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/* SPDX-License-Identifier: GPL-2.0 OR MIT */
2/**************************************************************************
3 *
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28/*
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30 */
31
32#define pr_fmt(fmt) "[TTM] " fmt
33
34#include <drm/ttm/ttm_module.h>
35#include <drm/ttm/ttm_bo_driver.h>
36#include <drm/ttm/ttm_placement.h>
37#include <linux/jiffies.h>
38#include <linux/slab.h>
39#include <linux/sched.h>
40#include <linux/mm.h>
41#include <linux/file.h>
42#include <linux/module.h>
43#include <linux/atomic.h>
44#include <linux/dma-resv.h>
45
46static void ttm_bo_global_kobj_release(struct kobject *kobj);
47
48/**
49 * ttm_global_mutex - protecting the global BO state
50 */
51DEFINE_MUTEX(ttm_global_mutex);
52unsigned ttm_bo_glob_use_count;
53struct ttm_bo_global ttm_bo_glob;
54EXPORT_SYMBOL(ttm_bo_glob);
55
56static struct attribute ttm_bo_count = {
57 .name = "bo_count",
58 .mode = S_IRUGO
59};
60
61/* default destructor */
62static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
63{
64 kfree(bo);
65}
66
67static inline int ttm_mem_type_from_place(const struct ttm_place *place,
68 uint32_t *mem_type)
69{
70 int pos;
71
72 pos = ffs(place->flags & TTM_PL_MASK_MEM);
73 if (unlikely(!pos))
74 return -EINVAL;
75
76 *mem_type = pos - 1;
77 return 0;
78}
79
80static void ttm_mem_type_debug(struct ttm_bo_device *bdev, struct drm_printer *p,
81 int mem_type)
82{
83 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
84
85 drm_printf(p, " has_type: %d\n", man->has_type);
86 drm_printf(p, " use_type: %d\n", man->use_type);
87 drm_printf(p, " flags: 0x%08X\n", man->flags);
88 drm_printf(p, " size: %llu\n", man->size);
89 drm_printf(p, " available_caching: 0x%08X\n", man->available_caching);
90 drm_printf(p, " default_caching: 0x%08X\n", man->default_caching);
91 if (mem_type != TTM_PL_SYSTEM)
92 (*man->func->debug)(man, p);
93}
94
95static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
96 struct ttm_placement *placement)
97{
98 struct drm_printer p = drm_debug_printer(TTM_PFX);
99 int i, ret, mem_type;
100
101 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n",
102 bo, bo->mem.num_pages, bo->mem.size >> 10,
103 bo->mem.size >> 20);
104 for (i = 0; i < placement->num_placement; i++) {
105 ret = ttm_mem_type_from_place(&placement->placement[i],
106 &mem_type);
107 if (ret)
108 return;
109 drm_printf(&p, " placement[%d]=0x%08X (%d)\n",
110 i, placement->placement[i].flags, mem_type);
111 ttm_mem_type_debug(bo->bdev, &p, mem_type);
112 }
113}
114
115static ssize_t ttm_bo_global_show(struct kobject *kobj,
116 struct attribute *attr,
117 char *buffer)
118{
119 struct ttm_bo_global *glob =
120 container_of(kobj, struct ttm_bo_global, kobj);
121
122 return snprintf(buffer, PAGE_SIZE, "%d\n",
123 atomic_read(&glob->bo_count));
124}
125
126static struct attribute *ttm_bo_global_attrs[] = {
127 &ttm_bo_count,
128 NULL
129};
130
131static const struct sysfs_ops ttm_bo_global_ops = {
132 .show = &ttm_bo_global_show
133};
134
135static struct kobj_type ttm_bo_glob_kobj_type = {
136 .release = &ttm_bo_global_kobj_release,
137 .sysfs_ops = &ttm_bo_global_ops,
138 .default_attrs = ttm_bo_global_attrs
139};
140
141
142static inline uint32_t ttm_bo_type_flags(unsigned type)
143{
144 return 1 << (type);
145}
146
147static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo,
148 struct ttm_mem_reg *mem)
149{
150 struct ttm_bo_device *bdev = bo->bdev;
151 struct ttm_mem_type_manager *man;
152
153 if (!list_empty(&bo->lru))
154 return;
155
156 if (mem->placement & TTM_PL_FLAG_NO_EVICT)
157 return;
158
159 man = &bdev->man[mem->mem_type];
160 list_add_tail(&bo->lru, &man->lru[bo->priority]);
161
162 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm &&
163 !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG |
164 TTM_PAGE_FLAG_SWAPPED))) {
165 list_add_tail(&bo->swap, &ttm_bo_glob.swap_lru[bo->priority]);
166 }
167}
168
169static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
170{
171 struct ttm_bo_device *bdev = bo->bdev;
172 bool notify = false;
173
174 if (!list_empty(&bo->swap)) {
175 list_del_init(&bo->swap);
176 notify = true;
177 }
178 if (!list_empty(&bo->lru)) {
179 list_del_init(&bo->lru);
180 notify = true;
181 }
182
183 if (notify && bdev->driver->del_from_lru_notify)
184 bdev->driver->del_from_lru_notify(bo);
185}
186
187static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
188 struct ttm_buffer_object *bo)
189{
190 if (!pos->first)
191 pos->first = bo;
192 pos->last = bo;
193}
194
195void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
196 struct ttm_lru_bulk_move *bulk)
197{
198 dma_resv_assert_held(bo->base.resv);
199
200 ttm_bo_del_from_lru(bo);
201 ttm_bo_add_mem_to_lru(bo, &bo->mem);
202
203 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
204 switch (bo->mem.mem_type) {
205 case TTM_PL_TT:
206 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
207 break;
208
209 case TTM_PL_VRAM:
210 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
211 break;
212 }
213 if (bo->ttm && !(bo->ttm->page_flags &
214 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
215 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
216 }
217}
218EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
219
220void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
221{
222 unsigned i;
223
224 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
225 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
226 struct ttm_mem_type_manager *man;
227
228 if (!pos->first)
229 continue;
230
231 dma_resv_assert_held(pos->first->base.resv);
232 dma_resv_assert_held(pos->last->base.resv);
233
234 man = &pos->first->bdev->man[TTM_PL_TT];
235 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
236 &pos->last->lru);
237 }
238
239 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
240 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
241 struct ttm_mem_type_manager *man;
242
243 if (!pos->first)
244 continue;
245
246 dma_resv_assert_held(pos->first->base.resv);
247 dma_resv_assert_held(pos->last->base.resv);
248
249 man = &pos->first->bdev->man[TTM_PL_VRAM];
250 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
251 &pos->last->lru);
252 }
253
254 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
255 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
256 struct list_head *lru;
257
258 if (!pos->first)
259 continue;
260
261 dma_resv_assert_held(pos->first->base.resv);
262 dma_resv_assert_held(pos->last->base.resv);
263
264 lru = &ttm_bo_glob.swap_lru[i];
265 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
266 }
267}
268EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
269
270static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
271 struct ttm_mem_reg *mem, bool evict,
272 struct ttm_operation_ctx *ctx)
273{
274 struct ttm_bo_device *bdev = bo->bdev;
275 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
276 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
277 int ret;
278
279 ret = ttm_mem_io_lock(old_man, true);
280 if (unlikely(ret != 0))
281 goto out_err;
282 ttm_bo_unmap_virtual_locked(bo);
283 ttm_mem_io_unlock(old_man);
284
285 /*
286 * Create and bind a ttm if required.
287 */
288
289 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
290 if (bo->ttm == NULL) {
291 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
292 ret = ttm_tt_create(bo, zero);
293 if (ret)
294 goto out_err;
295 }
296
297 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
298 if (ret)
299 goto out_err;
300
301 if (mem->mem_type != TTM_PL_SYSTEM) {
302 ret = ttm_tt_bind(bo->ttm, mem, ctx);
303 if (ret)
304 goto out_err;
305 }
306
307 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
308 if (bdev->driver->move_notify)
309 bdev->driver->move_notify(bo, evict, mem);
310 bo->mem = *mem;
311 goto moved;
312 }
313 }
314
315 if (bdev->driver->move_notify)
316 bdev->driver->move_notify(bo, evict, mem);
317
318 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
319 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
320 ret = ttm_bo_move_ttm(bo, ctx, mem);
321 else if (bdev->driver->move)
322 ret = bdev->driver->move(bo, evict, ctx, mem);
323 else
324 ret = ttm_bo_move_memcpy(bo, ctx, mem);
325
326 if (ret) {
327 if (bdev->driver->move_notify) {
328 swap(*mem, bo->mem);
329 bdev->driver->move_notify(bo, false, mem);
330 swap(*mem, bo->mem);
331 }
332
333 goto out_err;
334 }
335
336moved:
337 bo->evicted = false;
338
339 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
340 return 0;
341
342out_err:
343 new_man = &bdev->man[bo->mem.mem_type];
344 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
345 ttm_tt_destroy(bo->ttm);
346 bo->ttm = NULL;
347 }
348
349 return ret;
350}
351
352/**
353 * Call bo::reserved.
354 * Will release GPU memory type usage on destruction.
355 * This is the place to put in driver specific hooks to release
356 * driver private resources.
357 * Will release the bo::reserved lock.
358 */
359
360static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
361{
362 if (bo->bdev->driver->move_notify)
363 bo->bdev->driver->move_notify(bo, false, NULL);
364
365 ttm_tt_destroy(bo->ttm);
366 bo->ttm = NULL;
367 ttm_bo_mem_put(bo, &bo->mem);
368}
369
370static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
371{
372 int r;
373
374 if (bo->base.resv == &bo->base._resv)
375 return 0;
376
377 BUG_ON(!dma_resv_trylock(&bo->base._resv));
378
379 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
380 dma_resv_unlock(&bo->base._resv);
381 if (r)
382 return r;
383
384 if (bo->type != ttm_bo_type_sg) {
385 /* This works because the BO is about to be destroyed and nobody
386 * reference it any more. The only tricky case is the trylock on
387 * the resv object while holding the lru_lock.
388 */
389 spin_lock(&ttm_bo_glob.lru_lock);
390 bo->base.resv = &bo->base._resv;
391 spin_unlock(&ttm_bo_glob.lru_lock);
392 }
393
394 return r;
395}
396
397static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
398{
399 struct dma_resv *resv = &bo->base._resv;
400 struct dma_resv_list *fobj;
401 struct dma_fence *fence;
402 int i;
403
404 rcu_read_lock();
405 fobj = rcu_dereference(resv->fence);
406 fence = rcu_dereference(resv->fence_excl);
407 if (fence && !fence->ops->signaled)
408 dma_fence_enable_sw_signaling(fence);
409
410 for (i = 0; fobj && i < fobj->shared_count; ++i) {
411 fence = rcu_dereference(fobj->shared[i]);
412
413 if (!fence->ops->signaled)
414 dma_fence_enable_sw_signaling(fence);
415 }
416 rcu_read_unlock();
417}
418
419/**
420 * function ttm_bo_cleanup_refs
421 * If bo idle, remove from lru lists, and unref.
422 * If not idle, block if possible.
423 *
424 * Must be called with lru_lock and reservation held, this function
425 * will drop the lru lock and optionally the reservation lock before returning.
426 *
427 * @interruptible Any sleeps should occur interruptibly.
428 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
429 * @unlock_resv Unlock the reservation lock as well.
430 */
431
432static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
433 bool interruptible, bool no_wait_gpu,
434 bool unlock_resv)
435{
436 struct dma_resv *resv = &bo->base._resv;
437 int ret;
438
439 if (dma_resv_test_signaled_rcu(resv, true))
440 ret = 0;
441 else
442 ret = -EBUSY;
443
444 if (ret && !no_wait_gpu) {
445 long lret;
446
447 if (unlock_resv)
448 dma_resv_unlock(bo->base.resv);
449 spin_unlock(&ttm_bo_glob.lru_lock);
450
451 lret = dma_resv_wait_timeout_rcu(resv, true, interruptible,
452 30 * HZ);
453
454 if (lret < 0)
455 return lret;
456 else if (lret == 0)
457 return -EBUSY;
458
459 spin_lock(&ttm_bo_glob.lru_lock);
460 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
461 /*
462 * We raced, and lost, someone else holds the reservation now,
463 * and is probably busy in ttm_bo_cleanup_memtype_use.
464 *
465 * Even if it's not the case, because we finished waiting any
466 * delayed destruction would succeed, so just return success
467 * here.
468 */
469 spin_unlock(&ttm_bo_glob.lru_lock);
470 return 0;
471 }
472 ret = 0;
473 }
474
475 if (ret || unlikely(list_empty(&bo->ddestroy))) {
476 if (unlock_resv)
477 dma_resv_unlock(bo->base.resv);
478 spin_unlock(&ttm_bo_glob.lru_lock);
479 return ret;
480 }
481
482 ttm_bo_del_from_lru(bo);
483 list_del_init(&bo->ddestroy);
484 spin_unlock(&ttm_bo_glob.lru_lock);
485 ttm_bo_cleanup_memtype_use(bo);
486
487 if (unlock_resv)
488 dma_resv_unlock(bo->base.resv);
489
490 ttm_bo_put(bo);
491
492 return 0;
493}
494
495/**
496 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
497 * encountered buffers.
498 */
499static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
500{
501 struct ttm_bo_global *glob = &ttm_bo_glob;
502 struct list_head removed;
503 bool empty;
504
505 INIT_LIST_HEAD(&removed);
506
507 spin_lock(&glob->lru_lock);
508 while (!list_empty(&bdev->ddestroy)) {
509 struct ttm_buffer_object *bo;
510
511 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
512 ddestroy);
513 list_move_tail(&bo->ddestroy, &removed);
514 if (!ttm_bo_get_unless_zero(bo))
515 continue;
516
517 if (remove_all || bo->base.resv != &bo->base._resv) {
518 spin_unlock(&glob->lru_lock);
519 dma_resv_lock(bo->base.resv, NULL);
520
521 spin_lock(&glob->lru_lock);
522 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
523
524 } else if (dma_resv_trylock(bo->base.resv)) {
525 ttm_bo_cleanup_refs(bo, false, !remove_all, true);
526 } else {
527 spin_unlock(&glob->lru_lock);
528 }
529
530 ttm_bo_put(bo);
531 spin_lock(&glob->lru_lock);
532 }
533 list_splice_tail(&removed, &bdev->ddestroy);
534 empty = list_empty(&bdev->ddestroy);
535 spin_unlock(&glob->lru_lock);
536
537 return empty;
538}
539
540static void ttm_bo_delayed_workqueue(struct work_struct *work)
541{
542 struct ttm_bo_device *bdev =
543 container_of(work, struct ttm_bo_device, wq.work);
544
545 if (!ttm_bo_delayed_delete(bdev, false))
546 schedule_delayed_work(&bdev->wq,
547 ((HZ / 100) < 1) ? 1 : HZ / 100);
548}
549
550static void ttm_bo_release(struct kref *kref)
551{
552 struct ttm_buffer_object *bo =
553 container_of(kref, struct ttm_buffer_object, kref);
554 struct ttm_bo_device *bdev = bo->bdev;
555 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
556 size_t acc_size = bo->acc_size;
557 int ret;
558
559 if (!bo->deleted) {
560 ret = ttm_bo_individualize_resv(bo);
561 if (ret) {
562 /* Last resort, if we fail to allocate memory for the
563 * fences block for the BO to become idle
564 */
565 dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
566 30 * HZ);
567 }
568
569 if (bo->bdev->driver->release_notify)
570 bo->bdev->driver->release_notify(bo);
571
572 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
573 ttm_mem_io_lock(man, false);
574 ttm_mem_io_free_vm(bo);
575 ttm_mem_io_unlock(man);
576 }
577
578 if (!dma_resv_test_signaled_rcu(bo->base.resv, true) ||
579 !dma_resv_trylock(bo->base.resv)) {
580 /* The BO is not idle, resurrect it for delayed destroy */
581 ttm_bo_flush_all_fences(bo);
582 bo->deleted = true;
583
584 spin_lock(&ttm_bo_glob.lru_lock);
585
586 /*
587 * Make NO_EVICT bos immediately available to
588 * shrinkers, now that they are queued for
589 * destruction.
590 */
591 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
592 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
593 ttm_bo_del_from_lru(bo);
594 ttm_bo_add_mem_to_lru(bo, &bo->mem);
595 }
596
597 kref_init(&bo->kref);
598 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
599 spin_unlock(&ttm_bo_glob.lru_lock);
600
601 schedule_delayed_work(&bdev->wq,
602 ((HZ / 100) < 1) ? 1 : HZ / 100);
603 return;
604 }
605
606 spin_lock(&ttm_bo_glob.lru_lock);
607 ttm_bo_del_from_lru(bo);
608 list_del(&bo->ddestroy);
609 spin_unlock(&ttm_bo_glob.lru_lock);
610
611 ttm_bo_cleanup_memtype_use(bo);
612 dma_resv_unlock(bo->base.resv);
613
614 atomic_dec(&ttm_bo_glob.bo_count);
615 dma_fence_put(bo->moving);
616 if (!ttm_bo_uses_embedded_gem_object(bo))
617 dma_resv_fini(&bo->base._resv);
618 bo->destroy(bo);
619 ttm_mem_global_free(&ttm_mem_glob, acc_size);
620}
621
622void ttm_bo_put(struct ttm_buffer_object *bo)
623{
624 kref_put(&bo->kref, ttm_bo_release);
625}
626EXPORT_SYMBOL(ttm_bo_put);
627
628int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
629{
630 return cancel_delayed_work_sync(&bdev->wq);
631}
632EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
633
634void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
635{
636 if (resched)
637 schedule_delayed_work(&bdev->wq,
638 ((HZ / 100) < 1) ? 1 : HZ / 100);
639}
640EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
641
642static int ttm_bo_evict(struct ttm_buffer_object *bo,
643 struct ttm_operation_ctx *ctx)
644{
645 struct ttm_bo_device *bdev = bo->bdev;
646 struct ttm_mem_reg evict_mem;
647 struct ttm_placement placement;
648 int ret = 0;
649
650 dma_resv_assert_held(bo->base.resv);
651
652 placement.num_placement = 0;
653 placement.num_busy_placement = 0;
654 bdev->driver->evict_flags(bo, &placement);
655
656 if (!placement.num_placement && !placement.num_busy_placement) {
657 ret = ttm_bo_pipeline_gutting(bo);
658 if (ret)
659 return ret;
660
661 return ttm_tt_create(bo, false);
662 }
663
664 evict_mem = bo->mem;
665 evict_mem.mm_node = NULL;
666 evict_mem.bus.io_reserved_vm = false;
667 evict_mem.bus.io_reserved_count = 0;
668
669 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
670 if (ret) {
671 if (ret != -ERESTARTSYS) {
672 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
673 bo);
674 ttm_bo_mem_space_debug(bo, &placement);
675 }
676 goto out;
677 }
678
679 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
680 if (unlikely(ret)) {
681 if (ret != -ERESTARTSYS)
682 pr_err("Buffer eviction failed\n");
683 ttm_bo_mem_put(bo, &evict_mem);
684 goto out;
685 }
686 bo->evicted = true;
687out:
688 return ret;
689}
690
691bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
692 const struct ttm_place *place)
693{
694 /* Don't evict this BO if it's outside of the
695 * requested placement range
696 */
697 if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
698 (place->lpfn && place->lpfn <= bo->mem.start))
699 return false;
700
701 return true;
702}
703EXPORT_SYMBOL(ttm_bo_eviction_valuable);
704
705/**
706 * Check the target bo is allowable to be evicted or swapout, including cases:
707 *
708 * a. if share same reservation object with ctx->resv, have assumption
709 * reservation objects should already be locked, so not lock again and
710 * return true directly when either the opreation allow_reserved_eviction
711 * or the target bo already is in delayed free list;
712 *
713 * b. Otherwise, trylock it.
714 */
715static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
716 struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
717{
718 bool ret = false;
719
720 if (bo->base.resv == ctx->resv) {
721 dma_resv_assert_held(bo->base.resv);
722 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT)
723 ret = true;
724 *locked = false;
725 if (busy)
726 *busy = false;
727 } else {
728 ret = dma_resv_trylock(bo->base.resv);
729 *locked = ret;
730 if (busy)
731 *busy = !ret;
732 }
733
734 return ret;
735}
736
737/**
738 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
739 *
740 * @busy_bo: BO which couldn't be locked with trylock
741 * @ctx: operation context
742 * @ticket: acquire ticket
743 *
744 * Try to lock a busy buffer object to avoid failing eviction.
745 */
746static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
747 struct ttm_operation_ctx *ctx,
748 struct ww_acquire_ctx *ticket)
749{
750 int r;
751
752 if (!busy_bo || !ticket)
753 return -EBUSY;
754
755 if (ctx->interruptible)
756 r = dma_resv_lock_interruptible(busy_bo->base.resv,
757 ticket);
758 else
759 r = dma_resv_lock(busy_bo->base.resv, ticket);
760
761 /*
762 * TODO: It would be better to keep the BO locked until allocation is at
763 * least tried one more time, but that would mean a much larger rework
764 * of TTM.
765 */
766 if (!r)
767 dma_resv_unlock(busy_bo->base.resv);
768
769 return r == -EDEADLK ? -EBUSY : r;
770}
771
772static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
773 uint32_t mem_type,
774 const struct ttm_place *place,
775 struct ttm_operation_ctx *ctx,
776 struct ww_acquire_ctx *ticket)
777{
778 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
779 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
780 bool locked = false;
781 unsigned i;
782 int ret;
783
784 spin_lock(&ttm_bo_glob.lru_lock);
785 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
786 list_for_each_entry(bo, &man->lru[i], lru) {
787 bool busy;
788
789 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
790 &busy)) {
791 if (busy && !busy_bo && ticket !=
792 dma_resv_locking_ctx(bo->base.resv))
793 busy_bo = bo;
794 continue;
795 }
796
797 if (place && !bdev->driver->eviction_valuable(bo,
798 place)) {
799 if (locked)
800 dma_resv_unlock(bo->base.resv);
801 continue;
802 }
803 if (!ttm_bo_get_unless_zero(bo)) {
804 if (locked)
805 dma_resv_unlock(bo->base.resv);
806 continue;
807 }
808 break;
809 }
810
811 /* If the inner loop terminated early, we have our candidate */
812 if (&bo->lru != &man->lru[i])
813 break;
814
815 bo = NULL;
816 }
817
818 if (!bo) {
819 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
820 busy_bo = NULL;
821 spin_unlock(&ttm_bo_glob.lru_lock);
822 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
823 if (busy_bo)
824 ttm_bo_put(busy_bo);
825 return ret;
826 }
827
828 if (bo->deleted) {
829 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
830 ctx->no_wait_gpu, locked);
831 ttm_bo_put(bo);
832 return ret;
833 }
834
835 spin_unlock(&ttm_bo_glob.lru_lock);
836
837 ret = ttm_bo_evict(bo, ctx);
838 if (locked)
839 ttm_bo_unreserve(bo);
840
841 ttm_bo_put(bo);
842 return ret;
843}
844
845static int ttm_bo_mem_get(struct ttm_buffer_object *bo,
846 const struct ttm_place *place,
847 struct ttm_mem_reg *mem)
848{
849 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
850
851 mem->mm_node = NULL;
852 if (!man->func || !man->func->get_node)
853 return 0;
854
855 return man->func->get_node(man, bo, place, mem);
856}
857
858void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
859{
860 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
861
862 if (!man->func || !man->func->put_node)
863 return;
864
865 man->func->put_node(man, mem);
866 mem->mm_node = NULL;
867 mem->mem_type = TTM_PL_SYSTEM;
868}
869EXPORT_SYMBOL(ttm_bo_mem_put);
870
871/**
872 * Add the last move fence to the BO and reserve a new shared slot.
873 */
874static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
875 struct ttm_mem_type_manager *man,
876 struct ttm_mem_reg *mem,
877 bool no_wait_gpu)
878{
879 struct dma_fence *fence;
880 int ret;
881
882 spin_lock(&man->move_lock);
883 fence = dma_fence_get(man->move);
884 spin_unlock(&man->move_lock);
885
886 if (!fence)
887 return 0;
888
889 if (no_wait_gpu) {
890 dma_fence_put(fence);
891 return -EBUSY;
892 }
893
894 dma_resv_add_shared_fence(bo->base.resv, fence);
895
896 ret = dma_resv_reserve_shared(bo->base.resv, 1);
897 if (unlikely(ret)) {
898 dma_fence_put(fence);
899 return ret;
900 }
901
902 dma_fence_put(bo->moving);
903 bo->moving = fence;
904 return 0;
905}
906
907/**
908 * Repeatedly evict memory from the LRU for @mem_type until we create enough
909 * space, or we've evicted everything and there isn't enough space.
910 */
911static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
912 const struct ttm_place *place,
913 struct ttm_mem_reg *mem,
914 struct ttm_operation_ctx *ctx)
915{
916 struct ttm_bo_device *bdev = bo->bdev;
917 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
918 struct ww_acquire_ctx *ticket;
919 int ret;
920
921 ticket = dma_resv_locking_ctx(bo->base.resv);
922 do {
923 ret = ttm_bo_mem_get(bo, place, mem);
924 if (likely(!ret))
925 break;
926 if (unlikely(ret != -ENOSPC))
927 return ret;
928 ret = ttm_mem_evict_first(bdev, mem->mem_type, place, ctx,
929 ticket);
930 if (unlikely(ret != 0))
931 return ret;
932 } while (1);
933
934 return ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
935}
936
937static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
938 uint32_t cur_placement,
939 uint32_t proposed_placement)
940{
941 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
942 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
943
944 /**
945 * Keep current caching if possible.
946 */
947
948 if ((cur_placement & caching) != 0)
949 result |= (cur_placement & caching);
950 else if ((man->default_caching & caching) != 0)
951 result |= man->default_caching;
952 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
953 result |= TTM_PL_FLAG_CACHED;
954 else if ((TTM_PL_FLAG_WC & caching) != 0)
955 result |= TTM_PL_FLAG_WC;
956 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
957 result |= TTM_PL_FLAG_UNCACHED;
958
959 return result;
960}
961
962static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
963 uint32_t mem_type,
964 const struct ttm_place *place,
965 uint32_t *masked_placement)
966{
967 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
968
969 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
970 return false;
971
972 if ((place->flags & man->available_caching) == 0)
973 return false;
974
975 cur_flags |= (place->flags & man->available_caching);
976
977 *masked_placement = cur_flags;
978 return true;
979}
980
981/**
982 * ttm_bo_mem_placement - check if placement is compatible
983 * @bo: BO to find memory for
984 * @place: where to search
985 * @mem: the memory object to fill in
986 * @ctx: operation context
987 *
988 * Check if placement is compatible and fill in mem structure.
989 * Returns -EBUSY if placement won't work or negative error code.
990 * 0 when placement can be used.
991 */
992static int ttm_bo_mem_placement(struct ttm_buffer_object *bo,
993 const struct ttm_place *place,
994 struct ttm_mem_reg *mem,
995 struct ttm_operation_ctx *ctx)
996{
997 struct ttm_bo_device *bdev = bo->bdev;
998 uint32_t mem_type = TTM_PL_SYSTEM;
999 struct ttm_mem_type_manager *man;
1000 uint32_t cur_flags = 0;
1001 int ret;
1002
1003 ret = ttm_mem_type_from_place(place, &mem_type);
1004 if (ret)
1005 return ret;
1006
1007 man = &bdev->man[mem_type];
1008 if (!man->has_type || !man->use_type)
1009 return -EBUSY;
1010
1011 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
1012 return -EBUSY;
1013
1014 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags);
1015 /*
1016 * Use the access and other non-mapping-related flag bits from
1017 * the memory placement flags to the current flags
1018 */
1019 ttm_flag_masked(&cur_flags, place->flags, ~TTM_PL_MASK_MEMTYPE);
1020
1021 mem->mem_type = mem_type;
1022 mem->placement = cur_flags;
1023
1024 spin_lock(&ttm_bo_glob.lru_lock);
1025 ttm_bo_del_from_lru(bo);
1026 ttm_bo_add_mem_to_lru(bo, mem);
1027 spin_unlock(&ttm_bo_glob.lru_lock);
1028
1029 return 0;
1030}
1031
1032/**
1033 * Creates space for memory region @mem according to its type.
1034 *
1035 * This function first searches for free space in compatible memory types in
1036 * the priority order defined by the driver. If free space isn't found, then
1037 * ttm_bo_mem_force_space is attempted in priority order to evict and find
1038 * space.
1039 */
1040int ttm_bo_mem_space(struct ttm_buffer_object *bo,
1041 struct ttm_placement *placement,
1042 struct ttm_mem_reg *mem,
1043 struct ttm_operation_ctx *ctx)
1044{
1045 struct ttm_bo_device *bdev = bo->bdev;
1046 bool type_found = false;
1047 int i, ret;
1048
1049 ret = dma_resv_reserve_shared(bo->base.resv, 1);
1050 if (unlikely(ret))
1051 return ret;
1052
1053 for (i = 0; i < placement->num_placement; ++i) {
1054 const struct ttm_place *place = &placement->placement[i];
1055 struct ttm_mem_type_manager *man;
1056
1057 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
1058 if (ret == -EBUSY)
1059 continue;
1060 if (ret)
1061 goto error;
1062
1063 type_found = true;
1064 ret = ttm_bo_mem_get(bo, place, mem);
1065 if (ret == -ENOSPC)
1066 continue;
1067 if (unlikely(ret))
1068 goto error;
1069
1070 man = &bdev->man[mem->mem_type];
1071 ret = ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
1072 if (unlikely(ret)) {
1073 ttm_bo_mem_put(bo, mem);
1074 if (ret == -EBUSY)
1075 continue;
1076
1077 goto error;
1078 }
1079 return 0;
1080 }
1081
1082 for (i = 0; i < placement->num_busy_placement; ++i) {
1083 const struct ttm_place *place = &placement->busy_placement[i];
1084
1085 ret = ttm_bo_mem_placement(bo, place, mem, ctx);
1086 if (ret == -EBUSY)
1087 continue;
1088 if (ret)
1089 goto error;
1090
1091 type_found = true;
1092 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
1093 if (likely(!ret))
1094 return 0;
1095
1096 if (ret && ret != -EBUSY)
1097 goto error;
1098 }
1099
1100 ret = -ENOMEM;
1101 if (!type_found) {
1102 pr_err(TTM_PFX "No compatible memory type found\n");
1103 ret = -EINVAL;
1104 }
1105
1106error:
1107 if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) {
1108 spin_lock(&ttm_bo_glob.lru_lock);
1109 ttm_bo_move_to_lru_tail(bo, NULL);
1110 spin_unlock(&ttm_bo_glob.lru_lock);
1111 }
1112
1113 return ret;
1114}
1115EXPORT_SYMBOL(ttm_bo_mem_space);
1116
1117static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1118 struct ttm_placement *placement,
1119 struct ttm_operation_ctx *ctx)
1120{
1121 int ret = 0;
1122 struct ttm_mem_reg mem;
1123
1124 dma_resv_assert_held(bo->base.resv);
1125
1126 mem.num_pages = bo->num_pages;
1127 mem.size = mem.num_pages << PAGE_SHIFT;
1128 mem.page_alignment = bo->mem.page_alignment;
1129 mem.bus.io_reserved_vm = false;
1130 mem.bus.io_reserved_count = 0;
1131 mem.mm_node = NULL;
1132
1133 /*
1134 * Determine where to move the buffer.
1135 */
1136 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1137 if (ret)
1138 goto out_unlock;
1139 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1140out_unlock:
1141 if (ret)
1142 ttm_bo_mem_put(bo, &mem);
1143 return ret;
1144}
1145
1146static bool ttm_bo_places_compat(const struct ttm_place *places,
1147 unsigned num_placement,
1148 struct ttm_mem_reg *mem,
1149 uint32_t *new_flags)
1150{
1151 unsigned i;
1152
1153 for (i = 0; i < num_placement; i++) {
1154 const struct ttm_place *heap = &places[i];
1155
1156 if ((mem->start < heap->fpfn ||
1157 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1158 continue;
1159
1160 *new_flags = heap->flags;
1161 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1162 (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
1163 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1164 (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1165 return true;
1166 }
1167 return false;
1168}
1169
1170bool ttm_bo_mem_compat(struct ttm_placement *placement,
1171 struct ttm_mem_reg *mem,
1172 uint32_t *new_flags)
1173{
1174 if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1175 mem, new_flags))
1176 return true;
1177
1178 if ((placement->busy_placement != placement->placement ||
1179 placement->num_busy_placement > placement->num_placement) &&
1180 ttm_bo_places_compat(placement->busy_placement,
1181 placement->num_busy_placement,
1182 mem, new_flags))
1183 return true;
1184
1185 return false;
1186}
1187EXPORT_SYMBOL(ttm_bo_mem_compat);
1188
1189int ttm_bo_validate(struct ttm_buffer_object *bo,
1190 struct ttm_placement *placement,
1191 struct ttm_operation_ctx *ctx)
1192{
1193 int ret;
1194 uint32_t new_flags;
1195
1196 dma_resv_assert_held(bo->base.resv);
1197
1198 /*
1199 * Remove the backing store if no placement is given.
1200 */
1201 if (!placement->num_placement && !placement->num_busy_placement) {
1202 ret = ttm_bo_pipeline_gutting(bo);
1203 if (ret)
1204 return ret;
1205
1206 return ttm_tt_create(bo, false);
1207 }
1208
1209 /*
1210 * Check whether we need to move buffer.
1211 */
1212 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1213 ret = ttm_bo_move_buffer(bo, placement, ctx);
1214 if (ret)
1215 return ret;
1216 } else {
1217 /*
1218 * Use the access and other non-mapping-related flag bits from
1219 * the compatible memory placement flags to the active flags
1220 */
1221 ttm_flag_masked(&bo->mem.placement, new_flags,
1222 ~TTM_PL_MASK_MEMTYPE);
1223 }
1224 /*
1225 * We might need to add a TTM.
1226 */
1227 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1228 ret = ttm_tt_create(bo, true);
1229 if (ret)
1230 return ret;
1231 }
1232 return 0;
1233}
1234EXPORT_SYMBOL(ttm_bo_validate);
1235
1236int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1237 struct ttm_buffer_object *bo,
1238 unsigned long size,
1239 enum ttm_bo_type type,
1240 struct ttm_placement *placement,
1241 uint32_t page_alignment,
1242 struct ttm_operation_ctx *ctx,
1243 size_t acc_size,
1244 struct sg_table *sg,
1245 struct dma_resv *resv,
1246 void (*destroy) (struct ttm_buffer_object *))
1247{
1248 struct ttm_mem_global *mem_glob = &ttm_mem_glob;
1249 int ret = 0;
1250 unsigned long num_pages;
1251 bool locked;
1252
1253 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1254 if (ret) {
1255 pr_err("Out of kernel memory\n");
1256 if (destroy)
1257 (*destroy)(bo);
1258 else
1259 kfree(bo);
1260 return -ENOMEM;
1261 }
1262
1263 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1264 if (num_pages == 0) {
1265 pr_err("Illegal buffer object size\n");
1266 if (destroy)
1267 (*destroy)(bo);
1268 else
1269 kfree(bo);
1270 ttm_mem_global_free(mem_glob, acc_size);
1271 return -EINVAL;
1272 }
1273 bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1274
1275 kref_init(&bo->kref);
1276 INIT_LIST_HEAD(&bo->lru);
1277 INIT_LIST_HEAD(&bo->ddestroy);
1278 INIT_LIST_HEAD(&bo->swap);
1279 INIT_LIST_HEAD(&bo->io_reserve_lru);
1280 bo->bdev = bdev;
1281 bo->type = type;
1282 bo->num_pages = num_pages;
1283 bo->mem.size = num_pages << PAGE_SHIFT;
1284 bo->mem.mem_type = TTM_PL_SYSTEM;
1285 bo->mem.num_pages = bo->num_pages;
1286 bo->mem.mm_node = NULL;
1287 bo->mem.page_alignment = page_alignment;
1288 bo->mem.bus.io_reserved_vm = false;
1289 bo->mem.bus.io_reserved_count = 0;
1290 bo->moving = NULL;
1291 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1292 bo->acc_size = acc_size;
1293 bo->sg = sg;
1294 if (resv) {
1295 bo->base.resv = resv;
1296 dma_resv_assert_held(bo->base.resv);
1297 } else {
1298 bo->base.resv = &bo->base._resv;
1299 }
1300 if (!ttm_bo_uses_embedded_gem_object(bo)) {
1301 /*
1302 * bo.gem is not initialized, so we have to setup the
1303 * struct elements we want use regardless.
1304 */
1305 dma_resv_init(&bo->base._resv);
1306 drm_vma_node_reset(&bo->base.vma_node);
1307 }
1308 atomic_inc(&ttm_bo_glob.bo_count);
1309
1310 /*
1311 * For ttm_bo_type_device buffers, allocate
1312 * address space from the device.
1313 */
1314 if (bo->type == ttm_bo_type_device ||
1315 bo->type == ttm_bo_type_sg)
1316 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1317 bo->mem.num_pages);
1318
1319 /* passed reservation objects should already be locked,
1320 * since otherwise lockdep will be angered in radeon.
1321 */
1322 if (!resv) {
1323 locked = dma_resv_trylock(bo->base.resv);
1324 WARN_ON(!locked);
1325 }
1326
1327 if (likely(!ret))
1328 ret = ttm_bo_validate(bo, placement, ctx);
1329
1330 if (unlikely(ret)) {
1331 if (!resv)
1332 ttm_bo_unreserve(bo);
1333
1334 ttm_bo_put(bo);
1335 return ret;
1336 }
1337
1338 spin_lock(&ttm_bo_glob.lru_lock);
1339 ttm_bo_move_to_lru_tail(bo, NULL);
1340 spin_unlock(&ttm_bo_glob.lru_lock);
1341
1342 return ret;
1343}
1344EXPORT_SYMBOL(ttm_bo_init_reserved);
1345
1346int ttm_bo_init(struct ttm_bo_device *bdev,
1347 struct ttm_buffer_object *bo,
1348 unsigned long size,
1349 enum ttm_bo_type type,
1350 struct ttm_placement *placement,
1351 uint32_t page_alignment,
1352 bool interruptible,
1353 size_t acc_size,
1354 struct sg_table *sg,
1355 struct dma_resv *resv,
1356 void (*destroy) (struct ttm_buffer_object *))
1357{
1358 struct ttm_operation_ctx ctx = { interruptible, false };
1359 int ret;
1360
1361 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1362 page_alignment, &ctx, acc_size,
1363 sg, resv, destroy);
1364 if (ret)
1365 return ret;
1366
1367 if (!resv)
1368 ttm_bo_unreserve(bo);
1369
1370 return 0;
1371}
1372EXPORT_SYMBOL(ttm_bo_init);
1373
1374size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1375 unsigned long bo_size,
1376 unsigned struct_size)
1377{
1378 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1379 size_t size = 0;
1380
1381 size += ttm_round_pot(struct_size);
1382 size += ttm_round_pot(npages * sizeof(void *));
1383 size += ttm_round_pot(sizeof(struct ttm_tt));
1384 return size;
1385}
1386EXPORT_SYMBOL(ttm_bo_acc_size);
1387
1388size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1389 unsigned long bo_size,
1390 unsigned struct_size)
1391{
1392 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1393 size_t size = 0;
1394
1395 size += ttm_round_pot(struct_size);
1396 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1397 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1398 return size;
1399}
1400EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1401
1402int ttm_bo_create(struct ttm_bo_device *bdev,
1403 unsigned long size,
1404 enum ttm_bo_type type,
1405 struct ttm_placement *placement,
1406 uint32_t page_alignment,
1407 bool interruptible,
1408 struct ttm_buffer_object **p_bo)
1409{
1410 struct ttm_buffer_object *bo;
1411 size_t acc_size;
1412 int ret;
1413
1414 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1415 if (unlikely(bo == NULL))
1416 return -ENOMEM;
1417
1418 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1419 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1420 interruptible, acc_size,
1421 NULL, NULL, NULL);
1422 if (likely(ret == 0))
1423 *p_bo = bo;
1424
1425 return ret;
1426}
1427EXPORT_SYMBOL(ttm_bo_create);
1428
1429static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1430 unsigned mem_type)
1431{
1432 struct ttm_operation_ctx ctx = {
1433 .interruptible = false,
1434 .no_wait_gpu = false,
1435 .flags = TTM_OPT_FLAG_FORCE_ALLOC
1436 };
1437 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1438 struct ttm_bo_global *glob = &ttm_bo_glob;
1439 struct dma_fence *fence;
1440 int ret;
1441 unsigned i;
1442
1443 /*
1444 * Can't use standard list traversal since we're unlocking.
1445 */
1446
1447 spin_lock(&glob->lru_lock);
1448 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1449 while (!list_empty(&man->lru[i])) {
1450 spin_unlock(&glob->lru_lock);
1451 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx,
1452 NULL);
1453 if (ret)
1454 return ret;
1455 spin_lock(&glob->lru_lock);
1456 }
1457 }
1458 spin_unlock(&glob->lru_lock);
1459
1460 spin_lock(&man->move_lock);
1461 fence = dma_fence_get(man->move);
1462 spin_unlock(&man->move_lock);
1463
1464 if (fence) {
1465 ret = dma_fence_wait(fence, false);
1466 dma_fence_put(fence);
1467 if (ret)
1468 return ret;
1469 }
1470
1471 return 0;
1472}
1473
1474int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1475{
1476 struct ttm_mem_type_manager *man;
1477 int ret = -EINVAL;
1478
1479 if (mem_type >= TTM_NUM_MEM_TYPES) {
1480 pr_err("Illegal memory type %d\n", mem_type);
1481 return ret;
1482 }
1483 man = &bdev->man[mem_type];
1484
1485 if (!man->has_type) {
1486 pr_err("Trying to take down uninitialized memory manager type %u\n",
1487 mem_type);
1488 return ret;
1489 }
1490
1491 man->use_type = false;
1492 man->has_type = false;
1493
1494 ret = 0;
1495 if (mem_type > 0) {
1496 ret = ttm_bo_force_list_clean(bdev, mem_type);
1497 if (ret) {
1498 pr_err("Cleanup eviction failed\n");
1499 return ret;
1500 }
1501
1502 ret = (*man->func->takedown)(man);
1503 }
1504
1505 dma_fence_put(man->move);
1506 man->move = NULL;
1507
1508 return ret;
1509}
1510EXPORT_SYMBOL(ttm_bo_clean_mm);
1511
1512int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1513{
1514 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1515
1516 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1517 pr_err("Illegal memory manager memory type %u\n", mem_type);
1518 return -EINVAL;
1519 }
1520
1521 if (!man->has_type) {
1522 pr_err("Memory type %u has not been initialized\n", mem_type);
1523 return 0;
1524 }
1525
1526 return ttm_bo_force_list_clean(bdev, mem_type);
1527}
1528EXPORT_SYMBOL(ttm_bo_evict_mm);
1529
1530int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1531 unsigned long p_size)
1532{
1533 int ret;
1534 struct ttm_mem_type_manager *man;
1535 unsigned i;
1536
1537 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1538 man = &bdev->man[type];
1539 BUG_ON(man->has_type);
1540 man->use_io_reserve_lru = false;
1541 mutex_init(&man->io_reserve_mutex);
1542 spin_lock_init(&man->move_lock);
1543 INIT_LIST_HEAD(&man->io_reserve_lru);
1544
1545 ret = bdev->driver->init_mem_type(bdev, type, man);
1546 if (ret)
1547 return ret;
1548 man->bdev = bdev;
1549
1550 if (type != TTM_PL_SYSTEM) {
1551 ret = (*man->func->init)(man, p_size);
1552 if (ret)
1553 return ret;
1554 }
1555 man->has_type = true;
1556 man->use_type = true;
1557 man->size = p_size;
1558
1559 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1560 INIT_LIST_HEAD(&man->lru[i]);
1561 man->move = NULL;
1562
1563 return 0;
1564}
1565EXPORT_SYMBOL(ttm_bo_init_mm);
1566
1567static void ttm_bo_global_kobj_release(struct kobject *kobj)
1568{
1569 struct ttm_bo_global *glob =
1570 container_of(kobj, struct ttm_bo_global, kobj);
1571
1572 __free_page(glob->dummy_read_page);
1573}
1574
1575static void ttm_bo_global_release(void)
1576{
1577 struct ttm_bo_global *glob = &ttm_bo_glob;
1578
1579 mutex_lock(&ttm_global_mutex);
1580 if (--ttm_bo_glob_use_count > 0)
1581 goto out;
1582
1583 kobject_del(&glob->kobj);
1584 kobject_put(&glob->kobj);
1585 ttm_mem_global_release(&ttm_mem_glob);
1586 memset(glob, 0, sizeof(*glob));
1587out:
1588 mutex_unlock(&ttm_global_mutex);
1589}
1590
1591static int ttm_bo_global_init(void)
1592{
1593 struct ttm_bo_global *glob = &ttm_bo_glob;
1594 int ret = 0;
1595 unsigned i;
1596
1597 mutex_lock(&ttm_global_mutex);
1598 if (++ttm_bo_glob_use_count > 1)
1599 goto out;
1600
1601 ret = ttm_mem_global_init(&ttm_mem_glob);
1602 if (ret)
1603 goto out;
1604
1605 spin_lock_init(&glob->lru_lock);
1606 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1607
1608 if (unlikely(glob->dummy_read_page == NULL)) {
1609 ret = -ENOMEM;
1610 goto out;
1611 }
1612
1613 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1614 INIT_LIST_HEAD(&glob->swap_lru[i]);
1615 INIT_LIST_HEAD(&glob->device_list);
1616 atomic_set(&glob->bo_count, 0);
1617
1618 ret = kobject_init_and_add(
1619 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1620 if (unlikely(ret != 0))
1621 kobject_put(&glob->kobj);
1622out:
1623 mutex_unlock(&ttm_global_mutex);
1624 return ret;
1625}
1626
1627int ttm_bo_device_release(struct ttm_bo_device *bdev)
1628{
1629 struct ttm_bo_global *glob = &ttm_bo_glob;
1630 int ret = 0;
1631 unsigned i = TTM_NUM_MEM_TYPES;
1632 struct ttm_mem_type_manager *man;
1633
1634 while (i--) {
1635 man = &bdev->man[i];
1636 if (man->has_type) {
1637 man->use_type = false;
1638 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1639 ret = -EBUSY;
1640 pr_err("DRM memory manager type %d is not clean\n",
1641 i);
1642 }
1643 man->has_type = false;
1644 }
1645 }
1646
1647 mutex_lock(&ttm_global_mutex);
1648 list_del(&bdev->device_list);
1649 mutex_unlock(&ttm_global_mutex);
1650
1651 cancel_delayed_work_sync(&bdev->wq);
1652
1653 if (ttm_bo_delayed_delete(bdev, true))
1654 pr_debug("Delayed destroy list was clean\n");
1655
1656 spin_lock(&glob->lru_lock);
1657 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1658 if (list_empty(&bdev->man[0].lru[0]))
1659 pr_debug("Swap list %d was clean\n", i);
1660 spin_unlock(&glob->lru_lock);
1661
1662 if (!ret)
1663 ttm_bo_global_release();
1664
1665 return ret;
1666}
1667EXPORT_SYMBOL(ttm_bo_device_release);
1668
1669int ttm_bo_device_init(struct ttm_bo_device *bdev,
1670 struct ttm_bo_driver *driver,
1671 struct address_space *mapping,
1672 struct drm_vma_offset_manager *vma_manager,
1673 bool need_dma32)
1674{
1675 struct ttm_bo_global *glob = &ttm_bo_glob;
1676 int ret;
1677
1678 if (WARN_ON(vma_manager == NULL))
1679 return -EINVAL;
1680
1681 ret = ttm_bo_global_init();
1682 if (ret)
1683 return ret;
1684
1685 bdev->driver = driver;
1686
1687 memset(bdev->man, 0, sizeof(bdev->man));
1688
1689 /*
1690 * Initialize the system memory buffer type.
1691 * Other types need to be driver / IOCTL initialized.
1692 */
1693 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1694 if (unlikely(ret != 0))
1695 goto out_no_sys;
1696
1697 bdev->vma_manager = vma_manager;
1698 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1699 INIT_LIST_HEAD(&bdev->ddestroy);
1700 bdev->dev_mapping = mapping;
1701 bdev->need_dma32 = need_dma32;
1702 mutex_lock(&ttm_global_mutex);
1703 list_add_tail(&bdev->device_list, &glob->device_list);
1704 mutex_unlock(&ttm_global_mutex);
1705
1706 return 0;
1707out_no_sys:
1708 ttm_bo_global_release();
1709 return ret;
1710}
1711EXPORT_SYMBOL(ttm_bo_device_init);
1712
1713/*
1714 * buffer object vm functions.
1715 */
1716
1717void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1718{
1719 struct ttm_bo_device *bdev = bo->bdev;
1720
1721 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1722 ttm_mem_io_free_vm(bo);
1723}
1724
1725void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1726{
1727 struct ttm_bo_device *bdev = bo->bdev;
1728 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1729
1730 ttm_mem_io_lock(man, false);
1731 ttm_bo_unmap_virtual_locked(bo);
1732 ttm_mem_io_unlock(man);
1733}
1734
1735
1736EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1737
1738int ttm_bo_wait(struct ttm_buffer_object *bo,
1739 bool interruptible, bool no_wait)
1740{
1741 long timeout = 15 * HZ;
1742
1743 if (no_wait) {
1744 if (dma_resv_test_signaled_rcu(bo->base.resv, true))
1745 return 0;
1746 else
1747 return -EBUSY;
1748 }
1749
1750 timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true,
1751 interruptible, timeout);
1752 if (timeout < 0)
1753 return timeout;
1754
1755 if (timeout == 0)
1756 return -EBUSY;
1757
1758 dma_resv_add_excl_fence(bo->base.resv, NULL);
1759 return 0;
1760}
1761EXPORT_SYMBOL(ttm_bo_wait);
1762
1763/**
1764 * A buffer object shrink method that tries to swap out the first
1765 * buffer object on the bo_global::swap_lru list.
1766 */
1767int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1768{
1769 struct ttm_buffer_object *bo;
1770 int ret = -EBUSY;
1771 bool locked;
1772 unsigned i;
1773
1774 spin_lock(&glob->lru_lock);
1775 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1776 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1777 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
1778 NULL))
1779 continue;
1780
1781 if (!ttm_bo_get_unless_zero(bo)) {
1782 if (locked)
1783 dma_resv_unlock(bo->base.resv);
1784 continue;
1785 }
1786
1787 ret = 0;
1788 break;
1789 }
1790 if (!ret)
1791 break;
1792 }
1793
1794 if (ret) {
1795 spin_unlock(&glob->lru_lock);
1796 return ret;
1797 }
1798
1799 if (bo->deleted) {
1800 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1801 ttm_bo_put(bo);
1802 return ret;
1803 }
1804
1805 ttm_bo_del_from_lru(bo);
1806 spin_unlock(&glob->lru_lock);
1807
1808 /**
1809 * Move to system cached
1810 */
1811
1812 if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1813 bo->ttm->caching_state != tt_cached) {
1814 struct ttm_operation_ctx ctx = { false, false };
1815 struct ttm_mem_reg evict_mem;
1816
1817 evict_mem = bo->mem;
1818 evict_mem.mm_node = NULL;
1819 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1820 evict_mem.mem_type = TTM_PL_SYSTEM;
1821
1822 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1823 if (unlikely(ret != 0))
1824 goto out;
1825 }
1826
1827 /**
1828 * Make sure BO is idle.
1829 */
1830
1831 ret = ttm_bo_wait(bo, false, false);
1832 if (unlikely(ret != 0))
1833 goto out;
1834
1835 ttm_bo_unmap_virtual(bo);
1836
1837 /**
1838 * Swap out. Buffer will be swapped in again as soon as
1839 * anyone tries to access a ttm page.
1840 */
1841
1842 if (bo->bdev->driver->swap_notify)
1843 bo->bdev->driver->swap_notify(bo);
1844
1845 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1846out:
1847
1848 /**
1849 *
1850 * Unreserve without putting on LRU to avoid swapping out an
1851 * already swapped buffer.
1852 */
1853 if (locked)
1854 dma_resv_unlock(bo->base.resv);
1855 ttm_bo_put(bo);
1856 return ret;
1857}
1858EXPORT_SYMBOL(ttm_bo_swapout);
1859
1860void ttm_bo_swapout_all(void)
1861{
1862 struct ttm_operation_ctx ctx = {
1863 .interruptible = false,
1864 .no_wait_gpu = false
1865 };
1866
1867 while (ttm_bo_swapout(&ttm_bo_glob, &ctx) == 0);
1868}
1869EXPORT_SYMBOL(ttm_bo_swapout_all);