blob: 8b57173c1dd5b09a8bb0b3e24fedb8c8584ce447 [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>
78#include <linux/module.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 */
Andi Kleend42c6992005-07-06 19:56:03 +0200114struct 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;
639 pgoff_t pgoff;
640 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
647 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800649 vmstart = max(start, vma->vm_start);
650 vmend = min(end, vma->vm_end);
651
652 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
653 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
654 vma->anon_vma, vma->vm_file, pgoff, new_pol);
655 if (prev) {
656 vma = prev;
657 next = vma->vm_next;
658 continue;
659 }
660 if (vma->vm_start != vmstart) {
661 err = split_vma(vma->vm_mm, vma, vmstart, 1);
662 if (err)
663 goto out;
664 }
665 if (vma->vm_end != vmend) {
666 err = split_vma(vma->vm_mm, vma, vmend, 0);
667 if (err)
668 goto out;
669 }
670 err = policy_vma(vma, new_pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 if (err)
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800672 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800674
675 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 return err;
677}
678
Paul Jacksonc61afb12006-03-24 03:16:08 -0800679/*
680 * Update task->flags PF_MEMPOLICY bit: set iff non-default
681 * mempolicy. Allows more rapid checking of this (combined perhaps
682 * with other PF_* flag bits) on memory allocation hot code paths.
683 *
684 * If called from outside this file, the task 'p' should -only- be
685 * a newly forked child not yet visible on the task list, because
686 * manipulating the task flags of a visible task is not safe.
687 *
688 * The above limitation is why this routine has the funny name
689 * mpol_fix_fork_child_flag().
690 *
691 * It is also safe to call this with a task pointer of current,
692 * which the static wrapper mpol_set_task_struct_flag() does,
693 * for use within this file.
694 */
695
696void mpol_fix_fork_child_flag(struct task_struct *p)
697{
698 if (p->mempolicy)
699 p->flags |= PF_MEMPOLICY;
700 else
701 p->flags &= ~PF_MEMPOLICY;
702}
703
704static void mpol_set_task_struct_flag(void)
705{
706 mpol_fix_fork_child_flag(current);
707}
708
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700710static long do_set_mempolicy(unsigned short mode, unsigned short flags,
711 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712{
Miao Xie58568d22009-06-16 15:31:49 -0700713 struct mempolicy *new, *old;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700714 struct mm_struct *mm = current->mm;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700715 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700716 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700718 if (!scratch)
719 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700720
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700721 new = mpol_new(mode, flags, nodes);
722 if (IS_ERR(new)) {
723 ret = PTR_ERR(new);
724 goto out;
725 }
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700726 /*
727 * prevent changing our mempolicy while show_numa_maps()
728 * is using it.
729 * Note: do_set_mempolicy() can be called at init time
730 * with no 'mm'.
731 */
732 if (mm)
733 down_write(&mm->mmap_sem);
Miao Xie58568d22009-06-16 15:31:49 -0700734 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700735 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700736 if (ret) {
737 task_unlock(current);
738 if (mm)
739 up_write(&mm->mmap_sem);
740 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700741 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700742 }
743 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 current->mempolicy = new;
Paul Jacksonc61afb12006-03-24 03:16:08 -0800745 mpol_set_task_struct_flag();
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700746 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700747 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700748 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700749 task_unlock(current);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700750 if (mm)
751 up_write(&mm->mmap_sem);
752
Miao Xie58568d22009-06-16 15:31:49 -0700753 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700754 ret = 0;
755out:
756 NODEMASK_SCRATCH_FREE(scratch);
757 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758}
759
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700760/*
761 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700762 *
763 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700764 */
765static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700767 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700768 if (p == &default_policy)
769 return;
770
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700771 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700772 case MPOL_BIND:
773 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700775 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 break;
777 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700778 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700779 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700780 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 break;
782 default:
783 BUG();
784 }
785}
786
787static int lookup_node(struct mm_struct *mm, unsigned long addr)
788{
789 struct page *p;
790 int err;
791
792 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
793 if (err >= 0) {
794 err = page_to_nid(p);
795 put_page(p);
796 }
797 return err;
798}
799
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700801static long do_get_mempolicy(int *policy, nodemask_t *nmask,
802 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700804 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 struct mm_struct *mm = current->mm;
806 struct vm_area_struct *vma = NULL;
807 struct mempolicy *pol = current->mempolicy;
808
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700809 if (flags &
810 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700812
813 if (flags & MPOL_F_MEMS_ALLOWED) {
814 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
815 return -EINVAL;
816 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700817 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700818 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700819 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700820 return 0;
821 }
822
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700824 /*
825 * Do NOT fall back to task policy if the
826 * vma/shared policy at addr is NULL. We
827 * want to return MPOL_DEFAULT in this case.
828 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 down_read(&mm->mmap_sem);
830 vma = find_vma_intersection(mm, addr, addr+1);
831 if (!vma) {
832 up_read(&mm->mmap_sem);
833 return -EFAULT;
834 }
835 if (vma->vm_ops && vma->vm_ops->get_policy)
836 pol = vma->vm_ops->get_policy(vma, addr);
837 else
838 pol = vma->vm_policy;
839 } else if (addr)
840 return -EINVAL;
841
842 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700843 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
845 if (flags & MPOL_F_NODE) {
846 if (flags & MPOL_F_ADDR) {
847 err = lookup_node(mm, addr);
848 if (err < 0)
849 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700850 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700852 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700853 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 } else {
855 err = -EINVAL;
856 goto out;
857 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700858 } else {
859 *policy = pol == &default_policy ? MPOL_DEFAULT :
860 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700861 /*
862 * Internal mempolicy flags must be masked off before exposing
863 * the policy to userspace.
864 */
865 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700866 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867
868 if (vma) {
869 up_read(&current->mm->mmap_sem);
870 vma = NULL;
871 }
872
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700874 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700875 if (mpol_store_user_nodemask(pol)) {
876 *nmask = pol->w.user_nodemask;
877 } else {
878 task_lock(current);
879 get_policy_nodemask(pol, nmask);
880 task_unlock(current);
881 }
Miao Xie58568d22009-06-16 15:31:49 -0700882 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883
884 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700885 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 if (vma)
887 up_read(&current->mm->mmap_sem);
888 return err;
889}
890
Christoph Lameterb20a3502006-03-22 00:09:12 -0800891#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700892/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800893 * page migration
894 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800895static void migrate_page_add(struct page *page, struct list_head *pagelist,
896 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800897{
898 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800899 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800900 */
Nick Piggin62695a82008-10-18 20:26:09 -0700901 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
902 if (!isolate_lru_page(page)) {
903 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800904 inc_zone_page_state(page, NR_ISOLATED_ANON +
905 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700906 }
907 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800908}
909
Christoph Lameter742755a2006-06-23 02:03:55 -0700910static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700911{
Mel Gorman6484eb32009-06-16 15:31:54 -0700912 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700913}
914
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800915/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800916 * Migrate pages from one node to a target node.
917 * Returns error or the number of pages not migrated.
918 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700919static int migrate_to_node(struct mm_struct *mm, int source, int dest,
920 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800921{
922 nodemask_t nmask;
923 LIST_HEAD(pagelist);
924 int err = 0;
Vasiliy Kulikov0def08e2010-10-26 14:21:32 -0700925 struct vm_area_struct *vma;
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800926
927 nodes_clear(nmask);
928 node_set(source, nmask);
929
Vasiliy Kulikov0def08e2010-10-26 14:21:32 -0700930 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800931 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
Vasiliy Kulikov0def08e2010-10-26 14:21:32 -0700932 if (IS_ERR(vma))
933 return PTR_ERR(vma);
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800934
Minchan Kimcf608ac2010-10-26 14:21:29 -0700935 if (!list_empty(&pagelist)) {
Mel Gorman7f0f2492011-01-13 15:45:58 -0800936 err = migrate_pages(&pagelist, new_node_page, dest,
937 false, true);
Minchan Kimcf608ac2010-10-26 14:21:29 -0700938 if (err)
939 putback_lru_pages(&pagelist);
940 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700941
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800942 return err;
943}
944
945/*
946 * Move pages between the two nodesets so as to preserve the physical
947 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -0800948 *
949 * Returns the number of page that could not be moved.
950 */
951int do_migrate_pages(struct mm_struct *mm,
952 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
953{
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800954 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800955 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800956 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -0800957
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800958 err = migrate_prep();
959 if (err)
960 return err;
961
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700962 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -0800963
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700964 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
965 if (err)
966 goto out;
967
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -0800968 /*
969 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
970 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
971 * bit in 'tmp', and return that <source, dest> pair for migration.
972 * The pair of nodemasks 'to' and 'from' define the map.
973 *
974 * If no pair of bits is found that way, fallback to picking some
975 * pair of 'source' and 'dest' bits that are not the same. If the
976 * 'source' and 'dest' bits are the same, this represents a node
977 * that will be migrating to itself, so no pages need move.
978 *
979 * If no bits are left in 'tmp', or if all remaining bits left
980 * in 'tmp' correspond to the same bit in 'to', return false
981 * (nothing left to migrate).
982 *
983 * This lets us pick a pair of nodes to migrate between, such that
984 * if possible the dest node is not already occupied by some other
985 * source node, minimizing the risk of overloading the memory on a
986 * node that would happen if we migrated incoming memory to a node
987 * before migrating outgoing memory source that same node.
988 *
989 * A single scan of tmp is sufficient. As we go, we remember the
990 * most recent <s, d> pair that moved (s != d). If we find a pair
991 * that not only moved, but what's better, moved to an empty slot
992 * (d is not set in tmp), then we break out then, with that pair.
Justin P. Mattockae0e47f2011-03-01 15:06:02 +0100993 * Otherwise when we finish scanning from_tmp, we at least have the
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -0800994 * most recent <s, d> pair that moved. If we get all the way through
995 * the scan of tmp without finding any node that moved, much less
996 * moved to an empty node, then there is nothing left worth migrating.
997 */
Christoph Lameterd4984712006-01-08 01:00:55 -0800998
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800999 tmp = *from_nodes;
1000 while (!nodes_empty(tmp)) {
1001 int s,d;
1002 int source = -1;
1003 int dest = 0;
1004
1005 for_each_node_mask(s, tmp) {
1006 d = node_remap(s, *from_nodes, *to_nodes);
1007 if (s == d)
1008 continue;
1009
1010 source = s; /* Node moved. Memorize */
1011 dest = d;
1012
1013 /* dest not in remaining from nodes? */
1014 if (!node_isset(dest, tmp))
1015 break;
1016 }
1017 if (source == -1)
1018 break;
1019
1020 node_clear(source, tmp);
1021 err = migrate_to_node(mm, source, dest, flags);
1022 if (err > 0)
1023 busy += err;
1024 if (err < 0)
1025 break;
Christoph Lameter39743882006-01-08 01:00:51 -08001026 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001027out:
Christoph Lameter39743882006-01-08 01:00:51 -08001028 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001029 if (err < 0)
1030 return err;
1031 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001032
Christoph Lameter39743882006-01-08 01:00:51 -08001033}
1034
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001035/*
1036 * Allocate a new page for page migration based on vma policy.
1037 * Start assuming that page is mapped by vma pointed to by @private.
1038 * Search forward from there, if not. N.B., this assumes that the
1039 * list of pages handed to migrate_pages()--which is how we get here--
1040 * is in virtual address order.
1041 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001042static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001043{
1044 struct vm_area_struct *vma = (struct vm_area_struct *)private;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001045 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001046
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001047 while (vma) {
1048 address = page_address_in_vma(page, vma);
1049 if (address != -EFAULT)
1050 break;
1051 vma = vma->vm_next;
1052 }
1053
1054 /*
1055 * if !vma, alloc_page_vma() will use task or system default policy
1056 */
1057 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001058}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001059#else
1060
1061static void migrate_page_add(struct page *page, struct list_head *pagelist,
1062 unsigned long flags)
1063{
1064}
1065
1066int do_migrate_pages(struct mm_struct *mm,
1067 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1068{
1069 return -ENOSYS;
1070}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001071
Keith Owens69939742006-10-11 01:21:28 -07001072static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001073{
1074 return NULL;
1075}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001076#endif
1077
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001078static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001079 unsigned short mode, unsigned short mode_flags,
1080 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001081{
1082 struct vm_area_struct *vma;
1083 struct mm_struct *mm = current->mm;
1084 struct mempolicy *new;
1085 unsigned long end;
1086 int err;
1087 LIST_HEAD(pagelist);
1088
David Rientjesa3b51e02008-04-28 02:12:23 -07001089 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1090 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001091 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001092 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001093 return -EPERM;
1094
1095 if (start & ~PAGE_MASK)
1096 return -EINVAL;
1097
1098 if (mode == MPOL_DEFAULT)
1099 flags &= ~MPOL_MF_STRICT;
1100
1101 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1102 end = start + len;
1103
1104 if (end < start)
1105 return -EINVAL;
1106 if (end == start)
1107 return 0;
1108
David Rientjes028fec42008-04-28 02:12:25 -07001109 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001110 if (IS_ERR(new))
1111 return PTR_ERR(new);
1112
1113 /*
1114 * If we are using the default policy then operation
1115 * on discontinuous address spaces is okay after all
1116 */
1117 if (!new)
1118 flags |= MPOL_MF_DISCONTIG_OK;
1119
David Rientjes028fec42008-04-28 02:12:25 -07001120 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1121 start, start + len, mode, mode_flags,
1122 nmask ? nodes_addr(*nmask)[0] : -1);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001123
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001124 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1125
1126 err = migrate_prep();
1127 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001128 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001129 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001130 {
1131 NODEMASK_SCRATCH(scratch);
1132 if (scratch) {
1133 down_write(&mm->mmap_sem);
1134 task_lock(current);
1135 err = mpol_set_nodemask(new, nmask, scratch);
1136 task_unlock(current);
1137 if (err)
1138 up_write(&mm->mmap_sem);
1139 } else
1140 err = -ENOMEM;
1141 NODEMASK_SCRATCH_FREE(scratch);
1142 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001143 if (err)
1144 goto mpol_out;
1145
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001146 vma = check_range(mm, start, end, nmask,
1147 flags | MPOL_MF_INVERT, &pagelist);
1148
1149 err = PTR_ERR(vma);
1150 if (!IS_ERR(vma)) {
1151 int nr_failed = 0;
1152
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001153 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001154
Minchan Kimcf608ac2010-10-26 14:21:29 -07001155 if (!list_empty(&pagelist)) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001156 nr_failed = migrate_pages(&pagelist, new_vma_page,
Mel Gorman7f0f2492011-01-13 15:45:58 -08001157 (unsigned long)vma,
1158 false, true);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001159 if (nr_failed)
1160 putback_lru_pages(&pagelist);
1161 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001162
1163 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1164 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001165 } else
1166 putback_lru_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001167
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001168 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001169 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001170 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001171 return err;
1172}
1173
Christoph Lameter39743882006-01-08 01:00:51 -08001174/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001175 * User space interface with variable sized bitmaps for nodelists.
1176 */
1177
1178/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001179static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001180 unsigned long maxnode)
1181{
1182 unsigned long k;
1183 unsigned long nlongs;
1184 unsigned long endmask;
1185
1186 --maxnode;
1187 nodes_clear(*nodes);
1188 if (maxnode == 0 || !nmask)
1189 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001190 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001191 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001192
1193 nlongs = BITS_TO_LONGS(maxnode);
1194 if ((maxnode % BITS_PER_LONG) == 0)
1195 endmask = ~0UL;
1196 else
1197 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1198
1199 /* When the user specified more nodes than supported just check
1200 if the non supported part is all zero. */
1201 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1202 if (nlongs > PAGE_SIZE/sizeof(long))
1203 return -EINVAL;
1204 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1205 unsigned long t;
1206 if (get_user(t, nmask + k))
1207 return -EFAULT;
1208 if (k == nlongs - 1) {
1209 if (t & endmask)
1210 return -EINVAL;
1211 } else if (t)
1212 return -EINVAL;
1213 }
1214 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1215 endmask = ~0UL;
1216 }
1217
1218 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1219 return -EFAULT;
1220 nodes_addr(*nodes)[nlongs-1] &= endmask;
1221 return 0;
1222}
1223
1224/* Copy a kernel node mask to user space */
1225static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1226 nodemask_t *nodes)
1227{
1228 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1229 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1230
1231 if (copy > nbytes) {
1232 if (copy > PAGE_SIZE)
1233 return -EINVAL;
1234 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1235 return -EFAULT;
1236 copy = nbytes;
1237 }
1238 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1239}
1240
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001241SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1242 unsigned long, mode, unsigned long __user *, nmask,
1243 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001244{
1245 nodemask_t nodes;
1246 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001247 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001248
David Rientjes028fec42008-04-28 02:12:25 -07001249 mode_flags = mode & MPOL_MODE_FLAGS;
1250 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001251 if (mode >= MPOL_MAX)
1252 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001253 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1254 (mode_flags & MPOL_F_RELATIVE_NODES))
1255 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001256 err = get_nodes(&nodes, nmask, maxnode);
1257 if (err)
1258 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001259 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001260}
1261
1262/* Set the process memory policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001263SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1264 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001265{
1266 int err;
1267 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001268 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001269
David Rientjes028fec42008-04-28 02:12:25 -07001270 flags = mode & MPOL_MODE_FLAGS;
1271 mode &= ~MPOL_MODE_FLAGS;
1272 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001273 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001274 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1275 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001276 err = get_nodes(&nodes, nmask, maxnode);
1277 if (err)
1278 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001279 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001280}
1281
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001282SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1283 const unsigned long __user *, old_nodes,
1284 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001285{
David Howellsc69e8d92008-11-14 10:39:19 +11001286 const struct cred *cred = current_cred(), *tcred;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001287 struct mm_struct *mm = NULL;
Christoph Lameter39743882006-01-08 01:00:51 -08001288 struct task_struct *task;
Christoph Lameter39743882006-01-08 01:00:51 -08001289 nodemask_t task_nodes;
1290 int err;
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001291 nodemask_t *old;
1292 nodemask_t *new;
1293 NODEMASK_SCRATCH(scratch);
Christoph Lameter39743882006-01-08 01:00:51 -08001294
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001295 if (!scratch)
1296 return -ENOMEM;
Christoph Lameter39743882006-01-08 01:00:51 -08001297
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001298 old = &scratch->mask1;
1299 new = &scratch->mask2;
1300
1301 err = get_nodes(old, old_nodes, maxnode);
Christoph Lameter39743882006-01-08 01:00:51 -08001302 if (err)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001303 goto out;
1304
1305 err = get_nodes(new, new_nodes, maxnode);
1306 if (err)
1307 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001308
1309 /* Find the mm_struct */
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001310 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001311 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001312 if (!task) {
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001313 rcu_read_unlock();
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001314 err = -ESRCH;
1315 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001316 }
1317 mm = get_task_mm(task);
Zeng Zhaoming55cfaa32010-12-02 14:31:13 -08001318 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001319
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001320 err = -EINVAL;
Christoph Lameter39743882006-01-08 01:00:51 -08001321 if (!mm)
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001322 goto out;
Christoph Lameter39743882006-01-08 01:00:51 -08001323
1324 /*
1325 * Check if this process has the right to modify the specified
1326 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001327 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001328 * userid as the target process.
1329 */
David Howellsc69e8d92008-11-14 10:39:19 +11001330 rcu_read_lock();
1331 tcred = __task_cred(task);
David Howellsb6dff3e2008-11-14 10:39:16 +11001332 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1333 cred->uid != tcred->suid && cred->uid != tcred->uid &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001334 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001335 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001336 err = -EPERM;
1337 goto out;
1338 }
David Howellsc69e8d92008-11-14 10:39:19 +11001339 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001340
1341 task_nodes = cpuset_mems_allowed(task);
1342 /* Is the user allowed to access the target nodes? */
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001343 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001344 err = -EPERM;
1345 goto out;
1346 }
1347
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001348 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001349 err = -EINVAL;
1350 goto out;
1351 }
1352
David Quigley86c3a762006-06-23 02:04:02 -07001353 err = security_task_movememory(task);
1354 if (err)
1355 goto out;
1356
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001357 err = do_migrate_pages(mm, old, new,
Christoph Lameter74c00242006-03-14 19:50:21 -08001358 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter39743882006-01-08 01:00:51 -08001359out:
KOSAKI Motohiro596d7cf2010-08-09 17:19:01 -07001360 if (mm)
1361 mmput(mm);
1362 NODEMASK_SCRATCH_FREE(scratch);
1363
Christoph Lameter39743882006-01-08 01:00:51 -08001364 return err;
1365}
1366
1367
Christoph Lameter8bccd852005-10-29 18:16:59 -07001368/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001369SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1370 unsigned long __user *, nmask, unsigned long, maxnode,
1371 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001372{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001373 int err;
1374 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001375 nodemask_t nodes;
1376
1377 if (nmask != NULL && maxnode < MAX_NUMNODES)
1378 return -EINVAL;
1379
1380 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1381
1382 if (err)
1383 return err;
1384
1385 if (policy && put_user(pval, policy))
1386 return -EFAULT;
1387
1388 if (nmask)
1389 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1390
1391 return err;
1392}
1393
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394#ifdef CONFIG_COMPAT
1395
1396asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1397 compat_ulong_t __user *nmask,
1398 compat_ulong_t maxnode,
1399 compat_ulong_t addr, compat_ulong_t flags)
1400{
1401 long err;
1402 unsigned long __user *nm = NULL;
1403 unsigned long nr_bits, alloc_size;
1404 DECLARE_BITMAP(bm, MAX_NUMNODES);
1405
1406 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1407 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1408
1409 if (nmask)
1410 nm = compat_alloc_user_space(alloc_size);
1411
1412 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1413
1414 if (!err && nmask) {
1415 err = copy_from_user(bm, nm, alloc_size);
1416 /* ensure entire bitmap is zeroed */
1417 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1418 err |= compat_put_bitmap(nmask, bm, nr_bits);
1419 }
1420
1421 return err;
1422}
1423
1424asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1425 compat_ulong_t maxnode)
1426{
1427 long err = 0;
1428 unsigned long __user *nm = NULL;
1429 unsigned long nr_bits, alloc_size;
1430 DECLARE_BITMAP(bm, MAX_NUMNODES);
1431
1432 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1433 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1434
1435 if (nmask) {
1436 err = compat_get_bitmap(bm, nmask, nr_bits);
1437 nm = compat_alloc_user_space(alloc_size);
1438 err |= copy_to_user(nm, bm, alloc_size);
1439 }
1440
1441 if (err)
1442 return -EFAULT;
1443
1444 return sys_set_mempolicy(mode, nm, nr_bits+1);
1445}
1446
1447asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1448 compat_ulong_t mode, compat_ulong_t __user *nmask,
1449 compat_ulong_t maxnode, compat_ulong_t flags)
1450{
1451 long err = 0;
1452 unsigned long __user *nm = NULL;
1453 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001454 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455
1456 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1457 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1458
1459 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001460 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001462 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 }
1464
1465 if (err)
1466 return -EFAULT;
1467
1468 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1469}
1470
1471#endif
1472
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001473/*
1474 * get_vma_policy(@task, @vma, @addr)
1475 * @task - task for fallback if vma policy == default
1476 * @vma - virtual memory area whose policy is sought
1477 * @addr - address in @vma for shared policy lookup
1478 *
1479 * Returns effective policy for a VMA at specified address.
1480 * Falls back to @task or system default policy, as necessary.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001481 * Current or other task's task mempolicy and non-shared vma policies
1482 * are protected by the task's mmap_sem, which must be held for read by
1483 * the caller.
1484 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1485 * count--added by the get_policy() vm_op, as appropriate--to protect against
1486 * freeing by another task. It is the caller's responsibility to free the
1487 * extra reference for shared policies.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001488 */
Stephen Wilsond98f6cb2011-05-24 17:12:41 -07001489struct mempolicy *get_vma_policy(struct task_struct *task,
Christoph Lameter48fce342006-01-08 01:01:03 -08001490 struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001492 struct mempolicy *pol = task->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493
1494 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001495 if (vma->vm_ops && vma->vm_ops->get_policy) {
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001496 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1497 addr);
1498 if (vpol)
1499 pol = vpol;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001500 } else if (vma->vm_policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 pol = vma->vm_policy;
1502 }
1503 if (!pol)
1504 pol = &default_policy;
1505 return pol;
1506}
1507
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001508/*
1509 * Return a nodemask representing a mempolicy for filtering nodes for
1510 * page allocation
1511 */
1512static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001513{
1514 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001515 if (unlikely(policy->mode == MPOL_BIND) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001516 gfp_zone(gfp) >= policy_zone &&
1517 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1518 return &policy->v.nodes;
1519
1520 return NULL;
1521}
1522
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001523/* Return a zonelist indicated by gfp for node representing a mempolicy */
Andi Kleen2f5f9482011-03-04 17:36:29 -08001524static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1525 int nd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526{
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001527 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001529 if (!(policy->flags & MPOL_F_LOCAL))
1530 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 break;
1532 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001533 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001534 * Normally, MPOL_BIND allocations are node-local within the
1535 * allowed nodemask. However, if __GFP_THISNODE is set and the
Bob Liu6eb27e12010-05-24 14:32:00 -07001536 * current node isn't part of the mask, we use the zonelist for
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001537 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001538 */
Mel Gorman19770b32008-04-28 02:12:18 -07001539 if (unlikely(gfp & __GFP_THISNODE) &&
1540 unlikely(!node_isset(nd, policy->v.nodes)))
1541 nd = first_node(policy->v.nodes);
1542 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 BUG();
1545 }
Mel Gorman0e884602008-04-28 02:12:14 -07001546 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547}
1548
1549/* Do dynamic interleaving for a process */
1550static unsigned interleave_nodes(struct mempolicy *policy)
1551{
1552 unsigned nid, next;
1553 struct task_struct *me = current;
1554
1555 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001556 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001558 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001559 if (next < MAX_NUMNODES)
1560 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 return nid;
1562}
1563
Christoph Lameterdc85da12006-01-18 17:42:36 -08001564/*
1565 * Depending on the memory policy provide a node from which to allocate the
1566 * next slab entry.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001567 * @policy must be protected by freeing by the caller. If @policy is
1568 * the current task's mempolicy, this protection is implicit, as only the
1569 * task can change it's policy. The system default policy requires no
1570 * such protection.
Christoph Lameterdc85da12006-01-18 17:42:36 -08001571 */
1572unsigned slab_node(struct mempolicy *policy)
1573{
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001574 if (!policy || policy->flags & MPOL_F_LOCAL)
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001575 return numa_node_id();
Christoph Lameter765c4502006-09-27 01:50:08 -07001576
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001577 switch (policy->mode) {
1578 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001579 /*
1580 * handled MPOL_F_LOCAL above
1581 */
1582 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001583
Christoph Lameterdc85da12006-01-18 17:42:36 -08001584 case MPOL_INTERLEAVE:
1585 return interleave_nodes(policy);
1586
Mel Gormandd1a2392008-04-28 02:12:17 -07001587 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001588 /*
1589 * Follow bind policy behavior and start allocation at the
1590 * first node.
1591 */
Mel Gorman19770b32008-04-28 02:12:18 -07001592 struct zonelist *zonelist;
1593 struct zone *zone;
1594 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1595 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1596 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1597 &policy->v.nodes,
1598 &zone);
Eric Dumazet800416f2010-10-27 19:33:43 +02001599 return zone ? zone->node : numa_node_id();
Mel Gormandd1a2392008-04-28 02:12:17 -07001600 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001601
Christoph Lameterdc85da12006-01-18 17:42:36 -08001602 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001603 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001604 }
1605}
1606
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607/* Do static interleaving for a VMA with known offset. */
1608static unsigned offset_il_node(struct mempolicy *pol,
1609 struct vm_area_struct *vma, unsigned long off)
1610{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001611 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001612 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 int c;
1614 int nid = -1;
1615
David Rientjesf5b087b2008-04-28 02:12:27 -07001616 if (!nnodes)
1617 return numa_node_id();
1618 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 c = 0;
1620 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001621 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622 c++;
1623 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 return nid;
1625}
1626
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001627/* Determine a node number for interleave */
1628static inline unsigned interleave_nid(struct mempolicy *pol,
1629 struct vm_area_struct *vma, unsigned long addr, int shift)
1630{
1631 if (vma) {
1632 unsigned long off;
1633
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001634 /*
1635 * for small pages, there is no difference between
1636 * shift and PAGE_SHIFT, so the bit-shift is safe.
1637 * for huge pages, since vm_pgoff is in units of small
1638 * pages, we need to shift off the always 0 bits to get
1639 * a useful offset.
1640 */
1641 BUG_ON(shift < PAGE_SHIFT);
1642 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001643 off += (addr - vma->vm_start) >> shift;
1644 return offset_il_node(pol, vma, off);
1645 } else
1646 return interleave_nodes(pol);
1647}
1648
Michal Hocko778d3b02011-07-26 16:08:30 -07001649/*
1650 * Return the bit number of a random bit set in the nodemask.
1651 * (returns -1 if nodemask is empty)
1652 */
1653int node_random(const nodemask_t *maskp)
1654{
1655 int w, bit = -1;
1656
1657 w = nodes_weight(*maskp);
1658 if (w)
1659 bit = bitmap_ord_to_pos(maskp->bits,
1660 get_random_int() % w, MAX_NUMNODES);
1661 return bit;
1662}
1663
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001664#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001665/*
1666 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1667 * @vma = virtual memory area whose policy is sought
1668 * @addr = address in @vma for shared policy lookup and interleave policy
1669 * @gfp_flags = for requested zone
Mel Gorman19770b32008-04-28 02:12:18 -07001670 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1671 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001672 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001673 * Returns a zonelist suitable for a huge page allocation and a pointer
1674 * to the struct mempolicy for conditional unref after allocation.
1675 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1676 * @nodemask for filtering the zonelist.
Miao Xiec0ff7452010-05-24 14:32:08 -07001677 *
1678 * Must be protected by get_mems_allowed()
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001679 */
Mel Gorman396faf02007-07-17 04:03:13 -07001680struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001681 gfp_t gfp_flags, struct mempolicy **mpol,
1682 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001683{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001684 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001685
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001686 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001687 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001688
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001689 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1690 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001691 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001692 } else {
Andi Kleen2f5f9482011-03-04 17:36:29 -08001693 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001694 if ((*mpol)->mode == MPOL_BIND)
1695 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001696 }
1697 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001698}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001699
1700/*
1701 * init_nodemask_of_mempolicy
1702 *
1703 * If the current task's mempolicy is "default" [NULL], return 'false'
1704 * to indicate default policy. Otherwise, extract the policy nodemask
1705 * for 'bind' or 'interleave' policy into the argument nodemask, or
1706 * initialize the argument nodemask to contain the single node for
1707 * 'preferred' or 'local' policy and return 'true' to indicate presence
1708 * of non-default mempolicy.
1709 *
1710 * We don't bother with reference counting the mempolicy [mpol_get/put]
1711 * because the current task is examining it's own mempolicy and a task's
1712 * mempolicy is only ever changed by the task itself.
1713 *
1714 * N.B., it is the caller's responsibility to free a returned nodemask.
1715 */
1716bool init_nodemask_of_mempolicy(nodemask_t *mask)
1717{
1718 struct mempolicy *mempolicy;
1719 int nid;
1720
1721 if (!(mask && current->mempolicy))
1722 return false;
1723
Miao Xiec0ff7452010-05-24 14:32:08 -07001724 task_lock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001725 mempolicy = current->mempolicy;
1726 switch (mempolicy->mode) {
1727 case MPOL_PREFERRED:
1728 if (mempolicy->flags & MPOL_F_LOCAL)
1729 nid = numa_node_id();
1730 else
1731 nid = mempolicy->v.preferred_node;
1732 init_nodemask_of_node(mask, nid);
1733 break;
1734
1735 case MPOL_BIND:
1736 /* Fall through */
1737 case MPOL_INTERLEAVE:
1738 *mask = mempolicy->v.nodes;
1739 break;
1740
1741 default:
1742 BUG();
1743 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001744 task_unlock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001745
1746 return true;
1747}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001748#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001749
David Rientjes6f48d0eb2010-08-09 17:18:52 -07001750/*
1751 * mempolicy_nodemask_intersects
1752 *
1753 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1754 * policy. Otherwise, check for intersection between mask and the policy
1755 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1756 * policy, always return true since it may allocate elsewhere on fallback.
1757 *
1758 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1759 */
1760bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1761 const nodemask_t *mask)
1762{
1763 struct mempolicy *mempolicy;
1764 bool ret = true;
1765
1766 if (!mask)
1767 return ret;
1768 task_lock(tsk);
1769 mempolicy = tsk->mempolicy;
1770 if (!mempolicy)
1771 goto out;
1772
1773 switch (mempolicy->mode) {
1774 case MPOL_PREFERRED:
1775 /*
1776 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1777 * allocate from, they may fallback to other nodes when oom.
1778 * Thus, it's possible for tsk to have allocated memory from
1779 * nodes in mask.
1780 */
1781 break;
1782 case MPOL_BIND:
1783 case MPOL_INTERLEAVE:
1784 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1785 break;
1786 default:
1787 BUG();
1788 }
1789out:
1790 task_unlock(tsk);
1791 return ret;
1792}
1793
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794/* Allocate a page in interleaved policy.
1795 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001796static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1797 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798{
1799 struct zonelist *zl;
1800 struct page *page;
1801
Mel Gorman0e884602008-04-28 02:12:14 -07001802 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001804 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001805 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806 return page;
1807}
1808
1809/**
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001810 * alloc_pages_vma - Allocate a page for a VMA.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 *
1812 * @gfp:
1813 * %GFP_USER user allocation.
1814 * %GFP_KERNEL kernel allocations,
1815 * %GFP_HIGHMEM highmem/user allocations,
1816 * %GFP_FS allocation should not call back into a file system.
1817 * %GFP_ATOMIC don't sleep.
1818 *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001819 * @order:Order of the GFP allocation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 * @vma: Pointer to VMA or NULL if not available.
1821 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1822 *
1823 * This function allocates a page from the kernel page pool and applies
1824 * a NUMA policy associated with the VMA or the current process.
1825 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1826 * mm_struct of the VMA to prevent it from going away. Should be used for
1827 * all allocations for pages that will be mapped into
1828 * user space. Returns NULL when no page can be allocated.
1829 *
1830 * Should be called with the mm_sem of the vma hold.
1831 */
1832struct page *
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001833alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
Andi Kleen2f5f9482011-03-04 17:36:29 -08001834 unsigned long addr, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001836 struct mempolicy *pol = get_vma_policy(current, vma, addr);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001837 struct zonelist *zl;
Miao Xiec0ff7452010-05-24 14:32:08 -07001838 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839
Miao Xiec0ff7452010-05-24 14:32:08 -07001840 get_mems_allowed();
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001841 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001843
Andi Kleen8eac5632011-02-25 14:44:28 -08001844 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001845 mpol_cond_put(pol);
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001846 page = alloc_page_interleave(gfp, order, nid);
Miao Xiec0ff7452010-05-24 14:32:08 -07001847 put_mems_allowed();
1848 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 }
Andi Kleen2f5f9482011-03-04 17:36:29 -08001850 zl = policy_zonelist(gfp, pol, node);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001851 if (unlikely(mpol_needs_cond_ref(pol))) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001852 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001853 * slow path: ref counted shared policy
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001854 */
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001855 struct page *page = __alloc_pages_nodemask(gfp, order,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001856 zl, policy_nodemask(gfp, pol));
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001857 __mpol_put(pol);
Miao Xiec0ff7452010-05-24 14:32:08 -07001858 put_mems_allowed();
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001859 return page;
1860 }
1861 /*
1862 * fast path: default or task policy
1863 */
Andrea Arcangeli0bbbc0b2011-01-13 15:47:05 -08001864 page = __alloc_pages_nodemask(gfp, order, zl,
1865 policy_nodemask(gfp, pol));
Miao Xiec0ff7452010-05-24 14:32:08 -07001866 put_mems_allowed();
1867 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868}
1869
1870/**
1871 * alloc_pages_current - Allocate pages.
1872 *
1873 * @gfp:
1874 * %GFP_USER user allocation,
1875 * %GFP_KERNEL kernel allocation,
1876 * %GFP_HIGHMEM highmem allocation,
1877 * %GFP_FS don't call back into a file system.
1878 * %GFP_ATOMIC don't sleep.
1879 * @order: Power of two of allocation size in pages. 0 is a single page.
1880 *
1881 * Allocate a page from the kernel page pool. When not in
1882 * interrupt context and apply the current process NUMA policy.
1883 * Returns NULL when no page can be allocated.
1884 *
Paul Jacksoncf2a4732006-01-08 01:01:54 -08001885 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 * 1) it's ok to take cpuset_sem (can WAIT), and
1887 * 2) allocating for current task (not interrupt).
1888 */
Al Virodd0fc662005-10-07 07:46:04 +01001889struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890{
1891 struct mempolicy *pol = current->mempolicy;
Miao Xiec0ff7452010-05-24 14:32:08 -07001892 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893
Christoph Lameter9b819d22006-09-25 23:31:40 -07001894 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 pol = &default_policy;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001896
Miao Xiec0ff7452010-05-24 14:32:08 -07001897 get_mems_allowed();
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001898 /*
1899 * No reference counting needed for current->mempolicy
1900 * nor system default_policy
1901 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001902 if (pol->mode == MPOL_INTERLEAVE)
Miao Xiec0ff7452010-05-24 14:32:08 -07001903 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1904 else
1905 page = __alloc_pages_nodemask(gfp, order,
Andi Kleen5c4b4be2011-03-04 17:36:32 -08001906 policy_zonelist(gfp, pol, numa_node_id()),
1907 policy_nodemask(gfp, pol));
Miao Xiec0ff7452010-05-24 14:32:08 -07001908 put_mems_allowed();
1909 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910}
1911EXPORT_SYMBOL(alloc_pages_current);
1912
Paul Jackson42253992006-01-08 01:01:59 -08001913/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001914 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08001915 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1916 * with the mems_allowed returned by cpuset_mems_allowed(). This
1917 * keeps mempolicies cpuset relative after its cpuset moves. See
1918 * further kernel/cpuset.c update_nodemask().
Miao Xie708c1bb2010-05-24 14:32:07 -07001919 *
1920 * current's mempolicy may be rebinded by the other task(the task that changes
1921 * cpuset's mems), so we needn't do rebind work for current task.
Paul Jackson42253992006-01-08 01:01:59 -08001922 */
Paul Jackson42253992006-01-08 01:01:59 -08001923
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001924/* Slow path of a mempolicy duplicate */
1925struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926{
1927 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1928
1929 if (!new)
1930 return ERR_PTR(-ENOMEM);
Miao Xie708c1bb2010-05-24 14:32:07 -07001931
1932 /* task's mempolicy is protected by alloc_lock */
1933 if (old == current->mempolicy) {
1934 task_lock(current);
1935 *new = *old;
1936 task_unlock(current);
1937 } else
1938 *new = *old;
1939
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08001940 rcu_read_lock();
Paul Jackson42253992006-01-08 01:01:59 -08001941 if (current_cpuset_is_being_rebound()) {
1942 nodemask_t mems = cpuset_mems_allowed(current);
Miao Xie708c1bb2010-05-24 14:32:07 -07001943 if (new->flags & MPOL_F_REBINDING)
1944 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1945 else
1946 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
Paul Jackson42253992006-01-08 01:01:59 -08001947 }
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08001948 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 return new;
1951}
1952
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001953/*
1954 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1955 * eliminate the * MPOL_F_* flags that require conditional ref and
1956 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1957 * after return. Use the returned value.
1958 *
1959 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1960 * policy lookup, even if the policy needs/has extra ref on lookup.
1961 * shmem_readahead needs this.
1962 */
1963struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1964 struct mempolicy *frompol)
1965{
1966 if (!mpol_needs_cond_ref(frompol))
1967 return frompol;
1968
1969 *tompol = *frompol;
1970 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1971 __mpol_put(frompol);
1972 return tompol;
1973}
1974
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975/* Slow path of a mempolicy comparison */
1976int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1977{
1978 if (!a || !b)
1979 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001980 if (a->mode != b->mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 return 0;
Bob Liu19800502010-05-24 14:32:01 -07001982 if (a->flags != b->flags)
David Rientjesf5b087b2008-04-28 02:12:27 -07001983 return 0;
Bob Liu19800502010-05-24 14:32:01 -07001984 if (mpol_store_user_nodemask(a))
1985 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
1986 return 0;
1987
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001988 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07001989 case MPOL_BIND:
1990 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -07001992 return nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 case MPOL_PREFERRED:
Namhyung Kim75719662011-03-22 16:33:02 -07001994 return a->v.preferred_node == b->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 default:
1996 BUG();
1997 return 0;
1998 }
1999}
2000
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 * Shared memory backing store policy support.
2003 *
2004 * Remember policies even when nobody has shared memory mapped.
2005 * The policies are kept in Red-Black tree linked from the inode.
2006 * They are protected by the sp->lock spinlock, which should be held
2007 * for any accesses to the tree.
2008 */
2009
2010/* lookup first element intersecting start-end */
2011/* Caller holds sp->lock */
2012static struct sp_node *
2013sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2014{
2015 struct rb_node *n = sp->root.rb_node;
2016
2017 while (n) {
2018 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2019
2020 if (start >= p->end)
2021 n = n->rb_right;
2022 else if (end <= p->start)
2023 n = n->rb_left;
2024 else
2025 break;
2026 }
2027 if (!n)
2028 return NULL;
2029 for (;;) {
2030 struct sp_node *w = NULL;
2031 struct rb_node *prev = rb_prev(n);
2032 if (!prev)
2033 break;
2034 w = rb_entry(prev, struct sp_node, nd);
2035 if (w->end <= start)
2036 break;
2037 n = prev;
2038 }
2039 return rb_entry(n, struct sp_node, nd);
2040}
2041
2042/* Insert a new shared policy into the list. */
2043/* Caller holds sp->lock */
2044static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2045{
2046 struct rb_node **p = &sp->root.rb_node;
2047 struct rb_node *parent = NULL;
2048 struct sp_node *nd;
2049
2050 while (*p) {
2051 parent = *p;
2052 nd = rb_entry(parent, struct sp_node, nd);
2053 if (new->start < nd->start)
2054 p = &(*p)->rb_left;
2055 else if (new->end > nd->end)
2056 p = &(*p)->rb_right;
2057 else
2058 BUG();
2059 }
2060 rb_link_node(&new->nd, parent, p);
2061 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07002062 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002063 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064}
2065
2066/* Find shared policy intersecting idx */
2067struct mempolicy *
2068mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2069{
2070 struct mempolicy *pol = NULL;
2071 struct sp_node *sn;
2072
2073 if (!sp->root.rb_node)
2074 return NULL;
2075 spin_lock(&sp->lock);
2076 sn = sp_lookup(sp, idx, idx+1);
2077 if (sn) {
2078 mpol_get(sn->policy);
2079 pol = sn->policy;
2080 }
2081 spin_unlock(&sp->lock);
2082 return pol;
2083}
2084
2085static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2086{
Paul Mundt140d5a42007-07-15 23:38:16 -07002087 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 rb_erase(&n->nd, &sp->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002089 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 kmem_cache_free(sn_cache, n);
2091}
2092
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07002093static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2094 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095{
2096 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2097
2098 if (!n)
2099 return NULL;
2100 n->start = start;
2101 n->end = end;
2102 mpol_get(pol);
Lee Schermerhornaab0b102008-04-28 02:13:13 -07002103 pol->flags |= MPOL_F_SHARED; /* for unref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 n->policy = pol;
2105 return n;
2106}
2107
2108/* Replace a policy range. */
2109static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2110 unsigned long end, struct sp_node *new)
2111{
2112 struct sp_node *n, *new2 = NULL;
2113
2114restart:
2115 spin_lock(&sp->lock);
2116 n = sp_lookup(sp, start, end);
2117 /* Take care of old policies in the same range. */
2118 while (n && n->start < end) {
2119 struct rb_node *next = rb_next(&n->nd);
2120 if (n->start >= start) {
2121 if (n->end <= end)
2122 sp_delete(sp, n);
2123 else
2124 n->start = end;
2125 } else {
2126 /* Old policy spanning whole new range. */
2127 if (n->end > end) {
2128 if (!new2) {
2129 spin_unlock(&sp->lock);
2130 new2 = sp_alloc(end, n->end, n->policy);
2131 if (!new2)
2132 return -ENOMEM;
2133 goto restart;
2134 }
2135 n->end = start;
2136 sp_insert(sp, new2);
2137 new2 = NULL;
2138 break;
2139 } else
2140 n->end = start;
2141 }
2142 if (!next)
2143 break;
2144 n = rb_entry(next, struct sp_node, nd);
2145 }
2146 if (new)
2147 sp_insert(sp, new);
2148 spin_unlock(&sp->lock);
2149 if (new2) {
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002150 mpol_put(new2->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 kmem_cache_free(sn_cache, new2);
2152 }
2153 return 0;
2154}
2155
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002156/**
2157 * mpol_shared_policy_init - initialize shared policy for inode
2158 * @sp: pointer to inode shared policy
2159 * @mpol: struct mempolicy to install
2160 *
2161 * Install non-NULL @mpol in inode's shared policy rb-tree.
2162 * On entry, the current task has a reference on a non-NULL @mpol.
2163 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002164 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002165 */
2166void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08002167{
Miao Xie58568d22009-06-16 15:31:49 -07002168 int ret;
2169
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002170 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2171 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08002172
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002173 if (mpol) {
2174 struct vm_area_struct pvma;
2175 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002176 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002177
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002178 if (!scratch)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002179 goto put_mpol;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002180 /* contextualize the tmpfs mount point mempolicy */
2181 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002182 if (IS_ERR(new))
Dan Carpenter0cae3452010-05-25 23:42:58 -07002183 goto free_scratch; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002184
2185 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002186 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002187 task_unlock(current);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002188 if (ret)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002189 goto put_new;
Robin Holt7339ff82006-01-14 13:20:48 -08002190
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002191 /* Create pseudo-vma that contains just the policy */
2192 memset(&pvma, 0, sizeof(struct vm_area_struct));
2193 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2194 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002195
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002196put_new:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002197 mpol_put(new); /* drop initial ref */
Dan Carpenter0cae3452010-05-25 23:42:58 -07002198free_scratch:
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002199 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002200put_mpol:
2201 mpol_put(mpol); /* drop our incoming ref on sb mpol */
Robin Holt7339ff82006-01-14 13:20:48 -08002202 }
2203}
2204
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205int mpol_set_shared_policy(struct shared_policy *info,
2206 struct vm_area_struct *vma, struct mempolicy *npol)
2207{
2208 int err;
2209 struct sp_node *new = NULL;
2210 unsigned long sz = vma_pages(vma);
2211
David Rientjes028fec42008-04-28 02:12:25 -07002212 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002214 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002215 npol ? npol->flags : -1,
Paul Mundt140d5a42007-07-15 23:38:16 -07002216 npol ? nodes_addr(npol->v.nodes)[0] : -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217
2218 if (npol) {
2219 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2220 if (!new)
2221 return -ENOMEM;
2222 }
2223 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2224 if (err && new)
2225 kmem_cache_free(sn_cache, new);
2226 return err;
2227}
2228
2229/* Free a backing policy store on inode delete. */
2230void mpol_free_shared_policy(struct shared_policy *p)
2231{
2232 struct sp_node *n;
2233 struct rb_node *next;
2234
2235 if (!p->root.rb_node)
2236 return;
2237 spin_lock(&p->lock);
2238 next = rb_first(&p->root);
2239 while (next) {
2240 n = rb_entry(next, struct sp_node, nd);
2241 next = rb_next(&n->nd);
Andi Kleen90c50292005-07-27 11:43:50 -07002242 rb_erase(&n->nd, &p->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002243 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 kmem_cache_free(sn_cache, n);
2245 }
2246 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247}
2248
2249/* assumes fs == KERNEL_DS */
2250void __init numa_policy_init(void)
2251{
Paul Mundtb71636e2007-07-15 23:38:15 -07002252 nodemask_t interleave_nodes;
2253 unsigned long largest = 0;
2254 int nid, prefer = 0;
2255
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 policy_cache = kmem_cache_create("numa_policy",
2257 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002258 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
2260 sn_cache = kmem_cache_create("shared_policy_node",
2261 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002262 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
Paul Mundtb71636e2007-07-15 23:38:15 -07002264 /*
2265 * Set interleaving policy for system init. Interleaving is only
2266 * enabled across suitably sized nodes (default is >= 16MB), or
2267 * fall back to the largest node if they're all smaller.
2268 */
2269 nodes_clear(interleave_nodes);
Christoph Lameter56bbd652007-10-16 01:25:35 -07002270 for_each_node_state(nid, N_HIGH_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002271 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272
Paul Mundtb71636e2007-07-15 23:38:15 -07002273 /* Preserve the largest node */
2274 if (largest < total_pages) {
2275 largest = total_pages;
2276 prefer = nid;
2277 }
2278
2279 /* Interleave this node? */
2280 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2281 node_set(nid, interleave_nodes);
2282 }
2283
2284 /* All too small, use the largest */
2285 if (unlikely(nodes_empty(interleave_nodes)))
2286 node_set(prefer, interleave_nodes);
2287
David Rientjes028fec42008-04-28 02:12:25 -07002288 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 printk("numa_policy_init: interleaving failed\n");
2290}
2291
Christoph Lameter8bccd852005-10-29 18:16:59 -07002292/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293void numa_default_policy(void)
2294{
David Rientjes028fec42008-04-28 02:12:25 -07002295 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296}
Paul Jackson68860ec2005-10-30 15:02:36 -08002297
Paul Jackson42253992006-01-08 01:01:59 -08002298/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002299 * Parse and format mempolicy from/to strings
2300 */
2301
2302/*
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002303 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002304 * Used only for mpol_parse_str() and mpol_to_str()
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002305 */
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002306#define MPOL_LOCAL MPOL_MAX
2307static const char * const policy_modes[] =
2308{
2309 [MPOL_DEFAULT] = "default",
2310 [MPOL_PREFERRED] = "prefer",
2311 [MPOL_BIND] = "bind",
2312 [MPOL_INTERLEAVE] = "interleave",
2313 [MPOL_LOCAL] = "local"
2314};
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002315
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002316
2317#ifdef CONFIG_TMPFS
2318/**
2319 * mpol_parse_str - parse string to mempolicy
2320 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002321 * @mpol: pointer to struct mempolicy pointer, returned on success.
2322 * @no_context: flag whether to "contextualize" the mempolicy
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002323 *
2324 * Format of input:
2325 * <mode>[=<flags>][:<nodelist>]
2326 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002327 * if @no_context is true, save the input nodemask in w.user_nodemask in
2328 * the returned mempolicy. This will be used to "clone" the mempolicy in
2329 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2330 * mount option. Note that if 'static' or 'relative' mode flags were
2331 * specified, the input nodemask will already have been saved. Saving
2332 * it again is redundant, but safe.
2333 *
2334 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002335 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002336int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002337{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002338 struct mempolicy *new = NULL;
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002339 unsigned short mode;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002340 unsigned short uninitialized_var(mode_flags);
2341 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002342 char *nodelist = strchr(str, ':');
2343 char *flags = strchr(str, '=');
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002344 int err = 1;
2345
2346 if (nodelist) {
2347 /* NUL-terminate mode or flags string */
2348 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002349 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002350 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002351 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002352 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002353 } else
2354 nodes_clear(nodes);
2355
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002356 if (flags)
2357 *flags++ = '\0'; /* terminate mode string */
2358
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002359 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002360 if (!strcmp(str, policy_modes[mode])) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002361 break;
2362 }
2363 }
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002364 if (mode > MPOL_LOCAL)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002365 goto out;
2366
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002367 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002368 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002369 /*
2370 * Insist on a nodelist of one node only
2371 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002372 if (nodelist) {
2373 char *rest = nodelist;
2374 while (isdigit(*rest))
2375 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002376 if (*rest)
2377 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002378 }
2379 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002380 case MPOL_INTERLEAVE:
2381 /*
2382 * Default to online nodes with memory if no nodelist
2383 */
2384 if (!nodelist)
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002385 nodes = node_states[N_HIGH_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002386 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002387 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002388 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002389 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002390 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002391 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002392 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002393 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002394 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002395 case MPOL_DEFAULT:
2396 /*
2397 * Insist on a empty nodelist
2398 */
2399 if (!nodelist)
2400 err = 0;
2401 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002402 case MPOL_BIND:
2403 /*
2404 * Insist on a nodelist
2405 */
2406 if (!nodelist)
2407 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002408 }
2409
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002410 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002411 if (flags) {
2412 /*
2413 * Currently, we only support two mutually exclusive
2414 * mode flags.
2415 */
2416 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002417 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002418 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002419 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002420 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002421 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002422 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002423
2424 new = mpol_new(mode, mode_flags, &nodes);
2425 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002426 goto out;
2427
Lee Schermerhorne17f74a2010-05-24 14:32:02 -07002428 if (no_context) {
2429 /* save for contextualization */
2430 new->w.user_nodemask = nodes;
2431 } else {
Miao Xie58568d22009-06-16 15:31:49 -07002432 int ret;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002433 NODEMASK_SCRATCH(scratch);
2434 if (scratch) {
2435 task_lock(current);
2436 ret = mpol_set_nodemask(new, &nodes, scratch);
2437 task_unlock(current);
2438 } else
2439 ret = -ENOMEM;
2440 NODEMASK_SCRATCH_FREE(scratch);
2441 if (ret) {
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002442 mpol_put(new);
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002443 goto out;
Miao Xie58568d22009-06-16 15:31:49 -07002444 }
2445 }
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002446 err = 0;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002447
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002448out:
2449 /* Restore string for error message */
2450 if (nodelist)
2451 *--nodelist = ':';
2452 if (flags)
2453 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002454 if (!err)
2455 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002456 return err;
2457}
2458#endif /* CONFIG_TMPFS */
2459
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002460/**
2461 * mpol_to_str - format a mempolicy structure for printing
2462 * @buffer: to contain formatted mempolicy string
2463 * @maxlen: length of @buffer
2464 * @pol: pointer to mempolicy to be formatted
2465 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2466 *
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002467 * Convert a mempolicy into a string.
2468 * Returns the number of characters in buffer (if positive)
2469 * or an error (negative)
2470 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002471int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002472{
2473 char *p = buffer;
2474 int l;
2475 nodemask_t nodes;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002476 unsigned short mode;
David Rientjesf5b087b2008-04-28 02:12:27 -07002477 unsigned short flags = pol ? pol->flags : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002478
Lee Schermerhorn22919902008-04-28 02:13:22 -07002479 /*
2480 * Sanity check: room for longest mode, flag and some nodes
2481 */
2482 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2483
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002484 if (!pol || pol == &default_policy)
2485 mode = MPOL_DEFAULT;
2486 else
2487 mode = pol->mode;
2488
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002489 switch (mode) {
2490 case MPOL_DEFAULT:
2491 nodes_clear(nodes);
2492 break;
2493
2494 case MPOL_PREFERRED:
2495 nodes_clear(nodes);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002496 if (flags & MPOL_F_LOCAL)
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002497 mode = MPOL_LOCAL; /* pseudo-policy */
2498 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002499 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002500 break;
2501
2502 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07002503 /* Fall through */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002504 case MPOL_INTERLEAVE:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002505 if (no_context)
2506 nodes = pol->w.user_nodemask;
2507 else
2508 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002509 break;
2510
2511 default:
2512 BUG();
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002513 }
2514
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002515 l = strlen(policy_modes[mode]);
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002516 if (buffer + maxlen < p + l + 1)
2517 return -ENOSPC;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002518
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002519 strcpy(p, policy_modes[mode]);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002520 p += l;
2521
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002522 if (flags & MPOL_MODE_FLAGS) {
David Rientjesf5b087b2008-04-28 02:12:27 -07002523 if (buffer + maxlen < p + 2)
2524 return -ENOSPC;
2525 *p++ = '=';
2526
Lee Schermerhorn22919902008-04-28 02:13:22 -07002527 /*
2528 * Currently, the only defined flags are mutually exclusive
2529 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002530 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002531 p += snprintf(p, buffer + maxlen - p, "static");
2532 else if (flags & MPOL_F_RELATIVE_NODES)
2533 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002534 }
2535
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002536 if (!nodes_empty(nodes)) {
2537 if (buffer + maxlen < p + 2)
2538 return -ENOSPC;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002539 *p++ = ':';
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002540 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2541 }
2542 return p - buffer;
2543}