1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Fence mechanism for dma-buf and to allow for asynchronous dma access
  4 *
  5 * Copyright (C) 2012 Canonical Ltd
  6 * Copyright (C) 2012 Texas Instruments
  7 *
  8 * Authors:
  9 * Rob Clark <robdclark@gmail.com>
 10 * Maarten Lankhorst <maarten.lankhorst@canonical.com>
 11 */
 12
 13#include <linux/slab.h>
 14#include <linux/export.h>
 15#include <linux/atomic.h>
 16#include <linux/dma-fence.h>
 17#include <linux/sched/signal.h>
 18
 19#define CREATE_TRACE_POINTS
 20#include <trace/events/dma_fence.h>
 21
 22EXPORT_TRACEPOINT_SYMBOL(dma_fence_emit);
 23EXPORT_TRACEPOINT_SYMBOL(dma_fence_enable_signal);
 24EXPORT_TRACEPOINT_SYMBOL(dma_fence_signaled);
 25
 26static DEFINE_SPINLOCK(dma_fence_stub_lock);
 27static struct dma_fence dma_fence_stub;
 28
 29/*
 30 * fence context counter: each execution context should have its own
 31 * fence context, this allows checking if fences belong to the same
 32 * context or not. One device can have multiple separate contexts,
 33 * and they're used if some engine can run independently of another.
 34 */
 35static atomic64_t dma_fence_context_counter = ATOMIC64_INIT(1);
 36
 37/**
 38 * DOC: DMA fences overview
 39 *
 40 * DMA fences, represented by &struct dma_fence, are the kernel internal
 41 * synchronization primitive for DMA operations like GPU rendering, video
 42 * encoding/decoding, or displaying buffers on a screen.
 43 *
 44 * A fence is initialized using dma_fence_init() and completed using
 45 * dma_fence_signal(). Fences are associated with a context, allocated through
 46 * dma_fence_context_alloc(), and all fences on the same context are
 47 * fully ordered.
 48 *
 49 * Since the purposes of fences is to facilitate cross-device and
 50 * cross-application synchronization, there's multiple ways to use one:
 51 *
 52 * - Individual fences can be exposed as a &sync_file, accessed as a file
 53 *   descriptor from userspace, created by calling sync_file_create(). This is
 54 *   called explicit fencing, since userspace passes around explicit
 55 *   synchronization points.
 56 *
 57 * - Some subsystems also have their own explicit fencing primitives, like
 58 *   &drm_syncobj. Compared to &sync_file, a &drm_syncobj allows the underlying
 59 *   fence to be updated.
 60 *
 61 * - Then there's also implicit fencing, where the synchronization points are
 62 *   implicitly passed around as part of shared &dma_buf instances. Such
 63 *   implicit fences are stored in &struct dma_resv through the
 64 *   &dma_buf.resv pointer.
 65 */
 66
 67static const char *dma_fence_stub_get_name(struct dma_fence *fence)
 68{
 69        return "stub";
 70}
 71
 72static const struct dma_fence_ops dma_fence_stub_ops = {
 73	.get_driver_name = dma_fence_stub_get_name,
 74	.get_timeline_name = dma_fence_stub_get_name,
 75};
 76
 77/**
 78 * dma_fence_get_stub - return a signaled fence
 79 *
 80 * Return a stub fence which is already signaled.
 81 */
 82struct dma_fence *dma_fence_get_stub(void)
 83{
 84	spin_lock(&dma_fence_stub_lock);
 85	if (!dma_fence_stub.ops) {
 86		dma_fence_init(&dma_fence_stub,
 87			       &dma_fence_stub_ops,
 88			       &dma_fence_stub_lock,
 89			       0, 0);
 90		dma_fence_signal_locked(&dma_fence_stub);
 91	}
 92	spin_unlock(&dma_fence_stub_lock);
 93
 94	return dma_fence_get(&dma_fence_stub);
 95}
 96EXPORT_SYMBOL(dma_fence_get_stub);
 97
 98/**
 99 * dma_fence_context_alloc - allocate an array of fence contexts
100 * @num: amount of contexts to allocate
101 *
102 * This function will return the first index of the number of fence contexts
103 * allocated.  The fence context is used for setting &dma_fence.context to a
104 * unique number by passing the context to dma_fence_init().
105 */
106u64 dma_fence_context_alloc(unsigned num)
107{
108	WARN_ON(!num);
109	return atomic64_add_return(num, &dma_fence_context_counter) - num;
110}
111EXPORT_SYMBOL(dma_fence_context_alloc);
112
113/**
114 * dma_fence_signal_locked - signal completion of a fence
115 * @fence: the fence to signal
116 *
117 * Signal completion for software callbacks on a fence, this will unblock
118 * dma_fence_wait() calls and run all the callbacks added with
119 * dma_fence_add_callback(). Can be called multiple times, but since a fence
120 * can only go from the unsignaled to the signaled state and not back, it will
121 * only be effective the first time.
122 *
123 * Unlike dma_fence_signal(), this function must be called with &dma_fence.lock
124 * held.
125 *
126 * Returns 0 on success and a negative error value when @fence has been
127 * signalled already.
128 */
129int dma_fence_signal_locked(struct dma_fence *fence)
130{
131	struct dma_fence_cb *cur, *tmp;
132	struct list_head cb_list;
133
134	lockdep_assert_held(fence->lock);
135
136	if (unlikely(test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
137				      &fence->flags)))
138		return -EINVAL;
139
140	/* Stash the cb_list before replacing it with the timestamp */
141	list_replace(&fence->cb_list, &cb_list);
142
143	fence->timestamp = ktime_get();
144	set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
145	trace_dma_fence_signaled(fence);
146
147	list_for_each_entry_safe(cur, tmp, &cb_list, node) {
148		INIT_LIST_HEAD(&cur->node);
149		cur->func(fence, cur);
150	}
151
152	return 0;
153}
154EXPORT_SYMBOL(dma_fence_signal_locked);
155
156/**
157 * dma_fence_signal - signal completion of a fence
158 * @fence: the fence to signal
159 *
160 * Signal completion for software callbacks on a fence, this will unblock
161 * dma_fence_wait() calls and run all the callbacks added with
162 * dma_fence_add_callback(). Can be called multiple times, but since a fence
163 * can only go from the unsignaled to the signaled state and not back, it will
164 * only be effective the first time.
165 *
166 * Returns 0 on success and a negative error value when @fence has been
167 * signalled already.
168 */
169int dma_fence_signal(struct dma_fence *fence)
170{
171	unsigned long flags;
172	int ret;
173
174	if (!fence)
175		return -EINVAL;
176
177	spin_lock_irqsave(fence->lock, flags);
178	ret = dma_fence_signal_locked(fence);
179	spin_unlock_irqrestore(fence->lock, flags);
180
181	return ret;
182}
183EXPORT_SYMBOL(dma_fence_signal);
184
185/**
186 * dma_fence_wait_timeout - sleep until the fence gets signaled
187 * or until timeout elapses
188 * @fence: the fence to wait on
189 * @intr: if true, do an interruptible wait
190 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
191 *
192 * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or the
193 * remaining timeout in jiffies on success. Other error values may be
194 * returned on custom implementations.
195 *
196 * Performs a synchronous wait on this fence. It is assumed the caller
197 * directly or indirectly (buf-mgr between reservation and committing)
198 * holds a reference to the fence, otherwise the fence might be
199 * freed before return, resulting in undefined behavior.
200 *
201 * See also dma_fence_wait() and dma_fence_wait_any_timeout().
202 */
203signed long
204dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
205{
206	signed long ret;
207
208	if (WARN_ON(timeout < 0))
209		return -EINVAL;
210
211	trace_dma_fence_wait_start(fence);
212	if (fence->ops->wait)
213		ret = fence->ops->wait(fence, intr, timeout);
214	else
215		ret = dma_fence_default_wait(fence, intr, timeout);
216	trace_dma_fence_wait_end(fence);
217	return ret;
218}
219EXPORT_SYMBOL(dma_fence_wait_timeout);
220
221/**
222 * dma_fence_release - default relese function for fences
223 * @kref: &dma_fence.recfount
224 *
225 * This is the default release functions for &dma_fence. Drivers shouldn't call
226 * this directly, but instead call dma_fence_put().
227 */
228void dma_fence_release(struct kref *kref)
229{
230	struct dma_fence *fence =
231		container_of(kref, struct dma_fence, refcount);
232
233	trace_dma_fence_destroy(fence);
234
235	if (WARN(!list_empty(&fence->cb_list) &&
236		 !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags),
237		 "Fence %s:%s:%llx:%llx released with pending signals!\n",
238		 fence->ops->get_driver_name(fence),
239		 fence->ops->get_timeline_name(fence),
240		 fence->context, fence->seqno)) {
241		unsigned long flags;
242
243		/*
244		 * Failed to signal before release, likely a refcounting issue.
245		 *
246		 * This should never happen, but if it does make sure that we
247		 * don't leave chains dangling. We set the error flag first
248		 * so that the callbacks know this signal is due to an error.
249		 */
250		spin_lock_irqsave(fence->lock, flags);
251		fence->error = -EDEADLK;
252		dma_fence_signal_locked(fence);
253		spin_unlock_irqrestore(fence->lock, flags);
254	}
255
256	if (fence->ops->release)
257		fence->ops->release(fence);
258	else
259		dma_fence_free(fence);
260}
261EXPORT_SYMBOL(dma_fence_release);
262
263/**
264 * dma_fence_free - default release function for &dma_fence.
265 * @fence: fence to release
266 *
267 * This is the default implementation for &dma_fence_ops.release. It calls
268 * kfree_rcu() on @fence.
269 */
270void dma_fence_free(struct dma_fence *fence)
271{
272	kfree_rcu(fence, rcu);
273}
274EXPORT_SYMBOL(dma_fence_free);
275
276/**
277 * dma_fence_enable_sw_signaling - enable signaling on fence
278 * @fence: the fence to enable
279 *
280 * This will request for sw signaling to be enabled, to make the fence
281 * complete as soon as possible. This calls &dma_fence_ops.enable_signaling
282 * internally.
283 */
284void dma_fence_enable_sw_signaling(struct dma_fence *fence)
285{
286	unsigned long flags;
287
288	if (!test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
289			      &fence->flags) &&
290	    !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) &&
291	    fence->ops->enable_signaling) {
292		trace_dma_fence_enable_signal(fence);
293
294		spin_lock_irqsave(fence->lock, flags);
295
296		if (!fence->ops->enable_signaling(fence))
297			dma_fence_signal_locked(fence);
298
299		spin_unlock_irqrestore(fence->lock, flags);
300	}
301}
302EXPORT_SYMBOL(dma_fence_enable_sw_signaling);
303
304/**
305 * dma_fence_add_callback - add a callback to be called when the fence
306 * is signaled
307 * @fence: the fence to wait on
308 * @cb: the callback to register
309 * @func: the function to call
310 *
311 * @cb will be initialized by dma_fence_add_callback(), no initialization
312 * by the caller is required. Any number of callbacks can be registered
313 * to a fence, but a callback can only be registered to one fence at a time.
314 *
315 * Note that the callback can be called from an atomic context.  If
316 * fence is already signaled, this function will return -ENOENT (and
317 * *not* call the callback).
318 *
319 * Add a software callback to the fence. Same restrictions apply to
320 * refcount as it does to dma_fence_wait(), however the caller doesn't need to
321 * keep a refcount to fence afterward dma_fence_add_callback() has returned:
322 * when software access is enabled, the creator of the fence is required to keep
323 * the fence alive until after it signals with dma_fence_signal(). The callback
324 * itself can be called from irq context.
325 *
326 * Returns 0 in case of success, -ENOENT if the fence is already signaled
327 * and -EINVAL in case of error.
328 */
329int dma_fence_add_callback(struct dma_fence *fence, struct dma_fence_cb *cb,
330			   dma_fence_func_t func)
331{
332	unsigned long flags;
333	int ret = 0;
334	bool was_set;
335
336	if (WARN_ON(!fence || !func))
337		return -EINVAL;
338
339	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
340		INIT_LIST_HEAD(&cb->node);
341		return -ENOENT;
342	}
343
344	spin_lock_irqsave(fence->lock, flags);
345
346	was_set = test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
347				   &fence->flags);
348
349	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
350		ret = -ENOENT;
351	else if (!was_set && fence->ops->enable_signaling) {
352		trace_dma_fence_enable_signal(fence);
353
354		if (!fence->ops->enable_signaling(fence)) {
355			dma_fence_signal_locked(fence);
356			ret = -ENOENT;
357		}
358	}
359
360	if (!ret) {
361		cb->func = func;
362		list_add_tail(&cb->node, &fence->cb_list);
363	} else
364		INIT_LIST_HEAD(&cb->node);
365	spin_unlock_irqrestore(fence->lock, flags);
366
367	return ret;
368}
369EXPORT_SYMBOL(dma_fence_add_callback);
370
371/**
372 * dma_fence_get_status - returns the status upon completion
373 * @fence: the dma_fence to query
374 *
375 * This wraps dma_fence_get_status_locked() to return the error status
376 * condition on a signaled fence. See dma_fence_get_status_locked() for more
377 * details.
378 *
379 * Returns 0 if the fence has not yet been signaled, 1 if the fence has
380 * been signaled without an error condition, or a negative error code
381 * if the fence has been completed in err.
382 */
383int dma_fence_get_status(struct dma_fence *fence)
384{
385	unsigned long flags;
386	int status;
387
388	spin_lock_irqsave(fence->lock, flags);
389	status = dma_fence_get_status_locked(fence);
390	spin_unlock_irqrestore(fence->lock, flags);
391
392	return status;
393}
394EXPORT_SYMBOL(dma_fence_get_status);
395
396/**
397 * dma_fence_remove_callback - remove a callback from the signaling list
398 * @fence: the fence to wait on
399 * @cb: the callback to remove
400 *
401 * Remove a previously queued callback from the fence. This function returns
402 * true if the callback is successfully removed, or false if the fence has
403 * already been signaled.
404 *
405 * *WARNING*:
406 * Cancelling a callback should only be done if you really know what you're
407 * doing, since deadlocks and race conditions could occur all too easily. For
408 * this reason, it should only ever be done on hardware lockup recovery,
409 * with a reference held to the fence.
410 *
411 * Behaviour is undefined if @cb has not been added to @fence using
412 * dma_fence_add_callback() beforehand.
413 */
414bool
415dma_fence_remove_callback(struct dma_fence *fence, struct dma_fence_cb *cb)
416{
417	unsigned long flags;
418	bool ret;
419
420	spin_lock_irqsave(fence->lock, flags);
421
422	ret = !list_empty(&cb->node);
423	if (ret)
424		list_del_init(&cb->node);
425
426	spin_unlock_irqrestore(fence->lock, flags);
427
428	return ret;
429}
430EXPORT_SYMBOL(dma_fence_remove_callback);
431
432struct default_wait_cb {
433	struct dma_fence_cb base;
434	struct task_struct *task;
435};
436
437static void
438dma_fence_default_wait_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
439{
440	struct default_wait_cb *wait =
441		container_of(cb, struct default_wait_cb, base);
442
443	wake_up_state(wait->task, TASK_NORMAL);
444}
445
446/**
447 * dma_fence_default_wait - default sleep until the fence gets signaled
448 * or until timeout elapses
449 * @fence: the fence to wait on
450 * @intr: if true, do an interruptible wait
451 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
452 *
453 * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or the
454 * remaining timeout in jiffies on success. If timeout is zero the value one is
455 * returned if the fence is already signaled for consistency with other
456 * functions taking a jiffies timeout.
457 */
458signed long
459dma_fence_default_wait(struct dma_fence *fence, bool intr, signed long timeout)
460{
461	struct default_wait_cb cb;
462	unsigned long flags;
463	signed long ret = timeout ? timeout : 1;
464	bool was_set;
465
466	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
467		return ret;
468
469	spin_lock_irqsave(fence->lock, flags);
470
471	if (intr && signal_pending(current)) {
472		ret = -ERESTARTSYS;
473		goto out;
474	}
475
476	was_set = test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
477				   &fence->flags);
478
479	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
480		goto out;
481
482	if (!was_set && fence->ops->enable_signaling) {
483		trace_dma_fence_enable_signal(fence);
484
485		if (!fence->ops->enable_signaling(fence)) {
486			dma_fence_signal_locked(fence);
487			goto out;
488		}
489	}
490
491	if (!timeout) {
492		ret = 0;
493		goto out;
494	}
495
496	cb.base.func = dma_fence_default_wait_cb;
497	cb.task = current;
498	list_add(&cb.base.node, &fence->cb_list);
499
500	while (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) && ret > 0) {
501		if (intr)
502			__set_current_state(TASK_INTERRUPTIBLE);
503		else
504			__set_current_state(TASK_UNINTERRUPTIBLE);
505		spin_unlock_irqrestore(fence->lock, flags);
506
507		ret = schedule_timeout(ret);
508
509		spin_lock_irqsave(fence->lock, flags);
510		if (ret > 0 && intr && signal_pending(current))
511			ret = -ERESTARTSYS;
512	}
513
514	if (!list_empty(&cb.base.node))
515		list_del(&cb.base.node);
516	__set_current_state(TASK_RUNNING);
517
518out:
519	spin_unlock_irqrestore(fence->lock, flags);
520	return ret;
521}
522EXPORT_SYMBOL(dma_fence_default_wait);
523
524static bool
525dma_fence_test_signaled_any(struct dma_fence **fences, uint32_t count,
526			    uint32_t *idx)
527{
528	int i;
529
530	for (i = 0; i < count; ++i) {
531		struct dma_fence *fence = fences[i];
532		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
533			if (idx)
534				*idx = i;
535			return true;
536		}
537	}
538	return false;
539}
540
541/**
542 * dma_fence_wait_any_timeout - sleep until any fence gets signaled
543 * or until timeout elapses
544 * @fences: array of fences to wait on
545 * @count: number of fences to wait on
546 * @intr: if true, do an interruptible wait
547 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
548 * @idx: used to store the first signaled fence index, meaningful only on
549 *	positive return
550 *
551 * Returns -EINVAL on custom fence wait implementation, -ERESTARTSYS if
552 * interrupted, 0 if the wait timed out, or the remaining timeout in jiffies
553 * on success.
554 *
555 * Synchronous waits for the first fence in the array to be signaled. The
556 * caller needs to hold a reference to all fences in the array, otherwise a
557 * fence might be freed before return, resulting in undefined behavior.
558 *
559 * See also dma_fence_wait() and dma_fence_wait_timeout().
560 */
561signed long
562dma_fence_wait_any_timeout(struct dma_fence **fences, uint32_t count,
563			   bool intr, signed long timeout, uint32_t *idx)
564{
565	struct default_wait_cb *cb;
566	signed long ret = timeout;
567	unsigned i;
568
569	if (WARN_ON(!fences || !count || timeout < 0))
570		return -EINVAL;
571
572	if (timeout == 0) {
573		for (i = 0; i < count; ++i)
574			if (dma_fence_is_signaled(fences[i])) {
575				if (idx)
576					*idx = i;
577				return 1;
578			}
579
580		return 0;
581	}
582
583	cb = kcalloc(count, sizeof(struct default_wait_cb), GFP_KERNEL);
584	if (cb == NULL) {
585		ret = -ENOMEM;
586		goto err_free_cb;
587	}
588
589	for (i = 0; i < count; ++i) {
590		struct dma_fence *fence = fences[i];
591
592		cb[i].task = current;
593		if (dma_fence_add_callback(fence, &cb[i].base,
594					   dma_fence_default_wait_cb)) {
595			/* This fence is already signaled */
596			if (idx)
597				*idx = i;
598			goto fence_rm_cb;
599		}
600	}
601
602	while (ret > 0) {
603		if (intr)
604			set_current_state(TASK_INTERRUPTIBLE);
605		else
606			set_current_state(TASK_UNINTERRUPTIBLE);
607
608		if (dma_fence_test_signaled_any(fences, count, idx))
609			break;
610
611		ret = schedule_timeout(ret);
612
613		if (ret > 0 && intr && signal_pending(current))
614			ret = -ERESTARTSYS;
615	}
616
617	__set_current_state(TASK_RUNNING);
618
619fence_rm_cb:
620	while (i-- > 0)
621		dma_fence_remove_callback(fences[i], &cb[i].base);
622
623err_free_cb:
624	kfree(cb);
625
626	return ret;
627}
628EXPORT_SYMBOL(dma_fence_wait_any_timeout);
629
630/**
631 * dma_fence_init - Initialize a custom fence.
632 * @fence: the fence to initialize
633 * @ops: the dma_fence_ops for operations on this fence
634 * @lock: the irqsafe spinlock to use for locking this fence
635 * @context: the execution context this fence is run on
636 * @seqno: a linear increasing sequence number for this context
637 *
638 * Initializes an allocated fence, the caller doesn't have to keep its
639 * refcount after committing with this fence, but it will need to hold a
640 * refcount again if &dma_fence_ops.enable_signaling gets called.
641 *
642 * context and seqno are used for easy comparison between fences, allowing
643 * to check which fence is later by simply using dma_fence_later().
644 */
645void
646dma_fence_init(struct dma_fence *fence, const struct dma_fence_ops *ops,
647	       spinlock_t *lock, u64 context, u64 seqno)
648{
649	BUG_ON(!lock);
650	BUG_ON(!ops || !ops->get_driver_name || !ops->get_timeline_name);
651
652	kref_init(&fence->refcount);
653	fence->ops = ops;
654	INIT_LIST_HEAD(&fence->cb_list);
655	fence->lock = lock;
656	fence->context = context;
657	fence->seqno = seqno;
658	fence->flags = 0UL;
659	fence->error = 0;
660
661	trace_dma_fence_init(fence);
662}
663EXPORT_SYMBOL(dma_fence_init);