xref: /freebsd/share/man/man9/atomic.9 (revision 193d9e768ba63fcfb187cfd17f461f7d41345048)
1.\" Copyright (c) 2000-2001 John H. Baldwin <jhb@FreeBSD.org>
2.\" All rights reserved.
3.\"
4.\" Redistribution and use in source and binary forms, with or without
5.\" modification, are permitted provided that the following conditions
6.\" are met:
7.\" 1. Redistributions of source code must retain the above copyright
8.\"    notice, this list of conditions and the following disclaimer.
9.\" 2. Redistributions in binary form must reproduce the above copyright
10.\"    notice, this list of conditions and the following disclaimer in the
11.\"    documentation and/or other materials provided with the distribution.
12.\"
13.\" THIS SOFTWARE IS PROVIDED BY THE DEVELOPERS ``AS IS'' AND ANY EXPRESS OR
14.\" IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
15.\" OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
16.\" IN NO EVENT SHALL THE DEVELOPERS BE LIABLE FOR ANY DIRECT, INDIRECT,
17.\" INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
18.\" NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
19.\" DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
20.\" THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
21.\" (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
22.\" THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
23.\"
24.\" $FreeBSD$
25.\"
26.Dd Jan 3, 2017
27.Dt ATOMIC 9
28.Os
29.Sh NAME
30.Nm atomic_add ,
31.Nm atomic_clear ,
32.Nm atomic_cmpset ,
33.Nm atomic_fcmpset ,
34.Nm atomic_fetchadd ,
35.Nm atomic_load ,
36.Nm atomic_readandclear ,
37.Nm atomic_set ,
38.Nm atomic_subtract ,
39.Nm atomic_store
40.Nd atomic operations
41.Sh SYNOPSIS
42.In sys/types.h
43.In machine/atomic.h
44.Ft void
45.Fn atomic_add_[acq_|rel_]<type> "volatile <type> *p" "<type> v"
46.Ft void
47.Fn atomic_clear_[acq_|rel_]<type> "volatile <type> *p" "<type> v"
48.Ft int
49.Fo atomic_cmpset_[acq_|rel_]<type>
50.Fa "volatile <type> *dst"
51.Fa "<type> old"
52.Fa "<type> new"
53.Fc
54.Ft int
55.Fo atomic_fcmpset_[acq_|rel_]<type>
56.Fa "volatile <type> *dst"
57.Fa "<type> *old"
58.Fa "<type> new"
59.Fc
60.Ft <type>
61.Fn atomic_fetchadd_<type> "volatile <type> *p" "<type> v"
62.Ft <type>
63.Fn atomic_load_acq_<type> "volatile <type> *p"
64.Ft <type>
65.Fn atomic_readandclear_<type> "volatile <type> *p"
66.Ft void
67.Fn atomic_set_[acq_|rel_]<type> "volatile <type> *p" "<type> v"
68.Ft void
69.Fn atomic_subtract_[acq_|rel_]<type> "volatile <type> *p" "<type> v"
70.Ft void
71.Fn atomic_store_rel_<type> "volatile <type> *p" "<type> v"
72.Ft <type>
73.Fn atomic_swap_<type> "volatile <type> *p" "<type> v"
74.Ft int
75.Fn atomic_testandclear_<type> "volatile <type> *p" "u_int v"
76.Ft int
77.Fn atomic_testandset_<type> "volatile <type> *p" "u_int v"
78.Sh DESCRIPTION
79Each of the atomic operations is guaranteed to be atomic across multiple
80threads and in the presence of interrupts.
81They can be used to implement reference counts or as building blocks for more
82advanced synchronization primitives such as mutexes.
83.Ss Types
84Each atomic operation operates on a specific
85.Fa type .
86The type to use is indicated in the function name.
87The available types that can be used are:
88.Pp
89.Bl -tag -offset indent -width short -compact
90.It Li int
91unsigned integer
92.It Li long
93unsigned long integer
94.It Li ptr
95unsigned integer the size of a pointer
96.It Li 32
97unsigned 32-bit integer
98.It Li 64
99unsigned 64-bit integer
100.El
101.Pp
102For example, the function to atomically add two integers is called
103.Fn atomic_add_int .
104.Pp
105Certain architectures also provide operations for types smaller than
106.Dq Li int .
107.Pp
108.Bl -tag -offset indent -width short -compact
109.It Li char
110unsigned character
111.It Li short
112unsigned short integer
113.It Li 8
114unsigned 8-bit integer
115.It Li 16
116unsigned 16-bit integer
117.El
118.Pp
119These must not be used in MI code because the instructions to implement them
120efficiently might not be available.
121.Ss Acquire and Release Operations
122By default, a thread's accesses to different memory locations might not be
123performed in
124.Em program order ,
125that is, the order in which the accesses appear in the source code.
126To optimize the program's execution, both the compiler and processor might
127reorder the thread's accesses.
128However, both ensure that their reordering of the accesses is not visible to
129the thread.
130Otherwise, the traditional memory model that is expected by single-threaded
131programs would be violated.
132Nonetheless, other threads in a multithreaded program, such as the
133.Fx
134kernel, might observe the reordering.
135Moreover, in some cases, such as the implementation of synchronization between
136threads, arbitrary reordering might result in the incorrect execution of the
137program.
138To constrain the reordering that both the compiler and processor might perform
139on a thread's accesses, the thread should use atomic operations with
140.Em acquire
141and
142.Em release
143semantics.
144.Pp
145Most of the atomic operations on memory have three variants.
146The first variant performs the operation without imposing any ordering
147constraints on memory accesses to other locations.
148The second variant has acquire semantics, and the third variant has release
149semantics.
150In effect, operations with acquire and release semantics establish one-way
151barriers to reordering.
152.Pp
153When an atomic operation has acquire semantics, the effects of the operation
154must have completed before any subsequent load or store (by program order) is
155performed.
156Conversely, acquire semantics do not require that prior loads or stores have
157completed before the atomic operation is performed.
158To denote acquire semantics, the suffix
159.Dq Li _acq
160is inserted into the function name immediately prior to the
161.Dq Li _ Ns Aq Fa type
162suffix.
163For example, to subtract two integers ensuring that subsequent loads and
164stores happen after the subtraction is performed, use
165.Fn atomic_subtract_acq_int .
166.Pp
167When an atomic operation has release semantics, the effects of all prior
168loads or stores (by program order) must have completed before the operation
169is performed.
170Conversely, release semantics do not require that the effects of the
171atomic operation must have completed before any subsequent load or store is
172performed.
173To denote release semantics, the suffix
174.Dq Li _rel
175is inserted into the function name immediately prior to the
176.Dq Li _ Ns Aq Fa type
177suffix.
178For example, to add two long integers ensuring that all prior loads and
179stores happen before the addition, use
180.Fn atomic_add_rel_long .
181.Pp
182The one-way barriers provided by acquire and release operations allow the
183implementations of common synchronization primitives to express their
184ordering requirements without also imposing unnecessary ordering.
185For example, for a critical section guarded by a mutex, an acquire operation
186when the mutex is locked and a release operation when the mutex is unlocked
187will prevent any loads or stores from moving outside of the critical
188section.
189However, they will not prevent the compiler or processor from moving loads
190or stores into the critical section, which does not violate the semantics of
191a mutex.
192.Ss Multiple Processors
193In multiprocessor systems, the atomicity of the atomic operations on memory
194depends on support for cache coherence in the underlying architecture.
195In general, cache coherence on the default memory type,
196.Dv VM_MEMATTR_DEFAULT ,
197is guaranteed by all architectures that are supported by
198.Fx .
199For example, cache coherence is guaranteed on write-back memory by the
200.Tn amd64
201and
202.Tn i386
203architectures.
204However, on some architectures, cache coherence might not be enabled on all
205memory types.
206To determine if cache coherence is enabled for a non-default memory type,
207consult the architecture's documentation.
208.Ss Semantics
209This section describes the semantics of each operation using a C like notation.
210.Bl -hang
211.It Fn atomic_add p v
212.Bd -literal -compact
213*p += v;
214.Ed
215.It Fn atomic_clear p v
216.Bd -literal -compact
217*p &= ~v;
218.Ed
219.It Fn atomic_cmpset dst old new
220.Bd -literal -compact
221if (*dst == old) {
222	*dst = new;
223	return (1);
224} else
225	return (0);
226.Ed
227.El
228.Pp
229The
230.Fn atomic_cmpset
231functions are not implemented for the types
232.Dq Li char ,
233.Dq Li short ,
234.Dq Li 8 ,
235and
236.Dq Li 16 .
237.Bl -hang
238.It Fn atomic_fcmpset dst *old new
239.El
240.Pp
241On architectures implementing
242.Em Compare And Swap
243operation in hardware, the functionality can be described as
244.Bd -literal -offset indent -compact
245if (*dst == *old) {
246	*dst = new;
247	return (1);
248} else {
249	*old = *dst;
250	return (0);
251}
252.Ed
253On architectures which provide
254.Em Load Linked/Store Conditional
255primitive, the write to
256.Dv *dst
257might also fail for several reasons, most important of which
258is a parallel write to
259.Dv *dst
260cache line by other CPU.
261In this case
262.Fn atomic_fcmpset
263function also returns
264.Dv false ,
265despite
266.Dl *old == *dst .
267.Pp
268The
269.Fn atomic_fcmpset
270functions are not implemented for the types
271.Dq Li char ,
272.Dq Li short ,
273.Dq Li 8 ,
274and
275.Dq Li 16 .
276.Bl -hang
277.It Fn atomic_fetchadd p v
278.Bd -literal -compact
279tmp = *p;
280*p += v;
281return (tmp);
282.Ed
283.El
284.Pp
285The
286.Fn atomic_fetchadd
287functions are only implemented for the types
288.Dq Li int ,
289.Dq Li long
290and
291.Dq Li 32
292and do not have any variants with memory barriers at this time.
293.Bl -hang
294.It Fn atomic_load p
295.Bd -literal -compact
296return (*p);
297.Ed
298.El
299.Pp
300The
301.Fn atomic_load
302functions are only provided with acquire memory barriers.
303.Bl -hang
304.It Fn atomic_readandclear p
305.Bd -literal -compact
306tmp = *p;
307*p = 0;
308return (tmp);
309.Ed
310.El
311.Pp
312The
313.Fn atomic_readandclear
314functions are not implemented for the types
315.Dq Li char ,
316.Dq Li short ,
317.Dq Li ptr ,
318.Dq Li 8 ,
319and
320.Dq Li 16
321and do not have any variants with memory barriers at this time.
322.Bl -hang
323.It Fn atomic_set p v
324.Bd -literal -compact
325*p |= v;
326.Ed
327.It Fn atomic_subtract p v
328.Bd -literal -compact
329*p -= v;
330.Ed
331.It Fn atomic_store p v
332.Bd -literal -compact
333*p = v;
334.Ed
335.El
336.Pp
337The
338.Fn atomic_store
339functions are only provided with release memory barriers.
340.Bl -hang
341.It Fn atomic_swap p v
342.Bd -literal -compact
343tmp = *p;
344*p = v;
345return (tmp);
346.Ed
347.El
348.Pp
349The
350.Fn atomic_swap
351functions are not implemented for the types
352.Dq Li char ,
353.Dq Li short ,
354.Dq Li ptr ,
355.Dq Li 8 ,
356and
357.Dq Li 16
358and do not have any variants with memory barriers at this time.
359.Bl -hang
360.It Fn atomic_testandclear p v
361.Bd -literal -compact
362bit = 1 << (v % (sizeof(*p) * NBBY));
363tmp = (*p & bit) != 0;
364*p &= ~bit;
365return (tmp);
366.Ed
367.El
368.Bl -hang
369.It Fn atomic_testandset p v
370.Bd -literal -compact
371bit = 1 << (v % (sizeof(*p) * NBBY));
372tmp = (*p & bit) != 0;
373*p |= bit;
374return (tmp);
375.Ed
376.El
377.Pp
378The
379.Fn atomic_testandset
380and
381.Fn atomic_testandclear
382functions are only implemented for the types
383.Dq Li int ,
384.Dq Li long
385and
386.Dq Li 32
387and do not have any variants with memory barriers at this time.
388.Pp
389The type
390.Dq Li 64
391is currently not implemented for any of the atomic operations on the
392.Tn arm ,
393.Tn i386 ,
394and
395.Tn powerpc
396architectures.
397.Sh RETURN VALUES
398The
399.Fn atomic_cmpset
400function returns the result of the compare operation.
401The
402.Fn atomic_fcmpset
403function returns
404.Dv true
405if the operation succeeded.
406Otherwise it returns
407.Dv false
408and sets
409.Va *old
410to the found value.
411The
412.Fn atomic_fetchadd ,
413.Fn atomic_load ,
414.Fn atomic_readandclear ,
415and
416.Fn atomic_swap
417functions return the value at the specified address.
418The
419.Fn atomic_testandset
420and
421.Fn atomic_testandclear
422function returns the result of the test operation.
423.Sh EXAMPLES
424This example uses the
425.Fn atomic_cmpset_acq_ptr
426and
427.Fn atomic_set_ptr
428functions to obtain a sleep mutex and handle recursion.
429Since the
430.Va mtx_lock
431member of a
432.Vt "struct mtx"
433is a pointer, the
434.Dq Li ptr
435type is used.
436.Bd -literal
437/* Try to obtain mtx_lock once. */
438#define _obtain_lock(mp, tid)						\\
439	atomic_cmpset_acq_ptr(&(mp)->mtx_lock, MTX_UNOWNED, (tid))
440
441/* Get a sleep lock, deal with recursion inline. */
442#define _get_sleep_lock(mp, tid, opts, file, line) do {			\\
443	uintptr_t _tid = (uintptr_t)(tid);				\\
444									\\
445	if (!_obtain_lock(mp, tid)) {					\\
446		if (((mp)->mtx_lock & MTX_FLAGMASK) != _tid)		\\
447			_mtx_lock_sleep((mp), _tid, (opts), (file), (line));\\
448		else {							\\
449			atomic_set_ptr(&(mp)->mtx_lock, MTX_RECURSE);	\\
450			(mp)->mtx_recurse++;				\\
451		}							\\
452	}								\\
453} while (0)
454.Ed
455.Sh HISTORY
456The
457.Fn atomic_add ,
458.Fn atomic_clear ,
459.Fn atomic_set ,
460and
461.Fn atomic_subtract
462operations were first introduced in
463.Fx 3.0 .
464This first set only supported the types
465.Dq Li char ,
466.Dq Li short ,
467.Dq Li int ,
468and
469.Dq Li long .
470The
471.Fn atomic_cmpset ,
472.Fn atomic_load ,
473.Fn atomic_readandclear ,
474and
475.Fn atomic_store
476operations were added in
477.Fx 5.0 .
478The types
479.Dq Li 8 ,
480.Dq Li 16 ,
481.Dq Li 32 ,
482.Dq Li 64 ,
483and
484.Dq Li ptr
485and all of the acquire and release variants
486were added in
487.Fx 5.0
488as well.
489The
490.Fn atomic_fetchadd
491operations were added in
492.Fx 6.0 .
493The
494.Fn atomic_swap
495and
496.Fn atomic_testandset
497operations were added in
498.Fx 10.0 .
499.Fn atomic_testandclear
500operation was added in
501.Fx 11.0 .
502