1 /* SPDX-License-Identifier: GPL-2.0 */
6  * seqcount_t / seqlock_t - a reader-writer consistency mechanism with
7  * lockless readers (read-only retry loops), and no writer starvation.
12  * - Based on x86_64 vsyscall gettimeofday: Keith Owens, Andrea Arcangeli
13  * - Sequence counters with associated locks, (C) 2020 Linutronix GmbH
17 #include <linux/kcsan-checks.h>
26  * The seqlock seqcount_t interface does not prescribe a precise sequence of
27  * read begin/retry/end. For readers, typically there is a call to
31  * As a consequence, we take the following best-effort approach for raw usage
32  * via seqcount_t under KCSAN: upon beginning a seq-reader critical section,
34  * atomics; if there is a matching read_seqcount_retry() call, no following
45  * Write side critical sections must be serialized and non-preemptible.
52  * as the writer can invalidate a pointer that a reader is following.
55  * locking primitives, use a sequence counter with associated lock
59  * serialization and non-preemptibility requirements, use a sequential
75 	 * Make sure we are not reinitializing a held lock:  in __seqcount_init()
77 	lockdep_init_map(&s->dep_map, name, key, 0);  in __seqcount_init()
78 	s->sequence = 0;  in __seqcount_init()
87  * seqcount_init() - runtime initializer for seqcount_t
102 	seqcount_acquire_read(&l->dep_map, 0, 0, _RET_IP_);  in seqcount_lockdep_reader_access()
103 	seqcount_release(&l->dep_map, _RET_IP_);  in seqcount_lockdep_reader_access()
114  * SEQCNT_ZERO() - static initializer for seqcount_t
122  * A sequence counter which associates the lock used for writer
124  * that the write side critical section is properly serialized.
127  * preemption protection is enforced in the write side function.
135  * For PREEMPT_RT, seqcount_LOCKNAME_t write side critical sections cannot
136  * disable preemption. It can lead to higher latencies, and the write side
140  * To remain preemptible while avoiding a possible livelock caused by the
141  * reader preempting the writer, use a different technique: let the reader
142  * detect if a seqcount_LOCKNAME_t writer is in progress. If that is the
144  * lock. This will allow any possibly-preempted writer to make progress
147  * This lock-unlock technique must be implemented for all of PREEMPT_RT
157  * typedef seqcount_LOCKNAME_t - sequence counter with LOCKNAME associated
161  * A plain sequence counter with external writer synchronization by
164  * that the write side critical section is properly serialized.
170  * seqcount_LOCKNAME_init() - runtime initializer for seqcount_LOCKNAME_t
178 		seqcount_init(&____s->seqcount);			\
179 		__SEQ_LOCK(____s->lock = (_lock));			\
188  * SEQCOUNT_LOCKNAME()	- Instantiate seqcount_LOCKNAME_t and helpers
189  * seqprop_LOCKNAME_*()	- Property accessors for seqcount_LOCKNAME_t
207 	return &s->seqcount;						\
213 	unsigned seq = READ_ONCE(s->seqcount.sequence);			\
220 		__SEQ_LOCK(lockbase##_unlock(s->lock));			\
223 		 * Re-read the sequence counter since the (possibly	\
226 		seq = READ_ONCE(s->seqcount.sequence);			\
259 	return READ_ONCE(s->sequence);  in __seqprop_sequence()
274 …OUNT_LOCKNAME(raw_spinlock, raw_spinlock_t,  false,    s->lock,        raw_spin, raw_spin_lock(s->…
275 SEQCOUNT_LOCKNAME(spinlock,     spinlock_t,      __SEQ_RT, s->lock,        spin,     spin_lock(s->l…
276 SEQCOUNT_LOCKNAME(rwlock,       rwlock_t,        __SEQ_RT, s->lock,        read,     read_lock(s->l…
277 SEQCOUNT_LOCKNAME(mutex,        struct mutex,    true,     s->lock,        mutex,    mutex_lock(s->…
280  * SEQCNT_LOCKNAME_ZERO - static initializer for seqcount_LOCKNAME_t
312  * __read_seqcount_begin() - begin a seqcount_t read section w/o barrier
337  * raw_read_seqcount_begin() - begin a seqcount_t read section w/o lockdep
351  * read_seqcount_begin() - begin a seqcount_t read critical section
363  * raw_read_seqcount() - read the raw seqcount_t counter value
366  * raw_read_seqcount opens a read critical section of the given
383  * raw_seqcount_begin() - begin a seqcount_t read critical section w/o
387  * raw_seqcount_begin opens a read critical section of the given
389  * for the count to stabilize. If a writer is active when it begins, it
394  * small and has a high probability of success through other external
395  * means. It will save a single branching instruction.
409  * __read_seqcount_retry() - end a seqcount_t read section w/o barrier
421  * Return: true if a read section retry is required, else false
429 	return unlikely(READ_ONCE(s->sequence) != start);  in do___read_seqcount_retry()
433  * read_seqcount_retry() - end a seqcount_t read critical section
441  * Return: true if a read section retry is required, else false
453  * raw_write_seqcount_begin() - start a seqcount_t write section w/o lockdep
469 	s->sequence++;  in do_raw_write_seqcount_begin()
474  * raw_write_seqcount_end() - end a seqcount_t write section w/o lockdep
490 	s->sequence++;  in do_raw_write_seqcount_end()
495  * write_seqcount_begin_nested() - start a seqcount_t write section with
500  * See Documentation/locking/lockdep-design.rst
516 	seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_);  in do_write_seqcount_begin_nested()
520  * write_seqcount_begin() - start a seqcount_t write side critical section
523  * Context: sequence counter write side sections must be serialized and
524  * non-preemptible. Preemption will be automatically disabled if and
545  * write_seqcount_end() - end a seqcount_t write side critical section
548  * Context: Preemption will be automatically re-enabled if and only if
561 	seqcount_release(&s->dep_map, _RET_IP_);  in do_write_seqcount_end()
566  * raw_write_seqcount_barrier() - do a seqcount_t write barrier
571  * the two back-to-back wmb()s.
574  * via WRITE_ONCE): a) to ensure the writes become visible to other threads
577  * neither writes before and after the barrier are enclosed in a seq-writer
612 	s->sequence++;  in do_raw_write_seqcount_barrier()
614 	s->sequence++;  in do_raw_write_seqcount_barrier()
619  * write_seqcount_invalidate() - invalidate in-progress seqcount_t read
620  *                               side operations
623  * After write_seqcount_invalidate, no seqcount_t read side operations
633 	s->sequence+=2;  in do_write_seqcount_invalidate()
640  * A sequence counter variant where the counter even/odd value is used to
642  * typically NMIs, to safely interrupt the write side critical section.
652  * SEQCNT_LATCH_ZERO() - static initializer for seqcount_latch_t
660  * seqcount_latch_init() - runtime initializer for seqcount_latch_t
663 #define seqcount_latch_init(s) seqcount_init(&(s)->seqcount)
666  * raw_read_seqcount_latch() - pick even/odd latch data copy
669  * See raw_write_seqcount_latch() for details and a full reader/writer
680 	 * Due to the dependent load, a full smp_rmb() is not needed.  in raw_read_seqcount_latch()
682 	return READ_ONCE(s->seqcount.sequence);  in raw_read_seqcount_latch()
686  * read_seqcount_latch_retry() - end a seqcount_latch_t read section
690  * Return: true if a read section retry is required, else false
695 	return read_seqcount_retry(&s->seqcount, start);  in read_seqcount_latch_retry()
699  * raw_write_seqcount_latch() - redirect latch readers to even/odd copy
702  * The latch technique is a multiversion concurrency control method that allows
703  * queries during non-atomic modifications. If you can guarantee queries never
704  * interrupt the modification -- e.g. the concurrency is strictly between CPUs
705  * -- you most likely do not need this.
709  * latch allows the same for non-atomic updates. The trade-off is doubling the
714  * there is always one copy in a stable state, ready to give us an answer.
716  * The basic form is a data structure like::
723  * Where a modification, which is assumed to be externally serialized, does the
729  *		latch->seq.sequence++;
732  *		modify(latch->data[0], ...);
735  *		latch->seq.sequence++;
738  *		modify(latch->data[1], ...);
741  * The query will have a form like::
749  *			seq = raw_read_seqcount_latch(&latch->seq);
752  *			entry = data_query(latch->data[idx], ...);
755  *		} while (read_seqcount_latch_retry(&latch->seq, seq));
766  *	The non-requirement for atomic modifications does _NOT_ include
767  *	the publishing of new entries in the case where data is a dynamic
776  *	When data is a dynamic data structure; one should use regular RCU
782 	s->seqcount.sequence++;  in raw_write_seqcount_latch()
790  * and non-preemptibility.
793  *    - Comments on top of seqcount_t
794  *    - Documentation/locking/seqlock.rst
812  * seqlock_init() - dynamic initializer for seqlock_t
817 		spin_lock_init(&(sl)->lock);				\
818 		seqcount_spinlock_init(&(sl)->seqcount, &(sl)->lock);	\
822  * DEFINE_SEQLOCK(sl) - Define a statically allocated seqlock_t
829  * read_seqbegin() - start a seqlock_t read side critical section
836 	unsigned ret = read_seqcount_begin(&sl->seqcount);  in read_seqbegin()
838 	kcsan_atomic_next(0);  /* non-raw usage, assume closing read_seqretry() */  in read_seqbegin()
844  * read_seqretry() - end a seqlock_t read side section
848  * read_seqretry closes the read side critical section of given seqlock_t.
852  * Return: true if a read section retry is required, else false
862 	return read_seqcount_retry(&sl->seqcount, start);  in read_seqretry()
866  * For all seqlock_t write side functions, use the the internal
872  * write_seqlock() - start a seqlock_t write side critical section
875  * write_seqlock opens a write side critical section for the given
877  * that sequential lock. All seqlock_t write side sections are thus
878  * automatically serialized and non-preemptible.
880  * Context: if the seqlock_t read section, or other write side critical
886 	spin_lock(&sl->lock);  in write_seqlock()
887 	do_write_seqcount_begin(&sl->seqcount.seqcount);  in write_seqlock()
891  * write_sequnlock() - end a seqlock_t write side critical section
894  * write_sequnlock closes the (serialized and non-preemptible) write side
899 	do_write_seqcount_end(&sl->seqcount.seqcount);  in write_sequnlock()
900 	spin_unlock(&sl->lock);  in write_sequnlock()
904  * write_seqlock_bh() - start a softirqs-disabled seqlock_t write section
907  * _bh variant of write_seqlock(). Use only if the read side section, or
908  * other write side sections, can be invoked from softirq contexts.
912 	spin_lock_bh(&sl->lock);  in write_seqlock_bh()
913 	do_write_seqcount_begin(&sl->seqcount.seqcount);  in write_seqlock_bh()
917  * write_sequnlock_bh() - end a softirqs-disabled seqlock_t write section
920  * write_sequnlock_bh closes the serialized, non-preemptible, and
921  * softirqs-disabled, seqlock_t write side critical section opened with
926 	do_write_seqcount_end(&sl->seqcount.seqcount);  in write_sequnlock_bh()
927 	spin_unlock_bh(&sl->lock);  in write_sequnlock_bh()
931  * write_seqlock_irq() - start a non-interruptible seqlock_t write section
934  * _irq variant of write_seqlock(). Use only if the read side section, or
939 	spin_lock_irq(&sl->lock);  in write_seqlock_irq()
940 	do_write_seqcount_begin(&sl->seqcount.seqcount);  in write_seqlock_irq()
944  * write_sequnlock_irq() - end a non-interruptible seqlock_t write section
947  * write_sequnlock_irq closes the serialized and non-interruptible
948  * seqlock_t write side section opened with write_seqlock_irq().
952 	do_write_seqcount_end(&sl->seqcount.seqcount);  in write_sequnlock_irq()
953 	spin_unlock_irq(&sl->lock);  in write_sequnlock_irq()
960 	spin_lock_irqsave(&sl->lock, flags);  in __write_seqlock_irqsave()
961 	do_write_seqcount_begin(&sl->seqcount.seqcount);  in __write_seqlock_irqsave()
966  * write_seqlock_irqsave() - start a non-interruptible seqlock_t write
969  * @flags: Stack-allocated storage for saving caller's local interrupt
972  * _irqsave variant of write_seqlock(). Use it only if the read side
979  * write_sequnlock_irqrestore() - end non-interruptible seqlock_t write
984  * write_sequnlock_irqrestore closes the serialized and non-interruptible
990 	do_write_seqcount_end(&sl->seqcount.seqcount);  in write_sequnlock_irqrestore()
991 	spin_unlock_irqrestore(&sl->lock, flags);  in write_sequnlock_irqrestore()
995  * read_seqlock_excl() - begin a seqlock_t locking reader section
998  * read_seqlock_excl opens a seqlock_t locking reader critical section.  A
1002  * Locking readers act like a normal spin_lock()/spin_unlock().
1012 	spin_lock(&sl->lock);  in read_seqlock_excl()
1016  * read_sequnlock_excl() - end a seqlock_t locking reader critical section
1021 	spin_unlock(&sl->lock);  in read_sequnlock_excl()
1025  * read_seqlock_excl_bh() - start a seqlock_t locking reader section with
1030  * seqlock_t write side section, *or other read sections*, can be invoked
1035 	spin_lock_bh(&sl->lock);  in read_seqlock_excl_bh()
1039  * read_sequnlock_excl_bh() - stop a seqlock_t softirq-disabled locking
1045 	spin_unlock_bh(&sl->lock);  in read_sequnlock_excl_bh()
1049  * read_seqlock_excl_irq() - start a non-interruptible seqlock_t locking
1054  * write side section, *or other read sections*, can be invoked from a
1059 	spin_lock_irq(&sl->lock);  in read_seqlock_excl_irq()
1063  * read_sequnlock_excl_irq() - end an interrupts-disabled seqlock_t
1069 	spin_unlock_irq(&sl->lock);  in read_sequnlock_excl_irq()
1076 	spin_lock_irqsave(&sl->lock, flags);  in __read_seqlock_excl_irqsave()
1081  * read_seqlock_excl_irqsave() - start a non-interruptible seqlock_t
1084  * @flags: Stack-allocated storage for saving caller's local interrupt
1088  * write side section, *or other read sections*, can be invoked from a
1095  * read_sequnlock_excl_irqrestore() - end non-interruptible seqlock_t
1103 	spin_unlock_irqrestore(&sl->lock, flags);  in read_sequnlock_excl_irqrestore()
1107  * read_seqbegin_or_lock() - begin a seqlock_t lockless or locking reader
1110  * reader will become a *lockless* seqlock_t reader as in read_seqbegin().
1111  * If the passed value is odd, the reader will become a *locking* reader
1113  * caller *must* initialize and pass an even value to @seq; this way, a
1116  * read_seqbegin_or_lock is an API designed to optimistically try a normal
1119  * itself into a full seqlock_t locking reader.
1122  * (too much retry loops) in the case of a sharp spike in write side
1132  * parameter, which is overloaded as a return parameter. This returned
1146  * need_seqretry() - validate seqlock_t "locking or lockless" read section
1150  * Return: true if a read section retry is required, false otherwise
1158  * done_seqretry() - end seqlock_t "locking or lockless" reader section
1162  * done_seqretry finishes the seqlock_t read side critical section started
1172  * read_seqbegin_or_lock_irqsave() - begin a seqlock_t lockless reader, or
1173  *                                   a non-interruptible locking reader
1185  *   1. The saved local interrupts state in case of a locking reader, to
1189  *      overloaded as a return parameter. Check read_seqbegin_or_lock().
1205  * done_seqretry_irqrestore() - end a seqlock_t lockless reader, or a
1206  *				non-interruptible locking reader section
1209  * @flags: Caller's saved local interrupt state in case of a locking