libatomic-ops(3) Library providing user level atomic operations

SYNOPSIS

#include <atomic_ops.h>

cc ... -latomic_ops

Note that all operations have an additional barrier option that can be set explicitly.

void AO_load(AO_t *addr)
void AO_store(AO_t *addr, AO_t val)

int AO_test_and_set (AO_t *addr)

AO_t AO_fetch_and_add(AO_t *addr, AO_t incr)
AO_t AO_fetch_and_add1(AO_t *addr)
AO_t AO_fetch_and_sub1(AO_t *addr)

void AO_or(AO_t *p, AO_t incr)
int AO_compare_and_swap(AO_t *addr, AO_t old, AO_t new_val)

DESCRIPTION

libatomic-ops offers a programming interface to a comprehensive range of atomic operations at user level.

We define various atomic operations on memory in a machine-specific way. Unfortunately, this is complicated by the fact that these may or may not be combined with various memory barriers. Thus the actual operations we define have the form AO_<atomic-op>_<barrier> for all plausible combinations of <atomic-op> and <barrier>.

The valid barrier suffixes are

_release
Earlier operations may not be delayed past it.
_acquire
Later operations may not move ahead of it.
_read
Subsequent reads must follow this operation and preceding reads.
_write
Earlier writes precede both this operation and later writes.
_full
Ordered with respect to both earlier and later memops.
_release_write
Ordered with respect to earlier writes.
_acquire_read
Ordered with repsect to later reads.

This of course results in a mild combinatorial explosion.

The library will find the least expensive way to implement your operations on the applicable hardware. In many cases that will involve, for example, a stronger memory barrier, or a combination of hardware primitives.

Note that atomicity guarantees are valid only if both readers and writers use AO_ operations to access the shared value, while ordering constraints are intended to apply all memory operations. If a location can potentially be accessed simultaneously from multiple threads, and one of those accesses may be a write access, then all such accesses to that location should be through AO_ primitives. However if AO_ operations enforce sufficient ordering to ensure that a location x cannot be accessed concurrently, or can only be read concurrently, then x can be accessed via ordinary references and assignments.

All operations operate on an AO_t value, which is the natural word size for the architecture.

AO_load and AO_store load and store the specified pointer address.

AO_test_and_set atomically replaces an address with AO_TS_SET and returns the prior value. An AO_TS_t location can be reset with the AO_CLEAR macro, which usually uses AO_store_release

AO_fetch_and_add takes an address and a value to add.

AO_fetch_and_add1 and AO_fetch_and_sub1 are provided since they may have faster implemenations on some hardware

AO_or atomically ors an AO_t value into a memory location, but does not provide access to the original

AO_compare_and_swap takes an address, an old value and a new value and returns an int. non-zero indicates the compare and swap succeeded.

AUTHOR

This manual page was written by Ian Wienand <[email protected]>, based on comments in the source code. It was written for the Debian project (but may be used by others).