1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Fence mechanism for dma-buf 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#ifndef __LINUX_DMA_FENCE_H
14#define __LINUX_DMA_FENCE_H
15
16#include <linux/err.h>
17#include <linux/wait.h>
18#include <linux/list.h>
19#include <linux/bitops.h>
20#include <linux/kref.h>
21#include <linux/sched.h>
22#include <linux/printk.h>
23#include <linux/rcupdate.h>
24
25struct dma_fence;
26struct dma_fence_ops;
27struct dma_fence_cb;
28
29/**
30 * struct dma_fence - software synchronization primitive
31 * @refcount: refcount for this fence
32 * @ops: dma_fence_ops associated with this fence
33 * @rcu: used for releasing fence with kfree_rcu
34 * @cb_list: list of all callbacks to call
35 * @lock: spin_lock_irqsave used for locking
36 * @context: execution context this fence belongs to, returned by
37 * dma_fence_context_alloc()
38 * @seqno: the sequence number of this fence inside the execution context,
39 * can be compared to decide which fence would be signaled later.
40 * @flags: A mask of DMA_FENCE_FLAG_* defined below
41 * @timestamp: Timestamp when the fence was signaled.
42 * @error: Optional, only valid if < 0, must be set before calling
43 * dma_fence_signal, indicates that the fence has completed with an error.
44 *
45 * the flags member must be manipulated and read using the appropriate
46 * atomic ops (bit_*), so taking the spinlock will not be needed most
47 * of the time.
48 *
49 * DMA_FENCE_FLAG_SIGNALED_BIT - fence is already signaled
50 * DMA_FENCE_FLAG_TIMESTAMP_BIT - timestamp recorded for fence signaling
51 * DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT - enable_signaling might have been called
52 * DMA_FENCE_FLAG_USER_BITS - start of the unused bits, can be used by the
53 * implementer of the fence for its own purposes. Can be used in different
54 * ways by different fence implementers, so do not rely on this.
55 *
56 * Since atomic bitops are used, this is not guaranteed to be the case.
57 * Particularly, if the bit was set, but dma_fence_signal was called right
58 * before this bit was set, it would have been able to set the
59 * DMA_FENCE_FLAG_SIGNALED_BIT, before enable_signaling was called.
60 * Adding a check for DMA_FENCE_FLAG_SIGNALED_BIT after setting
61 * DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT closes this race, and makes sure that
62 * after dma_fence_signal was called, any enable_signaling call will have either
63 * been completed, or never called at all.
64 */
65struct dma_fence {
66 spinlock_t *lock;
67 const struct dma_fence_ops *ops;
68 /*
69 * We clear the callback list on kref_put so that by the time we
70 * release the fence it is unused. No one should be adding to the
71 * cb_list that they don't themselves hold a reference for.
72 *
73 * The lifetime of the timestamp is similarly tied to both the
74 * rcu freelist and the cb_list. The timestamp is only set upon
75 * signaling while simultaneously notifying the cb_list. Ergo, we
76 * only use either the cb_list of timestamp. Upon destruction,
77 * neither are accessible, and so we can use the rcu. This means
78 * that the cb_list is *only* valid until the signal bit is set,
79 * and to read either you *must* hold a reference to the fence,
80 * and not just the rcu_read_lock.
81 *
82 * Listed in chronological order.
83 */
84 union {
85 struct list_head cb_list;
86 /* @cb_list replaced by @timestamp on dma_fence_signal() */
87 ktime_t timestamp;
88 /* @timestamp replaced by @rcu on dma_fence_release() */
89 struct rcu_head rcu;
90 };
91 u64 context;
92 u64 seqno;
93 unsigned long flags;
94 struct kref refcount;
95 int error;
96};
97
98enum dma_fence_flag_bits {
99 DMA_FENCE_FLAG_SIGNALED_BIT,
100 DMA_FENCE_FLAG_TIMESTAMP_BIT,
101 DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
102 DMA_FENCE_FLAG_USER_BITS, /* must always be last member */
103};
104
105typedef void (*dma_fence_func_t)(struct dma_fence *fence,
106 struct dma_fence_cb *cb);
107
108/**
109 * struct dma_fence_cb - callback for dma_fence_add_callback()
110 * @node: used by dma_fence_add_callback() to append this struct to fence::cb_list
111 * @func: dma_fence_func_t to call
112 *
113 * This struct will be initialized by dma_fence_add_callback(), additional
114 * data can be passed along by embedding dma_fence_cb in another struct.
115 */
116struct dma_fence_cb {
117 struct list_head node;
118 dma_fence_func_t func;
119};
120
121/**
122 * struct dma_fence_ops - operations implemented for fence
123 *
124 */
125struct dma_fence_ops {
126 /**
127 * @use_64bit_seqno:
128 *
129 * True if this dma_fence implementation uses 64bit seqno, false
130 * otherwise.
131 */
132 bool use_64bit_seqno;
133
134 /**
135 * @get_driver_name:
136 *
137 * Returns the driver name. This is a callback to allow drivers to
138 * compute the name at runtime, without having it to store permanently
139 * for each fence, or build a cache of some sort.
140 *
141 * This callback is mandatory.
142 */
143 const char * (*get_driver_name)(struct dma_fence *fence);
144
145 /**
146 * @get_timeline_name:
147 *
148 * Return the name of the context this fence belongs to. This is a
149 * callback to allow drivers to compute the name at runtime, without
150 * having it to store permanently for each fence, or build a cache of
151 * some sort.
152 *
153 * This callback is mandatory.
154 */
155 const char * (*get_timeline_name)(struct dma_fence *fence);
156
157 /**
158 * @enable_signaling:
159 *
160 * Enable software signaling of fence.
161 *
162 * For fence implementations that have the capability for hw->hw
163 * signaling, they can implement this op to enable the necessary
164 * interrupts, or insert commands into cmdstream, etc, to avoid these
165 * costly operations for the common case where only hw->hw
166 * synchronization is required. This is called in the first
167 * dma_fence_wait() or dma_fence_add_callback() path to let the fence
168 * implementation know that there is another driver waiting on the
169 * signal (ie. hw->sw case).
170 *
171 * This function can be called from atomic context, but not
172 * from irq context, so normal spinlocks can be used.
173 *
174 * A return value of false indicates the fence already passed,
175 * or some failure occurred that made it impossible to enable
176 * signaling. True indicates successful enabling.
177 *
178 * &dma_fence.error may be set in enable_signaling, but only when false
179 * is returned.
180 *
181 * Since many implementations can call dma_fence_signal() even when before
182 * @enable_signaling has been called there's a race window, where the
183 * dma_fence_signal() might result in the final fence reference being
184 * released and its memory freed. To avoid this, implementations of this
185 * callback should grab their own reference using dma_fence_get(), to be
186 * released when the fence is signalled (through e.g. the interrupt
187 * handler).
188 *
189 * This callback is optional. If this callback is not present, then the
190 * driver must always have signaling enabled.
191 */
192 bool (*enable_signaling)(struct dma_fence *fence);
193
194 /**
195 * @signaled:
196 *
197 * Peek whether the fence is signaled, as a fastpath optimization for
198 * e.g. dma_fence_wait() or dma_fence_add_callback(). Note that this
199 * callback does not need to make any guarantees beyond that a fence
200 * once indicates as signalled must always return true from this
201 * callback. This callback may return false even if the fence has
202 * completed already, in this case information hasn't propogated throug
203 * the system yet. See also dma_fence_is_signaled().
204 *
205 * May set &dma_fence.error if returning true.
206 *
207 * This callback is optional.
208 */
209 bool (*signaled)(struct dma_fence *fence);
210
211 /**
212 * @wait:
213 *
214 * Custom wait implementation, defaults to dma_fence_default_wait() if
215 * not set.
216 *
217 * Deprecated and should not be used by new implementations. Only used
218 * by existing implementations which need special handling for their
219 * hardware reset procedure.
220 *
221 * Must return -ERESTARTSYS if the wait is intr = true and the wait was
222 * interrupted, and remaining jiffies if fence has signaled, or 0 if wait
223 * timed out. Can also return other error values on custom implementations,
224 * which should be treated as if the fence is signaled. For example a hardware
225 * lockup could be reported like that.
226 */
227 signed long (*wait)(struct dma_fence *fence,
228 bool intr, signed long timeout);
229
230 /**
231 * @release:
232 *
233 * Called on destruction of fence to release additional resources.
234 * Can be called from irq context. This callback is optional. If it is
235 * NULL, then dma_fence_free() is instead called as the default
236 * implementation.
237 */
238 void (*release)(struct dma_fence *fence);
239
240 /**
241 * @fence_value_str:
242 *
243 * Callback to fill in free-form debug info specific to this fence, like
244 * the sequence number.
245 *
246 * This callback is optional.
247 */
248 void (*fence_value_str)(struct dma_fence *fence, char *str, int size);
249
250 /**
251 * @timeline_value_str:
252 *
253 * Fills in the current value of the timeline as a string, like the
254 * sequence number. Note that the specific fence passed to this function
255 * should not matter, drivers should only use it to look up the
256 * corresponding timeline structures.
257 */
258 void (*timeline_value_str)(struct dma_fence *fence,
259 char *str, int size);
260
261 /**
262 * @set_deadline:
263 *
264 * Callback to allow a fence waiter to inform the fence signaler of
265 * an upcoming deadline, such as vblank, by which point the waiter
266 * would prefer the fence to be signaled by. This is intended to
267 * give feedback to the fence signaler to aid in power management
268 * decisions, such as boosting GPU frequency.
269 *
270 * This is called without &dma_fence.lock held, it can be called
271 * multiple times and from any context. Locking is up to the callee
272 * if it has some state to manage. If multiple deadlines are set,
273 * the expectation is to track the soonest one. If the deadline is
274 * before the current time, it should be interpreted as an immediate
275 * deadline.
276 *
277 * This callback is optional.
278 */
279 void (*set_deadline)(struct dma_fence *fence, ktime_t deadline);
280};
281
282void dma_fence_init(struct dma_fence *fence, const struct dma_fence_ops *ops,
283 spinlock_t *lock, u64 context, u64 seqno);
284
285void dma_fence_release(struct kref *kref);
286void dma_fence_free(struct dma_fence *fence);
287void dma_fence_describe(struct dma_fence *fence, struct seq_file *seq);
288
289/**
290 * dma_fence_put - decreases refcount of the fence
291 * @fence: fence to reduce refcount of
292 */
293static inline void dma_fence_put(struct dma_fence *fence)
294{
295 if (fence)
296 kref_put(kref: &fence->refcount, release: dma_fence_release);
297}
298
299/**
300 * dma_fence_get - increases refcount of the fence
301 * @fence: fence to increase refcount of
302 *
303 * Returns the same fence, with refcount increased by 1.
304 */
305static inline struct dma_fence *dma_fence_get(struct dma_fence *fence)
306{
307 if (fence)
308 kref_get(kref: &fence->refcount);
309 return fence;
310}
311
312/**
313 * dma_fence_get_rcu - get a fence from a dma_resv_list with
314 * rcu read lock
315 * @fence: fence to increase refcount of
316 *
317 * Function returns NULL if no refcount could be obtained, or the fence.
318 */
319static inline struct dma_fence *dma_fence_get_rcu(struct dma_fence *fence)
320{
321 if (kref_get_unless_zero(kref: &fence->refcount))
322 return fence;
323 else
324 return NULL;
325}
326
327/**
328 * dma_fence_get_rcu_safe - acquire a reference to an RCU tracked fence
329 * @fencep: pointer to fence to increase refcount of
330 *
331 * Function returns NULL if no refcount could be obtained, or the fence.
332 * This function handles acquiring a reference to a fence that may be
333 * reallocated within the RCU grace period (such as with SLAB_TYPESAFE_BY_RCU),
334 * so long as the caller is using RCU on the pointer to the fence.
335 *
336 * An alternative mechanism is to employ a seqlock to protect a bunch of
337 * fences, such as used by struct dma_resv. When using a seqlock,
338 * the seqlock must be taken before and checked after a reference to the
339 * fence is acquired (as shown here).
340 *
341 * The caller is required to hold the RCU read lock.
342 */
343static inline struct dma_fence *
344dma_fence_get_rcu_safe(struct dma_fence __rcu **fencep)
345{
346 do {
347 struct dma_fence *fence;
348
349 fence = rcu_dereference(*fencep);
350 if (!fence)
351 return NULL;
352
353 if (!dma_fence_get_rcu(fence))
354 continue;
355
356 /* The atomic_inc_not_zero() inside dma_fence_get_rcu()
357 * provides a full memory barrier upon success (such as now).
358 * This is paired with the write barrier from assigning
359 * to the __rcu protected fence pointer so that if that
360 * pointer still matches the current fence, we know we
361 * have successfully acquire a reference to it. If it no
362 * longer matches, we are holding a reference to some other
363 * reallocated pointer. This is possible if the allocator
364 * is using a freelist like SLAB_TYPESAFE_BY_RCU where the
365 * fence remains valid for the RCU grace period, but it
366 * may be reallocated. When using such allocators, we are
367 * responsible for ensuring the reference we get is to
368 * the right fence, as below.
369 */
370 if (fence == rcu_access_pointer(*fencep))
371 return rcu_pointer_handoff(fence);
372
373 dma_fence_put(fence);
374 } while (1);
375}
376
377#ifdef CONFIG_LOCKDEP
378bool dma_fence_begin_signalling(void);
379void dma_fence_end_signalling(bool cookie);
380void __dma_fence_might_wait(void);
381#else
382static inline bool dma_fence_begin_signalling(void)
383{
384 return true;
385}
386static inline void dma_fence_end_signalling(bool cookie) {}
387static inline void __dma_fence_might_wait(void) {}
388#endif
389
390int dma_fence_signal(struct dma_fence *fence);
391int dma_fence_signal_locked(struct dma_fence *fence);
392int dma_fence_signal_timestamp(struct dma_fence *fence, ktime_t timestamp);
393int dma_fence_signal_timestamp_locked(struct dma_fence *fence,
394 ktime_t timestamp);
395signed long dma_fence_default_wait(struct dma_fence *fence,
396 bool intr, signed long timeout);
397int dma_fence_add_callback(struct dma_fence *fence,
398 struct dma_fence_cb *cb,
399 dma_fence_func_t func);
400bool dma_fence_remove_callback(struct dma_fence *fence,
401 struct dma_fence_cb *cb);
402void dma_fence_enable_sw_signaling(struct dma_fence *fence);
403
404/**
405 * dma_fence_is_signaled_locked - Return an indication if the fence
406 * is signaled yet.
407 * @fence: the fence to check
408 *
409 * Returns true if the fence was already signaled, false if not. Since this
410 * function doesn't enable signaling, it is not guaranteed to ever return
411 * true if dma_fence_add_callback(), dma_fence_wait() or
412 * dma_fence_enable_sw_signaling() haven't been called before.
413 *
414 * This function requires &dma_fence.lock to be held.
415 *
416 * See also dma_fence_is_signaled().
417 */
418static inline bool
419dma_fence_is_signaled_locked(struct dma_fence *fence)
420{
421 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
422 return true;
423
424 if (fence->ops->signaled && fence->ops->signaled(fence)) {
425 dma_fence_signal_locked(fence);
426 return true;
427 }
428
429 return false;
430}
431
432/**
433 * dma_fence_is_signaled - Return an indication if the fence is signaled yet.
434 * @fence: the fence to check
435 *
436 * Returns true if the fence was already signaled, false if not. Since this
437 * function doesn't enable signaling, it is not guaranteed to ever return
438 * true if dma_fence_add_callback(), dma_fence_wait() or
439 * dma_fence_enable_sw_signaling() haven't been called before.
440 *
441 * It's recommended for seqno fences to call dma_fence_signal when the
442 * operation is complete, it makes it possible to prevent issues from
443 * wraparound between time of issue and time of use by checking the return
444 * value of this function before calling hardware-specific wait instructions.
445 *
446 * See also dma_fence_is_signaled_locked().
447 */
448static inline bool
449dma_fence_is_signaled(struct dma_fence *fence)
450{
451 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
452 return true;
453
454 if (fence->ops->signaled && fence->ops->signaled(fence)) {
455 dma_fence_signal(fence);
456 return true;
457 }
458
459 return false;
460}
461
462/**
463 * __dma_fence_is_later - return if f1 is chronologically later than f2
464 * @f1: the first fence's seqno
465 * @f2: the second fence's seqno from the same context
466 * @ops: dma_fence_ops associated with the seqno
467 *
468 * Returns true if f1 is chronologically later than f2. Both fences must be
469 * from the same context, since a seqno is not common across contexts.
470 */
471static inline bool __dma_fence_is_later(u64 f1, u64 f2,
472 const struct dma_fence_ops *ops)
473{
474 /* This is for backward compatibility with drivers which can only handle
475 * 32bit sequence numbers. Use a 64bit compare when the driver says to
476 * do so.
477 */
478 if (ops->use_64bit_seqno)
479 return f1 > f2;
480
481 return (int)(lower_32_bits(f1) - lower_32_bits(f2)) > 0;
482}
483
484/**
485 * dma_fence_is_later - return if f1 is chronologically later than f2
486 * @f1: the first fence from the same context
487 * @f2: the second fence from the same context
488 *
489 * Returns true if f1 is chronologically later than f2. Both fences must be
490 * from the same context, since a seqno is not re-used across contexts.
491 */
492static inline bool dma_fence_is_later(struct dma_fence *f1,
493 struct dma_fence *f2)
494{
495 if (WARN_ON(f1->context != f2->context))
496 return false;
497
498 return __dma_fence_is_later(f1: f1->seqno, f2: f2->seqno, ops: f1->ops);
499}
500
501/**
502 * dma_fence_later - return the chronologically later fence
503 * @f1: the first fence from the same context
504 * @f2: the second fence from the same context
505 *
506 * Returns NULL if both fences are signaled, otherwise the fence that would be
507 * signaled last. Both fences must be from the same context, since a seqno is
508 * not re-used across contexts.
509 */
510static inline struct dma_fence *dma_fence_later(struct dma_fence *f1,
511 struct dma_fence *f2)
512{
513 if (WARN_ON(f1->context != f2->context))
514 return NULL;
515
516 /*
517 * Can't check just DMA_FENCE_FLAG_SIGNALED_BIT here, it may never
518 * have been set if enable_signaling wasn't called, and enabling that
519 * here is overkill.
520 */
521 if (dma_fence_is_later(f1, f2))
522 return dma_fence_is_signaled(fence: f1) ? NULL : f1;
523 else
524 return dma_fence_is_signaled(fence: f2) ? NULL : f2;
525}
526
527/**
528 * dma_fence_get_status_locked - returns the status upon completion
529 * @fence: the dma_fence to query
530 *
531 * Drivers can supply an optional error status condition before they signal
532 * the fence (to indicate whether the fence was completed due to an error
533 * rather than success). The value of the status condition is only valid
534 * if the fence has been signaled, dma_fence_get_status_locked() first checks
535 * the signal state before reporting the error status.
536 *
537 * Returns 0 if the fence has not yet been signaled, 1 if the fence has
538 * been signaled without an error condition, or a negative error code
539 * if the fence has been completed in err.
540 */
541static inline int dma_fence_get_status_locked(struct dma_fence *fence)
542{
543 if (dma_fence_is_signaled_locked(fence))
544 return fence->error ?: 1;
545 else
546 return 0;
547}
548
549int dma_fence_get_status(struct dma_fence *fence);
550
551/**
552 * dma_fence_set_error - flag an error condition on the fence
553 * @fence: the dma_fence
554 * @error: the error to store
555 *
556 * Drivers can supply an optional error status condition before they signal
557 * the fence, to indicate that the fence was completed due to an error
558 * rather than success. This must be set before signaling (so that the value
559 * is visible before any waiters on the signal callback are woken). This
560 * helper exists to help catching erroneous setting of #dma_fence.error.
561 */
562static inline void dma_fence_set_error(struct dma_fence *fence,
563 int error)
564{
565 WARN_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
566 WARN_ON(error >= 0 || error < -MAX_ERRNO);
567
568 fence->error = error;
569}
570
571/**
572 * dma_fence_timestamp - helper to get the completion timestamp of a fence
573 * @fence: fence to get the timestamp from.
574 *
575 * After a fence is signaled the timestamp is updated with the signaling time,
576 * but setting the timestamp can race with tasks waiting for the signaling. This
577 * helper busy waits for the correct timestamp to appear.
578 */
579static inline ktime_t dma_fence_timestamp(struct dma_fence *fence)
580{
581 if (WARN_ON(!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)))
582 return ktime_get();
583
584 while (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags))
585 cpu_relax();
586
587 return fence->timestamp;
588}
589
590signed long dma_fence_wait_timeout(struct dma_fence *,
591 bool intr, signed long timeout);
592signed long dma_fence_wait_any_timeout(struct dma_fence **fences,
593 uint32_t count,
594 bool intr, signed long timeout,
595 uint32_t *idx);
596
597/**
598 * dma_fence_wait - sleep until the fence gets signaled
599 * @fence: the fence to wait on
600 * @intr: if true, do an interruptible wait
601 *
602 * This function will return -ERESTARTSYS if interrupted by a signal,
603 * or 0 if the fence was signaled. Other error values may be
604 * returned on custom implementations.
605 *
606 * Performs a synchronous wait on this fence. It is assumed the caller
607 * directly or indirectly holds a reference to the fence, otherwise the
608 * fence might be freed before return, resulting in undefined behavior.
609 *
610 * See also dma_fence_wait_timeout() and dma_fence_wait_any_timeout().
611 */
612static inline signed long dma_fence_wait(struct dma_fence *fence, bool intr)
613{
614 signed long ret;
615
616 /* Since dma_fence_wait_timeout cannot timeout with
617 * MAX_SCHEDULE_TIMEOUT, only valid return values are
618 * -ERESTARTSYS and MAX_SCHEDULE_TIMEOUT.
619 */
620 ret = dma_fence_wait_timeout(fence, intr, MAX_SCHEDULE_TIMEOUT);
621
622 return ret < 0 ? ret : 0;
623}
624
625void dma_fence_set_deadline(struct dma_fence *fence, ktime_t deadline);
626
627struct dma_fence *dma_fence_get_stub(void);
628struct dma_fence *dma_fence_allocate_private_stub(ktime_t timestamp);
629u64 dma_fence_context_alloc(unsigned num);
630
631extern const struct dma_fence_ops dma_fence_array_ops;
632extern const struct dma_fence_ops dma_fence_chain_ops;
633
634/**
635 * dma_fence_is_array - check if a fence is from the array subclass
636 * @fence: the fence to test
637 *
638 * Return true if it is a dma_fence_array and false otherwise.
639 */
640static inline bool dma_fence_is_array(struct dma_fence *fence)
641{
642 return fence->ops == &dma_fence_array_ops;
643}
644
645/**
646 * dma_fence_is_chain - check if a fence is from the chain subclass
647 * @fence: the fence to test
648 *
649 * Return true if it is a dma_fence_chain and false otherwise.
650 */
651static inline bool dma_fence_is_chain(struct dma_fence *fence)
652{
653 return fence->ops == &dma_fence_chain_ops;
654}
655
656/**
657 * dma_fence_is_container - check if a fence is a container for other fences
658 * @fence: the fence to test
659 *
660 * Return true if this fence is a container for other fences, false otherwise.
661 * This is important since we can't build up large fence structure or otherwise
662 * we run into recursion during operation on those fences.
663 */
664static inline bool dma_fence_is_container(struct dma_fence *fence)
665{
666 return dma_fence_is_array(fence) || dma_fence_is_chain(fence);
667}
668
669#endif /* __LINUX_DMA_FENCE_H */
670

source code of linux/include/linux/dma-fence.h