blob: 74310017296ee02317b79a2c28b25ef303fc7720 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
Christoph Lameter8bccd852005-10-29 18:16:59 -07005 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 * Subject to the GNU Public License, version 2.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
Christoph Lameter8bccd852005-10-29 18:16:59 -070021 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070022 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
Christoph Lameter8bccd852005-10-29 18:16:59 -070024 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070028 * preferred Try a specific node first before normal fallback.
David Rientjes00ef2d22013-02-22 16:35:36 -080029 * As a special case NUMA_NO_NODE here means do the allocation
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
Christoph Lameter8bccd852005-10-29 18:16:59 -070033 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070034 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
Linus Torvalds1da177e2005-04-16 15:20:36 -070066*/
67
68#include <linux/mempolicy.h>
69#include <linux/mm.h>
70#include <linux/highmem.h>
71#include <linux/hugetlb.h>
72#include <linux/kernel.h>
73#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074#include <linux/nodemask.h>
75#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070076#include <linux/slab.h>
77#include <linux/string.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040078#include <linux/export.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070079#include <linux/nsproxy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070080#include <linux/interrupt.h>
81#include <linux/init.h>
82#include <linux/compat.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080083#include <linux/swap.h>
Christoph Lameter1a75a6c2006-01-08 01:01:02 -080084#include <linux/seq_file.h>
85#include <linux/proc_fs.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080086#include <linux/migrate.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080087#include <linux/ksm.h>
Christoph Lameter95a402c2006-06-23 02:03:53 -070088#include <linux/rmap.h>
David Quigley86c3a762006-06-23 02:04:02 -070089#include <linux/security.h>
Adrian Bunkdbcb0f12007-10-16 01:26:26 -070090#include <linux/syscalls.h>
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -070091#include <linux/ctype.h>
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -080092#include <linux/mm_inline.h>
Lee Schermerhornb24f53a2012-10-25 14:16:32 +020093#include <linux/mmu_notifier.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080094
Linus Torvalds1da177e2005-04-16 15:20:36 -070095#include <asm/tlbflush.h>
96#include <asm/uaccess.h>
Michal Hocko778d3b02011-07-26 16:08:30 -070097#include <linux/random.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
Nick Piggin62695a82008-10-18 20:26:09 -070099#include "internal.h"
100
Christoph Lameter38e35862006-01-08 01:01:01 -0800101/* Internal flags */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800102#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
Christoph Lameter38e35862006-01-08 01:01:01 -0800103#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800104
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800105static struct kmem_cache *policy_cache;
106static struct kmem_cache *sn_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108/* Highest zone. An specific allocation for a zone below that is not
109 policied. */
Christoph Lameter62672762007-02-10 01:43:07 -0800110enum zone_type policy_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700112/*
113 * run-time system-wide default policy => local allocation
114 */
H Hartley Sweetene754d792011-10-31 17:09:23 -0700115static struct mempolicy default_policy = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 .refcnt = ATOMIC_INIT(1), /* never free it */
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700117 .mode = MPOL_PREFERRED,
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700118 .flags = MPOL_F_LOCAL,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119};
120
Mel Gorman5606e382012-11-02 18:19:13 +0000121static struct mempolicy preferred_node_policy[MAX_NUMNODES];
122
123static struct mempolicy *get_task_policy(struct task_struct *p)
124{
125 struct mempolicy *pol = p->mempolicy;
126 int node;
127
128 if (!pol) {
129 node = numa_node_id();
David Rientjes00ef2d22013-02-22 16:35:36 -0800130 if (node != NUMA_NO_NODE)
Mel Gorman5606e382012-11-02 18:19:13 +0000131 pol = &preferred_node_policy[node];
132
133 /* preferred_node_policy is not initialised early in boot */
134 if (!pol->mode)
135 pol = NULL;
136 }
137
138 return pol;
139}
140
David Rientjes37012942008-04-28 02:12:33 -0700141static const struct mempolicy_operations {
142 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
Miao Xie708c1bb2010-05-24 14:32:07 -0700143 /*
144 * If read-side task has no lock to protect task->mempolicy, write-side
145 * task will rebind the task->mempolicy by two step. The first step is
146 * setting all the newly nodes, and the second step is cleaning all the
147 * disallowed nodes. In this way, we can avoid finding no node to alloc
148 * page.
149 * If we have a lock to protect task->mempolicy in read-side, we do
150 * rebind directly.
151 *
152 * step:
153 * MPOL_REBIND_ONCE - do rebind work at once
154 * MPOL_REBIND_STEP1 - set all the newly nodes
155 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
156 */
157 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
158 enum mpol_rebind_step step);
David Rientjes37012942008-04-28 02:12:33 -0700159} mpol_ops[MPOL_MAX];
160
Mel Gorman19770b32008-04-28 02:12:18 -0700161/* Check that the nodemask contains at least one populated zone */
David Rientjes37012942008-04-28 02:12:33 -0700162static int is_valid_nodemask(const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163{
Lai Jiangshand3eb1572013-02-22 16:33:22 -0800164 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165}
166
David Rientjesf5b087b2008-04-28 02:12:27 -0700167static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
168{
Bob Liu6d556292010-05-24 14:31:59 -0700169 return pol->flags & MPOL_MODE_FLAGS;
David Rientjes4c50bc02008-04-28 02:12:30 -0700170}
171
172static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
173 const nodemask_t *rel)
174{
175 nodemask_t tmp;
176 nodes_fold(tmp, *orig, nodes_weight(*rel));
177 nodes_onto(*ret, tmp, *rel);
David Rientjesf5b087b2008-04-28 02:12:27 -0700178}
179
David Rientjes37012942008-04-28 02:12:33 -0700180static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
181{
182 if (nodes_empty(*nodes))
183 return -EINVAL;
184 pol->v.nodes = *nodes;
185 return 0;
186}
187
188static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
189{
190 if (!nodes)
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700191 pol->flags |= MPOL_F_LOCAL; /* local allocation */
David Rientjes37012942008-04-28 02:12:33 -0700192 else if (nodes_empty(*nodes))
193 return -EINVAL; /* no allowed nodes */
194 else
195 pol->v.preferred_node = first_node(*nodes);
196 return 0;
197}
198
199static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
200{
201 if (!is_valid_nodemask(nodes))
202 return -EINVAL;
203 pol->v.nodes = *nodes;
204 return 0;
205}
206
Miao Xie58568d22009-06-16 15:31:49 -0700207/*
208 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
209 * any, for the new policy. mpol_new() has already validated the nodes
210 * parameter with respect to the policy mode and flags. But, we need to
211 * handle an empty nodemask with MPOL_PREFERRED here.
212 *
213 * Must be called holding task's alloc_lock to protect task's mems_allowed
214 * and mempolicy. May also be called holding the mmap_semaphore for write.
215 */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700216static int mpol_set_nodemask(struct mempolicy *pol,
217 const nodemask_t *nodes, struct nodemask_scratch *nsc)
Miao Xie58568d22009-06-16 15:31:49 -0700218{
Miao Xie58568d22009-06-16 15:31:49 -0700219 int ret;
220
221 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
222 if (pol == NULL)
223 return 0;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -0800224 /* Check N_MEMORY */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700225 nodes_and(nsc->mask1,
Lai Jiangshan01f13bd2012-12-12 13:51:33 -0800226 cpuset_current_mems_allowed, node_states[N_MEMORY]);
Miao Xie58568d22009-06-16 15:31:49 -0700227
228 VM_BUG_ON(!nodes);
229 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
230 nodes = NULL; /* explicit local allocation */
231 else {
232 if (pol->flags & MPOL_F_RELATIVE_NODES)
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700233 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
Miao Xie58568d22009-06-16 15:31:49 -0700234 else
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700235 nodes_and(nsc->mask2, *nodes, nsc->mask1);
236
Miao Xie58568d22009-06-16 15:31:49 -0700237 if (mpol_store_user_nodemask(pol))
238 pol->w.user_nodemask = *nodes;
239 else
240 pol->w.cpuset_mems_allowed =
241 cpuset_current_mems_allowed;
242 }
243
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700244 if (nodes)
245 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
246 else
247 ret = mpol_ops[pol->mode].create(pol, NULL);
Miao Xie58568d22009-06-16 15:31:49 -0700248 return ret;
249}
250
251/*
252 * This function just creates a new policy, does some check and simple
253 * initialization. You must invoke mpol_set_nodemask() to set nodes.
254 */
David Rientjes028fec42008-04-28 02:12:25 -0700255static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
256 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257{
258 struct mempolicy *policy;
259
David Rientjes028fec42008-04-28 02:12:25 -0700260 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
David Rientjes00ef2d22013-02-22 16:35:36 -0800261 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
Paul Mundt140d5a42007-07-15 23:38:16 -0700262
David Rientjes3e1f0642008-04-28 02:12:34 -0700263 if (mode == MPOL_DEFAULT) {
264 if (nodes && !nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700265 return ERR_PTR(-EINVAL);
Lee Schermerhornd3a71032012-10-25 14:16:29 +0200266 return NULL;
David Rientjes37012942008-04-28 02:12:33 -0700267 }
David Rientjes3e1f0642008-04-28 02:12:34 -0700268 VM_BUG_ON(!nodes);
269
270 /*
271 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
272 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
273 * All other modes require a valid pointer to a non-empty nodemask.
274 */
275 if (mode == MPOL_PREFERRED) {
276 if (nodes_empty(*nodes)) {
277 if (((flags & MPOL_F_STATIC_NODES) ||
278 (flags & MPOL_F_RELATIVE_NODES)))
279 return ERR_PTR(-EINVAL);
David Rientjes3e1f0642008-04-28 02:12:34 -0700280 }
Peter Zijlstra479e2802012-10-25 14:16:28 +0200281 } else if (mode == MPOL_LOCAL) {
282 if (!nodes_empty(*nodes))
283 return ERR_PTR(-EINVAL);
284 mode = MPOL_PREFERRED;
David Rientjes3e1f0642008-04-28 02:12:34 -0700285 } else if (nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
288 if (!policy)
289 return ERR_PTR(-ENOMEM);
290 atomic_set(&policy->refcnt, 1);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700291 policy->mode = mode;
David Rientjes3e1f0642008-04-28 02:12:34 -0700292 policy->flags = flags;
David Rientjesf5b087b2008-04-28 02:12:27 -0700293
David Rientjes37012942008-04-28 02:12:33 -0700294 return policy;
295}
296
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700297/* Slow path of a mpol destructor. */
298void __mpol_put(struct mempolicy *p)
299{
300 if (!atomic_dec_and_test(&p->refcnt))
301 return;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700302 kmem_cache_free(policy_cache, p);
303}
304
Miao Xie708c1bb2010-05-24 14:32:07 -0700305static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
306 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700307{
308}
309
Miao Xie708c1bb2010-05-24 14:32:07 -0700310/*
311 * step:
312 * MPOL_REBIND_ONCE - do rebind work at once
313 * MPOL_REBIND_STEP1 - set all the newly nodes
314 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
315 */
316static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
317 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700318{
319 nodemask_t tmp;
320
321 if (pol->flags & MPOL_F_STATIC_NODES)
322 nodes_and(tmp, pol->w.user_nodemask, *nodes);
323 else if (pol->flags & MPOL_F_RELATIVE_NODES)
324 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
325 else {
Miao Xie708c1bb2010-05-24 14:32:07 -0700326 /*
327 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
328 * result
329 */
330 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
331 nodes_remap(tmp, pol->v.nodes,
332 pol->w.cpuset_mems_allowed, *nodes);
333 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
334 } else if (step == MPOL_REBIND_STEP2) {
335 tmp = pol->w.cpuset_mems_allowed;
336 pol->w.cpuset_mems_allowed = *nodes;
337 } else
338 BUG();
David Rientjes37012942008-04-28 02:12:33 -0700339 }
340
Miao Xie708c1bb2010-05-24 14:32:07 -0700341 if (nodes_empty(tmp))
342 tmp = *nodes;
343
344 if (step == MPOL_REBIND_STEP1)
345 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
346 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
347 pol->v.nodes = tmp;
348 else
349 BUG();
350
David Rientjes37012942008-04-28 02:12:33 -0700351 if (!node_isset(current->il_next, tmp)) {
352 current->il_next = next_node(current->il_next, tmp);
353 if (current->il_next >= MAX_NUMNODES)
354 current->il_next = first_node(tmp);
355 if (current->il_next >= MAX_NUMNODES)
356 current->il_next = numa_node_id();
357 }
358}
359
360static void mpol_rebind_preferred(struct mempolicy *pol,
Miao Xie708c1bb2010-05-24 14:32:07 -0700361 const nodemask_t *nodes,
362 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700363{
364 nodemask_t tmp;
365
David Rientjes37012942008-04-28 02:12:33 -0700366 if (pol->flags & MPOL_F_STATIC_NODES) {
367 int node = first_node(pol->w.user_nodemask);
368
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700369 if (node_isset(node, *nodes)) {
David Rientjes37012942008-04-28 02:12:33 -0700370 pol->v.preferred_node = node;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700371 pol->flags &= ~MPOL_F_LOCAL;
372 } else
373 pol->flags |= MPOL_F_LOCAL;
David Rientjes37012942008-04-28 02:12:33 -0700374 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
375 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
376 pol->v.preferred_node = first_node(tmp);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700377 } else if (!(pol->flags & MPOL_F_LOCAL)) {
David Rientjes37012942008-04-28 02:12:33 -0700378 pol->v.preferred_node = node_remap(pol->v.preferred_node,
379 pol->w.cpuset_mems_allowed,
380 *nodes);
381 pol->w.cpuset_mems_allowed = *nodes;
382 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383}
384
Miao Xie708c1bb2010-05-24 14:32:07 -0700385/*
386 * mpol_rebind_policy - Migrate a policy to a different set of nodes
387 *
388 * If read-side task has no lock to protect task->mempolicy, write-side
389 * task will rebind the task->mempolicy by two step. The first step is
390 * setting all the newly nodes, and the second step is cleaning all the
391 * disallowed nodes. In this way, we can avoid finding no node to alloc
392 * page.
393 * If we have a lock to protect task->mempolicy in read-side, we do
394 * rebind directly.
395 *
396 * step:
397 * MPOL_REBIND_ONCE - do rebind work at once
398 * MPOL_REBIND_STEP1 - set all the newly nodes
399 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
400 */
401static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
402 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700403{
David Rientjes1d0d2682008-04-28 02:12:32 -0700404 if (!pol)
405 return;
Wang Sheng-Hui89c522c2012-05-29 15:06:16 -0700406 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
David Rientjes1d0d2682008-04-28 02:12:32 -0700407 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
408 return;
Miao Xie708c1bb2010-05-24 14:32:07 -0700409
410 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
411 return;
412
413 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
414 BUG();
415
416 if (step == MPOL_REBIND_STEP1)
417 pol->flags |= MPOL_F_REBINDING;
418 else if (step == MPOL_REBIND_STEP2)
419 pol->flags &= ~MPOL_F_REBINDING;
420 else if (step >= MPOL_REBIND_NSTEP)
421 BUG();
422
423 mpol_ops[pol->mode].rebind(pol, newmask, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700424}
425
426/*
427 * Wrapper for mpol_rebind_policy() that just requires task
428 * pointer, and updates task mempolicy.
Miao Xie58568d22009-06-16 15:31:49 -0700429 *
430 * Called with task's alloc_lock held.
David Rientjes1d0d2682008-04-28 02:12:32 -0700431 */
432
Miao Xie708c1bb2010-05-24 14:32:07 -0700433void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
434 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700435{
Miao Xie708c1bb2010-05-24 14:32:07 -0700436 mpol_rebind_policy(tsk->mempolicy, new, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700437}
438
439/*
440 * Rebind each vma in mm to new nodemask.
441 *
442 * Call holding a reference to mm. Takes mm->mmap_sem during call.
443 */
444
445void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
446{
447 struct vm_area_struct *vma;
448
449 down_write(&mm->mmap_sem);
450 for (vma = mm->mmap; vma; vma = vma->vm_next)
Miao Xie708c1bb2010-05-24 14:32:07 -0700451 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
David Rientjes1d0d2682008-04-28 02:12:32 -0700452 up_write(&mm->mmap_sem);
453}
454
David Rientjes37012942008-04-28 02:12:33 -0700455static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
456 [MPOL_DEFAULT] = {
457 .rebind = mpol_rebind_default,
458 },
459 [MPOL_INTERLEAVE] = {
460 .create = mpol_new_interleave,
461 .rebind = mpol_rebind_nodemask,
462 },
463 [MPOL_PREFERRED] = {
464 .create = mpol_new_preferred,
465 .rebind = mpol_rebind_preferred,
466 },
467 [MPOL_BIND] = {
468 .create = mpol_new_bind,
469 .rebind = mpol_rebind_nodemask,
470 },
471};
472
Christoph Lameterfc301282006-01-18 17:42:29 -0800473static void migrate_page_add(struct page *page, struct list_head *pagelist,
474 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800475
Christoph Lameter38e35862006-01-08 01:01:01 -0800476/* Scan through pages checking if pages follow certain conditions. */
Nick Pigginb5810032005-10-29 18:16:12 -0700477static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800478 unsigned long addr, unsigned long end,
479 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800480 void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481{
Hugh Dickins91612e02005-06-21 17:15:07 -0700482 pte_t *orig_pte;
483 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700484 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700485
Hugh Dickins705e87c2005-10-29 18:16:27 -0700486 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700487 do {
Linus Torvalds6aab3412005-11-28 14:34:23 -0800488 struct page *page;
Andy Whitcroft25ba77c2006-12-06 20:33:03 -0800489 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700490
491 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800493 page = vm_normal_page(vma, addr, *pte);
494 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800496 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800497 * vm_normal_page() filters out zero pages, but there might
498 * still be PageReserved pages to skip, perhaps in a VDSO.
Nick Piggin053837f2006-01-18 17:42:27 -0800499 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800500 if (PageReserved(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800501 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800502 nid = page_to_nid(page);
Christoph Lameter38e35862006-01-08 01:01:01 -0800503 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
504 continue;
505
Stephen Wilsonb1f72d12011-05-24 17:12:43 -0700506 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameterfc301282006-01-18 17:42:29 -0800507 migrate_page_add(page, private, flags);
Christoph Lameter38e35862006-01-08 01:01:01 -0800508 else
509 break;
Hugh Dickins91612e02005-06-21 17:15:07 -0700510 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickins705e87c2005-10-29 18:16:27 -0700511 pte_unmap_unlock(orig_pte, ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700512 return addr != end;
513}
514
Nick Pigginb5810032005-10-29 18:16:12 -0700515static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800516 unsigned long addr, unsigned long end,
517 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800518 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700519{
520 pmd_t *pmd;
521 unsigned long next;
522
523 pmd = pmd_offset(pud, addr);
524 do {
525 next = pmd_addr_end(addr, end);
Kirill A. Shutemove1803772012-12-12 13:50:59 -0800526 split_huge_page_pmd(vma, addr, pmd);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -0700527 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
Hugh Dickins91612e02005-06-21 17:15:07 -0700528 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800529 if (check_pte_range(vma, pmd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800530 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700531 return -EIO;
532 } while (pmd++, addr = next, addr != end);
533 return 0;
534}
535
Nick Pigginb5810032005-10-29 18:16:12 -0700536static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800537 unsigned long addr, unsigned long end,
538 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800539 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700540{
541 pud_t *pud;
542 unsigned long next;
543
544 pud = pud_offset(pgd, addr);
545 do {
546 next = pud_addr_end(addr, end);
547 if (pud_none_or_clear_bad(pud))
548 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800549 if (check_pmd_range(vma, pud, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800550 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700551 return -EIO;
552 } while (pud++, addr = next, addr != end);
553 return 0;
554}
555
Nick Pigginb5810032005-10-29 18:16:12 -0700556static inline int check_pgd_range(struct vm_area_struct *vma,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800557 unsigned long addr, unsigned long end,
558 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800559 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700560{
561 pgd_t *pgd;
562 unsigned long next;
563
Nick Pigginb5810032005-10-29 18:16:12 -0700564 pgd = pgd_offset(vma->vm_mm, addr);
Hugh Dickins91612e02005-06-21 17:15:07 -0700565 do {
566 next = pgd_addr_end(addr, end);
567 if (pgd_none_or_clear_bad(pgd))
568 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800569 if (check_pud_range(vma, pgd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800570 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700571 return -EIO;
572 } while (pgd++, addr = next, addr != end);
573 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574}
575
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200576#ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
577/*
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200578 * This is used to mark a range of virtual addresses to be inaccessible.
579 * These are later cleared by a NUMA hinting fault. Depending on these
580 * faults, pages may be migrated for better NUMA placement.
581 *
582 * This is assuming that NUMA faults are handled using PROT_NONE. If
583 * an architecture makes a different choice, it will need further
584 * changes to the core.
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200585 */
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200586unsigned long change_prot_numa(struct vm_area_struct *vma,
587 unsigned long addr, unsigned long end)
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200588{
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200589 int nr_updated;
590 BUILD_BUG_ON(_PAGE_NUMA != _PAGE_PROTNONE);
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200591
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200592 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
Mel Gorman03c5a6e2012-11-02 14:52:48 +0000593 if (nr_updated)
594 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200595
Mel Gorman4b10e7d2012-10-25 14:16:32 +0200596 return nr_updated;
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200597}
598#else
599static unsigned long change_prot_numa(struct vm_area_struct *vma,
600 unsigned long addr, unsigned long end)
601{
602 return 0;
603}
604#endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
605
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800606/*
607 * Check if all pages in a range are on a set of nodes.
608 * If pagelist != NULL then isolate pages from the LRU and
609 * put them on the pagelist.
610 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611static struct vm_area_struct *
612check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Christoph Lameter38e35862006-01-08 01:01:01 -0800613 const nodemask_t *nodes, unsigned long flags, void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614{
615 int err;
616 struct vm_area_struct *first, *vma, *prev;
617
Nick Piggin053837f2006-01-18 17:42:27 -0800618
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 first = find_vma(mm, start);
620 if (!first)
621 return ERR_PTR(-EFAULT);
622 prev = NULL;
623 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200624 unsigned long endvma = vma->vm_end;
625
626 if (endvma > end)
627 endvma = end;
628 if (vma->vm_start > start)
629 start = vma->vm_start;
630
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800631 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
632 if (!vma->vm_next && vma->vm_end < end)
633 return ERR_PTR(-EFAULT);
634 if (prev && prev->vm_end < vma->vm_start)
635 return ERR_PTR(-EFAULT);
636 }
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800637
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200638 if (is_vm_hugetlb_page(vma))
639 goto next;
640
641 if (flags & MPOL_MF_LAZY) {
642 change_prot_numa(vma, start, endvma);
643 goto next;
644 }
645
646 if ((flags & MPOL_MF_STRICT) ||
647 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
648 vma_migratable(vma))) {
649
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800650 err = check_pgd_range(vma, start, endvma, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800651 flags, private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 if (err) {
653 first = ERR_PTR(err);
654 break;
655 }
656 }
Lee Schermerhornb24f53a2012-10-25 14:16:32 +0200657next:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 prev = vma;
659 }
660 return first;
661}
662
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700663/*
664 * Apply policy to a single VMA
665 * This must be called with the mmap_sem held for writing.
666 */
667static int vma_replace_policy(struct vm_area_struct *vma,
668 struct mempolicy *pol)
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700669{
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700670 int err;
671 struct mempolicy *old;
672 struct mempolicy *new;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700673
674 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
675 vma->vm_start, vma->vm_end, vma->vm_pgoff,
676 vma->vm_ops, vma->vm_file,
677 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
678
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700679 new = mpol_dup(pol);
680 if (IS_ERR(new))
681 return PTR_ERR(new);
682
683 if (vma->vm_ops && vma->vm_ops->set_policy) {
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700684 err = vma->vm_ops->set_policy(vma, new);
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700685 if (err)
686 goto err_out;
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700687 }
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700688
689 old = vma->vm_policy;
690 vma->vm_policy = new; /* protected by mmap_sem */
691 mpol_put(old);
692
693 return 0;
694 err_out:
695 mpol_put(new);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700696 return err;
697}
698
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800700static int mbind_range(struct mm_struct *mm, unsigned long start,
701 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702{
703 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800704 struct vm_area_struct *prev;
705 struct vm_area_struct *vma;
706 int err = 0;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800707 pgoff_t pgoff;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800708 unsigned long vmstart;
709 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
Linus Torvalds097d5912012-03-06 18:23:36 -0800711 vma = find_vma(mm, start);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800712 if (!vma || vma->vm_start > start)
713 return -EFAULT;
714
Linus Torvalds097d5912012-03-06 18:23:36 -0800715 prev = vma->vm_prev;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800716 if (start > vma->vm_start)
717 prev = vma;
718
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800719 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800721 vmstart = max(start, vma->vm_start);
722 vmend = min(end, vma->vm_end);
723
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800724 if (mpol_equal(vma_policy(vma), new_pol))
725 continue;
726
727 pgoff = vma->vm_pgoff +
728 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800729 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800730 vma->anon_vma, vma->vm_file, pgoff,
Caspar Zhang8aacc9f2011-09-14 16:20:58 -0700731 new_pol);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800732 if (prev) {
733 vma = prev;
734 next = vma->vm_next;
735 continue;
736 }
737 if (vma->vm_start != vmstart) {
738 err = split_vma(vma->vm_mm, vma, vmstart, 1);
739 if (err)
740 goto out;
741 }
742 if (vma->vm_end != vmend) {
743 err = split_vma(vma->vm_mm, vma, vmend, 0);
744 if (err)
745 goto out;
746 }
KOSAKI Motohiro869833f2012-10-08 16:29:16 -0700747 err = vma_replace_policy(vma, new_pol);
KOSAKI Motohiro8d34694c12012-10-08 16:29:14 -0700748 if (err)
749 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800751
752 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 return err;
754}
755
Paul Jacksonc61afb12006-03-24 03:16:08 -0800756/*
757 * Update task->flags PF_MEMPOLICY bit: set iff non-default
758 * mempolicy. Allows more rapid checking of this (combined perhaps
759 * with other PF_* flag bits) on memory allocation hot code paths.
760 *
761 * If called from outside this file, the task 'p' should -only- be
762 * a newly forked child not yet visible on the task list, because
763 * manipulating the task flags of a visible task is not safe.
764 *
765 * The above limitation is why this routine has the funny name
766 * mpol_fix_fork_child_flag().
767 *
768 * It is also safe to call this with a task pointer of current,
769 * which the static wrapper mpol_set_task_struct_flag() does,
770 * for use within this file.
771 */
772
773void mpol_fix_fork_child_flag(struct task_struct *p)
774{
775 if (p->mempolicy)
776 p->flags |= PF_MEMPOLICY;
777 else
778 p->flags &= ~PF_MEMPOLICY;
779}
780
781static void mpol_set_task_struct_flag(void)
782{
783 mpol_fix_fork_child_flag(current);
784}
785
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700787static long do_set_mempolicy(unsigned short mode, unsigned short flags,
788 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789{
Miao Xie58568d22009-06-16 15:31:49 -0700790 struct mempolicy *new, *old;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700791 struct mm_struct *mm = current->mm;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700792 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700793 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700795 if (!scratch)
796 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700797
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700798 new = mpol_new(mode, flags, nodes);
799 if (IS_ERR(new)) {
800 ret = PTR_ERR(new);
801 goto out;
802 }
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700803 /*
804 * prevent changing our mempolicy while show_numa_maps()
805 * is using it.
806 * Note: do_set_mempolicy() can be called at init time
807 * with no 'mm'.
808 */
809 if (mm)
810 down_write(&mm->mmap_sem);
Miao Xie58568d22009-06-16 15:31:49 -0700811 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700812 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700813 if (ret) {
814 task_unlock(current);
815 if (mm)
816 up_write(&mm->mmap_sem);
817 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700818 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700819 }
820 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 current->mempolicy = new;
Paul Jacksonc61afb12006-03-24 03:16:08 -0800822 mpol_set_task_struct_flag();
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700823 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700824 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700825 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700826 task_unlock(current);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700827 if (mm)
828 up_write(&mm->mmap_sem);
829
Miao Xie58568d22009-06-16 15:31:49 -0700830 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700831 ret = 0;
832out:
833 NODEMASK_SCRATCH_FREE(scratch);
834 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835}
836
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700837/*
838 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700839 *
840 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700841 */
842static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700844 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700845 if (p == &default_policy)
846 return;
847
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700848 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700849 case MPOL_BIND:
850 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700852 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 break;
854 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700855 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700856 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700857 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858 break;
859 default:
860 BUG();
861 }
862}
863
864static int lookup_node(struct mm_struct *mm, unsigned long addr)
865{
866 struct page *p;
867 int err;
868
869 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
870 if (err >= 0) {
871 err = page_to_nid(p);
872 put_page(p);
873 }
874 return err;
875}
876
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700878static long do_get_mempolicy(int *policy, nodemask_t *nmask,
879 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700881 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 struct mm_struct *mm = current->mm;
883 struct vm_area_struct *vma = NULL;
884 struct mempolicy *pol = current->mempolicy;
885
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700886 if (flags &
887 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700889
890 if (flags & MPOL_F_MEMS_ALLOWED) {
891 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
892 return -EINVAL;
893 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700894 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700895 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700896 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700897 return 0;
898 }
899
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700901 /*
902 * Do NOT fall back to task policy if the
903 * vma/shared policy at addr is NULL. We
904 * want to return MPOL_DEFAULT in this case.
905 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 down_read(&mm->mmap_sem);
907 vma = find_vma_intersection(mm, addr, addr+1);
908 if (!vma) {
909 up_read(&mm->mmap_sem);
910 return -EFAULT;
911 }
912 if (vma->vm_ops && vma->vm_ops->get_policy)
913 pol = vma->vm_ops->get_policy(vma, addr);
914 else
915 pol = vma->vm_policy;
916 } else if (addr)
917 return -EINVAL;
918
919 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700920 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
922 if (flags & MPOL_F_NODE) {
923 if (flags & MPOL_F_ADDR) {
924 err = lookup_node(mm, addr);
925 if (err < 0)
926 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700927 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700929 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700930 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 } else {
932 err = -EINVAL;
933 goto out;
934 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700935 } else {
936 *policy = pol == &default_policy ? MPOL_DEFAULT :
937 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700938 /*
939 * Internal mempolicy flags must be masked off before exposing
940 * the policy to userspace.
941 */
942 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700943 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944
945 if (vma) {
946 up_read(&current->mm->mmap_sem);
947 vma = NULL;
948 }
949
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700951 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700952 if (mpol_store_user_nodemask(pol)) {
953 *nmask = pol->w.user_nodemask;
954 } else {
955 task_lock(current);
956 get_policy_nodemask(pol, nmask);
957 task_unlock(current);
958 }
Miao Xie58568d22009-06-16 15:31:49 -0700959 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960
961 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700962 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 if (vma)
964 up_read(&current->mm->mmap_sem);
965 return err;
966}
967
Christoph Lameterb20a3502006-03-22 00:09:12 -0800968#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700969/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800970 * page migration
971 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800972static void migrate_page_add(struct page *page, struct list_head *pagelist,
973 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800974{
975 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800976 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800977 */
Nick Piggin62695a82008-10-18 20:26:09 -0700978 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
979 if (!isolate_lru_page(page)) {
980 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800981 inc_zone_page_state(page, NR_ISOLATED_ANON +
982 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700983 }
984 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800985}
986
Christoph Lameter742755a2006-06-23 02:03:55 -0700987static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700988{
Mel Gorman6484eb32009-06-16 15:31:54 -0700989 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700990}
991
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800992/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800993 * Migrate pages from one node to a target node.
994 * Returns error or the number of pages not migrated.
995 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700996static int migrate_to_node(struct mm_struct *mm, int source, int dest,
997 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800998{
999 nodemask_t nmask;
1000 LIST_HEAD(pagelist);
1001 int err = 0;
1002
1003 nodes_clear(nmask);
1004 node_set(source, nmask);
1005
Minchan Kim08270802012-10-08 16:33:38 -07001006 /*
1007 * This does not "check" the range but isolates all pages that
1008 * need migration. Between passing in the full user address
1009 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1010 */
1011 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1012 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001013 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1014
Minchan Kimcf608ac2010-10-26 14:21:29 -07001015 if (!list_empty(&pagelist)) {
Mel Gorman7f0f2492011-01-13 15:45:58 -08001016 err = migrate_pages(&pagelist, new_node_page, dest,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001017 MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001018 if (err)
1019 putback_lru_pages(&pagelist);
1020 }
Christoph Lameter95a402c2006-06-23 02:03:53 -07001021
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001022 return err;
1023}
1024
1025/*
1026 * Move pages between the two nodesets so as to preserve the physical
1027 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -08001028 *
1029 * Returns the number of page that could not be moved.
1030 */
Andrew Morton0ce72d42012-05-29 15:06:24 -07001031int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1032 const nodemask_t *to, int flags)
Christoph Lameter39743882006-01-08 01:00:51 -08001033{
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001034 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001035 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001036 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -08001037
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001038 err = migrate_prep();
1039 if (err)
1040 return err;
1041
Lee Schermerhorn53f25562008-04-28 02:13:20 -07001042 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -08001043
Andrew Morton0ce72d42012-05-29 15:06:24 -07001044 err = migrate_vmas(mm, from, to, flags);
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001045 if (err)
1046 goto out;
1047
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001048 /*
1049 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1050 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1051 * bit in 'tmp', and return that <source, dest> pair for migration.
1052 * The pair of nodemasks 'to' and 'from' define the map.
1053 *
1054 * If no pair of bits is found that way, fallback to picking some
1055 * pair of 'source' and 'dest' bits that are not the same. If the
1056 * 'source' and 'dest' bits are the same, this represents a node
1057 * that will be migrating to itself, so no pages need move.
1058 *
1059 * If no bits are left in 'tmp', or if all remaining bits left
1060 * in 'tmp' correspond to the same bit in 'to', return false
1061 * (nothing left to migrate).
1062 *
1063 * This lets us pick a pair of nodes to migrate between, such that
1064 * if possible the dest node is not already occupied by some other
1065 * source node, minimizing the risk of overloading the memory on a
1066 * node that would happen if we migrated incoming memory to a node
1067 * before migrating outgoing memory source that same node.
1068 *
1069 * A single scan of tmp is sufficient. As we go, we remember the
1070 * most recent <s, d> pair that moved (s != d). If we find a pair
1071 * that not only moved, but what's better, moved to an empty slot
1072 * (d is not set in tmp), then we break out then, with that pair.
Justin P. Mattockae0e47f2011-03-01 15:06:02 +01001073 * Otherwise when we finish scanning from_tmp, we at least have the
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001074 * most recent <s, d> pair that moved. If we get all the way through
1075 * the scan of tmp without finding any node that moved, much less
1076 * moved to an empty node, then there is nothing left worth migrating.
1077 */
Christoph Lameterd4984712006-01-08 01:00:55 -08001078
Andrew Morton0ce72d42012-05-29 15:06:24 -07001079 tmp = *from;
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001080 while (!nodes_empty(tmp)) {
1081 int s,d;
1082 int source = -1;
1083 int dest = 0;
1084
1085 for_each_node_mask(s, tmp) {
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001086
1087 /*
1088 * do_migrate_pages() tries to maintain the relative
1089 * node relationship of the pages established between
1090 * threads and memory areas.
1091 *
1092 * However if the number of source nodes is not equal to
1093 * the number of destination nodes we can not preserve
1094 * this node relative relationship. In that case, skip
1095 * copying memory from a node that is in the destination
1096 * mask.
1097 *
1098 * Example: [2,3,4] -> [3,4,5] moves everything.
1099 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1100 */
1101
Andrew Morton0ce72d42012-05-29 15:06:24 -07001102 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1103 (node_isset(s, *to)))
Larry Woodman4a5b18c2012-05-29 15:06:24 -07001104 continue;
1105
Andrew Morton0ce72d42012-05-29 15:06:24 -07001106 d = node_remap(s, *from, *to);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001107 if (s == d)
1108 continue;
1109
1110 source = s; /* Node moved. Memorize */
1111 dest = d;
1112
1113 /* dest not in remaining from nodes? */
1114 if (!node_isset(dest, tmp))
1115 break;
1116 }
1117 if (source == -1)
1118 break;
1119
1120 node_clear(source, tmp);
1121 err = migrate_to_node(mm, source, dest, flags);
1122 if (err > 0)
1123 busy += err;
1124 if (err < 0)
1125 break;
Christoph Lameter39743882006-01-08 01:00:51 -08001126 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001127out:
Christoph Lameter39743882006-01-08 01:00:51 -08001128 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001129 if (err < 0)
1130 return err;
1131 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001132
Christoph Lameter39743882006-01-08 01:00:51 -08001133}
1134
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001135/*
1136 * Allocate a new page for page migration based on vma policy.
1137 * Start assuming that page is mapped by vma pointed to by @private.
1138 * Search forward from there, if not. N.B., this assumes that the
1139 * list of pages handed to migrate_pages()--which is how we get here--
1140 * is in virtual address order.
1141 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001142static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001143{
1144 struct vm_area_struct *vma = (struct vm_area_struct *)private;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001145 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001146
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001147 while (vma) {
1148 address = page_address_in_vma(page, vma);
1149 if (address != -EFAULT)
1150 break;
1151 vma = vma->vm_next;
1152 }
1153
1154 /*
1155 * if !vma, alloc_page_vma() will use task or system default policy
1156 */
1157 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001158}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001159#else
1160
1161static void migrate_page_add(struct page *page, struct list_head *pagelist,
1162 unsigned long flags)
1163{
1164}
1165
Andrew Morton0ce72d42012-05-29 15:06:24 -07001166int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1167 const nodemask_t *to, int flags)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001168{
1169 return -ENOSYS;
1170}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001171
Keith Owens69939742006-10-11 01:21:28 -07001172static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001173{
1174 return NULL;
1175}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001176#endif
1177
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001178static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001179 unsigned short mode, unsigned short mode_flags,
1180 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001181{
1182 struct vm_area_struct *vma;
1183 struct mm_struct *mm = current->mm;
1184 struct mempolicy *new;
1185 unsigned long end;
1186 int err;
1187 LIST_HEAD(pagelist);
1188
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001189 if (flags & ~(unsigned long)MPOL_MF_VALID)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001190 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001191 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001192 return -EPERM;
1193
1194 if (start & ~PAGE_MASK)
1195 return -EINVAL;
1196
1197 if (mode == MPOL_DEFAULT)
1198 flags &= ~MPOL_MF_STRICT;
1199
1200 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1201 end = start + len;
1202
1203 if (end < start)
1204 return -EINVAL;
1205 if (end == start)
1206 return 0;
1207
David Rientjes028fec42008-04-28 02:12:25 -07001208 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001209 if (IS_ERR(new))
1210 return PTR_ERR(new);
1211
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001212 if (flags & MPOL_MF_LAZY)
1213 new->flags |= MPOL_F_MOF;
1214
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001215 /*
1216 * If we are using the default policy then operation
1217 * on discontinuous address spaces is okay after all
1218 */
1219 if (!new)
1220 flags |= MPOL_MF_DISCONTIG_OK;
1221
David Rientjes028fec42008-04-28 02:12:25 -07001222 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1223 start, start + len, mode, mode_flags,
David Rientjes00ef2d22013-02-22 16:35:36 -08001224 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001225
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001226 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1227
1228 err = migrate_prep();
1229 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001230 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001231 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001232 {
1233 NODEMASK_SCRATCH(scratch);
1234 if (scratch) {
1235 down_write(&mm->mmap_sem);
1236 task_lock(current);
1237 err = mpol_set_nodemask(new, nmask, scratch);
1238 task_unlock(current);
1239 if (err)
1240 up_write(&mm->mmap_sem);
1241 } else
1242 err = -ENOMEM;
1243 NODEMASK_SCRATCH_FREE(scratch);
1244 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001245 if (err)
1246 goto mpol_out;
1247
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001248 vma = check_range(mm, start, end, nmask,
1249 flags | MPOL_MF_INVERT, &pagelist);
1250
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001251 err = PTR_ERR(vma); /* maybe ... */
Mel Gormana7200942012-11-16 09:37:58 +00001252 if (!IS_ERR(vma))
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001253 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001254
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001255 if (!err) {
1256 int nr_failed = 0;
1257
Minchan Kimcf608ac2010-10-26 14:21:29 -07001258 if (!list_empty(&pagelist)) {
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001259 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001260 nr_failed = migrate_pages(&pagelist, new_vma_page,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001261 (unsigned long)vma,
1262 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001263 if (nr_failed)
1264 putback_lru_pages(&pagelist);
1265 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001266
Lee Schermerhornb24f53a2012-10-25 14:16:32 +02001267 if (nr_failed && (flags & MPOL_MF_STRICT))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001268 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001269 } else
1270 putback_lru_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001271
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001272 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001273 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001274 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001275 return err;
1276}
1277
Christoph Lameter39743882006-01-08 01:00:51 -08001278/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001279 * User space interface with variable sized bitmaps for nodelists.
1280 */
1281
1282/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001283static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001284 unsigned long maxnode)
1285{
1286 unsigned long k;
1287 unsigned long nlongs;
1288 unsigned long endmask;
1289
1290 --maxnode;
1291 nodes_clear(*nodes);
1292 if (maxnode == 0 || !nmask)
1293 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001294 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001295 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001296
1297 nlongs = BITS_TO_LONGS(maxnode);
1298 if ((maxnode % BITS_PER_LONG) == 0)
1299 endmask = ~0UL;
1300 else
1301 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1302
1303 /* When the user specified more nodes than supported just check
1304 if the non supported part is all zero. */
1305 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1306 if (nlongs > PAGE_SIZE/sizeof(long))
1307 return -EINVAL;
1308 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1309 unsigned long t;
1310 if (get_user(t, nmask + k))
1311 return -EFAULT;
1312 if (k == nlongs - 1) {
1313 if (t & endmask)
1314 return -EINVAL;
1315 } else if (t)
1316 return -EINVAL;
1317 }
1318 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1319 endmask = ~0UL;
1320 }
1321
1322 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1323 return -EFAULT;
1324 nodes_addr(*nodes)[nlongs-1] &= endmask;
1325 return 0;
1326}
1327
1328/* Copy a kernel node mask to user space */
1329static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1330 nodemask_t *nodes)
1331{
1332 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1333 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1334
1335 if (copy > nbytes) {
1336 if (copy > PAGE_SIZE)
1337 return -EINVAL;
1338 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1339 return -EFAULT;
1340 copy = nbytes;
1341 }
1342 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1343}
1344
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001345SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1346 unsigned long, mode, unsigned long __user *, nmask,
1347 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001348{
1349 nodemask_t nodes;
1350 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001351 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001352
David Rientjes028fec42008-04-28 02:12:25 -07001353 mode_flags = mode & MPOL_MODE_FLAGS;
1354 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001355 if (mode >= MPOL_MAX)
1356 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001357 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1358 (mode_flags & MPOL_F_RELATIVE_NODES))
1359 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001360 err = get_nodes(&nodes, nmask, maxnode);
1361 if (err)
1362 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001363 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001364}
1365
1366/* Set the process memory policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001367SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1368 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001369{
1370 int err;
1371 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001372 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001373
David Rientjes028fec42008-04-28 02:12:25 -07001374 flags = mode & MPOL_MODE_FLAGS;
1375 mode &= ~MPOL_MODE_FLAGS;
1376 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001377 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001378 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1379 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001380 err = get_nodes(&nodes, nmask, maxnode);
1381 if (err)
1382 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001383 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001384}
1385
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001386SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1387 const unsigned long __user *, old_nodes,
1388 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001389{
David Howellsc69e8d92008-11-14 10:39:19 +11001390 const struct cred *cred = current_cred(), *tcred;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001391 struct mm_struct *mm = NULL;
Christoph Lameter39743882006-01-08 01:00:51 -08001392 struct task_struct *task;
Christoph Lameter39743882006-01-08 01:00:51 -08001393 nodemask_t task_nodes;
1394 int err;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001395 nodemask_t *old;
1396 nodemask_t *new;
1397 NODEMASK_SCRATCH(scratch);
Christoph Lameter39743882006-01-08 01:00:51 -08001398
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001399 if (!scratch)
1400 return -ENOMEM;
Christoph Lameter39743882006-01-08 01:00:51 -08001401
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001402 old = &scratch->mask1;
1403 new = &scratch->mask2;
1404
1405 err = get_nodes(old, old_nodes, maxnode);
Christoph Lameter39743882006-01-08 01:00:51 -08001406 if (err)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001407 goto out;
1408
1409 err = get_nodes(new, new_nodes, maxnode);
1410 if (err)
1411 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001412
1413 /* Find the mm_struct */
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001414 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001415 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001416 if (!task) {
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001417 rcu_read_unlock();
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001418 err = -ESRCH;
1419 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001420 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001421 get_task_struct(task);
Christoph Lameter39743882006-01-08 01:00:51 -08001422
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001423 err = -EINVAL;
Christoph Lameter39743882006-01-08 01:00:51 -08001424
1425 /*
1426 * Check if this process has the right to modify the specified
1427 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001428 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001429 * userid as the target process.
1430 */
David Howellsc69e8d92008-11-14 10:39:19 +11001431 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001432 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1433 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001434 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001435 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001436 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001437 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001438 }
David Howellsc69e8d92008-11-14 10:39:19 +11001439 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001440
1441 task_nodes = cpuset_mems_allowed(task);
1442 /* Is the user allowed to access the target nodes? */
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001443 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001444 err = -EPERM;
Christoph Lameter3268c632012-03-21 16:34:06 -07001445 goto out_put;
Christoph Lameter39743882006-01-08 01:00:51 -08001446 }
1447
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08001448 if (!nodes_subset(*new, node_states[N_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001449 err = -EINVAL;
Christoph Lameter3268c632012-03-21 16:34:06 -07001450 goto out_put;
Christoph Lameter3b42d282007-08-31 00:12:08 -07001451 }
1452
David Quigley86c3a762006-06-23 02:04:02 -07001453 err = security_task_movememory(task);
1454 if (err)
Christoph Lameter3268c632012-03-21 16:34:06 -07001455 goto out_put;
David Quigley86c3a762006-06-23 02:04:02 -07001456
Christoph Lameter3268c632012-03-21 16:34:06 -07001457 mm = get_task_mm(task);
1458 put_task_struct(task);
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001459
1460 if (!mm) {
Christoph Lameter3268c632012-03-21 16:34:06 -07001461 err = -EINVAL;
Sasha Levinf2a9ef82012-04-25 16:01:52 -07001462 goto out;
1463 }
1464
1465 err = do_migrate_pages(mm, old, new,
1466 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter3268c632012-03-21 16:34:06 -07001467
1468 mmput(mm);
1469out:
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001470 NODEMASK_SCRATCH_FREE(scratch);
1471
Christoph Lameter39743882006-01-08 01:00:51 -08001472 return err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001473
1474out_put:
1475 put_task_struct(task);
1476 goto out;
1477
Christoph Lameter39743882006-01-08 01:00:51 -08001478}
1479
1480
Christoph Lameter8bccd852005-10-29 18:16:59 -07001481/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001482SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1483 unsigned long __user *, nmask, unsigned long, maxnode,
1484 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001485{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001486 int err;
1487 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001488 nodemask_t nodes;
1489
1490 if (nmask != NULL && maxnode < MAX_NUMNODES)
1491 return -EINVAL;
1492
1493 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1494
1495 if (err)
1496 return err;
1497
1498 if (policy && put_user(pval, policy))
1499 return -EFAULT;
1500
1501 if (nmask)
1502 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1503
1504 return err;
1505}
1506
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507#ifdef CONFIG_COMPAT
1508
1509asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1510 compat_ulong_t __user *nmask,
1511 compat_ulong_t maxnode,
1512 compat_ulong_t addr, compat_ulong_t flags)
1513{
1514 long err;
1515 unsigned long __user *nm = NULL;
1516 unsigned long nr_bits, alloc_size;
1517 DECLARE_BITMAP(bm, MAX_NUMNODES);
1518
1519 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1520 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1521
1522 if (nmask)
1523 nm = compat_alloc_user_space(alloc_size);
1524
1525 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1526
1527 if (!err && nmask) {
KAMEZAWA Hiroyuki2bbff6c2011-09-14 16:21:02 -07001528 unsigned long copy_size;
1529 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1530 err = copy_from_user(bm, nm, copy_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 /* ensure entire bitmap is zeroed */
1532 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1533 err |= compat_put_bitmap(nmask, bm, nr_bits);
1534 }
1535
1536 return err;
1537}
1538
1539asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1540 compat_ulong_t maxnode)
1541{
1542 long err = 0;
1543 unsigned long __user *nm = NULL;
1544 unsigned long nr_bits, alloc_size;
1545 DECLARE_BITMAP(bm, MAX_NUMNODES);
1546
1547 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1548 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1549
1550 if (nmask) {
1551 err = compat_get_bitmap(bm, nmask, nr_bits);
1552 nm = compat_alloc_user_space(alloc_size);
1553 err |= copy_to_user(nm, bm, alloc_size);
1554 }
1555
1556 if (err)
1557 return -EFAULT;
1558
1559 return sys_set_mempolicy(mode, nm, nr_bits+1);
1560}
1561
1562asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1563 compat_ulong_t mode, compat_ulong_t __user *nmask,
1564 compat_ulong_t maxnode, compat_ulong_t flags)
1565{
1566 long err = 0;
1567 unsigned long __user *nm = NULL;
1568 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001569 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570
1571 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1572 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1573
1574 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001575 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001577 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 }
1579
1580 if (err)
1581 return -EFAULT;
1582
1583 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1584}
1585
1586#endif
1587
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001588/*
1589 * get_vma_policy(@task, @vma, @addr)
1590 * @task - task for fallback if vma policy == default
1591 * @vma - virtual memory area whose policy is sought
1592 * @addr - address in @vma for shared policy lookup
1593 *
1594 * Returns effective policy for a VMA at specified address.
1595 * Falls back to @task or system default policy, as necessary.
David Rientjes32f85162012-10-16 17:31:23 -07001596 * Current or other task's task mempolicy and non-shared vma policies must be
1597 * protected by task_lock(task) by the caller.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001598 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1599 * count--added by the get_policy() vm_op, as appropriate--to protect against
1600 * freeing by another task. It is the caller's responsibility to free the
1601 * extra reference for shared policies.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001602 */
Stephen Wilsond98f6cb2011-05-24 17:12:41 -07001603struct mempolicy *get_vma_policy(struct task_struct *task,
Christoph Lameter48fce342006-01-08 01:01:03 -08001604 struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605{
Mel Gorman5606e382012-11-02 18:19:13 +00001606 struct mempolicy *pol = get_task_policy(task);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607
1608 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001609 if (vma->vm_ops && vma->vm_ops->get_policy) {
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001610 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1611 addr);
1612 if (vpol)
1613 pol = vpol;
Mel Gorman00442ad2012-10-08 16:29:20 -07001614 } else if (vma->vm_policy) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 pol = vma->vm_policy;
Mel Gorman00442ad2012-10-08 16:29:20 -07001616
1617 /*
1618 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1619 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1620 * count on these policies which will be dropped by
1621 * mpol_cond_put() later
1622 */
1623 if (mpol_needs_cond_ref(pol))
1624 mpol_get(pol);
1625 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 }
1627 if (!pol)
1628 pol = &default_policy;
1629 return pol;
1630}
1631
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001632static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1633{
1634 enum zone_type dynamic_policy_zone = policy_zone;
1635
1636 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1637
1638 /*
1639 * if policy->v.nodes has movable memory only,
1640 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1641 *
1642 * policy->v.nodes is intersect with node_states[N_MEMORY].
1643 * so if the following test faile, it implies
1644 * policy->v.nodes has movable memory only.
1645 */
1646 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1647 dynamic_policy_zone = ZONE_MOVABLE;
1648
1649 return zone >= dynamic_policy_zone;
1650}
1651
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001652/*
1653 * Return a nodemask representing a mempolicy for filtering nodes for
1654 * page allocation
1655 */
1656static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001657{
1658 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001659 if (unlikely(policy->mode == MPOL_BIND) &&
Lai Jiangshand3eb1572013-02-22 16:33:22 -08001660 apply_policy_zone(policy, gfp_zone(gfp)) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001661 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1662 return &policy->v.nodes;
1663
1664 return NULL;
1665}
1666
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001667/* Return a zonelist indicated by gfp for node representing a mempolicy */
Andi Kleen2f5f9482011-03-04 17:36:29 -08001668static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1669 int nd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670{
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001671 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001673 if (!(policy->flags & MPOL_F_LOCAL))
1674 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 break;
1676 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001677 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001678 * Normally, MPOL_BIND allocations are node-local within the
1679 * allowed nodemask. However, if __GFP_THISNODE is set and the
Bob Liu6eb27e12010-05-24 14:32:00 -07001680 * current node isn't part of the mask, we use the zonelist for
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001681 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001682 */
Mel Gorman19770b32008-04-28 02:12:18 -07001683 if (unlikely(gfp & __GFP_THISNODE) &&
1684 unlikely(!node_isset(nd, policy->v.nodes)))
1685 nd = first_node(policy->v.nodes);
1686 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 BUG();
1689 }
Mel Gorman0e884602008-04-28 02:12:14 -07001690 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691}
1692
1693/* Do dynamic interleaving for a process */
1694static unsigned interleave_nodes(struct mempolicy *policy)
1695{
1696 unsigned nid, next;
1697 struct task_struct *me = current;
1698
1699 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001700 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001702 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001703 if (next < MAX_NUMNODES)
1704 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 return nid;
1706}
1707
Christoph Lameterdc85da12006-01-18 17:42:36 -08001708/*
1709 * Depending on the memory policy provide a node from which to allocate the
1710 * next slab entry.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001711 * @policy must be protected by freeing by the caller. If @policy is
1712 * the current task's mempolicy, this protection is implicit, as only the
1713 * task can change it's policy. The system default policy requires no
1714 * such protection.
Christoph Lameterdc85da12006-01-18 17:42:36 -08001715 */
Andi Kleene7b691b2012-06-09 02:40:03 -07001716unsigned slab_node(void)
Christoph Lameterdc85da12006-01-18 17:42:36 -08001717{
Andi Kleene7b691b2012-06-09 02:40:03 -07001718 struct mempolicy *policy;
1719
1720 if (in_interrupt())
1721 return numa_node_id();
1722
1723 policy = current->mempolicy;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001724 if (!policy || policy->flags & MPOL_F_LOCAL)
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001725 return numa_node_id();
Christoph Lameter765c4502006-09-27 01:50:08 -07001726
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001727 switch (policy->mode) {
1728 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001729 /*
1730 * handled MPOL_F_LOCAL above
1731 */
1732 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001733
Christoph Lameterdc85da12006-01-18 17:42:36 -08001734 case MPOL_INTERLEAVE:
1735 return interleave_nodes(policy);
1736
Mel Gormandd1a2392008-04-28 02:12:17 -07001737 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001738 /*
1739 * Follow bind policy behavior and start allocation at the
1740 * first node.
1741 */
Mel Gorman19770b32008-04-28 02:12:18 -07001742 struct zonelist *zonelist;
1743 struct zone *zone;
1744 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1745 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1746 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1747 &policy->v.nodes,
1748 &zone);
Eric Dumazet800416f2010-10-27 19:33:43 +02001749 return zone ? zone->node : numa_node_id();
Mel Gormandd1a2392008-04-28 02:12:17 -07001750 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001751
Christoph Lameterdc85da12006-01-18 17:42:36 -08001752 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001753 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001754 }
1755}
1756
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757/* Do static interleaving for a VMA with known offset. */
1758static unsigned offset_il_node(struct mempolicy *pol,
1759 struct vm_area_struct *vma, unsigned long off)
1760{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001761 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001762 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 int c;
1764 int nid = -1;
1765
David Rientjesf5b087b2008-04-28 02:12:27 -07001766 if (!nnodes)
1767 return numa_node_id();
1768 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 c = 0;
1770 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001771 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 c++;
1773 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 return nid;
1775}
1776
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001777/* Determine a node number for interleave */
1778static inline unsigned interleave_nid(struct mempolicy *pol,
1779 struct vm_area_struct *vma, unsigned long addr, int shift)
1780{
1781 if (vma) {
1782 unsigned long off;
1783
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001784 /*
1785 * for small pages, there is no difference between
1786 * shift and PAGE_SHIFT, so the bit-shift is safe.
1787 * for huge pages, since vm_pgoff is in units of small
1788 * pages, we need to shift off the always 0 bits to get
1789 * a useful offset.
1790 */
1791 BUG_ON(shift < PAGE_SHIFT);
1792 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001793 off += (addr - vma->vm_start) >> shift;
1794 return offset_il_node(pol, vma, off);
1795 } else
1796 return interleave_nodes(pol);
1797}
1798
Michal Hocko778d3b02011-07-26 16:08:30 -07001799/*
1800 * Return the bit number of a random bit set in the nodemask.
1801 * (returns -1 if nodemask is empty)
1802 */
1803int node_random(const nodemask_t *maskp)
1804{
1805 int w, bit = -1;
1806
1807 w = nodes_weight(*maskp);
1808 if (w)
1809 bit = bitmap_ord_to_pos(maskp->bits,
1810 get_random_int() % w, MAX_NUMNODES);
1811 return bit;
1812}
1813
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001814#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001815/*
1816 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1817 * @vma = virtual memory area whose policy is sought
1818 * @addr = address in @vma for shared policy lookup and interleave policy
1819 * @gfp_flags = for requested zone
Mel Gorman19770b32008-04-28 02:12:18 -07001820 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1821 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001822 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001823 * Returns a zonelist suitable for a huge page allocation and a pointer
1824 * to the struct mempolicy for conditional unref after allocation.
1825 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1826 * @nodemask for filtering the zonelist.
Miao Xiec0ff7452010-05-24 14:32:08 -07001827 *
1828 * Must be protected by get_mems_allowed()
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001829 */
Mel Gorman396faf02007-07-17 04:03:13 -07001830struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001831 gfp_t gfp_flags, struct mempolicy **mpol,
1832 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001833{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001834 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001835
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001836 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001837 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001838
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001839 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1840 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001841 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001842 } else {
Andi Kleen2f5f9482011-03-04 17:36:29 -08001843 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001844 if ((*mpol)->mode == MPOL_BIND)
1845 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001846 }
1847 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001848}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001849
1850/*
1851 * init_nodemask_of_mempolicy
1852 *
1853 * If the current task's mempolicy is "default" [NULL], return 'false'
1854 * to indicate default policy. Otherwise, extract the policy nodemask
1855 * for 'bind' or 'interleave' policy into the argument nodemask, or
1856 * initialize the argument nodemask to contain the single node for
1857 * 'preferred' or 'local' policy and return 'true' to indicate presence
1858 * of non-default mempolicy.
1859 *
1860 * We don't bother with reference counting the mempolicy [mpol_get/put]
1861 * because the current task is examining it's own mempolicy and a task's
1862 * mempolicy is only ever changed by the task itself.
1863 *
1864 * N.B., it is the caller's responsibility to free a returned nodemask.
1865 */
1866bool init_nodemask_of_mempolicy(nodemask_t *mask)
1867{
1868 struct mempolicy *mempolicy;
1869 int nid;
1870
1871 if (!(mask && current->mempolicy))
1872 return false;
1873
Miao Xiec0ff7452010-05-24 14:32:08 -07001874 task_lock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001875 mempolicy = current->mempolicy;
1876 switch (mempolicy->mode) {
1877 case MPOL_PREFERRED:
1878 if (mempolicy->flags & MPOL_F_LOCAL)
1879 nid = numa_node_id();
1880 else
1881 nid = mempolicy->v.preferred_node;
1882 init_nodemask_of_node(mask, nid);
1883 break;
1884
1885 case MPOL_BIND:
1886 /* Fall through */
1887 case MPOL_INTERLEAVE:
1888 *mask = mempolicy->v.nodes;
1889 break;
1890
1891 default:
1892 BUG();
1893 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001894 task_unlock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001895
1896 return true;
1897}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001898#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001899
David Rientjes6f48d0eb2010-08-09 17:18:52 -07001900/*
1901 * mempolicy_nodemask_intersects
1902 *
1903 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1904 * policy. Otherwise, check for intersection between mask and the policy
1905 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1906 * policy, always return true since it may allocate elsewhere on fallback.
1907 *
1908 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1909 */
1910bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1911 const nodemask_t *mask)
1912{
1913 struct mempolicy *mempolicy;
1914 bool ret = true;
1915
1916 if (!mask)
1917 return ret;
1918 task_lock(tsk);
1919 mempolicy = tsk->mempolicy;
1920 if (!mempolicy)
1921 goto out;
1922
1923 switch (mempolicy->mode) {
1924 case MPOL_PREFERRED:
1925 /*
1926 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1927 * allocate from, they may fallback to other nodes when oom.
1928 * Thus, it's possible for tsk to have allocated memory from
1929 * nodes in mask.
1930 */
1931 break;
1932 case MPOL_BIND:
1933 case MPOL_INTERLEAVE:
1934 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1935 break;
1936 default:
1937 BUG();
1938 }
1939out:
1940 task_unlock(tsk);
1941 return ret;
1942}
1943
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944/* Allocate a page in interleaved policy.
1945 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001946static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1947 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948{
1949 struct zonelist *zl;
1950 struct page *page;
1951
Mel Gorman0e884602008-04-28 02:12:14 -07001952 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001954 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001955 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956 return page;
1957}
1958
1959/**
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001960 * alloc_pages_vma - Allocate a page for a VMA.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 *
1962 * @gfp:
1963 * %GFP_USER user allocation.
1964 * %GFP_KERNEL kernel allocations,
1965 * %GFP_HIGHMEM highmem/user allocations,
1966 * %GFP_FS allocation should not call back into a file system.
1967 * %GFP_ATOMIC don't sleep.
1968 *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001969 * @order:Order of the GFP allocation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 * @vma: Pointer to VMA or NULL if not available.
1971 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1972 *
1973 * This function allocates a page from the kernel page pool and applies
1974 * a NUMA policy associated with the VMA or the current process.
1975 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1976 * mm_struct of the VMA to prevent it from going away. Should be used for
1977 * all allocations for pages that will be mapped into
1978 * user space. Returns NULL when no page can be allocated.
1979 *
1980 * Should be called with the mm_sem of the vma hold.
1981 */
1982struct page *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001983alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
Andi Kleen2f5f9482011-03-04 17:36:29 -08001984 unsigned long addr, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985{
Mel Gormancc9a6c82012-03-21 16:34:11 -07001986 struct mempolicy *pol;
Miao Xiec0ff7452010-05-24 14:32:08 -07001987 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001988 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989
Mel Gormancc9a6c82012-03-21 16:34:11 -07001990retry_cpuset:
1991 pol = get_vma_policy(current, vma, addr);
1992 cpuset_mems_cookie = get_mems_allowed();
1993
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001994 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001996
Andi Kleen8eac5632011-02-25 14:44:28 -08001997 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001998 mpol_cond_put(pol);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001999 page = alloc_page_interleave(gfp, order, nid);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002000 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2001 goto retry_cpuset;
2002
Miao Xiec0ff7452010-05-24 14:32:08 -07002003 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 }
David Rientjes212a0a62012-12-11 16:02:51 -08002005 page = __alloc_pages_nodemask(gfp, order,
2006 policy_zonelist(gfp, pol, node),
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08002007 policy_nodemask(gfp, pol));
David Rientjes212a0a62012-12-11 16:02:51 -08002008 if (unlikely(mpol_needs_cond_ref(pol)))
2009 __mpol_put(pol);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002010 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2011 goto retry_cpuset;
Miao Xiec0ff7452010-05-24 14:32:08 -07002012 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013}
2014
2015/**
2016 * alloc_pages_current - Allocate pages.
2017 *
2018 * @gfp:
2019 * %GFP_USER user allocation,
2020 * %GFP_KERNEL kernel allocation,
2021 * %GFP_HIGHMEM highmem allocation,
2022 * %GFP_FS don't call back into a file system.
2023 * %GFP_ATOMIC don't sleep.
2024 * @order: Power of two of allocation size in pages. 0 is a single page.
2025 *
2026 * Allocate a page from the kernel page pool. When not in
2027 * interrupt context and apply the current process NUMA policy.
2028 * Returns NULL when no page can be allocated.
2029 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08002030 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 * 1) it's ok to take cpuset_sem (can WAIT), and
2032 * 2) allocating for current task (not interrupt).
2033 */
Al Virodd0fc662005-10-07 07:46:04 +01002034struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035{
Mel Gorman5606e382012-11-02 18:19:13 +00002036 struct mempolicy *pol = get_task_policy(current);
Miao Xiec0ff7452010-05-24 14:32:08 -07002037 struct page *page;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002038 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039
Christoph Lameter9b819d22006-09-25 23:31:40 -07002040 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 pol = &default_policy;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002042
Mel Gormancc9a6c82012-03-21 16:34:11 -07002043retry_cpuset:
2044 cpuset_mems_cookie = get_mems_allowed();
2045
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002046 /*
2047 * No reference counting needed for current->mempolicy
2048 * nor system default_policy
2049 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002050 if (pol->mode == MPOL_INTERLEAVE)
Miao Xiec0ff7452010-05-24 14:32:08 -07002051 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2052 else
2053 page = __alloc_pages_nodemask(gfp, order,
Andi Kleen5c4b4be2011-03-04 17:36:32 -08002054 policy_zonelist(gfp, pol, numa_node_id()),
2055 policy_nodemask(gfp, pol));
Mel Gormancc9a6c82012-03-21 16:34:11 -07002056
2057 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2058 goto retry_cpuset;
2059
Miao Xiec0ff7452010-05-24 14:32:08 -07002060 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061}
2062EXPORT_SYMBOL(alloc_pages_current);
2063
Paul Jackson42253992006-01-08 01:01:59 -08002064/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002065 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08002066 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2067 * with the mems_allowed returned by cpuset_mems_allowed(). This
2068 * keeps mempolicies cpuset relative after its cpuset moves. See
2069 * further kernel/cpuset.c update_nodemask().
Miao Xie708c1bb2010-05-24 14:32:07 -07002070 *
2071 * current's mempolicy may be rebinded by the other task(the task that changes
2072 * cpuset's mems), so we needn't do rebind work for current task.
Paul Jackson42253992006-01-08 01:01:59 -08002073 */
Paul Jackson42253992006-01-08 01:01:59 -08002074
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07002075/* Slow path of a mempolicy duplicate */
2076struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077{
2078 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2079
2080 if (!new)
2081 return ERR_PTR(-ENOMEM);
Miao Xie708c1bb2010-05-24 14:32:07 -07002082
2083 /* task's mempolicy is protected by alloc_lock */
2084 if (old == current->mempolicy) {
2085 task_lock(current);
2086 *new = *old;
2087 task_unlock(current);
2088 } else
2089 *new = *old;
2090
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08002091 rcu_read_lock();
Paul Jackson42253992006-01-08 01:01:59 -08002092 if (current_cpuset_is_being_rebound()) {
2093 nodemask_t mems = cpuset_mems_allowed(current);
Miao Xie708c1bb2010-05-24 14:32:07 -07002094 if (new->flags & MPOL_F_REBINDING)
2095 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2096 else
2097 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
Paul Jackson42253992006-01-08 01:01:59 -08002098 }
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08002099 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 return new;
2102}
2103
2104/* Slow path of a mempolicy comparison */
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002105bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106{
2107 if (!a || !b)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002108 return false;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002109 if (a->mode != b->mode)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002110 return false;
Bob Liu19800502010-05-24 14:32:01 -07002111 if (a->flags != b->flags)
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002112 return false;
Bob Liu19800502010-05-24 14:32:01 -07002113 if (mpol_store_user_nodemask(a))
2114 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002115 return false;
Bob Liu19800502010-05-24 14:32:01 -07002116
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002117 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07002118 case MPOL_BIND:
2119 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 case MPOL_INTERLEAVE:
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002121 return !!nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 case MPOL_PREFERRED:
Namhyung Kim75719662011-03-22 16:33:02 -07002123 return a->v.preferred_node == b->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 default:
2125 BUG();
KOSAKI Motohirofcfb4dc2012-01-10 15:08:21 -08002126 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 }
2128}
2129
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 * Shared memory backing store policy support.
2132 *
2133 * Remember policies even when nobody has shared memory mapped.
2134 * The policies are kept in Red-Black tree linked from the inode.
2135 * They are protected by the sp->lock spinlock, which should be held
2136 * for any accesses to the tree.
2137 */
2138
2139/* lookup first element intersecting start-end */
Mel Gorman42288fe2012-12-21 23:10:25 +00002140/* Caller holds sp->lock */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141static struct sp_node *
2142sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2143{
2144 struct rb_node *n = sp->root.rb_node;
2145
2146 while (n) {
2147 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2148
2149 if (start >= p->end)
2150 n = n->rb_right;
2151 else if (end <= p->start)
2152 n = n->rb_left;
2153 else
2154 break;
2155 }
2156 if (!n)
2157 return NULL;
2158 for (;;) {
2159 struct sp_node *w = NULL;
2160 struct rb_node *prev = rb_prev(n);
2161 if (!prev)
2162 break;
2163 w = rb_entry(prev, struct sp_node, nd);
2164 if (w->end <= start)
2165 break;
2166 n = prev;
2167 }
2168 return rb_entry(n, struct sp_node, nd);
2169}
2170
2171/* Insert a new shared policy into the list. */
2172/* Caller holds sp->lock */
2173static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2174{
2175 struct rb_node **p = &sp->root.rb_node;
2176 struct rb_node *parent = NULL;
2177 struct sp_node *nd;
2178
2179 while (*p) {
2180 parent = *p;
2181 nd = rb_entry(parent, struct sp_node, nd);
2182 if (new->start < nd->start)
2183 p = &(*p)->rb_left;
2184 else if (new->end > nd->end)
2185 p = &(*p)->rb_right;
2186 else
2187 BUG();
2188 }
2189 rb_link_node(&new->nd, parent, p);
2190 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07002191 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002192 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193}
2194
2195/* Find shared policy intersecting idx */
2196struct mempolicy *
2197mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2198{
2199 struct mempolicy *pol = NULL;
2200 struct sp_node *sn;
2201
2202 if (!sp->root.rb_node)
2203 return NULL;
Mel Gorman42288fe2012-12-21 23:10:25 +00002204 spin_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 sn = sp_lookup(sp, idx, idx+1);
2206 if (sn) {
2207 mpol_get(sn->policy);
2208 pol = sn->policy;
2209 }
Mel Gorman42288fe2012-12-21 23:10:25 +00002210 spin_unlock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 return pol;
2212}
2213
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002214static void sp_free(struct sp_node *n)
2215{
2216 mpol_put(n->policy);
2217 kmem_cache_free(sn_cache, n);
2218}
2219
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002220/**
2221 * mpol_misplaced - check whether current page node is valid in policy
2222 *
2223 * @page - page to be checked
2224 * @vma - vm area where page mapped
2225 * @addr - virtual address where page mapped
2226 *
2227 * Lookup current policy node id for vma,addr and "compare to" page's
2228 * node id.
2229 *
2230 * Returns:
2231 * -1 - not misplaced, page is in the right node
2232 * node - node id where the page should be
2233 *
2234 * Policy determination "mimics" alloc_page_vma().
2235 * Called from fault path where we know the vma and faulting address.
2236 */
2237int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2238{
2239 struct mempolicy *pol;
2240 struct zone *zone;
2241 int curnid = page_to_nid(page);
2242 unsigned long pgoff;
2243 int polnid = -1;
2244 int ret = -1;
2245
2246 BUG_ON(!vma);
2247
2248 pol = get_vma_policy(current, vma, addr);
2249 if (!(pol->flags & MPOL_F_MOF))
2250 goto out;
2251
2252 switch (pol->mode) {
2253 case MPOL_INTERLEAVE:
2254 BUG_ON(addr >= vma->vm_end);
2255 BUG_ON(addr < vma->vm_start);
2256
2257 pgoff = vma->vm_pgoff;
2258 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2259 polnid = offset_il_node(pol, vma, pgoff);
2260 break;
2261
2262 case MPOL_PREFERRED:
2263 if (pol->flags & MPOL_F_LOCAL)
2264 polnid = numa_node_id();
2265 else
2266 polnid = pol->v.preferred_node;
2267 break;
2268
2269 case MPOL_BIND:
2270 /*
2271 * allows binding to multiple nodes.
2272 * use current page if in policy nodemask,
2273 * else select nearest allowed node, if any.
2274 * If no allowed nodes, use current [!misplaced].
2275 */
2276 if (node_isset(curnid, pol->v.nodes))
2277 goto out;
2278 (void)first_zones_zonelist(
2279 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2280 gfp_zone(GFP_HIGHUSER),
2281 &pol->v.nodes, &zone);
2282 polnid = zone->node;
2283 break;
2284
2285 default:
2286 BUG();
2287 }
Mel Gorman5606e382012-11-02 18:19:13 +00002288
2289 /* Migrate the page towards the node whose CPU is referencing it */
Mel Gormane42c8ff2012-11-12 09:17:07 +00002290 if (pol->flags & MPOL_F_MORON) {
2291 int last_nid;
2292
Mel Gorman5606e382012-11-02 18:19:13 +00002293 polnid = numa_node_id();
2294
Mel Gormane42c8ff2012-11-12 09:17:07 +00002295 /*
2296 * Multi-stage node selection is used in conjunction
2297 * with a periodic migration fault to build a temporal
2298 * task<->page relation. By using a two-stage filter we
2299 * remove short/unlikely relations.
2300 *
2301 * Using P(p) ~ n_p / n_t as per frequentist
2302 * probability, we can equate a task's usage of a
2303 * particular page (n_p) per total usage of this
2304 * page (n_t) (in a given time-span) to a probability.
2305 *
2306 * Our periodic faults will sample this probability and
2307 * getting the same result twice in a row, given these
2308 * samples are fully independent, is then given by
2309 * P(n)^2, provided our sample period is sufficiently
2310 * short compared to the usage pattern.
2311 *
2312 * This quadric squishes small probabilities, making
2313 * it less likely we act on an unlikely task<->page
2314 * relation.
2315 */
Mel Gorman22b751c2013-02-22 16:34:59 -08002316 last_nid = page_nid_xchg_last(page, polnid);
Mel Gormane42c8ff2012-11-12 09:17:07 +00002317 if (last_nid != polnid)
2318 goto out;
2319 }
2320
Lee Schermerhorn771fb4d2012-10-25 14:16:30 +02002321 if (curnid != polnid)
2322 ret = polnid;
2323out:
2324 mpol_cond_put(pol);
2325
2326 return ret;
2327}
2328
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2330{
Paul Mundt140d5a42007-07-15 23:38:16 -07002331 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 rb_erase(&n->nd, &sp->root);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002333 sp_free(n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334}
2335
Mel Gorman42288fe2012-12-21 23:10:25 +00002336static void sp_node_init(struct sp_node *node, unsigned long start,
2337 unsigned long end, struct mempolicy *pol)
2338{
2339 node->start = start;
2340 node->end = end;
2341 node->policy = pol;
2342}
2343
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07002344static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2345 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346{
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002347 struct sp_node *n;
2348 struct mempolicy *newpol;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002350 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 if (!n)
2352 return NULL;
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002353
2354 newpol = mpol_dup(pol);
2355 if (IS_ERR(newpol)) {
2356 kmem_cache_free(sn_cache, n);
2357 return NULL;
2358 }
2359 newpol->flags |= MPOL_F_SHARED;
Mel Gorman42288fe2012-12-21 23:10:25 +00002360 sp_node_init(n, start, end, newpol);
KOSAKI Motohiro869833f2012-10-08 16:29:16 -07002361
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 return n;
2363}
2364
2365/* Replace a policy range. */
2366static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2367 unsigned long end, struct sp_node *new)
2368{
Mel Gormanb22d1272012-10-08 16:29:17 -07002369 struct sp_node *n;
Mel Gorman42288fe2012-12-21 23:10:25 +00002370 struct sp_node *n_new = NULL;
2371 struct mempolicy *mpol_new = NULL;
Mel Gormanb22d1272012-10-08 16:29:17 -07002372 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373
Mel Gorman42288fe2012-12-21 23:10:25 +00002374restart:
2375 spin_lock(&sp->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 n = sp_lookup(sp, start, end);
2377 /* Take care of old policies in the same range. */
2378 while (n && n->start < end) {
2379 struct rb_node *next = rb_next(&n->nd);
2380 if (n->start >= start) {
2381 if (n->end <= end)
2382 sp_delete(sp, n);
2383 else
2384 n->start = end;
2385 } else {
2386 /* Old policy spanning whole new range. */
2387 if (n->end > end) {
Mel Gorman42288fe2012-12-21 23:10:25 +00002388 if (!n_new)
2389 goto alloc_new;
2390
2391 *mpol_new = *n->policy;
2392 atomic_set(&mpol_new->refcnt, 1);
KOSAKI Motohiro78806392013-03-08 12:43:29 -08002393 sp_node_init(n_new, end, n->end, mpol_new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 n->end = start;
Hillf Danton5ca39572013-03-08 12:43:28 -08002395 sp_insert(sp, n_new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002396 n_new = NULL;
2397 mpol_new = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 break;
2399 } else
2400 n->end = start;
2401 }
2402 if (!next)
2403 break;
2404 n = rb_entry(next, struct sp_node, nd);
2405 }
2406 if (new)
2407 sp_insert(sp, new);
Mel Gorman42288fe2012-12-21 23:10:25 +00002408 spin_unlock(&sp->lock);
2409 ret = 0;
2410
2411err_out:
2412 if (mpol_new)
2413 mpol_put(mpol_new);
2414 if (n_new)
2415 kmem_cache_free(sn_cache, n_new);
2416
Mel Gormanb22d1272012-10-08 16:29:17 -07002417 return ret;
Mel Gorman42288fe2012-12-21 23:10:25 +00002418
2419alloc_new:
2420 spin_unlock(&sp->lock);
2421 ret = -ENOMEM;
2422 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2423 if (!n_new)
2424 goto err_out;
2425 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2426 if (!mpol_new)
2427 goto err_out;
2428 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429}
2430
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002431/**
2432 * mpol_shared_policy_init - initialize shared policy for inode
2433 * @sp: pointer to inode shared policy
2434 * @mpol: struct mempolicy to install
2435 *
2436 * Install non-NULL @mpol in inode's shared policy rb-tree.
2437 * On entry, the current task has a reference on a non-NULL @mpol.
2438 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002439 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002440 */
2441void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08002442{
Miao Xie58568d22009-06-16 15:31:49 -07002443 int ret;
2444
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002445 sp->root = RB_ROOT; /* empty tree == default mempolicy */
Mel Gorman42288fe2012-12-21 23:10:25 +00002446 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08002447
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002448 if (mpol) {
2449 struct vm_area_struct pvma;
2450 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002451 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002452
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002453 if (!scratch)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002454 goto put_mpol;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002455 /* contextualize the tmpfs mount point mempolicy */
2456 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002457 if (IS_ERR(new))
Dan Carpenter0cae3452010-05-25 23:42:58 -07002458 goto free_scratch; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002459
2460 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002461 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002462 task_unlock(current);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002463 if (ret)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002464 goto put_new;
Robin Holt7339ff82006-01-14 13:20:48 -08002465
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002466 /* Create pseudo-vma that contains just the policy */
2467 memset(&pvma, 0, sizeof(struct vm_area_struct));
2468 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2469 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002470
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002471put_new:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002472 mpol_put(new); /* drop initial ref */
Dan Carpenter0cae3452010-05-25 23:42:58 -07002473free_scratch:
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002474 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002475put_mpol:
2476 mpol_put(mpol); /* drop our incoming ref on sb mpol */
Robin Holt7339ff82006-01-14 13:20:48 -08002477 }
2478}
2479
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480int mpol_set_shared_policy(struct shared_policy *info,
2481 struct vm_area_struct *vma, struct mempolicy *npol)
2482{
2483 int err;
2484 struct sp_node *new = NULL;
2485 unsigned long sz = vma_pages(vma);
2486
David Rientjes028fec42008-04-28 02:12:25 -07002487 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002489 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002490 npol ? npol->flags : -1,
David Rientjes00ef2d22013-02-22 16:35:36 -08002491 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492
2493 if (npol) {
2494 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2495 if (!new)
2496 return -ENOMEM;
2497 }
2498 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2499 if (err && new)
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002500 sp_free(new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 return err;
2502}
2503
2504/* Free a backing policy store on inode delete. */
2505void mpol_free_shared_policy(struct shared_policy *p)
2506{
2507 struct sp_node *n;
2508 struct rb_node *next;
2509
2510 if (!p->root.rb_node)
2511 return;
Mel Gorman42288fe2012-12-21 23:10:25 +00002512 spin_lock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513 next = rb_first(&p->root);
2514 while (next) {
2515 n = rb_entry(next, struct sp_node, nd);
2516 next = rb_next(&n->nd);
KOSAKI Motohiro63f74ca2012-10-08 16:29:19 -07002517 sp_delete(p, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 }
Mel Gorman42288fe2012-12-21 23:10:25 +00002519 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520}
2521
Mel Gorman1a687c22012-11-22 11:16:36 +00002522#ifdef CONFIG_NUMA_BALANCING
2523static bool __initdata numabalancing_override;
2524
2525static void __init check_numabalancing_enable(void)
2526{
2527 bool numabalancing_default = false;
2528
2529 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2530 numabalancing_default = true;
2531
2532 if (nr_node_ids > 1 && !numabalancing_override) {
2533 printk(KERN_INFO "Enabling automatic NUMA balancing. "
2534 "Configure with numa_balancing= or sysctl");
2535 set_numabalancing_state(numabalancing_default);
2536 }
2537}
2538
2539static int __init setup_numabalancing(char *str)
2540{
2541 int ret = 0;
2542 if (!str)
2543 goto out;
2544 numabalancing_override = true;
2545
2546 if (!strcmp(str, "enable")) {
2547 set_numabalancing_state(true);
2548 ret = 1;
2549 } else if (!strcmp(str, "disable")) {
2550 set_numabalancing_state(false);
2551 ret = 1;
2552 }
2553out:
2554 if (!ret)
2555 printk(KERN_WARNING "Unable to parse numa_balancing=\n");
2556
2557 return ret;
2558}
2559__setup("numa_balancing=", setup_numabalancing);
2560#else
2561static inline void __init check_numabalancing_enable(void)
2562{
2563}
2564#endif /* CONFIG_NUMA_BALANCING */
2565
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566/* assumes fs == KERNEL_DS */
2567void __init numa_policy_init(void)
2568{
Paul Mundtb71636e2007-07-15 23:38:15 -07002569 nodemask_t interleave_nodes;
2570 unsigned long largest = 0;
2571 int nid, prefer = 0;
2572
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 policy_cache = kmem_cache_create("numa_policy",
2574 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002575 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576
2577 sn_cache = kmem_cache_create("shared_policy_node",
2578 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002579 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580
Mel Gorman5606e382012-11-02 18:19:13 +00002581 for_each_node(nid) {
2582 preferred_node_policy[nid] = (struct mempolicy) {
2583 .refcnt = ATOMIC_INIT(1),
2584 .mode = MPOL_PREFERRED,
2585 .flags = MPOL_F_MOF | MPOL_F_MORON,
2586 .v = { .preferred_node = nid, },
2587 };
2588 }
2589
Paul Mundtb71636e2007-07-15 23:38:15 -07002590 /*
2591 * Set interleaving policy for system init. Interleaving is only
2592 * enabled across suitably sized nodes (default is >= 16MB), or
2593 * fall back to the largest node if they're all smaller.
2594 */
2595 nodes_clear(interleave_nodes);
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002596 for_each_node_state(nid, N_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002597 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598
Paul Mundtb71636e2007-07-15 23:38:15 -07002599 /* Preserve the largest node */
2600 if (largest < total_pages) {
2601 largest = total_pages;
2602 prefer = nid;
2603 }
2604
2605 /* Interleave this node? */
2606 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2607 node_set(nid, interleave_nodes);
2608 }
2609
2610 /* All too small, use the largest */
2611 if (unlikely(nodes_empty(interleave_nodes)))
2612 node_set(prefer, interleave_nodes);
2613
David Rientjes028fec42008-04-28 02:12:25 -07002614 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615 printk("numa_policy_init: interleaving failed\n");
Mel Gorman1a687c22012-11-22 11:16:36 +00002616
2617 check_numabalancing_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618}
2619
Christoph Lameter8bccd852005-10-29 18:16:59 -07002620/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621void numa_default_policy(void)
2622{
David Rientjes028fec42008-04-28 02:12:25 -07002623 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624}
Paul Jackson68860ec2005-10-30 15:02:36 -08002625
Paul Jackson42253992006-01-08 01:01:59 -08002626/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002627 * Parse and format mempolicy from/to strings
2628 */
2629
2630/*
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002631 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002632 */
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002633static const char * const policy_modes[] =
2634{
2635 [MPOL_DEFAULT] = "default",
2636 [MPOL_PREFERRED] = "prefer",
2637 [MPOL_BIND] = "bind",
2638 [MPOL_INTERLEAVE] = "interleave",
Lee Schermerhornd3a71032012-10-25 14:16:29 +02002639 [MPOL_LOCAL] = "local",
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002640};
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002641
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002642
2643#ifdef CONFIG_TMPFS
2644/**
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002645 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002646 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002647 * @mpol: pointer to struct mempolicy pointer, returned on success.
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002648 *
2649 * Format of input:
2650 * <mode>[=<flags>][:<nodelist>]
2651 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002652 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002653 */
Hugh Dickinsa7a88b22013-01-02 02:04:23 -08002654int mpol_parse_str(char *str, struct mempolicy **mpol)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002655{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002656 struct mempolicy *new = NULL;
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002657 unsigned short mode;
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002658 unsigned short mode_flags;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002659 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002660 char *nodelist = strchr(str, ':');
2661 char *flags = strchr(str, '=');
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002662 int err = 1;
2663
2664 if (nodelist) {
2665 /* NUL-terminate mode or flags string */
2666 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002667 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002668 goto out;
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002669 if (!nodes_subset(nodes, node_states[N_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002670 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002671 } else
2672 nodes_clear(nodes);
2673
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002674 if (flags)
2675 *flags++ = '\0'; /* terminate mode string */
2676
Peter Zijlstra479e2802012-10-25 14:16:28 +02002677 for (mode = 0; mode < MPOL_MAX; mode++) {
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002678 if (!strcmp(str, policy_modes[mode])) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002679 break;
2680 }
2681 }
Mel Gormana7200942012-11-16 09:37:58 +00002682 if (mode >= MPOL_MAX)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002683 goto out;
2684
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002685 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002686 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002687 /*
2688 * Insist on a nodelist of one node only
2689 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002690 if (nodelist) {
2691 char *rest = nodelist;
2692 while (isdigit(*rest))
2693 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002694 if (*rest)
2695 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002696 }
2697 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002698 case MPOL_INTERLEAVE:
2699 /*
2700 * Default to online nodes with memory if no nodelist
2701 */
2702 if (!nodelist)
Lai Jiangshan01f13bd2012-12-12 13:51:33 -08002703 nodes = node_states[N_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002704 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002705 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002706 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002707 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002708 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002709 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002710 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002711 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002712 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002713 case MPOL_DEFAULT:
2714 /*
2715 * Insist on a empty nodelist
2716 */
2717 if (!nodelist)
2718 err = 0;
2719 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002720 case MPOL_BIND:
2721 /*
2722 * Insist on a nodelist
2723 */
2724 if (!nodelist)
2725 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002726 }
2727
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002728 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002729 if (flags) {
2730 /*
2731 * Currently, we only support two mutually exclusive
2732 * mode flags.
2733 */
2734 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002735 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002736 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002737 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002738 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002739 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002740 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002741
2742 new = mpol_new(mode, mode_flags, &nodes);
2743 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002744 goto out;
2745
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002746 /*
2747 * Save nodes for mpol_to_str() to show the tmpfs mount options
2748 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2749 */
2750 if (mode != MPOL_PREFERRED)
2751 new->v.nodes = nodes;
2752 else if (nodelist)
2753 new->v.preferred_node = first_node(nodes);
2754 else
2755 new->flags |= MPOL_F_LOCAL;
2756
2757 /*
2758 * Save nodes for contextualization: this will be used to "clone"
2759 * the mempolicy in a specific context [cpuset] at a later time.
2760 */
2761 new->w.user_nodemask = nodes;
2762
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002763 err = 0;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002764
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002765out:
2766 /* Restore string for error message */
2767 if (nodelist)
2768 *--nodelist = ':';
2769 if (flags)
2770 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002771 if (!err)
2772 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002773 return err;
2774}
2775#endif /* CONFIG_TMPFS */
2776
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002777/**
2778 * mpol_to_str - format a mempolicy structure for printing
2779 * @buffer: to contain formatted mempolicy string
2780 * @maxlen: length of @buffer
2781 * @pol: pointer to mempolicy to be formatted
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002782 *
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002783 * Convert a mempolicy into a string.
2784 * Returns the number of characters in buffer (if positive)
2785 * or an error (negative)
2786 */
Hugh Dickinsa7a88b22013-01-02 02:04:23 -08002787int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002788{
2789 char *p = buffer;
2790 int l;
2791 nodemask_t nodes;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002792 unsigned short mode;
David Rientjesf5b087b2008-04-28 02:12:27 -07002793 unsigned short flags = pol ? pol->flags : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002794
Lee Schermerhorn22919902008-04-28 02:13:22 -07002795 /*
2796 * Sanity check: room for longest mode, flag and some nodes
2797 */
2798 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2799
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002800 if (!pol || pol == &default_policy)
2801 mode = MPOL_DEFAULT;
2802 else
2803 mode = pol->mode;
2804
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002805 switch (mode) {
2806 case MPOL_DEFAULT:
2807 nodes_clear(nodes);
2808 break;
2809
2810 case MPOL_PREFERRED:
2811 nodes_clear(nodes);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002812 if (flags & MPOL_F_LOCAL)
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002813 mode = MPOL_LOCAL;
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002814 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002815 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002816 break;
2817
2818 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07002819 /* Fall through */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002820 case MPOL_INTERLEAVE:
Hugh Dickinsf2a07f42013-01-02 02:01:33 -08002821 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002822 break;
2823
2824 default:
Dave Jones80de7c32012-09-06 12:01:00 -04002825 return -EINVAL;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002826 }
2827
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002828 l = strlen(policy_modes[mode]);
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002829 if (buffer + maxlen < p + l + 1)
2830 return -ENOSPC;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002831
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002832 strcpy(p, policy_modes[mode]);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002833 p += l;
2834
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002835 if (flags & MPOL_MODE_FLAGS) {
David Rientjesf5b087b2008-04-28 02:12:27 -07002836 if (buffer + maxlen < p + 2)
2837 return -ENOSPC;
2838 *p++ = '=';
2839
Lee Schermerhorn22919902008-04-28 02:13:22 -07002840 /*
2841 * Currently, the only defined flags are mutually exclusive
2842 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002843 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002844 p += snprintf(p, buffer + maxlen - p, "static");
2845 else if (flags & MPOL_F_RELATIVE_NODES)
2846 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002847 }
2848
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002849 if (!nodes_empty(nodes)) {
2850 if (buffer + maxlen < p + 2)
2851 return -ENOSPC;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002852 *p++ = ':';
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002853 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2854 }
2855 return p - buffer;
2856}