blob: 08f40a2f3fe0aef1819636d5983b3242758463e9 [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);
122 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
123} mpol_ops[MPOL_MAX];
124
Mel Gorman19770b32008-04-28 02:12:18 -0700125/* Check that the nodemask contains at least one populated zone */
David Rientjes37012942008-04-28 02:12:33 -0700126static int is_valid_nodemask(const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127{
Mel Gorman19770b32008-04-28 02:12:18 -0700128 int nd, k;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
Mel Gorman19770b32008-04-28 02:12:18 -0700130 /* Check that there is something useful in this mask */
131 k = policy_zone;
132
133 for_each_node_mask(nd, *nodemask) {
134 struct zone *z;
135
136 for (k = 0; k <= policy_zone; k++) {
137 z = &NODE_DATA(nd)->node_zones[k];
138 if (z->present_pages > 0)
139 return 1;
Andi Kleendd942ae2006-02-17 01:39:16 +0100140 }
141 }
Mel Gorman19770b32008-04-28 02:12:18 -0700142
143 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144}
145
David Rientjesf5b087b2008-04-28 02:12:27 -0700146static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
147{
David Rientjes4c50bc02008-04-28 02:12:30 -0700148 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
149}
150
151static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
152 const nodemask_t *rel)
153{
154 nodemask_t tmp;
155 nodes_fold(tmp, *orig, nodes_weight(*rel));
156 nodes_onto(*ret, tmp, *rel);
David Rientjesf5b087b2008-04-28 02:12:27 -0700157}
158
David Rientjes37012942008-04-28 02:12:33 -0700159static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
160{
161 if (nodes_empty(*nodes))
162 return -EINVAL;
163 pol->v.nodes = *nodes;
164 return 0;
165}
166
167static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
168{
169 if (!nodes)
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700170 pol->flags |= MPOL_F_LOCAL; /* local allocation */
David Rientjes37012942008-04-28 02:12:33 -0700171 else if (nodes_empty(*nodes))
172 return -EINVAL; /* no allowed nodes */
173 else
174 pol->v.preferred_node = first_node(*nodes);
175 return 0;
176}
177
178static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
179{
180 if (!is_valid_nodemask(nodes))
181 return -EINVAL;
182 pol->v.nodes = *nodes;
183 return 0;
184}
185
Miao Xie58568d22009-06-16 15:31:49 -0700186/*
187 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
188 * any, for the new policy. mpol_new() has already validated the nodes
189 * parameter with respect to the policy mode and flags. But, we need to
190 * handle an empty nodemask with MPOL_PREFERRED here.
191 *
192 * Must be called holding task's alloc_lock to protect task's mems_allowed
193 * and mempolicy. May also be called holding the mmap_semaphore for write.
194 */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700195static int mpol_set_nodemask(struct mempolicy *pol,
196 const nodemask_t *nodes, struct nodemask_scratch *nsc)
Miao Xie58568d22009-06-16 15:31:49 -0700197{
Miao Xie58568d22009-06-16 15:31:49 -0700198 int ret;
199
200 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
201 if (pol == NULL)
202 return 0;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700203 /* Check N_HIGH_MEMORY */
204 nodes_and(nsc->mask1,
205 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
Miao Xie58568d22009-06-16 15:31:49 -0700206
207 VM_BUG_ON(!nodes);
208 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
209 nodes = NULL; /* explicit local allocation */
210 else {
211 if (pol->flags & MPOL_F_RELATIVE_NODES)
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700212 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
Miao Xie58568d22009-06-16 15:31:49 -0700213 else
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700214 nodes_and(nsc->mask2, *nodes, nsc->mask1);
215
Miao Xie58568d22009-06-16 15:31:49 -0700216 if (mpol_store_user_nodemask(pol))
217 pol->w.user_nodemask = *nodes;
218 else
219 pol->w.cpuset_mems_allowed =
220 cpuset_current_mems_allowed;
221 }
222
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700223 if (nodes)
224 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
225 else
226 ret = mpol_ops[pol->mode].create(pol, NULL);
Miao Xie58568d22009-06-16 15:31:49 -0700227 return ret;
228}
229
230/*
231 * This function just creates a new policy, does some check and simple
232 * initialization. You must invoke mpol_set_nodemask() to set nodes.
233 */
David Rientjes028fec42008-04-28 02:12:25 -0700234static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
235 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
237 struct mempolicy *policy;
238
David Rientjes028fec42008-04-28 02:12:25 -0700239 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
240 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
Paul Mundt140d5a42007-07-15 23:38:16 -0700241
David Rientjes3e1f0642008-04-28 02:12:34 -0700242 if (mode == MPOL_DEFAULT) {
243 if (nodes && !nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700244 return ERR_PTR(-EINVAL);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700245 return NULL; /* simply delete any existing policy */
David Rientjes37012942008-04-28 02:12:33 -0700246 }
David Rientjes3e1f0642008-04-28 02:12:34 -0700247 VM_BUG_ON(!nodes);
248
249 /*
250 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
251 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
252 * All other modes require a valid pointer to a non-empty nodemask.
253 */
254 if (mode == MPOL_PREFERRED) {
255 if (nodes_empty(*nodes)) {
256 if (((flags & MPOL_F_STATIC_NODES) ||
257 (flags & MPOL_F_RELATIVE_NODES)))
258 return ERR_PTR(-EINVAL);
David Rientjes3e1f0642008-04-28 02:12:34 -0700259 }
260 } else if (nodes_empty(*nodes))
261 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
263 if (!policy)
264 return ERR_PTR(-ENOMEM);
265 atomic_set(&policy->refcnt, 1);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700266 policy->mode = mode;
David Rientjes3e1f0642008-04-28 02:12:34 -0700267 policy->flags = flags;
David Rientjesf5b087b2008-04-28 02:12:27 -0700268
David Rientjes37012942008-04-28 02:12:33 -0700269 return policy;
270}
271
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700272/* Slow path of a mpol destructor. */
273void __mpol_put(struct mempolicy *p)
274{
275 if (!atomic_dec_and_test(&p->refcnt))
276 return;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700277 kmem_cache_free(policy_cache, p);
278}
279
David Rientjes37012942008-04-28 02:12:33 -0700280static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
281{
282}
283
284static void mpol_rebind_nodemask(struct mempolicy *pol,
285 const nodemask_t *nodes)
286{
287 nodemask_t tmp;
288
289 if (pol->flags & MPOL_F_STATIC_NODES)
290 nodes_and(tmp, pol->w.user_nodemask, *nodes);
291 else if (pol->flags & MPOL_F_RELATIVE_NODES)
292 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
293 else {
294 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
295 *nodes);
296 pol->w.cpuset_mems_allowed = *nodes;
297 }
298
299 pol->v.nodes = tmp;
300 if (!node_isset(current->il_next, tmp)) {
301 current->il_next = next_node(current->il_next, tmp);
302 if (current->il_next >= MAX_NUMNODES)
303 current->il_next = first_node(tmp);
304 if (current->il_next >= MAX_NUMNODES)
305 current->il_next = numa_node_id();
306 }
307}
308
309static void mpol_rebind_preferred(struct mempolicy *pol,
310 const nodemask_t *nodes)
311{
312 nodemask_t tmp;
313
David Rientjes37012942008-04-28 02:12:33 -0700314 if (pol->flags & MPOL_F_STATIC_NODES) {
315 int node = first_node(pol->w.user_nodemask);
316
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700317 if (node_isset(node, *nodes)) {
David Rientjes37012942008-04-28 02:12:33 -0700318 pol->v.preferred_node = node;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700319 pol->flags &= ~MPOL_F_LOCAL;
320 } else
321 pol->flags |= MPOL_F_LOCAL;
David Rientjes37012942008-04-28 02:12:33 -0700322 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
323 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
324 pol->v.preferred_node = first_node(tmp);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700325 } else if (!(pol->flags & MPOL_F_LOCAL)) {
David Rientjes37012942008-04-28 02:12:33 -0700326 pol->v.preferred_node = node_remap(pol->v.preferred_node,
327 pol->w.cpuset_mems_allowed,
328 *nodes);
329 pol->w.cpuset_mems_allowed = *nodes;
330 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331}
332
David Rientjes1d0d2682008-04-28 02:12:32 -0700333/* Migrate a policy to a different set of nodes */
334static void mpol_rebind_policy(struct mempolicy *pol,
335 const nodemask_t *newmask)
336{
David Rientjes1d0d2682008-04-28 02:12:32 -0700337 if (!pol)
338 return;
David Rientjes1d0d2682008-04-28 02:12:32 -0700339 if (!mpol_store_user_nodemask(pol) &&
340 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
341 return;
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700342 mpol_ops[pol->mode].rebind(pol, newmask);
David Rientjes1d0d2682008-04-28 02:12:32 -0700343}
344
345/*
346 * Wrapper for mpol_rebind_policy() that just requires task
347 * pointer, and updates task mempolicy.
Miao Xie58568d22009-06-16 15:31:49 -0700348 *
349 * Called with task's alloc_lock held.
David Rientjes1d0d2682008-04-28 02:12:32 -0700350 */
351
352void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
353{
354 mpol_rebind_policy(tsk->mempolicy, new);
355}
356
357/*
358 * Rebind each vma in mm to new nodemask.
359 *
360 * Call holding a reference to mm. Takes mm->mmap_sem during call.
361 */
362
363void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
364{
365 struct vm_area_struct *vma;
366
367 down_write(&mm->mmap_sem);
368 for (vma = mm->mmap; vma; vma = vma->vm_next)
369 mpol_rebind_policy(vma->vm_policy, new);
370 up_write(&mm->mmap_sem);
371}
372
David Rientjes37012942008-04-28 02:12:33 -0700373static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
374 [MPOL_DEFAULT] = {
375 .rebind = mpol_rebind_default,
376 },
377 [MPOL_INTERLEAVE] = {
378 .create = mpol_new_interleave,
379 .rebind = mpol_rebind_nodemask,
380 },
381 [MPOL_PREFERRED] = {
382 .create = mpol_new_preferred,
383 .rebind = mpol_rebind_preferred,
384 },
385 [MPOL_BIND] = {
386 .create = mpol_new_bind,
387 .rebind = mpol_rebind_nodemask,
388 },
389};
390
Christoph Lameter397874d2006-03-06 15:42:53 -0800391static void gather_stats(struct page *, void *, int pte_dirty);
Christoph Lameterfc301282006-01-18 17:42:29 -0800392static void migrate_page_add(struct page *page, struct list_head *pagelist,
393 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800394
Christoph Lameter38e35862006-01-08 01:01:01 -0800395/* Scan through pages checking if pages follow certain conditions. */
Nick Pigginb5810032005-10-29 18:16:12 -0700396static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800397 unsigned long addr, unsigned long end,
398 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800399 void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400{
Hugh Dickins91612e02005-06-21 17:15:07 -0700401 pte_t *orig_pte;
402 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700403 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700404
Hugh Dickins705e87c2005-10-29 18:16:27 -0700405 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700406 do {
Linus Torvalds6aab3412005-11-28 14:34:23 -0800407 struct page *page;
Andy Whitcroft25ba77c2006-12-06 20:33:03 -0800408 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700409
410 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800412 page = vm_normal_page(vma, addr, *pte);
413 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800415 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800416 * vm_normal_page() filters out zero pages, but there might
417 * still be PageReserved pages to skip, perhaps in a VDSO.
418 * And we cannot move PageKsm pages sensibly or safely yet.
Nick Piggin053837f2006-01-18 17:42:27 -0800419 */
Hugh Dickins62b61f62009-12-14 17:59:33 -0800420 if (PageReserved(page) || PageKsm(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800421 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800422 nid = page_to_nid(page);
Christoph Lameter38e35862006-01-08 01:01:01 -0800423 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
424 continue;
425
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800426 if (flags & MPOL_MF_STATS)
Christoph Lameter397874d2006-03-06 15:42:53 -0800427 gather_stats(page, private, pte_dirty(*pte));
Nick Piggin053837f2006-01-18 17:42:27 -0800428 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameterfc301282006-01-18 17:42:29 -0800429 migrate_page_add(page, private, flags);
Christoph Lameter38e35862006-01-08 01:01:01 -0800430 else
431 break;
Hugh Dickins91612e02005-06-21 17:15:07 -0700432 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickins705e87c2005-10-29 18:16:27 -0700433 pte_unmap_unlock(orig_pte, ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700434 return addr != end;
435}
436
Nick Pigginb5810032005-10-29 18:16:12 -0700437static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800438 unsigned long addr, unsigned long end,
439 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800440 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700441{
442 pmd_t *pmd;
443 unsigned long next;
444
445 pmd = pmd_offset(pud, addr);
446 do {
447 next = pmd_addr_end(addr, end);
448 if (pmd_none_or_clear_bad(pmd))
449 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800450 if (check_pte_range(vma, pmd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800451 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700452 return -EIO;
453 } while (pmd++, addr = next, addr != end);
454 return 0;
455}
456
Nick Pigginb5810032005-10-29 18:16:12 -0700457static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800458 unsigned long addr, unsigned long end,
459 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800460 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700461{
462 pud_t *pud;
463 unsigned long next;
464
465 pud = pud_offset(pgd, addr);
466 do {
467 next = pud_addr_end(addr, end);
468 if (pud_none_or_clear_bad(pud))
469 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800470 if (check_pmd_range(vma, pud, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800471 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700472 return -EIO;
473 } while (pud++, addr = next, addr != end);
474 return 0;
475}
476
Nick Pigginb5810032005-10-29 18:16:12 -0700477static inline int check_pgd_range(struct vm_area_struct *vma,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800478 unsigned long addr, unsigned long end,
479 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800480 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700481{
482 pgd_t *pgd;
483 unsigned long next;
484
Nick Pigginb5810032005-10-29 18:16:12 -0700485 pgd = pgd_offset(vma->vm_mm, addr);
Hugh Dickins91612e02005-06-21 17:15:07 -0700486 do {
487 next = pgd_addr_end(addr, end);
488 if (pgd_none_or_clear_bad(pgd))
489 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800490 if (check_pud_range(vma, pgd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800491 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700492 return -EIO;
493 } while (pgd++, addr = next, addr != end);
494 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495}
496
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800497/*
498 * Check if all pages in a range are on a set of nodes.
499 * If pagelist != NULL then isolate pages from the LRU and
500 * put them on the pagelist.
501 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502static struct vm_area_struct *
503check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Christoph Lameter38e35862006-01-08 01:01:01 -0800504 const nodemask_t *nodes, unsigned long flags, void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505{
506 int err;
507 struct vm_area_struct *first, *vma, *prev;
508
Nick Piggin053837f2006-01-18 17:42:27 -0800509
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 first = find_vma(mm, start);
511 if (!first)
512 return ERR_PTR(-EFAULT);
513 prev = NULL;
514 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800515 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
516 if (!vma->vm_next && vma->vm_end < end)
517 return ERR_PTR(-EFAULT);
518 if (prev && prev->vm_end < vma->vm_start)
519 return ERR_PTR(-EFAULT);
520 }
521 if (!is_vm_hugetlb_page(vma) &&
522 ((flags & MPOL_MF_STRICT) ||
523 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
524 vma_migratable(vma)))) {
Andi Kleen5b952b32005-09-13 01:25:08 -0700525 unsigned long endvma = vma->vm_end;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800526
Andi Kleen5b952b32005-09-13 01:25:08 -0700527 if (endvma > end)
528 endvma = end;
529 if (vma->vm_start > start)
530 start = vma->vm_start;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800531 err = check_pgd_range(vma, start, endvma, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800532 flags, private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 if (err) {
534 first = ERR_PTR(err);
535 break;
536 }
537 }
538 prev = vma;
539 }
540 return first;
541}
542
543/* Apply policy to a single VMA */
544static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
545{
546 int err = 0;
547 struct mempolicy *old = vma->vm_policy;
548
Paul Mundt140d5a42007-07-15 23:38:16 -0700549 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 vma->vm_start, vma->vm_end, vma->vm_pgoff,
551 vma->vm_ops, vma->vm_file,
552 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
553
554 if (vma->vm_ops && vma->vm_ops->set_policy)
555 err = vma->vm_ops->set_policy(vma, new);
556 if (!err) {
557 mpol_get(new);
558 vma->vm_policy = new;
Lee Schermerhornf0be3d32008-04-28 02:13:08 -0700559 mpol_put(old);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560 }
561 return err;
562}
563
564/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800565static int mbind_range(struct mm_struct *mm, unsigned long start,
566 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
568 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800569 struct vm_area_struct *prev;
570 struct vm_area_struct *vma;
571 int err = 0;
572 pgoff_t pgoff;
573 unsigned long vmstart;
574 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800576 vma = find_vma_prev(mm, start, &prev);
577 if (!vma || vma->vm_start > start)
578 return -EFAULT;
579
580 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800582 vmstart = max(start, vma->vm_start);
583 vmend = min(end, vma->vm_end);
584
585 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
586 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
587 vma->anon_vma, vma->vm_file, pgoff, new_pol);
588 if (prev) {
589 vma = prev;
590 next = vma->vm_next;
591 continue;
592 }
593 if (vma->vm_start != vmstart) {
594 err = split_vma(vma->vm_mm, vma, vmstart, 1);
595 if (err)
596 goto out;
597 }
598 if (vma->vm_end != vmend) {
599 err = split_vma(vma->vm_mm, vma, vmend, 0);
600 if (err)
601 goto out;
602 }
603 err = policy_vma(vma, new_pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 if (err)
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800605 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800607
608 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 return err;
610}
611
Paul Jacksonc61afb12006-03-24 03:16:08 -0800612/*
613 * Update task->flags PF_MEMPOLICY bit: set iff non-default
614 * mempolicy. Allows more rapid checking of this (combined perhaps
615 * with other PF_* flag bits) on memory allocation hot code paths.
616 *
617 * If called from outside this file, the task 'p' should -only- be
618 * a newly forked child not yet visible on the task list, because
619 * manipulating the task flags of a visible task is not safe.
620 *
621 * The above limitation is why this routine has the funny name
622 * mpol_fix_fork_child_flag().
623 *
624 * It is also safe to call this with a task pointer of current,
625 * which the static wrapper mpol_set_task_struct_flag() does,
626 * for use within this file.
627 */
628
629void mpol_fix_fork_child_flag(struct task_struct *p)
630{
631 if (p->mempolicy)
632 p->flags |= PF_MEMPOLICY;
633 else
634 p->flags &= ~PF_MEMPOLICY;
635}
636
637static void mpol_set_task_struct_flag(void)
638{
639 mpol_fix_fork_child_flag(current);
640}
641
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700643static long do_set_mempolicy(unsigned short mode, unsigned short flags,
644 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645{
Miao Xie58568d22009-06-16 15:31:49 -0700646 struct mempolicy *new, *old;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700647 struct mm_struct *mm = current->mm;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700648 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700649 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700651 if (!scratch)
652 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700653
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700654 new = mpol_new(mode, flags, nodes);
655 if (IS_ERR(new)) {
656 ret = PTR_ERR(new);
657 goto out;
658 }
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700659 /*
660 * prevent changing our mempolicy while show_numa_maps()
661 * is using it.
662 * Note: do_set_mempolicy() can be called at init time
663 * with no 'mm'.
664 */
665 if (mm)
666 down_write(&mm->mmap_sem);
Miao Xie58568d22009-06-16 15:31:49 -0700667 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700668 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700669 if (ret) {
670 task_unlock(current);
671 if (mm)
672 up_write(&mm->mmap_sem);
673 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700674 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700675 }
676 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 current->mempolicy = new;
Paul Jacksonc61afb12006-03-24 03:16:08 -0800678 mpol_set_task_struct_flag();
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700679 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700680 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700681 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700682 task_unlock(current);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700683 if (mm)
684 up_write(&mm->mmap_sem);
685
Miao Xie58568d22009-06-16 15:31:49 -0700686 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700687 ret = 0;
688out:
689 NODEMASK_SCRATCH_FREE(scratch);
690 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691}
692
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700693/*
694 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700695 *
696 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700697 */
698static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700700 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700701 if (p == &default_policy)
702 return;
703
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700704 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700705 case MPOL_BIND:
706 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700708 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709 break;
710 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700711 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700712 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700713 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 break;
715 default:
716 BUG();
717 }
718}
719
720static int lookup_node(struct mm_struct *mm, unsigned long addr)
721{
722 struct page *p;
723 int err;
724
725 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
726 if (err >= 0) {
727 err = page_to_nid(p);
728 put_page(p);
729 }
730 return err;
731}
732
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700734static long do_get_mempolicy(int *policy, nodemask_t *nmask,
735 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700737 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 struct mm_struct *mm = current->mm;
739 struct vm_area_struct *vma = NULL;
740 struct mempolicy *pol = current->mempolicy;
741
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700742 if (flags &
743 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700745
746 if (flags & MPOL_F_MEMS_ALLOWED) {
747 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
748 return -EINVAL;
749 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700750 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700751 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700752 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700753 return 0;
754 }
755
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700757 /*
758 * Do NOT fall back to task policy if the
759 * vma/shared policy at addr is NULL. We
760 * want to return MPOL_DEFAULT in this case.
761 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 down_read(&mm->mmap_sem);
763 vma = find_vma_intersection(mm, addr, addr+1);
764 if (!vma) {
765 up_read(&mm->mmap_sem);
766 return -EFAULT;
767 }
768 if (vma->vm_ops && vma->vm_ops->get_policy)
769 pol = vma->vm_ops->get_policy(vma, addr);
770 else
771 pol = vma->vm_policy;
772 } else if (addr)
773 return -EINVAL;
774
775 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700776 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777
778 if (flags & MPOL_F_NODE) {
779 if (flags & MPOL_F_ADDR) {
780 err = lookup_node(mm, addr);
781 if (err < 0)
782 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700783 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700785 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700786 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 } else {
788 err = -EINVAL;
789 goto out;
790 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700791 } else {
792 *policy = pol == &default_policy ? MPOL_DEFAULT :
793 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700794 /*
795 * Internal mempolicy flags must be masked off before exposing
796 * the policy to userspace.
797 */
798 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700799 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800
801 if (vma) {
802 up_read(&current->mm->mmap_sem);
803 vma = NULL;
804 }
805
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700807 if (nmask) {
Lee Schermerhornc6b6ef82010-03-23 13:35:41 -0700808 if (mpol_store_user_nodemask(pol)) {
809 *nmask = pol->w.user_nodemask;
810 } else {
811 task_lock(current);
812 get_policy_nodemask(pol, nmask);
813 task_unlock(current);
814 }
Miao Xie58568d22009-06-16 15:31:49 -0700815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816
817 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700818 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 if (vma)
820 up_read(&current->mm->mmap_sem);
821 return err;
822}
823
Christoph Lameterb20a3502006-03-22 00:09:12 -0800824#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700825/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800826 * page migration
827 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800828static void migrate_page_add(struct page *page, struct list_head *pagelist,
829 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800830{
831 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800832 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800833 */
Nick Piggin62695a82008-10-18 20:26:09 -0700834 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
835 if (!isolate_lru_page(page)) {
836 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800837 inc_zone_page_state(page, NR_ISOLATED_ANON +
838 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700839 }
840 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800841}
842
Christoph Lameter742755a2006-06-23 02:03:55 -0700843static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700844{
Mel Gorman6484eb32009-06-16 15:31:54 -0700845 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700846}
847
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800848/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800849 * Migrate pages from one node to a target node.
850 * Returns error or the number of pages not migrated.
851 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700852static int migrate_to_node(struct mm_struct *mm, int source, int dest,
853 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800854{
855 nodemask_t nmask;
856 LIST_HEAD(pagelist);
857 int err = 0;
858
859 nodes_clear(nmask);
860 node_set(source, nmask);
861
862 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
863 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
864
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700865 if (!list_empty(&pagelist))
Hugh Dickins62b61f62009-12-14 17:59:33 -0800866 err = migrate_pages(&pagelist, new_node_page, dest, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700867
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800868 return err;
869}
870
871/*
872 * Move pages between the two nodesets so as to preserve the physical
873 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -0800874 *
875 * Returns the number of page that could not be moved.
876 */
877int do_migrate_pages(struct mm_struct *mm,
878 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
879{
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800880 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800881 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800882 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -0800883
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800884 err = migrate_prep();
885 if (err)
886 return err;
887
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700888 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -0800889
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700890 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
891 if (err)
892 goto out;
893
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -0800894 /*
895 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
896 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
897 * bit in 'tmp', and return that <source, dest> pair for migration.
898 * The pair of nodemasks 'to' and 'from' define the map.
899 *
900 * If no pair of bits is found that way, fallback to picking some
901 * pair of 'source' and 'dest' bits that are not the same. If the
902 * 'source' and 'dest' bits are the same, this represents a node
903 * that will be migrating to itself, so no pages need move.
904 *
905 * If no bits are left in 'tmp', or if all remaining bits left
906 * in 'tmp' correspond to the same bit in 'to', return false
907 * (nothing left to migrate).
908 *
909 * This lets us pick a pair of nodes to migrate between, such that
910 * if possible the dest node is not already occupied by some other
911 * source node, minimizing the risk of overloading the memory on a
912 * node that would happen if we migrated incoming memory to a node
913 * before migrating outgoing memory source that same node.
914 *
915 * A single scan of tmp is sufficient. As we go, we remember the
916 * most recent <s, d> pair that moved (s != d). If we find a pair
917 * that not only moved, but what's better, moved to an empty slot
918 * (d is not set in tmp), then we break out then, with that pair.
919 * Otherwise when we finish scannng from_tmp, we at least have the
920 * most recent <s, d> pair that moved. If we get all the way through
921 * the scan of tmp without finding any node that moved, much less
922 * moved to an empty node, then there is nothing left worth migrating.
923 */
Christoph Lameterd4984712006-01-08 01:00:55 -0800924
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800925 tmp = *from_nodes;
926 while (!nodes_empty(tmp)) {
927 int s,d;
928 int source = -1;
929 int dest = 0;
930
931 for_each_node_mask(s, tmp) {
932 d = node_remap(s, *from_nodes, *to_nodes);
933 if (s == d)
934 continue;
935
936 source = s; /* Node moved. Memorize */
937 dest = d;
938
939 /* dest not in remaining from nodes? */
940 if (!node_isset(dest, tmp))
941 break;
942 }
943 if (source == -1)
944 break;
945
946 node_clear(source, tmp);
947 err = migrate_to_node(mm, source, dest, flags);
948 if (err > 0)
949 busy += err;
950 if (err < 0)
951 break;
Christoph Lameter39743882006-01-08 01:00:51 -0800952 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700953out:
Christoph Lameter39743882006-01-08 01:00:51 -0800954 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800955 if (err < 0)
956 return err;
957 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800958
Christoph Lameter39743882006-01-08 01:00:51 -0800959}
960
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800961/*
962 * Allocate a new page for page migration based on vma policy.
963 * Start assuming that page is mapped by vma pointed to by @private.
964 * Search forward from there, if not. N.B., this assumes that the
965 * list of pages handed to migrate_pages()--which is how we get here--
966 * is in virtual address order.
967 */
Christoph Lameter742755a2006-06-23 02:03:55 -0700968static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700969{
970 struct vm_area_struct *vma = (struct vm_area_struct *)private;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800971 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700972
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800973 while (vma) {
974 address = page_address_in_vma(page, vma);
975 if (address != -EFAULT)
976 break;
977 vma = vma->vm_next;
978 }
979
980 /*
981 * if !vma, alloc_page_vma() will use task or system default policy
982 */
983 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700984}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800985#else
986
987static void migrate_page_add(struct page *page, struct list_head *pagelist,
988 unsigned long flags)
989{
990}
991
992int do_migrate_pages(struct mm_struct *mm,
993 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
994{
995 return -ENOSYS;
996}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700997
Keith Owens69939742006-10-11 01:21:28 -0700998static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700999{
1000 return NULL;
1001}
Christoph Lameterb20a3502006-03-22 00:09:12 -08001002#endif
1003
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001004static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001005 unsigned short mode, unsigned short mode_flags,
1006 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001007{
1008 struct vm_area_struct *vma;
1009 struct mm_struct *mm = current->mm;
1010 struct mempolicy *new;
1011 unsigned long end;
1012 int err;
1013 LIST_HEAD(pagelist);
1014
David Rientjesa3b51e02008-04-28 02:12:23 -07001015 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1016 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001017 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001018 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001019 return -EPERM;
1020
1021 if (start & ~PAGE_MASK)
1022 return -EINVAL;
1023
1024 if (mode == MPOL_DEFAULT)
1025 flags &= ~MPOL_MF_STRICT;
1026
1027 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1028 end = start + len;
1029
1030 if (end < start)
1031 return -EINVAL;
1032 if (end == start)
1033 return 0;
1034
David Rientjes028fec42008-04-28 02:12:25 -07001035 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001036 if (IS_ERR(new))
1037 return PTR_ERR(new);
1038
1039 /*
1040 * If we are using the default policy then operation
1041 * on discontinuous address spaces is okay after all
1042 */
1043 if (!new)
1044 flags |= MPOL_MF_DISCONTIG_OK;
1045
David Rientjes028fec42008-04-28 02:12:25 -07001046 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1047 start, start + len, mode, mode_flags,
1048 nmask ? nodes_addr(*nmask)[0] : -1);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001049
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001050 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1051
1052 err = migrate_prep();
1053 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001054 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001055 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001056 {
1057 NODEMASK_SCRATCH(scratch);
1058 if (scratch) {
1059 down_write(&mm->mmap_sem);
1060 task_lock(current);
1061 err = mpol_set_nodemask(new, nmask, scratch);
1062 task_unlock(current);
1063 if (err)
1064 up_write(&mm->mmap_sem);
1065 } else
1066 err = -ENOMEM;
1067 NODEMASK_SCRATCH_FREE(scratch);
1068 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001069 if (err)
1070 goto mpol_out;
1071
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001072 vma = check_range(mm, start, end, nmask,
1073 flags | MPOL_MF_INVERT, &pagelist);
1074
1075 err = PTR_ERR(vma);
1076 if (!IS_ERR(vma)) {
1077 int nr_failed = 0;
1078
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001079 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001080
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001081 if (!list_empty(&pagelist))
Christoph Lameter95a402c2006-06-23 02:03:53 -07001082 nr_failed = migrate_pages(&pagelist, new_vma_page,
Hugh Dickins62b61f62009-12-14 17:59:33 -08001083 (unsigned long)vma, 0);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001084
1085 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1086 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001087 } else
1088 putback_lru_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001089
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001090 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001091 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001092 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001093 return err;
1094}
1095
Christoph Lameter39743882006-01-08 01:00:51 -08001096/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001097 * User space interface with variable sized bitmaps for nodelists.
1098 */
1099
1100/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001101static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001102 unsigned long maxnode)
1103{
1104 unsigned long k;
1105 unsigned long nlongs;
1106 unsigned long endmask;
1107
1108 --maxnode;
1109 nodes_clear(*nodes);
1110 if (maxnode == 0 || !nmask)
1111 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001112 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001113 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001114
1115 nlongs = BITS_TO_LONGS(maxnode);
1116 if ((maxnode % BITS_PER_LONG) == 0)
1117 endmask = ~0UL;
1118 else
1119 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1120
1121 /* When the user specified more nodes than supported just check
1122 if the non supported part is all zero. */
1123 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1124 if (nlongs > PAGE_SIZE/sizeof(long))
1125 return -EINVAL;
1126 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1127 unsigned long t;
1128 if (get_user(t, nmask + k))
1129 return -EFAULT;
1130 if (k == nlongs - 1) {
1131 if (t & endmask)
1132 return -EINVAL;
1133 } else if (t)
1134 return -EINVAL;
1135 }
1136 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1137 endmask = ~0UL;
1138 }
1139
1140 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1141 return -EFAULT;
1142 nodes_addr(*nodes)[nlongs-1] &= endmask;
1143 return 0;
1144}
1145
1146/* Copy a kernel node mask to user space */
1147static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1148 nodemask_t *nodes)
1149{
1150 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1151 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1152
1153 if (copy > nbytes) {
1154 if (copy > PAGE_SIZE)
1155 return -EINVAL;
1156 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1157 return -EFAULT;
1158 copy = nbytes;
1159 }
1160 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1161}
1162
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001163SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1164 unsigned long, mode, unsigned long __user *, nmask,
1165 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001166{
1167 nodemask_t nodes;
1168 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001169 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001170
David Rientjes028fec42008-04-28 02:12:25 -07001171 mode_flags = mode & MPOL_MODE_FLAGS;
1172 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001173 if (mode >= MPOL_MAX)
1174 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001175 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1176 (mode_flags & MPOL_F_RELATIVE_NODES))
1177 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001178 err = get_nodes(&nodes, nmask, maxnode);
1179 if (err)
1180 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001181 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001182}
1183
1184/* Set the process memory policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001185SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1186 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001187{
1188 int err;
1189 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001190 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001191
David Rientjes028fec42008-04-28 02:12:25 -07001192 flags = mode & MPOL_MODE_FLAGS;
1193 mode &= ~MPOL_MODE_FLAGS;
1194 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001195 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001196 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1197 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001198 err = get_nodes(&nodes, nmask, maxnode);
1199 if (err)
1200 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001201 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001202}
1203
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001204SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1205 const unsigned long __user *, old_nodes,
1206 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001207{
David Howellsc69e8d92008-11-14 10:39:19 +11001208 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter39743882006-01-08 01:00:51 -08001209 struct mm_struct *mm;
1210 struct task_struct *task;
1211 nodemask_t old;
1212 nodemask_t new;
1213 nodemask_t task_nodes;
1214 int err;
1215
1216 err = get_nodes(&old, old_nodes, maxnode);
1217 if (err)
1218 return err;
1219
1220 err = get_nodes(&new, new_nodes, maxnode);
1221 if (err)
1222 return err;
1223
1224 /* Find the mm_struct */
1225 read_lock(&tasklist_lock);
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001226 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001227 if (!task) {
1228 read_unlock(&tasklist_lock);
1229 return -ESRCH;
1230 }
1231 mm = get_task_mm(task);
1232 read_unlock(&tasklist_lock);
1233
1234 if (!mm)
1235 return -EINVAL;
1236
1237 /*
1238 * Check if this process has the right to modify the specified
1239 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001240 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001241 * userid as the target process.
1242 */
David Howellsc69e8d92008-11-14 10:39:19 +11001243 rcu_read_lock();
1244 tcred = __task_cred(task);
David Howellsb6dff3e2008-11-14 10:39:16 +11001245 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1246 cred->uid != tcred->suid && cred->uid != tcred->uid &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001247 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001248 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001249 err = -EPERM;
1250 goto out;
1251 }
David Howellsc69e8d92008-11-14 10:39:19 +11001252 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001253
1254 task_nodes = cpuset_mems_allowed(task);
1255 /* Is the user allowed to access the target nodes? */
Christoph Lameter74c00242006-03-14 19:50:21 -08001256 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001257 err = -EPERM;
1258 goto out;
1259 }
1260
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001261 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001262 err = -EINVAL;
1263 goto out;
1264 }
1265
David Quigley86c3a762006-06-23 02:04:02 -07001266 err = security_task_movememory(task);
1267 if (err)
1268 goto out;
1269
Christoph Lameter511030b2006-02-28 16:58:57 -08001270 err = do_migrate_pages(mm, &old, &new,
Christoph Lameter74c00242006-03-14 19:50:21 -08001271 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter39743882006-01-08 01:00:51 -08001272out:
1273 mmput(mm);
1274 return err;
1275}
1276
1277
Christoph Lameter8bccd852005-10-29 18:16:59 -07001278/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001279SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1280 unsigned long __user *, nmask, unsigned long, maxnode,
1281 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001282{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001283 int err;
1284 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001285 nodemask_t nodes;
1286
1287 if (nmask != NULL && maxnode < MAX_NUMNODES)
1288 return -EINVAL;
1289
1290 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1291
1292 if (err)
1293 return err;
1294
1295 if (policy && put_user(pval, policy))
1296 return -EFAULT;
1297
1298 if (nmask)
1299 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1300
1301 return err;
1302}
1303
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304#ifdef CONFIG_COMPAT
1305
1306asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1307 compat_ulong_t __user *nmask,
1308 compat_ulong_t maxnode,
1309 compat_ulong_t addr, compat_ulong_t flags)
1310{
1311 long err;
1312 unsigned long __user *nm = NULL;
1313 unsigned long nr_bits, alloc_size;
1314 DECLARE_BITMAP(bm, MAX_NUMNODES);
1315
1316 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1317 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1318
1319 if (nmask)
1320 nm = compat_alloc_user_space(alloc_size);
1321
1322 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1323
1324 if (!err && nmask) {
1325 err = copy_from_user(bm, nm, alloc_size);
1326 /* ensure entire bitmap is zeroed */
1327 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1328 err |= compat_put_bitmap(nmask, bm, nr_bits);
1329 }
1330
1331 return err;
1332}
1333
1334asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1335 compat_ulong_t maxnode)
1336{
1337 long err = 0;
1338 unsigned long __user *nm = NULL;
1339 unsigned long nr_bits, alloc_size;
1340 DECLARE_BITMAP(bm, MAX_NUMNODES);
1341
1342 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1343 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1344
1345 if (nmask) {
1346 err = compat_get_bitmap(bm, nmask, nr_bits);
1347 nm = compat_alloc_user_space(alloc_size);
1348 err |= copy_to_user(nm, bm, alloc_size);
1349 }
1350
1351 if (err)
1352 return -EFAULT;
1353
1354 return sys_set_mempolicy(mode, nm, nr_bits+1);
1355}
1356
1357asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1358 compat_ulong_t mode, compat_ulong_t __user *nmask,
1359 compat_ulong_t maxnode, compat_ulong_t flags)
1360{
1361 long err = 0;
1362 unsigned long __user *nm = NULL;
1363 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001364 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365
1366 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1367 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1368
1369 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001370 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001372 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 }
1374
1375 if (err)
1376 return -EFAULT;
1377
1378 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1379}
1380
1381#endif
1382
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001383/*
1384 * get_vma_policy(@task, @vma, @addr)
1385 * @task - task for fallback if vma policy == default
1386 * @vma - virtual memory area whose policy is sought
1387 * @addr - address in @vma for shared policy lookup
1388 *
1389 * Returns effective policy for a VMA at specified address.
1390 * Falls back to @task or system default policy, as necessary.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001391 * Current or other task's task mempolicy and non-shared vma policies
1392 * are protected by the task's mmap_sem, which must be held for read by
1393 * the caller.
1394 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1395 * count--added by the get_policy() vm_op, as appropriate--to protect against
1396 * freeing by another task. It is the caller's responsibility to free the
1397 * extra reference for shared policies.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001398 */
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001399static struct mempolicy *get_vma_policy(struct task_struct *task,
Christoph Lameter48fce342006-01-08 01:01:03 -08001400 struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001402 struct mempolicy *pol = task->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403
1404 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001405 if (vma->vm_ops && vma->vm_ops->get_policy) {
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001406 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1407 addr);
1408 if (vpol)
1409 pol = vpol;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001410 } else if (vma->vm_policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 pol = vma->vm_policy;
1412 }
1413 if (!pol)
1414 pol = &default_policy;
1415 return pol;
1416}
1417
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001418/*
1419 * Return a nodemask representing a mempolicy for filtering nodes for
1420 * page allocation
1421 */
1422static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001423{
1424 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001425 if (unlikely(policy->mode == MPOL_BIND) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001426 gfp_zone(gfp) >= policy_zone &&
1427 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1428 return &policy->v.nodes;
1429
1430 return NULL;
1431}
1432
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001433/* Return a zonelist indicated by gfp for node representing a mempolicy */
1434static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435{
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001436 int nd = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001438 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001440 if (!(policy->flags & MPOL_F_LOCAL))
1441 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 break;
1443 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001444 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001445 * Normally, MPOL_BIND allocations are node-local within the
1446 * allowed nodemask. However, if __GFP_THISNODE is set and the
1447 * current node is part of the mask, we use the zonelist for
1448 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001449 */
Mel Gorman19770b32008-04-28 02:12:18 -07001450 if (unlikely(gfp & __GFP_THISNODE) &&
1451 unlikely(!node_isset(nd, policy->v.nodes)))
1452 nd = first_node(policy->v.nodes);
1453 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 case MPOL_INTERLEAVE: /* should not happen */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455 break;
1456 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 BUG();
1458 }
Mel Gorman0e884602008-04-28 02:12:14 -07001459 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460}
1461
1462/* Do dynamic interleaving for a process */
1463static unsigned interleave_nodes(struct mempolicy *policy)
1464{
1465 unsigned nid, next;
1466 struct task_struct *me = current;
1467
1468 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001469 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001471 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001472 if (next < MAX_NUMNODES)
1473 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 return nid;
1475}
1476
Christoph Lameterdc85da12006-01-18 17:42:36 -08001477/*
1478 * Depending on the memory policy provide a node from which to allocate the
1479 * next slab entry.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001480 * @policy must be protected by freeing by the caller. If @policy is
1481 * the current task's mempolicy, this protection is implicit, as only the
1482 * task can change it's policy. The system default policy requires no
1483 * such protection.
Christoph Lameterdc85da12006-01-18 17:42:36 -08001484 */
1485unsigned slab_node(struct mempolicy *policy)
1486{
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001487 if (!policy || policy->flags & MPOL_F_LOCAL)
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001488 return numa_node_id();
Christoph Lameter765c4502006-09-27 01:50:08 -07001489
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001490 switch (policy->mode) {
1491 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001492 /*
1493 * handled MPOL_F_LOCAL above
1494 */
1495 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001496
Christoph Lameterdc85da12006-01-18 17:42:36 -08001497 case MPOL_INTERLEAVE:
1498 return interleave_nodes(policy);
1499
Mel Gormandd1a2392008-04-28 02:12:17 -07001500 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001501 /*
1502 * Follow bind policy behavior and start allocation at the
1503 * first node.
1504 */
Mel Gorman19770b32008-04-28 02:12:18 -07001505 struct zonelist *zonelist;
1506 struct zone *zone;
1507 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1508 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1509 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1510 &policy->v.nodes,
1511 &zone);
1512 return zone->node;
Mel Gormandd1a2392008-04-28 02:12:17 -07001513 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001514
Christoph Lameterdc85da12006-01-18 17:42:36 -08001515 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001516 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001517 }
1518}
1519
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520/* Do static interleaving for a VMA with known offset. */
1521static unsigned offset_il_node(struct mempolicy *pol,
1522 struct vm_area_struct *vma, unsigned long off)
1523{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001524 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001525 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 int c;
1527 int nid = -1;
1528
David Rientjesf5b087b2008-04-28 02:12:27 -07001529 if (!nnodes)
1530 return numa_node_id();
1531 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 c = 0;
1533 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001534 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535 c++;
1536 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 return nid;
1538}
1539
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001540/* Determine a node number for interleave */
1541static inline unsigned interleave_nid(struct mempolicy *pol,
1542 struct vm_area_struct *vma, unsigned long addr, int shift)
1543{
1544 if (vma) {
1545 unsigned long off;
1546
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001547 /*
1548 * for small pages, there is no difference between
1549 * shift and PAGE_SHIFT, so the bit-shift is safe.
1550 * for huge pages, since vm_pgoff is in units of small
1551 * pages, we need to shift off the always 0 bits to get
1552 * a useful offset.
1553 */
1554 BUG_ON(shift < PAGE_SHIFT);
1555 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001556 off += (addr - vma->vm_start) >> shift;
1557 return offset_il_node(pol, vma, off);
1558 } else
1559 return interleave_nodes(pol);
1560}
1561
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001562#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001563/*
1564 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1565 * @vma = virtual memory area whose policy is sought
1566 * @addr = address in @vma for shared policy lookup and interleave policy
1567 * @gfp_flags = for requested zone
Mel Gorman19770b32008-04-28 02:12:18 -07001568 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1569 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001570 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001571 * Returns a zonelist suitable for a huge page allocation and a pointer
1572 * to the struct mempolicy for conditional unref after allocation.
1573 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1574 * @nodemask for filtering the zonelist.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001575 */
Mel Gorman396faf02007-07-17 04:03:13 -07001576struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001577 gfp_t gfp_flags, struct mempolicy **mpol,
1578 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001579{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001580 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001581
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001582 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001583 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001584
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001585 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1586 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001587 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001588 } else {
1589 zl = policy_zonelist(gfp_flags, *mpol);
1590 if ((*mpol)->mode == MPOL_BIND)
1591 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001592 }
1593 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001594}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001595
1596/*
1597 * init_nodemask_of_mempolicy
1598 *
1599 * If the current task's mempolicy is "default" [NULL], return 'false'
1600 * to indicate default policy. Otherwise, extract the policy nodemask
1601 * for 'bind' or 'interleave' policy into the argument nodemask, or
1602 * initialize the argument nodemask to contain the single node for
1603 * 'preferred' or 'local' policy and return 'true' to indicate presence
1604 * of non-default mempolicy.
1605 *
1606 * We don't bother with reference counting the mempolicy [mpol_get/put]
1607 * because the current task is examining it's own mempolicy and a task's
1608 * mempolicy is only ever changed by the task itself.
1609 *
1610 * N.B., it is the caller's responsibility to free a returned nodemask.
1611 */
1612bool init_nodemask_of_mempolicy(nodemask_t *mask)
1613{
1614 struct mempolicy *mempolicy;
1615 int nid;
1616
1617 if (!(mask && current->mempolicy))
1618 return false;
1619
1620 mempolicy = current->mempolicy;
1621 switch (mempolicy->mode) {
1622 case MPOL_PREFERRED:
1623 if (mempolicy->flags & MPOL_F_LOCAL)
1624 nid = numa_node_id();
1625 else
1626 nid = mempolicy->v.preferred_node;
1627 init_nodemask_of_node(mask, nid);
1628 break;
1629
1630 case MPOL_BIND:
1631 /* Fall through */
1632 case MPOL_INTERLEAVE:
1633 *mask = mempolicy->v.nodes;
1634 break;
1635
1636 default:
1637 BUG();
1638 }
1639
1640 return true;
1641}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001642#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001643
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644/* Allocate a page in interleaved policy.
1645 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001646static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1647 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648{
1649 struct zonelist *zl;
1650 struct page *page;
1651
Mel Gorman0e884602008-04-28 02:12:14 -07001652 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001654 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001655 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 return page;
1657}
1658
1659/**
1660 * alloc_page_vma - Allocate a page for a VMA.
1661 *
1662 * @gfp:
1663 * %GFP_USER user allocation.
1664 * %GFP_KERNEL kernel allocations,
1665 * %GFP_HIGHMEM highmem/user allocations,
1666 * %GFP_FS allocation should not call back into a file system.
1667 * %GFP_ATOMIC don't sleep.
1668 *
1669 * @vma: Pointer to VMA or NULL if not available.
1670 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1671 *
1672 * This function allocates a page from the kernel page pool and applies
1673 * a NUMA policy associated with the VMA or the current process.
1674 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1675 * mm_struct of the VMA to prevent it from going away. Should be used for
1676 * all allocations for pages that will be mapped into
1677 * user space. Returns NULL when no page can be allocated.
1678 *
1679 * Should be called with the mm_sem of the vma hold.
1680 */
1681struct page *
Al Virodd0fc662005-10-07 07:46:04 +01001682alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001684 struct mempolicy *pol = get_vma_policy(current, vma, addr);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001685 struct zonelist *zl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001687 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001689
1690 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001691 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 return alloc_page_interleave(gfp, 0, nid);
1693 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001694 zl = policy_zonelist(gfp, pol);
1695 if (unlikely(mpol_needs_cond_ref(pol))) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001696 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001697 * slow path: ref counted shared policy
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001698 */
Mel Gorman19770b32008-04-28 02:12:18 -07001699 struct page *page = __alloc_pages_nodemask(gfp, 0,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001700 zl, policy_nodemask(gfp, pol));
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001701 __mpol_put(pol);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001702 return page;
1703 }
1704 /*
1705 * fast path: default or task policy
1706 */
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001707 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708}
1709
1710/**
1711 * alloc_pages_current - Allocate pages.
1712 *
1713 * @gfp:
1714 * %GFP_USER user allocation,
1715 * %GFP_KERNEL kernel allocation,
1716 * %GFP_HIGHMEM highmem allocation,
1717 * %GFP_FS don't call back into a file system.
1718 * %GFP_ATOMIC don't sleep.
1719 * @order: Power of two of allocation size in pages. 0 is a single page.
1720 *
1721 * Allocate a page from the kernel page pool. When not in
1722 * interrupt context and apply the current process NUMA policy.
1723 * Returns NULL when no page can be allocated.
1724 *
Paul Jacksoncf2a4732006-01-08 01:01:54 -08001725 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 * 1) it's ok to take cpuset_sem (can WAIT), and
1727 * 2) allocating for current task (not interrupt).
1728 */
Al Virodd0fc662005-10-07 07:46:04 +01001729struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730{
1731 struct mempolicy *pol = current->mempolicy;
1732
Christoph Lameter9b819d22006-09-25 23:31:40 -07001733 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 pol = &default_policy;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001735
1736 /*
1737 * No reference counting needed for current->mempolicy
1738 * nor system default_policy
1739 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001740 if (pol->mode == MPOL_INTERLEAVE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
Mel Gorman19770b32008-04-28 02:12:18 -07001742 return __alloc_pages_nodemask(gfp, order,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001743 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744}
1745EXPORT_SYMBOL(alloc_pages_current);
1746
Paul Jackson42253992006-01-08 01:01:59 -08001747/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001748 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08001749 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1750 * with the mems_allowed returned by cpuset_mems_allowed(). This
1751 * keeps mempolicies cpuset relative after its cpuset moves. See
1752 * further kernel/cpuset.c update_nodemask().
1753 */
Paul Jackson42253992006-01-08 01:01:59 -08001754
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001755/* Slow path of a mempolicy duplicate */
1756struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757{
1758 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1759
1760 if (!new)
1761 return ERR_PTR(-ENOMEM);
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08001762 rcu_read_lock();
Paul Jackson42253992006-01-08 01:01:59 -08001763 if (current_cpuset_is_being_rebound()) {
1764 nodemask_t mems = cpuset_mems_allowed(current);
1765 mpol_rebind_policy(old, &mems);
1766 }
Paul E. McKenney99ee4ca2010-03-03 17:50:17 -08001767 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 *new = *old;
1769 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 return new;
1771}
1772
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001773/*
1774 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1775 * eliminate the * MPOL_F_* flags that require conditional ref and
1776 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1777 * after return. Use the returned value.
1778 *
1779 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1780 * policy lookup, even if the policy needs/has extra ref on lookup.
1781 * shmem_readahead needs this.
1782 */
1783struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1784 struct mempolicy *frompol)
1785{
1786 if (!mpol_needs_cond_ref(frompol))
1787 return frompol;
1788
1789 *tompol = *frompol;
1790 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1791 __mpol_put(frompol);
1792 return tompol;
1793}
1794
David Rientjesf5b087b2008-04-28 02:12:27 -07001795static int mpol_match_intent(const struct mempolicy *a,
1796 const struct mempolicy *b)
1797{
1798 if (a->flags != b->flags)
1799 return 0;
1800 if (!mpol_store_user_nodemask(a))
1801 return 1;
1802 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1803}
1804
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805/* Slow path of a mempolicy comparison */
1806int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1807{
1808 if (!a || !b)
1809 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001810 if (a->mode != b->mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001812 if (a->mode != MPOL_DEFAULT && !mpol_match_intent(a, b))
David Rientjesf5b087b2008-04-28 02:12:27 -07001813 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001814 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07001815 case MPOL_BIND:
1816 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -07001818 return nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001820 return a->v.preferred_node == b->v.preferred_node &&
1821 a->flags == b->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 default:
1823 BUG();
1824 return 0;
1825 }
1826}
1827
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 * Shared memory backing store policy support.
1830 *
1831 * Remember policies even when nobody has shared memory mapped.
1832 * The policies are kept in Red-Black tree linked from the inode.
1833 * They are protected by the sp->lock spinlock, which should be held
1834 * for any accesses to the tree.
1835 */
1836
1837/* lookup first element intersecting start-end */
1838/* Caller holds sp->lock */
1839static struct sp_node *
1840sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1841{
1842 struct rb_node *n = sp->root.rb_node;
1843
1844 while (n) {
1845 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1846
1847 if (start >= p->end)
1848 n = n->rb_right;
1849 else if (end <= p->start)
1850 n = n->rb_left;
1851 else
1852 break;
1853 }
1854 if (!n)
1855 return NULL;
1856 for (;;) {
1857 struct sp_node *w = NULL;
1858 struct rb_node *prev = rb_prev(n);
1859 if (!prev)
1860 break;
1861 w = rb_entry(prev, struct sp_node, nd);
1862 if (w->end <= start)
1863 break;
1864 n = prev;
1865 }
1866 return rb_entry(n, struct sp_node, nd);
1867}
1868
1869/* Insert a new shared policy into the list. */
1870/* Caller holds sp->lock */
1871static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1872{
1873 struct rb_node **p = &sp->root.rb_node;
1874 struct rb_node *parent = NULL;
1875 struct sp_node *nd;
1876
1877 while (*p) {
1878 parent = *p;
1879 nd = rb_entry(parent, struct sp_node, nd);
1880 if (new->start < nd->start)
1881 p = &(*p)->rb_left;
1882 else if (new->end > nd->end)
1883 p = &(*p)->rb_right;
1884 else
1885 BUG();
1886 }
1887 rb_link_node(&new->nd, parent, p);
1888 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07001889 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001890 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891}
1892
1893/* Find shared policy intersecting idx */
1894struct mempolicy *
1895mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1896{
1897 struct mempolicy *pol = NULL;
1898 struct sp_node *sn;
1899
1900 if (!sp->root.rb_node)
1901 return NULL;
1902 spin_lock(&sp->lock);
1903 sn = sp_lookup(sp, idx, idx+1);
1904 if (sn) {
1905 mpol_get(sn->policy);
1906 pol = sn->policy;
1907 }
1908 spin_unlock(&sp->lock);
1909 return pol;
1910}
1911
1912static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1913{
Paul Mundt140d5a42007-07-15 23:38:16 -07001914 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 rb_erase(&n->nd, &sp->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001916 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 kmem_cache_free(sn_cache, n);
1918}
1919
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001920static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1921 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922{
1923 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1924
1925 if (!n)
1926 return NULL;
1927 n->start = start;
1928 n->end = end;
1929 mpol_get(pol);
Lee Schermerhornaab0b102008-04-28 02:13:13 -07001930 pol->flags |= MPOL_F_SHARED; /* for unref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931 n->policy = pol;
1932 return n;
1933}
1934
1935/* Replace a policy range. */
1936static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1937 unsigned long end, struct sp_node *new)
1938{
1939 struct sp_node *n, *new2 = NULL;
1940
1941restart:
1942 spin_lock(&sp->lock);
1943 n = sp_lookup(sp, start, end);
1944 /* Take care of old policies in the same range. */
1945 while (n && n->start < end) {
1946 struct rb_node *next = rb_next(&n->nd);
1947 if (n->start >= start) {
1948 if (n->end <= end)
1949 sp_delete(sp, n);
1950 else
1951 n->start = end;
1952 } else {
1953 /* Old policy spanning whole new range. */
1954 if (n->end > end) {
1955 if (!new2) {
1956 spin_unlock(&sp->lock);
1957 new2 = sp_alloc(end, n->end, n->policy);
1958 if (!new2)
1959 return -ENOMEM;
1960 goto restart;
1961 }
1962 n->end = start;
1963 sp_insert(sp, new2);
1964 new2 = NULL;
1965 break;
1966 } else
1967 n->end = start;
1968 }
1969 if (!next)
1970 break;
1971 n = rb_entry(next, struct sp_node, nd);
1972 }
1973 if (new)
1974 sp_insert(sp, new);
1975 spin_unlock(&sp->lock);
1976 if (new2) {
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001977 mpol_put(new2->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978 kmem_cache_free(sn_cache, new2);
1979 }
1980 return 0;
1981}
1982
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07001983/**
1984 * mpol_shared_policy_init - initialize shared policy for inode
1985 * @sp: pointer to inode shared policy
1986 * @mpol: struct mempolicy to install
1987 *
1988 * Install non-NULL @mpol in inode's shared policy rb-tree.
1989 * On entry, the current task has a reference on a non-NULL @mpol.
1990 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001991 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07001992 */
1993void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08001994{
Miao Xie58568d22009-06-16 15:31:49 -07001995 int ret;
1996
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07001997 sp->root = RB_ROOT; /* empty tree == default mempolicy */
1998 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08001999
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002000 if (mpol) {
2001 struct vm_area_struct pvma;
2002 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002003 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002004
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002005 if (!scratch)
2006 return;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002007 /* contextualize the tmpfs mount point mempolicy */
2008 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Miao Xie58568d22009-06-16 15:31:49 -07002009 if (IS_ERR(new)) {
2010 mpol_put(mpol); /* drop our ref on sb mpol */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002011 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002012 return; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002013 }
2014
2015 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002016 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002017 task_unlock(current);
2018 mpol_put(mpol); /* drop our ref on sb mpol */
2019 if (ret) {
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002020 NODEMASK_SCRATCH_FREE(scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002021 mpol_put(new);
2022 return;
2023 }
Robin Holt7339ff82006-01-14 13:20:48 -08002024
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002025 /* Create pseudo-vma that contains just the policy */
2026 memset(&pvma, 0, sizeof(struct vm_area_struct));
2027 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2028 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2029 mpol_put(new); /* drop initial ref */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002030 NODEMASK_SCRATCH_FREE(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002031 }
2032}
2033
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034int mpol_set_shared_policy(struct shared_policy *info,
2035 struct vm_area_struct *vma, struct mempolicy *npol)
2036{
2037 int err;
2038 struct sp_node *new = NULL;
2039 unsigned long sz = vma_pages(vma);
2040
David Rientjes028fec42008-04-28 02:12:25 -07002041 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002043 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002044 npol ? npol->flags : -1,
Paul Mundt140d5a42007-07-15 23:38:16 -07002045 npol ? nodes_addr(npol->v.nodes)[0] : -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046
2047 if (npol) {
2048 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2049 if (!new)
2050 return -ENOMEM;
2051 }
2052 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2053 if (err && new)
2054 kmem_cache_free(sn_cache, new);
2055 return err;
2056}
2057
2058/* Free a backing policy store on inode delete. */
2059void mpol_free_shared_policy(struct shared_policy *p)
2060{
2061 struct sp_node *n;
2062 struct rb_node *next;
2063
2064 if (!p->root.rb_node)
2065 return;
2066 spin_lock(&p->lock);
2067 next = rb_first(&p->root);
2068 while (next) {
2069 n = rb_entry(next, struct sp_node, nd);
2070 next = rb_next(&n->nd);
Andi Kleen90c50292005-07-27 11:43:50 -07002071 rb_erase(&n->nd, &p->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002072 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 kmem_cache_free(sn_cache, n);
2074 }
2075 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076}
2077
2078/* assumes fs == KERNEL_DS */
2079void __init numa_policy_init(void)
2080{
Paul Mundtb71636e2007-07-15 23:38:15 -07002081 nodemask_t interleave_nodes;
2082 unsigned long largest = 0;
2083 int nid, prefer = 0;
2084
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 policy_cache = kmem_cache_create("numa_policy",
2086 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002087 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088
2089 sn_cache = kmem_cache_create("shared_policy_node",
2090 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002091 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092
Paul Mundtb71636e2007-07-15 23:38:15 -07002093 /*
2094 * Set interleaving policy for system init. Interleaving is only
2095 * enabled across suitably sized nodes (default is >= 16MB), or
2096 * fall back to the largest node if they're all smaller.
2097 */
2098 nodes_clear(interleave_nodes);
Christoph Lameter56bbd652007-10-16 01:25:35 -07002099 for_each_node_state(nid, N_HIGH_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002100 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101
Paul Mundtb71636e2007-07-15 23:38:15 -07002102 /* Preserve the largest node */
2103 if (largest < total_pages) {
2104 largest = total_pages;
2105 prefer = nid;
2106 }
2107
2108 /* Interleave this node? */
2109 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2110 node_set(nid, interleave_nodes);
2111 }
2112
2113 /* All too small, use the largest */
2114 if (unlikely(nodes_empty(interleave_nodes)))
2115 node_set(prefer, interleave_nodes);
2116
David Rientjes028fec42008-04-28 02:12:25 -07002117 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118 printk("numa_policy_init: interleaving failed\n");
2119}
2120
Christoph Lameter8bccd852005-10-29 18:16:59 -07002121/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122void numa_default_policy(void)
2123{
David Rientjes028fec42008-04-28 02:12:25 -07002124 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125}
Paul Jackson68860ec2005-10-30 15:02:36 -08002126
Paul Jackson42253992006-01-08 01:01:59 -08002127/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002128 * Parse and format mempolicy from/to strings
2129 */
2130
2131/*
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002132 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002133 * Used only for mpol_parse_str() and mpol_to_str()
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002134 */
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002135#define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
Helge Deller15ad7cd2006-12-06 20:40:36 -08002136static const char * const policy_types[] =
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002137 { "default", "prefer", "bind", "interleave", "local" };
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002138
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002139
2140#ifdef CONFIG_TMPFS
2141/**
2142 * mpol_parse_str - parse string to mempolicy
2143 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002144 * @mpol: pointer to struct mempolicy pointer, returned on success.
2145 * @no_context: flag whether to "contextualize" the mempolicy
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002146 *
2147 * Format of input:
2148 * <mode>[=<flags>][:<nodelist>]
2149 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002150 * if @no_context is true, save the input nodemask in w.user_nodemask in
2151 * the returned mempolicy. This will be used to "clone" the mempolicy in
2152 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2153 * mount option. Note that if 'static' or 'relative' mode flags were
2154 * specified, the input nodemask will already have been saved. Saving
2155 * it again is redundant, but safe.
2156 *
2157 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002158 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002159int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002160{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002161 struct mempolicy *new = NULL;
2162 unsigned short uninitialized_var(mode);
2163 unsigned short uninitialized_var(mode_flags);
2164 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002165 char *nodelist = strchr(str, ':');
2166 char *flags = strchr(str, '=');
2167 int i;
2168 int err = 1;
2169
2170 if (nodelist) {
2171 /* NUL-terminate mode or flags string */
2172 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002173 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002174 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002175 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002176 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002177 } else
2178 nodes_clear(nodes);
2179
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002180 if (flags)
2181 *flags++ = '\0'; /* terminate mode string */
2182
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002183 for (i = 0; i <= MPOL_LOCAL; i++) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002184 if (!strcmp(str, policy_types[i])) {
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002185 mode = i;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002186 break;
2187 }
2188 }
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002189 if (i > MPOL_LOCAL)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002190 goto out;
2191
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002192 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002193 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002194 /*
2195 * Insist on a nodelist of one node only
2196 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002197 if (nodelist) {
2198 char *rest = nodelist;
2199 while (isdigit(*rest))
2200 rest++;
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002201 if (*rest)
2202 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002203 }
2204 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002205 case MPOL_INTERLEAVE:
2206 /*
2207 * Default to online nodes with memory if no nodelist
2208 */
2209 if (!nodelist)
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002210 nodes = node_states[N_HIGH_MEMORY];
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002211 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002212 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002213 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002214 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002215 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002216 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002217 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002218 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002219 break;
Ravikiran G Thirumalai413b43d2010-03-23 13:35:28 -07002220 case MPOL_DEFAULT:
2221 /*
2222 * Insist on a empty nodelist
2223 */
2224 if (!nodelist)
2225 err = 0;
2226 goto out;
KOSAKI Motohirod69b2e62010-03-23 13:35:30 -07002227 case MPOL_BIND:
2228 /*
2229 * Insist on a nodelist
2230 */
2231 if (!nodelist)
2232 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002233 }
2234
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002235 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002236 if (flags) {
2237 /*
2238 * Currently, we only support two mutually exclusive
2239 * mode flags.
2240 */
2241 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002242 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002243 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002244 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002245 else
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002246 goto out;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002247 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002248
2249 new = mpol_new(mode, mode_flags, &nodes);
2250 if (IS_ERR(new))
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002251 goto out;
2252
2253 {
Miao Xie58568d22009-06-16 15:31:49 -07002254 int ret;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002255 NODEMASK_SCRATCH(scratch);
2256 if (scratch) {
2257 task_lock(current);
2258 ret = mpol_set_nodemask(new, &nodes, scratch);
2259 task_unlock(current);
2260 } else
2261 ret = -ENOMEM;
2262 NODEMASK_SCRATCH_FREE(scratch);
2263 if (ret) {
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002264 mpol_put(new);
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002265 goto out;
Miao Xie58568d22009-06-16 15:31:49 -07002266 }
2267 }
KOSAKI Motohiro926f2ae2010-03-23 13:35:32 -07002268 err = 0;
2269 if (no_context) {
2270 /* save for contextualization */
2271 new->w.user_nodemask = nodes;
2272 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002273
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002274out:
2275 /* Restore string for error message */
2276 if (nodelist)
2277 *--nodelist = ':';
2278 if (flags)
2279 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002280 if (!err)
2281 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002282 return err;
2283}
2284#endif /* CONFIG_TMPFS */
2285
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002286/**
2287 * mpol_to_str - format a mempolicy structure for printing
2288 * @buffer: to contain formatted mempolicy string
2289 * @maxlen: length of @buffer
2290 * @pol: pointer to mempolicy to be formatted
2291 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2292 *
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002293 * Convert a mempolicy into a string.
2294 * Returns the number of characters in buffer (if positive)
2295 * or an error (negative)
2296 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002297int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002298{
2299 char *p = buffer;
2300 int l;
2301 nodemask_t nodes;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002302 unsigned short mode;
David Rientjesf5b087b2008-04-28 02:12:27 -07002303 unsigned short flags = pol ? pol->flags : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002304
Lee Schermerhorn22919902008-04-28 02:13:22 -07002305 /*
2306 * Sanity check: room for longest mode, flag and some nodes
2307 */
2308 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2309
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002310 if (!pol || pol == &default_policy)
2311 mode = MPOL_DEFAULT;
2312 else
2313 mode = pol->mode;
2314
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002315 switch (mode) {
2316 case MPOL_DEFAULT:
2317 nodes_clear(nodes);
2318 break;
2319
2320 case MPOL_PREFERRED:
2321 nodes_clear(nodes);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002322 if (flags & MPOL_F_LOCAL)
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002323 mode = MPOL_LOCAL; /* pseudo-policy */
2324 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002325 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002326 break;
2327
2328 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07002329 /* Fall through */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002330 case MPOL_INTERLEAVE:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002331 if (no_context)
2332 nodes = pol->w.user_nodemask;
2333 else
2334 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002335 break;
2336
2337 default:
2338 BUG();
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002339 }
2340
2341 l = strlen(policy_types[mode]);
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002342 if (buffer + maxlen < p + l + 1)
2343 return -ENOSPC;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002344
2345 strcpy(p, policy_types[mode]);
2346 p += l;
2347
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002348 if (flags & MPOL_MODE_FLAGS) {
David Rientjesf5b087b2008-04-28 02:12:27 -07002349 if (buffer + maxlen < p + 2)
2350 return -ENOSPC;
2351 *p++ = '=';
2352
Lee Schermerhorn22919902008-04-28 02:13:22 -07002353 /*
2354 * Currently, the only defined flags are mutually exclusive
2355 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002356 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002357 p += snprintf(p, buffer + maxlen - p, "static");
2358 else if (flags & MPOL_F_RELATIVE_NODES)
2359 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002360 }
2361
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002362 if (!nodes_empty(nodes)) {
2363 if (buffer + maxlen < p + 2)
2364 return -ENOSPC;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002365 *p++ = ':';
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002366 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2367 }
2368 return p - buffer;
2369}
2370
2371struct numa_maps {
2372 unsigned long pages;
2373 unsigned long anon;
Christoph Lameter397874d2006-03-06 15:42:53 -08002374 unsigned long active;
2375 unsigned long writeback;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002376 unsigned long mapcount_max;
Christoph Lameter397874d2006-03-06 15:42:53 -08002377 unsigned long dirty;
2378 unsigned long swapcache;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002379 unsigned long node[MAX_NUMNODES];
2380};
2381
Christoph Lameter397874d2006-03-06 15:42:53 -08002382static void gather_stats(struct page *page, void *private, int pte_dirty)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002383{
2384 struct numa_maps *md = private;
2385 int count = page_mapcount(page);
2386
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002387 md->pages++;
Christoph Lameter397874d2006-03-06 15:42:53 -08002388 if (pte_dirty || PageDirty(page))
2389 md->dirty++;
2390
2391 if (PageSwapCache(page))
2392 md->swapcache++;
2393
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002394 if (PageActive(page) || PageUnevictable(page))
Christoph Lameter397874d2006-03-06 15:42:53 -08002395 md->active++;
2396
2397 if (PageWriteback(page))
2398 md->writeback++;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002399
2400 if (PageAnon(page))
2401 md->anon++;
2402
Christoph Lameter397874d2006-03-06 15:42:53 -08002403 if (count > md->mapcount_max)
2404 md->mapcount_max = count;
2405
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002406 md->node[page_to_nid(page)]++;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002407}
2408
Andrew Morton7f709ed2006-03-07 21:55:22 -08002409#ifdef CONFIG_HUGETLB_PAGE
Christoph Lameter397874d2006-03-06 15:42:53 -08002410static void check_huge_range(struct vm_area_struct *vma,
2411 unsigned long start, unsigned long end,
2412 struct numa_maps *md)
2413{
2414 unsigned long addr;
2415 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002416 struct hstate *h = hstate_vma(vma);
2417 unsigned long sz = huge_page_size(h);
Christoph Lameter397874d2006-03-06 15:42:53 -08002418
Andi Kleena5516432008-07-23 21:27:41 -07002419 for (addr = start; addr < end; addr += sz) {
2420 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2421 addr & huge_page_mask(h));
Christoph Lameter397874d2006-03-06 15:42:53 -08002422 pte_t pte;
2423
2424 if (!ptep)
2425 continue;
2426
2427 pte = *ptep;
2428 if (pte_none(pte))
2429 continue;
2430
2431 page = pte_page(pte);
2432 if (!page)
2433 continue;
2434
2435 gather_stats(page, md, pte_dirty(*ptep));
2436 }
2437}
Andrew Morton7f709ed2006-03-07 21:55:22 -08002438#else
2439static inline void check_huge_range(struct vm_area_struct *vma,
2440 unsigned long start, unsigned long end,
2441 struct numa_maps *md)
2442{
2443}
2444#endif
Christoph Lameter397874d2006-03-06 15:42:53 -08002445
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002446/*
2447 * Display pages allocated per node and memory policy via /proc.
2448 */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002449int show_numa_map(struct seq_file *m, void *v)
2450{
Eric W. Biederman99f89552006-06-26 00:25:55 -07002451 struct proc_maps_private *priv = m->private;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002452 struct vm_area_struct *vma = v;
2453 struct numa_maps *md;
Christoph Lameter397874d2006-03-06 15:42:53 -08002454 struct file *file = vma->vm_file;
2455 struct mm_struct *mm = vma->vm_mm;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07002456 struct mempolicy *pol;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002457 int n;
2458 char buffer[50];
2459
Christoph Lameter397874d2006-03-06 15:42:53 -08002460 if (!mm)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002461 return 0;
2462
2463 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2464 if (!md)
2465 return 0;
2466
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07002467 pol = get_vma_policy(priv->task, vma, vma->vm_start);
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002468 mpol_to_str(buffer, sizeof(buffer), pol, 0);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002469 mpol_cond_put(pol);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002470
Christoph Lameter397874d2006-03-06 15:42:53 -08002471 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002472
Christoph Lameter397874d2006-03-06 15:42:53 -08002473 if (file) {
2474 seq_printf(m, " file=");
Jan Blunckc32c2f62008-02-14 19:38:43 -08002475 seq_path(m, &file->f_path, "\n\t= ");
Christoph Lameter397874d2006-03-06 15:42:53 -08002476 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2477 seq_printf(m, " heap");
2478 } else if (vma->vm_start <= mm->start_stack &&
2479 vma->vm_end >= mm->start_stack) {
2480 seq_printf(m, " stack");
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002481 }
Christoph Lameter397874d2006-03-06 15:42:53 -08002482
2483 if (is_vm_hugetlb_page(vma)) {
2484 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2485 seq_printf(m, " huge");
2486 } else {
2487 check_pgd_range(vma, vma->vm_start, vma->vm_end,
Christoph Lameter56bbd652007-10-16 01:25:35 -07002488 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
Christoph Lameter397874d2006-03-06 15:42:53 -08002489 }
2490
2491 if (!md->pages)
2492 goto out;
2493
2494 if (md->anon)
2495 seq_printf(m," anon=%lu",md->anon);
2496
2497 if (md->dirty)
2498 seq_printf(m," dirty=%lu",md->dirty);
2499
2500 if (md->pages != md->anon && md->pages != md->dirty)
2501 seq_printf(m, " mapped=%lu", md->pages);
2502
2503 if (md->mapcount_max > 1)
2504 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2505
2506 if (md->swapcache)
2507 seq_printf(m," swapcache=%lu", md->swapcache);
2508
2509 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2510 seq_printf(m," active=%lu", md->active);
2511
2512 if (md->writeback)
2513 seq_printf(m," writeback=%lu", md->writeback);
2514
Christoph Lameter56bbd652007-10-16 01:25:35 -07002515 for_each_node_state(n, N_HIGH_MEMORY)
Christoph Lameter397874d2006-03-06 15:42:53 -08002516 if (md->node[n])
2517 seq_printf(m, " N%d=%lu", n, md->node[n]);
2518out:
2519 seq_putc(m, '\n');
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002520 kfree(md);
2521
2522 if (m->count < m->size)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002523 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002524 return 0;
2525}