xref: /linux/drivers/dma-buf/dma-fence.c (revision 8da9704c8bb7d4b0a2b051a5a7eda9b049f5f766)
1 /*
2  * Fence mechanism for dma-buf and to allow for asynchronous dma access
3  *
4  * Copyright (C) 2012 Canonical Ltd
5  * Copyright (C) 2012 Texas Instruments
6  *
7  * Authors:
8  * Rob Clark <robdclark@gmail.com>
9  * Maarten Lankhorst <maarten.lankhorst@canonical.com>
10  *
11  * This program is free software; you can redistribute it and/or modify it
12  * under the terms of the GNU General Public License version 2 as published by
13  * the Free Software Foundation.
14  *
15  * This program is distributed in the hope that it will be useful, but WITHOUT
16  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
18  * more details.
19  */
20 
21 #include <linux/slab.h>
22 #include <linux/export.h>
23 #include <linux/atomic.h>
24 #include <linux/dma-fence.h>
25 
26 #define CREATE_TRACE_POINTS
27 #include <trace/events/dma_fence.h>
28 
29 EXPORT_TRACEPOINT_SYMBOL(dma_fence_annotate_wait_on);
30 EXPORT_TRACEPOINT_SYMBOL(dma_fence_emit);
31 
32 /*
33  * fence context counter: each execution context should have its own
34  * fence context, this allows checking if fences belong to the same
35  * context or not. One device can have multiple separate contexts,
36  * and they're used if some engine can run independently of another.
37  */
38 static atomic64_t dma_fence_context_counter = ATOMIC64_INIT(0);
39 
40 /**
41  * dma_fence_context_alloc - allocate an array of fence contexts
42  * @num:	[in]	amount of contexts to allocate
43  *
44  * This function will return the first index of the number of fences allocated.
45  * The fence context is used for setting fence->context to a unique number.
46  */
47 u64 dma_fence_context_alloc(unsigned num)
48 {
49 	BUG_ON(!num);
50 	return atomic64_add_return(num, &dma_fence_context_counter) - num;
51 }
52 EXPORT_SYMBOL(dma_fence_context_alloc);
53 
54 /**
55  * dma_fence_signal_locked - signal completion of a fence
56  * @fence: the fence to signal
57  *
58  * Signal completion for software callbacks on a fence, this will unblock
59  * dma_fence_wait() calls and run all the callbacks added with
60  * dma_fence_add_callback(). Can be called multiple times, but since a fence
61  * can only go from unsignaled to signaled state, it will only be effective
62  * the first time.
63  *
64  * Unlike dma_fence_signal, this function must be called with fence->lock held.
65  */
66 int dma_fence_signal_locked(struct dma_fence *fence)
67 {
68 	struct dma_fence_cb *cur, *tmp;
69 	int ret = 0;
70 
71 	lockdep_assert_held(fence->lock);
72 
73 	if (WARN_ON(!fence))
74 		return -EINVAL;
75 
76 	if (!ktime_to_ns(fence->timestamp)) {
77 		fence->timestamp = ktime_get();
78 		smp_mb__before_atomic();
79 	}
80 
81 	if (test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
82 		ret = -EINVAL;
83 
84 		/*
85 		 * we might have raced with the unlocked dma_fence_signal,
86 		 * still run through all callbacks
87 		 */
88 	} else
89 		trace_dma_fence_signaled(fence);
90 
91 	list_for_each_entry_safe(cur, tmp, &fence->cb_list, node) {
92 		list_del_init(&cur->node);
93 		cur->func(fence, cur);
94 	}
95 	return ret;
96 }
97 EXPORT_SYMBOL(dma_fence_signal_locked);
98 
99 /**
100  * dma_fence_signal - signal completion of a fence
101  * @fence: the fence to signal
102  *
103  * Signal completion for software callbacks on a fence, this will unblock
104  * dma_fence_wait() calls and run all the callbacks added with
105  * dma_fence_add_callback(). Can be called multiple times, but since a fence
106  * can only go from unsignaled to signaled state, it will only be effective
107  * the first time.
108  */
109 int dma_fence_signal(struct dma_fence *fence)
110 {
111 	unsigned long flags;
112 
113 	if (!fence)
114 		return -EINVAL;
115 
116 	if (!ktime_to_ns(fence->timestamp)) {
117 		fence->timestamp = ktime_get();
118 		smp_mb__before_atomic();
119 	}
120 
121 	if (test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
122 		return -EINVAL;
123 
124 	trace_dma_fence_signaled(fence);
125 
126 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &fence->flags)) {
127 		struct dma_fence_cb *cur, *tmp;
128 
129 		spin_lock_irqsave(fence->lock, flags);
130 		list_for_each_entry_safe(cur, tmp, &fence->cb_list, node) {
131 			list_del_init(&cur->node);
132 			cur->func(fence, cur);
133 		}
134 		spin_unlock_irqrestore(fence->lock, flags);
135 	}
136 	return 0;
137 }
138 EXPORT_SYMBOL(dma_fence_signal);
139 
140 /**
141  * dma_fence_wait_timeout - sleep until the fence gets signaled
142  * or until timeout elapses
143  * @fence:	[in]	the fence to wait on
144  * @intr:	[in]	if true, do an interruptible wait
145  * @timeout:	[in]	timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
146  *
147  * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or the
148  * remaining timeout in jiffies on success. Other error values may be
149  * returned on custom implementations.
150  *
151  * Performs a synchronous wait on this fence. It is assumed the caller
152  * directly or indirectly (buf-mgr between reservation and committing)
153  * holds a reference to the fence, otherwise the fence might be
154  * freed before return, resulting in undefined behavior.
155  */
156 signed long
157 dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
158 {
159 	signed long ret;
160 
161 	if (WARN_ON(timeout < 0))
162 		return -EINVAL;
163 
164 	trace_dma_fence_wait_start(fence);
165 	ret = fence->ops->wait(fence, intr, timeout);
166 	trace_dma_fence_wait_end(fence);
167 	return ret;
168 }
169 EXPORT_SYMBOL(dma_fence_wait_timeout);
170 
171 void dma_fence_release(struct kref *kref)
172 {
173 	struct dma_fence *fence =
174 		container_of(kref, struct dma_fence, refcount);
175 
176 	trace_dma_fence_destroy(fence);
177 
178 	BUG_ON(!list_empty(&fence->cb_list));
179 
180 	if (fence->ops->release)
181 		fence->ops->release(fence);
182 	else
183 		dma_fence_free(fence);
184 }
185 EXPORT_SYMBOL(dma_fence_release);
186 
187 void dma_fence_free(struct dma_fence *fence)
188 {
189 	kfree_rcu(fence, rcu);
190 }
191 EXPORT_SYMBOL(dma_fence_free);
192 
193 /**
194  * dma_fence_enable_sw_signaling - enable signaling on fence
195  * @fence:	[in]	the fence to enable
196  *
197  * this will request for sw signaling to be enabled, to make the fence
198  * complete as soon as possible
199  */
200 void dma_fence_enable_sw_signaling(struct dma_fence *fence)
201 {
202 	unsigned long flags;
203 
204 	if (!test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
205 			      &fence->flags) &&
206 	    !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
207 		trace_dma_fence_enable_signal(fence);
208 
209 		spin_lock_irqsave(fence->lock, flags);
210 
211 		if (!fence->ops->enable_signaling(fence))
212 			dma_fence_signal_locked(fence);
213 
214 		spin_unlock_irqrestore(fence->lock, flags);
215 	}
216 }
217 EXPORT_SYMBOL(dma_fence_enable_sw_signaling);
218 
219 /**
220  * dma_fence_add_callback - add a callback to be called when the fence
221  * is signaled
222  * @fence:	[in]	the fence to wait on
223  * @cb:		[in]	the callback to register
224  * @func:	[in]	the function to call
225  *
226  * cb will be initialized by dma_fence_add_callback, no initialization
227  * by the caller is required. Any number of callbacks can be registered
228  * to a fence, but a callback can only be registered to one fence at a time.
229  *
230  * Note that the callback can be called from an atomic context.  If
231  * fence is already signaled, this function will return -ENOENT (and
232  * *not* call the callback)
233  *
234  * Add a software callback to the fence. Same restrictions apply to
235  * refcount as it does to dma_fence_wait, however the caller doesn't need to
236  * keep a refcount to fence afterwards: when software access is enabled,
237  * the creator of the fence is required to keep the fence alive until
238  * after it signals with dma_fence_signal. The callback itself can be called
239  * from irq context.
240  *
241  */
242 int dma_fence_add_callback(struct dma_fence *fence, struct dma_fence_cb *cb,
243 			   dma_fence_func_t func)
244 {
245 	unsigned long flags;
246 	int ret = 0;
247 	bool was_set;
248 
249 	if (WARN_ON(!fence || !func))
250 		return -EINVAL;
251 
252 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
253 		INIT_LIST_HEAD(&cb->node);
254 		return -ENOENT;
255 	}
256 
257 	spin_lock_irqsave(fence->lock, flags);
258 
259 	was_set = test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
260 				   &fence->flags);
261 
262 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
263 		ret = -ENOENT;
264 	else if (!was_set) {
265 		trace_dma_fence_enable_signal(fence);
266 
267 		if (!fence->ops->enable_signaling(fence)) {
268 			dma_fence_signal_locked(fence);
269 			ret = -ENOENT;
270 		}
271 	}
272 
273 	if (!ret) {
274 		cb->func = func;
275 		list_add_tail(&cb->node, &fence->cb_list);
276 	} else
277 		INIT_LIST_HEAD(&cb->node);
278 	spin_unlock_irqrestore(fence->lock, flags);
279 
280 	return ret;
281 }
282 EXPORT_SYMBOL(dma_fence_add_callback);
283 
284 /**
285  * dma_fence_remove_callback - remove a callback from the signaling list
286  * @fence:	[in]	the fence to wait on
287  * @cb:		[in]	the callback to remove
288  *
289  * Remove a previously queued callback from the fence. This function returns
290  * true if the callback is successfully removed, or false if the fence has
291  * already been signaled.
292  *
293  * *WARNING*:
294  * Cancelling a callback should only be done if you really know what you're
295  * doing, since deadlocks and race conditions could occur all too easily. For
296  * this reason, it should only ever be done on hardware lockup recovery,
297  * with a reference held to the fence.
298  */
299 bool
300 dma_fence_remove_callback(struct dma_fence *fence, struct dma_fence_cb *cb)
301 {
302 	unsigned long flags;
303 	bool ret;
304 
305 	spin_lock_irqsave(fence->lock, flags);
306 
307 	ret = !list_empty(&cb->node);
308 	if (ret)
309 		list_del_init(&cb->node);
310 
311 	spin_unlock_irqrestore(fence->lock, flags);
312 
313 	return ret;
314 }
315 EXPORT_SYMBOL(dma_fence_remove_callback);
316 
317 struct default_wait_cb {
318 	struct dma_fence_cb base;
319 	struct task_struct *task;
320 };
321 
322 static void
323 dma_fence_default_wait_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
324 {
325 	struct default_wait_cb *wait =
326 		container_of(cb, struct default_wait_cb, base);
327 
328 	wake_up_state(wait->task, TASK_NORMAL);
329 }
330 
331 /**
332  * dma_fence_default_wait - default sleep until the fence gets signaled
333  * or until timeout elapses
334  * @fence:	[in]	the fence to wait on
335  * @intr:	[in]	if true, do an interruptible wait
336  * @timeout:	[in]	timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
337  *
338  * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or the
339  * remaining timeout in jiffies on success. If timeout is zero the value one is
340  * returned if the fence is already signaled for consistency with other
341  * functions taking a jiffies timeout.
342  */
343 signed long
344 dma_fence_default_wait(struct dma_fence *fence, bool intr, signed long timeout)
345 {
346 	struct default_wait_cb cb;
347 	unsigned long flags;
348 	signed long ret = timeout ? timeout : 1;
349 	bool was_set;
350 
351 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
352 		return ret;
353 
354 	spin_lock_irqsave(fence->lock, flags);
355 
356 	if (intr && signal_pending(current)) {
357 		ret = -ERESTARTSYS;
358 		goto out;
359 	}
360 
361 	was_set = test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
362 				   &fence->flags);
363 
364 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
365 		goto out;
366 
367 	if (!was_set) {
368 		trace_dma_fence_enable_signal(fence);
369 
370 		if (!fence->ops->enable_signaling(fence)) {
371 			dma_fence_signal_locked(fence);
372 			goto out;
373 		}
374 	}
375 
376 	cb.base.func = dma_fence_default_wait_cb;
377 	cb.task = current;
378 	list_add(&cb.base.node, &fence->cb_list);
379 
380 	while (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) && ret > 0) {
381 		if (intr)
382 			__set_current_state(TASK_INTERRUPTIBLE);
383 		else
384 			__set_current_state(TASK_UNINTERRUPTIBLE);
385 		spin_unlock_irqrestore(fence->lock, flags);
386 
387 		ret = schedule_timeout(ret);
388 
389 		spin_lock_irqsave(fence->lock, flags);
390 		if (ret > 0 && intr && signal_pending(current))
391 			ret = -ERESTARTSYS;
392 	}
393 
394 	if (!list_empty(&cb.base.node))
395 		list_del(&cb.base.node);
396 	__set_current_state(TASK_RUNNING);
397 
398 out:
399 	spin_unlock_irqrestore(fence->lock, flags);
400 	return ret;
401 }
402 EXPORT_SYMBOL(dma_fence_default_wait);
403 
404 static bool
405 dma_fence_test_signaled_any(struct dma_fence **fences, uint32_t count,
406 			    uint32_t *idx)
407 {
408 	int i;
409 
410 	for (i = 0; i < count; ++i) {
411 		struct dma_fence *fence = fences[i];
412 		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
413 			if (idx)
414 				*idx = i;
415 			return true;
416 		}
417 	}
418 	return false;
419 }
420 
421 /**
422  * dma_fence_wait_any_timeout - sleep until any fence gets signaled
423  * or until timeout elapses
424  * @fences:	[in]	array of fences to wait on
425  * @count:	[in]	number of fences to wait on
426  * @intr:	[in]	if true, do an interruptible wait
427  * @timeout:	[in]	timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
428  * @idx:       [out]	the first signaled fence index, meaningful only on
429  *			positive return
430  *
431  * Returns -EINVAL on custom fence wait implementation, -ERESTARTSYS if
432  * interrupted, 0 if the wait timed out, or the remaining timeout in jiffies
433  * on success.
434  *
435  * Synchronous waits for the first fence in the array to be signaled. The
436  * caller needs to hold a reference to all fences in the array, otherwise a
437  * fence might be freed before return, resulting in undefined behavior.
438  */
439 signed long
440 dma_fence_wait_any_timeout(struct dma_fence **fences, uint32_t count,
441 			   bool intr, signed long timeout, uint32_t *idx)
442 {
443 	struct default_wait_cb *cb;
444 	signed long ret = timeout;
445 	unsigned i;
446 
447 	if (WARN_ON(!fences || !count || timeout < 0))
448 		return -EINVAL;
449 
450 	if (timeout == 0) {
451 		for (i = 0; i < count; ++i)
452 			if (dma_fence_is_signaled(fences[i])) {
453 				if (idx)
454 					*idx = i;
455 				return 1;
456 			}
457 
458 		return 0;
459 	}
460 
461 	cb = kcalloc(count, sizeof(struct default_wait_cb), GFP_KERNEL);
462 	if (cb == NULL) {
463 		ret = -ENOMEM;
464 		goto err_free_cb;
465 	}
466 
467 	for (i = 0; i < count; ++i) {
468 		struct dma_fence *fence = fences[i];
469 
470 		if (fence->ops->wait != dma_fence_default_wait) {
471 			ret = -EINVAL;
472 			goto fence_rm_cb;
473 		}
474 
475 		cb[i].task = current;
476 		if (dma_fence_add_callback(fence, &cb[i].base,
477 					   dma_fence_default_wait_cb)) {
478 			/* This fence is already signaled */
479 			if (idx)
480 				*idx = i;
481 			goto fence_rm_cb;
482 		}
483 	}
484 
485 	while (ret > 0) {
486 		if (intr)
487 			set_current_state(TASK_INTERRUPTIBLE);
488 		else
489 			set_current_state(TASK_UNINTERRUPTIBLE);
490 
491 		if (dma_fence_test_signaled_any(fences, count, idx))
492 			break;
493 
494 		ret = schedule_timeout(ret);
495 
496 		if (ret > 0 && intr && signal_pending(current))
497 			ret = -ERESTARTSYS;
498 	}
499 
500 	__set_current_state(TASK_RUNNING);
501 
502 fence_rm_cb:
503 	while (i-- > 0)
504 		dma_fence_remove_callback(fences[i], &cb[i].base);
505 
506 err_free_cb:
507 	kfree(cb);
508 
509 	return ret;
510 }
511 EXPORT_SYMBOL(dma_fence_wait_any_timeout);
512 
513 /**
514  * dma_fence_init - Initialize a custom fence.
515  * @fence:	[in]	the fence to initialize
516  * @ops:	[in]	the dma_fence_ops for operations on this fence
517  * @lock:	[in]	the irqsafe spinlock to use for locking this fence
518  * @context:	[in]	the execution context this fence is run on
519  * @seqno:	[in]	a linear increasing sequence number for this context
520  *
521  * Initializes an allocated fence, the caller doesn't have to keep its
522  * refcount after committing with this fence, but it will need to hold a
523  * refcount again if dma_fence_ops.enable_signaling gets called. This can
524  * be used for other implementing other types of fence.
525  *
526  * context and seqno are used for easy comparison between fences, allowing
527  * to check which fence is later by simply using dma_fence_later.
528  */
529 void
530 dma_fence_init(struct dma_fence *fence, const struct dma_fence_ops *ops,
531 	       spinlock_t *lock, u64 context, unsigned seqno)
532 {
533 	BUG_ON(!lock);
534 	BUG_ON(!ops || !ops->wait || !ops->enable_signaling ||
535 	       !ops->get_driver_name || !ops->get_timeline_name);
536 
537 	kref_init(&fence->refcount);
538 	fence->ops = ops;
539 	INIT_LIST_HEAD(&fence->cb_list);
540 	fence->lock = lock;
541 	fence->context = context;
542 	fence->seqno = seqno;
543 	fence->flags = 0UL;
544 
545 	trace_dma_fence_init(fence);
546 }
547 EXPORT_SYMBOL(dma_fence_init);
548