blob: 5bc0a96beb51b550b69389998fc3e09053ee5a2f [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>
96
Nick Piggin62695a82008-10-18 20:26:09 -070097#include "internal.h"
98
Christoph Lameter38e35862006-01-08 01:01:01 -080099/* Internal flags */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800100#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
Christoph Lameter38e35862006-01-08 01:01:01 -0800101#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800102#define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
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 Lameter397874d2006-03-06 15:42:53 -0800460static void gather_stats(struct page *, void *, int pte_dirty);
Christoph Lameterfc301282006-01-18 17:42:29 -0800461static void migrate_page_add(struct page *page, struct list_head *pagelist,
462 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800463
Christoph Lameter38e35862006-01-08 01:01:01 -0800464/* Scan through pages checking if pages follow certain conditions. */
Nick Pigginb5810032005-10-29 18:16:12 -0700465static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800466 unsigned long addr, unsigned long end,
467 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800468 void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469{
Hugh Dickins91612e02005-06-21 17:15:07 -0700470 pte_t *orig_pte;
471 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700472 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700473
Hugh Dickins705e87c2005-10-29 18:16:27 -0700474 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700475 do {
Linus Torvalds6aab3412005-11-28 14:34:23 -0800476 struct page *page;
Andy Whitcroft25ba77c2006-12-06 20:33:03 -0800477 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700478
479 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800481 page = vm_normal_page(vma, addr, *pte);
482 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800484 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800485 * vm_normal_page() filters out zero pages, but there might
486 * still be PageReserved pages to skip, perhaps in a VDSO.
487 * And we cannot move PageKsm pages sensibly or safely yet.
Nick Piggin053837f2006-01-18 17:42:27 -0800488 */
Hugh Dickins62b61f62009-12-14 17:59:33 -0800489 if (PageReserved(page) || PageKsm(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800490 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800491 nid = page_to_nid(page);
Christoph Lameter38e35862006-01-08 01:01:01 -0800492 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
493 continue;
494
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800495 if (flags & MPOL_MF_STATS)
Christoph Lameter397874d2006-03-06 15:42:53 -0800496 gather_stats(page, private, pte_dirty(*pte));
Nick Piggin053837f2006-01-18 17:42:27 -0800497 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameterfc301282006-01-18 17:42:29 -0800498 migrate_page_add(page, private, flags);
Christoph Lameter38e35862006-01-08 01:01:01 -0800499 else
500 break;
Hugh Dickins91612e02005-06-21 17:15:07 -0700501 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickins705e87c2005-10-29 18:16:27 -0700502 pte_unmap_unlock(orig_pte, ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700503 return addr != end;
504}
505
Nick Pigginb5810032005-10-29 18:16:12 -0700506static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800507 unsigned long addr, unsigned long end,
508 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800509 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700510{
511 pmd_t *pmd;
512 unsigned long next;
513
514 pmd = pmd_offset(pud, addr);
515 do {
516 next = pmd_addr_end(addr, end);
517 if (pmd_none_or_clear_bad(pmd))
518 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800519 if (check_pte_range(vma, pmd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800520 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700521 return -EIO;
522 } while (pmd++, addr = next, addr != end);
523 return 0;
524}
525
Nick Pigginb5810032005-10-29 18:16:12 -0700526static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800527 unsigned long addr, unsigned long end,
528 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800529 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700530{
531 pud_t *pud;
532 unsigned long next;
533
534 pud = pud_offset(pgd, addr);
535 do {
536 next = pud_addr_end(addr, end);
537 if (pud_none_or_clear_bad(pud))
538 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800539 if (check_pmd_range(vma, pud, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800540 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700541 return -EIO;
542 } while (pud++, addr = next, addr != end);
543 return 0;
544}
545
Nick Pigginb5810032005-10-29 18:16:12 -0700546static inline int check_pgd_range(struct vm_area_struct *vma,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800547 unsigned long addr, unsigned long end,
548 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800549 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700550{
551 pgd_t *pgd;
552 unsigned long next;
553
Nick Pigginb5810032005-10-29 18:16:12 -0700554 pgd = pgd_offset(vma->vm_mm, addr);
Hugh Dickins91612e02005-06-21 17:15:07 -0700555 do {
556 next = pgd_addr_end(addr, end);
557 if (pgd_none_or_clear_bad(pgd))
558 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800559 if (check_pud_range(vma, pgd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800560 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700561 return -EIO;
562 } while (pgd++, addr = next, addr != end);
563 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564}
565
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800566/*
567 * Check if all pages in a range are on a set of nodes.
568 * If pagelist != NULL then isolate pages from the LRU and
569 * put them on the pagelist.
570 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571static struct vm_area_struct *
572check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Christoph Lameter38e35862006-01-08 01:01:01 -0800573 const nodemask_t *nodes, unsigned long flags, void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574{
575 int err;
576 struct vm_area_struct *first, *vma, *prev;
577
Nick Piggin053837f2006-01-18 17:42:27 -0800578
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 first = find_vma(mm, start);
580 if (!first)
581 return ERR_PTR(-EFAULT);
582 prev = NULL;
583 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800584 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
585 if (!vma->vm_next && vma->vm_end < end)
586 return ERR_PTR(-EFAULT);
587 if (prev && prev->vm_end < vma->vm_start)
588 return ERR_PTR(-EFAULT);
589 }
590 if (!is_vm_hugetlb_page(vma) &&
591 ((flags & MPOL_MF_STRICT) ||
592 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
593 vma_migratable(vma)))) {
Andi Kleen5b952b32005-09-13 01:25:08 -0700594 unsigned long endvma = vma->vm_end;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800595
Andi Kleen5b952b32005-09-13 01:25:08 -0700596 if (endvma > end)
597 endvma = end;
598 if (vma->vm_start > start)
599 start = vma->vm_start;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800600 err = check_pgd_range(vma, start, endvma, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800601 flags, private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 if (err) {
603 first = ERR_PTR(err);
604 break;
605 }
606 }
607 prev = vma;
608 }
609 return first;
610}
611
612/* Apply policy to a single VMA */
613static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
614{
615 int err = 0;
616 struct mempolicy *old = vma->vm_policy;
617
Paul Mundt140d5a42007-07-15 23:38:16 -0700618 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 vma->vm_start, vma->vm_end, vma->vm_pgoff,
620 vma->vm_ops, vma->vm_file,
621 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
622
623 if (vma->vm_ops && vma->vm_ops->set_policy)
624 err = vma->vm_ops->set_policy(vma, new);
625 if (!err) {
626 mpol_get(new);
627 vma->vm_policy = new;
Lee Schermerhornf0be3d32008-04-28 02:13:08 -0700628 mpol_put(old);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629 }
630 return err;
631}
632
633/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800634static int mbind_range(struct mm_struct *mm, unsigned long start,
635 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636{
637 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800638 struct vm_area_struct *prev;
639 struct vm_area_struct *vma;
640 int err = 0;
641 pgoff_t pgoff;
642 unsigned long vmstart;
643 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800645 vma = find_vma_prev(mm, start, &prev);
646 if (!vma || vma->vm_start > start)
647 return -EFAULT;
648
649 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800651 vmstart = max(start, vma->vm_start);
652 vmend = min(end, vma->vm_end);
653
654 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
655 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
656 vma->anon_vma, vma->vm_file, pgoff, new_pol);
657 if (prev) {
658 vma = prev;
659 next = vma->vm_next;
660 continue;
661 }
662 if (vma->vm_start != vmstart) {
663 err = split_vma(vma->vm_mm, vma, vmstart, 1);
664 if (err)
665 goto out;
666 }
667 if (vma->vm_end != vmend) {
668 err = split_vma(vma->vm_mm, vma, vmend, 0);
669 if (err)
670 goto out;
671 }
672 err = policy_vma(vma, new_pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 if (err)
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800674 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800676
677 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 return err;
679}
680
Paul Jacksonc61afb12006-03-24 03:16:08 -0800681/*
682 * Update task->flags PF_MEMPOLICY bit: set iff non-default
683 * mempolicy. Allows more rapid checking of this (combined perhaps
684 * with other PF_* flag bits) on memory allocation hot code paths.
685 *
686 * If called from outside this file, the task 'p' should -only- be
687 * a newly forked child not yet visible on the task list, because
688 * manipulating the task flags of a visible task is not safe.
689 *
690 * The above limitation is why this routine has the funny name
691 * mpol_fix_fork_child_flag().
692 *
693 * It is also safe to call this with a task pointer of current,
694 * which the static wrapper mpol_set_task_struct_flag() does,
695 * for use within this file.
696 */
697
698void mpol_fix_fork_child_flag(struct task_struct *p)
699{
700 if (p->mempolicy)
701 p->flags |= PF_MEMPOLICY;
702 else
703 p->flags &= ~PF_MEMPOLICY;
704}
705
706static void mpol_set_task_struct_flag(void)
707{
708 mpol_fix_fork_child_flag(current);
709}
710
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700712static long do_set_mempolicy(unsigned short mode, unsigned short flags,
713 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714{
Miao Xie58568d22009-06-16 15:31:49 -0700715 struct mempolicy *new, *old;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700716 struct mm_struct *mm = current->mm;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700717 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700718 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700720 if (!scratch)
721 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700722
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700723 new = mpol_new(mode, flags, nodes);
724 if (IS_ERR(new)) {
725 ret = PTR_ERR(new);
726 goto out;
727 }
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700728 /*
729 * prevent changing our mempolicy while show_numa_maps()
730 * is using it.
731 * Note: do_set_mempolicy() can be called at init time
732 * with no 'mm'.
733 */
734 if (mm)
735 down_write(&mm->mmap_sem);
Miao Xie58568d22009-06-16 15:31:49 -0700736 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700737 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700738 if (ret) {
739 task_unlock(current);
740 if (mm)
741 up_write(&mm->mmap_sem);
742 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700743 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700744 }
745 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 current->mempolicy = new;
Paul Jacksonc61afb12006-03-24 03:16:08 -0800747 mpol_set_task_struct_flag();
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700748 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700749 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700750 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700751 task_unlock(current);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700752 if (mm)
753 up_write(&mm->mmap_sem);
754
Miao Xie58568d22009-06-16 15:31:49 -0700755 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700756 ret = 0;
757out:
758 NODEMASK_SCRATCH_FREE(scratch);
759 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760}
761
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700762/*
763 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700764 *
765 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700766 */
767static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700769 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700770 if (p == &default_policy)
771 return;
772
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700773 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700774 case MPOL_BIND:
775 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700777 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 break;
779 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700780 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700781 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700782 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783 break;
784 default:
785 BUG();
786 }
787}
788
789static int lookup_node(struct mm_struct *mm, unsigned long addr)
790{
791 struct page *p;
792 int err;
793
794 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
795 if (err >= 0) {
796 err = page_to_nid(p);
797 put_page(p);
798 }
799 return err;
800}
801
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700803static long do_get_mempolicy(int *policy, nodemask_t *nmask,
804 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700806 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 struct mm_struct *mm = current->mm;
808 struct vm_area_struct *vma = NULL;
809 struct mempolicy *pol = current->mempolicy;
810
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700811 if (flags &
812 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700814
815 if (flags & MPOL_F_MEMS_ALLOWED) {
816 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
817 return -EINVAL;
818 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700819 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700820 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700821 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700822 return 0;
823 }
824
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700826 /*
827 * Do NOT fall back to task policy if the
828 * vma/shared policy at addr is NULL. We
829 * want to return MPOL_DEFAULT in this case.
830 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 down_read(&mm->mmap_sem);
832 vma = find_vma_intersection(mm, addr, addr+1);
833 if (!vma) {
834 up_read(&mm->mmap_sem);
835 return -EFAULT;
836 }
837 if (vma->vm_ops && vma->vm_ops->get_policy)
838 pol = vma->vm_ops->get_policy(vma, addr);
839 else
840 pol = vma->vm_policy;
841 } else if (addr)
842 return -EINVAL;
843
844 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700845 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846
847 if (flags & MPOL_F_NODE) {
848 if (flags & MPOL_F_ADDR) {
849 err = lookup_node(mm, addr);
850 if (err < 0)
851 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700852 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700854 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700855 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 } else {
857 err = -EINVAL;
858 goto out;
859 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700860 } else {
861 *policy = pol == &default_policy ? MPOL_DEFAULT :
862 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700863 /*
864 * Internal mempolicy flags must be masked off before exposing
865 * the policy to userspace.
866 */
867 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700868 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869
870 if (vma) {
871 up_read(&current->mm->mmap_sem);
872 vma = NULL;
873 }
874
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700876 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700877 if (mpol_store_user_nodemask(pol)) {
878 *nmask = pol->w.user_nodemask;
879 } else {
880 task_lock(current);
881 get_policy_nodemask(pol, nmask);
882 task_unlock(current);
883 }
Miao Xie58568d22009-06-16 15:31:49 -0700884 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885
886 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700887 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 if (vma)
889 up_read(&current->mm->mmap_sem);
890 return err;
891}
892
Christoph Lameterb20a3502006-03-22 00:09:12 -0800893#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700894/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800895 * page migration
896 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800897static void migrate_page_add(struct page *page, struct list_head *pagelist,
898 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800899{
900 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800901 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800902 */
Nick Piggin62695a82008-10-18 20:26:09 -0700903 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
904 if (!isolate_lru_page(page)) {
905 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800906 inc_zone_page_state(page, NR_ISOLATED_ANON +
907 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700908 }
909 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800910}
911
Christoph Lameter742755a2006-06-23 02:03:55 -0700912static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700913{
Mel Gorman6484eb32009-06-16 15:31:54 -0700914 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700915}
916
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800917/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800918 * Migrate pages from one node to a target node.
919 * Returns error or the number of pages not migrated.
920 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700921static int migrate_to_node(struct mm_struct *mm, int source, int dest,
922 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800923{
924 nodemask_t nmask;
925 LIST_HEAD(pagelist);
926 int err = 0;
927
928 nodes_clear(nmask);
929 node_set(source, nmask);
930
Greg Thelen6ec3a122010-05-24 14:32:33 -0700931 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800932 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
933
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700934 if (!list_empty(&pagelist))
Hugh Dickins62b61f62009-12-14 17:59:33 -0800935 err = migrate_pages(&pagelist, new_node_page, dest, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700936
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800937 return err;
938}
939
940/*
941 * Move pages between the two nodesets so as to preserve the physical
942 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -0800943 *
944 * Returns the number of page that could not be moved.
945 */
946int do_migrate_pages(struct mm_struct *mm,
947 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
948{
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800949 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800950 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800951 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -0800952
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800953 err = migrate_prep();
954 if (err)
955 return err;
956
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700957 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -0800958
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700959 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
960 if (err)
961 goto out;
962
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -0800963 /*
964 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
965 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
966 * bit in 'tmp', and return that <source, dest> pair for migration.
967 * The pair of nodemasks 'to' and 'from' define the map.
968 *
969 * If no pair of bits is found that way, fallback to picking some
970 * pair of 'source' and 'dest' bits that are not the same. If the
971 * 'source' and 'dest' bits are the same, this represents a node
972 * that will be migrating to itself, so no pages need move.
973 *
974 * If no bits are left in 'tmp', or if all remaining bits left
975 * in 'tmp' correspond to the same bit in 'to', return false
976 * (nothing left to migrate).
977 *
978 * This lets us pick a pair of nodes to migrate between, such that
979 * if possible the dest node is not already occupied by some other
980 * source node, minimizing the risk of overloading the memory on a
981 * node that would happen if we migrated incoming memory to a node
982 * before migrating outgoing memory source that same node.
983 *
984 * A single scan of tmp is sufficient. As we go, we remember the
985 * most recent <s, d> pair that moved (s != d). If we find a pair
986 * that not only moved, but what's better, moved to an empty slot
987 * (d is not set in tmp), then we break out then, with that pair.
988 * Otherwise when we finish scannng from_tmp, we at least have the
989 * most recent <s, d> pair that moved. If we get all the way through
990 * the scan of tmp without finding any node that moved, much less
991 * moved to an empty node, then there is nothing left worth migrating.
992 */
Christoph Lameterd4984712006-01-08 01:00:55 -0800993
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800994 tmp = *from_nodes;
995 while (!nodes_empty(tmp)) {
996 int s,d;
997 int source = -1;
998 int dest = 0;
999
1000 for_each_node_mask(s, tmp) {
1001 d = node_remap(s, *from_nodes, *to_nodes);
1002 if (s == d)
1003 continue;
1004
1005 source = s; /* Node moved. Memorize */
1006 dest = d;
1007
1008 /* dest not in remaining from nodes? */
1009 if (!node_isset(dest, tmp))
1010 break;
1011 }
1012 if (source == -1)
1013 break;
1014
1015 node_clear(source, tmp);
1016 err = migrate_to_node(mm, source, dest, flags);
1017 if (err > 0)
1018 busy += err;
1019 if (err < 0)
1020 break;
Christoph Lameter39743882006-01-08 01:00:51 -08001021 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001022out:
Christoph Lameter39743882006-01-08 01:00:51 -08001023 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001024 if (err < 0)
1025 return err;
1026 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001027
Christoph Lameter39743882006-01-08 01:00:51 -08001028}
1029
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001030/*
1031 * Allocate a new page for page migration based on vma policy.
1032 * Start assuming that page is mapped by vma pointed to by @private.
1033 * Search forward from there, if not. N.B., this assumes that the
1034 * list of pages handed to migrate_pages()--which is how we get here--
1035 * is in virtual address order.
1036 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001037static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001038{
1039 struct vm_area_struct *vma = (struct vm_area_struct *)private;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001040 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001041
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -08001042 while (vma) {
1043 address = page_address_in_vma(page, vma);
1044 if (address != -EFAULT)
1045 break;
1046 vma = vma->vm_next;
1047 }
1048
1049 /*
1050 * if !vma, alloc_page_vma() will use task or system default policy
1051 */
1052 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -07001053}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001054#else
1055
1056static void migrate_page_add(struct page *page, struct list_head *pagelist,
1057 unsigned long flags)
1058{
1059}
1060
1061int do_migrate_pages(struct mm_struct *mm,
1062 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1063{
1064 return -ENOSYS;
1065}
Christoph Lameter95a402c2006-06-23 02:03:53 -07001066
Keith Owens69939742006-10-11 01:21:28 -07001067static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -07001068{
1069 return NULL;
1070}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001071#endif
1072
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001073static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001074 unsigned short mode, unsigned short mode_flags,
1075 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001076{
1077 struct vm_area_struct *vma;
1078 struct mm_struct *mm = current->mm;
1079 struct mempolicy *new;
1080 unsigned long end;
1081 int err;
1082 LIST_HEAD(pagelist);
1083
David Rientjesa3b51e02008-04-28 02:12:23 -07001084 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1085 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001086 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001087 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001088 return -EPERM;
1089
1090 if (start & ~PAGE_MASK)
1091 return -EINVAL;
1092
1093 if (mode == MPOL_DEFAULT)
1094 flags &= ~MPOL_MF_STRICT;
1095
1096 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1097 end = start + len;
1098
1099 if (end < start)
1100 return -EINVAL;
1101 if (end == start)
1102 return 0;
1103
David Rientjes028fec42008-04-28 02:12:25 -07001104 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001105 if (IS_ERR(new))
1106 return PTR_ERR(new);
1107
1108 /*
1109 * If we are using the default policy then operation
1110 * on discontinuous address spaces is okay after all
1111 */
1112 if (!new)
1113 flags |= MPOL_MF_DISCONTIG_OK;
1114
David Rientjes028fec42008-04-28 02:12:25 -07001115 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1116 start, start + len, mode, mode_flags,
1117 nmask ? nodes_addr(*nmask)[0] : -1);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001118
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001119 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1120
1121 err = migrate_prep();
1122 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001123 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001124 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001125 {
1126 NODEMASK_SCRATCH(scratch);
1127 if (scratch) {
1128 down_write(&mm->mmap_sem);
1129 task_lock(current);
1130 err = mpol_set_nodemask(new, nmask, scratch);
1131 task_unlock(current);
1132 if (err)
1133 up_write(&mm->mmap_sem);
1134 } else
1135 err = -ENOMEM;
1136 NODEMASK_SCRATCH_FREE(scratch);
1137 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001138 if (err)
1139 goto mpol_out;
1140
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001141 vma = check_range(mm, start, end, nmask,
1142 flags | MPOL_MF_INVERT, &pagelist);
1143
1144 err = PTR_ERR(vma);
1145 if (!IS_ERR(vma)) {
1146 int nr_failed = 0;
1147
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001148 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001149
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001150 if (!list_empty(&pagelist))
Christoph Lameter95a402c2006-06-23 02:03:53 -07001151 nr_failed = migrate_pages(&pagelist, new_vma_page,
Hugh Dickins62b61f62009-12-14 17:59:33 -08001152 (unsigned long)vma, 0);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001153
1154 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1155 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001156 } else
1157 putback_lru_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001158
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001159 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001160 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001161 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001162 return err;
1163}
1164
Christoph Lameter39743882006-01-08 01:00:51 -08001165/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001166 * User space interface with variable sized bitmaps for nodelists.
1167 */
1168
1169/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001170static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001171 unsigned long maxnode)
1172{
1173 unsigned long k;
1174 unsigned long nlongs;
1175 unsigned long endmask;
1176
1177 --maxnode;
1178 nodes_clear(*nodes);
1179 if (maxnode == 0 || !nmask)
1180 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001181 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001182 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001183
1184 nlongs = BITS_TO_LONGS(maxnode);
1185 if ((maxnode % BITS_PER_LONG) == 0)
1186 endmask = ~0UL;
1187 else
1188 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1189
1190 /* When the user specified more nodes than supported just check
1191 if the non supported part is all zero. */
1192 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1193 if (nlongs > PAGE_SIZE/sizeof(long))
1194 return -EINVAL;
1195 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1196 unsigned long t;
1197 if (get_user(t, nmask + k))
1198 return -EFAULT;
1199 if (k == nlongs - 1) {
1200 if (t & endmask)
1201 return -EINVAL;
1202 } else if (t)
1203 return -EINVAL;
1204 }
1205 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1206 endmask = ~0UL;
1207 }
1208
1209 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1210 return -EFAULT;
1211 nodes_addr(*nodes)[nlongs-1] &= endmask;
1212 return 0;
1213}
1214
1215/* Copy a kernel node mask to user space */
1216static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1217 nodemask_t *nodes)
1218{
1219 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1220 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1221
1222 if (copy > nbytes) {
1223 if (copy > PAGE_SIZE)
1224 return -EINVAL;
1225 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1226 return -EFAULT;
1227 copy = nbytes;
1228 }
1229 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1230}
1231
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001232SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1233 unsigned long, mode, unsigned long __user *, nmask,
1234 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001235{
1236 nodemask_t nodes;
1237 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001238 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001239
David Rientjes028fec42008-04-28 02:12:25 -07001240 mode_flags = mode & MPOL_MODE_FLAGS;
1241 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001242 if (mode >= MPOL_MAX)
1243 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001244 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1245 (mode_flags & MPOL_F_RELATIVE_NODES))
1246 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001247 err = get_nodes(&nodes, nmask, maxnode);
1248 if (err)
1249 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001250 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001251}
1252
1253/* Set the process memory policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001254SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1255 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001256{
1257 int err;
1258 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001259 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001260
David Rientjes028fec42008-04-28 02:12:25 -07001261 flags = mode & MPOL_MODE_FLAGS;
1262 mode &= ~MPOL_MODE_FLAGS;
1263 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001264 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001265 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1266 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001267 err = get_nodes(&nodes, nmask, maxnode);
1268 if (err)
1269 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001270 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001271}
1272
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001273SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1274 const unsigned long __user *, old_nodes,
1275 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001276{
David Howellsc69e8d92008-11-14 10:39:19 +11001277 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter39743882006-01-08 01:00:51 -08001278 struct mm_struct *mm;
1279 struct task_struct *task;
1280 nodemask_t old;
1281 nodemask_t new;
1282 nodemask_t task_nodes;
1283 int err;
1284
1285 err = get_nodes(&old, old_nodes, maxnode);
1286 if (err)
1287 return err;
1288
1289 err = get_nodes(&new, new_nodes, maxnode);
1290 if (err)
1291 return err;
1292
1293 /* Find the mm_struct */
1294 read_lock(&tasklist_lock);
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001295 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001296 if (!task) {
1297 read_unlock(&tasklist_lock);
1298 return -ESRCH;
1299 }
1300 mm = get_task_mm(task);
1301 read_unlock(&tasklist_lock);
1302
1303 if (!mm)
1304 return -EINVAL;
1305
1306 /*
1307 * Check if this process has the right to modify the specified
1308 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001309 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001310 * userid as the target process.
1311 */
David Howellsc69e8d92008-11-14 10:39:19 +11001312 rcu_read_lock();
1313 tcred = __task_cred(task);
David Howellsb6dff3e2008-11-14 10:39:16 +11001314 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1315 cred->uid != tcred->suid && cred->uid != tcred->uid &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001316 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001317 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001318 err = -EPERM;
1319 goto out;
1320 }
David Howellsc69e8d92008-11-14 10:39:19 +11001321 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001322
1323 task_nodes = cpuset_mems_allowed(task);
1324 /* Is the user allowed to access the target nodes? */
Christoph Lameter74c00242006-03-14 19:50:21 -08001325 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001326 err = -EPERM;
1327 goto out;
1328 }
1329
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001330 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001331 err = -EINVAL;
1332 goto out;
1333 }
1334
David Quigley86c3a762006-06-23 02:04:02 -07001335 err = security_task_movememory(task);
1336 if (err)
1337 goto out;
1338
Christoph Lameter511030b2006-02-28 16:58:57 -08001339 err = do_migrate_pages(mm, &old, &new,
Christoph Lameter74c00242006-03-14 19:50:21 -08001340 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter39743882006-01-08 01:00:51 -08001341out:
1342 mmput(mm);
1343 return err;
1344}
1345
1346
Christoph Lameter8bccd852005-10-29 18:16:59 -07001347/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001348SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1349 unsigned long __user *, nmask, unsigned long, maxnode,
1350 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001351{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001352 int err;
1353 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001354 nodemask_t nodes;
1355
1356 if (nmask != NULL && maxnode < MAX_NUMNODES)
1357 return -EINVAL;
1358
1359 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1360
1361 if (err)
1362 return err;
1363
1364 if (policy && put_user(pval, policy))
1365 return -EFAULT;
1366
1367 if (nmask)
1368 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1369
1370 return err;
1371}
1372
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373#ifdef CONFIG_COMPAT
1374
1375asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1376 compat_ulong_t __user *nmask,
1377 compat_ulong_t maxnode,
1378 compat_ulong_t addr, compat_ulong_t flags)
1379{
1380 long err;
1381 unsigned long __user *nm = NULL;
1382 unsigned long nr_bits, alloc_size;
1383 DECLARE_BITMAP(bm, MAX_NUMNODES);
1384
1385 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1386 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1387
1388 if (nmask)
1389 nm = compat_alloc_user_space(alloc_size);
1390
1391 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1392
1393 if (!err && nmask) {
1394 err = copy_from_user(bm, nm, alloc_size);
1395 /* ensure entire bitmap is zeroed */
1396 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1397 err |= compat_put_bitmap(nmask, bm, nr_bits);
1398 }
1399
1400 return err;
1401}
1402
1403asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1404 compat_ulong_t maxnode)
1405{
1406 long err = 0;
1407 unsigned long __user *nm = NULL;
1408 unsigned long nr_bits, alloc_size;
1409 DECLARE_BITMAP(bm, MAX_NUMNODES);
1410
1411 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1412 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1413
1414 if (nmask) {
1415 err = compat_get_bitmap(bm, nmask, nr_bits);
1416 nm = compat_alloc_user_space(alloc_size);
1417 err |= copy_to_user(nm, bm, alloc_size);
1418 }
1419
1420 if (err)
1421 return -EFAULT;
1422
1423 return sys_set_mempolicy(mode, nm, nr_bits+1);
1424}
1425
1426asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1427 compat_ulong_t mode, compat_ulong_t __user *nmask,
1428 compat_ulong_t maxnode, compat_ulong_t flags)
1429{
1430 long err = 0;
1431 unsigned long __user *nm = NULL;
1432 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001433 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
1435 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1436 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1437
1438 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001439 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001441 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 }
1443
1444 if (err)
1445 return -EFAULT;
1446
1447 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1448}
1449
1450#endif
1451
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001452/*
1453 * get_vma_policy(@task, @vma, @addr)
1454 * @task - task for fallback if vma policy == default
1455 * @vma - virtual memory area whose policy is sought
1456 * @addr - address in @vma for shared policy lookup
1457 *
1458 * Returns effective policy for a VMA at specified address.
1459 * Falls back to @task or system default policy, as necessary.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001460 * Current or other task's task mempolicy and non-shared vma policies
1461 * are protected by the task's mmap_sem, which must be held for read by
1462 * the caller.
1463 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1464 * count--added by the get_policy() vm_op, as appropriate--to protect against
1465 * freeing by another task. It is the caller's responsibility to free the
1466 * extra reference for shared policies.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001467 */
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001468static struct mempolicy *get_vma_policy(struct task_struct *task,
Christoph Lameter48fce342006-01-08 01:01:03 -08001469 struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001471 struct mempolicy *pol = task->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472
1473 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001474 if (vma->vm_ops && vma->vm_ops->get_policy) {
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001475 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1476 addr);
1477 if (vpol)
1478 pol = vpol;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001479 } else if (vma->vm_policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 pol = vma->vm_policy;
1481 }
1482 if (!pol)
1483 pol = &default_policy;
1484 return pol;
1485}
1486
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001487/*
1488 * Return a nodemask representing a mempolicy for filtering nodes for
1489 * page allocation
1490 */
1491static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001492{
1493 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001494 if (unlikely(policy->mode == MPOL_BIND) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001495 gfp_zone(gfp) >= policy_zone &&
1496 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1497 return &policy->v.nodes;
1498
1499 return NULL;
1500}
1501
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001502/* Return a zonelist indicated by gfp for node representing a mempolicy */
1503static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504{
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001505 int nd = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001507 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001509 if (!(policy->flags & MPOL_F_LOCAL))
1510 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511 break;
1512 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001513 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001514 * Normally, MPOL_BIND allocations are node-local within the
1515 * allowed nodemask. However, if __GFP_THISNODE is set and the
Bob Liu6eb27e12010-05-24 14:32:00 -07001516 * current node isn't part of the mask, we use the zonelist for
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001517 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001518 */
Mel Gorman19770b32008-04-28 02:12:18 -07001519 if (unlikely(gfp & __GFP_THISNODE) &&
1520 unlikely(!node_isset(nd, policy->v.nodes)))
1521 nd = first_node(policy->v.nodes);
1522 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 BUG();
1525 }
Mel Gorman0e884602008-04-28 02:12:14 -07001526 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527}
1528
1529/* Do dynamic interleaving for a process */
1530static unsigned interleave_nodes(struct mempolicy *policy)
1531{
1532 unsigned nid, next;
1533 struct task_struct *me = current;
1534
1535 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001536 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001538 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001539 if (next < MAX_NUMNODES)
1540 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 return nid;
1542}
1543
Christoph Lameterdc85da12006-01-18 17:42:36 -08001544/*
1545 * Depending on the memory policy provide a node from which to allocate the
1546 * next slab entry.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001547 * @policy must be protected by freeing by the caller. If @policy is
1548 * the current task's mempolicy, this protection is implicit, as only the
1549 * task can change it's policy. The system default policy requires no
1550 * such protection.
Christoph Lameterdc85da12006-01-18 17:42:36 -08001551 */
1552unsigned slab_node(struct mempolicy *policy)
1553{
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001554 if (!policy || policy->flags & MPOL_F_LOCAL)
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001555 return numa_node_id();
Christoph Lameter765c4502006-09-27 01:50:08 -07001556
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001557 switch (policy->mode) {
1558 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001559 /*
1560 * handled MPOL_F_LOCAL above
1561 */
1562 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001563
Christoph Lameterdc85da12006-01-18 17:42:36 -08001564 case MPOL_INTERLEAVE:
1565 return interleave_nodes(policy);
1566
Mel Gormandd1a2392008-04-28 02:12:17 -07001567 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001568 /*
1569 * Follow bind policy behavior and start allocation at the
1570 * first node.
1571 */
Mel Gorman19770b32008-04-28 02:12:18 -07001572 struct zonelist *zonelist;
1573 struct zone *zone;
1574 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1575 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1576 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1577 &policy->v.nodes,
1578 &zone);
1579 return zone->node;
Mel Gormandd1a2392008-04-28 02:12:17 -07001580 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001581
Christoph Lameterdc85da12006-01-18 17:42:36 -08001582 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001583 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001584 }
1585}
1586
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587/* Do static interleaving for a VMA with known offset. */
1588static unsigned offset_il_node(struct mempolicy *pol,
1589 struct vm_area_struct *vma, unsigned long off)
1590{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001591 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001592 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 int c;
1594 int nid = -1;
1595
David Rientjesf5b087b2008-04-28 02:12:27 -07001596 if (!nnodes)
1597 return numa_node_id();
1598 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 c = 0;
1600 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001601 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 c++;
1603 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 return nid;
1605}
1606
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001607/* Determine a node number for interleave */
1608static inline unsigned interleave_nid(struct mempolicy *pol,
1609 struct vm_area_struct *vma, unsigned long addr, int shift)
1610{
1611 if (vma) {
1612 unsigned long off;
1613
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001614 /*
1615 * for small pages, there is no difference between
1616 * shift and PAGE_SHIFT, so the bit-shift is safe.
1617 * for huge pages, since vm_pgoff is in units of small
1618 * pages, we need to shift off the always 0 bits to get
1619 * a useful offset.
1620 */
1621 BUG_ON(shift < PAGE_SHIFT);
1622 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001623 off += (addr - vma->vm_start) >> shift;
1624 return offset_il_node(pol, vma, off);
1625 } else
1626 return interleave_nodes(pol);
1627}
1628
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001629#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001630/*
1631 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1632 * @vma = virtual memory area whose policy is sought
1633 * @addr = address in @vma for shared policy lookup and interleave policy
1634 * @gfp_flags = for requested zone
Mel Gorman19770b32008-04-28 02:12:18 -07001635 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1636 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001637 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001638 * Returns a zonelist suitable for a huge page allocation and a pointer
1639 * to the struct mempolicy for conditional unref after allocation.
1640 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1641 * @nodemask for filtering the zonelist.
Miao Xiec0ff7452010-05-24 14:32:08 -07001642 *
1643 * Must be protected by get_mems_allowed()
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001644 */
Mel Gorman396faf02007-07-17 04:03:13 -07001645struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001646 gfp_t gfp_flags, struct mempolicy **mpol,
1647 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001648{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001649 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001650
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001651 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001652 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001653
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001654 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1655 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001656 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001657 } else {
1658 zl = policy_zonelist(gfp_flags, *mpol);
1659 if ((*mpol)->mode == MPOL_BIND)
1660 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001661 }
1662 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001663}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001664
1665/*
1666 * init_nodemask_of_mempolicy
1667 *
1668 * If the current task's mempolicy is "default" [NULL], return 'false'
1669 * to indicate default policy. Otherwise, extract the policy nodemask
1670 * for 'bind' or 'interleave' policy into the argument nodemask, or
1671 * initialize the argument nodemask to contain the single node for
1672 * 'preferred' or 'local' policy and return 'true' to indicate presence
1673 * of non-default mempolicy.
1674 *
1675 * We don't bother with reference counting the mempolicy [mpol_get/put]
1676 * because the current task is examining it's own mempolicy and a task's
1677 * mempolicy is only ever changed by the task itself.
1678 *
1679 * N.B., it is the caller's responsibility to free a returned nodemask.
1680 */
1681bool init_nodemask_of_mempolicy(nodemask_t *mask)
1682{
1683 struct mempolicy *mempolicy;
1684 int nid;
1685
1686 if (!(mask && current->mempolicy))
1687 return false;
1688
Miao Xiec0ff7452010-05-24 14:32:08 -07001689 task_lock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001690 mempolicy = current->mempolicy;
1691 switch (mempolicy->mode) {
1692 case MPOL_PREFERRED:
1693 if (mempolicy->flags & MPOL_F_LOCAL)
1694 nid = numa_node_id();
1695 else
1696 nid = mempolicy->v.preferred_node;
1697 init_nodemask_of_node(mask, nid);
1698 break;
1699
1700 case MPOL_BIND:
1701 /* Fall through */
1702 case MPOL_INTERLEAVE:
1703 *mask = mempolicy->v.nodes;
1704 break;
1705
1706 default:
1707 BUG();
1708 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001709 task_unlock(current);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001710
1711 return true;
1712}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001713#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001714
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715/* Allocate a page in interleaved policy.
1716 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001717static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1718 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719{
1720 struct zonelist *zl;
1721 struct page *page;
1722
Mel Gorman0e884602008-04-28 02:12:14 -07001723 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001725 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001726 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 return page;
1728}
1729
1730/**
1731 * alloc_page_vma - Allocate a page for a VMA.
1732 *
1733 * @gfp:
1734 * %GFP_USER user allocation.
1735 * %GFP_KERNEL kernel allocations,
1736 * %GFP_HIGHMEM highmem/user allocations,
1737 * %GFP_FS allocation should not call back into a file system.
1738 * %GFP_ATOMIC don't sleep.
1739 *
1740 * @vma: Pointer to VMA or NULL if not available.
1741 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1742 *
1743 * This function allocates a page from the kernel page pool and applies
1744 * a NUMA policy associated with the VMA or the current process.
1745 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1746 * mm_struct of the VMA to prevent it from going away. Should be used for
1747 * all allocations for pages that will be mapped into
1748 * user space. Returns NULL when no page can be allocated.
1749 *
1750 * Should be called with the mm_sem of the vma hold.
1751 */
1752struct page *
Al Virodd0fc662005-10-07 07:46:04 +01001753alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001755 struct mempolicy *pol = get_vma_policy(current, vma, addr);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001756 struct zonelist *zl;
Miao Xiec0ff7452010-05-24 14:32:08 -07001757 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758
Miao Xiec0ff7452010-05-24 14:32:08 -07001759 get_mems_allowed();
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001760 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001762
1763 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001764 mpol_cond_put(pol);
Miao Xiec0ff7452010-05-24 14:32:08 -07001765 page = alloc_page_interleave(gfp, 0, nid);
1766 put_mems_allowed();
1767 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001769 zl = policy_zonelist(gfp, pol);
1770 if (unlikely(mpol_needs_cond_ref(pol))) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001771 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001772 * slow path: ref counted shared policy
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001773 */
Mel Gorman19770b32008-04-28 02:12:18 -07001774 struct page *page = __alloc_pages_nodemask(gfp, 0,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001775 zl, policy_nodemask(gfp, pol));
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001776 __mpol_put(pol);
Miao Xiec0ff7452010-05-24 14:32:08 -07001777 put_mems_allowed();
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001778 return page;
1779 }
1780 /*
1781 * fast path: default or task policy
1782 */
Miao Xiec0ff7452010-05-24 14:32:08 -07001783 page = __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1784 put_mems_allowed();
1785 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786}
1787
1788/**
1789 * alloc_pages_current - Allocate pages.
1790 *
1791 * @gfp:
1792 * %GFP_USER user allocation,
1793 * %GFP_KERNEL kernel allocation,
1794 * %GFP_HIGHMEM highmem allocation,
1795 * %GFP_FS don't call back into a file system.
1796 * %GFP_ATOMIC don't sleep.
1797 * @order: Power of two of allocation size in pages. 0 is a single page.
1798 *
1799 * Allocate a page from the kernel page pool. When not in
1800 * interrupt context and apply the current process NUMA policy.
1801 * Returns NULL when no page can be allocated.
1802 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001803 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804 * 1) it's ok to take cpuset_sem (can WAIT), and
1805 * 2) allocating for current task (not interrupt).
1806 */
Al Virodd0fc662005-10-07 07:46:04 +01001807struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808{
1809 struct mempolicy *pol = current->mempolicy;
Miao Xiec0ff7452010-05-24 14:32:08 -07001810 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Christoph Lameter9b819d22006-09-25 23:31:40 -07001812 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 pol = &default_policy;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001814
Miao Xiec0ff7452010-05-24 14:32:08 -07001815 get_mems_allowed();
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001816 /*
1817 * No reference counting needed for current->mempolicy
1818 * nor system default_policy
1819 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001820 if (pol->mode == MPOL_INTERLEAVE)
Miao Xiec0ff7452010-05-24 14:32:08 -07001821 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1822 else
1823 page = __alloc_pages_nodemask(gfp, order,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001824 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
Miao Xiec0ff7452010-05-24 14:32:08 -07001825 put_mems_allowed();
1826 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827}
1828EXPORT_SYMBOL(alloc_pages_current);
1829
Paul Jackson42253992006-01-08 01:01:59 -08001830/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001831 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08001832 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1833 * with the mems_allowed returned by cpuset_mems_allowed(). This
1834 * keeps mempolicies cpuset relative after its cpuset moves. See
1835 * further kernel/cpuset.c update_nodemask().
Miao Xie708c1bb2010-05-24 14:32:07 -07001836 *
1837 * current's mempolicy may be rebinded by the other task(the task that changes
1838 * cpuset's mems), so we needn't do rebind work for current task.
Paul Jackson42253992006-01-08 01:01:59 -08001839 */
Paul Jackson42253992006-01-08 01:01:59 -08001840
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001841/* Slow path of a mempolicy duplicate */
1842struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843{
1844 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1845
1846 if (!new)
1847 return ERR_PTR(-ENOMEM);
Miao Xie708c1bb2010-05-24 14:32:07 -07001848
1849 /* task's mempolicy is protected by alloc_lock */
1850 if (old == current->mempolicy) {
1851 task_lock(current);
1852 *new = *old;
1853 task_unlock(current);
1854 } else
1855 *new = *old;
1856
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08001857 rcu_read_lock();
Paul Jackson42253992006-01-08 01:01:59 -08001858 if (current_cpuset_is_being_rebound()) {
1859 nodemask_t mems = cpuset_mems_allowed(current);
Miao Xie708c1bb2010-05-24 14:32:07 -07001860 if (new->flags & MPOL_F_REBINDING)
1861 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1862 else
1863 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
Paul Jackson42253992006-01-08 01:01:59 -08001864 }
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08001865 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 return new;
1868}
1869
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001870/*
1871 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1872 * eliminate the * MPOL_F_* flags that require conditional ref and
1873 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1874 * after return. Use the returned value.
1875 *
1876 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1877 * policy lookup, even if the policy needs/has extra ref on lookup.
1878 * shmem_readahead needs this.
1879 */
1880struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1881 struct mempolicy *frompol)
1882{
1883 if (!mpol_needs_cond_ref(frompol))
1884 return frompol;
1885
1886 *tompol = *frompol;
1887 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1888 __mpol_put(frompol);
1889 return tompol;
1890}
1891
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892/* Slow path of a mempolicy comparison */
1893int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1894{
1895 if (!a || !b)
1896 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001897 if (a->mode != b->mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 return 0;
Bob Liu19800502010-05-24 14:32:01 -07001899 if (a->flags != b->flags)
David Rientjesf5b087b2008-04-28 02:12:27 -07001900 return 0;
Bob Liu19800502010-05-24 14:32:01 -07001901 if (mpol_store_user_nodemask(a))
1902 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
1903 return 0;
1904
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001905 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07001906 case MPOL_BIND:
1907 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -07001909 return nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001911 return a->v.preferred_node == b->v.preferred_node &&
1912 a->flags == b->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 default:
1914 BUG();
1915 return 0;
1916 }
1917}
1918
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 * Shared memory backing store policy support.
1921 *
1922 * Remember policies even when nobody has shared memory mapped.
1923 * The policies are kept in Red-Black tree linked from the inode.
1924 * They are protected by the sp->lock spinlock, which should be held
1925 * for any accesses to the tree.
1926 */
1927
1928/* lookup first element intersecting start-end */
1929/* Caller holds sp->lock */
1930static struct sp_node *
1931sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1932{
1933 struct rb_node *n = sp->root.rb_node;
1934
1935 while (n) {
1936 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1937
1938 if (start >= p->end)
1939 n = n->rb_right;
1940 else if (end <= p->start)
1941 n = n->rb_left;
1942 else
1943 break;
1944 }
1945 if (!n)
1946 return NULL;
1947 for (;;) {
1948 struct sp_node *w = NULL;
1949 struct rb_node *prev = rb_prev(n);
1950 if (!prev)
1951 break;
1952 w = rb_entry(prev, struct sp_node, nd);
1953 if (w->end <= start)
1954 break;
1955 n = prev;
1956 }
1957 return rb_entry(n, struct sp_node, nd);
1958}
1959
1960/* Insert a new shared policy into the list. */
1961/* Caller holds sp->lock */
1962static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1963{
1964 struct rb_node **p = &sp->root.rb_node;
1965 struct rb_node *parent = NULL;
1966 struct sp_node *nd;
1967
1968 while (*p) {
1969 parent = *p;
1970 nd = rb_entry(parent, struct sp_node, nd);
1971 if (new->start < nd->start)
1972 p = &(*p)->rb_left;
1973 else if (new->end > nd->end)
1974 p = &(*p)->rb_right;
1975 else
1976 BUG();
1977 }
1978 rb_link_node(&new->nd, parent, p);
1979 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07001980 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001981 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982}
1983
1984/* Find shared policy intersecting idx */
1985struct mempolicy *
1986mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1987{
1988 struct mempolicy *pol = NULL;
1989 struct sp_node *sn;
1990
1991 if (!sp->root.rb_node)
1992 return NULL;
1993 spin_lock(&sp->lock);
1994 sn = sp_lookup(sp, idx, idx+1);
1995 if (sn) {
1996 mpol_get(sn->policy);
1997 pol = sn->policy;
1998 }
1999 spin_unlock(&sp->lock);
2000 return pol;
2001}
2002
2003static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2004{
Paul Mundt140d5a42007-07-15 23:38:16 -07002005 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 rb_erase(&n->nd, &sp->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002007 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 kmem_cache_free(sn_cache, n);
2009}
2010
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07002011static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2012 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013{
2014 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2015
2016 if (!n)
2017 return NULL;
2018 n->start = start;
2019 n->end = end;
2020 mpol_get(pol);
Lee Schermerhornaab0b102008-04-28 02:13:13 -07002021 pol->flags |= MPOL_F_SHARED; /* for unref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 n->policy = pol;
2023 return n;
2024}
2025
2026/* Replace a policy range. */
2027static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2028 unsigned long end, struct sp_node *new)
2029{
2030 struct sp_node *n, *new2 = NULL;
2031
2032restart:
2033 spin_lock(&sp->lock);
2034 n = sp_lookup(sp, start, end);
2035 /* Take care of old policies in the same range. */
2036 while (n && n->start < end) {
2037 struct rb_node *next = rb_next(&n->nd);
2038 if (n->start >= start) {
2039 if (n->end <= end)
2040 sp_delete(sp, n);
2041 else
2042 n->start = end;
2043 } else {
2044 /* Old policy spanning whole new range. */
2045 if (n->end > end) {
2046 if (!new2) {
2047 spin_unlock(&sp->lock);
2048 new2 = sp_alloc(end, n->end, n->policy);
2049 if (!new2)
2050 return -ENOMEM;
2051 goto restart;
2052 }
2053 n->end = start;
2054 sp_insert(sp, new2);
2055 new2 = NULL;
2056 break;
2057 } else
2058 n->end = start;
2059 }
2060 if (!next)
2061 break;
2062 n = rb_entry(next, struct sp_node, nd);
2063 }
2064 if (new)
2065 sp_insert(sp, new);
2066 spin_unlock(&sp->lock);
2067 if (new2) {
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002068 mpol_put(new2->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 kmem_cache_free(sn_cache, new2);
2070 }
2071 return 0;
2072}
2073
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002074/**
2075 * mpol_shared_policy_init - initialize shared policy for inode
2076 * @sp: pointer to inode shared policy
2077 * @mpol: struct mempolicy to install
2078 *
2079 * Install non-NULL @mpol in inode's shared policy rb-tree.
2080 * On entry, the current task has a reference on a non-NULL @mpol.
2081 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002082 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002083 */
2084void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08002085{
Miao Xie58568d22009-06-16 15:31:49 -07002086 int ret;
2087
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002088 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2089 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08002090
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002091 if (mpol) {
2092 struct vm_area_struct pvma;
2093 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002094 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002095
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002096 if (!scratch)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002097 goto put_mpol;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002098 /* contextualize the tmpfs mount point mempolicy */
2099 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002100 if (IS_ERR(new))
Dan Carpenter0cae3452010-05-25 23:42:58 -07002101 goto free_scratch; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002102
2103 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002104 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002105 task_unlock(current);
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002106 if (ret)
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002107 goto put_new;
Robin Holt7339ff82006-01-14 13:20:48 -08002108
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002109 /* Create pseudo-vma that contains just the policy */
2110 memset(&pvma, 0, sizeof(struct vm_area_struct));
2111 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2112 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
Lee Schermerhorn15d77832010-05-24 14:32:04 -07002113
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002114put_new:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002115 mpol_put(new); /* drop initial ref */
Dan Carpenter0cae3452010-05-25 23:42:58 -07002116free_scratch:
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002117 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn5c0c1652010-06-29 15:05:30 -07002118put_mpol:
2119 mpol_put(mpol); /* drop our incoming ref on sb mpol */
Robin Holt7339ff82006-01-14 13:20:48 -08002120 }
2121}
2122
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123int mpol_set_shared_policy(struct shared_policy *info,
2124 struct vm_area_struct *vma, struct mempolicy *npol)
2125{
2126 int err;
2127 struct sp_node *new = NULL;
2128 unsigned long sz = vma_pages(vma);
2129
David Rientjes028fec42008-04-28 02:12:25 -07002130 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002132 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002133 npol ? npol->flags : -1,
Paul Mundt140d5a42007-07-15 23:38:16 -07002134 npol ? nodes_addr(npol->v.nodes)[0] : -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135
2136 if (npol) {
2137 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2138 if (!new)
2139 return -ENOMEM;
2140 }
2141 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2142 if (err && new)
2143 kmem_cache_free(sn_cache, new);
2144 return err;
2145}
2146
2147/* Free a backing policy store on inode delete. */
2148void mpol_free_shared_policy(struct shared_policy *p)
2149{
2150 struct sp_node *n;
2151 struct rb_node *next;
2152
2153 if (!p->root.rb_node)
2154 return;
2155 spin_lock(&p->lock);
2156 next = rb_first(&p->root);
2157 while (next) {
2158 n = rb_entry(next, struct sp_node, nd);
2159 next = rb_next(&n->nd);
Andi Kleen90c50292005-07-27 11:43:50 -07002160 rb_erase(&n->nd, &p->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002161 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 kmem_cache_free(sn_cache, n);
2163 }
2164 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165}
2166
2167/* assumes fs == KERNEL_DS */
2168void __init numa_policy_init(void)
2169{
Paul Mundtb71636e2007-07-15 23:38:15 -07002170 nodemask_t interleave_nodes;
2171 unsigned long largest = 0;
2172 int nid, prefer = 0;
2173
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 policy_cache = kmem_cache_create("numa_policy",
2175 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002176 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177
2178 sn_cache = kmem_cache_create("shared_policy_node",
2179 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002180 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181
Paul Mundtb71636e2007-07-15 23:38:15 -07002182 /*
2183 * Set interleaving policy for system init. Interleaving is only
2184 * enabled across suitably sized nodes (default is >= 16MB), or
2185 * fall back to the largest node if they're all smaller.
2186 */
2187 nodes_clear(interleave_nodes);
Christoph Lameter56bbd652007-10-16 01:25:35 -07002188 for_each_node_state(nid, N_HIGH_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002189 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190
Paul Mundtb71636e2007-07-15 23:38:15 -07002191 /* Preserve the largest node */
2192 if (largest < total_pages) {
2193 largest = total_pages;
2194 prefer = nid;
2195 }
2196
2197 /* Interleave this node? */
2198 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2199 node_set(nid, interleave_nodes);
2200 }
2201
2202 /* All too small, use the largest */
2203 if (unlikely(nodes_empty(interleave_nodes)))
2204 node_set(prefer, interleave_nodes);
2205
David Rientjes028fec42008-04-28 02:12:25 -07002206 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 printk("numa_policy_init: interleaving failed\n");
2208}
2209
Christoph Lameter8bccd852005-10-29 18:16:59 -07002210/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211void numa_default_policy(void)
2212{
David Rientjes028fec42008-04-28 02:12:25 -07002213 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214}
Paul Jackson68860ec2005-10-30 15:02:36 -08002215
Paul Jackson42253992006-01-08 01:01:59 -08002216/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002217 * Parse and format mempolicy from/to strings
2218 */
2219
2220/*
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002221 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002222 * Used only for mpol_parse_str() and mpol_to_str()
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002223 */
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002224#define MPOL_LOCAL MPOL_MAX
2225static const char * const policy_modes[] =
2226{
2227 [MPOL_DEFAULT] = "default",
2228 [MPOL_PREFERRED] = "prefer",
2229 [MPOL_BIND] = "bind",
2230 [MPOL_INTERLEAVE] = "interleave",
2231 [MPOL_LOCAL] = "local"
2232};
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002233
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002234
2235#ifdef CONFIG_TMPFS
2236/**
2237 * mpol_parse_str - parse string to mempolicy
2238 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002239 * @mpol: pointer to struct mempolicy pointer, returned on success.
2240 * @no_context: flag whether to "contextualize" the mempolicy
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002241 *
2242 * Format of input:
2243 * <mode>[=<flags>][:<nodelist>]
2244 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002245 * if @no_context is true, save the input nodemask in w.user_nodemask in
2246 * the returned mempolicy. This will be used to "clone" the mempolicy in
2247 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2248 * mount option. Note that if 'static' or 'relative' mode flags were
2249 * specified, the input nodemask will already have been saved. Saving
2250 * it again is redundant, but safe.
2251 *
2252 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002253 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002254int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002255{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002256 struct mempolicy *new = NULL;
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002257 unsigned short mode;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002258 unsigned short uninitialized_var(mode_flags);
2259 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002260 char *nodelist = strchr(str, ':');
2261 char *flags = strchr(str, '=');
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002262 int err = 1;
2263
2264 if (nodelist) {
2265 /* NUL-terminate mode or flags string */
2266 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002267 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002268 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002269 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002270 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002271 } else
2272 nodes_clear(nodes);
2273
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002274 if (flags)
2275 *flags++ = '\0'; /* terminate mode string */
2276
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002277 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002278 if (!strcmp(str, policy_modes[mode])) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002279 break;
2280 }
2281 }
Lee Schermerhornb4652e82010-05-24 14:32:03 -07002282 if (mode > MPOL_LOCAL)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002283 goto out;
2284
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002285 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002286 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002287 /*
2288 * Insist on a nodelist of one node only
2289 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002290 if (nodelist) {
2291 char *rest = nodelist;
2292 while (isdigit(*rest))
2293 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002294 if (*rest)
2295 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002296 }
2297 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002298 case MPOL_INTERLEAVE:
2299 /*
2300 * Default to online nodes with memory if no nodelist
2301 */
2302 if (!nodelist)
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002303 nodes = node_states[N_HIGH_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002304 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002305 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002306 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002307 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002308 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002309 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002310 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002311 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002312 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002313 case MPOL_DEFAULT:
2314 /*
2315 * Insist on a empty nodelist
2316 */
2317 if (!nodelist)
2318 err = 0;
2319 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002320 case MPOL_BIND:
2321 /*
2322 * Insist on a nodelist
2323 */
2324 if (!nodelist)
2325 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002326 }
2327
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002328 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002329 if (flags) {
2330 /*
2331 * Currently, we only support two mutually exclusive
2332 * mode flags.
2333 */
2334 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002335 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002336 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002337 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002338 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002339 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002340 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002341
2342 new = mpol_new(mode, mode_flags, &nodes);
2343 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002344 goto out;
2345
Lee Schermerhorne17f74a2010-05-24 14:32:02 -07002346 if (no_context) {
2347 /* save for contextualization */
2348 new->w.user_nodemask = nodes;
2349 } else {
Miao Xie58568d22009-06-16 15:31:49 -07002350 int ret;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002351 NODEMASK_SCRATCH(scratch);
2352 if (scratch) {
2353 task_lock(current);
2354 ret = mpol_set_nodemask(new, &nodes, scratch);
2355 task_unlock(current);
2356 } else
2357 ret = -ENOMEM;
2358 NODEMASK_SCRATCH_FREE(scratch);
2359 if (ret) {
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002360 mpol_put(new);
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002361 goto out;
Miao Xie58568d22009-06-16 15:31:49 -07002362 }
2363 }
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002364 err = 0;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002365
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002366out:
2367 /* Restore string for error message */
2368 if (nodelist)
2369 *--nodelist = ':';
2370 if (flags)
2371 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002372 if (!err)
2373 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002374 return err;
2375}
2376#endif /* CONFIG_TMPFS */
2377
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002378/**
2379 * mpol_to_str - format a mempolicy structure for printing
2380 * @buffer: to contain formatted mempolicy string
2381 * @maxlen: length of @buffer
2382 * @pol: pointer to mempolicy to be formatted
2383 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2384 *
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002385 * Convert a mempolicy into a string.
2386 * Returns the number of characters in buffer (if positive)
2387 * or an error (negative)
2388 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002389int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002390{
2391 char *p = buffer;
2392 int l;
2393 nodemask_t nodes;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002394 unsigned short mode;
David Rientjesf5b087b2008-04-28 02:12:27 -07002395 unsigned short flags = pol ? pol->flags : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002396
Lee Schermerhorn22919902008-04-28 02:13:22 -07002397 /*
2398 * Sanity check: room for longest mode, flag and some nodes
2399 */
2400 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2401
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002402 if (!pol || pol == &default_policy)
2403 mode = MPOL_DEFAULT;
2404 else
2405 mode = pol->mode;
2406
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002407 switch (mode) {
2408 case MPOL_DEFAULT:
2409 nodes_clear(nodes);
2410 break;
2411
2412 case MPOL_PREFERRED:
2413 nodes_clear(nodes);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002414 if (flags & MPOL_F_LOCAL)
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002415 mode = MPOL_LOCAL; /* pseudo-policy */
2416 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002417 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002418 break;
2419
2420 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07002421 /* Fall through */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002422 case MPOL_INTERLEAVE:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002423 if (no_context)
2424 nodes = pol->w.user_nodemask;
2425 else
2426 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002427 break;
2428
2429 default:
2430 BUG();
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002431 }
2432
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002433 l = strlen(policy_modes[mode]);
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002434 if (buffer + maxlen < p + l + 1)
2435 return -ENOSPC;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002436
Lee Schermerhorn345ace92010-05-24 14:32:04 -07002437 strcpy(p, policy_modes[mode]);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002438 p += l;
2439
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002440 if (flags & MPOL_MODE_FLAGS) {
David Rientjesf5b087b2008-04-28 02:12:27 -07002441 if (buffer + maxlen < p + 2)
2442 return -ENOSPC;
2443 *p++ = '=';
2444
Lee Schermerhorn22919902008-04-28 02:13:22 -07002445 /*
2446 * Currently, the only defined flags are mutually exclusive
2447 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002448 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002449 p += snprintf(p, buffer + maxlen - p, "static");
2450 else if (flags & MPOL_F_RELATIVE_NODES)
2451 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002452 }
2453
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002454 if (!nodes_empty(nodes)) {
2455 if (buffer + maxlen < p + 2)
2456 return -ENOSPC;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002457 *p++ = ':';
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002458 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2459 }
2460 return p - buffer;
2461}
2462
2463struct numa_maps {
2464 unsigned long pages;
2465 unsigned long anon;
Christoph Lameter397874d2006-03-06 15:42:53 -08002466 unsigned long active;
2467 unsigned long writeback;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002468 unsigned long mapcount_max;
Christoph Lameter397874d2006-03-06 15:42:53 -08002469 unsigned long dirty;
2470 unsigned long swapcache;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002471 unsigned long node[MAX_NUMNODES];
2472};
2473
Christoph Lameter397874d2006-03-06 15:42:53 -08002474static void gather_stats(struct page *page, void *private, int pte_dirty)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002475{
2476 struct numa_maps *md = private;
2477 int count = page_mapcount(page);
2478
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002479 md->pages++;
Christoph Lameter397874d2006-03-06 15:42:53 -08002480 if (pte_dirty || PageDirty(page))
2481 md->dirty++;
2482
2483 if (PageSwapCache(page))
2484 md->swapcache++;
2485
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002486 if (PageActive(page) || PageUnevictable(page))
Christoph Lameter397874d2006-03-06 15:42:53 -08002487 md->active++;
2488
2489 if (PageWriteback(page))
2490 md->writeback++;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002491
2492 if (PageAnon(page))
2493 md->anon++;
2494
Christoph Lameter397874d2006-03-06 15:42:53 -08002495 if (count > md->mapcount_max)
2496 md->mapcount_max = count;
2497
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002498 md->node[page_to_nid(page)]++;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002499}
2500
Andrew Morton7f709ed2006-03-07 21:55:22 -08002501#ifdef CONFIG_HUGETLB_PAGE
Christoph Lameter397874d2006-03-06 15:42:53 -08002502static void check_huge_range(struct vm_area_struct *vma,
2503 unsigned long start, unsigned long end,
2504 struct numa_maps *md)
2505{
2506 unsigned long addr;
2507 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002508 struct hstate *h = hstate_vma(vma);
2509 unsigned long sz = huge_page_size(h);
Christoph Lameter397874d2006-03-06 15:42:53 -08002510
Andi Kleena5516432008-07-23 21:27:41 -07002511 for (addr = start; addr < end; addr += sz) {
2512 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2513 addr & huge_page_mask(h));
Christoph Lameter397874d2006-03-06 15:42:53 -08002514 pte_t pte;
2515
2516 if (!ptep)
2517 continue;
2518
2519 pte = *ptep;
2520 if (pte_none(pte))
2521 continue;
2522
2523 page = pte_page(pte);
2524 if (!page)
2525 continue;
2526
2527 gather_stats(page, md, pte_dirty(*ptep));
2528 }
2529}
Andrew Morton7f709ed2006-03-07 21:55:22 -08002530#else
2531static inline void check_huge_range(struct vm_area_struct *vma,
2532 unsigned long start, unsigned long end,
2533 struct numa_maps *md)
2534{
2535}
2536#endif
Christoph Lameter397874d2006-03-06 15:42:53 -08002537
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002538/*
2539 * Display pages allocated per node and memory policy via /proc.
2540 */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002541int show_numa_map(struct seq_file *m, void *v)
2542{
Eric W. Biederman99f89552006-06-26 00:25:55 -07002543 struct proc_maps_private *priv = m->private;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002544 struct vm_area_struct *vma = v;
2545 struct numa_maps *md;
Christoph Lameter397874d2006-03-06 15:42:53 -08002546 struct file *file = vma->vm_file;
2547 struct mm_struct *mm = vma->vm_mm;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07002548 struct mempolicy *pol;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002549 int n;
2550 char buffer[50];
2551
Christoph Lameter397874d2006-03-06 15:42:53 -08002552 if (!mm)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002553 return 0;
2554
2555 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2556 if (!md)
2557 return 0;
2558
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07002559 pol = get_vma_policy(priv->task, vma, vma->vm_start);
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002560 mpol_to_str(buffer, sizeof(buffer), pol, 0);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002561 mpol_cond_put(pol);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002562
Christoph Lameter397874d2006-03-06 15:42:53 -08002563 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002564
Christoph Lameter397874d2006-03-06 15:42:53 -08002565 if (file) {
2566 seq_printf(m, " file=");
Jan Blunckc32c2f62008-02-14 19:38:43 -08002567 seq_path(m, &file->f_path, "\n\t= ");
Christoph Lameter397874d2006-03-06 15:42:53 -08002568 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2569 seq_printf(m, " heap");
2570 } else if (vma->vm_start <= mm->start_stack &&
2571 vma->vm_end >= mm->start_stack) {
2572 seq_printf(m, " stack");
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002573 }
Christoph Lameter397874d2006-03-06 15:42:53 -08002574
2575 if (is_vm_hugetlb_page(vma)) {
2576 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2577 seq_printf(m, " huge");
2578 } else {
2579 check_pgd_range(vma, vma->vm_start, vma->vm_end,
Christoph Lameter56bbd652007-10-16 01:25:35 -07002580 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
Christoph Lameter397874d2006-03-06 15:42:53 -08002581 }
2582
2583 if (!md->pages)
2584 goto out;
2585
2586 if (md->anon)
2587 seq_printf(m," anon=%lu",md->anon);
2588
2589 if (md->dirty)
2590 seq_printf(m," dirty=%lu",md->dirty);
2591
2592 if (md->pages != md->anon && md->pages != md->dirty)
2593 seq_printf(m, " mapped=%lu", md->pages);
2594
2595 if (md->mapcount_max > 1)
2596 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2597
2598 if (md->swapcache)
2599 seq_printf(m," swapcache=%lu", md->swapcache);
2600
2601 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2602 seq_printf(m," active=%lu", md->active);
2603
2604 if (md->writeback)
2605 seq_printf(m," writeback=%lu", md->writeback);
2606
Christoph Lameter56bbd652007-10-16 01:25:35 -07002607 for_each_node_state(n, N_HIGH_MEMORY)
Christoph Lameter397874d2006-03-06 15:42:53 -08002608 if (md->node[n])
2609 seq_printf(m, " N%d=%lu", n, md->node[n]);
2610out:
2611 seq_putc(m, '\n');
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002612 kfree(md);
2613
2614 if (m->count < m->size)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002615 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002616 return 0;
2617}