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 "ttm/ttm_module.h"
34#include "ttm/ttm_bo_driver.h"
35#include "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_setup_vm(struct ttm_buffer_object *bo);
49static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50static void ttm_bo_global_kobj_release(struct kobject *kobj);
51
52static struct attribute ttm_bo_count = {
53 .name = "bo_count",
54 .mode = S_IRUGO
55};
56
57static inline int ttm_mem_type_from_flags(uint32_t flags, uint32_t *mem_type)
58{
59 int i;
60
61 for (i = 0; i <= TTM_PL_PRIV5; i++)
62 if (flags & (1 << i)) {
63 *mem_type = i;
64 return 0;
65 }
66 return -EINVAL;
67}
68
69static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
70{
71 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
72
73 pr_err(" has_type: %d\n", man->has_type);
74 pr_err(" use_type: %d\n", man->use_type);
75 pr_err(" flags: 0x%08X\n", man->flags);
76 pr_err(" gpu_offset: 0x%08lX\n", man->gpu_offset);
77 pr_err(" size: %llu\n", man->size);
78 pr_err(" available_caching: 0x%08X\n", man->available_caching);
79 pr_err(" default_caching: 0x%08X\n", man->default_caching);
80 if (mem_type != TTM_PL_SYSTEM)
81 (*man->func->debug)(man, TTM_PFX);
82}
83
84static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
85 struct ttm_placement *placement)
86{
87 int i, ret, mem_type;
88
89 pr_err("No space for %p (%lu pages, %luK, %luM)\n",
90 bo, bo->mem.num_pages, bo->mem.size >> 10,
91 bo->mem.size >> 20);
92 for (i = 0; i < placement->num_placement; i++) {
93 ret = ttm_mem_type_from_flags(placement->placement[i],
94 &mem_type);
95 if (ret)
96 return;
97 pr_err(" placement[%d]=0x%08X (%d)\n",
98 i, placement->placement[i], mem_type);
99 ttm_mem_type_debug(bo->bdev, mem_type);
100 }
101}
102
103static ssize_t ttm_bo_global_show(struct kobject *kobj,
104 struct attribute *attr,
105 char *buffer)
106{
107 struct ttm_bo_global *glob =
108 container_of(kobj, struct ttm_bo_global, kobj);
109
110 return snprintf(buffer, PAGE_SIZE, "%lu\n",
111 (unsigned long) atomic_read(&glob->bo_count));
112}
113
114static struct attribute *ttm_bo_global_attrs[] = {
115 &ttm_bo_count,
116 NULL
117};
118
119static const struct sysfs_ops ttm_bo_global_ops = {
120 .show = &ttm_bo_global_show
121};
122
123static struct kobj_type ttm_bo_glob_kobj_type = {
124 .release = &ttm_bo_global_kobj_release,
125 .sysfs_ops = &ttm_bo_global_ops,
126 .default_attrs = ttm_bo_global_attrs
127};
128
129
130static inline uint32_t ttm_bo_type_flags(unsigned type)
131{
132 return 1 << (type);
133}
134
135static void ttm_bo_release_list(struct kref *list_kref)
136{
137 struct ttm_buffer_object *bo =
138 container_of(list_kref, struct ttm_buffer_object, list_kref);
139 struct ttm_bo_device *bdev = bo->bdev;
140 size_t acc_size = bo->acc_size;
141
142 BUG_ON(atomic_read(&bo->list_kref.refcount));
143 BUG_ON(atomic_read(&bo->kref.refcount));
144 BUG_ON(atomic_read(&bo->cpu_writers));
145 BUG_ON(bo->sync_obj != NULL);
146 BUG_ON(bo->mem.mm_node != NULL);
147 BUG_ON(!list_empty(&bo->lru));
148 BUG_ON(!list_empty(&bo->ddestroy));
149
150 if (bo->ttm)
151 ttm_tt_destroy(bo->ttm);
152 atomic_dec(&bo->glob->bo_count);
153 if (bo->destroy)
154 bo->destroy(bo);
155 else {
156 kfree(bo);
157 }
158 ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
159}
160
161int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo, bool interruptible)
162{
163 if (interruptible) {
164 return wait_event_interruptible(bo->event_queue,
165 atomic_read(&bo->reserved) == 0);
166 } else {
167 wait_event(bo->event_queue, atomic_read(&bo->reserved) == 0);
168 return 0;
169 }
170}
171EXPORT_SYMBOL(ttm_bo_wait_unreserved);
172
173void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
174{
175 struct ttm_bo_device *bdev = bo->bdev;
176 struct ttm_mem_type_manager *man;
177
178 BUG_ON(!atomic_read(&bo->reserved));
179
180 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
181
182 BUG_ON(!list_empty(&bo->lru));
183
184 man = &bdev->man[bo->mem.mem_type];
185 list_add_tail(&bo->lru, &man->lru);
186 kref_get(&bo->list_kref);
187
188 if (bo->ttm != NULL) {
189 list_add_tail(&bo->swap, &bo->glob->swap_lru);
190 kref_get(&bo->list_kref);
191 }
192 }
193}
194
195int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
196{
197 int put_count = 0;
198
199 if (!list_empty(&bo->swap)) {
200 list_del_init(&bo->swap);
201 ++put_count;
202 }
203 if (!list_empty(&bo->lru)) {
204 list_del_init(&bo->lru);
205 ++put_count;
206 }
207
208 /*
209 * TODO: Add a driver hook to delete from
210 * driver-specific LRU's here.
211 */
212
213 return put_count;
214}
215
216int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
217 bool interruptible,
218 bool no_wait, bool use_sequence, uint32_t sequence)
219{
220 struct ttm_bo_global *glob = bo->glob;
221 int ret;
222
223 while (unlikely(atomic_cmpxchg(&bo->reserved, 0, 1) != 0)) {
224 /**
225 * Deadlock avoidance for multi-bo reserving.
226 */
227 if (use_sequence && bo->seq_valid) {
228 /**
229 * We've already reserved this one.
230 */
231 if (unlikely(sequence == bo->val_seq))
232 return -EDEADLK;
233 /**
234 * Already reserved by a thread that will not back
235 * off for us. We need to back off.
236 */
237 if (unlikely(sequence - bo->val_seq < (1 << 31)))
238 return -EAGAIN;
239 }
240
241 if (no_wait)
242 return -EBUSY;
243
244 spin_unlock(&glob->lru_lock);
245 ret = ttm_bo_wait_unreserved(bo, interruptible);
246 spin_lock(&glob->lru_lock);
247
248 if (unlikely(ret))
249 return ret;
250 }
251
252 if (use_sequence) {
253 /**
254 * Wake up waiters that may need to recheck for deadlock,
255 * if we decreased the sequence number.
256 */
257 if (unlikely((bo->val_seq - sequence < (1 << 31))
258 || !bo->seq_valid))
259 wake_up_all(&bo->event_queue);
260
261 bo->val_seq = sequence;
262 bo->seq_valid = true;
263 } else {
264 bo->seq_valid = false;
265 }
266
267 return 0;
268}
269EXPORT_SYMBOL(ttm_bo_reserve);
270
271static void ttm_bo_ref_bug(struct kref *list_kref)
272{
273 BUG();
274}
275
276void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
277 bool never_free)
278{
279 kref_sub(&bo->list_kref, count,
280 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
281}
282
283int ttm_bo_reserve(struct ttm_buffer_object *bo,
284 bool interruptible,
285 bool no_wait, bool use_sequence, uint32_t sequence)
286{
287 struct ttm_bo_global *glob = bo->glob;
288 int put_count = 0;
289 int ret;
290
291 spin_lock(&glob->lru_lock);
292 ret = ttm_bo_reserve_locked(bo, interruptible, no_wait, use_sequence,
293 sequence);
294 if (likely(ret == 0))
295 put_count = ttm_bo_del_from_lru(bo);
296 spin_unlock(&glob->lru_lock);
297
298 ttm_bo_list_ref_sub(bo, put_count, true);
299
300 return ret;
301}
302
303void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo)
304{
305 ttm_bo_add_to_lru(bo);
306 atomic_set(&bo->reserved, 0);
307 wake_up_all(&bo->event_queue);
308}
309
310void ttm_bo_unreserve(struct ttm_buffer_object *bo)
311{
312 struct ttm_bo_global *glob = bo->glob;
313
314 spin_lock(&glob->lru_lock);
315 ttm_bo_unreserve_locked(bo);
316 spin_unlock(&glob->lru_lock);
317}
318EXPORT_SYMBOL(ttm_bo_unreserve);
319
320/*
321 * Call bo->mutex locked.
322 */
323static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
324{
325 struct ttm_bo_device *bdev = bo->bdev;
326 struct ttm_bo_global *glob = bo->glob;
327 int ret = 0;
328 uint32_t page_flags = 0;
329
330 TTM_ASSERT_LOCKED(&bo->mutex);
331 bo->ttm = NULL;
332
333 if (bdev->need_dma32)
334 page_flags |= TTM_PAGE_FLAG_DMA32;
335
336 switch (bo->type) {
337 case ttm_bo_type_device:
338 if (zero_alloc)
339 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
340 case ttm_bo_type_kernel:
341 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
342 page_flags, glob->dummy_read_page);
343 if (unlikely(bo->ttm == NULL))
344 ret = -ENOMEM;
345 break;
346 case ttm_bo_type_sg:
347 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
348 page_flags | TTM_PAGE_FLAG_SG,
349 glob->dummy_read_page);
350 if (unlikely(bo->ttm == NULL)) {
351 ret = -ENOMEM;
352 break;
353 }
354 bo->ttm->sg = bo->sg;
355 break;
356 default:
357 pr_err("Illegal buffer object type\n");
358 ret = -EINVAL;
359 break;
360 }
361
362 return ret;
363}
364
365static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
366 struct ttm_mem_reg *mem,
367 bool evict, bool interruptible,
368 bool no_wait_reserve, bool no_wait_gpu)
369{
370 struct ttm_bo_device *bdev = bo->bdev;
371 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
372 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
373 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
374 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
375 int ret = 0;
376
377 if (old_is_pci || new_is_pci ||
378 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
379 ret = ttm_mem_io_lock(old_man, true);
380 if (unlikely(ret != 0))
381 goto out_err;
382 ttm_bo_unmap_virtual_locked(bo);
383 ttm_mem_io_unlock(old_man);
384 }
385
386 /*
387 * Create and bind a ttm if required.
388 */
389
390 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
391 if (bo->ttm == NULL) {
392 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
393 ret = ttm_bo_add_ttm(bo, zero);
394 if (ret)
395 goto out_err;
396 }
397
398 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
399 if (ret)
400 goto out_err;
401
402 if (mem->mem_type != TTM_PL_SYSTEM) {
403 ret = ttm_tt_bind(bo->ttm, mem);
404 if (ret)
405 goto out_err;
406 }
407
408 if (bo->mem.mem_type == TTM_PL_SYSTEM) {
409 if (bdev->driver->move_notify)
410 bdev->driver->move_notify(bo, mem);
411 bo->mem = *mem;
412 mem->mm_node = NULL;
413 goto moved;
414 }
415 }
416
417 if (bdev->driver->move_notify)
418 bdev->driver->move_notify(bo, mem);
419
420 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
421 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
422 ret = ttm_bo_move_ttm(bo, evict, no_wait_reserve, no_wait_gpu, mem);
423 else if (bdev->driver->move)
424 ret = bdev->driver->move(bo, evict, interruptible,
425 no_wait_reserve, no_wait_gpu, mem);
426 else
427 ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, mem);
428
429 if (ret) {
430 if (bdev->driver->move_notify) {
431 struct ttm_mem_reg tmp_mem = *mem;
432 *mem = bo->mem;
433 bo->mem = tmp_mem;
434 bdev->driver->move_notify(bo, mem);
435 bo->mem = *mem;
436 }
437
438 goto out_err;
439 }
440
441moved:
442 if (bo->evicted) {
443 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
444 if (ret)
445 pr_err("Can not flush read caches\n");
446 bo->evicted = false;
447 }
448
449 if (bo->mem.mm_node) {
450 bo->offset = (bo->mem.start << PAGE_SHIFT) +
451 bdev->man[bo->mem.mem_type].gpu_offset;
452 bo->cur_placement = bo->mem.placement;
453 } else
454 bo->offset = 0;
455
456 return 0;
457
458out_err:
459 new_man = &bdev->man[bo->mem.mem_type];
460 if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
461 ttm_tt_unbind(bo->ttm);
462 ttm_tt_destroy(bo->ttm);
463 bo->ttm = NULL;
464 }
465
466 return ret;
467}
468
469/**
470 * Call bo::reserved.
471 * Will release GPU memory type usage on destruction.
472 * This is the place to put in driver specific hooks to release
473 * driver private resources.
474 * Will release the bo::reserved lock.
475 */
476
477static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
478{
479 if (bo->bdev->driver->move_notify)
480 bo->bdev->driver->move_notify(bo, NULL);
481
482 if (bo->ttm) {
483 ttm_tt_unbind(bo->ttm);
484 ttm_tt_destroy(bo->ttm);
485 bo->ttm = NULL;
486 }
487 ttm_bo_mem_put(bo, &bo->mem);
488
489 atomic_set(&bo->reserved, 0);
490
491 /*
492 * Make processes trying to reserve really pick it up.
493 */
494 smp_mb__after_atomic_dec();
495 wake_up_all(&bo->event_queue);
496}
497
498static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
499{
500 struct ttm_bo_device *bdev = bo->bdev;
501 struct ttm_bo_global *glob = bo->glob;
502 struct ttm_bo_driver *driver;
503 void *sync_obj = NULL;
504 void *sync_obj_arg;
505 int put_count;
506 int ret;
507
508 spin_lock(&bdev->fence_lock);
509 (void) ttm_bo_wait(bo, false, false, true);
510 if (!bo->sync_obj) {
511
512 spin_lock(&glob->lru_lock);
513
514 /**
515 * Lock inversion between bo:reserve and bdev::fence_lock here,
516 * but that's OK, since we're only trylocking.
517 */
518
519 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
520
521 if (unlikely(ret == -EBUSY))
522 goto queue;
523
524 spin_unlock(&bdev->fence_lock);
525 put_count = ttm_bo_del_from_lru(bo);
526
527 spin_unlock(&glob->lru_lock);
528 ttm_bo_cleanup_memtype_use(bo);
529
530 ttm_bo_list_ref_sub(bo, put_count, true);
531
532 return;
533 } else {
534 spin_lock(&glob->lru_lock);
535 }
536queue:
537 driver = bdev->driver;
538 if (bo->sync_obj)
539 sync_obj = driver->sync_obj_ref(bo->sync_obj);
540 sync_obj_arg = bo->sync_obj_arg;
541
542 kref_get(&bo->list_kref);
543 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
544 spin_unlock(&glob->lru_lock);
545 spin_unlock(&bdev->fence_lock);
546
547 if (sync_obj) {
548 driver->sync_obj_flush(sync_obj, sync_obj_arg);
549 driver->sync_obj_unref(&sync_obj);
550 }
551 schedule_delayed_work(&bdev->wq,
552 ((HZ / 100) < 1) ? 1 : HZ / 100);
553}
554
555/**
556 * function ttm_bo_cleanup_refs
557 * If bo idle, remove from delayed- and lru lists, and unref.
558 * If not idle, do nothing.
559 *
560 * @interruptible Any sleeps should occur interruptibly.
561 * @no_wait_reserve Never wait for reserve. Return -EBUSY instead.
562 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead.
563 */
564
565static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
566 bool interruptible,
567 bool no_wait_reserve,
568 bool no_wait_gpu)
569{
570 struct ttm_bo_device *bdev = bo->bdev;
571 struct ttm_bo_global *glob = bo->glob;
572 int put_count;
573 int ret = 0;
574
575retry:
576 spin_lock(&bdev->fence_lock);
577 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
578 spin_unlock(&bdev->fence_lock);
579
580 if (unlikely(ret != 0))
581 return ret;
582
583 spin_lock(&glob->lru_lock);
584
585 if (unlikely(list_empty(&bo->ddestroy))) {
586 spin_unlock(&glob->lru_lock);
587 return 0;
588 }
589
590 ret = ttm_bo_reserve_locked(bo, interruptible,
591 no_wait_reserve, false, 0);
592
593 if (unlikely(ret != 0)) {
594 spin_unlock(&glob->lru_lock);
595 return ret;
596 }
597
598 /**
599 * We can re-check for sync object without taking
600 * the bo::lock since setting the sync object requires
601 * also bo::reserved. A busy object at this point may
602 * be caused by another thread recently starting an accelerated
603 * eviction.
604 */
605
606 if (unlikely(bo->sync_obj)) {
607 atomic_set(&bo->reserved, 0);
608 wake_up_all(&bo->event_queue);
609 spin_unlock(&glob->lru_lock);
610 goto retry;
611 }
612
613 put_count = ttm_bo_del_from_lru(bo);
614 list_del_init(&bo->ddestroy);
615 ++put_count;
616
617 spin_unlock(&glob->lru_lock);
618 ttm_bo_cleanup_memtype_use(bo);
619
620 ttm_bo_list_ref_sub(bo, put_count, true);
621
622 return 0;
623}
624
625/**
626 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
627 * encountered buffers.
628 */
629
630static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
631{
632 struct ttm_bo_global *glob = bdev->glob;
633 struct ttm_buffer_object *entry = NULL;
634 int ret = 0;
635
636 spin_lock(&glob->lru_lock);
637 if (list_empty(&bdev->ddestroy))
638 goto out_unlock;
639
640 entry = list_first_entry(&bdev->ddestroy,
641 struct ttm_buffer_object, ddestroy);
642 kref_get(&entry->list_kref);
643
644 for (;;) {
645 struct ttm_buffer_object *nentry = NULL;
646
647 if (entry->ddestroy.next != &bdev->ddestroy) {
648 nentry = list_first_entry(&entry->ddestroy,
649 struct ttm_buffer_object, ddestroy);
650 kref_get(&nentry->list_kref);
651 }
652
653 spin_unlock(&glob->lru_lock);
654 ret = ttm_bo_cleanup_refs(entry, false, !remove_all,
655 !remove_all);
656 kref_put(&entry->list_kref, ttm_bo_release_list);
657 entry = nentry;
658
659 if (ret || !entry)
660 goto out;
661
662 spin_lock(&glob->lru_lock);
663 if (list_empty(&entry->ddestroy))
664 break;
665 }
666
667out_unlock:
668 spin_unlock(&glob->lru_lock);
669out:
670 if (entry)
671 kref_put(&entry->list_kref, ttm_bo_release_list);
672 return ret;
673}
674
675static void ttm_bo_delayed_workqueue(struct work_struct *work)
676{
677 struct ttm_bo_device *bdev =
678 container_of(work, struct ttm_bo_device, wq.work);
679
680 if (ttm_bo_delayed_delete(bdev, false)) {
681 schedule_delayed_work(&bdev->wq,
682 ((HZ / 100) < 1) ? 1 : HZ / 100);
683 }
684}
685
686static void ttm_bo_release(struct kref *kref)
687{
688 struct ttm_buffer_object *bo =
689 container_of(kref, struct ttm_buffer_object, kref);
690 struct ttm_bo_device *bdev = bo->bdev;
691 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
692
693 if (likely(bo->vm_node != NULL)) {
694 rb_erase(&bo->vm_rb, &bdev->addr_space_rb);
695 drm_mm_put_block(bo->vm_node);
696 bo->vm_node = NULL;
697 }
698 write_unlock(&bdev->vm_lock);
699 ttm_mem_io_lock(man, false);
700 ttm_mem_io_free_vm(bo);
701 ttm_mem_io_unlock(man);
702 ttm_bo_cleanup_refs_or_queue(bo);
703 kref_put(&bo->list_kref, ttm_bo_release_list);
704 write_lock(&bdev->vm_lock);
705}
706
707void ttm_bo_unref(struct ttm_buffer_object **p_bo)
708{
709 struct ttm_buffer_object *bo = *p_bo;
710 struct ttm_bo_device *bdev = bo->bdev;
711
712 *p_bo = NULL;
713 write_lock(&bdev->vm_lock);
714 kref_put(&bo->kref, ttm_bo_release);
715 write_unlock(&bdev->vm_lock);
716}
717EXPORT_SYMBOL(ttm_bo_unref);
718
719int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
720{
721 return cancel_delayed_work_sync(&bdev->wq);
722}
723EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
724
725void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
726{
727 if (resched)
728 schedule_delayed_work(&bdev->wq,
729 ((HZ / 100) < 1) ? 1 : HZ / 100);
730}
731EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
732
733static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
734 bool no_wait_reserve, bool no_wait_gpu)
735{
736 struct ttm_bo_device *bdev = bo->bdev;
737 struct ttm_mem_reg evict_mem;
738 struct ttm_placement placement;
739 int ret = 0;
740
741 spin_lock(&bdev->fence_lock);
742 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
743 spin_unlock(&bdev->fence_lock);
744
745 if (unlikely(ret != 0)) {
746 if (ret != -ERESTARTSYS) {
747 pr_err("Failed to expire sync object before buffer eviction\n");
748 }
749 goto out;
750 }
751
752 BUG_ON(!atomic_read(&bo->reserved));
753
754 evict_mem = bo->mem;
755 evict_mem.mm_node = NULL;
756 evict_mem.bus.io_reserved_vm = false;
757 evict_mem.bus.io_reserved_count = 0;
758
759 placement.fpfn = 0;
760 placement.lpfn = 0;
761 placement.num_placement = 0;
762 placement.num_busy_placement = 0;
763 bdev->driver->evict_flags(bo, &placement);
764 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
765 no_wait_reserve, no_wait_gpu);
766 if (ret) {
767 if (ret != -ERESTARTSYS) {
768 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
769 bo);
770 ttm_bo_mem_space_debug(bo, &placement);
771 }
772 goto out;
773 }
774
775 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
776 no_wait_reserve, no_wait_gpu);
777 if (ret) {
778 if (ret != -ERESTARTSYS)
779 pr_err("Buffer eviction failed\n");
780 ttm_bo_mem_put(bo, &evict_mem);
781 goto out;
782 }
783 bo->evicted = true;
784out:
785 return ret;
786}
787
788static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
789 uint32_t mem_type,
790 bool interruptible, bool no_wait_reserve,
791 bool no_wait_gpu)
792{
793 struct ttm_bo_global *glob = bdev->glob;
794 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
795 struct ttm_buffer_object *bo;
796 int ret, put_count = 0;
797
798retry:
799 spin_lock(&glob->lru_lock);
800 if (list_empty(&man->lru)) {
801 spin_unlock(&glob->lru_lock);
802 return -EBUSY;
803 }
804
805 bo = list_first_entry(&man->lru, struct ttm_buffer_object, lru);
806 kref_get(&bo->list_kref);
807
808 if (!list_empty(&bo->ddestroy)) {
809 spin_unlock(&glob->lru_lock);
810 ret = ttm_bo_cleanup_refs(bo, interruptible,
811 no_wait_reserve, no_wait_gpu);
812 kref_put(&bo->list_kref, ttm_bo_release_list);
813
814 if (likely(ret == 0 || ret == -ERESTARTSYS))
815 return ret;
816
817 goto retry;
818 }
819
820 ret = ttm_bo_reserve_locked(bo, false, no_wait_reserve, false, 0);
821
822 if (unlikely(ret == -EBUSY)) {
823 spin_unlock(&glob->lru_lock);
824 if (likely(!no_wait_gpu))
825 ret = ttm_bo_wait_unreserved(bo, interruptible);
826
827 kref_put(&bo->list_kref, ttm_bo_release_list);
828
829 /**
830 * We *need* to retry after releasing the lru lock.
831 */
832
833 if (unlikely(ret != 0))
834 return ret;
835 goto retry;
836 }
837
838 put_count = ttm_bo_del_from_lru(bo);
839 spin_unlock(&glob->lru_lock);
840
841 BUG_ON(ret != 0);
842
843 ttm_bo_list_ref_sub(bo, put_count, true);
844
845 ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
846 ttm_bo_unreserve(bo);
847
848 kref_put(&bo->list_kref, ttm_bo_release_list);
849 return ret;
850}
851
852void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
853{
854 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
855
856 if (mem->mm_node)
857 (*man->func->put_node)(man, mem);
858}
859EXPORT_SYMBOL(ttm_bo_mem_put);
860
861/**
862 * Repeatedly evict memory from the LRU for @mem_type until we create enough
863 * space, or we've evicted everything and there isn't enough space.
864 */
865static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
866 uint32_t mem_type,
867 struct ttm_placement *placement,
868 struct ttm_mem_reg *mem,
869 bool interruptible,
870 bool no_wait_reserve,
871 bool no_wait_gpu)
872{
873 struct ttm_bo_device *bdev = bo->bdev;
874 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
875 int ret;
876
877 do {
878 ret = (*man->func->get_node)(man, bo, placement, mem);
879 if (unlikely(ret != 0))
880 return ret;
881 if (mem->mm_node)
882 break;
883 ret = ttm_mem_evict_first(bdev, mem_type, interruptible,
884 no_wait_reserve, no_wait_gpu);
885 if (unlikely(ret != 0))
886 return ret;
887 } while (1);
888 if (mem->mm_node == NULL)
889 return -ENOMEM;
890 mem->mem_type = mem_type;
891 return 0;
892}
893
894static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
895 uint32_t cur_placement,
896 uint32_t proposed_placement)
897{
898 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
899 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
900
901 /**
902 * Keep current caching if possible.
903 */
904
905 if ((cur_placement & caching) != 0)
906 result |= (cur_placement & caching);
907 else if ((man->default_caching & caching) != 0)
908 result |= man->default_caching;
909 else if ((TTM_PL_FLAG_CACHED & caching) != 0)
910 result |= TTM_PL_FLAG_CACHED;
911 else if ((TTM_PL_FLAG_WC & caching) != 0)
912 result |= TTM_PL_FLAG_WC;
913 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
914 result |= TTM_PL_FLAG_UNCACHED;
915
916 return result;
917}
918
919static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
920 uint32_t mem_type,
921 uint32_t proposed_placement,
922 uint32_t *masked_placement)
923{
924 uint32_t cur_flags = ttm_bo_type_flags(mem_type);
925
926 if ((cur_flags & proposed_placement & TTM_PL_MASK_MEM) == 0)
927 return false;
928
929 if ((proposed_placement & man->available_caching) == 0)
930 return false;
931
932 cur_flags |= (proposed_placement & man->available_caching);
933
934 *masked_placement = cur_flags;
935 return true;
936}
937
938/**
939 * Creates space for memory region @mem according to its type.
940 *
941 * This function first searches for free space in compatible memory types in
942 * the priority order defined by the driver. If free space isn't found, then
943 * ttm_bo_mem_force_space is attempted in priority order to evict and find
944 * space.
945 */
946int ttm_bo_mem_space(struct ttm_buffer_object *bo,
947 struct ttm_placement *placement,
948 struct ttm_mem_reg *mem,
949 bool interruptible, bool no_wait_reserve,
950 bool no_wait_gpu)
951{
952 struct ttm_bo_device *bdev = bo->bdev;
953 struct ttm_mem_type_manager *man;
954 uint32_t mem_type = TTM_PL_SYSTEM;
955 uint32_t cur_flags = 0;
956 bool type_found = false;
957 bool type_ok = false;
958 bool has_erestartsys = false;
959 int i, ret;
960
961 mem->mm_node = NULL;
962 for (i = 0; i < placement->num_placement; ++i) {
963 ret = ttm_mem_type_from_flags(placement->placement[i],
964 &mem_type);
965 if (ret)
966 return ret;
967 man = &bdev->man[mem_type];
968
969 type_ok = ttm_bo_mt_compatible(man,
970 mem_type,
971 placement->placement[i],
972 &cur_flags);
973
974 if (!type_ok)
975 continue;
976
977 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
978 cur_flags);
979 /*
980 * Use the access and other non-mapping-related flag bits from
981 * the memory placement flags to the current flags
982 */
983 ttm_flag_masked(&cur_flags, placement->placement[i],
984 ~TTM_PL_MASK_MEMTYPE);
985
986 if (mem_type == TTM_PL_SYSTEM)
987 break;
988
989 if (man->has_type && man->use_type) {
990 type_found = true;
991 ret = (*man->func->get_node)(man, bo, placement, mem);
992 if (unlikely(ret))
993 return ret;
994 }
995 if (mem->mm_node)
996 break;
997 }
998
999 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
1000 mem->mem_type = mem_type;
1001 mem->placement = cur_flags;
1002 return 0;
1003 }
1004
1005 if (!type_found)
1006 return -EINVAL;
1007
1008 for (i = 0; i < placement->num_busy_placement; ++i) {
1009 ret = ttm_mem_type_from_flags(placement->busy_placement[i],
1010 &mem_type);
1011 if (ret)
1012 return ret;
1013 man = &bdev->man[mem_type];
1014 if (!man->has_type)
1015 continue;
1016 if (!ttm_bo_mt_compatible(man,
1017 mem_type,
1018 placement->busy_placement[i],
1019 &cur_flags))
1020 continue;
1021
1022 cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
1023 cur_flags);
1024 /*
1025 * Use the access and other non-mapping-related flag bits from
1026 * the memory placement flags to the current flags
1027 */
1028 ttm_flag_masked(&cur_flags, placement->busy_placement[i],
1029 ~TTM_PL_MASK_MEMTYPE);
1030
1031
1032 if (mem_type == TTM_PL_SYSTEM) {
1033 mem->mem_type = mem_type;
1034 mem->placement = cur_flags;
1035 mem->mm_node = NULL;
1036 return 0;
1037 }
1038
1039 ret = ttm_bo_mem_force_space(bo, mem_type, placement, mem,
1040 interruptible, no_wait_reserve, no_wait_gpu);
1041 if (ret == 0 && mem->mm_node) {
1042 mem->placement = cur_flags;
1043 return 0;
1044 }
1045 if (ret == -ERESTARTSYS)
1046 has_erestartsys = true;
1047 }
1048 ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1049 return ret;
1050}
1051EXPORT_SYMBOL(ttm_bo_mem_space);
1052
1053int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait)
1054{
1055 if ((atomic_read(&bo->cpu_writers) > 0) && no_wait)
1056 return -EBUSY;
1057
1058 return wait_event_interruptible(bo->event_queue,
1059 atomic_read(&bo->cpu_writers) == 0);
1060}
1061EXPORT_SYMBOL(ttm_bo_wait_cpu);
1062
1063int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1064 struct ttm_placement *placement,
1065 bool interruptible, bool no_wait_reserve,
1066 bool no_wait_gpu)
1067{
1068 int ret = 0;
1069 struct ttm_mem_reg mem;
1070 struct ttm_bo_device *bdev = bo->bdev;
1071
1072 BUG_ON(!atomic_read(&bo->reserved));
1073
1074 /*
1075 * FIXME: It's possible to pipeline buffer moves.
1076 * Have the driver move function wait for idle when necessary,
1077 * instead of doing it here.
1078 */
1079 spin_lock(&bdev->fence_lock);
1080 ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
1081 spin_unlock(&bdev->fence_lock);
1082 if (ret)
1083 return ret;
1084 mem.num_pages = bo->num_pages;
1085 mem.size = mem.num_pages << PAGE_SHIFT;
1086 mem.page_alignment = bo->mem.page_alignment;
1087 mem.bus.io_reserved_vm = false;
1088 mem.bus.io_reserved_count = 0;
1089 /*
1090 * Determine where to move the buffer.
1091 */
1092 ret = ttm_bo_mem_space(bo, placement, &mem, interruptible, no_wait_reserve, no_wait_gpu);
1093 if (ret)
1094 goto out_unlock;
1095 ret = ttm_bo_handle_move_mem(bo, &mem, false, interruptible, no_wait_reserve, no_wait_gpu);
1096out_unlock:
1097 if (ret && mem.mm_node)
1098 ttm_bo_mem_put(bo, &mem);
1099 return ret;
1100}
1101
1102static int ttm_bo_mem_compat(struct ttm_placement *placement,
1103 struct ttm_mem_reg *mem)
1104{
1105 int i;
1106
1107 if (mem->mm_node && placement->lpfn != 0 &&
1108 (mem->start < placement->fpfn ||
1109 mem->start + mem->num_pages > placement->lpfn))
1110 return -1;
1111
1112 for (i = 0; i < placement->num_placement; i++) {
1113 if ((placement->placement[i] & mem->placement &
1114 TTM_PL_MASK_CACHING) &&
1115 (placement->placement[i] & mem->placement &
1116 TTM_PL_MASK_MEM))
1117 return i;
1118 }
1119 return -1;
1120}
1121
1122int ttm_bo_validate(struct ttm_buffer_object *bo,
1123 struct ttm_placement *placement,
1124 bool interruptible, bool no_wait_reserve,
1125 bool no_wait_gpu)
1126{
1127 int ret;
1128
1129 BUG_ON(!atomic_read(&bo->reserved));
1130 /* Check that range is valid */
1131 if (placement->lpfn || placement->fpfn)
1132 if (placement->fpfn > placement->lpfn ||
1133 (placement->lpfn - placement->fpfn) < bo->num_pages)
1134 return -EINVAL;
1135 /*
1136 * Check whether we need to move buffer.
1137 */
1138 ret = ttm_bo_mem_compat(placement, &bo->mem);
1139 if (ret < 0) {
1140 ret = ttm_bo_move_buffer(bo, placement, interruptible, no_wait_reserve, no_wait_gpu);
1141 if (ret)
1142 return ret;
1143 } else {
1144 /*
1145 * Use the access and other non-mapping-related flag bits from
1146 * the compatible memory placement flags to the active flags
1147 */
1148 ttm_flag_masked(&bo->mem.placement, placement->placement[ret],
1149 ~TTM_PL_MASK_MEMTYPE);
1150 }
1151 /*
1152 * We might need to add a TTM.
1153 */
1154 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1155 ret = ttm_bo_add_ttm(bo, true);
1156 if (ret)
1157 return ret;
1158 }
1159 return 0;
1160}
1161EXPORT_SYMBOL(ttm_bo_validate);
1162
1163int ttm_bo_check_placement(struct ttm_buffer_object *bo,
1164 struct ttm_placement *placement)
1165{
1166 BUG_ON((placement->fpfn || placement->lpfn) &&
1167 (bo->mem.num_pages > (placement->lpfn - placement->fpfn)));
1168
1169 return 0;
1170}
1171
1172int ttm_bo_init(struct ttm_bo_device *bdev,
1173 struct ttm_buffer_object *bo,
1174 unsigned long size,
1175 enum ttm_bo_type type,
1176 struct ttm_placement *placement,
1177 uint32_t page_alignment,
1178 unsigned long buffer_start,
1179 bool interruptible,
1180 struct file *persistent_swap_storage,
1181 size_t acc_size,
1182 struct sg_table *sg,
1183 void (*destroy) (struct ttm_buffer_object *))
1184{
1185 int ret = 0;
1186 unsigned long num_pages;
1187 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1188
1189 ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1190 if (ret) {
1191 pr_err("Out of kernel memory\n");
1192 if (destroy)
1193 (*destroy)(bo);
1194 else
1195 kfree(bo);
1196 return -ENOMEM;
1197 }
1198
1199 size += buffer_start & ~PAGE_MASK;
1200 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1201 if (num_pages == 0) {
1202 pr_err("Illegal buffer object size\n");
1203 if (destroy)
1204 (*destroy)(bo);
1205 else
1206 kfree(bo);
1207 ttm_mem_global_free(mem_glob, acc_size);
1208 return -EINVAL;
1209 }
1210 bo->destroy = destroy;
1211
1212 kref_init(&bo->kref);
1213 kref_init(&bo->list_kref);
1214 atomic_set(&bo->cpu_writers, 0);
1215 atomic_set(&bo->reserved, 1);
1216 init_waitqueue_head(&bo->event_queue);
1217 INIT_LIST_HEAD(&bo->lru);
1218 INIT_LIST_HEAD(&bo->ddestroy);
1219 INIT_LIST_HEAD(&bo->swap);
1220 INIT_LIST_HEAD(&bo->io_reserve_lru);
1221 bo->bdev = bdev;
1222 bo->glob = bdev->glob;
1223 bo->type = type;
1224 bo->num_pages = num_pages;
1225 bo->mem.size = num_pages << PAGE_SHIFT;
1226 bo->mem.mem_type = TTM_PL_SYSTEM;
1227 bo->mem.num_pages = bo->num_pages;
1228 bo->mem.mm_node = NULL;
1229 bo->mem.page_alignment = page_alignment;
1230 bo->mem.bus.io_reserved_vm = false;
1231 bo->mem.bus.io_reserved_count = 0;
1232 bo->buffer_start = buffer_start & PAGE_MASK;
1233 bo->priv_flags = 0;
1234 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1235 bo->seq_valid = false;
1236 bo->persistent_swap_storage = persistent_swap_storage;
1237 bo->acc_size = acc_size;
1238 bo->sg = sg;
1239 atomic_inc(&bo->glob->bo_count);
1240
1241 ret = ttm_bo_check_placement(bo, placement);
1242 if (unlikely(ret != 0))
1243 goto out_err;
1244
1245 /*
1246 * For ttm_bo_type_device buffers, allocate
1247 * address space from the device.
1248 */
1249 if (bo->type == ttm_bo_type_device ||
1250 bo->type == ttm_bo_type_sg) {
1251 ret = ttm_bo_setup_vm(bo);
1252 if (ret)
1253 goto out_err;
1254 }
1255
1256 ret = ttm_bo_validate(bo, placement, interruptible, false, false);
1257 if (ret)
1258 goto out_err;
1259
1260 ttm_bo_unreserve(bo);
1261 return 0;
1262
1263out_err:
1264 ttm_bo_unreserve(bo);
1265 ttm_bo_unref(&bo);
1266
1267 return ret;
1268}
1269EXPORT_SYMBOL(ttm_bo_init);
1270
1271size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1272 unsigned long bo_size,
1273 unsigned struct_size)
1274{
1275 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1276 size_t size = 0;
1277
1278 size += ttm_round_pot(struct_size);
1279 size += PAGE_ALIGN(npages * sizeof(void *));
1280 size += ttm_round_pot(sizeof(struct ttm_tt));
1281 return size;
1282}
1283EXPORT_SYMBOL(ttm_bo_acc_size);
1284
1285size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1286 unsigned long bo_size,
1287 unsigned struct_size)
1288{
1289 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1290 size_t size = 0;
1291
1292 size += ttm_round_pot(struct_size);
1293 size += PAGE_ALIGN(npages * sizeof(void *));
1294 size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
1295 size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1296 return size;
1297}
1298EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1299
1300int ttm_bo_create(struct ttm_bo_device *bdev,
1301 unsigned long size,
1302 enum ttm_bo_type type,
1303 struct ttm_placement *placement,
1304 uint32_t page_alignment,
1305 unsigned long buffer_start,
1306 bool interruptible,
1307 struct file *persistent_swap_storage,
1308 struct ttm_buffer_object **p_bo)
1309{
1310 struct ttm_buffer_object *bo;
1311 size_t acc_size;
1312 int ret;
1313
1314 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1315 if (unlikely(bo == NULL))
1316 return -ENOMEM;
1317
1318 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1319 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1320 buffer_start, interruptible,
1321 persistent_swap_storage, acc_size, NULL, NULL);
1322 if (likely(ret == 0))
1323 *p_bo = bo;
1324
1325 return ret;
1326}
1327EXPORT_SYMBOL(ttm_bo_create);
1328
1329static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1330 unsigned mem_type, bool allow_errors)
1331{
1332 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1333 struct ttm_bo_global *glob = bdev->glob;
1334 int ret;
1335
1336 /*
1337 * Can't use standard list traversal since we're unlocking.
1338 */
1339
1340 spin_lock(&glob->lru_lock);
1341 while (!list_empty(&man->lru)) {
1342 spin_unlock(&glob->lru_lock);
1343 ret = ttm_mem_evict_first(bdev, mem_type, false, false, false);
1344 if (ret) {
1345 if (allow_errors) {
1346 return ret;
1347 } else {
1348 pr_err("Cleanup eviction failed\n");
1349 }
1350 }
1351 spin_lock(&glob->lru_lock);
1352 }
1353 spin_unlock(&glob->lru_lock);
1354 return 0;
1355}
1356
1357int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1358{
1359 struct ttm_mem_type_manager *man;
1360 int ret = -EINVAL;
1361
1362 if (mem_type >= TTM_NUM_MEM_TYPES) {
1363 pr_err("Illegal memory type %d\n", mem_type);
1364 return ret;
1365 }
1366 man = &bdev->man[mem_type];
1367
1368 if (!man->has_type) {
1369 pr_err("Trying to take down uninitialized memory manager type %u\n",
1370 mem_type);
1371 return ret;
1372 }
1373
1374 man->use_type = false;
1375 man->has_type = false;
1376
1377 ret = 0;
1378 if (mem_type > 0) {
1379 ttm_bo_force_list_clean(bdev, mem_type, false);
1380
1381 ret = (*man->func->takedown)(man);
1382 }
1383
1384 return ret;
1385}
1386EXPORT_SYMBOL(ttm_bo_clean_mm);
1387
1388int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1389{
1390 struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1391
1392 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1393 pr_err("Illegal memory manager memory type %u\n", mem_type);
1394 return -EINVAL;
1395 }
1396
1397 if (!man->has_type) {
1398 pr_err("Memory type %u has not been initialized\n", mem_type);
1399 return 0;
1400 }
1401
1402 return ttm_bo_force_list_clean(bdev, mem_type, true);
1403}
1404EXPORT_SYMBOL(ttm_bo_evict_mm);
1405
1406int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1407 unsigned long p_size)
1408{
1409 int ret = -EINVAL;
1410 struct ttm_mem_type_manager *man;
1411
1412 BUG_ON(type >= TTM_NUM_MEM_TYPES);
1413 man = &bdev->man[type];
1414 BUG_ON(man->has_type);
1415 man->io_reserve_fastpath = true;
1416 man->use_io_reserve_lru = false;
1417 mutex_init(&man->io_reserve_mutex);
1418 INIT_LIST_HEAD(&man->io_reserve_lru);
1419
1420 ret = bdev->driver->init_mem_type(bdev, type, man);
1421 if (ret)
1422 return ret;
1423 man->bdev = bdev;
1424
1425 ret = 0;
1426 if (type != TTM_PL_SYSTEM) {
1427 ret = (*man->func->init)(man, p_size);
1428 if (ret)
1429 return ret;
1430 }
1431 man->has_type = true;
1432 man->use_type = true;
1433 man->size = p_size;
1434
1435 INIT_LIST_HEAD(&man->lru);
1436
1437 return 0;
1438}
1439EXPORT_SYMBOL(ttm_bo_init_mm);
1440
1441static void ttm_bo_global_kobj_release(struct kobject *kobj)
1442{
1443 struct ttm_bo_global *glob =
1444 container_of(kobj, struct ttm_bo_global, kobj);
1445
1446 ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1447 __free_page(glob->dummy_read_page);
1448 kfree(glob);
1449}
1450
1451void ttm_bo_global_release(struct drm_global_reference *ref)
1452{
1453 struct ttm_bo_global *glob = ref->object;
1454
1455 kobject_del(&glob->kobj);
1456 kobject_put(&glob->kobj);
1457}
1458EXPORT_SYMBOL(ttm_bo_global_release);
1459
1460int ttm_bo_global_init(struct drm_global_reference *ref)
1461{
1462 struct ttm_bo_global_ref *bo_ref =
1463 container_of(ref, struct ttm_bo_global_ref, ref);
1464 struct ttm_bo_global *glob = ref->object;
1465 int ret;
1466
1467 mutex_init(&glob->device_list_mutex);
1468 spin_lock_init(&glob->lru_lock);
1469 glob->mem_glob = bo_ref->mem_glob;
1470 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1471
1472 if (unlikely(glob->dummy_read_page == NULL)) {
1473 ret = -ENOMEM;
1474 goto out_no_drp;
1475 }
1476
1477 INIT_LIST_HEAD(&glob->swap_lru);
1478 INIT_LIST_HEAD(&glob->device_list);
1479
1480 ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1481 ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1482 if (unlikely(ret != 0)) {
1483 pr_err("Could not register buffer object swapout\n");
1484 goto out_no_shrink;
1485 }
1486
1487 atomic_set(&glob->bo_count, 0);
1488
1489 ret = kobject_init_and_add(
1490 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1491 if (unlikely(ret != 0))
1492 kobject_put(&glob->kobj);
1493 return ret;
1494out_no_shrink:
1495 __free_page(glob->dummy_read_page);
1496out_no_drp:
1497 kfree(glob);
1498 return ret;
1499}
1500EXPORT_SYMBOL(ttm_bo_global_init);
1501
1502
1503int ttm_bo_device_release(struct ttm_bo_device *bdev)
1504{
1505 int ret = 0;
1506 unsigned i = TTM_NUM_MEM_TYPES;
1507 struct ttm_mem_type_manager *man;
1508 struct ttm_bo_global *glob = bdev->glob;
1509
1510 while (i--) {
1511 man = &bdev->man[i];
1512 if (man->has_type) {
1513 man->use_type = false;
1514 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1515 ret = -EBUSY;
1516 pr_err("DRM memory manager type %d is not clean\n",
1517 i);
1518 }
1519 man->has_type = false;
1520 }
1521 }
1522
1523 mutex_lock(&glob->device_list_mutex);
1524 list_del(&bdev->device_list);
1525 mutex_unlock(&glob->device_list_mutex);
1526
1527 cancel_delayed_work_sync(&bdev->wq);
1528
1529 while (ttm_bo_delayed_delete(bdev, true))
1530 ;
1531
1532 spin_lock(&glob->lru_lock);
1533 if (list_empty(&bdev->ddestroy))
1534 TTM_DEBUG("Delayed destroy list was clean\n");
1535
1536 if (list_empty(&bdev->man[0].lru))
1537 TTM_DEBUG("Swap list was clean\n");
1538 spin_unlock(&glob->lru_lock);
1539
1540 BUG_ON(!drm_mm_clean(&bdev->addr_space_mm));
1541 write_lock(&bdev->vm_lock);
1542 drm_mm_takedown(&bdev->addr_space_mm);
1543 write_unlock(&bdev->vm_lock);
1544
1545 return ret;
1546}
1547EXPORT_SYMBOL(ttm_bo_device_release);
1548
1549int ttm_bo_device_init(struct ttm_bo_device *bdev,
1550 struct ttm_bo_global *glob,
1551 struct ttm_bo_driver *driver,
1552 uint64_t file_page_offset,
1553 bool need_dma32)
1554{
1555 int ret = -EINVAL;
1556
1557 rwlock_init(&bdev->vm_lock);
1558 bdev->driver = driver;
1559
1560 memset(bdev->man, 0, sizeof(bdev->man));
1561
1562 /*
1563 * Initialize the system memory buffer type.
1564 * Other types need to be driver / IOCTL initialized.
1565 */
1566 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1567 if (unlikely(ret != 0))
1568 goto out_no_sys;
1569
1570 bdev->addr_space_rb = RB_ROOT;
1571 ret = drm_mm_init(&bdev->addr_space_mm, file_page_offset, 0x10000000);
1572 if (unlikely(ret != 0))
1573 goto out_no_addr_mm;
1574
1575 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1576 bdev->nice_mode = true;
1577 INIT_LIST_HEAD(&bdev->ddestroy);
1578 bdev->dev_mapping = NULL;
1579 bdev->glob = glob;
1580 bdev->need_dma32 = need_dma32;
1581 bdev->val_seq = 0;
1582 spin_lock_init(&bdev->fence_lock);
1583 mutex_lock(&glob->device_list_mutex);
1584 list_add_tail(&bdev->device_list, &glob->device_list);
1585 mutex_unlock(&glob->device_list_mutex);
1586
1587 return 0;
1588out_no_addr_mm:
1589 ttm_bo_clean_mm(bdev, 0);
1590out_no_sys:
1591 return ret;
1592}
1593EXPORT_SYMBOL(ttm_bo_device_init);
1594
1595/*
1596 * buffer object vm functions.
1597 */
1598
1599bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1600{
1601 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1602
1603 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1604 if (mem->mem_type == TTM_PL_SYSTEM)
1605 return false;
1606
1607 if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1608 return false;
1609
1610 if (mem->placement & TTM_PL_FLAG_CACHED)
1611 return false;
1612 }
1613 return true;
1614}
1615
1616void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1617{
1618 struct ttm_bo_device *bdev = bo->bdev;
1619 loff_t offset = (loff_t) bo->addr_space_offset;
1620 loff_t holelen = ((loff_t) bo->mem.num_pages) << PAGE_SHIFT;
1621
1622 if (!bdev->dev_mapping)
1623 return;
1624 unmap_mapping_range(bdev->dev_mapping, offset, holelen, 1);
1625 ttm_mem_io_free_vm(bo);
1626}
1627
1628void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1629{
1630 struct ttm_bo_device *bdev = bo->bdev;
1631 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1632
1633 ttm_mem_io_lock(man, false);
1634 ttm_bo_unmap_virtual_locked(bo);
1635 ttm_mem_io_unlock(man);
1636}
1637
1638
1639EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1640
1641static void ttm_bo_vm_insert_rb(struct ttm_buffer_object *bo)
1642{
1643 struct ttm_bo_device *bdev = bo->bdev;
1644 struct rb_node **cur = &bdev->addr_space_rb.rb_node;
1645 struct rb_node *parent = NULL;
1646 struct ttm_buffer_object *cur_bo;
1647 unsigned long offset = bo->vm_node->start;
1648 unsigned long cur_offset;
1649
1650 while (*cur) {
1651 parent = *cur;
1652 cur_bo = rb_entry(parent, struct ttm_buffer_object, vm_rb);
1653 cur_offset = cur_bo->vm_node->start;
1654 if (offset < cur_offset)
1655 cur = &parent->rb_left;
1656 else if (offset > cur_offset)
1657 cur = &parent->rb_right;
1658 else
1659 BUG();
1660 }
1661
1662 rb_link_node(&bo->vm_rb, parent, cur);
1663 rb_insert_color(&bo->vm_rb, &bdev->addr_space_rb);
1664}
1665
1666/**
1667 * ttm_bo_setup_vm:
1668 *
1669 * @bo: the buffer to allocate address space for
1670 *
1671 * Allocate address space in the drm device so that applications
1672 * can mmap the buffer and access the contents. This only
1673 * applies to ttm_bo_type_device objects as others are not
1674 * placed in the drm device address space.
1675 */
1676
1677static int ttm_bo_setup_vm(struct ttm_buffer_object *bo)
1678{
1679 struct ttm_bo_device *bdev = bo->bdev;
1680 int ret;
1681
1682retry_pre_get:
1683 ret = drm_mm_pre_get(&bdev->addr_space_mm);
1684 if (unlikely(ret != 0))
1685 return ret;
1686
1687 write_lock(&bdev->vm_lock);
1688 bo->vm_node = drm_mm_search_free(&bdev->addr_space_mm,
1689 bo->mem.num_pages, 0, 0);
1690
1691 if (unlikely(bo->vm_node == NULL)) {
1692 ret = -ENOMEM;
1693 goto out_unlock;
1694 }
1695
1696 bo->vm_node = drm_mm_get_block_atomic(bo->vm_node,
1697 bo->mem.num_pages, 0);
1698
1699 if (unlikely(bo->vm_node == NULL)) {
1700 write_unlock(&bdev->vm_lock);
1701 goto retry_pre_get;
1702 }
1703
1704 ttm_bo_vm_insert_rb(bo);
1705 write_unlock(&bdev->vm_lock);
1706 bo->addr_space_offset = ((uint64_t) bo->vm_node->start) << PAGE_SHIFT;
1707
1708 return 0;
1709out_unlock:
1710 write_unlock(&bdev->vm_lock);
1711 return ret;
1712}
1713
1714int ttm_bo_wait(struct ttm_buffer_object *bo,
1715 bool lazy, bool interruptible, bool no_wait)
1716{
1717 struct ttm_bo_driver *driver = bo->bdev->driver;
1718 struct ttm_bo_device *bdev = bo->bdev;
1719 void *sync_obj;
1720 void *sync_obj_arg;
1721 int ret = 0;
1722
1723 if (likely(bo->sync_obj == NULL))
1724 return 0;
1725
1726 while (bo->sync_obj) {
1727
1728 if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
1729 void *tmp_obj = bo->sync_obj;
1730 bo->sync_obj = NULL;
1731 clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
1732 spin_unlock(&bdev->fence_lock);
1733 driver->sync_obj_unref(&tmp_obj);
1734 spin_lock(&bdev->fence_lock);
1735 continue;
1736 }
1737
1738 if (no_wait)
1739 return -EBUSY;
1740
1741 sync_obj = driver->sync_obj_ref(bo->sync_obj);
1742 sync_obj_arg = bo->sync_obj_arg;
1743 spin_unlock(&bdev->fence_lock);
1744 ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
1745 lazy, interruptible);
1746 if (unlikely(ret != 0)) {
1747 driver->sync_obj_unref(&sync_obj);
1748 spin_lock(&bdev->fence_lock);
1749 return ret;
1750 }
1751 spin_lock(&bdev->fence_lock);
1752 if (likely(bo->sync_obj == sync_obj &&
1753 bo->sync_obj_arg == sync_obj_arg)) {
1754 void *tmp_obj = bo->sync_obj;
1755 bo->sync_obj = NULL;
1756 clear_bit(TTM_BO_PRIV_FLAG_MOVING,
1757 &bo->priv_flags);
1758 spin_unlock(&bdev->fence_lock);
1759 driver->sync_obj_unref(&sync_obj);
1760 driver->sync_obj_unref(&tmp_obj);
1761 spin_lock(&bdev->fence_lock);
1762 } else {
1763 spin_unlock(&bdev->fence_lock);
1764 driver->sync_obj_unref(&sync_obj);
1765 spin_lock(&bdev->fence_lock);
1766 }
1767 }
1768 return 0;
1769}
1770EXPORT_SYMBOL(ttm_bo_wait);
1771
1772int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1773{
1774 struct ttm_bo_device *bdev = bo->bdev;
1775 int ret = 0;
1776
1777 /*
1778 * Using ttm_bo_reserve makes sure the lru lists are updated.
1779 */
1780
1781 ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
1782 if (unlikely(ret != 0))
1783 return ret;
1784 spin_lock(&bdev->fence_lock);
1785 ret = ttm_bo_wait(bo, false, true, no_wait);
1786 spin_unlock(&bdev->fence_lock);
1787 if (likely(ret == 0))
1788 atomic_inc(&bo->cpu_writers);
1789 ttm_bo_unreserve(bo);
1790 return ret;
1791}
1792EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1793
1794void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1795{
1796 if (atomic_dec_and_test(&bo->cpu_writers))
1797 wake_up_all(&bo->event_queue);
1798}
1799EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1800
1801/**
1802 * A buffer object shrink method that tries to swap out the first
1803 * buffer object on the bo_global::swap_lru list.
1804 */
1805
1806static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1807{
1808 struct ttm_bo_global *glob =
1809 container_of(shrink, struct ttm_bo_global, shrink);
1810 struct ttm_buffer_object *bo;
1811 int ret = -EBUSY;
1812 int put_count;
1813 uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);
1814
1815 spin_lock(&glob->lru_lock);
1816 while (ret == -EBUSY) {
1817 if (unlikely(list_empty(&glob->swap_lru))) {
1818 spin_unlock(&glob->lru_lock);
1819 return -EBUSY;
1820 }
1821
1822 bo = list_first_entry(&glob->swap_lru,
1823 struct ttm_buffer_object, swap);
1824 kref_get(&bo->list_kref);
1825
1826 if (!list_empty(&bo->ddestroy)) {
1827 spin_unlock(&glob->lru_lock);
1828 (void) ttm_bo_cleanup_refs(bo, false, false, false);
1829 kref_put(&bo->list_kref, ttm_bo_release_list);
1830 spin_lock(&glob->lru_lock);
1831 continue;
1832 }
1833
1834 /**
1835 * Reserve buffer. Since we unlock while sleeping, we need
1836 * to re-check that nobody removed us from the swap-list while
1837 * we slept.
1838 */
1839
1840 ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
1841 if (unlikely(ret == -EBUSY)) {
1842 spin_unlock(&glob->lru_lock);
1843 ttm_bo_wait_unreserved(bo, false);
1844 kref_put(&bo->list_kref, ttm_bo_release_list);
1845 spin_lock(&glob->lru_lock);
1846 }
1847 }
1848
1849 BUG_ON(ret != 0);
1850 put_count = ttm_bo_del_from_lru(bo);
1851 spin_unlock(&glob->lru_lock);
1852
1853 ttm_bo_list_ref_sub(bo, put_count, true);
1854
1855 /**
1856 * Wait for GPU, then move to system cached.
1857 */
1858
1859 spin_lock(&bo->bdev->fence_lock);
1860 ret = ttm_bo_wait(bo, false, false, false);
1861 spin_unlock(&bo->bdev->fence_lock);
1862
1863 if (unlikely(ret != 0))
1864 goto out;
1865
1866 if ((bo->mem.placement & swap_placement) != swap_placement) {
1867 struct ttm_mem_reg evict_mem;
1868
1869 evict_mem = bo->mem;
1870 evict_mem.mm_node = NULL;
1871 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1872 evict_mem.mem_type = TTM_PL_SYSTEM;
1873
1874 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1875 false, false, false);
1876 if (unlikely(ret != 0))
1877 goto out;
1878 }
1879
1880 ttm_bo_unmap_virtual(bo);
1881
1882 /**
1883 * Swap out. Buffer will be swapped in again as soon as
1884 * anyone tries to access a ttm page.
1885 */
1886
1887 if (bo->bdev->driver->swap_notify)
1888 bo->bdev->driver->swap_notify(bo);
1889
1890 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1891out:
1892
1893 /**
1894 *
1895 * Unreserve without putting on LRU to avoid swapping out an
1896 * already swapped buffer.
1897 */
1898
1899 atomic_set(&bo->reserved, 0);
1900 wake_up_all(&bo->event_queue);
1901 kref_put(&bo->list_kref, ttm_bo_release_list);
1902 return ret;
1903}
1904
1905void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1906{
1907 while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1908 ;
1909}
1910EXPORT_SYMBOL(ttm_bo_swapout_all);