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  *
88  * Update the BOs bulk move object, making sure that resources are added/removed
89  * as well. A bulk move allows to move many resource on the LRU at once,
90  * resulting in much less overhead of maintaining the LRU.
91  * The only requirement is that the resources stay together on the LRU and are
92  * never separated. This is enforces by setting the bulk_move structure on a BO.
93  * ttm_lru_bulk_move_tail() should be used to move all resources to the tail of
94  * their LRU list.
95  */
ttm_bo_set_bulk_move(struct ttm_buffer_object * bo,struct ttm_lru_bulk_move * bulk)96 void ttm_bo_set_bulk_move(struct ttm_buffer_object *bo,
97 			  struct ttm_lru_bulk_move *bulk)
98 {
99 	dma_resv_assert_held(bo->base.resv);
100 
101 	if (bo->bulk_move == bulk)
102 		return;
103 
104 	spin_lock(&bo->bdev->lru_lock);
105 	if (bo->resource)
106 		ttm_resource_del_bulk_move(bo->resource, bo);
107 	bo->bulk_move = bulk;
108 	if (bo->resource)
109 		ttm_resource_add_bulk_move(bo->resource, bo);
110 	spin_unlock(&bo->bdev->lru_lock);
111 }
112 EXPORT_SYMBOL(ttm_bo_set_bulk_move);
113 
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)114 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
115 				  struct ttm_resource *mem, bool evict,
116 				  struct ttm_operation_ctx *ctx,
117 				  struct ttm_place *hop)
118 {
119 	struct ttm_device *bdev = bo->bdev;
120 	bool old_use_tt, new_use_tt;
121 	int ret;
122 
123 	old_use_tt = bo->resource &&
124 		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 	ttm_bo_put(bo);
299 
300 	return 0;
301 }
302 
303 /*
304  * Block for the dma_resv object to become idle, lock the buffer and clean up
305  * the resource and tt object.
306  */
ttm_bo_delayed_delete(struct work_struct * work)307 static void ttm_bo_delayed_delete(struct work_struct *work)
308 {
309 	struct ttm_buffer_object *bo;
310 
311 	bo = container_of(work, typeof(*bo), delayed_delete);
312 
313 	dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP, false,
314 			      MAX_SCHEDULE_TIMEOUT);
315 	dma_resv_lock(bo->base.resv, NULL);
316 	ttm_bo_cleanup_memtype_use(bo);
317 	dma_resv_unlock(bo->base.resv);
318 	ttm_bo_put(bo);
319 }
320 
ttm_bo_release(struct kref * kref)321 static void ttm_bo_release(struct kref *kref)
322 {
323 	struct ttm_buffer_object *bo =
324 	    container_of(kref, struct ttm_buffer_object, kref);
325 	struct ttm_device *bdev = bo->bdev;
326 	int ret;
327 
328 	WARN_ON_ONCE(bo->pin_count);
329 	WARN_ON_ONCE(bo->bulk_move);
330 
331 	if (!bo->deleted) {
332 		ret = ttm_bo_individualize_resv(bo);
333 		if (ret) {
334 			/* Last resort, if we fail to allocate memory for the
335 			 * fences block for the BO to become idle
336 			 */
337 			dma_resv_wait_timeout(bo->base.resv,
338 					      DMA_RESV_USAGE_BOOKKEEP, false,
339 					      30 * HZ);
340 		}
341 
342 		if (bo->bdev->funcs->release_notify)
343 			bo->bdev->funcs->release_notify(bo);
344 
345 		drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
346 		ttm_mem_io_free(bdev, bo->resource);
347 
348 		if (!dma_resv_test_signaled(bo->base.resv,
349 					    DMA_RESV_USAGE_BOOKKEEP) ||
350 		    !dma_resv_trylock(bo->base.resv)) {
351 			/* The BO is not idle, resurrect it for delayed destroy */
352 			ttm_bo_flush_all_fences(bo);
353 			bo->deleted = true;
354 
355 			spin_lock(&bo->bdev->lru_lock);
356 
357 			/*
358 			 * Make pinned bos immediately available to
359 			 * shrinkers, now that they are queued for
360 			 * destruction.
361 			 *
362 			 * FIXME: QXL is triggering this. Can be removed when the
363 			 * driver is fixed.
364 			 */
365 			if (bo->pin_count) {
366 				bo->pin_count = 0;
367 				ttm_resource_move_to_lru_tail(bo->resource);
368 			}
369 
370 			kref_init(&bo->kref);
371 			spin_unlock(&bo->bdev->lru_lock);
372 
373 			INIT_WORK(&bo->delayed_delete, ttm_bo_delayed_delete);
374 			queue_work(bdev->wq, &bo->delayed_delete);
375 			return;
376 		}
377 
378 		ttm_bo_cleanup_memtype_use(bo);
379 		dma_resv_unlock(bo->base.resv);
380 	}
381 
382 	atomic_dec(&ttm_glob.bo_count);
383 	bo->destroy(bo);
384 }
385 
386 /**
387  * ttm_bo_put
388  *
389  * @bo: The buffer object.
390  *
391  * Unreference a buffer object.
392  */
ttm_bo_put(struct ttm_buffer_object * bo)393 void ttm_bo_put(struct ttm_buffer_object *bo)
394 {
395 	kref_put(&bo->kref, ttm_bo_release);
396 }
397 EXPORT_SYMBOL(ttm_bo_put);
398 
ttm_bo_bounce_temp_buffer(struct ttm_buffer_object * bo,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx,struct ttm_place * hop)399 static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
400 				     struct ttm_resource **mem,
401 				     struct ttm_operation_ctx *ctx,
402 				     struct ttm_place *hop)
403 {
404 	struct ttm_placement hop_placement;
405 	struct ttm_resource *hop_mem;
406 	int ret;
407 
408 	hop_placement.num_placement = hop_placement.num_busy_placement = 1;
409 	hop_placement.placement = hop_placement.busy_placement = hop;
410 
411 	/* find space in the bounce domain */
412 	ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
413 	if (ret)
414 		return ret;
415 	/* move to the bounce domain */
416 	ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
417 	if (ret) {
418 		ttm_resource_free(bo, &hop_mem);
419 		return ret;
420 	}
421 	return 0;
422 }
423 
ttm_bo_evict(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)424 static int ttm_bo_evict(struct ttm_buffer_object *bo,
425 			struct ttm_operation_ctx *ctx)
426 {
427 	struct ttm_device *bdev = bo->bdev;
428 	struct ttm_resource *evict_mem;
429 	struct ttm_placement placement;
430 	struct ttm_place hop;
431 	int ret = 0;
432 
433 	memset(&hop, 0, sizeof(hop));
434 
435 	dma_resv_assert_held(bo->base.resv);
436 
437 	placement.num_placement = 0;
438 	placement.num_busy_placement = 0;
439 	bdev->funcs->evict_flags(bo, &placement);
440 
441 	if (!placement.num_placement && !placement.num_busy_placement) {
442 		ret = ttm_bo_wait_ctx(bo, ctx);
443 		if (ret)
444 			return ret;
445 
446 		/*
447 		 * Since we've already synced, this frees backing store
448 		 * immediately.
449 		 */
450 		return ttm_bo_pipeline_gutting(bo);
451 	}
452 
453 	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
454 	if (ret) {
455 		if (ret != -ERESTARTSYS) {
456 			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
457 			       bo);
458 			ttm_bo_mem_space_debug(bo, &placement);
459 		}
460 		goto out;
461 	}
462 
463 bounce:
464 	ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
465 	if (ret == -EMULTIHOP) {
466 		ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
467 		if (ret) {
468 			pr_err("Buffer eviction failed\n");
469 			ttm_resource_free(bo, &evict_mem);
470 			goto out;
471 		}
472 		/* try and move to final place now. */
473 		goto bounce;
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->base.resv == ctx->resv) {
522 		dma_resv_assert_held(bo->base.resv);
523 		if (ctx->allow_res_evict)
524 			ret = true;
525 		*locked = false;
526 		if (busy)
527 			*busy = false;
528 	} else {
529 		ret = dma_resv_trylock(bo->base.resv);
530 		*locked = ret;
531 		if (busy)
532 			*busy = !ret;
533 	}
534 
535 	if (ret && place && (bo->resource->mem_type != place->mem_type ||
536 		!bo->bdev->funcs->eviction_valuable(bo, place))) {
537 		ret = false;
538 		if (*locked) {
539 			dma_resv_unlock(bo->base.resv);
540 			*locked = false;
541 		}
542 	}
543 
544 	return ret;
545 }
546 
547 /**
548  * ttm_mem_evict_wait_busy - wait for a busy BO to become available
549  *
550  * @busy_bo: BO which couldn't be locked with trylock
551  * @ctx: operation context
552  * @ticket: acquire ticket
553  *
554  * Try to lock a busy buffer object to avoid failing eviction.
555  */
ttm_mem_evict_wait_busy(struct ttm_buffer_object * busy_bo,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)556 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
557 				   struct ttm_operation_ctx *ctx,
558 				   struct ww_acquire_ctx *ticket)
559 {
560 	int r;
561 
562 	if (!busy_bo || !ticket)
563 		return -EBUSY;
564 
565 	if (ctx->interruptible)
566 		r = dma_resv_lock_interruptible(busy_bo->base.resv,
567 							  ticket);
568 	else
569 		r = dma_resv_lock(busy_bo->base.resv, ticket);
570 
571 	/*
572 	 * TODO: It would be better to keep the BO locked until allocation is at
573 	 * least tried one more time, but that would mean a much larger rework
574 	 * of TTM.
575 	 */
576 	if (!r)
577 		dma_resv_unlock(busy_bo->base.resv);
578 
579 	return r == -EDEADLK ? -EBUSY : r;
580 }
581 
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)582 int ttm_mem_evict_first(struct ttm_device *bdev,
583 			struct ttm_resource_manager *man,
584 			const struct ttm_place *place,
585 			struct ttm_operation_ctx *ctx,
586 			struct ww_acquire_ctx *ticket)
587 {
588 	struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
589 	struct ttm_resource_cursor cursor;
590 	struct ttm_resource *res;
591 	bool locked = false;
592 	int ret;
593 
594 	spin_lock(&bdev->lru_lock);
595 	ttm_resource_manager_for_each_res(man, &cursor, res) {
596 		bool busy;
597 
598 		if (!ttm_bo_evict_swapout_allowable(res->bo, ctx, place,
599 						    &locked, &busy)) {
600 			if (busy && !busy_bo && ticket !=
601 			    dma_resv_locking_ctx(res->bo->base.resv))
602 				busy_bo = res->bo;
603 			continue;
604 		}
605 
606 		if (ttm_bo_get_unless_zero(res->bo)) {
607 			bo = res->bo;
608 			break;
609 		}
610 		if (locked)
611 			dma_resv_unlock(res->bo->base.resv);
612 	}
613 
614 	if (!bo) {
615 		if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
616 			busy_bo = NULL;
617 		spin_unlock(&bdev->lru_lock);
618 		ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
619 		if (busy_bo)
620 			ttm_bo_put(busy_bo);
621 		return ret;
622 	}
623 
624 	if (bo->deleted) {
625 		ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
626 					  ctx->no_wait_gpu, locked);
627 		ttm_bo_put(bo);
628 		return ret;
629 	}
630 
631 	spin_unlock(&bdev->lru_lock);
632 
633 	ret = ttm_bo_evict(bo, ctx);
634 	if (locked)
635 		ttm_bo_unreserve(bo);
636 	else
637 		ttm_bo_move_to_lru_tail_unlocked(bo);
638 
639 	ttm_bo_put(bo);
640 	return ret;
641 }
642 
643 /**
644  * ttm_bo_pin - Pin the buffer object.
645  * @bo: The buffer object to pin
646  *
647  * Make sure the buffer is not evicted any more during memory pressure.
648  * @bo must be unpinned again by calling ttm_bo_unpin().
649  */
ttm_bo_pin(struct ttm_buffer_object * bo)650 void ttm_bo_pin(struct ttm_buffer_object *bo)
651 {
652 	dma_resv_assert_held(bo->base.resv);
653 	WARN_ON_ONCE(!kref_read(&bo->kref));
654 	spin_lock(&bo->bdev->lru_lock);
655 	if (bo->resource)
656 		ttm_resource_del_bulk_move(bo->resource, bo);
657 	++bo->pin_count;
658 	spin_unlock(&bo->bdev->lru_lock);
659 }
660 EXPORT_SYMBOL(ttm_bo_pin);
661 
662 /**
663  * ttm_bo_unpin - Unpin the buffer object.
664  * @bo: The buffer object to unpin
665  *
666  * Allows the buffer object to be evicted again during memory pressure.
667  */
ttm_bo_unpin(struct ttm_buffer_object * bo)668 void ttm_bo_unpin(struct ttm_buffer_object *bo)
669 {
670 	dma_resv_assert_held(bo->base.resv);
671 	WARN_ON_ONCE(!kref_read(&bo->kref));
672 	if (WARN_ON_ONCE(!bo->pin_count))
673 		return;
674 
675 	spin_lock(&bo->bdev->lru_lock);
676 	--bo->pin_count;
677 	if (bo->resource)
678 		ttm_resource_add_bulk_move(bo->resource, bo);
679 	spin_unlock(&bo->bdev->lru_lock);
680 }
681 EXPORT_SYMBOL(ttm_bo_unpin);
682 
683 /*
684  * Add the last move fence to the BO as kernel dependency and reserve a new
685  * fence slot.
686  */
ttm_bo_add_move_fence(struct ttm_buffer_object * bo,struct ttm_resource_manager * man,struct ttm_resource * mem,bool no_wait_gpu)687 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
688 				 struct ttm_resource_manager *man,
689 				 struct ttm_resource *mem,
690 				 bool no_wait_gpu)
691 {
692 	struct dma_fence *fence;
693 	int ret;
694 
695 	spin_lock(&man->move_lock);
696 	fence = dma_fence_get(man->move);
697 	spin_unlock(&man->move_lock);
698 
699 	if (!fence)
700 		return 0;
701 
702 	if (no_wait_gpu) {
703 		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
704 		dma_fence_put(fence);
705 		return ret;
706 	}
707 
708 	dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
709 
710 	ret = dma_resv_reserve_fences(bo->base.resv, 1);
711 	dma_fence_put(fence);
712 	return ret;
713 }
714 
715 /*
716  * Repeatedly evict memory from the LRU for @mem_type until we create enough
717  * space, or we've evicted everything and there isn't enough space.
718  */
ttm_bo_mem_force_space(struct ttm_buffer_object * bo,const struct ttm_place * place,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx)719 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
720 				  const struct ttm_place *place,
721 				  struct ttm_resource **mem,
722 				  struct ttm_operation_ctx *ctx)
723 {
724 	struct ttm_device *bdev = bo->bdev;
725 	struct ttm_resource_manager *man;
726 	struct ww_acquire_ctx *ticket;
727 	int ret;
728 
729 	man = ttm_manager_type(bdev, place->mem_type);
730 	ticket = dma_resv_locking_ctx(bo->base.resv);
731 	do {
732 		ret = ttm_resource_alloc(bo, place, mem);
733 		if (likely(!ret))
734 			break;
735 		if (unlikely(ret != -ENOSPC))
736 			return ret;
737 		ret = ttm_mem_evict_first(bdev, man, place, ctx,
738 					  ticket);
739 		if (unlikely(ret != 0))
740 			return ret;
741 	} while (1);
742 
743 	return ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
744 }
745 
746 /**
747  * ttm_bo_mem_space
748  *
749  * @bo: Pointer to a struct ttm_buffer_object. the data of which
750  * we want to allocate space for.
751  * @proposed_placement: Proposed new placement for the buffer object.
752  * @mem: A struct ttm_resource.
753  * @ctx: if and how to sleep, lock buffers and alloc memory
754  *
755  * Allocate memory space for the buffer object pointed to by @bo, using
756  * the placement flags in @placement, potentially evicting other idle buffer objects.
757  * This function may sleep while waiting for space to become available.
758  * Returns:
759  * -EBUSY: No space available (only if no_wait == 1).
760  * -ENOMEM: Could not allocate memory for the buffer object, either due to
761  * fragmentation or concurrent allocators.
762  * -ERESTARTSYS: An interruptible sleep was interrupted by a signal.
763  */
ttm_bo_mem_space(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_resource ** mem,struct ttm_operation_ctx * ctx)764 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
765 			struct ttm_placement *placement,
766 			struct ttm_resource **mem,
767 			struct ttm_operation_ctx *ctx)
768 {
769 	struct ttm_device *bdev = bo->bdev;
770 	bool type_found = false;
771 	int i, ret;
772 
773 	ret = dma_resv_reserve_fences(bo->base.resv, 1);
774 	if (unlikely(ret))
775 		return ret;
776 
777 	for (i = 0; i < placement->num_placement; ++i) {
778 		const struct ttm_place *place = &placement->placement[i];
779 		struct ttm_resource_manager *man;
780 
781 		man = ttm_manager_type(bdev, place->mem_type);
782 		if (!man || !ttm_resource_manager_used(man))
783 			continue;
784 
785 		type_found = true;
786 		ret = ttm_resource_alloc(bo, place, mem);
787 		if (ret == -ENOSPC)
788 			continue;
789 		if (unlikely(ret))
790 			goto error;
791 
792 		ret = ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
793 		if (unlikely(ret)) {
794 			ttm_resource_free(bo, mem);
795 			if (ret == -EBUSY)
796 				continue;
797 
798 			goto error;
799 		}
800 		return 0;
801 	}
802 
803 	for (i = 0; i < placement->num_busy_placement; ++i) {
804 		const struct ttm_place *place = &placement->busy_placement[i];
805 		struct ttm_resource_manager *man;
806 
807 		man = ttm_manager_type(bdev, place->mem_type);
808 		if (!man || !ttm_resource_manager_used(man))
809 			continue;
810 
811 		type_found = true;
812 		ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
813 		if (likely(!ret))
814 			return 0;
815 
816 		if (ret && ret != -EBUSY)
817 			goto error;
818 	}
819 
820 	ret = -ENOMEM;
821 	if (!type_found) {
822 		pr_err(TTM_PFX "No compatible memory type found\n");
823 		ret = -EINVAL;
824 	}
825 
826 error:
827 	return ret;
828 }
829 EXPORT_SYMBOL(ttm_bo_mem_space);
830 
ttm_bo_move_buffer(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)831 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
832 			      struct ttm_placement *placement,
833 			      struct ttm_operation_ctx *ctx)
834 {
835 	struct ttm_resource *mem;
836 	struct ttm_place hop;
837 	int ret;
838 
839 	dma_resv_assert_held(bo->base.resv);
840 
841 	/*
842 	 * Determine where to move the buffer.
843 	 *
844 	 * If driver determines move is going to need
845 	 * an extra step then it will return -EMULTIHOP
846 	 * and the buffer will be moved to the temporary
847 	 * stop and the driver will be called to make
848 	 * the second hop.
849 	 */
850 	ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
851 	if (ret)
852 		return ret;
853 bounce:
854 	ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
855 	if (ret == -EMULTIHOP) {
856 		ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
857 		if (ret)
858 			goto out;
859 		/* try and move to final place now. */
860 		goto bounce;
861 	}
862 out:
863 	if (ret)
864 		ttm_resource_free(bo, &mem);
865 	return ret;
866 }
867 
868 /**
869  * ttm_bo_validate
870  *
871  * @bo: The buffer object.
872  * @placement: Proposed placement for the buffer object.
873  * @ctx: validation parameters.
874  *
875  * Changes placement and caching policy of the buffer object
876  * according proposed placement.
877  * Returns
878  * -EINVAL on invalid proposed placement.
879  * -ENOMEM on out-of-memory condition.
880  * -EBUSY if no_wait is true and buffer busy.
881  * -ERESTARTSYS if interrupted by a signal.
882  */
ttm_bo_validate(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)883 int ttm_bo_validate(struct ttm_buffer_object *bo,
884 		    struct ttm_placement *placement,
885 		    struct ttm_operation_ctx *ctx)
886 {
887 	int ret;
888 
889 	dma_resv_assert_held(bo->base.resv);
890 
891 	/*
892 	 * Remove the backing store if no placement is given.
893 	 */
894 	if (!placement->num_placement && !placement->num_busy_placement)
895 		return ttm_bo_pipeline_gutting(bo);
896 
897 	/*
898 	 * Check whether we need to move buffer.
899 	 */
900 	if (!bo->resource || !ttm_resource_compat(bo->resource, placement)) {
901 		ret = ttm_bo_move_buffer(bo, placement, ctx);
902 		if (ret)
903 			return ret;
904 	}
905 	/*
906 	 * We might need to add a TTM.
907 	 */
908 	if (!bo->resource || bo->resource->mem_type == TTM_PL_SYSTEM) {
909 		ret = ttm_tt_create(bo, true);
910 		if (ret)
911 			return ret;
912 	}
913 	return 0;
914 }
915 EXPORT_SYMBOL(ttm_bo_validate);
916 
917 /**
918  * ttm_bo_init_reserved
919  *
920  * @bdev: Pointer to a ttm_device struct.
921  * @bo: Pointer to a ttm_buffer_object to be initialized.
922  * @type: Requested type of buffer object.
923  * @placement: Initial placement for buffer object.
924  * @alignment: Data alignment in pages.
925  * @ctx: TTM operation context for memory allocation.
926  * @sg: Scatter-gather table.
927  * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
928  * @destroy: Destroy function. Use NULL for kfree().
929  *
930  * This function initializes a pre-allocated struct ttm_buffer_object.
931  * As this object may be part of a larger structure, this function,
932  * together with the @destroy function, enables driver-specific objects
933  * derived from a ttm_buffer_object.
934  *
935  * On successful return, the caller owns an object kref to @bo. The kref and
936  * list_kref are usually set to 1, but note that in some situations, other
937  * tasks may already be holding references to @bo as well.
938  * Furthermore, if resv == NULL, the buffer's reservation lock will be held,
939  * and it is the caller's responsibility to call ttm_bo_unreserve.
940  *
941  * If a failure occurs, the function will call the @destroy function. Thus,
942  * after a failure, dereferencing @bo is illegal and will likely cause memory
943  * corruption.
944  *
945  * Returns
946  * -ENOMEM: Out of memory.
947  * -EINVAL: Invalid placement flags.
948  * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
949  */
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 *))950 int ttm_bo_init_reserved(struct ttm_device *bdev, struct ttm_buffer_object *bo,
951 			 enum ttm_bo_type type, struct ttm_placement *placement,
952 			 uint32_t alignment, struct ttm_operation_ctx *ctx,
953 			 struct sg_table *sg, struct dma_resv *resv,
954 			 void (*destroy) (struct ttm_buffer_object *))
955 {
956 	static const struct ttm_place sys_mem = { .mem_type = TTM_PL_SYSTEM };
957 	int ret;
958 
959 	kref_init(&bo->kref);
960 	bo->bdev = bdev;
961 	bo->type = type;
962 	bo->page_alignment = alignment;
963 	bo->destroy = destroy;
964 	bo->pin_count = 0;
965 	bo->sg = sg;
966 	bo->bulk_move = NULL;
967 	if (resv)
968 		bo->base.resv = resv;
969 	else
970 		bo->base.resv = &bo->base._resv;
971 	atomic_inc(&ttm_glob.bo_count);
972 
973 	ret = ttm_resource_alloc(bo, &sys_mem, &bo->resource);
974 	if (unlikely(ret)) {
975 		ttm_bo_put(bo);
976 		return ret;
977 	}
978 
979 	/*
980 	 * For ttm_bo_type_device buffers, allocate
981 	 * address space from the device.
982 	 */
983 	if (bo->type == ttm_bo_type_device || bo->type == ttm_bo_type_sg) {
984 		ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
985 					 PFN_UP(bo->base.size));
986 		if (ret)
987 			goto err_put;
988 	}
989 
990 	/* passed reservation objects should already be locked,
991 	 * since otherwise lockdep will be angered in radeon.
992 	 */
993 	if (!resv)
994 		WARN_ON(!dma_resv_trylock(bo->base.resv));
995 	else
996 		dma_resv_assert_held(resv);
997 
998 	ret = ttm_bo_validate(bo, placement, ctx);
999 	if (unlikely(ret))
1000 		goto err_unlock;
1001 
1002 	return 0;
1003 
1004 err_unlock:
1005 	if (!resv)
1006 		dma_resv_unlock(bo->base.resv);
1007 
1008 err_put:
1009 	ttm_bo_put(bo);
1010 	return ret;
1011 }
1012 EXPORT_SYMBOL(ttm_bo_init_reserved);
1013 
1014 /**
1015  * ttm_bo_init_validate
1016  *
1017  * @bdev: Pointer to a ttm_device struct.
1018  * @bo: Pointer to a ttm_buffer_object to be initialized.
1019  * @type: Requested type of buffer object.
1020  * @placement: Initial placement for buffer object.
1021  * @alignment: Data alignment in pages.
1022  * @interruptible: If needing to sleep to wait for GPU resources,
1023  * sleep interruptible.
1024  * pinned in physical memory. If this behaviour is not desired, this member
1025  * holds a pointer to a persistent shmem object. Typically, this would
1026  * point to the shmem object backing a GEM object if TTM is used to back a
1027  * GEM user interface.
1028  * @sg: Scatter-gather table.
1029  * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
1030  * @destroy: Destroy function. Use NULL for kfree().
1031  *
1032  * This function initializes a pre-allocated struct ttm_buffer_object.
1033  * As this object may be part of a larger structure, this function,
1034  * together with the @destroy function,
1035  * enables driver-specific objects derived from a ttm_buffer_object.
1036  *
1037  * On successful return, the caller owns an object kref to @bo. The kref and
1038  * list_kref are usually set to 1, but note that in some situations, other
1039  * tasks may already be holding references to @bo as well.
1040  *
1041  * If a failure occurs, the function will call the @destroy function, Thus,
1042  * after a failure, dereferencing @bo is illegal and will likely cause memory
1043  * corruption.
1044  *
1045  * Returns
1046  * -ENOMEM: Out of memory.
1047  * -EINVAL: Invalid placement flags.
1048  * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
1049  */
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 *))1050 int ttm_bo_init_validate(struct ttm_device *bdev, struct ttm_buffer_object *bo,
1051 			 enum ttm_bo_type type, struct ttm_placement *placement,
1052 			 uint32_t alignment, bool interruptible,
1053 			 struct sg_table *sg, struct dma_resv *resv,
1054 			 void (*destroy) (struct ttm_buffer_object *))
1055 {
1056 	struct ttm_operation_ctx ctx = { interruptible, false };
1057 	int ret;
1058 
1059 	ret = ttm_bo_init_reserved(bdev, bo, type, placement, alignment, &ctx,
1060 				   sg, resv, destroy);
1061 	if (ret)
1062 		return ret;
1063 
1064 	if (!resv)
1065 		ttm_bo_unreserve(bo);
1066 
1067 	return 0;
1068 }
1069 EXPORT_SYMBOL(ttm_bo_init_validate);
1070 
1071 /*
1072  * buffer object vm functions.
1073  */
1074 
1075 /**
1076  * ttm_bo_unmap_virtual
1077  *
1078  * @bo: tear down the virtual mappings for this BO
1079  */
ttm_bo_unmap_virtual(struct ttm_buffer_object * bo)1080 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1081 {
1082 	struct ttm_device *bdev = bo->bdev;
1083 
1084 	drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1085 	ttm_mem_io_free(bdev, bo->resource);
1086 }
1087 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1088 
1089 /**
1090  * ttm_bo_wait_ctx - wait for buffer idle.
1091  *
1092  * @bo:  The buffer object.
1093  * @ctx: defines how to wait
1094  *
1095  * Waits for the buffer to be idle. Used timeout depends on the context.
1096  * Returns -EBUSY if wait timed outt, -ERESTARTSYS if interrupted by a signal or
1097  * zero on success.
1098  */
ttm_bo_wait_ctx(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)1099 int ttm_bo_wait_ctx(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx)
1100 {
1101 	long ret;
1102 
1103 	if (ctx->no_wait_gpu) {
1104 		if (dma_resv_test_signaled(bo->base.resv,
1105 					   DMA_RESV_USAGE_BOOKKEEP))
1106 			return 0;
1107 		else
1108 			return -EBUSY;
1109 	}
1110 
1111 	ret = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
1112 				    ctx->interruptible, 15 * HZ);
1113 	if (unlikely(ret < 0))
1114 		return ret;
1115 	if (unlikely(ret == 0))
1116 		return -EBUSY;
1117 	return 0;
1118 }
1119 EXPORT_SYMBOL(ttm_bo_wait_ctx);
1120 
ttm_bo_swapout(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,gfp_t gfp_flags)1121 int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1122 		   gfp_t gfp_flags)
1123 {
1124 	struct ttm_place place;
1125 	bool locked;
1126 	long ret;
1127 
1128 	/*
1129 	 * While the bo may already reside in SYSTEM placement, set
1130 	 * SYSTEM as new placement to cover also the move further below.
1131 	 * The driver may use the fact that we're moving from SYSTEM
1132 	 * as an indication that we're about to swap out.
1133 	 */
1134 	memset(&place, 0, sizeof(place));
1135 	place.mem_type = bo->resource->mem_type;
1136 	if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1137 		return -EBUSY;
1138 
1139 	if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1140 	    bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1141 	    bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED ||
1142 	    !ttm_bo_get_unless_zero(bo)) {
1143 		if (locked)
1144 			dma_resv_unlock(bo->base.resv);
1145 		return -EBUSY;
1146 	}
1147 
1148 	if (bo->deleted) {
1149 		ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1150 		ttm_bo_put(bo);
1151 		return ret == -EBUSY ? -ENOSPC : ret;
1152 	}
1153 
1154 	/* TODO: Cleanup the locking */
1155 	spin_unlock(&bo->bdev->lru_lock);
1156 
1157 	/*
1158 	 * Move to system cached
1159 	 */
1160 	if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1161 		struct ttm_operation_ctx ctx = { false, false };
1162 		struct ttm_resource *evict_mem;
1163 		struct ttm_place hop;
1164 
1165 		memset(&hop, 0, sizeof(hop));
1166 		place.mem_type = TTM_PL_SYSTEM;
1167 		ret = ttm_resource_alloc(bo, &place, &evict_mem);
1168 		if (unlikely(ret))
1169 			goto out;
1170 
1171 		ret = ttm_bo_handle_move_mem(bo, evict_mem, true, &ctx, &hop);
1172 		if (unlikely(ret != 0)) {
1173 			WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1174 			goto out;
1175 		}
1176 	}
1177 
1178 	/*
1179 	 * Make sure BO is idle.
1180 	 */
1181 	ret = ttm_bo_wait_ctx(bo, ctx);
1182 	if (unlikely(ret != 0))
1183 		goto out;
1184 
1185 	ttm_bo_unmap_virtual(bo);
1186 
1187 	/*
1188 	 * Swap out. Buffer will be swapped in again as soon as
1189 	 * anyone tries to access a ttm page.
1190 	 */
1191 	if (bo->bdev->funcs->swap_notify)
1192 		bo->bdev->funcs->swap_notify(bo);
1193 
1194 	if (ttm_tt_is_populated(bo->ttm))
1195 		ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1196 out:
1197 
1198 	/*
1199 	 * Unreserve without putting on LRU to avoid swapping out an
1200 	 * already swapped buffer.
1201 	 */
1202 	if (locked)
1203 		dma_resv_unlock(bo->base.resv);
1204 	ttm_bo_put(bo);
1205 	return ret == -EBUSY ? -ENOSPC : ret;
1206 }
1207 
ttm_bo_tt_destroy(struct ttm_buffer_object * bo)1208 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1209 {
1210 	if (bo->ttm == NULL)
1211 		return;
1212 
1213 	ttm_tt_unpopulate(bo->bdev, bo->ttm);
1214 	ttm_tt_destroy(bo->bdev, bo->ttm);
1215 	bo->ttm = NULL;
1216 }
1217