blob: 6aaaeb38730cf0ff1d5389cd73a6a8d2fb6bd413 [file] [log] [blame]
Lee Schermerhorn42b88e62007-08-22 14:01:06 -07001
2What is Linux Memory Policy?
3
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
9support.
10
11Memory policies should not be confused with cpusets (Documentation/cpusets.txt)
12which is an administrative mechanism for restricting the nodes from which
13memory may be allocated by a set of processes. Memory policies are a
14programming interface that a NUMA-aware application can take advantage of. When
15both cpusets and policies are applied to a task, the restrictions of the cpuset
16takes priority. See "MEMORY POLICIES AND CPUSETS" below for more details.
17
18MEMORY POLICY CONCEPTS
19
20Scope of Memory Policies
21
22The Linux kernel supports _scopes_ of memory policy, described here from
23most general to most specific:
24
25 System Default Policy: this policy is "hard coded" into the kernel. It
26 is the policy that governs all page allocations that aren't controlled
27 by one of the more specific policy scopes discussed below. When the
28 system is "up and running", the system default policy will use "local
29 allocation" described below. However, during boot up, the system
30 default policy will be set to interleave allocations across all nodes
31 with "sufficient" memory, so as not to overload the initial boot node
32 with boot-time allocations.
33
34 Task/Process Policy: this is an optional, per-task policy. When defined
35 for a specific task, this policy controls all page allocations made by or
36 on behalf of the task that aren't controlled by a more specific scope.
37 If a task does not define a task policy, then all page allocations that
38 would have been controlled by the task policy "fall back" to the System
39 Default Policy.
40
41 The task policy applies to the entire address space of a task. Thus,
42 it is inheritable, and indeed is inherited, across both fork()
43 [clone() w/o the CLONE_VM flag] and exec*(). This allows a parent task
44 to establish the task policy for a child task exec()'d from an
45 executable image that has no awareness of memory policy. See the
46 MEMORY POLICY APIS section, below, for an overview of the system call
47 that a task may use to set/change it's task/process policy.
48
49 In a multi-threaded task, task policies apply only to the thread
50 [Linux kernel task] that installs the policy and any threads
51 subsequently created by that thread. Any sibling threads existing
52 at the time a new task policy is installed retain their current
53 policy.
54
55 A task policy applies only to pages allocated after the policy is
56 installed. Any pages already faulted in by the task when the task
57 changes its task policy remain where they were allocated based on
58 the policy at the time they were allocated.
59
60 VMA Policy: A "VMA" or "Virtual Memory Area" refers to a range of a task's
Matt LaPlanted9195882008-07-25 19:45:33 -070061 virtual address space. A task may define a specific policy for a range
Lee Schermerhorn42b88e62007-08-22 14:01:06 -070062 of its virtual address space. See the MEMORY POLICIES APIS section,
63 below, for an overview of the mbind() system call used to set a VMA
64 policy.
65
66 A VMA policy will govern the allocation of pages that back this region of
67 the address space. Any regions of the task's address space that don't
68 have an explicit VMA policy will fall back to the task policy, which may
69 itself fall back to the System Default Policy.
70
71 VMA policies have a few complicating details:
72
73 VMA policy applies ONLY to anonymous pages. These include pages
74 allocated for anonymous segments, such as the task stack and heap, and
75 any regions of the address space mmap()ed with the MAP_ANONYMOUS flag.
76 If a VMA policy is applied to a file mapping, it will be ignored if
77 the mapping used the MAP_SHARED flag. If the file mapping used the
78 MAP_PRIVATE flag, the VMA policy will only be applied when an
79 anonymous page is allocated on an attempt to write to the mapping--
80 i.e., at Copy-On-Write.
81
82 VMA policies are shared between all tasks that share a virtual address
83 space--a.k.a. threads--independent of when the policy is installed; and
84 they are inherited across fork(). However, because VMA policies refer
85 to a specific region of a task's address space, and because the address
86 space is discarded and recreated on exec*(), VMA policies are NOT
87 inheritable across exec(). Thus, only NUMA-aware applications may
88 use VMA policies.
89
90 A task may install a new VMA policy on a sub-range of a previously
91 mmap()ed region. When this happens, Linux splits the existing virtual
92 memory area into 2 or 3 VMAs, each with it's own policy.
93
94 By default, VMA policy applies only to pages allocated after the policy
95 is installed. Any pages already faulted into the VMA range remain
96 where they were allocated based on the policy at the time they were
97 allocated. However, since 2.6.16, Linux supports page migration via
98 the mbind() system call, so that page contents can be moved to match
99 a newly installed policy.
100
101 Shared Policy: Conceptually, shared policies apply to "memory objects"
102 mapped shared into one or more tasks' distinct address spaces. An
103 application installs a shared policies the same way as VMA policies--using
104 the mbind() system call specifying a range of virtual addresses that map
105 the shared object. However, unlike VMA policies, which can be considered
106 to be an attribute of a range of a task's address space, shared policies
107 apply directly to the shared object. Thus, all tasks that attach to the
108 object share the policy, and all pages allocated for the shared object,
109 by any task, will obey the shared policy.
110
111 As of 2.6.22, only shared memory segments, created by shmget() or
112 mmap(MAP_ANONYMOUS|MAP_SHARED), support shared policy. When shared
113 policy support was added to Linux, the associated data structures were
114 added to hugetlbfs shmem segments. At the time, hugetlbfs did not
115 support allocation at fault time--a.k.a lazy allocation--so hugetlbfs
116 shmem segments were never "hooked up" to the shared policy support.
117 Although hugetlbfs segments now support lazy allocation, their support
118 for shared policy has not been completed.
119
120 As mentioned above [re: VMA policies], allocations of page cache
121 pages for regular files mmap()ed with MAP_SHARED ignore any VMA
122 policy installed on the virtual address range backed by the shared
123 file mapping. Rather, shared page cache pages, including pages backing
124 private mappings that have not yet been written by the task, follow
125 task policy, if any, else System Default Policy.
126
127 The shared policy infrastructure supports different policies on subset
128 ranges of the shared object. However, Linux still splits the VMA of
129 the task that installs the policy for each range of distinct policy.
130 Thus, different tasks that attach to a shared memory segment can have
131 different VMA configurations mapping that one shared object. This
132 can be seen by examining the /proc/<pid>/numa_maps of tasks sharing
133 a shared memory region, when one task has installed shared policy on
134 one or more ranges of the region.
135
136Components of Memory Policies
137
David Rientjes65d66fc2008-04-28 02:12:31 -0700138 A Linux memory policy consists of a "mode", optional mode flags, and an
139 optional set of nodes. The mode determines the behavior of the policy,
140 the optional mode flags determine the behavior of the mode, and the
141 optional set of nodes can be viewed as the arguments to the policy
142 behavior.
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700143
144 Internally, memory policies are implemented by a reference counted
145 structure, struct mempolicy. Details of this structure will be discussed
146 in context, below, as required to explain the behavior.
147
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700148 Linux memory policy supports the following 4 behavioral modes:
149
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700150 Default Mode--MPOL_DEFAULT: This mode is only used in the memory
151 policy APIs. Internally, MPOL_DEFAULT is converted to the NULL
152 memory policy in all policy scopes. Any existing non-default policy
153 will simply be removed when MPOL_DEFAULT is specified. As a result,
154 MPOL_DEFAULT means "fall back to the next most specific policy scope."
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700155
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700156 For example, a NULL or default task policy will fall back to the
157 system default policy. A NULL or default vma policy will fall
158 back to the task policy.
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700159
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700160 When specified in one of the memory policy APIs, the Default mode
161 does not use the optional set of nodes.
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700162
David Rientjes65d66fc2008-04-28 02:12:31 -0700163 It is an error for the set of nodes specified for this policy to
164 be non-empty.
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700165
166 MPOL_BIND: This mode specifies that memory must come from the
Mel Gorman19770b32008-04-28 02:12:18 -0700167 set of nodes specified by the policy. Memory will be allocated from
168 the node in the set with sufficient free memory that is closest to
169 the node where the allocation takes place.
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700170
171 MPOL_PREFERRED: This mode specifies that the allocation should be
172 attempted from the single node specified in the policy. If that
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700173 allocation fails, the kernel will search other nodes, in order of
174 increasing distance from the preferred node based on information
175 provided by the platform firmware.
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700176 containing the cpu where the allocation takes place.
177
178 Internally, the Preferred policy uses a single node--the
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700179 preferred_node member of struct mempolicy. When the internal
180 mode flag MPOL_F_LOCAL is set, the preferred_node is ignored and
181 the policy is interpreted as local allocation. "Local" allocation
182 policy can be viewed as a Preferred policy that starts at the node
183 containing the cpu where the allocation takes place.
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700184
David Rientjes3e1f0642008-04-28 02:12:34 -0700185 It is possible for the user to specify that local allocation is
186 always preferred by passing an empty nodemask with this mode.
187 If an empty nodemask is passed, the policy cannot use the
188 MPOL_F_STATIC_NODES or MPOL_F_RELATIVE_NODES flags described
189 below.
190
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700191 MPOL_INTERLEAVED: This mode specifies that page allocations be
192 interleaved, on a page granularity, across the nodes specified in
193 the policy. This mode also behaves slightly differently, based on
194 the context where it is used:
195
196 For allocation of anonymous pages and shared memory pages,
197 Interleave mode indexes the set of nodes specified by the policy
198 using the page offset of the faulting address into the segment
199 [VMA] containing the address modulo the number of nodes specified
200 by the policy. It then attempts to allocate a page, starting at
201 the selected node, as if the node had been specified by a Preferred
202 policy or had been selected by a local allocation. That is,
203 allocation will follow the per node zonelist.
204
205 For allocation of page cache pages, Interleave mode indexes the set
206 of nodes specified by the policy using a node counter maintained
207 per task. This counter wraps around to the lowest specified node
208 after it reaches the highest specified node. This will tend to
209 spread the pages out over the nodes specified by the policy based
210 on the order in which they are allocated, rather than based on any
211 page offset into an address range or file. During system boot up,
212 the temporary interleaved system default policy works in this
213 mode.
214
David Rientjes65d66fc2008-04-28 02:12:31 -0700215 Linux memory policy supports the following optional mode flags:
216
217 MPOL_F_STATIC_NODES: This flag specifies that the nodemask passed by
218 the user should not be remapped if the task or VMA's set of allowed
219 nodes changes after the memory policy has been defined.
220
221 Without this flag, anytime a mempolicy is rebound because of a
222 change in the set of allowed nodes, the node (Preferred) or
223 nodemask (Bind, Interleave) is remapped to the new set of
224 allowed nodes. This may result in nodes being used that were
225 previously undesired.
226
227 With this flag, if the user-specified nodes overlap with the
228 nodes allowed by the task's cpuset, then the memory policy is
229 applied to their intersection. If the two sets of nodes do not
230 overlap, the Default policy is used.
231
232 For example, consider a task that is attached to a cpuset with
233 mems 1-3 that sets an Interleave policy over the same set. If
234 the cpuset's mems change to 3-5, the Interleave will now occur
235 over nodes 3, 4, and 5. With this flag, however, since only node
236 3 is allowed from the user's nodemask, the "interleave" only
237 occurs over that node. If no nodes from the user's nodemask are
238 now allowed, the Default behavior is used.
239
David Rientjes3e1f0642008-04-28 02:12:34 -0700240 MPOL_F_STATIC_NODES cannot be combined with the
241 MPOL_F_RELATIVE_NODES flag. It also cannot be used for
242 MPOL_PREFERRED policies that were created with an empty nodemask
243 (local allocation).
David Rientjes65d66fc2008-04-28 02:12:31 -0700244
245 MPOL_F_RELATIVE_NODES: This flag specifies that the nodemask passed
246 by the user will be mapped relative to the set of the task or VMA's
247 set of allowed nodes. The kernel stores the user-passed nodemask,
248 and if the allowed nodes changes, then that original nodemask will
249 be remapped relative to the new set of allowed nodes.
250
251 Without this flag (and without MPOL_F_STATIC_NODES), anytime a
252 mempolicy is rebound because of a change in the set of allowed
253 nodes, the node (Preferred) or nodemask (Bind, Interleave) is
254 remapped to the new set of allowed nodes. That remap may not
255 preserve the relative nature of the user's passed nodemask to its
256 set of allowed nodes upon successive rebinds: a nodemask of
257 1,3,5 may be remapped to 7-9 and then to 1-3 if the set of
258 allowed nodes is restored to its original state.
259
260 With this flag, the remap is done so that the node numbers from
261 the user's passed nodemask are relative to the set of allowed
262 nodes. In other words, if nodes 0, 2, and 4 are set in the user's
263 nodemask, the policy will be effected over the first (and in the
264 Bind or Interleave case, the third and fifth) nodes in the set of
265 allowed nodes. The nodemask passed by the user represents nodes
266 relative to task or VMA's set of allowed nodes.
267
268 If the user's nodemask includes nodes that are outside the range
269 of the new set of allowed nodes (for example, node 5 is set in
270 the user's nodemask when the set of allowed nodes is only 0-3),
271 then the remap wraps around to the beginning of the nodemask and,
272 if not already set, sets the node in the mempolicy nodemask.
273
274 For example, consider a task that is attached to a cpuset with
275 mems 2-5 that sets an Interleave policy over the same set with
276 MPOL_F_RELATIVE_NODES. If the cpuset's mems change to 3-7, the
277 interleave now occurs over nodes 3,5-6. If the cpuset's mems
278 then change to 0,2-3,5, then the interleave occurs over nodes
279 0,3,5.
280
281 Thanks to the consistent remapping, applications preparing
282 nodemasks to specify memory policies using this flag should
283 disregard their current, actual cpuset imposed memory placement
284 and prepare the nodemask as if they were always located on
285 memory nodes 0 to N-1, where N is the number of memory nodes the
286 policy is intended to manage. Let the kernel then remap to the
287 set of memory nodes allowed by the task's cpuset, as that may
288 change over time.
289
David Rientjes3e1f0642008-04-28 02:12:34 -0700290 MPOL_F_RELATIVE_NODES cannot be combined with the
291 MPOL_F_STATIC_NODES flag. It also cannot be used for
292 MPOL_PREFERRED policies that were created with an empty nodemask
293 (local allocation).
David Rientjes65d66fc2008-04-28 02:12:31 -0700294
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700295MEMORY POLICY REFERENCE COUNTING
296
297To resolve use/free races, struct mempolicy contains an atomic reference
298count field. Internal interfaces, mpol_get()/mpol_put() increment and
299decrement this reference count, respectively. mpol_put() will only free
300the structure back to the mempolicy kmem cache when the reference count
301goes to zero.
302
303When a new memory policy is allocated, it's reference count is initialized
304to '1', representing the reference held by the task that is installing the
305new policy. When a pointer to a memory policy structure is stored in another
306structure, another reference is added, as the task's reference will be dropped
307on completion of the policy installation.
308
309During run-time "usage" of the policy, we attempt to minimize atomic operations
310on the reference count, as this can lead to cache lines bouncing between cpus
311and NUMA nodes. "Usage" here means one of the following:
312
3131) querying of the policy, either by the task itself [using the get_mempolicy()
314 API discussed below] or by another task using the /proc/<pid>/numa_maps
315 interface.
316
3172) examination of the policy to determine the policy mode and associated node
318 or node lists, if any, for page allocation. This is considered a "hot
319 path". Note that for MPOL_BIND, the "usage" extends across the entire
320 allocation process, which may sleep during page reclaimation, because the
321 BIND policy nodemask is used, by reference, to filter ineligible nodes.
322
323We can avoid taking an extra reference during the usages listed above as
324follows:
325
3261) we never need to get/free the system default policy as this is never
327 changed nor freed, once the system is up and running.
328
3292) for querying the policy, we do not need to take an extra reference on the
330 target task's task policy nor vma policies because we always acquire the
331 task's mm's mmap_sem for read during the query. The set_mempolicy() and
332 mbind() APIs [see below] always acquire the mmap_sem for write when
333 installing or replacing task or vma policies. Thus, there is no possibility
334 of a task or thread freeing a policy while another task or thread is
335 querying it.
336
3373) Page allocation usage of task or vma policy occurs in the fault path where
338 we hold them mmap_sem for read. Again, because replacing the task or vma
339 policy requires that the mmap_sem be held for write, the policy can't be
340 freed out from under us while we're using it for page allocation.
341
3424) Shared policies require special consideration. One task can replace a
343 shared memory policy while another task, with a distinct mmap_sem, is
344 querying or allocating a page based on the policy. To resolve this
345 potential race, the shared policy infrastructure adds an extra reference
346 to the shared policy during lookup while holding a spin lock on the shared
347 policy management structure. This requires that we drop this extra
348 reference when we're finished "using" the policy. We must drop the
349 extra reference on shared policies in the same query/allocation paths
350 used for non-shared policies. For this reason, shared policies are marked
351 as such, and the extra reference is dropped "conditionally"--i.e., only
352 for shared policies.
353
354 Because of this extra reference counting, and because we must lookup
355 shared policies in a tree structure under spinlock, shared policies are
Matt LaPlanted9195882008-07-25 19:45:33 -0700356 more expensive to use in the page allocation path. This is especially
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700357 true for shared policies on shared memory regions shared by tasks running
358 on different NUMA nodes. This extra overhead can be avoided by always
359 falling back to task or system default policy for shared memory regions,
360 or by prefaulting the entire shared memory region into memory and locking
361 it down. However, this might not be appropriate for all applications.
362
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700363MEMORY POLICY APIs
364
365Linux supports 3 system calls for controlling memory policy. These APIS
366always affect only the calling task, the calling task's address space, or
367some shared object mapped into the calling task's address space.
368
369 Note: the headers that define these APIs and the parameter data types
370 for user space applications reside in a package that is not part of
371 the Linux kernel. The kernel system call interfaces, with the 'sys_'
372 prefix, are defined in <linux/syscalls.h>; the mode and flag
373 definitions are defined in <linux/mempolicy.h>.
374
375Set [Task] Memory Policy:
376
377 long set_mempolicy(int mode, const unsigned long *nmask,
378 unsigned long maxnode);
379
380 Set's the calling task's "task/process memory policy" to mode
381 specified by the 'mode' argument and the set of nodes defined
382 by 'nmask'. 'nmask' points to a bit mask of node ids containing
David Rientjes65d66fc2008-04-28 02:12:31 -0700383 at least 'maxnode' ids. Optional mode flags may be passed by
384 combining the 'mode' argument with the flag (for example:
385 MPOL_INTERLEAVE | MPOL_F_STATIC_NODES).
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700386
387 See the set_mempolicy(2) man page for more details
388
389
390Get [Task] Memory Policy or Related Information
391
392 long get_mempolicy(int *mode,
393 const unsigned long *nmask, unsigned long maxnode,
394 void *addr, int flags);
395
396 Queries the "task/process memory policy" of the calling task, or
397 the policy or location of a specified virtual address, depending
398 on the 'flags' argument.
399
400 See the get_mempolicy(2) man page for more details
401
402
403Install VMA/Shared Policy for a Range of Task's Address Space
404
405 long mbind(void *start, unsigned long len, int mode,
406 const unsigned long *nmask, unsigned long maxnode,
407 unsigned flags);
408
409 mbind() installs the policy specified by (mode, nmask, maxnodes) as
410 a VMA policy for the range of the calling task's address space
411 specified by the 'start' and 'len' arguments. Additional actions
412 may be requested via the 'flags' argument.
413
414 See the mbind(2) man page for more details.
415
416MEMORY POLICY COMMAND LINE INTERFACE
417
418Although not strictly part of the Linux implementation of memory policy,
419a command line tool, numactl(8), exists that allows one to:
420
421+ set the task policy for a specified program via set_mempolicy(2), fork(2) and
422 exec(2)
423
424+ set the shared policy for a shared memory segment via mbind(2)
425
426The numactl(8) tool is packages with the run-time version of the library
427containing the memory policy system call wrappers. Some distributions
428package the headers and compile-time libraries in a separate development
429package.
430
431
432MEMORY POLICIES AND CPUSETS
433
434Memory policies work within cpusets as described above. For memory policies
435that require a node or set of nodes, the nodes are restricted to the set of
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700436nodes whose memories are allowed by the cpuset constraints. If the nodemask
David Rientjes65d66fc2008-04-28 02:12:31 -0700437specified for the policy contains nodes that are not allowed by the cpuset and
438MPOL_F_RELATIVE_NODES is not used, the intersection of the set of nodes
439specified for the policy and the set of nodes with memory is used. If the
440result is the empty set, the policy is considered invalid and cannot be
441installed. If MPOL_F_RELATIVE_NODES is used, the policy's nodes are mapped
442onto and folded into the task's set of allowed nodes as previously described.
Lee Schermerhorn42b88e62007-08-22 14:01:06 -0700443
David Rientjes65d66fc2008-04-28 02:12:31 -0700444The interaction of memory policies and cpusets can be problematic when tasks
445in two cpusets share access to a memory region, such as shared memory segments
446created by shmget() of mmap() with the MAP_ANONYMOUS and MAP_SHARED flags, and
447any of the tasks install shared policy on the region, only nodes whose
448memories are allowed in both cpusets may be used in the policies. Obtaining
449this information requires "stepping outside" the memory policy APIs to use the
450cpuset information and requires that one know in what cpusets other task might
451be attaching to the shared region. Furthermore, if the cpusets' allowed
452memory sets are disjoint, "local" allocation is the only valid policy.