1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef __LINUX_SEQLOCK_H
3 #define __LINUX_SEQLOCK_H
6 * seqcount_t / seqlock_t - a reader-writer consistency mechanism with
7 * lockless readers (read-only retry loops), and no writer starvation.
9 * See Documentation/locking/seqlock.rst
12 * - Based on x86_64 vsyscall gettimeofday: Keith Owens, Andrea Arcangeli
13 * - Sequence counters with associated locks, (C) 2020 Linutronix GmbH
16 #include <linux/compiler.h>
17 #include <linux/kcsan-checks.h>
18 #include <linux/lockdep.h>
19 #include <linux/mutex.h>
20 #include <linux/ww_mutex.h>
21 #include <linux/preempt.h>
22 #include <linux/spinlock.h>
24 #include <asm/processor.h>
27 * The seqlock seqcount_t interface does not prescribe a precise sequence of
28 * read begin/retry/end. For readers, typically there is a call to
29 * read_seqcount_begin() and read_seqcount_retry(), however, there are more
30 * esoteric cases which do not follow this pattern.
32 * As a consequence, we take the following best-effort approach for raw usage
33 * via seqcount_t under KCSAN: upon beginning a seq-reader critical section,
34 * pessimistically mark the next KCSAN_SEQLOCK_REGION_MAX memory accesses as
35 * atomics; if there is a matching read_seqcount_retry() call, no following
36 * memory operations are considered atomic. Usage of the seqlock_t interface
39 #define KCSAN_SEQLOCK_REGION_MAX 1000
42 * Sequence counters (seqcount_t)
44 * This is the raw counting mechanism, without any writer protection.
46 * Write side critical sections must be serialized and non-preemptible.
48 * If readers can be invoked from hardirq or softirq contexts,
49 * interrupts or bottom halves must also be respectively disabled before
50 * entering the write section.
52 * This mechanism can't be used if the protected data contains pointers,
53 * as the writer can invalidate a pointer that a reader is following.
55 * If the write serialization mechanism is one of the common kernel
56 * locking primitives, use a sequence counter with associated lock
57 * (seqcount_LOCKNAME_t) instead.
59 * If it's desired to automatically handle the sequence counter writer
60 * serialization and non-preemptibility requirements, use a sequential
61 * lock (seqlock_t) instead.
63 * See Documentation/locking/seqlock.rst
65 typedef struct seqcount {
67 #ifdef CONFIG_DEBUG_LOCK_ALLOC
68 struct lockdep_map dep_map;
72 static inline void __seqcount_init(seqcount_t *s, const char *name,
73 struct lock_class_key *key)
76 * Make sure we are not reinitializing a held lock:
78 lockdep_init_map(&s->dep_map, name, key, 0);
82 #ifdef CONFIG_DEBUG_LOCK_ALLOC
84 # define SEQCOUNT_DEP_MAP_INIT(lockname) \
85 .dep_map = { .name = #lockname }
88 * seqcount_init() - runtime initializer for seqcount_t
89 * @s: Pointer to the seqcount_t instance
91 # define seqcount_init(s) \
93 static struct lock_class_key __key; \
94 __seqcount_init((s), #s, &__key); \
97 static inline void seqcount_lockdep_reader_access(const seqcount_t *s)
99 seqcount_t *l = (seqcount_t *)s;
102 local_irq_save(flags);
103 seqcount_acquire_read(&l->dep_map, 0, 0, _RET_IP_);
104 seqcount_release(&l->dep_map, _RET_IP_);
105 local_irq_restore(flags);
109 # define SEQCOUNT_DEP_MAP_INIT(lockname)
110 # define seqcount_init(s) __seqcount_init(s, NULL, NULL)
111 # define seqcount_lockdep_reader_access(x)
115 * SEQCNT_ZERO() - static initializer for seqcount_t
116 * @name: Name of the seqcount_t instance
118 #define SEQCNT_ZERO(name) { .sequence = 0, SEQCOUNT_DEP_MAP_INIT(name) }
121 * Sequence counters with associated locks (seqcount_LOCKNAME_t)
123 * A sequence counter which associates the lock used for writer
124 * serialization at initialization time. This enables lockdep to validate
125 * that the write side critical section is properly serialized.
127 * For associated locks which do not implicitly disable preemption,
128 * preemption protection is enforced in the write side function.
130 * Lockdep is never used in any for the raw write variants.
132 * See Documentation/locking/seqlock.rst
136 * For PREEMPT_RT, seqcount_LOCKNAME_t write side critical sections cannot
137 * disable preemption. It can lead to higher latencies, and the write side
138 * sections will not be able to acquire locks which become sleeping locks
141 * To remain preemptible while avoiding a possible livelock caused by the
142 * reader preempting the writer, use a different technique: let the reader
143 * detect if a seqcount_LOCKNAME_t writer is in progress. If that is the
144 * case, acquire then release the associated LOCKNAME writer serialization
145 * lock. This will allow any possibly-preempted writer to make progress
146 * until the end of its writer serialization lock critical section.
148 * This lock-unlock technique must be implemented for all of PREEMPT_RT
149 * sleeping locks. See Documentation/locking/locktypes.rst
151 #if defined(CONFIG_LOCKDEP) || defined(CONFIG_PREEMPT_RT)
152 #define __SEQ_LOCK(expr) expr
154 #define __SEQ_LOCK(expr)
158 * typedef seqcount_LOCKNAME_t - sequence counter with LOCKNAME associated
159 * @seqcount: The real sequence counter
160 * @lock: Pointer to the associated lock
162 * A plain sequence counter with external writer synchronization by
163 * LOCKNAME @lock. The lock is associated to the sequence counter in the
164 * static initializer or init function. This enables lockdep to validate
165 * that the write side critical section is properly serialized.
167 * LOCKNAME: raw_spinlock, spinlock, rwlock, mutex, or ww_mutex.
171 * seqcount_LOCKNAME_init() - runtime initializer for seqcount_LOCKNAME_t
172 * @s: Pointer to the seqcount_LOCKNAME_t instance
173 * @lock: Pointer to the associated lock
176 #define seqcount_LOCKNAME_init(s, _lock, lockname) \
178 seqcount_##lockname##_t *____s = (s); \
179 seqcount_init(&____s->seqcount); \
180 __SEQ_LOCK(____s->lock = (_lock)); \
183 #define seqcount_raw_spinlock_init(s, lock) seqcount_LOCKNAME_init(s, lock, raw_spinlock)
184 #define seqcount_spinlock_init(s, lock) seqcount_LOCKNAME_init(s, lock, spinlock)
185 #define seqcount_rwlock_init(s, lock) seqcount_LOCKNAME_init(s, lock, rwlock);
186 #define seqcount_mutex_init(s, lock) seqcount_LOCKNAME_init(s, lock, mutex);
187 #define seqcount_ww_mutex_init(s, lock) seqcount_LOCKNAME_init(s, lock, ww_mutex);
190 * SEQCOUNT_LOCKNAME() - Instantiate seqcount_LOCKNAME_t and helpers
191 * seqprop_LOCKNAME_*() - Property accessors for seqcount_LOCKNAME_t
193 * @lockname: "LOCKNAME" part of seqcount_LOCKNAME_t
194 * @locktype: LOCKNAME canonical C data type
195 * @preemptible: preemptibility of above locktype
196 * @lockmember: argument for lockdep_assert_held()
197 * @lockbase: associated lock release function (prefix only)
198 * @lock_acquire: associated lock acquisition function (full call)
200 #define SEQCOUNT_LOCKNAME(lockname, locktype, preemptible, lockmember, lockbase, lock_acquire) \
201 typedef struct seqcount_##lockname { \
202 seqcount_t seqcount; \
203 __SEQ_LOCK(locktype *lock); \
204 } seqcount_##lockname##_t; \
206 static __always_inline seqcount_t * \
207 __seqprop_##lockname##_ptr(seqcount_##lockname##_t *s) \
209 return &s->seqcount; \
212 static __always_inline unsigned \
213 __seqprop_##lockname##_sequence(const seqcount_##lockname##_t *s) \
215 unsigned seq = READ_ONCE(s->seqcount.sequence); \
217 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) \
220 if (preemptible && unlikely(seq & 1)) { \
221 __SEQ_LOCK(lock_acquire); \
222 __SEQ_LOCK(lockbase##_unlock(s->lock)); \
225 * Re-read the sequence counter since the (possibly \
226 * preempted) writer made progress. \
228 seq = READ_ONCE(s->seqcount.sequence); \
234 static __always_inline bool \
235 __seqprop_##lockname##_preemptible(const seqcount_##lockname##_t *s) \
237 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) \
238 return preemptible; \
240 /* PREEMPT_RT relies on the above LOCK+UNLOCK */ \
244 static __always_inline void \
245 __seqprop_##lockname##_assert(const seqcount_##lockname##_t *s) \
247 __SEQ_LOCK(lockdep_assert_held(lockmember)); \
251 * __seqprop() for seqcount_t
254 static inline seqcount_t *__seqprop_ptr(seqcount_t *s)
259 static inline unsigned __seqprop_sequence(const seqcount_t *s)
261 return READ_ONCE(s->sequence);
264 static inline bool __seqprop_preemptible(const seqcount_t *s)
269 static inline void __seqprop_assert(const seqcount_t *s)
271 lockdep_assert_preemption_disabled();
274 #define __SEQ_RT IS_ENABLED(CONFIG_PREEMPT_RT)
276 SEQCOUNT_LOCKNAME(raw_spinlock, raw_spinlock_t, false, s->lock, raw_spin, raw_spin_lock(s->lock))
277 SEQCOUNT_LOCKNAME(spinlock, spinlock_t, __SEQ_RT, s->lock, spin, spin_lock(s->lock))
278 SEQCOUNT_LOCKNAME(rwlock, rwlock_t, __SEQ_RT, s->lock, read, read_lock(s->lock))
279 SEQCOUNT_LOCKNAME(mutex, struct mutex, true, s->lock, mutex, mutex_lock(s->lock))
280 SEQCOUNT_LOCKNAME(ww_mutex, struct ww_mutex, true, &s->lock->base, ww_mutex, ww_mutex_lock(s->lock, NULL))
283 * SEQCNT_LOCKNAME_ZERO - static initializer for seqcount_LOCKNAME_t
284 * @name: Name of the seqcount_LOCKNAME_t instance
285 * @lock: Pointer to the associated LOCKNAME
288 #define SEQCOUNT_LOCKNAME_ZERO(seq_name, assoc_lock) { \
289 .seqcount = SEQCNT_ZERO(seq_name.seqcount), \
290 __SEQ_LOCK(.lock = (assoc_lock)) \
293 #define SEQCNT_RAW_SPINLOCK_ZERO(name, lock) SEQCOUNT_LOCKNAME_ZERO(name, lock)
294 #define SEQCNT_SPINLOCK_ZERO(name, lock) SEQCOUNT_LOCKNAME_ZERO(name, lock)
295 #define SEQCNT_RWLOCK_ZERO(name, lock) SEQCOUNT_LOCKNAME_ZERO(name, lock)
296 #define SEQCNT_MUTEX_ZERO(name, lock) SEQCOUNT_LOCKNAME_ZERO(name, lock)
297 #define SEQCNT_WW_MUTEX_ZERO(name, lock) SEQCOUNT_LOCKNAME_ZERO(name, lock)
299 #define __seqprop_case(s, lockname, prop) \
300 seqcount_##lockname##_t: __seqprop_##lockname##_##prop((void *)(s))
302 #define __seqprop(s, prop) _Generic(*(s), \
303 seqcount_t: __seqprop_##prop((void *)(s)), \
304 __seqprop_case((s), raw_spinlock, prop), \
305 __seqprop_case((s), spinlock, prop), \
306 __seqprop_case((s), rwlock, prop), \
307 __seqprop_case((s), mutex, prop), \
308 __seqprop_case((s), ww_mutex, prop))
310 #define __seqcount_ptr(s) __seqprop(s, ptr)
311 #define __seqcount_sequence(s) __seqprop(s, sequence)
312 #define __seqcount_lock_preemptible(s) __seqprop(s, preemptible)
313 #define __seqcount_assert_lock_held(s) __seqprop(s, assert)
316 * __read_seqcount_begin() - begin a seqcount_t read section w/o barrier
317 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
319 * __read_seqcount_begin is like read_seqcount_begin, but has no smp_rmb()
320 * barrier. Callers should ensure that smp_rmb() or equivalent ordering is
321 * provided before actually loading any of the variables that are to be
322 * protected in this critical section.
324 * Use carefully, only in critical code, and comment how the barrier is
327 * Return: count to be passed to read_seqcount_retry()
329 #define __read_seqcount_begin(s) \
333 while ((seq = __seqcount_sequence(s)) & 1) \
336 kcsan_atomic_next(KCSAN_SEQLOCK_REGION_MAX); \
341 * raw_read_seqcount_begin() - begin a seqcount_t read section w/o lockdep
342 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
344 * Return: count to be passed to read_seqcount_retry()
346 #define raw_read_seqcount_begin(s) \
348 unsigned seq = __read_seqcount_begin(s); \
355 * read_seqcount_begin() - begin a seqcount_t read critical section
356 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
358 * Return: count to be passed to read_seqcount_retry()
360 #define read_seqcount_begin(s) \
362 seqcount_lockdep_reader_access(__seqcount_ptr(s)); \
363 raw_read_seqcount_begin(s); \
367 * raw_read_seqcount() - read the raw seqcount_t counter value
368 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
370 * raw_read_seqcount opens a read critical section of the given
371 * seqcount_t, without any lockdep checking, and without checking or
372 * masking the sequence counter LSB. Calling code is responsible for
375 * Return: count to be passed to read_seqcount_retry()
377 #define raw_read_seqcount(s) \
379 unsigned seq = __seqcount_sequence(s); \
382 kcsan_atomic_next(KCSAN_SEQLOCK_REGION_MAX); \
387 * raw_seqcount_begin() - begin a seqcount_t read critical section w/o
388 * lockdep and w/o counter stabilization
389 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
391 * raw_seqcount_begin opens a read critical section of the given
392 * seqcount_t. Unlike read_seqcount_begin(), this function will not wait
393 * for the count to stabilize. If a writer is active when it begins, it
394 * will fail the read_seqcount_retry() at the end of the read critical
395 * section instead of stabilizing at the beginning of it.
397 * Use this only in special kernel hot paths where the read section is
398 * small and has a high probability of success through other external
399 * means. It will save a single branching instruction.
401 * Return: count to be passed to read_seqcount_retry()
403 #define raw_seqcount_begin(s) \
406 * If the counter is odd, let read_seqcount_retry() fail \
407 * by decrementing the counter. \
409 raw_read_seqcount(s) & ~1; \
413 * __read_seqcount_retry() - end a seqcount_t read section w/o barrier
414 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
415 * @start: count, from read_seqcount_begin()
417 * __read_seqcount_retry is like read_seqcount_retry, but has no smp_rmb()
418 * barrier. Callers should ensure that smp_rmb() or equivalent ordering is
419 * provided before actually loading any of the variables that are to be
420 * protected in this critical section.
422 * Use carefully, only in critical code, and comment how the barrier is
425 * Return: true if a read section retry is required, else false
427 #define __read_seqcount_retry(s, start) \
428 __read_seqcount_t_retry(__seqcount_ptr(s), start)
430 static inline int __read_seqcount_t_retry(const seqcount_t *s, unsigned start)
432 kcsan_atomic_next(0);
433 return unlikely(READ_ONCE(s->sequence) != start);
437 * read_seqcount_retry() - end a seqcount_t read critical section
438 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
439 * @start: count, from read_seqcount_begin()
441 * read_seqcount_retry closes the read critical section of given
442 * seqcount_t. If the critical section was invalid, it must be ignored
443 * (and typically retried).
445 * Return: true if a read section retry is required, else false
447 #define read_seqcount_retry(s, start) \
448 read_seqcount_t_retry(__seqcount_ptr(s), start)
450 static inline int read_seqcount_t_retry(const seqcount_t *s, unsigned start)
453 return __read_seqcount_t_retry(s, start);
457 * raw_write_seqcount_begin() - start a seqcount_t write section w/o lockdep
458 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
460 #define raw_write_seqcount_begin(s) \
462 if (__seqcount_lock_preemptible(s)) \
465 raw_write_seqcount_t_begin(__seqcount_ptr(s)); \
468 static inline void raw_write_seqcount_t_begin(seqcount_t *s)
470 kcsan_nestable_atomic_begin();
476 * raw_write_seqcount_end() - end a seqcount_t write section w/o lockdep
477 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
479 #define raw_write_seqcount_end(s) \
481 raw_write_seqcount_t_end(__seqcount_ptr(s)); \
483 if (__seqcount_lock_preemptible(s)) \
487 static inline void raw_write_seqcount_t_end(seqcount_t *s)
491 kcsan_nestable_atomic_end();
495 * write_seqcount_begin_nested() - start a seqcount_t write section with
496 * custom lockdep nesting level
497 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
498 * @subclass: lockdep nesting level
500 * See Documentation/locking/lockdep-design.rst
502 #define write_seqcount_begin_nested(s, subclass) \
504 __seqcount_assert_lock_held(s); \
506 if (__seqcount_lock_preemptible(s)) \
509 write_seqcount_t_begin_nested(__seqcount_ptr(s), subclass); \
512 static inline void write_seqcount_t_begin_nested(seqcount_t *s, int subclass)
514 raw_write_seqcount_t_begin(s);
515 seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_);
519 * write_seqcount_begin() - start a seqcount_t write side critical section
520 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
522 * write_seqcount_begin opens a write side critical section of the given
525 * Context: seqcount_t write side critical sections must be serialized and
526 * non-preemptible. If readers can be invoked from hardirq or softirq
527 * context, interrupts or bottom halves must be respectively disabled.
529 #define write_seqcount_begin(s) \
531 __seqcount_assert_lock_held(s); \
533 if (__seqcount_lock_preemptible(s)) \
536 write_seqcount_t_begin(__seqcount_ptr(s)); \
539 static inline void write_seqcount_t_begin(seqcount_t *s)
541 write_seqcount_t_begin_nested(s, 0);
545 * write_seqcount_end() - end a seqcount_t write side critical section
546 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
548 * The write section must've been opened with write_seqcount_begin().
550 #define write_seqcount_end(s) \
552 write_seqcount_t_end(__seqcount_ptr(s)); \
554 if (__seqcount_lock_preemptible(s)) \
558 static inline void write_seqcount_t_end(seqcount_t *s)
560 seqcount_release(&s->dep_map, _RET_IP_);
561 raw_write_seqcount_t_end(s);
565 * raw_write_seqcount_barrier() - do a seqcount_t write barrier
566 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
568 * This can be used to provide an ordering guarantee instead of the usual
569 * consistency guarantee. It is one wmb cheaper, because it can collapse
570 * the two back-to-back wmb()s.
572 * Note that writes surrounding the barrier should be declared atomic (e.g.
573 * via WRITE_ONCE): a) to ensure the writes become visible to other threads
574 * atomically, avoiding compiler optimizations; b) to document which writes are
575 * meant to propagate to the reader critical section. This is necessary because
576 * neither writes before and after the barrier are enclosed in a seq-writer
577 * critical section that would ensure readers are aware of ongoing writes::
580 * bool X = true, Y = false;
587 * int s = read_seqcount_begin(&seq);
591 * } while (read_seqcount_retry(&seq, s));
598 * WRITE_ONCE(Y, true);
600 * raw_write_seqcount_barrier(seq);
602 * WRITE_ONCE(X, false);
605 #define raw_write_seqcount_barrier(s) \
606 raw_write_seqcount_t_barrier(__seqcount_ptr(s))
608 static inline void raw_write_seqcount_t_barrier(seqcount_t *s)
610 kcsan_nestable_atomic_begin();
614 kcsan_nestable_atomic_end();
618 * write_seqcount_invalidate() - invalidate in-progress seqcount_t read
620 * @s: Pointer to seqcount_t or any of the seqcount_LOCKNAME_t variants
622 * After write_seqcount_invalidate, no seqcount_t read side operations
623 * will complete successfully and see data older than this.
625 #define write_seqcount_invalidate(s) \
626 write_seqcount_t_invalidate(__seqcount_ptr(s))
628 static inline void write_seqcount_t_invalidate(seqcount_t *s)
631 kcsan_nestable_atomic_begin();
633 kcsan_nestable_atomic_end();
637 * Latch sequence counters (seqcount_latch_t)
639 * A sequence counter variant where the counter even/odd value is used to
640 * switch between two copies of protected data. This allows the read path,
641 * typically NMIs, to safely interrupt the write side critical section.
643 * As the write sections are fully preemptible, no special handling for
644 * PREEMPT_RT is needed.
651 * SEQCNT_LATCH_ZERO() - static initializer for seqcount_latch_t
652 * @seq_name: Name of the seqcount_latch_t instance
654 #define SEQCNT_LATCH_ZERO(seq_name) { \
655 .seqcount = SEQCNT_ZERO(seq_name.seqcount), \
659 * seqcount_latch_init() - runtime initializer for seqcount_latch_t
660 * @s: Pointer to the seqcount_latch_t instance
662 static inline void seqcount_latch_init(seqcount_latch_t *s)
664 seqcount_init(&s->seqcount);
668 * raw_read_seqcount_latch() - pick even/odd latch data copy
669 * @s: Pointer to seqcount_latch_t
671 * See raw_write_seqcount_latch() for details and a full reader/writer
674 * Return: sequence counter raw value. Use the lowest bit as an index for
675 * picking which data copy to read. The full counter must then be checked
676 * with read_seqcount_latch_retry().
678 static inline unsigned raw_read_seqcount_latch(const seqcount_latch_t *s)
681 * Pairs with the first smp_wmb() in raw_write_seqcount_latch().
682 * Due to the dependent load, a full smp_rmb() is not needed.
684 return READ_ONCE(s->seqcount.sequence);
688 * read_seqcount_latch_retry() - end a seqcount_latch_t read section
689 * @s: Pointer to seqcount_latch_t
690 * @start: count, from raw_read_seqcount_latch()
692 * Return: true if a read section retry is required, else false
695 read_seqcount_latch_retry(const seqcount_latch_t *s, unsigned start)
697 return read_seqcount_retry(&s->seqcount, start);
701 * raw_write_seqcount_latch() - redirect latch readers to even/odd copy
702 * @s: Pointer to seqcount_latch_t
704 * The latch technique is a multiversion concurrency control method that allows
705 * queries during non-atomic modifications. If you can guarantee queries never
706 * interrupt the modification -- e.g. the concurrency is strictly between CPUs
707 * -- you most likely do not need this.
709 * Where the traditional RCU/lockless data structures rely on atomic
710 * modifications to ensure queries observe either the old or the new state the
711 * latch allows the same for non-atomic updates. The trade-off is doubling the
712 * cost of storage; we have to maintain two copies of the entire data
715 * Very simply put: we first modify one copy and then the other. This ensures
716 * there is always one copy in a stable state, ready to give us an answer.
718 * The basic form is a data structure like::
720 * struct latch_struct {
721 * seqcount_latch_t seq;
722 * struct data_struct data[2];
725 * Where a modification, which is assumed to be externally serialized, does the
728 * void latch_modify(struct latch_struct *latch, ...)
730 * smp_wmb(); // Ensure that the last data[1] update is visible
731 * latch->seq.sequence++;
732 * smp_wmb(); // Ensure that the seqcount update is visible
734 * modify(latch->data[0], ...);
736 * smp_wmb(); // Ensure that the data[0] update is visible
737 * latch->seq.sequence++;
738 * smp_wmb(); // Ensure that the seqcount update is visible
740 * modify(latch->data[1], ...);
743 * The query will have a form like::
745 * struct entry *latch_query(struct latch_struct *latch, ...)
747 * struct entry *entry;
751 * seq = raw_read_seqcount_latch(&latch->seq);
754 * entry = data_query(latch->data[idx], ...);
756 * // This includes needed smp_rmb()
757 * } while (read_seqcount_latch_retry(&latch->seq, seq));
762 * So during the modification, queries are first redirected to data[1]. Then we
763 * modify data[0]. When that is complete, we redirect queries back to data[0]
764 * and we can modify data[1].
768 * The non-requirement for atomic modifications does _NOT_ include
769 * the publishing of new entries in the case where data is a dynamic
772 * An iteration might start in data[0] and get suspended long enough
773 * to miss an entire modification sequence, once it resumes it might
774 * observe the new entry.
778 * When data is a dynamic data structure; one should use regular RCU
779 * patterns to manage the lifetimes of the objects within.
781 static inline void raw_write_seqcount_latch(seqcount_latch_t *s)
783 smp_wmb(); /* prior stores before incrementing "sequence" */
784 s->seqcount.sequence++;
785 smp_wmb(); /* increment "sequence" before following stores */
789 * Sequential locks (seqlock_t)
791 * Sequence counters with an embedded spinlock for writer serialization
792 * and non-preemptibility.
794 * For more info, see:
795 * - Comments on top of seqcount_t
796 * - Documentation/locking/seqlock.rst
800 * Make sure that readers don't starve writers on PREEMPT_RT: use
801 * seqcount_spinlock_t instead of seqcount_t. Check __SEQ_LOCK().
803 seqcount_spinlock_t seqcount;
807 #define __SEQLOCK_UNLOCKED(lockname) \
809 .seqcount = SEQCNT_SPINLOCK_ZERO(lockname, &(lockname).lock), \
810 .lock = __SPIN_LOCK_UNLOCKED(lockname) \
814 * seqlock_init() - dynamic initializer for seqlock_t
815 * @sl: Pointer to the seqlock_t instance
817 #define seqlock_init(sl) \
819 spin_lock_init(&(sl)->lock); \
820 seqcount_spinlock_init(&(sl)->seqcount, &(sl)->lock); \
824 * DEFINE_SEQLOCK(sl) - Define a statically allocated seqlock_t
825 * @sl: Name of the seqlock_t instance
827 #define DEFINE_SEQLOCK(sl) \
828 seqlock_t sl = __SEQLOCK_UNLOCKED(sl)
831 * read_seqbegin() - start a seqlock_t read side critical section
832 * @sl: Pointer to seqlock_t
834 * Return: count, to be passed to read_seqretry()
836 static inline unsigned read_seqbegin(const seqlock_t *sl)
838 unsigned ret = read_seqcount_begin(&sl->seqcount);
840 kcsan_atomic_next(0); /* non-raw usage, assume closing read_seqretry() */
841 kcsan_flat_atomic_begin();
846 * read_seqretry() - end a seqlock_t read side section
847 * @sl: Pointer to seqlock_t
848 * @start: count, from read_seqbegin()
850 * read_seqretry closes the read side critical section of given seqlock_t.
851 * If the critical section was invalid, it must be ignored (and typically
854 * Return: true if a read section retry is required, else false
856 static inline unsigned read_seqretry(const seqlock_t *sl, unsigned start)
859 * Assume not nested: read_seqretry() may be called multiple times when
860 * completing read critical section.
862 kcsan_flat_atomic_end();
864 return read_seqcount_retry(&sl->seqcount, start);
868 * For all seqlock_t write side functions, use write_seqcount_*t*_begin()
869 * instead of the generic write_seqcount_begin(). This way, no redundant
870 * lockdep_assert_held() checks are added.
874 * write_seqlock() - start a seqlock_t write side critical section
875 * @sl: Pointer to seqlock_t
877 * write_seqlock opens a write side critical section for the given
878 * seqlock_t. It also implicitly acquires the spinlock_t embedded inside
879 * that sequential lock. All seqlock_t write side sections are thus
880 * automatically serialized and non-preemptible.
882 * Context: if the seqlock_t read section, or other write side critical
883 * sections, can be invoked from hardirq or softirq contexts, use the
884 * _irqsave or _bh variants of this function instead.
886 static inline void write_seqlock(seqlock_t *sl)
888 spin_lock(&sl->lock);
889 write_seqcount_t_begin(&sl->seqcount.seqcount);
893 * write_sequnlock() - end a seqlock_t write side critical section
894 * @sl: Pointer to seqlock_t
896 * write_sequnlock closes the (serialized and non-preemptible) write side
897 * critical section of given seqlock_t.
899 static inline void write_sequnlock(seqlock_t *sl)
901 write_seqcount_t_end(&sl->seqcount.seqcount);
902 spin_unlock(&sl->lock);
906 * write_seqlock_bh() - start a softirqs-disabled seqlock_t write section
907 * @sl: Pointer to seqlock_t
909 * _bh variant of write_seqlock(). Use only if the read side section, or
910 * other write side sections, can be invoked from softirq contexts.
912 static inline void write_seqlock_bh(seqlock_t *sl)
914 spin_lock_bh(&sl->lock);
915 write_seqcount_t_begin(&sl->seqcount.seqcount);
919 * write_sequnlock_bh() - end a softirqs-disabled seqlock_t write section
920 * @sl: Pointer to seqlock_t
922 * write_sequnlock_bh closes the serialized, non-preemptible, and
923 * softirqs-disabled, seqlock_t write side critical section opened with
924 * write_seqlock_bh().
926 static inline void write_sequnlock_bh(seqlock_t *sl)
928 write_seqcount_t_end(&sl->seqcount.seqcount);
929 spin_unlock_bh(&sl->lock);
933 * write_seqlock_irq() - start a non-interruptible seqlock_t write section
934 * @sl: Pointer to seqlock_t
936 * _irq variant of write_seqlock(). Use only if the read side section, or
937 * other write sections, can be invoked from hardirq contexts.
939 static inline void write_seqlock_irq(seqlock_t *sl)
941 spin_lock_irq(&sl->lock);
942 write_seqcount_t_begin(&sl->seqcount.seqcount);
946 * write_sequnlock_irq() - end a non-interruptible seqlock_t write section
947 * @sl: Pointer to seqlock_t
949 * write_sequnlock_irq closes the serialized and non-interruptible
950 * seqlock_t write side section opened with write_seqlock_irq().
952 static inline void write_sequnlock_irq(seqlock_t *sl)
954 write_seqcount_t_end(&sl->seqcount.seqcount);
955 spin_unlock_irq(&sl->lock);
958 static inline unsigned long __write_seqlock_irqsave(seqlock_t *sl)
962 spin_lock_irqsave(&sl->lock, flags);
963 write_seqcount_t_begin(&sl->seqcount.seqcount);
968 * write_seqlock_irqsave() - start a non-interruptible seqlock_t write
970 * @lock: Pointer to seqlock_t
971 * @flags: Stack-allocated storage for saving caller's local interrupt
972 * state, to be passed to write_sequnlock_irqrestore().
974 * _irqsave variant of write_seqlock(). Use it only if the read side
975 * section, or other write sections, can be invoked from hardirq context.
977 #define write_seqlock_irqsave(lock, flags) \
978 do { flags = __write_seqlock_irqsave(lock); } while (0)
981 * write_sequnlock_irqrestore() - end non-interruptible seqlock_t write
983 * @sl: Pointer to seqlock_t
984 * @flags: Caller's saved interrupt state, from write_seqlock_irqsave()
986 * write_sequnlock_irqrestore closes the serialized and non-interruptible
987 * seqlock_t write section previously opened with write_seqlock_irqsave().
990 write_sequnlock_irqrestore(seqlock_t *sl, unsigned long flags)
992 write_seqcount_t_end(&sl->seqcount.seqcount);
993 spin_unlock_irqrestore(&sl->lock, flags);
997 * read_seqlock_excl() - begin a seqlock_t locking reader section
998 * @sl: Pointer to seqlock_t
1000 * read_seqlock_excl opens a seqlock_t locking reader critical section. A
1001 * locking reader exclusively locks out *both* other writers *and* other
1002 * locking readers, but it does not update the embedded sequence number.
1004 * Locking readers act like a normal spin_lock()/spin_unlock().
1006 * Context: if the seqlock_t write section, *or other read sections*, can
1007 * be invoked from hardirq or softirq contexts, use the _irqsave or _bh
1008 * variant of this function instead.
1010 * The opened read section must be closed with read_sequnlock_excl().
1012 static inline void read_seqlock_excl(seqlock_t *sl)
1014 spin_lock(&sl->lock);
1018 * read_sequnlock_excl() - end a seqlock_t locking reader critical section
1019 * @sl: Pointer to seqlock_t
1021 static inline void read_sequnlock_excl(seqlock_t *sl)
1023 spin_unlock(&sl->lock);
1027 * read_seqlock_excl_bh() - start a seqlock_t locking reader section with
1029 * @sl: Pointer to seqlock_t
1031 * _bh variant of read_seqlock_excl(). Use this variant only if the
1032 * seqlock_t write side section, *or other read sections*, can be invoked
1033 * from softirq contexts.
1035 static inline void read_seqlock_excl_bh(seqlock_t *sl)
1037 spin_lock_bh(&sl->lock);
1041 * read_sequnlock_excl_bh() - stop a seqlock_t softirq-disabled locking
1043 * @sl: Pointer to seqlock_t
1045 static inline void read_sequnlock_excl_bh(seqlock_t *sl)
1047 spin_unlock_bh(&sl->lock);
1051 * read_seqlock_excl_irq() - start a non-interruptible seqlock_t locking
1053 * @sl: Pointer to seqlock_t
1055 * _irq variant of read_seqlock_excl(). Use this only if the seqlock_t
1056 * write side section, *or other read sections*, can be invoked from a
1059 static inline void read_seqlock_excl_irq(seqlock_t *sl)
1061 spin_lock_irq(&sl->lock);
1065 * read_sequnlock_excl_irq() - end an interrupts-disabled seqlock_t
1066 * locking reader section
1067 * @sl: Pointer to seqlock_t
1069 static inline void read_sequnlock_excl_irq(seqlock_t *sl)
1071 spin_unlock_irq(&sl->lock);
1074 static inline unsigned long __read_seqlock_excl_irqsave(seqlock_t *sl)
1076 unsigned long flags;
1078 spin_lock_irqsave(&sl->lock, flags);
1083 * read_seqlock_excl_irqsave() - start a non-interruptible seqlock_t
1084 * locking reader section
1085 * @lock: Pointer to seqlock_t
1086 * @flags: Stack-allocated storage for saving caller's local interrupt
1087 * state, to be passed to read_sequnlock_excl_irqrestore().
1089 * _irqsave variant of read_seqlock_excl(). Use this only if the seqlock_t
1090 * write side section, *or other read sections*, can be invoked from a
1093 #define read_seqlock_excl_irqsave(lock, flags) \
1094 do { flags = __read_seqlock_excl_irqsave(lock); } while (0)
1097 * read_sequnlock_excl_irqrestore() - end non-interruptible seqlock_t
1098 * locking reader section
1099 * @sl: Pointer to seqlock_t
1100 * @flags: Caller saved interrupt state, from read_seqlock_excl_irqsave()
1103 read_sequnlock_excl_irqrestore(seqlock_t *sl, unsigned long flags)
1105 spin_unlock_irqrestore(&sl->lock, flags);
1109 * read_seqbegin_or_lock() - begin a seqlock_t lockless or locking reader
1110 * @lock: Pointer to seqlock_t
1111 * @seq : Marker and return parameter. If the passed value is even, the
1112 * reader will become a *lockless* seqlock_t reader as in read_seqbegin().
1113 * If the passed value is odd, the reader will become a *locking* reader
1114 * as in read_seqlock_excl(). In the first call to this function, the
1115 * caller *must* initialize and pass an even value to @seq; this way, a
1116 * lockless read can be optimistically tried first.
1118 * read_seqbegin_or_lock is an API designed to optimistically try a normal
1119 * lockless seqlock_t read section first. If an odd counter is found, the
1120 * lockless read trial has failed, and the next read iteration transforms
1121 * itself into a full seqlock_t locking reader.
1123 * This is typically used to avoid seqlock_t lockless readers starvation
1124 * (too much retry loops) in the case of a sharp spike in write side
1127 * Context: if the seqlock_t write section, *or other read sections*, can
1128 * be invoked from hardirq or softirq contexts, use the _irqsave or _bh
1129 * variant of this function instead.
1131 * Check Documentation/locking/seqlock.rst for template example code.
1133 * Return: the encountered sequence counter value, through the @seq
1134 * parameter, which is overloaded as a return parameter. This returned
1135 * value must be checked with need_seqretry(). If the read section need to
1136 * be retried, this returned value must also be passed as the @seq
1137 * parameter of the next read_seqbegin_or_lock() iteration.
1139 static inline void read_seqbegin_or_lock(seqlock_t *lock, int *seq)
1141 if (!(*seq & 1)) /* Even */
1142 *seq = read_seqbegin(lock);
1144 read_seqlock_excl(lock);
1148 * need_seqretry() - validate seqlock_t "locking or lockless" read section
1149 * @lock: Pointer to seqlock_t
1150 * @seq: sequence count, from read_seqbegin_or_lock()
1152 * Return: true if a read section retry is required, false otherwise
1154 static inline int need_seqretry(seqlock_t *lock, int seq)
1156 return !(seq & 1) && read_seqretry(lock, seq);
1160 * done_seqretry() - end seqlock_t "locking or lockless" reader section
1161 * @lock: Pointer to seqlock_t
1162 * @seq: count, from read_seqbegin_or_lock()
1164 * done_seqretry finishes the seqlock_t read side critical section started
1165 * with read_seqbegin_or_lock() and validated by need_seqretry().
1167 static inline void done_seqretry(seqlock_t *lock, int seq)
1170 read_sequnlock_excl(lock);
1174 * read_seqbegin_or_lock_irqsave() - begin a seqlock_t lockless reader, or
1175 * a non-interruptible locking reader
1176 * @lock: Pointer to seqlock_t
1177 * @seq: Marker and return parameter. Check read_seqbegin_or_lock().
1179 * This is the _irqsave variant of read_seqbegin_or_lock(). Use it only if
1180 * the seqlock_t write section, *or other read sections*, can be invoked
1181 * from hardirq context.
1183 * Note: Interrupts will be disabled only for "locking reader" mode.
1187 * 1. The saved local interrupts state in case of a locking reader, to
1188 * be passed to done_seqretry_irqrestore().
1190 * 2. The encountered sequence counter value, returned through @seq
1191 * overloaded as a return parameter. Check read_seqbegin_or_lock().
1193 static inline unsigned long
1194 read_seqbegin_or_lock_irqsave(seqlock_t *lock, int *seq)
1196 unsigned long flags = 0;
1198 if (!(*seq & 1)) /* Even */
1199 *seq = read_seqbegin(lock);
1201 read_seqlock_excl_irqsave(lock, flags);
1207 * done_seqretry_irqrestore() - end a seqlock_t lockless reader, or a
1208 * non-interruptible locking reader section
1209 * @lock: Pointer to seqlock_t
1210 * @seq: Count, from read_seqbegin_or_lock_irqsave()
1211 * @flags: Caller's saved local interrupt state in case of a locking
1212 * reader, also from read_seqbegin_or_lock_irqsave()
1214 * This is the _irqrestore variant of done_seqretry(). The read section
1215 * must've been opened with read_seqbegin_or_lock_irqsave(), and validated
1216 * by need_seqretry().
1219 done_seqretry_irqrestore(seqlock_t *lock, int seq, unsigned long flags)
1222 read_sequnlock_excl_irqrestore(lock, flags);
1224 #endif /* __LINUX_SEQLOCK_H */