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/drm_print.h>
35#include <drm/ttm/ttm_allocation.h>
36#include <drm/ttm/ttm_bo.h>
37#include <drm/ttm/ttm_placement.h>
38#include <drm/ttm/ttm_tt.h>
39
40#include <linux/export.h>
41#include <linux/jiffies.h>
42#include <linux/slab.h>
43#include <linux/sched.h>
44#include <linux/mm.h>
45#include <linux/file.h>
46#include <linux/module.h>
47#include <linux/atomic.h>
48#include <linux/cgroup_dmem.h>
49#include <linux/dma-resv.h>
50
51#include "ttm_module.h"
52#include "ttm_bo_internal.h"
53
54static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
55 struct ttm_placement *placement)
56{
57 struct drm_printer p = drm_dbg_printer(NULL, category: DRM_UT_CORE, TTM_PFX);
58 struct ttm_resource_manager *man;
59 int i, mem_type;
60
61 for (i = 0; i < placement->num_placement; i++) {
62 mem_type = placement->placement[i].mem_type;
63 drm_printf(p: &p, f: " placement[%d]=0x%08X (%d)\n",
64 i, placement->placement[i].flags, mem_type);
65 man = ttm_manager_type(bdev: bo->bdev, mem_type);
66 ttm_resource_manager_debug(man, p: &p);
67 }
68}
69
70/**
71 * ttm_bo_move_to_lru_tail
72 *
73 * @bo: The buffer object.
74 *
75 * Move this BO to the tail of all lru lists used to lookup and reserve an
76 * object. This function must be called with struct ttm_global::lru_lock
77 * held, and is used to make a BO less likely to be considered for eviction.
78 */
79void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
80{
81 dma_resv_assert_held(bo->base.resv);
82
83 if (bo->resource)
84 ttm_resource_move_to_lru_tail(res: bo->resource);
85}
86EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
87
88/**
89 * ttm_bo_set_bulk_move - update BOs bulk move object
90 *
91 * @bo: The buffer object.
92 * @bulk: bulk move structure
93 *
94 * Update the BOs bulk move object, making sure that resources are added/removed
95 * as well. A bulk move allows to move many resource on the LRU at once,
96 * resulting in much less overhead of maintaining the LRU.
97 * The only requirement is that the resources stay together on the LRU and are
98 * never separated. This is enforces by setting the bulk_move structure on a BO.
99 * ttm_lru_bulk_move_tail() should be used to move all resources to the tail of
100 * their LRU list.
101 */
102void ttm_bo_set_bulk_move(struct ttm_buffer_object *bo,
103 struct ttm_lru_bulk_move *bulk)
104{
105 dma_resv_assert_held(bo->base.resv);
106
107 if (bo->bulk_move == bulk)
108 return;
109
110 spin_lock(lock: &bo->bdev->lru_lock);
111 if (bo->resource)
112 ttm_resource_del_bulk_move(res: bo->resource, bo);
113 bo->bulk_move = bulk;
114 if (bo->resource)
115 ttm_resource_add_bulk_move(res: bo->resource, bo);
116 spin_unlock(lock: &bo->bdev->lru_lock);
117}
118EXPORT_SYMBOL(ttm_bo_set_bulk_move);
119
120static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
121 struct ttm_resource *mem, bool evict,
122 struct ttm_operation_ctx *ctx,
123 struct ttm_place *hop)
124{
125 struct ttm_device *bdev = bo->bdev;
126 bool old_use_tt, new_use_tt;
127 int ret;
128
129 old_use_tt = !bo->resource || ttm_manager_type(bdev, mem_type: bo->resource->mem_type)->use_tt;
130 new_use_tt = ttm_manager_type(bdev, mem_type: mem->mem_type)->use_tt;
131
132 ttm_bo_unmap_virtual(bo);
133
134 /*
135 * Create and bind a ttm if required.
136 */
137
138 if (new_use_tt) {
139 /* Zero init the new TTM structure if the old location should
140 * have used one as well.
141 */
142 ret = ttm_tt_create(bo, zero_alloc: old_use_tt);
143 if (ret)
144 goto out_err;
145
146 if (mem->mem_type != TTM_PL_SYSTEM) {
147 ret = ttm_bo_populate(bo, ctx);
148 if (ret)
149 goto out_err;
150 }
151 }
152
153 ret = dma_resv_reserve_fences(obj: bo->base.resv, num_fences: 1);
154 if (ret)
155 goto out_err;
156
157 ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
158 if (ret) {
159 if (ret == -EMULTIHOP)
160 return ret;
161 goto out_err;
162 }
163
164 ctx->bytes_moved += bo->base.size;
165 return 0;
166
167out_err:
168 if (!old_use_tt)
169 ttm_bo_tt_destroy(bo);
170
171 return ret;
172}
173
174/*
175 * Call bo::reserved.
176 * Will release GPU memory type usage on destruction.
177 * This is the place to put in driver specific hooks to release
178 * driver private resources.
179 * Will release the bo::reserved lock.
180 */
181
182static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
183{
184 if (bo->bdev->funcs->delete_mem_notify)
185 bo->bdev->funcs->delete_mem_notify(bo);
186
187 ttm_bo_tt_destroy(bo);
188 ttm_resource_free(bo, res: &bo->resource);
189}
190
191static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
192{
193 int r;
194
195 if (bo->base.resv == &bo->base._resv)
196 return 0;
197
198 BUG_ON(!dma_resv_trylock(&bo->base._resv));
199
200 r = dma_resv_copy_fences(dst: &bo->base._resv, src: bo->base.resv);
201 dma_resv_unlock(obj: &bo->base._resv);
202 if (r)
203 return r;
204
205 if (bo->type != ttm_bo_type_sg) {
206 /* This works because the BO is about to be destroyed and nobody
207 * reference it any more. The only tricky case is the trylock on
208 * the resv object while holding the lru_lock.
209 */
210 spin_lock(lock: &bo->bdev->lru_lock);
211 bo->base.resv = &bo->base._resv;
212 spin_unlock(lock: &bo->bdev->lru_lock);
213 }
214
215 return r;
216}
217
218static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
219{
220 struct dma_resv *resv = &bo->base._resv;
221 struct dma_resv_iter cursor;
222 struct dma_fence *fence;
223
224 dma_resv_iter_begin(cursor: &cursor, obj: resv, usage: DMA_RESV_USAGE_BOOKKEEP);
225 dma_resv_for_each_fence_unlocked(&cursor, fence) {
226 if (!fence->ops->signaled)
227 dma_fence_enable_sw_signaling(fence);
228 }
229 dma_resv_iter_end(cursor: &cursor);
230}
231
232/*
233 * Block for the dma_resv object to become idle, lock the buffer and clean up
234 * the resource and tt object.
235 */
236static void ttm_bo_delayed_delete(struct work_struct *work)
237{
238 struct ttm_buffer_object *bo;
239
240 bo = container_of(work, typeof(*bo), delayed_delete);
241
242 dma_resv_wait_timeout(obj: &bo->base._resv, usage: DMA_RESV_USAGE_BOOKKEEP, intr: false,
243 MAX_SCHEDULE_TIMEOUT);
244 dma_resv_lock(obj: bo->base.resv, NULL);
245 ttm_bo_cleanup_memtype_use(bo);
246 dma_resv_unlock(obj: bo->base.resv);
247 ttm_bo_put(bo);
248}
249
250static void ttm_bo_release(struct kref *kref)
251{
252 struct ttm_buffer_object *bo =
253 container_of(kref, struct ttm_buffer_object, kref);
254 struct ttm_device *bdev = bo->bdev;
255 int ret;
256
257 WARN_ON_ONCE(bo->pin_count);
258 WARN_ON_ONCE(bo->bulk_move);
259
260 if (!bo->deleted) {
261 ret = ttm_bo_individualize_resv(bo);
262 if (ret) {
263 /* Last resort, if we fail to allocate memory for the
264 * fences block for the BO to become idle
265 */
266 dma_resv_wait_timeout(obj: bo->base.resv,
267 usage: DMA_RESV_USAGE_BOOKKEEP, intr: false,
268 timeout: 30 * HZ);
269 }
270
271 if (bo->bdev->funcs->release_notify)
272 bo->bdev->funcs->release_notify(bo);
273
274 drm_vma_offset_remove(mgr: bdev->vma_manager, node: &bo->base.vma_node);
275 ttm_mem_io_free(bdev, mem: bo->resource);
276
277 if (!dma_resv_test_signaled(obj: &bo->base._resv,
278 usage: DMA_RESV_USAGE_BOOKKEEP) ||
279 (want_init_on_free() && (bo->ttm != NULL)) ||
280 bo->type == ttm_bo_type_sg ||
281 !dma_resv_trylock(obj: bo->base.resv)) {
282 /* The BO is not idle, resurrect it for delayed destroy */
283 ttm_bo_flush_all_fences(bo);
284 bo->deleted = true;
285
286 spin_lock(lock: &bo->bdev->lru_lock);
287
288 /*
289 * Make pinned bos immediately available to
290 * shrinkers, now that they are queued for
291 * destruction.
292 *
293 * FIXME: QXL is triggering this. Can be removed when the
294 * driver is fixed.
295 */
296 if (bo->pin_count) {
297 bo->pin_count = 0;
298 ttm_resource_move_to_lru_tail(res: bo->resource);
299 }
300
301 kref_init(kref: &bo->kref);
302 spin_unlock(lock: &bo->bdev->lru_lock);
303
304 INIT_WORK(&bo->delayed_delete, ttm_bo_delayed_delete);
305
306 /* Schedule the worker on the closest NUMA node. This
307 * improves performance since system memory might be
308 * cleared on free and that is best done on a CPU core
309 * close to it.
310 */
311 queue_work_node(node: bdev->pool.nid, wq: bdev->wq, work: &bo->delayed_delete);
312 return;
313 }
314
315 ttm_bo_cleanup_memtype_use(bo);
316 dma_resv_unlock(obj: bo->base.resv);
317 }
318
319 atomic_dec(v: &ttm_glob.bo_count);
320 bo->destroy(bo);
321}
322
323/* TODO: remove! */
324void ttm_bo_put(struct ttm_buffer_object *bo)
325{
326 kref_put(kref: &bo->kref, release: ttm_bo_release);
327}
328
329void ttm_bo_fini(struct ttm_buffer_object *bo)
330{
331 ttm_bo_put(bo);
332}
333EXPORT_SYMBOL(ttm_bo_fini);
334
335static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
336 struct ttm_operation_ctx *ctx,
337 struct ttm_place *hop)
338{
339 struct ttm_placement hop_placement;
340 struct ttm_resource *hop_mem;
341 int ret;
342
343 hop_placement.num_placement = 1;
344 hop_placement.placement = hop;
345
346 /* find space in the bounce domain */
347 ret = ttm_bo_mem_space(bo, placement: &hop_placement, mem: &hop_mem, ctx);
348 if (ret)
349 return ret;
350 /* move to the bounce domain */
351 ret = ttm_bo_handle_move_mem(bo, mem: hop_mem, evict: false, ctx, NULL);
352 if (ret) {
353 ttm_resource_free(bo, res: &hop_mem);
354 return ret;
355 }
356 return 0;
357}
358
359static int ttm_bo_evict(struct ttm_buffer_object *bo,
360 struct ttm_operation_ctx *ctx)
361{
362 struct ttm_device *bdev = bo->bdev;
363 struct ttm_resource *evict_mem;
364 struct ttm_placement placement;
365 struct ttm_place hop;
366 int ret = 0;
367
368 memset(&hop, 0, sizeof(hop));
369
370 dma_resv_assert_held(bo->base.resv);
371
372 placement.num_placement = 0;
373 bdev->funcs->evict_flags(bo, &placement);
374
375 if (!placement.num_placement) {
376 ret = ttm_bo_wait_ctx(bo, ctx);
377 if (ret)
378 return ret;
379
380 /*
381 * Since we've already synced, this frees backing store
382 * immediately.
383 */
384 return ttm_bo_pipeline_gutting(bo);
385 }
386
387 ret = ttm_bo_mem_space(bo, placement: &placement, mem: &evict_mem, ctx);
388 if (ret) {
389 if (ret != -ERESTARTSYS) {
390 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
391 bo);
392 ttm_bo_mem_space_debug(bo, placement: &placement);
393 }
394 goto out;
395 }
396
397 do {
398 ret = ttm_bo_handle_move_mem(bo, mem: evict_mem, evict: true, ctx, hop: &hop);
399 if (ret != -EMULTIHOP)
400 break;
401
402 ret = ttm_bo_bounce_temp_buffer(bo, ctx, hop: &hop);
403 } while (!ret);
404
405 if (ret) {
406 ttm_resource_free(bo, res: &evict_mem);
407 if (ret != -ERESTARTSYS && ret != -EINTR)
408 pr_err("Buffer eviction failed\n");
409 }
410out:
411 return ret;
412}
413
414/**
415 * ttm_bo_eviction_valuable
416 *
417 * @bo: The buffer object to evict
418 * @place: the placement we need to make room for
419 *
420 * Check if it is valuable to evict the BO to make room for the given placement.
421 */
422bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
423 const struct ttm_place *place)
424{
425 struct ttm_resource *res = bo->resource;
426 struct ttm_device *bdev = bo->bdev;
427
428 dma_resv_assert_held(bo->base.resv);
429 if (bo->resource->mem_type == TTM_PL_SYSTEM)
430 return true;
431
432 /* Don't evict this BO if it's outside of the
433 * requested placement range
434 */
435 return ttm_resource_intersects(bdev, res, place, size: bo->base.size);
436}
437EXPORT_SYMBOL(ttm_bo_eviction_valuable);
438
439/**
440 * ttm_bo_evict_first() - Evict the first bo on the manager's LRU list.
441 * @bdev: The ttm device.
442 * @man: The manager whose bo to evict.
443 * @ctx: The TTM operation ctx governing the eviction.
444 *
445 * Return: 0 if successful or the resource disappeared. Negative error code on error.
446 */
447int ttm_bo_evict_first(struct ttm_device *bdev, struct ttm_resource_manager *man,
448 struct ttm_operation_ctx *ctx)
449{
450 struct ttm_resource_cursor cursor;
451 struct ttm_buffer_object *bo;
452 struct ttm_resource *res;
453 unsigned int mem_type;
454 int ret = 0;
455
456 spin_lock(lock: &bdev->lru_lock);
457 ttm_resource_cursor_init(cursor: &cursor, man);
458 res = ttm_resource_manager_first(cursor: &cursor);
459 ttm_resource_cursor_fini(cursor: &cursor);
460 if (!res) {
461 ret = -ENOENT;
462 goto out_no_ref;
463 }
464 bo = res->bo;
465 if (!ttm_bo_get_unless_zero(bo))
466 goto out_no_ref;
467 mem_type = res->mem_type;
468 spin_unlock(lock: &bdev->lru_lock);
469 ret = ttm_bo_reserve(bo, interruptible: ctx->interruptible, no_wait: ctx->no_wait_gpu, NULL);
470 if (ret)
471 goto out_no_lock;
472 if (!bo->resource || bo->resource->mem_type != mem_type)
473 goto out_bo_moved;
474
475 if (bo->deleted) {
476 ret = ttm_bo_wait_ctx(bo, ctx);
477 if (!ret)
478 ttm_bo_cleanup_memtype_use(bo);
479 } else {
480 ret = ttm_bo_evict(bo, ctx);
481 }
482out_bo_moved:
483 dma_resv_unlock(obj: bo->base.resv);
484out_no_lock:
485 ttm_bo_put(bo);
486 return ret;
487
488out_no_ref:
489 spin_unlock(lock: &bdev->lru_lock);
490 return ret;
491}
492
493/**
494 * struct ttm_bo_evict_walk - Parameters for the evict walk.
495 */
496struct ttm_bo_evict_walk {
497 /** @walk: The walk base parameters. */
498 struct ttm_lru_walk walk;
499 /** @place: The place passed to the resource allocation. */
500 const struct ttm_place *place;
501 /** @evictor: The buffer object we're trying to make room for. */
502 struct ttm_buffer_object *evictor;
503 /** @res: The allocated resource if any. */
504 struct ttm_resource **res;
505 /** @evicted: Number of successful evictions. */
506 unsigned long evicted;
507
508 /** @limit_pool: Which pool limit we should test against */
509 struct dmem_cgroup_pool_state *limit_pool;
510 /** @try_low: Whether we should attempt to evict BO's with low watermark threshold */
511 bool try_low;
512 /** @hit_low: If we cannot evict a bo when @try_low is false (first pass) */
513 bool hit_low;
514};
515
516static s64 ttm_bo_evict_cb(struct ttm_lru_walk *walk, struct ttm_buffer_object *bo)
517{
518 struct ttm_bo_evict_walk *evict_walk =
519 container_of(walk, typeof(*evict_walk), walk);
520 s64 lret;
521
522 if (!dmem_cgroup_state_evict_valuable(limit_pool: evict_walk->limit_pool, test_pool: bo->resource->css,
523 ignore_low: evict_walk->try_low, ret_hit_low: &evict_walk->hit_low))
524 return 0;
525
526 if (bo->pin_count || !bo->bdev->funcs->eviction_valuable(bo, evict_walk->place))
527 return 0;
528
529 if (bo->deleted) {
530 lret = ttm_bo_wait_ctx(bo, ctx: walk->arg.ctx);
531 if (!lret)
532 ttm_bo_cleanup_memtype_use(bo);
533 } else {
534 lret = ttm_bo_evict(bo, ctx: walk->arg.ctx);
535 }
536
537 if (lret)
538 goto out;
539
540 evict_walk->evicted++;
541 if (evict_walk->res)
542 lret = ttm_resource_alloc(bo: evict_walk->evictor, place: evict_walk->place,
543 res: evict_walk->res, NULL);
544 if (lret == 0)
545 return 1;
546out:
547 /* Errors that should terminate the walk. */
548 if (lret == -ENOSPC)
549 return -EBUSY;
550
551 return lret;
552}
553
554static const struct ttm_lru_walk_ops ttm_evict_walk_ops = {
555 .process_bo = ttm_bo_evict_cb,
556};
557
558static int ttm_bo_evict_alloc(struct ttm_device *bdev,
559 struct ttm_resource_manager *man,
560 const struct ttm_place *place,
561 struct ttm_buffer_object *evictor,
562 struct ttm_operation_ctx *ctx,
563 struct ww_acquire_ctx *ticket,
564 struct ttm_resource **res,
565 struct dmem_cgroup_pool_state *limit_pool)
566{
567 struct ttm_bo_evict_walk evict_walk = {
568 .walk = {
569 .ops = &ttm_evict_walk_ops,
570 .arg = {
571 .ctx = ctx,
572 .ticket = ticket,
573 }
574 },
575 .place = place,
576 .evictor = evictor,
577 .res = res,
578 .limit_pool = limit_pool,
579 };
580 s64 lret;
581
582 evict_walk.walk.arg.trylock_only = true;
583 lret = ttm_lru_walk_for_evict(walk: &evict_walk.walk, bdev, man, target: 1);
584
585 /* One more attempt if we hit low limit? */
586 if (!lret && evict_walk.hit_low) {
587 evict_walk.try_low = true;
588 lret = ttm_lru_walk_for_evict(walk: &evict_walk.walk, bdev, man, target: 1);
589 }
590 if (lret || !ticket)
591 goto out;
592
593 /* Reset low limit */
594 evict_walk.try_low = evict_walk.hit_low = false;
595 /* If ticket-locking, repeat while making progress. */
596 evict_walk.walk.arg.trylock_only = false;
597
598retry:
599 do {
600 /* The walk may clear the evict_walk.walk.ticket field */
601 evict_walk.walk.arg.ticket = ticket;
602 evict_walk.evicted = 0;
603 lret = ttm_lru_walk_for_evict(walk: &evict_walk.walk, bdev, man, target: 1);
604 } while (!lret && evict_walk.evicted);
605
606 /* We hit the low limit? Try once more */
607 if (!lret && evict_walk.hit_low && !evict_walk.try_low) {
608 evict_walk.try_low = true;
609 goto retry;
610 }
611out:
612 if (lret < 0)
613 return lret;
614 if (lret == 0)
615 return -EBUSY;
616 return 0;
617}
618
619/**
620 * ttm_bo_pin - Pin the buffer object.
621 * @bo: The buffer object to pin
622 *
623 * Make sure the buffer is not evicted any more during memory pressure.
624 * @bo must be unpinned again by calling ttm_bo_unpin().
625 */
626void ttm_bo_pin(struct ttm_buffer_object *bo)
627{
628 dma_resv_assert_held(bo->base.resv);
629 WARN_ON_ONCE(!kref_read(&bo->kref));
630 spin_lock(lock: &bo->bdev->lru_lock);
631 if (bo->resource)
632 ttm_resource_del_bulk_move(res: bo->resource, bo);
633 if (!bo->pin_count++ && bo->resource)
634 ttm_resource_move_to_lru_tail(res: bo->resource);
635 spin_unlock(lock: &bo->bdev->lru_lock);
636}
637EXPORT_SYMBOL(ttm_bo_pin);
638
639/**
640 * ttm_bo_unpin - Unpin the buffer object.
641 * @bo: The buffer object to unpin
642 *
643 * Allows the buffer object to be evicted again during memory pressure.
644 */
645void ttm_bo_unpin(struct ttm_buffer_object *bo)
646{
647 dma_resv_assert_held(bo->base.resv);
648 WARN_ON_ONCE(!kref_read(&bo->kref));
649 if (WARN_ON_ONCE(!bo->pin_count))
650 return;
651
652 spin_lock(lock: &bo->bdev->lru_lock);
653 if (!--bo->pin_count && bo->resource) {
654 ttm_resource_add_bulk_move(res: bo->resource, bo);
655 ttm_resource_move_to_lru_tail(res: bo->resource);
656 }
657 spin_unlock(lock: &bo->bdev->lru_lock);
658}
659EXPORT_SYMBOL(ttm_bo_unpin);
660
661/*
662 * Add the pipelined eviction fencesto the BO as kernel dependency and reserve new
663 * fence slots.
664 */
665static int ttm_bo_add_pipelined_eviction_fences(struct ttm_buffer_object *bo,
666 struct ttm_resource_manager *man,
667 bool no_wait_gpu)
668{
669 struct dma_fence *fence;
670 int i;
671
672 spin_lock(lock: &man->eviction_lock);
673 for (i = 0; i < TTM_NUM_MOVE_FENCES; i++) {
674 fence = man->eviction_fences[i];
675 if (!fence)
676 continue;
677
678 if (no_wait_gpu) {
679 if (!dma_fence_is_signaled(fence)) {
680 spin_unlock(lock: &man->eviction_lock);
681 return -EBUSY;
682 }
683 } else {
684 dma_resv_add_fence(obj: bo->base.resv, fence, usage: DMA_RESV_USAGE_KERNEL);
685 }
686 }
687 spin_unlock(lock: &man->eviction_lock);
688
689 /* TODO: this call should be removed. */
690 return dma_resv_reserve_fences(obj: bo->base.resv, num_fences: 1);
691}
692
693/**
694 * ttm_bo_alloc_resource - Allocate backing store for a BO
695 *
696 * @bo: Pointer to a struct ttm_buffer_object of which we want a resource for
697 * @placement: Proposed new placement for the buffer object
698 * @ctx: if and how to sleep, lock buffers and alloc memory
699 * @force_space: If we should evict buffers to force space
700 * @res: The resulting struct ttm_resource.
701 *
702 * Allocates a resource for the buffer object pointed to by @bo, using the
703 * placement flags in @placement, potentially evicting other buffer objects when
704 * @force_space is true.
705 * This function may sleep while waiting for resources to become available.
706 * Returns:
707 * -EBUSY: No space available (only if no_wait == true).
708 * -ENOSPC: Could not allocate space for the buffer object, either due to
709 * fragmentation or concurrent allocators.
710 * -ERESTARTSYS: An interruptible sleep was interrupted by a signal.
711 */
712static int ttm_bo_alloc_resource(struct ttm_buffer_object *bo,
713 struct ttm_placement *placement,
714 struct ttm_operation_ctx *ctx,
715 bool force_space,
716 struct ttm_resource **res)
717{
718 struct ttm_device *bdev = bo->bdev;
719 struct ww_acquire_ctx *ticket;
720 int i, ret;
721
722 ticket = dma_resv_locking_ctx(obj: bo->base.resv);
723 ret = dma_resv_reserve_fences(obj: bo->base.resv, TTM_NUM_MOVE_FENCES);
724 if (unlikely(ret))
725 return ret;
726
727 for (i = 0; i < placement->num_placement; ++i) {
728 const struct ttm_place *place = &placement->placement[i];
729 struct dmem_cgroup_pool_state *limit_pool = NULL;
730 struct ttm_resource_manager *man;
731 bool may_evict;
732
733 man = ttm_manager_type(bdev, mem_type: place->mem_type);
734 if (!man || !ttm_resource_manager_used(man))
735 continue;
736
737 if (place->flags & (force_space ? TTM_PL_FLAG_DESIRED :
738 TTM_PL_FLAG_FALLBACK))
739 continue;
740
741 may_evict = (force_space && place->mem_type != TTM_PL_SYSTEM);
742 ret = ttm_resource_alloc(bo, place, res, ret_limit_pool: force_space ? &limit_pool : NULL);
743 if (ret) {
744 if (ret != -ENOSPC && ret != -EAGAIN) {
745 dmem_cgroup_pool_state_put(pool: limit_pool);
746 return ret;
747 }
748 if (!may_evict) {
749 dmem_cgroup_pool_state_put(pool: limit_pool);
750 continue;
751 }
752
753 ret = ttm_bo_evict_alloc(bdev, man, place, evictor: bo, ctx,
754 ticket, res, limit_pool);
755 dmem_cgroup_pool_state_put(pool: limit_pool);
756 if (ret == -EBUSY)
757 continue;
758 if (ret)
759 return ret;
760 }
761
762 ret = ttm_bo_add_pipelined_eviction_fences(bo, man, no_wait_gpu: ctx->no_wait_gpu);
763 if (unlikely(ret)) {
764 ttm_resource_free(bo, res);
765 if (ret == -EBUSY)
766 continue;
767
768 return ret;
769 }
770 return 0;
771 }
772
773 return -ENOSPC;
774}
775
776/*
777 * ttm_bo_mem_space - Wrapper around ttm_bo_alloc_resource
778 *
779 * @bo: Pointer to a struct ttm_buffer_object of which we want a resource for
780 * @placement: Proposed new placement for the buffer object
781 * @res: The resulting struct ttm_resource.
782 * @ctx: if and how to sleep, lock buffers and alloc memory
783 *
784 * Tries both idle allocation and forcefully eviction of buffers. See
785 * ttm_bo_alloc_resource for details.
786 */
787int ttm_bo_mem_space(struct ttm_buffer_object *bo,
788 struct ttm_placement *placement,
789 struct ttm_resource **res,
790 struct ttm_operation_ctx *ctx)
791{
792 bool force_space = false;
793 int ret;
794
795 do {
796 ret = ttm_bo_alloc_resource(bo, placement, ctx,
797 force_space, res);
798 force_space = !force_space;
799 } while (ret == -ENOSPC && force_space);
800
801 return ret;
802}
803EXPORT_SYMBOL(ttm_bo_mem_space);
804
805/**
806 * ttm_bo_validate
807 *
808 * @bo: The buffer object.
809 * @placement: Proposed placement for the buffer object.
810 * @ctx: validation parameters.
811 *
812 * Changes placement and caching policy of the buffer object
813 * according proposed placement.
814 * Returns
815 * -EINVAL on invalid proposed placement.
816 * -ENOMEM on out-of-memory condition.
817 * -EBUSY if no_wait is true and buffer busy.
818 * -ERESTARTSYS if interrupted by a signal.
819 */
820int ttm_bo_validate(struct ttm_buffer_object *bo,
821 struct ttm_placement *placement,
822 struct ttm_operation_ctx *ctx)
823{
824 struct ttm_resource *res;
825 struct ttm_place hop;
826 bool force_space;
827 int ret;
828
829 dma_resv_assert_held(bo->base.resv);
830
831 /*
832 * Remove the backing store if no placement is given.
833 */
834 if (!placement->num_placement)
835 return ttm_bo_pipeline_gutting(bo);
836
837 force_space = false;
838 do {
839 /* Check whether we need to move buffer. */
840 if (bo->resource &&
841 ttm_resource_compatible(res: bo->resource, placement,
842 evicting: force_space))
843 return 0;
844
845 /* Moving of pinned BOs is forbidden */
846 if (bo->pin_count)
847 return -EINVAL;
848
849 /*
850 * Determine where to move the buffer.
851 *
852 * If driver determines move is going to need
853 * an extra step then it will return -EMULTIHOP
854 * and the buffer will be moved to the temporary
855 * stop and the driver will be called to make
856 * the second hop.
857 */
858 ret = ttm_bo_alloc_resource(bo, placement, ctx, force_space,
859 res: &res);
860 force_space = !force_space;
861 if (ret == -ENOSPC)
862 continue;
863 if (ret)
864 return ret;
865
866bounce:
867 ret = ttm_bo_handle_move_mem(bo, mem: res, evict: false, ctx, hop: &hop);
868 if (ret == -EMULTIHOP) {
869 ret = ttm_bo_bounce_temp_buffer(bo, ctx, hop: &hop);
870 /* try and move to final place now. */
871 if (!ret)
872 goto bounce;
873 }
874 if (ret) {
875 ttm_resource_free(bo, res: &res);
876 return ret;
877 }
878
879 } while (ret && force_space);
880
881 /* For backward compatibility with userspace */
882 if (ret == -ENOSPC)
883 return bo->bdev->alloc_flags & TTM_ALLOCATION_PROPAGATE_ENOSPC ?
884 ret : -ENOMEM;
885
886 /*
887 * We might need to add a TTM.
888 */
889 if (!bo->resource || bo->resource->mem_type == TTM_PL_SYSTEM) {
890 ret = ttm_tt_create(bo, zero_alloc: true);
891 if (ret)
892 return ret;
893 }
894 return 0;
895}
896EXPORT_SYMBOL(ttm_bo_validate);
897
898/**
899 * ttm_bo_init_reserved
900 *
901 * @bdev: Pointer to a ttm_device struct.
902 * @bo: Pointer to a ttm_buffer_object to be initialized.
903 * @type: Requested type of buffer object.
904 * @placement: Initial placement for buffer object.
905 * @alignment: Data alignment in pages.
906 * @ctx: TTM operation context for memory allocation.
907 * @sg: Scatter-gather table.
908 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
909 * @destroy: Destroy function. Use NULL for kfree().
910 *
911 * This function initializes a pre-allocated struct ttm_buffer_object.
912 * As this object may be part of a larger structure, this function,
913 * together with the @destroy function, enables driver-specific objects
914 * derived from a ttm_buffer_object.
915 *
916 * On successful return, the caller owns an object kref to @bo. The kref and
917 * list_kref are usually set to 1, but note that in some situations, other
918 * tasks may already be holding references to @bo as well.
919 * Furthermore, if resv == NULL, the buffer's reservation lock will be held,
920 * and it is the caller's responsibility to call ttm_bo_unreserve.
921 *
922 * If a failure occurs, the function will call the @destroy function. Thus,
923 * after a failure, dereferencing @bo is illegal and will likely cause memory
924 * corruption.
925 *
926 * Returns
927 * -ENOMEM: Out of memory.
928 * -EINVAL: Invalid placement flags.
929 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
930 */
931int ttm_bo_init_reserved(struct ttm_device *bdev, struct ttm_buffer_object *bo,
932 enum ttm_bo_type type, struct ttm_placement *placement,
933 uint32_t alignment, struct ttm_operation_ctx *ctx,
934 struct sg_table *sg, struct dma_resv *resv,
935 void (*destroy) (struct ttm_buffer_object *))
936{
937 int ret;
938
939 kref_init(kref: &bo->kref);
940 bo->bdev = bdev;
941 bo->type = type;
942 bo->page_alignment = alignment;
943 bo->destroy = destroy;
944 bo->pin_count = 0;
945 bo->sg = sg;
946 bo->bulk_move = NULL;
947 if (resv)
948 bo->base.resv = resv;
949 else
950 bo->base.resv = &bo->base._resv;
951 atomic_inc(v: &ttm_glob.bo_count);
952
953 /*
954 * For ttm_bo_type_device buffers, allocate
955 * address space from the device.
956 */
957 if (bo->type == ttm_bo_type_device || bo->type == ttm_bo_type_sg) {
958 ret = drm_vma_offset_add(mgr: bdev->vma_manager, node: &bo->base.vma_node,
959 PFN_UP(bo->base.size));
960 if (ret)
961 goto err_put;
962 }
963
964 /* passed reservation objects should already be locked,
965 * since otherwise lockdep will be angered in radeon.
966 */
967 if (!resv)
968 WARN_ON(!dma_resv_trylock(bo->base.resv));
969 else
970 dma_resv_assert_held(resv);
971
972 ret = ttm_bo_validate(bo, placement, ctx);
973 if (unlikely(ret))
974 goto err_unlock;
975
976 return 0;
977
978err_unlock:
979 if (!resv)
980 dma_resv_unlock(obj: bo->base.resv);
981
982err_put:
983 ttm_bo_put(bo);
984 return ret;
985}
986EXPORT_SYMBOL(ttm_bo_init_reserved);
987
988/**
989 * ttm_bo_init_validate
990 *
991 * @bdev: Pointer to a ttm_device struct.
992 * @bo: Pointer to a ttm_buffer_object to be initialized.
993 * @type: Requested type of buffer object.
994 * @placement: Initial placement for buffer object.
995 * @alignment: Data alignment in pages.
996 * @interruptible: If needing to sleep to wait for GPU resources,
997 * sleep interruptible.
998 * pinned in physical memory. If this behaviour is not desired, this member
999 * holds a pointer to a persistent shmem object. Typically, this would
1000 * point to the shmem object backing a GEM object if TTM is used to back a
1001 * GEM user interface.
1002 * @sg: Scatter-gather table.
1003 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
1004 * @destroy: Destroy function. Use NULL for kfree().
1005 *
1006 * This function initializes a pre-allocated struct ttm_buffer_object.
1007 * As this object may be part of a larger structure, this function,
1008 * together with the @destroy function,
1009 * enables driver-specific objects derived from a ttm_buffer_object.
1010 *
1011 * On successful return, the caller owns an object kref to @bo. The kref and
1012 * list_kref are usually set to 1, but note that in some situations, other
1013 * tasks may already be holding references to @bo as well.
1014 *
1015 * If a failure occurs, the function will call the @destroy function, Thus,
1016 * after a failure, dereferencing @bo is illegal and will likely cause memory
1017 * corruption.
1018 *
1019 * Returns
1020 * -ENOMEM: Out of memory.
1021 * -EINVAL: Invalid placement flags.
1022 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
1023 */
1024int ttm_bo_init_validate(struct ttm_device *bdev, struct ttm_buffer_object *bo,
1025 enum ttm_bo_type type, struct ttm_placement *placement,
1026 uint32_t alignment, bool interruptible,
1027 struct sg_table *sg, struct dma_resv *resv,
1028 void (*destroy) (struct ttm_buffer_object *))
1029{
1030 struct ttm_operation_ctx ctx = { interruptible, false };
1031 int ret;
1032
1033 ret = ttm_bo_init_reserved(bdev, bo, type, placement, alignment, &ctx,
1034 sg, resv, destroy);
1035 if (ret)
1036 return ret;
1037
1038 if (!resv)
1039 ttm_bo_unreserve(bo);
1040
1041 return 0;
1042}
1043EXPORT_SYMBOL(ttm_bo_init_validate);
1044
1045/*
1046 * buffer object vm functions.
1047 */
1048
1049/**
1050 * ttm_bo_unmap_virtual
1051 *
1052 * @bo: tear down the virtual mappings for this BO
1053 */
1054void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1055{
1056 struct ttm_device *bdev = bo->bdev;
1057
1058 drm_vma_node_unmap(node: &bo->base.vma_node, file_mapping: bdev->dev_mapping);
1059 ttm_mem_io_free(bdev, mem: bo->resource);
1060}
1061EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1062
1063/**
1064 * ttm_bo_wait_ctx - wait for buffer idle.
1065 *
1066 * @bo: The buffer object.
1067 * @ctx: defines how to wait
1068 *
1069 * Waits for the buffer to be idle. Used timeout depends on the context.
1070 * Returns -EBUSY if wait timed outt, -ERESTARTSYS if interrupted by a signal or
1071 * zero on success.
1072 */
1073int ttm_bo_wait_ctx(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx)
1074{
1075 long ret;
1076
1077 if (ctx->no_wait_gpu) {
1078 if (dma_resv_test_signaled(obj: bo->base.resv,
1079 usage: DMA_RESV_USAGE_BOOKKEEP))
1080 return 0;
1081 else
1082 return -EBUSY;
1083 }
1084
1085 ret = dma_resv_wait_timeout(obj: bo->base.resv, usage: DMA_RESV_USAGE_BOOKKEEP,
1086 intr: ctx->interruptible, timeout: 15 * HZ);
1087 if (unlikely(ret < 0))
1088 return ret;
1089 if (unlikely(ret == 0))
1090 return -EBUSY;
1091 return 0;
1092}
1093EXPORT_SYMBOL(ttm_bo_wait_ctx);
1094
1095/**
1096 * struct ttm_bo_swapout_walk - Parameters for the swapout walk
1097 */
1098struct ttm_bo_swapout_walk {
1099 /** @walk: The walk base parameters. */
1100 struct ttm_lru_walk walk;
1101 /** @gfp_flags: The gfp flags to use for ttm_tt_swapout() */
1102 gfp_t gfp_flags;
1103 /** @hit_low: Whether we should attempt to swap BO's with low watermark threshold */
1104 /** @evict_low: If we cannot swap a bo when @try_low is false (first pass) */
1105 bool hit_low, evict_low;
1106};
1107
1108static s64
1109ttm_bo_swapout_cb(struct ttm_lru_walk *walk, struct ttm_buffer_object *bo)
1110{
1111 struct ttm_place place = {.mem_type = bo->resource->mem_type};
1112 struct ttm_bo_swapout_walk *swapout_walk =
1113 container_of(walk, typeof(*swapout_walk), walk);
1114 struct ttm_operation_ctx *ctx = walk->arg.ctx;
1115 s64 ret;
1116
1117 /*
1118 * While the bo may already reside in SYSTEM placement, set
1119 * SYSTEM as new placement to cover also the move further below.
1120 * The driver may use the fact that we're moving from SYSTEM
1121 * as an indication that we're about to swap out.
1122 */
1123 if (bo->pin_count || !bo->bdev->funcs->eviction_valuable(bo, &place)) {
1124 ret = -EBUSY;
1125 goto out;
1126 }
1127
1128 if (!bo->ttm || !ttm_tt_is_populated(tt: bo->ttm) ||
1129 bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1130 bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED) {
1131 ret = -EBUSY;
1132 goto out;
1133 }
1134
1135 if (bo->deleted) {
1136 pgoff_t num_pages = bo->ttm->num_pages;
1137
1138 ret = ttm_bo_wait_ctx(bo, ctx);
1139 if (ret)
1140 goto out;
1141
1142 ttm_bo_cleanup_memtype_use(bo);
1143 ret = num_pages;
1144 goto out;
1145 }
1146
1147 /*
1148 * Move to system cached
1149 */
1150 if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1151 struct ttm_resource *evict_mem;
1152 struct ttm_place hop;
1153
1154 memset(&hop, 0, sizeof(hop));
1155 place.mem_type = TTM_PL_SYSTEM;
1156 ret = ttm_resource_alloc(bo, place: &place, res: &evict_mem, NULL);
1157 if (ret)
1158 goto out;
1159
1160 ret = ttm_bo_handle_move_mem(bo, mem: evict_mem, evict: true, ctx, hop: &hop);
1161 if (ret) {
1162 WARN(ret == -EMULTIHOP,
1163 "Unexpected multihop in swapout - likely driver bug.\n");
1164 ttm_resource_free(bo, res: &evict_mem);
1165 goto out;
1166 }
1167 }
1168
1169 /*
1170 * Make sure BO is idle.
1171 */
1172 ret = ttm_bo_wait_ctx(bo, ctx);
1173 if (ret)
1174 goto out;
1175
1176 ttm_bo_unmap_virtual(bo);
1177 if (bo->bdev->funcs->swap_notify)
1178 bo->bdev->funcs->swap_notify(bo);
1179
1180 if (ttm_tt_is_populated(tt: bo->ttm)) {
1181 spin_lock(lock: &bo->bdev->lru_lock);
1182 ttm_resource_del_bulk_move(res: bo->resource, bo);
1183 spin_unlock(lock: &bo->bdev->lru_lock);
1184
1185 ret = ttm_tt_swapout(bdev: bo->bdev, ttm: bo->ttm, gfp_flags: swapout_walk->gfp_flags);
1186
1187 spin_lock(lock: &bo->bdev->lru_lock);
1188 if (ret)
1189 ttm_resource_add_bulk_move(res: bo->resource, bo);
1190 ttm_resource_move_to_lru_tail(res: bo->resource);
1191 spin_unlock(lock: &bo->bdev->lru_lock);
1192 }
1193
1194out:
1195 /* Consider -ENOMEM and -ENOSPC non-fatal. */
1196 if (ret == -ENOMEM || ret == -ENOSPC)
1197 ret = -EBUSY;
1198
1199 return ret;
1200}
1201
1202const struct ttm_lru_walk_ops ttm_swap_ops = {
1203 .process_bo = ttm_bo_swapout_cb,
1204};
1205
1206/**
1207 * ttm_bo_swapout() - Swap out buffer objects on the LRU list to shmem.
1208 * @bdev: The ttm device.
1209 * @ctx: The ttm_operation_ctx governing the swapout operation.
1210 * @man: The resource manager whose resources / buffer objects are
1211 * goint to be swapped out.
1212 * @gfp_flags: The gfp flags used for shmem page allocations.
1213 * @target: The desired number of bytes to swap out.
1214 *
1215 * Return: The number of bytes actually swapped out, or negative error code
1216 * on error.
1217 */
1218s64 ttm_bo_swapout(struct ttm_device *bdev, struct ttm_operation_ctx *ctx,
1219 struct ttm_resource_manager *man, gfp_t gfp_flags,
1220 s64 target)
1221{
1222 struct ttm_bo_swapout_walk swapout_walk = {
1223 .walk = {
1224 .ops = &ttm_swap_ops,
1225 .arg = {
1226 .ctx = ctx,
1227 .trylock_only = true,
1228 },
1229 },
1230 .gfp_flags = gfp_flags,
1231 };
1232
1233 return ttm_lru_walk_for_evict(walk: &swapout_walk.walk, bdev, man, target);
1234}
1235
1236void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1237{
1238 if (bo->ttm == NULL)
1239 return;
1240
1241 ttm_tt_unpopulate(bdev: bo->bdev, ttm: bo->ttm);
1242 ttm_tt_destroy(bdev: bo->bdev, ttm: bo->ttm);
1243 bo->ttm = NULL;
1244}
1245
1246/**
1247 * ttm_bo_populate() - Ensure that a buffer object has backing pages
1248 * @bo: The buffer object
1249 * @ctx: The ttm_operation_ctx governing the operation.
1250 *
1251 * For buffer objects in a memory type whose manager uses
1252 * struct ttm_tt for backing pages, ensure those backing pages
1253 * are present and with valid content. The bo's resource is also
1254 * placed on the correct LRU list if it was previously swapped
1255 * out.
1256 *
1257 * Return: 0 if successful, negative error code on failure.
1258 * Note: May return -EINTR or -ERESTARTSYS if @ctx::interruptible
1259 * is set to true.
1260 */
1261int ttm_bo_populate(struct ttm_buffer_object *bo,
1262 struct ttm_operation_ctx *ctx)
1263{
1264 struct ttm_tt *tt = bo->ttm;
1265 bool swapped;
1266 int ret;
1267
1268 dma_resv_assert_held(bo->base.resv);
1269
1270 if (!tt)
1271 return 0;
1272
1273 swapped = ttm_tt_is_swapped(tt);
1274 ret = ttm_tt_populate(bdev: bo->bdev, ttm: tt, ctx);
1275 if (ret)
1276 return ret;
1277
1278 if (swapped && !ttm_tt_is_swapped(tt) && !bo->pin_count &&
1279 bo->resource) {
1280 spin_lock(lock: &bo->bdev->lru_lock);
1281 ttm_resource_add_bulk_move(res: bo->resource, bo);
1282 ttm_resource_move_to_lru_tail(res: bo->resource);
1283 spin_unlock(lock: &bo->bdev->lru_lock);
1284 }
1285
1286 return 0;
1287}
1288EXPORT_SYMBOL(ttm_bo_populate);
1289
1290int ttm_bo_setup_export(struct ttm_buffer_object *bo,
1291 struct ttm_operation_ctx *ctx)
1292{
1293 int ret;
1294
1295 ret = ttm_bo_reserve(bo, interruptible: false, no_wait: false, NULL);
1296 if (ret != 0)
1297 return ret;
1298
1299 ret = ttm_bo_populate(bo, ctx);
1300 ttm_bo_unreserve(bo);
1301 return ret;
1302}
1303EXPORT_SYMBOL(ttm_bo_setup_export);
1304

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