1 /*
2 * Copyright (C) 2012-2014 Canonical Ltd (Maarten Lankhorst)
3 *
4 * Based on bo.c which bears the following copyright notice,
5 * but is dual licensed:
6 *
7 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
8 * All Rights Reserved.
9 *
10 * Permission is hereby granted, free of charge, to any person obtaining a
11 * copy of this software and associated documentation files (the
12 * "Software"), to deal in the Software without restriction, including
13 * without limitation the rights to use, copy, modify, merge, publish,
14 * distribute, sub license, and/or sell copies of the Software, and to
15 * permit persons to whom the Software is furnished to do so, subject to
16 * the following conditions:
17 *
18 * The above copyright notice and this permission notice (including the
19 * next paragraph) shall be included in all copies or substantial portions
20 * of the Software.
21 *
22 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
23 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
24 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
25 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
26 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
27 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
28 * USE OR OTHER DEALINGS IN THE SOFTWARE.
29 *
30 **************************************************************************/
31 /*
32 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
33 */
34
35 #include <linux/dma-resv.h>
36 #include <linux/export.h>
37 #include <linux/mm.h>
38 #include <linux/sched/mm.h>
39 #include <linux/mmu_notifier.h>
40
41 /**
42 * DOC: Reservation Object Overview
43 *
44 * The reservation object provides a mechanism to manage shared and
45 * exclusive fences associated with a buffer. A reservation object
46 * can have attached one exclusive fence (normally associated with
47 * write operations) or N shared fences (read operations). The RCU
48 * mechanism is used to protect read access to fences from locked
49 * write-side updates.
50 */
51
52 DEFINE_WD_CLASS(reservation_ww_class);
53 EXPORT_SYMBOL(reservation_ww_class);
54
55 /**
56 * dma_resv_list_alloc - allocate fence list
57 * @shared_max: number of fences we need space for
58 *
59 * Allocate a new dma_resv_list and make sure to correctly initialize
60 * shared_max.
61 */
dma_resv_list_alloc(unsigned int shared_max)62 static struct dma_resv_list *dma_resv_list_alloc(unsigned int shared_max)
63 {
64 struct dma_resv_list *list;
65
66 list = kmalloc(offsetof(typeof(*list), shared[shared_max]), GFP_KERNEL);
67 if (!list)
68 return NULL;
69
70 list->shared_max = (ksize(list) - offsetof(typeof(*list), shared)) /
71 sizeof(*list->shared);
72
73 return list;
74 }
75
76 /**
77 * dma_resv_list_free - free fence list
78 * @list: list to free
79 *
80 * Free a dma_resv_list and make sure to drop all references.
81 */
dma_resv_list_free(struct dma_resv_list * list)82 static void dma_resv_list_free(struct dma_resv_list *list)
83 {
84 unsigned int i;
85
86 if (!list)
87 return;
88
89 for (i = 0; i < list->shared_count; ++i)
90 dma_fence_put(rcu_dereference_protected(list->shared[i], true));
91
92 kfree_rcu(list, rcu);
93 }
94
95 #if IS_ENABLED(CONFIG_LOCKDEP)
dma_resv_lockdep(void)96 static int __init dma_resv_lockdep(void)
97 {
98 struct mm_struct *mm = mm_alloc();
99 struct ww_acquire_ctx ctx;
100 struct dma_resv obj;
101 struct address_space mapping;
102 int ret;
103
104 if (!mm)
105 return -ENOMEM;
106
107 dma_resv_init(&obj);
108 address_space_init_once(&mapping);
109
110 mmap_read_lock(mm);
111 ww_acquire_init(&ctx, &reservation_ww_class);
112 ret = dma_resv_lock(&obj, &ctx);
113 if (ret == -EDEADLK)
114 dma_resv_lock_slow(&obj, &ctx);
115 fs_reclaim_acquire(GFP_KERNEL);
116 /* for unmap_mapping_range on trylocked buffer objects in shrinkers */
117 i_mmap_lock_write(&mapping);
118 i_mmap_unlock_write(&mapping);
119 #ifdef CONFIG_MMU_NOTIFIER
120 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
121 __dma_fence_might_wait();
122 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
123 #else
124 __dma_fence_might_wait();
125 #endif
126 fs_reclaim_release(GFP_KERNEL);
127 ww_mutex_unlock(&obj.lock);
128 ww_acquire_fini(&ctx);
129 mmap_read_unlock(mm);
130
131 mmput(mm);
132
133 return 0;
134 }
135 subsys_initcall(dma_resv_lockdep);
136 #endif
137
138 /**
139 * dma_resv_init - initialize a reservation object
140 * @obj: the reservation object
141 */
dma_resv_init(struct dma_resv * obj)142 void dma_resv_init(struct dma_resv *obj)
143 {
144 ww_mutex_init(&obj->lock, &reservation_ww_class);
145 seqcount_ww_mutex_init(&obj->seq, &obj->lock);
146
147 RCU_INIT_POINTER(obj->fence, NULL);
148 RCU_INIT_POINTER(obj->fence_excl, NULL);
149 }
150 EXPORT_SYMBOL(dma_resv_init);
151
152 /**
153 * dma_resv_fini - destroys a reservation object
154 * @obj: the reservation object
155 */
dma_resv_fini(struct dma_resv * obj)156 void dma_resv_fini(struct dma_resv *obj)
157 {
158 struct dma_resv_list *fobj;
159 struct dma_fence *excl;
160
161 /*
162 * This object should be dead and all references must have
163 * been released to it, so no need to be protected with rcu.
164 */
165 excl = rcu_dereference_protected(obj->fence_excl, 1);
166 if (excl)
167 dma_fence_put(excl);
168
169 fobj = rcu_dereference_protected(obj->fence, 1);
170 dma_resv_list_free(fobj);
171 ww_mutex_destroy(&obj->lock);
172 }
173 EXPORT_SYMBOL(dma_resv_fini);
174
175 /**
176 * dma_resv_reserve_shared - Reserve space to add shared fences to
177 * a dma_resv.
178 * @obj: reservation object
179 * @num_fences: number of fences we want to add
180 *
181 * Should be called before dma_resv_add_shared_fence(). Must
182 * be called with obj->lock held.
183 *
184 * RETURNS
185 * Zero for success, or -errno
186 */
dma_resv_reserve_shared(struct dma_resv * obj,unsigned int num_fences)187 int dma_resv_reserve_shared(struct dma_resv *obj, unsigned int num_fences)
188 {
189 struct dma_resv_list *old, *new;
190 unsigned int i, j, k, max;
191
192 dma_resv_assert_held(obj);
193
194 old = dma_resv_get_list(obj);
195
196 if (old && old->shared_max) {
197 if ((old->shared_count + num_fences) <= old->shared_max)
198 return 0;
199 else
200 max = max(old->shared_count + num_fences,
201 old->shared_max * 2);
202 } else {
203 max = max(4ul, roundup_pow_of_two(num_fences));
204 }
205
206 new = dma_resv_list_alloc(max);
207 if (!new)
208 return -ENOMEM;
209
210 /*
211 * no need to bump fence refcounts, rcu_read access
212 * requires the use of kref_get_unless_zero, and the
213 * references from the old struct are carried over to
214 * the new.
215 */
216 for (i = 0, j = 0, k = max; i < (old ? old->shared_count : 0); ++i) {
217 struct dma_fence *fence;
218
219 fence = rcu_dereference_protected(old->shared[i],
220 dma_resv_held(obj));
221 if (dma_fence_is_signaled(fence))
222 RCU_INIT_POINTER(new->shared[--k], fence);
223 else
224 RCU_INIT_POINTER(new->shared[j++], fence);
225 }
226 new->shared_count = j;
227
228 /*
229 * We are not changing the effective set of fences here so can
230 * merely update the pointer to the new array; both existing
231 * readers and new readers will see exactly the same set of
232 * active (unsignaled) shared fences. Individual fences and the
233 * old array are protected by RCU and so will not vanish under
234 * the gaze of the rcu_read_lock() readers.
235 */
236 rcu_assign_pointer(obj->fence, new);
237
238 if (!old)
239 return 0;
240
241 /* Drop the references to the signaled fences */
242 for (i = k; i < max; ++i) {
243 struct dma_fence *fence;
244
245 fence = rcu_dereference_protected(new->shared[i],
246 dma_resv_held(obj));
247 dma_fence_put(fence);
248 }
249 kfree_rcu(old, rcu);
250
251 return 0;
252 }
253 EXPORT_SYMBOL(dma_resv_reserve_shared);
254
255 /**
256 * dma_resv_add_shared_fence - Add a fence to a shared slot
257 * @obj: the reservation object
258 * @fence: the shared fence to add
259 *
260 * Add a fence to a shared slot, obj->lock must be held, and
261 * dma_resv_reserve_shared() has been called.
262 */
dma_resv_add_shared_fence(struct dma_resv * obj,struct dma_fence * fence)263 void dma_resv_add_shared_fence(struct dma_resv *obj, struct dma_fence *fence)
264 {
265 struct dma_resv_list *fobj;
266 struct dma_fence *old;
267 unsigned int i, count;
268
269 dma_fence_get(fence);
270
271 dma_resv_assert_held(obj);
272
273 fobj = dma_resv_get_list(obj);
274 count = fobj->shared_count;
275
276 write_seqcount_begin(&obj->seq);
277
278 for (i = 0; i < count; ++i) {
279
280 old = rcu_dereference_protected(fobj->shared[i],
281 dma_resv_held(obj));
282 if (old->context == fence->context ||
283 dma_fence_is_signaled(old))
284 goto replace;
285 }
286
287 BUG_ON(fobj->shared_count >= fobj->shared_max);
288 old = NULL;
289 count++;
290
291 replace:
292 RCU_INIT_POINTER(fobj->shared[i], fence);
293 /* pointer update must be visible before we extend the shared_count */
294 smp_store_mb(fobj->shared_count, count);
295
296 write_seqcount_end(&obj->seq);
297 dma_fence_put(old);
298 }
299 EXPORT_SYMBOL(dma_resv_add_shared_fence);
300
301 /**
302 * dma_resv_add_excl_fence - Add an exclusive fence.
303 * @obj: the reservation object
304 * @fence: the shared fence to add
305 *
306 * Add a fence to the exclusive slot. The obj->lock must be held.
307 */
dma_resv_add_excl_fence(struct dma_resv * obj,struct dma_fence * fence)308 void dma_resv_add_excl_fence(struct dma_resv *obj, struct dma_fence *fence)
309 {
310 struct dma_fence *old_fence = dma_resv_get_excl(obj);
311 struct dma_resv_list *old;
312 u32 i = 0;
313
314 dma_resv_assert_held(obj);
315
316 old = dma_resv_get_list(obj);
317 if (old)
318 i = old->shared_count;
319
320 if (fence)
321 dma_fence_get(fence);
322
323 write_seqcount_begin(&obj->seq);
324 /* write_seqcount_begin provides the necessary memory barrier */
325 RCU_INIT_POINTER(obj->fence_excl, fence);
326 if (old)
327 old->shared_count = 0;
328 write_seqcount_end(&obj->seq);
329
330 /* inplace update, no shared fences */
331 while (i--)
332 dma_fence_put(rcu_dereference_protected(old->shared[i],
333 dma_resv_held(obj)));
334
335 dma_fence_put(old_fence);
336 }
337 EXPORT_SYMBOL(dma_resv_add_excl_fence);
338
339 /**
340 * dma_resv_copy_fences - Copy all fences from src to dst.
341 * @dst: the destination reservation object
342 * @src: the source reservation object
343 *
344 * Copy all fences from src to dst. dst-lock must be held.
345 */
dma_resv_copy_fences(struct dma_resv * dst,struct dma_resv * src)346 int dma_resv_copy_fences(struct dma_resv *dst, struct dma_resv *src)
347 {
348 struct dma_resv_list *src_list, *dst_list;
349 struct dma_fence *old, *new;
350 unsigned i;
351
352 dma_resv_assert_held(dst);
353
354 rcu_read_lock();
355 src_list = rcu_dereference(src->fence);
356
357 retry:
358 if (src_list) {
359 unsigned shared_count = src_list->shared_count;
360
361 rcu_read_unlock();
362
363 dst_list = dma_resv_list_alloc(shared_count);
364 if (!dst_list)
365 return -ENOMEM;
366
367 rcu_read_lock();
368 src_list = rcu_dereference(src->fence);
369 if (!src_list || src_list->shared_count > shared_count) {
370 kfree(dst_list);
371 goto retry;
372 }
373
374 dst_list->shared_count = 0;
375 for (i = 0; i < src_list->shared_count; ++i) {
376 struct dma_fence *fence;
377
378 fence = rcu_dereference(src_list->shared[i]);
379 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
380 &fence->flags))
381 continue;
382
383 if (!dma_fence_get_rcu(fence)) {
384 dma_resv_list_free(dst_list);
385 src_list = rcu_dereference(src->fence);
386 goto retry;
387 }
388
389 if (dma_fence_is_signaled(fence)) {
390 dma_fence_put(fence);
391 continue;
392 }
393
394 rcu_assign_pointer(dst_list->shared[dst_list->shared_count++], fence);
395 }
396 } else {
397 dst_list = NULL;
398 }
399
400 new = dma_fence_get_rcu_safe(&src->fence_excl);
401 rcu_read_unlock();
402
403 src_list = dma_resv_get_list(dst);
404 old = dma_resv_get_excl(dst);
405
406 write_seqcount_begin(&dst->seq);
407 /* write_seqcount_begin provides the necessary memory barrier */
408 RCU_INIT_POINTER(dst->fence_excl, new);
409 RCU_INIT_POINTER(dst->fence, dst_list);
410 write_seqcount_end(&dst->seq);
411
412 dma_resv_list_free(src_list);
413 dma_fence_put(old);
414
415 return 0;
416 }
417 EXPORT_SYMBOL(dma_resv_copy_fences);
418
419 /**
420 * dma_resv_get_fences_rcu - Get an object's shared and exclusive
421 * fences without update side lock held
422 * @obj: the reservation object
423 * @pfence_excl: the returned exclusive fence (or NULL)
424 * @pshared_count: the number of shared fences returned
425 * @pshared: the array of shared fence ptrs returned (array is krealloc'd to
426 * the required size, and must be freed by caller)
427 *
428 * Retrieve all fences from the reservation object. If the pointer for the
429 * exclusive fence is not specified the fence is put into the array of the
430 * shared fences as well. Returns either zero or -ENOMEM.
431 */
dma_resv_get_fences_rcu(struct dma_resv * obj,struct dma_fence ** pfence_excl,unsigned * pshared_count,struct dma_fence *** pshared)432 int dma_resv_get_fences_rcu(struct dma_resv *obj,
433 struct dma_fence **pfence_excl,
434 unsigned *pshared_count,
435 struct dma_fence ***pshared)
436 {
437 struct dma_fence **shared = NULL;
438 struct dma_fence *fence_excl;
439 unsigned int shared_count;
440 int ret = 1;
441
442 do {
443 struct dma_resv_list *fobj;
444 unsigned int i, seq;
445 size_t sz = 0;
446
447 shared_count = i = 0;
448
449 rcu_read_lock();
450 seq = read_seqcount_begin(&obj->seq);
451
452 fence_excl = rcu_dereference(obj->fence_excl);
453 if (fence_excl && !dma_fence_get_rcu(fence_excl))
454 goto unlock;
455
456 fobj = rcu_dereference(obj->fence);
457 if (fobj)
458 sz += sizeof(*shared) * fobj->shared_max;
459
460 if (!pfence_excl && fence_excl)
461 sz += sizeof(*shared);
462
463 if (sz) {
464 struct dma_fence **nshared;
465
466 nshared = krealloc(shared, sz,
467 GFP_NOWAIT | __GFP_NOWARN);
468 if (!nshared) {
469 rcu_read_unlock();
470
471 dma_fence_put(fence_excl);
472 fence_excl = NULL;
473
474 nshared = krealloc(shared, sz, GFP_KERNEL);
475 if (nshared) {
476 shared = nshared;
477 continue;
478 }
479
480 ret = -ENOMEM;
481 break;
482 }
483 shared = nshared;
484 shared_count = fobj ? fobj->shared_count : 0;
485 for (i = 0; i < shared_count; ++i) {
486 shared[i] = rcu_dereference(fobj->shared[i]);
487 if (!dma_fence_get_rcu(shared[i]))
488 break;
489 }
490 }
491
492 if (i != shared_count || read_seqcount_retry(&obj->seq, seq)) {
493 while (i--)
494 dma_fence_put(shared[i]);
495 dma_fence_put(fence_excl);
496 goto unlock;
497 }
498
499 ret = 0;
500 unlock:
501 rcu_read_unlock();
502 } while (ret);
503
504 if (pfence_excl)
505 *pfence_excl = fence_excl;
506 else if (fence_excl)
507 shared[shared_count++] = fence_excl;
508
509 if (!shared_count) {
510 kfree(shared);
511 shared = NULL;
512 }
513
514 *pshared_count = shared_count;
515 *pshared = shared;
516 return ret;
517 }
518 EXPORT_SYMBOL_GPL(dma_resv_get_fences_rcu);
519
520 /**
521 * dma_resv_wait_timeout_rcu - Wait on reservation's objects
522 * shared and/or exclusive fences.
523 * @obj: the reservation object
524 * @wait_all: if true, wait on all fences, else wait on just exclusive fence
525 * @intr: if true, do interruptible wait
526 * @timeout: timeout value in jiffies or zero to return immediately
527 *
528 * RETURNS
529 * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or
530 * greater than zer on success.
531 */
dma_resv_wait_timeout_rcu(struct dma_resv * obj,bool wait_all,bool intr,unsigned long timeout)532 long dma_resv_wait_timeout_rcu(struct dma_resv *obj,
533 bool wait_all, bool intr,
534 unsigned long timeout)
535 {
536 struct dma_fence *fence;
537 unsigned seq, shared_count;
538 long ret = timeout ? timeout : 1;
539 int i;
540
541 retry:
542 shared_count = 0;
543 seq = read_seqcount_begin(&obj->seq);
544 rcu_read_lock();
545 i = -1;
546
547 fence = rcu_dereference(obj->fence_excl);
548 if (fence && !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
549 if (!dma_fence_get_rcu(fence))
550 goto unlock_retry;
551
552 if (dma_fence_is_signaled(fence)) {
553 dma_fence_put(fence);
554 fence = NULL;
555 }
556
557 } else {
558 fence = NULL;
559 }
560
561 if (wait_all) {
562 struct dma_resv_list *fobj = rcu_dereference(obj->fence);
563
564 if (fobj)
565 shared_count = fobj->shared_count;
566
567 for (i = 0; !fence && i < shared_count; ++i) {
568 struct dma_fence *lfence = rcu_dereference(fobj->shared[i]);
569
570 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
571 &lfence->flags))
572 continue;
573
574 if (!dma_fence_get_rcu(lfence))
575 goto unlock_retry;
576
577 if (dma_fence_is_signaled(lfence)) {
578 dma_fence_put(lfence);
579 continue;
580 }
581
582 fence = lfence;
583 break;
584 }
585 }
586
587 rcu_read_unlock();
588 if (fence) {
589 if (read_seqcount_retry(&obj->seq, seq)) {
590 dma_fence_put(fence);
591 goto retry;
592 }
593
594 ret = dma_fence_wait_timeout(fence, intr, ret);
595 dma_fence_put(fence);
596 if (ret > 0 && wait_all && (i + 1 < shared_count))
597 goto retry;
598 }
599 return ret;
600
601 unlock_retry:
602 rcu_read_unlock();
603 goto retry;
604 }
605 EXPORT_SYMBOL_GPL(dma_resv_wait_timeout_rcu);
606
607
dma_resv_test_signaled_single(struct dma_fence * passed_fence)608 static inline int dma_resv_test_signaled_single(struct dma_fence *passed_fence)
609 {
610 struct dma_fence *fence, *lfence = passed_fence;
611 int ret = 1;
612
613 if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &lfence->flags)) {
614 fence = dma_fence_get_rcu(lfence);
615 if (!fence)
616 return -1;
617
618 ret = !!dma_fence_is_signaled(fence);
619 dma_fence_put(fence);
620 }
621 return ret;
622 }
623
624 /**
625 * dma_resv_test_signaled_rcu - Test if a reservation object's
626 * fences have been signaled.
627 * @obj: the reservation object
628 * @test_all: if true, test all fences, otherwise only test the exclusive
629 * fence
630 *
631 * RETURNS
632 * true if all fences signaled, else false
633 */
dma_resv_test_signaled_rcu(struct dma_resv * obj,bool test_all)634 bool dma_resv_test_signaled_rcu(struct dma_resv *obj, bool test_all)
635 {
636 unsigned seq, shared_count;
637 int ret;
638
639 rcu_read_lock();
640 retry:
641 ret = true;
642 shared_count = 0;
643 seq = read_seqcount_begin(&obj->seq);
644
645 if (test_all) {
646 unsigned i;
647
648 struct dma_resv_list *fobj = rcu_dereference(obj->fence);
649
650 if (fobj)
651 shared_count = fobj->shared_count;
652
653 for (i = 0; i < shared_count; ++i) {
654 struct dma_fence *fence = rcu_dereference(fobj->shared[i]);
655
656 ret = dma_resv_test_signaled_single(fence);
657 if (ret < 0)
658 goto retry;
659 else if (!ret)
660 break;
661 }
662
663 if (read_seqcount_retry(&obj->seq, seq))
664 goto retry;
665 }
666
667 if (!shared_count) {
668 struct dma_fence *fence_excl = rcu_dereference(obj->fence_excl);
669
670 if (fence_excl) {
671 ret = dma_resv_test_signaled_single(fence_excl);
672 if (ret < 0)
673 goto retry;
674
675 if (read_seqcount_retry(&obj->seq, seq))
676 goto retry;
677 }
678 }
679
680 rcu_read_unlock();
681 return ret;
682 }
683 EXPORT_SYMBOL_GPL(dma_resv_test_signaled_rcu);
684