• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_module.h>
35 #include <drm/ttm/ttm_bo_driver.h>
36 #include <drm/ttm/ttm_placement.h>
37 #include <linux/jiffies.h>
38 #include <linux/slab.h>
39 #include <linux/sched.h>
40 #include <linux/mm.h>
41 #include <linux/file.h>
42 #include <linux/module.h>
43 #include <linux/atomic.h>
44 #include <linux/dma-resv.h>
45 
46 static void ttm_bo_global_kobj_release(struct kobject *kobj);
47 
48 /**
49  * ttm_global_mutex - protecting the global BO state
50  */
51 DEFINE_MUTEX(ttm_global_mutex);
52 unsigned ttm_bo_glob_use_count;
53 struct ttm_bo_global ttm_bo_glob;
54 EXPORT_SYMBOL(ttm_bo_glob);
55 
56 static struct attribute ttm_bo_count = {
57 	.name = "bo_count",
58 	.mode = S_IRUGO
59 };
60 
61 /* default destructor */
ttm_bo_default_destroy(struct ttm_buffer_object * bo)62 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
63 {
64 	kfree(bo);
65 }
66 
ttm_bo_mem_space_debug(struct ttm_buffer_object * bo,struct ttm_placement * placement)67 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
68 					struct ttm_placement *placement)
69 {
70 	struct drm_printer p = drm_debug_printer(TTM_PFX);
71 	struct ttm_resource_manager *man;
72 	int i, mem_type;
73 
74 	drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n",
75 		   bo, bo->mem.num_pages, bo->mem.size >> 10,
76 		   bo->mem.size >> 20);
77 	for (i = 0; i < placement->num_placement; i++) {
78 		mem_type = placement->placement[i].mem_type;
79 		drm_printf(&p, "  placement[%d]=0x%08X (%d)\n",
80 			   i, placement->placement[i].flags, mem_type);
81 		man = ttm_manager_type(bo->bdev, mem_type);
82 		ttm_resource_manager_debug(man, &p);
83 	}
84 }
85 
ttm_bo_global_show(struct kobject * kobj,struct attribute * attr,char * buffer)86 static ssize_t ttm_bo_global_show(struct kobject *kobj,
87 				  struct attribute *attr,
88 				  char *buffer)
89 {
90 	struct ttm_bo_global *glob =
91 		container_of(kobj, struct ttm_bo_global, kobj);
92 
93 	return snprintf(buffer, PAGE_SIZE, "%d\n",
94 				atomic_read(&glob->bo_count));
95 }
96 
97 static struct attribute *ttm_bo_global_attrs[] = {
98 	&ttm_bo_count,
99 	NULL
100 };
101 
102 static const struct sysfs_ops ttm_bo_global_ops = {
103 	.show = &ttm_bo_global_show
104 };
105 
106 static struct kobj_type ttm_bo_glob_kobj_type  = {
107 	.release = &ttm_bo_global_kobj_release,
108 	.sysfs_ops = &ttm_bo_global_ops,
109 	.default_attrs = ttm_bo_global_attrs
110 };
111 
ttm_bo_add_mem_to_lru(struct ttm_buffer_object * bo,struct ttm_resource * mem)112 static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo,
113 				  struct ttm_resource *mem)
114 {
115 	struct ttm_bo_device *bdev = bo->bdev;
116 	struct ttm_resource_manager *man;
117 
118 	if (!list_empty(&bo->lru))
119 		return;
120 
121 	if (mem->placement & TTM_PL_FLAG_NO_EVICT)
122 		return;
123 
124 	man = ttm_manager_type(bdev, mem->mem_type);
125 	list_add_tail(&bo->lru, &man->lru[bo->priority]);
126 
127 	if (man->use_tt && bo->ttm &&
128 	    !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG |
129 				     TTM_PAGE_FLAG_SWAPPED))) {
130 		list_add_tail(&bo->swap, &ttm_bo_glob.swap_lru[bo->priority]);
131 	}
132 }
133 
ttm_bo_del_from_lru(struct ttm_buffer_object * bo)134 static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
135 {
136 	struct ttm_bo_device *bdev = bo->bdev;
137 	bool notify = false;
138 
139 	if (!list_empty(&bo->swap)) {
140 		list_del_init(&bo->swap);
141 		notify = true;
142 	}
143 	if (!list_empty(&bo->lru)) {
144 		list_del_init(&bo->lru);
145 		notify = true;
146 	}
147 
148 	if (notify && bdev->driver->del_from_lru_notify)
149 		bdev->driver->del_from_lru_notify(bo);
150 }
151 
ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos * pos,struct ttm_buffer_object * bo)152 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
153 				     struct ttm_buffer_object *bo)
154 {
155 	if (!pos->first)
156 		pos->first = bo;
157 	pos->last = bo;
158 }
159 
ttm_bo_move_to_lru_tail(struct ttm_buffer_object * bo,struct ttm_lru_bulk_move * bulk)160 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
161 			     struct ttm_lru_bulk_move *bulk)
162 {
163 	dma_resv_assert_held(bo->base.resv);
164 
165 	ttm_bo_del_from_lru(bo);
166 	ttm_bo_add_mem_to_lru(bo, &bo->mem);
167 
168 	if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
169 		switch (bo->mem.mem_type) {
170 		case TTM_PL_TT:
171 			ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
172 			break;
173 
174 		case TTM_PL_VRAM:
175 			ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
176 			break;
177 		}
178 		if (bo->ttm && !(bo->ttm->page_flags &
179 				 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
180 			ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
181 	}
182 }
183 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
184 
ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move * bulk)185 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
186 {
187 	unsigned i;
188 
189 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
190 		struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
191 		struct ttm_resource_manager *man;
192 
193 		if (!pos->first)
194 			continue;
195 
196 		dma_resv_assert_held(pos->first->base.resv);
197 		dma_resv_assert_held(pos->last->base.resv);
198 
199 		man = ttm_manager_type(pos->first->bdev, TTM_PL_TT);
200 		list_bulk_move_tail(&man->lru[i], &pos->first->lru,
201 				    &pos->last->lru);
202 	}
203 
204 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
205 		struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
206 		struct ttm_resource_manager *man;
207 
208 		if (!pos->first)
209 			continue;
210 
211 		dma_resv_assert_held(pos->first->base.resv);
212 		dma_resv_assert_held(pos->last->base.resv);
213 
214 		man = ttm_manager_type(pos->first->bdev, TTM_PL_VRAM);
215 		list_bulk_move_tail(&man->lru[i], &pos->first->lru,
216 				    &pos->last->lru);
217 	}
218 
219 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
220 		struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
221 		struct list_head *lru;
222 
223 		if (!pos->first)
224 			continue;
225 
226 		dma_resv_assert_held(pos->first->base.resv);
227 		dma_resv_assert_held(pos->last->base.resv);
228 
229 		lru = &ttm_bo_glob.swap_lru[i];
230 		list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
231 	}
232 }
233 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
234 
ttm_bo_handle_move_mem(struct ttm_buffer_object * bo,struct ttm_resource * mem,bool evict,struct ttm_operation_ctx * ctx)235 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
236 				  struct ttm_resource *mem, bool evict,
237 				  struct ttm_operation_ctx *ctx)
238 {
239 	struct ttm_bo_device *bdev = bo->bdev;
240 	struct ttm_resource_manager *old_man = ttm_manager_type(bdev, bo->mem.mem_type);
241 	struct ttm_resource_manager *new_man = ttm_manager_type(bdev, mem->mem_type);
242 	int ret;
243 
244 	ttm_bo_unmap_virtual(bo);
245 
246 	/*
247 	 * Create and bind a ttm if required.
248 	 */
249 
250 	if (new_man->use_tt) {
251 		/* Zero init the new TTM structure if the old location should
252 		 * have used one as well.
253 		 */
254 		ret = ttm_tt_create(bo, old_man->use_tt);
255 		if (ret)
256 			goto out_err;
257 
258 		ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
259 		if (ret)
260 			goto out_err;
261 
262 		if (mem->mem_type != TTM_PL_SYSTEM) {
263 			ret = ttm_tt_populate(bdev, bo->ttm, ctx);
264 			if (ret)
265 				goto out_err;
266 
267 			ret = ttm_bo_tt_bind(bo, mem);
268 			if (ret)
269 				goto out_err;
270 		}
271 
272 		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
273 			if (bdev->driver->move_notify)
274 				bdev->driver->move_notify(bo, evict, mem);
275 			bo->mem = *mem;
276 			goto moved;
277 		}
278 	}
279 
280 	if (bdev->driver->move_notify)
281 		bdev->driver->move_notify(bo, evict, mem);
282 
283 	if (old_man->use_tt && new_man->use_tt)
284 		ret = ttm_bo_move_ttm(bo, ctx, mem);
285 	else if (bdev->driver->move)
286 		ret = bdev->driver->move(bo, evict, ctx, mem);
287 	else
288 		ret = ttm_bo_move_memcpy(bo, ctx, mem);
289 
290 	if (ret) {
291 		if (bdev->driver->move_notify) {
292 			swap(*mem, bo->mem);
293 			bdev->driver->move_notify(bo, false, mem);
294 			swap(*mem, bo->mem);
295 		}
296 
297 		goto out_err;
298 	}
299 
300 moved:
301 	ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
302 	return 0;
303 
304 out_err:
305 	new_man = ttm_manager_type(bdev, bo->mem.mem_type);
306 	if (!new_man->use_tt)
307 		ttm_bo_tt_destroy(bo);
308 
309 	return ret;
310 }
311 
312 /**
313  * Call bo::reserved.
314  * Will release GPU memory type usage on destruction.
315  * This is the place to put in driver specific hooks to release
316  * driver private resources.
317  * Will release the bo::reserved lock.
318  */
319 
ttm_bo_cleanup_memtype_use(struct ttm_buffer_object * bo)320 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
321 {
322 	if (bo->bdev->driver->move_notify)
323 		bo->bdev->driver->move_notify(bo, false, NULL);
324 
325 	ttm_bo_tt_destroy(bo);
326 	ttm_resource_free(bo, &bo->mem);
327 }
328 
ttm_bo_individualize_resv(struct ttm_buffer_object * bo)329 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
330 {
331 	int r;
332 
333 	if (bo->base.resv == &bo->base._resv)
334 		return 0;
335 
336 	BUG_ON(!dma_resv_trylock(&bo->base._resv));
337 
338 	r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
339 	dma_resv_unlock(&bo->base._resv);
340 	if (r)
341 		return r;
342 
343 	if (bo->type != ttm_bo_type_sg) {
344 		/* This works because the BO is about to be destroyed and nobody
345 		 * reference it any more. The only tricky case is the trylock on
346 		 * the resv object while holding the lru_lock.
347 		 */
348 		spin_lock(&ttm_bo_glob.lru_lock);
349 		bo->base.resv = &bo->base._resv;
350 		spin_unlock(&ttm_bo_glob.lru_lock);
351 	}
352 
353 	return r;
354 }
355 
ttm_bo_flush_all_fences(struct ttm_buffer_object * bo)356 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
357 {
358 	struct dma_resv *resv = &bo->base._resv;
359 	struct dma_resv_list *fobj;
360 	struct dma_fence *fence;
361 	int i;
362 
363 	rcu_read_lock();
364 	fobj = rcu_dereference(resv->fence);
365 	fence = rcu_dereference(resv->fence_excl);
366 	if (fence && !fence->ops->signaled)
367 		dma_fence_enable_sw_signaling(fence);
368 
369 	for (i = 0; fobj && i < fobj->shared_count; ++i) {
370 		fence = rcu_dereference(fobj->shared[i]);
371 
372 		if (!fence->ops->signaled)
373 			dma_fence_enable_sw_signaling(fence);
374 	}
375 	rcu_read_unlock();
376 }
377 
378 /**
379  * function ttm_bo_cleanup_refs
380  * If bo idle, remove from lru lists, and unref.
381  * If not idle, block if possible.
382  *
383  * Must be called with lru_lock and reservation held, this function
384  * will drop the lru lock and optionally the reservation lock before returning.
385  *
386  * @interruptible         Any sleeps should occur interruptibly.
387  * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
388  * @unlock_resv           Unlock the reservation lock as well.
389  */
390 
ttm_bo_cleanup_refs(struct ttm_buffer_object * bo,bool interruptible,bool no_wait_gpu,bool unlock_resv)391 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
392 			       bool interruptible, bool no_wait_gpu,
393 			       bool unlock_resv)
394 {
395 	struct dma_resv *resv = &bo->base._resv;
396 	int ret;
397 
398 	if (dma_resv_test_signaled_rcu(resv, true))
399 		ret = 0;
400 	else
401 		ret = -EBUSY;
402 
403 	if (ret && !no_wait_gpu) {
404 		long lret;
405 
406 		if (unlock_resv)
407 			dma_resv_unlock(bo->base.resv);
408 		spin_unlock(&ttm_bo_glob.lru_lock);
409 
410 		lret = dma_resv_wait_timeout_rcu(resv, true, interruptible,
411 						 30 * HZ);
412 
413 		if (lret < 0)
414 			return lret;
415 		else if (lret == 0)
416 			return -EBUSY;
417 
418 		spin_lock(&ttm_bo_glob.lru_lock);
419 		if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
420 			/*
421 			 * We raced, and lost, someone else holds the reservation now,
422 			 * and is probably busy in ttm_bo_cleanup_memtype_use.
423 			 *
424 			 * Even if it's not the case, because we finished waiting any
425 			 * delayed destruction would succeed, so just return success
426 			 * here.
427 			 */
428 			spin_unlock(&ttm_bo_glob.lru_lock);
429 			return 0;
430 		}
431 		ret = 0;
432 	}
433 
434 	if (ret || unlikely(list_empty(&bo->ddestroy))) {
435 		if (unlock_resv)
436 			dma_resv_unlock(bo->base.resv);
437 		spin_unlock(&ttm_bo_glob.lru_lock);
438 		return ret;
439 	}
440 
441 	ttm_bo_del_from_lru(bo);
442 	list_del_init(&bo->ddestroy);
443 	spin_unlock(&ttm_bo_glob.lru_lock);
444 	ttm_bo_cleanup_memtype_use(bo);
445 
446 	if (unlock_resv)
447 		dma_resv_unlock(bo->base.resv);
448 
449 	ttm_bo_put(bo);
450 
451 	return 0;
452 }
453 
454 /**
455  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
456  * encountered buffers.
457  */
ttm_bo_delayed_delete(struct ttm_bo_device * bdev,bool remove_all)458 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
459 {
460 	struct ttm_bo_global *glob = &ttm_bo_glob;
461 	struct list_head removed;
462 	bool empty;
463 
464 	INIT_LIST_HEAD(&removed);
465 
466 	spin_lock(&glob->lru_lock);
467 	while (!list_empty(&bdev->ddestroy)) {
468 		struct ttm_buffer_object *bo;
469 
470 		bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
471 				      ddestroy);
472 		list_move_tail(&bo->ddestroy, &removed);
473 		if (!ttm_bo_get_unless_zero(bo))
474 			continue;
475 
476 		if (remove_all || bo->base.resv != &bo->base._resv) {
477 			spin_unlock(&glob->lru_lock);
478 			dma_resv_lock(bo->base.resv, NULL);
479 
480 			spin_lock(&glob->lru_lock);
481 			ttm_bo_cleanup_refs(bo, false, !remove_all, true);
482 
483 		} else if (dma_resv_trylock(bo->base.resv)) {
484 			ttm_bo_cleanup_refs(bo, false, !remove_all, true);
485 		} else {
486 			spin_unlock(&glob->lru_lock);
487 		}
488 
489 		ttm_bo_put(bo);
490 		spin_lock(&glob->lru_lock);
491 	}
492 	list_splice_tail(&removed, &bdev->ddestroy);
493 	empty = list_empty(&bdev->ddestroy);
494 	spin_unlock(&glob->lru_lock);
495 
496 	return empty;
497 }
498 
ttm_bo_delayed_workqueue(struct work_struct * work)499 static void ttm_bo_delayed_workqueue(struct work_struct *work)
500 {
501 	struct ttm_bo_device *bdev =
502 	    container_of(work, struct ttm_bo_device, wq.work);
503 
504 	if (!ttm_bo_delayed_delete(bdev, false))
505 		schedule_delayed_work(&bdev->wq,
506 				      ((HZ / 100) < 1) ? 1 : HZ / 100);
507 }
508 
ttm_bo_release(struct kref * kref)509 static void ttm_bo_release(struct kref *kref)
510 {
511 	struct ttm_buffer_object *bo =
512 	    container_of(kref, struct ttm_buffer_object, kref);
513 	struct ttm_bo_device *bdev = bo->bdev;
514 	size_t acc_size = bo->acc_size;
515 	int ret;
516 
517 	if (!bo->deleted) {
518 		ret = ttm_bo_individualize_resv(bo);
519 		if (ret) {
520 			/* Last resort, if we fail to allocate memory for the
521 			 * fences block for the BO to become idle
522 			 */
523 			dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
524 						  30 * HZ);
525 		}
526 
527 		if (bo->bdev->driver->release_notify)
528 			bo->bdev->driver->release_notify(bo);
529 
530 		drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
531 		ttm_mem_io_free(bdev, &bo->mem);
532 	}
533 
534 	if (!dma_resv_test_signaled_rcu(bo->base.resv, true) ||
535 	    !dma_resv_trylock(bo->base.resv)) {
536 		/* The BO is not idle, resurrect it for delayed destroy */
537 		ttm_bo_flush_all_fences(bo);
538 		bo->deleted = true;
539 
540 		spin_lock(&ttm_bo_glob.lru_lock);
541 
542 		/*
543 		 * Make NO_EVICT bos immediately available to
544 		 * shrinkers, now that they are queued for
545 		 * destruction.
546 		 */
547 		if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
548 			bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
549 			ttm_bo_del_from_lru(bo);
550 			ttm_bo_add_mem_to_lru(bo, &bo->mem);
551 		}
552 
553 		kref_init(&bo->kref);
554 		list_add_tail(&bo->ddestroy, &bdev->ddestroy);
555 		spin_unlock(&ttm_bo_glob.lru_lock);
556 
557 		schedule_delayed_work(&bdev->wq,
558 				      ((HZ / 100) < 1) ? 1 : HZ / 100);
559 		return;
560 	}
561 
562 	spin_lock(&ttm_bo_glob.lru_lock);
563 	ttm_bo_del_from_lru(bo);
564 	list_del(&bo->ddestroy);
565 	spin_unlock(&ttm_bo_glob.lru_lock);
566 
567 	ttm_bo_cleanup_memtype_use(bo);
568 	dma_resv_unlock(bo->base.resv);
569 
570 	atomic_dec(&ttm_bo_glob.bo_count);
571 	dma_fence_put(bo->moving);
572 	if (!ttm_bo_uses_embedded_gem_object(bo))
573 		dma_resv_fini(&bo->base._resv);
574 	bo->destroy(bo);
575 	ttm_mem_global_free(&ttm_mem_glob, acc_size);
576 }
577 
ttm_bo_put(struct ttm_buffer_object * bo)578 void ttm_bo_put(struct ttm_buffer_object *bo)
579 {
580 	kref_put(&bo->kref, ttm_bo_release);
581 }
582 EXPORT_SYMBOL(ttm_bo_put);
583 
ttm_bo_lock_delayed_workqueue(struct ttm_bo_device * bdev)584 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
585 {
586 	return cancel_delayed_work_sync(&bdev->wq);
587 }
588 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
589 
ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device * bdev,int resched)590 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
591 {
592 	if (resched)
593 		schedule_delayed_work(&bdev->wq,
594 				      ((HZ / 100) < 1) ? 1 : HZ / 100);
595 }
596 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
597 
ttm_bo_evict(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx)598 static int ttm_bo_evict(struct ttm_buffer_object *bo,
599 			struct ttm_operation_ctx *ctx)
600 {
601 	struct ttm_bo_device *bdev = bo->bdev;
602 	struct ttm_resource evict_mem;
603 	struct ttm_placement placement;
604 	int ret = 0;
605 
606 	dma_resv_assert_held(bo->base.resv);
607 
608 	placement.num_placement = 0;
609 	placement.num_busy_placement = 0;
610 	bdev->driver->evict_flags(bo, &placement);
611 
612 	if (!placement.num_placement && !placement.num_busy_placement) {
613 		ttm_bo_wait(bo, false, false);
614 
615 		ttm_bo_cleanup_memtype_use(bo);
616 		return ttm_tt_create(bo, false);
617 	}
618 
619 	evict_mem = bo->mem;
620 	evict_mem.mm_node = NULL;
621 	evict_mem.bus.offset = 0;
622 	evict_mem.bus.addr = NULL;
623 
624 	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
625 	if (ret) {
626 		if (ret != -ERESTARTSYS) {
627 			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
628 			       bo);
629 			ttm_bo_mem_space_debug(bo, &placement);
630 		}
631 		goto out;
632 	}
633 
634 	ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
635 	if (unlikely(ret)) {
636 		if (ret != -ERESTARTSYS)
637 			pr_err("Buffer eviction failed\n");
638 		ttm_resource_free(bo, &evict_mem);
639 	}
640 out:
641 	return ret;
642 }
643 
ttm_bo_eviction_valuable(struct ttm_buffer_object * bo,const struct ttm_place * place)644 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
645 			      const struct ttm_place *place)
646 {
647 	/* Don't evict this BO if it's outside of the
648 	 * requested placement range
649 	 */
650 	if (place->fpfn >= (bo->mem.start + bo->mem.num_pages) ||
651 	    (place->lpfn && place->lpfn <= bo->mem.start))
652 		return false;
653 
654 	return true;
655 }
656 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
657 
658 /**
659  * Check the target bo is allowable to be evicted or swapout, including cases:
660  *
661  * a. if share same reservation object with ctx->resv, have assumption
662  * reservation objects should already be locked, so not lock again and
663  * return true directly when either the opreation allow_reserved_eviction
664  * or the target bo already is in delayed free list;
665  *
666  * b. Otherwise, trylock it.
667  */
ttm_bo_evict_swapout_allowable(struct ttm_buffer_object * bo,struct ttm_operation_ctx * ctx,bool * locked,bool * busy)668 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
669 			struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
670 {
671 	bool ret = false;
672 
673 	if (bo->base.resv == ctx->resv) {
674 		dma_resv_assert_held(bo->base.resv);
675 		if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT)
676 			ret = true;
677 		*locked = false;
678 		if (busy)
679 			*busy = false;
680 	} else {
681 		ret = dma_resv_trylock(bo->base.resv);
682 		*locked = ret;
683 		if (busy)
684 			*busy = !ret;
685 	}
686 
687 	return ret;
688 }
689 
690 /**
691  * ttm_mem_evict_wait_busy - wait for a busy BO to become available
692  *
693  * @busy_bo: BO which couldn't be locked with trylock
694  * @ctx: operation context
695  * @ticket: acquire ticket
696  *
697  * Try to lock a busy buffer object to avoid failing eviction.
698  */
ttm_mem_evict_wait_busy(struct ttm_buffer_object * busy_bo,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)699 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
700 				   struct ttm_operation_ctx *ctx,
701 				   struct ww_acquire_ctx *ticket)
702 {
703 	int r;
704 
705 	if (!busy_bo || !ticket)
706 		return -EBUSY;
707 
708 	if (ctx->interruptible)
709 		r = dma_resv_lock_interruptible(busy_bo->base.resv,
710 							  ticket);
711 	else
712 		r = dma_resv_lock(busy_bo->base.resv, ticket);
713 
714 	/*
715 	 * TODO: It would be better to keep the BO locked until allocation is at
716 	 * least tried one more time, but that would mean a much larger rework
717 	 * of TTM.
718 	 */
719 	if (!r)
720 		dma_resv_unlock(busy_bo->base.resv);
721 
722 	return r == -EDEADLK ? -EBUSY : r;
723 }
724 
ttm_mem_evict_first(struct ttm_bo_device * bdev,struct ttm_resource_manager * man,const struct ttm_place * place,struct ttm_operation_ctx * ctx,struct ww_acquire_ctx * ticket)725 int ttm_mem_evict_first(struct ttm_bo_device *bdev,
726 			struct ttm_resource_manager *man,
727 			const struct ttm_place *place,
728 			struct ttm_operation_ctx *ctx,
729 			struct ww_acquire_ctx *ticket)
730 {
731 	struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
732 	bool locked = false;
733 	unsigned i;
734 	int ret;
735 
736 	spin_lock(&ttm_bo_glob.lru_lock);
737 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
738 		list_for_each_entry(bo, &man->lru[i], lru) {
739 			bool busy;
740 
741 			if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
742 							    &busy)) {
743 				if (busy && !busy_bo && ticket !=
744 				    dma_resv_locking_ctx(bo->base.resv))
745 					busy_bo = bo;
746 				continue;
747 			}
748 
749 			if (place && !bdev->driver->eviction_valuable(bo,
750 								      place)) {
751 				if (locked)
752 					dma_resv_unlock(bo->base.resv);
753 				continue;
754 			}
755 			if (!ttm_bo_get_unless_zero(bo)) {
756 				if (locked)
757 					dma_resv_unlock(bo->base.resv);
758 				continue;
759 			}
760 			break;
761 		}
762 
763 		/* If the inner loop terminated early, we have our candidate */
764 		if (&bo->lru != &man->lru[i])
765 			break;
766 
767 		bo = NULL;
768 	}
769 
770 	if (!bo) {
771 		if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
772 			busy_bo = NULL;
773 		spin_unlock(&ttm_bo_glob.lru_lock);
774 		ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
775 		if (busy_bo)
776 			ttm_bo_put(busy_bo);
777 		return ret;
778 	}
779 
780 	if (bo->deleted) {
781 		ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
782 					  ctx->no_wait_gpu, locked);
783 		ttm_bo_put(bo);
784 		return ret;
785 	}
786 
787 	spin_unlock(&ttm_bo_glob.lru_lock);
788 
789 	ret = ttm_bo_evict(bo, ctx);
790 	if (locked)
791 		ttm_bo_unreserve(bo);
792 	else
793 		ttm_bo_move_to_lru_tail_unlocked(bo);
794 
795 	ttm_bo_put(bo);
796 	return ret;
797 }
798 
799 /**
800  * Add the last move fence to the BO and reserve a new shared slot.
801  */
ttm_bo_add_move_fence(struct ttm_buffer_object * bo,struct ttm_resource_manager * man,struct ttm_resource * mem,bool no_wait_gpu)802 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
803 				 struct ttm_resource_manager *man,
804 				 struct ttm_resource *mem,
805 				 bool no_wait_gpu)
806 {
807 	struct dma_fence *fence;
808 	int ret;
809 
810 	spin_lock(&man->move_lock);
811 	fence = dma_fence_get(man->move);
812 	spin_unlock(&man->move_lock);
813 
814 	if (!fence)
815 		return 0;
816 
817 	if (no_wait_gpu) {
818 		dma_fence_put(fence);
819 		return -EBUSY;
820 	}
821 
822 	dma_resv_add_shared_fence(bo->base.resv, fence);
823 
824 	ret = dma_resv_reserve_shared(bo->base.resv, 1);
825 	if (unlikely(ret)) {
826 		dma_fence_put(fence);
827 		return ret;
828 	}
829 
830 	dma_fence_put(bo->moving);
831 	bo->moving = fence;
832 	return 0;
833 }
834 
835 /**
836  * Repeatedly evict memory from the LRU for @mem_type until we create enough
837  * space, or we've evicted everything and there isn't enough space.
838  */
ttm_bo_mem_force_space(struct ttm_buffer_object * bo,const struct ttm_place * place,struct ttm_resource * mem,struct ttm_operation_ctx * ctx)839 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
840 				  const struct ttm_place *place,
841 				  struct ttm_resource *mem,
842 				  struct ttm_operation_ctx *ctx)
843 {
844 	struct ttm_bo_device *bdev = bo->bdev;
845 	struct ttm_resource_manager *man = ttm_manager_type(bdev, mem->mem_type);
846 	struct ww_acquire_ctx *ticket;
847 	int ret;
848 
849 	ticket = dma_resv_locking_ctx(bo->base.resv);
850 	do {
851 		ret = ttm_resource_alloc(bo, place, mem);
852 		if (likely(!ret))
853 			break;
854 		if (unlikely(ret != -ENOSPC))
855 			return ret;
856 		ret = ttm_mem_evict_first(bdev, man, place, ctx,
857 					  ticket);
858 		if (unlikely(ret != 0))
859 			return ret;
860 	} while (1);
861 
862 	return ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
863 }
864 
ttm_bo_select_caching(struct ttm_resource_manager * man,uint32_t cur_placement,uint32_t proposed_placement)865 static uint32_t ttm_bo_select_caching(struct ttm_resource_manager *man,
866 				      uint32_t cur_placement,
867 				      uint32_t proposed_placement)
868 {
869 	uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
870 	uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
871 
872 	/**
873 	 * Keep current caching if possible.
874 	 */
875 
876 	if ((cur_placement & caching) != 0)
877 		result |= (cur_placement & caching);
878 	else if ((TTM_PL_FLAG_CACHED & caching) != 0)
879 		result |= TTM_PL_FLAG_CACHED;
880 	else if ((TTM_PL_FLAG_WC & caching) != 0)
881 		result |= TTM_PL_FLAG_WC;
882 	else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
883 		result |= TTM_PL_FLAG_UNCACHED;
884 
885 	return result;
886 }
887 
888 /**
889  * ttm_bo_mem_placement - check if placement is compatible
890  * @bo: BO to find memory for
891  * @place: where to search
892  * @mem: the memory object to fill in
893  * @ctx: operation context
894  *
895  * Check if placement is compatible and fill in mem structure.
896  * Returns -EBUSY if placement won't work or negative error code.
897  * 0 when placement can be used.
898  */
ttm_bo_mem_placement(struct ttm_buffer_object * bo,const struct ttm_place * place,struct ttm_resource * mem,struct ttm_operation_ctx * ctx)899 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo,
900 				const struct ttm_place *place,
901 				struct ttm_resource *mem,
902 				struct ttm_operation_ctx *ctx)
903 {
904 	struct ttm_bo_device *bdev = bo->bdev;
905 	struct ttm_resource_manager *man;
906 	uint32_t cur_flags = 0;
907 
908 	man = ttm_manager_type(bdev, place->mem_type);
909 	if (!man || !ttm_resource_manager_used(man))
910 		return -EBUSY;
911 
912 	cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
913 					  place->flags);
914 	cur_flags |= place->flags & ~TTM_PL_MASK_CACHING;
915 
916 	mem->mem_type = place->mem_type;
917 	mem->placement = cur_flags;
918 
919 	spin_lock(&ttm_bo_glob.lru_lock);
920 	ttm_bo_del_from_lru(bo);
921 	ttm_bo_add_mem_to_lru(bo, mem);
922 	spin_unlock(&ttm_bo_glob.lru_lock);
923 
924 	return 0;
925 }
926 
927 /**
928  * Creates space for memory region @mem according to its type.
929  *
930  * This function first searches for free space in compatible memory types in
931  * the priority order defined by the driver.  If free space isn't found, then
932  * ttm_bo_mem_force_space is attempted in priority order to evict and find
933  * space.
934  */
ttm_bo_mem_space(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_resource * mem,struct ttm_operation_ctx * ctx)935 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
936 			struct ttm_placement *placement,
937 			struct ttm_resource *mem,
938 			struct ttm_operation_ctx *ctx)
939 {
940 	struct ttm_bo_device *bdev = bo->bdev;
941 	bool type_found = false;
942 	int i, ret;
943 
944 	ret = dma_resv_reserve_shared(bo->base.resv, 1);
945 	if (unlikely(ret))
946 		return ret;
947 
948 	for (i = 0; i < placement->num_placement; ++i) {
949 		const struct ttm_place *place = &placement->placement[i];
950 		struct ttm_resource_manager *man;
951 
952 		ret = ttm_bo_mem_placement(bo, place, mem, ctx);
953 		if (ret)
954 			continue;
955 
956 		type_found = true;
957 		ret = ttm_resource_alloc(bo, place, mem);
958 		if (ret == -ENOSPC)
959 			continue;
960 		if (unlikely(ret))
961 			goto error;
962 
963 		man = ttm_manager_type(bdev, mem->mem_type);
964 		ret = ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
965 		if (unlikely(ret)) {
966 			ttm_resource_free(bo, mem);
967 			if (ret == -EBUSY)
968 				continue;
969 
970 			goto error;
971 		}
972 		return 0;
973 	}
974 
975 	for (i = 0; i < placement->num_busy_placement; ++i) {
976 		const struct ttm_place *place = &placement->busy_placement[i];
977 
978 		ret = ttm_bo_mem_placement(bo, place, mem, ctx);
979 		if (ret)
980 			continue;
981 
982 		type_found = true;
983 		ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
984 		if (likely(!ret))
985 			return 0;
986 
987 		if (ret && ret != -EBUSY)
988 			goto error;
989 	}
990 
991 	ret = -ENOMEM;
992 	if (!type_found) {
993 		pr_err(TTM_PFX "No compatible memory type found\n");
994 		ret = -EINVAL;
995 	}
996 
997 error:
998 	if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) {
999 		ttm_bo_move_to_lru_tail_unlocked(bo);
1000 	}
1001 
1002 	return ret;
1003 }
1004 EXPORT_SYMBOL(ttm_bo_mem_space);
1005 
ttm_bo_move_buffer(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)1006 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1007 			      struct ttm_placement *placement,
1008 			      struct ttm_operation_ctx *ctx)
1009 {
1010 	int ret = 0;
1011 	struct ttm_resource mem;
1012 
1013 	dma_resv_assert_held(bo->base.resv);
1014 
1015 	mem.num_pages = bo->num_pages;
1016 	mem.size = mem.num_pages << PAGE_SHIFT;
1017 	mem.page_alignment = bo->mem.page_alignment;
1018 	mem.bus.offset = 0;
1019 	mem.bus.addr = NULL;
1020 	mem.mm_node = NULL;
1021 
1022 	/*
1023 	 * Determine where to move the buffer.
1024 	 */
1025 	ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1026 	if (ret)
1027 		goto out_unlock;
1028 	ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1029 out_unlock:
1030 	if (ret)
1031 		ttm_resource_free(bo, &mem);
1032 	return ret;
1033 }
1034 
ttm_bo_places_compat(const struct ttm_place * places,unsigned num_placement,struct ttm_resource * mem,uint32_t * new_flags)1035 static bool ttm_bo_places_compat(const struct ttm_place *places,
1036 				 unsigned num_placement,
1037 				 struct ttm_resource *mem,
1038 				 uint32_t *new_flags)
1039 {
1040 	unsigned i;
1041 
1042 	for (i = 0; i < num_placement; i++) {
1043 		const struct ttm_place *heap = &places[i];
1044 
1045 		if ((mem->start < heap->fpfn ||
1046 		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1047 			continue;
1048 
1049 		*new_flags = heap->flags;
1050 		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1051 		    (mem->mem_type == heap->mem_type) &&
1052 		    (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1053 		     (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1054 			return true;
1055 	}
1056 	return false;
1057 }
1058 
ttm_bo_mem_compat(struct ttm_placement * placement,struct ttm_resource * mem,uint32_t * new_flags)1059 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1060 		       struct ttm_resource *mem,
1061 		       uint32_t *new_flags)
1062 {
1063 	if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1064 				 mem, new_flags))
1065 		return true;
1066 
1067 	if ((placement->busy_placement != placement->placement ||
1068 	     placement->num_busy_placement > placement->num_placement) &&
1069 	    ttm_bo_places_compat(placement->busy_placement,
1070 				 placement->num_busy_placement,
1071 				 mem, new_flags))
1072 		return true;
1073 
1074 	return false;
1075 }
1076 EXPORT_SYMBOL(ttm_bo_mem_compat);
1077 
ttm_bo_validate(struct ttm_buffer_object * bo,struct ttm_placement * placement,struct ttm_operation_ctx * ctx)1078 int ttm_bo_validate(struct ttm_buffer_object *bo,
1079 		    struct ttm_placement *placement,
1080 		    struct ttm_operation_ctx *ctx)
1081 {
1082 	int ret;
1083 	uint32_t new_flags;
1084 
1085 	dma_resv_assert_held(bo->base.resv);
1086 
1087 	/*
1088 	 * Remove the backing store if no placement is given.
1089 	 */
1090 	if (!placement->num_placement && !placement->num_busy_placement) {
1091 		ret = ttm_bo_pipeline_gutting(bo);
1092 		if (ret)
1093 			return ret;
1094 
1095 		return ttm_tt_create(bo, false);
1096 	}
1097 
1098 	/*
1099 	 * Check whether we need to move buffer.
1100 	 */
1101 	if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1102 		ret = ttm_bo_move_buffer(bo, placement, ctx);
1103 		if (ret)
1104 			return ret;
1105 	} else {
1106 		bo->mem.placement &= TTM_PL_MASK_CACHING;
1107 		bo->mem.placement |= new_flags & ~TTM_PL_MASK_CACHING;
1108 	}
1109 	/*
1110 	 * We might need to add a TTM.
1111 	 */
1112 	if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1113 		ret = ttm_tt_create(bo, true);
1114 		if (ret)
1115 			return ret;
1116 	}
1117 	return 0;
1118 }
1119 EXPORT_SYMBOL(ttm_bo_validate);
1120 
ttm_bo_init_reserved(struct ttm_bo_device * bdev,struct ttm_buffer_object * bo,unsigned long size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,struct ttm_operation_ctx * ctx,size_t acc_size,struct sg_table * sg,struct dma_resv * resv,void (* destroy)(struct ttm_buffer_object *))1121 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1122 			 struct ttm_buffer_object *bo,
1123 			 unsigned long size,
1124 			 enum ttm_bo_type type,
1125 			 struct ttm_placement *placement,
1126 			 uint32_t page_alignment,
1127 			 struct ttm_operation_ctx *ctx,
1128 			 size_t acc_size,
1129 			 struct sg_table *sg,
1130 			 struct dma_resv *resv,
1131 			 void (*destroy) (struct ttm_buffer_object *))
1132 {
1133 	struct ttm_mem_global *mem_glob = &ttm_mem_glob;
1134 	int ret = 0;
1135 	unsigned long num_pages;
1136 	bool locked;
1137 
1138 	ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1139 	if (ret) {
1140 		pr_err("Out of kernel memory\n");
1141 		if (destroy)
1142 			(*destroy)(bo);
1143 		else
1144 			kfree(bo);
1145 		return -ENOMEM;
1146 	}
1147 
1148 	num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1149 	if (num_pages == 0) {
1150 		pr_err("Illegal buffer object size\n");
1151 		if (destroy)
1152 			(*destroy)(bo);
1153 		else
1154 			kfree(bo);
1155 		ttm_mem_global_free(mem_glob, acc_size);
1156 		return -EINVAL;
1157 	}
1158 	bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1159 
1160 	kref_init(&bo->kref);
1161 	INIT_LIST_HEAD(&bo->lru);
1162 	INIT_LIST_HEAD(&bo->ddestroy);
1163 	INIT_LIST_HEAD(&bo->swap);
1164 	bo->bdev = bdev;
1165 	bo->type = type;
1166 	bo->num_pages = num_pages;
1167 	bo->mem.size = num_pages << PAGE_SHIFT;
1168 	bo->mem.mem_type = TTM_PL_SYSTEM;
1169 	bo->mem.num_pages = bo->num_pages;
1170 	bo->mem.mm_node = NULL;
1171 	bo->mem.page_alignment = page_alignment;
1172 	bo->mem.bus.offset = 0;
1173 	bo->mem.bus.addr = NULL;
1174 	bo->moving = NULL;
1175 	bo->mem.placement = TTM_PL_FLAG_CACHED;
1176 	bo->acc_size = acc_size;
1177 	bo->sg = sg;
1178 	if (resv) {
1179 		bo->base.resv = resv;
1180 		dma_resv_assert_held(bo->base.resv);
1181 	} else {
1182 		bo->base.resv = &bo->base._resv;
1183 	}
1184 	if (!ttm_bo_uses_embedded_gem_object(bo)) {
1185 		/*
1186 		 * bo.gem is not initialized, so we have to setup the
1187 		 * struct elements we want use regardless.
1188 		 */
1189 		dma_resv_init(&bo->base._resv);
1190 		drm_vma_node_reset(&bo->base.vma_node);
1191 	}
1192 	atomic_inc(&ttm_bo_glob.bo_count);
1193 
1194 	/*
1195 	 * For ttm_bo_type_device buffers, allocate
1196 	 * address space from the device.
1197 	 */
1198 	if (bo->type == ttm_bo_type_device ||
1199 	    bo->type == ttm_bo_type_sg)
1200 		ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1201 					 bo->mem.num_pages);
1202 
1203 	/* passed reservation objects should already be locked,
1204 	 * since otherwise lockdep will be angered in radeon.
1205 	 */
1206 	if (!resv) {
1207 		locked = dma_resv_trylock(bo->base.resv);
1208 		WARN_ON(!locked);
1209 	}
1210 
1211 	if (likely(!ret))
1212 		ret = ttm_bo_validate(bo, placement, ctx);
1213 
1214 	if (unlikely(ret)) {
1215 		if (!resv)
1216 			ttm_bo_unreserve(bo);
1217 
1218 		ttm_bo_put(bo);
1219 		return ret;
1220 	}
1221 
1222 	ttm_bo_move_to_lru_tail_unlocked(bo);
1223 
1224 	return ret;
1225 }
1226 EXPORT_SYMBOL(ttm_bo_init_reserved);
1227 
ttm_bo_init(struct ttm_bo_device * bdev,struct ttm_buffer_object * bo,unsigned long size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,bool interruptible,size_t acc_size,struct sg_table * sg,struct dma_resv * resv,void (* destroy)(struct ttm_buffer_object *))1228 int ttm_bo_init(struct ttm_bo_device *bdev,
1229 		struct ttm_buffer_object *bo,
1230 		unsigned long size,
1231 		enum ttm_bo_type type,
1232 		struct ttm_placement *placement,
1233 		uint32_t page_alignment,
1234 		bool interruptible,
1235 		size_t acc_size,
1236 		struct sg_table *sg,
1237 		struct dma_resv *resv,
1238 		void (*destroy) (struct ttm_buffer_object *))
1239 {
1240 	struct ttm_operation_ctx ctx = { interruptible, false };
1241 	int ret;
1242 
1243 	ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1244 				   page_alignment, &ctx, acc_size,
1245 				   sg, resv, destroy);
1246 	if (ret)
1247 		return ret;
1248 
1249 	if (!resv)
1250 		ttm_bo_unreserve(bo);
1251 
1252 	return 0;
1253 }
1254 EXPORT_SYMBOL(ttm_bo_init);
1255 
ttm_bo_acc_size(struct ttm_bo_device * bdev,unsigned long bo_size,unsigned struct_size)1256 static size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1257 			      unsigned long bo_size,
1258 			      unsigned struct_size)
1259 {
1260 	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1261 	size_t size = 0;
1262 
1263 	size += ttm_round_pot(struct_size);
1264 	size += ttm_round_pot(npages * sizeof(void *));
1265 	size += ttm_round_pot(sizeof(struct ttm_tt));
1266 	return size;
1267 }
1268 
ttm_bo_dma_acc_size(struct ttm_bo_device * bdev,unsigned long bo_size,unsigned struct_size)1269 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1270 			   unsigned long bo_size,
1271 			   unsigned struct_size)
1272 {
1273 	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1274 	size_t size = 0;
1275 
1276 	size += ttm_round_pot(struct_size);
1277 	size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1278 	size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1279 	return size;
1280 }
1281 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1282 
ttm_bo_create(struct ttm_bo_device * bdev,unsigned long size,enum ttm_bo_type type,struct ttm_placement * placement,uint32_t page_alignment,bool interruptible,struct ttm_buffer_object ** p_bo)1283 int ttm_bo_create(struct ttm_bo_device *bdev,
1284 			unsigned long size,
1285 			enum ttm_bo_type type,
1286 			struct ttm_placement *placement,
1287 			uint32_t page_alignment,
1288 			bool interruptible,
1289 			struct ttm_buffer_object **p_bo)
1290 {
1291 	struct ttm_buffer_object *bo;
1292 	size_t acc_size;
1293 	int ret;
1294 
1295 	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1296 	if (unlikely(bo == NULL))
1297 		return -ENOMEM;
1298 
1299 	acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1300 	ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1301 			  interruptible, acc_size,
1302 			  NULL, NULL, NULL);
1303 	if (likely(ret == 0))
1304 		*p_bo = bo;
1305 
1306 	return ret;
1307 }
1308 EXPORT_SYMBOL(ttm_bo_create);
1309 
ttm_bo_evict_mm(struct ttm_bo_device * bdev,unsigned mem_type)1310 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1311 {
1312 	struct ttm_resource_manager *man = ttm_manager_type(bdev, mem_type);
1313 
1314 	if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1315 		pr_err("Illegal memory manager memory type %u\n", mem_type);
1316 		return -EINVAL;
1317 	}
1318 
1319 	if (!man) {
1320 		pr_err("Memory type %u has not been initialized\n", mem_type);
1321 		return 0;
1322 	}
1323 
1324 	return ttm_resource_manager_force_list_clean(bdev, man);
1325 }
1326 EXPORT_SYMBOL(ttm_bo_evict_mm);
1327 
ttm_bo_global_kobj_release(struct kobject * kobj)1328 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1329 {
1330 	struct ttm_bo_global *glob =
1331 		container_of(kobj, struct ttm_bo_global, kobj);
1332 
1333 	__free_page(glob->dummy_read_page);
1334 }
1335 
ttm_bo_global_release(void)1336 static void ttm_bo_global_release(void)
1337 {
1338 	struct ttm_bo_global *glob = &ttm_bo_glob;
1339 
1340 	mutex_lock(&ttm_global_mutex);
1341 	if (--ttm_bo_glob_use_count > 0)
1342 		goto out;
1343 
1344 	kobject_del(&glob->kobj);
1345 	kobject_put(&glob->kobj);
1346 	ttm_mem_global_release(&ttm_mem_glob);
1347 	memset(glob, 0, sizeof(*glob));
1348 out:
1349 	mutex_unlock(&ttm_global_mutex);
1350 }
1351 
ttm_bo_global_init(void)1352 static int ttm_bo_global_init(void)
1353 {
1354 	struct ttm_bo_global *glob = &ttm_bo_glob;
1355 	int ret = 0;
1356 	unsigned i;
1357 
1358 	mutex_lock(&ttm_global_mutex);
1359 	if (++ttm_bo_glob_use_count > 1)
1360 		goto out;
1361 
1362 	ret = ttm_mem_global_init(&ttm_mem_glob);
1363 	if (ret)
1364 		goto out;
1365 
1366 	spin_lock_init(&glob->lru_lock);
1367 	glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1368 
1369 	if (unlikely(glob->dummy_read_page == NULL)) {
1370 		ret = -ENOMEM;
1371 		goto out;
1372 	}
1373 
1374 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1375 		INIT_LIST_HEAD(&glob->swap_lru[i]);
1376 	INIT_LIST_HEAD(&glob->device_list);
1377 	atomic_set(&glob->bo_count, 0);
1378 
1379 	ret = kobject_init_and_add(
1380 		&glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1381 	if (unlikely(ret != 0))
1382 		kobject_put(&glob->kobj);
1383 out:
1384 	mutex_unlock(&ttm_global_mutex);
1385 	return ret;
1386 }
1387 
ttm_bo_device_release(struct ttm_bo_device * bdev)1388 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1389 {
1390 	struct ttm_bo_global *glob = &ttm_bo_glob;
1391 	int ret = 0;
1392 	unsigned i;
1393 	struct ttm_resource_manager *man;
1394 
1395 	man = ttm_manager_type(bdev, TTM_PL_SYSTEM);
1396 	ttm_resource_manager_set_used(man, false);
1397 	ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, NULL);
1398 
1399 	mutex_lock(&ttm_global_mutex);
1400 	list_del(&bdev->device_list);
1401 	mutex_unlock(&ttm_global_mutex);
1402 
1403 	cancel_delayed_work_sync(&bdev->wq);
1404 
1405 	if (ttm_bo_delayed_delete(bdev, true))
1406 		pr_debug("Delayed destroy list was clean\n");
1407 
1408 	spin_lock(&glob->lru_lock);
1409 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1410 		if (list_empty(&man->lru[0]))
1411 			pr_debug("Swap list %d was clean\n", i);
1412 	spin_unlock(&glob->lru_lock);
1413 
1414 	if (!ret)
1415 		ttm_bo_global_release();
1416 
1417 	return ret;
1418 }
1419 EXPORT_SYMBOL(ttm_bo_device_release);
1420 
ttm_bo_init_sysman(struct ttm_bo_device * bdev)1421 static void ttm_bo_init_sysman(struct ttm_bo_device *bdev)
1422 {
1423 	struct ttm_resource_manager *man = &bdev->sysman;
1424 
1425 	/*
1426 	 * Initialize the system memory buffer type.
1427 	 * Other types need to be driver / IOCTL initialized.
1428 	 */
1429 	man->use_tt = true;
1430 
1431 	ttm_resource_manager_init(man, 0);
1432 	ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, man);
1433 	ttm_resource_manager_set_used(man, true);
1434 }
1435 
ttm_bo_device_init(struct ttm_bo_device * bdev,struct ttm_bo_driver * driver,struct address_space * mapping,struct drm_vma_offset_manager * vma_manager,bool need_dma32)1436 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1437 		       struct ttm_bo_driver *driver,
1438 		       struct address_space *mapping,
1439 		       struct drm_vma_offset_manager *vma_manager,
1440 		       bool need_dma32)
1441 {
1442 	struct ttm_bo_global *glob = &ttm_bo_glob;
1443 	int ret;
1444 
1445 	if (WARN_ON(vma_manager == NULL))
1446 		return -EINVAL;
1447 
1448 	ret = ttm_bo_global_init();
1449 	if (ret)
1450 		return ret;
1451 
1452 	bdev->driver = driver;
1453 
1454 	ttm_bo_init_sysman(bdev);
1455 
1456 	bdev->vma_manager = vma_manager;
1457 	INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1458 	INIT_LIST_HEAD(&bdev->ddestroy);
1459 	bdev->dev_mapping = mapping;
1460 	bdev->need_dma32 = need_dma32;
1461 	mutex_lock(&ttm_global_mutex);
1462 	list_add_tail(&bdev->device_list, &glob->device_list);
1463 	mutex_unlock(&ttm_global_mutex);
1464 
1465 	return 0;
1466 }
1467 EXPORT_SYMBOL(ttm_bo_device_init);
1468 
1469 /*
1470  * buffer object vm functions.
1471  */
1472 
ttm_bo_unmap_virtual(struct ttm_buffer_object * bo)1473 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1474 {
1475 	struct ttm_bo_device *bdev = bo->bdev;
1476 
1477 	drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1478 	ttm_mem_io_free(bdev, &bo->mem);
1479 }
1480 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1481 
ttm_bo_wait(struct ttm_buffer_object * bo,bool interruptible,bool no_wait)1482 int ttm_bo_wait(struct ttm_buffer_object *bo,
1483 		bool interruptible, bool no_wait)
1484 {
1485 	long timeout = 15 * HZ;
1486 
1487 	if (no_wait) {
1488 		if (dma_resv_test_signaled_rcu(bo->base.resv, true))
1489 			return 0;
1490 		else
1491 			return -EBUSY;
1492 	}
1493 
1494 	timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true,
1495 						      interruptible, timeout);
1496 	if (timeout < 0)
1497 		return timeout;
1498 
1499 	if (timeout == 0)
1500 		return -EBUSY;
1501 
1502 	dma_resv_add_excl_fence(bo->base.resv, NULL);
1503 	return 0;
1504 }
1505 EXPORT_SYMBOL(ttm_bo_wait);
1506 
1507 /**
1508  * A buffer object shrink method that tries to swap out the first
1509  * buffer object on the bo_global::swap_lru list.
1510  */
ttm_bo_swapout(struct ttm_bo_global * glob,struct ttm_operation_ctx * ctx)1511 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1512 {
1513 	struct ttm_buffer_object *bo;
1514 	int ret = -EBUSY;
1515 	bool locked;
1516 	unsigned i;
1517 
1518 	spin_lock(&glob->lru_lock);
1519 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1520 		list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1521 			if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
1522 							    NULL))
1523 				continue;
1524 
1525 			if (!ttm_bo_get_unless_zero(bo)) {
1526 				if (locked)
1527 					dma_resv_unlock(bo->base.resv);
1528 				continue;
1529 			}
1530 
1531 			ret = 0;
1532 			break;
1533 		}
1534 		if (!ret)
1535 			break;
1536 	}
1537 
1538 	if (ret) {
1539 		spin_unlock(&glob->lru_lock);
1540 		return ret;
1541 	}
1542 
1543 	if (bo->deleted) {
1544 		ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1545 		ttm_bo_put(bo);
1546 		return ret;
1547 	}
1548 
1549 	ttm_bo_del_from_lru(bo);
1550 	spin_unlock(&glob->lru_lock);
1551 
1552 	/**
1553 	 * Move to system cached
1554 	 */
1555 
1556 	if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1557 	    bo->ttm->caching_state != tt_cached) {
1558 		struct ttm_operation_ctx ctx = { false, false };
1559 		struct ttm_resource evict_mem;
1560 
1561 		evict_mem = bo->mem;
1562 		evict_mem.mm_node = NULL;
1563 		evict_mem.placement = TTM_PL_FLAG_CACHED;
1564 		evict_mem.mem_type = TTM_PL_SYSTEM;
1565 
1566 		ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1567 		if (unlikely(ret != 0))
1568 			goto out;
1569 	}
1570 
1571 	/**
1572 	 * Make sure BO is idle.
1573 	 */
1574 
1575 	ret = ttm_bo_wait(bo, false, false);
1576 	if (unlikely(ret != 0))
1577 		goto out;
1578 
1579 	ttm_bo_unmap_virtual(bo);
1580 
1581 	/**
1582 	 * Swap out. Buffer will be swapped in again as soon as
1583 	 * anyone tries to access a ttm page.
1584 	 */
1585 
1586 	if (bo->bdev->driver->swap_notify)
1587 		bo->bdev->driver->swap_notify(bo);
1588 
1589 	ret = ttm_tt_swapout(bo->bdev, bo->ttm, bo->persistent_swap_storage);
1590 out:
1591 
1592 	/**
1593 	 *
1594 	 * Unreserve without putting on LRU to avoid swapping out an
1595 	 * already swapped buffer.
1596 	 */
1597 	if (locked)
1598 		dma_resv_unlock(bo->base.resv);
1599 	ttm_bo_put(bo);
1600 	return ret;
1601 }
1602 EXPORT_SYMBOL(ttm_bo_swapout);
1603 
ttm_bo_swapout_all(void)1604 void ttm_bo_swapout_all(void)
1605 {
1606 	struct ttm_operation_ctx ctx = {
1607 		.interruptible = false,
1608 		.no_wait_gpu = false
1609 	};
1610 
1611 	while (ttm_bo_swapout(&ttm_bo_glob, &ctx) == 0);
1612 }
1613 EXPORT_SYMBOL(ttm_bo_swapout_all);
1614 
ttm_bo_tt_destroy(struct ttm_buffer_object * bo)1615 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1616 {
1617 	if (bo->ttm == NULL)
1618 		return;
1619 
1620 	ttm_tt_destroy(bo->bdev, bo->ttm);
1621 	bo->ttm = NULL;
1622 }
1623 
ttm_bo_tt_bind(struct ttm_buffer_object * bo,struct ttm_resource * mem)1624 int ttm_bo_tt_bind(struct ttm_buffer_object *bo, struct ttm_resource *mem)
1625 {
1626 	return bo->bdev->driver->ttm_tt_bind(bo->bdev, bo->ttm, mem);
1627 }
1628 
ttm_bo_tt_unbind(struct ttm_buffer_object * bo)1629 void ttm_bo_tt_unbind(struct ttm_buffer_object *bo)
1630 {
1631 	bo->bdev->driver->ttm_tt_unbind(bo->bdev, bo->ttm);
1632 }
1633