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
ttm_bo_mem_space_debug(struct ttm_buffer_object * bo,struct ttm_placement * placement)49 static 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 */
ttm_bo_move_to_lru_tail(struct ttm_buffer_object * bo)74 void 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 }
81 EXPORT_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 */
ttm_bo_set_bulk_move(struct ttm_buffer_object * bo,struct ttm_lru_bulk_move * bulk)97 void 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 }
113 EXPORT_SYMBOL(ttm_bo_set_bulk_move);
114
ttm_bo_handle_move_mem(struct ttm_buffer_object * bo,struct ttm_resource * mem,bool evict,struct ttm_operation_ctx * ctx,struct ttm_place * hop)115 static 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
162 out_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
ttm_bo_cleanup_memtype_use(struct ttm_buffer_object * bo)177 static 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
ttm_bo_individualize_resv(struct ttm_buffer_object * bo)186 static 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
ttm_bo_flush_all_fences(struct ttm_buffer_object * bo)213 static 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
ttm_bo_cleanup_refs(struct ttm_buffer_object * bo,bool interruptible,bool no_wait_gpu,bool unlock_resv)241 static 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 */
ttm_bo_delayed_delete(struct work_struct * work)305 static 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
ttm_bo_release(struct kref * kref)319 static 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 queue_work(bdev->wq, &bo->delayed_delete);
374 return;
375 }
376
377 ttm_bo_cleanup_memtype_use(bo);
378 dma_resv_unlock(bo->base.resv);
379 }
380
381 atomic_dec(&ttm_glob.bo_count);
382 bo->destroy(bo);
383 }
384
385 /**
386 * ttm_bo_put
387 *
388 * @bo: The buffer object.
389 *
390 * Unreference a buffer object.
391 */
ttm_bo_put(struct ttm_buffer_object * bo)392 void ttm_bo_put(struct ttm_buffer_object *bo)
393 {
394 kref_put(&bo->kref, ttm_bo_release);
395 }
396 EXPORT_SYMBOL(ttm_bo_put);
397
ttm_bo_bounce_temp_buffer(struct ttm_buffer_object * bo,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx,struct ttm_place * hop)398 static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
399 struct ttm_resource **mem,
400 struct ttm_operation_ctx *ctx,
401 struct ttm_place *hop)
402 {
403 struct ttm_placement hop_placement;
404 struct ttm_resource *hop_mem;
405 int ret;
406
407 hop_placement.num_placement = hop_placement.num_busy_placement = 1;
408 hop_placement.placement = hop_placement.busy_placement = hop;
409
410 /* find space in the bounce domain */
411 ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
412 if (ret)
413 return ret;
414 /* move to the bounce domain */
415 ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
416 if (ret) {
417 ttm_resource_free(bo, &hop_mem);
418 return ret;
419 }
420 return 0;
421 }
422
ttm_bo_evict(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)423 static int ttm_bo_evict(struct ttm_buffer_object *bo,
424 struct ttm_operation_ctx *ctx)
425 {
426 struct ttm_device *bdev = bo->bdev;
427 struct ttm_resource *evict_mem;
428 struct ttm_placement placement;
429 struct ttm_place hop;
430 int ret = 0;
431
432 memset(&hop, 0, sizeof(hop));
433
434 dma_resv_assert_held(bo->base.resv);
435
436 placement.num_placement = 0;
437 placement.num_busy_placement = 0;
438 bdev->funcs->evict_flags(bo, &placement);
439
440 if (!placement.num_placement && !placement.num_busy_placement) {
441 ret = ttm_bo_wait_ctx(bo, ctx);
442 if (ret)
443 return ret;
444
445 /*
446 * Since we've already synced, this frees backing store
447 * immediately.
448 */
449 return ttm_bo_pipeline_gutting(bo);
450 }
451
452 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
453 if (ret) {
454 if (ret != -ERESTARTSYS) {
455 pr_err("Failed to find memory space for buffer 0x%p eviction\n",
456 bo);
457 ttm_bo_mem_space_debug(bo, &placement);
458 }
459 goto out;
460 }
461
462 do {
463 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
464 if (ret != -EMULTIHOP)
465 break;
466
467 ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
468 } while (!ret);
469
470 if (ret) {
471 ttm_resource_free(bo, &evict_mem);
472 if (ret != -ERESTARTSYS && ret != -EINTR)
473 pr_err("Buffer eviction failed\n");
474 }
475 out:
476 return ret;
477 }
478
479 /**
480 * ttm_bo_eviction_valuable
481 *
482 * @bo: The buffer object to evict
483 * @place: the placement we need to make room for
484 *
485 * Check if it is valuable to evict the BO to make room for the given placement.
486 */
ttm_bo_eviction_valuable(struct ttm_buffer_object * bo,const struct ttm_place * place)487 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
488 const struct ttm_place *place)
489 {
490 struct ttm_resource *res = bo->resource;
491 struct ttm_device *bdev = bo->bdev;
492
493 dma_resv_assert_held(bo->base.resv);
494 if (bo->resource->mem_type == TTM_PL_SYSTEM)
495 return true;
496
497 /* Don't evict this BO if it's outside of the
498 * requested placement range
499 */
500 return ttm_resource_intersects(bdev, res, place, bo->base.size);
501 }
502 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
503
504 /*
505 * Check the target bo is allowable to be evicted or swapout, including cases:
506 *
507 * a. if share same reservation object with ctx->resv, have assumption
508 * reservation objects should already be locked, so not lock again and
509 * return true directly when either the opreation allow_reserved_eviction
510 * or the target bo already is in delayed free list;
511 *
512 * b. Otherwise, trylock it.
513 */
ttm_bo_evict_swapout_allowable(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,const struct ttm_place * place,bool * locked,bool * busy)514 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
515 struct ttm_operation_ctx *ctx,
516 const struct ttm_place *place,
517 bool *locked, bool *busy)
518 {
519 bool ret = false;
520
521 if (bo->pin_count) {
522 *locked = false;
523 if (busy)
524 *busy = false;
525 return false;
526 }
527
528 if (bo->base.resv == ctx->resv) {
529 dma_resv_assert_held(bo->base.resv);
530 if (ctx->allow_res_evict)
531 ret = true;
532 *locked = false;
533 if (busy)
534 *busy = false;
535 } else {
536 ret = dma_resv_trylock(bo->base.resv);
537 *locked = ret;
538 if (busy)
539 *busy = !ret;
540 }
541
542 if (ret && place && (bo->resource->mem_type != place->mem_type ||
543 !bo->bdev->funcs->eviction_valuable(bo, place))) {
544 ret = false;
545 if (*locked) {
546 dma_resv_unlock(bo->base.resv);
547 *locked = false;
548 }
549 }
550
551 return ret;
552 }
553
554 /**
555 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
556 *
557 * @busy_bo: BO which couldn't be locked with trylock
558 * @ctx: operation context
559 * @ticket: acquire ticket
560 *
561 * Try to lock a busy buffer object to avoid failing eviction.
562 */
ttm_mem_evict_wait_busy(struct ttm_buffer_object * busy_bo,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)563 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
564 struct ttm_operation_ctx *ctx,
565 struct ww_acquire_ctx *ticket)
566 {
567 int r;
568
569 if (!busy_bo || !ticket)
570 return -EBUSY;
571
572 if (ctx->interruptible)
573 r = dma_resv_lock_interruptible(busy_bo->base.resv,
574 ticket);
575 else
576 r = dma_resv_lock(busy_bo->base.resv, ticket);
577
578 /*
579 * TODO: It would be better to keep the BO locked until allocation is at
580 * least tried one more time, but that would mean a much larger rework
581 * of TTM.
582 */
583 if (!r)
584 dma_resv_unlock(busy_bo->base.resv);
585
586 return r == -EDEADLK ? -EBUSY : r;
587 }
588
ttm_mem_evict_first(struct ttm_device * bdev,struct ttm_resource_manager * man,const struct ttm_place * place,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)589 int ttm_mem_evict_first(struct ttm_device *bdev,
590 struct ttm_resource_manager *man,
591 const struct ttm_place *place,
592 struct ttm_operation_ctx *ctx,
593 struct ww_acquire_ctx *ticket)
594 {
595 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
596 struct ttm_resource_cursor cursor;
597 struct ttm_resource *res;
598 bool locked = false;
599 int ret;
600
601 spin_lock(&bdev->lru_lock);
602 ttm_resource_manager_for_each_res(man, &cursor, res) {
603 bool busy;
604
605 if (!ttm_bo_evict_swapout_allowable(res->bo, ctx, place,
606 &locked, &busy)) {
607 if (busy && !busy_bo && ticket !=
608 dma_resv_locking_ctx(res->bo->base.resv))
609 busy_bo = res->bo;
610 continue;
611 }
612
613 if (ttm_bo_get_unless_zero(res->bo)) {
614 bo = res->bo;
615 break;
616 }
617 if (locked)
618 dma_resv_unlock(res->bo->base.resv);
619 }
620
621 if (!bo) {
622 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
623 busy_bo = NULL;
624 spin_unlock(&bdev->lru_lock);
625 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
626 if (busy_bo)
627 ttm_bo_put(busy_bo);
628 return ret;
629 }
630
631 if (bo->deleted) {
632 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
633 ctx->no_wait_gpu, locked);
634 ttm_bo_put(bo);
635 return ret;
636 }
637
638 spin_unlock(&bdev->lru_lock);
639
640 ret = ttm_bo_evict(bo, ctx);
641 if (locked)
642 ttm_bo_unreserve(bo);
643 else
644 ttm_bo_move_to_lru_tail_unlocked(bo);
645
646 ttm_bo_put(bo);
647 return ret;
648 }
649
650 /**
651 * ttm_bo_pin - Pin the buffer object.
652 * @bo: The buffer object to pin
653 *
654 * Make sure the buffer is not evicted any more during memory pressure.
655 * @bo must be unpinned again by calling ttm_bo_unpin().
656 */
ttm_bo_pin(struct ttm_buffer_object * bo)657 void ttm_bo_pin(struct ttm_buffer_object *bo)
658 {
659 dma_resv_assert_held(bo->base.resv);
660 WARN_ON_ONCE(!kref_read(&bo->kref));
661 spin_lock(&bo->bdev->lru_lock);
662 if (bo->resource)
663 ttm_resource_del_bulk_move(bo->resource, bo);
664 ++bo->pin_count;
665 spin_unlock(&bo->bdev->lru_lock);
666 }
667 EXPORT_SYMBOL(ttm_bo_pin);
668
669 /**
670 * ttm_bo_unpin - Unpin the buffer object.
671 * @bo: The buffer object to unpin
672 *
673 * Allows the buffer object to be evicted again during memory pressure.
674 */
ttm_bo_unpin(struct ttm_buffer_object * bo)675 void ttm_bo_unpin(struct ttm_buffer_object *bo)
676 {
677 dma_resv_assert_held(bo->base.resv);
678 WARN_ON_ONCE(!kref_read(&bo->kref));
679 if (WARN_ON_ONCE(!bo->pin_count))
680 return;
681
682 spin_lock(&bo->bdev->lru_lock);
683 --bo->pin_count;
684 if (bo->resource)
685 ttm_resource_add_bulk_move(bo->resource, bo);
686 spin_unlock(&bo->bdev->lru_lock);
687 }
688 EXPORT_SYMBOL(ttm_bo_unpin);
689
690 /*
691 * Add the last move fence to the BO as kernel dependency and reserve a new
692 * fence slot.
693 */
ttm_bo_add_move_fence(struct ttm_buffer_object * bo,struct ttm_resource_manager * man,struct ttm_resource * mem,bool no_wait_gpu)694 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
695 struct ttm_resource_manager *man,
696 struct ttm_resource *mem,
697 bool no_wait_gpu)
698 {
699 struct dma_fence *fence;
700 int ret;
701
702 spin_lock(&man->move_lock);
703 fence = dma_fence_get(man->move);
704 spin_unlock(&man->move_lock);
705
706 if (!fence)
707 return 0;
708
709 if (no_wait_gpu) {
710 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
711 dma_fence_put(fence);
712 return ret;
713 }
714
715 dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
716
717 ret = dma_resv_reserve_fences(bo->base.resv, 1);
718 dma_fence_put(fence);
719 return ret;
720 }
721
722 /*
723 * Repeatedly evict memory from the LRU for @mem_type until we create enough
724 * space, or we've evicted everything and there isn't enough space.
725 */
ttm_bo_mem_force_space(struct ttm_buffer_object * bo,const struct ttm_place * place,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx)726 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
727 const struct ttm_place *place,
728 struct ttm_resource **mem,
729 struct ttm_operation_ctx *ctx)
730 {
731 struct ttm_device *bdev = bo->bdev;
732 struct ttm_resource_manager *man;
733 struct ww_acquire_ctx *ticket;
734 int ret;
735
736 man = ttm_manager_type(bdev, place->mem_type);
737 ticket = dma_resv_locking_ctx(bo->base.resv);
738 do {
739 ret = ttm_resource_alloc(bo, place, mem);
740 if (likely(!ret))
741 break;
742 if (unlikely(ret != -ENOSPC))
743 return ret;
744 ret = ttm_mem_evict_first(bdev, man, place, ctx,
745 ticket);
746 if (unlikely(ret != 0))
747 return ret;
748 } while (1);
749
750 return ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
751 }
752
753 /**
754 * ttm_bo_mem_space
755 *
756 * @bo: Pointer to a struct ttm_buffer_object. the data of which
757 * we want to allocate space for.
758 * @placement: Proposed new placement for the buffer object.
759 * @mem: A struct ttm_resource.
760 * @ctx: if and how to sleep, lock buffers and alloc memory
761 *
762 * Allocate memory space for the buffer object pointed to by @bo, using
763 * the placement flags in @placement, potentially evicting other idle buffer objects.
764 * This function may sleep while waiting for space to become available.
765 * Returns:
766 * -EBUSY: No space available (only if no_wait == 1).
767 * -ENOSPC: Could not allocate space for the buffer object, either due to
768 * fragmentation or concurrent allocators.
769 * -ERESTARTSYS: An interruptible sleep was interrupted by a signal.
770 */
ttm_bo_mem_space(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx)771 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
772 struct ttm_placement *placement,
773 struct ttm_resource **mem,
774 struct ttm_operation_ctx *ctx)
775 {
776 struct ttm_device *bdev = bo->bdev;
777 bool type_found = false;
778 int i, ret;
779
780 ret = dma_resv_reserve_fences(bo->base.resv, 1);
781 if (unlikely(ret))
782 return ret;
783
784 for (i = 0; i < placement->num_placement; ++i) {
785 const struct ttm_place *place = &placement->placement[i];
786 struct ttm_resource_manager *man;
787
788 man = ttm_manager_type(bdev, place->mem_type);
789 if (!man || !ttm_resource_manager_used(man))
790 continue;
791
792 type_found = true;
793 ret = ttm_resource_alloc(bo, place, mem);
794 if (ret == -ENOSPC)
795 continue;
796 if (unlikely(ret))
797 goto error;
798
799 ret = ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
800 if (unlikely(ret)) {
801 ttm_resource_free(bo, mem);
802 if (ret == -EBUSY)
803 continue;
804
805 goto error;
806 }
807 return 0;
808 }
809
810 for (i = 0; i < placement->num_busy_placement; ++i) {
811 const struct ttm_place *place = &placement->busy_placement[i];
812 struct ttm_resource_manager *man;
813
814 man = ttm_manager_type(bdev, place->mem_type);
815 if (!man || !ttm_resource_manager_used(man))
816 continue;
817
818 type_found = true;
819 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
820 if (likely(!ret))
821 return 0;
822
823 if (ret && ret != -EBUSY)
824 goto error;
825 }
826
827 ret = -ENOSPC;
828 if (!type_found) {
829 pr_err(TTM_PFX "No compatible memory type found\n");
830 ret = -EINVAL;
831 }
832
833 error:
834 return ret;
835 }
836 EXPORT_SYMBOL(ttm_bo_mem_space);
837
ttm_bo_move_buffer(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)838 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
839 struct ttm_placement *placement,
840 struct ttm_operation_ctx *ctx)
841 {
842 struct ttm_resource *mem;
843 struct ttm_place hop;
844 int ret;
845
846 dma_resv_assert_held(bo->base.resv);
847
848 /*
849 * Determine where to move the buffer.
850 *
851 * If driver determines move is going to need
852 * an extra step then it will return -EMULTIHOP
853 * and the buffer will be moved to the temporary
854 * stop and the driver will be called to make
855 * the second hop.
856 */
857 ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
858 if (ret)
859 return ret;
860 bounce:
861 ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
862 if (ret == -EMULTIHOP) {
863 ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
864 if (ret)
865 goto out;
866 /* try and move to final place now. */
867 goto bounce;
868 }
869 out:
870 if (ret)
871 ttm_resource_free(bo, &mem);
872 return ret;
873 }
874
875 /**
876 * ttm_bo_validate
877 *
878 * @bo: The buffer object.
879 * @placement: Proposed placement for the buffer object.
880 * @ctx: validation parameters.
881 *
882 * Changes placement and caching policy of the buffer object
883 * according proposed placement.
884 * Returns
885 * -EINVAL on invalid proposed placement.
886 * -ENOMEM on out-of-memory condition.
887 * -EBUSY if no_wait is true and buffer busy.
888 * -ERESTARTSYS if interrupted by a signal.
889 */
ttm_bo_validate(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)890 int ttm_bo_validate(struct ttm_buffer_object *bo,
891 struct ttm_placement *placement,
892 struct ttm_operation_ctx *ctx)
893 {
894 int ret;
895
896 dma_resv_assert_held(bo->base.resv);
897
898 /*
899 * Remove the backing store if no placement is given.
900 */
901 if (!placement->num_placement && !placement->num_busy_placement)
902 return ttm_bo_pipeline_gutting(bo);
903
904 /* Check whether we need to move buffer. */
905 if (bo->resource && ttm_resource_compat(bo->resource, placement))
906 return 0;
907
908 /* Moving of pinned BOs is forbidden */
909 if (bo->pin_count)
910 return -EINVAL;
911
912 ret = ttm_bo_move_buffer(bo, placement, ctx);
913 /* For backward compatibility with userspace */
914 if (ret == -ENOSPC)
915 return -ENOMEM;
916 if (ret)
917 return ret;
918
919 /*
920 * We might need to add a TTM.
921 */
922 if (!bo->resource || bo->resource->mem_type == TTM_PL_SYSTEM) {
923 ret = ttm_tt_create(bo, true);
924 if (ret)
925 return ret;
926 }
927 return 0;
928 }
929 EXPORT_SYMBOL(ttm_bo_validate);
930
931 /**
932 * ttm_bo_init_reserved
933 *
934 * @bdev: Pointer to a ttm_device struct.
935 * @bo: Pointer to a ttm_buffer_object to be initialized.
936 * @type: Requested type of buffer object.
937 * @placement: Initial placement for buffer object.
938 * @alignment: Data alignment in pages.
939 * @ctx: TTM operation context for memory allocation.
940 * @sg: Scatter-gather table.
941 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
942 * @destroy: Destroy function. Use NULL for kfree().
943 *
944 * This function initializes a pre-allocated struct ttm_buffer_object.
945 * As this object may be part of a larger structure, this function,
946 * together with the @destroy function, enables driver-specific objects
947 * derived from a ttm_buffer_object.
948 *
949 * On successful return, the caller owns an object kref to @bo. The kref and
950 * list_kref are usually set to 1, but note that in some situations, other
951 * tasks may already be holding references to @bo as well.
952 * Furthermore, if resv == NULL, the buffer's reservation lock will be held,
953 * and it is the caller's responsibility to call ttm_bo_unreserve.
954 *
955 * If a failure occurs, the function will call the @destroy function. Thus,
956 * after a failure, dereferencing @bo is illegal and will likely cause memory
957 * corruption.
958 *
959 * Returns
960 * -ENOMEM: Out of memory.
961 * -EINVAL: Invalid placement flags.
962 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
963 */
ttm_bo_init_reserved(struct ttm_device * bdev,struct ttm_buffer_object * bo,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t alignment,struct ttm_operation_ctx * ctx,struct sg_table * sg,struct dma_resv * resv,void (* destroy)(struct ttm_buffer_object *))964 int ttm_bo_init_reserved(struct ttm_device *bdev, struct ttm_buffer_object *bo,
965 enum ttm_bo_type type, struct ttm_placement *placement,
966 uint32_t alignment, struct ttm_operation_ctx *ctx,
967 struct sg_table *sg, struct dma_resv *resv,
968 void (*destroy) (struct ttm_buffer_object *))
969 {
970 int ret;
971
972 kref_init(&bo->kref);
973 bo->bdev = bdev;
974 bo->type = type;
975 bo->page_alignment = alignment;
976 bo->destroy = destroy;
977 bo->pin_count = 0;
978 bo->sg = sg;
979 bo->bulk_move = NULL;
980 if (resv)
981 bo->base.resv = resv;
982 else
983 bo->base.resv = &bo->base._resv;
984 atomic_inc(&ttm_glob.bo_count);
985
986 /*
987 * For ttm_bo_type_device buffers, allocate
988 * address space from the device.
989 */
990 if (bo->type == ttm_bo_type_device || bo->type == ttm_bo_type_sg) {
991 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
992 PFN_UP(bo->base.size));
993 if (ret)
994 goto err_put;
995 }
996
997 /* passed reservation objects should already be locked,
998 * since otherwise lockdep will be angered in radeon.
999 */
1000 if (!resv)
1001 WARN_ON(!dma_resv_trylock(bo->base.resv));
1002 else
1003 dma_resv_assert_held(resv);
1004
1005 ret = ttm_bo_validate(bo, placement, ctx);
1006 if (unlikely(ret))
1007 goto err_unlock;
1008
1009 return 0;
1010
1011 err_unlock:
1012 if (!resv)
1013 dma_resv_unlock(bo->base.resv);
1014
1015 err_put:
1016 ttm_bo_put(bo);
1017 return ret;
1018 }
1019 EXPORT_SYMBOL(ttm_bo_init_reserved);
1020
1021 /**
1022 * ttm_bo_init_validate
1023 *
1024 * @bdev: Pointer to a ttm_device struct.
1025 * @bo: Pointer to a ttm_buffer_object to be initialized.
1026 * @type: Requested type of buffer object.
1027 * @placement: Initial placement for buffer object.
1028 * @alignment: Data alignment in pages.
1029 * @interruptible: If needing to sleep to wait for GPU resources,
1030 * sleep interruptible.
1031 * pinned in physical memory. If this behaviour is not desired, this member
1032 * holds a pointer to a persistent shmem object. Typically, this would
1033 * point to the shmem object backing a GEM object if TTM is used to back a
1034 * GEM user interface.
1035 * @sg: Scatter-gather table.
1036 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
1037 * @destroy: Destroy function. Use NULL for kfree().
1038 *
1039 * This function initializes a pre-allocated struct ttm_buffer_object.
1040 * As this object may be part of a larger structure, this function,
1041 * together with the @destroy function,
1042 * enables driver-specific objects derived from a ttm_buffer_object.
1043 *
1044 * On successful return, the caller owns an object kref to @bo. The kref and
1045 * list_kref are usually set to 1, but note that in some situations, other
1046 * tasks may already be holding references to @bo as well.
1047 *
1048 * If a failure occurs, the function will call the @destroy function, Thus,
1049 * after a failure, dereferencing @bo is illegal and will likely cause memory
1050 * corruption.
1051 *
1052 * Returns
1053 * -ENOMEM: Out of memory.
1054 * -EINVAL: Invalid placement flags.
1055 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
1056 */
ttm_bo_init_validate(struct ttm_device * bdev,struct ttm_buffer_object * bo,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t alignment,bool interruptible,struct sg_table * sg,struct dma_resv * resv,void (* destroy)(struct ttm_buffer_object *))1057 int ttm_bo_init_validate(struct ttm_device *bdev, struct ttm_buffer_object *bo,
1058 enum ttm_bo_type type, struct ttm_placement *placement,
1059 uint32_t alignment, bool interruptible,
1060 struct sg_table *sg, struct dma_resv *resv,
1061 void (*destroy) (struct ttm_buffer_object *))
1062 {
1063 struct ttm_operation_ctx ctx = { interruptible, false };
1064 int ret;
1065
1066 ret = ttm_bo_init_reserved(bdev, bo, type, placement, alignment, &ctx,
1067 sg, resv, destroy);
1068 if (ret)
1069 return ret;
1070
1071 if (!resv)
1072 ttm_bo_unreserve(bo);
1073
1074 return 0;
1075 }
1076 EXPORT_SYMBOL(ttm_bo_init_validate);
1077
1078 /*
1079 * buffer object vm functions.
1080 */
1081
1082 /**
1083 * ttm_bo_unmap_virtual
1084 *
1085 * @bo: tear down the virtual mappings for this BO
1086 */
ttm_bo_unmap_virtual(struct ttm_buffer_object * bo)1087 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1088 {
1089 struct ttm_device *bdev = bo->bdev;
1090
1091 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1092 ttm_mem_io_free(bdev, bo->resource);
1093 }
1094 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1095
1096 /**
1097 * ttm_bo_wait_ctx - wait for buffer idle.
1098 *
1099 * @bo: The buffer object.
1100 * @ctx: defines how to wait
1101 *
1102 * Waits for the buffer to be idle. Used timeout depends on the context.
1103 * Returns -EBUSY if wait timed outt, -ERESTARTSYS if interrupted by a signal or
1104 * zero on success.
1105 */
ttm_bo_wait_ctx(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)1106 int ttm_bo_wait_ctx(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx)
1107 {
1108 long ret;
1109
1110 if (ctx->no_wait_gpu) {
1111 if (dma_resv_test_signaled(bo->base.resv,
1112 DMA_RESV_USAGE_BOOKKEEP))
1113 return 0;
1114 else
1115 return -EBUSY;
1116 }
1117
1118 ret = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
1119 ctx->interruptible, 15 * HZ);
1120 if (unlikely(ret < 0))
1121 return ret;
1122 if (unlikely(ret == 0))
1123 return -EBUSY;
1124 return 0;
1125 }
1126 EXPORT_SYMBOL(ttm_bo_wait_ctx);
1127
ttm_bo_swapout(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,gfp_t gfp_flags)1128 int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1129 gfp_t gfp_flags)
1130 {
1131 struct ttm_place place;
1132 bool locked;
1133 long ret;
1134
1135 /*
1136 * While the bo may already reside in SYSTEM placement, set
1137 * SYSTEM as new placement to cover also the move further below.
1138 * The driver may use the fact that we're moving from SYSTEM
1139 * as an indication that we're about to swap out.
1140 */
1141 memset(&place, 0, sizeof(place));
1142 place.mem_type = bo->resource->mem_type;
1143 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1144 return -EBUSY;
1145
1146 if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1147 bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1148 bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED ||
1149 !ttm_bo_get_unless_zero(bo)) {
1150 if (locked)
1151 dma_resv_unlock(bo->base.resv);
1152 return -EBUSY;
1153 }
1154
1155 if (bo->deleted) {
1156 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1157 ttm_bo_put(bo);
1158 return ret == -EBUSY ? -ENOSPC : ret;
1159 }
1160
1161 /* TODO: Cleanup the locking */
1162 spin_unlock(&bo->bdev->lru_lock);
1163
1164 /*
1165 * Move to system cached
1166 */
1167 if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1168 struct ttm_resource *evict_mem;
1169 struct ttm_place hop;
1170
1171 memset(&hop, 0, sizeof(hop));
1172 place.mem_type = TTM_PL_SYSTEM;
1173 ret = ttm_resource_alloc(bo, &place, &evict_mem);
1174 if (unlikely(ret))
1175 goto out;
1176
1177 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
1178 if (unlikely(ret != 0)) {
1179 WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1180 ttm_resource_free(bo, &evict_mem);
1181 goto out;
1182 }
1183 }
1184
1185 /*
1186 * Make sure BO is idle.
1187 */
1188 ret = ttm_bo_wait_ctx(bo, ctx);
1189 if (unlikely(ret != 0))
1190 goto out;
1191
1192 ttm_bo_unmap_virtual(bo);
1193
1194 /*
1195 * Swap out. Buffer will be swapped in again as soon as
1196 * anyone tries to access a ttm page.
1197 */
1198 if (bo->bdev->funcs->swap_notify)
1199 bo->bdev->funcs->swap_notify(bo);
1200
1201 if (ttm_tt_is_populated(bo->ttm))
1202 ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1203 out:
1204
1205 /*
1206 * Unreserve without putting on LRU to avoid swapping out an
1207 * already swapped buffer.
1208 */
1209 if (locked)
1210 dma_resv_unlock(bo->base.resv);
1211 ttm_bo_put(bo);
1212 return ret == -EBUSY ? -ENOSPC : ret;
1213 }
1214
ttm_bo_tt_destroy(struct ttm_buffer_object * bo)1215 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1216 {
1217 if (bo->ttm == NULL)
1218 return;
1219
1220 ttm_tt_unpopulate(bo->bdev, bo->ttm);
1221 ttm_tt_destroy(bo->bdev, bo->ttm);
1222 bo->ttm = NULL;
1223 }
1224