blob: 9695a9a3ab902f8fef6c4885bfb1a91c6ed77916 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
Christoph Lameter8bccd852005-10-29 18:16:59 -07005 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 * Subject to the GNU Public License, version 2.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
Christoph Lameter8bccd852005-10-29 18:16:59 -070021 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070022 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
Christoph Lameter8bccd852005-10-29 18:16:59 -070024 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070028 * preferred Try a specific node first before normal fallback.
David Rientjes00ef2d22013-02-22 16:35:36 -080029 * As a special case NUMA_NO_NODE here means do the allocation
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
Christoph Lameter8bccd852005-10-29 18:16:59 -070033 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070034 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
Linus Torvalds1da177e2005-04-16 15:20:36 -070066*/
67
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070068#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69
Linus Torvalds1da177e2005-04-16 15:20:36 -070070#include <linux/mempolicy.h>
71#include <linux/mm.h>
72#include <linux/highmem.h>
73#include <linux/hugetlb.h>
74#include <linux/kernel.h>
75#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070076#include <linux/nodemask.h>
77#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070078#include <linux/slab.h>
79#include <linux/string.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040080#include <linux/export.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070081#include <linux/nsproxy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070082#include <linux/interrupt.h>
83#include <linux/init.h>
84#include <linux/compat.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080085#include <linux/swap.h>
Christoph Lameter1a75a6c2006-01-08 01:01:02 -080086#include <linux/seq_file.h>
87#include <linux/proc_fs.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080088#include <linux/migrate.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080089#include <linux/ksm.h>
Christoph Lameter95a402c2006-06-23 02:03:53 -070090#include <linux/rmap.h>
David Quigley86c3a762006-06-23 02:04:02 -070091#include <linux/security.h>
Adrian Bunkdbcb0f12007-10-16 01:26:26 -070092#include <linux/syscalls.h>
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -070093#include <linux/ctype.h>
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -080094#include <linux/mm_inline.h>
Lee Schermerhornb24f53a2012-10-25 14:16:32 +020095#include <linux/mmu_notifier.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070096#include <linux/printk.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080097
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <asm/tlbflush.h>
99#include <asm/uaccess.h>
Michal Hocko778d3b02011-07-26 16:08:30 -0700100#include <linux/random.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101
Nick Piggin62695a82008-10-18 20:26:09 -0700102#include "internal.h"
103
Christoph Lameter38e35862006-01-08 01:01:01 -0800104/* Internal flags */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800105#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
Christoph Lameter38e35862006-01-08 01:01:01 -0800106#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800107
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800108static struct kmem_cache *policy_cache;
109static struct kmem_cache *sn_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111/* Highest zone. An specific allocation for a zone below that is not
112 policied. */
Christoph Lameter62672762007-02-10 01:43:07 -0800113enum zone_type policy_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700115/*
116 * run-time system-wide default policy => local allocation
117 */
H Hartley Sweetene754d792011-10-31 17:09:23 -0700118static struct mempolicy default_policy = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 .refcnt = ATOMIC_INIT(1), /* never free it */
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700120 .mode = MPOL_PREFERRED,
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700121 .flags = MPOL_F_LOCAL,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122};
123
Mel Gorman5606e382012-11-02 18:19:13 +0000124static struct mempolicy preferred_node_policy[MAX_NUMNODES];
125
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -0700126struct mempolicy *get_task_policy(struct task_struct *p)
Mel Gorman5606e382012-11-02 18:19:13 +0000127{
128 struct mempolicy *pol = p->mempolicy;
Oleg Nesterovf15ca782014-10-09 15:27:43 -0700129 int node;
Mel Gorman5606e382012-11-02 18:19:13 +0000130
Oleg Nesterovf15ca782014-10-09 15:27:43 -0700131 if (pol)
132 return pol;
Mel Gorman5606e382012-11-02 18:19:13 +0000133
Oleg Nesterovf15ca782014-10-09 15:27:43 -0700134 node = numa_node_id();
135 if (node != NUMA_NO_NODE) {
136 pol = &preferred_node_policy[node];
137 /* preferred_node_policy is not initialised early in boot */
138 if (pol->mode)
139 return pol;
Mel Gorman5606e382012-11-02 18:19:13 +0000140 }
141
Oleg Nesterovf15ca782014-10-09 15:27:43 -0700142 return &default_policy;
Mel Gorman5606e382012-11-02 18:19:13 +0000143}
144
David Rientjes37012942008-04-28 02:12:33 -0700145static const struct mempolicy_operations {
146 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
Miao Xie708c1bb2010-05-24 14:32:07 -0700147 /*
148 * If read-side task has no lock to protect task->mempolicy, write-side
149 * task will rebind the task->mempolicy by two step. The first step is
150 * setting all the newly nodes, and the second step is cleaning all the
151 * disallowed nodes. In this way, we can avoid finding no node to alloc
152 * page.
153 * If we have a lock to protect task->mempolicy in read-side, we do
154 * rebind directly.
155 *
156 * step:
157 * MPOL_REBIND_ONCE - do rebind work at once
158 * MPOL_REBIND_STEP1 - set all the newly nodes
159 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
160 */
161 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
162 enum mpol_rebind_step step);
David Rientjes37012942008-04-28 02:12:33 -0700163} mpol_ops[MPOL_MAX];
164
Mel Gorman19770b32008-04-28 02:12:18 -0700165/* Check that the nodemask contains at least one populated zone */
David Rientjes37012942008-04-28 02:12:33 -0700166static int is_valid_nodemask(const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167{
Lai Jiangshand3eb1572013-02-22 16:33:22 -0800168 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
David Rientjesf5b087b2008-04-28 02:12:27 -0700171static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
172{
Bob Liu6d556292010-05-24 14:31:59 -0700173 return pol->flags & MPOL_MODE_FLAGS;
David Rientjes4c50bc02008-04-28 02:12:30 -0700174}
175
176static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
177 const nodemask_t *rel)
178{
179 nodemask_t tmp;
180 nodes_fold(tmp, *orig, nodes_weight(*rel));
181 nodes_onto(*ret, tmp, *rel);
David Rientjesf5b087b2008-04-28 02:12:27 -0700182}
183
David Rientjes37012942008-04-28 02:12:33 -0700184static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
185{
186 if (nodes_empty(*nodes))
187 return -EINVAL;
188 pol->v.nodes = *nodes;
189 return 0;
190}
191
192static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
193{
194 if (!nodes)
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700195 pol->flags |= MPOL_F_LOCAL; /* local allocation */
David Rientjes37012942008-04-28 02:12:33 -0700196 else if (nodes_empty(*nodes))
197 return -EINVAL; /* no allowed nodes */
198 else
199 pol->v.preferred_node = first_node(*nodes);
200 return 0;
201}
202
203static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
204{
205 if (!is_valid_nodemask(nodes))
206 return -EINVAL;
207 pol->v.nodes = *nodes;
208 return 0;
209}
210
Miao Xie58568d22009-06-16 15:31:49 -0700211/*
212 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
213 * any, for the new policy. mpol_new() has already validated the nodes
214 * parameter with respect to the policy mode and flags. But, we need to
215 * handle an empty nodemask with MPOL_PREFERRED here.
216 *
217 * Must be called holding task's alloc_lock to protect task's mems_allowed
218 * and mempolicy. May also be called holding the mmap_semaphore for write.
219 */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700220static int mpol_set_nodemask(struct mempolicy *pol,
221 const nodemask_t *nodes, struct nodemask_scratch *nsc)
Miao Xie58568d22009-06-16 15:31:49 -0700222{
Miao Xie58568d22009-06-16 15:31:49 -0700223 int ret;
224
225 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
226 if (pol == NULL)
227 return 0;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -0800228 /* Check N_MEMORY */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700229 nodes_and(nsc->mask1,
Lai Jiangshan01f13bd2012-12-12 13:51:33 -0800230 cpuset_current_mems_allowed, node_states[N_MEMORY]);
Miao Xie58568d22009-06-16 15:31:49 -0700231
232 VM_BUG_ON(!nodes);
233 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
234 nodes = NULL; /* explicit local allocation */
235 else {
236 if (pol->flags & MPOL_F_RELATIVE_NODES)
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700237 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
Miao Xie58568d22009-06-16 15:31:49 -0700238 else
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700239 nodes_and(nsc->mask2, *nodes, nsc->mask1);
240
Miao Xie58568d22009-06-16 15:31:49 -0700241 if (mpol_store_user_nodemask(pol))
242 pol->w.user_nodemask = *nodes;
243 else
244 pol->w.cpuset_mems_allowed =
245 cpuset_current_mems_allowed;
246 }
247
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700248 if (nodes)
249 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
250 else
251 ret = mpol_ops[pol->mode].create(pol, NULL);
Miao Xie58568d22009-06-16 15:31:49 -0700252 return ret;
253}
254
255/*
256 * This function just creates a new policy, does some check and simple
257 * initialization. You must invoke mpol_set_nodemask() to set nodes.
258 */
David Rientjes028fec42008-04-28 02:12:25 -0700259static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
260 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261{
262 struct mempolicy *policy;
263
David Rientjes028fec42008-04-28 02:12:25 -0700264 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
David Rientjes00ef2d22013-02-22 16:35:36 -0800265 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
Paul Mundt140d5a42007-07-15 23:38:16 -0700266
David Rientjes3e1f0642008-04-28 02:12:34 -0700267 if (mode == MPOL_DEFAULT) {
268 if (nodes && !nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700269 return ERR_PTR(-EINVAL);
Lee Schermerhornd3a71032012-10-25 14:16:29 +0200270 return NULL;
David Rientjes37012942008-04-28 02:12:33 -0700271 }
David Rientjes3e1f0642008-04-28 02:12:34 -0700272 VM_BUG_ON(!nodes);
273
274 /*
275 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
276 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
277 * All other modes require a valid pointer to a non-empty nodemask.
278 */
279 if (mode == MPOL_PREFERRED) {
280 if (nodes_empty(*nodes)) {
281 if (((flags & MPOL_F_STATIC_NODES) ||
282 (flags & MPOL_F_RELATIVE_NODES)))
283 return ERR_PTR(-EINVAL);
David Rientjes3e1f0642008-04-28 02:12:34 -0700284 }
Peter Zijlstra479e2802012-10-25 14:16:28 +0200285 } else if (mode == MPOL_LOCAL) {
286 if (!nodes_empty(*nodes))
287 return ERR_PTR(-EINVAL);
288 mode = MPOL_PREFERRED;
David Rientjes3e1f0642008-04-28 02:12:34 -0700289 } else if (nodes_empty(*nodes))
290 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
292 if (!policy)
293 return ERR_PTR(-ENOMEM);
294 atomic_set(&policy->refcnt, 1);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700295 policy->mode = mode;
David Rientjes3e1f0642008-04-28 02:12:34 -0700296 policy->flags = flags;
David Rientjesf5b087b2008-04-28 02:12:27 -0700297
David Rientjes37012942008-04-28 02:12:33 -0700298 return policy;
299}
300
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700301/* Slow path of a mpol destructor. */
302void __mpol_put(struct mempolicy *p)
303{
304 if (!atomic_dec_and_test(&p->refcnt))
305 return;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700306 kmem_cache_free(policy_cache, p);
307}
308
Miao Xie708c1bb2010-05-24 14:32:07 -0700309static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
310 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700311{
312}
313
Miao Xie708c1bb2010-05-24 14:32:07 -0700314/*
315 * step:
316 * MPOL_REBIND_ONCE - do rebind work at once
317 * MPOL_REBIND_STEP1 - set all the newly nodes
318 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
319 */
320static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
321 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700322{
323 nodemask_t tmp;
324
325 if (pol->flags & MPOL_F_STATIC_NODES)
326 nodes_and(tmp, pol->w.user_nodemask, *nodes);
327 else if (pol->flags & MPOL_F_RELATIVE_NODES)
328 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
329 else {
Miao Xie708c1bb2010-05-24 14:32:07 -0700330 /*
331 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
332 * result
333 */
334 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
335 nodes_remap(tmp, pol->v.nodes,
336 pol->w.cpuset_mems_allowed, *nodes);
337 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
338 } else if (step == MPOL_REBIND_STEP2) {
339 tmp = pol->w.cpuset_mems_allowed;
340 pol->w.cpuset_mems_allowed = *nodes;
341 } else
342 BUG();
David Rientjes37012942008-04-28 02:12:33 -0700343 }
344
Miao Xie708c1bb2010-05-24 14:32:07 -0700345 if (nodes_empty(tmp))
346 tmp = *nodes;
347
348 if (step == MPOL_REBIND_STEP1)
349 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
350 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
351 pol->v.nodes = tmp;
352 else
353 BUG();
354
David Rientjes37012942008-04-28 02:12:33 -0700355 if (!node_isset(current->il_next, tmp)) {
356 current->il_next = next_node(current->il_next, tmp);
357 if (current->il_next >= MAX_NUMNODES)
358 current->il_next = first_node(tmp);
359 if (current->il_next >= MAX_NUMNODES)
360 current->il_next = numa_node_id();
361 }
362}
363
364static void mpol_rebind_preferred(struct mempolicy *pol,
Miao Xie708c1bb2010-05-24 14:32:07 -0700365 const nodemask_t *nodes,
366 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700367{
368 nodemask_t tmp;
369
David Rientjes37012942008-04-28 02:12:33 -0700370 if (pol->flags & MPOL_F_STATIC_NODES) {
371 int node = first_node(pol->w.user_nodemask);
372
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700373 if (node_isset(node, *nodes)) {
David Rientjes37012942008-04-28 02:12:33 -0700374 pol->v.preferred_node = node;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700375 pol->flags &= ~MPOL_F_LOCAL;
376 } else
377 pol->flags |= MPOL_F_LOCAL;
David Rientjes37012942008-04-28 02:12:33 -0700378 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
379 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
380 pol->v.preferred_node = first_node(tmp);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700381 } else if (!(pol->flags & MPOL_F_LOCAL)) {
David Rientjes37012942008-04-28 02:12:33 -0700382 pol->v.preferred_node = node_remap(pol->v.preferred_node,
383 pol->w.cpuset_mems_allowed,
384 *nodes);
385 pol->w.cpuset_mems_allowed = *nodes;
386 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387}
388
Miao Xie708c1bb2010-05-24 14:32:07 -0700389/*
390 * mpol_rebind_policy - Migrate a policy to a different set of nodes
391 *
392 * If read-side task has no lock to protect task->mempolicy, write-side
393 * task will rebind the task->mempolicy by two step. The first step is
394 * setting all the newly nodes, and the second step is cleaning all the
395 * disallowed nodes. In this way, we can avoid finding no node to alloc
396 * page.
397 * If we have a lock to protect task->mempolicy in read-side, we do
398 * rebind directly.
399 *
400 * step:
401 * MPOL_REBIND_ONCE - do rebind work at once
402 * MPOL_REBIND_STEP1 - set all the newly nodes
403 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
404 */
405static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
406 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700407{
David Rientjes1d0d2682008-04-28 02:12:32 -0700408 if (!pol)
409 return;
Wang Sheng-Hui89c522c2012-05-29 15:06:16 -0700410 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
David Rientjes1d0d2682008-04-28 02:12:32 -0700411 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
412 return;
Miao Xie708c1bb2010-05-24 14:32:07 -0700413
414 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
415 return;
416
417 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
418 BUG();
419
420 if (step == MPOL_REBIND_STEP1)
421 pol->flags |= MPOL_F_REBINDING;
422 else if (step == MPOL_REBIND_STEP2)
423 pol->flags &= ~MPOL_F_REBINDING;
424 else if (step >= MPOL_REBIND_NSTEP)
425 BUG();
426
427 mpol_ops[pol->mode].rebind(pol, newmask, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700428}
429
430/*
431 * Wrapper for mpol_rebind_policy() that just requires task
432 * pointer, and updates task mempolicy.
Miao Xie58568d22009-06-16 15:31:49 -0700433 *
434 * Called with task's alloc_lock held.
David Rientjes1d0d2682008-04-28 02:12:32 -0700435 */
436
Miao Xie708c1bb2010-05-24 14:32:07 -0700437void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
438 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700439{
Miao Xie708c1bb2010-05-24 14:32:07 -0700440 mpol_rebind_policy(tsk->mempolicy, new, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700441}
442
443/*
444 * Rebind each vma in mm to new nodemask.
445 *
446 * Call holding a reference to mm. Takes mm->mmap_sem during call.
447 */
448
449void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
450{
451 struct vm_area_struct *vma;
452
453 down_write(&mm->mmap_sem);
454 for (vma = mm->mmap; vma; vma = vma->vm_next)
Miao Xie708c1bb2010-05-24 14:32:07 -0700455 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
David Rientjes1d0d2682008-04-28 02:12:32 -0700456 up_write(&mm->mmap_sem);
457}
458
David Rientjes37012942008-04-28 02:12:33 -0700459static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
460 [MPOL_DEFAULT] = {
461 .rebind = mpol_rebind_default,
462 },
463 [MPOL_INTERLEAVE] = {
464 .create = mpol_new_interleave,
465 .rebind = mpol_rebind_nodemask,
466 },
467 [MPOL_PREFERRED] = {
468 .create = mpol_new_preferred,
469 .rebind = mpol_rebind_preferred,
470 },
471 [MPOL_BIND] = {
472 .create = mpol_new_bind,
473 .rebind = mpol_rebind_nodemask,
474 },
475};
476
Christoph Lameterfc301282006-01-18 17:42:29 -0800477static void migrate_page_add(struct page *page, struct list_head *pagelist,
478 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800479
Naoya Horiguchi98094942013-09-11 14:22:14 -0700480/*
481 * Scan through pages checking if pages follow certain conditions,
482 * and move them to the pagelist if they do.
483 */
484static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800485 unsigned long addr, unsigned long end,
486 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800487 void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488{
Hugh Dickins91612e02005-06-21 17:15:07 -0700489 pte_t *orig_pte;
490 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700491 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700492
Hugh Dickins705e87c2005-10-29 18:16:27 -0700493 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700494 do {
Linus Torvalds6aab3412005-11-28 14:34:23 -0800495 struct page *page;
Andy Whitcroft25ba77c2006-12-06 20:33:03 -0800496 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700497
498 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800500 page = vm_normal_page(vma, addr, *pte);
501 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800503 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800504 * vm_normal_page() filters out zero pages, but there might
505 * still be PageReserved pages to skip, perhaps in a VDSO.
Nick Piggin053837f2006-01-18 17:42:27 -0800506 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800507 if (PageReserved(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800508 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800509 nid = page_to_nid(page);
Christoph Lameter38e35862006-01-08 01:01:01 -0800510 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
511 continue;
512
Stephen Wilsonb1f72d12011-05-24 17:12:43 -0700513 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameterfc301282006-01-18 17:42:29 -0800514 migrate_page_add(page, private, flags);
Christoph Lameter38e35862006-01-08 01:01:01 -0800515 else
516 break;
Hugh Dickins91612e02005-06-21 17:15:07 -0700517 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickins705e87c2005-10-29 18:16:27 -0700518 pte_unmap_unlock(orig_pte, ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700519 return addr != end;
520}
521
Naoya Horiguchi98094942013-09-11 14:22:14 -0700522static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
523 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700524 void *private)
525{
526#ifdef CONFIG_HUGETLB_PAGE
527 int nid;
528 struct page *page;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800529 spinlock_t *ptl;
Naoya Horiguchid4c54912014-06-06 10:00:01 -0400530 pte_t entry;
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700531
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800532 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
Naoya Horiguchid4c54912014-06-06 10:00:01 -0400533 entry = huge_ptep_get((pte_t *)pmd);
534 if (!pte_present(entry))
535 goto unlock;
536 page = pte_page(entry);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700537 nid = page_to_nid(page);
538 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
539 goto unlock;
540 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
541 if (flags & (MPOL_MF_MOVE_ALL) ||
542 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
543 isolate_huge_page(page, private);
544unlock:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800545 spin_unlock(ptl);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700546#else
547 BUG();
548#endif
549}
550
Naoya Horiguchi98094942013-09-11 14:22:14 -0700551static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800552 unsigned long addr, unsigned long end,
553 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800554 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700555{
556 pmd_t *pmd;
557 unsigned long next;
558
559 pmd = pmd_offset(pud, addr);
560 do {
561 next = pmd_addr_end(addr, end);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700562 if (!pmd_present(*pmd))
563 continue;
564 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
Naoya Horiguchi98094942013-09-11 14:22:14 -0700565 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700566 flags, private);
567 continue;
568 }
Kirill A. Shutemove1803772012-12-12 13:50:59 -0800569 split_huge_page_pmd(vma, addr, pmd);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -0700570 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
Hugh Dickins91612e02005-06-21 17:15:07 -0700571 continue;
Naoya Horiguchi98094942013-09-11 14:22:14 -0700572 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800573 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700574 return -EIO;
575 } while (pmd++, addr = next, addr != end);
576 return 0;
577}
578
Naoya Horiguchi98094942013-09-11 14:22:14 -0700579static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800580 unsigned long addr, unsigned long end,
581 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800582 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700583{
584 pud_t *pud;
585 unsigned long next;
586
587 pud = pud_offset(pgd, addr);
588 do {
589 next = pud_addr_end(addr, end);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700590 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
591 continue;
Hugh Dickins91612e02005-06-21 17:15:07 -0700592 if (pud_none_or_clear_bad(pud))
593 continue;
Naoya Horiguchi98094942013-09-11 14:22:14 -0700594 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800595 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700596 return -EIO;
597 } while (pud++, addr = next, addr != end);
598 return 0;
599}
600
Naoya Horiguchi98094942013-09-11 14:22:14 -0700601static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800602 unsigned long addr, unsigned long end,
603 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800604 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700605{
606 pgd_t *pgd;
607 unsigned long next;
608
Nick Pigginb5810032005-10-29 18:16:12 -0700609 pgd = pgd_offset(vma->vm_mm, addr);
Hugh Dickins91612e02005-06-21 17:15:07 -0700610 do {
611 next = pgd_addr_end(addr, end);
612 if (pgd_none_or_clear_bad(pgd))
613 continue;
Naoya Horiguchi98094942013-09-11 14:22:14 -0700614 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800615 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700616 return -EIO;
617 } while (pgd++, addr = next, addr != end);
618 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619}
620
Aneesh Kumar K.V58772312013-12-06 00:08:22 +0530621#ifdef CONFIG_NUMA_BALANCING
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200622/*
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200623 * This is used to mark a range of virtual addresses to be inaccessible.
624 * These are later cleared by a NUMA hinting fault. Depending on these
625 * faults, pages may be migrated for better NUMA placement.
626 *
627 * This is assuming that NUMA faults are handled using PROT_NONE. If
628 * an architecture makes a different choice, it will need further
629 * changes to the core.
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200630 */
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200631unsigned long change_prot_numa(struct vm_area_struct *vma,
632 unsigned long addr, unsigned long end)
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200633{
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200634 int nr_updated;
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200635
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200636 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
Mel Gorman03c5a6e2012-11-02 14:52:48 +0000637 if (nr_updated)
638 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200639
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200640 return nr_updated;
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200641}
642#else
643static unsigned long change_prot_numa(struct vm_area_struct *vma,
644 unsigned long addr, unsigned long end)
645{
646 return 0;
647}
Aneesh Kumar K.V58772312013-12-06 00:08:22 +0530648#endif /* CONFIG_NUMA_BALANCING */
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200649
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800650/*
Naoya Horiguchi98094942013-09-11 14:22:14 -0700651 * Walk through page tables and collect pages to be migrated.
652 *
653 * If pages found in a given range are on a set of nodes (determined by
654 * @nodes and @flags,) it's isolated and queued to the pagelist which is
655 * passed via @private.)
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800656 */
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700657static int
Naoya Horiguchi98094942013-09-11 14:22:14 -0700658queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Christoph Lameter38e35862006-01-08 01:01:01 -0800659 const nodemask_t *nodes, unsigned long flags, void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660{
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700661 int err = 0;
662 struct vm_area_struct *vma, *prev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700664 vma = find_vma(mm, start);
665 if (!vma)
666 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 prev = NULL;
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700668 for (; vma && vma->vm_start < end; vma = vma->vm_next) {
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200669 unsigned long endvma = vma->vm_end;
670
671 if (endvma > end)
672 endvma = end;
673 if (vma->vm_start > start)
674 start = vma->vm_start;
675
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800676 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
677 if (!vma->vm_next && vma->vm_end < end)
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700678 return -EFAULT;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800679 if (prev && prev->vm_end < vma->vm_start)
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700680 return -EFAULT;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800681 }
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800682
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200683 if (flags & MPOL_MF_LAZY) {
684 change_prot_numa(vma, start, endvma);
685 goto next;
686 }
687
688 if ((flags & MPOL_MF_STRICT) ||
689 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
690 vma_migratable(vma))) {
691
Naoya Horiguchi98094942013-09-11 14:22:14 -0700692 err = queue_pages_pgd_range(vma, start, endvma, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800693 flags, private);
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700694 if (err)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 }
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200697next:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 prev = vma;
699 }
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700700 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701}
702
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700703/*
704 * Apply policy to a single VMA
705 * This must be called with the mmap_sem held for writing.
706 */
707static int vma_replace_policy(struct vm_area_struct *vma,
708 struct mempolicy *pol)
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700709{
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700710 int err;
711 struct mempolicy *old;
712 struct mempolicy *new;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700713
714 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
715 vma->vm_start, vma->vm_end, vma->vm_pgoff,
716 vma->vm_ops, vma->vm_file,
717 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
718
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700719 new = mpol_dup(pol);
720 if (IS_ERR(new))
721 return PTR_ERR(new);
722
723 if (vma->vm_ops && vma->vm_ops->set_policy) {
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700724 err = vma->vm_ops->set_policy(vma, new);
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700725 if (err)
726 goto err_out;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700727 }
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700728
729 old = vma->vm_policy;
730 vma->vm_policy = new; /* protected by mmap_sem */
731 mpol_put(old);
732
733 return 0;
734 err_out:
735 mpol_put(new);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700736 return err;
737}
738
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800740static int mbind_range(struct mm_struct *mm, unsigned long start,
741 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742{
743 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800744 struct vm_area_struct *prev;
745 struct vm_area_struct *vma;
746 int err = 0;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800747 pgoff_t pgoff;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800748 unsigned long vmstart;
749 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750
Linus Torvalds097d5912012-03-06 18:23:36 -0800751 vma = find_vma(mm, start);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800752 if (!vma || vma->vm_start > start)
753 return -EFAULT;
754
Linus Torvalds097d5912012-03-06 18:23:36 -0800755 prev = vma->vm_prev;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800756 if (start > vma->vm_start)
757 prev = vma;
758
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800759 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800761 vmstart = max(start, vma->vm_start);
762 vmend = min(end, vma->vm_end);
763
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800764 if (mpol_equal(vma_policy(vma), new_pol))
765 continue;
766
767 pgoff = vma->vm_pgoff +
768 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800769 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800770 vma->anon_vma, vma->vm_file, pgoff,
Caspar Zhang8aacc9f2011-09-14 16:20:58 -0700771 new_pol);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800772 if (prev) {
773 vma = prev;
774 next = vma->vm_next;
Oleg Nesterov3964acd2013-07-31 13:53:28 -0700775 if (mpol_equal(vma_policy(vma), new_pol))
776 continue;
777 /* vma_merge() joined vma && vma->next, case 8 */
778 goto replace;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800779 }
780 if (vma->vm_start != vmstart) {
781 err = split_vma(vma->vm_mm, vma, vmstart, 1);
782 if (err)
783 goto out;
784 }
785 if (vma->vm_end != vmend) {
786 err = split_vma(vma->vm_mm, vma, vmend, 0);
787 if (err)
788 goto out;
789 }
Oleg Nesterov3964acd2013-07-31 13:53:28 -0700790 replace:
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700791 err = vma_replace_policy(vma, new_pol);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700792 if (err)
793 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800795
796 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 return err;
798}
799
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700801static long do_set_mempolicy(unsigned short mode, unsigned short flags,
802 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803{
Miao Xie58568d22009-06-16 15:31:49 -0700804 struct mempolicy *new, *old;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700805 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700806 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700808 if (!scratch)
809 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700810
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700811 new = mpol_new(mode, flags, nodes);
812 if (IS_ERR(new)) {
813 ret = PTR_ERR(new);
814 goto out;
815 }
Oleg Nesterov2c7c3a72014-10-09 15:27:55 -0700816
Miao Xie58568d22009-06-16 15:31:49 -0700817 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700818 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700819 if (ret) {
820 task_unlock(current);
Miao Xie58568d22009-06-16 15:31:49 -0700821 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700822 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700823 }
824 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 current->mempolicy = new;
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700826 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700827 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700828 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700829 task_unlock(current);
Miao Xie58568d22009-06-16 15:31:49 -0700830 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700831 ret = 0;
832out:
833 NODEMASK_SCRATCH_FREE(scratch);
834 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835}
836
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700837/*
838 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700839 *
840 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700841 */
842static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700844 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700845 if (p == &default_policy)
846 return;
847
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700848 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700849 case MPOL_BIND:
850 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700852 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 break;
854 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700855 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700856 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700857 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858 break;
859 default:
860 BUG();
861 }
862}
863
864static int lookup_node(struct mm_struct *mm, unsigned long addr)
865{
866 struct page *p;
867 int err;
868
869 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
870 if (err >= 0) {
871 err = page_to_nid(p);
872 put_page(p);
873 }
874 return err;
875}
876
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700878static long do_get_mempolicy(int *policy, nodemask_t *nmask,
879 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700881 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 struct mm_struct *mm = current->mm;
883 struct vm_area_struct *vma = NULL;
884 struct mempolicy *pol = current->mempolicy;
885
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700886 if (flags &
887 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700889
890 if (flags & MPOL_F_MEMS_ALLOWED) {
891 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
892 return -EINVAL;
893 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700894 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700895 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700896 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700897 return 0;
898 }
899
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700901 /*
902 * Do NOT fall back to task policy if the
903 * vma/shared policy at addr is NULL. We
904 * want to return MPOL_DEFAULT in this case.
905 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 down_read(&mm->mmap_sem);
907 vma = find_vma_intersection(mm, addr, addr+1);
908 if (!vma) {
909 up_read(&mm->mmap_sem);
910 return -EFAULT;
911 }
912 if (vma->vm_ops && vma->vm_ops->get_policy)
913 pol = vma->vm_ops->get_policy(vma, addr);
914 else
915 pol = vma->vm_policy;
916 } else if (addr)
917 return -EINVAL;
918
919 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700920 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
922 if (flags & MPOL_F_NODE) {
923 if (flags & MPOL_F_ADDR) {
924 err = lookup_node(mm, addr);
925 if (err < 0)
926 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700927 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700929 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700930 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 } else {
932 err = -EINVAL;
933 goto out;
934 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700935 } else {
936 *policy = pol == &default_policy ? MPOL_DEFAULT :
937 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700938 /*
939 * Internal mempolicy flags must be masked off before exposing
940 * the policy to userspace.
941 */
942 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700943 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944
945 if (vma) {
946 up_read(&current->mm->mmap_sem);
947 vma = NULL;
948 }
949
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700951 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700952 if (mpol_store_user_nodemask(pol)) {
953 *nmask = pol->w.user_nodemask;
954 } else {
955 task_lock(current);
956 get_policy_nodemask(pol, nmask);
957 task_unlock(current);
958 }
Miao Xie58568d22009-06-16 15:31:49 -0700959 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960
961 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700962 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 if (vma)
964 up_read(&current->mm->mmap_sem);
965 return err;
966}
967
Christoph Lameterb20a3502006-03-22 00:09:12 -0800968#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700969/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800970 * page migration
971 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800972static void migrate_page_add(struct page *page, struct list_head *pagelist,
973 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800974{
975 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800976 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800977 */
Nick Piggin62695a82008-10-18 20:26:09 -0700978 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
979 if (!isolate_lru_page(page)) {
980 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800981 inc_zone_page_state(page, NR_ISOLATED_ANON +
982 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700983 }
984 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800985}
986
Christoph Lameter742755a2006-06-23 02:03:55 -0700987static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700988{
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700989 if (PageHuge(page))
990 return alloc_huge_page_node(page_hstate(compound_head(page)),
991 node);
992 else
993 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700994}
995
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800996/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800997 * Migrate pages from one node to a target node.
998 * Returns error or the number of pages not migrated.
999 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001000static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1001 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001002{
1003 nodemask_t nmask;
1004 LIST_HEAD(pagelist);
1005 int err = 0;
1006
1007 nodes_clear(nmask);
1008 node_set(source, nmask);
1009
Minchan Kim08270802012-10-08 16:33:38 -07001010 /*
1011 * This does not "check" the range but isolates all pages that
1012 * need migration. Between passing in the full user address
1013 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1014 */
1015 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
Naoya Horiguchi98094942013-09-11 14:22:14 -07001016 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001017 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1018
Minchan Kimcf608ac2010-10-26 14:21:29 -07001019 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001020 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001021 MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001022 if (err)
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -07001023 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001024 }
Christoph Lameter95a402c2006-06-23 02:03:53 -07001025
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001026 return err;
1027}
1028
1029/*
1030 * Move pages between the two nodesets so as to preserve the physical
1031 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -08001032 *
1033 * Returns the number of page that could not be moved.
1034 */
Andrew Morton0ce72d42012-05-29 15:06:24 -07001035int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1036 const nodemask_t *to, int flags)
Christoph Lameter39743882006-01-08 01:00:51 -08001037{
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001038 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001039 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001040 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -08001041
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001042 err = migrate_prep();
1043 if (err)
1044 return err;
1045
Lee Schermerhorn53f25562008-04-28 02:13:20 -07001046 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -08001047
Andrew Morton0ce72d42012-05-29 15:06:24 -07001048 err = migrate_vmas(mm, from, to, flags);
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001049 if (err)
1050 goto out;
1051
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001052 /*
1053 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1054 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1055 * bit in 'tmp', and return that <source, dest> pair for migration.
1056 * The pair of nodemasks 'to' and 'from' define the map.
1057 *
1058 * If no pair of bits is found that way, fallback to picking some
1059 * pair of 'source' and 'dest' bits that are not the same. If the
1060 * 'source' and 'dest' bits are the same, this represents a node
1061 * that will be migrating to itself, so no pages need move.
1062 *
1063 * If no bits are left in 'tmp', or if all remaining bits left
1064 * in 'tmp' correspond to the same bit in 'to', return false
1065 * (nothing left to migrate).
1066 *
1067 * This lets us pick a pair of nodes to migrate between, such that
1068 * if possible the dest node is not already occupied by some other
1069 * source node, minimizing the risk of overloading the memory on a
1070 * node that would happen if we migrated incoming memory to a node
1071 * before migrating outgoing memory source that same node.
1072 *
1073 * A single scan of tmp is sufficient. As we go, we remember the
1074 * most recent <s, d> pair that moved (s != d). If we find a pair
1075 * that not only moved, but what's better, moved to an empty slot
1076 * (d is not set in tmp), then we break out then, with that pair.
Justin P. Mattockae0e47f2011-03-01 15:06:02 +01001077 * Otherwise when we finish scanning from_tmp, we at least have the
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001078 * most recent <s, d> pair that moved. If we get all the way through
1079 * the scan of tmp without finding any node that moved, much less
1080 * moved to an empty node, then there is nothing left worth migrating.
1081 */
Christoph Lameterd4984712006-01-08 01:00:55 -08001082
Andrew Morton0ce72d42012-05-29 15:06:24 -07001083 tmp = *from;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001084 while (!nodes_empty(tmp)) {
1085 int s,d;
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001086 int source = NUMA_NO_NODE;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001087 int dest = 0;
1088
1089 for_each_node_mask(s, tmp) {
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001090
1091 /*
1092 * do_migrate_pages() tries to maintain the relative
1093 * node relationship of the pages established between
1094 * threads and memory areas.
1095 *
1096 * However if the number of source nodes is not equal to
1097 * the number of destination nodes we can not preserve
1098 * this node relative relationship. In that case, skip
1099 * copying memory from a node that is in the destination
1100 * mask.
1101 *
1102 * Example: [2,3,4] -> [3,4,5] moves everything.
1103 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1104 */
1105
Andrew Morton0ce72d42012-05-29 15:06:24 -07001106 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1107 (node_isset(s, *to)))
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001108 continue;
1109
Andrew Morton0ce72d42012-05-29 15:06:24 -07001110 d = node_remap(s, *from, *to);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001111 if (s == d)
1112 continue;
1113
1114 source = s; /* Node moved. Memorize */
1115 dest = d;
1116
1117 /* dest not in remaining from nodes? */
1118 if (!node_isset(dest, tmp))
1119 break;
1120 }
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001121 if (source == NUMA_NO_NODE)
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001122 break;
1123
1124 node_clear(source, tmp);
1125 err = migrate_to_node(mm, source, dest, flags);
1126 if (err > 0)
1127 busy += err;
1128 if (err < 0)
1129 break;
Christoph Lameter39743882006-01-08 01:00:51 -08001130 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001131out:
Christoph Lameter39743882006-01-08 01:00:51 -08001132 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001133 if (err < 0)
1134 return err;
1135 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001136
Christoph Lameter39743882006-01-08 01:00:51 -08001137}
1138
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001139/*
1140 * Allocate a new page for page migration based on vma policy.
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001141 * Start by assuming the page is mapped by the same vma as contains @start.
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001142 * Search forward from there, if not. N.B., this assumes that the
1143 * list of pages handed to migrate_pages()--which is how we get here--
1144 * is in virtual address order.
1145 */
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001146static struct page *new_page(struct page *page, unsigned long start, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001147{
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001148 struct vm_area_struct *vma;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001149 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001150
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001151 vma = find_vma(current->mm, start);
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001152 while (vma) {
1153 address = page_address_in_vma(page, vma);
1154 if (address != -EFAULT)
1155 break;
1156 vma = vma->vm_next;
1157 }
1158
Wanpeng Li11c731e2013-12-18 17:08:56 -08001159 if (PageHuge(page)) {
Michal Hockocc817172014-01-23 15:53:15 -08001160 BUG_ON(!vma);
1161 return alloc_huge_page_noerr(vma, address, 1);
Wanpeng Li11c731e2013-12-18 17:08:56 -08001162 }
1163 /*
1164 * if !vma, alloc_page_vma() will use task or system default policy
1165 */
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001166 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001167}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001168#else
1169
1170static void migrate_page_add(struct page *page, struct list_head *pagelist,
1171 unsigned long flags)
1172{
1173}
1174
Andrew Morton0ce72d42012-05-29 15:06:24 -07001175int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1176 const nodemask_t *to, int flags)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001177{
1178 return -ENOSYS;
1179}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001180
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001181static struct page *new_page(struct page *page, unsigned long start, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001182{
1183 return NULL;
1184}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001185#endif
1186
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001187static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001188 unsigned short mode, unsigned short mode_flags,
1189 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001190{
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001191 struct mm_struct *mm = current->mm;
1192 struct mempolicy *new;
1193 unsigned long end;
1194 int err;
1195 LIST_HEAD(pagelist);
1196
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001197 if (flags & ~(unsigned long)MPOL_MF_VALID)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001198 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001199 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001200 return -EPERM;
1201
1202 if (start & ~PAGE_MASK)
1203 return -EINVAL;
1204
1205 if (mode == MPOL_DEFAULT)
1206 flags &= ~MPOL_MF_STRICT;
1207
1208 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1209 end = start + len;
1210
1211 if (end < start)
1212 return -EINVAL;
1213 if (end == start)
1214 return 0;
1215
David Rientjes028fec42008-04-28 02:12:25 -07001216 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001217 if (IS_ERR(new))
1218 return PTR_ERR(new);
1219
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001220 if (flags & MPOL_MF_LAZY)
1221 new->flags |= MPOL_F_MOF;
1222
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001223 /*
1224 * If we are using the default policy then operation
1225 * on discontinuous address spaces is okay after all
1226 */
1227 if (!new)
1228 flags |= MPOL_MF_DISCONTIG_OK;
1229
David Rientjes028fec42008-04-28 02:12:25 -07001230 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1231 start, start + len, mode, mode_flags,
David Rientjes00ef2d22013-02-22 16:35:36 -08001232 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001233
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001234 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1235
1236 err = migrate_prep();
1237 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001238 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001239 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001240 {
1241 NODEMASK_SCRATCH(scratch);
1242 if (scratch) {
1243 down_write(&mm->mmap_sem);
1244 task_lock(current);
1245 err = mpol_set_nodemask(new, nmask, scratch);
1246 task_unlock(current);
1247 if (err)
1248 up_write(&mm->mmap_sem);
1249 } else
1250 err = -ENOMEM;
1251 NODEMASK_SCRATCH_FREE(scratch);
1252 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001253 if (err)
1254 goto mpol_out;
1255
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001256 err = queue_pages_range(mm, start, end, nmask,
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001257 flags | MPOL_MF_INVERT, &pagelist);
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001258 if (!err)
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001259 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001260
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001261 if (!err) {
1262 int nr_failed = 0;
1263
Minchan Kimcf608ac2010-10-26 14:21:29 -07001264 if (!list_empty(&pagelist)) {
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001265 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001266 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1267 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001268 if (nr_failed)
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001269 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001270 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001271
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001272 if (nr_failed && (flags & MPOL_MF_STRICT))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001273 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001274 } else
Joonsoo Kimb0e5fd72013-12-18 17:08:51 -08001275 putback_movable_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001276
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001277 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001278 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001279 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001280 return err;
1281}
1282
Christoph Lameter39743882006-01-08 01:00:51 -08001283/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001284 * User space interface with variable sized bitmaps for nodelists.
1285 */
1286
1287/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001288static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001289 unsigned long maxnode)
1290{
1291 unsigned long k;
1292 unsigned long nlongs;
1293 unsigned long endmask;
1294
1295 --maxnode;
1296 nodes_clear(*nodes);
1297 if (maxnode == 0 || !nmask)
1298 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001299 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001300 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001301
1302 nlongs = BITS_TO_LONGS(maxnode);
1303 if ((maxnode % BITS_PER_LONG) == 0)
1304 endmask = ~0UL;
1305 else
1306 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1307
1308 /* When the user specified more nodes than supported just check
1309 if the non supported part is all zero. */
1310 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1311 if (nlongs > PAGE_SIZE/sizeof(long))
1312 return -EINVAL;
1313 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1314 unsigned long t;
1315 if (get_user(t, nmask + k))
1316 return -EFAULT;
1317 if (k == nlongs - 1) {
1318 if (t & endmask)
1319 return -EINVAL;
1320 } else if (t)
1321 return -EINVAL;
1322 }
1323 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1324 endmask = ~0UL;
1325 }
1326
1327 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1328 return -EFAULT;
1329 nodes_addr(*nodes)[nlongs-1] &= endmask;
1330 return 0;
1331}
1332
1333/* Copy a kernel node mask to user space */
1334static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1335 nodemask_t *nodes)
1336{
1337 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1338 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1339
1340 if (copy > nbytes) {
1341 if (copy > PAGE_SIZE)
1342 return -EINVAL;
1343 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1344 return -EFAULT;
1345 copy = nbytes;
1346 }
1347 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1348}
1349
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001350SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
Rasmus Villemoesf7f28ca2014-06-04 16:07:57 -07001351 unsigned long, mode, const unsigned long __user *, nmask,
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001352 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001353{
1354 nodemask_t nodes;
1355 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001356 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001357
David Rientjes028fec42008-04-28 02:12:25 -07001358 mode_flags = mode & MPOL_MODE_FLAGS;
1359 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001360 if (mode >= MPOL_MAX)
1361 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001362 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1363 (mode_flags & MPOL_F_RELATIVE_NODES))
1364 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001365 err = get_nodes(&nodes, nmask, maxnode);
1366 if (err)
1367 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001368 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001369}
1370
1371/* Set the process memory policy */
Rasmus Villemoes23c89022014-06-04 16:07:58 -07001372SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001373 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001374{
1375 int err;
1376 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001377 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001378
David Rientjes028fec42008-04-28 02:12:25 -07001379 flags = mode & MPOL_MODE_FLAGS;
1380 mode &= ~MPOL_MODE_FLAGS;
1381 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001382 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001383 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1384 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001385 err = get_nodes(&nodes, nmask, maxnode);
1386 if (err)
1387 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001388 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001389}
1390
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001391SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1392 const unsigned long __user *, old_nodes,
1393 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001394{
David Howellsc69e8d92008-11-14 10:39:19 +11001395 const struct cred *cred = current_cred(), *tcred;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001396 struct mm_struct *mm = NULL;
Christoph Lameter39743882006-01-08 01:00:51 -08001397 struct task_struct *task;
Christoph Lameter39743882006-01-08 01:00:51 -08001398 nodemask_t task_nodes;
1399 int err;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001400 nodemask_t *old;
1401 nodemask_t *new;
1402 NODEMASK_SCRATCH(scratch);
Christoph Lameter39743882006-01-08 01:00:51 -08001403
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001404 if (!scratch)
1405 return -ENOMEM;
Christoph Lameter39743882006-01-08 01:00:51 -08001406
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001407 old = &scratch->mask1;
1408 new = &scratch->mask2;
1409
1410 err = get_nodes(old, old_nodes, maxnode);
Christoph Lameter39743882006-01-08 01:00:51 -08001411 if (err)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001412 goto out;
1413
1414 err = get_nodes(new, new_nodes, maxnode);
1415 if (err)
1416 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001417
1418 /* Find the mm_struct */
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001419 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001420 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001421 if (!task) {
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001422 rcu_read_unlock();
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001423 err = -ESRCH;
1424 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001425 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001426 get_task_struct(task);
Christoph Lameter39743882006-01-08 01:00:51 -08001427
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001428 err = -EINVAL;
Christoph Lameter39743882006-01-08 01:00:51 -08001429
1430 /*
1431 * Check if this process has the right to modify the specified
1432 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001433 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001434 * userid as the target process.
1435 */
David Howellsc69e8d92008-11-14 10:39:19 +11001436 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001437 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1438 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001439 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001440 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001441 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001442 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001443 }
David Howellsc69e8d92008-11-14 10:39:19 +11001444 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001445
1446 task_nodes = cpuset_mems_allowed(task);
1447 /* Is the user allowed to access the target nodes? */
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001448 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001449 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001450 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001451 }
1452
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08001453 if (!nodes_subset(*new, node_states[N_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001454 err = -EINVAL;
Christoph Lameter3268c632012-03-21 16:34:06 -07001455 goto out_put;
Christoph Lameter3b42d282007-08-31 00:12:08 -07001456 }
1457
David Quigley86c3a762006-06-23 02:04:02 -07001458 err = security_task_movememory(task);
1459 if (err)
Christoph Lameter3268c632012-03-21 16:34:06 -07001460 goto out_put;
David Quigley86c3a762006-06-23 02:04:02 -07001461
Christoph Lameter3268c632012-03-21 16:34:06 -07001462 mm = get_task_mm(task);
1463 put_task_struct(task);
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001464
1465 if (!mm) {
Christoph Lameter3268c632012-03-21 16:34:06 -07001466 err = -EINVAL;
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001467 goto out;
1468 }
1469
1470 err = do_migrate_pages(mm, old, new,
1471 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter3268c632012-03-21 16:34:06 -07001472
1473 mmput(mm);
1474out:
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001475 NODEMASK_SCRATCH_FREE(scratch);
1476
Christoph Lameter39743882006-01-08 01:00:51 -08001477 return err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001478
1479out_put:
1480 put_task_struct(task);
1481 goto out;
1482
Christoph Lameter39743882006-01-08 01:00:51 -08001483}
1484
1485
Christoph Lameter8bccd852005-10-29 18:16:59 -07001486/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001487SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1488 unsigned long __user *, nmask, unsigned long, maxnode,
1489 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001490{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001491 int err;
1492 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001493 nodemask_t nodes;
1494
1495 if (nmask != NULL && maxnode < MAX_NUMNODES)
1496 return -EINVAL;
1497
1498 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1499
1500 if (err)
1501 return err;
1502
1503 if (policy && put_user(pval, policy))
1504 return -EFAULT;
1505
1506 if (nmask)
1507 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1508
1509 return err;
1510}
1511
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512#ifdef CONFIG_COMPAT
1513
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001514COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1515 compat_ulong_t __user *, nmask,
1516 compat_ulong_t, maxnode,
1517 compat_ulong_t, addr, compat_ulong_t, flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518{
1519 long err;
1520 unsigned long __user *nm = NULL;
1521 unsigned long nr_bits, alloc_size;
1522 DECLARE_BITMAP(bm, MAX_NUMNODES);
1523
1524 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1525 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1526
1527 if (nmask)
1528 nm = compat_alloc_user_space(alloc_size);
1529
1530 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1531
1532 if (!err && nmask) {
KAMEZAWA Hiroyuki2bbff6c2011-09-14 16:21:02 -07001533 unsigned long copy_size;
1534 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1535 err = copy_from_user(bm, nm, copy_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536 /* ensure entire bitmap is zeroed */
1537 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1538 err |= compat_put_bitmap(nmask, bm, nr_bits);
1539 }
1540
1541 return err;
1542}
1543
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001544COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1545 compat_ulong_t, maxnode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546{
1547 long err = 0;
1548 unsigned long __user *nm = NULL;
1549 unsigned long nr_bits, alloc_size;
1550 DECLARE_BITMAP(bm, MAX_NUMNODES);
1551
1552 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1553 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1554
1555 if (nmask) {
1556 err = compat_get_bitmap(bm, nmask, nr_bits);
1557 nm = compat_alloc_user_space(alloc_size);
1558 err |= copy_to_user(nm, bm, alloc_size);
1559 }
1560
1561 if (err)
1562 return -EFAULT;
1563
1564 return sys_set_mempolicy(mode, nm, nr_bits+1);
1565}
1566
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001567COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1568 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1569 compat_ulong_t, maxnode, compat_ulong_t, flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570{
1571 long err = 0;
1572 unsigned long __user *nm = NULL;
1573 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001574 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575
1576 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1577 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1578
1579 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001580 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001582 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 }
1584
1585 if (err)
1586 return -EFAULT;
1587
1588 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1589}
1590
1591#endif
1592
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001593struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1594 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595{
Oleg Nesterov8d902742014-10-09 15:27:45 -07001596 struct mempolicy *pol = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597
1598 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001599 if (vma->vm_ops && vma->vm_ops->get_policy) {
Oleg Nesterov8d902742014-10-09 15:27:45 -07001600 pol = vma->vm_ops->get_policy(vma, addr);
Mel Gorman00442ad2012-10-08 16:29:20 -07001601 } else if (vma->vm_policy) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 pol = vma->vm_policy;
Mel Gorman00442ad2012-10-08 16:29:20 -07001603
1604 /*
1605 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1606 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1607 * count on these policies which will be dropped by
1608 * mpol_cond_put() later
1609 */
1610 if (mpol_needs_cond_ref(pol))
1611 mpol_get(pol);
1612 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 }
Oleg Nesterovf15ca782014-10-09 15:27:43 -07001614
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001615 return pol;
1616}
1617
1618/*
1619 * get_vma_policy(@task, @vma, @addr)
1620 * @task: task for fallback if vma policy == default
1621 * @vma: virtual memory area whose policy is sought
1622 * @addr: address in @vma for shared policy lookup
1623 *
1624 * Returns effective policy for a VMA at specified address.
1625 * Falls back to @task or system default policy, as necessary.
1626 * Current or other task's task mempolicy and non-shared vma policies must be
1627 * protected by task_lock(task) by the caller.
1628 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1629 * count--added by the get_policy() vm_op, as appropriate--to protect against
1630 * freeing by another task. It is the caller's responsibility to free the
1631 * extra reference for shared policies.
1632 */
1633struct mempolicy *get_vma_policy(struct task_struct *task,
1634 struct vm_area_struct *vma, unsigned long addr)
1635{
1636 struct mempolicy *pol = __get_vma_policy(vma, addr);
1637
Oleg Nesterov8d902742014-10-09 15:27:45 -07001638 if (!pol)
1639 pol = get_task_policy(task);
1640
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 return pol;
1642}
1643
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001644bool vma_policy_mof(struct vm_area_struct *vma)
Mel Gormanfc3147242013-10-07 11:29:09 +01001645{
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001646 struct mempolicy *pol;
Oleg Nesterovf15ca782014-10-09 15:27:43 -07001647
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001648 if (vma->vm_ops && vma->vm_ops->get_policy) {
1649 bool ret = false;
Mel Gormanfc3147242013-10-07 11:29:09 +01001650
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001651 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1652 if (pol && (pol->flags & MPOL_F_MOF))
1653 ret = true;
1654 mpol_cond_put(pol);
Mel Gormanfc3147242013-10-07 11:29:09 +01001655
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001656 return ret;
Mel Gormanfc3147242013-10-07 11:29:09 +01001657 }
1658
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001659 pol = vma->vm_policy;
Oleg Nesterov8d902742014-10-09 15:27:45 -07001660 if (!pol)
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001661 pol = get_task_policy(current);
Oleg Nesterov8d902742014-10-09 15:27:45 -07001662
Mel Gormanfc3147242013-10-07 11:29:09 +01001663 return pol->flags & MPOL_F_MOF;
1664}
1665
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001666static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1667{
1668 enum zone_type dynamic_policy_zone = policy_zone;
1669
1670 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1671
1672 /*
1673 * if policy->v.nodes has movable memory only,
1674 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1675 *
1676 * policy->v.nodes is intersect with node_states[N_MEMORY].
1677 * so if the following test faile, it implies
1678 * policy->v.nodes has movable memory only.
1679 */
1680 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1681 dynamic_policy_zone = ZONE_MOVABLE;
1682
1683 return zone >= dynamic_policy_zone;
1684}
1685
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001686/*
1687 * Return a nodemask representing a mempolicy for filtering nodes for
1688 * page allocation
1689 */
1690static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001691{
1692 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001693 if (unlikely(policy->mode == MPOL_BIND) &&
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001694 apply_policy_zone(policy, gfp_zone(gfp)) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001695 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1696 return &policy->v.nodes;
1697
1698 return NULL;
1699}
1700
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001701/* Return a zonelist indicated by gfp for node representing a mempolicy */
Andi Kleen2f5f9482011-03-04 17:36:29 -08001702static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1703 int nd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704{
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001705 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001707 if (!(policy->flags & MPOL_F_LOCAL))
1708 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 break;
1710 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001711 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001712 * Normally, MPOL_BIND allocations are node-local within the
1713 * allowed nodemask. However, if __GFP_THISNODE is set and the
Bob Liu6eb27e12010-05-24 14:32:00 -07001714 * current node isn't part of the mask, we use the zonelist for
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001715 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001716 */
Mel Gorman19770b32008-04-28 02:12:18 -07001717 if (unlikely(gfp & __GFP_THISNODE) &&
1718 unlikely(!node_isset(nd, policy->v.nodes)))
1719 nd = first_node(policy->v.nodes);
1720 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722 BUG();
1723 }
Mel Gorman0e884602008-04-28 02:12:14 -07001724 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725}
1726
1727/* Do dynamic interleaving for a process */
1728static unsigned interleave_nodes(struct mempolicy *policy)
1729{
1730 unsigned nid, next;
1731 struct task_struct *me = current;
1732
1733 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001734 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001736 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001737 if (next < MAX_NUMNODES)
1738 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739 return nid;
1740}
1741
Christoph Lameterdc85da12006-01-18 17:42:36 -08001742/*
1743 * Depending on the memory policy provide a node from which to allocate the
1744 * next slab entry.
1745 */
David Rientjes2a389612014-04-07 15:37:29 -07001746unsigned int mempolicy_slab_node(void)
Christoph Lameterdc85da12006-01-18 17:42:36 -08001747{
Andi Kleene7b691b2012-06-09 02:40:03 -07001748 struct mempolicy *policy;
David Rientjes2a389612014-04-07 15:37:29 -07001749 int node = numa_mem_id();
Andi Kleene7b691b2012-06-09 02:40:03 -07001750
1751 if (in_interrupt())
David Rientjes2a389612014-04-07 15:37:29 -07001752 return node;
Andi Kleene7b691b2012-06-09 02:40:03 -07001753
1754 policy = current->mempolicy;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001755 if (!policy || policy->flags & MPOL_F_LOCAL)
David Rientjes2a389612014-04-07 15:37:29 -07001756 return node;
Christoph Lameter765c4502006-09-27 01:50:08 -07001757
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001758 switch (policy->mode) {
1759 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001760 /*
1761 * handled MPOL_F_LOCAL above
1762 */
1763 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001764
Christoph Lameterdc85da12006-01-18 17:42:36 -08001765 case MPOL_INTERLEAVE:
1766 return interleave_nodes(policy);
1767
Mel Gormandd1a2392008-04-28 02:12:17 -07001768 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001769 /*
1770 * Follow bind policy behavior and start allocation at the
1771 * first node.
1772 */
Mel Gorman19770b32008-04-28 02:12:18 -07001773 struct zonelist *zonelist;
1774 struct zone *zone;
1775 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
David Rientjes2a389612014-04-07 15:37:29 -07001776 zonelist = &NODE_DATA(node)->node_zonelists[0];
Mel Gorman19770b32008-04-28 02:12:18 -07001777 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1778 &policy->v.nodes,
1779 &zone);
David Rientjes2a389612014-04-07 15:37:29 -07001780 return zone ? zone->node : node;
Mel Gormandd1a2392008-04-28 02:12:17 -07001781 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001782
Christoph Lameterdc85da12006-01-18 17:42:36 -08001783 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001784 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001785 }
1786}
1787
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788/* Do static interleaving for a VMA with known offset. */
1789static unsigned offset_il_node(struct mempolicy *pol,
1790 struct vm_area_struct *vma, unsigned long off)
1791{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001792 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001793 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 int c;
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001795 int nid = NUMA_NO_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796
David Rientjesf5b087b2008-04-28 02:12:27 -07001797 if (!nnodes)
1798 return numa_node_id();
1799 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 c = 0;
1801 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001802 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 c++;
1804 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 return nid;
1806}
1807
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001808/* Determine a node number for interleave */
1809static inline unsigned interleave_nid(struct mempolicy *pol,
1810 struct vm_area_struct *vma, unsigned long addr, int shift)
1811{
1812 if (vma) {
1813 unsigned long off;
1814
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001815 /*
1816 * for small pages, there is no difference between
1817 * shift and PAGE_SHIFT, so the bit-shift is safe.
1818 * for huge pages, since vm_pgoff is in units of small
1819 * pages, we need to shift off the always 0 bits to get
1820 * a useful offset.
1821 */
1822 BUG_ON(shift < PAGE_SHIFT);
1823 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001824 off += (addr - vma->vm_start) >> shift;
1825 return offset_il_node(pol, vma, off);
1826 } else
1827 return interleave_nodes(pol);
1828}
1829
Michal Hocko778d3b02011-07-26 16:08:30 -07001830/*
1831 * Return the bit number of a random bit set in the nodemask.
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001832 * (returns NUMA_NO_NODE if nodemask is empty)
Michal Hocko778d3b02011-07-26 16:08:30 -07001833 */
1834int node_random(const nodemask_t *maskp)
1835{
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001836 int w, bit = NUMA_NO_NODE;
Michal Hocko778d3b02011-07-26 16:08:30 -07001837
1838 w = nodes_weight(*maskp);
1839 if (w)
1840 bit = bitmap_ord_to_pos(maskp->bits,
1841 get_random_int() % w, MAX_NUMNODES);
1842 return bit;
1843}
1844
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001845#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001846/*
1847 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
Fabian Frederickb46e14a2014-06-04 16:08:18 -07001848 * @vma: virtual memory area whose policy is sought
1849 * @addr: address in @vma for shared policy lookup and interleave policy
1850 * @gfp_flags: for requested zone
1851 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1852 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001853 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001854 * Returns a zonelist suitable for a huge page allocation and a pointer
1855 * to the struct mempolicy for conditional unref after allocation.
1856 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1857 * @nodemask for filtering the zonelist.
Miao Xiec0ff7452010-05-24 14:32:08 -07001858 *
Mel Gormand26914d2014-04-03 14:47:24 -07001859 * Must be protected by read_mems_allowed_begin()
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001860 */
Mel Gorman396faf02007-07-17 04:03:13 -07001861struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001862 gfp_t gfp_flags, struct mempolicy **mpol,
1863 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001864{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001865 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001866
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001867 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001868 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001869
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001870 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1871 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001872 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001873 } else {
Andi Kleen2f5f9482011-03-04 17:36:29 -08001874 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001875 if ((*mpol)->mode == MPOL_BIND)
1876 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001877 }
1878 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001879}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001880
1881/*
1882 * init_nodemask_of_mempolicy
1883 *
1884 * If the current task's mempolicy is "default" [NULL], return 'false'
1885 * to indicate default policy. Otherwise, extract the policy nodemask
1886 * for 'bind' or 'interleave' policy into the argument nodemask, or
1887 * initialize the argument nodemask to contain the single node for
1888 * 'preferred' or 'local' policy and return 'true' to indicate presence
1889 * of non-default mempolicy.
1890 *
1891 * We don't bother with reference counting the mempolicy [mpol_get/put]
1892 * because the current task is examining it's own mempolicy and a task's
1893 * mempolicy is only ever changed by the task itself.
1894 *
1895 * N.B., it is the caller's responsibility to free a returned nodemask.
1896 */
1897bool init_nodemask_of_mempolicy(nodemask_t *mask)
1898{
1899 struct mempolicy *mempolicy;
1900 int nid;
1901
1902 if (!(mask && current->mempolicy))
1903 return false;
1904
Miao Xiec0ff7452010-05-24 14:32:08 -07001905 task_lock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001906 mempolicy = current->mempolicy;
1907 switch (mempolicy->mode) {
1908 case MPOL_PREFERRED:
1909 if (mempolicy->flags & MPOL_F_LOCAL)
1910 nid = numa_node_id();
1911 else
1912 nid = mempolicy->v.preferred_node;
1913 init_nodemask_of_node(mask, nid);
1914 break;
1915
1916 case MPOL_BIND:
1917 /* Fall through */
1918 case MPOL_INTERLEAVE:
1919 *mask = mempolicy->v.nodes;
1920 break;
1921
1922 default:
1923 BUG();
1924 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001925 task_unlock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001926
1927 return true;
1928}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001929#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001930
David Rientjes6f48d0eb2010-08-09 17:18:52 -07001931/*
1932 * mempolicy_nodemask_intersects
1933 *
1934 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1935 * policy. Otherwise, check for intersection between mask and the policy
1936 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1937 * policy, always return true since it may allocate elsewhere on fallback.
1938 *
1939 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1940 */
1941bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1942 const nodemask_t *mask)
1943{
1944 struct mempolicy *mempolicy;
1945 bool ret = true;
1946
1947 if (!mask)
1948 return ret;
1949 task_lock(tsk);
1950 mempolicy = tsk->mempolicy;
1951 if (!mempolicy)
1952 goto out;
1953
1954 switch (mempolicy->mode) {
1955 case MPOL_PREFERRED:
1956 /*
1957 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1958 * allocate from, they may fallback to other nodes when oom.
1959 * Thus, it's possible for tsk to have allocated memory from
1960 * nodes in mask.
1961 */
1962 break;
1963 case MPOL_BIND:
1964 case MPOL_INTERLEAVE:
1965 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1966 break;
1967 default:
1968 BUG();
1969 }
1970out:
1971 task_unlock(tsk);
1972 return ret;
1973}
1974
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975/* Allocate a page in interleaved policy.
1976 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001977static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1978 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979{
1980 struct zonelist *zl;
1981 struct page *page;
1982
Mel Gorman0e884602008-04-28 02:12:14 -07001983 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001985 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001986 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987 return page;
1988}
1989
1990/**
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001991 * alloc_pages_vma - Allocate a page for a VMA.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992 *
1993 * @gfp:
1994 * %GFP_USER user allocation.
1995 * %GFP_KERNEL kernel allocations,
1996 * %GFP_HIGHMEM highmem/user allocations,
1997 * %GFP_FS allocation should not call back into a file system.
1998 * %GFP_ATOMIC don't sleep.
1999 *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002000 * @order:Order of the GFP allocation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 * @vma: Pointer to VMA or NULL if not available.
2002 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2003 *
2004 * This function allocates a page from the kernel page pool and applies
2005 * a NUMA policy associated with the VMA or the current process.
2006 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2007 * mm_struct of the VMA to prevent it from going away. Should be used for
2008 * all allocations for pages that will be mapped into
2009 * user space. Returns NULL when no page can be allocated.
2010 *
2011 * Should be called with the mm_sem of the vma hold.
2012 */
2013struct page *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002014alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
Andi Kleen2f5f9482011-03-04 17:36:29 -08002015 unsigned long addr, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016{
Mel Gormancc9a6c82012-03-21 16:34:11 -07002017 struct mempolicy *pol;
Miao Xiec0ff7452010-05-24 14:32:08 -07002018 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002019 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020
Mel Gormancc9a6c82012-03-21 16:34:11 -07002021retry_cpuset:
2022 pol = get_vma_policy(current, vma, addr);
Mel Gormand26914d2014-04-03 14:47:24 -07002023 cpuset_mems_cookie = read_mems_allowed_begin();
Mel Gormancc9a6c82012-03-21 16:34:11 -07002024
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002025 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08002027
Andi Kleen8eac5632011-02-25 14:44:28 -08002028 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002029 mpol_cond_put(pol);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002030 page = alloc_page_interleave(gfp, order, nid);
Mel Gormand26914d2014-04-03 14:47:24 -07002031 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07002032 goto retry_cpuset;
2033
Miao Xiec0ff7452010-05-24 14:32:08 -07002034 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035 }
David Rientjes212a0a62012-12-11 16:02:51 -08002036 page = __alloc_pages_nodemask(gfp, order,
2037 policy_zonelist(gfp, pol, node),
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002038 policy_nodemask(gfp, pol));
Oleg Nesterov23867402014-10-09 15:27:41 -07002039 mpol_cond_put(pol);
Mel Gormand26914d2014-04-03 14:47:24 -07002040 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07002041 goto retry_cpuset;
Miao Xiec0ff7452010-05-24 14:32:08 -07002042 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043}
2044
2045/**
2046 * alloc_pages_current - Allocate pages.
2047 *
2048 * @gfp:
2049 * %GFP_USER user allocation,
2050 * %GFP_KERNEL kernel allocation,
2051 * %GFP_HIGHMEM highmem allocation,
2052 * %GFP_FS don't call back into a file system.
2053 * %GFP_ATOMIC don't sleep.
2054 * @order: Power of two of allocation size in pages. 0 is a single page.
2055 *
2056 * Allocate a page from the kernel page pool. When not in
2057 * interrupt context and apply the current process NUMA policy.
2058 * Returns NULL when no page can be allocated.
2059 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08002060 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 * 1) it's ok to take cpuset_sem (can WAIT), and
2062 * 2) allocating for current task (not interrupt).
2063 */
Al Virodd0fc662005-10-07 07:46:04 +01002064struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065{
Oleg Nesterov8d902742014-10-09 15:27:45 -07002066 struct mempolicy *pol = &default_policy;
Miao Xiec0ff7452010-05-24 14:32:08 -07002067 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002068 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069
Oleg Nesterov8d902742014-10-09 15:27:45 -07002070 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2071 pol = get_task_policy(current);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002072
Mel Gormancc9a6c82012-03-21 16:34:11 -07002073retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07002074 cpuset_mems_cookie = read_mems_allowed_begin();
Mel Gormancc9a6c82012-03-21 16:34:11 -07002075
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002076 /*
2077 * No reference counting needed for current->mempolicy
2078 * nor system default_policy
2079 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002080 if (pol->mode == MPOL_INTERLEAVE)
Miao Xiec0ff7452010-05-24 14:32:08 -07002081 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2082 else
2083 page = __alloc_pages_nodemask(gfp, order,
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002084 policy_zonelist(gfp, pol, numa_node_id()),
2085 policy_nodemask(gfp, pol));
Mel Gormancc9a6c82012-03-21 16:34:11 -07002086
Mel Gormand26914d2014-04-03 14:47:24 -07002087 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07002088 goto retry_cpuset;
2089
Miao Xiec0ff7452010-05-24 14:32:08 -07002090 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091}
2092EXPORT_SYMBOL(alloc_pages_current);
2093
Oleg Nesterovef0855d2013-09-11 14:20:14 -07002094int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2095{
2096 struct mempolicy *pol = mpol_dup(vma_policy(src));
2097
2098 if (IS_ERR(pol))
2099 return PTR_ERR(pol);
2100 dst->vm_policy = pol;
2101 return 0;
2102}
2103
Paul Jackson42253992006-01-08 01:01:59 -08002104/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002105 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08002106 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2107 * with the mems_allowed returned by cpuset_mems_allowed(). This
2108 * keeps mempolicies cpuset relative after its cpuset moves. See
2109 * further kernel/cpuset.c update_nodemask().
Miao Xie708c1bb2010-05-24 14:32:07 -07002110 *
2111 * current's mempolicy may be rebinded by the other task(the task that changes
2112 * cpuset's mems), so we needn't do rebind work for current task.
Paul Jackson42253992006-01-08 01:01:59 -08002113 */
Paul Jackson42253992006-01-08 01:01:59 -08002114
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002115/* Slow path of a mempolicy duplicate */
2116struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117{
2118 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2119
2120 if (!new)
2121 return ERR_PTR(-ENOMEM);
Miao Xie708c1bb2010-05-24 14:32:07 -07002122
2123 /* task's mempolicy is protected by alloc_lock */
2124 if (old == current->mempolicy) {
2125 task_lock(current);
2126 *new = *old;
2127 task_unlock(current);
2128 } else
2129 *new = *old;
2130
Paul Jackson42253992006-01-08 01:01:59 -08002131 if (current_cpuset_is_being_rebound()) {
2132 nodemask_t mems = cpuset_mems_allowed(current);
Miao Xie708c1bb2010-05-24 14:32:07 -07002133 if (new->flags & MPOL_F_REBINDING)
2134 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2135 else
2136 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
Paul Jackson42253992006-01-08 01:01:59 -08002137 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 return new;
2140}
2141
2142/* Slow path of a mempolicy comparison */
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002143bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144{
2145 if (!a || !b)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002146 return false;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002147 if (a->mode != b->mode)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002148 return false;
Bob Liu19800502010-05-24 14:32:01 -07002149 if (a->flags != b->flags)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002150 return false;
Bob Liu19800502010-05-24 14:32:01 -07002151 if (mpol_store_user_nodemask(a))
2152 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002153 return false;
Bob Liu19800502010-05-24 14:32:01 -07002154
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002155 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07002156 case MPOL_BIND:
2157 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 case MPOL_INTERLEAVE:
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002159 return !!nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160 case MPOL_PREFERRED:
Namhyung Kim75719662011-03-22 16:33:02 -07002161 return a->v.preferred_node == b->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 default:
2163 BUG();
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002164 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 }
2166}
2167
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169 * Shared memory backing store policy support.
2170 *
2171 * Remember policies even when nobody has shared memory mapped.
2172 * The policies are kept in Red-Black tree linked from the inode.
2173 * They are protected by the sp->lock spinlock, which should be held
2174 * for any accesses to the tree.
2175 */
2176
2177/* lookup first element intersecting start-end */
Mel Gorman42288fe2012-12-21 23:10:25 +00002178/* Caller holds sp->lock */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179static struct sp_node *
2180sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2181{
2182 struct rb_node *n = sp->root.rb_node;
2183
2184 while (n) {
2185 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2186
2187 if (start >= p->end)
2188 n = n->rb_right;
2189 else if (end <= p->start)
2190 n = n->rb_left;
2191 else
2192 break;
2193 }
2194 if (!n)
2195 return NULL;
2196 for (;;) {
2197 struct sp_node *w = NULL;
2198 struct rb_node *prev = rb_prev(n);
2199 if (!prev)
2200 break;
2201 w = rb_entry(prev, struct sp_node, nd);
2202 if (w->end <= start)
2203 break;
2204 n = prev;
2205 }
2206 return rb_entry(n, struct sp_node, nd);
2207}
2208
2209/* Insert a new shared policy into the list. */
2210/* Caller holds sp->lock */
2211static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2212{
2213 struct rb_node **p = &sp->root.rb_node;
2214 struct rb_node *parent = NULL;
2215 struct sp_node *nd;
2216
2217 while (*p) {
2218 parent = *p;
2219 nd = rb_entry(parent, struct sp_node, nd);
2220 if (new->start < nd->start)
2221 p = &(*p)->rb_left;
2222 else if (new->end > nd->end)
2223 p = &(*p)->rb_right;
2224 else
2225 BUG();
2226 }
2227 rb_link_node(&new->nd, parent, p);
2228 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07002229 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002230 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231}
2232
2233/* Find shared policy intersecting idx */
2234struct mempolicy *
2235mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2236{
2237 struct mempolicy *pol = NULL;
2238 struct sp_node *sn;
2239
2240 if (!sp->root.rb_node)
2241 return NULL;
Mel Gorman42288fe2012-12-21 23:10:25 +00002242 spin_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 sn = sp_lookup(sp, idx, idx+1);
2244 if (sn) {
2245 mpol_get(sn->policy);
2246 pol = sn->policy;
2247 }
Mel Gorman42288fe2012-12-21 23:10:25 +00002248 spin_unlock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 return pol;
2250}
2251
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002252static void sp_free(struct sp_node *n)
2253{
2254 mpol_put(n->policy);
2255 kmem_cache_free(sn_cache, n);
2256}
2257
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002258/**
2259 * mpol_misplaced - check whether current page node is valid in policy
2260 *
Fabian Frederickb46e14a2014-06-04 16:08:18 -07002261 * @page: page to be checked
2262 * @vma: vm area where page mapped
2263 * @addr: virtual address where page mapped
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002264 *
2265 * Lookup current policy node id for vma,addr and "compare to" page's
2266 * node id.
2267 *
2268 * Returns:
2269 * -1 - not misplaced, page is in the right node
2270 * node - node id where the page should be
2271 *
2272 * Policy determination "mimics" alloc_page_vma().
2273 * Called from fault path where we know the vma and faulting address.
2274 */
2275int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2276{
2277 struct mempolicy *pol;
2278 struct zone *zone;
2279 int curnid = page_to_nid(page);
2280 unsigned long pgoff;
Peter Zijlstra90572892013-10-07 11:29:20 +01002281 int thiscpu = raw_smp_processor_id();
2282 int thisnid = cpu_to_node(thiscpu);
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002283 int polnid = -1;
2284 int ret = -1;
2285
2286 BUG_ON(!vma);
2287
2288 pol = get_vma_policy(current, vma, addr);
2289 if (!(pol->flags & MPOL_F_MOF))
2290 goto out;
2291
2292 switch (pol->mode) {
2293 case MPOL_INTERLEAVE:
2294 BUG_ON(addr >= vma->vm_end);
2295 BUG_ON(addr < vma->vm_start);
2296
2297 pgoff = vma->vm_pgoff;
2298 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2299 polnid = offset_il_node(pol, vma, pgoff);
2300 break;
2301
2302 case MPOL_PREFERRED:
2303 if (pol->flags & MPOL_F_LOCAL)
2304 polnid = numa_node_id();
2305 else
2306 polnid = pol->v.preferred_node;
2307 break;
2308
2309 case MPOL_BIND:
2310 /*
2311 * allows binding to multiple nodes.
2312 * use current page if in policy nodemask,
2313 * else select nearest allowed node, if any.
2314 * If no allowed nodes, use current [!misplaced].
2315 */
2316 if (node_isset(curnid, pol->v.nodes))
2317 goto out;
2318 (void)first_zones_zonelist(
2319 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2320 gfp_zone(GFP_HIGHUSER),
2321 &pol->v.nodes, &zone);
2322 polnid = zone->node;
2323 break;
2324
2325 default:
2326 BUG();
2327 }
Mel Gorman5606e382012-11-02 18:19:13 +00002328
2329 /* Migrate the page towards the node whose CPU is referencing it */
Mel Gormane42c8ff2012-11-12 09:17:07 +00002330 if (pol->flags & MPOL_F_MORON) {
Peter Zijlstra90572892013-10-07 11:29:20 +01002331 polnid = thisnid;
Mel Gorman5606e382012-11-02 18:19:13 +00002332
Rik van Riel10f39042014-01-27 17:03:44 -05002333 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
Rik van Rielde1c9ce2013-10-07 11:29:39 +01002334 goto out;
Mel Gormane42c8ff2012-11-12 09:17:07 +00002335 }
2336
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002337 if (curnid != polnid)
2338 ret = polnid;
2339out:
2340 mpol_cond_put(pol);
2341
2342 return ret;
2343}
2344
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2346{
Paul Mundt140d5a42007-07-15 23:38:16 -07002347 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 rb_erase(&n->nd, &sp->root);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002349 sp_free(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350}
2351
Mel Gorman42288fe2012-12-21 23:10:25 +00002352static void sp_node_init(struct sp_node *node, unsigned long start,
2353 unsigned long end, struct mempolicy *pol)
2354{
2355 node->start = start;
2356 node->end = end;
2357 node->policy = pol;
2358}
2359
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07002360static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2361 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362{
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002363 struct sp_node *n;
2364 struct mempolicy *newpol;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002366 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 if (!n)
2368 return NULL;
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002369
2370 newpol = mpol_dup(pol);
2371 if (IS_ERR(newpol)) {
2372 kmem_cache_free(sn_cache, n);
2373 return NULL;
2374 }
2375 newpol->flags |= MPOL_F_SHARED;
Mel Gorman42288fe2012-12-21 23:10:25 +00002376 sp_node_init(n, start, end, newpol);
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002377
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 return n;
2379}
2380
2381/* Replace a policy range. */
2382static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2383 unsigned long end, struct sp_node *new)
2384{
Mel Gormanb22d1272012-10-08 16:29:17 -07002385 struct sp_node *n;
Mel Gorman42288fe2012-12-21 23:10:25 +00002386 struct sp_node *n_new = NULL;
2387 struct mempolicy *mpol_new = NULL;
Mel Gormanb22d1272012-10-08 16:29:17 -07002388 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389
Mel Gorman42288fe2012-12-21 23:10:25 +00002390restart:
2391 spin_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 n = sp_lookup(sp, start, end);
2393 /* Take care of old policies in the same range. */
2394 while (n && n->start < end) {
2395 struct rb_node *next = rb_next(&n->nd);
2396 if (n->start >= start) {
2397 if (n->end <= end)
2398 sp_delete(sp, n);
2399 else
2400 n->start = end;
2401 } else {
2402 /* Old policy spanning whole new range. */
2403 if (n->end > end) {
Mel Gorman42288fe2012-12-21 23:10:25 +00002404 if (!n_new)
2405 goto alloc_new;
2406
2407 *mpol_new = *n->policy;
2408 atomic_set(&mpol_new->refcnt, 1);
KOSAKI Motohiro78806392013-03-08 12:43:29 -08002409 sp_node_init(n_new, end, n->end, mpol_new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 n->end = start;
Hillf Danton5ca39572013-03-08 12:43:28 -08002411 sp_insert(sp, n_new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002412 n_new = NULL;
2413 mpol_new = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 break;
2415 } else
2416 n->end = start;
2417 }
2418 if (!next)
2419 break;
2420 n = rb_entry(next, struct sp_node, nd);
2421 }
2422 if (new)
2423 sp_insert(sp, new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002424 spin_unlock(&sp->lock);
2425 ret = 0;
2426
2427err_out:
2428 if (mpol_new)
2429 mpol_put(mpol_new);
2430 if (n_new)
2431 kmem_cache_free(sn_cache, n_new);
2432
Mel Gormanb22d1272012-10-08 16:29:17 -07002433 return ret;
Mel Gorman42288fe2012-12-21 23:10:25 +00002434
2435alloc_new:
2436 spin_unlock(&sp->lock);
2437 ret = -ENOMEM;
2438 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2439 if (!n_new)
2440 goto err_out;
2441 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2442 if (!mpol_new)
2443 goto err_out;
2444 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445}
2446
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002447/**
2448 * mpol_shared_policy_init - initialize shared policy for inode
2449 * @sp: pointer to inode shared policy
2450 * @mpol: struct mempolicy to install
2451 *
2452 * Install non-NULL @mpol in inode's shared policy rb-tree.
2453 * On entry, the current task has a reference on a non-NULL @mpol.
2454 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002455 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002456 */
2457void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08002458{
Miao Xie58568d22009-06-16 15:31:49 -07002459 int ret;
2460
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002461 sp->root = RB_ROOT; /* empty tree == default mempolicy */
Mel Gorman42288fe2012-12-21 23:10:25 +00002462 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08002463
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002464 if (mpol) {
2465 struct vm_area_struct pvma;
2466 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002467 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002468
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002469 if (!scratch)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002470 goto put_mpol;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002471 /* contextualize the tmpfs mount point mempolicy */
2472 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002473 if (IS_ERR(new))
Dan Carpenter0cae3452010-05-25 23:42:58 -07002474 goto free_scratch; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002475
2476 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002477 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002478 task_unlock(current);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002479 if (ret)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002480 goto put_new;
Robin Holt7339ff82006-01-14 13:20:48 -08002481
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002482 /* Create pseudo-vma that contains just the policy */
2483 memset(&pvma, 0, sizeof(struct vm_area_struct));
2484 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2485 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002486
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002487put_new:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002488 mpol_put(new); /* drop initial ref */
Dan Carpenter0cae3452010-05-25 23:42:58 -07002489free_scratch:
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002490 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002491put_mpol:
2492 mpol_put(mpol); /* drop our incoming ref on sb mpol */
Robin Holt7339ff82006-01-14 13:20:48 -08002493 }
2494}
2495
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496int mpol_set_shared_policy(struct shared_policy *info,
2497 struct vm_area_struct *vma, struct mempolicy *npol)
2498{
2499 int err;
2500 struct sp_node *new = NULL;
2501 unsigned long sz = vma_pages(vma);
2502
David Rientjes028fec42008-04-28 02:12:25 -07002503 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002505 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002506 npol ? npol->flags : -1,
David Rientjes00ef2d22013-02-22 16:35:36 -08002507 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508
2509 if (npol) {
2510 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2511 if (!new)
2512 return -ENOMEM;
2513 }
2514 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2515 if (err && new)
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002516 sp_free(new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 return err;
2518}
2519
2520/* Free a backing policy store on inode delete. */
2521void mpol_free_shared_policy(struct shared_policy *p)
2522{
2523 struct sp_node *n;
2524 struct rb_node *next;
2525
2526 if (!p->root.rb_node)
2527 return;
Mel Gorman42288fe2012-12-21 23:10:25 +00002528 spin_lock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 next = rb_first(&p->root);
2530 while (next) {
2531 n = rb_entry(next, struct sp_node, nd);
2532 next = rb_next(&n->nd);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002533 sp_delete(p, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 }
Mel Gorman42288fe2012-12-21 23:10:25 +00002535 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536}
2537
Mel Gorman1a687c22012-11-22 11:16:36 +00002538#ifdef CONFIG_NUMA_BALANCING
Mel Gormanc2976632014-01-29 14:05:42 -08002539static int __initdata numabalancing_override;
Mel Gorman1a687c22012-11-22 11:16:36 +00002540
2541static void __init check_numabalancing_enable(void)
2542{
2543 bool numabalancing_default = false;
2544
2545 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2546 numabalancing_default = true;
2547
Mel Gormanc2976632014-01-29 14:05:42 -08002548 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2549 if (numabalancing_override)
2550 set_numabalancing_state(numabalancing_override == 1);
2551
Mel Gorman1a687c22012-11-22 11:16:36 +00002552 if (nr_node_ids > 1 && !numabalancing_override) {
Andrew Morton4a404be2014-01-29 14:05:43 -08002553 pr_info("%s automatic NUMA balancing. "
Mel Gormanc2976632014-01-29 14:05:42 -08002554 "Configure with numa_balancing= or the "
2555 "kernel.numa_balancing sysctl",
2556 numabalancing_default ? "Enabling" : "Disabling");
Mel Gorman1a687c22012-11-22 11:16:36 +00002557 set_numabalancing_state(numabalancing_default);
2558 }
2559}
2560
2561static int __init setup_numabalancing(char *str)
2562{
2563 int ret = 0;
2564 if (!str)
2565 goto out;
Mel Gorman1a687c22012-11-22 11:16:36 +00002566
2567 if (!strcmp(str, "enable")) {
Mel Gormanc2976632014-01-29 14:05:42 -08002568 numabalancing_override = 1;
Mel Gorman1a687c22012-11-22 11:16:36 +00002569 ret = 1;
2570 } else if (!strcmp(str, "disable")) {
Mel Gormanc2976632014-01-29 14:05:42 -08002571 numabalancing_override = -1;
Mel Gorman1a687c22012-11-22 11:16:36 +00002572 ret = 1;
2573 }
2574out:
2575 if (!ret)
Andrew Morton4a404be2014-01-29 14:05:43 -08002576 pr_warn("Unable to parse numa_balancing=\n");
Mel Gorman1a687c22012-11-22 11:16:36 +00002577
2578 return ret;
2579}
2580__setup("numa_balancing=", setup_numabalancing);
2581#else
2582static inline void __init check_numabalancing_enable(void)
2583{
2584}
2585#endif /* CONFIG_NUMA_BALANCING */
2586
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587/* assumes fs == KERNEL_DS */
2588void __init numa_policy_init(void)
2589{
Paul Mundtb71636e2007-07-15 23:38:15 -07002590 nodemask_t interleave_nodes;
2591 unsigned long largest = 0;
2592 int nid, prefer = 0;
2593
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 policy_cache = kmem_cache_create("numa_policy",
2595 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002596 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597
2598 sn_cache = kmem_cache_create("shared_policy_node",
2599 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002600 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601
Mel Gorman5606e382012-11-02 18:19:13 +00002602 for_each_node(nid) {
2603 preferred_node_policy[nid] = (struct mempolicy) {
2604 .refcnt = ATOMIC_INIT(1),
2605 .mode = MPOL_PREFERRED,
2606 .flags = MPOL_F_MOF | MPOL_F_MORON,
2607 .v = { .preferred_node = nid, },
2608 };
2609 }
2610
Paul Mundtb71636e2007-07-15 23:38:15 -07002611 /*
2612 * Set interleaving policy for system init. Interleaving is only
2613 * enabled across suitably sized nodes (default is >= 16MB), or
2614 * fall back to the largest node if they're all smaller.
2615 */
2616 nodes_clear(interleave_nodes);
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002617 for_each_node_state(nid, N_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002618 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619
Paul Mundtb71636e2007-07-15 23:38:15 -07002620 /* Preserve the largest node */
2621 if (largest < total_pages) {
2622 largest = total_pages;
2623 prefer = nid;
2624 }
2625
2626 /* Interleave this node? */
2627 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2628 node_set(nid, interleave_nodes);
2629 }
2630
2631 /* All too small, use the largest */
2632 if (unlikely(nodes_empty(interleave_nodes)))
2633 node_set(prefer, interleave_nodes);
2634
David Rientjes028fec42008-04-28 02:12:25 -07002635 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -07002636 pr_err("%s: interleaving failed\n", __func__);
Mel Gorman1a687c22012-11-22 11:16:36 +00002637
2638 check_numabalancing_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639}
2640
Christoph Lameter8bccd852005-10-29 18:16:59 -07002641/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642void numa_default_policy(void)
2643{
David Rientjes028fec42008-04-28 02:12:25 -07002644 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645}
Paul Jackson68860ec2005-10-30 15:02:36 -08002646
Paul Jackson42253992006-01-08 01:01:59 -08002647/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002648 * Parse and format mempolicy from/to strings
2649 */
2650
2651/*
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002652 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002653 */
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002654static const char * const policy_modes[] =
2655{
2656 [MPOL_DEFAULT] = "default",
2657 [MPOL_PREFERRED] = "prefer",
2658 [MPOL_BIND] = "bind",
2659 [MPOL_INTERLEAVE] = "interleave",
Lee Schermerhornd3a71032012-10-25 14:16:29 +02002660 [MPOL_LOCAL] = "local",
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002661};
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002662
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002663
2664#ifdef CONFIG_TMPFS
2665/**
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002666 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002667 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002668 * @mpol: pointer to struct mempolicy pointer, returned on success.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002669 *
2670 * Format of input:
2671 * <mode>[=<flags>][:<nodelist>]
2672 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002673 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002674 */
Hugh Dickinsa7a88b22013-01-02 02:04:23 -08002675int mpol_parse_str(char *str, struct mempolicy **mpol)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002676{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002677 struct mempolicy *new = NULL;
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002678 unsigned short mode;
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002679 unsigned short mode_flags;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002680 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002681 char *nodelist = strchr(str, ':');
2682 char *flags = strchr(str, '=');
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002683 int err = 1;
2684
2685 if (nodelist) {
2686 /* NUL-terminate mode or flags string */
2687 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002688 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002689 goto out;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002690 if (!nodes_subset(nodes, node_states[N_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002691 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002692 } else
2693 nodes_clear(nodes);
2694
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002695 if (flags)
2696 *flags++ = '\0'; /* terminate mode string */
2697
Peter Zijlstra479e2802012-10-25 14:16:28 +02002698 for (mode = 0; mode < MPOL_MAX; mode++) {
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002699 if (!strcmp(str, policy_modes[mode])) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002700 break;
2701 }
2702 }
Mel Gormana7200942012-11-16 09:37:58 +00002703 if (mode >= MPOL_MAX)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002704 goto out;
2705
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002706 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002707 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002708 /*
2709 * Insist on a nodelist of one node only
2710 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002711 if (nodelist) {
2712 char *rest = nodelist;
2713 while (isdigit(*rest))
2714 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002715 if (*rest)
2716 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002717 }
2718 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002719 case MPOL_INTERLEAVE:
2720 /*
2721 * Default to online nodes with memory if no nodelist
2722 */
2723 if (!nodelist)
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002724 nodes = node_states[N_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002725 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002726 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002727 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002728 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002729 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002730 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002731 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002732 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002733 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002734 case MPOL_DEFAULT:
2735 /*
2736 * Insist on a empty nodelist
2737 */
2738 if (!nodelist)
2739 err = 0;
2740 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002741 case MPOL_BIND:
2742 /*
2743 * Insist on a nodelist
2744 */
2745 if (!nodelist)
2746 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002747 }
2748
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002749 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002750 if (flags) {
2751 /*
2752 * Currently, we only support two mutually exclusive
2753 * mode flags.
2754 */
2755 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002756 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002757 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002758 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002759 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002760 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002761 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002762
2763 new = mpol_new(mode, mode_flags, &nodes);
2764 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002765 goto out;
2766
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002767 /*
2768 * Save nodes for mpol_to_str() to show the tmpfs mount options
2769 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2770 */
2771 if (mode != MPOL_PREFERRED)
2772 new->v.nodes = nodes;
2773 else if (nodelist)
2774 new->v.preferred_node = first_node(nodes);
2775 else
2776 new->flags |= MPOL_F_LOCAL;
2777
2778 /*
2779 * Save nodes for contextualization: this will be used to "clone"
2780 * the mempolicy in a specific context [cpuset] at a later time.
2781 */
2782 new->w.user_nodemask = nodes;
2783
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002784 err = 0;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002785
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002786out:
2787 /* Restore string for error message */
2788 if (nodelist)
2789 *--nodelist = ':';
2790 if (flags)
2791 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002792 if (!err)
2793 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002794 return err;
2795}
2796#endif /* CONFIG_TMPFS */
2797
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002798/**
2799 * mpol_to_str - format a mempolicy structure for printing
2800 * @buffer: to contain formatted mempolicy string
2801 * @maxlen: length of @buffer
2802 * @pol: pointer to mempolicy to be formatted
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002803 *
David Rientjes948927e2013-11-12 15:07:28 -08002804 * Convert @pol into a string. If @buffer is too short, truncate the string.
2805 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2806 * longest flag, "relative", and to display at least a few node ids.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002807 */
David Rientjes948927e2013-11-12 15:07:28 -08002808void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002809{
2810 char *p = buffer;
David Rientjes948927e2013-11-12 15:07:28 -08002811 nodemask_t nodes = NODE_MASK_NONE;
2812 unsigned short mode = MPOL_DEFAULT;
2813 unsigned short flags = 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002814
David Rientjes8790c712014-01-30 15:46:08 -08002815 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002816 mode = pol->mode;
David Rientjes948927e2013-11-12 15:07:28 -08002817 flags = pol->flags;
2818 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002819
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002820 switch (mode) {
2821 case MPOL_DEFAULT:
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002822 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002823 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002824 if (flags & MPOL_F_LOCAL)
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002825 mode = MPOL_LOCAL;
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002826 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002827 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002828 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002829 case MPOL_BIND:
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002830 case MPOL_INTERLEAVE:
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002831 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002832 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002833 default:
David Rientjes948927e2013-11-12 15:07:28 -08002834 WARN_ON_ONCE(1);
2835 snprintf(p, maxlen, "unknown");
2836 return;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002837 }
2838
David Rientjesb7a9f422013-11-21 14:32:06 -08002839 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002840
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002841 if (flags & MPOL_MODE_FLAGS) {
David Rientjes948927e2013-11-12 15:07:28 -08002842 p += snprintf(p, buffer + maxlen - p, "=");
David Rientjesf5b087b2008-04-28 02:12:27 -07002843
Lee Schermerhorn22919902008-04-28 02:13:22 -07002844 /*
2845 * Currently, the only defined flags are mutually exclusive
2846 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002847 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002848 p += snprintf(p, buffer + maxlen - p, "static");
2849 else if (flags & MPOL_F_RELATIVE_NODES)
2850 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002851 }
2852
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002853 if (!nodes_empty(nodes)) {
David Rientjes948927e2013-11-12 15:07:28 -08002854 p += snprintf(p, buffer + maxlen - p, ":");
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002855 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2856 }
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002857}