File-copy from v4.4.100

This is the result of 'cp' from a linux-stable tree with the 'v4.4.100'
tag checked out (commit 26d6298789e695c9f627ce49a7bbd2286405798a) on
git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git

Please refer to that tree for all history prior to this point.

Change-Id: I8a9ee2aea93cd29c52c847d0ce33091a73ae6afe
diff --git a/Documentation/vm/numa_memory_policy.txt b/Documentation/vm/numa_memory_policy.txt
new file mode 100644
index 0000000..badb050
--- /dev/null
+++ b/Documentation/vm/numa_memory_policy.txt
@@ -0,0 +1,452 @@
+
+What is Linux Memory Policy?
+
+In the Linux kernel, "memory policy" determines from which node the kernel will
+allocate memory in a NUMA system or in an emulated NUMA system.  Linux has
+supported platforms with Non-Uniform Memory Access architectures since 2.4.?.
+The current memory policy support was added to Linux 2.6 around May 2004.  This
+document attempts to describe the concepts and APIs of the 2.6 memory policy
+support.
+
+Memory policies should not be confused with cpusets
+(Documentation/cgroups/cpusets.txt)
+which is an administrative mechanism for restricting the nodes from which
+memory may be allocated by a set of processes. Memory policies are a
+programming interface that a NUMA-aware application can take advantage of.  When
+both cpusets and policies are applied to a task, the restrictions of the cpuset
+takes priority.  See "MEMORY POLICIES AND CPUSETS" below for more details.
+
+MEMORY POLICY CONCEPTS
+
+Scope of Memory Policies
+
+The Linux kernel supports _scopes_ of memory policy, described here from
+most general to most specific:
+
+    System Default Policy:  this policy is "hard coded" into the kernel.  It
+    is the policy that governs all page allocations that aren't controlled
+    by one of the more specific policy scopes discussed below.  When the
+    system is "up and running", the system default policy will use "local
+    allocation" described below.  However, during boot up, the system
+    default policy will be set to interleave allocations across all nodes
+    with "sufficient" memory, so as not to overload the initial boot node
+    with boot-time allocations.
+
+    Task/Process Policy:  this is an optional, per-task policy.  When defined
+    for a specific task, this policy controls all page allocations made by or
+    on behalf of the task that aren't controlled by a more specific scope.
+    If a task does not define a task policy, then all page allocations that
+    would have been controlled by the task policy "fall back" to the System
+    Default Policy.
+
+	The task policy applies to the entire address space of a task. Thus,
+	it is inheritable, and indeed is inherited, across both fork()
+	[clone() w/o the CLONE_VM flag] and exec*().  This allows a parent task
+	to establish the task policy for a child task exec()'d from an
+	executable image that has no awareness of memory policy.  See the
+	MEMORY POLICY APIS section, below, for an overview of the system call
+	that a task may use to set/change its task/process policy.
+
+	In a multi-threaded task, task policies apply only to the thread
+	[Linux kernel task] that installs the policy and any threads
+	subsequently created by that thread.  Any sibling threads existing
+	at the time a new task policy is installed retain their current
+	policy.
+
+	A task policy applies only to pages allocated after the policy is
+	installed.  Any pages already faulted in by the task when the task
+	changes its task policy remain where they were allocated based on
+	the policy at the time they were allocated.
+
+    VMA Policy:  A "VMA" or "Virtual Memory Area" refers to a range of a task's
+    virtual address space.  A task may define a specific policy for a range
+    of its virtual address space.   See the MEMORY POLICIES APIS section,
+    below, for an overview of the mbind() system call used to set a VMA
+    policy.
+
+    A VMA policy will govern the allocation of pages that back this region of
+    the address space.  Any regions of the task's address space that don't
+    have an explicit VMA policy will fall back to the task policy, which may
+    itself fall back to the System Default Policy.
+
+    VMA policies have a few complicating details:
+
+	VMA policy applies ONLY to anonymous pages.  These include pages
+	allocated for anonymous segments, such as the task stack and heap, and
+	any regions of the address space mmap()ed with the MAP_ANONYMOUS flag.
+	If a VMA policy is applied to a file mapping, it will be ignored if
+	the mapping used the MAP_SHARED flag.  If the file mapping used the
+	MAP_PRIVATE flag, the VMA policy will only be applied when an
+	anonymous page is allocated on an attempt to write to the mapping--
+	i.e., at Copy-On-Write.
+
+	VMA policies are shared between all tasks that share a virtual address
+	space--a.k.a. threads--independent of when the policy is installed; and
+	they are inherited across fork().  However, because VMA policies refer
+	to a specific region of a task's address space, and because the address
+	space is discarded and recreated on exec*(), VMA policies are NOT
+	inheritable across exec().  Thus, only NUMA-aware applications may
+	use VMA policies.
+
+	A task may install a new VMA policy on a sub-range of a previously
+	mmap()ed region.  When this happens, Linux splits the existing virtual
+	memory area into 2 or 3 VMAs, each with it's own policy.
+
+	By default, VMA policy applies only to pages allocated after the policy
+	is installed.  Any pages already faulted into the VMA range remain
+	where they were allocated based on the policy at the time they were
+	allocated.  However, since 2.6.16, Linux supports page migration via
+	the mbind() system call, so that page contents can be moved to match
+	a newly installed policy.
+
+    Shared Policy:  Conceptually, shared policies apply to "memory objects"
+    mapped shared into one or more tasks' distinct address spaces.  An
+    application installs a shared policies the same way as VMA policies--using
+    the mbind() system call specifying a range of virtual addresses that map
+    the shared object.  However, unlike VMA policies, which can be considered
+    to be an attribute of a range of a task's address space, shared policies
+    apply directly to the shared object.  Thus, all tasks that attach to the
+    object share the policy, and all pages allocated for the shared object,
+    by any task, will obey the shared policy.
+
+	As of 2.6.22, only shared memory segments, created by shmget() or
+	mmap(MAP_ANONYMOUS|MAP_SHARED), support shared policy.  When shared
+	policy support was added to Linux, the associated data structures were
+	added to hugetlbfs shmem segments.  At the time, hugetlbfs did not
+	support allocation at fault time--a.k.a lazy allocation--so hugetlbfs
+	shmem segments were never "hooked up" to the shared policy support.
+	Although hugetlbfs segments now support lazy allocation, their support
+	for shared policy has not been completed.
+
+	As mentioned above [re: VMA policies], allocations of page cache
+	pages for regular files mmap()ed with MAP_SHARED ignore any VMA
+	policy installed on the virtual address range backed by the shared
+	file mapping.  Rather, shared page cache pages, including pages backing
+	private mappings that have not yet been written by the task, follow
+	task policy, if any, else System Default Policy.
+
+	The shared policy infrastructure supports different policies on subset
+	ranges of the shared object.  However, Linux still splits the VMA of
+	the task that installs the policy for each range of distinct policy.
+	Thus, different tasks that attach to a shared memory segment can have
+	different VMA configurations mapping that one shared object.  This
+	can be seen by examining the /proc/<pid>/numa_maps of tasks sharing
+	a shared memory region, when one task has installed shared policy on
+	one or more ranges of the region.
+
+Components of Memory Policies
+
+    A Linux memory policy consists of a "mode", optional mode flags, and an
+    optional set of nodes.  The mode determines the behavior of the policy,
+    the optional mode flags determine the behavior of the mode, and the
+    optional set of nodes can be viewed as the arguments to the policy
+    behavior.
+
+   Internally, memory policies are implemented by a reference counted
+   structure, struct mempolicy.  Details of this structure will be discussed
+   in context, below, as required to explain the behavior.
+
+   Linux memory policy supports the following 4 behavioral modes:
+
+	Default Mode--MPOL_DEFAULT:  This mode is only used in the memory
+	policy APIs.  Internally, MPOL_DEFAULT is converted to the NULL
+	memory policy in all policy scopes.  Any existing non-default policy
+	will simply be removed when MPOL_DEFAULT is specified.  As a result,
+	MPOL_DEFAULT means "fall back to the next most specific policy scope."
+
+	    For example, a NULL or default task policy will fall back to the
+	    system default policy.  A NULL or default vma policy will fall
+	    back to the task policy.
+
+	    When specified in one of the memory policy APIs, the Default mode
+	    does not use the optional set of nodes.
+
+	    It is an error for the set of nodes specified for this policy to
+	    be non-empty.
+
+	MPOL_BIND:  This mode specifies that memory must come from the
+	set of nodes specified by the policy.  Memory will be allocated from
+	the node in the set with sufficient free memory that is closest to
+	the node where the allocation takes place.
+
+	MPOL_PREFERRED:  This mode specifies that the allocation should be
+	attempted from the single node specified in the policy.  If that
+	allocation fails, the kernel will search other nodes, in order of
+	increasing distance from the preferred node based on information
+	provided by the platform firmware.
+
+	    Internally, the Preferred policy uses a single node--the
+	    preferred_node member of struct mempolicy.  When the internal
+	    mode flag MPOL_F_LOCAL is set, the preferred_node is ignored and
+	    the policy is interpreted as local allocation.  "Local" allocation
+	    policy can be viewed as a Preferred policy that starts at the node
+	    containing the cpu where the allocation takes place.
+
+	    It is possible for the user to specify that local allocation is
+	    always preferred by passing an empty nodemask with this mode.
+	    If an empty nodemask is passed, the policy cannot use the
+	    MPOL_F_STATIC_NODES or MPOL_F_RELATIVE_NODES flags described
+	    below.
+
+	MPOL_INTERLEAVED:  This mode specifies that page allocations be
+	interleaved, on a page granularity, across the nodes specified in
+	the policy.  This mode also behaves slightly differently, based on
+	the context where it is used:
+
+	    For allocation of anonymous pages and shared memory pages,
+	    Interleave mode indexes the set of nodes specified by the policy
+	    using the page offset of the faulting address into the segment
+	    [VMA] containing the address modulo the number of nodes specified
+	    by the policy.  It then attempts to allocate a page, starting at
+	    the selected node, as if the node had been specified by a Preferred
+	    policy or had been selected by a local allocation.  That is,
+	    allocation will follow the per node zonelist.
+
+	    For allocation of page cache pages, Interleave mode indexes the set
+	    of nodes specified by the policy using a node counter maintained
+	    per task.  This counter wraps around to the lowest specified node
+	    after it reaches the highest specified node.  This will tend to
+	    spread the pages out over the nodes specified by the policy based
+	    on the order in which they are allocated, rather than based on any
+	    page offset into an address range or file.  During system boot up,
+	    the temporary interleaved system default policy works in this
+	    mode.
+
+   Linux memory policy supports the following optional mode flags:
+
+	MPOL_F_STATIC_NODES:  This flag specifies that the nodemask passed by
+	the user should not be remapped if the task or VMA's set of allowed
+	nodes changes after the memory policy has been defined.
+
+	    Without this flag, anytime a mempolicy is rebound because of a
+	    change in the set of allowed nodes, the node (Preferred) or
+	    nodemask (Bind, Interleave) is remapped to the new set of
+	    allowed nodes.  This may result in nodes being used that were
+	    previously undesired.
+
+	    With this flag, if the user-specified nodes overlap with the
+	    nodes allowed by the task's cpuset, then the memory policy is
+	    applied to their intersection.  If the two sets of nodes do not
+	    overlap, the Default policy is used.
+
+	    For example, consider a task that is attached to a cpuset with
+	    mems 1-3 that sets an Interleave policy over the same set.  If
+	    the cpuset's mems change to 3-5, the Interleave will now occur
+	    over nodes 3, 4, and 5.  With this flag, however, since only node
+	    3 is allowed from the user's nodemask, the "interleave" only
+	    occurs over that node.  If no nodes from the user's nodemask are
+	    now allowed, the Default behavior is used.
+
+	    MPOL_F_STATIC_NODES cannot be combined with the
+	    MPOL_F_RELATIVE_NODES flag.  It also cannot be used for
+	    MPOL_PREFERRED policies that were created with an empty nodemask
+	    (local allocation).
+
+	MPOL_F_RELATIVE_NODES:  This flag specifies that the nodemask passed
+	by the user will be mapped relative to the set of the task or VMA's
+	set of allowed nodes.  The kernel stores the user-passed nodemask,
+	and if the allowed nodes changes, then that original nodemask will
+	be remapped relative to the new set of allowed nodes.
+
+	    Without this flag (and without MPOL_F_STATIC_NODES), anytime a
+	    mempolicy is rebound because of a change in the set of allowed
+	    nodes, the node (Preferred) or nodemask (Bind, Interleave) is
+	    remapped to the new set of allowed nodes.  That remap may not
+	    preserve the relative nature of the user's passed nodemask to its
+	    set of allowed nodes upon successive rebinds: a nodemask of
+	    1,3,5 may be remapped to 7-9 and then to 1-3 if the set of
+	    allowed nodes is restored to its original state.
+
+	    With this flag, the remap is done so that the node numbers from
+	    the user's passed nodemask are relative to the set of allowed
+	    nodes.  In other words, if nodes 0, 2, and 4 are set in the user's
+	    nodemask, the policy will be effected over the first (and in the
+	    Bind or Interleave case, the third and fifth) nodes in the set of
+	    allowed nodes.  The nodemask passed by the user represents nodes
+	    relative to task or VMA's set of allowed nodes.
+
+	    If the user's nodemask includes nodes that are outside the range
+	    of the new set of allowed nodes (for example, node 5 is set in
+	    the user's nodemask when the set of allowed nodes is only 0-3),
+	    then the remap wraps around to the beginning of the nodemask and,
+	    if not already set, sets the node in the mempolicy nodemask.
+
+	    For example, consider a task that is attached to a cpuset with
+	    mems 2-5 that sets an Interleave policy over the same set with
+	    MPOL_F_RELATIVE_NODES.  If the cpuset's mems change to 3-7, the
+	    interleave now occurs over nodes 3,5-7.  If the cpuset's mems
+	    then change to 0,2-3,5, then the interleave occurs over nodes
+	    0,2-3,5.
+
+	    Thanks to the consistent remapping, applications preparing
+	    nodemasks to specify memory policies using this flag should
+	    disregard their current, actual cpuset imposed memory placement
+	    and prepare the nodemask as if they were always located on
+	    memory nodes 0 to N-1, where N is the number of memory nodes the
+	    policy is intended to manage.  Let the kernel then remap to the
+	    set of memory nodes allowed by the task's cpuset, as that may
+	    change over time.
+
+	    MPOL_F_RELATIVE_NODES cannot be combined with the
+	    MPOL_F_STATIC_NODES flag.  It also cannot be used for
+	    MPOL_PREFERRED policies that were created with an empty nodemask
+	    (local allocation).
+
+MEMORY POLICY REFERENCE COUNTING
+
+To resolve use/free races, struct mempolicy contains an atomic reference
+count field.  Internal interfaces, mpol_get()/mpol_put() increment and
+decrement this reference count, respectively.  mpol_put() will only free
+the structure back to the mempolicy kmem cache when the reference count
+goes to zero.
+
+When a new memory policy is allocated, its reference count is initialized
+to '1', representing the reference held by the task that is installing the
+new policy.  When a pointer to a memory policy structure is stored in another
+structure, another reference is added, as the task's reference will be dropped
+on completion of the policy installation.
+
+During run-time "usage" of the policy, we attempt to minimize atomic operations
+on the reference count, as this can lead to cache lines bouncing between cpus
+and NUMA nodes.  "Usage" here means one of the following:
+
+1) querying of the policy, either by the task itself [using the get_mempolicy()
+   API discussed below] or by another task using the /proc/<pid>/numa_maps
+   interface.
+
+2) examination of the policy to determine the policy mode and associated node
+   or node lists, if any, for page allocation.  This is considered a "hot
+   path".  Note that for MPOL_BIND, the "usage" extends across the entire
+   allocation process, which may sleep during page reclaimation, because the
+   BIND policy nodemask is used, by reference, to filter ineligible nodes.
+
+We can avoid taking an extra reference during the usages listed above as
+follows:
+
+1) we never need to get/free the system default policy as this is never
+   changed nor freed, once the system is up and running.
+
+2) for querying the policy, we do not need to take an extra reference on the
+   target task's task policy nor vma policies because we always acquire the
+   task's mm's mmap_sem for read during the query.  The set_mempolicy() and
+   mbind() APIs [see below] always acquire the mmap_sem for write when
+   installing or replacing task or vma policies.  Thus, there is no possibility
+   of a task or thread freeing a policy while another task or thread is
+   querying it.
+
+3) Page allocation usage of task or vma policy occurs in the fault path where
+   we hold them mmap_sem for read.  Again, because replacing the task or vma
+   policy requires that the mmap_sem be held for write, the policy can't be
+   freed out from under us while we're using it for page allocation.
+
+4) Shared policies require special consideration.  One task can replace a
+   shared memory policy while another task, with a distinct mmap_sem, is
+   querying or allocating a page based on the policy.  To resolve this
+   potential race, the shared policy infrastructure adds an extra reference
+   to the shared policy during lookup while holding a spin lock on the shared
+   policy management structure.  This requires that we drop this extra
+   reference when we're finished "using" the policy.  We must drop the
+   extra reference on shared policies in the same query/allocation paths
+   used for non-shared policies.  For this reason, shared policies are marked
+   as such, and the extra reference is dropped "conditionally"--i.e., only
+   for shared policies.
+
+   Because of this extra reference counting, and because we must lookup
+   shared policies in a tree structure under spinlock, shared policies are
+   more expensive to use in the page allocation path.  This is especially
+   true for shared policies on shared memory regions shared by tasks running
+   on different NUMA nodes.  This extra overhead can be avoided by always
+   falling back to task or system default policy for shared memory regions,
+   or by prefaulting the entire shared memory region into memory and locking
+   it down.  However, this might not be appropriate for all applications.
+
+MEMORY POLICY APIs
+
+Linux supports 3 system calls for controlling memory policy.  These APIS
+always affect only the calling task, the calling task's address space, or
+some shared object mapped into the calling task's address space.
+
+	Note:  the headers that define these APIs and the parameter data types
+	for user space applications reside in a package that is not part of
+	the Linux kernel.  The kernel system call interfaces, with the 'sys_'
+	prefix, are defined in <linux/syscalls.h>; the mode and flag
+	definitions are defined in <linux/mempolicy.h>.
+
+Set [Task] Memory Policy:
+
+	long set_mempolicy(int mode, const unsigned long *nmask,
+					unsigned long maxnode);
+
+	Set's the calling task's "task/process memory policy" to mode
+	specified by the 'mode' argument and the set of nodes defined
+	by 'nmask'.  'nmask' points to a bit mask of node ids containing
+	at least 'maxnode' ids.  Optional mode flags may be passed by
+	combining the 'mode' argument with the flag (for example:
+	MPOL_INTERLEAVE | MPOL_F_STATIC_NODES).
+
+	See the set_mempolicy(2) man page for more details
+
+
+Get [Task] Memory Policy or Related Information
+
+	long get_mempolicy(int *mode,
+			   const unsigned long *nmask, unsigned long maxnode,
+			   void *addr, int flags);
+
+	Queries the "task/process memory policy" of the calling task, or
+	the policy or location of a specified virtual address, depending
+	on the 'flags' argument.
+
+	See the get_mempolicy(2) man page for more details
+
+
+Install VMA/Shared Policy for a Range of Task's Address Space
+
+	long mbind(void *start, unsigned long len, int mode,
+		   const unsigned long *nmask, unsigned long maxnode,
+		   unsigned flags);
+
+	mbind() installs the policy specified by (mode, nmask, maxnodes) as
+	a VMA policy for the range of the calling task's address space
+	specified by the 'start' and 'len' arguments.  Additional actions
+	may be requested via the 'flags' argument.
+
+	See the mbind(2) man page for more details.
+
+MEMORY POLICY COMMAND LINE INTERFACE
+
+Although not strictly part of the Linux implementation of memory policy,
+a command line tool, numactl(8), exists that allows one to:
+
++ set the task policy for a specified program via set_mempolicy(2), fork(2) and
+  exec(2)
+
++ set the shared policy for a shared memory segment via mbind(2)
+
+The numactl(8) tool is packaged with the run-time version of the library
+containing the memory policy system call wrappers.  Some distributions
+package the headers and compile-time libraries in a separate development
+package.
+
+
+MEMORY POLICIES AND CPUSETS
+
+Memory policies work within cpusets as described above.  For memory policies
+that require a node or set of nodes, the nodes are restricted to the set of
+nodes whose memories are allowed by the cpuset constraints.  If the nodemask
+specified for the policy contains nodes that are not allowed by the cpuset and
+MPOL_F_RELATIVE_NODES is not used, the intersection of the set of nodes
+specified for the policy and the set of nodes with memory is used.  If the
+result is the empty set, the policy is considered invalid and cannot be
+installed.  If MPOL_F_RELATIVE_NODES is used, the policy's nodes are mapped
+onto and folded into the task's set of allowed nodes as previously described.
+
+The interaction of memory policies and cpusets can be problematic when tasks
+in two cpusets share access to a memory region, such as shared memory segments
+created by shmget() of mmap() with the MAP_ANONYMOUS and MAP_SHARED flags, and
+any of the tasks install shared policy on the region, only nodes whose
+memories are allowed in both cpusets may be used in the policies.  Obtaining
+this information requires "stepping outside" the memory policy APIs to use the
+cpuset information and requires that one know in what cpusets other task might
+be attaching to the shared region.  Furthermore, if the cpusets' allowed
+memory sets are disjoint, "local" allocation is the only valid policy.