blob: c4403cdf3433ddc79515dac25e5eedd0b39fa320 [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
68#include <linux/mempolicy.h>
69#include <linux/mm.h>
70#include <linux/highmem.h>
71#include <linux/hugetlb.h>
72#include <linux/kernel.h>
73#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074#include <linux/nodemask.h>
75#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070076#include <linux/slab.h>
77#include <linux/string.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040078#include <linux/export.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070079#include <linux/nsproxy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070080#include <linux/interrupt.h>
81#include <linux/init.h>
82#include <linux/compat.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080083#include <linux/swap.h>
Christoph Lameter1a75a6c2006-01-08 01:01:02 -080084#include <linux/seq_file.h>
85#include <linux/proc_fs.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080086#include <linux/migrate.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080087#include <linux/ksm.h>
Christoph Lameter95a402c2006-06-23 02:03:53 -070088#include <linux/rmap.h>
David Quigley86c3a762006-06-23 02:04:02 -070089#include <linux/security.h>
Adrian Bunkdbcb0f12007-10-16 01:26:26 -070090#include <linux/syscalls.h>
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -070091#include <linux/ctype.h>
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -080092#include <linux/mm_inline.h>
Lee Schermerhornb24f53a2012-10-25 14:16:32 +020093#include <linux/mmu_notifier.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080094
Linus Torvalds1da177e2005-04-16 15:20:36 -070095#include <asm/tlbflush.h>
96#include <asm/uaccess.h>
Michal Hocko778d3b02011-07-26 16:08:30 -070097#include <linux/random.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
Nick Piggin62695a82008-10-18 20:26:09 -070099#include "internal.h"
100
Christoph Lameter38e35862006-01-08 01:01:01 -0800101/* Internal flags */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800102#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
Christoph Lameter38e35862006-01-08 01:01:01 -0800103#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800104
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800105static struct kmem_cache *policy_cache;
106static struct kmem_cache *sn_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108/* Highest zone. An specific allocation for a zone below that is not
109 policied. */
Christoph Lameter62672762007-02-10 01:43:07 -0800110enum zone_type policy_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700112/*
113 * run-time system-wide default policy => local allocation
114 */
H Hartley Sweetene754d792011-10-31 17:09:23 -0700115static struct mempolicy default_policy = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 .refcnt = ATOMIC_INIT(1), /* never free it */
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700117 .mode = MPOL_PREFERRED,
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700118 .flags = MPOL_F_LOCAL,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119};
120
Mel Gorman5606e382012-11-02 18:19:13 +0000121static struct mempolicy preferred_node_policy[MAX_NUMNODES];
122
123static struct mempolicy *get_task_policy(struct task_struct *p)
124{
125 struct mempolicy *pol = p->mempolicy;
Mel Gorman5606e382012-11-02 18:19:13 +0000126
127 if (!pol) {
Jianguo Wu1da6f0e2013-09-11 14:21:25 -0700128 int node = numa_node_id();
Mel Gorman5606e382012-11-02 18:19:13 +0000129
Jianguo Wu1da6f0e2013-09-11 14:21:25 -0700130 if (node != NUMA_NO_NODE) {
131 pol = &preferred_node_policy[node];
132 /*
133 * preferred_node_policy is not initialised early in
134 * boot
135 */
136 if (!pol->mode)
137 pol = NULL;
138 }
Mel Gorman5606e382012-11-02 18:19:13 +0000139 }
140
141 return pol;
142}
143
David Rientjes37012942008-04-28 02:12:33 -0700144static const struct mempolicy_operations {
145 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
Miao Xie708c1bb2010-05-24 14:32:07 -0700146 /*
147 * If read-side task has no lock to protect task->mempolicy, write-side
148 * task will rebind the task->mempolicy by two step. The first step is
149 * setting all the newly nodes, and the second step is cleaning all the
150 * disallowed nodes. In this way, we can avoid finding no node to alloc
151 * page.
152 * If we have a lock to protect task->mempolicy in read-side, we do
153 * rebind directly.
154 *
155 * step:
156 * MPOL_REBIND_ONCE - do rebind work at once
157 * MPOL_REBIND_STEP1 - set all the newly nodes
158 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
159 */
160 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
161 enum mpol_rebind_step step);
David Rientjes37012942008-04-28 02:12:33 -0700162} mpol_ops[MPOL_MAX];
163
Mel Gorman19770b32008-04-28 02:12:18 -0700164/* Check that the nodemask contains at least one populated zone */
David Rientjes37012942008-04-28 02:12:33 -0700165static int is_valid_nodemask(const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166{
Lai Jiangshand3eb1572013-02-22 16:33:22 -0800167 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168}
169
David Rientjesf5b087b2008-04-28 02:12:27 -0700170static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
171{
Bob Liu6d556292010-05-24 14:31:59 -0700172 return pol->flags & MPOL_MODE_FLAGS;
David Rientjes4c50bc02008-04-28 02:12:30 -0700173}
174
175static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
176 const nodemask_t *rel)
177{
178 nodemask_t tmp;
179 nodes_fold(tmp, *orig, nodes_weight(*rel));
180 nodes_onto(*ret, tmp, *rel);
David Rientjesf5b087b2008-04-28 02:12:27 -0700181}
182
David Rientjes37012942008-04-28 02:12:33 -0700183static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
184{
185 if (nodes_empty(*nodes))
186 return -EINVAL;
187 pol->v.nodes = *nodes;
188 return 0;
189}
190
191static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
192{
193 if (!nodes)
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700194 pol->flags |= MPOL_F_LOCAL; /* local allocation */
David Rientjes37012942008-04-28 02:12:33 -0700195 else if (nodes_empty(*nodes))
196 return -EINVAL; /* no allowed nodes */
197 else
198 pol->v.preferred_node = first_node(*nodes);
199 return 0;
200}
201
202static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
203{
204 if (!is_valid_nodemask(nodes))
205 return -EINVAL;
206 pol->v.nodes = *nodes;
207 return 0;
208}
209
Miao Xie58568d22009-06-16 15:31:49 -0700210/*
211 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
212 * any, for the new policy. mpol_new() has already validated the nodes
213 * parameter with respect to the policy mode and flags. But, we need to
214 * handle an empty nodemask with MPOL_PREFERRED here.
215 *
216 * Must be called holding task's alloc_lock to protect task's mems_allowed
217 * and mempolicy. May also be called holding the mmap_semaphore for write.
218 */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700219static int mpol_set_nodemask(struct mempolicy *pol,
220 const nodemask_t *nodes, struct nodemask_scratch *nsc)
Miao Xie58568d22009-06-16 15:31:49 -0700221{
Miao Xie58568d22009-06-16 15:31:49 -0700222 int ret;
223
224 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
225 if (pol == NULL)
226 return 0;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -0800227 /* Check N_MEMORY */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700228 nodes_and(nsc->mask1,
Lai Jiangshan01f13bd2012-12-12 13:51:33 -0800229 cpuset_current_mems_allowed, node_states[N_MEMORY]);
Miao Xie58568d22009-06-16 15:31:49 -0700230
231 VM_BUG_ON(!nodes);
232 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
233 nodes = NULL; /* explicit local allocation */
234 else {
235 if (pol->flags & MPOL_F_RELATIVE_NODES)
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700236 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
Miao Xie58568d22009-06-16 15:31:49 -0700237 else
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700238 nodes_and(nsc->mask2, *nodes, nsc->mask1);
239
Miao Xie58568d22009-06-16 15:31:49 -0700240 if (mpol_store_user_nodemask(pol))
241 pol->w.user_nodemask = *nodes;
242 else
243 pol->w.cpuset_mems_allowed =
244 cpuset_current_mems_allowed;
245 }
246
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700247 if (nodes)
248 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
249 else
250 ret = mpol_ops[pol->mode].create(pol, NULL);
Miao Xie58568d22009-06-16 15:31:49 -0700251 return ret;
252}
253
254/*
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
257 */
David Rientjes028fec42008-04-28 02:12:25 -0700258static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
259 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260{
261 struct mempolicy *policy;
262
David Rientjes028fec42008-04-28 02:12:25 -0700263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
David Rientjes00ef2d22013-02-22 16:35:36 -0800264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
Paul Mundt140d5a42007-07-15 23:38:16 -0700265
David Rientjes3e1f0642008-04-28 02:12:34 -0700266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700268 return ERR_PTR(-EINVAL);
Lee Schermerhornd3a71032012-10-25 14:16:29 +0200269 return NULL;
David Rientjes37012942008-04-28 02:12:33 -0700270 }
David Rientjes3e1f0642008-04-28 02:12:34 -0700271 VM_BUG_ON(!nodes);
272
273 /*
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
277 */
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
David Rientjes3e1f0642008-04-28 02:12:34 -0700283 }
Peter Zijlstra479e2802012-10-25 14:16:28 +0200284 } else if (mode == MPOL_LOCAL) {
285 if (!nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
287 mode = MPOL_PREFERRED;
David Rientjes3e1f0642008-04-28 02:12:34 -0700288 } else if (nodes_empty(*nodes))
289 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
291 if (!policy)
292 return ERR_PTR(-ENOMEM);
293 atomic_set(&policy->refcnt, 1);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700294 policy->mode = mode;
David Rientjes3e1f0642008-04-28 02:12:34 -0700295 policy->flags = flags;
David Rientjesf5b087b2008-04-28 02:12:27 -0700296
David Rientjes37012942008-04-28 02:12:33 -0700297 return policy;
298}
299
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700300/* Slow path of a mpol destructor. */
301void __mpol_put(struct mempolicy *p)
302{
303 if (!atomic_dec_and_test(&p->refcnt))
304 return;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700305 kmem_cache_free(policy_cache, p);
306}
307
Miao Xie708c1bb2010-05-24 14:32:07 -0700308static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700310{
311}
312
Miao Xie708c1bb2010-05-24 14:32:07 -0700313/*
314 * step:
315 * MPOL_REBIND_ONCE - do rebind work at once
316 * MPOL_REBIND_STEP1 - set all the newly nodes
317 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
318 */
319static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
320 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700321{
322 nodemask_t tmp;
323
324 if (pol->flags & MPOL_F_STATIC_NODES)
325 nodes_and(tmp, pol->w.user_nodemask, *nodes);
326 else if (pol->flags & MPOL_F_RELATIVE_NODES)
327 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
328 else {
Miao Xie708c1bb2010-05-24 14:32:07 -0700329 /*
330 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
331 * result
332 */
333 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
334 nodes_remap(tmp, pol->v.nodes,
335 pol->w.cpuset_mems_allowed, *nodes);
336 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
337 } else if (step == MPOL_REBIND_STEP2) {
338 tmp = pol->w.cpuset_mems_allowed;
339 pol->w.cpuset_mems_allowed = *nodes;
340 } else
341 BUG();
David Rientjes37012942008-04-28 02:12:33 -0700342 }
343
Miao Xie708c1bb2010-05-24 14:32:07 -0700344 if (nodes_empty(tmp))
345 tmp = *nodes;
346
347 if (step == MPOL_REBIND_STEP1)
348 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
349 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
350 pol->v.nodes = tmp;
351 else
352 BUG();
353
David Rientjes37012942008-04-28 02:12:33 -0700354 if (!node_isset(current->il_next, tmp)) {
355 current->il_next = next_node(current->il_next, tmp);
356 if (current->il_next >= MAX_NUMNODES)
357 current->il_next = first_node(tmp);
358 if (current->il_next >= MAX_NUMNODES)
359 current->il_next = numa_node_id();
360 }
361}
362
363static void mpol_rebind_preferred(struct mempolicy *pol,
Miao Xie708c1bb2010-05-24 14:32:07 -0700364 const nodemask_t *nodes,
365 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700366{
367 nodemask_t tmp;
368
David Rientjes37012942008-04-28 02:12:33 -0700369 if (pol->flags & MPOL_F_STATIC_NODES) {
370 int node = first_node(pol->w.user_nodemask);
371
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700372 if (node_isset(node, *nodes)) {
David Rientjes37012942008-04-28 02:12:33 -0700373 pol->v.preferred_node = node;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700374 pol->flags &= ~MPOL_F_LOCAL;
375 } else
376 pol->flags |= MPOL_F_LOCAL;
David Rientjes37012942008-04-28 02:12:33 -0700377 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
378 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
379 pol->v.preferred_node = first_node(tmp);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700380 } else if (!(pol->flags & MPOL_F_LOCAL)) {
David Rientjes37012942008-04-28 02:12:33 -0700381 pol->v.preferred_node = node_remap(pol->v.preferred_node,
382 pol->w.cpuset_mems_allowed,
383 *nodes);
384 pol->w.cpuset_mems_allowed = *nodes;
385 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386}
387
Miao Xie708c1bb2010-05-24 14:32:07 -0700388/*
389 * mpol_rebind_policy - Migrate a policy to a different set of nodes
390 *
391 * If read-side task has no lock to protect task->mempolicy, write-side
392 * task will rebind the task->mempolicy by two step. The first step is
393 * setting all the newly nodes, and the second step is cleaning all the
394 * disallowed nodes. In this way, we can avoid finding no node to alloc
395 * page.
396 * If we have a lock to protect task->mempolicy in read-side, we do
397 * rebind directly.
398 *
399 * step:
400 * MPOL_REBIND_ONCE - do rebind work at once
401 * MPOL_REBIND_STEP1 - set all the newly nodes
402 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
403 */
404static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
405 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700406{
David Rientjes1d0d2682008-04-28 02:12:32 -0700407 if (!pol)
408 return;
Wang Sheng-Hui89c522c2012-05-29 15:06:16 -0700409 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
David Rientjes1d0d2682008-04-28 02:12:32 -0700410 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
411 return;
Miao Xie708c1bb2010-05-24 14:32:07 -0700412
413 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
414 return;
415
416 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
417 BUG();
418
419 if (step == MPOL_REBIND_STEP1)
420 pol->flags |= MPOL_F_REBINDING;
421 else if (step == MPOL_REBIND_STEP2)
422 pol->flags &= ~MPOL_F_REBINDING;
423 else if (step >= MPOL_REBIND_NSTEP)
424 BUG();
425
426 mpol_ops[pol->mode].rebind(pol, newmask, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700427}
428
429/*
430 * Wrapper for mpol_rebind_policy() that just requires task
431 * pointer, and updates task mempolicy.
Miao Xie58568d22009-06-16 15:31:49 -0700432 *
433 * Called with task's alloc_lock held.
David Rientjes1d0d2682008-04-28 02:12:32 -0700434 */
435
Miao Xie708c1bb2010-05-24 14:32:07 -0700436void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
437 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700438{
Miao Xie708c1bb2010-05-24 14:32:07 -0700439 mpol_rebind_policy(tsk->mempolicy, new, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700440}
441
442/*
443 * Rebind each vma in mm to new nodemask.
444 *
445 * Call holding a reference to mm. Takes mm->mmap_sem during call.
446 */
447
448void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
449{
450 struct vm_area_struct *vma;
451
452 down_write(&mm->mmap_sem);
453 for (vma = mm->mmap; vma; vma = vma->vm_next)
Miao Xie708c1bb2010-05-24 14:32:07 -0700454 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
David Rientjes1d0d2682008-04-28 02:12:32 -0700455 up_write(&mm->mmap_sem);
456}
457
David Rientjes37012942008-04-28 02:12:33 -0700458static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
459 [MPOL_DEFAULT] = {
460 .rebind = mpol_rebind_default,
461 },
462 [MPOL_INTERLEAVE] = {
463 .create = mpol_new_interleave,
464 .rebind = mpol_rebind_nodemask,
465 },
466 [MPOL_PREFERRED] = {
467 .create = mpol_new_preferred,
468 .rebind = mpol_rebind_preferred,
469 },
470 [MPOL_BIND] = {
471 .create = mpol_new_bind,
472 .rebind = mpol_rebind_nodemask,
473 },
474};
475
Christoph Lameterfc301282006-01-18 17:42:29 -0800476static void migrate_page_add(struct page *page, struct list_head *pagelist,
477 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800478
Naoya Horiguchi98094942013-09-11 14:22:14 -0700479/*
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
482 */
483static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800484 unsigned long addr, unsigned long end,
485 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800486 void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487{
Hugh Dickins91612e02005-06-21 17:15:07 -0700488 pte_t *orig_pte;
489 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700490 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700491
Hugh Dickins705e87c2005-10-29 18:16:27 -0700492 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700493 do {
Linus Torvalds6aab3412005-11-28 14:34:23 -0800494 struct page *page;
Andy Whitcroft25ba77c2006-12-06 20:33:03 -0800495 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700496
497 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800499 page = vm_normal_page(vma, addr, *pte);
500 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800502 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800503 * vm_normal_page() filters out zero pages, but there might
504 * still be PageReserved pages to skip, perhaps in a VDSO.
Nick Piggin053837f2006-01-18 17:42:27 -0800505 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800506 if (PageReserved(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800507 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800508 nid = page_to_nid(page);
Christoph Lameter38e35862006-01-08 01:01:01 -0800509 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
510 continue;
511
Stephen Wilsonb1f72d12011-05-24 17:12:43 -0700512 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameterfc301282006-01-18 17:42:29 -0800513 migrate_page_add(page, private, flags);
Christoph Lameter38e35862006-01-08 01:01:01 -0800514 else
515 break;
Hugh Dickins91612e02005-06-21 17:15:07 -0700516 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickins705e87c2005-10-29 18:16:27 -0700517 pte_unmap_unlock(orig_pte, ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700518 return addr != end;
519}
520
Naoya Horiguchi98094942013-09-11 14:22:14 -0700521static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
522 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700523 void *private)
524{
525#ifdef CONFIG_HUGETLB_PAGE
526 int nid;
527 struct page *page;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800528 spinlock_t *ptl;
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700529
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800530 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700531 page = pte_page(huge_ptep_get((pte_t *)pmd));
532 nid = page_to_nid(page);
533 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
534 goto unlock;
535 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
536 if (flags & (MPOL_MF_MOVE_ALL) ||
537 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
538 isolate_huge_page(page, private);
539unlock:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800540 spin_unlock(ptl);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700541#else
542 BUG();
543#endif
544}
545
Naoya Horiguchi98094942013-09-11 14:22:14 -0700546static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800547 unsigned long addr, unsigned long end,
548 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800549 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700550{
551 pmd_t *pmd;
552 unsigned long next;
553
554 pmd = pmd_offset(pud, addr);
555 do {
556 next = pmd_addr_end(addr, end);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700557 if (!pmd_present(*pmd))
558 continue;
559 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
Naoya Horiguchi98094942013-09-11 14:22:14 -0700560 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700561 flags, private);
562 continue;
563 }
Kirill A. Shutemove1803772012-12-12 13:50:59 -0800564 split_huge_page_pmd(vma, addr, pmd);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -0700565 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
Hugh Dickins91612e02005-06-21 17:15:07 -0700566 continue;
Naoya Horiguchi98094942013-09-11 14:22:14 -0700567 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800568 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700569 return -EIO;
570 } while (pmd++, addr = next, addr != end);
571 return 0;
572}
573
Naoya Horiguchi98094942013-09-11 14:22:14 -0700574static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800575 unsigned long addr, unsigned long end,
576 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800577 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700578{
579 pud_t *pud;
580 unsigned long next;
581
582 pud = pud_offset(pgd, addr);
583 do {
584 next = pud_addr_end(addr, end);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700585 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
586 continue;
Hugh Dickins91612e02005-06-21 17:15:07 -0700587 if (pud_none_or_clear_bad(pud))
588 continue;
Naoya Horiguchi98094942013-09-11 14:22:14 -0700589 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800590 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700591 return -EIO;
592 } while (pud++, addr = next, addr != end);
593 return 0;
594}
595
Naoya Horiguchi98094942013-09-11 14:22:14 -0700596static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800597 unsigned long addr, unsigned long end,
598 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800599 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700600{
601 pgd_t *pgd;
602 unsigned long next;
603
Nick Pigginb5810032005-10-29 18:16:12 -0700604 pgd = pgd_offset(vma->vm_mm, addr);
Hugh Dickins91612e02005-06-21 17:15:07 -0700605 do {
606 next = pgd_addr_end(addr, end);
607 if (pgd_none_or_clear_bad(pgd))
608 continue;
Naoya Horiguchi98094942013-09-11 14:22:14 -0700609 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800610 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700611 return -EIO;
612 } while (pgd++, addr = next, addr != end);
613 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614}
615
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200616#ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
617/*
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200618 * This is used to mark a range of virtual addresses to be inaccessible.
619 * These are later cleared by a NUMA hinting fault. Depending on these
620 * faults, pages may be migrated for better NUMA placement.
621 *
622 * This is assuming that NUMA faults are handled using PROT_NONE. If
623 * an architecture makes a different choice, it will need further
624 * changes to the core.
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200625 */
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200626unsigned long change_prot_numa(struct vm_area_struct *vma,
627 unsigned long addr, unsigned long end)
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200628{
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200629 int nr_updated;
630 BUILD_BUG_ON(_PAGE_NUMA != _PAGE_PROTNONE);
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200631
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200632 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
Mel Gorman03c5a6e2012-11-02 14:52:48 +0000633 if (nr_updated)
634 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200635
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200636 return nr_updated;
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200637}
638#else
639static unsigned long change_prot_numa(struct vm_area_struct *vma,
640 unsigned long addr, unsigned long end)
641{
642 return 0;
643}
644#endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
645
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800646/*
Naoya Horiguchi98094942013-09-11 14:22:14 -0700647 * Walk through page tables and collect pages to be migrated.
648 *
649 * If pages found in a given range are on a set of nodes (determined by
650 * @nodes and @flags,) it's isolated and queued to the pagelist which is
651 * passed via @private.)
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800652 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653static struct vm_area_struct *
Naoya Horiguchi98094942013-09-11 14:22:14 -0700654queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Christoph Lameter38e35862006-01-08 01:01:01 -0800655 const nodemask_t *nodes, unsigned long flags, void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656{
657 int err;
658 struct vm_area_struct *first, *vma, *prev;
659
Nick Piggin053837f2006-01-18 17:42:27 -0800660
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 first = find_vma(mm, start);
662 if (!first)
663 return ERR_PTR(-EFAULT);
664 prev = NULL;
665 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200666 unsigned long endvma = vma->vm_end;
667
668 if (endvma > end)
669 endvma = end;
670 if (vma->vm_start > start)
671 start = vma->vm_start;
672
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800673 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
674 if (!vma->vm_next && vma->vm_end < end)
675 return ERR_PTR(-EFAULT);
676 if (prev && prev->vm_end < vma->vm_start)
677 return ERR_PTR(-EFAULT);
678 }
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800679
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200680 if (flags & MPOL_MF_LAZY) {
681 change_prot_numa(vma, start, endvma);
682 goto next;
683 }
684
685 if ((flags & MPOL_MF_STRICT) ||
686 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
687 vma_migratable(vma))) {
688
Naoya Horiguchi98094942013-09-11 14:22:14 -0700689 err = queue_pages_pgd_range(vma, start, endvma, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800690 flags, private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 if (err) {
692 first = ERR_PTR(err);
693 break;
694 }
695 }
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200696next:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 prev = vma;
698 }
699 return first;
700}
701
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700702/*
703 * Apply policy to a single VMA
704 * This must be called with the mmap_sem held for writing.
705 */
706static int vma_replace_policy(struct vm_area_struct *vma,
707 struct mempolicy *pol)
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700708{
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700709 int err;
710 struct mempolicy *old;
711 struct mempolicy *new;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700712
713 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
714 vma->vm_start, vma->vm_end, vma->vm_pgoff,
715 vma->vm_ops, vma->vm_file,
716 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
717
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700718 new = mpol_dup(pol);
719 if (IS_ERR(new))
720 return PTR_ERR(new);
721
722 if (vma->vm_ops && vma->vm_ops->set_policy) {
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700723 err = vma->vm_ops->set_policy(vma, new);
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700724 if (err)
725 goto err_out;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700726 }
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700727
728 old = vma->vm_policy;
729 vma->vm_policy = new; /* protected by mmap_sem */
730 mpol_put(old);
731
732 return 0;
733 err_out:
734 mpol_put(new);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700735 return err;
736}
737
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800739static int mbind_range(struct mm_struct *mm, unsigned long start,
740 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741{
742 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800743 struct vm_area_struct *prev;
744 struct vm_area_struct *vma;
745 int err = 0;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800746 pgoff_t pgoff;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800747 unsigned long vmstart;
748 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749
Linus Torvalds097d5912012-03-06 18:23:36 -0800750 vma = find_vma(mm, start);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800751 if (!vma || vma->vm_start > start)
752 return -EFAULT;
753
Linus Torvalds097d5912012-03-06 18:23:36 -0800754 prev = vma->vm_prev;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800755 if (start > vma->vm_start)
756 prev = vma;
757
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800758 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800760 vmstart = max(start, vma->vm_start);
761 vmend = min(end, vma->vm_end);
762
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800763 if (mpol_equal(vma_policy(vma), new_pol))
764 continue;
765
766 pgoff = vma->vm_pgoff +
767 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800768 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800769 vma->anon_vma, vma->vm_file, pgoff,
Caspar Zhang8aacc9f2011-09-14 16:20:58 -0700770 new_pol);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800771 if (prev) {
772 vma = prev;
773 next = vma->vm_next;
Oleg Nesterov3964acd2013-07-31 13:53:28 -0700774 if (mpol_equal(vma_policy(vma), new_pol))
775 continue;
776 /* vma_merge() joined vma && vma->next, case 8 */
777 goto replace;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800778 }
779 if (vma->vm_start != vmstart) {
780 err = split_vma(vma->vm_mm, vma, vmstart, 1);
781 if (err)
782 goto out;
783 }
784 if (vma->vm_end != vmend) {
785 err = split_vma(vma->vm_mm, vma, vmend, 0);
786 if (err)
787 goto out;
788 }
Oleg Nesterov3964acd2013-07-31 13:53:28 -0700789 replace:
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700790 err = vma_replace_policy(vma, new_pol);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700791 if (err)
792 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800794
795 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 return err;
797}
798
Paul Jacksonc61afb12006-03-24 03:16:08 -0800799/*
800 * Update task->flags PF_MEMPOLICY bit: set iff non-default
801 * mempolicy. Allows more rapid checking of this (combined perhaps
802 * with other PF_* flag bits) on memory allocation hot code paths.
803 *
804 * If called from outside this file, the task 'p' should -only- be
805 * a newly forked child not yet visible on the task list, because
806 * manipulating the task flags of a visible task is not safe.
807 *
808 * The above limitation is why this routine has the funny name
809 * mpol_fix_fork_child_flag().
810 *
811 * It is also safe to call this with a task pointer of current,
812 * which the static wrapper mpol_set_task_struct_flag() does,
813 * for use within this file.
814 */
815
816void mpol_fix_fork_child_flag(struct task_struct *p)
817{
818 if (p->mempolicy)
819 p->flags |= PF_MEMPOLICY;
820 else
821 p->flags &= ~PF_MEMPOLICY;
822}
823
824static void mpol_set_task_struct_flag(void)
825{
826 mpol_fix_fork_child_flag(current);
827}
828
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700830static long do_set_mempolicy(unsigned short mode, unsigned short flags,
831 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832{
Miao Xie58568d22009-06-16 15:31:49 -0700833 struct mempolicy *new, *old;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700834 struct mm_struct *mm = current->mm;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700835 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700836 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700838 if (!scratch)
839 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700840
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700841 new = mpol_new(mode, flags, nodes);
842 if (IS_ERR(new)) {
843 ret = PTR_ERR(new);
844 goto out;
845 }
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700846 /*
847 * prevent changing our mempolicy while show_numa_maps()
848 * is using it.
849 * Note: do_set_mempolicy() can be called at init time
850 * with no 'mm'.
851 */
852 if (mm)
853 down_write(&mm->mmap_sem);
Miao Xie58568d22009-06-16 15:31:49 -0700854 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700855 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700856 if (ret) {
857 task_unlock(current);
858 if (mm)
859 up_write(&mm->mmap_sem);
860 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700861 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700862 }
863 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 current->mempolicy = new;
Paul Jacksonc61afb12006-03-24 03:16:08 -0800865 mpol_set_task_struct_flag();
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700866 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700867 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700868 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700869 task_unlock(current);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700870 if (mm)
871 up_write(&mm->mmap_sem);
872
Miao Xie58568d22009-06-16 15:31:49 -0700873 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700874 ret = 0;
875out:
876 NODEMASK_SCRATCH_FREE(scratch);
877 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878}
879
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700880/*
881 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700882 *
883 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700884 */
885static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700887 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700888 if (p == &default_policy)
889 return;
890
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700891 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700892 case MPOL_BIND:
893 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700895 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 break;
897 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700898 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700899 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700900 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 break;
902 default:
903 BUG();
904 }
905}
906
907static int lookup_node(struct mm_struct *mm, unsigned long addr)
908{
909 struct page *p;
910 int err;
911
912 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
913 if (err >= 0) {
914 err = page_to_nid(p);
915 put_page(p);
916 }
917 return err;
918}
919
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700921static long do_get_mempolicy(int *policy, nodemask_t *nmask,
922 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700924 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 struct mm_struct *mm = current->mm;
926 struct vm_area_struct *vma = NULL;
927 struct mempolicy *pol = current->mempolicy;
928
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700929 if (flags &
930 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700932
933 if (flags & MPOL_F_MEMS_ALLOWED) {
934 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
935 return -EINVAL;
936 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700937 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700938 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700939 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700940 return 0;
941 }
942
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700944 /*
945 * Do NOT fall back to task policy if the
946 * vma/shared policy at addr is NULL. We
947 * want to return MPOL_DEFAULT in this case.
948 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 down_read(&mm->mmap_sem);
950 vma = find_vma_intersection(mm, addr, addr+1);
951 if (!vma) {
952 up_read(&mm->mmap_sem);
953 return -EFAULT;
954 }
955 if (vma->vm_ops && vma->vm_ops->get_policy)
956 pol = vma->vm_ops->get_policy(vma, addr);
957 else
958 pol = vma->vm_policy;
959 } else if (addr)
960 return -EINVAL;
961
962 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700963 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964
965 if (flags & MPOL_F_NODE) {
966 if (flags & MPOL_F_ADDR) {
967 err = lookup_node(mm, addr);
968 if (err < 0)
969 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700970 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700972 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700973 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 } else {
975 err = -EINVAL;
976 goto out;
977 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700978 } else {
979 *policy = pol == &default_policy ? MPOL_DEFAULT :
980 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700981 /*
982 * Internal mempolicy flags must be masked off before exposing
983 * the policy to userspace.
984 */
985 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700986 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987
988 if (vma) {
989 up_read(&current->mm->mmap_sem);
990 vma = NULL;
991 }
992
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700994 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700995 if (mpol_store_user_nodemask(pol)) {
996 *nmask = pol->w.user_nodemask;
997 } else {
998 task_lock(current);
999 get_policy_nodemask(pol, nmask);
1000 task_unlock(current);
1001 }
Miao Xie58568d22009-06-16 15:31:49 -07001002 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003
1004 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001005 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 if (vma)
1007 up_read(&current->mm->mmap_sem);
1008 return err;
1009}
1010
Christoph Lameterb20a3502006-03-22 00:09:12 -08001011#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -07001012/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001013 * page migration
1014 */
Christoph Lameterfc301282006-01-18 17:42:29 -08001015static void migrate_page_add(struct page *page, struct list_head *pagelist,
1016 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001017{
1018 /*
Christoph Lameterfc301282006-01-18 17:42:29 -08001019 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001020 */
Nick Piggin62695a82008-10-18 20:26:09 -07001021 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
1022 if (!isolate_lru_page(page)) {
1023 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001024 inc_zone_page_state(page, NR_ISOLATED_ANON +
1025 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -07001026 }
1027 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001028}
1029
Christoph Lameter742755a2006-06-23 02:03:55 -07001030static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001031{
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -07001032 if (PageHuge(page))
1033 return alloc_huge_page_node(page_hstate(compound_head(page)),
1034 node);
1035 else
1036 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001037}
1038
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001039/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001040 * Migrate pages from one node to a target node.
1041 * Returns error or the number of pages not migrated.
1042 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001043static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1044 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001045{
1046 nodemask_t nmask;
1047 LIST_HEAD(pagelist);
1048 int err = 0;
1049
1050 nodes_clear(nmask);
1051 node_set(source, nmask);
1052
Minchan Kim08270802012-10-08 16:33:38 -07001053 /*
1054 * This does not "check" the range but isolates all pages that
1055 * need migration. Between passing in the full user address
1056 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1057 */
1058 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
Naoya Horiguchi98094942013-09-11 14:22:14 -07001059 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001060 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1061
Minchan Kimcf608ac2010-10-26 14:21:29 -07001062 if (!list_empty(&pagelist)) {
Mel Gorman7f0f2492011-01-13 15:45:58 -08001063 err = migrate_pages(&pagelist, new_node_page, dest,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001064 MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001065 if (err)
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -07001066 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001067 }
Christoph Lameter95a402c2006-06-23 02:03:53 -07001068
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001069 return err;
1070}
1071
1072/*
1073 * Move pages between the two nodesets so as to preserve the physical
1074 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -08001075 *
1076 * Returns the number of page that could not be moved.
1077 */
Andrew Morton0ce72d42012-05-29 15:06:24 -07001078int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1079 const nodemask_t *to, int flags)
Christoph Lameter39743882006-01-08 01:00:51 -08001080{
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001081 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001082 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001083 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -08001084
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001085 err = migrate_prep();
1086 if (err)
1087 return err;
1088
Lee Schermerhorn53f25562008-04-28 02:13:20 -07001089 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -08001090
Andrew Morton0ce72d42012-05-29 15:06:24 -07001091 err = migrate_vmas(mm, from, to, flags);
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001092 if (err)
1093 goto out;
1094
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001095 /*
1096 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1097 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1098 * bit in 'tmp', and return that <source, dest> pair for migration.
1099 * The pair of nodemasks 'to' and 'from' define the map.
1100 *
1101 * If no pair of bits is found that way, fallback to picking some
1102 * pair of 'source' and 'dest' bits that are not the same. If the
1103 * 'source' and 'dest' bits are the same, this represents a node
1104 * that will be migrating to itself, so no pages need move.
1105 *
1106 * If no bits are left in 'tmp', or if all remaining bits left
1107 * in 'tmp' correspond to the same bit in 'to', return false
1108 * (nothing left to migrate).
1109 *
1110 * This lets us pick a pair of nodes to migrate between, such that
1111 * if possible the dest node is not already occupied by some other
1112 * source node, minimizing the risk of overloading the memory on a
1113 * node that would happen if we migrated incoming memory to a node
1114 * before migrating outgoing memory source that same node.
1115 *
1116 * A single scan of tmp is sufficient. As we go, we remember the
1117 * most recent <s, d> pair that moved (s != d). If we find a pair
1118 * that not only moved, but what's better, moved to an empty slot
1119 * (d is not set in tmp), then we break out then, with that pair.
Justin P. Mattockae0e47f2011-03-01 15:06:02 +01001120 * Otherwise when we finish scanning from_tmp, we at least have the
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001121 * most recent <s, d> pair that moved. If we get all the way through
1122 * the scan of tmp without finding any node that moved, much less
1123 * moved to an empty node, then there is nothing left worth migrating.
1124 */
Christoph Lameterd4984712006-01-08 01:00:55 -08001125
Andrew Morton0ce72d42012-05-29 15:06:24 -07001126 tmp = *from;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001127 while (!nodes_empty(tmp)) {
1128 int s,d;
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001129 int source = NUMA_NO_NODE;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001130 int dest = 0;
1131
1132 for_each_node_mask(s, tmp) {
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001133
1134 /*
1135 * do_migrate_pages() tries to maintain the relative
1136 * node relationship of the pages established between
1137 * threads and memory areas.
1138 *
1139 * However if the number of source nodes is not equal to
1140 * the number of destination nodes we can not preserve
1141 * this node relative relationship. In that case, skip
1142 * copying memory from a node that is in the destination
1143 * mask.
1144 *
1145 * Example: [2,3,4] -> [3,4,5] moves everything.
1146 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1147 */
1148
Andrew Morton0ce72d42012-05-29 15:06:24 -07001149 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1150 (node_isset(s, *to)))
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001151 continue;
1152
Andrew Morton0ce72d42012-05-29 15:06:24 -07001153 d = node_remap(s, *from, *to);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001154 if (s == d)
1155 continue;
1156
1157 source = s; /* Node moved. Memorize */
1158 dest = d;
1159
1160 /* dest not in remaining from nodes? */
1161 if (!node_isset(dest, tmp))
1162 break;
1163 }
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001164 if (source == NUMA_NO_NODE)
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001165 break;
1166
1167 node_clear(source, tmp);
1168 err = migrate_to_node(mm, source, dest, flags);
1169 if (err > 0)
1170 busy += err;
1171 if (err < 0)
1172 break;
Christoph Lameter39743882006-01-08 01:00:51 -08001173 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001174out:
Christoph Lameter39743882006-01-08 01:00:51 -08001175 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001176 if (err < 0)
1177 return err;
1178 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001179
Christoph Lameter39743882006-01-08 01:00:51 -08001180}
1181
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001182/*
1183 * Allocate a new page for page migration based on vma policy.
1184 * Start assuming that page is mapped by vma pointed to by @private.
1185 * Search forward from there, if not. N.B., this assumes that the
1186 * list of pages handed to migrate_pages()--which is how we get here--
1187 * is in virtual address order.
1188 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001189static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001190{
1191 struct vm_area_struct *vma = (struct vm_area_struct *)private;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001192 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001193
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001194 while (vma) {
1195 address = page_address_in_vma(page, vma);
1196 if (address != -EFAULT)
1197 break;
1198 vma = vma->vm_next;
1199 }
Naoya Horiguchi0bf598d2013-09-11 14:22:16 -07001200 /*
1201 * queue_pages_range() confirms that @page belongs to some vma,
1202 * so vma shouldn't be NULL.
1203 */
1204 BUG_ON(!vma);
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001205
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001206 if (PageHuge(page))
1207 return alloc_huge_page_noerr(vma, address, 1);
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001208 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001209}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001210#else
1211
1212static void migrate_page_add(struct page *page, struct list_head *pagelist,
1213 unsigned long flags)
1214{
1215}
1216
Andrew Morton0ce72d42012-05-29 15:06:24 -07001217int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1218 const nodemask_t *to, int flags)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001219{
1220 return -ENOSYS;
1221}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001222
Keith Owens69939742006-10-11 01:21:28 -07001223static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001224{
1225 return NULL;
1226}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001227#endif
1228
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001229static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001230 unsigned short mode, unsigned short mode_flags,
1231 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001232{
1233 struct vm_area_struct *vma;
1234 struct mm_struct *mm = current->mm;
1235 struct mempolicy *new;
1236 unsigned long end;
1237 int err;
1238 LIST_HEAD(pagelist);
1239
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001240 if (flags & ~(unsigned long)MPOL_MF_VALID)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001241 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001242 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001243 return -EPERM;
1244
1245 if (start & ~PAGE_MASK)
1246 return -EINVAL;
1247
1248 if (mode == MPOL_DEFAULT)
1249 flags &= ~MPOL_MF_STRICT;
1250
1251 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1252 end = start + len;
1253
1254 if (end < start)
1255 return -EINVAL;
1256 if (end == start)
1257 return 0;
1258
David Rientjes028fec42008-04-28 02:12:25 -07001259 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001260 if (IS_ERR(new))
1261 return PTR_ERR(new);
1262
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001263 if (flags & MPOL_MF_LAZY)
1264 new->flags |= MPOL_F_MOF;
1265
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001266 /*
1267 * If we are using the default policy then operation
1268 * on discontinuous address spaces is okay after all
1269 */
1270 if (!new)
1271 flags |= MPOL_MF_DISCONTIG_OK;
1272
David Rientjes028fec42008-04-28 02:12:25 -07001273 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1274 start, start + len, mode, mode_flags,
David Rientjes00ef2d22013-02-22 16:35:36 -08001275 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001276
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001277 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1278
1279 err = migrate_prep();
1280 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001281 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001282 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001283 {
1284 NODEMASK_SCRATCH(scratch);
1285 if (scratch) {
1286 down_write(&mm->mmap_sem);
1287 task_lock(current);
1288 err = mpol_set_nodemask(new, nmask, scratch);
1289 task_unlock(current);
1290 if (err)
1291 up_write(&mm->mmap_sem);
1292 } else
1293 err = -ENOMEM;
1294 NODEMASK_SCRATCH_FREE(scratch);
1295 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001296 if (err)
1297 goto mpol_out;
1298
Naoya Horiguchi98094942013-09-11 14:22:14 -07001299 vma = queue_pages_range(mm, start, end, nmask,
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001300 flags | MPOL_MF_INVERT, &pagelist);
1301
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001302 err = PTR_ERR(vma); /* maybe ... */
Mel Gormana7200942012-11-16 09:37:58 +00001303 if (!IS_ERR(vma))
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001304 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001305
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001306 if (!err) {
1307 int nr_failed = 0;
1308
Minchan Kimcf608ac2010-10-26 14:21:29 -07001309 if (!list_empty(&pagelist)) {
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001310 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001311 nr_failed = migrate_pages(&pagelist, new_vma_page,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001312 (unsigned long)vma,
1313 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001314 if (nr_failed)
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001315 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001316 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001317
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001318 if (nr_failed && (flags & MPOL_MF_STRICT))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001319 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001320 } else
1321 putback_lru_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001322
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001323 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001324 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001325 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001326 return err;
1327}
1328
Christoph Lameter39743882006-01-08 01:00:51 -08001329/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001330 * User space interface with variable sized bitmaps for nodelists.
1331 */
1332
1333/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001334static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001335 unsigned long maxnode)
1336{
1337 unsigned long k;
1338 unsigned long nlongs;
1339 unsigned long endmask;
1340
1341 --maxnode;
1342 nodes_clear(*nodes);
1343 if (maxnode == 0 || !nmask)
1344 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001345 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001346 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001347
1348 nlongs = BITS_TO_LONGS(maxnode);
1349 if ((maxnode % BITS_PER_LONG) == 0)
1350 endmask = ~0UL;
1351 else
1352 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1353
1354 /* When the user specified more nodes than supported just check
1355 if the non supported part is all zero. */
1356 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1357 if (nlongs > PAGE_SIZE/sizeof(long))
1358 return -EINVAL;
1359 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1360 unsigned long t;
1361 if (get_user(t, nmask + k))
1362 return -EFAULT;
1363 if (k == nlongs - 1) {
1364 if (t & endmask)
1365 return -EINVAL;
1366 } else if (t)
1367 return -EINVAL;
1368 }
1369 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1370 endmask = ~0UL;
1371 }
1372
1373 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1374 return -EFAULT;
1375 nodes_addr(*nodes)[nlongs-1] &= endmask;
1376 return 0;
1377}
1378
1379/* Copy a kernel node mask to user space */
1380static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1381 nodemask_t *nodes)
1382{
1383 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1384 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1385
1386 if (copy > nbytes) {
1387 if (copy > PAGE_SIZE)
1388 return -EINVAL;
1389 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1390 return -EFAULT;
1391 copy = nbytes;
1392 }
1393 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1394}
1395
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001396SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1397 unsigned long, mode, unsigned long __user *, nmask,
1398 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001399{
1400 nodemask_t nodes;
1401 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001402 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001403
David Rientjes028fec42008-04-28 02:12:25 -07001404 mode_flags = mode & MPOL_MODE_FLAGS;
1405 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001406 if (mode >= MPOL_MAX)
1407 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001408 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1409 (mode_flags & MPOL_F_RELATIVE_NODES))
1410 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001411 err = get_nodes(&nodes, nmask, maxnode);
1412 if (err)
1413 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001414 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001415}
1416
1417/* Set the process memory policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001418SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1419 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001420{
1421 int err;
1422 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001423 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001424
David Rientjes028fec42008-04-28 02:12:25 -07001425 flags = mode & MPOL_MODE_FLAGS;
1426 mode &= ~MPOL_MODE_FLAGS;
1427 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001428 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001429 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1430 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001431 err = get_nodes(&nodes, nmask, maxnode);
1432 if (err)
1433 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001434 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001435}
1436
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001437SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1438 const unsigned long __user *, old_nodes,
1439 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001440{
David Howellsc69e8d92008-11-14 10:39:19 +11001441 const struct cred *cred = current_cred(), *tcred;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001442 struct mm_struct *mm = NULL;
Christoph Lameter39743882006-01-08 01:00:51 -08001443 struct task_struct *task;
Christoph Lameter39743882006-01-08 01:00:51 -08001444 nodemask_t task_nodes;
1445 int err;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001446 nodemask_t *old;
1447 nodemask_t *new;
1448 NODEMASK_SCRATCH(scratch);
Christoph Lameter39743882006-01-08 01:00:51 -08001449
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001450 if (!scratch)
1451 return -ENOMEM;
Christoph Lameter39743882006-01-08 01:00:51 -08001452
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001453 old = &scratch->mask1;
1454 new = &scratch->mask2;
1455
1456 err = get_nodes(old, old_nodes, maxnode);
Christoph Lameter39743882006-01-08 01:00:51 -08001457 if (err)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001458 goto out;
1459
1460 err = get_nodes(new, new_nodes, maxnode);
1461 if (err)
1462 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001463
1464 /* Find the mm_struct */
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001465 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001466 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001467 if (!task) {
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001468 rcu_read_unlock();
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001469 err = -ESRCH;
1470 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001471 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001472 get_task_struct(task);
Christoph Lameter39743882006-01-08 01:00:51 -08001473
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001474 err = -EINVAL;
Christoph Lameter39743882006-01-08 01:00:51 -08001475
1476 /*
1477 * Check if this process has the right to modify the specified
1478 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001479 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001480 * userid as the target process.
1481 */
David Howellsc69e8d92008-11-14 10:39:19 +11001482 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001483 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1484 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001485 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001486 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001487 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001488 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001489 }
David Howellsc69e8d92008-11-14 10:39:19 +11001490 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001491
1492 task_nodes = cpuset_mems_allowed(task);
1493 /* Is the user allowed to access the target nodes? */
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001494 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001495 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001496 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001497 }
1498
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08001499 if (!nodes_subset(*new, node_states[N_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001500 err = -EINVAL;
Christoph Lameter3268c632012-03-21 16:34:06 -07001501 goto out_put;
Christoph Lameter3b42d282007-08-31 00:12:08 -07001502 }
1503
David Quigley86c3a762006-06-23 02:04:02 -07001504 err = security_task_movememory(task);
1505 if (err)
Christoph Lameter3268c632012-03-21 16:34:06 -07001506 goto out_put;
David Quigley86c3a762006-06-23 02:04:02 -07001507
Christoph Lameter3268c632012-03-21 16:34:06 -07001508 mm = get_task_mm(task);
1509 put_task_struct(task);
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001510
1511 if (!mm) {
Christoph Lameter3268c632012-03-21 16:34:06 -07001512 err = -EINVAL;
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001513 goto out;
1514 }
1515
1516 err = do_migrate_pages(mm, old, new,
1517 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter3268c632012-03-21 16:34:06 -07001518
1519 mmput(mm);
1520out:
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001521 NODEMASK_SCRATCH_FREE(scratch);
1522
Christoph Lameter39743882006-01-08 01:00:51 -08001523 return err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001524
1525out_put:
1526 put_task_struct(task);
1527 goto out;
1528
Christoph Lameter39743882006-01-08 01:00:51 -08001529}
1530
1531
Christoph Lameter8bccd852005-10-29 18:16:59 -07001532/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001533SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1534 unsigned long __user *, nmask, unsigned long, maxnode,
1535 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001536{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001537 int err;
1538 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001539 nodemask_t nodes;
1540
1541 if (nmask != NULL && maxnode < MAX_NUMNODES)
1542 return -EINVAL;
1543
1544 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1545
1546 if (err)
1547 return err;
1548
1549 if (policy && put_user(pval, policy))
1550 return -EFAULT;
1551
1552 if (nmask)
1553 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1554
1555 return err;
1556}
1557
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558#ifdef CONFIG_COMPAT
1559
1560asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1561 compat_ulong_t __user *nmask,
1562 compat_ulong_t maxnode,
1563 compat_ulong_t addr, compat_ulong_t flags)
1564{
1565 long err;
1566 unsigned long __user *nm = NULL;
1567 unsigned long nr_bits, alloc_size;
1568 DECLARE_BITMAP(bm, MAX_NUMNODES);
1569
1570 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1571 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1572
1573 if (nmask)
1574 nm = compat_alloc_user_space(alloc_size);
1575
1576 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1577
1578 if (!err && nmask) {
KAMEZAWA Hiroyuki2bbff6c2011-09-14 16:21:02 -07001579 unsigned long copy_size;
1580 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1581 err = copy_from_user(bm, nm, copy_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582 /* ensure entire bitmap is zeroed */
1583 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1584 err |= compat_put_bitmap(nmask, bm, nr_bits);
1585 }
1586
1587 return err;
1588}
1589
1590asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1591 compat_ulong_t maxnode)
1592{
1593 long err = 0;
1594 unsigned long __user *nm = NULL;
1595 unsigned long nr_bits, alloc_size;
1596 DECLARE_BITMAP(bm, MAX_NUMNODES);
1597
1598 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1599 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1600
1601 if (nmask) {
1602 err = compat_get_bitmap(bm, nmask, nr_bits);
1603 nm = compat_alloc_user_space(alloc_size);
1604 err |= copy_to_user(nm, bm, alloc_size);
1605 }
1606
1607 if (err)
1608 return -EFAULT;
1609
1610 return sys_set_mempolicy(mode, nm, nr_bits+1);
1611}
1612
1613asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1614 compat_ulong_t mode, compat_ulong_t __user *nmask,
1615 compat_ulong_t maxnode, compat_ulong_t flags)
1616{
1617 long err = 0;
1618 unsigned long __user *nm = NULL;
1619 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001620 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
1622 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1623 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1624
1625 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001626 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001628 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 }
1630
1631 if (err)
1632 return -EFAULT;
1633
1634 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1635}
1636
1637#endif
1638
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001639/*
1640 * get_vma_policy(@task, @vma, @addr)
1641 * @task - task for fallback if vma policy == default
1642 * @vma - virtual memory area whose policy is sought
1643 * @addr - address in @vma for shared policy lookup
1644 *
1645 * Returns effective policy for a VMA at specified address.
1646 * Falls back to @task or system default policy, as necessary.
David Rientjes32f85162012-10-16 17:31:23 -07001647 * Current or other task's task mempolicy and non-shared vma policies must be
1648 * protected by task_lock(task) by the caller.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001649 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1650 * count--added by the get_policy() vm_op, as appropriate--to protect against
1651 * freeing by another task. It is the caller's responsibility to free the
1652 * extra reference for shared policies.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001653 */
Stephen Wilsond98f6cb2011-05-24 17:12:41 -07001654struct mempolicy *get_vma_policy(struct task_struct *task,
Christoph Lameter48fce342006-01-08 01:01:03 -08001655 struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656{
Mel Gorman5606e382012-11-02 18:19:13 +00001657 struct mempolicy *pol = get_task_policy(task);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658
1659 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001660 if (vma->vm_ops && vma->vm_ops->get_policy) {
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001661 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1662 addr);
1663 if (vpol)
1664 pol = vpol;
Mel Gorman00442ad2012-10-08 16:29:20 -07001665 } else if (vma->vm_policy) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 pol = vma->vm_policy;
Mel Gorman00442ad2012-10-08 16:29:20 -07001667
1668 /*
1669 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1670 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1671 * count on these policies which will be dropped by
1672 * mpol_cond_put() later
1673 */
1674 if (mpol_needs_cond_ref(pol))
1675 mpol_get(pol);
1676 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 }
1678 if (!pol)
1679 pol = &default_policy;
1680 return pol;
1681}
1682
Mel Gormanfc3147242013-10-07 11:29:09 +01001683bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1684{
1685 struct mempolicy *pol = get_task_policy(task);
1686 if (vma) {
1687 if (vma->vm_ops && vma->vm_ops->get_policy) {
1688 bool ret = false;
1689
1690 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1691 if (pol && (pol->flags & MPOL_F_MOF))
1692 ret = true;
1693 mpol_cond_put(pol);
1694
1695 return ret;
1696 } else if (vma->vm_policy) {
1697 pol = vma->vm_policy;
1698 }
1699 }
1700
1701 if (!pol)
1702 return default_policy.flags & MPOL_F_MOF;
1703
1704 return pol->flags & MPOL_F_MOF;
1705}
1706
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001707static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1708{
1709 enum zone_type dynamic_policy_zone = policy_zone;
1710
1711 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1712
1713 /*
1714 * if policy->v.nodes has movable memory only,
1715 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1716 *
1717 * policy->v.nodes is intersect with node_states[N_MEMORY].
1718 * so if the following test faile, it implies
1719 * policy->v.nodes has movable memory only.
1720 */
1721 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1722 dynamic_policy_zone = ZONE_MOVABLE;
1723
1724 return zone >= dynamic_policy_zone;
1725}
1726
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001727/*
1728 * Return a nodemask representing a mempolicy for filtering nodes for
1729 * page allocation
1730 */
1731static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001732{
1733 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001734 if (unlikely(policy->mode == MPOL_BIND) &&
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001735 apply_policy_zone(policy, gfp_zone(gfp)) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001736 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1737 return &policy->v.nodes;
1738
1739 return NULL;
1740}
1741
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001742/* Return a zonelist indicated by gfp for node representing a mempolicy */
Andi Kleen2f5f9482011-03-04 17:36:29 -08001743static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1744 int nd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745{
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001746 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001748 if (!(policy->flags & MPOL_F_LOCAL))
1749 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 break;
1751 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001752 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001753 * Normally, MPOL_BIND allocations are node-local within the
1754 * allowed nodemask. However, if __GFP_THISNODE is set and the
Bob Liu6eb27e12010-05-24 14:32:00 -07001755 * current node isn't part of the mask, we use the zonelist for
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001756 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001757 */
Mel Gorman19770b32008-04-28 02:12:18 -07001758 if (unlikely(gfp & __GFP_THISNODE) &&
1759 unlikely(!node_isset(nd, policy->v.nodes)))
1760 nd = first_node(policy->v.nodes);
1761 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 BUG();
1764 }
Mel Gorman0e884602008-04-28 02:12:14 -07001765 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766}
1767
1768/* Do dynamic interleaving for a process */
1769static unsigned interleave_nodes(struct mempolicy *policy)
1770{
1771 unsigned nid, next;
1772 struct task_struct *me = current;
1773
1774 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001775 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001777 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001778 if (next < MAX_NUMNODES)
1779 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 return nid;
1781}
1782
Christoph Lameterdc85da12006-01-18 17:42:36 -08001783/*
1784 * Depending on the memory policy provide a node from which to allocate the
1785 * next slab entry.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001786 * @policy must be protected by freeing by the caller. If @policy is
1787 * the current task's mempolicy, this protection is implicit, as only the
1788 * task can change it's policy. The system default policy requires no
1789 * such protection.
Christoph Lameterdc85da12006-01-18 17:42:36 -08001790 */
Andi Kleene7b691b2012-06-09 02:40:03 -07001791unsigned slab_node(void)
Christoph Lameterdc85da12006-01-18 17:42:36 -08001792{
Andi Kleene7b691b2012-06-09 02:40:03 -07001793 struct mempolicy *policy;
1794
1795 if (in_interrupt())
1796 return numa_node_id();
1797
1798 policy = current->mempolicy;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001799 if (!policy || policy->flags & MPOL_F_LOCAL)
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001800 return numa_node_id();
Christoph Lameter765c4502006-09-27 01:50:08 -07001801
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001802 switch (policy->mode) {
1803 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001804 /*
1805 * handled MPOL_F_LOCAL above
1806 */
1807 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001808
Christoph Lameterdc85da12006-01-18 17:42:36 -08001809 case MPOL_INTERLEAVE:
1810 return interleave_nodes(policy);
1811
Mel Gormandd1a2392008-04-28 02:12:17 -07001812 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001813 /*
1814 * Follow bind policy behavior and start allocation at the
1815 * first node.
1816 */
Mel Gorman19770b32008-04-28 02:12:18 -07001817 struct zonelist *zonelist;
1818 struct zone *zone;
1819 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1820 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1821 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1822 &policy->v.nodes,
1823 &zone);
Eric Dumazet800416f2010-10-27 19:33:43 +02001824 return zone ? zone->node : numa_node_id();
Mel Gormandd1a2392008-04-28 02:12:17 -07001825 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001826
Christoph Lameterdc85da12006-01-18 17:42:36 -08001827 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001828 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001829 }
1830}
1831
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832/* Do static interleaving for a VMA with known offset. */
1833static unsigned offset_il_node(struct mempolicy *pol,
1834 struct vm_area_struct *vma, unsigned long off)
1835{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001836 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001837 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 int c;
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001839 int nid = NUMA_NO_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840
David Rientjesf5b087b2008-04-28 02:12:27 -07001841 if (!nnodes)
1842 return numa_node_id();
1843 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 c = 0;
1845 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001846 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 c++;
1848 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 return nid;
1850}
1851
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001852/* Determine a node number for interleave */
1853static inline unsigned interleave_nid(struct mempolicy *pol,
1854 struct vm_area_struct *vma, unsigned long addr, int shift)
1855{
1856 if (vma) {
1857 unsigned long off;
1858
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001859 /*
1860 * for small pages, there is no difference between
1861 * shift and PAGE_SHIFT, so the bit-shift is safe.
1862 * for huge pages, since vm_pgoff is in units of small
1863 * pages, we need to shift off the always 0 bits to get
1864 * a useful offset.
1865 */
1866 BUG_ON(shift < PAGE_SHIFT);
1867 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001868 off += (addr - vma->vm_start) >> shift;
1869 return offset_il_node(pol, vma, off);
1870 } else
1871 return interleave_nodes(pol);
1872}
1873
Michal Hocko778d3b02011-07-26 16:08:30 -07001874/*
1875 * Return the bit number of a random bit set in the nodemask.
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001876 * (returns NUMA_NO_NODE if nodemask is empty)
Michal Hocko778d3b02011-07-26 16:08:30 -07001877 */
1878int node_random(const nodemask_t *maskp)
1879{
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001880 int w, bit = NUMA_NO_NODE;
Michal Hocko778d3b02011-07-26 16:08:30 -07001881
1882 w = nodes_weight(*maskp);
1883 if (w)
1884 bit = bitmap_ord_to_pos(maskp->bits,
1885 get_random_int() % w, MAX_NUMNODES);
1886 return bit;
1887}
1888
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001889#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001890/*
1891 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1892 * @vma = virtual memory area whose policy is sought
1893 * @addr = address in @vma for shared policy lookup and interleave policy
1894 * @gfp_flags = for requested zone
Mel Gorman19770b32008-04-28 02:12:18 -07001895 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1896 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001897 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001898 * Returns a zonelist suitable for a huge page allocation and a pointer
1899 * to the struct mempolicy for conditional unref after allocation.
1900 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1901 * @nodemask for filtering the zonelist.
Miao Xiec0ff7452010-05-24 14:32:08 -07001902 *
1903 * Must be protected by get_mems_allowed()
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001904 */
Mel Gorman396faf02007-07-17 04:03:13 -07001905struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001906 gfp_t gfp_flags, struct mempolicy **mpol,
1907 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001908{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001909 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001910
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001911 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001912 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001913
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001914 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1915 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001916 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001917 } else {
Andi Kleen2f5f9482011-03-04 17:36:29 -08001918 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001919 if ((*mpol)->mode == MPOL_BIND)
1920 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001921 }
1922 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001923}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001924
1925/*
1926 * init_nodemask_of_mempolicy
1927 *
1928 * If the current task's mempolicy is "default" [NULL], return 'false'
1929 * to indicate default policy. Otherwise, extract the policy nodemask
1930 * for 'bind' or 'interleave' policy into the argument nodemask, or
1931 * initialize the argument nodemask to contain the single node for
1932 * 'preferred' or 'local' policy and return 'true' to indicate presence
1933 * of non-default mempolicy.
1934 *
1935 * We don't bother with reference counting the mempolicy [mpol_get/put]
1936 * because the current task is examining it's own mempolicy and a task's
1937 * mempolicy is only ever changed by the task itself.
1938 *
1939 * N.B., it is the caller's responsibility to free a returned nodemask.
1940 */
1941bool init_nodemask_of_mempolicy(nodemask_t *mask)
1942{
1943 struct mempolicy *mempolicy;
1944 int nid;
1945
1946 if (!(mask && current->mempolicy))
1947 return false;
1948
Miao Xiec0ff7452010-05-24 14:32:08 -07001949 task_lock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001950 mempolicy = current->mempolicy;
1951 switch (mempolicy->mode) {
1952 case MPOL_PREFERRED:
1953 if (mempolicy->flags & MPOL_F_LOCAL)
1954 nid = numa_node_id();
1955 else
1956 nid = mempolicy->v.preferred_node;
1957 init_nodemask_of_node(mask, nid);
1958 break;
1959
1960 case MPOL_BIND:
1961 /* Fall through */
1962 case MPOL_INTERLEAVE:
1963 *mask = mempolicy->v.nodes;
1964 break;
1965
1966 default:
1967 BUG();
1968 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001969 task_unlock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001970
1971 return true;
1972}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001973#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001974
David Rientjes6f48d0eb2010-08-09 17:18:52 -07001975/*
1976 * mempolicy_nodemask_intersects
1977 *
1978 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1979 * policy. Otherwise, check for intersection between mask and the policy
1980 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1981 * policy, always return true since it may allocate elsewhere on fallback.
1982 *
1983 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1984 */
1985bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1986 const nodemask_t *mask)
1987{
1988 struct mempolicy *mempolicy;
1989 bool ret = true;
1990
1991 if (!mask)
1992 return ret;
1993 task_lock(tsk);
1994 mempolicy = tsk->mempolicy;
1995 if (!mempolicy)
1996 goto out;
1997
1998 switch (mempolicy->mode) {
1999 case MPOL_PREFERRED:
2000 /*
2001 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2002 * allocate from, they may fallback to other nodes when oom.
2003 * Thus, it's possible for tsk to have allocated memory from
2004 * nodes in mask.
2005 */
2006 break;
2007 case MPOL_BIND:
2008 case MPOL_INTERLEAVE:
2009 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2010 break;
2011 default:
2012 BUG();
2013 }
2014out:
2015 task_unlock(tsk);
2016 return ret;
2017}
2018
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019/* Allocate a page in interleaved policy.
2020 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07002021static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2022 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023{
2024 struct zonelist *zl;
2025 struct page *page;
2026
Mel Gorman0e884602008-04-28 02:12:14 -07002027 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07002029 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07002030 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 return page;
2032}
2033
2034/**
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002035 * alloc_pages_vma - Allocate a page for a VMA.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 *
2037 * @gfp:
2038 * %GFP_USER user allocation.
2039 * %GFP_KERNEL kernel allocations,
2040 * %GFP_HIGHMEM highmem/user allocations,
2041 * %GFP_FS allocation should not call back into a file system.
2042 * %GFP_ATOMIC don't sleep.
2043 *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002044 * @order:Order of the GFP allocation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 * @vma: Pointer to VMA or NULL if not available.
2046 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2047 *
2048 * This function allocates a page from the kernel page pool and applies
2049 * a NUMA policy associated with the VMA or the current process.
2050 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2051 * mm_struct of the VMA to prevent it from going away. Should be used for
2052 * all allocations for pages that will be mapped into
2053 * user space. Returns NULL when no page can be allocated.
2054 *
2055 * Should be called with the mm_sem of the vma hold.
2056 */
2057struct page *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002058alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
Andi Kleen2f5f9482011-03-04 17:36:29 -08002059 unsigned long addr, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060{
Mel Gormancc9a6c82012-03-21 16:34:11 -07002061 struct mempolicy *pol;
Miao Xiec0ff7452010-05-24 14:32:08 -07002062 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002063 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064
Mel Gormancc9a6c82012-03-21 16:34:11 -07002065retry_cpuset:
2066 pol = get_vma_policy(current, vma, addr);
2067 cpuset_mems_cookie = get_mems_allowed();
2068
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002069 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08002071
Andi Kleen8eac5632011-02-25 14:44:28 -08002072 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002073 mpol_cond_put(pol);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002074 page = alloc_page_interleave(gfp, order, nid);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002075 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2076 goto retry_cpuset;
2077
Miao Xiec0ff7452010-05-24 14:32:08 -07002078 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 }
David Rientjes212a0a62012-12-11 16:02:51 -08002080 page = __alloc_pages_nodemask(gfp, order,
2081 policy_zonelist(gfp, pol, node),
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002082 policy_nodemask(gfp, pol));
David Rientjes212a0a62012-12-11 16:02:51 -08002083 if (unlikely(mpol_needs_cond_ref(pol)))
2084 __mpol_put(pol);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002085 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2086 goto retry_cpuset;
Miao Xiec0ff7452010-05-24 14:32:08 -07002087 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088}
2089
2090/**
2091 * alloc_pages_current - Allocate pages.
2092 *
2093 * @gfp:
2094 * %GFP_USER user allocation,
2095 * %GFP_KERNEL kernel allocation,
2096 * %GFP_HIGHMEM highmem allocation,
2097 * %GFP_FS don't call back into a file system.
2098 * %GFP_ATOMIC don't sleep.
2099 * @order: Power of two of allocation size in pages. 0 is a single page.
2100 *
2101 * Allocate a page from the kernel page pool. When not in
2102 * interrupt context and apply the current process NUMA policy.
2103 * Returns NULL when no page can be allocated.
2104 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08002105 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 * 1) it's ok to take cpuset_sem (can WAIT), and
2107 * 2) allocating for current task (not interrupt).
2108 */
Al Virodd0fc662005-10-07 07:46:04 +01002109struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110{
Mel Gorman5606e382012-11-02 18:19:13 +00002111 struct mempolicy *pol = get_task_policy(current);
Miao Xiec0ff7452010-05-24 14:32:08 -07002112 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002113 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114
Christoph Lameter9b819d22006-09-25 23:31:40 -07002115 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116 pol = &default_policy;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002117
Mel Gormancc9a6c82012-03-21 16:34:11 -07002118retry_cpuset:
2119 cpuset_mems_cookie = get_mems_allowed();
2120
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002121 /*
2122 * No reference counting needed for current->mempolicy
2123 * nor system default_policy
2124 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002125 if (pol->mode == MPOL_INTERLEAVE)
Miao Xiec0ff7452010-05-24 14:32:08 -07002126 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2127 else
2128 page = __alloc_pages_nodemask(gfp, order,
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002129 policy_zonelist(gfp, pol, numa_node_id()),
2130 policy_nodemask(gfp, pol));
Mel Gormancc9a6c82012-03-21 16:34:11 -07002131
2132 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2133 goto retry_cpuset;
2134
Miao Xiec0ff7452010-05-24 14:32:08 -07002135 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136}
2137EXPORT_SYMBOL(alloc_pages_current);
2138
Oleg Nesterovef0855d2013-09-11 14:20:14 -07002139int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2140{
2141 struct mempolicy *pol = mpol_dup(vma_policy(src));
2142
2143 if (IS_ERR(pol))
2144 return PTR_ERR(pol);
2145 dst->vm_policy = pol;
2146 return 0;
2147}
2148
Paul Jackson42253992006-01-08 01:01:59 -08002149/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002150 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08002151 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2152 * with the mems_allowed returned by cpuset_mems_allowed(). This
2153 * keeps mempolicies cpuset relative after its cpuset moves. See
2154 * further kernel/cpuset.c update_nodemask().
Miao Xie708c1bb2010-05-24 14:32:07 -07002155 *
2156 * current's mempolicy may be rebinded by the other task(the task that changes
2157 * cpuset's mems), so we needn't do rebind work for current task.
Paul Jackson42253992006-01-08 01:01:59 -08002158 */
Paul Jackson42253992006-01-08 01:01:59 -08002159
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002160/* Slow path of a mempolicy duplicate */
2161struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162{
2163 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2164
2165 if (!new)
2166 return ERR_PTR(-ENOMEM);
Miao Xie708c1bb2010-05-24 14:32:07 -07002167
2168 /* task's mempolicy is protected by alloc_lock */
2169 if (old == current->mempolicy) {
2170 task_lock(current);
2171 *new = *old;
2172 task_unlock(current);
2173 } else
2174 *new = *old;
2175
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08002176 rcu_read_lock();
Paul Jackson42253992006-01-08 01:01:59 -08002177 if (current_cpuset_is_being_rebound()) {
2178 nodemask_t mems = cpuset_mems_allowed(current);
Miao Xie708c1bb2010-05-24 14:32:07 -07002179 if (new->flags & MPOL_F_REBINDING)
2180 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2181 else
2182 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
Paul Jackson42253992006-01-08 01:01:59 -08002183 }
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08002184 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186 return new;
2187}
2188
2189/* Slow path of a mempolicy comparison */
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002190bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191{
2192 if (!a || !b)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002193 return false;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002194 if (a->mode != b->mode)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002195 return false;
Bob Liu19800502010-05-24 14:32:01 -07002196 if (a->flags != b->flags)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002197 return false;
Bob Liu19800502010-05-24 14:32:01 -07002198 if (mpol_store_user_nodemask(a))
2199 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002200 return false;
Bob Liu19800502010-05-24 14:32:01 -07002201
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002202 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07002203 case MPOL_BIND:
2204 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 case MPOL_INTERLEAVE:
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002206 return !!nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 case MPOL_PREFERRED:
Namhyung Kim75719662011-03-22 16:33:02 -07002208 return a->v.preferred_node == b->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 default:
2210 BUG();
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002211 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 }
2213}
2214
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 * Shared memory backing store policy support.
2217 *
2218 * Remember policies even when nobody has shared memory mapped.
2219 * The policies are kept in Red-Black tree linked from the inode.
2220 * They are protected by the sp->lock spinlock, which should be held
2221 * for any accesses to the tree.
2222 */
2223
2224/* lookup first element intersecting start-end */
Mel Gorman42288fe2012-12-21 23:10:25 +00002225/* Caller holds sp->lock */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226static struct sp_node *
2227sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2228{
2229 struct rb_node *n = sp->root.rb_node;
2230
2231 while (n) {
2232 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2233
2234 if (start >= p->end)
2235 n = n->rb_right;
2236 else if (end <= p->start)
2237 n = n->rb_left;
2238 else
2239 break;
2240 }
2241 if (!n)
2242 return NULL;
2243 for (;;) {
2244 struct sp_node *w = NULL;
2245 struct rb_node *prev = rb_prev(n);
2246 if (!prev)
2247 break;
2248 w = rb_entry(prev, struct sp_node, nd);
2249 if (w->end <= start)
2250 break;
2251 n = prev;
2252 }
2253 return rb_entry(n, struct sp_node, nd);
2254}
2255
2256/* Insert a new shared policy into the list. */
2257/* Caller holds sp->lock */
2258static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2259{
2260 struct rb_node **p = &sp->root.rb_node;
2261 struct rb_node *parent = NULL;
2262 struct sp_node *nd;
2263
2264 while (*p) {
2265 parent = *p;
2266 nd = rb_entry(parent, struct sp_node, nd);
2267 if (new->start < nd->start)
2268 p = &(*p)->rb_left;
2269 else if (new->end > nd->end)
2270 p = &(*p)->rb_right;
2271 else
2272 BUG();
2273 }
2274 rb_link_node(&new->nd, parent, p);
2275 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07002276 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002277 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278}
2279
2280/* Find shared policy intersecting idx */
2281struct mempolicy *
2282mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2283{
2284 struct mempolicy *pol = NULL;
2285 struct sp_node *sn;
2286
2287 if (!sp->root.rb_node)
2288 return NULL;
Mel Gorman42288fe2012-12-21 23:10:25 +00002289 spin_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 sn = sp_lookup(sp, idx, idx+1);
2291 if (sn) {
2292 mpol_get(sn->policy);
2293 pol = sn->policy;
2294 }
Mel Gorman42288fe2012-12-21 23:10:25 +00002295 spin_unlock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 return pol;
2297}
2298
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002299static void sp_free(struct sp_node *n)
2300{
2301 mpol_put(n->policy);
2302 kmem_cache_free(sn_cache, n);
2303}
2304
Rik van Rielde1c9ce2013-10-07 11:29:39 +01002305#ifdef CONFIG_NUMA_BALANCING
2306static bool numa_migrate_deferred(struct task_struct *p, int last_cpupid)
2307{
2308 /* Never defer a private fault */
2309 if (cpupid_match_pid(p, last_cpupid))
2310 return false;
2311
2312 if (p->numa_migrate_deferred) {
2313 p->numa_migrate_deferred--;
2314 return true;
2315 }
2316 return false;
2317}
2318
2319static inline void defer_numa_migrate(struct task_struct *p)
2320{
2321 p->numa_migrate_deferred = sysctl_numa_balancing_migrate_deferred;
2322}
2323#else
2324static inline bool numa_migrate_deferred(struct task_struct *p, int last_cpupid)
2325{
2326 return false;
2327}
2328
2329static inline void defer_numa_migrate(struct task_struct *p)
2330{
2331}
2332#endif /* CONFIG_NUMA_BALANCING */
2333
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002334/**
2335 * mpol_misplaced - check whether current page node is valid in policy
2336 *
2337 * @page - page to be checked
2338 * @vma - vm area where page mapped
2339 * @addr - virtual address where page mapped
2340 *
2341 * Lookup current policy node id for vma,addr and "compare to" page's
2342 * node id.
2343 *
2344 * Returns:
2345 * -1 - not misplaced, page is in the right node
2346 * node - node id where the page should be
2347 *
2348 * Policy determination "mimics" alloc_page_vma().
2349 * Called from fault path where we know the vma and faulting address.
2350 */
2351int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2352{
2353 struct mempolicy *pol;
2354 struct zone *zone;
2355 int curnid = page_to_nid(page);
2356 unsigned long pgoff;
Peter Zijlstra90572892013-10-07 11:29:20 +01002357 int thiscpu = raw_smp_processor_id();
2358 int thisnid = cpu_to_node(thiscpu);
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002359 int polnid = -1;
2360 int ret = -1;
2361
2362 BUG_ON(!vma);
2363
2364 pol = get_vma_policy(current, vma, addr);
2365 if (!(pol->flags & MPOL_F_MOF))
2366 goto out;
2367
2368 switch (pol->mode) {
2369 case MPOL_INTERLEAVE:
2370 BUG_ON(addr >= vma->vm_end);
2371 BUG_ON(addr < vma->vm_start);
2372
2373 pgoff = vma->vm_pgoff;
2374 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2375 polnid = offset_il_node(pol, vma, pgoff);
2376 break;
2377
2378 case MPOL_PREFERRED:
2379 if (pol->flags & MPOL_F_LOCAL)
2380 polnid = numa_node_id();
2381 else
2382 polnid = pol->v.preferred_node;
2383 break;
2384
2385 case MPOL_BIND:
2386 /*
2387 * allows binding to multiple nodes.
2388 * use current page if in policy nodemask,
2389 * else select nearest allowed node, if any.
2390 * If no allowed nodes, use current [!misplaced].
2391 */
2392 if (node_isset(curnid, pol->v.nodes))
2393 goto out;
2394 (void)first_zones_zonelist(
2395 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2396 gfp_zone(GFP_HIGHUSER),
2397 &pol->v.nodes, &zone);
2398 polnid = zone->node;
2399 break;
2400
2401 default:
2402 BUG();
2403 }
Mel Gorman5606e382012-11-02 18:19:13 +00002404
2405 /* Migrate the page towards the node whose CPU is referencing it */
Mel Gormane42c8ff2012-11-12 09:17:07 +00002406 if (pol->flags & MPOL_F_MORON) {
Peter Zijlstra90572892013-10-07 11:29:20 +01002407 int last_cpupid;
2408 int this_cpupid;
Mel Gormane42c8ff2012-11-12 09:17:07 +00002409
Peter Zijlstra90572892013-10-07 11:29:20 +01002410 polnid = thisnid;
2411 this_cpupid = cpu_pid_to_cpupid(thiscpu, current->pid);
Mel Gorman5606e382012-11-02 18:19:13 +00002412
Mel Gormane42c8ff2012-11-12 09:17:07 +00002413 /*
2414 * Multi-stage node selection is used in conjunction
2415 * with a periodic migration fault to build a temporal
2416 * task<->page relation. By using a two-stage filter we
2417 * remove short/unlikely relations.
2418 *
2419 * Using P(p) ~ n_p / n_t as per frequentist
2420 * probability, we can equate a task's usage of a
2421 * particular page (n_p) per total usage of this
2422 * page (n_t) (in a given time-span) to a probability.
2423 *
2424 * Our periodic faults will sample this probability and
2425 * getting the same result twice in a row, given these
2426 * samples are fully independent, is then given by
2427 * P(n)^2, provided our sample period is sufficiently
2428 * short compared to the usage pattern.
2429 *
2430 * This quadric squishes small probabilities, making
2431 * it less likely we act on an unlikely task<->page
2432 * relation.
2433 */
Peter Zijlstra90572892013-10-07 11:29:20 +01002434 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
Rik van Rielde1c9ce2013-10-07 11:29:39 +01002435 if (!cpupid_pid_unset(last_cpupid) && cpupid_to_nid(last_cpupid) != thisnid) {
2436
2437 /* See sysctl_numa_balancing_migrate_deferred comment */
2438 if (!cpupid_match_pid(current, last_cpupid))
2439 defer_numa_migrate(current);
2440
2441 goto out;
2442 }
2443
2444 /*
2445 * The quadratic filter above reduces extraneous migration
2446 * of shared pages somewhat. This code reduces it even more,
2447 * reducing the overhead of page migrations of shared pages.
2448 * This makes workloads with shared pages rely more on
2449 * "move task near its memory", and less on "move memory
2450 * towards its task", which is exactly what we want.
2451 */
2452 if (numa_migrate_deferred(current, last_cpupid))
Mel Gormane42c8ff2012-11-12 09:17:07 +00002453 goto out;
2454 }
2455
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002456 if (curnid != polnid)
2457 ret = polnid;
2458out:
2459 mpol_cond_put(pol);
2460
2461 return ret;
2462}
2463
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2465{
Paul Mundt140d5a42007-07-15 23:38:16 -07002466 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 rb_erase(&n->nd, &sp->root);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002468 sp_free(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469}
2470
Mel Gorman42288fe2012-12-21 23:10:25 +00002471static void sp_node_init(struct sp_node *node, unsigned long start,
2472 unsigned long end, struct mempolicy *pol)
2473{
2474 node->start = start;
2475 node->end = end;
2476 node->policy = pol;
2477}
2478
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07002479static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2480 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481{
KOSAKI Motohiro869833f22012-10-08 16:29:16 -07002482 struct sp_node *n;
2483 struct mempolicy *newpol;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484
KOSAKI Motohiro869833f22012-10-08 16:29:16 -07002485 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486 if (!n)
2487 return NULL;
KOSAKI Motohiro869833f22012-10-08 16:29:16 -07002488
2489 newpol = mpol_dup(pol);
2490 if (IS_ERR(newpol)) {
2491 kmem_cache_free(sn_cache, n);
2492 return NULL;
2493 }
2494 newpol->flags |= MPOL_F_SHARED;
Mel Gorman42288fe2012-12-21 23:10:25 +00002495 sp_node_init(n, start, end, newpol);
KOSAKI Motohiro869833f22012-10-08 16:29:16 -07002496
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 return n;
2498}
2499
2500/* Replace a policy range. */
2501static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2502 unsigned long end, struct sp_node *new)
2503{
Mel Gormanb22d1272012-10-08 16:29:17 -07002504 struct sp_node *n;
Mel Gorman42288fe2012-12-21 23:10:25 +00002505 struct sp_node *n_new = NULL;
2506 struct mempolicy *mpol_new = NULL;
Mel Gormanb22d1272012-10-08 16:29:17 -07002507 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508
Mel Gorman42288fe2012-12-21 23:10:25 +00002509restart:
2510 spin_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 n = sp_lookup(sp, start, end);
2512 /* Take care of old policies in the same range. */
2513 while (n && n->start < end) {
2514 struct rb_node *next = rb_next(&n->nd);
2515 if (n->start >= start) {
2516 if (n->end <= end)
2517 sp_delete(sp, n);
2518 else
2519 n->start = end;
2520 } else {
2521 /* Old policy spanning whole new range. */
2522 if (n->end > end) {
Mel Gorman42288fe2012-12-21 23:10:25 +00002523 if (!n_new)
2524 goto alloc_new;
2525
2526 *mpol_new = *n->policy;
2527 atomic_set(&mpol_new->refcnt, 1);
KOSAKI Motohiro78806392013-03-08 12:43:29 -08002528 sp_node_init(n_new, end, n->end, mpol_new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 n->end = start;
Hillf Danton5ca39572013-03-08 12:43:28 -08002530 sp_insert(sp, n_new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002531 n_new = NULL;
2532 mpol_new = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 break;
2534 } else
2535 n->end = start;
2536 }
2537 if (!next)
2538 break;
2539 n = rb_entry(next, struct sp_node, nd);
2540 }
2541 if (new)
2542 sp_insert(sp, new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002543 spin_unlock(&sp->lock);
2544 ret = 0;
2545
2546err_out:
2547 if (mpol_new)
2548 mpol_put(mpol_new);
2549 if (n_new)
2550 kmem_cache_free(sn_cache, n_new);
2551
Mel Gormanb22d1272012-10-08 16:29:17 -07002552 return ret;
Mel Gorman42288fe2012-12-21 23:10:25 +00002553
2554alloc_new:
2555 spin_unlock(&sp->lock);
2556 ret = -ENOMEM;
2557 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2558 if (!n_new)
2559 goto err_out;
2560 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2561 if (!mpol_new)
2562 goto err_out;
2563 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564}
2565
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002566/**
2567 * mpol_shared_policy_init - initialize shared policy for inode
2568 * @sp: pointer to inode shared policy
2569 * @mpol: struct mempolicy to install
2570 *
2571 * Install non-NULL @mpol in inode's shared policy rb-tree.
2572 * On entry, the current task has a reference on a non-NULL @mpol.
2573 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002574 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002575 */
2576void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08002577{
Miao Xie58568d22009-06-16 15:31:49 -07002578 int ret;
2579
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002580 sp->root = RB_ROOT; /* empty tree == default mempolicy */
Mel Gorman42288fe2012-12-21 23:10:25 +00002581 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08002582
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002583 if (mpol) {
2584 struct vm_area_struct pvma;
2585 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002586 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002587
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002588 if (!scratch)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002589 goto put_mpol;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002590 /* contextualize the tmpfs mount point mempolicy */
2591 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002592 if (IS_ERR(new))
Dan Carpenter0cae3452010-05-25 23:42:58 -07002593 goto free_scratch; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002594
2595 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002596 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002597 task_unlock(current);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002598 if (ret)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002599 goto put_new;
Robin Holt7339ff82006-01-14 13:20:48 -08002600
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002601 /* Create pseudo-vma that contains just the policy */
2602 memset(&pvma, 0, sizeof(struct vm_area_struct));
2603 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2604 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002605
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002606put_new:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002607 mpol_put(new); /* drop initial ref */
Dan Carpenter0cae3452010-05-25 23:42:58 -07002608free_scratch:
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002609 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002610put_mpol:
2611 mpol_put(mpol); /* drop our incoming ref on sb mpol */
Robin Holt7339ff82006-01-14 13:20:48 -08002612 }
2613}
2614
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615int mpol_set_shared_policy(struct shared_policy *info,
2616 struct vm_area_struct *vma, struct mempolicy *npol)
2617{
2618 int err;
2619 struct sp_node *new = NULL;
2620 unsigned long sz = vma_pages(vma);
2621
David Rientjes028fec42008-04-28 02:12:25 -07002622 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002624 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002625 npol ? npol->flags : -1,
David Rientjes00ef2d22013-02-22 16:35:36 -08002626 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627
2628 if (npol) {
2629 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2630 if (!new)
2631 return -ENOMEM;
2632 }
2633 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2634 if (err && new)
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002635 sp_free(new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 return err;
2637}
2638
2639/* Free a backing policy store on inode delete. */
2640void mpol_free_shared_policy(struct shared_policy *p)
2641{
2642 struct sp_node *n;
2643 struct rb_node *next;
2644
2645 if (!p->root.rb_node)
2646 return;
Mel Gorman42288fe2012-12-21 23:10:25 +00002647 spin_lock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 next = rb_first(&p->root);
2649 while (next) {
2650 n = rb_entry(next, struct sp_node, nd);
2651 next = rb_next(&n->nd);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002652 sp_delete(p, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 }
Mel Gorman42288fe2012-12-21 23:10:25 +00002654 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655}
2656
Mel Gorman1a687c22012-11-22 11:16:36 +00002657#ifdef CONFIG_NUMA_BALANCING
2658static bool __initdata numabalancing_override;
2659
2660static void __init check_numabalancing_enable(void)
2661{
2662 bool numabalancing_default = false;
2663
2664 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2665 numabalancing_default = true;
2666
2667 if (nr_node_ids > 1 && !numabalancing_override) {
2668 printk(KERN_INFO "Enabling automatic NUMA balancing. "
2669 "Configure with numa_balancing= or sysctl");
2670 set_numabalancing_state(numabalancing_default);
2671 }
2672}
2673
2674static int __init setup_numabalancing(char *str)
2675{
2676 int ret = 0;
2677 if (!str)
2678 goto out;
2679 numabalancing_override = true;
2680
2681 if (!strcmp(str, "enable")) {
2682 set_numabalancing_state(true);
2683 ret = 1;
2684 } else if (!strcmp(str, "disable")) {
2685 set_numabalancing_state(false);
2686 ret = 1;
2687 }
2688out:
2689 if (!ret)
2690 printk(KERN_WARNING "Unable to parse numa_balancing=\n");
2691
2692 return ret;
2693}
2694__setup("numa_balancing=", setup_numabalancing);
2695#else
2696static inline void __init check_numabalancing_enable(void)
2697{
2698}
2699#endif /* CONFIG_NUMA_BALANCING */
2700
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701/* assumes fs == KERNEL_DS */
2702void __init numa_policy_init(void)
2703{
Paul Mundtb71636e2007-07-15 23:38:15 -07002704 nodemask_t interleave_nodes;
2705 unsigned long largest = 0;
2706 int nid, prefer = 0;
2707
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 policy_cache = kmem_cache_create("numa_policy",
2709 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002710 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711
2712 sn_cache = kmem_cache_create("shared_policy_node",
2713 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002714 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715
Mel Gorman5606e382012-11-02 18:19:13 +00002716 for_each_node(nid) {
2717 preferred_node_policy[nid] = (struct mempolicy) {
2718 .refcnt = ATOMIC_INIT(1),
2719 .mode = MPOL_PREFERRED,
2720 .flags = MPOL_F_MOF | MPOL_F_MORON,
2721 .v = { .preferred_node = nid, },
2722 };
2723 }
2724
Paul Mundtb71636e2007-07-15 23:38:15 -07002725 /*
2726 * Set interleaving policy for system init. Interleaving is only
2727 * enabled across suitably sized nodes (default is >= 16MB), or
2728 * fall back to the largest node if they're all smaller.
2729 */
2730 nodes_clear(interleave_nodes);
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002731 for_each_node_state(nid, N_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002732 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733
Paul Mundtb71636e2007-07-15 23:38:15 -07002734 /* Preserve the largest node */
2735 if (largest < total_pages) {
2736 largest = total_pages;
2737 prefer = nid;
2738 }
2739
2740 /* Interleave this node? */
2741 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2742 node_set(nid, interleave_nodes);
2743 }
2744
2745 /* All too small, use the largest */
2746 if (unlikely(nodes_empty(interleave_nodes)))
2747 node_set(prefer, interleave_nodes);
2748
David Rientjes028fec42008-04-28 02:12:25 -07002749 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 printk("numa_policy_init: interleaving failed\n");
Mel Gorman1a687c22012-11-22 11:16:36 +00002751
2752 check_numabalancing_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753}
2754
Christoph Lameter8bccd852005-10-29 18:16:59 -07002755/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756void numa_default_policy(void)
2757{
David Rientjes028fec42008-04-28 02:12:25 -07002758 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759}
Paul Jackson68860ec2005-10-30 15:02:36 -08002760
Paul Jackson42253992006-01-08 01:01:59 -08002761/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002762 * Parse and format mempolicy from/to strings
2763 */
2764
2765/*
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002766 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002767 */
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002768static const char * const policy_modes[] =
2769{
2770 [MPOL_DEFAULT] = "default",
2771 [MPOL_PREFERRED] = "prefer",
2772 [MPOL_BIND] = "bind",
2773 [MPOL_INTERLEAVE] = "interleave",
Lee Schermerhornd3a71032012-10-25 14:16:29 +02002774 [MPOL_LOCAL] = "local",
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002775};
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002776
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002777
2778#ifdef CONFIG_TMPFS
2779/**
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002780 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002781 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002782 * @mpol: pointer to struct mempolicy pointer, returned on success.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002783 *
2784 * Format of input:
2785 * <mode>[=<flags>][:<nodelist>]
2786 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002787 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002788 */
Hugh Dickinsa7a88b22013-01-02 02:04:23 -08002789int mpol_parse_str(char *str, struct mempolicy **mpol)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002790{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002791 struct mempolicy *new = NULL;
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002792 unsigned short mode;
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002793 unsigned short mode_flags;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002794 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002795 char *nodelist = strchr(str, ':');
2796 char *flags = strchr(str, '=');
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002797 int err = 1;
2798
2799 if (nodelist) {
2800 /* NUL-terminate mode or flags string */
2801 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002802 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002803 goto out;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002804 if (!nodes_subset(nodes, node_states[N_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002805 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002806 } else
2807 nodes_clear(nodes);
2808
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002809 if (flags)
2810 *flags++ = '\0'; /* terminate mode string */
2811
Peter Zijlstra479e2802012-10-25 14:16:28 +02002812 for (mode = 0; mode < MPOL_MAX; mode++) {
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002813 if (!strcmp(str, policy_modes[mode])) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002814 break;
2815 }
2816 }
Mel Gormana7200942012-11-16 09:37:58 +00002817 if (mode >= MPOL_MAX)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002818 goto out;
2819
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002820 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002821 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002822 /*
2823 * Insist on a nodelist of one node only
2824 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002825 if (nodelist) {
2826 char *rest = nodelist;
2827 while (isdigit(*rest))
2828 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002829 if (*rest)
2830 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002831 }
2832 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002833 case MPOL_INTERLEAVE:
2834 /*
2835 * Default to online nodes with memory if no nodelist
2836 */
2837 if (!nodelist)
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002838 nodes = node_states[N_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002839 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002840 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002841 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002842 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002843 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002844 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002845 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002846 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002847 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002848 case MPOL_DEFAULT:
2849 /*
2850 * Insist on a empty nodelist
2851 */
2852 if (!nodelist)
2853 err = 0;
2854 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002855 case MPOL_BIND:
2856 /*
2857 * Insist on a nodelist
2858 */
2859 if (!nodelist)
2860 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002861 }
2862
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002863 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002864 if (flags) {
2865 /*
2866 * Currently, we only support two mutually exclusive
2867 * mode flags.
2868 */
2869 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002870 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002871 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002872 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002873 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002874 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002875 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002876
2877 new = mpol_new(mode, mode_flags, &nodes);
2878 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002879 goto out;
2880
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002881 /*
2882 * Save nodes for mpol_to_str() to show the tmpfs mount options
2883 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2884 */
2885 if (mode != MPOL_PREFERRED)
2886 new->v.nodes = nodes;
2887 else if (nodelist)
2888 new->v.preferred_node = first_node(nodes);
2889 else
2890 new->flags |= MPOL_F_LOCAL;
2891
2892 /*
2893 * Save nodes for contextualization: this will be used to "clone"
2894 * the mempolicy in a specific context [cpuset] at a later time.
2895 */
2896 new->w.user_nodemask = nodes;
2897
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002898 err = 0;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002899
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002900out:
2901 /* Restore string for error message */
2902 if (nodelist)
2903 *--nodelist = ':';
2904 if (flags)
2905 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002906 if (!err)
2907 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002908 return err;
2909}
2910#endif /* CONFIG_TMPFS */
2911
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002912/**
2913 * mpol_to_str - format a mempolicy structure for printing
2914 * @buffer: to contain formatted mempolicy string
2915 * @maxlen: length of @buffer
2916 * @pol: pointer to mempolicy to be formatted
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002917 *
David Rientjes948927e2013-11-12 15:07:28 -08002918 * Convert @pol into a string. If @buffer is too short, truncate the string.
2919 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2920 * longest flag, "relative", and to display at least a few node ids.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002921 */
David Rientjes948927e2013-11-12 15:07:28 -08002922void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002923{
2924 char *p = buffer;
David Rientjes948927e2013-11-12 15:07:28 -08002925 nodemask_t nodes = NODE_MASK_NONE;
2926 unsigned short mode = MPOL_DEFAULT;
2927 unsigned short flags = 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002928
David Rientjes948927e2013-11-12 15:07:28 -08002929 if (pol && pol != &default_policy) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002930 mode = pol->mode;
David Rientjes948927e2013-11-12 15:07:28 -08002931 flags = pol->flags;
2932 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002933
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002934 switch (mode) {
2935 case MPOL_DEFAULT:
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002936 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002937 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002938 if (flags & MPOL_F_LOCAL)
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002939 mode = MPOL_LOCAL;
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002940 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002941 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002942 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002943 case MPOL_BIND:
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002944 case MPOL_INTERLEAVE:
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002945 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002946 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002947 default:
David Rientjes948927e2013-11-12 15:07:28 -08002948 WARN_ON_ONCE(1);
2949 snprintf(p, maxlen, "unknown");
2950 return;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002951 }
2952
David Rientjes948927e2013-11-12 15:07:28 -08002953 p += snprintf(p, maxlen, policy_modes[mode]);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002954
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002955 if (flags & MPOL_MODE_FLAGS) {
David Rientjes948927e2013-11-12 15:07:28 -08002956 p += snprintf(p, buffer + maxlen - p, "=");
David Rientjesf5b087b2008-04-28 02:12:27 -07002957
Lee Schermerhorn22919902008-04-28 02:13:22 -07002958 /*
2959 * Currently, the only defined flags are mutually exclusive
2960 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002961 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002962 p += snprintf(p, buffer + maxlen - p, "static");
2963 else if (flags & MPOL_F_RELATIVE_NODES)
2964 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002965 }
2966
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002967 if (!nodes_empty(nodes)) {
David Rientjes948927e2013-11-12 15:07:28 -08002968 p += snprintf(p, buffer + maxlen - p, ":");
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002969 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2970 }
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002971}