2What is Linux Memory Policy?
4In the Linux kernel, "memory policy" determines from which node the kernel will
5allocate memory in a NUMA system or in an emulated NUMA system. Linux has
6supported platforms with Non-Uniform Memory Access architectures since 2.4.?.
7The current memory policy support was added to Linux 2.6 around May 2004. This
8document attempts to describe the concepts and APIs of the 2.6 memory policy
11Memory policies should not be confused with cpusets
13which is an administrative mechanism for restricting the nodes from which
14memory may be allocated by a set of processes. Memory policies are a
15programming interface that a NUMA-aware application can take advantage of. When
16both cpusets and policies are applied to a task, the restrictions of the cpuset
17takes priority. See "MEMORY POLICIES AND CPUSETS" below for more details.
21Scope of Memory Policies
23The Linux kernel supports _scopes_ of memory policy, described here from
24most general to most specific:
26    System Default Policy: this policy is "hard coded" into the kernel. It
27    is the policy that governs all page allocations that aren't controlled
28    by one of the more specific policy scopes discussed below. When the
29    system is "up and running", the system default policy will use "local
30    allocation" described below. However, during boot up, the system
31    default policy will be set to interleave allocations across all nodes
32    with "sufficient" memory, so as not to overload the initial boot node
33    with boot-time allocations.
35    Task/Process Policy: this is an optional, per-task policy. When defined
36    for a specific task, this policy controls all page allocations made by or
37    on behalf of the task that aren't controlled by a more specific scope.
38    If a task does not define a task policy, then all page allocations that
39    would have been controlled by the task policy "fall back" to the System
40    Default Policy.
42    The task policy applies to the entire address space of a task. Thus,
43    it is inheritable, and indeed is inherited, across both fork()
44    [clone() w/o the CLONE_VM flag] and exec*(). This allows a parent task
45    to establish the task policy for a child task exec()'d from an
46    executable image that has no awareness of memory policy. See the
47    MEMORY POLICY APIS section, below, for an overview of the system call
48    that a task may use to set/change its task/process policy.
50    In a multi-threaded task, task policies apply only to the thread
51    [Linux kernel task] that installs the policy and any threads
52    subsequently created by that thread. Any sibling threads existing
53    at the time a new task policy is installed retain their current
54    policy.
56    A task policy applies only to pages allocated after the policy is
57    installed. Any pages already faulted in by the task when the task
58    changes its task policy remain where they were allocated based on
59    the policy at the time they were allocated.
61    VMA Policy: A "VMA" or "Virtual Memory Area" refers to a range of a task's
62    virtual address space. A task may define a specific policy for a range
63    of its virtual address space. See the MEMORY POLICIES APIS section,
64    below, for an overview of the mbind() system call used to set a VMA
65    policy.
67    A VMA policy will govern the allocation of pages that back this region of
68    the address space. Any regions of the task's address space that don't
69    have an explicit VMA policy will fall back to the task policy, which may
70    itself fall back to the System Default Policy.
72    VMA policies have a few complicating details:
74    VMA policy applies ONLY to anonymous pages. These include pages
75    allocated for anonymous segments, such as the task stack and heap, and
76    any regions of the address space mmap()ed with the MAP_ANONYMOUS flag.
77    If a VMA policy is applied to a file mapping, it will be ignored if
78    the mapping used the MAP_SHARED flag. If the file mapping used the
79    MAP_PRIVATE flag, the VMA policy will only be applied when an
80    anonymous page is allocated on an attempt to write to the mapping--
81    i.e., at Copy-On-Write.
83    VMA policies are shared between all tasks that share a virtual address
84    space--a.k.a. threads--independent of when the policy is installed; and
85    they are inherited across fork(). However, because VMA policies refer
86    to a specific region of a task's address space, and because the address
87    space is discarded and recreated on exec*(), VMA policies are NOT
88    inheritable across exec(). Thus, only NUMA-aware applications may
89    use VMA policies.
91    A task may install a new VMA policy on a sub-range of a previously
92    mmap()ed region. When this happens, Linux splits the existing virtual
93    memory area into 2 or 3 VMAs, each with it's own policy.
95    By default, VMA policy applies only to pages allocated after the policy
96    is installed. Any pages already faulted into the VMA range remain
97    where they were allocated based on the policy at the time they were
98    allocated. However, since 2.6.16, Linux supports page migration via
99    the mbind() system call, so that page contents can be moved to match
100    a newly installed policy.
102    Shared Policy: Conceptually, shared policies apply to "memory objects"
103    mapped shared into one or more tasks' distinct address spaces. An
104    application installs a shared policies the same way as VMA policies--using
105    the mbind() system call specifying a range of virtual addresses that map
106    the shared object. However, unlike VMA policies, which can be considered
107    to be an attribute of a range of a task's address space, shared policies
108    apply directly to the shared object. Thus, all tasks that attach to the
109    object share the policy, and all pages allocated for the shared object,
110    by any task, will obey the shared policy.
112    As of 2.6.22, only shared memory segments, created by shmget() or
113    mmap(MAP_ANONYMOUS|MAP_SHARED), support shared policy. When shared
114    policy support was added to Linux, the associated data structures were
115    added to hugetlbfs shmem segments. At the time, hugetlbfs did not
116    support allocation at fault time--a.k.a lazy allocation--so hugetlbfs
117    shmem segments were never "hooked up" to the shared policy support.
118    Although hugetlbfs segments now support lazy allocation, their support
119    for shared policy has not been completed.
121    As mentioned above [re: VMA policies], allocations of page cache
122    pages for regular files mmap()ed with MAP_SHARED ignore any VMA
123    policy installed on the virtual address range backed by the shared
124    file mapping. Rather, shared page cache pages, including pages backing
125    private mappings that have not yet been written by the task, follow
126    task policy, if any, else System Default Policy.
128    The shared policy infrastructure supports different policies on subset
129    ranges of the shared object. However, Linux still splits the VMA of
130    the task that installs the policy for each range of distinct policy.
131    Thus, different tasks that attach to a shared memory segment can have
132    different VMA configurations mapping that one shared object. This
133    can be seen by examining the /proc/<pid>/numa_maps of tasks sharing
134    a shared memory region, when one task has installed shared policy on
135    one or more ranges of the region.
137Components of Memory Policies
139    A Linux memory policy consists of a "mode", optional mode flags, and an
140    optional set of nodes. The mode determines the behavior of the policy,
141    the optional mode flags determine the behavior of the mode, and the
142    optional set of nodes can be viewed as the arguments to the policy
143    behavior.
145   Internally, memory policies are implemented by a reference counted
146   structure, struct mempolicy. Details of this structure will be discussed
147   in context, below, as required to explain the behavior.
149   Linux memory policy supports the following 4 behavioral modes:
151    Default Mode--MPOL_DEFAULT: This mode is only used in the memory
152    policy APIs. Internally, MPOL_DEFAULT is converted to the NULL
153    memory policy in all policy scopes. Any existing non-default policy
154    will simply be removed when MPOL_DEFAULT is specified. As a result,
155    MPOL_DEFAULT means "fall back to the next most specific policy scope."
157        For example, a NULL or default task policy will fall back to the
158        system default policy. A NULL or default vma policy will fall
159        back to the task policy.
161        When specified in one of the memory policy APIs, the Default mode
162        does not use the optional set of nodes.
164        It is an error for the set of nodes specified for this policy to
165        be non-empty.
167    MPOL_BIND: This mode specifies that memory must come from the
168    set of nodes specified by the policy. Memory will be allocated from
169    the node in the set with sufficient free memory that is closest to
170    the node where the allocation takes place.
172    MPOL_PREFERRED: This mode specifies that the allocation should be
173    attempted from the single node specified in the policy. If that
174    allocation fails, the kernel will search other nodes, in order of
175    increasing distance from the preferred node based on information
176    provided by the platform firmware.
177    containing the cpu where the allocation takes place.
179        Internally, the Preferred policy uses a single node--the
180        preferred_node member of struct mempolicy. When the internal
181        mode flag MPOL_F_LOCAL is set, the preferred_node is ignored and
182        the policy is interpreted as local allocation. "Local" allocation
183        policy can be viewed as a Preferred policy that starts at the node
184        containing the cpu where the allocation takes place.
186        It is possible for the user to specify that local allocation is
187        always preferred by passing an empty nodemask with this mode.
188        If an empty nodemask is passed, the policy cannot use the
189        MPOL_F_STATIC_NODES or MPOL_F_RELATIVE_NODES flags described
190        below.
192    MPOL_INTERLEAVED: This mode specifies that page allocations be
193    interleaved, on a page granularity, across the nodes specified in
194    the policy. This mode also behaves slightly differently, based on
195    the context where it is used:
197        For allocation of anonymous pages and shared memory pages,
198        Interleave mode indexes the set of nodes specified by the policy
199        using the page offset of the faulting address into the segment
200        [VMA] containing the address modulo the number of nodes specified
201        by the policy. It then attempts to allocate a page, starting at
202        the selected node, as if the node had been specified by a Preferred
203        policy or had been selected by a local allocation. That is,
204        allocation will follow the per node zonelist.
206        For allocation of page cache pages, Interleave mode indexes the set
207        of nodes specified by the policy using a node counter maintained
208        per task. This counter wraps around to the lowest specified node
209        after it reaches the highest specified node. This will tend to
210        spread the pages out over the nodes specified by the policy based
211        on the order in which they are allocated, rather than based on any
212        page offset into an address range or file. During system boot up,
213        the temporary interleaved system default policy works in this
214        mode.
216   Linux memory policy supports the following optional mode flags:
218    MPOL_F_STATIC_NODES: This flag specifies that the nodemask passed by
219    the user should not be remapped if the task or VMA's set of allowed
220    nodes changes after the memory policy has been defined.
222        Without this flag, anytime a mempolicy is rebound because of a
223        change in the set of allowed nodes, the node (Preferred) or
224        nodemask (Bind, Interleave) is remapped to the new set of
225        allowed nodes. This may result in nodes being used that were
226        previously undesired.
228        With this flag, if the user-specified nodes overlap with the
229        nodes allowed by the task's cpuset, then the memory policy is
230        applied to their intersection. If the two sets of nodes do not
231        overlap, the Default policy is used.
233        For example, consider a task that is attached to a cpuset with
234        mems 1-3 that sets an Interleave policy over the same set. If
235        the cpuset's mems change to 3-5, the Interleave will now occur
236        over nodes 3, 4, and 5. With this flag, however, since only node
237        3 is allowed from the user's nodemask, the "interleave" only
238        occurs over that node. If no nodes from the user's nodemask are
239        now allowed, the Default behavior is used.
241        MPOL_F_STATIC_NODES cannot be combined with the
242        MPOL_F_RELATIVE_NODES flag. It also cannot be used for
243        MPOL_PREFERRED policies that were created with an empty nodemask
244        (local allocation).
246    MPOL_F_RELATIVE_NODES: This flag specifies that the nodemask passed
247    by the user will be mapped relative to the set of the task or VMA's
248    set of allowed nodes. The kernel stores the user-passed nodemask,
249    and if the allowed nodes changes, then that original nodemask will
250    be remapped relative to the new set of allowed nodes.
252        Without this flag (and without MPOL_F_STATIC_NODES), anytime a
253        mempolicy is rebound because of a change in the set of allowed
254        nodes, the node (Preferred) or nodemask (Bind, Interleave) is
255        remapped to the new set of allowed nodes. That remap may not
256        preserve the relative nature of the user's passed nodemask to its
257        set of allowed nodes upon successive rebinds: a nodemask of
258        1,3,5 may be remapped to 7-9 and then to 1-3 if the set of
259        allowed nodes is restored to its original state.
261        With this flag, the remap is done so that the node numbers from
262        the user's passed nodemask are relative to the set of allowed
263        nodes. In other words, if nodes 0, 2, and 4 are set in the user's
264        nodemask, the policy will be effected over the first (and in the
265        Bind or Interleave case, the third and fifth) nodes in the set of
266        allowed nodes. The nodemask passed by the user represents nodes
267        relative to task or VMA's set of allowed nodes.
269        If the user's nodemask includes nodes that are outside the range
270        of the new set of allowed nodes (for example, node 5 is set in
271        the user's nodemask when the set of allowed nodes is only 0-3),
272        then the remap wraps around to the beginning of the nodemask and,
273        if not already set, sets the node in the mempolicy nodemask.
275        For example, consider a task that is attached to a cpuset with
276        mems 2-5 that sets an Interleave policy over the same set with
277        MPOL_F_RELATIVE_NODES. If the cpuset's mems change to 3-7, the
278        interleave now occurs over nodes 3,5-6. If the cpuset's mems
279        then change to 0,2-3,5, then the interleave occurs over nodes
280        0,3,5.
282        Thanks to the consistent remapping, applications preparing
283        nodemasks to specify memory policies using this flag should
284        disregard their current, actual cpuset imposed memory placement
285        and prepare the nodemask as if they were always located on
286        memory nodes 0 to N-1, where N is the number of memory nodes the
287        policy is intended to manage. Let the kernel then remap to the
288        set of memory nodes allowed by the task's cpuset, as that may
289        change over time.
291        MPOL_F_RELATIVE_NODES cannot be combined with the
292        MPOL_F_STATIC_NODES flag. It also cannot be used for
293        MPOL_PREFERRED policies that were created with an empty nodemask
294        (local allocation).
298To resolve use/free races, struct mempolicy contains an atomic reference
299count field. Internal interfaces, mpol_get()/mpol_put() increment and
300decrement this reference count, respectively. mpol_put() will only free
301the structure back to the mempolicy kmem cache when the reference count
302goes to zero.
304When a new memory policy is allocated, its reference count is initialized
305to '1', representing the reference held by the task that is installing the
306new policy. When a pointer to a memory policy structure is stored in another
307structure, another reference is added, as the task's reference will be dropped
308on completion of the policy installation.
310During run-time "usage" of the policy, we attempt to minimize atomic operations
311on the reference count, as this can lead to cache lines bouncing between cpus
312and NUMA nodes. "Usage" here means one of the following:
3141) querying of the policy, either by the task itself [using the get_mempolicy()
315   API discussed below] or by another task using the /proc/<pid>/numa_maps
316   interface.
3182) examination of the policy to determine the policy mode and associated node
319   or node lists, if any, for page allocation. This is considered a "hot
320   path". Note that for MPOL_BIND, the "usage" extends across the entire
321   allocation process, which may sleep during page reclaimation, because the
322   BIND policy nodemask is used, by reference, to filter ineligible nodes.
324We can avoid taking an extra reference during the usages listed above as
3271) we never need to get/free the system default policy as this is never
328   changed nor freed, once the system is up and running.
3302) for querying the policy, we do not need to take an extra reference on the
331   target task's task policy nor vma policies because we always acquire the
332   task's mm's mmap_sem for read during the query. The set_mempolicy() and
333   mbind() APIs [see below] always acquire the mmap_sem for write when
334   installing or replacing task or vma policies. Thus, there is no possibility
335   of a task or thread freeing a policy while another task or thread is
336   querying it.
3383) Page allocation usage of task or vma policy occurs in the fault path where
339   we hold them mmap_sem for read. Again, because replacing the task or vma
340   policy requires that the mmap_sem be held for write, the policy can't be
341   freed out from under us while we're using it for page allocation.
3434) Shared policies require special consideration. One task can replace a
344   shared memory policy while another task, with a distinct mmap_sem, is
345   querying or allocating a page based on the policy. To resolve this
346   potential race, the shared policy infrastructure adds an extra reference
347   to the shared policy during lookup while holding a spin lock on the shared
348   policy management structure. This requires that we drop this extra
349   reference when we're finished "using" the policy. We must drop the
350   extra reference on shared policies in the same query/allocation paths
351   used for non-shared policies. For this reason, shared policies are marked
352   as such, and the extra reference is dropped "conditionally"--i.e., only
353   for shared policies.
355   Because of this extra reference counting, and because we must lookup
356   shared policies in a tree structure under spinlock, shared policies are
357   more expensive to use in the page allocation path. This is especially
358   true for shared policies on shared memory regions shared by tasks running
359   on different NUMA nodes. This extra overhead can be avoided by always
360   falling back to task or system default policy for shared memory regions,
361   or by prefaulting the entire shared memory region into memory and locking
362   it down. However, this might not be appropriate for all applications.
366Linux supports 3 system calls for controlling memory policy. These APIS
367always affect only the calling task, the calling task's address space, or
368some shared object mapped into the calling task's address space.
370    Note: the headers that define these APIs and the parameter data types
371    for user space applications reside in a package that is not part of
372    the Linux kernel. The kernel system call interfaces, with the 'sys_'
373    prefix, are defined in <linux/syscalls.h>; the mode and flag
374    definitions are defined in <linux/mempolicy.h>.
376Set [Task] Memory Policy:
378    long set_mempolicy(int mode, const unsigned long *nmask,
379                    unsigned long maxnode);
381    Set's the calling task's "task/process memory policy" to mode
382    specified by the 'mode' argument and the set of nodes defined
383    by 'nmask'. 'nmask' points to a bit mask of node ids containing
384    at least 'maxnode' ids. Optional mode flags may be passed by
385    combining the 'mode' argument with the flag (for example:
388    See the set_mempolicy(2) man page for more details
391Get [Task] Memory Policy or Related Information
393    long get_mempolicy(int *mode,
394               const unsigned long *nmask, unsigned long maxnode,
395               void *addr, int flags);
397    Queries the "task/process memory policy" of the calling task, or
398    the policy or location of a specified virtual address, depending
399    on the 'flags' argument.
401    See the get_mempolicy(2) man page for more details
404Install VMA/Shared Policy for a Range of Task's Address Space
406    long mbind(void *start, unsigned long len, int mode,
407           const unsigned long *nmask, unsigned long maxnode,
408           unsigned flags);
410    mbind() installs the policy specified by (mode, nmask, maxnodes) as
411    a VMA policy for the range of the calling task's address space
412    specified by the 'start' and 'len' arguments. Additional actions
413    may be requested via the 'flags' argument.
415    See the mbind(2) man page for more details.
419Although not strictly part of the Linux implementation of memory policy,
420a command line tool, numactl(8), exists that allows one to:
422+ set the task policy for a specified program via set_mempolicy(2), fork(2) and
423  exec(2)
425+ set the shared policy for a shared memory segment via mbind(2)
427The numactl(8) tool is packaged with the run-time version of the library
428containing the memory policy system call wrappers. Some distributions
429package the headers and compile-time libraries in a separate development
435Memory policies work within cpusets as described above. For memory policies
436that require a node or set of nodes, the nodes are restricted to the set of
437nodes whose memories are allowed by the cpuset constraints. If the nodemask
438specified for the policy contains nodes that are not allowed by the cpuset and
439MPOL_F_RELATIVE_NODES is not used, the intersection of the set of nodes
440specified for the policy and the set of nodes with memory is used. If the
441result is the empty set, the policy is considered invalid and cannot be
442installed. If MPOL_F_RELATIVE_NODES is used, the policy's nodes are mapped
443onto and folded into the task's set of allowed nodes as previously described.
445The interaction of memory policies and cpusets can be problematic when tasks
446in two cpusets share access to a memory region, such as shared memory segments
447created by shmget() of mmap() with the MAP_ANONYMOUS and MAP_SHARED flags, and
448any of the tasks install shared policy on the region, only nodes whose
449memories are allowed in both cpusets may be used in the policies. Obtaining
450this information requires "stepping outside" the memory policy APIs to use the
451cpuset information and requires that one know in what cpusets other task might
452be attaching to the shared region. Furthermore, if the cpusets' allowed
453memory sets are disjoint, "local" allocation is the only valid policy.

Archive Download this file