man2/set_mempolicy.2

   1 .\" SPDX-License-Identifier: Linux-man-pages-copyleft-var
   2 .\"
   3 .\" Copyright 2003,2004 Andi Kleen, SuSE Labs.
   4 .\" and Copyright 2007 Lee Schermerhorn, Hewlett Packard
   5 .\"
   6 .\" 2006-02-03, mtk, substantial wording changes and other improvements
   7 .\" 2007-08-27, Lee Schermerhorn <Lee.Schermerhorn@hp.com>
   8 .\"     more precise specification of behavior.
   9 .\"
  10 .TH set_mempolicy 2 (date) "Linux man-pages (unreleased)"
  11 .SH NAME
  12 set_mempolicy \- set default NUMA memory policy for a thread and its children
  13 .SH LIBRARY
  14 NUMA (Non-Uniform Memory Access) policy library
  15 .RI ( libnuma ", " \-lnuma )
  16 .SH SYNOPSIS
  17 .nf
  18 .B "#include <numaif.h>"
  19 .P
  20 .BI "long set_mempolicy(int " mode ", const unsigned long *" nodemask ,
  21 .BI "                   unsigned long " maxnode );
  22 .fi
  23 .SH DESCRIPTION
  24 .BR set_mempolicy ()
  25 sets the NUMA memory policy of the calling thread,
  26 which consists of a policy mode and zero or more nodes,
  27 to the values specified by the
  28 .IR mode ,
  29 .IR nodemask ,
  30 and
  31 .I maxnode
  32 arguments.
  33 .P
  34 A NUMA machine has different
  35 memory controllers with different distances to specific CPUs.
  36 The memory policy defines from which node memory is allocated for
  37 the thread.
  38 .P
  39 This system call defines the default policy for the thread.
  40 The thread policy governs allocation of pages in the process's
  41 address space outside of memory ranges
  42 controlled by a more specific policy set by
  43 .BR mbind (2).
  44 The thread default policy also controls allocation of any pages for
  45 memory-mapped files mapped using the
  46 .BR mmap (2)
  47 call with the
  48 .B MAP_PRIVATE
  49 flag and that are only read (loaded) from by the thread
  50 and of memory-mapped files mapped using the
  51 .BR mmap (2)
  52 call with the
  53 .B MAP_SHARED
  54 flag, regardless of the access type.
  55 The policy is applied only when a new page is allocated
  56 for the thread.
  57 For anonymous memory this is when the page is first
  58 touched by the thread.
  59 .P
  60 The
  61 .I mode
  62 argument must specify one of
  63 .BR MPOL_DEFAULT ,
  64 .BR MPOL_BIND ,
  65 .BR MPOL_INTERLEAVE ,
  66 .BR MPOL_PREFERRED ,
  67 or
  68 .B MPOL_LOCAL
  69 (which are described in detail below).
  70 All modes except
  71 .B MPOL_DEFAULT
  72 require the caller to specify the node or nodes to which the mode applies,
  73 via the
  74 .I nodemask
  75 argument.
  76 .P
  77 The
  78 .I mode
  79 argument may also include an optional
  80 .IR "mode flag" .
  81 The supported
  82 .I "mode flags"
  83 are:
  84 .TP
  85 .BR MPOL_F_NUMA_BALANCING " (since Linux 5.12)"
  86 .\" commit bda420b985054a3badafef23807c4b4fa38a3dff
  87 When
  88 .I mode
  89 is
  90 .BR MPOL_BIND ,
  91 enable the kernel NUMA balancing for the task if it is supported by the kernel.
  92 If the flag isn't supported by the kernel, or is used with
  93 .I mode
  94 other than
  95 .BR MPOL_BIND ,
  96 \-1 is returned and
  97 .I errno
  98 is set to
  99 .BR EINVAL .
 100 .TP
 101 .BR MPOL_F_RELATIVE_NODES " (since Linux 2.6.26)"
 102 A nonempty
 103 .I nodemask
 104 specifies node IDs that are relative to the
 105 set of node IDs allowed by the process's current cpuset.
 106 .TP
 107 .BR MPOL_F_STATIC_NODES " (since Linux 2.6.26)"
 108 A nonempty
 109 .I nodemask
 110 specifies physical node IDs.
 111 Linux will not remap the
 112 .I nodemask
 113 when the process moves to a different cpuset context,
 114 nor when the set of nodes allowed by the process's
 115 current cpuset context changes.
 116 .P
 117 .I nodemask
 118 points to a bit mask of node IDs that contains up to
 119 .I maxnode
 120 bits.
 121 The bit mask size is rounded to the next multiple of
 122 .IR "sizeof(unsigned long)" ,
 123 but the kernel will use bits only up to
 124 .IR maxnode .
 125 A NULL value of
 126 .I nodemask
 127 or a
 128 .I maxnode
 129 value of zero specifies the empty set of nodes.
 130 If the value of
 131 .I maxnode
 132 is zero,
 133 the
 134 .I nodemask
 135 argument is ignored.
 136 .P
 137 Where a
 138 .I nodemask
 139 is required, it must contain at least one node that is on-line,
 140 allowed by the process's current cpuset context,
 141 (unless the
 142 .B MPOL_F_STATIC_NODES
 143 mode flag is specified),
 144 and contains memory.
 145 If the
 146 .B MPOL_F_STATIC_NODES
 147 is set in
 148 .I mode
 149 and a required
 150 .I nodemask
 151 contains no nodes that are allowed by the process's current cpuset context,
 152 the memory policy reverts to
 153 .IR "local allocation" .
 154 This effectively overrides the specified policy until the process's
 155 cpuset context includes one or more of the nodes specified by
 156 .IR nodemask .
 157 .P
 158 The
 159 .I mode
 160 argument must include one of the following values:
 161 .TP
 162 .B MPOL_DEFAULT
 163 This mode specifies that any nondefault thread memory policy be removed,
 164 so that the memory policy "falls back" to the system default policy.
 165 The system default policy is "local allocation"\[em]that is,
 166 allocate memory on the node of the CPU that triggered the allocation.
 167 .I nodemask
 168 must be specified as NULL.
 169 If the "local node" contains no free memory, the system will
 170 attempt to allocate memory from a "near by" node.
 171 .TP
 172 .B MPOL_BIND
 173 This mode defines a strict policy that restricts memory allocation to the
 174 nodes specified in
 175 .IR nodemask .
 176 If
 177 .I nodemask
 178 specifies more than one node, page allocations will come from
 179 the node with the lowest numeric node ID first, until that node
 180 contains no free memory.
 181 Allocations will then come from the node with the next highest
 182 node ID specified in
 183 .I nodemask
 184 and so forth, until none of the specified nodes contain free memory.
 185 Pages will not be allocated from any node not specified in the
 186 .IR nodemask .
 187 .TP
 188 .B MPOL_INTERLEAVE
 189 This mode interleaves page allocations across the nodes specified in
 190 .I nodemask
 191 in numeric node ID order.
 192 This optimizes for bandwidth instead of latency
 193 by spreading out pages and memory accesses to those pages across
 194 multiple nodes.
 195 However, accesses to a single page will still be limited to
 196 the memory bandwidth of a single node.
 197 .\" NOTE:  the following sentence doesn't make sense in the context
 198 .\" of set_mempolicy() -- no memory area specified.
 199 .\" To be effective the memory area should be fairly large,
 200 .\" at least 1 MB or bigger.
 201 .TP
 202 .B MPOL_PREFERRED
 203 This mode sets the preferred node for allocation.
 204 The kernel will try to allocate pages from this node first
 205 and fall back to "near by" nodes if the preferred node is low on free
 206 memory.
 207 If
 208 .I nodemask
 209 specifies more than one node ID, the first node in the
 210 mask will be selected as the preferred node.
 211 If the
 212 .I nodemask
 213 and
 214 .I maxnode
 215 arguments specify the empty set, then the policy
 216 specifies "local allocation"
 217 (like the system default policy discussed above).
 218 .TP
 219 .BR MPOL_LOCAL " (since Linux 3.8)"
 220 .\" commit 479e2802d09f1e18a97262c4c6f8f17ae5884bd8
 221 .\" commit f2a07f40dbc603c15f8b06e6ec7f768af67b424f
 222 This mode specifies "local allocation"; the memory is allocated on
 223 the node of the CPU that triggered the allocation (the "local node").
 224 The
 225 .I nodemask
 226 and
 227 .I maxnode
 228 arguments must specify the empty set.
 229 If the "local node" is low on free memory,
 230 the kernel will try to allocate memory from other nodes.
 231 The kernel will allocate memory from the "local node"
 232 whenever memory for this node is available.
 233 If the "local node" is not allowed by the process's current cpuset context,
 234 the kernel will try to allocate memory from other nodes.
 235 The kernel will allocate memory from the "local node" whenever
 236 it becomes allowed by the process's current cpuset context.
 237 .P
 238 The thread memory policy is preserved across an
 239 .BR execve (2),
 240 and is inherited by child threads created using
 241 .BR fork (2)
 242 or
 243 .BR clone (2).
 244 .SH RETURN VALUE
 245 On success,
 246 .BR set_mempolicy ()
 247 returns 0;
 248 on error, \-1 is returned and
 249 .I errno
 250 is set to indicate the error.
 251 .SH ERRORS
 252 .TP
 253 .B EFAULT
 254 Part of all of the memory range specified by
 255 .I nodemask
 256 and
 257 .I maxnode
 258 points outside your accessible address space.
 259 .TP
 260 .B EINVAL
 261 .I mode
 262 is invalid.
 263 Or,
 264 .I mode
 265 is
 266 .B MPOL_DEFAULT
 267 and
 268 .I nodemask
 269 is nonempty,
 270 or
 271 .I mode
 272 is
 273 .B MPOL_BIND
 274 or
 275 .B MPOL_INTERLEAVE
 276 and
 277 .I nodemask
 278 is empty.
 279 Or,
 280 .I maxnode
 281 specifies more than a page worth of bits.
 282 Or,
 283 .I nodemask
 284 specifies one or more node IDs that are
 285 greater than the maximum supported node ID.
 286 Or, none of the node IDs specified by
 287 .I nodemask
 288 are on-line and allowed by the process's current cpuset context,
 289 or none of the specified nodes contain memory.
 290 Or, the
 291 .I mode
 292 argument specified both
 293 .B MPOL_F_STATIC_NODES
 294 and
 295 .BR MPOL_F_RELATIVE_NODES .
 296 Or, the
 297 .B MPOL_F_NUMA_BALANCING
 298 isn't supported by the kernel, or is used with
 299 .I mode
 300 other than
 301 .BR MPOL_BIND .
 302 .TP
 303 .B ENOMEM
 304 Insufficient kernel memory was available.
 305 .SH STANDARDS
 306 Linux.
 307 .SH HISTORY
 308 Linux 2.6.7.
 309 .SH NOTES
 310 Memory policy is not remembered if the page is swapped out.
 311 When such a page is paged back in, it will use the policy of
 312 the thread or memory range that is in effect at the time the
 313 page is allocated.
 314 .P
 315 For information on library support, see
 316 .BR numa (7).
 317 .SH SEE ALSO
 318 .BR get_mempolicy (2),
 319 .BR getcpu (2),
 320 .BR mbind (2),
 321 .BR mmap (2),
 322 .BR numa (3),
 323 .BR cpuset (7),
 324 .BR numa (7),
 325 .BR numactl (8)