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_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <linux/jiffies.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 #include <linux/mm.h>
40 #include <linux/file.h>
41 #include <linux/module.h>
42 #include <linux/atomic.h>
43 #include <linux/dma-resv.h>
44
45 #include "ttm_module.h"
46
47 /* default destructor */
ttm_bo_default_destroy(struct ttm_buffer_object * bo)48 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
49 {
50 kfree(bo);
51 }
52
ttm_bo_mem_space_debug(struct ttm_buffer_object * bo,struct ttm_placement * placement)53 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
54 struct ttm_placement *placement)
55 {
56 struct drm_printer p = drm_debug_printer(TTM_PFX);
57 struct ttm_resource_manager *man;
58 int i, mem_type;
59
60 drm_printf(&p, "No space for %p (%lu pages, %zuK, %zuM)\n",
61 bo, bo->resource->num_pages, bo->base.size >> 10,
62 bo->base.size >> 20);
63 for (i = 0; i < placement->num_placement; i++) {
64 mem_type = placement->placement[i].mem_type;
65 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
66 i, placement->placement[i].flags, mem_type);
67 man = ttm_manager_type(bo->bdev, mem_type);
68 ttm_resource_manager_debug(man, &p);
69 }
70 }
71
ttm_bo_move_to_pinned(struct ttm_buffer_object * bo)72 static inline void ttm_bo_move_to_pinned(struct ttm_buffer_object *bo)
73 {
74 struct ttm_device *bdev = bo->bdev;
75
76 list_move_tail(&bo->lru, &bdev->pinned);
77
78 if (bdev->funcs->del_from_lru_notify)
79 bdev->funcs->del_from_lru_notify(bo);
80 }
81
ttm_bo_del_from_lru(struct ttm_buffer_object * bo)82 static inline void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
83 {
84 struct ttm_device *bdev = bo->bdev;
85
86 list_del_init(&bo->lru);
87
88 if (bdev->funcs->del_from_lru_notify)
89 bdev->funcs->del_from_lru_notify(bo);
90 }
91
ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos * pos,struct ttm_buffer_object * bo)92 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
93 struct ttm_buffer_object *bo)
94 {
95 if (!pos->first)
96 pos->first = bo;
97 pos->last = bo;
98 }
99
ttm_bo_move_to_lru_tail(struct ttm_buffer_object * bo,struct ttm_resource * mem,struct ttm_lru_bulk_move * bulk)100 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
101 struct ttm_resource *mem,
102 struct ttm_lru_bulk_move *bulk)
103 {
104 struct ttm_device *bdev = bo->bdev;
105 struct ttm_resource_manager *man;
106
107 if (!bo->deleted)
108 dma_resv_assert_held(bo->base.resv);
109
110 if (bo->pin_count) {
111 ttm_bo_move_to_pinned(bo);
112 return;
113 }
114
115 if (!mem)
116 return;
117
118 man = ttm_manager_type(bdev, mem->mem_type);
119 list_move_tail(&bo->lru, &man->lru[bo->priority]);
120
121 if (bdev->funcs->del_from_lru_notify)
122 bdev->funcs->del_from_lru_notify(bo);
123
124 if (bulk && !bo->pin_count) {
125 switch (bo->resource->mem_type) {
126 case TTM_PL_TT:
127 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
128 break;
129
130 case TTM_PL_VRAM:
131 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
132 break;
133 }
134 }
135 }
136 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
137
ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move * bulk)138 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
139 {
140 unsigned i;
141
142 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
143 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
144 struct ttm_resource_manager *man;
145
146 if (!pos->first)
147 continue;
148
149 dma_resv_assert_held(pos->first->base.resv);
150 dma_resv_assert_held(pos->last->base.resv);
151
152 man = ttm_manager_type(pos->first->bdev, TTM_PL_TT);
153 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
154 &pos->last->lru);
155 }
156
157 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
158 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
159 struct ttm_resource_manager *man;
160
161 if (!pos->first)
162 continue;
163
164 dma_resv_assert_held(pos->first->base.resv);
165 dma_resv_assert_held(pos->last->base.resv);
166
167 man = ttm_manager_type(pos->first->bdev, TTM_PL_VRAM);
168 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
169 &pos->last->lru);
170 }
171 }
172 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
173
ttm_bo_handle_move_mem(struct ttm_buffer_object * bo,struct ttm_resource * mem,bool evict,struct ttm_operation_ctx * ctx,struct ttm_place * hop)174 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
175 struct ttm_resource *mem, bool evict,
176 struct ttm_operation_ctx *ctx,
177 struct ttm_place *hop)
178 {
179 struct ttm_resource_manager *old_man, *new_man;
180 struct ttm_device *bdev = bo->bdev;
181 int ret;
182
183 old_man = ttm_manager_type(bdev, bo->resource->mem_type);
184 new_man = ttm_manager_type(bdev, mem->mem_type);
185
186 ttm_bo_unmap_virtual(bo);
187
188 /*
189 * Create and bind a ttm if required.
190 */
191
192 if (new_man->use_tt) {
193 /* Zero init the new TTM structure if the old location should
194 * have used one as well.
195 */
196 ret = ttm_tt_create(bo, old_man->use_tt);
197 if (ret)
198 goto out_err;
199
200 if (mem->mem_type != TTM_PL_SYSTEM) {
201 ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
202 if (ret)
203 goto out_err;
204 }
205 }
206
207 ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
208 if (ret) {
209 if (ret == -EMULTIHOP)
210 return ret;
211 goto out_err;
212 }
213
214 ctx->bytes_moved += bo->base.size;
215 return 0;
216
217 out_err:
218 new_man = ttm_manager_type(bdev, bo->resource->mem_type);
219 if (!new_man->use_tt)
220 ttm_bo_tt_destroy(bo);
221
222 return ret;
223 }
224
225 /*
226 * Call bo::reserved.
227 * Will release GPU memory type usage on destruction.
228 * This is the place to put in driver specific hooks to release
229 * driver private resources.
230 * Will release the bo::reserved lock.
231 */
232
ttm_bo_cleanup_memtype_use(struct ttm_buffer_object * bo)233 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
234 {
235 if (bo->bdev->funcs->delete_mem_notify)
236 bo->bdev->funcs->delete_mem_notify(bo);
237
238 ttm_bo_tt_destroy(bo);
239 ttm_resource_free(bo, &bo->resource);
240 }
241
ttm_bo_individualize_resv(struct ttm_buffer_object * bo)242 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
243 {
244 int r;
245
246 if (bo->base.resv == &bo->base._resv)
247 return 0;
248
249 BUG_ON(!dma_resv_trylock(&bo->base._resv));
250
251 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
252 dma_resv_unlock(&bo->base._resv);
253 if (r)
254 return r;
255
256 if (bo->type != ttm_bo_type_sg) {
257 /* This works because the BO is about to be destroyed and nobody
258 * reference it any more. The only tricky case is the trylock on
259 * the resv object while holding the lru_lock.
260 */
261 spin_lock(&bo->bdev->lru_lock);
262 bo->base.resv = &bo->base._resv;
263 spin_unlock(&bo->bdev->lru_lock);
264 }
265
266 return r;
267 }
268
ttm_bo_flush_all_fences(struct ttm_buffer_object * bo)269 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
270 {
271 struct dma_resv *resv = &bo->base._resv;
272 struct dma_resv_iter cursor;
273 struct dma_fence *fence;
274
275 dma_resv_iter_begin(&cursor, resv, true);
276 dma_resv_for_each_fence_unlocked(&cursor, fence) {
277 if (!fence->ops->signaled)
278 dma_fence_enable_sw_signaling(fence);
279 }
280 dma_resv_iter_end(&cursor);
281 }
282
283 /**
284 * ttm_bo_cleanup_refs
285 * If bo idle, remove from lru lists, and unref.
286 * If not idle, block if possible.
287 *
288 * Must be called with lru_lock and reservation held, this function
289 * will drop the lru lock and optionally the reservation lock before returning.
290 *
291 * @bo: The buffer object to clean-up
292 * @interruptible: Any sleeps should occur interruptibly.
293 * @no_wait_gpu: Never wait for gpu. Return -EBUSY instead.
294 * @unlock_resv: Unlock the reservation lock as well.
295 */
296
ttm_bo_cleanup_refs(struct ttm_buffer_object * bo,bool interruptible,bool no_wait_gpu,bool unlock_resv)297 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
298 bool interruptible, bool no_wait_gpu,
299 bool unlock_resv)
300 {
301 struct dma_resv *resv = &bo->base._resv;
302 int ret;
303
304 if (dma_resv_test_signaled(resv, true))
305 ret = 0;
306 else
307 ret = -EBUSY;
308
309 if (ret && !no_wait_gpu) {
310 long lret;
311
312 if (unlock_resv)
313 dma_resv_unlock(bo->base.resv);
314 spin_unlock(&bo->bdev->lru_lock);
315
316 lret = dma_resv_wait_timeout(resv, true, interruptible,
317 30 * HZ);
318
319 if (lret < 0)
320 return lret;
321 else if (lret == 0)
322 return -EBUSY;
323
324 spin_lock(&bo->bdev->lru_lock);
325 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
326 /*
327 * We raced, and lost, someone else holds the reservation now,
328 * and is probably busy in ttm_bo_cleanup_memtype_use.
329 *
330 * Even if it's not the case, because we finished waiting any
331 * delayed destruction would succeed, so just return success
332 * here.
333 */
334 spin_unlock(&bo->bdev->lru_lock);
335 return 0;
336 }
337 ret = 0;
338 }
339
340 if (ret || unlikely(list_empty(&bo->ddestroy))) {
341 if (unlock_resv)
342 dma_resv_unlock(bo->base.resv);
343 spin_unlock(&bo->bdev->lru_lock);
344 return ret;
345 }
346
347 ttm_bo_move_to_pinned(bo);
348 list_del_init(&bo->ddestroy);
349 spin_unlock(&bo->bdev->lru_lock);
350 ttm_bo_cleanup_memtype_use(bo);
351
352 if (unlock_resv)
353 dma_resv_unlock(bo->base.resv);
354
355 ttm_bo_put(bo);
356
357 return 0;
358 }
359
360 /*
361 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
362 * encountered buffers.
363 */
ttm_bo_delayed_delete(struct ttm_device * bdev,bool remove_all)364 bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all)
365 {
366 struct list_head removed;
367 bool empty;
368
369 INIT_LIST_HEAD(&removed);
370
371 spin_lock(&bdev->lru_lock);
372 while (!list_empty(&bdev->ddestroy)) {
373 struct ttm_buffer_object *bo;
374
375 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
376 ddestroy);
377 list_move_tail(&bo->ddestroy, &removed);
378 if (!ttm_bo_get_unless_zero(bo))
379 continue;
380
381 if (remove_all || bo->base.resv != &bo->base._resv) {
382 spin_unlock(&bdev->lru_lock);
383 dma_resv_lock(bo->base.resv, NULL);
384
385 spin_lock(&bdev->lru_lock);
386 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
387
388 } else if (dma_resv_trylock(bo->base.resv)) {
389 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
390 } else {
391 spin_unlock(&bdev->lru_lock);
392 }
393
394 ttm_bo_put(bo);
395 spin_lock(&bdev->lru_lock);
396 }
397 list_splice_tail(&removed, &bdev->ddestroy);
398 empty = list_empty(&bdev->ddestroy);
399 spin_unlock(&bdev->lru_lock);
400
401 return empty;
402 }
403
ttm_bo_release(struct kref * kref)404 static void ttm_bo_release(struct kref *kref)
405 {
406 struct ttm_buffer_object *bo =
407 container_of(kref, struct ttm_buffer_object, kref);
408 struct ttm_device *bdev = bo->bdev;
409 int ret;
410
411 WARN_ON_ONCE(bo->pin_count);
412
413 if (!bo->deleted) {
414 ret = ttm_bo_individualize_resv(bo);
415 if (ret) {
416 /* Last resort, if we fail to allocate memory for the
417 * fences block for the BO to become idle
418 */
419 dma_resv_wait_timeout(bo->base.resv, true, false,
420 30 * HZ);
421 }
422
423 if (bo->bdev->funcs->release_notify)
424 bo->bdev->funcs->release_notify(bo);
425
426 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
427 ttm_mem_io_free(bdev, bo->resource);
428 }
429
430 if (!dma_resv_test_signaled(bo->base.resv, true) ||
431 !dma_resv_trylock(bo->base.resv)) {
432 /* The BO is not idle, resurrect it for delayed destroy */
433 ttm_bo_flush_all_fences(bo);
434 bo->deleted = true;
435
436 spin_lock(&bo->bdev->lru_lock);
437
438 /*
439 * Make pinned bos immediately available to
440 * shrinkers, now that they are queued for
441 * destruction.
442 *
443 * FIXME: QXL is triggering this. Can be removed when the
444 * driver is fixed.
445 */
446 if (bo->pin_count) {
447 bo->pin_count = 0;
448 ttm_bo_move_to_lru_tail(bo, bo->resource, NULL);
449 }
450
451 kref_init(&bo->kref);
452 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
453 spin_unlock(&bo->bdev->lru_lock);
454
455 schedule_delayed_work(&bdev->wq,
456 ((HZ / 100) < 1) ? 1 : HZ / 100);
457 return;
458 }
459
460 spin_lock(&bo->bdev->lru_lock);
461 ttm_bo_del_from_lru(bo);
462 list_del(&bo->ddestroy);
463 spin_unlock(&bo->bdev->lru_lock);
464
465 ttm_bo_cleanup_memtype_use(bo);
466 dma_resv_unlock(bo->base.resv);
467
468 atomic_dec(&ttm_glob.bo_count);
469 dma_fence_put(bo->moving);
470 bo->destroy(bo);
471 }
472
ttm_bo_put(struct ttm_buffer_object * bo)473 void ttm_bo_put(struct ttm_buffer_object *bo)
474 {
475 kref_put(&bo->kref, ttm_bo_release);
476 }
477 EXPORT_SYMBOL(ttm_bo_put);
478
ttm_bo_lock_delayed_workqueue(struct ttm_device * bdev)479 int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev)
480 {
481 return cancel_delayed_work_sync(&bdev->wq);
482 }
483 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
484
ttm_bo_unlock_delayed_workqueue(struct ttm_device * bdev,int resched)485 void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched)
486 {
487 if (resched)
488 schedule_delayed_work(&bdev->wq,
489 ((HZ / 100) < 1) ? 1 : HZ / 100);
490 }
491 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
492
ttm_bo_bounce_temp_buffer(struct ttm_buffer_object * bo,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx,struct ttm_place * hop)493 static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
494 struct ttm_resource **mem,
495 struct ttm_operation_ctx *ctx,
496 struct ttm_place *hop)
497 {
498 struct ttm_placement hop_placement;
499 struct ttm_resource *hop_mem;
500 int ret;
501
502 hop_placement.num_placement = hop_placement.num_busy_placement = 1;
503 hop_placement.placement = hop_placement.busy_placement = hop;
504
505 /* find space in the bounce domain */
506 ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
507 if (ret)
508 return ret;
509 /* move to the bounce domain */
510 ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
511 if (ret) {
512 ttm_resource_free(bo, &hop_mem);
513 return ret;
514 }
515 return 0;
516 }
517
ttm_bo_evict(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)518 static int ttm_bo_evict(struct ttm_buffer_object *bo,
519 struct ttm_operation_ctx *ctx)
520 {
521 struct ttm_device *bdev = bo->bdev;
522 struct ttm_resource *evict_mem;
523 struct ttm_placement placement;
524 struct ttm_place hop;
525 int ret = 0;
526
527 memset(&hop, 0, sizeof(hop));
528
529 dma_resv_assert_held(bo->base.resv);
530
531 placement.num_placement = 0;
532 placement.num_busy_placement = 0;
533 bdev->funcs->evict_flags(bo, &placement);
534
535 if (!placement.num_placement && !placement.num_busy_placement) {
536 ret = ttm_bo_wait(bo, true, false);
537 if (ret)
538 return ret;
539
540 /*
541 * Since we've already synced, this frees backing store
542 * immediately.
543 */
544 return ttm_bo_pipeline_gutting(bo);
545 }
546
547 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
548 if (ret) {
549 if (ret != -ERESTARTSYS) {
550 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
551 bo);
552 ttm_bo_mem_space_debug(bo, &placement);
553 }
554 goto out;
555 }
556
557 bounce:
558 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
559 if (ret == -EMULTIHOP) {
560 ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
561 if (ret) {
562 pr_err("Buffer eviction failed\n");
563 ttm_resource_free(bo, &evict_mem);
564 goto out;
565 }
566 /* try and move to final place now. */
567 goto bounce;
568 }
569 out:
570 return ret;
571 }
572
ttm_bo_eviction_valuable(struct ttm_buffer_object * bo,const struct ttm_place * place)573 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
574 const struct ttm_place *place)
575 {
576 dma_resv_assert_held(bo->base.resv);
577 if (bo->resource->mem_type == TTM_PL_SYSTEM)
578 return true;
579
580 /* Don't evict this BO if it's outside of the
581 * requested placement range
582 */
583 if (place->fpfn >= (bo->resource->start + bo->resource->num_pages) ||
584 (place->lpfn && place->lpfn <= bo->resource->start))
585 return false;
586
587 return true;
588 }
589 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
590
591 /*
592 * Check the target bo is allowable to be evicted or swapout, including cases:
593 *
594 * a. if share same reservation object with ctx->resv, have assumption
595 * reservation objects should already be locked, so not lock again and
596 * return true directly when either the opreation allow_reserved_eviction
597 * or the target bo already is in delayed free list;
598 *
599 * b. Otherwise, trylock it.
600 */
ttm_bo_evict_swapout_allowable(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,const struct ttm_place * place,bool * locked,bool * busy)601 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
602 struct ttm_operation_ctx *ctx,
603 const struct ttm_place *place,
604 bool *locked, bool *busy)
605 {
606 bool ret = false;
607
608 if (bo->base.resv == ctx->resv) {
609 dma_resv_assert_held(bo->base.resv);
610 if (ctx->allow_res_evict)
611 ret = true;
612 *locked = false;
613 if (busy)
614 *busy = false;
615 } else {
616 ret = dma_resv_trylock(bo->base.resv);
617 *locked = ret;
618 if (busy)
619 *busy = !ret;
620 }
621
622 if (ret && place && (bo->resource->mem_type != place->mem_type ||
623 !bo->bdev->funcs->eviction_valuable(bo, place))) {
624 ret = false;
625 if (*locked) {
626 dma_resv_unlock(bo->base.resv);
627 *locked = false;
628 }
629 }
630
631 return ret;
632 }
633
634 /**
635 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
636 *
637 * @busy_bo: BO which couldn't be locked with trylock
638 * @ctx: operation context
639 * @ticket: acquire ticket
640 *
641 * Try to lock a busy buffer object to avoid failing eviction.
642 */
ttm_mem_evict_wait_busy(struct ttm_buffer_object * busy_bo,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)643 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
644 struct ttm_operation_ctx *ctx,
645 struct ww_acquire_ctx *ticket)
646 {
647 int r;
648
649 if (!busy_bo || !ticket)
650 return -EBUSY;
651
652 if (ctx->interruptible)
653 r = dma_resv_lock_interruptible(busy_bo->base.resv,
654 ticket);
655 else
656 r = dma_resv_lock(busy_bo->base.resv, ticket);
657
658 /*
659 * TODO: It would be better to keep the BO locked until allocation is at
660 * least tried one more time, but that would mean a much larger rework
661 * of TTM.
662 */
663 if (!r)
664 dma_resv_unlock(busy_bo->base.resv);
665
666 return r == -EDEADLK ? -EBUSY : r;
667 }
668
ttm_mem_evict_first(struct ttm_device * bdev,struct ttm_resource_manager * man,const struct ttm_place * place,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)669 int ttm_mem_evict_first(struct ttm_device *bdev,
670 struct ttm_resource_manager *man,
671 const struct ttm_place *place,
672 struct ttm_operation_ctx *ctx,
673 struct ww_acquire_ctx *ticket)
674 {
675 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
676 bool locked = false;
677 unsigned i;
678 int ret;
679
680 spin_lock(&bdev->lru_lock);
681 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
682 list_for_each_entry(bo, &man->lru[i], lru) {
683 bool busy;
684
685 if (!ttm_bo_evict_swapout_allowable(bo, ctx, place,
686 &locked, &busy)) {
687 if (busy && !busy_bo && ticket !=
688 dma_resv_locking_ctx(bo->base.resv))
689 busy_bo = bo;
690 continue;
691 }
692
693 if (!ttm_bo_get_unless_zero(bo)) {
694 if (locked)
695 dma_resv_unlock(bo->base.resv);
696 continue;
697 }
698 break;
699 }
700
701 /* If the inner loop terminated early, we have our candidate */
702 if (&bo->lru != &man->lru[i])
703 break;
704
705 bo = NULL;
706 }
707
708 if (!bo) {
709 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
710 busy_bo = NULL;
711 spin_unlock(&bdev->lru_lock);
712 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
713 if (busy_bo)
714 ttm_bo_put(busy_bo);
715 return ret;
716 }
717
718 if (bo->deleted) {
719 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
720 ctx->no_wait_gpu, locked);
721 ttm_bo_put(bo);
722 return ret;
723 }
724
725 spin_unlock(&bdev->lru_lock);
726
727 ret = ttm_bo_evict(bo, ctx);
728 if (locked)
729 ttm_bo_unreserve(bo);
730
731 ttm_bo_put(bo);
732 return ret;
733 }
734
735 /*
736 * Add the last move fence to the BO and reserve a new shared slot. We only use
737 * a shared slot to avoid unecessary sync and rely on the subsequent bo move to
738 * either stall or use an exclusive fence respectively set bo->moving.
739 */
ttm_bo_add_move_fence(struct ttm_buffer_object * bo,struct ttm_resource_manager * man,struct ttm_resource * mem,bool no_wait_gpu)740 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
741 struct ttm_resource_manager *man,
742 struct ttm_resource *mem,
743 bool no_wait_gpu)
744 {
745 struct dma_fence *fence;
746 int ret;
747
748 spin_lock(&man->move_lock);
749 fence = dma_fence_get(man->move);
750 spin_unlock(&man->move_lock);
751
752 if (!fence)
753 return 0;
754
755 if (no_wait_gpu) {
756 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
757 dma_fence_put(fence);
758 return ret;
759 }
760
761 dma_resv_add_shared_fence(bo->base.resv, fence);
762
763 ret = dma_resv_reserve_shared(bo->base.resv, 1);
764 if (unlikely(ret)) {
765 dma_fence_put(fence);
766 return ret;
767 }
768
769 dma_fence_put(bo->moving);
770 bo->moving = fence;
771 return 0;
772 }
773
774 /*
775 * Repeatedly evict memory from the LRU for @mem_type until we create enough
776 * space, or we've evicted everything and there isn't enough space.
777 */
ttm_bo_mem_force_space(struct ttm_buffer_object * bo,const struct ttm_place * place,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx)778 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
779 const struct ttm_place *place,
780 struct ttm_resource **mem,
781 struct ttm_operation_ctx *ctx)
782 {
783 struct ttm_device *bdev = bo->bdev;
784 struct ttm_resource_manager *man;
785 struct ww_acquire_ctx *ticket;
786 int ret;
787
788 man = ttm_manager_type(bdev, place->mem_type);
789 ticket = dma_resv_locking_ctx(bo->base.resv);
790 do {
791 ret = ttm_resource_alloc(bo, place, mem);
792 if (likely(!ret))
793 break;
794 if (unlikely(ret != -ENOSPC))
795 return ret;
796 ret = ttm_mem_evict_first(bdev, man, place, ctx,
797 ticket);
798 if (unlikely(ret != 0))
799 return ret;
800 } while (1);
801
802 return ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
803 }
804
805 /*
806 * Creates space for memory region @mem according to its type.
807 *
808 * This function first searches for free space in compatible memory types in
809 * the priority order defined by the driver. If free space isn't found, then
810 * ttm_bo_mem_force_space is attempted in priority order to evict and find
811 * space.
812 */
ttm_bo_mem_space(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx)813 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
814 struct ttm_placement *placement,
815 struct ttm_resource **mem,
816 struct ttm_operation_ctx *ctx)
817 {
818 struct ttm_device *bdev = bo->bdev;
819 bool type_found = false;
820 int i, ret;
821
822 ret = dma_resv_reserve_shared(bo->base.resv, 1);
823 if (unlikely(ret))
824 return ret;
825
826 for (i = 0; i < placement->num_placement; ++i) {
827 const struct ttm_place *place = &placement->placement[i];
828 struct ttm_resource_manager *man;
829
830 man = ttm_manager_type(bdev, place->mem_type);
831 if (!man || !ttm_resource_manager_used(man))
832 continue;
833
834 type_found = true;
835 ret = ttm_resource_alloc(bo, place, mem);
836 if (ret == -ENOSPC)
837 continue;
838 if (unlikely(ret))
839 goto error;
840
841 ret = ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
842 if (unlikely(ret)) {
843 ttm_resource_free(bo, mem);
844 if (ret == -EBUSY)
845 continue;
846
847 goto error;
848 }
849 return 0;
850 }
851
852 for (i = 0; i < placement->num_busy_placement; ++i) {
853 const struct ttm_place *place = &placement->busy_placement[i];
854 struct ttm_resource_manager *man;
855
856 man = ttm_manager_type(bdev, place->mem_type);
857 if (!man || !ttm_resource_manager_used(man))
858 continue;
859
860 type_found = true;
861 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
862 if (likely(!ret))
863 return 0;
864
865 if (ret && ret != -EBUSY)
866 goto error;
867 }
868
869 ret = -ENOMEM;
870 if (!type_found) {
871 pr_err(TTM_PFX "No compatible memory type found\n");
872 ret = -EINVAL;
873 }
874
875 error:
876 if (bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
877 ttm_bo_move_to_lru_tail_unlocked(bo);
878
879 return ret;
880 }
881 EXPORT_SYMBOL(ttm_bo_mem_space);
882
ttm_bo_move_buffer(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)883 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
884 struct ttm_placement *placement,
885 struct ttm_operation_ctx *ctx)
886 {
887 struct ttm_resource *mem;
888 struct ttm_place hop;
889 int ret;
890
891 dma_resv_assert_held(bo->base.resv);
892
893 /*
894 * Determine where to move the buffer.
895 *
896 * If driver determines move is going to need
897 * an extra step then it will return -EMULTIHOP
898 * and the buffer will be moved to the temporary
899 * stop and the driver will be called to make
900 * the second hop.
901 */
902 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
903 if (ret)
904 return ret;
905 bounce:
906 ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
907 if (ret == -EMULTIHOP) {
908 ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
909 if (ret)
910 goto out;
911 /* try and move to final place now. */
912 goto bounce;
913 }
914 out:
915 if (ret)
916 ttm_resource_free(bo, &mem);
917 return ret;
918 }
919
ttm_bo_validate(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)920 int ttm_bo_validate(struct ttm_buffer_object *bo,
921 struct ttm_placement *placement,
922 struct ttm_operation_ctx *ctx)
923 {
924 int ret;
925
926 dma_resv_assert_held(bo->base.resv);
927
928 /*
929 * Remove the backing store if no placement is given.
930 */
931 if (!placement->num_placement && !placement->num_busy_placement)
932 return ttm_bo_pipeline_gutting(bo);
933
934 /*
935 * Check whether we need to move buffer.
936 */
937 if (!ttm_resource_compat(bo->resource, placement)) {
938 ret = ttm_bo_move_buffer(bo, placement, ctx);
939 if (ret)
940 return ret;
941 }
942 /*
943 * We might need to add a TTM.
944 */
945 if (bo->resource->mem_type == TTM_PL_SYSTEM) {
946 ret = ttm_tt_create(bo, true);
947 if (ret)
948 return ret;
949 }
950 return 0;
951 }
952 EXPORT_SYMBOL(ttm_bo_validate);
953
ttm_bo_init_reserved(struct ttm_device * bdev,struct ttm_buffer_object * bo,size_t size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,struct ttm_operation_ctx * ctx,struct sg_table * sg,struct dma_resv * resv,void (* destroy)(struct ttm_buffer_object *))954 int ttm_bo_init_reserved(struct ttm_device *bdev,
955 struct ttm_buffer_object *bo,
956 size_t size,
957 enum ttm_bo_type type,
958 struct ttm_placement *placement,
959 uint32_t page_alignment,
960 struct ttm_operation_ctx *ctx,
961 struct sg_table *sg,
962 struct dma_resv *resv,
963 void (*destroy) (struct ttm_buffer_object *))
964 {
965 static const struct ttm_place sys_mem = { .mem_type = TTM_PL_SYSTEM };
966 bool locked;
967 int ret;
968
969 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
970
971 kref_init(&bo->kref);
972 INIT_LIST_HEAD(&bo->lru);
973 INIT_LIST_HEAD(&bo->ddestroy);
974 bo->bdev = bdev;
975 bo->type = type;
976 bo->page_alignment = page_alignment;
977 bo->moving = NULL;
978 bo->pin_count = 0;
979 bo->sg = sg;
980 if (resv) {
981 bo->base.resv = resv;
982 dma_resv_assert_held(bo->base.resv);
983 } else {
984 bo->base.resv = &bo->base._resv;
985 }
986 atomic_inc(&ttm_glob.bo_count);
987
988 ret = ttm_resource_alloc(bo, &sys_mem, &bo->resource);
989 if (unlikely(ret)) {
990 ttm_bo_put(bo);
991 return ret;
992 }
993
994 /*
995 * For ttm_bo_type_device buffers, allocate
996 * address space from the device.
997 */
998 if (bo->type == ttm_bo_type_device ||
999 bo->type == ttm_bo_type_sg)
1000 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1001 bo->resource->num_pages);
1002
1003 /* passed reservation objects should already be locked,
1004 * since otherwise lockdep will be angered in radeon.
1005 */
1006 if (!resv) {
1007 locked = dma_resv_trylock(bo->base.resv);
1008 WARN_ON(!locked);
1009 }
1010
1011 if (likely(!ret))
1012 ret = ttm_bo_validate(bo, placement, ctx);
1013
1014 if (unlikely(ret)) {
1015 if (!resv)
1016 ttm_bo_unreserve(bo);
1017
1018 ttm_bo_put(bo);
1019 return ret;
1020 }
1021
1022 ttm_bo_move_to_lru_tail_unlocked(bo);
1023
1024 return ret;
1025 }
1026 EXPORT_SYMBOL(ttm_bo_init_reserved);
1027
ttm_bo_init(struct ttm_device * bdev,struct ttm_buffer_object * bo,size_t size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,bool interruptible,struct sg_table * sg,struct dma_resv * resv,void (* destroy)(struct ttm_buffer_object *))1028 int ttm_bo_init(struct ttm_device *bdev,
1029 struct ttm_buffer_object *bo,
1030 size_t size,
1031 enum ttm_bo_type type,
1032 struct ttm_placement *placement,
1033 uint32_t page_alignment,
1034 bool interruptible,
1035 struct sg_table *sg,
1036 struct dma_resv *resv,
1037 void (*destroy) (struct ttm_buffer_object *))
1038 {
1039 struct ttm_operation_ctx ctx = { interruptible, false };
1040 int ret;
1041
1042 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1043 page_alignment, &ctx, sg, resv, destroy);
1044 if (ret)
1045 return ret;
1046
1047 if (!resv)
1048 ttm_bo_unreserve(bo);
1049
1050 return 0;
1051 }
1052 EXPORT_SYMBOL(ttm_bo_init);
1053
1054 /*
1055 * buffer object vm functions.
1056 */
1057
ttm_bo_unmap_virtual(struct ttm_buffer_object * bo)1058 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1059 {
1060 struct ttm_device *bdev = bo->bdev;
1061
1062 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1063 ttm_mem_io_free(bdev, bo->resource);
1064 }
1065 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1066
ttm_bo_wait(struct ttm_buffer_object * bo,bool interruptible,bool no_wait)1067 int ttm_bo_wait(struct ttm_buffer_object *bo,
1068 bool interruptible, bool no_wait)
1069 {
1070 long timeout = 15 * HZ;
1071
1072 if (no_wait) {
1073 if (dma_resv_test_signaled(bo->base.resv, true))
1074 return 0;
1075 else
1076 return -EBUSY;
1077 }
1078
1079 timeout = dma_resv_wait_timeout(bo->base.resv, true, interruptible,
1080 timeout);
1081 if (timeout < 0)
1082 return timeout;
1083
1084 if (timeout == 0)
1085 return -EBUSY;
1086
1087 dma_resv_add_excl_fence(bo->base.resv, NULL);
1088 return 0;
1089 }
1090 EXPORT_SYMBOL(ttm_bo_wait);
1091
ttm_bo_swapout(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,gfp_t gfp_flags)1092 int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1093 gfp_t gfp_flags)
1094 {
1095 struct ttm_place place;
1096 bool locked;
1097 int ret;
1098
1099 /*
1100 * While the bo may already reside in SYSTEM placement, set
1101 * SYSTEM as new placement to cover also the move further below.
1102 * The driver may use the fact that we're moving from SYSTEM
1103 * as an indication that we're about to swap out.
1104 */
1105 memset(&place, 0, sizeof(place));
1106 place.mem_type = bo->resource->mem_type;
1107 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1108 return -EBUSY;
1109
1110 if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1111 bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1112 bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED ||
1113 !ttm_bo_get_unless_zero(bo)) {
1114 if (locked)
1115 dma_resv_unlock(bo->base.resv);
1116 return -EBUSY;
1117 }
1118
1119 if (bo->deleted) {
1120 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1121 ttm_bo_put(bo);
1122 return ret == -EBUSY ? -ENOSPC : ret;
1123 }
1124
1125 ttm_bo_move_to_pinned(bo);
1126 /* TODO: Cleanup the locking */
1127 spin_unlock(&bo->bdev->lru_lock);
1128
1129 /*
1130 * Move to system cached
1131 */
1132 if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1133 struct ttm_operation_ctx ctx = { false, false };
1134 struct ttm_resource *evict_mem;
1135 struct ttm_place hop;
1136
1137 memset(&hop, 0, sizeof(hop));
1138 place.mem_type = TTM_PL_SYSTEM;
1139 ret = ttm_resource_alloc(bo, &place, &evict_mem);
1140 if (unlikely(ret))
1141 goto out;
1142
1143 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, &ctx, &hop);
1144 if (unlikely(ret != 0)) {
1145 WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1146 goto out;
1147 }
1148 }
1149
1150 /*
1151 * Make sure BO is idle.
1152 */
1153 ret = ttm_bo_wait(bo, false, false);
1154 if (unlikely(ret != 0))
1155 goto out;
1156
1157 ttm_bo_unmap_virtual(bo);
1158
1159 /*
1160 * Swap out. Buffer will be swapped in again as soon as
1161 * anyone tries to access a ttm page.
1162 */
1163 if (bo->bdev->funcs->swap_notify)
1164 bo->bdev->funcs->swap_notify(bo);
1165
1166 if (ttm_tt_is_populated(bo->ttm))
1167 ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1168 out:
1169
1170 /*
1171 * Unreserve without putting on LRU to avoid swapping out an
1172 * already swapped buffer.
1173 */
1174 if (locked)
1175 dma_resv_unlock(bo->base.resv);
1176 ttm_bo_put(bo);
1177 return ret == -EBUSY ? -ENOSPC : ret;
1178 }
1179
ttm_bo_tt_destroy(struct ttm_buffer_object * bo)1180 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1181 {
1182 if (bo->ttm == NULL)
1183 return;
1184
1185 ttm_tt_unpopulate(bo->bdev, bo->ttm);
1186 ttm_tt_destroy(bo->bdev, bo->ttm);
1187 bo->ttm = NULL;
1188 }
1189