1 // SPDX-License-Identifier: GPL-2.0-only
3 * Fence mechanism for dma-buf and to allow for asynchronous dma access
5 * Copyright (C) 2012 Canonical Ltd
6 * Copyright (C) 2012 Texas Instruments
9 * Rob Clark <robdclark@gmail.com>
10 * Maarten Lankhorst <maarten.lankhorst@canonical.com>
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>
19 #define CREATE_TRACE_POINTS
20 #include <trace/events/dma_fence.h>
22 EXPORT_TRACEPOINT_SYMBOL(dma_fence_emit);
23 EXPORT_TRACEPOINT_SYMBOL(dma_fence_enable_signal);
24 EXPORT_TRACEPOINT_SYMBOL(dma_fence_signaled);
26 static DEFINE_SPINLOCK(dma_fence_stub_lock);
27 static struct dma_fence dma_fence_stub;
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.
35 static atomic64_t dma_fence_context_counter = ATOMIC64_INIT(1);
38 * DOC: DMA fences overview
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.
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
49 * Since the purposes of fences is to facilitate cross-device and
50 * cross-application synchronization, there's multiple ways to use one:
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.
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.
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 reservation_object through the
64 * &dma_buf.resv pointer.
67 static const char *dma_fence_stub_get_name(struct dma_fence *fence)
72 static 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,
78 * dma_fence_get_stub - return a signaled fence
80 * Return a stub fence which is already signaled.
82 struct dma_fence *dma_fence_get_stub(void)
84 spin_lock(&dma_fence_stub_lock);
85 if (!dma_fence_stub.ops) {
86 dma_fence_init(&dma_fence_stub,
90 dma_fence_signal_locked(&dma_fence_stub);
92 spin_unlock(&dma_fence_stub_lock);
94 return dma_fence_get(&dma_fence_stub);
96 EXPORT_SYMBOL(dma_fence_get_stub);
99 * dma_fence_context_alloc - allocate an array of fence contexts
100 * @num: amount of contexts to allocate
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().
106 u64 dma_fence_context_alloc(unsigned num)
109 return atomic64_add_return(num, &dma_fence_context_counter) - num;
111 EXPORT_SYMBOL(dma_fence_context_alloc);
114 * dma_fence_signal_locked - signal completion of a fence
115 * @fence: the fence to signal
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.
123 * Unlike dma_fence_signal(), this function must be called with &dma_fence.lock
126 * Returns 0 on success and a negative error value when @fence has been
129 int dma_fence_signal_locked(struct dma_fence *fence)
131 struct dma_fence_cb *cur, *tmp;
134 lockdep_assert_held(fence->lock);
139 if (test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
143 * we might have raced with the unlocked dma_fence_signal,
144 * still run through all callbacks
147 fence->timestamp = ktime_get();
148 set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
149 trace_dma_fence_signaled(fence);
152 list_for_each_entry_safe(cur, tmp, &fence->cb_list, node) {
153 list_del_init(&cur->node);
154 cur->func(fence, cur);
158 EXPORT_SYMBOL(dma_fence_signal_locked);
161 * dma_fence_signal - signal completion of a fence
162 * @fence: the fence to signal
164 * Signal completion for software callbacks on a fence, this will unblock
165 * dma_fence_wait() calls and run all the callbacks added with
166 * dma_fence_add_callback(). Can be called multiple times, but since a fence
167 * can only go from the unsignaled to the signaled state and not back, it will
168 * only be effective the first time.
170 * Returns 0 on success and a negative error value when @fence has been
173 int dma_fence_signal(struct dma_fence *fence)
180 if (test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
183 fence->timestamp = ktime_get();
184 set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
185 trace_dma_fence_signaled(fence);
187 if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &fence->flags)) {
188 struct dma_fence_cb *cur, *tmp;
190 spin_lock_irqsave(fence->lock, flags);
191 list_for_each_entry_safe(cur, tmp, &fence->cb_list, node) {
192 list_del_init(&cur->node);
193 cur->func(fence, cur);
195 spin_unlock_irqrestore(fence->lock, flags);
199 EXPORT_SYMBOL(dma_fence_signal);
202 * dma_fence_wait_timeout - sleep until the fence gets signaled
203 * or until timeout elapses
204 * @fence: the fence to wait on
205 * @intr: if true, do an interruptible wait
206 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
208 * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or the
209 * remaining timeout in jiffies on success. Other error values may be
210 * returned on custom implementations.
212 * Performs a synchronous wait on this fence. It is assumed the caller
213 * directly or indirectly (buf-mgr between reservation and committing)
214 * holds a reference to the fence, otherwise the fence might be
215 * freed before return, resulting in undefined behavior.
217 * See also dma_fence_wait() and dma_fence_wait_any_timeout().
220 dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
224 if (WARN_ON(timeout < 0))
227 trace_dma_fence_wait_start(fence);
228 if (fence->ops->wait)
229 ret = fence->ops->wait(fence, intr, timeout);
231 ret = dma_fence_default_wait(fence, intr, timeout);
232 trace_dma_fence_wait_end(fence);
235 EXPORT_SYMBOL(dma_fence_wait_timeout);
238 * dma_fence_release - default relese function for fences
239 * @kref: &dma_fence.recfount
241 * This is the default release functions for &dma_fence. Drivers shouldn't call
242 * this directly, but instead call dma_fence_put().
244 void dma_fence_release(struct kref *kref)
246 struct dma_fence *fence =
247 container_of(kref, struct dma_fence, refcount);
249 trace_dma_fence_destroy(fence);
251 if (WARN(!list_empty(&fence->cb_list),
252 "Fence %s:%s:%llx:%llx released with pending signals!\n",
253 fence->ops->get_driver_name(fence),
254 fence->ops->get_timeline_name(fence),
255 fence->context, fence->seqno)) {
259 * Failed to signal before release, likely a refcounting issue.
261 * This should never happen, but if it does make sure that we
262 * don't leave chains dangling. We set the error flag first
263 * so that the callbacks know this signal is due to an error.
265 spin_lock_irqsave(fence->lock, flags);
266 fence->error = -EDEADLK;
267 dma_fence_signal_locked(fence);
268 spin_unlock_irqrestore(fence->lock, flags);
271 if (fence->ops->release)
272 fence->ops->release(fence);
274 dma_fence_free(fence);
276 EXPORT_SYMBOL(dma_fence_release);
279 * dma_fence_free - default release function for &dma_fence.
280 * @fence: fence to release
282 * This is the default implementation for &dma_fence_ops.release. It calls
283 * kfree_rcu() on @fence.
285 void dma_fence_free(struct dma_fence *fence)
287 kfree_rcu(fence, rcu);
289 EXPORT_SYMBOL(dma_fence_free);
292 * dma_fence_enable_sw_signaling - enable signaling on fence
293 * @fence: the fence to enable
295 * This will request for sw signaling to be enabled, to make the fence
296 * complete as soon as possible. This calls &dma_fence_ops.enable_signaling
299 void dma_fence_enable_sw_signaling(struct dma_fence *fence)
303 if (!test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
305 !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) &&
306 fence->ops->enable_signaling) {
307 trace_dma_fence_enable_signal(fence);
309 spin_lock_irqsave(fence->lock, flags);
311 if (!fence->ops->enable_signaling(fence))
312 dma_fence_signal_locked(fence);
314 spin_unlock_irqrestore(fence->lock, flags);
317 EXPORT_SYMBOL(dma_fence_enable_sw_signaling);
320 * dma_fence_add_callback - add a callback to be called when the fence
322 * @fence: the fence to wait on
323 * @cb: the callback to register
324 * @func: the function to call
326 * @cb will be initialized by dma_fence_add_callback(), no initialization
327 * by the caller is required. Any number of callbacks can be registered
328 * to a fence, but a callback can only be registered to one fence at a time.
330 * Note that the callback can be called from an atomic context. If
331 * fence is already signaled, this function will return -ENOENT (and
332 * *not* call the callback).
334 * Add a software callback to the fence. Same restrictions apply to
335 * refcount as it does to dma_fence_wait(), however the caller doesn't need to
336 * keep a refcount to fence afterward dma_fence_add_callback() has returned:
337 * when software access is enabled, the creator of the fence is required to keep
338 * the fence alive until after it signals with dma_fence_signal(). The callback
339 * itself can be called from irq context.
341 * Returns 0 in case of success, -ENOENT if the fence is already signaled
342 * and -EINVAL in case of error.
344 int dma_fence_add_callback(struct dma_fence *fence, struct dma_fence_cb *cb,
345 dma_fence_func_t func)
351 if (WARN_ON(!fence || !func))
354 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
355 INIT_LIST_HEAD(&cb->node);
359 spin_lock_irqsave(fence->lock, flags);
361 was_set = test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
364 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
366 else if (!was_set && fence->ops->enable_signaling) {
367 trace_dma_fence_enable_signal(fence);
369 if (!fence->ops->enable_signaling(fence)) {
370 dma_fence_signal_locked(fence);
377 list_add_tail(&cb->node, &fence->cb_list);
379 INIT_LIST_HEAD(&cb->node);
380 spin_unlock_irqrestore(fence->lock, flags);
384 EXPORT_SYMBOL(dma_fence_add_callback);
387 * dma_fence_get_status - returns the status upon completion
388 * @fence: the dma_fence to query
390 * This wraps dma_fence_get_status_locked() to return the error status
391 * condition on a signaled fence. See dma_fence_get_status_locked() for more
394 * Returns 0 if the fence has not yet been signaled, 1 if the fence has
395 * been signaled without an error condition, or a negative error code
396 * if the fence has been completed in err.
398 int dma_fence_get_status(struct dma_fence *fence)
403 spin_lock_irqsave(fence->lock, flags);
404 status = dma_fence_get_status_locked(fence);
405 spin_unlock_irqrestore(fence->lock, flags);
409 EXPORT_SYMBOL(dma_fence_get_status);
412 * dma_fence_remove_callback - remove a callback from the signaling list
413 * @fence: the fence to wait on
414 * @cb: the callback to remove
416 * Remove a previously queued callback from the fence. This function returns
417 * true if the callback is successfully removed, or false if the fence has
418 * already been signaled.
421 * Cancelling a callback should only be done if you really know what you're
422 * doing, since deadlocks and race conditions could occur all too easily. For
423 * this reason, it should only ever be done on hardware lockup recovery,
424 * with a reference held to the fence.
426 * Behaviour is undefined if @cb has not been added to @fence using
427 * dma_fence_add_callback() beforehand.
430 dma_fence_remove_callback(struct dma_fence *fence, struct dma_fence_cb *cb)
435 spin_lock_irqsave(fence->lock, flags);
437 ret = !list_empty(&cb->node);
439 list_del_init(&cb->node);
441 spin_unlock_irqrestore(fence->lock, flags);
445 EXPORT_SYMBOL(dma_fence_remove_callback);
447 struct default_wait_cb {
448 struct dma_fence_cb base;
449 struct task_struct *task;
453 dma_fence_default_wait_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
455 struct default_wait_cb *wait =
456 container_of(cb, struct default_wait_cb, base);
458 wake_up_state(wait->task, TASK_NORMAL);
462 * dma_fence_default_wait - default sleep until the fence gets signaled
463 * or until timeout elapses
464 * @fence: the fence to wait on
465 * @intr: if true, do an interruptible wait
466 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
468 * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or the
469 * remaining timeout in jiffies on success. If timeout is zero the value one is
470 * returned if the fence is already signaled for consistency with other
471 * functions taking a jiffies timeout.
474 dma_fence_default_wait(struct dma_fence *fence, bool intr, signed long timeout)
476 struct default_wait_cb cb;
478 signed long ret = timeout ? timeout : 1;
481 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
484 spin_lock_irqsave(fence->lock, flags);
486 if (intr && signal_pending(current)) {
491 was_set = test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
494 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
497 if (!was_set && fence->ops->enable_signaling) {
498 trace_dma_fence_enable_signal(fence);
500 if (!fence->ops->enable_signaling(fence)) {
501 dma_fence_signal_locked(fence);
511 cb.base.func = dma_fence_default_wait_cb;
513 list_add(&cb.base.node, &fence->cb_list);
515 while (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) && ret > 0) {
517 __set_current_state(TASK_INTERRUPTIBLE);
519 __set_current_state(TASK_UNINTERRUPTIBLE);
520 spin_unlock_irqrestore(fence->lock, flags);
522 ret = schedule_timeout(ret);
524 spin_lock_irqsave(fence->lock, flags);
525 if (ret > 0 && intr && signal_pending(current))
529 if (!list_empty(&cb.base.node))
530 list_del(&cb.base.node);
531 __set_current_state(TASK_RUNNING);
534 spin_unlock_irqrestore(fence->lock, flags);
537 EXPORT_SYMBOL(dma_fence_default_wait);
540 dma_fence_test_signaled_any(struct dma_fence **fences, uint32_t count,
545 for (i = 0; i < count; ++i) {
546 struct dma_fence *fence = fences[i];
547 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
557 * dma_fence_wait_any_timeout - sleep until any fence gets signaled
558 * or until timeout elapses
559 * @fences: array of fences to wait on
560 * @count: number of fences to wait on
561 * @intr: if true, do an interruptible wait
562 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
563 * @idx: used to store the first signaled fence index, meaningful only on
566 * Returns -EINVAL on custom fence wait implementation, -ERESTARTSYS if
567 * interrupted, 0 if the wait timed out, or the remaining timeout in jiffies
570 * Synchronous waits for the first fence in the array to be signaled. The
571 * caller needs to hold a reference to all fences in the array, otherwise a
572 * fence might be freed before return, resulting in undefined behavior.
574 * See also dma_fence_wait() and dma_fence_wait_timeout().
577 dma_fence_wait_any_timeout(struct dma_fence **fences, uint32_t count,
578 bool intr, signed long timeout, uint32_t *idx)
580 struct default_wait_cb *cb;
581 signed long ret = timeout;
584 if (WARN_ON(!fences || !count || timeout < 0))
588 for (i = 0; i < count; ++i)
589 if (dma_fence_is_signaled(fences[i])) {
598 cb = kcalloc(count, sizeof(struct default_wait_cb), GFP_KERNEL);
604 for (i = 0; i < count; ++i) {
605 struct dma_fence *fence = fences[i];
607 cb[i].task = current;
608 if (dma_fence_add_callback(fence, &cb[i].base,
609 dma_fence_default_wait_cb)) {
610 /* This fence is already signaled */
619 set_current_state(TASK_INTERRUPTIBLE);
621 set_current_state(TASK_UNINTERRUPTIBLE);
623 if (dma_fence_test_signaled_any(fences, count, idx))
626 ret = schedule_timeout(ret);
628 if (ret > 0 && intr && signal_pending(current))
632 __set_current_state(TASK_RUNNING);
636 dma_fence_remove_callback(fences[i], &cb[i].base);
643 EXPORT_SYMBOL(dma_fence_wait_any_timeout);
646 * dma_fence_init - Initialize a custom fence.
647 * @fence: the fence to initialize
648 * @ops: the dma_fence_ops for operations on this fence
649 * @lock: the irqsafe spinlock to use for locking this fence
650 * @context: the execution context this fence is run on
651 * @seqno: a linear increasing sequence number for this context
653 * Initializes an allocated fence, the caller doesn't have to keep its
654 * refcount after committing with this fence, but it will need to hold a
655 * refcount again if &dma_fence_ops.enable_signaling gets called.
657 * context and seqno are used for easy comparison between fences, allowing
658 * to check which fence is later by simply using dma_fence_later().
661 dma_fence_init(struct dma_fence *fence, const struct dma_fence_ops *ops,
662 spinlock_t *lock, u64 context, u64 seqno)
665 BUG_ON(!ops || !ops->get_driver_name || !ops->get_timeline_name);
667 kref_init(&fence->refcount);
669 INIT_LIST_HEAD(&fence->cb_list);
671 fence->context = context;
672 fence->seqno = seqno;
676 trace_dma_fence_init(fence);
678 EXPORT_SYMBOL(dma_fence_init);