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_dbg_printer(NULL, category: DRM_UT_CORE, 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: &p, f: " placement[%d]=0x%08X (%d)\n",
59 i, placement->placement[i].flags, mem_type);
60 man = ttm_manager_type(bdev: bo->bdev, mem_type);
61 ttm_resource_manager_debug(man, p: &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(res: 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(lock: &bo->bdev->lru_lock);
106 if (bo->resource)
107 ttm_resource_del_bulk_move(res: bo->resource, bo);
108 bo->bulk_move = bulk;
109 if (bo->resource)
110 ttm_resource_add_bulk_move(res: bo->resource, bo);
111 spin_unlock(lock: &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, mem_type: bo->resource->mem_type)->use_tt;
125 new_use_tt = ttm_manager_type(bdev, mem_type: 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, zero_alloc: old_use_tt);
138 if (ret)
139 goto out_err;
140
141 if (mem->mem_type != TTM_PL_SYSTEM) {
142 ret = ttm_tt_populate(bdev: bo->bdev, ttm: bo->ttm, ctx);
143 if (ret)
144 goto out_err;
145 }
146 }
147
148 ret = dma_resv_reserve_fences(obj: bo->base.resv, num_fences: 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, res: &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(dst: &bo->base._resv, src: bo->base.resv);
196 dma_resv_unlock(obj: &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(lock: &bo->bdev->lru_lock);
206 bo->base.resv = &bo->base._resv;
207 spin_unlock(lock: &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: &cursor, obj: resv, usage: 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: &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(obj: resv, usage: 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(obj: bo->base.resv);
258 spin_unlock(lock: &bo->bdev->lru_lock);
259
260 lret = dma_resv_wait_timeout(obj: resv, usage: DMA_RESV_USAGE_BOOKKEEP,
261 intr: interruptible,
262 timeout: 30 * HZ);
263
264 if (lret < 0)
265 return lret;
266 else if (lret == 0)
267 return -EBUSY;
268
269 spin_lock(lock: &bo->bdev->lru_lock);
270 if (unlock_resv && !dma_resv_trylock(obj: 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(lock: &bo->bdev->lru_lock);
280 return 0;
281 }
282 ret = 0;
283 }
284
285 if (ret) {
286 if (unlock_resv)
287 dma_resv_unlock(obj: bo->base.resv);
288 spin_unlock(lock: &bo->bdev->lru_lock);
289 return ret;
290 }
291
292 spin_unlock(lock: &bo->bdev->lru_lock);
293 ttm_bo_cleanup_memtype_use(bo);
294
295 if (unlock_resv)
296 dma_resv_unlock(obj: 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(obj: bo->base.resv, usage: DMA_RESV_USAGE_BOOKKEEP, intr: false,
312 MAX_SCHEDULE_TIMEOUT);
313 dma_resv_lock(obj: bo->base.resv, NULL);
314 ttm_bo_cleanup_memtype_use(bo);
315 dma_resv_unlock(obj: 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(obj: bo->base.resv,
336 usage: DMA_RESV_USAGE_BOOKKEEP, intr: false,
337 timeout: 30 * HZ);
338 }
339
340 if (bo->bdev->funcs->release_notify)
341 bo->bdev->funcs->release_notify(bo);
342
343 drm_vma_offset_remove(mgr: bdev->vma_manager, node: &bo->base.vma_node);
344 ttm_mem_io_free(bdev, mem: bo->resource);
345
346 if (!dma_resv_test_signaled(obj: bo->base.resv,
347 usage: DMA_RESV_USAGE_BOOKKEEP) ||
348 (want_init_on_free() && (bo->ttm != NULL)) ||
349 !dma_resv_trylock(obj: 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(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(res: bo->resource);
367 }
368
369 kref_init(kref: &bo->kref);
370 spin_unlock(lock: &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(node: bdev->pool.nid, wq: bdev->wq, work: &bo->delayed_delete);
380 return;
381 }
382
383 ttm_bo_cleanup_memtype_use(bo);
384 dma_resv_unlock(obj: bo->base.resv);
385 }
386
387 atomic_dec(v: &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(kref: &bo->kref, release: 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 = 1;
414 hop_placement.placement = hop;
415
416 /* find space in the bounce domain */
417 ret = ttm_bo_mem_space(bo, placement: &hop_placement, mem: &hop_mem, ctx);
418 if (ret)
419 return ret;
420 /* move to the bounce domain */
421 ret = ttm_bo_handle_move_mem(bo, mem: hop_mem, evict: false, ctx, NULL);
422 if (ret) {
423 ttm_resource_free(bo, res: &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 bdev->funcs->evict_flags(bo, &placement);
444
445 if (!placement.num_placement) {
446 ret = ttm_bo_wait_ctx(bo, ctx);
447 if (ret)
448 return ret;
449
450 /*
451 * Since we've already synced, this frees backing store
452 * immediately.
453 */
454 return ttm_bo_pipeline_gutting(bo);
455 }
456
457 ret = ttm_bo_mem_space(bo, placement: &placement, mem: &evict_mem, ctx);
458 if (ret) {
459 if (ret != -ERESTARTSYS) {
460 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
461 bo);
462 ttm_bo_mem_space_debug(bo, placement: &placement);
463 }
464 goto out;
465 }
466
467 do {
468 ret = ttm_bo_handle_move_mem(bo, mem: evict_mem, evict: true, ctx, hop: &hop);
469 if (ret != -EMULTIHOP)
470 break;
471
472 ret = ttm_bo_bounce_temp_buffer(bo, mem: &evict_mem, ctx, hop: &hop);
473 } while (!ret);
474
475 if (ret) {
476 ttm_resource_free(bo, res: &evict_mem);
477 if (ret != -ERESTARTSYS && ret != -EINTR)
478 pr_err("Buffer eviction failed\n");
479 }
480out:
481 return ret;
482}
483
484/**
485 * ttm_bo_eviction_valuable
486 *
487 * @bo: The buffer object to evict
488 * @place: the placement we need to make room for
489 *
490 * Check if it is valuable to evict the BO to make room for the given placement.
491 */
492bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
493 const struct ttm_place *place)
494{
495 struct ttm_resource *res = bo->resource;
496 struct ttm_device *bdev = bo->bdev;
497
498 dma_resv_assert_held(bo->base.resv);
499 if (bo->resource->mem_type == TTM_PL_SYSTEM)
500 return true;
501
502 /* Don't evict this BO if it's outside of the
503 * requested placement range
504 */
505 return ttm_resource_intersects(bdev, res, place, size: bo->base.size);
506}
507EXPORT_SYMBOL(ttm_bo_eviction_valuable);
508
509/*
510 * Check the target bo is allowable to be evicted or swapout, including cases:
511 *
512 * a. if share same reservation object with ctx->resv, have assumption
513 * reservation objects should already be locked, so not lock again and
514 * return true directly when either the opreation allow_reserved_eviction
515 * or the target bo already is in delayed free list;
516 *
517 * b. Otherwise, trylock it.
518 */
519static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
520 struct ttm_operation_ctx *ctx,
521 const struct ttm_place *place,
522 bool *locked, bool *busy)
523{
524 bool ret = false;
525
526 if (bo->pin_count) {
527 *locked = false;
528 if (busy)
529 *busy = false;
530 return false;
531 }
532
533 if (bo->base.resv == ctx->resv) {
534 dma_resv_assert_held(bo->base.resv);
535 if (ctx->allow_res_evict)
536 ret = true;
537 *locked = false;
538 if (busy)
539 *busy = false;
540 } else {
541 ret = dma_resv_trylock(obj: bo->base.resv);
542 *locked = ret;
543 if (busy)
544 *busy = !ret;
545 }
546
547 if (ret && place && (bo->resource->mem_type != place->mem_type ||
548 !bo->bdev->funcs->eviction_valuable(bo, place))) {
549 ret = false;
550 if (*locked) {
551 dma_resv_unlock(obj: bo->base.resv);
552 *locked = false;
553 }
554 }
555
556 return ret;
557}
558
559/**
560 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
561 *
562 * @busy_bo: BO which couldn't be locked with trylock
563 * @ctx: operation context
564 * @ticket: acquire ticket
565 *
566 * Try to lock a busy buffer object to avoid failing eviction.
567 */
568static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
569 struct ttm_operation_ctx *ctx,
570 struct ww_acquire_ctx *ticket)
571{
572 int r;
573
574 if (!busy_bo || !ticket)
575 return -EBUSY;
576
577 if (ctx->interruptible)
578 r = dma_resv_lock_interruptible(obj: busy_bo->base.resv,
579 ctx: ticket);
580 else
581 r = dma_resv_lock(obj: busy_bo->base.resv, ctx: ticket);
582
583 /*
584 * TODO: It would be better to keep the BO locked until allocation is at
585 * least tried one more time, but that would mean a much larger rework
586 * of TTM.
587 */
588 if (!r)
589 dma_resv_unlock(obj: busy_bo->base.resv);
590
591 return r == -EDEADLK ? -EBUSY : r;
592}
593
594int ttm_mem_evict_first(struct ttm_device *bdev,
595 struct ttm_resource_manager *man,
596 const struct ttm_place *place,
597 struct ttm_operation_ctx *ctx,
598 struct ww_acquire_ctx *ticket)
599{
600 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
601 struct ttm_resource_cursor cursor;
602 struct ttm_resource *res;
603 bool locked = false;
604 int ret;
605
606 spin_lock(lock: &bdev->lru_lock);
607 ttm_resource_manager_for_each_res(man, &cursor, res) {
608 bool busy;
609
610 if (!ttm_bo_evict_swapout_allowable(bo: res->bo, ctx, place,
611 locked: &locked, busy: &busy)) {
612 if (busy && !busy_bo && ticket !=
613 dma_resv_locking_ctx(obj: res->bo->base.resv))
614 busy_bo = res->bo;
615 continue;
616 }
617
618 if (ttm_bo_get_unless_zero(bo: res->bo)) {
619 bo = res->bo;
620 break;
621 }
622 if (locked)
623 dma_resv_unlock(obj: res->bo->base.resv);
624 }
625
626 if (!bo) {
627 if (busy_bo && !ttm_bo_get_unless_zero(bo: busy_bo))
628 busy_bo = NULL;
629 spin_unlock(lock: &bdev->lru_lock);
630 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
631 if (busy_bo)
632 ttm_bo_put(busy_bo);
633 return ret;
634 }
635
636 if (bo->deleted) {
637 ret = ttm_bo_cleanup_refs(bo, interruptible: ctx->interruptible,
638 no_wait_gpu: ctx->no_wait_gpu, unlock_resv: locked);
639 ttm_bo_put(bo);
640 return ret;
641 }
642
643 spin_unlock(lock: &bdev->lru_lock);
644
645 ret = ttm_bo_evict(bo, ctx);
646 if (locked)
647 ttm_bo_unreserve(bo);
648 else
649 ttm_bo_move_to_lru_tail_unlocked(bo);
650
651 ttm_bo_put(bo);
652 return ret;
653}
654
655/**
656 * ttm_bo_pin - Pin the buffer object.
657 * @bo: The buffer object to pin
658 *
659 * Make sure the buffer is not evicted any more during memory pressure.
660 * @bo must be unpinned again by calling ttm_bo_unpin().
661 */
662void ttm_bo_pin(struct ttm_buffer_object *bo)
663{
664 dma_resv_assert_held(bo->base.resv);
665 WARN_ON_ONCE(!kref_read(&bo->kref));
666 spin_lock(lock: &bo->bdev->lru_lock);
667 if (bo->resource)
668 ttm_resource_del_bulk_move(res: bo->resource, bo);
669 ++bo->pin_count;
670 spin_unlock(lock: &bo->bdev->lru_lock);
671}
672EXPORT_SYMBOL(ttm_bo_pin);
673
674/**
675 * ttm_bo_unpin - Unpin the buffer object.
676 * @bo: The buffer object to unpin
677 *
678 * Allows the buffer object to be evicted again during memory pressure.
679 */
680void ttm_bo_unpin(struct ttm_buffer_object *bo)
681{
682 dma_resv_assert_held(bo->base.resv);
683 WARN_ON_ONCE(!kref_read(&bo->kref));
684 if (WARN_ON_ONCE(!bo->pin_count))
685 return;
686
687 spin_lock(lock: &bo->bdev->lru_lock);
688 --bo->pin_count;
689 if (bo->resource)
690 ttm_resource_add_bulk_move(res: bo->resource, bo);
691 spin_unlock(lock: &bo->bdev->lru_lock);
692}
693EXPORT_SYMBOL(ttm_bo_unpin);
694
695/*
696 * Add the last move fence to the BO as kernel dependency and reserve a new
697 * fence slot.
698 */
699static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
700 struct ttm_resource_manager *man,
701 struct ttm_resource *mem,
702 bool no_wait_gpu)
703{
704 struct dma_fence *fence;
705 int ret;
706
707 spin_lock(lock: &man->move_lock);
708 fence = dma_fence_get(fence: man->move);
709 spin_unlock(lock: &man->move_lock);
710
711 if (!fence)
712 return 0;
713
714 if (no_wait_gpu) {
715 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
716 dma_fence_put(fence);
717 return ret;
718 }
719
720 dma_resv_add_fence(obj: bo->base.resv, fence, usage: DMA_RESV_USAGE_KERNEL);
721
722 ret = dma_resv_reserve_fences(obj: bo->base.resv, num_fences: 1);
723 dma_fence_put(fence);
724 return ret;
725}
726
727/*
728 * Repeatedly evict memory from the LRU for @mem_type until we create enough
729 * space, or we've evicted everything and there isn't enough space.
730 */
731static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
732 const struct ttm_place *place,
733 struct ttm_resource **mem,
734 struct ttm_operation_ctx *ctx)
735{
736 struct ttm_device *bdev = bo->bdev;
737 struct ttm_resource_manager *man;
738 struct ww_acquire_ctx *ticket;
739 int ret;
740
741 man = ttm_manager_type(bdev, mem_type: place->mem_type);
742 ticket = dma_resv_locking_ctx(obj: bo->base.resv);
743 do {
744 ret = ttm_resource_alloc(bo, place, res: mem);
745 if (likely(!ret))
746 break;
747 if (unlikely(ret != -ENOSPC))
748 return ret;
749 ret = ttm_mem_evict_first(bdev, man, place, ctx,
750 ticket);
751 if (unlikely(ret != 0))
752 return ret;
753 } while (1);
754
755 return ttm_bo_add_move_fence(bo, man, mem: *mem, no_wait_gpu: ctx->no_wait_gpu);
756}
757
758/**
759 * ttm_bo_mem_space
760 *
761 * @bo: Pointer to a struct ttm_buffer_object. the data of which
762 * we want to allocate space for.
763 * @placement: Proposed new placement for the buffer object.
764 * @mem: A struct ttm_resource.
765 * @ctx: if and how to sleep, lock buffers and alloc memory
766 *
767 * Allocate memory space for the buffer object pointed to by @bo, using
768 * the placement flags in @placement, potentially evicting other idle buffer objects.
769 * This function may sleep while waiting for space to become available.
770 * Returns:
771 * -EBUSY: No space available (only if no_wait == 1).
772 * -ENOSPC: Could not allocate space for the buffer object, either due to
773 * fragmentation or concurrent allocators.
774 * -ERESTARTSYS: An interruptible sleep was interrupted by a signal.
775 */
776int ttm_bo_mem_space(struct ttm_buffer_object *bo,
777 struct ttm_placement *placement,
778 struct ttm_resource **mem,
779 struct ttm_operation_ctx *ctx)
780{
781 struct ttm_device *bdev = bo->bdev;
782 bool type_found = false;
783 int i, ret;
784
785 ret = dma_resv_reserve_fences(obj: bo->base.resv, num_fences: 1);
786 if (unlikely(ret))
787 return ret;
788
789 for (i = 0; i < placement->num_placement; ++i) {
790 const struct ttm_place *place = &placement->placement[i];
791 struct ttm_resource_manager *man;
792
793 if (place->flags & TTM_PL_FLAG_FALLBACK)
794 continue;
795
796 man = ttm_manager_type(bdev, mem_type: place->mem_type);
797 if (!man || !ttm_resource_manager_used(man))
798 continue;
799
800 type_found = true;
801 ret = ttm_resource_alloc(bo, place, res: mem);
802 if (ret == -ENOSPC)
803 continue;
804 if (unlikely(ret))
805 goto error;
806
807 ret = ttm_bo_add_move_fence(bo, man, mem: *mem, no_wait_gpu: ctx->no_wait_gpu);
808 if (unlikely(ret)) {
809 ttm_resource_free(bo, res: mem);
810 if (ret == -EBUSY)
811 continue;
812
813 goto error;
814 }
815 return 0;
816 }
817
818 for (i = 0; i < placement->num_placement; ++i) {
819 const struct ttm_place *place = &placement->placement[i];
820 struct ttm_resource_manager *man;
821
822 if (place->flags & TTM_PL_FLAG_DESIRED)
823 continue;
824
825 man = ttm_manager_type(bdev, mem_type: place->mem_type);
826 if (!man || !ttm_resource_manager_used(man))
827 continue;
828
829 type_found = true;
830 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
831 if (likely(!ret))
832 return 0;
833
834 if (ret && ret != -EBUSY)
835 goto error;
836 }
837
838 ret = -ENOSPC;
839 if (!type_found) {
840 pr_err(TTM_PFX "No compatible memory type found\n");
841 ret = -EINVAL;
842 }
843
844error:
845 return ret;
846}
847EXPORT_SYMBOL(ttm_bo_mem_space);
848
849static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
850 struct ttm_placement *placement,
851 struct ttm_operation_ctx *ctx)
852{
853 struct ttm_resource *mem;
854 struct ttm_place hop;
855 int ret;
856
857 dma_resv_assert_held(bo->base.resv);
858
859 /*
860 * Determine where to move the buffer.
861 *
862 * If driver determines move is going to need
863 * an extra step then it will return -EMULTIHOP
864 * and the buffer will be moved to the temporary
865 * stop and the driver will be called to make
866 * the second hop.
867 */
868 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
869 if (ret)
870 return ret;
871bounce:
872 ret = ttm_bo_handle_move_mem(bo, mem, evict: false, ctx, hop: &hop);
873 if (ret == -EMULTIHOP) {
874 ret = ttm_bo_bounce_temp_buffer(bo, mem: &mem, ctx, hop: &hop);
875 if (ret)
876 goto out;
877 /* try and move to final place now. */
878 goto bounce;
879 }
880out:
881 if (ret)
882 ttm_resource_free(bo, res: &mem);
883 return ret;
884}
885
886/**
887 * ttm_bo_validate
888 *
889 * @bo: The buffer object.
890 * @placement: Proposed placement for the buffer object.
891 * @ctx: validation parameters.
892 *
893 * Changes placement and caching policy of the buffer object
894 * according proposed placement.
895 * Returns
896 * -EINVAL on invalid proposed placement.
897 * -ENOMEM on out-of-memory condition.
898 * -EBUSY if no_wait is true and buffer busy.
899 * -ERESTARTSYS if interrupted by a signal.
900 */
901int ttm_bo_validate(struct ttm_buffer_object *bo,
902 struct ttm_placement *placement,
903 struct ttm_operation_ctx *ctx)
904{
905 int ret;
906
907 dma_resv_assert_held(bo->base.resv);
908
909 /*
910 * Remove the backing store if no placement is given.
911 */
912 if (!placement->num_placement)
913 return ttm_bo_pipeline_gutting(bo);
914
915 /* Check whether we need to move buffer. */
916 if (bo->resource && ttm_resource_compatible(res: bo->resource, placement))
917 return 0;
918
919 /* Moving of pinned BOs is forbidden */
920 if (bo->pin_count)
921 return -EINVAL;
922
923 ret = ttm_bo_move_buffer(bo, placement, ctx);
924 /* For backward compatibility with userspace */
925 if (ret == -ENOSPC)
926 return -ENOMEM;
927 if (ret)
928 return ret;
929
930 /*
931 * We might need to add a TTM.
932 */
933 if (!bo->resource || bo->resource->mem_type == TTM_PL_SYSTEM) {
934 ret = ttm_tt_create(bo, zero_alloc: true);
935 if (ret)
936 return ret;
937 }
938 return 0;
939}
940EXPORT_SYMBOL(ttm_bo_validate);
941
942/**
943 * ttm_bo_init_reserved
944 *
945 * @bdev: Pointer to a ttm_device struct.
946 * @bo: Pointer to a ttm_buffer_object to be initialized.
947 * @type: Requested type of buffer object.
948 * @placement: Initial placement for buffer object.
949 * @alignment: Data alignment in pages.
950 * @ctx: TTM operation context for memory allocation.
951 * @sg: Scatter-gather table.
952 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
953 * @destroy: Destroy function. Use NULL for kfree().
954 *
955 * This function initializes a pre-allocated struct ttm_buffer_object.
956 * As this object may be part of a larger structure, this function,
957 * together with the @destroy function, enables driver-specific objects
958 * derived from a ttm_buffer_object.
959 *
960 * On successful return, the caller owns an object kref to @bo. The kref and
961 * list_kref are usually set to 1, but note that in some situations, other
962 * tasks may already be holding references to @bo as well.
963 * Furthermore, if resv == NULL, the buffer's reservation lock will be held,
964 * and it is the caller's responsibility to call ttm_bo_unreserve.
965 *
966 * If a failure occurs, the function will call the @destroy function. Thus,
967 * after a failure, dereferencing @bo is illegal and will likely cause memory
968 * corruption.
969 *
970 * Returns
971 * -ENOMEM: Out of memory.
972 * -EINVAL: Invalid placement flags.
973 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
974 */
975int ttm_bo_init_reserved(struct ttm_device *bdev, struct ttm_buffer_object *bo,
976 enum ttm_bo_type type, struct ttm_placement *placement,
977 uint32_t alignment, struct ttm_operation_ctx *ctx,
978 struct sg_table *sg, struct dma_resv *resv,
979 void (*destroy) (struct ttm_buffer_object *))
980{
981 int ret;
982
983 kref_init(kref: &bo->kref);
984 bo->bdev = bdev;
985 bo->type = type;
986 bo->page_alignment = alignment;
987 bo->destroy = destroy;
988 bo->pin_count = 0;
989 bo->sg = sg;
990 bo->bulk_move = NULL;
991 if (resv)
992 bo->base.resv = resv;
993 else
994 bo->base.resv = &bo->base._resv;
995 atomic_inc(v: &ttm_glob.bo_count);
996
997 /*
998 * For ttm_bo_type_device buffers, allocate
999 * address space from the device.
1000 */
1001 if (bo->type == ttm_bo_type_device || bo->type == ttm_bo_type_sg) {
1002 ret = drm_vma_offset_add(mgr: bdev->vma_manager, node: &bo->base.vma_node,
1003 PFN_UP(bo->base.size));
1004 if (ret)
1005 goto err_put;
1006 }
1007
1008 /* passed reservation objects should already be locked,
1009 * since otherwise lockdep will be angered in radeon.
1010 */
1011 if (!resv)
1012 WARN_ON(!dma_resv_trylock(bo->base.resv));
1013 else
1014 dma_resv_assert_held(resv);
1015
1016 ret = ttm_bo_validate(bo, placement, ctx);
1017 if (unlikely(ret))
1018 goto err_unlock;
1019
1020 return 0;
1021
1022err_unlock:
1023 if (!resv)
1024 dma_resv_unlock(obj: bo->base.resv);
1025
1026err_put:
1027 ttm_bo_put(bo);
1028 return ret;
1029}
1030EXPORT_SYMBOL(ttm_bo_init_reserved);
1031
1032/**
1033 * ttm_bo_init_validate
1034 *
1035 * @bdev: Pointer to a ttm_device struct.
1036 * @bo: Pointer to a ttm_buffer_object to be initialized.
1037 * @type: Requested type of buffer object.
1038 * @placement: Initial placement for buffer object.
1039 * @alignment: Data alignment in pages.
1040 * @interruptible: If needing to sleep to wait for GPU resources,
1041 * sleep interruptible.
1042 * pinned in physical memory. If this behaviour is not desired, this member
1043 * holds a pointer to a persistent shmem object. Typically, this would
1044 * point to the shmem object backing a GEM object if TTM is used to back a
1045 * GEM user interface.
1046 * @sg: Scatter-gather table.
1047 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
1048 * @destroy: Destroy function. Use NULL for kfree().
1049 *
1050 * This function initializes a pre-allocated struct ttm_buffer_object.
1051 * As this object may be part of a larger structure, this function,
1052 * together with the @destroy function,
1053 * enables driver-specific objects derived from a ttm_buffer_object.
1054 *
1055 * On successful return, the caller owns an object kref to @bo. The kref and
1056 * list_kref are usually set to 1, but note that in some situations, other
1057 * tasks may already be holding references to @bo as well.
1058 *
1059 * If a failure occurs, the function will call the @destroy function, Thus,
1060 * after a failure, dereferencing @bo is illegal and will likely cause memory
1061 * corruption.
1062 *
1063 * Returns
1064 * -ENOMEM: Out of memory.
1065 * -EINVAL: Invalid placement flags.
1066 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
1067 */
1068int ttm_bo_init_validate(struct ttm_device *bdev, struct ttm_buffer_object *bo,
1069 enum ttm_bo_type type, struct ttm_placement *placement,
1070 uint32_t alignment, bool interruptible,
1071 struct sg_table *sg, struct dma_resv *resv,
1072 void (*destroy) (struct ttm_buffer_object *))
1073{
1074 struct ttm_operation_ctx ctx = { interruptible, false };
1075 int ret;
1076
1077 ret = ttm_bo_init_reserved(bdev, bo, type, placement, alignment, &ctx,
1078 sg, resv, destroy);
1079 if (ret)
1080 return ret;
1081
1082 if (!resv)
1083 ttm_bo_unreserve(bo);
1084
1085 return 0;
1086}
1087EXPORT_SYMBOL(ttm_bo_init_validate);
1088
1089/*
1090 * buffer object vm functions.
1091 */
1092
1093/**
1094 * ttm_bo_unmap_virtual
1095 *
1096 * @bo: tear down the virtual mappings for this BO
1097 */
1098void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1099{
1100 struct ttm_device *bdev = bo->bdev;
1101
1102 drm_vma_node_unmap(node: &bo->base.vma_node, file_mapping: bdev->dev_mapping);
1103 ttm_mem_io_free(bdev, mem: bo->resource);
1104}
1105EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1106
1107/**
1108 * ttm_bo_wait_ctx - wait for buffer idle.
1109 *
1110 * @bo: The buffer object.
1111 * @ctx: defines how to wait
1112 *
1113 * Waits for the buffer to be idle. Used timeout depends on the context.
1114 * Returns -EBUSY if wait timed outt, -ERESTARTSYS if interrupted by a signal or
1115 * zero on success.
1116 */
1117int ttm_bo_wait_ctx(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx)
1118{
1119 long ret;
1120
1121 if (ctx->no_wait_gpu) {
1122 if (dma_resv_test_signaled(obj: bo->base.resv,
1123 usage: DMA_RESV_USAGE_BOOKKEEP))
1124 return 0;
1125 else
1126 return -EBUSY;
1127 }
1128
1129 ret = dma_resv_wait_timeout(obj: bo->base.resv, usage: DMA_RESV_USAGE_BOOKKEEP,
1130 intr: ctx->interruptible, timeout: 15 * HZ);
1131 if (unlikely(ret < 0))
1132 return ret;
1133 if (unlikely(ret == 0))
1134 return -EBUSY;
1135 return 0;
1136}
1137EXPORT_SYMBOL(ttm_bo_wait_ctx);
1138
1139int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1140 gfp_t gfp_flags)
1141{
1142 struct ttm_place place;
1143 bool locked;
1144 long ret;
1145
1146 /*
1147 * While the bo may already reside in SYSTEM placement, set
1148 * SYSTEM as new placement to cover also the move further below.
1149 * The driver may use the fact that we're moving from SYSTEM
1150 * as an indication that we're about to swap out.
1151 */
1152 memset(&place, 0, sizeof(place));
1153 place.mem_type = bo->resource->mem_type;
1154 if (!ttm_bo_evict_swapout_allowable(bo, ctx, place: &place, locked: &locked, NULL))
1155 return -EBUSY;
1156
1157 if (!bo->ttm || !ttm_tt_is_populated(tt: bo->ttm) ||
1158 bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1159 bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED ||
1160 !ttm_bo_get_unless_zero(bo)) {
1161 if (locked)
1162 dma_resv_unlock(obj: bo->base.resv);
1163 return -EBUSY;
1164 }
1165
1166 if (bo->deleted) {
1167 ret = ttm_bo_cleanup_refs(bo, interruptible: false, no_wait_gpu: false, unlock_resv: locked);
1168 ttm_bo_put(bo);
1169 return ret == -EBUSY ? -ENOSPC : ret;
1170 }
1171
1172 /* TODO: Cleanup the locking */
1173 spin_unlock(lock: &bo->bdev->lru_lock);
1174
1175 /*
1176 * Move to system cached
1177 */
1178 if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1179 struct ttm_resource *evict_mem;
1180 struct ttm_place hop;
1181
1182 memset(&hop, 0, sizeof(hop));
1183 place.mem_type = TTM_PL_SYSTEM;
1184 ret = ttm_resource_alloc(bo, place: &place, res: &evict_mem);
1185 if (unlikely(ret))
1186 goto out;
1187
1188 ret = ttm_bo_handle_move_mem(bo, mem: evict_mem, evict: true, ctx, hop: &hop);
1189 if (unlikely(ret != 0)) {
1190 WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1191 ttm_resource_free(bo, res: &evict_mem);
1192 goto out;
1193 }
1194 }
1195
1196 /*
1197 * Make sure BO is idle.
1198 */
1199 ret = ttm_bo_wait_ctx(bo, ctx);
1200 if (unlikely(ret != 0))
1201 goto out;
1202
1203 ttm_bo_unmap_virtual(bo);
1204
1205 /*
1206 * Swap out. Buffer will be swapped in again as soon as
1207 * anyone tries to access a ttm page.
1208 */
1209 if (bo->bdev->funcs->swap_notify)
1210 bo->bdev->funcs->swap_notify(bo);
1211
1212 if (ttm_tt_is_populated(tt: bo->ttm))
1213 ret = ttm_tt_swapout(bdev: bo->bdev, ttm: bo->ttm, gfp_flags);
1214out:
1215
1216 /*
1217 * Unreserve without putting on LRU to avoid swapping out an
1218 * already swapped buffer.
1219 */
1220 if (locked)
1221 dma_resv_unlock(obj: bo->base.resv);
1222 ttm_bo_put(bo);
1223 return ret == -EBUSY ? -ENOSPC : ret;
1224}
1225
1226void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1227{
1228 if (bo->ttm == NULL)
1229 return;
1230
1231 ttm_tt_unpopulate(bdev: bo->bdev, ttm: bo->ttm);
1232 ttm_tt_destroy(bdev: bo->bdev, ttm: bo->ttm);
1233 bo->ttm = NULL;
1234}
1235

source code of linux/drivers/gpu/drm/ttm/ttm_bo.c