1/*
2 * Fence mechanism for dma-buf 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#ifndef __LINUX_DMA_FENCE_H
22#define __LINUX_DMA_FENCE_H
23
24#include <linux/err.h>
25#include <linux/wait.h>
26#include <linux/list.h>
27#include <linux/bitops.h>
28#include <linux/kref.h>
29#include <linux/sched.h>
30#include <linux/printk.h>
31#include <linux/rcupdate.h>
32
33struct dma_fence;
34struct dma_fence_ops;
35struct dma_fence_cb;
36
37/**
38 * struct dma_fence - software synchronization primitive
39 * @refcount: refcount for this fence
40 * @ops: dma_fence_ops associated with this fence
41 * @rcu: used for releasing fence with kfree_rcu
42 * @cb_list: list of all callbacks to call
43 * @lock: spin_lock_irqsave used for locking
44 * @context: execution context this fence belongs to, returned by
45 * dma_fence_context_alloc()
46 * @seqno: the sequence number of this fence inside the execution context,
47 * can be compared to decide which fence would be signaled later.
48 * @flags: A mask of DMA_FENCE_FLAG_* defined below
49 * @timestamp: Timestamp when the fence was signaled.
50 * @error: Optional, only valid if < 0, must be set before calling
51 * dma_fence_signal, indicates that the fence has completed with an error.
52 *
53 * the flags member must be manipulated and read using the appropriate
54 * atomic ops (bit_*), so taking the spinlock will not be needed most
55 * of the time.
56 *
57 * DMA_FENCE_FLAG_SIGNALED_BIT - fence is already signaled
58 * DMA_FENCE_FLAG_TIMESTAMP_BIT - timestamp recorded for fence signaling
59 * DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT - enable_signaling might have been called
60 * DMA_FENCE_FLAG_USER_BITS - start of the unused bits, can be used by the
61 * implementer of the fence for its own purposes. Can be used in different
62 * ways by different fence implementers, so do not rely on this.
63 *
64 * Since atomic bitops are used, this is not guaranteed to be the case.
65 * Particularly, if the bit was set, but dma_fence_signal was called right
66 * before this bit was set, it would have been able to set the
67 * DMA_FENCE_FLAG_SIGNALED_BIT, before enable_signaling was called.
68 * Adding a check for DMA_FENCE_FLAG_SIGNALED_BIT after setting
69 * DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT closes this race, and makes sure that
70 * after dma_fence_signal was called, any enable_signaling call will have either
71 * been completed, or never called at all.
72 */
73struct dma_fence {
74 struct kref refcount;
75 const struct dma_fence_ops *ops;
76 struct rcu_head rcu;
77 struct list_head cb_list;
78 spinlock_t *lock;
79 u64 context;
80 u64 seqno;
81 unsigned long flags;
82 ktime_t timestamp;
83 int error;
84};
85
86enum dma_fence_flag_bits {
87 DMA_FENCE_FLAG_SIGNALED_BIT,
88 DMA_FENCE_FLAG_TIMESTAMP_BIT,
89 DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
90 DMA_FENCE_FLAG_USER_BITS, /* must always be last member */
91};
92
93typedef void (*dma_fence_func_t)(struct dma_fence *fence,
94 struct dma_fence_cb *cb);
95
96/**
97 * struct dma_fence_cb - callback for dma_fence_add_callback()
98 * @node: used by dma_fence_add_callback() to append this struct to fence::cb_list
99 * @func: dma_fence_func_t to call
100 *
101 * This struct will be initialized by dma_fence_add_callback(), additional
102 * data can be passed along by embedding dma_fence_cb in another struct.
103 */
104struct dma_fence_cb {
105 struct list_head node;
106 dma_fence_func_t func;
107};
108
109/**
110 * struct dma_fence_ops - operations implemented for fence
111 *
112 */
113struct dma_fence_ops {
114 /**
115 * @get_driver_name:
116 *
117 * Returns the driver name. This is a callback to allow drivers to
118 * compute the name at runtime, without having it to store permanently
119 * for each fence, or build a cache of some sort.
120 *
121 * This callback is mandatory.
122 */
123 const char * (*get_driver_name)(struct dma_fence *fence);
124
125 /**
126 * @get_timeline_name:
127 *
128 * Return the name of the context this fence belongs to. This is a
129 * callback to allow drivers to compute the name at runtime, without
130 * having it to store permanently for each fence, or build a cache of
131 * some sort.
132 *
133 * This callback is mandatory.
134 */
135 const char * (*get_timeline_name)(struct dma_fence *fence);
136
137 /**
138 * @enable_signaling:
139 *
140 * Enable software signaling of fence.
141 *
142 * For fence implementations that have the capability for hw->hw
143 * signaling, they can implement this op to enable the necessary
144 * interrupts, or insert commands into cmdstream, etc, to avoid these
145 * costly operations for the common case where only hw->hw
146 * synchronization is required. This is called in the first
147 * dma_fence_wait() or dma_fence_add_callback() path to let the fence
148 * implementation know that there is another driver waiting on the
149 * signal (ie. hw->sw case).
150 *
151 * This function can be called from atomic context, but not
152 * from irq context, so normal spinlocks can be used.
153 *
154 * A return value of false indicates the fence already passed,
155 * or some failure occurred that made it impossible to enable
156 * signaling. True indicates successful enabling.
157 *
158 * &dma_fence.error may be set in enable_signaling, but only when false
159 * is returned.
160 *
161 * Since many implementations can call dma_fence_signal() even when before
162 * @enable_signaling has been called there's a race window, where the
163 * dma_fence_signal() might result in the final fence reference being
164 * released and its memory freed. To avoid this, implementations of this
165 * callback should grab their own reference using dma_fence_get(), to be
166 * released when the fence is signalled (through e.g. the interrupt
167 * handler).
168 *
169 * This callback is optional. If this callback is not present, then the
170 * driver must always have signaling enabled.
171 */
172 bool (*enable_signaling)(struct dma_fence *fence);
173
174 /**
175 * @signaled:
176 *
177 * Peek whether the fence is signaled, as a fastpath optimization for
178 * e.g. dma_fence_wait() or dma_fence_add_callback(). Note that this
179 * callback does not need to make any guarantees beyond that a fence
180 * once indicates as signalled must always return true from this
181 * callback. This callback may return false even if the fence has
182 * completed already, in this case information hasn't propogated throug
183 * the system yet. See also dma_fence_is_signaled().
184 *
185 * May set &dma_fence.error if returning true.
186 *
187 * This callback is optional.
188 */
189 bool (*signaled)(struct dma_fence *fence);
190
191 /**
192 * @wait:
193 *
194 * Custom wait implementation, defaults to dma_fence_default_wait() if
195 * not set.
196 *
197 * The dma_fence_default_wait implementation should work for any fence, as long
198 * as @enable_signaling works correctly. This hook allows drivers to
199 * have an optimized version for the case where a process context is
200 * already available, e.g. if @enable_signaling for the general case
201 * needs to set up a worker thread.
202 *
203 * Must return -ERESTARTSYS if the wait is intr = true and the wait was
204 * interrupted, and remaining jiffies if fence has signaled, or 0 if wait
205 * timed out. Can also return other error values on custom implementations,
206 * which should be treated as if the fence is signaled. For example a hardware
207 * lockup could be reported like that.
208 *
209 * This callback is optional.
210 */
211 signed long (*wait)(struct dma_fence *fence,
212 bool intr, signed long timeout);
213
214 /**
215 * @release:
216 *
217 * Called on destruction of fence to release additional resources.
218 * Can be called from irq context. This callback is optional. If it is
219 * NULL, then dma_fence_free() is instead called as the default
220 * implementation.
221 */
222 void (*release)(struct dma_fence *fence);
223
224 /**
225 * @fence_value_str:
226 *
227 * Callback to fill in free-form debug info specific to this fence, like
228 * the sequence number.
229 *
230 * This callback is optional.
231 */
232 void (*fence_value_str)(struct dma_fence *fence, char *str, int size);
233
234 /**
235 * @timeline_value_str:
236 *
237 * Fills in the current value of the timeline as a string, like the
238 * sequence number. Note that the specific fence passed to this function
239 * should not matter, drivers should only use it to look up the
240 * corresponding timeline structures.
241 */
242 void (*timeline_value_str)(struct dma_fence *fence,
243 char *str, int size);
244};
245
246void dma_fence_init(struct dma_fence *fence, const struct dma_fence_ops *ops,
247 spinlock_t *lock, u64 context, u64 seqno);
248
249void dma_fence_release(struct kref *kref);
250void dma_fence_free(struct dma_fence *fence);
251
252/**
253 * dma_fence_put - decreases refcount of the fence
254 * @fence: fence to reduce refcount of
255 */
256static inline void dma_fence_put(struct dma_fence *fence)
257{
258 if (fence)
259 kref_put(&fence->refcount, dma_fence_release);
260}
261
262/**
263 * dma_fence_get - increases refcount of the fence
264 * @fence: fence to increase refcount of
265 *
266 * Returns the same fence, with refcount increased by 1.
267 */
268static inline struct dma_fence *dma_fence_get(struct dma_fence *fence)
269{
270 if (fence)
271 kref_get(&fence->refcount);
272 return fence;
273}
274
275/**
276 * dma_fence_get_rcu - get a fence from a reservation_object_list with
277 * rcu read lock
278 * @fence: fence to increase refcount of
279 *
280 * Function returns NULL if no refcount could be obtained, or the fence.
281 */
282static inline struct dma_fence *dma_fence_get_rcu(struct dma_fence *fence)
283{
284 if (kref_get_unless_zero(&fence->refcount))
285 return fence;
286 else
287 return NULL;
288}
289
290/**
291 * dma_fence_get_rcu_safe - acquire a reference to an RCU tracked fence
292 * @fencep: pointer to fence to increase refcount of
293 *
294 * Function returns NULL if no refcount could be obtained, or the fence.
295 * This function handles acquiring a reference to a fence that may be
296 * reallocated within the RCU grace period (such as with SLAB_TYPESAFE_BY_RCU),
297 * so long as the caller is using RCU on the pointer to the fence.
298 *
299 * An alternative mechanism is to employ a seqlock to protect a bunch of
300 * fences, such as used by struct reservation_object. When using a seqlock,
301 * the seqlock must be taken before and checked after a reference to the
302 * fence is acquired (as shown here).
303 *
304 * The caller is required to hold the RCU read lock.
305 */
306static inline struct dma_fence *
307dma_fence_get_rcu_safe(struct dma_fence __rcu **fencep)
308{
309 do {
310 struct dma_fence *fence;
311
312 fence = rcu_dereference(*fencep);
313 if (!fence)
314 return NULL;
315
316 if (!dma_fence_get_rcu(fence))
317 continue;
318
319 /* The atomic_inc_not_zero() inside dma_fence_get_rcu()
320 * provides a full memory barrier upon success (such as now).
321 * This is paired with the write barrier from assigning
322 * to the __rcu protected fence pointer so that if that
323 * pointer still matches the current fence, we know we
324 * have successfully acquire a reference to it. If it no
325 * longer matches, we are holding a reference to some other
326 * reallocated pointer. This is possible if the allocator
327 * is using a freelist like SLAB_TYPESAFE_BY_RCU where the
328 * fence remains valid for the RCU grace period, but it
329 * may be reallocated. When using such allocators, we are
330 * responsible for ensuring the reference we get is to
331 * the right fence, as below.
332 */
333 if (fence == rcu_access_pointer(*fencep))
334 return rcu_pointer_handoff(fence);
335
336 dma_fence_put(fence);
337 } while (1);
338}
339
340int dma_fence_signal(struct dma_fence *fence);
341int dma_fence_signal_locked(struct dma_fence *fence);
342signed long dma_fence_default_wait(struct dma_fence *fence,
343 bool intr, signed long timeout);
344int dma_fence_add_callback(struct dma_fence *fence,
345 struct dma_fence_cb *cb,
346 dma_fence_func_t func);
347bool dma_fence_remove_callback(struct dma_fence *fence,
348 struct dma_fence_cb *cb);
349void dma_fence_enable_sw_signaling(struct dma_fence *fence);
350
351/**
352 * dma_fence_is_signaled_locked - Return an indication if the fence
353 * is signaled yet.
354 * @fence: the fence to check
355 *
356 * Returns true if the fence was already signaled, false if not. Since this
357 * function doesn't enable signaling, it is not guaranteed to ever return
358 * true if dma_fence_add_callback(), dma_fence_wait() or
359 * dma_fence_enable_sw_signaling() haven't been called before.
360 *
361 * This function requires &dma_fence.lock to be held.
362 *
363 * See also dma_fence_is_signaled().
364 */
365static inline bool
366dma_fence_is_signaled_locked(struct dma_fence *fence)
367{
368 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
369 return true;
370
371 if (fence->ops->signaled && fence->ops->signaled(fence)) {
372 dma_fence_signal_locked(fence);
373 return true;
374 }
375
376 return false;
377}
378
379/**
380 * dma_fence_is_signaled - Return an indication if the fence is signaled yet.
381 * @fence: the fence to check
382 *
383 * Returns true if the fence was already signaled, false if not. Since this
384 * function doesn't enable signaling, it is not guaranteed to ever return
385 * true if dma_fence_add_callback(), dma_fence_wait() or
386 * dma_fence_enable_sw_signaling() haven't been called before.
387 *
388 * It's recommended for seqno fences to call dma_fence_signal when the
389 * operation is complete, it makes it possible to prevent issues from
390 * wraparound between time of issue and time of use by checking the return
391 * value of this function before calling hardware-specific wait instructions.
392 *
393 * See also dma_fence_is_signaled_locked().
394 */
395static inline bool
396dma_fence_is_signaled(struct dma_fence *fence)
397{
398 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
399 return true;
400
401 if (fence->ops->signaled && fence->ops->signaled(fence)) {
402 dma_fence_signal(fence);
403 return true;
404 }
405
406 return false;
407}
408
409/**
410 * __dma_fence_is_later - return if f1 is chronologically later than f2
411 * @f1: the first fence's seqno
412 * @f2: the second fence's seqno from the same context
413 *
414 * Returns true if f1 is chronologically later than f2. Both fences must be
415 * from the same context, since a seqno is not common across contexts.
416 */
417static inline bool __dma_fence_is_later(u64 f1, u64 f2)
418{
419 /* This is for backward compatibility with drivers which can only handle
420 * 32bit sequence numbers. Use a 64bit compare when any of the higher
421 * bits are none zero, otherwise use a 32bit compare with wrap around
422 * handling.
423 */
424 if (upper_32_bits(f1) || upper_32_bits(f2))
425 return f1 > f2;
426
427 return (int)(lower_32_bits(f1) - lower_32_bits(f2)) > 0;
428}
429
430/**
431 * dma_fence_is_later - return if f1 is chronologically later than f2
432 * @f1: the first fence from the same context
433 * @f2: the second fence from the same context
434 *
435 * Returns true if f1 is chronologically later than f2. Both fences must be
436 * from the same context, since a seqno is not re-used across contexts.
437 */
438static inline bool dma_fence_is_later(struct dma_fence *f1,
439 struct dma_fence *f2)
440{
441 if (WARN_ON(f1->context != f2->context))
442 return false;
443
444 return __dma_fence_is_later(f1->seqno, f2->seqno);
445}
446
447/**
448 * dma_fence_later - return the chronologically later fence
449 * @f1: the first fence from the same context
450 * @f2: the second fence from the same context
451 *
452 * Returns NULL if both fences are signaled, otherwise the fence that would be
453 * signaled last. Both fences must be from the same context, since a seqno is
454 * not re-used across contexts.
455 */
456static inline struct dma_fence *dma_fence_later(struct dma_fence *f1,
457 struct dma_fence *f2)
458{
459 if (WARN_ON(f1->context != f2->context))
460 return NULL;
461
462 /*
463 * Can't check just DMA_FENCE_FLAG_SIGNALED_BIT here, it may never
464 * have been set if enable_signaling wasn't called, and enabling that
465 * here is overkill.
466 */
467 if (dma_fence_is_later(f1, f2))
468 return dma_fence_is_signaled(f1) ? NULL : f1;
469 else
470 return dma_fence_is_signaled(f2) ? NULL : f2;
471}
472
473/**
474 * dma_fence_get_status_locked - returns the status upon completion
475 * @fence: the dma_fence to query
476 *
477 * Drivers can supply an optional error status condition before they signal
478 * the fence (to indicate whether the fence was completed due to an error
479 * rather than success). The value of the status condition is only valid
480 * if the fence has been signaled, dma_fence_get_status_locked() first checks
481 * the signal state before reporting the error status.
482 *
483 * Returns 0 if the fence has not yet been signaled, 1 if the fence has
484 * been signaled without an error condition, or a negative error code
485 * if the fence has been completed in err.
486 */
487static inline int dma_fence_get_status_locked(struct dma_fence *fence)
488{
489 if (dma_fence_is_signaled_locked(fence))
490 return fence->error ?: 1;
491 else
492 return 0;
493}
494
495int dma_fence_get_status(struct dma_fence *fence);
496
497/**
498 * dma_fence_set_error - flag an error condition on the fence
499 * @fence: the dma_fence
500 * @error: the error to store
501 *
502 * Drivers can supply an optional error status condition before they signal
503 * the fence, to indicate that the fence was completed due to an error
504 * rather than success. This must be set before signaling (so that the value
505 * is visible before any waiters on the signal callback are woken). This
506 * helper exists to help catching erroneous setting of #dma_fence.error.
507 */
508static inline void dma_fence_set_error(struct dma_fence *fence,
509 int error)
510{
511 WARN_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
512 WARN_ON(error >= 0 || error < -MAX_ERRNO);
513
514 fence->error = error;
515}
516
517signed long dma_fence_wait_timeout(struct dma_fence *,
518 bool intr, signed long timeout);
519signed long dma_fence_wait_any_timeout(struct dma_fence **fences,
520 uint32_t count,
521 bool intr, signed long timeout,
522 uint32_t *idx);
523
524/**
525 * dma_fence_wait - sleep until the fence gets signaled
526 * @fence: the fence to wait on
527 * @intr: if true, do an interruptible wait
528 *
529 * This function will return -ERESTARTSYS if interrupted by a signal,
530 * or 0 if the fence was signaled. Other error values may be
531 * returned on custom implementations.
532 *
533 * Performs a synchronous wait on this fence. It is assumed the caller
534 * directly or indirectly holds a reference to the fence, otherwise the
535 * fence might be freed before return, resulting in undefined behavior.
536 *
537 * See also dma_fence_wait_timeout() and dma_fence_wait_any_timeout().
538 */
539static inline signed long dma_fence_wait(struct dma_fence *fence, bool intr)
540{
541 signed long ret;
542
543 /* Since dma_fence_wait_timeout cannot timeout with
544 * MAX_SCHEDULE_TIMEOUT, only valid return values are
545 * -ERESTARTSYS and MAX_SCHEDULE_TIMEOUT.
546 */
547 ret = dma_fence_wait_timeout(fence, intr, MAX_SCHEDULE_TIMEOUT);
548
549 return ret < 0 ? ret : 0;
550}
551
552struct dma_fence *dma_fence_get_stub(void);
553u64 dma_fence_context_alloc(unsigned num);
554
555#define DMA_FENCE_TRACE(f, fmt, args...) \
556 do { \
557 struct dma_fence *__ff = (f); \
558 if (IS_ENABLED(CONFIG_DMA_FENCE_TRACE)) \
559 pr_info("f %llu#%llu: " fmt, \
560 __ff->context, __ff->seqno, ##args); \
561 } while (0)
562
563#define DMA_FENCE_WARN(f, fmt, args...) \
564 do { \
565 struct dma_fence *__ff = (f); \
566 pr_warn("f %llu#%llu: " fmt, __ff->context, __ff->seqno,\
567 ##args); \
568 } while (0)
569
570#define DMA_FENCE_ERR(f, fmt, args...) \
571 do { \
572 struct dma_fence *__ff = (f); \
573 pr_err("f %llu#%llu: " fmt, __ff->context, __ff->seqno, \
574 ##args); \
575 } while (0)
576
577#endif /* __LINUX_DMA_FENCE_H */
578