| blob | 7d4096437a74c2455b386c29a9bd84bf4469a4a8 |
1 .\" Copyright (C) 2002 Robert Love
2 .\" and Copyright (C) 2006, 2015 Michael Kerrisk
3 .\"
4 .\" %%%LICENSE_START(GPLv2+_DOC_FULL)
5 .\" This is free documentation; you can redistribute it and/or
6 .\" modify it under the terms of the GNU General Public License as
7 .\" published by the Free Software Foundation; either version 2 of
8 .\" the License, or (at your option) any later version.
9 .\"
10 .\" The GNU General Public License's references to "object code"
11 .\" and "executables" are to be interpreted as the output of any
12 .\" document formatting or typesetting system, including
13 .\" intermediate and printed output.
14 .\"
15 .\" This manual is distributed in the hope that it will be useful,
16 .\" but WITHOUT ANY WARRANTY; without even the implied warranty of
17 .\" MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 .\" GNU General Public License for more details.
19 .\"
20 .\" You should have received a copy of the GNU General Public
21 .\" License along with this manual; if not, see
22 .\" <http://www.gnu.org/licenses/>.
23 .\" %%%LICENSE_END
24 .\"
25 .\" 2002-11-19 Robert Love <rml@tech9.net> - initial version
26 .\" 2004-04-20 mtk - fixed description of return value
27 .\" 2004-04-22 aeb - added glibc prototype history
28 .\" 2005-05-03 mtk - noted that sched_setaffinity may cause thread
29 .\" migration and that CPU affinity is a per-thread attribute.
30 .\" 2006-02-03 mtk -- Major rewrite
31 .\" 2008-11-12, mtk, removed CPU_*() macro descriptions to a
32 .\" separate CPU_SET(3) page.
33 .\"
34 .TH SCHED_SETAFFINITY 2 2017-09-15 "Linux" "Linux Programmer's Manual"
35 .SH NAME
36 sched_setaffinity, sched_getaffinity \- \
37 set and get a thread's CPU affinity mask
38 .SH SYNOPSIS
39 .nf
40 .BR "#define _GNU_SOURCE" " /* See feature_test_macros(7) */"
41 .B #include <sched.h>
42 .PP
43 .BI "int sched_setaffinity(pid_t " pid ", size_t " cpusetsize ,
44 .BI " const cpu_set_t *" mask );
45 .PP
46 .BI "int sched_getaffinity(pid_t " pid ", size_t " cpusetsize ,
47 .BI " cpu_set_t *" mask );
48 .fi
49 .SH DESCRIPTION
50 A thread's CPU affinity mask determines the set of CPUs on which
51 it is eligible to run.
52 On a multiprocessor system, setting the CPU affinity mask
53 can be used to obtain performance benefits.
54 For example,
55 by dedicating one CPU to a particular thread
56 (i.e., setting the affinity mask of that thread to specify a single CPU,
57 and setting the affinity mask of all other threads to exclude that CPU),
58 it is possible to ensure maximum execution speed for that thread.
59 Restricting a thread to run on a single CPU also avoids
60 the performance cost caused by the cache invalidation that occurs
61 when a thread ceases to execute on one CPU and then
62 recommences execution on a different CPU.
63 .PP
64 A CPU affinity mask is represented by the
65 .I cpu_set_t
66 structure, a "CPU set", pointed to by
67 .IR mask .
68 A set of macros for manipulating CPU sets is described in
69 .BR CPU_SET (3).
70 .PP
71 .BR sched_setaffinity ()
72 sets the CPU affinity mask of the thread whose ID is
73 .I pid
74 to the value specified by
75 .IR mask .
76 If
77 .I pid
78 is zero, then the calling thread is used.
79 The argument
80 .I cpusetsize
81 is the length (in bytes) of the data pointed to by
82 .IR mask .
83 Normally this argument would be specified as
84 .IR "sizeof(cpu_set_t)" .
85 .PP
86 If the thread specified by
87 .I pid
88 is not currently running on one of the CPUs specified in
89 .IR mask ,
90 then that thread is migrated to one of the CPUs specified in
91 .IR mask .
92 .PP
93 .BR sched_getaffinity ()
94 writes the affinity mask of the thread whose ID is
95 .I pid
96 into the
97 .I cpu_set_t
98 structure pointed to by
99 .IR mask .
100 The
101 .I cpusetsize
102 argument specifies the size (in bytes) of
103 .IR mask .
104 If
105 .I pid
106 is zero, then the mask of the calling thread is returned.
107 .SH RETURN VALUE
108 On success,
109 .BR sched_setaffinity ()
110 and
111 .BR sched_getaffinity ()
112 return 0.
113 On error, \-1 is returned, and
114 .I errno
115 is set appropriately.
116 .SH ERRORS
117 .TP
118 .B EFAULT
119 A supplied memory address was invalid.
120 .TP
121 .B EINVAL
122 The affinity bit mask
123 .I mask
124 contains no processors that are currently physically on the system
125 and permitted to the thread according to any restrictions that
126 may be imposed by
127 .I cpuset
128 cgroups or the "cpuset" mechanism described in
129 .BR cpuset (7).
130 .TP
131 .B EINVAL
132 .RB ( sched_getaffinity ()
133 and, in kernels before 2.6.9,
134 .BR sched_setaffinity ())
135 .I cpusetsize
136 is smaller than the size of the affinity mask used by the kernel.
137 .TP
138 .B EPERM
139 .RB ( sched_setaffinity ())
140 The calling thread does not have appropriate privileges.
141 The caller needs an effective user ID equal to the real user ID
142 or effective user ID of the thread identified by
143 .IR pid ,
144 or it must possess the
145 .B CAP_SYS_NICE
146 capability in the user namespace of the thread
147 .IR pid .
148 .TP
149 .B ESRCH
150 The thread whose ID is \fIpid\fP could not be found.
151 .SH VERSIONS
152 The CPU affinity system calls were introduced in Linux kernel 2.5.8.
153 The system call wrappers were introduced in glibc 2.3.
154 Initially, the glibc interfaces included a
155 .I cpusetsize
156 argument, typed as
157 .IR "unsigned int" .
158 In glibc 2.3.3, the
159 .I cpusetsize
160 argument was removed, but was then restored in glibc 2.3.4, with type
161 .IR size_t .
162 .SH CONFORMING TO
163 These system calls are Linux-specific.
164 .SH NOTES
165 After a call to
166 .BR sched_setaffinity (),
167 the set of CPUs on which the thread will actually run is
168 the intersection of the set specified in the
169 .I mask
170 argument and the set of CPUs actually present on the system.
171 The system may further restrict the set of CPUs on which the thread
172 runs if the "cpuset" mechanism described in
173 .BR cpuset (7)
174 is being used.
175 These restrictions on the actual set of CPUs on which the thread
176 will run are silently imposed by the kernel.
177 .PP
178 There are various ways of determining the number of CPUs
179 available on the system, including: inspecting the contents of
180 .IR /proc/cpuinfo ;
181 using
182 .BR sysconf (3)
183 to obtain the values of the
184 .BR _SC_NPROCESSORS_CONF
185 and
186 .BR _SC_NPROCESSORS_ONLN
187 parameters; and inspecting the list of CPU directories under
188 .IR /sys/devices/system/cpu/ .
189 .PP
190 .BR sched (7)
191 has a description of the Linux scheduling scheme.
192 .PP
193 The affinity mask is a per-thread attribute that can be
194 adjusted independently for each of the threads in a thread group.
195 The value returned from a call to
196 .BR gettid (2)
197 can be passed in the argument
198 .IR pid .
199 Specifying
200 .I pid
201 as 0 will set the attribute for the calling thread,
202 and passing the value returned from a call to
203 .BR getpid (2)
204 will set the attribute for the main thread of the thread group.
205 (If you are using the POSIX threads API, then use
206 .BR pthread_setaffinity_np (3)
207 instead of
208 .BR sched_setaffinity ().)
209 .PP
210 The
211 .I isolcpus
212 boot option can be used to isolate one or more CPUs at boot time,
213 so that no processes are scheduled onto those CPUs.
214 Following the use of this boot option,
215 the only way to schedule processes onto the isolated CPUs is via
216 .BR sched_setaffinity ()
217 or the
218 .BR cpuset (7)
219 mechanism.
220 For further information, see the kernel source file
221 .IR Documentation/admin-guide/kernel-parameters.txt .
222 As noted in that file,
223 .I isolcpus
224 is the preferred mechanism of isolating CPUs
225 (versus the alternative of manually setting the CPU affinity
226 of all processes on the system).
227 .PP
228 A child created via
229 .BR fork (2)
230 inherits its parent's CPU affinity mask.
231 The affinity mask is preserved across an
232 .BR execve (2).
233 .SS C library/kernel differences
234 This manual page describes the glibc interface for the CPU affinity calls.
235 The actual system call interface is slightly different, with the
236 .I mask
237 being typed as
238 .IR "unsigned long\ *" ,
239 reflecting the fact that the underlying implementation of CPU
240 sets is a simple bit mask.
241 On success, the raw
242 .BR sched_getaffinity ()
243 system call returns the size (in bytes) of the
244 .I cpumask_t
245 data type that is used internally by the kernel to
246 represent the CPU set bit mask.
247 .SS Handling systems with large CPU affinity masks
248 The underlying system calls (which represent CPU masks as bit masks of type
249 .IR "unsigned long\ *" )
250 impose no restriction on the size of the CPU mask.
251 However, the
252 .I cpu_set_t
253 data type used by glibc has a fixed size of 128 bytes,
254 meaning that the maximum CPU number that can be represented is 1023.
255 .\" FIXME . See https://sourceware.org/bugzilla/show_bug.cgi?id=15630
256 .\" and https://sourceware.org/ml/libc-alpha/2013-07/msg00288.html
257 If the kernel CPU affinity mask is larger than 1024,
258 then calls of the form:
259 .PP
260 sched_getaffinity(pid, sizeof(cpu_set_t), &mask);
261 .PP
262 fail with the error
263 .BR EINVAL ,
264 the error produced by the underlying system call for the case where the
265 .I mask
266 size specified in
267 .I cpusetsize
268 is smaller than the size of the affinity mask used by the kernel.
269 (Depending on the system CPU topology, the kernel affinity mask can
270 be substantially larger than the number of active CPUs in the system.)
271 .PP
272 When working on systems with large kernel CPU affinity masks,
273 one must dynamically allocate the
274 .I mask
275 argument (see
276 .BR CPU_ALLOC (3)).
277 Currently, the only way to do this is by probing for the size
278 of the required mask using
279 .BR sched_getaffinity ()
280 calls with increasing mask sizes (until the call does not fail with the error
281 .BR EINVAL ).
282 .PP
283 Be aware that
284 .BR CPU_ALLOC (3)
285 may allocate a slightly larger CPU set than requested
286 (because CPU sets are implemented as bit masks allocated in units of
287 .IR sizeof(long) ).
288 Consequently,
289 .BR sched_getaffinity ()
290 can set bits beyond the requested allocation size, because the kernel
291 sees a few additional bits.
292 Therefore, the caller should iterate over the bits in the returned set,
293 counting those which are set, and stop upon reaching the value returned by
294 .BR CPU_COUNT (3)
295 (rather than iterating over the number of bits
296 requested to be allocated).
297 .SH EXAMPLE
298 The program below creates a child process.
299 The parent and child then each assign themselves to a specified CPU
300 and execute identical loops that consume some CPU time.
301 Before terminating, the parent waits for the child to complete.
302 The program takes three command-line arguments:
303 the CPU number for the parent,
304 the CPU number for the child,
305 and the number of loop iterations that both processes should perform.
306 .PP
307 As the sample runs below demonstrate, the amount of real and CPU time
308 consumed when running the program will depend on intra-core caching effects
309 and whether the processes are using the same CPU.
310 .PP
311 We first employ
312 .BR lscpu (1)
313 to determine that this (x86)
314 system has two cores, each with two CPUs:
315 .PP
316 .in +4n
317 .EX
318 $ \fBlscpu | grep -i 'core.*:|socket'\fP
319 Thread(s) per core: 2
320 Core(s) per socket: 2
321 Socket(s): 1
322 .EE
323 .in
324 .PP
325 We then time the operation of the example program for three cases:
326 both processes running on the same CPU;
327 both processes running on different CPUs on the same core;
328 and both processes running on different CPUs on different cores.
329 .PP
330 .in +4n
331 .EX
332 $ \fBtime \-p ./a.out 0 0 100000000\fP
333 real 14.75
334 user 3.02
335 sys 11.73
336 $ \fBtime \-p ./a.out 0 1 100000000\fP
337 real 11.52
338 user 3.98
339 sys 19.06
340 $ \fBtime \-p ./a.out 0 3 100000000\fP
341 real 7.89
342 user 3.29
343 sys 12.07
344 .EE
345 .in
346 .SS Program source
347 \&
348 .EX
349 #define _GNU_SOURCE
350 #include <sched.h>
351 #include <stdio.h>
352 #include <stdlib.h>
353 #include <unistd.h>
354 #include <sys/wait.h>
356 #define errExit(msg) do { perror(msg); exit(EXIT_FAILURE); \\
357 } while (0)
359 int
360 main(int argc, char *argv[])
361 {
362 cpu_set_t set;
363 int parentCPU, childCPU;
364 int nloops, j;
366 if (argc != 4) {
367 fprintf(stderr, "Usage: %s parent\-cpu child\-cpu num\-loops\\n",
368 argv[0]);
369 exit(EXIT_FAILURE);
370 }
372 parentCPU = atoi(argv[1]);
373 childCPU = atoi(argv[2]);
374 nloops = atoi(argv[3]);
376 CPU_ZERO(&set);
378 switch (fork()) {
379 case \-1: /* Error */
380 errExit("fork");
382 case 0: /* Child */
383 CPU_SET(childCPU, &set);
385 if (sched_setaffinity(getpid(), sizeof(set), &set) == \-1)
386 errExit("sched_setaffinity");
388 for (j = 0; j < nloops; j++)
389 getppid();
391 exit(EXIT_SUCCESS);
393 default: /* Parent */
394 CPU_SET(parentCPU, &set);
396 if (sched_setaffinity(getpid(), sizeof(set), &set) == \-1)
397 errExit("sched_setaffinity");
399 for (j = 0; j < nloops; j++)
400 getppid();
402 wait(NULL); /* Wait for child to terminate */
403 exit(EXIT_SUCCESS);
404 }
405 }
406 .EE
407 .SH SEE ALSO
408 .ad l
409 .nh
410 .BR lscpu (1),
411 .BR nproc (1),
412 .BR taskset (1),
413 .BR clone (2),
414 .BR getcpu (2),
415 .BR getpriority (2),
416 .BR gettid (2),
417 .BR nice (2),
418 .BR sched_get_priority_max (2),
419 .BR sched_get_priority_min (2),
420 .BR sched_getscheduler (2),
421 .BR sched_setscheduler (2),
422 .BR setpriority (2),
423 .BR CPU_SET (3),
424 .BR get_nprocs (3),
425 .BR pthread_setaffinity_np (3),
426 .BR sched_getcpu (3),
427 .BR capabilities (7),
428 .BR cpuset (7),
429 .BR sched (7),
430 .BR numactl (8)