blob: 36cc01bc950a6e5f8ef94c0d0506f463e01791d9 [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;
Kirill A. Shutemov248db922016-01-15 16:54:14 -0800492 int nid, ret;
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
Kirill A. Shutemov248db922016-01-15 16:54:14 -0800496 if (pmd_trans_huge(*pmd)) {
497 ptl = pmd_lock(walk->mm, pmd);
498 if (pmd_trans_huge(*pmd)) {
499 page = pmd_page(*pmd);
500 if (is_huge_zero_page(page)) {
501 spin_unlock(ptl);
502 split_huge_pmd(vma, pmd, addr);
503 } else {
504 get_page(page);
505 spin_unlock(ptl);
506 lock_page(page);
507 ret = split_huge_page(page);
508 unlock_page(page);
509 put_page(page);
510 if (ret)
511 return 0;
512 }
513 } else {
514 spin_unlock(ptl);
515 }
516 }
Hugh Dickins91612e02005-06-21 17:15:07 -0700517
Kirill A. Shutemov248db922016-01-15 16:54:14 -0800518retry:
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800519 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
520 for (; addr != end; pte++, addr += PAGE_SIZE) {
Hugh Dickins91612e02005-06-21 17:15:07 -0700521 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800523 page = vm_normal_page(vma, addr, *pte);
524 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800526 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800527 * vm_normal_page() filters out zero pages, but there might
528 * still be PageReserved pages to skip, perhaps in a VDSO.
Nick Piggin053837f2006-01-18 17:42:27 -0800529 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800530 if (PageReserved(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800531 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800532 nid = page_to_nid(page);
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800533 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
Christoph Lameter38e35862006-01-08 01:01:01 -0800534 continue;
Kirill A. Shutemov0a2e2802016-03-09 14:08:30 -0800535 if (PageTransCompound(page) && PageAnon(page)) {
Kirill A. Shutemov248db922016-01-15 16:54:14 -0800536 get_page(page);
537 pte_unmap_unlock(pte, ptl);
538 lock_page(page);
539 ret = split_huge_page(page);
540 unlock_page(page);
541 put_page(page);
542 /* Failed to split -- skip. */
543 if (ret) {
544 pte = pte_offset_map_lock(walk->mm, pmd,
545 addr, &ptl);
546 continue;
547 }
548 goto retry;
549 }
Christoph Lameter38e35862006-01-08 01:01:01 -0800550
Kirill A. Shutemov77bf45e2016-02-05 15:36:33 -0800551 migrate_page_add(page, qp->pagelist, flags);
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800552 }
553 pte_unmap_unlock(pte - 1, ptl);
554 cond_resched();
555 return 0;
Hugh Dickins91612e02005-06-21 17:15:07 -0700556}
557
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800558static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
559 unsigned long addr, unsigned long end,
560 struct mm_walk *walk)
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700561{
562#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800563 struct queue_pages *qp = walk->private;
564 unsigned long flags = qp->flags;
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700565 int nid;
566 struct page *page;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800567 spinlock_t *ptl;
Naoya Horiguchid4c54912014-06-06 10:00:01 -0400568 pte_t entry;
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700569
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800570 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
571 entry = huge_ptep_get(pte);
Naoya Horiguchid4c54912014-06-06 10:00:01 -0400572 if (!pte_present(entry))
573 goto unlock;
574 page = pte_page(entry);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700575 nid = page_to_nid(page);
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800576 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700577 goto unlock;
578 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
579 if (flags & (MPOL_MF_MOVE_ALL) ||
580 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800581 isolate_huge_page(page, qp->pagelist);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700582unlock:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800583 spin_unlock(ptl);
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700584#else
585 BUG();
586#endif
Hugh Dickins91612e02005-06-21 17:15:07 -0700587 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588}
589
Aneesh Kumar K.V58772312013-12-06 00:08:22 +0530590#ifdef CONFIG_NUMA_BALANCING
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200591/*
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200592 * This is used to mark a range of virtual addresses to be inaccessible.
593 * These are later cleared by a NUMA hinting fault. Depending on these
594 * faults, pages may be migrated for better NUMA placement.
595 *
596 * This is assuming that NUMA faults are handled using PROT_NONE. If
597 * an architecture makes a different choice, it will need further
598 * changes to the core.
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200599 */
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200600unsigned long change_prot_numa(struct vm_area_struct *vma,
601 unsigned long addr, unsigned long end)
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200602{
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200603 int nr_updated;
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200604
Mel Gorman4d942462015-02-12 14:58:28 -0800605 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
Mel Gorman03c5a6e2012-11-02 14:52:48 +0000606 if (nr_updated)
607 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200608
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200609 return nr_updated;
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200610}
611#else
612static unsigned long change_prot_numa(struct vm_area_struct *vma,
613 unsigned long addr, unsigned long end)
614{
615 return 0;
616}
Aneesh Kumar K.V58772312013-12-06 00:08:22 +0530617#endif /* CONFIG_NUMA_BALANCING */
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200618
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800619static int queue_pages_test_walk(unsigned long start, unsigned long end,
620 struct mm_walk *walk)
621{
622 struct vm_area_struct *vma = walk->vma;
623 struct queue_pages *qp = walk->private;
624 unsigned long endvma = vma->vm_end;
625 unsigned long flags = qp->flags;
626
Kirill A. Shutemov77bf45e2016-02-05 15:36:33 -0800627 if (!vma_migratable(vma))
Naoya Horiguchi48684a62015-02-11 15:28:06 -0800628 return 1;
629
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800630 if (endvma > end)
631 endvma = end;
632 if (vma->vm_start > start)
633 start = vma->vm_start;
634
635 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
636 if (!vma->vm_next && vma->vm_end < end)
637 return -EFAULT;
638 if (qp->prev && qp->prev->vm_end < vma->vm_start)
639 return -EFAULT;
640 }
641
642 qp->prev = vma;
643
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800644 if (flags & MPOL_MF_LAZY) {
645 /* Similar to task_numa_work, skip inaccessible VMAs */
Liang Chen4355c012016-03-15 14:56:42 -0700646 if (!is_vm_hugetlb_page(vma) &&
647 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
648 !(vma->vm_flags & VM_MIXEDMAP))
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800649 change_prot_numa(vma, start, endvma);
650 return 1;
651 }
652
Kirill A. Shutemov77bf45e2016-02-05 15:36:33 -0800653 /* queue pages from current vma */
654 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800655 return 0;
656 return 1;
657}
658
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800659/*
Naoya Horiguchi98094942013-09-11 14:22:14 -0700660 * Walk through page tables and collect pages to be migrated.
661 *
662 * If pages found in a given range are on a set of nodes (determined by
663 * @nodes and @flags,) it's isolated and queued to the pagelist which is
664 * passed via @private.)
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800665 */
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -0700666static int
Naoya Horiguchi98094942013-09-11 14:22:14 -0700667queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800668 nodemask_t *nodes, unsigned long flags,
669 struct list_head *pagelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670{
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800671 struct queue_pages qp = {
672 .pagelist = pagelist,
673 .flags = flags,
674 .nmask = nodes,
675 .prev = NULL,
676 };
677 struct mm_walk queue_pages_walk = {
678 .hugetlb_entry = queue_pages_hugetlb,
679 .pmd_entry = queue_pages_pte_range,
680 .test_walk = queue_pages_test_walk,
681 .mm = mm,
682 .private = &qp,
683 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684
Naoya Horiguchi6f4576e2015-02-11 15:28:03 -0800685 return walk_page_range(start, end, &queue_pages_walk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686}
687
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700688/*
689 * Apply policy to a single VMA
690 * This must be called with the mmap_sem held for writing.
691 */
692static int vma_replace_policy(struct vm_area_struct *vma,
693 struct mempolicy *pol)
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700694{
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700695 int err;
696 struct mempolicy *old;
697 struct mempolicy *new;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700698
699 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
700 vma->vm_start, vma->vm_end, vma->vm_pgoff,
701 vma->vm_ops, vma->vm_file,
702 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
703
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700704 new = mpol_dup(pol);
705 if (IS_ERR(new))
706 return PTR_ERR(new);
707
708 if (vma->vm_ops && vma->vm_ops->set_policy) {
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700709 err = vma->vm_ops->set_policy(vma, new);
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700710 if (err)
711 goto err_out;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700712 }
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700713
714 old = vma->vm_policy;
715 vma->vm_policy = new; /* protected by mmap_sem */
716 mpol_put(old);
717
718 return 0;
719 err_out:
720 mpol_put(new);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700721 return err;
722}
723
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800725static int mbind_range(struct mm_struct *mm, unsigned long start,
726 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727{
728 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800729 struct vm_area_struct *prev;
730 struct vm_area_struct *vma;
731 int err = 0;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800732 pgoff_t pgoff;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800733 unsigned long vmstart;
734 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735
Linus Torvalds097d5912012-03-06 18:23:36 -0800736 vma = find_vma(mm, start);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800737 if (!vma || vma->vm_start > start)
738 return -EFAULT;
739
Linus Torvalds097d5912012-03-06 18:23:36 -0800740 prev = vma->vm_prev;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800741 if (start > vma->vm_start)
742 prev = vma;
743
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800744 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800746 vmstart = max(start, vma->vm_start);
747 vmend = min(end, vma->vm_end);
748
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800749 if (mpol_equal(vma_policy(vma), new_pol))
750 continue;
751
752 pgoff = vma->vm_pgoff +
753 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800754 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
Andrea Arcangeli19a809a2015-09-04 15:46:24 -0700755 vma->anon_vma, vma->vm_file, pgoff,
756 new_pol, vma->vm_userfaultfd_ctx);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800757 if (prev) {
758 vma = prev;
759 next = vma->vm_next;
Oleg Nesterov3964acd2013-07-31 13:53:28 -0700760 if (mpol_equal(vma_policy(vma), new_pol))
761 continue;
762 /* vma_merge() joined vma && vma->next, case 8 */
763 goto replace;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800764 }
765 if (vma->vm_start != vmstart) {
766 err = split_vma(vma->vm_mm, vma, vmstart, 1);
767 if (err)
768 goto out;
769 }
770 if (vma->vm_end != vmend) {
771 err = split_vma(vma->vm_mm, vma, vmend, 0);
772 if (err)
773 goto out;
774 }
Oleg Nesterov3964acd2013-07-31 13:53:28 -0700775 replace:
KOSAKI Motohiro869833f22012-10-08 16:29:16 -0700776 err = vma_replace_policy(vma, new_pol);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700777 if (err)
778 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800780
781 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 return err;
783}
784
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700786static long do_set_mempolicy(unsigned short mode, unsigned short flags,
787 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788{
Miao Xie58568d22009-06-16 15:31:49 -0700789 struct mempolicy *new, *old;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700790 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700791 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700793 if (!scratch)
794 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700795
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700796 new = mpol_new(mode, flags, nodes);
797 if (IS_ERR(new)) {
798 ret = PTR_ERR(new);
799 goto out;
800 }
Oleg Nesterov2c7c3a72014-10-09 15:27:55 -0700801
Miao Xie58568d22009-06-16 15:31:49 -0700802 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700803 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700804 if (ret) {
805 task_unlock(current);
Miao Xie58568d22009-06-16 15:31:49 -0700806 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700807 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700808 }
809 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 current->mempolicy = new;
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700811 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700812 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700813 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700814 task_unlock(current);
Miao Xie58568d22009-06-16 15:31:49 -0700815 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700816 ret = 0;
817out:
818 NODEMASK_SCRATCH_FREE(scratch);
819 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820}
821
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700822/*
823 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700824 *
825 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700826 */
827static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700829 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700830 if (p == &default_policy)
831 return;
832
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700833 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700834 case MPOL_BIND:
835 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700837 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 break;
839 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700840 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700841 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700842 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 break;
844 default:
845 BUG();
846 }
847}
848
Dave Hansend4edcf02016-02-12 13:01:56 -0800849static int lookup_node(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850{
851 struct page *p;
852 int err;
853
Dave Hansend4edcf02016-02-12 13:01:56 -0800854 err = get_user_pages(addr & PAGE_MASK, 1, 0, 0, &p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 if (err >= 0) {
856 err = page_to_nid(p);
857 put_page(p);
858 }
859 return err;
860}
861
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700863static long do_get_mempolicy(int *policy, nodemask_t *nmask,
864 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700866 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 struct mm_struct *mm = current->mm;
868 struct vm_area_struct *vma = NULL;
869 struct mempolicy *pol = current->mempolicy;
870
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700871 if (flags &
872 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700874
875 if (flags & MPOL_F_MEMS_ALLOWED) {
876 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
877 return -EINVAL;
878 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700879 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700880 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700881 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700882 return 0;
883 }
884
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700886 /*
887 * Do NOT fall back to task policy if the
888 * vma/shared policy at addr is NULL. We
889 * want to return MPOL_DEFAULT in this case.
890 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 down_read(&mm->mmap_sem);
892 vma = find_vma_intersection(mm, addr, addr+1);
893 if (!vma) {
894 up_read(&mm->mmap_sem);
895 return -EFAULT;
896 }
897 if (vma->vm_ops && vma->vm_ops->get_policy)
898 pol = vma->vm_ops->get_policy(vma, addr);
899 else
900 pol = vma->vm_policy;
901 } else if (addr)
902 return -EINVAL;
903
904 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700905 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906
907 if (flags & MPOL_F_NODE) {
908 if (flags & MPOL_F_ADDR) {
Dave Hansend4edcf02016-02-12 13:01:56 -0800909 err = lookup_node(addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 if (err < 0)
911 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700912 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700914 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700915 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 } else {
917 err = -EINVAL;
918 goto out;
919 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700920 } else {
921 *policy = pol == &default_policy ? MPOL_DEFAULT :
922 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700923 /*
924 * Internal mempolicy flags must be masked off before exposing
925 * the policy to userspace.
926 */
927 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700928 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
930 if (vma) {
931 up_read(&current->mm->mmap_sem);
932 vma = NULL;
933 }
934
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700936 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700937 if (mpol_store_user_nodemask(pol)) {
938 *nmask = pol->w.user_nodemask;
939 } else {
940 task_lock(current);
941 get_policy_nodemask(pol, nmask);
942 task_unlock(current);
943 }
Miao Xie58568d22009-06-16 15:31:49 -0700944 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945
946 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700947 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 if (vma)
949 up_read(&current->mm->mmap_sem);
950 return err;
951}
952
Christoph Lameterb20a3502006-03-22 00:09:12 -0800953#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700954/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800955 * page migration
956 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800957static void migrate_page_add(struct page *page, struct list_head *pagelist,
958 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800959{
960 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800961 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800962 */
Nick Piggin62695a82008-10-18 20:26:09 -0700963 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
964 if (!isolate_lru_page(page)) {
965 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800966 inc_zone_page_state(page, NR_ISOLATED_ANON +
967 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700968 }
969 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800970}
971
Christoph Lameter742755a2006-06-23 02:03:55 -0700972static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700973{
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -0700974 if (PageHuge(page))
975 return alloc_huge_page_node(page_hstate(compound_head(page)),
976 node);
977 else
Vlastimil Babka96db8002015-09-08 15:03:50 -0700978 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
David Rientjesb360edb2015-04-14 15:46:52 -0700979 __GFP_THISNODE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700980}
981
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800982/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800983 * Migrate pages from one node to a target node.
984 * Returns error or the number of pages not migrated.
985 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700986static int migrate_to_node(struct mm_struct *mm, int source, int dest,
987 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800988{
989 nodemask_t nmask;
990 LIST_HEAD(pagelist);
991 int err = 0;
992
993 nodes_clear(nmask);
994 node_set(source, nmask);
995
Minchan Kim08270802012-10-08 16:33:38 -0700996 /*
997 * This does not "check" the range but isolates all pages that
998 * need migration. Between passing in the full user address
999 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1000 */
1001 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
Naoya Horiguchi98094942013-09-11 14:22:14 -07001002 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001003 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1004
Minchan Kimcf608ac2010-10-26 14:21:29 -07001005 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001006 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001007 MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001008 if (err)
Naoya Horiguchie2d8cf42013-09-11 14:22:03 -07001009 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001010 }
Christoph Lameter95a402c2006-06-23 02:03:53 -07001011
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001012 return err;
1013}
1014
1015/*
1016 * Move pages between the two nodesets so as to preserve the physical
1017 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -08001018 *
1019 * Returns the number of page that could not be moved.
1020 */
Andrew Morton0ce72d42012-05-29 15:06:24 -07001021int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1022 const nodemask_t *to, int flags)
Christoph Lameter39743882006-01-08 01:00:51 -08001023{
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001024 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001025 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001026 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -08001027
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001028 err = migrate_prep();
1029 if (err)
1030 return err;
1031
Lee Schermerhorn53f25562008-04-28 02:13:20 -07001032 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -08001033
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001034 /*
1035 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1036 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1037 * bit in 'tmp', and return that <source, dest> pair for migration.
1038 * The pair of nodemasks 'to' and 'from' define the map.
1039 *
1040 * If no pair of bits is found that way, fallback to picking some
1041 * pair of 'source' and 'dest' bits that are not the same. If the
1042 * 'source' and 'dest' bits are the same, this represents a node
1043 * that will be migrating to itself, so no pages need move.
1044 *
1045 * If no bits are left in 'tmp', or if all remaining bits left
1046 * in 'tmp' correspond to the same bit in 'to', return false
1047 * (nothing left to migrate).
1048 *
1049 * This lets us pick a pair of nodes to migrate between, such that
1050 * if possible the dest node is not already occupied by some other
1051 * source node, minimizing the risk of overloading the memory on a
1052 * node that would happen if we migrated incoming memory to a node
1053 * before migrating outgoing memory source that same node.
1054 *
1055 * A single scan of tmp is sufficient. As we go, we remember the
1056 * most recent <s, d> pair that moved (s != d). If we find a pair
1057 * that not only moved, but what's better, moved to an empty slot
1058 * (d is not set in tmp), then we break out then, with that pair.
Justin P. Mattockae0e47f2011-03-01 15:06:02 +01001059 * Otherwise when we finish scanning from_tmp, we at least have the
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001060 * most recent <s, d> pair that moved. If we get all the way through
1061 * the scan of tmp without finding any node that moved, much less
1062 * moved to an empty node, then there is nothing left worth migrating.
1063 */
Christoph Lameterd4984712006-01-08 01:00:55 -08001064
Andrew Morton0ce72d42012-05-29 15:06:24 -07001065 tmp = *from;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001066 while (!nodes_empty(tmp)) {
1067 int s,d;
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001068 int source = NUMA_NO_NODE;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001069 int dest = 0;
1070
1071 for_each_node_mask(s, tmp) {
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001072
1073 /*
1074 * do_migrate_pages() tries to maintain the relative
1075 * node relationship of the pages established between
1076 * threads and memory areas.
1077 *
1078 * However if the number of source nodes is not equal to
1079 * the number of destination nodes we can not preserve
1080 * this node relative relationship. In that case, skip
1081 * copying memory from a node that is in the destination
1082 * mask.
1083 *
1084 * Example: [2,3,4] -> [3,4,5] moves everything.
1085 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1086 */
1087
Andrew Morton0ce72d42012-05-29 15:06:24 -07001088 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1089 (node_isset(s, *to)))
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001090 continue;
1091
Andrew Morton0ce72d42012-05-29 15:06:24 -07001092 d = node_remap(s, *from, *to);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001093 if (s == d)
1094 continue;
1095
1096 source = s; /* Node moved. Memorize */
1097 dest = d;
1098
1099 /* dest not in remaining from nodes? */
1100 if (!node_isset(dest, tmp))
1101 break;
1102 }
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001103 if (source == NUMA_NO_NODE)
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001104 break;
1105
1106 node_clear(source, tmp);
1107 err = migrate_to_node(mm, source, dest, flags);
1108 if (err > 0)
1109 busy += err;
1110 if (err < 0)
1111 break;
Christoph Lameter39743882006-01-08 01:00:51 -08001112 }
1113 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001114 if (err < 0)
1115 return err;
1116 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001117
Christoph Lameter39743882006-01-08 01:00:51 -08001118}
1119
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001120/*
1121 * Allocate a new page for page migration based on vma policy.
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001122 * Start by assuming the page is mapped by the same vma as contains @start.
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001123 * Search forward from there, if not. N.B., this assumes that the
1124 * list of pages handed to migrate_pages()--which is how we get here--
1125 * is in virtual address order.
1126 */
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001127static struct page *new_page(struct page *page, unsigned long start, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001128{
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001129 struct vm_area_struct *vma;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001130 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001131
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001132 vma = find_vma(current->mm, start);
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001133 while (vma) {
1134 address = page_address_in_vma(page, vma);
1135 if (address != -EFAULT)
1136 break;
1137 vma = vma->vm_next;
1138 }
1139
Wanpeng Li11c731e2013-12-18 17:08:56 -08001140 if (PageHuge(page)) {
Michal Hockocc817172014-01-23 15:53:15 -08001141 BUG_ON(!vma);
1142 return alloc_huge_page_noerr(vma, address, 1);
Wanpeng Li11c731e2013-12-18 17:08:56 -08001143 }
1144 /*
1145 * if !vma, alloc_page_vma() will use task or system default policy
1146 */
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001147 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001148}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001149#else
1150
1151static void migrate_page_add(struct page *page, struct list_head *pagelist,
1152 unsigned long flags)
1153{
1154}
1155
Andrew Morton0ce72d42012-05-29 15:06:24 -07001156int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1157 const nodemask_t *to, int flags)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001158{
1159 return -ENOSYS;
1160}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001161
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001162static struct page *new_page(struct page *page, unsigned long start, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001163{
1164 return NULL;
1165}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001166#endif
1167
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001168static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001169 unsigned short mode, unsigned short mode_flags,
1170 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001171{
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001172 struct mm_struct *mm = current->mm;
1173 struct mempolicy *new;
1174 unsigned long end;
1175 int err;
1176 LIST_HEAD(pagelist);
1177
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001178 if (flags & ~(unsigned long)MPOL_MF_VALID)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001179 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001180 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001181 return -EPERM;
1182
1183 if (start & ~PAGE_MASK)
1184 return -EINVAL;
1185
1186 if (mode == MPOL_DEFAULT)
1187 flags &= ~MPOL_MF_STRICT;
1188
1189 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1190 end = start + len;
1191
1192 if (end < start)
1193 return -EINVAL;
1194 if (end == start)
1195 return 0;
1196
David Rientjes028fec42008-04-28 02:12:25 -07001197 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001198 if (IS_ERR(new))
1199 return PTR_ERR(new);
1200
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001201 if (flags & MPOL_MF_LAZY)
1202 new->flags |= MPOL_F_MOF;
1203
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001204 /*
1205 * If we are using the default policy then operation
1206 * on discontinuous address spaces is okay after all
1207 */
1208 if (!new)
1209 flags |= MPOL_MF_DISCONTIG_OK;
1210
David Rientjes028fec42008-04-28 02:12:25 -07001211 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1212 start, start + len, mode, mode_flags,
David Rientjes00ef2d22013-02-22 16:35:36 -08001213 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001214
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001215 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1216
1217 err = migrate_prep();
1218 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001219 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001220 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001221 {
1222 NODEMASK_SCRATCH(scratch);
1223 if (scratch) {
1224 down_write(&mm->mmap_sem);
1225 task_lock(current);
1226 err = mpol_set_nodemask(new, nmask, scratch);
1227 task_unlock(current);
1228 if (err)
1229 up_write(&mm->mmap_sem);
1230 } else
1231 err = -ENOMEM;
1232 NODEMASK_SCRATCH_FREE(scratch);
1233 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001234 if (err)
1235 goto mpol_out;
1236
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001237 err = queue_pages_range(mm, start, end, nmask,
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001238 flags | MPOL_MF_INVERT, &pagelist);
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001239 if (!err)
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001240 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001241
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001242 if (!err) {
1243 int nr_failed = 0;
1244
Minchan Kimcf608ac2010-10-26 14:21:29 -07001245 if (!list_empty(&pagelist)) {
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001246 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
Hugh Dickinsd05f0cdcb2014-06-23 13:22:07 -07001247 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1248 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001249 if (nr_failed)
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001250 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001251 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001252
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001253 if (nr_failed && (flags & MPOL_MF_STRICT))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001254 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001255 } else
Joonsoo Kimb0e5fd72013-12-18 17:08:51 -08001256 putback_movable_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001257
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001258 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001259 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001260 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001261 return err;
1262}
1263
Christoph Lameter39743882006-01-08 01:00:51 -08001264/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001265 * User space interface with variable sized bitmaps for nodelists.
1266 */
1267
1268/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001269static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001270 unsigned long maxnode)
1271{
1272 unsigned long k;
1273 unsigned long nlongs;
1274 unsigned long endmask;
1275
1276 --maxnode;
1277 nodes_clear(*nodes);
1278 if (maxnode == 0 || !nmask)
1279 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001280 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001281 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001282
1283 nlongs = BITS_TO_LONGS(maxnode);
1284 if ((maxnode % BITS_PER_LONG) == 0)
1285 endmask = ~0UL;
1286 else
1287 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1288
1289 /* When the user specified more nodes than supported just check
1290 if the non supported part is all zero. */
1291 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1292 if (nlongs > PAGE_SIZE/sizeof(long))
1293 return -EINVAL;
1294 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1295 unsigned long t;
1296 if (get_user(t, nmask + k))
1297 return -EFAULT;
1298 if (k == nlongs - 1) {
1299 if (t & endmask)
1300 return -EINVAL;
1301 } else if (t)
1302 return -EINVAL;
1303 }
1304 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1305 endmask = ~0UL;
1306 }
1307
1308 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1309 return -EFAULT;
1310 nodes_addr(*nodes)[nlongs-1] &= endmask;
1311 return 0;
1312}
1313
1314/* Copy a kernel node mask to user space */
1315static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1316 nodemask_t *nodes)
1317{
1318 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1319 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1320
1321 if (copy > nbytes) {
1322 if (copy > PAGE_SIZE)
1323 return -EINVAL;
1324 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1325 return -EFAULT;
1326 copy = nbytes;
1327 }
1328 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1329}
1330
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001331SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
Rasmus Villemoesf7f28ca2014-06-04 16:07:57 -07001332 unsigned long, mode, const unsigned long __user *, nmask,
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001333 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001334{
1335 nodemask_t nodes;
1336 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001337 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001338
David Rientjes028fec42008-04-28 02:12:25 -07001339 mode_flags = mode & MPOL_MODE_FLAGS;
1340 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001341 if (mode >= MPOL_MAX)
1342 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001343 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1344 (mode_flags & MPOL_F_RELATIVE_NODES))
1345 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001346 err = get_nodes(&nodes, nmask, maxnode);
1347 if (err)
1348 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001349 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001350}
1351
1352/* Set the process memory policy */
Rasmus Villemoes23c89022014-06-04 16:07:58 -07001353SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001354 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001355{
1356 int err;
1357 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001358 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001359
David Rientjes028fec42008-04-28 02:12:25 -07001360 flags = mode & MPOL_MODE_FLAGS;
1361 mode &= ~MPOL_MODE_FLAGS;
1362 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001363 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001364 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1365 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001366 err = get_nodes(&nodes, nmask, maxnode);
1367 if (err)
1368 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001369 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001370}
1371
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001372SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1373 const unsigned long __user *, old_nodes,
1374 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001375{
David Howellsc69e8d92008-11-14 10:39:19 +11001376 const struct cred *cred = current_cred(), *tcred;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001377 struct mm_struct *mm = NULL;
Christoph Lameter39743882006-01-08 01:00:51 -08001378 struct task_struct *task;
Christoph Lameter39743882006-01-08 01:00:51 -08001379 nodemask_t task_nodes;
1380 int err;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001381 nodemask_t *old;
1382 nodemask_t *new;
1383 NODEMASK_SCRATCH(scratch);
Christoph Lameter39743882006-01-08 01:00:51 -08001384
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001385 if (!scratch)
1386 return -ENOMEM;
Christoph Lameter39743882006-01-08 01:00:51 -08001387
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001388 old = &scratch->mask1;
1389 new = &scratch->mask2;
1390
1391 err = get_nodes(old, old_nodes, maxnode);
Christoph Lameter39743882006-01-08 01:00:51 -08001392 if (err)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001393 goto out;
1394
1395 err = get_nodes(new, new_nodes, maxnode);
1396 if (err)
1397 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001398
1399 /* Find the mm_struct */
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001400 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001401 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001402 if (!task) {
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001403 rcu_read_unlock();
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001404 err = -ESRCH;
1405 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001406 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001407 get_task_struct(task);
Christoph Lameter39743882006-01-08 01:00:51 -08001408
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001409 err = -EINVAL;
Christoph Lameter39743882006-01-08 01:00:51 -08001410
1411 /*
1412 * Check if this process has the right to modify the specified
1413 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001414 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001415 * userid as the target process.
1416 */
David Howellsc69e8d92008-11-14 10:39:19 +11001417 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001418 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1419 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001420 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001421 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001422 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001423 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001424 }
David Howellsc69e8d92008-11-14 10:39:19 +11001425 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001426
1427 task_nodes = cpuset_mems_allowed(task);
1428 /* Is the user allowed to access the target nodes? */
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001429 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001430 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001431 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001432 }
1433
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08001434 if (!nodes_subset(*new, node_states[N_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001435 err = -EINVAL;
Christoph Lameter3268c632012-03-21 16:34:06 -07001436 goto out_put;
Christoph Lameter3b42d282007-08-31 00:12:08 -07001437 }
1438
David Quigley86c3a762006-06-23 02:04:02 -07001439 err = security_task_movememory(task);
1440 if (err)
Christoph Lameter3268c632012-03-21 16:34:06 -07001441 goto out_put;
David Quigley86c3a762006-06-23 02:04:02 -07001442
Christoph Lameter3268c632012-03-21 16:34:06 -07001443 mm = get_task_mm(task);
1444 put_task_struct(task);
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001445
1446 if (!mm) {
Christoph Lameter3268c632012-03-21 16:34:06 -07001447 err = -EINVAL;
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001448 goto out;
1449 }
1450
1451 err = do_migrate_pages(mm, old, new,
1452 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter3268c632012-03-21 16:34:06 -07001453
1454 mmput(mm);
1455out:
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001456 NODEMASK_SCRATCH_FREE(scratch);
1457
Christoph Lameter39743882006-01-08 01:00:51 -08001458 return err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001459
1460out_put:
1461 put_task_struct(task);
1462 goto out;
1463
Christoph Lameter39743882006-01-08 01:00:51 -08001464}
1465
1466
Christoph Lameter8bccd852005-10-29 18:16:59 -07001467/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001468SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1469 unsigned long __user *, nmask, unsigned long, maxnode,
1470 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001471{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001472 int err;
1473 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001474 nodemask_t nodes;
1475
1476 if (nmask != NULL && maxnode < MAX_NUMNODES)
1477 return -EINVAL;
1478
1479 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1480
1481 if (err)
1482 return err;
1483
1484 if (policy && put_user(pval, policy))
1485 return -EFAULT;
1486
1487 if (nmask)
1488 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1489
1490 return err;
1491}
1492
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493#ifdef CONFIG_COMPAT
1494
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001495COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1496 compat_ulong_t __user *, nmask,
1497 compat_ulong_t, maxnode,
1498 compat_ulong_t, addr, compat_ulong_t, flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499{
1500 long err;
1501 unsigned long __user *nm = NULL;
1502 unsigned long nr_bits, alloc_size;
1503 DECLARE_BITMAP(bm, MAX_NUMNODES);
1504
1505 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1506 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1507
1508 if (nmask)
1509 nm = compat_alloc_user_space(alloc_size);
1510
1511 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1512
1513 if (!err && nmask) {
KAMEZAWA Hiroyuki2bbff6c2011-09-14 16:21:02 -07001514 unsigned long copy_size;
1515 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1516 err = copy_from_user(bm, nm, copy_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 /* ensure entire bitmap is zeroed */
1518 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1519 err |= compat_put_bitmap(nmask, bm, nr_bits);
1520 }
1521
1522 return err;
1523}
1524
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001525COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1526 compat_ulong_t, maxnode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527{
1528 long err = 0;
1529 unsigned long __user *nm = NULL;
1530 unsigned long nr_bits, alloc_size;
1531 DECLARE_BITMAP(bm, MAX_NUMNODES);
1532
1533 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1534 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1535
1536 if (nmask) {
1537 err = compat_get_bitmap(bm, nmask, nr_bits);
1538 nm = compat_alloc_user_space(alloc_size);
1539 err |= copy_to_user(nm, bm, alloc_size);
1540 }
1541
1542 if (err)
1543 return -EFAULT;
1544
1545 return sys_set_mempolicy(mode, nm, nr_bits+1);
1546}
1547
Heiko Carstensc93e0f62014-03-03 16:32:26 +01001548COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1549 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1550 compat_ulong_t, maxnode, compat_ulong_t, flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551{
1552 long err = 0;
1553 unsigned long __user *nm = NULL;
1554 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001555 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556
1557 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1558 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1559
1560 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001561 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001563 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 }
1565
1566 if (err)
1567 return -EFAULT;
1568
1569 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1570}
1571
1572#endif
1573
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001574struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1575 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576{
Oleg Nesterov8d902742014-10-09 15:27:45 -07001577 struct mempolicy *pol = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578
1579 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001580 if (vma->vm_ops && vma->vm_ops->get_policy) {
Oleg Nesterov8d902742014-10-09 15:27:45 -07001581 pol = vma->vm_ops->get_policy(vma, addr);
Mel Gorman00442ad2012-10-08 16:29:20 -07001582 } else if (vma->vm_policy) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 pol = vma->vm_policy;
Mel Gorman00442ad2012-10-08 16:29:20 -07001584
1585 /*
1586 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1587 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1588 * count on these policies which will be dropped by
1589 * mpol_cond_put() later
1590 */
1591 if (mpol_needs_cond_ref(pol))
1592 mpol_get(pol);
1593 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 }
Oleg Nesterovf15ca782014-10-09 15:27:43 -07001595
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001596 return pol;
1597}
1598
1599/*
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001600 * get_vma_policy(@vma, @addr)
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001601 * @vma: virtual memory area whose policy is sought
1602 * @addr: address in @vma for shared policy lookup
1603 *
1604 * Returns effective policy for a VMA at specified address.
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001605 * Falls back to current->mempolicy or system default policy, as necessary.
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001606 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1607 * count--added by the get_policy() vm_op, as appropriate--to protect against
1608 * freeing by another task. It is the caller's responsibility to free the
1609 * extra reference for shared policies.
1610 */
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001611static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1612 unsigned long addr)
Oleg Nesterov74d2c3a2014-10-09 15:27:50 -07001613{
1614 struct mempolicy *pol = __get_vma_policy(vma, addr);
1615
Oleg Nesterov8d902742014-10-09 15:27:45 -07001616 if (!pol)
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001617 pol = get_task_policy(current);
Oleg Nesterov8d902742014-10-09 15:27:45 -07001618
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 return pol;
1620}
1621
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001622bool vma_policy_mof(struct vm_area_struct *vma)
Mel Gormanfc3147242013-10-07 11:29:09 +01001623{
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001624 struct mempolicy *pol;
Oleg Nesterovf15ca782014-10-09 15:27:43 -07001625
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001626 if (vma->vm_ops && vma->vm_ops->get_policy) {
1627 bool ret = false;
Mel Gormanfc3147242013-10-07 11:29:09 +01001628
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001629 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1630 if (pol && (pol->flags & MPOL_F_MOF))
1631 ret = true;
1632 mpol_cond_put(pol);
Mel Gormanfc3147242013-10-07 11:29:09 +01001633
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001634 return ret;
Mel Gormanfc3147242013-10-07 11:29:09 +01001635 }
1636
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001637 pol = vma->vm_policy;
Oleg Nesterov8d902742014-10-09 15:27:45 -07001638 if (!pol)
Oleg Nesterov6b6482b2014-10-09 15:27:48 -07001639 pol = get_task_policy(current);
Oleg Nesterov8d902742014-10-09 15:27:45 -07001640
Mel Gormanfc3147242013-10-07 11:29:09 +01001641 return pol->flags & MPOL_F_MOF;
1642}
1643
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001644static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1645{
1646 enum zone_type dynamic_policy_zone = policy_zone;
1647
1648 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1649
1650 /*
1651 * if policy->v.nodes has movable memory only,
1652 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1653 *
1654 * policy->v.nodes is intersect with node_states[N_MEMORY].
1655 * so if the following test faile, it implies
1656 * policy->v.nodes has movable memory only.
1657 */
1658 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1659 dynamic_policy_zone = ZONE_MOVABLE;
1660
1661 return zone >= dynamic_policy_zone;
1662}
1663
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001664/*
1665 * Return a nodemask representing a mempolicy for filtering nodes for
1666 * page allocation
1667 */
1668static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001669{
1670 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001671 if (unlikely(policy->mode == MPOL_BIND) &&
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001672 apply_policy_zone(policy, gfp_zone(gfp)) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001673 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1674 return &policy->v.nodes;
1675
1676 return NULL;
1677}
1678
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001679/* Return a zonelist indicated by gfp for node representing a mempolicy */
Andi Kleen2f5f9482011-03-04 17:36:29 -08001680static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1681 int nd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682{
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001683 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001685 if (!(policy->flags & MPOL_F_LOCAL))
1686 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 break;
1688 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001689 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001690 * Normally, MPOL_BIND allocations are node-local within the
1691 * allowed nodemask. However, if __GFP_THISNODE is set and the
Bob Liu6eb27e12010-05-24 14:32:00 -07001692 * current node isn't part of the mask, we use the zonelist for
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001693 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001694 */
Mel Gorman19770b32008-04-28 02:12:18 -07001695 if (unlikely(gfp & __GFP_THISNODE) &&
1696 unlikely(!node_isset(nd, policy->v.nodes)))
1697 nd = first_node(policy->v.nodes);
1698 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 BUG();
1701 }
Mel Gorman0e884602008-04-28 02:12:14 -07001702 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703}
1704
1705/* Do dynamic interleaving for a process */
1706static unsigned interleave_nodes(struct mempolicy *policy)
1707{
1708 unsigned nid, next;
1709 struct task_struct *me = current;
1710
1711 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001712 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001714 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001715 if (next < MAX_NUMNODES)
1716 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 return nid;
1718}
1719
Christoph Lameterdc85da12006-01-18 17:42:36 -08001720/*
1721 * Depending on the memory policy provide a node from which to allocate the
1722 * next slab entry.
1723 */
David Rientjes2a389612014-04-07 15:37:29 -07001724unsigned int mempolicy_slab_node(void)
Christoph Lameterdc85da12006-01-18 17:42:36 -08001725{
Andi Kleene7b691b2012-06-09 02:40:03 -07001726 struct mempolicy *policy;
David Rientjes2a389612014-04-07 15:37:29 -07001727 int node = numa_mem_id();
Andi Kleene7b691b2012-06-09 02:40:03 -07001728
1729 if (in_interrupt())
David Rientjes2a389612014-04-07 15:37:29 -07001730 return node;
Andi Kleene7b691b2012-06-09 02:40:03 -07001731
1732 policy = current->mempolicy;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001733 if (!policy || policy->flags & MPOL_F_LOCAL)
David Rientjes2a389612014-04-07 15:37:29 -07001734 return node;
Christoph Lameter765c4502006-09-27 01:50:08 -07001735
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001736 switch (policy->mode) {
1737 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001738 /*
1739 * handled MPOL_F_LOCAL above
1740 */
1741 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001742
Christoph Lameterdc85da12006-01-18 17:42:36 -08001743 case MPOL_INTERLEAVE:
1744 return interleave_nodes(policy);
1745
Mel Gormandd1a2392008-04-28 02:12:17 -07001746 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001747 /*
1748 * Follow bind policy behavior and start allocation at the
1749 * first node.
1750 */
Mel Gorman19770b32008-04-28 02:12:18 -07001751 struct zonelist *zonelist;
1752 struct zone *zone;
1753 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
David Rientjes2a389612014-04-07 15:37:29 -07001754 zonelist = &NODE_DATA(node)->node_zonelists[0];
Mel Gorman19770b32008-04-28 02:12:18 -07001755 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1756 &policy->v.nodes,
1757 &zone);
David Rientjes2a389612014-04-07 15:37:29 -07001758 return zone ? zone->node : node;
Mel Gormandd1a2392008-04-28 02:12:17 -07001759 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001760
Christoph Lameterdc85da12006-01-18 17:42:36 -08001761 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001762 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001763 }
1764}
1765
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766/* Do static interleaving for a VMA with known offset. */
1767static unsigned offset_il_node(struct mempolicy *pol,
1768 struct vm_area_struct *vma, unsigned long off)
1769{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001770 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001771 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 int c;
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001773 int nid = NUMA_NO_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774
David Rientjesf5b087b2008-04-28 02:12:27 -07001775 if (!nnodes)
1776 return numa_node_id();
1777 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778 c = 0;
1779 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001780 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 c++;
1782 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 return nid;
1784}
1785
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001786/* Determine a node number for interleave */
1787static inline unsigned interleave_nid(struct mempolicy *pol,
1788 struct vm_area_struct *vma, unsigned long addr, int shift)
1789{
1790 if (vma) {
1791 unsigned long off;
1792
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001793 /*
1794 * for small pages, there is no difference between
1795 * shift and PAGE_SHIFT, so the bit-shift is safe.
1796 * for huge pages, since vm_pgoff is in units of small
1797 * pages, we need to shift off the always 0 bits to get
1798 * a useful offset.
1799 */
1800 BUG_ON(shift < PAGE_SHIFT);
1801 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001802 off += (addr - vma->vm_start) >> shift;
1803 return offset_il_node(pol, vma, off);
1804 } else
1805 return interleave_nodes(pol);
1806}
1807
Michal Hocko778d3b02011-07-26 16:08:30 -07001808/*
1809 * Return the bit number of a random bit set in the nodemask.
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001810 * (returns NUMA_NO_NODE if nodemask is empty)
Michal Hocko778d3b02011-07-26 16:08:30 -07001811 */
1812int node_random(const nodemask_t *maskp)
1813{
Jianguo Wub76ac7e2013-11-12 15:07:39 -08001814 int w, bit = NUMA_NO_NODE;
Michal Hocko778d3b02011-07-26 16:08:30 -07001815
1816 w = nodes_weight(*maskp);
1817 if (w)
1818 bit = bitmap_ord_to_pos(maskp->bits,
1819 get_random_int() % w, MAX_NUMNODES);
1820 return bit;
1821}
1822
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001823#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001824/*
1825 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
Fabian Frederickb46e14a2014-06-04 16:08:18 -07001826 * @vma: virtual memory area whose policy is sought
1827 * @addr: address in @vma for shared policy lookup and interleave policy
1828 * @gfp_flags: for requested zone
1829 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1830 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001831 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001832 * Returns a zonelist suitable for a huge page allocation and a pointer
1833 * to the struct mempolicy for conditional unref after allocation.
1834 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1835 * @nodemask for filtering the zonelist.
Miao Xiec0ff7452010-05-24 14:32:08 -07001836 *
Mel Gormand26914d2014-04-03 14:47:24 -07001837 * Must be protected by read_mems_allowed_begin()
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001838 */
Mel Gorman396faf02007-07-17 04:03:13 -07001839struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001840 gfp_t gfp_flags, struct mempolicy **mpol,
1841 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001842{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001843 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001844
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07001845 *mpol = get_vma_policy(vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001846 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001847
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001848 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1849 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001850 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001851 } else {
Andi Kleen2f5f9482011-03-04 17:36:29 -08001852 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001853 if ((*mpol)->mode == MPOL_BIND)
1854 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001855 }
1856 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001857}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001858
1859/*
1860 * init_nodemask_of_mempolicy
1861 *
1862 * If the current task's mempolicy is "default" [NULL], return 'false'
1863 * to indicate default policy. Otherwise, extract the policy nodemask
1864 * for 'bind' or 'interleave' policy into the argument nodemask, or
1865 * initialize the argument nodemask to contain the single node for
1866 * 'preferred' or 'local' policy and return 'true' to indicate presence
1867 * of non-default mempolicy.
1868 *
1869 * We don't bother with reference counting the mempolicy [mpol_get/put]
1870 * because the current task is examining it's own mempolicy and a task's
1871 * mempolicy is only ever changed by the task itself.
1872 *
1873 * N.B., it is the caller's responsibility to free a returned nodemask.
1874 */
1875bool init_nodemask_of_mempolicy(nodemask_t *mask)
1876{
1877 struct mempolicy *mempolicy;
1878 int nid;
1879
1880 if (!(mask && current->mempolicy))
1881 return false;
1882
Miao Xiec0ff7452010-05-24 14:32:08 -07001883 task_lock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001884 mempolicy = current->mempolicy;
1885 switch (mempolicy->mode) {
1886 case MPOL_PREFERRED:
1887 if (mempolicy->flags & MPOL_F_LOCAL)
1888 nid = numa_node_id();
1889 else
1890 nid = mempolicy->v.preferred_node;
1891 init_nodemask_of_node(mask, nid);
1892 break;
1893
1894 case MPOL_BIND:
1895 /* Fall through */
1896 case MPOL_INTERLEAVE:
1897 *mask = mempolicy->v.nodes;
1898 break;
1899
1900 default:
1901 BUG();
1902 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001903 task_unlock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001904
1905 return true;
1906}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001907#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001908
David Rientjes6f48d0eb2010-08-09 17:18:52 -07001909/*
1910 * mempolicy_nodemask_intersects
1911 *
1912 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1913 * policy. Otherwise, check for intersection between mask and the policy
1914 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1915 * policy, always return true since it may allocate elsewhere on fallback.
1916 *
1917 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1918 */
1919bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1920 const nodemask_t *mask)
1921{
1922 struct mempolicy *mempolicy;
1923 bool ret = true;
1924
1925 if (!mask)
1926 return ret;
1927 task_lock(tsk);
1928 mempolicy = tsk->mempolicy;
1929 if (!mempolicy)
1930 goto out;
1931
1932 switch (mempolicy->mode) {
1933 case MPOL_PREFERRED:
1934 /*
1935 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1936 * allocate from, they may fallback to other nodes when oom.
1937 * Thus, it's possible for tsk to have allocated memory from
1938 * nodes in mask.
1939 */
1940 break;
1941 case MPOL_BIND:
1942 case MPOL_INTERLEAVE:
1943 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1944 break;
1945 default:
1946 BUG();
1947 }
1948out:
1949 task_unlock(tsk);
1950 return ret;
1951}
1952
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953/* Allocate a page in interleaved policy.
1954 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001955static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1956 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957{
1958 struct zonelist *zl;
1959 struct page *page;
1960
Mel Gorman0e884602008-04-28 02:12:14 -07001961 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001963 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001964 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 return page;
1966}
1967
1968/**
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001969 * alloc_pages_vma - Allocate a page for a VMA.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 *
1971 * @gfp:
1972 * %GFP_USER user allocation.
1973 * %GFP_KERNEL kernel allocations,
1974 * %GFP_HIGHMEM highmem/user allocations,
1975 * %GFP_FS allocation should not call back into a file system.
1976 * %GFP_ATOMIC don't sleep.
1977 *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001978 * @order:Order of the GFP allocation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 * @vma: Pointer to VMA or NULL if not available.
1980 * @addr: Virtual Address of the allocation. Must be inside the VMA.
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001981 * @node: Which node to prefer for allocation (modulo policy).
1982 * @hugepage: for hugepages try only the preferred node if possible
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 *
1984 * This function allocates a page from the kernel page pool and applies
1985 * a NUMA policy associated with the VMA or the current process.
1986 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1987 * mm_struct of the VMA to prevent it from going away. Should be used for
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001988 * all allocations for pages that will be mapped into user space. Returns
1989 * NULL when no page can be allocated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 */
1991struct page *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001992alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001993 unsigned long addr, int node, bool hugepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994{
Mel Gormancc9a6c82012-03-21 16:34:11 -07001995 struct mempolicy *pol;
Miao Xiec0ff7452010-05-24 14:32:08 -07001996 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001997 unsigned int cpuset_mems_cookie;
Vlastimil Babkabe97a412015-02-11 15:27:15 -08001998 struct zonelist *zl;
1999 nodemask_t *nmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000
Mel Gormancc9a6c82012-03-21 16:34:11 -07002001retry_cpuset:
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07002002 pol = get_vma_policy(vma, addr);
Mel Gormand26914d2014-04-03 14:47:24 -07002003 cpuset_mems_cookie = read_mems_allowed_begin();
Mel Gormancc9a6c82012-03-21 16:34:11 -07002004
Vlastimil Babkabe97a412015-02-11 15:27:15 -08002005 if (pol->mode == MPOL_INTERLEAVE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08002007
Andi Kleen8eac5632011-02-25 14:44:28 -08002008 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002009 mpol_cond_put(pol);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002010 page = alloc_page_interleave(gfp, order, nid);
Vlastimil Babkabe97a412015-02-11 15:27:15 -08002011 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 }
Vlastimil Babkabe97a412015-02-11 15:27:15 -08002013
Vlastimil Babka0867a572015-06-24 16:58:48 -07002014 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2015 int hpage_node = node;
2016
2017 /*
2018 * For hugepage allocation and non-interleave policy which
2019 * allows the current node (or other explicitly preferred
2020 * node) we only try to allocate from the current/preferred
2021 * node and don't fall back to other nodes, as the cost of
2022 * remote accesses would likely offset THP benefits.
2023 *
2024 * If the policy is interleave, or does not allow the current
2025 * node in its nodemask, we allocate the standard way.
2026 */
2027 if (pol->mode == MPOL_PREFERRED &&
2028 !(pol->flags & MPOL_F_LOCAL))
2029 hpage_node = pol->v.preferred_node;
2030
2031 nmask = policy_nodemask(gfp, pol);
2032 if (!nmask || node_isset(hpage_node, *nmask)) {
2033 mpol_cond_put(pol);
Vlastimil Babka96db8002015-09-08 15:03:50 -07002034 page = __alloc_pages_node(hpage_node,
Vlastimil Babka0867a572015-06-24 16:58:48 -07002035 gfp | __GFP_THISNODE, order);
2036 goto out;
2037 }
2038 }
2039
Vlastimil Babkabe97a412015-02-11 15:27:15 -08002040 nmask = policy_nodemask(gfp, pol);
2041 zl = policy_zonelist(gfp, pol, node);
Oleg Nesterov23867402014-10-09 15:27:41 -07002042 mpol_cond_put(pol);
Vlastimil Babkabe97a412015-02-11 15:27:15 -08002043 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2044out:
Mel Gormand26914d2014-04-03 14:47:24 -07002045 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07002046 goto retry_cpuset;
Miao Xiec0ff7452010-05-24 14:32:08 -07002047 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048}
2049
2050/**
2051 * alloc_pages_current - Allocate pages.
2052 *
2053 * @gfp:
2054 * %GFP_USER user allocation,
2055 * %GFP_KERNEL kernel allocation,
2056 * %GFP_HIGHMEM highmem allocation,
2057 * %GFP_FS don't call back into a file system.
2058 * %GFP_ATOMIC don't sleep.
2059 * @order: Power of two of allocation size in pages. 0 is a single page.
2060 *
2061 * Allocate a page from the kernel page pool. When not in
2062 * interrupt context and apply the current process NUMA policy.
2063 * Returns NULL when no page can be allocated.
2064 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08002065 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07002066 * 1) it's ok to take cpuset_sem (can WAIT), and
2067 * 2) allocating for current task (not interrupt).
2068 */
Al Virodd0fc662005-10-07 07:46:04 +01002069struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070{
Oleg Nesterov8d902742014-10-09 15:27:45 -07002071 struct mempolicy *pol = &default_policy;
Miao Xiec0ff7452010-05-24 14:32:08 -07002072 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002073 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074
Oleg Nesterov8d902742014-10-09 15:27:45 -07002075 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2076 pol = get_task_policy(current);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002077
Mel Gormancc9a6c82012-03-21 16:34:11 -07002078retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07002079 cpuset_mems_cookie = read_mems_allowed_begin();
Mel Gormancc9a6c82012-03-21 16:34:11 -07002080
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002081 /*
2082 * No reference counting needed for current->mempolicy
2083 * nor system default_policy
2084 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002085 if (pol->mode == MPOL_INTERLEAVE)
Miao Xiec0ff7452010-05-24 14:32:08 -07002086 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2087 else
2088 page = __alloc_pages_nodemask(gfp, order,
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002089 policy_zonelist(gfp, pol, numa_node_id()),
2090 policy_nodemask(gfp, pol));
Mel Gormancc9a6c82012-03-21 16:34:11 -07002091
Mel Gormand26914d2014-04-03 14:47:24 -07002092 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07002093 goto retry_cpuset;
2094
Miao Xiec0ff7452010-05-24 14:32:08 -07002095 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096}
2097EXPORT_SYMBOL(alloc_pages_current);
2098
Oleg Nesterovef0855d2013-09-11 14:20:14 -07002099int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2100{
2101 struct mempolicy *pol = mpol_dup(vma_policy(src));
2102
2103 if (IS_ERR(pol))
2104 return PTR_ERR(pol);
2105 dst->vm_policy = pol;
2106 return 0;
2107}
2108
Paul Jackson42253992006-01-08 01:01:59 -08002109/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002110 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08002111 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2112 * with the mems_allowed returned by cpuset_mems_allowed(). This
2113 * keeps mempolicies cpuset relative after its cpuset moves. See
2114 * further kernel/cpuset.c update_nodemask().
Miao Xie708c1bb2010-05-24 14:32:07 -07002115 *
2116 * current's mempolicy may be rebinded by the other task(the task that changes
2117 * cpuset's mems), so we needn't do rebind work for current task.
Paul Jackson42253992006-01-08 01:01:59 -08002118 */
Paul Jackson42253992006-01-08 01:01:59 -08002119
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002120/* Slow path of a mempolicy duplicate */
2121struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122{
2123 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2124
2125 if (!new)
2126 return ERR_PTR(-ENOMEM);
Miao Xie708c1bb2010-05-24 14:32:07 -07002127
2128 /* task's mempolicy is protected by alloc_lock */
2129 if (old == current->mempolicy) {
2130 task_lock(current);
2131 *new = *old;
2132 task_unlock(current);
2133 } else
2134 *new = *old;
2135
Paul Jackson42253992006-01-08 01:01:59 -08002136 if (current_cpuset_is_being_rebound()) {
2137 nodemask_t mems = cpuset_mems_allowed(current);
Miao Xie708c1bb2010-05-24 14:32:07 -07002138 if (new->flags & MPOL_F_REBINDING)
2139 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2140 else
2141 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
Paul Jackson42253992006-01-08 01:01:59 -08002142 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 return new;
2145}
2146
2147/* Slow path of a mempolicy comparison */
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002148bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149{
2150 if (!a || !b)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002151 return false;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002152 if (a->mode != b->mode)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002153 return false;
Bob Liu19800502010-05-24 14:32:01 -07002154 if (a->flags != b->flags)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002155 return false;
Bob Liu19800502010-05-24 14:32:01 -07002156 if (mpol_store_user_nodemask(a))
2157 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002158 return false;
Bob Liu19800502010-05-24 14:32:01 -07002159
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002160 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07002161 case MPOL_BIND:
2162 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 case MPOL_INTERLEAVE:
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002164 return !!nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 case MPOL_PREFERRED:
Namhyung Kim75719662011-03-22 16:33:02 -07002166 return a->v.preferred_node == b->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 default:
2168 BUG();
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002169 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 }
2171}
2172
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 * Shared memory backing store policy support.
2175 *
2176 * Remember policies even when nobody has shared memory mapped.
2177 * The policies are kept in Red-Black tree linked from the inode.
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002178 * They are protected by the sp->lock rwlock, which should be held
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 * for any accesses to the tree.
2180 */
2181
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002182/*
2183 * lookup first element intersecting start-end. Caller holds sp->lock for
2184 * reading or for writing
2185 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186static struct sp_node *
2187sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2188{
2189 struct rb_node *n = sp->root.rb_node;
2190
2191 while (n) {
2192 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2193
2194 if (start >= p->end)
2195 n = n->rb_right;
2196 else if (end <= p->start)
2197 n = n->rb_left;
2198 else
2199 break;
2200 }
2201 if (!n)
2202 return NULL;
2203 for (;;) {
2204 struct sp_node *w = NULL;
2205 struct rb_node *prev = rb_prev(n);
2206 if (!prev)
2207 break;
2208 w = rb_entry(prev, struct sp_node, nd);
2209 if (w->end <= start)
2210 break;
2211 n = prev;
2212 }
2213 return rb_entry(n, struct sp_node, nd);
2214}
2215
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002216/*
2217 * Insert a new shared policy into the list. Caller holds sp->lock for
2218 * writing.
2219 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2221{
2222 struct rb_node **p = &sp->root.rb_node;
2223 struct rb_node *parent = NULL;
2224 struct sp_node *nd;
2225
2226 while (*p) {
2227 parent = *p;
2228 nd = rb_entry(parent, struct sp_node, nd);
2229 if (new->start < nd->start)
2230 p = &(*p)->rb_left;
2231 else if (new->end > nd->end)
2232 p = &(*p)->rb_right;
2233 else
2234 BUG();
2235 }
2236 rb_link_node(&new->nd, parent, p);
2237 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07002238 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002239 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240}
2241
2242/* Find shared policy intersecting idx */
2243struct mempolicy *
2244mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2245{
2246 struct mempolicy *pol = NULL;
2247 struct sp_node *sn;
2248
2249 if (!sp->root.rb_node)
2250 return NULL;
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002251 read_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 sn = sp_lookup(sp, idx, idx+1);
2253 if (sn) {
2254 mpol_get(sn->policy);
2255 pol = sn->policy;
2256 }
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002257 read_unlock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 return pol;
2259}
2260
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002261static void sp_free(struct sp_node *n)
2262{
2263 mpol_put(n->policy);
2264 kmem_cache_free(sn_cache, n);
2265}
2266
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002267/**
2268 * mpol_misplaced - check whether current page node is valid in policy
2269 *
Fabian Frederickb46e14a2014-06-04 16:08:18 -07002270 * @page: page to be checked
2271 * @vma: vm area where page mapped
2272 * @addr: virtual address where page mapped
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002273 *
2274 * Lookup current policy node id for vma,addr and "compare to" page's
2275 * node id.
2276 *
2277 * Returns:
2278 * -1 - not misplaced, page is in the right node
2279 * node - node id where the page should be
2280 *
2281 * Policy determination "mimics" alloc_page_vma().
2282 * Called from fault path where we know the vma and faulting address.
2283 */
2284int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2285{
2286 struct mempolicy *pol;
2287 struct zone *zone;
2288 int curnid = page_to_nid(page);
2289 unsigned long pgoff;
Peter Zijlstra90572892013-10-07 11:29:20 +01002290 int thiscpu = raw_smp_processor_id();
2291 int thisnid = cpu_to_node(thiscpu);
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002292 int polnid = -1;
2293 int ret = -1;
2294
2295 BUG_ON(!vma);
2296
Oleg Nesterovdd6eecb2014-10-09 15:27:57 -07002297 pol = get_vma_policy(vma, addr);
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002298 if (!(pol->flags & MPOL_F_MOF))
2299 goto out;
2300
2301 switch (pol->mode) {
2302 case MPOL_INTERLEAVE:
2303 BUG_ON(addr >= vma->vm_end);
2304 BUG_ON(addr < vma->vm_start);
2305
2306 pgoff = vma->vm_pgoff;
2307 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2308 polnid = offset_il_node(pol, vma, pgoff);
2309 break;
2310
2311 case MPOL_PREFERRED:
2312 if (pol->flags & MPOL_F_LOCAL)
2313 polnid = numa_node_id();
2314 else
2315 polnid = pol->v.preferred_node;
2316 break;
2317
2318 case MPOL_BIND:
2319 /*
2320 * allows binding to multiple nodes.
2321 * use current page if in policy nodemask,
2322 * else select nearest allowed node, if any.
2323 * If no allowed nodes, use current [!misplaced].
2324 */
2325 if (node_isset(curnid, pol->v.nodes))
2326 goto out;
2327 (void)first_zones_zonelist(
2328 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2329 gfp_zone(GFP_HIGHUSER),
2330 &pol->v.nodes, &zone);
2331 polnid = zone->node;
2332 break;
2333
2334 default:
2335 BUG();
2336 }
Mel Gorman5606e382012-11-02 18:19:13 +00002337
2338 /* Migrate the page towards the node whose CPU is referencing it */
Mel Gormane42c8ff2012-11-12 09:17:07 +00002339 if (pol->flags & MPOL_F_MORON) {
Peter Zijlstra90572892013-10-07 11:29:20 +01002340 polnid = thisnid;
Mel Gorman5606e382012-11-02 18:19:13 +00002341
Rik van Riel10f39042014-01-27 17:03:44 -05002342 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
Rik van Rielde1c9ce2013-10-07 11:29:39 +01002343 goto out;
Mel Gormane42c8ff2012-11-12 09:17:07 +00002344 }
2345
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002346 if (curnid != polnid)
2347 ret = polnid;
2348out:
2349 mpol_cond_put(pol);
2350
2351 return ret;
2352}
2353
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2355{
Paul Mundt140d5a42007-07-15 23:38:16 -07002356 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 rb_erase(&n->nd, &sp->root);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002358 sp_free(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359}
2360
Mel Gorman42288fe2012-12-21 23:10:25 +00002361static void sp_node_init(struct sp_node *node, unsigned long start,
2362 unsigned long end, struct mempolicy *pol)
2363{
2364 node->start = start;
2365 node->end = end;
2366 node->policy = pol;
2367}
2368
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07002369static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2370 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371{
KOSAKI Motohiro869833f22012-10-08 16:29:16 -07002372 struct sp_node *n;
2373 struct mempolicy *newpol;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374
KOSAKI Motohiro869833f22012-10-08 16:29:16 -07002375 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 if (!n)
2377 return NULL;
KOSAKI Motohiro869833f22012-10-08 16:29:16 -07002378
2379 newpol = mpol_dup(pol);
2380 if (IS_ERR(newpol)) {
2381 kmem_cache_free(sn_cache, n);
2382 return NULL;
2383 }
2384 newpol->flags |= MPOL_F_SHARED;
Mel Gorman42288fe2012-12-21 23:10:25 +00002385 sp_node_init(n, start, end, newpol);
KOSAKI Motohiro869833f22012-10-08 16:29:16 -07002386
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387 return n;
2388}
2389
2390/* Replace a policy range. */
2391static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2392 unsigned long end, struct sp_node *new)
2393{
Mel Gormanb22d1272012-10-08 16:29:17 -07002394 struct sp_node *n;
Mel Gorman42288fe2012-12-21 23:10:25 +00002395 struct sp_node *n_new = NULL;
2396 struct mempolicy *mpol_new = NULL;
Mel Gormanb22d1272012-10-08 16:29:17 -07002397 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398
Mel Gorman42288fe2012-12-21 23:10:25 +00002399restart:
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002400 write_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 n = sp_lookup(sp, start, end);
2402 /* Take care of old policies in the same range. */
2403 while (n && n->start < end) {
2404 struct rb_node *next = rb_next(&n->nd);
2405 if (n->start >= start) {
2406 if (n->end <= end)
2407 sp_delete(sp, n);
2408 else
2409 n->start = end;
2410 } else {
2411 /* Old policy spanning whole new range. */
2412 if (n->end > end) {
Mel Gorman42288fe2012-12-21 23:10:25 +00002413 if (!n_new)
2414 goto alloc_new;
2415
2416 *mpol_new = *n->policy;
2417 atomic_set(&mpol_new->refcnt, 1);
KOSAKI Motohiro78806392013-03-08 12:43:29 -08002418 sp_node_init(n_new, end, n->end, mpol_new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 n->end = start;
Hillf Danton5ca39572013-03-08 12:43:28 -08002420 sp_insert(sp, n_new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002421 n_new = NULL;
2422 mpol_new = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423 break;
2424 } else
2425 n->end = start;
2426 }
2427 if (!next)
2428 break;
2429 n = rb_entry(next, struct sp_node, nd);
2430 }
2431 if (new)
2432 sp_insert(sp, new);
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002433 write_unlock(&sp->lock);
Mel Gorman42288fe2012-12-21 23:10:25 +00002434 ret = 0;
2435
2436err_out:
2437 if (mpol_new)
2438 mpol_put(mpol_new);
2439 if (n_new)
2440 kmem_cache_free(sn_cache, n_new);
2441
Mel Gormanb22d1272012-10-08 16:29:17 -07002442 return ret;
Mel Gorman42288fe2012-12-21 23:10:25 +00002443
2444alloc_new:
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002445 write_unlock(&sp->lock);
Mel Gorman42288fe2012-12-21 23:10:25 +00002446 ret = -ENOMEM;
2447 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2448 if (!n_new)
2449 goto err_out;
2450 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2451 if (!mpol_new)
2452 goto err_out;
2453 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454}
2455
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002456/**
2457 * mpol_shared_policy_init - initialize shared policy for inode
2458 * @sp: pointer to inode shared policy
2459 * @mpol: struct mempolicy to install
2460 *
2461 * Install non-NULL @mpol in inode's shared policy rb-tree.
2462 * On entry, the current task has a reference on a non-NULL @mpol.
2463 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002464 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002465 */
2466void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08002467{
Miao Xie58568d22009-06-16 15:31:49 -07002468 int ret;
2469
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002470 sp->root = RB_ROOT; /* empty tree == default mempolicy */
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002471 rwlock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08002472
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002473 if (mpol) {
2474 struct vm_area_struct pvma;
2475 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002476 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002477
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002478 if (!scratch)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002479 goto put_mpol;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002480 /* contextualize the tmpfs mount point mempolicy */
2481 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002482 if (IS_ERR(new))
Dan Carpenter0cae3452010-05-25 23:42:58 -07002483 goto free_scratch; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002484
2485 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002486 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002487 task_unlock(current);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002488 if (ret)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002489 goto put_new;
Robin Holt7339ff82006-01-14 13:20:48 -08002490
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002491 /* Create pseudo-vma that contains just the policy */
2492 memset(&pvma, 0, sizeof(struct vm_area_struct));
2493 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2494 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002495
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002496put_new:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002497 mpol_put(new); /* drop initial ref */
Dan Carpenter0cae3452010-05-25 23:42:58 -07002498free_scratch:
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002499 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002500put_mpol:
2501 mpol_put(mpol); /* drop our incoming ref on sb mpol */
Robin Holt7339ff82006-01-14 13:20:48 -08002502 }
2503}
2504
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505int mpol_set_shared_policy(struct shared_policy *info,
2506 struct vm_area_struct *vma, struct mempolicy *npol)
2507{
2508 int err;
2509 struct sp_node *new = NULL;
2510 unsigned long sz = vma_pages(vma);
2511
David Rientjes028fec42008-04-28 02:12:25 -07002512 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002514 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002515 npol ? npol->flags : -1,
David Rientjes00ef2d22013-02-22 16:35:36 -08002516 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517
2518 if (npol) {
2519 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2520 if (!new)
2521 return -ENOMEM;
2522 }
2523 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2524 if (err && new)
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002525 sp_free(new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 return err;
2527}
2528
2529/* Free a backing policy store on inode delete. */
2530void mpol_free_shared_policy(struct shared_policy *p)
2531{
2532 struct sp_node *n;
2533 struct rb_node *next;
2534
2535 if (!p->root.rb_node)
2536 return;
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002537 write_lock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538 next = rb_first(&p->root);
2539 while (next) {
2540 n = rb_entry(next, struct sp_node, nd);
2541 next = rb_next(&n->nd);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002542 sp_delete(p, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543 }
Nathan Zimmer4a8c7bb2016-01-14 15:18:36 -08002544 write_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545}
2546
Mel Gorman1a687c22012-11-22 11:16:36 +00002547#ifdef CONFIG_NUMA_BALANCING
Mel Gormanc2976632014-01-29 14:05:42 -08002548static int __initdata numabalancing_override;
Mel Gorman1a687c22012-11-22 11:16:36 +00002549
2550static void __init check_numabalancing_enable(void)
2551{
2552 bool numabalancing_default = false;
2553
2554 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2555 numabalancing_default = true;
2556
Mel Gormanc2976632014-01-29 14:05:42 -08002557 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2558 if (numabalancing_override)
2559 set_numabalancing_state(numabalancing_override == 1);
2560
Mel Gormanb0dc2b92015-05-14 15:17:09 -07002561 if (num_online_nodes() > 1 && !numabalancing_override) {
Joe Perches756a025f02016-03-17 14:19:47 -07002562 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
Mel Gormanc2976632014-01-29 14:05:42 -08002563 numabalancing_default ? "Enabling" : "Disabling");
Mel Gorman1a687c22012-11-22 11:16:36 +00002564 set_numabalancing_state(numabalancing_default);
2565 }
2566}
2567
2568static int __init setup_numabalancing(char *str)
2569{
2570 int ret = 0;
2571 if (!str)
2572 goto out;
Mel Gorman1a687c22012-11-22 11:16:36 +00002573
2574 if (!strcmp(str, "enable")) {
Mel Gormanc2976632014-01-29 14:05:42 -08002575 numabalancing_override = 1;
Mel Gorman1a687c22012-11-22 11:16:36 +00002576 ret = 1;
2577 } else if (!strcmp(str, "disable")) {
Mel Gormanc2976632014-01-29 14:05:42 -08002578 numabalancing_override = -1;
Mel Gorman1a687c22012-11-22 11:16:36 +00002579 ret = 1;
2580 }
2581out:
2582 if (!ret)
Andrew Morton4a404be2014-01-29 14:05:43 -08002583 pr_warn("Unable to parse numa_balancing=\n");
Mel Gorman1a687c22012-11-22 11:16:36 +00002584
2585 return ret;
2586}
2587__setup("numa_balancing=", setup_numabalancing);
2588#else
2589static inline void __init check_numabalancing_enable(void)
2590{
2591}
2592#endif /* CONFIG_NUMA_BALANCING */
2593
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594/* assumes fs == KERNEL_DS */
2595void __init numa_policy_init(void)
2596{
Paul Mundtb71636e2007-07-15 23:38:15 -07002597 nodemask_t interleave_nodes;
2598 unsigned long largest = 0;
2599 int nid, prefer = 0;
2600
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 policy_cache = kmem_cache_create("numa_policy",
2602 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002603 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604
2605 sn_cache = kmem_cache_create("shared_policy_node",
2606 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002607 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608
Mel Gorman5606e382012-11-02 18:19:13 +00002609 for_each_node(nid) {
2610 preferred_node_policy[nid] = (struct mempolicy) {
2611 .refcnt = ATOMIC_INIT(1),
2612 .mode = MPOL_PREFERRED,
2613 .flags = MPOL_F_MOF | MPOL_F_MORON,
2614 .v = { .preferred_node = nid, },
2615 };
2616 }
2617
Paul Mundtb71636e2007-07-15 23:38:15 -07002618 /*
2619 * Set interleaving policy for system init. Interleaving is only
2620 * enabled across suitably sized nodes (default is >= 16MB), or
2621 * fall back to the largest node if they're all smaller.
2622 */
2623 nodes_clear(interleave_nodes);
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002624 for_each_node_state(nid, N_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002625 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626
Paul Mundtb71636e2007-07-15 23:38:15 -07002627 /* Preserve the largest node */
2628 if (largest < total_pages) {
2629 largest = total_pages;
2630 prefer = nid;
2631 }
2632
2633 /* Interleave this node? */
2634 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2635 node_set(nid, interleave_nodes);
2636 }
2637
2638 /* All too small, use the largest */
2639 if (unlikely(nodes_empty(interleave_nodes)))
2640 node_set(prefer, interleave_nodes);
2641
David Rientjes028fec42008-04-28 02:12:25 -07002642 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -07002643 pr_err("%s: interleaving failed\n", __func__);
Mel Gorman1a687c22012-11-22 11:16:36 +00002644
2645 check_numabalancing_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646}
2647
Christoph Lameter8bccd852005-10-29 18:16:59 -07002648/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649void numa_default_policy(void)
2650{
David Rientjes028fec42008-04-28 02:12:25 -07002651 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652}
Paul Jackson68860ec2005-10-30 15:02:36 -08002653
Paul Jackson42253992006-01-08 01:01:59 -08002654/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002655 * Parse and format mempolicy from/to strings
2656 */
2657
2658/*
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002659 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002660 */
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002661static const char * const policy_modes[] =
2662{
2663 [MPOL_DEFAULT] = "default",
2664 [MPOL_PREFERRED] = "prefer",
2665 [MPOL_BIND] = "bind",
2666 [MPOL_INTERLEAVE] = "interleave",
Lee Schermerhornd3a71032012-10-25 14:16:29 +02002667 [MPOL_LOCAL] = "local",
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002668};
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002669
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002670
2671#ifdef CONFIG_TMPFS
2672/**
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002673 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002674 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002675 * @mpol: pointer to struct mempolicy pointer, returned on success.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002676 *
2677 * Format of input:
2678 * <mode>[=<flags>][:<nodelist>]
2679 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002680 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002681 */
Hugh Dickinsa7a88b22013-01-02 02:04:23 -08002682int mpol_parse_str(char *str, struct mempolicy **mpol)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002683{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002684 struct mempolicy *new = NULL;
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002685 unsigned short mode;
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002686 unsigned short mode_flags;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002687 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002688 char *nodelist = strchr(str, ':');
2689 char *flags = strchr(str, '=');
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002690 int err = 1;
2691
2692 if (nodelist) {
2693 /* NUL-terminate mode or flags string */
2694 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002695 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002696 goto out;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002697 if (!nodes_subset(nodes, node_states[N_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002698 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002699 } else
2700 nodes_clear(nodes);
2701
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002702 if (flags)
2703 *flags++ = '\0'; /* terminate mode string */
2704
Peter Zijlstra479e2802012-10-25 14:16:28 +02002705 for (mode = 0; mode < MPOL_MAX; mode++) {
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002706 if (!strcmp(str, policy_modes[mode])) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002707 break;
2708 }
2709 }
Mel Gormana7200942012-11-16 09:37:58 +00002710 if (mode >= MPOL_MAX)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002711 goto out;
2712
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002713 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002714 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002715 /*
2716 * Insist on a nodelist of one node only
2717 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002718 if (nodelist) {
2719 char *rest = nodelist;
2720 while (isdigit(*rest))
2721 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002722 if (*rest)
2723 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002724 }
2725 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002726 case MPOL_INTERLEAVE:
2727 /*
2728 * Default to online nodes with memory if no nodelist
2729 */
2730 if (!nodelist)
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002731 nodes = node_states[N_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002732 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002733 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002734 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002735 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002736 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002737 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002738 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002739 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002740 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002741 case MPOL_DEFAULT:
2742 /*
2743 * Insist on a empty nodelist
2744 */
2745 if (!nodelist)
2746 err = 0;
2747 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002748 case MPOL_BIND:
2749 /*
2750 * Insist on a nodelist
2751 */
2752 if (!nodelist)
2753 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002754 }
2755
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002756 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002757 if (flags) {
2758 /*
2759 * Currently, we only support two mutually exclusive
2760 * mode flags.
2761 */
2762 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002763 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002764 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002765 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002766 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002767 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002768 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002769
2770 new = mpol_new(mode, mode_flags, &nodes);
2771 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002772 goto out;
2773
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002774 /*
2775 * Save nodes for mpol_to_str() to show the tmpfs mount options
2776 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2777 */
2778 if (mode != MPOL_PREFERRED)
2779 new->v.nodes = nodes;
2780 else if (nodelist)
2781 new->v.preferred_node = first_node(nodes);
2782 else
2783 new->flags |= MPOL_F_LOCAL;
2784
2785 /*
2786 * Save nodes for contextualization: this will be used to "clone"
2787 * the mempolicy in a specific context [cpuset] at a later time.
2788 */
2789 new->w.user_nodemask = nodes;
2790
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002791 err = 0;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002792
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002793out:
2794 /* Restore string for error message */
2795 if (nodelist)
2796 *--nodelist = ':';
2797 if (flags)
2798 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002799 if (!err)
2800 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002801 return err;
2802}
2803#endif /* CONFIG_TMPFS */
2804
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002805/**
2806 * mpol_to_str - format a mempolicy structure for printing
2807 * @buffer: to contain formatted mempolicy string
2808 * @maxlen: length of @buffer
2809 * @pol: pointer to mempolicy to be formatted
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002810 *
David Rientjes948927e2013-11-12 15:07:28 -08002811 * Convert @pol into a string. If @buffer is too short, truncate the string.
2812 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2813 * longest flag, "relative", and to display at least a few node ids.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002814 */
David Rientjes948927e2013-11-12 15:07:28 -08002815void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002816{
2817 char *p = buffer;
David Rientjes948927e2013-11-12 15:07:28 -08002818 nodemask_t nodes = NODE_MASK_NONE;
2819 unsigned short mode = MPOL_DEFAULT;
2820 unsigned short flags = 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002821
David Rientjes8790c712014-01-30 15:46:08 -08002822 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002823 mode = pol->mode;
David Rientjes948927e2013-11-12 15:07:28 -08002824 flags = pol->flags;
2825 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002826
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002827 switch (mode) {
2828 case MPOL_DEFAULT:
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002829 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002830 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002831 if (flags & MPOL_F_LOCAL)
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002832 mode = MPOL_LOCAL;
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002833 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002834 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002835 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002836 case MPOL_BIND:
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002837 case MPOL_INTERLEAVE:
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002838 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002839 break;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002840 default:
David Rientjes948927e2013-11-12 15:07:28 -08002841 WARN_ON_ONCE(1);
2842 snprintf(p, maxlen, "unknown");
2843 return;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002844 }
2845
David Rientjesb7a9f422013-11-21 14:32:06 -08002846 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002847
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002848 if (flags & MPOL_MODE_FLAGS) {
David Rientjes948927e2013-11-12 15:07:28 -08002849 p += snprintf(p, buffer + maxlen - p, "=");
David Rientjesf5b087b2008-04-28 02:12:27 -07002850
Lee Schermerhorn22919902008-04-28 02:13:22 -07002851 /*
2852 * Currently, the only defined flags are mutually exclusive
2853 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002854 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002855 p += snprintf(p, buffer + maxlen - p, "static");
2856 else if (flags & MPOL_F_RELATIVE_NODES)
2857 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002858 }
2859
Tejun Heo9e763e02015-02-13 14:38:02 -08002860 if (!nodes_empty(nodes))
2861 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2862 nodemask_pr_args(&nodes));
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002863}