1.. _numa_memory_policy:
2
3==================
4NUMA Memory Policy
5==================
6
7What is NUMA Memory Policy?
8============================
9
10In the Linux kernel, "memory policy" determines from which node the kernel will
11allocate memory in a NUMA system or in an emulated NUMA system.  Linux has
12supported platforms with Non-Uniform Memory Access architectures since 2.4.?.
13The current memory policy support was added to Linux 2.6 around May 2004.  This
14document attempts to describe the concepts and APIs of the 2.6 memory policy
15support.
16
17Memory policies should not be confused with cpusets
18(``Documentation/admin-guide/cgroup-v1/cpusets.rst``)
19which is an administrative mechanism for restricting the nodes from which
20memory may be allocated by a set of processes. Memory policies are a
21programming interface that a NUMA-aware application can take advantage of.  When
22both cpusets and policies are applied to a task, the restrictions of the cpuset
23takes priority.  See :ref:`Memory Policies and cpusets <mem_pol_and_cpusets>`
24below for more details.
25
26Memory Policy Concepts
27======================
28
29Scope of Memory Policies
30------------------------
31
32The Linux kernel supports _scopes_ of memory policy, described here from
33most general to most specific:
34
35System Default Policy
36	this policy is "hard coded" into the kernel.  It is the policy
37	that governs all page allocations that aren't controlled by
38	one of the more specific policy scopes discussed below.  When
39	the system is "up and running", the system default policy will
40	use "local allocation" described below.  However, during boot
41	up, the system default policy will be set to interleave
42	allocations across all nodes with "sufficient" memory, so as
43	not to overload the initial boot node with boot-time
44	allocations.
45
46Task/Process Policy
47	this is an optional, per-task policy.  When defined for a
48	specific task, this policy controls all page allocations made
49	by or on behalf of the task that aren't controlled by a more
50	specific scope. If a task does not define a task policy, then
51	all page allocations that would have been controlled by the
52	task policy "fall back" to the System Default Policy.
53
54	The task policy applies to the entire address space of a task. Thus,
55	it is inheritable, and indeed is inherited, across both fork()
56	[clone() w/o the CLONE_VM flag] and exec*().  This allows a parent task
57	to establish the task policy for a child task exec()'d from an
58	executable image that has no awareness of memory policy.  See the
59	:ref:`Memory Policy APIs <memory_policy_apis>` section,
60	below, for an overview of the system call
61	that a task may use to set/change its task/process policy.
62
63	In a multi-threaded task, task policies apply only to the thread
64	[Linux kernel task] that installs the policy and any threads
65	subsequently created by that thread.  Any sibling threads existing
66	at the time a new task policy is installed retain their current
67	policy.
68
69	A task policy applies only to pages allocated after the policy is
70	installed.  Any pages already faulted in by the task when the task
71	changes its task policy remain where they were allocated based on
72	the policy at the time they were allocated.
73
74.. _vma_policy:
75
76VMA Policy
77	A "VMA" or "Virtual Memory Area" refers to a range of a task's
78	virtual address space.  A task may define a specific policy for a range
79	of its virtual address space.   See the
80	:ref:`Memory Policy APIs <memory_policy_apis>` section,
81	below, for an overview of the mbind() system call used to set a VMA
82	policy.
83
84	A VMA policy will govern the allocation of pages that back
85	this region of the address space.  Any regions of the task's
86	address space that don't have an explicit VMA policy will fall
87	back to the task policy, which may itself fall back to the
88	System Default Policy.
89
90	VMA policies have a few complicating details:
91
92	* VMA policy applies ONLY to anonymous pages.  These include
93	  pages allocated for anonymous segments, such as the task
94	  stack and heap, and any regions of the address space
95	  mmap()ed with the MAP_ANONYMOUS flag.  If a VMA policy is
96	  applied to a file mapping, it will be ignored if the mapping
97	  used the MAP_SHARED flag.  If the file mapping used the
98	  MAP_PRIVATE flag, the VMA policy will only be applied when
99	  an anonymous page is allocated on an attempt to write to the
100	  mapping-- i.e., at Copy-On-Write.
101
102	* VMA policies are shared between all tasks that share a
103	  virtual address space--a.k.a. threads--independent of when
104	  the policy is installed; and they are inherited across
105	  fork().  However, because VMA policies refer to a specific
106	  region of a task's address space, and because the address
107	  space is discarded and recreated on exec*(), VMA policies
108	  are NOT inheritable across exec().  Thus, only NUMA-aware
109	  applications may use VMA policies.
110
111	* A task may install a new VMA policy on a sub-range of a
112	  previously mmap()ed region.  When this happens, Linux splits
113	  the existing virtual memory area into 2 or 3 VMAs, each with
114	  it's own policy.
115
116	* By default, VMA policy applies only to pages allocated after
117	  the policy is installed.  Any pages already faulted into the
118	  VMA range remain where they were allocated based on the
119	  policy at the time they were allocated.  However, since
120	  2.6.16, Linux supports page migration via the mbind() system
121	  call, so that page contents can be moved to match a newly
122	  installed policy.
123
124Shared Policy
125	Conceptually, shared policies apply to "memory objects" mapped
126	shared into one or more tasks' distinct address spaces.  An
127	application installs shared policies the same way as VMA
128	policies--using the mbind() system call specifying a range of
129	virtual addresses that map the shared object.  However, unlike
130	VMA policies, which can be considered to be an attribute of a
131	range of a task's address space, shared policies apply
132	directly to the shared object.  Thus, all tasks that attach to
133	the object share the policy, and all pages allocated for the
134	shared object, by any task, will obey the shared policy.
135
136	As of 2.6.22, only shared memory segments, created by shmget() or
137	mmap(MAP_ANONYMOUS|MAP_SHARED), support shared policy.  When shared
138	policy support was added to Linux, the associated data structures were
139	added to hugetlbfs shmem segments.  At the time, hugetlbfs did not
140	support allocation at fault time--a.k.a lazy allocation--so hugetlbfs
141	shmem segments were never "hooked up" to the shared policy support.
142	Although hugetlbfs segments now support lazy allocation, their support
143	for shared policy has not been completed.
144
145	As mentioned above in :ref:`VMA policies <vma_policy>` section,
146	allocations of page cache pages for regular files mmap()ed
147	with MAP_SHARED ignore any VMA policy installed on the virtual
148	address range backed by the shared file mapping.  Rather,
149	shared page cache pages, including pages backing private
150	mappings that have not yet been written by the task, follow
151	task policy, if any, else System Default Policy.
152
153	The shared policy infrastructure supports different policies on subset
154	ranges of the shared object.  However, Linux still splits the VMA of
155	the task that installs the policy for each range of distinct policy.
156	Thus, different tasks that attach to a shared memory segment can have
157	different VMA configurations mapping that one shared object.  This
158	can be seen by examining the /proc/<pid>/numa_maps of tasks sharing
159	a shared memory region, when one task has installed shared policy on
160	one or more ranges of the region.
161
162Components of Memory Policies
163-----------------------------
164
165A NUMA memory policy consists of a "mode", optional mode flags, and
166an optional set of nodes.  The mode determines the behavior of the
167policy, the optional mode flags determine the behavior of the mode,
168and the optional set of nodes can be viewed as the arguments to the
169policy behavior.
170
171Internally, memory policies are implemented by a reference counted
172structure, struct mempolicy.  Details of this structure will be
173discussed in context, below, as required to explain the behavior.
174
175NUMA memory policy supports the following 4 behavioral modes:
176
177Default Mode--MPOL_DEFAULT
178	This mode is only used in the memory policy APIs.  Internally,
179	MPOL_DEFAULT is converted to the NULL memory policy in all
180	policy scopes.  Any existing non-default policy will simply be
181	removed when MPOL_DEFAULT is specified.  As a result,
182	MPOL_DEFAULT means "fall back to the next most specific policy
183	scope."
184
185	For example, a NULL or default task policy will fall back to the
186	system default policy.  A NULL or default vma policy will fall
187	back to the task policy.
188
189	When specified in one of the memory policy APIs, the Default mode
190	does not use the optional set of nodes.
191
192	It is an error for the set of nodes specified for this policy to
193	be non-empty.
194
195MPOL_BIND
196	This mode specifies that memory must come from the set of
197	nodes specified by the policy.  Memory will be allocated from
198	the node in the set with sufficient free memory that is
199	closest to the node where the allocation takes place.
200
201MPOL_PREFERRED
202	This mode specifies that the allocation should be attempted
203	from the single node specified in the policy.  If that
204	allocation fails, the kernel will search other nodes, in order
205	of increasing distance from the preferred node based on
206	information provided by the platform firmware.
207
208	Internally, the Preferred policy uses a single node--the
209	preferred_node member of struct mempolicy.  When the internal
210	mode flag MPOL_F_LOCAL is set, the preferred_node is ignored
211	and the policy is interpreted as local allocation.  "Local"
212	allocation policy can be viewed as a Preferred policy that
213	starts at the node containing the cpu where the allocation
214	takes place.
215
216	It is possible for the user to specify that local allocation
217	is always preferred by passing an empty nodemask with this
218	mode.  If an empty nodemask is passed, the policy cannot use
219	the MPOL_F_STATIC_NODES or MPOL_F_RELATIVE_NODES flags
220	described below.
221
222MPOL_INTERLEAVED
223	This mode specifies that page allocations be interleaved, on a
224	page granularity, across the nodes specified in the policy.
225	This mode also behaves slightly differently, based on the
226	context where it is used:
227
228	For allocation of anonymous pages and shared memory pages,
229	Interleave mode indexes the set of nodes specified by the
230	policy using the page offset of the faulting address into the
231	segment [VMA] containing the address modulo the number of
232	nodes specified by the policy.  It then attempts to allocate a
233	page, starting at the selected node, as if the node had been
234	specified by a Preferred policy or had been selected by a
235	local allocation.  That is, allocation will follow the per
236	node zonelist.
237
238	For allocation of page cache pages, Interleave mode indexes
239	the set of nodes specified by the policy using a node counter
240	maintained per task.  This counter wraps around to the lowest
241	specified node after it reaches the highest specified node.
242	This will tend to spread the pages out over the nodes
243	specified by the policy based on the order in which they are
244	allocated, rather than based on any page offset into an
245	address range or file.  During system boot up, the temporary
246	interleaved system default policy works in this mode.
247
248MPOL_PREFERRED_MANY
249	This mode specifices that the allocation should be preferrably
250	satisfied from the nodemask specified in the policy. If there is
251	a memory pressure on all nodes in the nodemask, the allocation
252	can fall back to all existing numa nodes. This is effectively
253	MPOL_PREFERRED allowed for a mask rather than a single node.
254
255NUMA memory policy supports the following optional mode flags:
256
257MPOL_F_STATIC_NODES
258	This flag specifies that the nodemask passed by
259	the user should not be remapped if the task or VMA's set of allowed
260	nodes changes after the memory policy has been defined.
261
262	Without this flag, any time a mempolicy is rebound because of a
263        change in the set of allowed nodes, the preferred nodemask (Preferred
264        Many), preferred node (Preferred) or nodemask (Bind, Interleave) is
265        remapped to the new set of allowed nodes.  This may result in nodes
266        being used that were previously undesired.
267
268	With this flag, if the user-specified nodes overlap with the
269	nodes allowed by the task's cpuset, then the memory policy is
270	applied to their intersection.  If the two sets of nodes do not
271	overlap, the Default policy is used.
272
273	For example, consider a task that is attached to a cpuset with
274	mems 1-3 that sets an Interleave policy over the same set.  If
275	the cpuset's mems change to 3-5, the Interleave will now occur
276	over nodes 3, 4, and 5.  With this flag, however, since only node
277	3 is allowed from the user's nodemask, the "interleave" only
278	occurs over that node.  If no nodes from the user's nodemask are
279	now allowed, the Default behavior is used.
280
281	MPOL_F_STATIC_NODES cannot be combined with the
282	MPOL_F_RELATIVE_NODES flag.  It also cannot be used for
283	MPOL_PREFERRED policies that were created with an empty nodemask
284	(local allocation).
285
286MPOL_F_RELATIVE_NODES
287	This flag specifies that the nodemask passed
288	by the user will be mapped relative to the set of the task or VMA's
289	set of allowed nodes.  The kernel stores the user-passed nodemask,
290	and if the allowed nodes changes, then that original nodemask will
291	be remapped relative to the new set of allowed nodes.
292
293	Without this flag (and without MPOL_F_STATIC_NODES), anytime a
294	mempolicy is rebound because of a change in the set of allowed
295	nodes, the node (Preferred) or nodemask (Bind, Interleave) is
296	remapped to the new set of allowed nodes.  That remap may not
297	preserve the relative nature of the user's passed nodemask to its
298	set of allowed nodes upon successive rebinds: a nodemask of
299	1,3,5 may be remapped to 7-9 and then to 1-3 if the set of
300	allowed nodes is restored to its original state.
301
302	With this flag, the remap is done so that the node numbers from
303	the user's passed nodemask are relative to the set of allowed
304	nodes.  In other words, if nodes 0, 2, and 4 are set in the user's
305	nodemask, the policy will be effected over the first (and in the
306	Bind or Interleave case, the third and fifth) nodes in the set of
307	allowed nodes.  The nodemask passed by the user represents nodes
308	relative to task or VMA's set of allowed nodes.
309
310	If the user's nodemask includes nodes that are outside the range
311	of the new set of allowed nodes (for example, node 5 is set in
312	the user's nodemask when the set of allowed nodes is only 0-3),
313	then the remap wraps around to the beginning of the nodemask and,
314	if not already set, sets the node in the mempolicy nodemask.
315
316	For example, consider a task that is attached to a cpuset with
317	mems 2-5 that sets an Interleave policy over the same set with
318	MPOL_F_RELATIVE_NODES.  If the cpuset's mems change to 3-7, the
319	interleave now occurs over nodes 3,5-7.  If the cpuset's mems
320	then change to 0,2-3,5, then the interleave occurs over nodes
321	0,2-3,5.
322
323	Thanks to the consistent remapping, applications preparing
324	nodemasks to specify memory policies using this flag should
325	disregard their current, actual cpuset imposed memory placement
326	and prepare the nodemask as if they were always located on
327	memory nodes 0 to N-1, where N is the number of memory nodes the
328	policy is intended to manage.  Let the kernel then remap to the
329	set of memory nodes allowed by the task's cpuset, as that may
330	change over time.
331
332	MPOL_F_RELATIVE_NODES cannot be combined with the
333	MPOL_F_STATIC_NODES flag.  It also cannot be used for
334	MPOL_PREFERRED policies that were created with an empty nodemask
335	(local allocation).
336
337Memory Policy Reference Counting
338================================
339
340To resolve use/free races, struct mempolicy contains an atomic reference
341count field.  Internal interfaces, mpol_get()/mpol_put() increment and
342decrement this reference count, respectively.  mpol_put() will only free
343the structure back to the mempolicy kmem cache when the reference count
344goes to zero.
345
346When a new memory policy is allocated, its reference count is initialized
347to '1', representing the reference held by the task that is installing the
348new policy.  When a pointer to a memory policy structure is stored in another
349structure, another reference is added, as the task's reference will be dropped
350on completion of the policy installation.
351
352During run-time "usage" of the policy, we attempt to minimize atomic operations
353on the reference count, as this can lead to cache lines bouncing between cpus
354and NUMA nodes.  "Usage" here means one of the following:
355
3561) querying of the policy, either by the task itself [using the get_mempolicy()
357   API discussed below] or by another task using the /proc/<pid>/numa_maps
358   interface.
359
3602) examination of the policy to determine the policy mode and associated node
361   or node lists, if any, for page allocation.  This is considered a "hot
362   path".  Note that for MPOL_BIND, the "usage" extends across the entire
363   allocation process, which may sleep during page reclaimation, because the
364   BIND policy nodemask is used, by reference, to filter ineligible nodes.
365
366We can avoid taking an extra reference during the usages listed above as
367follows:
368
3691) we never need to get/free the system default policy as this is never
370   changed nor freed, once the system is up and running.
371
3722) for querying the policy, we do not need to take an extra reference on the
373   target task's task policy nor vma policies because we always acquire the
374   task's mm's mmap_lock for read during the query.  The set_mempolicy() and
375   mbind() APIs [see below] always acquire the mmap_lock for write when
376   installing or replacing task or vma policies.  Thus, there is no possibility
377   of a task or thread freeing a policy while another task or thread is
378   querying it.
379
3803) Page allocation usage of task or vma policy occurs in the fault path where
381   we hold them mmap_lock for read.  Again, because replacing the task or vma
382   policy requires that the mmap_lock be held for write, the policy can't be
383   freed out from under us while we're using it for page allocation.
384
3854) Shared policies require special consideration.  One task can replace a
386   shared memory policy while another task, with a distinct mmap_lock, is
387   querying or allocating a page based on the policy.  To resolve this
388   potential race, the shared policy infrastructure adds an extra reference
389   to the shared policy during lookup while holding a spin lock on the shared
390   policy management structure.  This requires that we drop this extra
391   reference when we're finished "using" the policy.  We must drop the
392   extra reference on shared policies in the same query/allocation paths
393   used for non-shared policies.  For this reason, shared policies are marked
394   as such, and the extra reference is dropped "conditionally"--i.e., only
395   for shared policies.
396
397   Because of this extra reference counting, and because we must lookup
398   shared policies in a tree structure under spinlock, shared policies are
399   more expensive to use in the page allocation path.  This is especially
400   true for shared policies on shared memory regions shared by tasks running
401   on different NUMA nodes.  This extra overhead can be avoided by always
402   falling back to task or system default policy for shared memory regions,
403   or by prefaulting the entire shared memory region into memory and locking
404   it down.  However, this might not be appropriate for all applications.
405
406.. _memory_policy_apis:
407
408Memory Policy APIs
409==================
410
411Linux supports 3 system calls for controlling memory policy.  These APIS
412always affect only the calling task, the calling task's address space, or
413some shared object mapped into the calling task's address space.
414
415.. note::
416   the headers that define these APIs and the parameter data types for
417   user space applications reside in a package that is not part of the
418   Linux kernel.  The kernel system call interfaces, with the 'sys\_'
419   prefix, are defined in <linux/syscalls.h>; the mode and flag
420   definitions are defined in <linux/mempolicy.h>.
421
422Set [Task] Memory Policy::
423
424	long set_mempolicy(int mode, const unsigned long *nmask,
425					unsigned long maxnode);
426
427Set's the calling task's "task/process memory policy" to mode
428specified by the 'mode' argument and the set of nodes defined by
429'nmask'.  'nmask' points to a bit mask of node ids containing at least
430'maxnode' ids.  Optional mode flags may be passed by combining the
431'mode' argument with the flag (for example: MPOL_INTERLEAVE |
432MPOL_F_STATIC_NODES).
433
434See the set_mempolicy(2) man page for more details
435
436
437Get [Task] Memory Policy or Related Information::
438
439	long get_mempolicy(int *mode,
440			   const unsigned long *nmask, unsigned long maxnode,
441			   void *addr, int flags);
442
443Queries the "task/process memory policy" of the calling task, or the
444policy or location of a specified virtual address, depending on the
445'flags' argument.
446
447See the get_mempolicy(2) man page for more details
448
449
450Install VMA/Shared Policy for a Range of Task's Address Space::
451
452	long mbind(void *start, unsigned long len, int mode,
453		   const unsigned long *nmask, unsigned long maxnode,
454		   unsigned flags);
455
456mbind() installs the policy specified by (mode, nmask, maxnodes) as a
457VMA policy for the range of the calling task's address space specified
458by the 'start' and 'len' arguments.  Additional actions may be
459requested via the 'flags' argument.
460
461See the mbind(2) man page for more details.
462
463Memory Policy Command Line Interface
464====================================
465
466Although not strictly part of the Linux implementation of memory policy,
467a command line tool, numactl(8), exists that allows one to:
468
469+ set the task policy for a specified program via set_mempolicy(2), fork(2) and
470  exec(2)
471
472+ set the shared policy for a shared memory segment via mbind(2)
473
474The numactl(8) tool is packaged with the run-time version of the library
475containing the memory policy system call wrappers.  Some distributions
476package the headers and compile-time libraries in a separate development
477package.
478
479.. _mem_pol_and_cpusets:
480
481Memory Policies and cpusets
482===========================
483
484Memory policies work within cpusets as described above.  For memory policies
485that require a node or set of nodes, the nodes are restricted to the set of
486nodes whose memories are allowed by the cpuset constraints.  If the nodemask
487specified for the policy contains nodes that are not allowed by the cpuset and
488MPOL_F_RELATIVE_NODES is not used, the intersection of the set of nodes
489specified for the policy and the set of nodes with memory is used.  If the
490result is the empty set, the policy is considered invalid and cannot be
491installed.  If MPOL_F_RELATIVE_NODES is used, the policy's nodes are mapped
492onto and folded into the task's set of allowed nodes as previously described.
493
494The interaction of memory policies and cpusets can be problematic when tasks
495in two cpusets share access to a memory region, such as shared memory segments
496created by shmget() of mmap() with the MAP_ANONYMOUS and MAP_SHARED flags, and
497any of the tasks install shared policy on the region, only nodes whose
498memories are allowed in both cpusets may be used in the policies.  Obtaining
499this information requires "stepping outside" the memory policy APIs to use the
500cpuset information and requires that one know in what cpusets other task might
501be attaching to the shared region.  Furthermore, if the cpusets' allowed
502memory sets are disjoint, "local" allocation is the only valid policy.
503