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.