return ret;
865 ret = ttm_mem_evict_first(bdev, man, place, ctx,
866 ticket);
867 if (unlikely(ret != 0))
868 return ret;
869 } while (1);
870
871 return ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
872 }
873
ttm_bo_select_caching(struct ttm_resource_manager * man,uint32_t cur_placement,uint32_t proposed_placement)874 static uint32_t ttm_bo_select_caching(struct ttm_resource_manager *man,
875 uint32_t cur_placement,
876 uint32_t proposed_placement)
877 {
878 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
879 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
880
881 /**
882 * Keep current caching if possible.
883 */
884
885 if ((cur_placement & caching) != 0)
886 result |= (cur_placement & caching);
887 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
888 result |= TTM_PL_FLAG_CACHED;
889 else if ((TTM_PL_FLAG_WC & caching) != 0)
890 result |= TTM_PL_FLAG_WC;
891 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
892 result |= TTM_PL_FLAG_UNCACHED;
893
894 return result;
895 }
896
897 /**
898 * ttm_bo_mem_placement - check if placement is compatible
899 * @bo: BO to find memory for
900 * @place: where to search
901 * @mem: the memory object to fill in
902 * @ctx: operation context
903 *
904 * Check if placement is compatible and fill in mem structure.
905 * Returns -EBUSY if placement won't work or negative error code.
906 * 0 when placement can be used.
907 */
ttm_bo_mem_placement(struct ttm_buffer_object * bo,const struct ttm_place * place,struct ttm_resource * mem,struct ttm_operation_ctx * ctx)908 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo,
909 const struct ttm_place *place,
910 struct ttm_resource *mem,
911 struct ttm_operation_ctx *ctx)
912 {
913 struct ttm_bo_device *bdev = bo->bdev;
914 struct ttm_resource_manager *man;
915 uint32_t cur_flags = 0;
916
917 man = ttm_manager_type(bdev, place->mem_type);
918 if (!man || !ttm_resource_manager_used(man))
919 return -EBUSY;
920
921 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
922 place->flags);
923 cur_flags |= place->flags & ~TTM_PL_MASK_CACHING;
924
925 mem->mem_type = place->mem_type;
926 mem->placement = cur_flags;
927
928 spin_lock(&ttm_bo_glob.lru_lock);
929 ttm_bo_del_from_lru(bo);
930 ttm_bo_add_mem_to_lru(bo, mem);
931 spin_unlock(&ttm_bo_glob.lru_lock);
932
933 return 0;
934 }
935
936 /**
937 * Creates space for memory region @mem according to its type.
938 *
939 * This function first searches for free space in compatible memory types in
940 * the priority order defined by the driver. If free space isn't found, then
941 * ttm_bo_mem_force_space is attempted in priority order to evict and find
942 * space.
943 */
ttm_bo_mem_space(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_resource * mem,struct ttm_operation_ctx * ctx)944 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
945 struct ttm_placement *placement,
946 struct ttm_resource *mem,
947 struct ttm_operation_ctx *ctx)
948 {
949 struct ttm_bo_device *bdev = bo->bdev;
950 bool type_found = false;
951 int i, ret;
952
953 ret = dma_resv_reserve_shared(bo->base.resv, 1);
954 if (unlikely(ret))
955 return ret;
956
957 for (i = 0; i < placement->num_placement; ++i) {
958 const struct ttm_place *place = &placement->placement[i];
959 struct ttm_resource_manager *man;
960
961 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
962 if (ret)
963 continue;
964
965 type_found = true;
966 ret = ttm_resource_alloc(bo, place, mem);
967 if (ret == -ENOSPC)
968 continue;
969 if (unlikely(ret))
970 goto error;
971
972 man = ttm_manager_type(bdev, mem->mem_type);
973 ret = ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
974 if (unlikely(ret)) {
975 ttm_resource_free(bo, mem);
976 if (ret == -EBUSY)
977 continue;
978
979 goto error;
980 }
981 return 0;
982 }
983
984 for (i = 0; i < placement->num_busy_placement; ++i) {
985 const struct ttm_place *place = &placement->busy_placement[i];
986
987 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
988 if (ret)
989 continue;
990
991 type_found = true;
992 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
993 if (likely(!ret))
994 return 0;
995
996 if (ret && ret != -EBUSY)
997 goto error;
998 }
999
1000 ret = -ENOMEM;
1001 if (!type_found) {
1002 pr_err(TTM_PFX "No compatible memory type found\n");
1003 ret = -EINVAL;
1004 }
1005
1006 error:
1007 if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) {
1008 ttm_bo_move_to_lru_tail_unlocked(bo);
1009 }
1010
1011 return ret;
1012 }
1013 EXPORT_SYMBOL(ttm_bo_mem_space);
1014
ttm_bo_move_buffer(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)1015 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1016 struct ttm_placement *placement,
1017 struct ttm_operation_ctx *ctx)
1018 {
1019 int ret = 0;
1020 struct ttm_resource mem;
1021
1022 dma_resv_assert_held(bo->base.resv);
1023
1024 mem.num_pages = bo->num_pages;
1025 mem.size = mem.num_pages << PAGE_SHIFT;
1026 mem.page_alignment = bo->mem.page_alignment;
1027 mem.bus.offset = 0;
1028 mem.bus.addr = NULL;
1029 mem.mm_node = NULL;
1030
1031 /*
1032 * Determine where to move the buffer.
1033 */
1034 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1035 if (ret)
1036 goto out_unlock;
1037 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1038 out_unlock:
1039 if (ret)
1040 ttm_resource_free(bo, &mem);
1041 return ret;
1042 }
1043
ttm_bo_places_compat(const struct ttm_place * places,unsigned num_placement,struct ttm_resource * mem,uint32_t * new_flags)1044 static bool ttm_bo_places_compat(const struct ttm_place *places,
1045 unsigned num_placement,
1046 struct ttm_resource *mem,
1047 uint32_t *new_flags)
1048 {
1049 unsigned i;
1050
1051 for (i = 0; i < num_placement; i++) {
1052 const struct ttm_place *heap = &places[i];
1053
1054 if ((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 (mem->mem_type == heap->mem_type) &&
1061 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1062 (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1063 return true;
1064 }
1065 return false;
1066 }
1067
ttm_bo_mem_compat(struct ttm_placement * placement,struct ttm_resource * mem,uint32_t * new_flags)1068 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1069 struct ttm_resource *mem,
1070 uint32_t *new_flags)
1071 {
1072 if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1073 mem, new_flags))
1074 return true;
1075
1076 if ((placement->busy_placement != placement->placement ||
1077 placement->num_busy_placement > placement->num_placement) &&
1078 ttm_bo_places_compat(placement->busy_placement,
1079 placement->num_busy_placement,
1080 mem, new_flags))
1081 return true;
1082
1083 return false;
1084 }
1085 EXPORT_SYMBOL(ttm_bo_mem_compat);
1086
ttm_bo_validate(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)1087 int ttm_bo_validate(struct ttm_buffer_object *bo,
1088 struct ttm_placement *placement,
1089 struct ttm_operation_ctx *ctx)
1090 {
1091 int ret;
1092 uint32_t new_flags;
1093
1094 dma_resv_assert_held(bo->base.resv);
1095
1096 /*
1097 * Remove the backing store if no placement is given.
1098 */
1099 if (!placement->num_placement && !placement->num_busy_placement) {
1100 ret = ttm_bo_pipeline_gutting(bo);
1101 if (ret)
1102 return ret;
1103
1104 return ttm_tt_create(bo, false);
1105 }
1106
1107 /*
1108 * Check whether we need to move buffer.
1109 */
1110 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1111 ret = ttm_bo_move_buffer(bo, placement, ctx);
1112 if (ret)
1113 return ret;
1114 } else {
1115 bo->mem.placement &= TTM_PL_MASK_CACHING;
1116 bo->mem.placement |= new_flags & ~TTM_PL_MASK_CACHING;
1117 }
1118 /*
1119 * We might need to add a TTM.
1120 */
1121 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1122 ret = ttm_tt_create(bo, true);
1123 if (ret)
1124 return ret;
1125 }
1126 return 0;
1127 }
1128 EXPORT_SYMBOL(ttm_bo_validate);
1129
ttm_bo_init_reserved(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,struct ttm_operation_ctx * ctx,size_t acc_size,struct sg_table * sg,struct dma_resv * resv,void (* destroy)(struct ttm_buffer_object *))1130 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1131 struct ttm_buffer_object *bo,
1132 unsigned long size,
1133 enum ttm_bo_type type,
1134 struct ttm_placement *placement,
1135 uint32_t page_alignment,
1136 struct ttm_operation_ctx *ctx,
1137 size_t acc_size,
1138 struct sg_table *sg,
1139 struct dma_resv *resv,
1140 void (*destroy) (struct ttm_buffer_object *))
1141 {
1142 struct ttm_mem_global *mem_glob = &ttm_mem_glob;
1143 int ret = 0;
1144 unsigned long num_pages;
1145 bool locked;
1146
1147 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1148 if (ret) {
1149 pr_err("Out of kernel memory\n");
1150 if (destroy)
1151 (*destroy)(bo);
1152 else
1153 kfree(bo);
1154 return -ENOMEM;
1155 }
1156
1157 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1158 if (num_pages == 0) {
1159 pr_err("Illegal buffer object size\n");
1160 if (destroy)
1161 (*destroy)(bo);
1162 else
1163 kfree(bo);
1164 ttm_mem_global_free(mem_glob, acc_size);
1165 return -EINVAL;
1166 }
1167 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1168
1169 kref_init(&bo->kref);
1170 INIT_LIST_HEAD(&bo->lru);
1171 INIT_LIST_HEAD(&bo->ddestroy);
1172 INIT_LIST_HEAD(&bo->swap);
1173 bo->bdev = bdev;
1174 bo->type = type;
1175 bo->num_pages = num_pages;
1176 bo->mem.size = num_pages << PAGE_SHIFT;
1177 bo->mem.mem_type = TTM_PL_SYSTEM;
1178 bo->mem.num_pages = bo->num_pages;
1179 bo->mem.mm_node = NULL;
1180 bo->mem.page_alignment = page_alignment;
1181 bo->mem.bus.offset = 0;
1182 bo->mem.bus.addr = NULL;
1183 bo->moving = NULL;
1184 bo->mem.placement = TTM_PL_FLAG_CACHED;
1185 bo->acc_size = acc_size;
1186 bo->pin_count = 0;
1187 bo->sg = sg;
1188 if (resv) {
1189 bo->base.resv = resv;
1190 dma_resv_assert_held(bo->base.resv);
1191 } else {
1192 bo->base.resv = &bo->base._resv;
1193 }
1194 if (!ttm_bo_uses_embedded_gem_object(bo)) {
1195 /*
1196 * bo.gem is not initialized, so we have to setup the
1197 * struct elements we want use regardless.
1198 */
1199 dma_resv_init(&bo->base._resv);
1200 drm_vma_node_reset(&bo->base.vma_node);
1201 }
1202 atomic_inc(&ttm_bo_glob.bo_count);
1203
1204 /*
1205 * For ttm_bo_type_device buffers, allocate
1206 * address space from the device.
1207 */
1208 if (bo->type == ttm_bo_type_device ||
1209 bo->type == ttm_bo_type_sg)
1210 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1211 bo->mem.num_pages);
1212
1213 /* passed reservation objects should already be locked,
1214 * since otherwise lockdep will be angered in radeon.
1215 */
1216 if (!resv) {
1217 locked = dma_resv_trylock(bo->base.resv);
1218 WARN_ON(!locked);
1219 }
1220
1221 if (likely(!ret))
1222 ret = ttm_bo_validate(bo, placement, ctx);
1223
1224 if (unlikely(ret)) {
1225 if (!resv)
1226 ttm_bo_unreserve(bo);
1227
1228 ttm_bo_put(bo);
1229 return ret;
1230 }
1231
1232 ttm_bo_move_to_lru_tail_unlocked(bo);
1233
1234 return ret;
1235 }
1236 EXPORT_SYMBOL(ttm_bo_init_reserved);
1237
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,size_t acc_size,struct sg_table * sg,struct dma_resv * resv,void (* destroy)(struct ttm_buffer_object *))1238 int ttm_bo_init(struct ttm_bo_device *bdev,
1239 struct ttm_buffer_object *bo,
1240 unsigned long size,
1241 enum ttm_bo_type type,
1242 struct ttm_placement *placement,
1243 uint32_t page_alignment,
1244 bool interruptible,
1245 size_t acc_size,
1246 struct sg_table *sg,
1247 struct dma_resv *resv,
1248 void (*destroy) (struct ttm_buffer_object *))
1249 {
1250 struct ttm_operation_ctx ctx = { interruptible, false };
1251 int ret;
1252
1253 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1254 page_alignment, &ctx, acc_size,
1255 sg, resv, destroy);
1256 if (ret)
1257 return ret;
1258
1259 if (!resv)
1260 ttm_bo_unreserve(bo);
1261
1262 return 0;
1263 }
1264 EXPORT_SYMBOL(ttm_bo_init);
1265
ttm_bo_acc_size(struct ttm_bo_device * bdev,unsigned long bo_size,unsigned struct_size)1266 static size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1267 unsigned long bo_size,
1268 unsigned struct_size)
1269 {
1270 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1271 size_t size = 0;
1272
1273 size += ttm_round_pot(struct_size);
1274 size += ttm_round_pot(npages * sizeof(void *));
1275 size += ttm_round_pot(sizeof(struct ttm_tt));
1276 return size;
1277 }
1278
ttm_bo_dma_acc_size(struct ttm_bo_device * bdev,unsigned long bo_size,unsigned struct_size)1279 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1280 unsigned long bo_size,
1281 unsigned struct_size)
1282 {
1283 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1284 size_t size = 0;
1285
1286 size += ttm_round_pot(struct_size);
1287 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1288 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1289 return size;
1290 }
1291 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1292
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 ttm_buffer_object ** p_bo)1293 int ttm_bo_create(struct ttm_bo_device *bdev,
1294 unsigned long size,
1295 enum ttm_bo_type type,
1296 struct ttm_placement *placement,
1297 uint32_t page_alignment,
1298 bool interruptible,
1299 struct ttm_buffer_object **p_bo)
1300 {
1301 struct ttm_buffer_object *bo;
1302 size_t acc_size;
1303 int ret;
1304
1305 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1306 if (unlikely(bo == NULL))
1307 return -ENOMEM;
1308
1309 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1310 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1311 interruptible, acc_size,
1312 NULL, NULL, NULL);
1313 if (likely(ret == 0))
1314 *p_bo = bo;
1315
1316 return ret;
1317 }
1318 EXPORT_SYMBOL(ttm_bo_create);
1319
ttm_bo_evict_mm(struct ttm_bo_device * bdev,unsigned mem_type)1320 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1321 {
1322 struct ttm_resource_manager *man = ttm_manager_type(bdev, mem_type);
1323
1324 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1325 pr_err("Illegal memory manager memory type %u\n", mem_type);
1326 return -EINVAL;
1327 }
1328
1329 if (!man) {
1330 pr_err("Memory type %u has not been initialized\n", mem_type);
1331 return 0;
1332 }
1333
1334 return ttm_resource_manager_force_list_clean(bdev, man);
1335 }
1336 EXPORT_SYMBOL(ttm_bo_evict_mm);
1337
ttm_bo_global_kobj_release(struct kobject * kobj)1338 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1339 {
1340 struct ttm_bo_global *glob =
1341 container_of(kobj, struct ttm_bo_global, kobj);
1342
1343 __free_page(glob->dummy_read_page);
1344 }
1345
ttm_bo_global_release(void)1346 static void ttm_bo_global_release(void)
1347 {
1348 struct ttm_bo_global *glob = &ttm_bo_glob;
1349
1350 mutex_lock(&ttm_global_mutex);
1351 if (--ttm_bo_glob_use_count > 0)
1352 goto out;
1353
1354 kobject_del(&glob->kobj);
1355 kobject_put(&glob->kobj);
1356 ttm_mem_global_release(&ttm_mem_glob);
1357 memset(glob, 0, sizeof(*glob));
1358 out:
1359 mutex_unlock(&ttm_global_mutex);
1360 }
1361
ttm_bo_global_init(void)1362 static int ttm_bo_global_init(void)
1363 {
1364 struct ttm_bo_global *glob = &ttm_bo_glob;
1365 int ret = 0;
1366 unsigned i;
1367
1368 mutex_lock(&ttm_global_mutex);
1369 if (++ttm_bo_glob_use_count > 1)
1370 goto out;
1371
1372 ret = ttm_mem_global_init(&ttm_mem_glob);
1373 if (ret)
1374 goto out;
1375
1376 spin_lock_init(&glob->lru_lock);
1377 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1378
1379 if (unlikely(glob->dummy_read_page == NULL)) {
1380 ret = -ENOMEM;
1381 goto out;
1382 }
1383
1384 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1385 INIT_LIST_HEAD(&glob->swap_lru[i]);
1386 INIT_LIST_HEAD(&glob->device_list);
1387 atomic_set(&glob->bo_count, 0);
1388
1389 ret = kobject_init_and_add(
1390 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1391 if (unlikely(ret != 0))
1392 kobject_put(&glob->kobj);
1393 out:
1394 mutex_unlock(&ttm_global_mutex);
1395 return ret;
1396 }
1397
ttm_bo_device_release(struct ttm_bo_device * bdev)1398 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1399 {
1400 struct ttm_bo_global *glob = &ttm_bo_glob;
1401 int ret = 0;
1402 unsigned i;
1403 struct ttm_resource_manager *man;
1404
1405 man = ttm_manager_type(bdev, TTM_PL_SYSTEM);
1406 ttm_resource_manager_set_used(man, false);
1407 ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, NULL);
1408
1409 mutex_lock(&ttm_global_mutex);
1410 list_del(&bdev->device_list);
1411 mutex_unlock(&ttm_global_mutex);
1412
1413 cancel_delayed_work_sync(&bdev->wq);
1414
1415 if (ttm_bo_delayed_delete(bdev, true))
1416 pr_debug("Delayed destroy list was clean\n");
1417
1418 spin_lock(&glob->lru_lock);
1419 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1420 if (list_empty(&man->lru[0]))
1421 pr_debug("Swap list %d was clean\n", i);
1422 spin_unlock(&glob->lru_lock);
1423
1424 if (!ret)
1425 ttm_bo_global_release();
1426
1427 return ret;
1428 }
1429 EXPORT_SYMBOL(ttm_bo_device_release);
1430
ttm_bo_init_sysman(struct ttm_bo_device * bdev)1431 static void ttm_bo_init_sysman(struct ttm_bo_device *bdev)
1432 {
1433 struct ttm_resource_manager *man = &bdev->sysman;
1434
1435 /*
1436 * Initialize the system memory buffer type.
1437 * Other types need to be driver / IOCTL initialized.
1438 */
1439 man->use_tt = true;
1440
1441 ttm_resource_manager_init(man, 0);
1442 ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, man);
1443 ttm_resource_manager_set_used(man, true);
1444 }
1445
ttm_bo_device_init(struct ttm_bo_device * bdev,struct ttm_bo_driver * driver,struct address_space * mapping,struct drm_vma_offset_manager * vma_manager,bool need_dma32)1446 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1447 struct ttm_bo_driver *driver,
1448 struct address_space *mapping,
1449 struct drm_vma_offset_manager *vma_manager,
1450 bool need_dma32)
1451 {
1452 struct ttm_bo_global *glob = &ttm_bo_glob;
1453 int ret;
1454
1455 if (WARN_ON(vma_manager == NULL))
1456 return -EINVAL;
1457
1458 ret = ttm_bo_global_init();
1459 if (ret)
1460 return ret;
1461
1462 bdev->driver = driver;
1463
1464 ttm_bo_init_sysman(bdev);
1465
1466 bdev->vma_manager = vma_manager;
1467 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1468 INIT_LIST_HEAD(&bdev->ddestroy);
1469 bdev->dev_mapping = mapping;
1470 bdev->need_dma32 = need_dma32;
1471 mutex_lock(&ttm_global_mutex);
1472 list_add_tail(&bdev->device_list, &glob->device_list);
1473 mutex_unlock(&ttm_global_mutex);
1474
1475 return 0;
1476 }
1477 EXPORT_SYMBOL(ttm_bo_device_init);
1478
1479 /*
1480 * buffer object vm functions.
1481 */
1482
ttm_bo_unmap_virtual(struct ttm_buffer_object * bo)1483 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1484 {
1485 struct ttm_bo_device *bdev = bo->bdev;
1486
1487 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1488 ttm_mem_io_free(bdev, &bo->mem);
1489 }
1490 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1491
ttm_bo_wait(struct ttm_buffer_object * bo,bool interruptible,bool no_wait)1492 int ttm_bo_wait(struct ttm_buffer_object *bo,
1493 bool interruptible, bool no_wait)
1494 {
1495 long timeout = 15 * HZ;
1496
1497 if (no_wait) {
1498 if (dma_resv_test_signaled_rcu(bo->base.resv, true))
1499 return 0;
1500 else
1501 return -EBUSY;
1502 }
1503
1504 timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true,
1505 interruptible, timeout);
1506 if (timeout < 0)
1507 return timeout;
1508
1509 if (timeout == 0)
1510 return -EBUSY;
1511
1512 dma_resv_add_excl_fence(bo->base.resv, NULL);
1513 return 0;
1514 }
1515 EXPORT_SYMBOL(ttm_bo_wait);
1516
1517 /**
1518 * A buffer object shrink method that tries to swap out the first
1519 * buffer object on the bo_global::swap_lru list.
1520 */
ttm_bo_swapout(struct ttm_bo_global * glob,struct ttm_operation_ctx * ctx)1521 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1522 {
1523 struct ttm_buffer_object *bo;
1524 int ret = -EBUSY;
1525 bool locked;
1526 unsigned i;
1527
1528 spin_lock(&glob->lru_lock);
1529 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1530 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1531 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
1532 NULL))
1533 continue;
1534
1535 if (!ttm_bo_get_unless_zero(bo)) {
1536 if (locked)
1537 dma_resv_unlock(bo->base.resv);
1538 continue;
1539 }
1540
1541 ret = 0;
1542 break;
1543 }
1544 if (!ret)
1545 break;
1546 }
1547
1548 if (ret) {
1549 spin_unlock(&glob->lru_lock);
1550 return ret;
1551 }
1552
1553 if (bo->deleted) {
1554 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1555 ttm_bo_put(bo);
1556 return ret;
1557 }
1558
1559 ttm_bo_del_from_lru(bo);
1560 spin_unlock(&glob->lru_lock);
1561
1562 /**
1563 * Move to system cached
1564 */
1565
1566 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1567 bo->ttm->caching_state != tt_cached) {
1568 struct ttm_operation_ctx ctx = { false, false };
1569 struct ttm_resource evict_mem;
1570
1571 evict_mem = bo->mem;
1572 evict_mem.mm_node = NULL;
1573 evict_mem.placement = TTM_PL_FLAG_CACHED;
1574 evict_mem.mem_type = TTM_PL_SYSTEM;
1575
1576 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1577 if (unlikely(ret != 0))
1578 goto out;
1579 }
1580
1581 /**
1582 * Make sure BO is idle.
1583 */
1584
1585 ret = ttm_bo_wait(bo, false, false);
1586 if (unlikely(ret != 0))
1587 goto out;
1588
1589 ttm_bo_unmap_virtual(bo);
1590
1591 /**
1592 * Swap out. Buffer will be swapped in again as soon as
1593 * anyone tries to access a ttm page.
1594 */
1595
1596 if (bo->bdev->driver->swap_notify)
1597 bo->bdev->driver->swap_notify(bo);
1598
1599 ret = ttm_tt_swapout(bo->bdev, bo->ttm, bo->persistent_swap_storage);
1600 out:
1601
1602 /**
1603 *
1604 * Unreserve without putting on LRU to avoid swapping out an
1605 * already swapped buffer.
1606 */
1607 if (locked)
1608 dma_resv_unlock(bo->base.resv);
1609 ttm_bo_put(bo);
1610 return ret;
1611 }
1612 EXPORT_SYMBOL(ttm_bo_swapout);
1613
ttm_bo_swapout_all(void)1614 void ttm_bo_swapout_all(void)
1615 {
1616 struct ttm_operation_ctx ctx = {
1617 .interruptible = false,
1618 .no_wait_gpu = false
1619 };
1620
1621 while (ttm_bo_swapout(&ttm_bo_glob, &ctx) == 0);
1622 }
1623 EXPORT_SYMBOL(ttm_bo_swapout_all);
1624
ttm_bo_tt_destroy(struct ttm_buffer_object * bo)1625 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1626 {
1627 if (bo->ttm == NULL)
1628 return;
1629
1630 ttm_tt_destroy(bo->bdev, bo->ttm);
1631 bo->ttm = NULL;
1632 }
1633
ttm_bo_tt_bind(struct ttm_buffer_object * bo,struct ttm_resource * mem)1634 int ttm_bo_tt_bind(struct ttm_buffer_object *bo, struct ttm_resource *mem)
1635 {
1636 return bo->bdev->driver->ttm_tt_bind(bo->bdev, bo->ttm, mem);
1637 }
1638
ttm_bo_tt_unbind(struct ttm_buffer_object * bo)1639 void ttm_bo_tt_unbind(struct ttm_buffer_object *bo)
1640 {
1641 bo->bdev->driver->ttm_tt_unbind(bo->bdev, bo->ttm);
1642 }
1643