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>
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
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,
35  * atomics; if there is a matching read_seqcount_retry() call, no following
45 	 * Make sure we are not reinitializing a held lock:  in __seqcount_init()
47 	lockdep_init_map(&s->dep_map, name, key, 0);  in __seqcount_init()
48 	s->sequence = 0;  in __seqcount_init()
57  * seqcount_init() - runtime initializer for seqcount_t
72 	seqcount_acquire_read(&l->dep_map, 0, 0, _RET_IP_);  in seqcount_lockdep_reader_access()
73 	seqcount_release(&l->dep_map, _RET_IP_);  in seqcount_lockdep_reader_access()
84  * SEQCNT_ZERO() - static initializer for seqcount_t
92  * A sequence counter which associates the lock used for writer
94  * that the write side critical section is properly serialized.
97  * preemption protection is enforced in the write side function.
105  * typedef seqcount_LOCKNAME_t - sequence counter with LOCKNAME associated
109  * A plain sequence counter with external writer synchronization by
112  * that the write side critical section is properly serialized.
118  * seqcount_LOCKNAME_init() - runtime initializer for seqcount_LOCKNAME_t
126 		seqcount_init(&____s->seqcount);			\
127 		__SEQ_LOCK(____s->lock = (_lock));			\
136  * SEQCOUNT_LOCKNAME()	- Instantiate seqcount_LOCKNAME_t and helpers
137  * seqprop_LOCKNAME_*()	- Property accessors for seqcount_LOCKNAME_t
148 	return &s->seqcount;						\
154 	return &s->seqcount;						\
160 	unsigned seq = READ_ONCE(s->seqcount.sequence);			\
166 		__SEQ_LOCK(lockbase##_lock(s->lock));			\
167 		__SEQ_LOCK(lockbase##_unlock(s->lock));			\
170 		 * Re-read the sequence counter since the (possibly	\
173 		seq = READ_ONCE(s->seqcount.sequence);			\
192 	__SEQ_LOCK(lockdep_assert_held(s->lock));			\
211 	return READ_ONCE(s->sequence);  in __seqprop_sequence()
233  * SEQCNT_LOCKNAME_ZERO - static initializer for seqcount_LOCKNAME_t  in SEQCOUNT_LOCKNAME()
266  * __read_seqcount_begin() - begin a seqcount_t read section w/o barrier
291  * raw_read_seqcount_begin() - begin a seqcount_t read section w/o lockdep
305  * read_seqcount_begin() - begin a seqcount_t read critical section
317  * raw_read_seqcount() - read the raw seqcount_t counter value
320  * raw_read_seqcount opens a read critical section of the given
337  * raw_seqcount_begin() - begin a seqcount_t read critical section w/o
341  * raw_seqcount_begin opens a read critical section of the given
343  * for the count to stabilize. If a writer is active when it begins, it
348  * small and has a high probability of success through other external
349  * means. It will save a single branching instruction.
363  * __read_seqcount_retry() - end a seqcount_t read section w/o barrier
375  * Return: true if a read section retry is required, else false
383 	return unlikely(READ_ONCE(s->sequence) != start);
387  * read_seqcount_retry() - end a seqcount_t read critical section
395  * Return: true if a read section retry is required, else false
407  * raw_write_seqcount_begin() - start a seqcount_t write section w/o lockdep
423 	s->sequence++;  in do_raw_write_seqcount_begin()
428  * raw_write_seqcount_end() - end a seqcount_t write section w/o lockdep
444 	s->sequence++;  in do_raw_write_seqcount_end()
449  * write_seqcount_begin_nested() - start a seqcount_t write section with
454  * See Documentation/locking/lockdep-design.rst
469 	seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_);  in do_write_seqcount_begin_nested()
474  * write_seqcount_begin() - start a seqcount_t write side critical section
477  * Context: sequence counter write side sections must be serialized and
478  * non-preemptible. Preemption will be automatically disabled if and
499  * write_seqcount_end() - end a seqcount_t write side critical section
502  * Context: Preemption will be automatically re-enabled if and only if
515 	seqcount_release(&s->dep_map, _RET_IP_);  in do_write_seqcount_end()
520  * raw_write_seqcount_barrier() - do a seqcount_t write barrier
525  * the two back-to-back wmb()s.
528  * via WRITE_ONCE): a) to ensure the writes become visible to other threads
531  * neither writes before nor after the barrier are enclosed in a seq-writer
566 	s->sequence++;  in do_raw_write_seqcount_barrier()
568 	s->sequence++;  in do_raw_write_seqcount_barrier()
573  * write_seqcount_invalidate() - invalidate in-progress seqcount_t read
574  *                               side operations
577  * After write_seqcount_invalidate, no seqcount_t read side operations
587 	s->sequence+=2;  in do_write_seqcount_invalidate()
594  * A sequence counter variant where the counter even/odd value is used to
596  * typically NMIs, to safely interrupt the write side critical section.
606  * SEQCNT_LATCH_ZERO() - static initializer for seqcount_latch_t
614  * seqcount_latch_init() - runtime initializer for seqcount_latch_t
617 #define seqcount_latch_init(s) seqcount_init(&(s)->seqcount)
620  * raw_read_seqcount_latch() - pick even/odd latch data copy
623  * See raw_write_seqcount_latch() for details and a full reader/writer
634 	 * Due to the dependent load, a full smp_rmb() is not needed.  in raw_read_seqcount_latch()
636 	return READ_ONCE(s->seqcount.sequence);  in raw_read_seqcount_latch()
640  * raw_read_seqcount_latch_retry() - end a seqcount_latch_t read section
644  * Return: true if a read section retry is required, else false
650 	return unlikely(READ_ONCE(s->seqcount.sequence) != start);  in raw_read_seqcount_latch_retry()
654  * raw_write_seqcount_latch() - redirect latch readers to even/odd copy
657  * The latch technique is a multiversion concurrency control method that allows
658  * queries during non-atomic modifications. If you can guarantee queries never
659  * interrupt the modification -- e.g. the concurrency is strictly between CPUs
660  * -- you most likely do not need this.
664  * latch allows the same for non-atomic updates. The trade-off is doubling the
669  * there is always one copy in a stable state, ready to give us an answer.
671  * The basic form is a data structure like::
678  * Where a modification, which is assumed to be externally serialized, does the
684  *		latch->seq.sequence++;
687  *		modify(latch->data[0], ...);
690  *		latch->seq.sequence++;
693  *		modify(latch->data[1], ...);
696  * The query will have a form like::
704  *			seq = raw_read_seqcount_latch(&latch->seq);
707  *			entry = data_query(latch->data[idx], ...);
710  *		} while (raw_read_seqcount_latch_retry(&latch->seq, seq));
721  *	The non-requirement for atomic modifications does _NOT_ include
722  *	the publishing of new entries in the case where data is a dynamic
731  *	When data is a dynamic data structure; one should use regular RCU
737 	s->seqcount.sequence++;  in raw_write_seqcount_latch()
748  * seqlock_init() - dynamic initializer for seqlock_t
753 		spin_lock_init(&(sl)->lock);				\
754 		seqcount_spinlock_init(&(sl)->seqcount, &(sl)->lock);	\
758  * DEFINE_SEQLOCK(sl) - Define a statically allocated seqlock_t
765  * read_seqbegin() - start a seqlock_t read side critical section
772 	unsigned ret = read_seqcount_begin(&sl->seqcount);  in read_seqbegin()
774 	kcsan_atomic_next(0);  /* non-raw usage, assume closing read_seqretry() */  in read_seqbegin()
780  * read_seqretry() - end a seqlock_t read side section
784  * read_seqretry closes the read side critical section of given seqlock_t.
788  * Return: true if a read section retry is required, else false
798 	return read_seqcount_retry(&sl->seqcount, start);  in read_seqretry()
802  * For all seqlock_t write side functions, use the internal
808  * write_seqlock() - start a seqlock_t write side critical section
811  * write_seqlock opens a write side critical section for the given
813  * that sequential lock. All seqlock_t write side sections are thus
814  * automatically serialized and non-preemptible.
816  * Context: if the seqlock_t read section, or other write side critical
822 	spin_lock(&sl->lock);  in write_seqlock()
823 	do_write_seqcount_begin(&sl->seqcount.seqcount);  in write_seqlock()
827  * write_sequnlock() - end a seqlock_t write side critical section
830  * write_sequnlock closes the (serialized and non-preemptible) write side
835 	do_write_seqcount_end(&sl->seqcount.seqcount);  in write_sequnlock()
836 	spin_unlock(&sl->lock);  in write_sequnlock()
840  * write_seqlock_bh() - start a softirqs-disabled seqlock_t write section
843  * _bh variant of write_seqlock(). Use only if the read side section, or
844  * other write side sections, can be invoked from softirq contexts.
848 	spin_lock_bh(&sl->lock);  in write_seqlock_bh()
849 	do_write_seqcount_begin(&sl->seqcount.seqcount);  in write_seqlock_bh()
853  * write_sequnlock_bh() - end a softirqs-disabled seqlock_t write section
856  * write_sequnlock_bh closes the serialized, non-preemptible, and
857  * softirqs-disabled, seqlock_t write side critical section opened with
862 	do_write_seqcount_end(&sl->seqcount.seqcount);  in write_sequnlock_bh()
863 	spin_unlock_bh(&sl->lock);  in write_sequnlock_bh()
867  * write_seqlock_irq() - start a non-interruptible seqlock_t write section
870  * _irq variant of write_seqlock(). Use only if the read side section, or
875 	spin_lock_irq(&sl->lock);  in write_seqlock_irq()
876 	do_write_seqcount_begin(&sl->seqcount.seqcount);  in write_seqlock_irq()
880  * write_sequnlock_irq() - end a non-interruptible seqlock_t write section
883  * write_sequnlock_irq closes the serialized and non-interruptible
884  * seqlock_t write side section opened with write_seqlock_irq().
888 	do_write_seqcount_end(&sl->seqcount.seqcount);  in write_sequnlock_irq()
889 	spin_unlock_irq(&sl->lock);  in write_sequnlock_irq()
896 	spin_lock_irqsave(&sl->lock, flags);  in __write_seqlock_irqsave()
897 	do_write_seqcount_begin(&sl->seqcount.seqcount);  in __write_seqlock_irqsave()
902  * write_seqlock_irqsave() - start a non-interruptible seqlock_t write
905  * @flags: Stack-allocated storage for saving caller's local interrupt
908  * _irqsave variant of write_seqlock(). Use it only if the read side
915  * write_sequnlock_irqrestore() - end non-interruptible seqlock_t write
920  * write_sequnlock_irqrestore closes the serialized and non-interruptible
926 	do_write_seqcount_end(&sl->seqcount.seqcount);  in write_sequnlock_irqrestore()
927 	spin_unlock_irqrestore(&sl->lock, flags);  in write_sequnlock_irqrestore()
931  * read_seqlock_excl() - begin a seqlock_t locking reader section
934  * read_seqlock_excl opens a seqlock_t locking reader critical section.  A
938  * Locking readers act like a normal spin_lock()/spin_unlock().
948 	spin_lock(&sl->lock);  in read_seqlock_excl()
952  * read_sequnlock_excl() - end a seqlock_t locking reader critical section
957 	spin_unlock(&sl->lock);  in read_sequnlock_excl()
961  * read_seqlock_excl_bh() - start a seqlock_t locking reader section with
966  * seqlock_t write side section, *or other read sections*, can be invoked
971 	spin_lock_bh(&sl->lock);  in read_seqlock_excl_bh()
975  * read_sequnlock_excl_bh() - stop a seqlock_t softirq-disabled locking
981 	spin_unlock_bh(&sl->lock);  in read_sequnlock_excl_bh()
985  * read_seqlock_excl_irq() - start a non-interruptible seqlock_t locking
990  * write side section, *or other read sections*, can be invoked from a
995 	spin_lock_irq(&sl->lock);  in read_seqlock_excl_irq()
999  * read_sequnlock_excl_irq() - end an interrupts-disabled seqlock_t
1005 	spin_unlock_irq(&sl->lock);  in read_sequnlock_excl_irq()
1012 	spin_lock_irqsave(&sl->lock, flags);  in __read_seqlock_excl_irqsave()
1017  * read_seqlock_excl_irqsave() - start a non-interruptible seqlock_t
1020  * @flags: Stack-allocated storage for saving caller's local interrupt
1024  * write side section, *or other read sections*, can be invoked from a
1031  * read_sequnlock_excl_irqrestore() - end non-interruptible seqlock_t
1039 	spin_unlock_irqrestore(&sl->lock, flags);  in read_sequnlock_excl_irqrestore()
1043  * read_seqbegin_or_lock() - begin a seqlock_t lockless or locking reader
1046  * reader will become a *lockless* seqlock_t reader as in read_seqbegin().
1047  * If the passed value is odd, the reader will become a *locking* reader
1049  * caller *must* initialize and pass an even value to @seq; this way, a
1052  * read_seqbegin_or_lock is an API designed to optimistically try a normal
1055  * itself into a full seqlock_t locking reader.
1058  * (too much retry loops) in the case of a sharp spike in write side
1068  * parameter, which is overloaded as a return parameter. This returned
1082  * need_seqretry() - validate seqlock_t "locking or lockless" read section
1086  * Return: true if a read section retry is required, false otherwise
1094  * done_seqretry() - end seqlock_t "locking or lockless" reader section
1098  * done_seqretry finishes the seqlock_t read side critical section started
1108  * read_seqbegin_or_lock_irqsave() - begin a seqlock_t lockless reader, or
1109  *                                   a non-interruptible locking reader
1121  *   1. The saved local interrupts state in case of a locking reader, to
1125  *      overloaded as a return parameter. Check read_seqbegin_or_lock().
1141  * done_seqretry_irqrestore() - end a seqlock_t lockless reader, or a
1142  *				non-interruptible locking reader section
1145  * @flags: Caller's saved local interrupt state in case of a locking