5 What is NUMA Memory Policy?
6 ============================
8 In the Linux kernel, "memory policy" determines from which node the kernel will
9 allocate memory in a NUMA system or in an emulated NUMA system. Linux has
10 supported platforms with Non-Uniform Memory Access architectures since 2.4.?.
11 The current memory policy support was added to Linux 2.6 around May 2004. This
12 document attempts to describe the concepts and APIs of the 2.6 memory policy
15 Memory policies should not be confused with cpusets
16 (``Documentation/admin-guide/cgroup-v1/cpusets.rst``)
17 which is an administrative mechanism for restricting the nodes from which
18 memory may be allocated by a set of processes. Memory policies are a
19 programming interface that a NUMA-aware application can take advantage of. When
20 both cpusets and policies are applied to a task, the restrictions of the cpuset
21 takes priority. See :ref:`Memory Policies and cpusets <mem_pol_and_cpusets>`
22 below for more details.
24 Memory Policy Concepts
25 ======================
27 Scope of Memory Policies
28 ------------------------
30 The Linux kernel supports _scopes_ of memory policy, described here from
31 most general to most specific:
34 this policy is "hard coded" into the kernel. It is the policy
35 that governs all page allocations that aren't controlled by
36 one of the more specific policy scopes discussed below. When
37 the system is "up and running", the system default policy will
38 use "local allocation" described below. However, during boot
39 up, the system default policy will be set to interleave
40 allocations across all nodes with "sufficient" memory, so as
41 not to overload the initial boot node with boot-time
45 this is an optional, per-task policy. When defined for a
46 specific task, this policy controls all page allocations made
47 by or on behalf of the task that aren't controlled by a more
48 specific scope. If a task does not define a task policy, then
49 all page allocations that would have been controlled by the
50 task policy "fall back" to the System Default Policy.
52 The task policy applies to the entire address space of a task. Thus,
53 it is inheritable, and indeed is inherited, across both fork()
54 [clone() w/o the CLONE_VM flag] and exec*(). This allows a parent task
55 to establish the task policy for a child task exec()'d from an
56 executable image that has no awareness of memory policy. See the
57 :ref:`Memory Policy APIs <memory_policy_apis>` section,
58 below, for an overview of the system call
59 that a task may use to set/change its task/process policy.
61 In a multi-threaded task, task policies apply only to the thread
62 [Linux kernel task] that installs the policy and any threads
63 subsequently created by that thread. Any sibling threads existing
64 at the time a new task policy is installed retain their current
67 A task policy applies only to pages allocated after the policy is
68 installed. Any pages already faulted in by the task when the task
69 changes its task policy remain where they were allocated based on
70 the policy at the time they were allocated.
75 A "VMA" or "Virtual Memory Area" refers to a range of a task's
76 virtual address space. A task may define a specific policy for a range
77 of its virtual address space. See the
78 :ref:`Memory Policy APIs <memory_policy_apis>` section,
79 below, for an overview of the mbind() system call used to set a VMA
82 A VMA policy will govern the allocation of pages that back
83 this region of the address space. Any regions of the task's
84 address space that don't have an explicit VMA policy will fall
85 back to the task policy, which may itself fall back to the
86 System Default Policy.
88 VMA policies have a few complicating details:
90 * VMA policy applies ONLY to anonymous pages. These include
91 pages allocated for anonymous segments, such as the task
92 stack and heap, and any regions of the address space
93 mmap()ed with the MAP_ANONYMOUS flag. If a VMA policy is
94 applied to a file mapping, it will be ignored if the mapping
95 used the MAP_SHARED flag. If the file mapping used the
96 MAP_PRIVATE flag, the VMA policy will only be applied when
97 an anonymous page is allocated on an attempt to write to the
98 mapping-- i.e., at Copy-On-Write.
100 * VMA policies are shared between all tasks that share a
101 virtual address space--a.k.a. threads--independent of when
102 the policy is installed; and they are inherited across
103 fork(). However, because VMA policies refer to a specific
104 region of a task's address space, and because the address
105 space is discarded and recreated on exec*(), VMA policies
106 are NOT inheritable across exec(). Thus, only NUMA-aware
107 applications may use VMA policies.
109 * A task may install a new VMA policy on a sub-range of a
110 previously mmap()ed region. When this happens, Linux splits
111 the existing virtual memory area into 2 or 3 VMAs, each with
114 * By default, VMA policy applies only to pages allocated after
115 the policy is installed. Any pages already faulted into the
116 VMA range remain where they were allocated based on the
117 policy at the time they were allocated. However, since
118 2.6.16, Linux supports page migration via the mbind() system
119 call, so that page contents can be moved to match a newly
123 Conceptually, shared policies apply to "memory objects" mapped
124 shared into one or more tasks' distinct address spaces. An
125 application installs shared policies the same way as VMA
126 policies--using the mbind() system call specifying a range of
127 virtual addresses that map the shared object. However, unlike
128 VMA policies, which can be considered to be an attribute of a
129 range of a task's address space, shared policies apply
130 directly to the shared object. Thus, all tasks that attach to
131 the object share the policy, and all pages allocated for the
132 shared object, by any task, will obey the shared policy.
134 As of 2.6.22, only shared memory segments, created by shmget() or
135 mmap(MAP_ANONYMOUS|MAP_SHARED), support shared policy. When shared
136 policy support was added to Linux, the associated data structures were
137 added to hugetlbfs shmem segments. At the time, hugetlbfs did not
138 support allocation at fault time--a.k.a lazy allocation--so hugetlbfs
139 shmem segments were never "hooked up" to the shared policy support.
140 Although hugetlbfs segments now support lazy allocation, their support
141 for shared policy has not been completed.
143 As mentioned above in :ref:`VMA policies <vma_policy>` section,
144 allocations of page cache pages for regular files mmap()ed
145 with MAP_SHARED ignore any VMA policy installed on the virtual
146 address range backed by the shared file mapping. Rather,
147 shared page cache pages, including pages backing private
148 mappings that have not yet been written by the task, follow
149 task policy, if any, else System Default Policy.
151 The shared policy infrastructure supports different policies on subset
152 ranges of the shared object. However, Linux still splits the VMA of
153 the task that installs the policy for each range of distinct policy.
154 Thus, different tasks that attach to a shared memory segment can have
155 different VMA configurations mapping that one shared object. This
156 can be seen by examining the /proc/<pid>/numa_maps of tasks sharing
157 a shared memory region, when one task has installed shared policy on
158 one or more ranges of the region.
160 Components of Memory Policies
161 -----------------------------
163 A NUMA memory policy consists of a "mode", optional mode flags, and
164 an optional set of nodes. The mode determines the behavior of the
165 policy, the optional mode flags determine the behavior of the mode,
166 and the optional set of nodes can be viewed as the arguments to the
169 Internally, memory policies are implemented by a reference counted
170 structure, struct mempolicy. Details of this structure will be
171 discussed in context, below, as required to explain the behavior.
173 NUMA memory policy supports the following 4 behavioral modes:
175 Default Mode--MPOL_DEFAULT
176 This mode is only used in the memory policy APIs. Internally,
177 MPOL_DEFAULT is converted to the NULL memory policy in all
178 policy scopes. Any existing non-default policy will simply be
179 removed when MPOL_DEFAULT is specified. As a result,
180 MPOL_DEFAULT means "fall back to the next most specific policy
183 For example, a NULL or default task policy will fall back to the
184 system default policy. A NULL or default vma policy will fall
185 back to the task policy.
187 When specified in one of the memory policy APIs, the Default mode
188 does not use the optional set of nodes.
190 It is an error for the set of nodes specified for this policy to
194 This mode specifies that memory must come from the set of
195 nodes specified by the policy. Memory will be allocated from
196 the node in the set with sufficient free memory that is
197 closest to the node where the allocation takes place.
200 This mode specifies that the allocation should be attempted
201 from the single node specified in the policy. If that
202 allocation fails, the kernel will search other nodes, in order
203 of increasing distance from the preferred node based on
204 information provided by the platform firmware.
206 Internally, the Preferred policy uses a single node--the
207 preferred_node member of struct mempolicy. When the internal
208 mode flag MPOL_F_LOCAL is set, the preferred_node is ignored
209 and the policy is interpreted as local allocation. "Local"
210 allocation policy can be viewed as a Preferred policy that
211 starts at the node containing the cpu where the allocation
214 It is possible for the user to specify that local allocation
215 is always preferred by passing an empty nodemask with this
216 mode. If an empty nodemask is passed, the policy cannot use
217 the MPOL_F_STATIC_NODES or MPOL_F_RELATIVE_NODES flags
221 This mode specifies that page allocations be interleaved, on a
222 page granularity, across the nodes specified in the policy.
223 This mode also behaves slightly differently, based on the
224 context where it is used:
226 For allocation of anonymous pages and shared memory pages,
227 Interleave mode indexes the set of nodes specified by the
228 policy using the page offset of the faulting address into the
229 segment [VMA] containing the address modulo the number of
230 nodes specified by the policy. It then attempts to allocate a
231 page, starting at the selected node, as if the node had been
232 specified by a Preferred policy or had been selected by a
233 local allocation. That is, allocation will follow the per
236 For allocation of page cache pages, Interleave mode indexes
237 the set of nodes specified by the policy using a node counter
238 maintained per task. This counter wraps around to the lowest
239 specified node after it reaches the highest specified node.
240 This will tend to spread the pages out over the nodes
241 specified by the policy based on the order in which they are
242 allocated, rather than based on any page offset into an
243 address range or file. During system boot up, the temporary
244 interleaved system default policy works in this mode.
247 This mode specifies that the allocation should be preferably
248 satisfied from the nodemask specified in the policy. If there is
249 a memory pressure on all nodes in the nodemask, the allocation
250 can fall back to all existing numa nodes. This is effectively
251 MPOL_PREFERRED allowed for a mask rather than a single node.
253 MPOL_WEIGHTED_INTERLEAVE
254 This mode operates the same as MPOL_INTERLEAVE, except that
255 interleaving behavior is executed based on weights set in
256 /sys/kernel/mm/mempolicy/weighted_interleave/
258 Weighted interleave allocates pages on nodes according to a
259 weight. For example if nodes [0,1] are weighted [5,2], 5 pages
260 will be allocated on node0 for every 2 pages allocated on node1.
262 NUMA memory policy supports the following optional mode flags:
265 This flag specifies that the nodemask passed by
266 the user should not be remapped if the task or VMA's set of allowed
267 nodes changes after the memory policy has been defined.
269 Without this flag, any time a mempolicy is rebound because of a
270 change in the set of allowed nodes, the preferred nodemask (Preferred
271 Many), preferred node (Preferred) or nodemask (Bind, Interleave) is
272 remapped to the new set of allowed nodes. This may result in nodes
273 being used that were previously undesired.
275 With this flag, if the user-specified nodes overlap with the
276 nodes allowed by the task's cpuset, then the memory policy is
277 applied to their intersection. If the two sets of nodes do not
278 overlap, the Default policy is used.
280 For example, consider a task that is attached to a cpuset with
281 mems 1-3 that sets an Interleave policy over the same set. If
282 the cpuset's mems change to 3-5, the Interleave will now occur
283 over nodes 3, 4, and 5. With this flag, however, since only node
284 3 is allowed from the user's nodemask, the "interleave" only
285 occurs over that node. If no nodes from the user's nodemask are
286 now allowed, the Default behavior is used.
288 MPOL_F_STATIC_NODES cannot be combined with the
289 MPOL_F_RELATIVE_NODES flag. It also cannot be used for
290 MPOL_PREFERRED policies that were created with an empty nodemask
293 MPOL_F_RELATIVE_NODES
294 This flag specifies that the nodemask passed
295 by the user will be mapped relative to the set of the task or VMA's
296 set of allowed nodes. The kernel stores the user-passed nodemask,
297 and if the allowed nodes changes, then that original nodemask will
298 be remapped relative to the new set of allowed nodes.
300 Without this flag (and without MPOL_F_STATIC_NODES), anytime a
301 mempolicy is rebound because of a change in the set of allowed
302 nodes, the node (Preferred) or nodemask (Bind, Interleave) is
303 remapped to the new set of allowed nodes. That remap may not
304 preserve the relative nature of the user's passed nodemask to its
305 set of allowed nodes upon successive rebinds: a nodemask of
306 1,3,5 may be remapped to 7-9 and then to 1-3 if the set of
307 allowed nodes is restored to its original state.
309 With this flag, the remap is done so that the node numbers from
310 the user's passed nodemask are relative to the set of allowed
311 nodes. In other words, if nodes 0, 2, and 4 are set in the user's
312 nodemask, the policy will be effected over the first (and in the
313 Bind or Interleave case, the third and fifth) nodes in the set of
314 allowed nodes. The nodemask passed by the user represents nodes
315 relative to task or VMA's set of allowed nodes.
317 If the user's nodemask includes nodes that are outside the range
318 of the new set of allowed nodes (for example, node 5 is set in
319 the user's nodemask when the set of allowed nodes is only 0-3),
320 then the remap wraps around to the beginning of the nodemask and,
321 if not already set, sets the node in the mempolicy nodemask.
323 For example, consider a task that is attached to a cpuset with
324 mems 2-5 that sets an Interleave policy over the same set with
325 MPOL_F_RELATIVE_NODES. If the cpuset's mems change to 3-7, the
326 interleave now occurs over nodes 3,5-7. If the cpuset's mems
327 then change to 0,2-3,5, then the interleave occurs over nodes
330 Thanks to the consistent remapping, applications preparing
331 nodemasks to specify memory policies using this flag should
332 disregard their current, actual cpuset imposed memory placement
333 and prepare the nodemask as if they were always located on
334 memory nodes 0 to N-1, where N is the number of memory nodes the
335 policy is intended to manage. Let the kernel then remap to the
336 set of memory nodes allowed by the task's cpuset, as that may
339 MPOL_F_RELATIVE_NODES cannot be combined with the
340 MPOL_F_STATIC_NODES flag. It also cannot be used for
341 MPOL_PREFERRED policies that were created with an empty nodemask
344 Memory Policy Reference Counting
345 ================================
347 To resolve use/free races, struct mempolicy contains an atomic reference
348 count field. Internal interfaces, mpol_get()/mpol_put() increment and
349 decrement this reference count, respectively. mpol_put() will only free
350 the structure back to the mempolicy kmem cache when the reference count
353 When a new memory policy is allocated, its reference count is initialized
354 to '1', representing the reference held by the task that is installing the
355 new policy. When a pointer to a memory policy structure is stored in another
356 structure, another reference is added, as the task's reference will be dropped
357 on completion of the policy installation.
359 During run-time "usage" of the policy, we attempt to minimize atomic operations
360 on the reference count, as this can lead to cache lines bouncing between cpus
361 and NUMA nodes. "Usage" here means one of the following:
363 1) querying of the policy, either by the task itself [using the get_mempolicy()
364 API discussed below] or by another task using the /proc/<pid>/numa_maps
367 2) examination of the policy to determine the policy mode and associated node
368 or node lists, if any, for page allocation. This is considered a "hot
369 path". Note that for MPOL_BIND, the "usage" extends across the entire
370 allocation process, which may sleep during page reclamation, because the
371 BIND policy nodemask is used, by reference, to filter ineligible nodes.
373 We can avoid taking an extra reference during the usages listed above as
376 1) we never need to get/free the system default policy as this is never
377 changed nor freed, once the system is up and running.
379 2) for querying the policy, we do not need to take an extra reference on the
380 target task's task policy nor vma policies because we always acquire the
381 task's mm's mmap_lock for read during the query. The set_mempolicy() and
382 mbind() APIs [see below] always acquire the mmap_lock for write when
383 installing or replacing task or vma policies. Thus, there is no possibility
384 of a task or thread freeing a policy while another task or thread is
387 3) Page allocation usage of task or vma policy occurs in the fault path where
388 we hold them mmap_lock for read. Again, because replacing the task or vma
389 policy requires that the mmap_lock be held for write, the policy can't be
390 freed out from under us while we're using it for page allocation.
392 4) Shared policies require special consideration. One task can replace a
393 shared memory policy while another task, with a distinct mmap_lock, is
394 querying or allocating a page based on the policy. To resolve this
395 potential race, the shared policy infrastructure adds an extra reference
396 to the shared policy during lookup while holding a spin lock on the shared
397 policy management structure. This requires that we drop this extra
398 reference when we're finished "using" the policy. We must drop the
399 extra reference on shared policies in the same query/allocation paths
400 used for non-shared policies. For this reason, shared policies are marked
401 as such, and the extra reference is dropped "conditionally"--i.e., only
404 Because of this extra reference counting, and because we must lookup
405 shared policies in a tree structure under spinlock, shared policies are
406 more expensive to use in the page allocation path. This is especially
407 true for shared policies on shared memory regions shared by tasks running
408 on different NUMA nodes. This extra overhead can be avoided by always
409 falling back to task or system default policy for shared memory regions,
410 or by prefaulting the entire shared memory region into memory and locking
411 it down. However, this might not be appropriate for all applications.
413 .. _memory_policy_apis:
418 Linux supports 4 system calls for controlling memory policy. These APIS
419 always affect only the calling task, the calling task's address space, or
420 some shared object mapped into the calling task's address space.
423 the headers that define these APIs and the parameter data types for
424 user space applications reside in a package that is not part of the
425 Linux kernel. The kernel system call interfaces, with the 'sys\_'
426 prefix, are defined in <linux/syscalls.h>; the mode and flag
427 definitions are defined in <linux/mempolicy.h>.
429 Set [Task] Memory Policy::
431 long set_mempolicy(int mode, const unsigned long *nmask,
432 unsigned long maxnode);
434 Set's the calling task's "task/process memory policy" to mode
435 specified by the 'mode' argument and the set of nodes defined by
436 'nmask'. 'nmask' points to a bit mask of node ids containing at least
437 'maxnode' ids. Optional mode flags may be passed by combining the
438 'mode' argument with the flag (for example: MPOL_INTERLEAVE |
439 MPOL_F_STATIC_NODES).
441 See the set_mempolicy(2) man page for more details
444 Get [Task] Memory Policy or Related Information::
446 long get_mempolicy(int *mode,
447 const unsigned long *nmask, unsigned long maxnode,
448 void *addr, int flags);
450 Queries the "task/process memory policy" of the calling task, or the
451 policy or location of a specified virtual address, depending on the
454 See the get_mempolicy(2) man page for more details
457 Install VMA/Shared Policy for a Range of Task's Address Space::
459 long mbind(void *start, unsigned long len, int mode,
460 const unsigned long *nmask, unsigned long maxnode,
463 mbind() installs the policy specified by (mode, nmask, maxnodes) as a
464 VMA policy for the range of the calling task's address space specified
465 by the 'start' and 'len' arguments. Additional actions may be
466 requested via the 'flags' argument.
468 See the mbind(2) man page for more details.
470 Set home node for a Range of Task's Address Spacec::
472 long sys_set_mempolicy_home_node(unsigned long start, unsigned long len,
473 unsigned long home_node,
474 unsigned long flags);
476 sys_set_mempolicy_home_node set the home node for a VMA policy present in the
477 task's address range. The system call updates the home node only for the existing
478 mempolicy range. Other address ranges are ignored. A home node is the NUMA node
479 closest to which page allocation will come from. Specifying the home node override
480 the default allocation policy to allocate memory close to the local node for an
484 Memory Policy Command Line Interface
485 ====================================
487 Although not strictly part of the Linux implementation of memory policy,
488 a command line tool, numactl(8), exists that allows one to:
490 + set the task policy for a specified program via set_mempolicy(2), fork(2) and
493 + set the shared policy for a shared memory segment via mbind(2)
495 The numactl(8) tool is packaged with the run-time version of the library
496 containing the memory policy system call wrappers. Some distributions
497 package the headers and compile-time libraries in a separate development
500 .. _mem_pol_and_cpusets:
502 Memory Policies and cpusets
503 ===========================
505 Memory policies work within cpusets as described above. For memory policies
506 that require a node or set of nodes, the nodes are restricted to the set of
507 nodes whose memories are allowed by the cpuset constraints. If the nodemask
508 specified for the policy contains nodes that are not allowed by the cpuset and
509 MPOL_F_RELATIVE_NODES is not used, the intersection of the set of nodes
510 specified for the policy and the set of nodes with memory is used. If the
511 result is the empty set, the policy is considered invalid and cannot be
512 installed. If MPOL_F_RELATIVE_NODES is used, the policy's nodes are mapped
513 onto and folded into the task's set of allowed nodes as previously described.
515 The interaction of memory policies and cpusets can be problematic when tasks
516 in two cpusets share access to a memory region, such as shared memory segments
517 created by shmget() of mmap() with the MAP_ANONYMOUS and MAP_SHARED flags, and
518 any of the tasks install shared policy on the region, only nodes whose
519 memories are allowed in both cpusets may be used in the policies. Obtaining
520 this information requires "stepping outside" the memory policy APIs to use the
521 cpuset information and requires that one know in what cpusets other task might
522 be attaching to the shared region. Furthermore, if the cpusets' allowed
523 memory sets are disjoint, "local" allocation is the only valid policy.