blob: c3fdbcb17658ce405131e5b0310e1857fd6558bc [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.
29 * As a special case node -1 here means do the allocation
30 * 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>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080093
Linus Torvalds1da177e2005-04-16 15:20:36 -070094#include <asm/tlbflush.h>
95#include <asm/uaccess.h>
Michal Hocko778d3b02011-07-26 16:08:30 -070096#include <linux/random.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Nick Piggin62695a82008-10-18 20:26:09 -070098#include "internal.h"
99
Christoph Lameter38e35862006-01-08 01:01:01 -0800100/* Internal flags */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800101#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
Christoph Lameter38e35862006-01-08 01:01:01 -0800102#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800103
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800104static struct kmem_cache *policy_cache;
105static struct kmem_cache *sn_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107/* Highest zone. An specific allocation for a zone below that is not
108 policied. */
Christoph Lameter62672762007-02-10 01:43:07 -0800109enum zone_type policy_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700111/*
112 * run-time system-wide default policy => local allocation
113 */
H Hartley Sweetene754d792011-10-31 17:09:23 -0700114static struct mempolicy default_policy = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 .refcnt = ATOMIC_INIT(1), /* never free it */
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700116 .mode = MPOL_PREFERRED,
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700117 .flags = MPOL_F_LOCAL,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118};
119
David Rientjes37012942008-04-28 02:12:33 -0700120static const struct mempolicy_operations {
121 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
Miao Xie708c1bb2010-05-24 14:32:07 -0700122 /*
123 * If read-side task has no lock to protect task->mempolicy, write-side
124 * task will rebind the task->mempolicy by two step. The first step is
125 * setting all the newly nodes, and the second step is cleaning all the
126 * disallowed nodes. In this way, we can avoid finding no node to alloc
127 * page.
128 * If we have a lock to protect task->mempolicy in read-side, we do
129 * rebind directly.
130 *
131 * step:
132 * MPOL_REBIND_ONCE - do rebind work at once
133 * MPOL_REBIND_STEP1 - set all the newly nodes
134 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
135 */
136 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
137 enum mpol_rebind_step step);
David Rientjes37012942008-04-28 02:12:33 -0700138} mpol_ops[MPOL_MAX];
139
Mel Gorman19770b32008-04-28 02:12:18 -0700140/* Check that the nodemask contains at least one populated zone */
David Rientjes37012942008-04-28 02:12:33 -0700141static int is_valid_nodemask(const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142{
Mel Gorman19770b32008-04-28 02:12:18 -0700143 int nd, k;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
Mel Gorman19770b32008-04-28 02:12:18 -0700145 for_each_node_mask(nd, *nodemask) {
146 struct zone *z;
147
148 for (k = 0; k <= policy_zone; k++) {
149 z = &NODE_DATA(nd)->node_zones[k];
150 if (z->present_pages > 0)
151 return 1;
Andi Kleendd942ae2006-02-17 01:39:16 +0100152 }
153 }
Mel Gorman19770b32008-04-28 02:12:18 -0700154
155 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156}
157
David Rientjesf5b087b2008-04-28 02:12:27 -0700158static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
159{
Bob Liu6d556292010-05-24 14:31:59 -0700160 return pol->flags & MPOL_MODE_FLAGS;
David Rientjes4c50bc02008-04-28 02:12:30 -0700161}
162
163static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
164 const nodemask_t *rel)
165{
166 nodemask_t tmp;
167 nodes_fold(tmp, *orig, nodes_weight(*rel));
168 nodes_onto(*ret, tmp, *rel);
David Rientjesf5b087b2008-04-28 02:12:27 -0700169}
170
David Rientjes37012942008-04-28 02:12:33 -0700171static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
172{
173 if (nodes_empty(*nodes))
174 return -EINVAL;
175 pol->v.nodes = *nodes;
176 return 0;
177}
178
179static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
180{
181 if (!nodes)
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700182 pol->flags |= MPOL_F_LOCAL; /* local allocation */
David Rientjes37012942008-04-28 02:12:33 -0700183 else if (nodes_empty(*nodes))
184 return -EINVAL; /* no allowed nodes */
185 else
186 pol->v.preferred_node = first_node(*nodes);
187 return 0;
188}
189
190static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
191{
192 if (!is_valid_nodemask(nodes))
193 return -EINVAL;
194 pol->v.nodes = *nodes;
195 return 0;
196}
197
Miao Xie58568d22009-06-16 15:31:49 -0700198/*
199 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
200 * any, for the new policy. mpol_new() has already validated the nodes
201 * parameter with respect to the policy mode and flags. But, we need to
202 * handle an empty nodemask with MPOL_PREFERRED here.
203 *
204 * Must be called holding task's alloc_lock to protect task's mems_allowed
205 * and mempolicy. May also be called holding the mmap_semaphore for write.
206 */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700207static int mpol_set_nodemask(struct mempolicy *pol,
208 const nodemask_t *nodes, struct nodemask_scratch *nsc)
Miao Xie58568d22009-06-16 15:31:49 -0700209{
Miao Xie58568d22009-06-16 15:31:49 -0700210 int ret;
211
212 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
213 if (pol == NULL)
214 return 0;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700215 /* Check N_HIGH_MEMORY */
216 nodes_and(nsc->mask1,
217 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
Miao Xie58568d22009-06-16 15:31:49 -0700218
219 VM_BUG_ON(!nodes);
220 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
221 nodes = NULL; /* explicit local allocation */
222 else {
223 if (pol->flags & MPOL_F_RELATIVE_NODES)
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700224 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
Miao Xie58568d22009-06-16 15:31:49 -0700225 else
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700226 nodes_and(nsc->mask2, *nodes, nsc->mask1);
227
Miao Xie58568d22009-06-16 15:31:49 -0700228 if (mpol_store_user_nodemask(pol))
229 pol->w.user_nodemask = *nodes;
230 else
231 pol->w.cpuset_mems_allowed =
232 cpuset_current_mems_allowed;
233 }
234
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700235 if (nodes)
236 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
237 else
238 ret = mpol_ops[pol->mode].create(pol, NULL);
Miao Xie58568d22009-06-16 15:31:49 -0700239 return ret;
240}
241
242/*
243 * This function just creates a new policy, does some check and simple
244 * initialization. You must invoke mpol_set_nodemask() to set nodes.
245 */
David Rientjes028fec42008-04-28 02:12:25 -0700246static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
247 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248{
249 struct mempolicy *policy;
250
David Rientjes028fec42008-04-28 02:12:25 -0700251 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
252 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
Paul Mundt140d5a42007-07-15 23:38:16 -0700253
David Rientjes3e1f0642008-04-28 02:12:34 -0700254 if (mode == MPOL_DEFAULT) {
255 if (nodes && !nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700256 return ERR_PTR(-EINVAL);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700257 return NULL; /* simply delete any existing policy */
David Rientjes37012942008-04-28 02:12:33 -0700258 }
David Rientjes3e1f0642008-04-28 02:12:34 -0700259 VM_BUG_ON(!nodes);
260
261 /*
262 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
263 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
264 * All other modes require a valid pointer to a non-empty nodemask.
265 */
266 if (mode == MPOL_PREFERRED) {
267 if (nodes_empty(*nodes)) {
268 if (((flags & MPOL_F_STATIC_NODES) ||
269 (flags & MPOL_F_RELATIVE_NODES)))
270 return ERR_PTR(-EINVAL);
David Rientjes3e1f0642008-04-28 02:12:34 -0700271 }
272 } else if (nodes_empty(*nodes))
273 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
275 if (!policy)
276 return ERR_PTR(-ENOMEM);
277 atomic_set(&policy->refcnt, 1);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700278 policy->mode = mode;
David Rientjes3e1f0642008-04-28 02:12:34 -0700279 policy->flags = flags;
David Rientjesf5b087b2008-04-28 02:12:27 -0700280
David Rientjes37012942008-04-28 02:12:33 -0700281 return policy;
282}
283
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700284/* Slow path of a mpol destructor. */
285void __mpol_put(struct mempolicy *p)
286{
287 if (!atomic_dec_and_test(&p->refcnt))
288 return;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700289 kmem_cache_free(policy_cache, p);
290}
291
Miao Xie708c1bb2010-05-24 14:32:07 -0700292static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
293 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700294{
295}
296
Miao Xie708c1bb2010-05-24 14:32:07 -0700297/*
298 * step:
299 * MPOL_REBIND_ONCE - do rebind work at once
300 * MPOL_REBIND_STEP1 - set all the newly nodes
301 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
302 */
303static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700305{
306 nodemask_t tmp;
307
308 if (pol->flags & MPOL_F_STATIC_NODES)
309 nodes_and(tmp, pol->w.user_nodemask, *nodes);
310 else if (pol->flags & MPOL_F_RELATIVE_NODES)
311 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
312 else {
Miao Xie708c1bb2010-05-24 14:32:07 -0700313 /*
314 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
315 * result
316 */
317 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
318 nodes_remap(tmp, pol->v.nodes,
319 pol->w.cpuset_mems_allowed, *nodes);
320 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
321 } else if (step == MPOL_REBIND_STEP2) {
322 tmp = pol->w.cpuset_mems_allowed;
323 pol->w.cpuset_mems_allowed = *nodes;
324 } else
325 BUG();
David Rientjes37012942008-04-28 02:12:33 -0700326 }
327
Miao Xie708c1bb2010-05-24 14:32:07 -0700328 if (nodes_empty(tmp))
329 tmp = *nodes;
330
331 if (step == MPOL_REBIND_STEP1)
332 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
333 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
334 pol->v.nodes = tmp;
335 else
336 BUG();
337
David Rientjes37012942008-04-28 02:12:33 -0700338 if (!node_isset(current->il_next, tmp)) {
339 current->il_next = next_node(current->il_next, tmp);
340 if (current->il_next >= MAX_NUMNODES)
341 current->il_next = first_node(tmp);
342 if (current->il_next >= MAX_NUMNODES)
343 current->il_next = numa_node_id();
344 }
345}
346
347static void mpol_rebind_preferred(struct mempolicy *pol,
Miao Xie708c1bb2010-05-24 14:32:07 -0700348 const nodemask_t *nodes,
349 enum mpol_rebind_step step)
David Rientjes37012942008-04-28 02:12:33 -0700350{
351 nodemask_t tmp;
352
David Rientjes37012942008-04-28 02:12:33 -0700353 if (pol->flags & MPOL_F_STATIC_NODES) {
354 int node = first_node(pol->w.user_nodemask);
355
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700356 if (node_isset(node, *nodes)) {
David Rientjes37012942008-04-28 02:12:33 -0700357 pol->v.preferred_node = node;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700358 pol->flags &= ~MPOL_F_LOCAL;
359 } else
360 pol->flags |= MPOL_F_LOCAL;
David Rientjes37012942008-04-28 02:12:33 -0700361 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
362 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
363 pol->v.preferred_node = first_node(tmp);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700364 } else if (!(pol->flags & MPOL_F_LOCAL)) {
David Rientjes37012942008-04-28 02:12:33 -0700365 pol->v.preferred_node = node_remap(pol->v.preferred_node,
366 pol->w.cpuset_mems_allowed,
367 *nodes);
368 pol->w.cpuset_mems_allowed = *nodes;
369 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370}
371
Miao Xie708c1bb2010-05-24 14:32:07 -0700372/*
373 * mpol_rebind_policy - Migrate a policy to a different set of nodes
374 *
375 * If read-side task has no lock to protect task->mempolicy, write-side
376 * task will rebind the task->mempolicy by two step. The first step is
377 * setting all the newly nodes, and the second step is cleaning all the
378 * disallowed nodes. In this way, we can avoid finding no node to alloc
379 * page.
380 * If we have a lock to protect task->mempolicy in read-side, we do
381 * rebind directly.
382 *
383 * step:
384 * MPOL_REBIND_ONCE - do rebind work at once
385 * MPOL_REBIND_STEP1 - set all the newly nodes
386 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
387 */
388static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
389 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700390{
David Rientjes1d0d2682008-04-28 02:12:32 -0700391 if (!pol)
392 return;
Miao Xie708c1bb2010-05-24 14:32:07 -0700393 if (!mpol_store_user_nodemask(pol) && step == 0 &&
David Rientjes1d0d2682008-04-28 02:12:32 -0700394 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
395 return;
Miao Xie708c1bb2010-05-24 14:32:07 -0700396
397 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
398 return;
399
400 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
401 BUG();
402
403 if (step == MPOL_REBIND_STEP1)
404 pol->flags |= MPOL_F_REBINDING;
405 else if (step == MPOL_REBIND_STEP2)
406 pol->flags &= ~MPOL_F_REBINDING;
407 else if (step >= MPOL_REBIND_NSTEP)
408 BUG();
409
410 mpol_ops[pol->mode].rebind(pol, newmask, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700411}
412
413/*
414 * Wrapper for mpol_rebind_policy() that just requires task
415 * pointer, and updates task mempolicy.
Miao Xie58568d22009-06-16 15:31:49 -0700416 *
417 * Called with task's alloc_lock held.
David Rientjes1d0d2682008-04-28 02:12:32 -0700418 */
419
Miao Xie708c1bb2010-05-24 14:32:07 -0700420void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
421 enum mpol_rebind_step step)
David Rientjes1d0d2682008-04-28 02:12:32 -0700422{
Miao Xie708c1bb2010-05-24 14:32:07 -0700423 mpol_rebind_policy(tsk->mempolicy, new, step);
David Rientjes1d0d2682008-04-28 02:12:32 -0700424}
425
426/*
427 * Rebind each vma in mm to new nodemask.
428 *
429 * Call holding a reference to mm. Takes mm->mmap_sem during call.
430 */
431
432void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
433{
434 struct vm_area_struct *vma;
435
436 down_write(&mm->mmap_sem);
437 for (vma = mm->mmap; vma; vma = vma->vm_next)
Miao Xie708c1bb2010-05-24 14:32:07 -0700438 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
David Rientjes1d0d2682008-04-28 02:12:32 -0700439 up_write(&mm->mmap_sem);
440}
441
David Rientjes37012942008-04-28 02:12:33 -0700442static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
443 [MPOL_DEFAULT] = {
444 .rebind = mpol_rebind_default,
445 },
446 [MPOL_INTERLEAVE] = {
447 .create = mpol_new_interleave,
448 .rebind = mpol_rebind_nodemask,
449 },
450 [MPOL_PREFERRED] = {
451 .create = mpol_new_preferred,
452 .rebind = mpol_rebind_preferred,
453 },
454 [MPOL_BIND] = {
455 .create = mpol_new_bind,
456 .rebind = mpol_rebind_nodemask,
457 },
458};
459
Christoph Lameterfc301282006-01-18 17:42:29 -0800460static void migrate_page_add(struct page *page, struct list_head *pagelist,
461 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800462
Christoph Lameter38e35862006-01-08 01:01:01 -0800463/* Scan through pages checking if pages follow certain conditions. */
Nick Pigginb5810032005-10-29 18:16:12 -0700464static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800465 unsigned long addr, unsigned long end,
466 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800467 void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468{
Hugh Dickins91612e02005-06-21 17:15:07 -0700469 pte_t *orig_pte;
470 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700471 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700472
Hugh Dickins705e87c2005-10-29 18:16:27 -0700473 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700474 do {
Linus Torvalds6aab3412005-11-28 14:34:23 -0800475 struct page *page;
Andy Whitcroft25ba77c2006-12-06 20:33:03 -0800476 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700477
478 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800480 page = vm_normal_page(vma, addr, *pte);
481 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800483 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800484 * vm_normal_page() filters out zero pages, but there might
485 * still be PageReserved pages to skip, perhaps in a VDSO.
486 * And we cannot move PageKsm pages sensibly or safely yet.
Nick Piggin053837f2006-01-18 17:42:27 -0800487 */
Hugh Dickins62b61f62009-12-14 17:59:33 -0800488 if (PageReserved(page) || PageKsm(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800489 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800490 nid = page_to_nid(page);
Christoph Lameter38e35862006-01-08 01:01:01 -0800491 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
492 continue;
493
Stephen Wilsonb1f72d12011-05-24 17:12:43 -0700494 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameterfc301282006-01-18 17:42:29 -0800495 migrate_page_add(page, private, flags);
Christoph Lameter38e35862006-01-08 01:01:01 -0800496 else
497 break;
Hugh Dickins91612e02005-06-21 17:15:07 -0700498 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickins705e87c2005-10-29 18:16:27 -0700499 pte_unmap_unlock(orig_pte, ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700500 return addr != end;
501}
502
Nick Pigginb5810032005-10-29 18:16:12 -0700503static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800504 unsigned long addr, unsigned long end,
505 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800506 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700507{
508 pmd_t *pmd;
509 unsigned long next;
510
511 pmd = pmd_offset(pud, addr);
512 do {
513 next = pmd_addr_end(addr, end);
Andrea Arcangelibae9c192011-01-13 15:46:46 -0800514 split_huge_page_pmd(vma->vm_mm, pmd);
Hugh Dickins91612e02005-06-21 17:15:07 -0700515 if (pmd_none_or_clear_bad(pmd))
516 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800517 if (check_pte_range(vma, pmd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800518 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700519 return -EIO;
520 } while (pmd++, addr = next, addr != end);
521 return 0;
522}
523
Nick Pigginb5810032005-10-29 18:16:12 -0700524static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800525 unsigned long addr, unsigned long end,
526 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800527 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700528{
529 pud_t *pud;
530 unsigned long next;
531
532 pud = pud_offset(pgd, addr);
533 do {
534 next = pud_addr_end(addr, end);
535 if (pud_none_or_clear_bad(pud))
536 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800537 if (check_pmd_range(vma, pud, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800538 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700539 return -EIO;
540 } while (pud++, addr = next, addr != end);
541 return 0;
542}
543
Nick Pigginb5810032005-10-29 18:16:12 -0700544static inline int check_pgd_range(struct vm_area_struct *vma,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800545 unsigned long addr, unsigned long end,
546 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800547 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700548{
549 pgd_t *pgd;
550 unsigned long next;
551
Nick Pigginb5810032005-10-29 18:16:12 -0700552 pgd = pgd_offset(vma->vm_mm, addr);
Hugh Dickins91612e02005-06-21 17:15:07 -0700553 do {
554 next = pgd_addr_end(addr, end);
555 if (pgd_none_or_clear_bad(pgd))
556 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800557 if (check_pud_range(vma, pgd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800558 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700559 return -EIO;
560 } while (pgd++, addr = next, addr != end);
561 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562}
563
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800564/*
565 * Check if all pages in a range are on a set of nodes.
566 * If pagelist != NULL then isolate pages from the LRU and
567 * put them on the pagelist.
568 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569static struct vm_area_struct *
570check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Christoph Lameter38e35862006-01-08 01:01:01 -0800571 const nodemask_t *nodes, unsigned long flags, void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572{
573 int err;
574 struct vm_area_struct *first, *vma, *prev;
575
Nick Piggin053837f2006-01-18 17:42:27 -0800576
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 first = find_vma(mm, start);
578 if (!first)
579 return ERR_PTR(-EFAULT);
580 prev = NULL;
581 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800582 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
583 if (!vma->vm_next && vma->vm_end < end)
584 return ERR_PTR(-EFAULT);
585 if (prev && prev->vm_end < vma->vm_start)
586 return ERR_PTR(-EFAULT);
587 }
588 if (!is_vm_hugetlb_page(vma) &&
589 ((flags & MPOL_MF_STRICT) ||
590 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
591 vma_migratable(vma)))) {
Andi Kleen5b952b32005-09-13 01:25:08 -0700592 unsigned long endvma = vma->vm_end;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800593
Andi Kleen5b952b32005-09-13 01:25:08 -0700594 if (endvma > end)
595 endvma = end;
596 if (vma->vm_start > start)
597 start = vma->vm_start;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800598 err = check_pgd_range(vma, start, endvma, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800599 flags, private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 if (err) {
601 first = ERR_PTR(err);
602 break;
603 }
604 }
605 prev = vma;
606 }
607 return first;
608}
609
610/* Apply policy to a single VMA */
611static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
612{
613 int err = 0;
614 struct mempolicy *old = vma->vm_policy;
615
Paul Mundt140d5a42007-07-15 23:38:16 -0700616 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 vma->vm_start, vma->vm_end, vma->vm_pgoff,
618 vma->vm_ops, vma->vm_file,
619 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
620
621 if (vma->vm_ops && vma->vm_ops->set_policy)
622 err = vma->vm_ops->set_policy(vma, new);
623 if (!err) {
624 mpol_get(new);
625 vma->vm_policy = new;
Lee Schermerhornf0be3d32008-04-28 02:13:08 -0700626 mpol_put(old);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 }
628 return err;
629}
630
631/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800632static int mbind_range(struct mm_struct *mm, unsigned long start,
633 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634{
635 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800636 struct vm_area_struct *prev;
637 struct vm_area_struct *vma;
638 int err = 0;
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800639 pgoff_t pgoff;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800640 unsigned long vmstart;
641 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800643 vma = find_vma_prev(mm, start, &prev);
644 if (!vma || vma->vm_start > start)
645 return -EFAULT;
646
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800647 if (start > vma->vm_start)
648 prev = vma;
649
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800650 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800652 vmstart = max(start, vma->vm_start);
653 vmend = min(end, vma->vm_end);
654
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800655 if (mpol_equal(vma_policy(vma), new_pol))
656 continue;
657
658 pgoff = vma->vm_pgoff +
659 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800660 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
KOSAKI Motohiroe26a5112011-12-28 15:57:11 -0800661 vma->anon_vma, vma->vm_file, pgoff,
Caspar Zhang8aacc9f2011-09-14 16:20:58 -0700662 new_pol);
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800663 if (prev) {
664 vma = prev;
665 next = vma->vm_next;
666 continue;
667 }
668 if (vma->vm_start != vmstart) {
669 err = split_vma(vma->vm_mm, vma, vmstart, 1);
670 if (err)
671 goto out;
672 }
673 if (vma->vm_end != vmend) {
674 err = split_vma(vma->vm_mm, vma, vmend, 0);
675 if (err)
676 goto out;
677 }
678 err = policy_vma(vma, new_pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 if (err)
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800680 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800682
683 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 return err;
685}
686
Paul Jacksonc61afb12006-03-24 03:16:08 -0800687/*
688 * Update task->flags PF_MEMPOLICY bit: set iff non-default
689 * mempolicy. Allows more rapid checking of this (combined perhaps
690 * with other PF_* flag bits) on memory allocation hot code paths.
691 *
692 * If called from outside this file, the task 'p' should -only- be
693 * a newly forked child not yet visible on the task list, because
694 * manipulating the task flags of a visible task is not safe.
695 *
696 * The above limitation is why this routine has the funny name
697 * mpol_fix_fork_child_flag().
698 *
699 * It is also safe to call this with a task pointer of current,
700 * which the static wrapper mpol_set_task_struct_flag() does,
701 * for use within this file.
702 */
703
704void mpol_fix_fork_child_flag(struct task_struct *p)
705{
706 if (p->mempolicy)
707 p->flags |= PF_MEMPOLICY;
708 else
709 p->flags &= ~PF_MEMPOLICY;
710}
711
712static void mpol_set_task_struct_flag(void)
713{
714 mpol_fix_fork_child_flag(current);
715}
716
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700718static long do_set_mempolicy(unsigned short mode, unsigned short flags,
719 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720{
Miao Xie58568d22009-06-16 15:31:49 -0700721 struct mempolicy *new, *old;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700722 struct mm_struct *mm = current->mm;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700723 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700724 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700726 if (!scratch)
727 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700728
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700729 new = mpol_new(mode, flags, nodes);
730 if (IS_ERR(new)) {
731 ret = PTR_ERR(new);
732 goto out;
733 }
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700734 /*
735 * prevent changing our mempolicy while show_numa_maps()
736 * is using it.
737 * Note: do_set_mempolicy() can be called at init time
738 * with no 'mm'.
739 */
740 if (mm)
741 down_write(&mm->mmap_sem);
Miao Xie58568d22009-06-16 15:31:49 -0700742 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700743 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700744 if (ret) {
745 task_unlock(current);
746 if (mm)
747 up_write(&mm->mmap_sem);
748 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700749 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700750 }
751 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 current->mempolicy = new;
Paul Jacksonc61afb12006-03-24 03:16:08 -0800753 mpol_set_task_struct_flag();
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700754 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700755 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700756 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700757 task_unlock(current);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700758 if (mm)
759 up_write(&mm->mmap_sem);
760
Miao Xie58568d22009-06-16 15:31:49 -0700761 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700762 ret = 0;
763out:
764 NODEMASK_SCRATCH_FREE(scratch);
765 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766}
767
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700768/*
769 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700770 *
771 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700772 */
773static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700775 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700776 if (p == &default_policy)
777 return;
778
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700779 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700780 case MPOL_BIND:
781 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700783 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 break;
785 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700786 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700787 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700788 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 break;
790 default:
791 BUG();
792 }
793}
794
795static int lookup_node(struct mm_struct *mm, unsigned long addr)
796{
797 struct page *p;
798 int err;
799
800 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
801 if (err >= 0) {
802 err = page_to_nid(p);
803 put_page(p);
804 }
805 return err;
806}
807
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700809static long do_get_mempolicy(int *policy, nodemask_t *nmask,
810 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700812 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 struct mm_struct *mm = current->mm;
814 struct vm_area_struct *vma = NULL;
815 struct mempolicy *pol = current->mempolicy;
816
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700817 if (flags &
818 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700820
821 if (flags & MPOL_F_MEMS_ALLOWED) {
822 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
823 return -EINVAL;
824 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700825 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700826 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700827 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700828 return 0;
829 }
830
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700832 /*
833 * Do NOT fall back to task policy if the
834 * vma/shared policy at addr is NULL. We
835 * want to return MPOL_DEFAULT in this case.
836 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 down_read(&mm->mmap_sem);
838 vma = find_vma_intersection(mm, addr, addr+1);
839 if (!vma) {
840 up_read(&mm->mmap_sem);
841 return -EFAULT;
842 }
843 if (vma->vm_ops && vma->vm_ops->get_policy)
844 pol = vma->vm_ops->get_policy(vma, addr);
845 else
846 pol = vma->vm_policy;
847 } else if (addr)
848 return -EINVAL;
849
850 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700851 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852
853 if (flags & MPOL_F_NODE) {
854 if (flags & MPOL_F_ADDR) {
855 err = lookup_node(mm, addr);
856 if (err < 0)
857 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700858 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700860 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700861 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 } else {
863 err = -EINVAL;
864 goto out;
865 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700866 } else {
867 *policy = pol == &default_policy ? MPOL_DEFAULT :
868 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700869 /*
870 * Internal mempolicy flags must be masked off before exposing
871 * the policy to userspace.
872 */
873 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700874 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875
876 if (vma) {
877 up_read(&current->mm->mmap_sem);
878 vma = NULL;
879 }
880
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700882 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700883 if (mpol_store_user_nodemask(pol)) {
884 *nmask = pol->w.user_nodemask;
885 } else {
886 task_lock(current);
887 get_policy_nodemask(pol, nmask);
888 task_unlock(current);
889 }
Miao Xie58568d22009-06-16 15:31:49 -0700890 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891
892 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700893 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 if (vma)
895 up_read(&current->mm->mmap_sem);
896 return err;
897}
898
Christoph Lameterb20a3502006-03-22 00:09:12 -0800899#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700900/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800901 * page migration
902 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800903static void migrate_page_add(struct page *page, struct list_head *pagelist,
904 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800905{
906 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800907 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800908 */
Nick Piggin62695a82008-10-18 20:26:09 -0700909 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
910 if (!isolate_lru_page(page)) {
911 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800912 inc_zone_page_state(page, NR_ISOLATED_ANON +
913 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700914 }
915 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800916}
917
Christoph Lameter742755a2006-06-23 02:03:55 -0700918static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700919{
Mel Gorman6484eb32009-06-16 15:31:54 -0700920 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700921}
922
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800923/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800924 * Migrate pages from one node to a target node.
925 * Returns error or the number of pages not migrated.
926 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700927static int migrate_to_node(struct mm_struct *mm, int source, int dest,
928 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800929{
930 nodemask_t nmask;
931 LIST_HEAD(pagelist);
932 int err = 0;
Vasiliy Kulikov0def08e2010-10-26 14:21:32 -0700933 struct vm_area_struct *vma;
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800934
935 nodes_clear(nmask);
936 node_set(source, nmask);
937
Vasiliy Kulikov0def08e2010-10-26 14:21:32 -0700938 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800939 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
Vasiliy Kulikov0def08e2010-10-26 14:21:32 -0700940 if (IS_ERR(vma))
941 return PTR_ERR(vma);
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800942
Minchan Kimcf608ac2010-10-26 14:21:29 -0700943 if (!list_empty(&pagelist)) {
Mel Gorman7f0f2492011-01-13 15:45:58 -0800944 err = migrate_pages(&pagelist, new_node_page, dest,
945 false, true);
Minchan Kimcf608ac2010-10-26 14:21:29 -0700946 if (err)
947 putback_lru_pages(&pagelist);
948 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700949
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800950 return err;
951}
952
953/*
954 * Move pages between the two nodesets so as to preserve the physical
955 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -0800956 *
957 * Returns the number of page that could not be moved.
958 */
959int do_migrate_pages(struct mm_struct *mm,
960 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
961{
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800962 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800963 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800964 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -0800965
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800966 err = migrate_prep();
967 if (err)
968 return err;
969
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700970 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -0800971
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700972 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
973 if (err)
974 goto out;
975
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -0800976 /*
977 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
978 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
979 * bit in 'tmp', and return that <source, dest> pair for migration.
980 * The pair of nodemasks 'to' and 'from' define the map.
981 *
982 * If no pair of bits is found that way, fallback to picking some
983 * pair of 'source' and 'dest' bits that are not the same. If the
984 * 'source' and 'dest' bits are the same, this represents a node
985 * that will be migrating to itself, so no pages need move.
986 *
987 * If no bits are left in 'tmp', or if all remaining bits left
988 * in 'tmp' correspond to the same bit in 'to', return false
989 * (nothing left to migrate).
990 *
991 * This lets us pick a pair of nodes to migrate between, such that
992 * if possible the dest node is not already occupied by some other
993 * source node, minimizing the risk of overloading the memory on a
994 * node that would happen if we migrated incoming memory to a node
995 * before migrating outgoing memory source that same node.
996 *
997 * A single scan of tmp is sufficient. As we go, we remember the
998 * most recent <s, d> pair that moved (s != d). If we find a pair
999 * that not only moved, but what's better, moved to an empty slot
1000 * (d is not set in tmp), then we break out then, with that pair.
Justin P. Mattockae0e47f2011-03-01 15:06:02 +01001001 * Otherwise when we finish scanning from_tmp, we at least have the
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -08001002 * most recent <s, d> pair that moved. If we get all the way through
1003 * the scan of tmp without finding any node that moved, much less
1004 * moved to an empty node, then there is nothing left worth migrating.
1005 */
Christoph Lameterd4984712006-01-08 01:00:55 -08001006
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001007 tmp = *from_nodes;
1008 while (!nodes_empty(tmp)) {
1009 int s,d;
1010 int source = -1;
1011 int dest = 0;
1012
1013 for_each_node_mask(s, tmp) {
1014 d = node_remap(s, *from_nodes, *to_nodes);
1015 if (s == d)
1016 continue;
1017
1018 source = s; /* Node moved. Memorize */
1019 dest = d;
1020
1021 /* dest not in remaining from nodes? */
1022 if (!node_isset(dest, tmp))
1023 break;
1024 }
1025 if (source == -1)
1026 break;
1027
1028 node_clear(source, tmp);
1029 err = migrate_to_node(mm, source, dest, flags);
1030 if (err > 0)
1031 busy += err;
1032 if (err < 0)
1033 break;
Christoph Lameter39743882006-01-08 01:00:51 -08001034 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001035out:
Christoph Lameter39743882006-01-08 01:00:51 -08001036 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001037 if (err < 0)
1038 return err;
1039 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001040
Christoph Lameter39743882006-01-08 01:00:51 -08001041}
1042
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001043/*
1044 * Allocate a new page for page migration based on vma policy.
1045 * Start assuming that page is mapped by vma pointed to by @private.
1046 * Search forward from there, if not. N.B., this assumes that the
1047 * list of pages handed to migrate_pages()--which is how we get here--
1048 * is in virtual address order.
1049 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001050static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001051{
1052 struct vm_area_struct *vma = (struct vm_area_struct *)private;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001053 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001054
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001055 while (vma) {
1056 address = page_address_in_vma(page, vma);
1057 if (address != -EFAULT)
1058 break;
1059 vma = vma->vm_next;
1060 }
1061
1062 /*
1063 * if !vma, alloc_page_vma() will use task or system default policy
1064 */
1065 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001066}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001067#else
1068
1069static void migrate_page_add(struct page *page, struct list_head *pagelist,
1070 unsigned long flags)
1071{
1072}
1073
1074int do_migrate_pages(struct mm_struct *mm,
1075 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1076{
1077 return -ENOSYS;
1078}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001079
Keith Owens69939742006-10-11 01:21:28 -07001080static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001081{
1082 return NULL;
1083}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001084#endif
1085
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001086static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001087 unsigned short mode, unsigned short mode_flags,
1088 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001089{
1090 struct vm_area_struct *vma;
1091 struct mm_struct *mm = current->mm;
1092 struct mempolicy *new;
1093 unsigned long end;
1094 int err;
1095 LIST_HEAD(pagelist);
1096
David Rientjesa3b51e02008-04-28 02:12:23 -07001097 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1098 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001099 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001100 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001101 return -EPERM;
1102
1103 if (start & ~PAGE_MASK)
1104 return -EINVAL;
1105
1106 if (mode == MPOL_DEFAULT)
1107 flags &= ~MPOL_MF_STRICT;
1108
1109 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1110 end = start + len;
1111
1112 if (end < start)
1113 return -EINVAL;
1114 if (end == start)
1115 return 0;
1116
David Rientjes028fec42008-04-28 02:12:25 -07001117 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001118 if (IS_ERR(new))
1119 return PTR_ERR(new);
1120
1121 /*
1122 * If we are using the default policy then operation
1123 * on discontinuous address spaces is okay after all
1124 */
1125 if (!new)
1126 flags |= MPOL_MF_DISCONTIG_OK;
1127
David Rientjes028fec42008-04-28 02:12:25 -07001128 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1129 start, start + len, mode, mode_flags,
1130 nmask ? nodes_addr(*nmask)[0] : -1);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001131
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001132 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1133
1134 err = migrate_prep();
1135 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001136 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001137 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001138 {
1139 NODEMASK_SCRATCH(scratch);
1140 if (scratch) {
1141 down_write(&mm->mmap_sem);
1142 task_lock(current);
1143 err = mpol_set_nodemask(new, nmask, scratch);
1144 task_unlock(current);
1145 if (err)
1146 up_write(&mm->mmap_sem);
1147 } else
1148 err = -ENOMEM;
1149 NODEMASK_SCRATCH_FREE(scratch);
1150 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001151 if (err)
1152 goto mpol_out;
1153
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001154 vma = check_range(mm, start, end, nmask,
1155 flags | MPOL_MF_INVERT, &pagelist);
1156
1157 err = PTR_ERR(vma);
1158 if (!IS_ERR(vma)) {
1159 int nr_failed = 0;
1160
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001161 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001162
Minchan Kimcf608ac2010-10-26 14:21:29 -07001163 if (!list_empty(&pagelist)) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001164 nr_failed = migrate_pages(&pagelist, new_vma_page,
Mel Gorman7f0f2492011-01-13 15:45:58 -08001165 (unsigned long)vma,
1166 false, true);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001167 if (nr_failed)
1168 putback_lru_pages(&pagelist);
1169 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001170
1171 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1172 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001173 } else
1174 putback_lru_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001175
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001176 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001177 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001178 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001179 return err;
1180}
1181
Christoph Lameter39743882006-01-08 01:00:51 -08001182/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001183 * User space interface with variable sized bitmaps for nodelists.
1184 */
1185
1186/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001187static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001188 unsigned long maxnode)
1189{
1190 unsigned long k;
1191 unsigned long nlongs;
1192 unsigned long endmask;
1193
1194 --maxnode;
1195 nodes_clear(*nodes);
1196 if (maxnode == 0 || !nmask)
1197 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001198 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001199 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001200
1201 nlongs = BITS_TO_LONGS(maxnode);
1202 if ((maxnode % BITS_PER_LONG) == 0)
1203 endmask = ~0UL;
1204 else
1205 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1206
1207 /* When the user specified more nodes than supported just check
1208 if the non supported part is all zero. */
1209 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1210 if (nlongs > PAGE_SIZE/sizeof(long))
1211 return -EINVAL;
1212 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1213 unsigned long t;
1214 if (get_user(t, nmask + k))
1215 return -EFAULT;
1216 if (k == nlongs - 1) {
1217 if (t & endmask)
1218 return -EINVAL;
1219 } else if (t)
1220 return -EINVAL;
1221 }
1222 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1223 endmask = ~0UL;
1224 }
1225
1226 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1227 return -EFAULT;
1228 nodes_addr(*nodes)[nlongs-1] &= endmask;
1229 return 0;
1230}
1231
1232/* Copy a kernel node mask to user space */
1233static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1234 nodemask_t *nodes)
1235{
1236 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1237 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1238
1239 if (copy > nbytes) {
1240 if (copy > PAGE_SIZE)
1241 return -EINVAL;
1242 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1243 return -EFAULT;
1244 copy = nbytes;
1245 }
1246 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1247}
1248
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001249SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1250 unsigned long, mode, unsigned long __user *, nmask,
1251 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001252{
1253 nodemask_t nodes;
1254 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001255 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001256
David Rientjes028fec42008-04-28 02:12:25 -07001257 mode_flags = mode & MPOL_MODE_FLAGS;
1258 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001259 if (mode >= MPOL_MAX)
1260 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001261 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1262 (mode_flags & MPOL_F_RELATIVE_NODES))
1263 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001264 err = get_nodes(&nodes, nmask, maxnode);
1265 if (err)
1266 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001267 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001268}
1269
1270/* Set the process memory policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001271SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1272 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001273{
1274 int err;
1275 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001276 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001277
David Rientjes028fec42008-04-28 02:12:25 -07001278 flags = mode & MPOL_MODE_FLAGS;
1279 mode &= ~MPOL_MODE_FLAGS;
1280 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001281 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001282 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1283 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001284 err = get_nodes(&nodes, nmask, maxnode);
1285 if (err)
1286 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001287 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001288}
1289
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001290SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1291 const unsigned long __user *, old_nodes,
1292 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001293{
David Howellsc69e8d92008-11-14 10:39:19 +11001294 const struct cred *cred = current_cred(), *tcred;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001295 struct mm_struct *mm = NULL;
Christoph Lameter39743882006-01-08 01:00:51 -08001296 struct task_struct *task;
Christoph Lameter39743882006-01-08 01:00:51 -08001297 nodemask_t task_nodes;
1298 int err;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001299 nodemask_t *old;
1300 nodemask_t *new;
1301 NODEMASK_SCRATCH(scratch);
Christoph Lameter39743882006-01-08 01:00:51 -08001302
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001303 if (!scratch)
1304 return -ENOMEM;
Christoph Lameter39743882006-01-08 01:00:51 -08001305
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001306 old = &scratch->mask1;
1307 new = &scratch->mask2;
1308
1309 err = get_nodes(old, old_nodes, maxnode);
Christoph Lameter39743882006-01-08 01:00:51 -08001310 if (err)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001311 goto out;
1312
1313 err = get_nodes(new, new_nodes, maxnode);
1314 if (err)
1315 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001316
1317 /* Find the mm_struct */
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001318 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001319 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001320 if (!task) {
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001321 rcu_read_unlock();
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001322 err = -ESRCH;
1323 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001324 }
1325 mm = get_task_mm(task);
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001326 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001327
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001328 err = -EINVAL;
Christoph Lameter39743882006-01-08 01:00:51 -08001329 if (!mm)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001330 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001331
1332 /*
1333 * Check if this process has the right to modify the specified
1334 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001335 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001336 * userid as the target process.
1337 */
David Howellsc69e8d92008-11-14 10:39:19 +11001338 rcu_read_lock();
1339 tcred = __task_cred(task);
David Howellsb6dff3e2008-11-14 10:39:16 +11001340 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1341 cred->uid != tcred->suid && cred->uid != tcred->uid &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001342 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001343 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001344 err = -EPERM;
1345 goto out;
1346 }
David Howellsc69e8d92008-11-14 10:39:19 +11001347 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001348
1349 task_nodes = cpuset_mems_allowed(task);
1350 /* Is the user allowed to access the target nodes? */
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001351 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001352 err = -EPERM;
1353 goto out;
1354 }
1355
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001356 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001357 err = -EINVAL;
1358 goto out;
1359 }
1360
David Quigley86c3a762006-06-23 02:04:02 -07001361 err = security_task_movememory(task);
1362 if (err)
1363 goto out;
1364
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001365 err = do_migrate_pages(mm, old, new,
Christoph Lameter74c00242006-03-14 19:50:21 -08001366 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter39743882006-01-08 01:00:51 -08001367out:
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001368 if (mm)
1369 mmput(mm);
1370 NODEMASK_SCRATCH_FREE(scratch);
1371
Christoph Lameter39743882006-01-08 01:00:51 -08001372 return err;
1373}
1374
1375
Christoph Lameter8bccd852005-10-29 18:16:59 -07001376/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001377SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1378 unsigned long __user *, nmask, unsigned long, maxnode,
1379 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001380{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001381 int err;
1382 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001383 nodemask_t nodes;
1384
1385 if (nmask != NULL && maxnode < MAX_NUMNODES)
1386 return -EINVAL;
1387
1388 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1389
1390 if (err)
1391 return err;
1392
1393 if (policy && put_user(pval, policy))
1394 return -EFAULT;
1395
1396 if (nmask)
1397 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1398
1399 return err;
1400}
1401
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402#ifdef CONFIG_COMPAT
1403
1404asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1405 compat_ulong_t __user *nmask,
1406 compat_ulong_t maxnode,
1407 compat_ulong_t addr, compat_ulong_t flags)
1408{
1409 long err;
1410 unsigned long __user *nm = NULL;
1411 unsigned long nr_bits, alloc_size;
1412 DECLARE_BITMAP(bm, MAX_NUMNODES);
1413
1414 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1415 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1416
1417 if (nmask)
1418 nm = compat_alloc_user_space(alloc_size);
1419
1420 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1421
1422 if (!err && nmask) {
KAMEZAWA Hiroyuki2bbff6c2011-09-14 16:21:02 -07001423 unsigned long copy_size;
1424 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1425 err = copy_from_user(bm, nm, copy_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 /* ensure entire bitmap is zeroed */
1427 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1428 err |= compat_put_bitmap(nmask, bm, nr_bits);
1429 }
1430
1431 return err;
1432}
1433
1434asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1435 compat_ulong_t maxnode)
1436{
1437 long err = 0;
1438 unsigned long __user *nm = NULL;
1439 unsigned long nr_bits, alloc_size;
1440 DECLARE_BITMAP(bm, MAX_NUMNODES);
1441
1442 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1443 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1444
1445 if (nmask) {
1446 err = compat_get_bitmap(bm, nmask, nr_bits);
1447 nm = compat_alloc_user_space(alloc_size);
1448 err |= copy_to_user(nm, bm, alloc_size);
1449 }
1450
1451 if (err)
1452 return -EFAULT;
1453
1454 return sys_set_mempolicy(mode, nm, nr_bits+1);
1455}
1456
1457asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1458 compat_ulong_t mode, compat_ulong_t __user *nmask,
1459 compat_ulong_t maxnode, compat_ulong_t flags)
1460{
1461 long err = 0;
1462 unsigned long __user *nm = NULL;
1463 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001464 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
1466 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1467 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1468
1469 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001470 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001472 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 }
1474
1475 if (err)
1476 return -EFAULT;
1477
1478 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1479}
1480
1481#endif
1482
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001483/*
1484 * get_vma_policy(@task, @vma, @addr)
1485 * @task - task for fallback if vma policy == default
1486 * @vma - virtual memory area whose policy is sought
1487 * @addr - address in @vma for shared policy lookup
1488 *
1489 * Returns effective policy for a VMA at specified address.
1490 * Falls back to @task or system default policy, as necessary.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001491 * Current or other task's task mempolicy and non-shared vma policies
1492 * are protected by the task's mmap_sem, which must be held for read by
1493 * the caller.
1494 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1495 * count--added by the get_policy() vm_op, as appropriate--to protect against
1496 * freeing by another task. It is the caller's responsibility to free the
1497 * extra reference for shared policies.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001498 */
Stephen Wilsond98f6cb2011-05-24 17:12:41 -07001499struct mempolicy *get_vma_policy(struct task_struct *task,
Christoph Lameter48fce342006-01-08 01:01:03 -08001500 struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001502 struct mempolicy *pol = task->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503
1504 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001505 if (vma->vm_ops && vma->vm_ops->get_policy) {
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001506 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1507 addr);
1508 if (vpol)
1509 pol = vpol;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001510 } else if (vma->vm_policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511 pol = vma->vm_policy;
1512 }
1513 if (!pol)
1514 pol = &default_policy;
1515 return pol;
1516}
1517
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001518/*
1519 * Return a nodemask representing a mempolicy for filtering nodes for
1520 * page allocation
1521 */
1522static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001523{
1524 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001525 if (unlikely(policy->mode == MPOL_BIND) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001526 gfp_zone(gfp) >= policy_zone &&
1527 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1528 return &policy->v.nodes;
1529
1530 return NULL;
1531}
1532
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001533/* Return a zonelist indicated by gfp for node representing a mempolicy */
Andi Kleen2f5f9482011-03-04 17:36:29 -08001534static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1535 int nd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536{
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001537 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001539 if (!(policy->flags & MPOL_F_LOCAL))
1540 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 break;
1542 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001543 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001544 * Normally, MPOL_BIND allocations are node-local within the
1545 * allowed nodemask. However, if __GFP_THISNODE is set and the
Bob Liu6eb27e12010-05-24 14:32:00 -07001546 * current node isn't part of the mask, we use the zonelist for
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001547 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001548 */
Mel Gorman19770b32008-04-28 02:12:18 -07001549 if (unlikely(gfp & __GFP_THISNODE) &&
1550 unlikely(!node_isset(nd, policy->v.nodes)))
1551 nd = first_node(policy->v.nodes);
1552 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 BUG();
1555 }
Mel Gorman0e884602008-04-28 02:12:14 -07001556 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557}
1558
1559/* Do dynamic interleaving for a process */
1560static unsigned interleave_nodes(struct mempolicy *policy)
1561{
1562 unsigned nid, next;
1563 struct task_struct *me = current;
1564
1565 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001566 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001568 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001569 if (next < MAX_NUMNODES)
1570 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 return nid;
1572}
1573
Christoph Lameterdc85da12006-01-18 17:42:36 -08001574/*
1575 * Depending on the memory policy provide a node from which to allocate the
1576 * next slab entry.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001577 * @policy must be protected by freeing by the caller. If @policy is
1578 * the current task's mempolicy, this protection is implicit, as only the
1579 * task can change it's policy. The system default policy requires no
1580 * such protection.
Christoph Lameterdc85da12006-01-18 17:42:36 -08001581 */
1582unsigned slab_node(struct mempolicy *policy)
1583{
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001584 if (!policy || policy->flags & MPOL_F_LOCAL)
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001585 return numa_node_id();
Christoph Lameter765c4502006-09-27 01:50:08 -07001586
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001587 switch (policy->mode) {
1588 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001589 /*
1590 * handled MPOL_F_LOCAL above
1591 */
1592 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001593
Christoph Lameterdc85da12006-01-18 17:42:36 -08001594 case MPOL_INTERLEAVE:
1595 return interleave_nodes(policy);
1596
Mel Gormandd1a2392008-04-28 02:12:17 -07001597 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001598 /*
1599 * Follow bind policy behavior and start allocation at the
1600 * first node.
1601 */
Mel Gorman19770b32008-04-28 02:12:18 -07001602 struct zonelist *zonelist;
1603 struct zone *zone;
1604 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1605 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1606 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1607 &policy->v.nodes,
1608 &zone);
Eric Dumazet800416f2010-10-27 19:33:43 +02001609 return zone ? zone->node : numa_node_id();
Mel Gormandd1a2392008-04-28 02:12:17 -07001610 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001611
Christoph Lameterdc85da12006-01-18 17:42:36 -08001612 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001613 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001614 }
1615}
1616
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617/* Do static interleaving for a VMA with known offset. */
1618static unsigned offset_il_node(struct mempolicy *pol,
1619 struct vm_area_struct *vma, unsigned long off)
1620{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001621 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001622 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 int c;
1624 int nid = -1;
1625
David Rientjesf5b087b2008-04-28 02:12:27 -07001626 if (!nnodes)
1627 return numa_node_id();
1628 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 c = 0;
1630 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001631 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 c++;
1633 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634 return nid;
1635}
1636
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001637/* Determine a node number for interleave */
1638static inline unsigned interleave_nid(struct mempolicy *pol,
1639 struct vm_area_struct *vma, unsigned long addr, int shift)
1640{
1641 if (vma) {
1642 unsigned long off;
1643
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001644 /*
1645 * for small pages, there is no difference between
1646 * shift and PAGE_SHIFT, so the bit-shift is safe.
1647 * for huge pages, since vm_pgoff is in units of small
1648 * pages, we need to shift off the always 0 bits to get
1649 * a useful offset.
1650 */
1651 BUG_ON(shift < PAGE_SHIFT);
1652 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001653 off += (addr - vma->vm_start) >> shift;
1654 return offset_il_node(pol, vma, off);
1655 } else
1656 return interleave_nodes(pol);
1657}
1658
Michal Hocko778d3b02011-07-26 16:08:30 -07001659/*
1660 * Return the bit number of a random bit set in the nodemask.
1661 * (returns -1 if nodemask is empty)
1662 */
1663int node_random(const nodemask_t *maskp)
1664{
1665 int w, bit = -1;
1666
1667 w = nodes_weight(*maskp);
1668 if (w)
1669 bit = bitmap_ord_to_pos(maskp->bits,
1670 get_random_int() % w, MAX_NUMNODES);
1671 return bit;
1672}
1673
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001674#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001675/*
1676 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1677 * @vma = virtual memory area whose policy is sought
1678 * @addr = address in @vma for shared policy lookup and interleave policy
1679 * @gfp_flags = for requested zone
Mel Gorman19770b32008-04-28 02:12:18 -07001680 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1681 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001682 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001683 * Returns a zonelist suitable for a huge page allocation and a pointer
1684 * to the struct mempolicy for conditional unref after allocation.
1685 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1686 * @nodemask for filtering the zonelist.
Miao Xiec0ff7452010-05-24 14:32:08 -07001687 *
1688 * Must be protected by get_mems_allowed()
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001689 */
Mel Gorman396faf02007-07-17 04:03:13 -07001690struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001691 gfp_t gfp_flags, struct mempolicy **mpol,
1692 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001693{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001694 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001695
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001696 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001697 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001698
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001699 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1700 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001701 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001702 } else {
Andi Kleen2f5f9482011-03-04 17:36:29 -08001703 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001704 if ((*mpol)->mode == MPOL_BIND)
1705 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001706 }
1707 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001708}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001709
1710/*
1711 * init_nodemask_of_mempolicy
1712 *
1713 * If the current task's mempolicy is "default" [NULL], return 'false'
1714 * to indicate default policy. Otherwise, extract the policy nodemask
1715 * for 'bind' or 'interleave' policy into the argument nodemask, or
1716 * initialize the argument nodemask to contain the single node for
1717 * 'preferred' or 'local' policy and return 'true' to indicate presence
1718 * of non-default mempolicy.
1719 *
1720 * We don't bother with reference counting the mempolicy [mpol_get/put]
1721 * because the current task is examining it's own mempolicy and a task's
1722 * mempolicy is only ever changed by the task itself.
1723 *
1724 * N.B., it is the caller's responsibility to free a returned nodemask.
1725 */
1726bool init_nodemask_of_mempolicy(nodemask_t *mask)
1727{
1728 struct mempolicy *mempolicy;
1729 int nid;
1730
1731 if (!(mask && current->mempolicy))
1732 return false;
1733
Miao Xiec0ff7452010-05-24 14:32:08 -07001734 task_lock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001735 mempolicy = current->mempolicy;
1736 switch (mempolicy->mode) {
1737 case MPOL_PREFERRED:
1738 if (mempolicy->flags & MPOL_F_LOCAL)
1739 nid = numa_node_id();
1740 else
1741 nid = mempolicy->v.preferred_node;
1742 init_nodemask_of_node(mask, nid);
1743 break;
1744
1745 case MPOL_BIND:
1746 /* Fall through */
1747 case MPOL_INTERLEAVE:
1748 *mask = mempolicy->v.nodes;
1749 break;
1750
1751 default:
1752 BUG();
1753 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001754 task_unlock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001755
1756 return true;
1757}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001758#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001759
David Rientjes6f48d0eb2010-08-09 17:18:52 -07001760/*
1761 * mempolicy_nodemask_intersects
1762 *
1763 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1764 * policy. Otherwise, check for intersection between mask and the policy
1765 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1766 * policy, always return true since it may allocate elsewhere on fallback.
1767 *
1768 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1769 */
1770bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1771 const nodemask_t *mask)
1772{
1773 struct mempolicy *mempolicy;
1774 bool ret = true;
1775
1776 if (!mask)
1777 return ret;
1778 task_lock(tsk);
1779 mempolicy = tsk->mempolicy;
1780 if (!mempolicy)
1781 goto out;
1782
1783 switch (mempolicy->mode) {
1784 case MPOL_PREFERRED:
1785 /*
1786 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1787 * allocate from, they may fallback to other nodes when oom.
1788 * Thus, it's possible for tsk to have allocated memory from
1789 * nodes in mask.
1790 */
1791 break;
1792 case MPOL_BIND:
1793 case MPOL_INTERLEAVE:
1794 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1795 break;
1796 default:
1797 BUG();
1798 }
1799out:
1800 task_unlock(tsk);
1801 return ret;
1802}
1803
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804/* Allocate a page in interleaved policy.
1805 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001806static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1807 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808{
1809 struct zonelist *zl;
1810 struct page *page;
1811
Mel Gorman0e884602008-04-28 02:12:14 -07001812 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001814 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001815 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 return page;
1817}
1818
1819/**
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001820 * alloc_pages_vma - Allocate a page for a VMA.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 *
1822 * @gfp:
1823 * %GFP_USER user allocation.
1824 * %GFP_KERNEL kernel allocations,
1825 * %GFP_HIGHMEM highmem/user allocations,
1826 * %GFP_FS allocation should not call back into a file system.
1827 * %GFP_ATOMIC don't sleep.
1828 *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001829 * @order:Order of the GFP allocation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 * @vma: Pointer to VMA or NULL if not available.
1831 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1832 *
1833 * This function allocates a page from the kernel page pool and applies
1834 * a NUMA policy associated with the VMA or the current process.
1835 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1836 * mm_struct of the VMA to prevent it from going away. Should be used for
1837 * all allocations for pages that will be mapped into
1838 * user space. Returns NULL when no page can be allocated.
1839 *
1840 * Should be called with the mm_sem of the vma hold.
1841 */
1842struct page *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001843alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
Andi Kleen2f5f9482011-03-04 17:36:29 -08001844 unsigned long addr, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001846 struct mempolicy *pol = get_vma_policy(current, vma, addr);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001847 struct zonelist *zl;
Miao Xiec0ff7452010-05-24 14:32:08 -07001848 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849
Miao Xiec0ff7452010-05-24 14:32:08 -07001850 get_mems_allowed();
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001851 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001853
Andi Kleen8eac5632011-02-25 14:44:28 -08001854 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001855 mpol_cond_put(pol);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001856 page = alloc_page_interleave(gfp, order, nid);
Miao Xiec0ff7452010-05-24 14:32:08 -07001857 put_mems_allowed();
1858 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 }
Andi Kleen2f5f9482011-03-04 17:36:29 -08001860 zl = policy_zonelist(gfp, pol, node);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001861 if (unlikely(mpol_needs_cond_ref(pol))) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001862 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001863 * slow path: ref counted shared policy
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001864 */
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001865 struct page *page = __alloc_pages_nodemask(gfp, order,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001866 zl, policy_nodemask(gfp, pol));
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001867 __mpol_put(pol);
Miao Xiec0ff7452010-05-24 14:32:08 -07001868 put_mems_allowed();
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001869 return page;
1870 }
1871 /*
1872 * fast path: default or task policy
1873 */
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001874 page = __alloc_pages_nodemask(gfp, order, zl,
1875 policy_nodemask(gfp, pol));
Miao Xiec0ff7452010-05-24 14:32:08 -07001876 put_mems_allowed();
1877 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878}
1879
1880/**
1881 * alloc_pages_current - Allocate pages.
1882 *
1883 * @gfp:
1884 * %GFP_USER user allocation,
1885 * %GFP_KERNEL kernel allocation,
1886 * %GFP_HIGHMEM highmem allocation,
1887 * %GFP_FS don't call back into a file system.
1888 * %GFP_ATOMIC don't sleep.
1889 * @order: Power of two of allocation size in pages. 0 is a single page.
1890 *
1891 * Allocate a page from the kernel page pool. When not in
1892 * interrupt context and apply the current process NUMA policy.
1893 * Returns NULL when no page can be allocated.
1894 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001895 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 * 1) it's ok to take cpuset_sem (can WAIT), and
1897 * 2) allocating for current task (not interrupt).
1898 */
Al Virodd0fc662005-10-07 07:46:04 +01001899struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900{
1901 struct mempolicy *pol = current->mempolicy;
Miao Xiec0ff7452010-05-24 14:32:08 -07001902 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903
Christoph Lameter9b819d22006-09-25 23:31:40 -07001904 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 pol = &default_policy;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001906
Miao Xiec0ff7452010-05-24 14:32:08 -07001907 get_mems_allowed();
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001908 /*
1909 * No reference counting needed for current->mempolicy
1910 * nor system default_policy
1911 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001912 if (pol->mode == MPOL_INTERLEAVE)
Miao Xiec0ff7452010-05-24 14:32:08 -07001913 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1914 else
1915 page = __alloc_pages_nodemask(gfp, order,
Andi Kleen5c4b4be2011-03-04 17:36:32 -08001916 policy_zonelist(gfp, pol, numa_node_id()),
1917 policy_nodemask(gfp, pol));
Miao Xiec0ff7452010-05-24 14:32:08 -07001918 put_mems_allowed();
1919 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920}
1921EXPORT_SYMBOL(alloc_pages_current);
1922
Paul Jackson42253992006-01-08 01:01:59 -08001923/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001924 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08001925 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1926 * with the mems_allowed returned by cpuset_mems_allowed(). This
1927 * keeps mempolicies cpuset relative after its cpuset moves. See
1928 * further kernel/cpuset.c update_nodemask().
Miao Xie708c1bb2010-05-24 14:32:07 -07001929 *
1930 * current's mempolicy may be rebinded by the other task(the task that changes
1931 * cpuset's mems), so we needn't do rebind work for current task.
Paul Jackson42253992006-01-08 01:01:59 -08001932 */
Paul Jackson42253992006-01-08 01:01:59 -08001933
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001934/* Slow path of a mempolicy duplicate */
1935struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936{
1937 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1938
1939 if (!new)
1940 return ERR_PTR(-ENOMEM);
Miao Xie708c1bb2010-05-24 14:32:07 -07001941
1942 /* task's mempolicy is protected by alloc_lock */
1943 if (old == current->mempolicy) {
1944 task_lock(current);
1945 *new = *old;
1946 task_unlock(current);
1947 } else
1948 *new = *old;
1949
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08001950 rcu_read_lock();
Paul Jackson42253992006-01-08 01:01:59 -08001951 if (current_cpuset_is_being_rebound()) {
1952 nodemask_t mems = cpuset_mems_allowed(current);
Miao Xie708c1bb2010-05-24 14:32:07 -07001953 if (new->flags & MPOL_F_REBINDING)
1954 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1955 else
1956 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
Paul Jackson42253992006-01-08 01:01:59 -08001957 }
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08001958 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 return new;
1961}
1962
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001963/*
1964 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1965 * eliminate the * MPOL_F_* flags that require conditional ref and
1966 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1967 * after return. Use the returned value.
1968 *
1969 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1970 * policy lookup, even if the policy needs/has extra ref on lookup.
1971 * shmem_readahead needs this.
1972 */
1973struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1974 struct mempolicy *frompol)
1975{
1976 if (!mpol_needs_cond_ref(frompol))
1977 return frompol;
1978
1979 *tompol = *frompol;
1980 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1981 __mpol_put(frompol);
1982 return tompol;
1983}
1984
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985/* Slow path of a mempolicy comparison */
1986int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1987{
1988 if (!a || !b)
1989 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001990 if (a->mode != b->mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 return 0;
Bob Liu19800502010-05-24 14:32:01 -07001992 if (a->flags != b->flags)
David Rientjesf5b087b2008-04-28 02:12:27 -07001993 return 0;
Bob Liu19800502010-05-24 14:32:01 -07001994 if (mpol_store_user_nodemask(a))
1995 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
1996 return 0;
1997
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001998 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07001999 case MPOL_BIND:
2000 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -07002002 return nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 case MPOL_PREFERRED:
Namhyung Kim75719662011-03-22 16:33:02 -07002004 return a->v.preferred_node == b->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 default:
2006 BUG();
2007 return 0;
2008 }
2009}
2010
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 * Shared memory backing store policy support.
2013 *
2014 * Remember policies even when nobody has shared memory mapped.
2015 * The policies are kept in Red-Black tree linked from the inode.
2016 * They are protected by the sp->lock spinlock, which should be held
2017 * for any accesses to the tree.
2018 */
2019
2020/* lookup first element intersecting start-end */
2021/* Caller holds sp->lock */
2022static struct sp_node *
2023sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2024{
2025 struct rb_node *n = sp->root.rb_node;
2026
2027 while (n) {
2028 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2029
2030 if (start >= p->end)
2031 n = n->rb_right;
2032 else if (end <= p->start)
2033 n = n->rb_left;
2034 else
2035 break;
2036 }
2037 if (!n)
2038 return NULL;
2039 for (;;) {
2040 struct sp_node *w = NULL;
2041 struct rb_node *prev = rb_prev(n);
2042 if (!prev)
2043 break;
2044 w = rb_entry(prev, struct sp_node, nd);
2045 if (w->end <= start)
2046 break;
2047 n = prev;
2048 }
2049 return rb_entry(n, struct sp_node, nd);
2050}
2051
2052/* Insert a new shared policy into the list. */
2053/* Caller holds sp->lock */
2054static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2055{
2056 struct rb_node **p = &sp->root.rb_node;
2057 struct rb_node *parent = NULL;
2058 struct sp_node *nd;
2059
2060 while (*p) {
2061 parent = *p;
2062 nd = rb_entry(parent, struct sp_node, nd);
2063 if (new->start < nd->start)
2064 p = &(*p)->rb_left;
2065 else if (new->end > nd->end)
2066 p = &(*p)->rb_right;
2067 else
2068 BUG();
2069 }
2070 rb_link_node(&new->nd, parent, p);
2071 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07002072 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002073 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074}
2075
2076/* Find shared policy intersecting idx */
2077struct mempolicy *
2078mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2079{
2080 struct mempolicy *pol = NULL;
2081 struct sp_node *sn;
2082
2083 if (!sp->root.rb_node)
2084 return NULL;
2085 spin_lock(&sp->lock);
2086 sn = sp_lookup(sp, idx, idx+1);
2087 if (sn) {
2088 mpol_get(sn->policy);
2089 pol = sn->policy;
2090 }
2091 spin_unlock(&sp->lock);
2092 return pol;
2093}
2094
2095static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2096{
Paul Mundt140d5a42007-07-15 23:38:16 -07002097 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 rb_erase(&n->nd, &sp->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002099 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 kmem_cache_free(sn_cache, n);
2101}
2102
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07002103static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2104 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105{
2106 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2107
2108 if (!n)
2109 return NULL;
2110 n->start = start;
2111 n->end = end;
2112 mpol_get(pol);
Lee Schermerhornaab0b102008-04-28 02:13:13 -07002113 pol->flags |= MPOL_F_SHARED; /* for unref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 n->policy = pol;
2115 return n;
2116}
2117
2118/* Replace a policy range. */
2119static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2120 unsigned long end, struct sp_node *new)
2121{
2122 struct sp_node *n, *new2 = NULL;
2123
2124restart:
2125 spin_lock(&sp->lock);
2126 n = sp_lookup(sp, start, end);
2127 /* Take care of old policies in the same range. */
2128 while (n && n->start < end) {
2129 struct rb_node *next = rb_next(&n->nd);
2130 if (n->start >= start) {
2131 if (n->end <= end)
2132 sp_delete(sp, n);
2133 else
2134 n->start = end;
2135 } else {
2136 /* Old policy spanning whole new range. */
2137 if (n->end > end) {
2138 if (!new2) {
2139 spin_unlock(&sp->lock);
2140 new2 = sp_alloc(end, n->end, n->policy);
2141 if (!new2)
2142 return -ENOMEM;
2143 goto restart;
2144 }
2145 n->end = start;
2146 sp_insert(sp, new2);
2147 new2 = NULL;
2148 break;
2149 } else
2150 n->end = start;
2151 }
2152 if (!next)
2153 break;
2154 n = rb_entry(next, struct sp_node, nd);
2155 }
2156 if (new)
2157 sp_insert(sp, new);
2158 spin_unlock(&sp->lock);
2159 if (new2) {
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002160 mpol_put(new2->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161 kmem_cache_free(sn_cache, new2);
2162 }
2163 return 0;
2164}
2165
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002166/**
2167 * mpol_shared_policy_init - initialize shared policy for inode
2168 * @sp: pointer to inode shared policy
2169 * @mpol: struct mempolicy to install
2170 *
2171 * Install non-NULL @mpol in inode's shared policy rb-tree.
2172 * On entry, the current task has a reference on a non-NULL @mpol.
2173 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002174 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002175 */
2176void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08002177{
Miao Xie58568d22009-06-16 15:31:49 -07002178 int ret;
2179
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002180 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2181 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08002182
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002183 if (mpol) {
2184 struct vm_area_struct pvma;
2185 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002186 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002187
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002188 if (!scratch)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002189 goto put_mpol;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002190 /* contextualize the tmpfs mount point mempolicy */
2191 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002192 if (IS_ERR(new))
Dan Carpenter0cae3452010-05-25 23:42:58 -07002193 goto free_scratch; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002194
2195 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002196 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002197 task_unlock(current);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002198 if (ret)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002199 goto put_new;
Robin Holt7339ff82006-01-14 13:20:48 -08002200
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002201 /* Create pseudo-vma that contains just the policy */
2202 memset(&pvma, 0, sizeof(struct vm_area_struct));
2203 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2204 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002205
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002206put_new:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002207 mpol_put(new); /* drop initial ref */
Dan Carpenter0cae3452010-05-25 23:42:58 -07002208free_scratch:
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002209 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002210put_mpol:
2211 mpol_put(mpol); /* drop our incoming ref on sb mpol */
Robin Holt7339ff82006-01-14 13:20:48 -08002212 }
2213}
2214
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215int mpol_set_shared_policy(struct shared_policy *info,
2216 struct vm_area_struct *vma, struct mempolicy *npol)
2217{
2218 int err;
2219 struct sp_node *new = NULL;
2220 unsigned long sz = vma_pages(vma);
2221
David Rientjes028fec42008-04-28 02:12:25 -07002222 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002224 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002225 npol ? npol->flags : -1,
Paul Mundt140d5a42007-07-15 23:38:16 -07002226 npol ? nodes_addr(npol->v.nodes)[0] : -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227
2228 if (npol) {
2229 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2230 if (!new)
2231 return -ENOMEM;
2232 }
2233 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2234 if (err && new)
2235 kmem_cache_free(sn_cache, new);
2236 return err;
2237}
2238
2239/* Free a backing policy store on inode delete. */
2240void mpol_free_shared_policy(struct shared_policy *p)
2241{
2242 struct sp_node *n;
2243 struct rb_node *next;
2244
2245 if (!p->root.rb_node)
2246 return;
2247 spin_lock(&p->lock);
2248 next = rb_first(&p->root);
2249 while (next) {
2250 n = rb_entry(next, struct sp_node, nd);
2251 next = rb_next(&n->nd);
Andi Kleen90c50292005-07-27 11:43:50 -07002252 rb_erase(&n->nd, &p->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002253 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 kmem_cache_free(sn_cache, n);
2255 }
2256 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257}
2258
2259/* assumes fs == KERNEL_DS */
2260void __init numa_policy_init(void)
2261{
Paul Mundtb71636e2007-07-15 23:38:15 -07002262 nodemask_t interleave_nodes;
2263 unsigned long largest = 0;
2264 int nid, prefer = 0;
2265
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 policy_cache = kmem_cache_create("numa_policy",
2267 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002268 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269
2270 sn_cache = kmem_cache_create("shared_policy_node",
2271 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002272 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273
Paul Mundtb71636e2007-07-15 23:38:15 -07002274 /*
2275 * Set interleaving policy for system init. Interleaving is only
2276 * enabled across suitably sized nodes (default is >= 16MB), or
2277 * fall back to the largest node if they're all smaller.
2278 */
2279 nodes_clear(interleave_nodes);
Christoph Lameter56bbd652007-10-16 01:25:35 -07002280 for_each_node_state(nid, N_HIGH_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002281 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282
Paul Mundtb71636e2007-07-15 23:38:15 -07002283 /* Preserve the largest node */
2284 if (largest < total_pages) {
2285 largest = total_pages;
2286 prefer = nid;
2287 }
2288
2289 /* Interleave this node? */
2290 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2291 node_set(nid, interleave_nodes);
2292 }
2293
2294 /* All too small, use the largest */
2295 if (unlikely(nodes_empty(interleave_nodes)))
2296 node_set(prefer, interleave_nodes);
2297
David Rientjes028fec42008-04-28 02:12:25 -07002298 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 printk("numa_policy_init: interleaving failed\n");
2300}
2301
Christoph Lameter8bccd852005-10-29 18:16:59 -07002302/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303void numa_default_policy(void)
2304{
David Rientjes028fec42008-04-28 02:12:25 -07002305 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306}
Paul Jackson68860ec2005-10-30 15:02:36 -08002307
Paul Jackson42253992006-01-08 01:01:59 -08002308/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002309 * Parse and format mempolicy from/to strings
2310 */
2311
2312/*
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002313 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002314 * Used only for mpol_parse_str() and mpol_to_str()
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002315 */
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002316#define MPOL_LOCAL MPOL_MAX
2317static const char * const policy_modes[] =
2318{
2319 [MPOL_DEFAULT] = "default",
2320 [MPOL_PREFERRED] = "prefer",
2321 [MPOL_BIND] = "bind",
2322 [MPOL_INTERLEAVE] = "interleave",
2323 [MPOL_LOCAL] = "local"
2324};
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002325
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002326
2327#ifdef CONFIG_TMPFS
2328/**
2329 * mpol_parse_str - parse string to mempolicy
2330 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002331 * @mpol: pointer to struct mempolicy pointer, returned on success.
2332 * @no_context: flag whether to "contextualize" the mempolicy
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002333 *
2334 * Format of input:
2335 * <mode>[=<flags>][:<nodelist>]
2336 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002337 * if @no_context is true, save the input nodemask in w.user_nodemask in
2338 * the returned mempolicy. This will be used to "clone" the mempolicy in
2339 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2340 * mount option. Note that if 'static' or 'relative' mode flags were
2341 * specified, the input nodemask will already have been saved. Saving
2342 * it again is redundant, but safe.
2343 *
2344 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002345 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002346int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002347{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002348 struct mempolicy *new = NULL;
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002349 unsigned short mode;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002350 unsigned short uninitialized_var(mode_flags);
2351 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002352 char *nodelist = strchr(str, ':');
2353 char *flags = strchr(str, '=');
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002354 int err = 1;
2355
2356 if (nodelist) {
2357 /* NUL-terminate mode or flags string */
2358 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002359 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002360 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002361 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002362 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002363 } else
2364 nodes_clear(nodes);
2365
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002366 if (flags)
2367 *flags++ = '\0'; /* terminate mode string */
2368
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002369 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002370 if (!strcmp(str, policy_modes[mode])) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002371 break;
2372 }
2373 }
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002374 if (mode > MPOL_LOCAL)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002375 goto out;
2376
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002377 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002378 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002379 /*
2380 * Insist on a nodelist of one node only
2381 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002382 if (nodelist) {
2383 char *rest = nodelist;
2384 while (isdigit(*rest))
2385 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002386 if (*rest)
2387 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002388 }
2389 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002390 case MPOL_INTERLEAVE:
2391 /*
2392 * Default to online nodes with memory if no nodelist
2393 */
2394 if (!nodelist)
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002395 nodes = node_states[N_HIGH_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002396 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002397 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002398 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002399 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002400 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002401 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002402 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002403 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002404 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002405 case MPOL_DEFAULT:
2406 /*
2407 * Insist on a empty nodelist
2408 */
2409 if (!nodelist)
2410 err = 0;
2411 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002412 case MPOL_BIND:
2413 /*
2414 * Insist on a nodelist
2415 */
2416 if (!nodelist)
2417 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002418 }
2419
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002420 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002421 if (flags) {
2422 /*
2423 * Currently, we only support two mutually exclusive
2424 * mode flags.
2425 */
2426 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002427 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002428 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002429 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002430 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002431 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002432 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002433
2434 new = mpol_new(mode, mode_flags, &nodes);
2435 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002436 goto out;
2437
Lee Schermerhorne17f74a2010-05-24 14:32:02 -07002438 if (no_context) {
2439 /* save for contextualization */
2440 new->w.user_nodemask = nodes;
2441 } else {
Miao Xie58568d22009-06-16 15:31:49 -07002442 int ret;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002443 NODEMASK_SCRATCH(scratch);
2444 if (scratch) {
2445 task_lock(current);
2446 ret = mpol_set_nodemask(new, &nodes, scratch);
2447 task_unlock(current);
2448 } else
2449 ret = -ENOMEM;
2450 NODEMASK_SCRATCH_FREE(scratch);
2451 if (ret) {
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002452 mpol_put(new);
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002453 goto out;
Miao Xie58568d22009-06-16 15:31:49 -07002454 }
2455 }
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002456 err = 0;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002457
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002458out:
2459 /* Restore string for error message */
2460 if (nodelist)
2461 *--nodelist = ':';
2462 if (flags)
2463 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002464 if (!err)
2465 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002466 return err;
2467}
2468#endif /* CONFIG_TMPFS */
2469
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002470/**
2471 * mpol_to_str - format a mempolicy structure for printing
2472 * @buffer: to contain formatted mempolicy string
2473 * @maxlen: length of @buffer
2474 * @pol: pointer to mempolicy to be formatted
2475 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2476 *
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002477 * Convert a mempolicy into a string.
2478 * Returns the number of characters in buffer (if positive)
2479 * or an error (negative)
2480 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002481int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002482{
2483 char *p = buffer;
2484 int l;
2485 nodemask_t nodes;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002486 unsigned short mode;
David Rientjesf5b087b2008-04-28 02:12:27 -07002487 unsigned short flags = pol ? pol->flags : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002488
Lee Schermerhorn22919902008-04-28 02:13:22 -07002489 /*
2490 * Sanity check: room for longest mode, flag and some nodes
2491 */
2492 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2493
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002494 if (!pol || pol == &default_policy)
2495 mode = MPOL_DEFAULT;
2496 else
2497 mode = pol->mode;
2498
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002499 switch (mode) {
2500 case MPOL_DEFAULT:
2501 nodes_clear(nodes);
2502 break;
2503
2504 case MPOL_PREFERRED:
2505 nodes_clear(nodes);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002506 if (flags & MPOL_F_LOCAL)
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002507 mode = MPOL_LOCAL; /* pseudo-policy */
2508 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002509 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002510 break;
2511
2512 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07002513 /* Fall through */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002514 case MPOL_INTERLEAVE:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002515 if (no_context)
2516 nodes = pol->w.user_nodemask;
2517 else
2518 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002519 break;
2520
2521 default:
2522 BUG();
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002523 }
2524
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002525 l = strlen(policy_modes[mode]);
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002526 if (buffer + maxlen < p + l + 1)
2527 return -ENOSPC;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002528
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002529 strcpy(p, policy_modes[mode]);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002530 p += l;
2531
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002532 if (flags & MPOL_MODE_FLAGS) {
David Rientjesf5b087b2008-04-28 02:12:27 -07002533 if (buffer + maxlen < p + 2)
2534 return -ENOSPC;
2535 *p++ = '=';
2536
Lee Schermerhorn22919902008-04-28 02:13:22 -07002537 /*
2538 * Currently, the only defined flags are mutually exclusive
2539 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002540 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002541 p += snprintf(p, buffer + maxlen - p, "static");
2542 else if (flags & MPOL_F_RELATIVE_NODES)
2543 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002544 }
2545
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002546 if (!nodes_empty(nodes)) {
2547 if (buffer + maxlen < p + 2)
2548 return -ENOSPC;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002549 *p++ = ':';
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002550 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2551 }
2552 return p - buffer;
2553}