Loading...
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.h>
35#include <drm/ttm/ttm_placement.h>
36#include <drm/ttm/ttm_tt.h>
37
38#include <linux/jiffies.h>
39#include <linux/slab.h>
40#include <linux/sched.h>
41#include <linux/mm.h>
42#include <linux/file.h>
43#include <linux/module.h>
44#include <linux/atomic.h>
45#include <linux/dma-resv.h>
46
47#include "ttm_module.h"
48
49static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
50 struct ttm_placement *placement)
51{
52 struct drm_printer p = drm_debug_printer(TTM_PFX);
53 struct ttm_resource_manager *man;
54 int i, mem_type;
55
56 for (i = 0; i < placement->num_placement; i++) {
57 mem_type = placement->placement[i].mem_type;
58 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
59 i, placement->placement[i].flags, mem_type);
60 man = ttm_manager_type(bo->bdev, mem_type);
61 ttm_resource_manager_debug(man, &p);
62 }
63}
64
65/**
66 * ttm_bo_move_to_lru_tail
67 *
68 * @bo: The buffer object.
69 *
70 * Move this BO to the tail of all lru lists used to lookup and reserve an
71 * object. This function must be called with struct ttm_global::lru_lock
72 * held, and is used to make a BO less likely to be considered for eviction.
73 */
74void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
75{
76 dma_resv_assert_held(bo->base.resv);
77
78 if (bo->resource)
79 ttm_resource_move_to_lru_tail(bo->resource);
80}
81EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
82
83/**
84 * ttm_bo_set_bulk_move - update BOs bulk move object
85 *
86 * @bo: The buffer object.
87 * @bulk: bulk move structure
88 *
89 * Update the BOs bulk move object, making sure that resources are added/removed
90 * as well. A bulk move allows to move many resource on the LRU at once,
91 * resulting in much less overhead of maintaining the LRU.
92 * The only requirement is that the resources stay together on the LRU and are
93 * never separated. This is enforces by setting the bulk_move structure on a BO.
94 * ttm_lru_bulk_move_tail() should be used to move all resources to the tail of
95 * their LRU list.
96 */
97void ttm_bo_set_bulk_move(struct ttm_buffer_object *bo,
98 struct ttm_lru_bulk_move *bulk)
99{
100 dma_resv_assert_held(bo->base.resv);
101
102 if (bo->bulk_move == bulk)
103 return;
104
105 spin_lock(&bo->bdev->lru_lock);
106 if (bo->resource)
107 ttm_resource_del_bulk_move(bo->resource, bo);
108 bo->bulk_move = bulk;
109 if (bo->resource)
110 ttm_resource_add_bulk_move(bo->resource, bo);
111 spin_unlock(&bo->bdev->lru_lock);
112}
113EXPORT_SYMBOL(ttm_bo_set_bulk_move);
114
115static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
116 struct ttm_resource *mem, bool evict,
117 struct ttm_operation_ctx *ctx,
118 struct ttm_place *hop)
119{
120 struct ttm_device *bdev = bo->bdev;
121 bool old_use_tt, new_use_tt;
122 int ret;
123
124 old_use_tt = !bo->resource || ttm_manager_type(bdev, bo->resource->mem_type)->use_tt;
125 new_use_tt = ttm_manager_type(bdev, mem->mem_type)->use_tt;
126
127 ttm_bo_unmap_virtual(bo);
128
129 /*
130 * Create and bind a ttm if required.
131 */
132
133 if (new_use_tt) {
134 /* Zero init the new TTM structure if the old location should
135 * have used one as well.
136 */
137 ret = ttm_tt_create(bo, old_use_tt);
138 if (ret)
139 goto out_err;
140
141 if (mem->mem_type != TTM_PL_SYSTEM) {
142 ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
143 if (ret)
144 goto out_err;
145 }
146 }
147
148 ret = dma_resv_reserve_fences(bo->base.resv, 1);
149 if (ret)
150 goto out_err;
151
152 ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
153 if (ret) {
154 if (ret == -EMULTIHOP)
155 return ret;
156 goto out_err;
157 }
158
159 ctx->bytes_moved += bo->base.size;
160 return 0;
161
162out_err:
163 if (!old_use_tt)
164 ttm_bo_tt_destroy(bo);
165
166 return ret;
167}
168
169/*
170 * Call bo::reserved.
171 * Will release GPU memory type usage on destruction.
172 * This is the place to put in driver specific hooks to release
173 * driver private resources.
174 * Will release the bo::reserved lock.
175 */
176
177static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
178{
179 if (bo->bdev->funcs->delete_mem_notify)
180 bo->bdev->funcs->delete_mem_notify(bo);
181
182 ttm_bo_tt_destroy(bo);
183 ttm_resource_free(bo, &bo->resource);
184}
185
186static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
187{
188 int r;
189
190 if (bo->base.resv == &bo->base._resv)
191 return 0;
192
193 BUG_ON(!dma_resv_trylock(&bo->base._resv));
194
195 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
196 dma_resv_unlock(&bo->base._resv);
197 if (r)
198 return r;
199
200 if (bo->type != ttm_bo_type_sg) {
201 /* This works because the BO is about to be destroyed and nobody
202 * reference it any more. The only tricky case is the trylock on
203 * the resv object while holding the lru_lock.
204 */
205 spin_lock(&bo->bdev->lru_lock);
206 bo->base.resv = &bo->base._resv;
207 spin_unlock(&bo->bdev->lru_lock);
208 }
209
210 return r;
211}
212
213static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
214{
215 struct dma_resv *resv = &bo->base._resv;
216 struct dma_resv_iter cursor;
217 struct dma_fence *fence;
218
219 dma_resv_iter_begin(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP);
220 dma_resv_for_each_fence_unlocked(&cursor, fence) {
221 if (!fence->ops->signaled)
222 dma_fence_enable_sw_signaling(fence);
223 }
224 dma_resv_iter_end(&cursor);
225}
226
227/**
228 * ttm_bo_cleanup_refs
229 * If bo idle, remove from lru lists, and unref.
230 * If not idle, block if possible.
231 *
232 * Must be called with lru_lock and reservation held, this function
233 * will drop the lru lock and optionally the reservation lock before returning.
234 *
235 * @bo: The buffer object to clean-up
236 * @interruptible: Any sleeps should occur interruptibly.
237 * @no_wait_gpu: Never wait for gpu. Return -EBUSY instead.
238 * @unlock_resv: Unlock the reservation lock as well.
239 */
240
241static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
242 bool interruptible, bool no_wait_gpu,
243 bool unlock_resv)
244{
245 struct dma_resv *resv = &bo->base._resv;
246 int ret;
247
248 if (dma_resv_test_signaled(resv, DMA_RESV_USAGE_BOOKKEEP))
249 ret = 0;
250 else
251 ret = -EBUSY;
252
253 if (ret && !no_wait_gpu) {
254 long lret;
255
256 if (unlock_resv)
257 dma_resv_unlock(bo->base.resv);
258 spin_unlock(&bo->bdev->lru_lock);
259
260 lret = dma_resv_wait_timeout(resv, DMA_RESV_USAGE_BOOKKEEP,
261 interruptible,
262 30 * HZ);
263
264 if (lret < 0)
265 return lret;
266 else if (lret == 0)
267 return -EBUSY;
268
269 spin_lock(&bo->bdev->lru_lock);
270 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
271 /*
272 * We raced, and lost, someone else holds the reservation now,
273 * and is probably busy in ttm_bo_cleanup_memtype_use.
274 *
275 * Even if it's not the case, because we finished waiting any
276 * delayed destruction would succeed, so just return success
277 * here.
278 */
279 spin_unlock(&bo->bdev->lru_lock);
280 return 0;
281 }
282 ret = 0;
283 }
284
285 if (ret) {
286 if (unlock_resv)
287 dma_resv_unlock(bo->base.resv);
288 spin_unlock(&bo->bdev->lru_lock);
289 return ret;
290 }
291
292 spin_unlock(&bo->bdev->lru_lock);
293 ttm_bo_cleanup_memtype_use(bo);
294
295 if (unlock_resv)
296 dma_resv_unlock(bo->base.resv);
297
298 return 0;
299}
300
301/*
302 * Block for the dma_resv object to become idle, lock the buffer and clean up
303 * the resource and tt object.
304 */
305static void ttm_bo_delayed_delete(struct work_struct *work)
306{
307 struct ttm_buffer_object *bo;
308
309 bo = container_of(work, typeof(*bo), delayed_delete);
310
311 dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP, false,
312 MAX_SCHEDULE_TIMEOUT);
313 dma_resv_lock(bo->base.resv, NULL);
314 ttm_bo_cleanup_memtype_use(bo);
315 dma_resv_unlock(bo->base.resv);
316 ttm_bo_put(bo);
317}
318
319static void ttm_bo_release(struct kref *kref)
320{
321 struct ttm_buffer_object *bo =
322 container_of(kref, struct ttm_buffer_object, kref);
323 struct ttm_device *bdev = bo->bdev;
324 int ret;
325
326 WARN_ON_ONCE(bo->pin_count);
327 WARN_ON_ONCE(bo->bulk_move);
328
329 if (!bo->deleted) {
330 ret = ttm_bo_individualize_resv(bo);
331 if (ret) {
332 /* Last resort, if we fail to allocate memory for the
333 * fences block for the BO to become idle
334 */
335 dma_resv_wait_timeout(bo->base.resv,
336 DMA_RESV_USAGE_BOOKKEEP, false,
337 30 * HZ);
338 }
339
340 if (bo->bdev->funcs->release_notify)
341 bo->bdev->funcs->release_notify(bo);
342
343 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
344 ttm_mem_io_free(bdev, bo->resource);
345
346 if (!dma_resv_test_signaled(bo->base.resv,
347 DMA_RESV_USAGE_BOOKKEEP) ||
348 (want_init_on_free() && (bo->ttm != NULL)) ||
349 !dma_resv_trylock(bo->base.resv)) {
350 /* The BO is not idle, resurrect it for delayed destroy */
351 ttm_bo_flush_all_fences(bo);
352 bo->deleted = true;
353
354 spin_lock(&bo->bdev->lru_lock);
355
356 /*
357 * Make pinned bos immediately available to
358 * shrinkers, now that they are queued for
359 * destruction.
360 *
361 * FIXME: QXL is triggering this. Can be removed when the
362 * driver is fixed.
363 */
364 if (bo->pin_count) {
365 bo->pin_count = 0;
366 ttm_resource_move_to_lru_tail(bo->resource);
367 }
368
369 kref_init(&bo->kref);
370 spin_unlock(&bo->bdev->lru_lock);
371
372 INIT_WORK(&bo->delayed_delete, ttm_bo_delayed_delete);
373
374 /* Schedule the worker on the closest NUMA node. This
375 * improves performance since system memory might be
376 * cleared on free and that is best done on a CPU core
377 * close to it.
378 */
379 queue_work_node(bdev->pool.nid, bdev->wq, &bo->delayed_delete);
380 return;
381 }
382
383 ttm_bo_cleanup_memtype_use(bo);
384 dma_resv_unlock(bo->base.resv);
385 }
386
387 atomic_dec(&ttm_glob.bo_count);
388 bo->destroy(bo);
389}
390
391/**
392 * ttm_bo_put
393 *
394 * @bo: The buffer object.
395 *
396 * Unreference a buffer object.
397 */
398void ttm_bo_put(struct ttm_buffer_object *bo)
399{
400 kref_put(&bo->kref, ttm_bo_release);
401}
402EXPORT_SYMBOL(ttm_bo_put);
403
404static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
405 struct ttm_resource **mem,
406 struct ttm_operation_ctx *ctx,
407 struct ttm_place *hop)
408{
409 struct ttm_placement hop_placement;
410 struct ttm_resource *hop_mem;
411 int ret;
412
413 hop_placement.num_placement = hop_placement.num_busy_placement = 1;
414 hop_placement.placement = hop_placement.busy_placement = hop;
415
416 /* find space in the bounce domain */
417 ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
418 if (ret)
419 return ret;
420 /* move to the bounce domain */
421 ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
422 if (ret) {
423 ttm_resource_free(bo, &hop_mem);
424 return ret;
425 }
426 return 0;
427}
428
429static int ttm_bo_evict(struct ttm_buffer_object *bo,
430 struct ttm_operation_ctx *ctx)
431{
432 struct ttm_device *bdev = bo->bdev;
433 struct ttm_resource *evict_mem;
434 struct ttm_placement placement;
435 struct ttm_place hop;
436 int ret = 0;
437
438 memset(&hop, 0, sizeof(hop));
439
440 dma_resv_assert_held(bo->base.resv);
441
442 placement.num_placement = 0;
443 placement.num_busy_placement = 0;
444 bdev->funcs->evict_flags(bo, &placement);
445
446 if (!placement.num_placement && !placement.num_busy_placement) {
447 ret = ttm_bo_wait_ctx(bo, ctx);
448 if (ret)
449 return ret;
450
451 /*
452 * Since we've already synced, this frees backing store
453 * immediately.
454 */
455 return ttm_bo_pipeline_gutting(bo);
456 }
457
458 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
459 if (ret) {
460 if (ret != -ERESTARTSYS) {
461 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
462 bo);
463 ttm_bo_mem_space_debug(bo, &placement);
464 }
465 goto out;
466 }
467
468 do {
469 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
470 if (ret != -EMULTIHOP)
471 break;
472
473 ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
474 } while (!ret);
475
476 if (ret) {
477 ttm_resource_free(bo, &evict_mem);
478 if (ret != -ERESTARTSYS && ret != -EINTR)
479 pr_err("Buffer eviction failed\n");
480 }
481out:
482 return ret;
483}
484
485/**
486 * ttm_bo_eviction_valuable
487 *
488 * @bo: The buffer object to evict
489 * @place: the placement we need to make room for
490 *
491 * Check if it is valuable to evict the BO to make room for the given placement.
492 */
493bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
494 const struct ttm_place *place)
495{
496 struct ttm_resource *res = bo->resource;
497 struct ttm_device *bdev = bo->bdev;
498
499 dma_resv_assert_held(bo->base.resv);
500 if (bo->resource->mem_type == TTM_PL_SYSTEM)
501 return true;
502
503 /* Don't evict this BO if it's outside of the
504 * requested placement range
505 */
506 return ttm_resource_intersects(bdev, res, place, bo->base.size);
507}
508EXPORT_SYMBOL(ttm_bo_eviction_valuable);
509
510/*
511 * Check the target bo is allowable to be evicted or swapout, including cases:
512 *
513 * a. if share same reservation object with ctx->resv, have assumption
514 * reservation objects should already be locked, so not lock again and
515 * return true directly when either the opreation allow_reserved_eviction
516 * or the target bo already is in delayed free list;
517 *
518 * b. Otherwise, trylock it.
519 */
520static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
521 struct ttm_operation_ctx *ctx,
522 const struct ttm_place *place,
523 bool *locked, bool *busy)
524{
525 bool ret = false;
526
527 if (bo->pin_count) {
528 *locked = false;
529 if (busy)
530 *busy = false;
531 return false;
532 }
533
534 if (bo->base.resv == ctx->resv) {
535 dma_resv_assert_held(bo->base.resv);
536 if (ctx->allow_res_evict)
537 ret = true;
538 *locked = false;
539 if (busy)
540 *busy = false;
541 } else {
542 ret = dma_resv_trylock(bo->base.resv);
543 *locked = ret;
544 if (busy)
545 *busy = !ret;
546 }
547
548 if (ret && place && (bo->resource->mem_type != place->mem_type ||
549 !bo->bdev->funcs->eviction_valuable(bo, place))) {
550 ret = false;
551 if (*locked) {
552 dma_resv_unlock(bo->base.resv);
553 *locked = false;
554 }
555 }
556
557 return ret;
558}
559
560/**
561 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
562 *
563 * @busy_bo: BO which couldn't be locked with trylock
564 * @ctx: operation context
565 * @ticket: acquire ticket
566 *
567 * Try to lock a busy buffer object to avoid failing eviction.
568 */
569static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
570 struct ttm_operation_ctx *ctx,
571 struct ww_acquire_ctx *ticket)
572{
573 int r;
574
575 if (!busy_bo || !ticket)
576 return -EBUSY;
577
578 if (ctx->interruptible)
579 r = dma_resv_lock_interruptible(busy_bo->base.resv,
580 ticket);
581 else
582 r = dma_resv_lock(busy_bo->base.resv, ticket);
583
584 /*
585 * TODO: It would be better to keep the BO locked until allocation is at
586 * least tried one more time, but that would mean a much larger rework
587 * of TTM.
588 */
589 if (!r)
590 dma_resv_unlock(busy_bo->base.resv);
591
592 return r == -EDEADLK ? -EBUSY : r;
593}
594
595int ttm_mem_evict_first(struct ttm_device *bdev,
596 struct ttm_resource_manager *man,
597 const struct ttm_place *place,
598 struct ttm_operation_ctx *ctx,
599 struct ww_acquire_ctx *ticket)
600{
601 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
602 struct ttm_resource_cursor cursor;
603 struct ttm_resource *res;
604 bool locked = false;
605 int ret;
606
607 spin_lock(&bdev->lru_lock);
608 ttm_resource_manager_for_each_res(man, &cursor, res) {
609 bool busy;
610
611 if (!ttm_bo_evict_swapout_allowable(res->bo, ctx, place,
612 &locked, &busy)) {
613 if (busy && !busy_bo && ticket !=
614 dma_resv_locking_ctx(res->bo->base.resv))
615 busy_bo = res->bo;
616 continue;
617 }
618
619 if (ttm_bo_get_unless_zero(res->bo)) {
620 bo = res->bo;
621 break;
622 }
623 if (locked)
624 dma_resv_unlock(res->bo->base.resv);
625 }
626
627 if (!bo) {
628 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
629 busy_bo = NULL;
630 spin_unlock(&bdev->lru_lock);
631 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
632 if (busy_bo)
633 ttm_bo_put(busy_bo);
634 return ret;
635 }
636
637 if (bo->deleted) {
638 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
639 ctx->no_wait_gpu, locked);
640 ttm_bo_put(bo);
641 return ret;
642 }
643
644 spin_unlock(&bdev->lru_lock);
645
646 ret = ttm_bo_evict(bo, ctx);
647 if (locked)
648 ttm_bo_unreserve(bo);
649 else
650 ttm_bo_move_to_lru_tail_unlocked(bo);
651
652 ttm_bo_put(bo);
653 return ret;
654}
655
656/**
657 * ttm_bo_pin - Pin the buffer object.
658 * @bo: The buffer object to pin
659 *
660 * Make sure the buffer is not evicted any more during memory pressure.
661 * @bo must be unpinned again by calling ttm_bo_unpin().
662 */
663void ttm_bo_pin(struct ttm_buffer_object *bo)
664{
665 dma_resv_assert_held(bo->base.resv);
666 WARN_ON_ONCE(!kref_read(&bo->kref));
667 spin_lock(&bo->bdev->lru_lock);
668 if (bo->resource)
669 ttm_resource_del_bulk_move(bo->resource, bo);
670 ++bo->pin_count;
671 spin_unlock(&bo->bdev->lru_lock);
672}
673EXPORT_SYMBOL(ttm_bo_pin);
674
675/**
676 * ttm_bo_unpin - Unpin the buffer object.
677 * @bo: The buffer object to unpin
678 *
679 * Allows the buffer object to be evicted again during memory pressure.
680 */
681void ttm_bo_unpin(struct ttm_buffer_object *bo)
682{
683 dma_resv_assert_held(bo->base.resv);
684 WARN_ON_ONCE(!kref_read(&bo->kref));
685 if (WARN_ON_ONCE(!bo->pin_count))
686 return;
687
688 spin_lock(&bo->bdev->lru_lock);
689 --bo->pin_count;
690 if (bo->resource)
691 ttm_resource_add_bulk_move(bo->resource, bo);
692 spin_unlock(&bo->bdev->lru_lock);
693}
694EXPORT_SYMBOL(ttm_bo_unpin);
695
696/*
697 * Add the last move fence to the BO as kernel dependency and reserve a new
698 * fence slot.
699 */
700static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
701 struct ttm_resource_manager *man,
702 struct ttm_resource *mem,
703 bool no_wait_gpu)
704{
705 struct dma_fence *fence;
706 int ret;
707
708 spin_lock(&man->move_lock);
709 fence = dma_fence_get(man->move);
710 spin_unlock(&man->move_lock);
711
712 if (!fence)
713 return 0;
714
715 if (no_wait_gpu) {
716 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
717 dma_fence_put(fence);
718 return ret;
719 }
720
721 dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
722
723 ret = dma_resv_reserve_fences(bo->base.resv, 1);
724 dma_fence_put(fence);
725 return ret;
726}
727
728/*
729 * Repeatedly evict memory from the LRU for @mem_type until we create enough
730 * space, or we've evicted everything and there isn't enough space.
731 */
732static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
733 const struct ttm_place *place,
734 struct ttm_resource **mem,
735 struct ttm_operation_ctx *ctx)
736{
737 struct ttm_device *bdev = bo->bdev;
738 struct ttm_resource_manager *man;
739 struct ww_acquire_ctx *ticket;
740 int ret;
741
742 man = ttm_manager_type(bdev, place->mem_type);
743 ticket = dma_resv_locking_ctx(bo->base.resv);
744 do {
745 ret = ttm_resource_alloc(bo, place, mem);
746 if (likely(!ret))
747 break;
748 if (unlikely(ret != -ENOSPC))
749 return ret;
750 ret = ttm_mem_evict_first(bdev, man, place, ctx,
751 ticket);
752 if (unlikely(ret != 0))
753 return ret;
754 } while (1);
755
756 return ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
757}
758
759/**
760 * ttm_bo_mem_space
761 *
762 * @bo: Pointer to a struct ttm_buffer_object. the data of which
763 * we want to allocate space for.
764 * @placement: Proposed new placement for the buffer object.
765 * @mem: A struct ttm_resource.
766 * @ctx: if and how to sleep, lock buffers and alloc memory
767 *
768 * Allocate memory space for the buffer object pointed to by @bo, using
769 * the placement flags in @placement, potentially evicting other idle buffer objects.
770 * This function may sleep while waiting for space to become available.
771 * Returns:
772 * -EBUSY: No space available (only if no_wait == 1).
773 * -ENOMEM: Could not allocate memory for the buffer object, either due to
774 * fragmentation or concurrent allocators.
775 * -ERESTARTSYS: An interruptible sleep was interrupted by a signal.
776 */
777int ttm_bo_mem_space(struct ttm_buffer_object *bo,
778 struct ttm_placement *placement,
779 struct ttm_resource **mem,
780 struct ttm_operation_ctx *ctx)
781{
782 struct ttm_device *bdev = bo->bdev;
783 bool type_found = false;
784 int i, ret;
785
786 ret = dma_resv_reserve_fences(bo->base.resv, 1);
787 if (unlikely(ret))
788 return ret;
789
790 for (i = 0; i < placement->num_placement; ++i) {
791 const struct ttm_place *place = &placement->placement[i];
792 struct ttm_resource_manager *man;
793
794 man = ttm_manager_type(bdev, place->mem_type);
795 if (!man || !ttm_resource_manager_used(man))
796 continue;
797
798 type_found = true;
799 ret = ttm_resource_alloc(bo, place, mem);
800 if (ret == -ENOSPC)
801 continue;
802 if (unlikely(ret))
803 goto error;
804
805 ret = ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
806 if (unlikely(ret)) {
807 ttm_resource_free(bo, mem);
808 if (ret == -EBUSY)
809 continue;
810
811 goto error;
812 }
813 return 0;
814 }
815
816 for (i = 0; i < placement->num_busy_placement; ++i) {
817 const struct ttm_place *place = &placement->busy_placement[i];
818 struct ttm_resource_manager *man;
819
820 man = ttm_manager_type(bdev, place->mem_type);
821 if (!man || !ttm_resource_manager_used(man))
822 continue;
823
824 type_found = true;
825 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
826 if (likely(!ret))
827 return 0;
828
829 if (ret && ret != -EBUSY)
830 goto error;
831 }
832
833 ret = -ENOMEM;
834 if (!type_found) {
835 pr_err(TTM_PFX "No compatible memory type found\n");
836 ret = -EINVAL;
837 }
838
839error:
840 return ret;
841}
842EXPORT_SYMBOL(ttm_bo_mem_space);
843
844static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
845 struct ttm_placement *placement,
846 struct ttm_operation_ctx *ctx)
847{
848 struct ttm_resource *mem;
849 struct ttm_place hop;
850 int ret;
851
852 dma_resv_assert_held(bo->base.resv);
853
854 /*
855 * Determine where to move the buffer.
856 *
857 * If driver determines move is going to need
858 * an extra step then it will return -EMULTIHOP
859 * and the buffer will be moved to the temporary
860 * stop and the driver will be called to make
861 * the second hop.
862 */
863 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
864 if (ret)
865 return ret;
866bounce:
867 ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
868 if (ret == -EMULTIHOP) {
869 ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
870 if (ret)
871 goto out;
872 /* try and move to final place now. */
873 goto bounce;
874 }
875out:
876 if (ret)
877 ttm_resource_free(bo, &mem);
878 return ret;
879}
880
881/**
882 * ttm_bo_validate
883 *
884 * @bo: The buffer object.
885 * @placement: Proposed placement for the buffer object.
886 * @ctx: validation parameters.
887 *
888 * Changes placement and caching policy of the buffer object
889 * according proposed placement.
890 * Returns
891 * -EINVAL on invalid proposed placement.
892 * -ENOMEM on out-of-memory condition.
893 * -EBUSY if no_wait is true and buffer busy.
894 * -ERESTARTSYS if interrupted by a signal.
895 */
896int ttm_bo_validate(struct ttm_buffer_object *bo,
897 struct ttm_placement *placement,
898 struct ttm_operation_ctx *ctx)
899{
900 int ret;
901
902 dma_resv_assert_held(bo->base.resv);
903
904 /*
905 * Remove the backing store if no placement is given.
906 */
907 if (!placement->num_placement && !placement->num_busy_placement)
908 return ttm_bo_pipeline_gutting(bo);
909
910 /* Check whether we need to move buffer. */
911 if (bo->resource && ttm_resource_compat(bo->resource, placement))
912 return 0;
913
914 /* Moving of pinned BOs is forbidden */
915 if (bo->pin_count)
916 return -EINVAL;
917
918 ret = ttm_bo_move_buffer(bo, placement, ctx);
919 if (ret)
920 return ret;
921
922 /*
923 * We might need to add a TTM.
924 */
925 if (!bo->resource || bo->resource->mem_type == TTM_PL_SYSTEM) {
926 ret = ttm_tt_create(bo, true);
927 if (ret)
928 return ret;
929 }
930 return 0;
931}
932EXPORT_SYMBOL(ttm_bo_validate);
933
934/**
935 * ttm_bo_init_reserved
936 *
937 * @bdev: Pointer to a ttm_device struct.
938 * @bo: Pointer to a ttm_buffer_object to be initialized.
939 * @type: Requested type of buffer object.
940 * @placement: Initial placement for buffer object.
941 * @alignment: Data alignment in pages.
942 * @ctx: TTM operation context for memory allocation.
943 * @sg: Scatter-gather table.
944 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
945 * @destroy: Destroy function. Use NULL for kfree().
946 *
947 * This function initializes a pre-allocated struct ttm_buffer_object.
948 * As this object may be part of a larger structure, this function,
949 * together with the @destroy function, enables driver-specific objects
950 * derived from a ttm_buffer_object.
951 *
952 * On successful return, the caller owns an object kref to @bo. The kref and
953 * list_kref are usually set to 1, but note that in some situations, other
954 * tasks may already be holding references to @bo as well.
955 * Furthermore, if resv == NULL, the buffer's reservation lock will be held,
956 * and it is the caller's responsibility to call ttm_bo_unreserve.
957 *
958 * If a failure occurs, the function will call the @destroy function. Thus,
959 * after a failure, dereferencing @bo is illegal and will likely cause memory
960 * corruption.
961 *
962 * Returns
963 * -ENOMEM: Out of memory.
964 * -EINVAL: Invalid placement flags.
965 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
966 */
967int ttm_bo_init_reserved(struct ttm_device *bdev, struct ttm_buffer_object *bo,
968 enum ttm_bo_type type, struct ttm_placement *placement,
969 uint32_t alignment, struct ttm_operation_ctx *ctx,
970 struct sg_table *sg, struct dma_resv *resv,
971 void (*destroy) (struct ttm_buffer_object *))
972{
973 int ret;
974
975 kref_init(&bo->kref);
976 bo->bdev = bdev;
977 bo->type = type;
978 bo->page_alignment = alignment;
979 bo->destroy = destroy;
980 bo->pin_count = 0;
981 bo->sg = sg;
982 bo->bulk_move = NULL;
983 if (resv)
984 bo->base.resv = resv;
985 else
986 bo->base.resv = &bo->base._resv;
987 atomic_inc(&ttm_glob.bo_count);
988
989 /*
990 * For ttm_bo_type_device buffers, allocate
991 * address space from the device.
992 */
993 if (bo->type == ttm_bo_type_device || bo->type == ttm_bo_type_sg) {
994 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
995 PFN_UP(bo->base.size));
996 if (ret)
997 goto err_put;
998 }
999
1000 /* passed reservation objects should already be locked,
1001 * since otherwise lockdep will be angered in radeon.
1002 */
1003 if (!resv)
1004 WARN_ON(!dma_resv_trylock(bo->base.resv));
1005 else
1006 dma_resv_assert_held(resv);
1007
1008 ret = ttm_bo_validate(bo, placement, ctx);
1009 if (unlikely(ret))
1010 goto err_unlock;
1011
1012 return 0;
1013
1014err_unlock:
1015 if (!resv)
1016 dma_resv_unlock(bo->base.resv);
1017
1018err_put:
1019 ttm_bo_put(bo);
1020 return ret;
1021}
1022EXPORT_SYMBOL(ttm_bo_init_reserved);
1023
1024/**
1025 * ttm_bo_init_validate
1026 *
1027 * @bdev: Pointer to a ttm_device struct.
1028 * @bo: Pointer to a ttm_buffer_object to be initialized.
1029 * @type: Requested type of buffer object.
1030 * @placement: Initial placement for buffer object.
1031 * @alignment: Data alignment in pages.
1032 * @interruptible: If needing to sleep to wait for GPU resources,
1033 * sleep interruptible.
1034 * pinned in physical memory. If this behaviour is not desired, this member
1035 * holds a pointer to a persistent shmem object. Typically, this would
1036 * point to the shmem object backing a GEM object if TTM is used to back a
1037 * GEM user interface.
1038 * @sg: Scatter-gather table.
1039 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
1040 * @destroy: Destroy function. Use NULL for kfree().
1041 *
1042 * This function initializes a pre-allocated struct ttm_buffer_object.
1043 * As this object may be part of a larger structure, this function,
1044 * together with the @destroy function,
1045 * enables driver-specific objects derived from a ttm_buffer_object.
1046 *
1047 * On successful return, the caller owns an object kref to @bo. The kref and
1048 * list_kref are usually set to 1, but note that in some situations, other
1049 * tasks may already be holding references to @bo as well.
1050 *
1051 * If a failure occurs, the function will call the @destroy function, Thus,
1052 * after a failure, dereferencing @bo is illegal and will likely cause memory
1053 * corruption.
1054 *
1055 * Returns
1056 * -ENOMEM: Out of memory.
1057 * -EINVAL: Invalid placement flags.
1058 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
1059 */
1060int ttm_bo_init_validate(struct ttm_device *bdev, struct ttm_buffer_object *bo,
1061 enum ttm_bo_type type, struct ttm_placement *placement,
1062 uint32_t alignment, bool interruptible,
1063 struct sg_table *sg, struct dma_resv *resv,
1064 void (*destroy) (struct ttm_buffer_object *))
1065{
1066 struct ttm_operation_ctx ctx = { interruptible, false };
1067 int ret;
1068
1069 ret = ttm_bo_init_reserved(bdev, bo, type, placement, alignment, &ctx,
1070 sg, resv, destroy);
1071 if (ret)
1072 return ret;
1073
1074 if (!resv)
1075 ttm_bo_unreserve(bo);
1076
1077 return 0;
1078}
1079EXPORT_SYMBOL(ttm_bo_init_validate);
1080
1081/*
1082 * buffer object vm functions.
1083 */
1084
1085/**
1086 * ttm_bo_unmap_virtual
1087 *
1088 * @bo: tear down the virtual mappings for this BO
1089 */
1090void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1091{
1092 struct ttm_device *bdev = bo->bdev;
1093
1094 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1095 ttm_mem_io_free(bdev, bo->resource);
1096}
1097EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1098
1099/**
1100 * ttm_bo_wait_ctx - wait for buffer idle.
1101 *
1102 * @bo: The buffer object.
1103 * @ctx: defines how to wait
1104 *
1105 * Waits for the buffer to be idle. Used timeout depends on the context.
1106 * Returns -EBUSY if wait timed outt, -ERESTARTSYS if interrupted by a signal or
1107 * zero on success.
1108 */
1109int ttm_bo_wait_ctx(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx)
1110{
1111 long ret;
1112
1113 if (ctx->no_wait_gpu) {
1114 if (dma_resv_test_signaled(bo->base.resv,
1115 DMA_RESV_USAGE_BOOKKEEP))
1116 return 0;
1117 else
1118 return -EBUSY;
1119 }
1120
1121 ret = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
1122 ctx->interruptible, 15 * HZ);
1123 if (unlikely(ret < 0))
1124 return ret;
1125 if (unlikely(ret == 0))
1126 return -EBUSY;
1127 return 0;
1128}
1129EXPORT_SYMBOL(ttm_bo_wait_ctx);
1130
1131int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1132 gfp_t gfp_flags)
1133{
1134 struct ttm_place place;
1135 bool locked;
1136 long ret;
1137
1138 /*
1139 * While the bo may already reside in SYSTEM placement, set
1140 * SYSTEM as new placement to cover also the move further below.
1141 * The driver may use the fact that we're moving from SYSTEM
1142 * as an indication that we're about to swap out.
1143 */
1144 memset(&place, 0, sizeof(place));
1145 place.mem_type = bo->resource->mem_type;
1146 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1147 return -EBUSY;
1148
1149 if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1150 bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1151 bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED ||
1152 !ttm_bo_get_unless_zero(bo)) {
1153 if (locked)
1154 dma_resv_unlock(bo->base.resv);
1155 return -EBUSY;
1156 }
1157
1158 if (bo->deleted) {
1159 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1160 ttm_bo_put(bo);
1161 return ret == -EBUSY ? -ENOSPC : ret;
1162 }
1163
1164 /* TODO: Cleanup the locking */
1165 spin_unlock(&bo->bdev->lru_lock);
1166
1167 /*
1168 * Move to system cached
1169 */
1170 if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1171 struct ttm_resource *evict_mem;
1172 struct ttm_place hop;
1173
1174 memset(&hop, 0, sizeof(hop));
1175 place.mem_type = TTM_PL_SYSTEM;
1176 ret = ttm_resource_alloc(bo, &place, &evict_mem);
1177 if (unlikely(ret))
1178 goto out;
1179
1180 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
1181 if (unlikely(ret != 0)) {
1182 WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1183 ttm_resource_free(bo, &evict_mem);
1184 goto out;
1185 }
1186 }
1187
1188 /*
1189 * Make sure BO is idle.
1190 */
1191 ret = ttm_bo_wait_ctx(bo, ctx);
1192 if (unlikely(ret != 0))
1193 goto out;
1194
1195 ttm_bo_unmap_virtual(bo);
1196
1197 /*
1198 * Swap out. Buffer will be swapped in again as soon as
1199 * anyone tries to access a ttm page.
1200 */
1201 if (bo->bdev->funcs->swap_notify)
1202 bo->bdev->funcs->swap_notify(bo);
1203
1204 if (ttm_tt_is_populated(bo->ttm))
1205 ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1206out:
1207
1208 /*
1209 * Unreserve without putting on LRU to avoid swapping out an
1210 * already swapped buffer.
1211 */
1212 if (locked)
1213 dma_resv_unlock(bo->base.resv);
1214 ttm_bo_put(bo);
1215 return ret == -EBUSY ? -ENOSPC : ret;
1216}
1217
1218void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1219{
1220 if (bo->ttm == NULL)
1221 return;
1222
1223 ttm_tt_unpopulate(bo->bdev, bo->ttm);
1224 ttm_tt_destroy(bo->bdev, bo->ttm);
1225 bo->ttm = NULL;
1226}
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
48static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
49static void ttm_bo_global_kobj_release(struct kobject *kobj);
50
51static struct attribute ttm_bo_count = {
52 .name = "bo_count",
53 .mode = S_IRUGO
54};
55
56static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
57{
58 int i;
59
60 for (i = 0; i <= TTM_PL_PRIV5; i++)
61 if (flags & (1 << i)) {
62 *mem_type = i;
63 return 0;
64 }
65 return -EINVAL;
66}
67
68static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
69{
70 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
71
72 pr_err(" has_type: %d\n", man->has_type);
73 pr_err(" use_type: %d\n", man->use_type);
74 pr_err(" flags: 0x%08X\n", man->flags);
75 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset);
76 pr_err(" size: %llu\n", man->size);
77 pr_err(" available_caching: 0x%08X\n", man->available_caching);
78 pr_err(" default_caching: 0x%08X\n", man->default_caching);
79 if (mem_type != TTM_PL_SYSTEM)
80 (*man->func->debug)(man, TTM_PFX);
81}
82
83static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
84 struct ttm_placement *placement)
85{
86 int i, ret, mem_type;
87
88 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
89 bo, bo->mem.num_pages, bo->mem.size >> 10,
90 bo->mem.size >> 20);
91 for (i = 0; i < placement->num_placement; i++) {
92 ret = ttm_mem_type_from_flags(placement->placement[i],
93 &mem_type);
94 if (ret)
95 return;
96 pr_err(" placement[%d]=0x%08X (%d)\n",
97 i, placement->placement[i], mem_type);
98 ttm_mem_type_debug(bo->bdev, mem_type);
99 }
100}
101
102static ssize_t ttm_bo_global_show(struct kobject *kobj,
103 struct attribute *attr,
104 char *buffer)
105{
106 struct ttm_bo_global *glob =
107 container_of(kobj, struct ttm_bo_global, kobj);
108
109 return snprintf(buffer, PAGE_SIZE, "%lu\n",
110 (unsigned long) atomic_read(&glob->bo_count));
111}
112
113static struct attribute *ttm_bo_global_attrs[] = {
114 &ttm_bo_count,
115 NULL
116};
117
118static const struct sysfs_ops ttm_bo_global_ops = {
119 .show = &ttm_bo_global_show
120};
121
122static struct kobj_type ttm_bo_glob_kobj_type = {
123 .release = &ttm_bo_global_kobj_release,
124 .sysfs_ops = &ttm_bo_global_ops,
125 .default_attrs = ttm_bo_global_attrs
126};
127
128
129static inline uint32_t ttm_bo_type_flags(unsigned type)
130{
131 return 1 << (type);
132}
133
134static void ttm_bo_release_list(struct kref *list_kref)
135{
136 struct ttm_buffer_object *bo =
137 container_of(list_kref, struct ttm_buffer_object, list_kref);
138 struct ttm_bo_device *bdev = bo->bdev;
139 size_t acc_size = bo->acc_size;
140
141 BUG_ON(atomic_read(&bo->list_kref.refcount));
142 BUG_ON(atomic_read(&bo->kref.refcount));
143 BUG_ON(atomic_read(&bo->cpu_writers));
144 BUG_ON(bo->sync_obj != NULL);
145 BUG_ON(bo->mem.mm_node != NULL);
146 BUG_ON(!list_empty(&bo->lru));
147 BUG_ON(!list_empty(&bo->ddestroy));
148
149 if (bo->ttm)
150 ttm_tt_destroy(bo->ttm);
151 atomic_dec(&bo->glob->bo_count);
152 if (bo->resv == &bo->ttm_resv)
153 reservation_object_fini(&bo->ttm_resv);
154 mutex_destroy(&bo->wu_mutex);
155 if (bo->destroy)
156 bo->destroy(bo);
157 else {
158 kfree(bo);
159 }
160 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
161}
162
163void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
164{
165 struct ttm_bo_device *bdev = bo->bdev;
166 struct ttm_mem_type_manager *man;
167
168 lockdep_assert_held(&bo->resv->lock.base);
169
170 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
171
172 BUG_ON(!list_empty(&bo->lru));
173
174 man = &bdev->man[bo->mem.mem_type];
175 list_add_tail(&bo->lru, &man->lru);
176 kref_get(&bo->list_kref);
177
178 if (bo->ttm != NULL) {
179 list_add_tail(&bo->swap, &bo->glob->swap_lru);
180 kref_get(&bo->list_kref);
181 }
182 }
183}
184EXPORT_SYMBOL(ttm_bo_add_to_lru);
185
186int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
187{
188 int put_count = 0;
189
190 if (!list_empty(&bo->swap)) {
191 list_del_init(&bo->swap);
192 ++put_count;
193 }
194 if (!list_empty(&bo->lru)) {
195 list_del_init(&bo->lru);
196 ++put_count;
197 }
198
199 /*
200 * TODO: Add a driver hook to delete from
201 * driver-specific LRU's here.
202 */
203
204 return put_count;
205}
206
207static void ttm_bo_ref_bug(struct kref *list_kref)
208{
209 BUG();
210}
211
212void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
213 bool never_free)
214{
215 kref_sub(&bo->list_kref, count,
216 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
217}
218
219void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
220{
221 int put_count;
222
223 spin_lock(&bo->glob->lru_lock);
224 put_count = ttm_bo_del_from_lru(bo);
225 spin_unlock(&bo->glob->lru_lock);
226 ttm_bo_list_ref_sub(bo, put_count, true);
227}
228EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
229
230/*
231 * Call bo->mutex locked.
232 */
233static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
234{
235 struct ttm_bo_device *bdev = bo->bdev;
236 struct ttm_bo_global *glob = bo->glob;
237 int ret = 0;
238 uint32_t page_flags = 0;
239
240 TTM_ASSERT_LOCKED(&bo->mutex);
241 bo->ttm = NULL;
242
243 if (bdev->need_dma32)
244 page_flags |= TTM_PAGE_FLAG_DMA32;
245
246 switch (bo->type) {
247 case ttm_bo_type_device:
248 if (zero_alloc)
249 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
250 case ttm_bo_type_kernel:
251 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
252 page_flags, glob->dummy_read_page);
253 if (unlikely(bo->ttm == NULL))
254 ret = -ENOMEM;
255 break;
256 case ttm_bo_type_sg:
257 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
258 page_flags | TTM_PAGE_FLAG_SG,
259 glob->dummy_read_page);
260 if (unlikely(bo->ttm == NULL)) {
261 ret = -ENOMEM;
262 break;
263 }
264 bo->ttm->sg = bo->sg;
265 break;
266 default:
267 pr_err("Illegal buffer object type\n");
268 ret = -EINVAL;
269 break;
270 }
271
272 return ret;
273}
274
275static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
276 struct ttm_mem_reg *mem,
277 bool evict, bool interruptible,
278 bool no_wait_gpu)
279{
280 struct ttm_bo_device *bdev = bo->bdev;
281 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
282 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
283 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
284 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
285 int ret = 0;
286
287 if (old_is_pci || new_is_pci ||
288 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
289 ret = ttm_mem_io_lock(old_man, true);
290 if (unlikely(ret != 0))
291 goto out_err;
292 ttm_bo_unmap_virtual_locked(bo);
293 ttm_mem_io_unlock(old_man);
294 }
295
296 /*
297 * Create and bind a ttm if required.
298 */
299
300 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
301 if (bo->ttm == NULL) {
302 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
303 ret = ttm_bo_add_ttm(bo, zero);
304 if (ret)
305 goto out_err;
306 }
307
308 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
309 if (ret)
310 goto out_err;
311
312 if (mem->mem_type != TTM_PL_SYSTEM) {
313 ret = ttm_tt_bind(bo->ttm, mem);
314 if (ret)
315 goto out_err;
316 }
317
318 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
319 if (bdev->driver->move_notify)
320 bdev->driver->move_notify(bo, mem);
321 bo->mem = *mem;
322 mem->mm_node = NULL;
323 goto moved;
324 }
325 }
326
327 if (bdev->driver->move_notify)
328 bdev->driver->move_notify(bo, mem);
329
330 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
331 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
332 ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
333 else if (bdev->driver->move)
334 ret = bdev->driver->move(bo, evict, interruptible,
335 no_wait_gpu, mem);
336 else
337 ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
338
339 if (ret) {
340 if (bdev->driver->move_notify) {
341 struct ttm_mem_reg tmp_mem = *mem;
342 *mem = bo->mem;
343 bo->mem = tmp_mem;
344 bdev->driver->move_notify(bo, mem);
345 bo->mem = *mem;
346 *mem = tmp_mem;
347 }
348
349 goto out_err;
350 }
351
352moved:
353 if (bo->evicted) {
354 if (bdev->driver->invalidate_caches) {
355 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
356 if (ret)
357 pr_err("Can not flush read caches\n");
358 }
359 bo->evicted = false;
360 }
361
362 if (bo->mem.mm_node) {
363 bo->offset = (bo->mem.start << PAGE_SHIFT) +
364 bdev->man[bo->mem.mem_type].gpu_offset;
365 bo->cur_placement = bo->mem.placement;
366 } else
367 bo->offset = 0;
368
369 return 0;
370
371out_err:
372 new_man = &bdev->man[bo->mem.mem_type];
373 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
374 ttm_tt_unbind(bo->ttm);
375 ttm_tt_destroy(bo->ttm);
376 bo->ttm = NULL;
377 }
378
379 return ret;
380}
381
382/**
383 * Call bo::reserved.
384 * Will release GPU memory type usage on destruction.
385 * This is the place to put in driver specific hooks to release
386 * driver private resources.
387 * Will release the bo::reserved lock.
388 */
389
390static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
391{
392 if (bo->bdev->driver->move_notify)
393 bo->bdev->driver->move_notify(bo, NULL);
394
395 if (bo->ttm) {
396 ttm_tt_unbind(bo->ttm);
397 ttm_tt_destroy(bo->ttm);
398 bo->ttm = NULL;
399 }
400 ttm_bo_mem_put(bo, &bo->mem);
401
402 ww_mutex_unlock (&bo->resv->lock);
403}
404
405static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
406{
407 struct ttm_bo_device *bdev = bo->bdev;
408 struct ttm_bo_global *glob = bo->glob;
409 struct ttm_bo_driver *driver = bdev->driver;
410 void *sync_obj = NULL;
411 int put_count;
412 int ret;
413
414 spin_lock(&glob->lru_lock);
415 ret = __ttm_bo_reserve(bo, false, true, false, 0);
416
417 spin_lock(&bdev->fence_lock);
418 (void) ttm_bo_wait(bo, false, false, true);
419 if (!ret && !bo->sync_obj) {
420 spin_unlock(&bdev->fence_lock);
421 put_count = ttm_bo_del_from_lru(bo);
422
423 spin_unlock(&glob->lru_lock);
424 ttm_bo_cleanup_memtype_use(bo);
425
426 ttm_bo_list_ref_sub(bo, put_count, true);
427
428 return;
429 }
430 if (bo->sync_obj)
431 sync_obj = driver->sync_obj_ref(bo->sync_obj);
432 spin_unlock(&bdev->fence_lock);
433
434 if (!ret) {
435
436 /*
437 * Make NO_EVICT bos immediately available to
438 * shrinkers, now that they are queued for
439 * destruction.
440 */
441 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
442 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
443 ttm_bo_add_to_lru(bo);
444 }
445
446 __ttm_bo_unreserve(bo);
447 }
448
449 kref_get(&bo->list_kref);
450 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
451 spin_unlock(&glob->lru_lock);
452
453 if (sync_obj) {
454 driver->sync_obj_flush(sync_obj);
455 driver->sync_obj_unref(&sync_obj);
456 }
457 schedule_delayed_work(&bdev->wq,
458 ((HZ / 100) < 1) ? 1 : HZ / 100);
459}
460
461/**
462 * function ttm_bo_cleanup_refs_and_unlock
463 * If bo idle, remove from delayed- and lru lists, and unref.
464 * If not idle, do nothing.
465 *
466 * Must be called with lru_lock and reservation held, this function
467 * will drop both before returning.
468 *
469 * @interruptible Any sleeps should occur interruptibly.
470 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
471 */
472
473static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
474 bool interruptible,
475 bool no_wait_gpu)
476{
477 struct ttm_bo_device *bdev = bo->bdev;
478 struct ttm_bo_driver *driver = bdev->driver;
479 struct ttm_bo_global *glob = bo->glob;
480 int put_count;
481 int ret;
482
483 spin_lock(&bdev->fence_lock);
484 ret = ttm_bo_wait(bo, false, false, true);
485
486 if (ret && !no_wait_gpu) {
487 void *sync_obj;
488
489 /*
490 * Take a reference to the fence and unreserve,
491 * at this point the buffer should be dead, so
492 * no new sync objects can be attached.
493 */
494 sync_obj = driver->sync_obj_ref(bo->sync_obj);
495 spin_unlock(&bdev->fence_lock);
496
497 __ttm_bo_unreserve(bo);
498 spin_unlock(&glob->lru_lock);
499
500 ret = driver->sync_obj_wait(sync_obj, false, interruptible);
501 driver->sync_obj_unref(&sync_obj);
502 if (ret)
503 return ret;
504
505 /*
506 * remove sync_obj with ttm_bo_wait, the wait should be
507 * finished, and no new wait object should have been added.
508 */
509 spin_lock(&bdev->fence_lock);
510 ret = ttm_bo_wait(bo, false, false, true);
511 WARN_ON(ret);
512 spin_unlock(&bdev->fence_lock);
513 if (ret)
514 return ret;
515
516 spin_lock(&glob->lru_lock);
517 ret = __ttm_bo_reserve(bo, false, true, false, 0);
518
519 /*
520 * We raced, and lost, someone else holds the reservation now,
521 * and is probably busy in ttm_bo_cleanup_memtype_use.
522 *
523 * Even if it's not the case, because we finished waiting any
524 * delayed destruction would succeed, so just return success
525 * here.
526 */
527 if (ret) {
528 spin_unlock(&glob->lru_lock);
529 return 0;
530 }
531 } else
532 spin_unlock(&bdev->fence_lock);
533
534 if (ret || unlikely(list_empty(&bo->ddestroy))) {
535 __ttm_bo_unreserve(bo);
536 spin_unlock(&glob->lru_lock);
537 return ret;
538 }
539
540 put_count = ttm_bo_del_from_lru(bo);
541 list_del_init(&bo->ddestroy);
542 ++put_count;
543
544 spin_unlock(&glob->lru_lock);
545 ttm_bo_cleanup_memtype_use(bo);
546
547 ttm_bo_list_ref_sub(bo, put_count, true);
548
549 return 0;
550}
551
552/**
553 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
554 * encountered buffers.
555 */
556
557static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
558{
559 struct ttm_bo_global *glob = bdev->glob;
560 struct ttm_buffer_object *entry = NULL;
561 int ret = 0;
562
563 spin_lock(&glob->lru_lock);
564 if (list_empty(&bdev->ddestroy))
565 goto out_unlock;
566
567 entry = list_first_entry(&bdev->ddestroy,
568 struct ttm_buffer_object, ddestroy);
569 kref_get(&entry->list_kref);
570
571 for (;;) {
572 struct ttm_buffer_object *nentry = NULL;
573
574 if (entry->ddestroy.next != &bdev->ddestroy) {
575 nentry = list_first_entry(&entry->ddestroy,
576 struct ttm_buffer_object, ddestroy);
577 kref_get(&nentry->list_kref);
578 }
579
580 ret = __ttm_bo_reserve(entry, false, true, false, 0);
581 if (remove_all && ret) {
582 spin_unlock(&glob->lru_lock);
583 ret = __ttm_bo_reserve(entry, false, false,
584 false, 0);
585 spin_lock(&glob->lru_lock);
586 }
587
588 if (!ret)
589 ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
590 !remove_all);
591 else
592 spin_unlock(&glob->lru_lock);
593
594 kref_put(&entry->list_kref, ttm_bo_release_list);
595 entry = nentry;
596
597 if (ret || !entry)
598 goto out;
599
600 spin_lock(&glob->lru_lock);
601 if (list_empty(&entry->ddestroy))
602 break;
603 }
604
605out_unlock:
606 spin_unlock(&glob->lru_lock);
607out:
608 if (entry)
609 kref_put(&entry->list_kref, ttm_bo_release_list);
610 return ret;
611}
612
613static void ttm_bo_delayed_workqueue(struct work_struct *work)
614{
615 struct ttm_bo_device *bdev =
616 container_of(work, struct ttm_bo_device, wq.work);
617
618 if (ttm_bo_delayed_delete(bdev, false)) {
619 schedule_delayed_work(&bdev->wq,
620 ((HZ / 100) < 1) ? 1 : HZ / 100);
621 }
622}
623
624static void ttm_bo_release(struct kref *kref)
625{
626 struct ttm_buffer_object *bo =
627 container_of(kref, struct ttm_buffer_object, kref);
628 struct ttm_bo_device *bdev = bo->bdev;
629 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
630
631 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
632 ttm_mem_io_lock(man, false);
633 ttm_mem_io_free_vm(bo);
634 ttm_mem_io_unlock(man);
635 ttm_bo_cleanup_refs_or_queue(bo);
636 kref_put(&bo->list_kref, ttm_bo_release_list);
637}
638
639void ttm_bo_unref(struct ttm_buffer_object **p_bo)
640{
641 struct ttm_buffer_object *bo = *p_bo;
642
643 *p_bo = NULL;
644 kref_put(&bo->kref, ttm_bo_release);
645}
646EXPORT_SYMBOL(ttm_bo_unref);
647
648int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
649{
650 return cancel_delayed_work_sync(&bdev->wq);
651}
652EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
653
654void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
655{
656 if (resched)
657 schedule_delayed_work(&bdev->wq,
658 ((HZ / 100) < 1) ? 1 : HZ / 100);
659}
660EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
661
662static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
663 bool no_wait_gpu)
664{
665 struct ttm_bo_device *bdev = bo->bdev;
666 struct ttm_mem_reg evict_mem;
667 struct ttm_placement placement;
668 int ret = 0;
669
670 spin_lock(&bdev->fence_lock);
671 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
672 spin_unlock(&bdev->fence_lock);
673
674 if (unlikely(ret != 0)) {
675 if (ret != -ERESTARTSYS) {
676 pr_err("Failed to expire sync object before buffer eviction\n");
677 }
678 goto out;
679 }
680
681 lockdep_assert_held(&bo->resv->lock.base);
682
683 evict_mem = bo->mem;
684 evict_mem.mm_node = NULL;
685 evict_mem.bus.io_reserved_vm = false;
686 evict_mem.bus.io_reserved_count = 0;
687
688 placement.fpfn = 0;
689 placement.lpfn = 0;
690 placement.num_placement = 0;
691 placement.num_busy_placement = 0;
692 bdev->driver->evict_flags(bo, &placement);
693 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
694 no_wait_gpu);
695 if (ret) {
696 if (ret != -ERESTARTSYS) {
697 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
698 bo);
699 ttm_bo_mem_space_debug(bo, &placement);
700 }
701 goto out;
702 }
703
704 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
705 no_wait_gpu);
706 if (ret) {
707 if (ret != -ERESTARTSYS)
708 pr_err("Buffer eviction failed\n");
709 ttm_bo_mem_put(bo, &evict_mem);
710 goto out;
711 }
712 bo->evicted = true;
713out:
714 return ret;
715}
716
717static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
718 uint32_t mem_type,
719 bool interruptible,
720 bool no_wait_gpu)
721{
722 struct ttm_bo_global *glob = bdev->glob;
723 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
724 struct ttm_buffer_object *bo;
725 int ret = -EBUSY, put_count;
726
727 spin_lock(&glob->lru_lock);
728 list_for_each_entry(bo, &man->lru, lru) {
729 ret = __ttm_bo_reserve(bo, false, true, false, 0);
730 if (!ret)
731 break;
732 }
733
734 if (ret) {
735 spin_unlock(&glob->lru_lock);
736 return ret;
737 }
738
739 kref_get(&bo->list_kref);
740
741 if (!list_empty(&bo->ddestroy)) {
742 ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
743 no_wait_gpu);
744 kref_put(&bo->list_kref, ttm_bo_release_list);
745 return ret;
746 }
747
748 put_count = ttm_bo_del_from_lru(bo);
749 spin_unlock(&glob->lru_lock);
750
751 BUG_ON(ret != 0);
752
753 ttm_bo_list_ref_sub(bo, put_count, true);
754
755 ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
756 ttm_bo_unreserve(bo);
757
758 kref_put(&bo->list_kref, ttm_bo_release_list);
759 return ret;
760}
761
762void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
763{
764 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
765
766 if (mem->mm_node)
767 (*man->func->put_node)(man, mem);
768}
769EXPORT_SYMBOL(ttm_bo_mem_put);
770
771/**
772 * Repeatedly evict memory from the LRU for @mem_type until we create enough
773 * space, or we've evicted everything and there isn't enough space.
774 */
775static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
776 uint32_t mem_type,
777 struct ttm_placement *placement,
778 struct ttm_mem_reg *mem,
779 bool interruptible,
780 bool no_wait_gpu)
781{
782 struct ttm_bo_device *bdev = bo->bdev;
783 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
784 int ret;
785
786 do {
787 ret = (*man->func->get_node)(man, bo, placement, mem);
788 if (unlikely(ret != 0))
789 return ret;
790 if (mem->mm_node)
791 break;
792 ret = ttm_mem_evict_first(bdev, mem_type,
793 interruptible, no_wait_gpu);
794 if (unlikely(ret != 0))
795 return ret;
796 } while (1);
797 if (mem->mm_node == NULL)
798 return -ENOMEM;
799 mem->mem_type = mem_type;
800 return 0;
801}
802
803static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
804 uint32_t cur_placement,
805 uint32_t proposed_placement)
806{
807 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
808 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
809
810 /**
811 * Keep current caching if possible.
812 */
813
814 if ((cur_placement & caching) != 0)
815 result |= (cur_placement & caching);
816 else if ((man->default_caching & caching) != 0)
817 result |= man->default_caching;
818 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
819 result |= TTM_PL_FLAG_CACHED;
820 else if ((TTM_PL_FLAG_WC & caching) != 0)
821 result |= TTM_PL_FLAG_WC;
822 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
823 result |= TTM_PL_FLAG_UNCACHED;
824
825 return result;
826}
827
828static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
829 uint32_t mem_type,
830 uint32_t proposed_placement,
831 uint32_t *masked_placement)
832{
833 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
834
835 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
836 return false;
837
838 if ((proposed_placement & man->available_caching) == 0)
839 return false;
840
841 cur_flags |= (proposed_placement & man->available_caching);
842
843 *masked_placement = cur_flags;
844 return true;
845}
846
847/**
848 * Creates space for memory region @mem according to its type.
849 *
850 * This function first searches for free space in compatible memory types in
851 * the priority order defined by the driver. If free space isn't found, then
852 * ttm_bo_mem_force_space is attempted in priority order to evict and find
853 * space.
854 */
855int ttm_bo_mem_space(struct ttm_buffer_object *bo,
856 struct ttm_placement *placement,
857 struct ttm_mem_reg *mem,
858 bool interruptible,
859 bool no_wait_gpu)
860{
861 struct ttm_bo_device *bdev = bo->bdev;
862 struct ttm_mem_type_manager *man;
863 uint32_t mem_type = TTM_PL_SYSTEM;
864 uint32_t cur_flags = 0;
865 bool type_found = false;
866 bool type_ok = false;
867 bool has_erestartsys = false;
868 int i, ret;
869
870 mem->mm_node = NULL;
871 for (i = 0; i < placement->num_placement; ++i) {
872 ret = ttm_mem_type_from_flags(placement->placement[i],
873 &mem_type);
874 if (ret)
875 return ret;
876 man = &bdev->man[mem_type];
877
878 type_ok = ttm_bo_mt_compatible(man,
879 mem_type,
880 placement->placement[i],
881 &cur_flags);
882
883 if (!type_ok)
884 continue;
885
886 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
887 cur_flags);
888 /*
889 * Use the access and other non-mapping-related flag bits from
890 * the memory placement flags to the current flags
891 */
892 ttm_flag_masked(&cur_flags, placement->placement[i],
893 ~TTM_PL_MASK_MEMTYPE);
894
895 if (mem_type == TTM_PL_SYSTEM)
896 break;
897
898 if (man->has_type && man->use_type) {
899 type_found = true;
900 ret = (*man->func->get_node)(man, bo, placement, mem);
901 if (unlikely(ret))
902 return ret;
903 }
904 if (mem->mm_node)
905 break;
906 }
907
908 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
909 mem->mem_type = mem_type;
910 mem->placement = cur_flags;
911 return 0;
912 }
913
914 if (!type_found)
915 return -EINVAL;
916
917 for (i = 0; i < placement->num_busy_placement; ++i) {
918 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
919 &mem_type);
920 if (ret)
921 return ret;
922 man = &bdev->man[mem_type];
923 if (!man->has_type)
924 continue;
925 if (!ttm_bo_mt_compatible(man,
926 mem_type,
927 placement->busy_placement[i],
928 &cur_flags))
929 continue;
930
931 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
932 cur_flags);
933 /*
934 * Use the access and other non-mapping-related flag bits from
935 * the memory placement flags to the current flags
936 */
937 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
938 ~TTM_PL_MASK_MEMTYPE);
939
940
941 if (mem_type == TTM_PL_SYSTEM) {
942 mem->mem_type = mem_type;
943 mem->placement = cur_flags;
944 mem->mm_node = NULL;
945 return 0;
946 }
947
948 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
949 interruptible, no_wait_gpu);
950 if (ret == 0 && mem->mm_node) {
951 mem->placement = cur_flags;
952 return 0;
953 }
954 if (ret == -ERESTARTSYS)
955 has_erestartsys = true;
956 }
957 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
958 return ret;
959}
960EXPORT_SYMBOL(ttm_bo_mem_space);
961
962static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
963 struct ttm_placement *placement,
964 bool interruptible,
965 bool no_wait_gpu)
966{
967 int ret = 0;
968 struct ttm_mem_reg mem;
969 struct ttm_bo_device *bdev = bo->bdev;
970
971 lockdep_assert_held(&bo->resv->lock.base);
972
973 /*
974 * FIXME: It's possible to pipeline buffer moves.
975 * Have the driver move function wait for idle when necessary,
976 * instead of doing it here.
977 */
978 spin_lock(&bdev->fence_lock);
979 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
980 spin_unlock(&bdev->fence_lock);
981 if (ret)
982 return ret;
983 mem.num_pages = bo->num_pages;
984 mem.size = mem.num_pages << PAGE_SHIFT;
985 mem.page_alignment = bo->mem.page_alignment;
986 mem.bus.io_reserved_vm = false;
987 mem.bus.io_reserved_count = 0;
988 /*
989 * Determine where to move the buffer.
990 */
991 ret = ttm_bo_mem_space(bo, placement, &mem,
992 interruptible, no_wait_gpu);
993 if (ret)
994 goto out_unlock;
995 ret = ttm_bo_handle_move_mem(bo, &mem, false,
996 interruptible, no_wait_gpu);
997out_unlock:
998 if (ret && mem.mm_node)
999 ttm_bo_mem_put(bo, &mem);
1000 return ret;
1001}
1002
1003static bool ttm_bo_mem_compat(struct ttm_placement *placement,
1004 struct ttm_mem_reg *mem,
1005 uint32_t *new_flags)
1006{
1007 int i;
1008
1009 if (mem->mm_node && placement->lpfn != 0 &&
1010 (mem->start < placement->fpfn ||
1011 mem->start + mem->num_pages > placement->lpfn))
1012 return false;
1013
1014 for (i = 0; i < placement->num_placement; i++) {
1015 *new_flags = placement->placement[i];
1016 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1017 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1018 return true;
1019 }
1020
1021 for (i = 0; i < placement->num_busy_placement; i++) {
1022 *new_flags = placement->busy_placement[i];
1023 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1024 (*new_flags & mem->placement & TTM_PL_MASK_MEM))
1025 return true;
1026 }
1027
1028 return false;
1029}
1030
1031int ttm_bo_validate(struct ttm_buffer_object *bo,
1032 struct ttm_placement *placement,
1033 bool interruptible,
1034 bool no_wait_gpu)
1035{
1036 int ret;
1037 uint32_t new_flags;
1038
1039 lockdep_assert_held(&bo->resv->lock.base);
1040 /* Check that range is valid */
1041 if (placement->lpfn || placement->fpfn)
1042 if (placement->fpfn > placement->lpfn ||
1043 (placement->lpfn - placement->fpfn) < bo->num_pages)
1044 return -EINVAL;
1045 /*
1046 * Check whether we need to move buffer.
1047 */
1048 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1049 ret = ttm_bo_move_buffer(bo, placement, interruptible,
1050 no_wait_gpu);
1051 if (ret)
1052 return ret;
1053 } else {
1054 /*
1055 * Use the access and other non-mapping-related flag bits from
1056 * the compatible memory placement flags to the active flags
1057 */
1058 ttm_flag_masked(&bo->mem.placement, new_flags,
1059 ~TTM_PL_MASK_MEMTYPE);
1060 }
1061 /*
1062 * We might need to add a TTM.
1063 */
1064 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1065 ret = ttm_bo_add_ttm(bo, true);
1066 if (ret)
1067 return ret;
1068 }
1069 return 0;
1070}
1071EXPORT_SYMBOL(ttm_bo_validate);
1072
1073int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1074 struct ttm_placement *placement)
1075{
1076 BUG_ON((placement->fpfn || placement->lpfn) &&
1077 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1078
1079 return 0;
1080}
1081
1082int ttm_bo_init(struct ttm_bo_device *bdev,
1083 struct ttm_buffer_object *bo,
1084 unsigned long size,
1085 enum ttm_bo_type type,
1086 struct ttm_placement *placement,
1087 uint32_t page_alignment,
1088 bool interruptible,
1089 struct file *persistent_swap_storage,
1090 size_t acc_size,
1091 struct sg_table *sg,
1092 void (*destroy) (struct ttm_buffer_object *))
1093{
1094 int ret = 0;
1095 unsigned long num_pages;
1096 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1097 bool locked;
1098
1099 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1100 if (ret) {
1101 pr_err("Out of kernel memory\n");
1102 if (destroy)
1103 (*destroy)(bo);
1104 else
1105 kfree(bo);
1106 return -ENOMEM;
1107 }
1108
1109 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1110 if (num_pages == 0) {
1111 pr_err("Illegal buffer object size\n");
1112 if (destroy)
1113 (*destroy)(bo);
1114 else
1115 kfree(bo);
1116 ttm_mem_global_free(mem_glob, acc_size);
1117 return -EINVAL;
1118 }
1119 bo->destroy = destroy;
1120
1121 kref_init(&bo->kref);
1122 kref_init(&bo->list_kref);
1123 atomic_set(&bo->cpu_writers, 0);
1124 INIT_LIST_HEAD(&bo->lru);
1125 INIT_LIST_HEAD(&bo->ddestroy);
1126 INIT_LIST_HEAD(&bo->swap);
1127 INIT_LIST_HEAD(&bo->io_reserve_lru);
1128 mutex_init(&bo->wu_mutex);
1129 bo->bdev = bdev;
1130 bo->glob = bdev->glob;
1131 bo->type = type;
1132 bo->num_pages = num_pages;
1133 bo->mem.size = num_pages << PAGE_SHIFT;
1134 bo->mem.mem_type = TTM_PL_SYSTEM;
1135 bo->mem.num_pages = bo->num_pages;
1136 bo->mem.mm_node = NULL;
1137 bo->mem.page_alignment = page_alignment;
1138 bo->mem.bus.io_reserved_vm = false;
1139 bo->mem.bus.io_reserved_count = 0;
1140 bo->priv_flags = 0;
1141 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1142 bo->persistent_swap_storage = persistent_swap_storage;
1143 bo->acc_size = acc_size;
1144 bo->sg = sg;
1145 bo->resv = &bo->ttm_resv;
1146 reservation_object_init(bo->resv);
1147 atomic_inc(&bo->glob->bo_count);
1148 drm_vma_node_reset(&bo->vma_node);
1149
1150 ret = ttm_bo_check_placement(bo, placement);
1151
1152 /*
1153 * For ttm_bo_type_device buffers, allocate
1154 * address space from the device.
1155 */
1156 if (likely(!ret) &&
1157 (bo->type == ttm_bo_type_device ||
1158 bo->type == ttm_bo_type_sg))
1159 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1160 bo->mem.num_pages);
1161
1162 locked = ww_mutex_trylock(&bo->resv->lock);
1163 WARN_ON(!locked);
1164
1165 if (likely(!ret))
1166 ret = ttm_bo_validate(bo, placement, interruptible, false);
1167
1168 ttm_bo_unreserve(bo);
1169
1170 if (unlikely(ret))
1171 ttm_bo_unref(&bo);
1172
1173 return ret;
1174}
1175EXPORT_SYMBOL(ttm_bo_init);
1176
1177size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1178 unsigned long bo_size,
1179 unsigned struct_size)
1180{
1181 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1182 size_t size = 0;
1183
1184 size += ttm_round_pot(struct_size);
1185 size += PAGE_ALIGN(npages * sizeof(void *));
1186 size += ttm_round_pot(sizeof(struct ttm_tt));
1187 return size;
1188}
1189EXPORT_SYMBOL(ttm_bo_acc_size);
1190
1191size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1192 unsigned long bo_size,
1193 unsigned struct_size)
1194{
1195 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1196 size_t size = 0;
1197
1198 size += ttm_round_pot(struct_size);
1199 size += PAGE_ALIGN(npages * sizeof(void *));
1200 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1201 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1202 return size;
1203}
1204EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1205
1206int ttm_bo_create(struct ttm_bo_device *bdev,
1207 unsigned long size,
1208 enum ttm_bo_type type,
1209 struct ttm_placement *placement,
1210 uint32_t page_alignment,
1211 bool interruptible,
1212 struct file *persistent_swap_storage,
1213 struct ttm_buffer_object **p_bo)
1214{
1215 struct ttm_buffer_object *bo;
1216 size_t acc_size;
1217 int ret;
1218
1219 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1220 if (unlikely(bo == NULL))
1221 return -ENOMEM;
1222
1223 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1224 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1225 interruptible, persistent_swap_storage, acc_size,
1226 NULL, NULL);
1227 if (likely(ret == 0))
1228 *p_bo = bo;
1229
1230 return ret;
1231}
1232EXPORT_SYMBOL(ttm_bo_create);
1233
1234static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1235 unsigned mem_type, bool allow_errors)
1236{
1237 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1238 struct ttm_bo_global *glob = bdev->glob;
1239 int ret;
1240
1241 /*
1242 * Can't use standard list traversal since we're unlocking.
1243 */
1244
1245 spin_lock(&glob->lru_lock);
1246 while (!list_empty(&man->lru)) {
1247 spin_unlock(&glob->lru_lock);
1248 ret = ttm_mem_evict_first(bdev, mem_type, false, false);
1249 if (ret) {
1250 if (allow_errors) {
1251 return ret;
1252 } else {
1253 pr_err("Cleanup eviction failed\n");
1254 }
1255 }
1256 spin_lock(&glob->lru_lock);
1257 }
1258 spin_unlock(&glob->lru_lock);
1259 return 0;
1260}
1261
1262int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1263{
1264 struct ttm_mem_type_manager *man;
1265 int ret = -EINVAL;
1266
1267 if (mem_type >= TTM_NUM_MEM_TYPES) {
1268 pr_err("Illegal memory type %d\n", mem_type);
1269 return ret;
1270 }
1271 man = &bdev->man[mem_type];
1272
1273 if (!man->has_type) {
1274 pr_err("Trying to take down uninitialized memory manager type %u\n",
1275 mem_type);
1276 return ret;
1277 }
1278
1279 man->use_type = false;
1280 man->has_type = false;
1281
1282 ret = 0;
1283 if (mem_type > 0) {
1284 ttm_bo_force_list_clean(bdev, mem_type, false);
1285
1286 ret = (*man->func->takedown)(man);
1287 }
1288
1289 return ret;
1290}
1291EXPORT_SYMBOL(ttm_bo_clean_mm);
1292
1293int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1294{
1295 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1296
1297 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1298 pr_err("Illegal memory manager memory type %u\n", mem_type);
1299 return -EINVAL;
1300 }
1301
1302 if (!man->has_type) {
1303 pr_err("Memory type %u has not been initialized\n", mem_type);
1304 return 0;
1305 }
1306
1307 return ttm_bo_force_list_clean(bdev, mem_type, true);
1308}
1309EXPORT_SYMBOL(ttm_bo_evict_mm);
1310
1311int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1312 unsigned long p_size)
1313{
1314 int ret = -EINVAL;
1315 struct ttm_mem_type_manager *man;
1316
1317 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1318 man = &bdev->man[type];
1319 BUG_ON(man->has_type);
1320 man->io_reserve_fastpath = true;
1321 man->use_io_reserve_lru = false;
1322 mutex_init(&man->io_reserve_mutex);
1323 INIT_LIST_HEAD(&man->io_reserve_lru);
1324
1325 ret = bdev->driver->init_mem_type(bdev, type, man);
1326 if (ret)
1327 return ret;
1328 man->bdev = bdev;
1329
1330 ret = 0;
1331 if (type != TTM_PL_SYSTEM) {
1332 ret = (*man->func->init)(man, p_size);
1333 if (ret)
1334 return ret;
1335 }
1336 man->has_type = true;
1337 man->use_type = true;
1338 man->size = p_size;
1339
1340 INIT_LIST_HEAD(&man->lru);
1341
1342 return 0;
1343}
1344EXPORT_SYMBOL(ttm_bo_init_mm);
1345
1346static void ttm_bo_global_kobj_release(struct kobject *kobj)
1347{
1348 struct ttm_bo_global *glob =
1349 container_of(kobj, struct ttm_bo_global, kobj);
1350
1351 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1352 __free_page(glob->dummy_read_page);
1353 kfree(glob);
1354}
1355
1356void ttm_bo_global_release(struct drm_global_reference *ref)
1357{
1358 struct ttm_bo_global *glob = ref->object;
1359
1360 kobject_del(&glob->kobj);
1361 kobject_put(&glob->kobj);
1362}
1363EXPORT_SYMBOL(ttm_bo_global_release);
1364
1365int ttm_bo_global_init(struct drm_global_reference *ref)
1366{
1367 struct ttm_bo_global_ref *bo_ref =
1368 container_of(ref, struct ttm_bo_global_ref, ref);
1369 struct ttm_bo_global *glob = ref->object;
1370 int ret;
1371
1372 mutex_init(&glob->device_list_mutex);
1373 spin_lock_init(&glob->lru_lock);
1374 glob->mem_glob = bo_ref->mem_glob;
1375 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1376
1377 if (unlikely(glob->dummy_read_page == NULL)) {
1378 ret = -ENOMEM;
1379 goto out_no_drp;
1380 }
1381
1382 INIT_LIST_HEAD(&glob->swap_lru);
1383 INIT_LIST_HEAD(&glob->device_list);
1384
1385 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1386 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1387 if (unlikely(ret != 0)) {
1388 pr_err("Could not register buffer object swapout\n");
1389 goto out_no_shrink;
1390 }
1391
1392 atomic_set(&glob->bo_count, 0);
1393
1394 ret = kobject_init_and_add(
1395 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1396 if (unlikely(ret != 0))
1397 kobject_put(&glob->kobj);
1398 return ret;
1399out_no_shrink:
1400 __free_page(glob->dummy_read_page);
1401out_no_drp:
1402 kfree(glob);
1403 return ret;
1404}
1405EXPORT_SYMBOL(ttm_bo_global_init);
1406
1407
1408int ttm_bo_device_release(struct ttm_bo_device *bdev)
1409{
1410 int ret = 0;
1411 unsigned i = TTM_NUM_MEM_TYPES;
1412 struct ttm_mem_type_manager *man;
1413 struct ttm_bo_global *glob = bdev->glob;
1414
1415 while (i--) {
1416 man = &bdev->man[i];
1417 if (man->has_type) {
1418 man->use_type = false;
1419 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1420 ret = -EBUSY;
1421 pr_err("DRM memory manager type %d is not clean\n",
1422 i);
1423 }
1424 man->has_type = false;
1425 }
1426 }
1427
1428 mutex_lock(&glob->device_list_mutex);
1429 list_del(&bdev->device_list);
1430 mutex_unlock(&glob->device_list_mutex);
1431
1432 cancel_delayed_work_sync(&bdev->wq);
1433
1434 while (ttm_bo_delayed_delete(bdev, true))
1435 ;
1436
1437 spin_lock(&glob->lru_lock);
1438 if (list_empty(&bdev->ddestroy))
1439 TTM_DEBUG("Delayed destroy list was clean\n");
1440
1441 if (list_empty(&bdev->man[0].lru))
1442 TTM_DEBUG("Swap list was clean\n");
1443 spin_unlock(&glob->lru_lock);
1444
1445 drm_vma_offset_manager_destroy(&bdev->vma_manager);
1446
1447 return ret;
1448}
1449EXPORT_SYMBOL(ttm_bo_device_release);
1450
1451int ttm_bo_device_init(struct ttm_bo_device *bdev,
1452 struct ttm_bo_global *glob,
1453 struct ttm_bo_driver *driver,
1454 struct address_space *mapping,
1455 uint64_t file_page_offset,
1456 bool need_dma32)
1457{
1458 int ret = -EINVAL;
1459
1460 bdev->driver = driver;
1461
1462 memset(bdev->man, 0, sizeof(bdev->man));
1463
1464 /*
1465 * Initialize the system memory buffer type.
1466 * Other types need to be driver / IOCTL initialized.
1467 */
1468 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1469 if (unlikely(ret != 0))
1470 goto out_no_sys;
1471
1472 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1473 0x10000000);
1474 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1475 INIT_LIST_HEAD(&bdev->ddestroy);
1476 bdev->dev_mapping = mapping;
1477 bdev->glob = glob;
1478 bdev->need_dma32 = need_dma32;
1479 bdev->val_seq = 0;
1480 spin_lock_init(&bdev->fence_lock);
1481 mutex_lock(&glob->device_list_mutex);
1482 list_add_tail(&bdev->device_list, &glob->device_list);
1483 mutex_unlock(&glob->device_list_mutex);
1484
1485 return 0;
1486out_no_sys:
1487 return ret;
1488}
1489EXPORT_SYMBOL(ttm_bo_device_init);
1490
1491/*
1492 * buffer object vm functions.
1493 */
1494
1495bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1496{
1497 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1498
1499 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1500 if (mem->mem_type == TTM_PL_SYSTEM)
1501 return false;
1502
1503 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1504 return false;
1505
1506 if (mem->placement & TTM_PL_FLAG_CACHED)
1507 return false;
1508 }
1509 return true;
1510}
1511
1512void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1513{
1514 struct ttm_bo_device *bdev = bo->bdev;
1515
1516 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1517 ttm_mem_io_free_vm(bo);
1518}
1519
1520void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1521{
1522 struct ttm_bo_device *bdev = bo->bdev;
1523 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1524
1525 ttm_mem_io_lock(man, false);
1526 ttm_bo_unmap_virtual_locked(bo);
1527 ttm_mem_io_unlock(man);
1528}
1529
1530
1531EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1532
1533
1534int ttm_bo_wait(struct ttm_buffer_object *bo,
1535 bool lazy, bool interruptible, bool no_wait)
1536{
1537 struct ttm_bo_driver *driver = bo->bdev->driver;
1538 struct ttm_bo_device *bdev = bo->bdev;
1539 void *sync_obj;
1540 int ret = 0;
1541
1542 if (likely(bo->sync_obj == NULL))
1543 return 0;
1544
1545 while (bo->sync_obj) {
1546
1547 if (driver->sync_obj_signaled(bo->sync_obj)) {
1548 void *tmp_obj = bo->sync_obj;
1549 bo->sync_obj = NULL;
1550 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1551 spin_unlock(&bdev->fence_lock);
1552 driver->sync_obj_unref(&tmp_obj);
1553 spin_lock(&bdev->fence_lock);
1554 continue;
1555 }
1556
1557 if (no_wait)
1558 return -EBUSY;
1559
1560 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1561 spin_unlock(&bdev->fence_lock);
1562 ret = driver->sync_obj_wait(sync_obj,
1563 lazy, interruptible);
1564 if (unlikely(ret != 0)) {
1565 driver->sync_obj_unref(&sync_obj);
1566 spin_lock(&bdev->fence_lock);
1567 return ret;
1568 }
1569 spin_lock(&bdev->fence_lock);
1570 if (likely(bo->sync_obj == sync_obj)) {
1571 void *tmp_obj = bo->sync_obj;
1572 bo->sync_obj = NULL;
1573 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1574 &bo->priv_flags);
1575 spin_unlock(&bdev->fence_lock);
1576 driver->sync_obj_unref(&sync_obj);
1577 driver->sync_obj_unref(&tmp_obj);
1578 spin_lock(&bdev->fence_lock);
1579 } else {
1580 spin_unlock(&bdev->fence_lock);
1581 driver->sync_obj_unref(&sync_obj);
1582 spin_lock(&bdev->fence_lock);
1583 }
1584 }
1585 return 0;
1586}
1587EXPORT_SYMBOL(ttm_bo_wait);
1588
1589int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1590{
1591 struct ttm_bo_device *bdev = bo->bdev;
1592 int ret = 0;
1593
1594 /*
1595 * Using ttm_bo_reserve makes sure the lru lists are updated.
1596 */
1597
1598 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1599 if (unlikely(ret != 0))
1600 return ret;
1601 spin_lock(&bdev->fence_lock);
1602 ret = ttm_bo_wait(bo, false, true, no_wait);
1603 spin_unlock(&bdev->fence_lock);
1604 if (likely(ret == 0))
1605 atomic_inc(&bo->cpu_writers);
1606 ttm_bo_unreserve(bo);
1607 return ret;
1608}
1609EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1610
1611void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1612{
1613 atomic_dec(&bo->cpu_writers);
1614}
1615EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1616
1617/**
1618 * A buffer object shrink method that tries to swap out the first
1619 * buffer object on the bo_global::swap_lru list.
1620 */
1621
1622static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1623{
1624 struct ttm_bo_global *glob =
1625 container_of(shrink, struct ttm_bo_global, shrink);
1626 struct ttm_buffer_object *bo;
1627 int ret = -EBUSY;
1628 int put_count;
1629 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1630
1631 spin_lock(&glob->lru_lock);
1632 list_for_each_entry(bo, &glob->swap_lru, swap) {
1633 ret = __ttm_bo_reserve(bo, false, true, false, 0);
1634 if (!ret)
1635 break;
1636 }
1637
1638 if (ret) {
1639 spin_unlock(&glob->lru_lock);
1640 return ret;
1641 }
1642
1643 kref_get(&bo->list_kref);
1644
1645 if (!list_empty(&bo->ddestroy)) {
1646 ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1647 kref_put(&bo->list_kref, ttm_bo_release_list);
1648 return ret;
1649 }
1650
1651 put_count = ttm_bo_del_from_lru(bo);
1652 spin_unlock(&glob->lru_lock);
1653
1654 ttm_bo_list_ref_sub(bo, put_count, true);
1655
1656 /**
1657 * Wait for GPU, then move to system cached.
1658 */
1659
1660 spin_lock(&bo->bdev->fence_lock);
1661 ret = ttm_bo_wait(bo, false, false, false);
1662 spin_unlock(&bo->bdev->fence_lock);
1663
1664 if (unlikely(ret != 0))
1665 goto out;
1666
1667 if ((bo->mem.placement & swap_placement) != swap_placement) {
1668 struct ttm_mem_reg evict_mem;
1669
1670 evict_mem = bo->mem;
1671 evict_mem.mm_node = NULL;
1672 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1673 evict_mem.mem_type = TTM_PL_SYSTEM;
1674
1675 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1676 false, false);
1677 if (unlikely(ret != 0))
1678 goto out;
1679 }
1680
1681 ttm_bo_unmap_virtual(bo);
1682
1683 /**
1684 * Swap out. Buffer will be swapped in again as soon as
1685 * anyone tries to access a ttm page.
1686 */
1687
1688 if (bo->bdev->driver->swap_notify)
1689 bo->bdev->driver->swap_notify(bo);
1690
1691 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1692out:
1693
1694 /**
1695 *
1696 * Unreserve without putting on LRU to avoid swapping out an
1697 * already swapped buffer.
1698 */
1699
1700 __ttm_bo_unreserve(bo);
1701 kref_put(&bo->list_kref, ttm_bo_release_list);
1702 return ret;
1703}
1704
1705void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1706{
1707 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1708 ;
1709}
1710EXPORT_SYMBOL(ttm_bo_swapout_all);
1711
1712/**
1713 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1714 * unreserved
1715 *
1716 * @bo: Pointer to buffer
1717 */
1718int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1719{
1720 int ret;
1721
1722 /*
1723 * In the absense of a wait_unlocked API,
1724 * Use the bo::wu_mutex to avoid triggering livelocks due to
1725 * concurrent use of this function. Note that this use of
1726 * bo::wu_mutex can go away if we change locking order to
1727 * mmap_sem -> bo::reserve.
1728 */
1729 ret = mutex_lock_interruptible(&bo->wu_mutex);
1730 if (unlikely(ret != 0))
1731 return -ERESTARTSYS;
1732 if (!ww_mutex_is_locked(&bo->resv->lock))
1733 goto out_unlock;
1734 ret = __ttm_bo_reserve(bo, true, false, false, NULL);
1735 if (unlikely(ret != 0))
1736 goto out_unlock;
1737 __ttm_bo_unreserve(bo);
1738
1739out_unlock:
1740 mutex_unlock(&bo->wu_mutex);
1741 return ret;
1742}