blob: 747743237d9f4d3ead6117d4ee152c00659cd362 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
Christoph Lameter8bccd852005-10-29 18:16:59 -07005 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 * Subject to the GNU Public License, version 2.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
Christoph Lameter8bccd852005-10-29 18:16:59 -070021 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070022 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
Christoph Lameter8bccd852005-10-29 18:16:59 -070024 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070028 * preferred Try a specific node first before normal fallback.
David Rientjes00ef2d22013-02-22 16:35:36 -080029 * As a special case NUMA_NO_NODE here means do the allocation
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
Christoph Lameter8bccd852005-10-29 18:16:59 -070033 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070034 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
Linus Torvalds1da177e2005-04-16 15:20:36 -070066*/
67
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070068#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69
Linus Torvalds1da177e2005-04-16 15:20:36 -070070#include <linux/mempolicy.h>
71#include <linux/mm.h>
72#include <linux/highmem.h>
73#include <linux/hugetlb.h>
74#include <linux/kernel.h>
75#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070076#include <linux/nodemask.h>
77#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070078#include <linux/slab.h>
79#include <linux/string.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040080#include <linux/export.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070081#include <linux/nsproxy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070082#include <linux/interrupt.h>
83#include <linux/init.h>
84#include <linux/compat.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080085#include <linux/swap.h>
Christoph Lameter1a75a6c2006-01-08 01:01:02 -080086#include <linux/seq_file.h>
87#include <linux/proc_fs.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080088#include <linux/migrate.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080089#include <linux/ksm.h>
Christoph Lameter95a402c2006-06-23 02:03:53 -070090#include <linux/rmap.h>
David Quigley86c3a762006-06-23 02:04:02 -070091#include <linux/security.h>
Adrian Bunkdbcb0f12007-10-16 01:26:26 -070092#include <linux/syscalls.h>
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -070093#include <linux/ctype.h>
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -080094#include <linux/mm_inline.h>
Lee Schermerhornb24f53a2012-10-25 14:16:32 +020095#include <linux/mmu_notifier.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070096#include <linux/printk.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080097
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <asm/tlbflush.h>
99#include <asm/uaccess.h>
Michal Hocko778d3b02011-07-26 16:08:30 -0700100#include <linux/random.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101
Nick Piggin62695a82008-10-18 20:26:09 -0700102#include "internal.h"
103
Christoph Lameter38e35862006-01-08 01:01:01 -0800104/* Internal flags */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800105#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
Christoph Lameter38e35862006-01-08 01:01:01 -0800106#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800107
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800108static struct kmem_cache *policy_cache;
109static struct kmem_cache *sn_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111/* Highest zone. An specific allocation for a zone below that is not
112 policied. */
Christoph Lameter62672762007-02-10 01:43:07 -0800113enum zone_type policy_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700115/*
116 * run-time system-wide default policy => local allocation
117 */
H Hartley Sweetene754d792011-10-31 17:09:23 -0700118static struct mempolicy default_policy = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 .refcnt = ATOMIC_INIT(1), /* never free it */
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700120 .mode = MPOL_PREFERRED,
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700121 .flags = MPOL_F_LOCAL,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122};
123
Mel Gorman5606e382012-11-02 18:19:13 +0000124static struct mempolicy preferred_node_policy[MAX_NUMNODES];
125
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -0700126struct mempolicy *get_task_policy(struct task_struct *p)
Mel Gorman5606e382012-11-02 18:19:13 +0000127{
128 struct mempolicy *pol = p->mempolicy;
Oleg Nesterovf15ca782014-10-09 15:27:43 -0700129 int node;
Mel Gorman5606e382012-11-02 18:19:13 +0000130
Oleg Nesterovf15ca782014-10-09 15:27:43 -0700131 if (pol)
132 return pol;
Mel Gorman5606e382012-11-02 18:19:13 +0000133
Oleg Nesterovf15ca782014-10-09 15:27:43 -0700134 node = numa_node_id();
135 if (node != NUMA_NO_NODE) {
136 pol = &preferred_node_policy[node];
137 /* preferred_node_policy is not initialised early in boot */
138 if (pol->mode)
139 return pol;
Mel Gorman5606e382012-11-02 18:19:13 +0000140 }
141
Oleg Nesterovf15ca782014-10-09 15:27:43 -0700142 return &default_policy;
Mel Gorman5606e382012-11-02 18:19:13 +0000143}
144
David Rientjes37012942008-04-28 02:12:33 -0700145static const struct mempolicy_operations {
146 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
Miao Xie708c1bb2010-05-24 14:32:07 -0700147 /*
148 * If read-side task has no lock to protect task->mempolicy, write-side
149 * task will rebind the task->mempolicy by two step. The first step is
150 * setting all the newly nodes, and the second step is cleaning all the
151 * disallowed nodes. In this way, we can avoid finding no node to alloc
152 * page.
153 * If we have a lock to protect task->mempolicy in read-side, we do
154 * rebind directly.
155 *
156 * step:
157 * MPOL_REBIND_ONCE - do rebind work at once
158 * MPOL_REBIND_STEP1 - set all the newly nodes
159 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
160 */
161 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
162 enum mpol_rebind_step step);
David Rientjes37012942008-04-28 02:12:33 -0700163} mpol_ops[MPOL_MAX];
164
David Rientjesf5b087b2008-04-28 02:12:27 -0700165static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
166{
Bob Liu6d556292010-05-24 14:31:59 -0700167 return pol->flags & MPOL_MODE_FLAGS;
David Rientjes4c50bc02008-04-28 02:12:30 -0700168}
169
170static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
171 const nodemask_t *rel)
172{
173 nodemask_t tmp;
174 nodes_fold(tmp, *orig, nodes_weight(*rel));
175 nodes_onto(*ret, tmp, *rel);
David Rientjesf5b087b2008-04-28 02:12:27 -0700176}
177
David Rientjes37012942008-04-28 02:12:33 -0700178static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
179{
180 if (nodes_empty(*nodes))
181 return -EINVAL;
182 pol->v.nodes = *nodes;
183 return 0;
184}
185
186static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
187{
188 if (!nodes)
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700189 pol->flags |= MPOL_F_LOCAL; /* local allocation */
David Rientjes37012942008-04-28 02:12:33 -0700190 else if (nodes_empty(*nodes))
191 return -EINVAL; /* no allowed nodes */
192 else
193 pol->v.preferred_node = first_node(*nodes);
194 return 0;
195}
196
197static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
198{
Zhihui Zhang859f7ef2014-12-18 16:17:09 -0800199 if (nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700200 return -EINVAL;
201 pol->v.nodes = *nodes;
202 return 0;
203}
204
Miao Xie58568d22009-06-16 15:31:49 -0700205/*
206 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
207 * any, for the new policy. mpol_new() has already validated the nodes
208 * parameter with respect to the policy mode and flags. But, we need to
209 * handle an empty nodemask with MPOL_PREFERRED here.
210 *
211 * Must be called holding task's alloc_lock to protect task's mems_allowed
212 * and mempolicy. May also be called holding the mmap_semaphore for write.
213 */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700214static int mpol_set_nodemask(struct mempolicy *pol,
215 const nodemask_t *nodes, struct nodemask_scratch *nsc)
Miao Xie58568d22009-06-16 15:31:49 -0700216{
Miao Xie58568d22009-06-16 15:31:49 -0700217 int ret;
218
219 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
220 if (pol == NULL)
221 return 0;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -0800222 /* Check N_MEMORY */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700223 nodes_and(nsc->mask1,
Lai Jiangshan01f13bd2012-12-12 13:51:33 -0800224 cpuset_current_mems_allowed, node_states[N_MEMORY]);
Miao Xie58568d22009-06-16 15:31:49 -0700225
226 VM_BUG_ON(!nodes);
227 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
228 nodes = NULL; /* explicit local allocation */
229 else {
230 if (pol->flags & MPOL_F_RELATIVE_NODES)
Zhihui Zhang859f7ef2014-12-18 16:17:09 -0800231 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
Miao Xie58568d22009-06-16 15:31:49 -0700232 else
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700233 nodes_and(nsc->mask2, *nodes, nsc->mask1);
234
Miao Xie58568d22009-06-16 15:31:49 -0700235 if (mpol_store_user_nodemask(pol))
236 pol->w.user_nodemask = *nodes;
237 else
238 pol->w.cpuset_mems_allowed =
239 cpuset_current_mems_allowed;
240 }
241
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700242 if (nodes)
243 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
244 else
245 ret = mpol_ops[pol->mode].create(pol, NULL);
Miao Xie58568d22009-06-16 15:31:49 -0700246 return ret;
247}
248
249/*
250 * This function just creates a new policy, does some check and simple
251 * initialization. You must invoke mpol_set_nodemask() to set nodes.
252 */
David Rientjes028fec42008-04-28 02:12:25 -0700253static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
254 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255{
256 struct mempolicy *policy;
257
David Rientjes028fec42008-04-28 02:12:25 -0700258 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
David Rientjes00ef2d22013-02-22 16:35:36 -0800259 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
Paul Mundt140d5a42007-07-15 23:38:16 -0700260
David Rientjes3e1f0642008-04-28 02:12:34 -0700261 if (mode == MPOL_DEFAULT) {
262 if (nodes && !nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700263 return ERR_PTR(-EINVAL);
Lee Schermerhornd3a71032012-10-25 14:16:29 +0200264 return NULL;
David Rientjes37012942008-04-28 02:12:33 -0700265 }
David Rientjes3e1f0642008-04-28 02:12:34 -0700266 VM_BUG_ON(!nodes);
267
268 /*
269 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
270 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
271 * All other modes require a valid pointer to a non-empty nodemask.
272 */
273 if (mode == MPOL_PREFERRED) {
274 if (nodes_empty(*nodes)) {
275 if (((flags & MPOL_F_STATIC_NODES) ||
276 (flags & MPOL_F_RELATIVE_NODES)))
277 return ERR_PTR(-EINVAL);
David Rientjes3e1f0642008-04-28 02:12:34 -0700278 }
Peter Zijlstra479e2802012-10-25 14:16:28 +0200279 } else if (mode == MPOL_LOCAL) {
280 if (!nodes_empty(*nodes))
281 return ERR_PTR(-EINVAL);
282 mode = MPOL_PREFERRED;
David Rientjes3e1f0642008-04-28 02:12:34 -0700283 } else if (nodes_empty(*nodes))
284 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
286 if (!policy)
287 return ERR_PTR(-ENOMEM);
288 atomic_set(&policy->refcnt, 1);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700289 policy->mode = mode;
David Rientjes3e1f0642008-04-28 02:12:34 -0700290 policy->flags = flags;
David Rientjesf5b087b2008-04-28 02:12:27 -0700291
David Rientjes37012942008-04-28 02:12:33 -0700292 return policy;
293}
294
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700295/* Slow path of a mpol destructor. */
296void __mpol_put(struct mempolicy *p)
297{
298 if (!atomic_dec_and_test(&p->refcnt))
299 return;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700300 kmem_cache_free(policy_cache, p);
301}
302
Miao Xie708c1bb2010-05-24 14:32:07 -0700303static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700305{
306}
307
Miao Xie708c1bb2010-05-24 14:32:07 -0700308/*
309 * step:
310 * MPOL_REBIND_ONCE - do rebind work at once
311 * MPOL_REBIND_STEP1 - set all the newly nodes
312 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
313 */
314static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
315 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700316{
317 nodemask_t tmp;
318
319 if (pol->flags & MPOL_F_STATIC_NODES)
320 nodes_and(tmp, pol->w.user_nodemask, *nodes);
321 else if (pol->flags & MPOL_F_RELATIVE_NODES)
322 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
323 else {
Miao Xie708c1bb2010-05-24 14:32:07 -0700324 /*
325 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
326 * result
327 */
328 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
329 nodes_remap(tmp, pol->v.nodes,
330 pol->w.cpuset_mems_allowed, *nodes);
331 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
332 } else if (step == MPOL_REBIND_STEP2) {
333 tmp = pol->w.cpuset_mems_allowed;
334 pol->w.cpuset_mems_allowed = *nodes;
335 } else
336 BUG();
David Rientjes37012942008-04-28 02:12:33 -0700337 }
338
Miao Xie708c1bb2010-05-24 14:32:07 -0700339 if (nodes_empty(tmp))
340 tmp = *nodes;
341
342 if (step == MPOL_REBIND_STEP1)
343 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
344 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
345 pol->v.nodes = tmp;
346 else
347 BUG();
348
David Rientjes37012942008-04-28 02:12:33 -0700349 if (!node_isset(current->il_next, tmp)) {
350 current->il_next = next_node(current->il_next, tmp);
351 if (current->il_next >= MAX_NUMNODES)
352 current->il_next = first_node(tmp);
353 if (current->il_next >= MAX_NUMNODES)
354 current->il_next = numa_node_id();
355 }
356}
357
358static void mpol_rebind_preferred(struct mempolicy *pol,
Miao Xie708c1bb2010-05-24 14:32:07 -0700359 const nodemask_t *nodes,
360 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700361{
362 nodemask_t tmp;
363
David Rientjes37012942008-04-28 02:12:33 -0700364 if (pol->flags & MPOL_F_STATIC_NODES) {
365 int node = first_node(pol->w.user_nodemask);
366
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700367 if (node_isset(node, *nodes)) {
David Rientjes37012942008-04-28 02:12:33 -0700368 pol->v.preferred_node = node;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700369 pol->flags &= ~MPOL_F_LOCAL;
370 } else
371 pol->flags |= MPOL_F_LOCAL;
David Rientjes37012942008-04-28 02:12:33 -0700372 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
373 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
374 pol->v.preferred_node = first_node(tmp);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700375 } else if (!(pol->flags & MPOL_F_LOCAL)) {
David Rientjes37012942008-04-28 02:12:33 -0700376 pol->v.preferred_node = node_remap(pol->v.preferred_node,
377 pol->w.cpuset_mems_allowed,
378 *nodes);
379 pol->w.cpuset_mems_allowed = *nodes;
380 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381}
382
Miao Xie708c1bb2010-05-24 14:32:07 -0700383/*
384 * mpol_rebind_policy - Migrate a policy to a different set of nodes
385 *
386 * If read-side task has no lock to protect task->mempolicy, write-side
387 * task will rebind the task->mempolicy by two step. The first step is
388 * setting all the newly nodes, and the second step is cleaning all the
389 * disallowed nodes. In this way, we can avoid finding no node to alloc
390 * page.
391 * If we have a lock to protect task->mempolicy in read-side, we do
392 * rebind directly.
393 *
394 * step:
395 * MPOL_REBIND_ONCE - do rebind work at once
396 * MPOL_REBIND_STEP1 - set all the newly nodes
397 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
398 */
399static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
400 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700401{
David Rientjes1d0d2682008-04-28 02:12:32 -0700402 if (!pol)
403 return;
Wang Sheng-Hui89c522c2012-05-29 15:06:16 -0700404 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
David Rientjes1d0d2682008-04-28 02:12:32 -0700405 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
406 return;
Miao Xie708c1bb2010-05-24 14:32:07 -0700407
408 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
409 return;
410
411 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
412 BUG();
413
414 if (step == MPOL_REBIND_STEP1)
415 pol->flags |= MPOL_F_REBINDING;
416 else if (step == MPOL_REBIND_STEP2)
417 pol->flags &= ~MPOL_F_REBINDING;
418 else if (step >= MPOL_REBIND_NSTEP)
419 BUG();
420
421 mpol_ops[pol->mode].rebind(pol, newmask, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700422}
423
424/*
425 * Wrapper for mpol_rebind_policy() that just requires task
426 * pointer, and updates task mempolicy.
Miao Xie58568d22009-06-16 15:31:49 -0700427 *
428 * Called with task's alloc_lock held.
David Rientjes1d0d2682008-04-28 02:12:32 -0700429 */
430
Miao Xie708c1bb2010-05-24 14:32:07 -0700431void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
432 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700433{
Miao Xie708c1bb2010-05-24 14:32:07 -0700434 mpol_rebind_policy(tsk->mempolicy, new, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700435}
436
437/*
438 * Rebind each vma in mm to new nodemask.
439 *
440 * Call holding a reference to mm. Takes mm->mmap_sem during call.
441 */
442
443void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
444{
445 struct vm_area_struct *vma;
446
447 down_write(&mm->mmap_sem);
448 for (vma = mm->mmap; vma; vma = vma->vm_next)
Miao Xie708c1bb2010-05-24 14:32:07 -0700449 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
David Rientjes1d0d2682008-04-28 02:12:32 -0700450 up_write(&mm->mmap_sem);
451}
452
David Rientjes37012942008-04-28 02:12:33 -0700453static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
454 [MPOL_DEFAULT] = {
455 .rebind = mpol_rebind_default,
456 },
457 [MPOL_INTERLEAVE] = {
458 .create = mpol_new_interleave,
459 .rebind = mpol_rebind_nodemask,
460 },
461 [MPOL_PREFERRED] = {
462 .create = mpol_new_preferred,
463 .rebind = mpol_rebind_preferred,
464 },
465 [MPOL_BIND] = {
466 .create = mpol_new_bind,
467 .rebind = mpol_rebind_nodemask,
468 },
469};
470
Christoph Lameterfc301282006-01-18 17:42:29 -0800471static void migrate_page_add(struct page *page, struct list_head *pagelist,
472 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800473
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800474struct queue_pages {
475 struct list_head *pagelist;
476 unsigned long flags;
477 nodemask_t *nmask;
478 struct vm_area_struct *prev;
479};
480
Naoya Horiguchi98094942013-09-11 14:22:14 -0700481/*
482 * Scan through pages checking if pages follow certain conditions,
483 * and move them to the pagelist if they do.
484 */
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800485static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
486 unsigned long end, struct mm_walk *walk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487{
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800488 struct vm_area_struct *vma = walk->vma;
489 struct page *page;
490 struct queue_pages *qp = walk->private;
491 unsigned long flags = qp->flags;
492 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700493 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700494 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700495
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800496 split_huge_page_pmd(vma, addr, pmd);
497 if (pmd_trans_unstable(pmd))
498 return 0;
Hugh Dickins91612e02005-06-21 17:15:07 -0700499
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800500 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
501 for (; addr != end; pte++, addr += PAGE_SIZE) {
Hugh Dickins91612e02005-06-21 17:15:07 -0700502 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800504 page = vm_normal_page(vma, addr, *pte);
505 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800507 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800508 * vm_normal_page() filters out zero pages, but there might
509 * still be PageReserved pages to skip, perhaps in a VDSO.
Nick Piggin053837f2006-01-18 17:42:27 -0800510 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800511 if (PageReserved(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800512 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800513 nid = page_to_nid(page);
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800514 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
Christoph Lameter38e35862006-01-08 01:01:01 -0800515 continue;
516
Stephen Wilsonb1f72d12011-05-24 17:12:43 -0700517 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800518 migrate_page_add(page, qp->pagelist, flags);
519 }
520 pte_unmap_unlock(pte - 1, ptl);
521 cond_resched();
522 return 0;
Hugh Dickins91612e02005-06-21 17:15:07 -0700523}
524
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800525static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
526 unsigned long addr, unsigned long end,
527 struct mm_walk *walk)
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700528{
529#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800530 struct queue_pages *qp = walk->private;
531 unsigned long flags = qp->flags;
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700532 int nid;
533 struct page *page;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800534 spinlock_t *ptl;
Naoya Horiguchid4c54912014-06-06 10:00:01 -0400535 pte_t entry;
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700536
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800537 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
538 entry = huge_ptep_get(pte);
Naoya Horiguchid4c54912014-06-06 10:00:01 -0400539 if (!pte_present(entry))
540 goto unlock;
541 page = pte_page(entry);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700542 nid = page_to_nid(page);
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800543 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700544 goto unlock;
545 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
546 if (flags & (MPOL_MF_MOVE_ALL) ||
547 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800548 isolate_huge_page(page, qp->pagelist);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700549unlock:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800550 spin_unlock(ptl);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700551#else
552 BUG();
553#endif
Hugh Dickins91612e02005-06-21 17:15:07 -0700554 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555}
556
Aneesh Kumar K.V58772312013-12-06 00:08:22 +0530557#ifdef CONFIG_NUMA_BALANCING
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200558/*
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200559 * This is used to mark a range of virtual addresses to be inaccessible.
560 * These are later cleared by a NUMA hinting fault. Depending on these
561 * faults, pages may be migrated for better NUMA placement.
562 *
563 * This is assuming that NUMA faults are handled using PROT_NONE. If
564 * an architecture makes a different choice, it will need further
565 * changes to the core.
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200566 */
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200567unsigned long change_prot_numa(struct vm_area_struct *vma,
568 unsigned long addr, unsigned long end)
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200569{
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200570 int nr_updated;
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200571
Mel Gorman4d942462015-02-12 14:58:28 -0800572 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
Mel Gorman03c5a6e2012-11-02 14:52:48 +0000573 if (nr_updated)
574 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200575
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200576 return nr_updated;
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200577}
578#else
579static unsigned long change_prot_numa(struct vm_area_struct *vma,
580 unsigned long addr, unsigned long end)
581{
582 return 0;
583}
Aneesh Kumar K.V58772312013-12-06 00:08:22 +0530584#endif /* CONFIG_NUMA_BALANCING */
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200585
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800586static int queue_pages_test_walk(unsigned long start, unsigned long end,
587 struct mm_walk *walk)
588{
589 struct vm_area_struct *vma = walk->vma;
590 struct queue_pages *qp = walk->private;
591 unsigned long endvma = vma->vm_end;
592 unsigned long flags = qp->flags;
593
Naoya Horiguchi48684a62015-02-11 15:28:06 -0800594 if (vma->vm_flags & VM_PFNMAP)
595 return 1;
596
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800597 if (endvma > end)
598 endvma = end;
599 if (vma->vm_start > start)
600 start = vma->vm_start;
601
602 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
603 if (!vma->vm_next && vma->vm_end < end)
604 return -EFAULT;
605 if (qp->prev && qp->prev->vm_end < vma->vm_start)
606 return -EFAULT;
607 }
608
609 qp->prev = vma;
610
611 if (vma->vm_flags & VM_PFNMAP)
612 return 1;
613
614 if (flags & MPOL_MF_LAZY) {
615 /* Similar to task_numa_work, skip inaccessible VMAs */
616 if (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))
617 change_prot_numa(vma, start, endvma);
618 return 1;
619 }
620
621 if ((flags & MPOL_MF_STRICT) ||
622 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
623 vma_migratable(vma)))
624 /* queue pages from current vma */
625 return 0;
626 return 1;
627}
628
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800629/*
Naoya Horiguchi98094942013-09-11 14:22:14 -0700630 * Walk through page tables and collect pages to be migrated.
631 *
632 * If pages found in a given range are on a set of nodes (determined by
633 * @nodes and @flags,) it's isolated and queued to the pagelist which is
634 * passed via @private.)
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800635 */
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700636static int
Naoya Horiguchi98094942013-09-11 14:22:14 -0700637queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800638 nodemask_t *nodes, unsigned long flags,
639 struct list_head *pagelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640{
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800641 struct queue_pages qp = {
642 .pagelist = pagelist,
643 .flags = flags,
644 .nmask = nodes,
645 .prev = NULL,
646 };
647 struct mm_walk queue_pages_walk = {
648 .hugetlb_entry = queue_pages_hugetlb,
649 .pmd_entry = queue_pages_pte_range,
650 .test_walk = queue_pages_test_walk,
651 .mm = mm,
652 .private = &qp,
653 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800655 return walk_page_range(start, end, &queue_pages_walk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656}
657
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700658/*
659 * Apply policy to a single VMA
660 * This must be called with the mmap_sem held for writing.
661 */
662static int vma_replace_policy(struct vm_area_struct *vma,
663 struct mempolicy *pol)
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700664{
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700665 int err;
666 struct mempolicy *old;
667 struct mempolicy *new;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700668
669 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
670 vma->vm_start, vma->vm_end, vma->vm_pgoff,
671 vma->vm_ops, vma->vm_file,
672 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
673
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700674 new = mpol_dup(pol);
675 if (IS_ERR(new))
676 return PTR_ERR(new);
677
678 if (vma->vm_ops && vma->vm_ops->set_policy) {
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700679 err = vma->vm_ops->set_policy(vma, new);
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700680 if (err)
681 goto err_out;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700682 }
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700683
684 old = vma->vm_policy;
685 vma->vm_policy = new; /* protected by mmap_sem */
686 mpol_put(old);
687
688 return 0;
689 err_out:
690 mpol_put(new);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700691 return err;
692}
693
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800695static int mbind_range(struct mm_struct *mm, unsigned long start,
696 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697{
698 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800699 struct vm_area_struct *prev;
700 struct vm_area_struct *vma;
701 int err = 0;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800702 pgoff_t pgoff;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800703 unsigned long vmstart;
704 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
Linus Torvalds097d5912012-03-06 18:23:36 -0800706 vma = find_vma(mm, start);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800707 if (!vma || vma->vm_start > start)
708 return -EFAULT;
709
Linus Torvalds097d5912012-03-06 18:23:36 -0800710 prev = vma->vm_prev;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800711 if (start > vma->vm_start)
712 prev = vma;
713
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800714 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800716 vmstart = max(start, vma->vm_start);
717 vmend = min(end, vma->vm_end);
718
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800719 if (mpol_equal(vma_policy(vma), new_pol))
720 continue;
721
722 pgoff = vma->vm_pgoff +
723 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800724 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800725 vma->anon_vma, vma->vm_file, pgoff,
Caspar Zhang8aacc9f2011-09-14 16:20:58 -0700726 new_pol);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800727 if (prev) {
728 vma = prev;
729 next = vma->vm_next;
Oleg Nesterov3964acd2013-07-31 13:53:28 -0700730 if (mpol_equal(vma_policy(vma), new_pol))
731 continue;
732 /* vma_merge() joined vma && vma->next, case 8 */
733 goto replace;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800734 }
735 if (vma->vm_start != vmstart) {
736 err = split_vma(vma->vm_mm, vma, vmstart, 1);
737 if (err)
738 goto out;
739 }
740 if (vma->vm_end != vmend) {
741 err = split_vma(vma->vm_mm, vma, vmend, 0);
742 if (err)
743 goto out;
744 }
Oleg Nesterov3964acd2013-07-31 13:53:28 -0700745 replace:
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700746 err = vma_replace_policy(vma, new_pol);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700747 if (err)
748 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800750
751 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 return err;
753}
754
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700756static long do_set_mempolicy(unsigned short mode, unsigned short flags,
757 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758{
Miao Xie58568d22009-06-16 15:31:49 -0700759 struct mempolicy *new, *old;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700760 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700761 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700763 if (!scratch)
764 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700765
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700766 new = mpol_new(mode, flags, nodes);
767 if (IS_ERR(new)) {
768 ret = PTR_ERR(new);
769 goto out;
770 }
Oleg Nesterov2c7c3a72014-10-09 15:27:55 -0700771
Miao Xie58568d22009-06-16 15:31:49 -0700772 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700773 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700774 if (ret) {
775 task_unlock(current);
Miao Xie58568d22009-06-16 15:31:49 -0700776 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700777 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700778 }
779 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 current->mempolicy = new;
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700781 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700782 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700783 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700784 task_unlock(current);
Miao Xie58568d22009-06-16 15:31:49 -0700785 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700786 ret = 0;
787out:
788 NODEMASK_SCRATCH_FREE(scratch);
789 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790}
791
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700792/*
793 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700794 *
795 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700796 */
797static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700799 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700800 if (p == &default_policy)
801 return;
802
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700803 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700804 case MPOL_BIND:
805 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700807 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 break;
809 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700810 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700811 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700812 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 break;
814 default:
815 BUG();
816 }
817}
818
819static int lookup_node(struct mm_struct *mm, unsigned long addr)
820{
821 struct page *p;
822 int err;
823
824 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
825 if (err >= 0) {
826 err = page_to_nid(p);
827 put_page(p);
828 }
829 return err;
830}
831
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700833static long do_get_mempolicy(int *policy, nodemask_t *nmask,
834 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700836 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 struct mm_struct *mm = current->mm;
838 struct vm_area_struct *vma = NULL;
839 struct mempolicy *pol = current->mempolicy;
840
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700841 if (flags &
842 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700844
845 if (flags & MPOL_F_MEMS_ALLOWED) {
846 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
847 return -EINVAL;
848 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700849 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700850 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700851 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700852 return 0;
853 }
854
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700856 /*
857 * Do NOT fall back to task policy if the
858 * vma/shared policy at addr is NULL. We
859 * want to return MPOL_DEFAULT in this case.
860 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 down_read(&mm->mmap_sem);
862 vma = find_vma_intersection(mm, addr, addr+1);
863 if (!vma) {
864 up_read(&mm->mmap_sem);
865 return -EFAULT;
866 }
867 if (vma->vm_ops && vma->vm_ops->get_policy)
868 pol = vma->vm_ops->get_policy(vma, addr);
869 else
870 pol = vma->vm_policy;
871 } else if (addr)
872 return -EINVAL;
873
874 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700875 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876
877 if (flags & MPOL_F_NODE) {
878 if (flags & MPOL_F_ADDR) {
879 err = lookup_node(mm, addr);
880 if (err < 0)
881 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700882 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700884 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700885 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 } else {
887 err = -EINVAL;
888 goto out;
889 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700890 } else {
891 *policy = pol == &default_policy ? MPOL_DEFAULT :
892 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700893 /*
894 * Internal mempolicy flags must be masked off before exposing
895 * the policy to userspace.
896 */
897 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700898 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899
900 if (vma) {
901 up_read(&current->mm->mmap_sem);
902 vma = NULL;
903 }
904
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700906 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700907 if (mpol_store_user_nodemask(pol)) {
908 *nmask = pol->w.user_nodemask;
909 } else {
910 task_lock(current);
911 get_policy_nodemask(pol, nmask);
912 task_unlock(current);
913 }
Miao Xie58568d22009-06-16 15:31:49 -0700914 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915
916 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700917 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 if (vma)
919 up_read(&current->mm->mmap_sem);
920 return err;
921}
922
Christoph Lameterb20a3502006-03-22 00:09:12 -0800923#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700924/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800925 * page migration
926 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800927static void migrate_page_add(struct page *page, struct list_head *pagelist,
928 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800929{
930 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800931 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800932 */
Nick Piggin62695a82008-10-18 20:26:09 -0700933 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
934 if (!isolate_lru_page(page)) {
935 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800936 inc_zone_page_state(page, NR_ISOLATED_ANON +
937 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700938 }
939 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800940}
941
Christoph Lameter742755a2006-06-23 02:03:55 -0700942static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700943{
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700944 if (PageHuge(page))
945 return alloc_huge_page_node(page_hstate(compound_head(page)),
946 node);
947 else
David Rientjesb360edb2015-04-14 15:46:52 -0700948 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE |
949 __GFP_THISNODE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700950}
951
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800952/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800953 * Migrate pages from one node to a target node.
954 * Returns error or the number of pages not migrated.
955 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700956static int migrate_to_node(struct mm_struct *mm, int source, int dest,
957 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800958{
959 nodemask_t nmask;
960 LIST_HEAD(pagelist);
961 int err = 0;
962
963 nodes_clear(nmask);
964 node_set(source, nmask);
965
Minchan Kim08270802012-10-08 16:33:38 -0700966 /*
967 * This does not "check" the range but isolates all pages that
968 * need migration. Between passing in the full user address
969 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
970 */
971 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
Naoya Horiguchi98094942013-09-11 14:22:14 -0700972 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800973 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
974
Minchan Kimcf608ac2010-10-26 14:21:29 -0700975 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -0700976 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800977 MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -0700978 if (err)
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700979 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -0700980 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700981
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800982 return err;
983}
984
985/*
986 * Move pages between the two nodesets so as to preserve the physical
987 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -0800988 *
989 * Returns the number of page that could not be moved.
990 */
Andrew Morton0ce72d42012-05-29 15:06:24 -0700991int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
992 const nodemask_t *to, int flags)
Christoph Lameter39743882006-01-08 01:00:51 -0800993{
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800994 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800995 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800996 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -0800997
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800998 err = migrate_prep();
999 if (err)
1000 return err;
1001
Lee Schermerhorn53f25562008-04-28 02:13:20 -07001002 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -08001003
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001004 /*
1005 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1006 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1007 * bit in 'tmp', and return that <source, dest> pair for migration.
1008 * The pair of nodemasks 'to' and 'from' define the map.
1009 *
1010 * If no pair of bits is found that way, fallback to picking some
1011 * pair of 'source' and 'dest' bits that are not the same. If the
1012 * 'source' and 'dest' bits are the same, this represents a node
1013 * that will be migrating to itself, so no pages need move.
1014 *
1015 * If no bits are left in 'tmp', or if all remaining bits left
1016 * in 'tmp' correspond to the same bit in 'to', return false
1017 * (nothing left to migrate).
1018 *
1019 * This lets us pick a pair of nodes to migrate between, such that
1020 * if possible the dest node is not already occupied by some other
1021 * source node, minimizing the risk of overloading the memory on a
1022 * node that would happen if we migrated incoming memory to a node
1023 * before migrating outgoing memory source that same node.
1024 *
1025 * A single scan of tmp is sufficient. As we go, we remember the
1026 * most recent <s, d> pair that moved (s != d). If we find a pair
1027 * that not only moved, but what's better, moved to an empty slot
1028 * (d is not set in tmp), then we break out then, with that pair.
Justin P. Mattockae0e47f2011-03-01 15:06:02 +01001029 * Otherwise when we finish scanning from_tmp, we at least have the
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001030 * most recent <s, d> pair that moved. If we get all the way through
1031 * the scan of tmp without finding any node that moved, much less
1032 * moved to an empty node, then there is nothing left worth migrating.
1033 */
Christoph Lameterd4984712006-01-08 01:00:55 -08001034
Andrew Morton0ce72d42012-05-29 15:06:24 -07001035 tmp = *from;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001036 while (!nodes_empty(tmp)) {
1037 int s,d;
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001038 int source = NUMA_NO_NODE;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001039 int dest = 0;
1040
1041 for_each_node_mask(s, tmp) {
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001042
1043 /*
1044 * do_migrate_pages() tries to maintain the relative
1045 * node relationship of the pages established between
1046 * threads and memory areas.
1047 *
1048 * However if the number of source nodes is not equal to
1049 * the number of destination nodes we can not preserve
1050 * this node relative relationship. In that case, skip
1051 * copying memory from a node that is in the destination
1052 * mask.
1053 *
1054 * Example: [2,3,4] -> [3,4,5] moves everything.
1055 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1056 */
1057
Andrew Morton0ce72d42012-05-29 15:06:24 -07001058 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1059 (node_isset(s, *to)))
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001060 continue;
1061
Andrew Morton0ce72d42012-05-29 15:06:24 -07001062 d = node_remap(s, *from, *to);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001063 if (s == d)
1064 continue;
1065
1066 source = s; /* Node moved. Memorize */
1067 dest = d;
1068
1069 /* dest not in remaining from nodes? */
1070 if (!node_isset(dest, tmp))
1071 break;
1072 }
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001073 if (source == NUMA_NO_NODE)
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001074 break;
1075
1076 node_clear(source, tmp);
1077 err = migrate_to_node(mm, source, dest, flags);
1078 if (err > 0)
1079 busy += err;
1080 if (err < 0)
1081 break;
Christoph Lameter39743882006-01-08 01:00:51 -08001082 }
1083 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001084 if (err < 0)
1085 return err;
1086 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001087
Christoph Lameter39743882006-01-08 01:00:51 -08001088}
1089
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001090/*
1091 * Allocate a new page for page migration based on vma policy.
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001092 * Start by assuming the page is mapped by the same vma as contains @start.
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001093 * Search forward from there, if not. N.B., this assumes that the
1094 * list of pages handed to migrate_pages()--which is how we get here--
1095 * is in virtual address order.
1096 */
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001097static struct page *new_page(struct page *page, unsigned long start, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001098{
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001099 struct vm_area_struct *vma;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001100 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001101
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001102 vma = find_vma(current->mm, start);
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001103 while (vma) {
1104 address = page_address_in_vma(page, vma);
1105 if (address != -EFAULT)
1106 break;
1107 vma = vma->vm_next;
1108 }
1109
Wanpeng Li11c731e2013-12-18 17:08:56 -08001110 if (PageHuge(page)) {
Michal Hockocc817172014-01-23 15:53:15 -08001111 BUG_ON(!vma);
1112 return alloc_huge_page_noerr(vma, address, 1);
Wanpeng Li11c731e2013-12-18 17:08:56 -08001113 }
1114 /*
1115 * if !vma, alloc_page_vma() will use task or system default policy
1116 */
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001117 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001118}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001119#else
1120
1121static void migrate_page_add(struct page *page, struct list_head *pagelist,
1122 unsigned long flags)
1123{
1124}
1125
Andrew Morton0ce72d42012-05-29 15:06:24 -07001126int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1127 const nodemask_t *to, int flags)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001128{
1129 return -ENOSYS;
1130}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001131
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001132static struct page *new_page(struct page *page, unsigned long start, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001133{
1134 return NULL;
1135}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001136#endif
1137
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001138static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001139 unsigned short mode, unsigned short mode_flags,
1140 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001141{
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001142 struct mm_struct *mm = current->mm;
1143 struct mempolicy *new;
1144 unsigned long end;
1145 int err;
1146 LIST_HEAD(pagelist);
1147
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001148 if (flags & ~(unsigned long)MPOL_MF_VALID)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001149 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001150 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001151 return -EPERM;
1152
1153 if (start & ~PAGE_MASK)
1154 return -EINVAL;
1155
1156 if (mode == MPOL_DEFAULT)
1157 flags &= ~MPOL_MF_STRICT;
1158
1159 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1160 end = start + len;
1161
1162 if (end < start)
1163 return -EINVAL;
1164 if (end == start)
1165 return 0;
1166
David Rientjes028fec42008-04-28 02:12:25 -07001167 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001168 if (IS_ERR(new))
1169 return PTR_ERR(new);
1170
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001171 if (flags & MPOL_MF_LAZY)
1172 new->flags |= MPOL_F_MOF;
1173
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001174 /*
1175 * If we are using the default policy then operation
1176 * on discontinuous address spaces is okay after all
1177 */
1178 if (!new)
1179 flags |= MPOL_MF_DISCONTIG_OK;
1180
David Rientjes028fec42008-04-28 02:12:25 -07001181 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1182 start, start + len, mode, mode_flags,
David Rientjes00ef2d22013-02-22 16:35:36 -08001183 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001184
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001185 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1186
1187 err = migrate_prep();
1188 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001189 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001190 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001191 {
1192 NODEMASK_SCRATCH(scratch);
1193 if (scratch) {
1194 down_write(&mm->mmap_sem);
1195 task_lock(current);
1196 err = mpol_set_nodemask(new, nmask, scratch);
1197 task_unlock(current);
1198 if (err)
1199 up_write(&mm->mmap_sem);
1200 } else
1201 err = -ENOMEM;
1202 NODEMASK_SCRATCH_FREE(scratch);
1203 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001204 if (err)
1205 goto mpol_out;
1206
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001207 err = queue_pages_range(mm, start, end, nmask,
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001208 flags | MPOL_MF_INVERT, &pagelist);
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001209 if (!err)
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001210 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001211
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001212 if (!err) {
1213 int nr_failed = 0;
1214
Minchan Kimcf608ac2010-10-26 14:21:29 -07001215 if (!list_empty(&pagelist)) {
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001216 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001217 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1218 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001219 if (nr_failed)
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001220 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001221 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001222
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001223 if (nr_failed && (flags & MPOL_MF_STRICT))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001224 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001225 } else
Joonsoo Kimb0e5fd72013-12-18 17:08:51 -08001226 putback_movable_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001227
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001228 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001229 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001230 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001231 return err;
1232}
1233
Christoph Lameter39743882006-01-08 01:00:51 -08001234/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001235 * User space interface with variable sized bitmaps for nodelists.
1236 */
1237
1238/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001239static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001240 unsigned long maxnode)
1241{
1242 unsigned long k;
1243 unsigned long nlongs;
1244 unsigned long endmask;
1245
1246 --maxnode;
1247 nodes_clear(*nodes);
1248 if (maxnode == 0 || !nmask)
1249 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001250 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001251 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001252
1253 nlongs = BITS_TO_LONGS(maxnode);
1254 if ((maxnode % BITS_PER_LONG) == 0)
1255 endmask = ~0UL;
1256 else
1257 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1258
1259 /* When the user specified more nodes than supported just check
1260 if the non supported part is all zero. */
1261 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1262 if (nlongs > PAGE_SIZE/sizeof(long))
1263 return -EINVAL;
1264 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1265 unsigned long t;
1266 if (get_user(t, nmask + k))
1267 return -EFAULT;
1268 if (k == nlongs - 1) {
1269 if (t & endmask)
1270 return -EINVAL;
1271 } else if (t)
1272 return -EINVAL;
1273 }
1274 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1275 endmask = ~0UL;
1276 }
1277
1278 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1279 return -EFAULT;
1280 nodes_addr(*nodes)[nlongs-1] &= endmask;
1281 return 0;
1282}
1283
1284/* Copy a kernel node mask to user space */
1285static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1286 nodemask_t *nodes)
1287{
1288 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1289 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1290
1291 if (copy > nbytes) {
1292 if (copy > PAGE_SIZE)
1293 return -EINVAL;
1294 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1295 return -EFAULT;
1296 copy = nbytes;
1297 }
1298 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1299}
1300
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001301SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
Rasmus Villemoesf7f28ca2014-06-04 16:07:57 -07001302 unsigned long, mode, const unsigned long __user *, nmask,
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001303 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001304{
1305 nodemask_t nodes;
1306 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001307 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001308
David Rientjes028fec42008-04-28 02:12:25 -07001309 mode_flags = mode & MPOL_MODE_FLAGS;
1310 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001311 if (mode >= MPOL_MAX)
1312 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001313 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1314 (mode_flags & MPOL_F_RELATIVE_NODES))
1315 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001316 err = get_nodes(&nodes, nmask, maxnode);
1317 if (err)
1318 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001319 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001320}
1321
1322/* Set the process memory policy */
Rasmus Villemoes23c89022014-06-04 16:07:58 -07001323SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001324 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001325{
1326 int err;
1327 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001328 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001329
David Rientjes028fec42008-04-28 02:12:25 -07001330 flags = mode & MPOL_MODE_FLAGS;
1331 mode &= ~MPOL_MODE_FLAGS;
1332 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001333 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001334 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1335 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001336 err = get_nodes(&nodes, nmask, maxnode);
1337 if (err)
1338 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001339 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001340}
1341
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001342SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1343 const unsigned long __user *, old_nodes,
1344 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001345{
David Howellsc69e8d92008-11-14 10:39:19 +11001346 const struct cred *cred = current_cred(), *tcred;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001347 struct mm_struct *mm = NULL;
Christoph Lameter39743882006-01-08 01:00:51 -08001348 struct task_struct *task;
Christoph Lameter39743882006-01-08 01:00:51 -08001349 nodemask_t task_nodes;
1350 int err;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001351 nodemask_t *old;
1352 nodemask_t *new;
1353 NODEMASK_SCRATCH(scratch);
Christoph Lameter39743882006-01-08 01:00:51 -08001354
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001355 if (!scratch)
1356 return -ENOMEM;
Christoph Lameter39743882006-01-08 01:00:51 -08001357
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001358 old = &scratch->mask1;
1359 new = &scratch->mask2;
1360
1361 err = get_nodes(old, old_nodes, maxnode);
Christoph Lameter39743882006-01-08 01:00:51 -08001362 if (err)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001363 goto out;
1364
1365 err = get_nodes(new, new_nodes, maxnode);
1366 if (err)
1367 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001368
1369 /* Find the mm_struct */
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001370 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001371 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001372 if (!task) {
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001373 rcu_read_unlock();
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001374 err = -ESRCH;
1375 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001376 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001377 get_task_struct(task);
Christoph Lameter39743882006-01-08 01:00:51 -08001378
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001379 err = -EINVAL;
Christoph Lameter39743882006-01-08 01:00:51 -08001380
1381 /*
1382 * Check if this process has the right to modify the specified
1383 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001384 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001385 * userid as the target process.
1386 */
David Howellsc69e8d92008-11-14 10:39:19 +11001387 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001388 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1389 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001390 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001391 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001392 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001393 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001394 }
David Howellsc69e8d92008-11-14 10:39:19 +11001395 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001396
1397 task_nodes = cpuset_mems_allowed(task);
1398 /* Is the user allowed to access the target nodes? */
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001399 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001400 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001401 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001402 }
1403
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08001404 if (!nodes_subset(*new, node_states[N_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001405 err = -EINVAL;
Christoph Lameter3268c632012-03-21 16:34:06 -07001406 goto out_put;
Christoph Lameter3b42d282007-08-31 00:12:08 -07001407 }
1408
David Quigley86c3a762006-06-23 02:04:02 -07001409 err = security_task_movememory(task);
1410 if (err)
Christoph Lameter3268c632012-03-21 16:34:06 -07001411 goto out_put;
David Quigley86c3a762006-06-23 02:04:02 -07001412
Christoph Lameter3268c632012-03-21 16:34:06 -07001413 mm = get_task_mm(task);
1414 put_task_struct(task);
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001415
1416 if (!mm) {
Christoph Lameter3268c632012-03-21 16:34:06 -07001417 err = -EINVAL;
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001418 goto out;
1419 }
1420
1421 err = do_migrate_pages(mm, old, new,
1422 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter3268c632012-03-21 16:34:06 -07001423
1424 mmput(mm);
1425out:
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001426 NODEMASK_SCRATCH_FREE(scratch);
1427
Christoph Lameter39743882006-01-08 01:00:51 -08001428 return err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001429
1430out_put:
1431 put_task_struct(task);
1432 goto out;
1433
Christoph Lameter39743882006-01-08 01:00:51 -08001434}
1435
1436
Christoph Lameter8bccd852005-10-29 18:16:59 -07001437/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001438SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1439 unsigned long __user *, nmask, unsigned long, maxnode,
1440 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001441{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001442 int err;
1443 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001444 nodemask_t nodes;
1445
1446 if (nmask != NULL && maxnode < MAX_NUMNODES)
1447 return -EINVAL;
1448
1449 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1450
1451 if (err)
1452 return err;
1453
1454 if (policy && put_user(pval, policy))
1455 return -EFAULT;
1456
1457 if (nmask)
1458 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1459
1460 return err;
1461}
1462
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463#ifdef CONFIG_COMPAT
1464
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001465COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1466 compat_ulong_t __user *, nmask,
1467 compat_ulong_t, maxnode,
1468 compat_ulong_t, addr, compat_ulong_t, flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469{
1470 long err;
1471 unsigned long __user *nm = NULL;
1472 unsigned long nr_bits, alloc_size;
1473 DECLARE_BITMAP(bm, MAX_NUMNODES);
1474
1475 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1476 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1477
1478 if (nmask)
1479 nm = compat_alloc_user_space(alloc_size);
1480
1481 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1482
1483 if (!err && nmask) {
KAMEZAWA Hiroyuki2bbff6c2011-09-14 16:21:02 -07001484 unsigned long copy_size;
1485 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1486 err = copy_from_user(bm, nm, copy_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 /* ensure entire bitmap is zeroed */
1488 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1489 err |= compat_put_bitmap(nmask, bm, nr_bits);
1490 }
1491
1492 return err;
1493}
1494
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001495COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1496 compat_ulong_t, maxnode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497{
1498 long err = 0;
1499 unsigned long __user *nm = NULL;
1500 unsigned long nr_bits, alloc_size;
1501 DECLARE_BITMAP(bm, MAX_NUMNODES);
1502
1503 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1504 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1505
1506 if (nmask) {
1507 err = compat_get_bitmap(bm, nmask, nr_bits);
1508 nm = compat_alloc_user_space(alloc_size);
1509 err |= copy_to_user(nm, bm, alloc_size);
1510 }
1511
1512 if (err)
1513 return -EFAULT;
1514
1515 return sys_set_mempolicy(mode, nm, nr_bits+1);
1516}
1517
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001518COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1519 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1520 compat_ulong_t, maxnode, compat_ulong_t, flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521{
1522 long err = 0;
1523 unsigned long __user *nm = NULL;
1524 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001525 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526
1527 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1528 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1529
1530 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001531 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001533 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 }
1535
1536 if (err)
1537 return -EFAULT;
1538
1539 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1540}
1541
1542#endif
1543
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001544struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1545 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546{
Oleg Nesterov8d90274b2014-10-09 15:27:45 -07001547 struct mempolicy *pol = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548
1549 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001550 if (vma->vm_ops && vma->vm_ops->get_policy) {
Oleg Nesterov8d90274b2014-10-09 15:27:45 -07001551 pol = vma->vm_ops->get_policy(vma, addr);
Mel Gorman00442ad2012-10-08 16:29:20 -07001552 } else if (vma->vm_policy) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 pol = vma->vm_policy;
Mel Gorman00442ad2012-10-08 16:29:20 -07001554
1555 /*
1556 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1557 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1558 * count on these policies which will be dropped by
1559 * mpol_cond_put() later
1560 */
1561 if (mpol_needs_cond_ref(pol))
1562 mpol_get(pol);
1563 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 }
Oleg Nesterovf15ca782014-10-09 15:27:43 -07001565
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001566 return pol;
1567}
1568
1569/*
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001570 * get_vma_policy(@vma, @addr)
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001571 * @vma: virtual memory area whose policy is sought
1572 * @addr: address in @vma for shared policy lookup
1573 *
1574 * Returns effective policy for a VMA at specified address.
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001575 * Falls back to current->mempolicy or system default policy, as necessary.
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001576 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1577 * count--added by the get_policy() vm_op, as appropriate--to protect against
1578 * freeing by another task. It is the caller's responsibility to free the
1579 * extra reference for shared policies.
1580 */
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001581static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1582 unsigned long addr)
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001583{
1584 struct mempolicy *pol = __get_vma_policy(vma, addr);
1585
Oleg Nesterov8d90274b2014-10-09 15:27:45 -07001586 if (!pol)
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001587 pol = get_task_policy(current);
Oleg Nesterov8d90274b2014-10-09 15:27:45 -07001588
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 return pol;
1590}
1591
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001592bool vma_policy_mof(struct vm_area_struct *vma)
Mel Gormanfc3147242013-10-07 11:29:09 +01001593{
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001594 struct mempolicy *pol;
Oleg Nesterovf15ca782014-10-09 15:27:43 -07001595
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001596 if (vma->vm_ops && vma->vm_ops->get_policy) {
1597 bool ret = false;
Mel Gormanfc3147242013-10-07 11:29:09 +01001598
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001599 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1600 if (pol && (pol->flags & MPOL_F_MOF))
1601 ret = true;
1602 mpol_cond_put(pol);
Mel Gormanfc3147242013-10-07 11:29:09 +01001603
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001604 return ret;
Mel Gormanfc3147242013-10-07 11:29:09 +01001605 }
1606
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001607 pol = vma->vm_policy;
Oleg Nesterov8d90274b2014-10-09 15:27:45 -07001608 if (!pol)
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001609 pol = get_task_policy(current);
Oleg Nesterov8d90274b2014-10-09 15:27:45 -07001610
Mel Gormanfc3147242013-10-07 11:29:09 +01001611 return pol->flags & MPOL_F_MOF;
1612}
1613
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001614static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1615{
1616 enum zone_type dynamic_policy_zone = policy_zone;
1617
1618 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1619
1620 /*
1621 * if policy->v.nodes has movable memory only,
1622 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1623 *
1624 * policy->v.nodes is intersect with node_states[N_MEMORY].
1625 * so if the following test faile, it implies
1626 * policy->v.nodes has movable memory only.
1627 */
1628 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1629 dynamic_policy_zone = ZONE_MOVABLE;
1630
1631 return zone >= dynamic_policy_zone;
1632}
1633
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001634/*
1635 * Return a nodemask representing a mempolicy for filtering nodes for
1636 * page allocation
1637 */
1638static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001639{
1640 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001641 if (unlikely(policy->mode == MPOL_BIND) &&
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001642 apply_policy_zone(policy, gfp_zone(gfp)) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001643 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1644 return &policy->v.nodes;
1645
1646 return NULL;
1647}
1648
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001649/* Return a zonelist indicated by gfp for node representing a mempolicy */
Andi Kleen2f5f9482011-03-04 17:36:29 -08001650static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1651 int nd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652{
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001653 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001655 if (!(policy->flags & MPOL_F_LOCAL))
1656 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 break;
1658 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001659 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001660 * Normally, MPOL_BIND allocations are node-local within the
1661 * allowed nodemask. However, if __GFP_THISNODE is set and the
Bob Liu6eb27e12010-05-24 14:32:00 -07001662 * current node isn't part of the mask, we use the zonelist for
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001663 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001664 */
Mel Gorman19770b32008-04-28 02:12:18 -07001665 if (unlikely(gfp & __GFP_THISNODE) &&
1666 unlikely(!node_isset(nd, policy->v.nodes)))
1667 nd = first_node(policy->v.nodes);
1668 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 BUG();
1671 }
Mel Gorman0e884602008-04-28 02:12:14 -07001672 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673}
1674
1675/* Do dynamic interleaving for a process */
1676static unsigned interleave_nodes(struct mempolicy *policy)
1677{
1678 unsigned nid, next;
1679 struct task_struct *me = current;
1680
1681 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001682 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001684 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001685 if (next < MAX_NUMNODES)
1686 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 return nid;
1688}
1689
Christoph Lameterdc85da12006-01-18 17:42:36 -08001690/*
1691 * Depending on the memory policy provide a node from which to allocate the
1692 * next slab entry.
1693 */
David Rientjes2a389612014-04-07 15:37:29 -07001694unsigned int mempolicy_slab_node(void)
Christoph Lameterdc85da12006-01-18 17:42:36 -08001695{
Andi Kleene7b691b2012-06-09 02:40:03 -07001696 struct mempolicy *policy;
David Rientjes2a389612014-04-07 15:37:29 -07001697 int node = numa_mem_id();
Andi Kleene7b691b2012-06-09 02:40:03 -07001698
1699 if (in_interrupt())
David Rientjes2a389612014-04-07 15:37:29 -07001700 return node;
Andi Kleene7b691b2012-06-09 02:40:03 -07001701
1702 policy = current->mempolicy;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001703 if (!policy || policy->flags & MPOL_F_LOCAL)
David Rientjes2a389612014-04-07 15:37:29 -07001704 return node;
Christoph Lameter765c4502006-09-27 01:50:08 -07001705
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001706 switch (policy->mode) {
1707 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001708 /*
1709 * handled MPOL_F_LOCAL above
1710 */
1711 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001712
Christoph Lameterdc85da12006-01-18 17:42:36 -08001713 case MPOL_INTERLEAVE:
1714 return interleave_nodes(policy);
1715
Mel Gormandd1a2392008-04-28 02:12:17 -07001716 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001717 /*
1718 * Follow bind policy behavior and start allocation at the
1719 * first node.
1720 */
Mel Gorman19770b32008-04-28 02:12:18 -07001721 struct zonelist *zonelist;
1722 struct zone *zone;
1723 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
David Rientjes2a389612014-04-07 15:37:29 -07001724 zonelist = &NODE_DATA(node)->node_zonelists[0];
Mel Gorman19770b32008-04-28 02:12:18 -07001725 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1726 &policy->v.nodes,
1727 &zone);
David Rientjes2a389612014-04-07 15:37:29 -07001728 return zone ? zone->node : node;
Mel Gormandd1a2392008-04-28 02:12:17 -07001729 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001730
Christoph Lameterdc85da12006-01-18 17:42:36 -08001731 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001732 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001733 }
1734}
1735
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736/* Do static interleaving for a VMA with known offset. */
1737static unsigned offset_il_node(struct mempolicy *pol,
1738 struct vm_area_struct *vma, unsigned long off)
1739{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001740 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001741 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 int c;
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001743 int nid = NUMA_NO_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744
David Rientjesf5b087b2008-04-28 02:12:27 -07001745 if (!nnodes)
1746 return numa_node_id();
1747 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 c = 0;
1749 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001750 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 c++;
1752 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753 return nid;
1754}
1755
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001756/* Determine a node number for interleave */
1757static inline unsigned interleave_nid(struct mempolicy *pol,
1758 struct vm_area_struct *vma, unsigned long addr, int shift)
1759{
1760 if (vma) {
1761 unsigned long off;
1762
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001763 /*
1764 * for small pages, there is no difference between
1765 * shift and PAGE_SHIFT, so the bit-shift is safe.
1766 * for huge pages, since vm_pgoff is in units of small
1767 * pages, we need to shift off the always 0 bits to get
1768 * a useful offset.
1769 */
1770 BUG_ON(shift < PAGE_SHIFT);
1771 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001772 off += (addr - vma->vm_start) >> shift;
1773 return offset_il_node(pol, vma, off);
1774 } else
1775 return interleave_nodes(pol);
1776}
1777
Michal Hocko778d3b02011-07-26 16:08:30 -07001778/*
1779 * Return the bit number of a random bit set in the nodemask.
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001780 * (returns NUMA_NO_NODE if nodemask is empty)
Michal Hocko778d3b02011-07-26 16:08:30 -07001781 */
1782int node_random(const nodemask_t *maskp)
1783{
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001784 int w, bit = NUMA_NO_NODE;
Michal Hocko778d3b02011-07-26 16:08:30 -07001785
1786 w = nodes_weight(*maskp);
1787 if (w)
1788 bit = bitmap_ord_to_pos(maskp->bits,
1789 get_random_int() % w, MAX_NUMNODES);
1790 return bit;
1791}
1792
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001793#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001794/*
1795 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
Fabian Frederickb46e14a2014-06-04 16:08:18 -07001796 * @vma: virtual memory area whose policy is sought
1797 * @addr: address in @vma for shared policy lookup and interleave policy
1798 * @gfp_flags: for requested zone
1799 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1800 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001801 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001802 * Returns a zonelist suitable for a huge page allocation and a pointer
1803 * to the struct mempolicy for conditional unref after allocation.
1804 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1805 * @nodemask for filtering the zonelist.
Miao Xiec0ff7452010-05-24 14:32:08 -07001806 *
Mel Gormand26914d2014-04-03 14:47:24 -07001807 * Must be protected by read_mems_allowed_begin()
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001808 */
Mel Gorman396faf02007-07-17 04:03:13 -07001809struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001810 gfp_t gfp_flags, struct mempolicy **mpol,
1811 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001812{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001813 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001814
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001815 *mpol = get_vma_policy(vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001816 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001817
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001818 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1819 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001820 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001821 } else {
Andi Kleen2f5f9482011-03-04 17:36:29 -08001822 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001823 if ((*mpol)->mode == MPOL_BIND)
1824 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001825 }
1826 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001827}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001828
1829/*
1830 * init_nodemask_of_mempolicy
1831 *
1832 * If the current task's mempolicy is "default" [NULL], return 'false'
1833 * to indicate default policy. Otherwise, extract the policy nodemask
1834 * for 'bind' or 'interleave' policy into the argument nodemask, or
1835 * initialize the argument nodemask to contain the single node for
1836 * 'preferred' or 'local' policy and return 'true' to indicate presence
1837 * of non-default mempolicy.
1838 *
1839 * We don't bother with reference counting the mempolicy [mpol_get/put]
1840 * because the current task is examining it's own mempolicy and a task's
1841 * mempolicy is only ever changed by the task itself.
1842 *
1843 * N.B., it is the caller's responsibility to free a returned nodemask.
1844 */
1845bool init_nodemask_of_mempolicy(nodemask_t *mask)
1846{
1847 struct mempolicy *mempolicy;
1848 int nid;
1849
1850 if (!(mask && current->mempolicy))
1851 return false;
1852
Miao Xiec0ff7452010-05-24 14:32:08 -07001853 task_lock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001854 mempolicy = current->mempolicy;
1855 switch (mempolicy->mode) {
1856 case MPOL_PREFERRED:
1857 if (mempolicy->flags & MPOL_F_LOCAL)
1858 nid = numa_node_id();
1859 else
1860 nid = mempolicy->v.preferred_node;
1861 init_nodemask_of_node(mask, nid);
1862 break;
1863
1864 case MPOL_BIND:
1865 /* Fall through */
1866 case MPOL_INTERLEAVE:
1867 *mask = mempolicy->v.nodes;
1868 break;
1869
1870 default:
1871 BUG();
1872 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001873 task_unlock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001874
1875 return true;
1876}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001877#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001878
David Rientjes6f48d0eb2010-08-09 17:18:52 -07001879/*
1880 * mempolicy_nodemask_intersects
1881 *
1882 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1883 * policy. Otherwise, check for intersection between mask and the policy
1884 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1885 * policy, always return true since it may allocate elsewhere on fallback.
1886 *
1887 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1888 */
1889bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1890 const nodemask_t *mask)
1891{
1892 struct mempolicy *mempolicy;
1893 bool ret = true;
1894
1895 if (!mask)
1896 return ret;
1897 task_lock(tsk);
1898 mempolicy = tsk->mempolicy;
1899 if (!mempolicy)
1900 goto out;
1901
1902 switch (mempolicy->mode) {
1903 case MPOL_PREFERRED:
1904 /*
1905 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1906 * allocate from, they may fallback to other nodes when oom.
1907 * Thus, it's possible for tsk to have allocated memory from
1908 * nodes in mask.
1909 */
1910 break;
1911 case MPOL_BIND:
1912 case MPOL_INTERLEAVE:
1913 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1914 break;
1915 default:
1916 BUG();
1917 }
1918out:
1919 task_unlock(tsk);
1920 return ret;
1921}
1922
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923/* Allocate a page in interleaved policy.
1924 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001925static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1926 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927{
1928 struct zonelist *zl;
1929 struct page *page;
1930
Mel Gorman0e884602008-04-28 02:12:14 -07001931 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001933 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001934 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 return page;
1936}
1937
1938/**
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001939 * alloc_pages_vma - Allocate a page for a VMA.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 *
1941 * @gfp:
1942 * %GFP_USER user allocation.
1943 * %GFP_KERNEL kernel allocations,
1944 * %GFP_HIGHMEM highmem/user allocations,
1945 * %GFP_FS allocation should not call back into a file system.
1946 * %GFP_ATOMIC don't sleep.
1947 *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001948 * @order:Order of the GFP allocation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 * @vma: Pointer to VMA or NULL if not available.
1950 * @addr: Virtual Address of the allocation. Must be inside the VMA.
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001951 * @node: Which node to prefer for allocation (modulo policy).
1952 * @hugepage: for hugepages try only the preferred node if possible
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 *
1954 * This function allocates a page from the kernel page pool and applies
1955 * a NUMA policy associated with the VMA or the current process.
1956 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1957 * mm_struct of the VMA to prevent it from going away. Should be used for
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001958 * all allocations for pages that will be mapped into user space. Returns
1959 * NULL when no page can be allocated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 */
1961struct page *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001962alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001963 unsigned long addr, int node, bool hugepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964{
Mel Gormancc9a6c82012-03-21 16:34:11 -07001965 struct mempolicy *pol;
Miao Xiec0ff7452010-05-24 14:32:08 -07001966 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001967 unsigned int cpuset_mems_cookie;
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001968 struct zonelist *zl;
1969 nodemask_t *nmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970
Mel Gormancc9a6c82012-03-21 16:34:11 -07001971retry_cpuset:
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001972 pol = get_vma_policy(vma, addr);
Mel Gormand26914d2014-04-03 14:47:24 -07001973 cpuset_mems_cookie = read_mems_allowed_begin();
Mel Gormancc9a6c82012-03-21 16:34:11 -07001974
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001975 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage &&
1976 pol->mode != MPOL_INTERLEAVE)) {
1977 /*
1978 * For hugepage allocation and non-interleave policy which
1979 * allows the current node, we only try to allocate from the
1980 * current node and don't fall back to other nodes, as the
1981 * cost of remote accesses would likely offset THP benefits.
1982 *
1983 * If the policy is interleave, or does not allow the current
1984 * node in its nodemask, we allocate the standard way.
1985 */
1986 nmask = policy_nodemask(gfp, pol);
1987 if (!nmask || node_isset(node, *nmask)) {
1988 mpol_cond_put(pol);
David Rientjes52650472015-04-14 15:46:58 -07001989 page = alloc_pages_exact_node(node,
1990 gfp | __GFP_THISNODE, order);
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001991 goto out;
1992 }
1993 }
1994
1995 if (pol->mode == MPOL_INTERLEAVE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001997
Andi Kleen8eac5632011-02-25 14:44:28 -08001998 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001999 mpol_cond_put(pol);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002000 page = alloc_page_interleave(gfp, order, nid);
Vlastimil Babkabe97a412015-02-11 15:27:15 -08002001 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 }
Vlastimil Babkabe97a412015-02-11 15:27:15 -08002003
2004 nmask = policy_nodemask(gfp, pol);
2005 zl = policy_zonelist(gfp, pol, node);
Oleg Nesterov23867402014-10-09 15:27:41 -07002006 mpol_cond_put(pol);
Vlastimil Babkabe97a412015-02-11 15:27:15 -08002007 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2008out:
Mel Gormand26914d2014-04-03 14:47:24 -07002009 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07002010 goto retry_cpuset;
Miao Xiec0ff7452010-05-24 14:32:08 -07002011 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012}
2013
2014/**
2015 * alloc_pages_current - Allocate pages.
2016 *
2017 * @gfp:
2018 * %GFP_USER user allocation,
2019 * %GFP_KERNEL kernel allocation,
2020 * %GFP_HIGHMEM highmem allocation,
2021 * %GFP_FS don't call back into a file system.
2022 * %GFP_ATOMIC don't sleep.
2023 * @order: Power of two of allocation size in pages. 0 is a single page.
2024 *
2025 * Allocate a page from the kernel page pool. When not in
2026 * interrupt context and apply the current process NUMA policy.
2027 * Returns NULL when no page can be allocated.
2028 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08002029 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 * 1) it's ok to take cpuset_sem (can WAIT), and
2031 * 2) allocating for current task (not interrupt).
2032 */
Al Virodd0fc662005-10-07 07:46:04 +01002033struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034{
Oleg Nesterov8d90274b2014-10-09 15:27:45 -07002035 struct mempolicy *pol = &default_policy;
Miao Xiec0ff7452010-05-24 14:32:08 -07002036 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002037 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038
Oleg Nesterov8d90274b2014-10-09 15:27:45 -07002039 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2040 pol = get_task_policy(current);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002041
Mel Gormancc9a6c82012-03-21 16:34:11 -07002042retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07002043 cpuset_mems_cookie = read_mems_allowed_begin();
Mel Gormancc9a6c82012-03-21 16:34:11 -07002044
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002045 /*
2046 * No reference counting needed for current->mempolicy
2047 * nor system default_policy
2048 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002049 if (pol->mode == MPOL_INTERLEAVE)
Miao Xiec0ff7452010-05-24 14:32:08 -07002050 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2051 else
2052 page = __alloc_pages_nodemask(gfp, order,
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002053 policy_zonelist(gfp, pol, numa_node_id()),
2054 policy_nodemask(gfp, pol));
Mel Gormancc9a6c82012-03-21 16:34:11 -07002055
Mel Gormand26914d2014-04-03 14:47:24 -07002056 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07002057 goto retry_cpuset;
2058
Miao Xiec0ff7452010-05-24 14:32:08 -07002059 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060}
2061EXPORT_SYMBOL(alloc_pages_current);
2062
Oleg Nesterovef0855d2013-09-11 14:20:14 -07002063int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2064{
2065 struct mempolicy *pol = mpol_dup(vma_policy(src));
2066
2067 if (IS_ERR(pol))
2068 return PTR_ERR(pol);
2069 dst->vm_policy = pol;
2070 return 0;
2071}
2072
Paul Jackson42253992006-01-08 01:01:59 -08002073/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002074 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08002075 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2076 * with the mems_allowed returned by cpuset_mems_allowed(). This
2077 * keeps mempolicies cpuset relative after its cpuset moves. See
2078 * further kernel/cpuset.c update_nodemask().
Miao Xie708c1bb2010-05-24 14:32:07 -07002079 *
2080 * current's mempolicy may be rebinded by the other task(the task that changes
2081 * cpuset's mems), so we needn't do rebind work for current task.
Paul Jackson42253992006-01-08 01:01:59 -08002082 */
Paul Jackson42253992006-01-08 01:01:59 -08002083
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002084/* Slow path of a mempolicy duplicate */
2085struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086{
2087 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2088
2089 if (!new)
2090 return ERR_PTR(-ENOMEM);
Miao Xie708c1bb2010-05-24 14:32:07 -07002091
2092 /* task's mempolicy is protected by alloc_lock */
2093 if (old == current->mempolicy) {
2094 task_lock(current);
2095 *new = *old;
2096 task_unlock(current);
2097 } else
2098 *new = *old;
2099
Paul Jackson42253992006-01-08 01:01:59 -08002100 if (current_cpuset_is_being_rebound()) {
2101 nodemask_t mems = cpuset_mems_allowed(current);
Miao Xie708c1bb2010-05-24 14:32:07 -07002102 if (new->flags & MPOL_F_REBINDING)
2103 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2104 else
2105 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
Paul Jackson42253992006-01-08 01:01:59 -08002106 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 return new;
2109}
2110
2111/* Slow path of a mempolicy comparison */
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002112bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113{
2114 if (!a || !b)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002115 return false;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002116 if (a->mode != b->mode)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002117 return false;
Bob Liu19800502010-05-24 14:32:01 -07002118 if (a->flags != b->flags)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002119 return false;
Bob Liu19800502010-05-24 14:32:01 -07002120 if (mpol_store_user_nodemask(a))
2121 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002122 return false;
Bob Liu19800502010-05-24 14:32:01 -07002123
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002124 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07002125 case MPOL_BIND:
2126 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 case MPOL_INTERLEAVE:
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002128 return !!nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 case MPOL_PREFERRED:
Namhyung Kim75719662011-03-22 16:33:02 -07002130 return a->v.preferred_node == b->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 default:
2132 BUG();
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002133 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 }
2135}
2136
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 * Shared memory backing store policy support.
2139 *
2140 * Remember policies even when nobody has shared memory mapped.
2141 * The policies are kept in Red-Black tree linked from the inode.
2142 * They are protected by the sp->lock spinlock, which should be held
2143 * for any accesses to the tree.
2144 */
2145
2146/* lookup first element intersecting start-end */
Mel Gorman42288fe2012-12-21 23:10:25 +00002147/* Caller holds sp->lock */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148static struct sp_node *
2149sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2150{
2151 struct rb_node *n = sp->root.rb_node;
2152
2153 while (n) {
2154 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2155
2156 if (start >= p->end)
2157 n = n->rb_right;
2158 else if (end <= p->start)
2159 n = n->rb_left;
2160 else
2161 break;
2162 }
2163 if (!n)
2164 return NULL;
2165 for (;;) {
2166 struct sp_node *w = NULL;
2167 struct rb_node *prev = rb_prev(n);
2168 if (!prev)
2169 break;
2170 w = rb_entry(prev, struct sp_node, nd);
2171 if (w->end <= start)
2172 break;
2173 n = prev;
2174 }
2175 return rb_entry(n, struct sp_node, nd);
2176}
2177
2178/* Insert a new shared policy into the list. */
2179/* Caller holds sp->lock */
2180static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2181{
2182 struct rb_node **p = &sp->root.rb_node;
2183 struct rb_node *parent = NULL;
2184 struct sp_node *nd;
2185
2186 while (*p) {
2187 parent = *p;
2188 nd = rb_entry(parent, struct sp_node, nd);
2189 if (new->start < nd->start)
2190 p = &(*p)->rb_left;
2191 else if (new->end > nd->end)
2192 p = &(*p)->rb_right;
2193 else
2194 BUG();
2195 }
2196 rb_link_node(&new->nd, parent, p);
2197 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07002198 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002199 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200}
2201
2202/* Find shared policy intersecting idx */
2203struct mempolicy *
2204mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2205{
2206 struct mempolicy *pol = NULL;
2207 struct sp_node *sn;
2208
2209 if (!sp->root.rb_node)
2210 return NULL;
Mel Gorman42288fe2012-12-21 23:10:25 +00002211 spin_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 sn = sp_lookup(sp, idx, idx+1);
2213 if (sn) {
2214 mpol_get(sn->policy);
2215 pol = sn->policy;
2216 }
Mel Gorman42288fe2012-12-21 23:10:25 +00002217 spin_unlock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 return pol;
2219}
2220
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002221static void sp_free(struct sp_node *n)
2222{
2223 mpol_put(n->policy);
2224 kmem_cache_free(sn_cache, n);
2225}
2226
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002227/**
2228 * mpol_misplaced - check whether current page node is valid in policy
2229 *
Fabian Frederickb46e14a2014-06-04 16:08:18 -07002230 * @page: page to be checked
2231 * @vma: vm area where page mapped
2232 * @addr: virtual address where page mapped
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002233 *
2234 * Lookup current policy node id for vma,addr and "compare to" page's
2235 * node id.
2236 *
2237 * Returns:
2238 * -1 - not misplaced, page is in the right node
2239 * node - node id where the page should be
2240 *
2241 * Policy determination "mimics" alloc_page_vma().
2242 * Called from fault path where we know the vma and faulting address.
2243 */
2244int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2245{
2246 struct mempolicy *pol;
2247 struct zone *zone;
2248 int curnid = page_to_nid(page);
2249 unsigned long pgoff;
Peter Zijlstra90572892013-10-07 11:29:20 +01002250 int thiscpu = raw_smp_processor_id();
2251 int thisnid = cpu_to_node(thiscpu);
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002252 int polnid = -1;
2253 int ret = -1;
2254
2255 BUG_ON(!vma);
2256
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07002257 pol = get_vma_policy(vma, addr);
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002258 if (!(pol->flags & MPOL_F_MOF))
2259 goto out;
2260
2261 switch (pol->mode) {
2262 case MPOL_INTERLEAVE:
2263 BUG_ON(addr >= vma->vm_end);
2264 BUG_ON(addr < vma->vm_start);
2265
2266 pgoff = vma->vm_pgoff;
2267 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2268 polnid = offset_il_node(pol, vma, pgoff);
2269 break;
2270
2271 case MPOL_PREFERRED:
2272 if (pol->flags & MPOL_F_LOCAL)
2273 polnid = numa_node_id();
2274 else
2275 polnid = pol->v.preferred_node;
2276 break;
2277
2278 case MPOL_BIND:
2279 /*
2280 * allows binding to multiple nodes.
2281 * use current page if in policy nodemask,
2282 * else select nearest allowed node, if any.
2283 * If no allowed nodes, use current [!misplaced].
2284 */
2285 if (node_isset(curnid, pol->v.nodes))
2286 goto out;
2287 (void)first_zones_zonelist(
2288 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2289 gfp_zone(GFP_HIGHUSER),
2290 &pol->v.nodes, &zone);
2291 polnid = zone->node;
2292 break;
2293
2294 default:
2295 BUG();
2296 }
Mel Gorman5606e382012-11-02 18:19:13 +00002297
2298 /* Migrate the page towards the node whose CPU is referencing it */
Mel Gormane42c8ff2012-11-12 09:17:07 +00002299 if (pol->flags & MPOL_F_MORON) {
Peter Zijlstra90572892013-10-07 11:29:20 +01002300 polnid = thisnid;
Mel Gorman5606e382012-11-02 18:19:13 +00002301
Rik van Riel10f39042014-01-27 17:03:44 -05002302 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
Rik van Rielde1c9ce2013-10-07 11:29:39 +01002303 goto out;
Mel Gormane42c8ff2012-11-12 09:17:07 +00002304 }
2305
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002306 if (curnid != polnid)
2307 ret = polnid;
2308out:
2309 mpol_cond_put(pol);
2310
2311 return ret;
2312}
2313
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2315{
Paul Mundt140d5a42007-07-15 23:38:16 -07002316 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 rb_erase(&n->nd, &sp->root);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002318 sp_free(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319}
2320
Mel Gorman42288fe2012-12-21 23:10:25 +00002321static void sp_node_init(struct sp_node *node, unsigned long start,
2322 unsigned long end, struct mempolicy *pol)
2323{
2324 node->start = start;
2325 node->end = end;
2326 node->policy = pol;
2327}
2328
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07002329static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2330 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331{
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002332 struct sp_node *n;
2333 struct mempolicy *newpol;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002335 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 if (!n)
2337 return NULL;
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002338
2339 newpol = mpol_dup(pol);
2340 if (IS_ERR(newpol)) {
2341 kmem_cache_free(sn_cache, n);
2342 return NULL;
2343 }
2344 newpol->flags |= MPOL_F_SHARED;
Mel Gorman42288fe2012-12-21 23:10:25 +00002345 sp_node_init(n, start, end, newpol);
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002346
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 return n;
2348}
2349
2350/* Replace a policy range. */
2351static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2352 unsigned long end, struct sp_node *new)
2353{
Mel Gormanb22d1272012-10-08 16:29:17 -07002354 struct sp_node *n;
Mel Gorman42288fe2012-12-21 23:10:25 +00002355 struct sp_node *n_new = NULL;
2356 struct mempolicy *mpol_new = NULL;
Mel Gormanb22d1272012-10-08 16:29:17 -07002357 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
Mel Gorman42288fe2012-12-21 23:10:25 +00002359restart:
2360 spin_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 n = sp_lookup(sp, start, end);
2362 /* Take care of old policies in the same range. */
2363 while (n && n->start < end) {
2364 struct rb_node *next = rb_next(&n->nd);
2365 if (n->start >= start) {
2366 if (n->end <= end)
2367 sp_delete(sp, n);
2368 else
2369 n->start = end;
2370 } else {
2371 /* Old policy spanning whole new range. */
2372 if (n->end > end) {
Mel Gorman42288fe2012-12-21 23:10:25 +00002373 if (!n_new)
2374 goto alloc_new;
2375
2376 *mpol_new = *n->policy;
2377 atomic_set(&mpol_new->refcnt, 1);
KOSAKI Motohiro78806392013-03-08 12:43:29 -08002378 sp_node_init(n_new, end, n->end, mpol_new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379 n->end = start;
Hillf Danton5ca39572013-03-08 12:43:28 -08002380 sp_insert(sp, n_new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002381 n_new = NULL;
2382 mpol_new = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 break;
2384 } else
2385 n->end = start;
2386 }
2387 if (!next)
2388 break;
2389 n = rb_entry(next, struct sp_node, nd);
2390 }
2391 if (new)
2392 sp_insert(sp, new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002393 spin_unlock(&sp->lock);
2394 ret = 0;
2395
2396err_out:
2397 if (mpol_new)
2398 mpol_put(mpol_new);
2399 if (n_new)
2400 kmem_cache_free(sn_cache, n_new);
2401
Mel Gormanb22d1272012-10-08 16:29:17 -07002402 return ret;
Mel Gorman42288fe2012-12-21 23:10:25 +00002403
2404alloc_new:
2405 spin_unlock(&sp->lock);
2406 ret = -ENOMEM;
2407 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2408 if (!n_new)
2409 goto err_out;
2410 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2411 if (!mpol_new)
2412 goto err_out;
2413 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414}
2415
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002416/**
2417 * mpol_shared_policy_init - initialize shared policy for inode
2418 * @sp: pointer to inode shared policy
2419 * @mpol: struct mempolicy to install
2420 *
2421 * Install non-NULL @mpol in inode's shared policy rb-tree.
2422 * On entry, the current task has a reference on a non-NULL @mpol.
2423 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002424 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002425 */
2426void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08002427{
Miao Xie58568d22009-06-16 15:31:49 -07002428 int ret;
2429
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002430 sp->root = RB_ROOT; /* empty tree == default mempolicy */
Mel Gorman42288fe2012-12-21 23:10:25 +00002431 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08002432
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002433 if (mpol) {
2434 struct vm_area_struct pvma;
2435 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002436 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002437
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002438 if (!scratch)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002439 goto put_mpol;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002440 /* contextualize the tmpfs mount point mempolicy */
2441 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002442 if (IS_ERR(new))
Dan Carpenter0cae3452010-05-25 23:42:58 -07002443 goto free_scratch; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002444
2445 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002446 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002447 task_unlock(current);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002448 if (ret)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002449 goto put_new;
Robin Holt7339ff82006-01-14 13:20:48 -08002450
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002451 /* Create pseudo-vma that contains just the policy */
2452 memset(&pvma, 0, sizeof(struct vm_area_struct));
2453 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2454 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002455
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002456put_new:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002457 mpol_put(new); /* drop initial ref */
Dan Carpenter0cae3452010-05-25 23:42:58 -07002458free_scratch:
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002459 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002460put_mpol:
2461 mpol_put(mpol); /* drop our incoming ref on sb mpol */
Robin Holt7339ff82006-01-14 13:20:48 -08002462 }
2463}
2464
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465int mpol_set_shared_policy(struct shared_policy *info,
2466 struct vm_area_struct *vma, struct mempolicy *npol)
2467{
2468 int err;
2469 struct sp_node *new = NULL;
2470 unsigned long sz = vma_pages(vma);
2471
David Rientjes028fec42008-04-28 02:12:25 -07002472 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002474 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002475 npol ? npol->flags : -1,
David Rientjes00ef2d22013-02-22 16:35:36 -08002476 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477
2478 if (npol) {
2479 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2480 if (!new)
2481 return -ENOMEM;
2482 }
2483 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2484 if (err && new)
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002485 sp_free(new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486 return err;
2487}
2488
2489/* Free a backing policy store on inode delete. */
2490void mpol_free_shared_policy(struct shared_policy *p)
2491{
2492 struct sp_node *n;
2493 struct rb_node *next;
2494
2495 if (!p->root.rb_node)
2496 return;
Mel Gorman42288fe2012-12-21 23:10:25 +00002497 spin_lock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 next = rb_first(&p->root);
2499 while (next) {
2500 n = rb_entry(next, struct sp_node, nd);
2501 next = rb_next(&n->nd);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002502 sp_delete(p, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 }
Mel Gorman42288fe2012-12-21 23:10:25 +00002504 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505}
2506
Mel Gorman1a687c22012-11-22 11:16:36 +00002507#ifdef CONFIG_NUMA_BALANCING
Mel Gormanc2976632014-01-29 14:05:42 -08002508static int __initdata numabalancing_override;
Mel Gorman1a687c22012-11-22 11:16:36 +00002509
2510static void __init check_numabalancing_enable(void)
2511{
2512 bool numabalancing_default = false;
2513
2514 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2515 numabalancing_default = true;
2516
Mel Gormanc2976632014-01-29 14:05:42 -08002517 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2518 if (numabalancing_override)
2519 set_numabalancing_state(numabalancing_override == 1);
2520
Mel Gormanb0dc2b92015-05-14 15:17:09 -07002521 if (num_online_nodes() > 1 && !numabalancing_override) {
Andrew Morton4a404be2014-01-29 14:05:43 -08002522 pr_info("%s automatic NUMA balancing. "
Mel Gormanc2976632014-01-29 14:05:42 -08002523 "Configure with numa_balancing= or the "
2524 "kernel.numa_balancing sysctl",
2525 numabalancing_default ? "Enabling" : "Disabling");
Mel Gorman1a687c22012-11-22 11:16:36 +00002526 set_numabalancing_state(numabalancing_default);
2527 }
2528}
2529
2530static int __init setup_numabalancing(char *str)
2531{
2532 int ret = 0;
2533 if (!str)
2534 goto out;
Mel Gorman1a687c22012-11-22 11:16:36 +00002535
2536 if (!strcmp(str, "enable")) {
Mel Gormanc2976632014-01-29 14:05:42 -08002537 numabalancing_override = 1;
Mel Gorman1a687c22012-11-22 11:16:36 +00002538 ret = 1;
2539 } else if (!strcmp(str, "disable")) {
Mel Gormanc2976632014-01-29 14:05:42 -08002540 numabalancing_override = -1;
Mel Gorman1a687c22012-11-22 11:16:36 +00002541 ret = 1;
2542 }
2543out:
2544 if (!ret)
Andrew Morton4a404be2014-01-29 14:05:43 -08002545 pr_warn("Unable to parse numa_balancing=\n");
Mel Gorman1a687c22012-11-22 11:16:36 +00002546
2547 return ret;
2548}
2549__setup("numa_balancing=", setup_numabalancing);
2550#else
2551static inline void __init check_numabalancing_enable(void)
2552{
2553}
2554#endif /* CONFIG_NUMA_BALANCING */
2555
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556/* assumes fs == KERNEL_DS */
2557void __init numa_policy_init(void)
2558{
Paul Mundtb71636e2007-07-15 23:38:15 -07002559 nodemask_t interleave_nodes;
2560 unsigned long largest = 0;
2561 int nid, prefer = 0;
2562
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 policy_cache = kmem_cache_create("numa_policy",
2564 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002565 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566
2567 sn_cache = kmem_cache_create("shared_policy_node",
2568 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002569 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570
Mel Gorman5606e382012-11-02 18:19:13 +00002571 for_each_node(nid) {
2572 preferred_node_policy[nid] = (struct mempolicy) {
2573 .refcnt = ATOMIC_INIT(1),
2574 .mode = MPOL_PREFERRED,
2575 .flags = MPOL_F_MOF | MPOL_F_MORON,
2576 .v = { .preferred_node = nid, },
2577 };
2578 }
2579
Paul Mundtb71636e2007-07-15 23:38:15 -07002580 /*
2581 * Set interleaving policy for system init. Interleaving is only
2582 * enabled across suitably sized nodes (default is >= 16MB), or
2583 * fall back to the largest node if they're all smaller.
2584 */
2585 nodes_clear(interleave_nodes);
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002586 for_each_node_state(nid, N_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002587 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588
Paul Mundtb71636e2007-07-15 23:38:15 -07002589 /* Preserve the largest node */
2590 if (largest < total_pages) {
2591 largest = total_pages;
2592 prefer = nid;
2593 }
2594
2595 /* Interleave this node? */
2596 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2597 node_set(nid, interleave_nodes);
2598 }
2599
2600 /* All too small, use the largest */
2601 if (unlikely(nodes_empty(interleave_nodes)))
2602 node_set(prefer, interleave_nodes);
2603
David Rientjes028fec42008-04-28 02:12:25 -07002604 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -07002605 pr_err("%s: interleaving failed\n", __func__);
Mel Gorman1a687c22012-11-22 11:16:36 +00002606
2607 check_numabalancing_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608}
2609
Christoph Lameter8bccd852005-10-29 18:16:59 -07002610/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611void numa_default_policy(void)
2612{
David Rientjes028fec42008-04-28 02:12:25 -07002613 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614}
Paul Jackson68860ec2005-10-30 15:02:36 -08002615
Paul Jackson42253992006-01-08 01:01:59 -08002616/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002617 * Parse and format mempolicy from/to strings
2618 */
2619
2620/*
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002621 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002622 */
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002623static const char * const policy_modes[] =
2624{
2625 [MPOL_DEFAULT] = "default",
2626 [MPOL_PREFERRED] = "prefer",
2627 [MPOL_BIND] = "bind",
2628 [MPOL_INTERLEAVE] = "interleave",
Lee Schermerhornd3a71032012-10-25 14:16:29 +02002629 [MPOL_LOCAL] = "local",
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002630};
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002631
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002632
2633#ifdef CONFIG_TMPFS
2634/**
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002635 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002636 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002637 * @mpol: pointer to struct mempolicy pointer, returned on success.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002638 *
2639 * Format of input:
2640 * <mode>[=<flags>][:<nodelist>]
2641 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002642 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002643 */
Hugh Dickinsa7a88b22013-01-02 02:04:23 -08002644int mpol_parse_str(char *str, struct mempolicy **mpol)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002645{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002646 struct mempolicy *new = NULL;
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002647 unsigned short mode;
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002648 unsigned short mode_flags;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002649 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002650 char *nodelist = strchr(str, ':');
2651 char *flags = strchr(str, '=');
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002652 int err = 1;
2653
2654 if (nodelist) {
2655 /* NUL-terminate mode or flags string */
2656 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002657 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002658 goto out;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002659 if (!nodes_subset(nodes, node_states[N_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002660 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002661 } else
2662 nodes_clear(nodes);
2663
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002664 if (flags)
2665 *flags++ = '\0'; /* terminate mode string */
2666
Peter Zijlstra479e2802012-10-25 14:16:28 +02002667 for (mode = 0; mode < MPOL_MAX; mode++) {
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002668 if (!strcmp(str, policy_modes[mode])) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002669 break;
2670 }
2671 }
Mel Gormana7200942012-11-16 09:37:58 +00002672 if (mode >= MPOL_MAX)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002673 goto out;
2674
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002675 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002676 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002677 /*
2678 * Insist on a nodelist of one node only
2679 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002680 if (nodelist) {
2681 char *rest = nodelist;
2682 while (isdigit(*rest))
2683 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002684 if (*rest)
2685 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002686 }
2687 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002688 case MPOL_INTERLEAVE:
2689 /*
2690 * Default to online nodes with memory if no nodelist
2691 */
2692 if (!nodelist)
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002693 nodes = node_states[N_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002694 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002695 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002696 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002697 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002698 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002699 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002700 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002701 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002702 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002703 case MPOL_DEFAULT:
2704 /*
2705 * Insist on a empty nodelist
2706 */
2707 if (!nodelist)
2708 err = 0;
2709 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002710 case MPOL_BIND:
2711 /*
2712 * Insist on a nodelist
2713 */
2714 if (!nodelist)
2715 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002716 }
2717
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002718 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002719 if (flags) {
2720 /*
2721 * Currently, we only support two mutually exclusive
2722 * mode flags.
2723 */
2724 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002725 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002726 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002727 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002728 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002729 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002730 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002731
2732 new = mpol_new(mode, mode_flags, &nodes);
2733 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002734 goto out;
2735
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002736 /*
2737 * Save nodes for mpol_to_str() to show the tmpfs mount options
2738 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2739 */
2740 if (mode != MPOL_PREFERRED)
2741 new->v.nodes = nodes;
2742 else if (nodelist)
2743 new->v.preferred_node = first_node(nodes);
2744 else
2745 new->flags |= MPOL_F_LOCAL;
2746
2747 /*
2748 * Save nodes for contextualization: this will be used to "clone"
2749 * the mempolicy in a specific context [cpuset] at a later time.
2750 */
2751 new->w.user_nodemask = nodes;
2752
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002753 err = 0;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002754
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002755out:
2756 /* Restore string for error message */
2757 if (nodelist)
2758 *--nodelist = ':';
2759 if (flags)
2760 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002761 if (!err)
2762 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002763 return err;
2764}
2765#endif /* CONFIG_TMPFS */
2766
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002767/**
2768 * mpol_to_str - format a mempolicy structure for printing
2769 * @buffer: to contain formatted mempolicy string
2770 * @maxlen: length of @buffer
2771 * @pol: pointer to mempolicy to be formatted
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002772 *
David Rientjes948927e2013-11-12 15:07:28 -08002773 * Convert @pol into a string. If @buffer is too short, truncate the string.
2774 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2775 * longest flag, "relative", and to display at least a few node ids.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002776 */
David Rientjes948927e2013-11-12 15:07:28 -08002777void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002778{
2779 char *p = buffer;
David Rientjes948927e2013-11-12 15:07:28 -08002780 nodemask_t nodes = NODE_MASK_NONE;
2781 unsigned short mode = MPOL_DEFAULT;
2782 unsigned short flags = 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002783
David Rientjes8790c712014-01-30 15:46:08 -08002784 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002785 mode = pol->mode;
David Rientjes948927e2013-11-12 15:07:28 -08002786 flags = pol->flags;
2787 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002788
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002789 switch (mode) {
2790 case MPOL_DEFAULT:
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002791 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002792 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002793 if (flags & MPOL_F_LOCAL)
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002794 mode = MPOL_LOCAL;
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002795 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002796 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002797 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002798 case MPOL_BIND:
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002799 case MPOL_INTERLEAVE:
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002800 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002801 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002802 default:
David Rientjes948927e2013-11-12 15:07:28 -08002803 WARN_ON_ONCE(1);
2804 snprintf(p, maxlen, "unknown");
2805 return;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002806 }
2807
David Rientjesb7a9f422013-11-21 14:32:06 -08002808 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002809
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002810 if (flags & MPOL_MODE_FLAGS) {
David Rientjes948927e2013-11-12 15:07:28 -08002811 p += snprintf(p, buffer + maxlen - p, "=");
David Rientjesf5b087b2008-04-28 02:12:27 -07002812
Lee Schermerhorn22919902008-04-28 02:13:22 -07002813 /*
2814 * Currently, the only defined flags are mutually exclusive
2815 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002816 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002817 p += snprintf(p, buffer + maxlen - p, "static");
2818 else if (flags & MPOL_F_RELATIVE_NODES)
2819 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002820 }
2821
Tejun Heo9e763e02015-02-13 14:38:02 -08002822 if (!nodes_empty(nodes))
2823 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2824 nodemask_pr_args(&nodes));
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002825}