blob: bda230e52acd94b0475640c4fd6b97895c6d3b9a [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>
76#include <linux/gfp.h>
77#include <linux/slab.h>
78#include <linux/string.h>
79#include <linux/module.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070080#include <linux/nsproxy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070081#include <linux/interrupt.h>
82#include <linux/init.h>
83#include <linux/compat.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080084#include <linux/swap.h>
Christoph Lameter1a75a6c2006-01-08 01:01:02 -080085#include <linux/seq_file.h>
86#include <linux/proc_fs.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080087#include <linux/migrate.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080088#include <linux/ksm.h>
Christoph Lameter95a402c2006-06-23 02:03:53 -070089#include <linux/rmap.h>
David Quigley86c3a762006-06-23 02:04:02 -070090#include <linux/security.h>
Adrian Bunkdbcb0f12007-10-16 01:26:26 -070091#include <linux/syscalls.h>
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -070092#include <linux/ctype.h>
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -080093#include <linux/mm_inline.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080094
Linus Torvalds1da177e2005-04-16 15:20:36 -070095#include <asm/tlbflush.h>
96#include <asm/uaccess.h>
97
Nick Piggin62695a82008-10-18 20:26:09 -070098#include "internal.h"
99
Christoph Lameter38e35862006-01-08 01:01:01 -0800100/* Internal flags */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800101#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
Christoph Lameter38e35862006-01-08 01:01:01 -0800102#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800103#define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800104
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800105static struct kmem_cache *policy_cache;
106static struct kmem_cache *sn_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108/* Highest zone. An specific allocation for a zone below that is not
109 policied. */
Christoph Lameter62672762007-02-10 01:43:07 -0800110enum zone_type policy_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700112/*
113 * run-time system-wide default policy => local allocation
114 */
Andi Kleend42c6992005-07-06 19:56:03 +0200115struct mempolicy default_policy = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 .refcnt = ATOMIC_INIT(1), /* never free it */
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700117 .mode = MPOL_PREFERRED,
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700118 .flags = MPOL_F_LOCAL,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119};
120
David Rientjes37012942008-04-28 02:12:33 -0700121static const struct mempolicy_operations {
122 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
123 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
124} mpol_ops[MPOL_MAX];
125
Mel Gorman19770b32008-04-28 02:12:18 -0700126/* Check that the nodemask contains at least one populated zone */
David Rientjes37012942008-04-28 02:12:33 -0700127static int is_valid_nodemask(const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128{
Mel Gorman19770b32008-04-28 02:12:18 -0700129 int nd, k;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
Mel Gorman19770b32008-04-28 02:12:18 -0700131 /* Check that there is something useful in this mask */
132 k = policy_zone;
133
134 for_each_node_mask(nd, *nodemask) {
135 struct zone *z;
136
137 for (k = 0; k <= policy_zone; k++) {
138 z = &NODE_DATA(nd)->node_zones[k];
139 if (z->present_pages > 0)
140 return 1;
Andi Kleendd942ae2006-02-17 01:39:16 +0100141 }
142 }
Mel Gorman19770b32008-04-28 02:12:18 -0700143
144 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145}
146
David Rientjesf5b087b2008-04-28 02:12:27 -0700147static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
148{
David Rientjes4c50bc02008-04-28 02:12:30 -0700149 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
150}
151
152static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
153 const nodemask_t *rel)
154{
155 nodemask_t tmp;
156 nodes_fold(tmp, *orig, nodes_weight(*rel));
157 nodes_onto(*ret, tmp, *rel);
David Rientjesf5b087b2008-04-28 02:12:27 -0700158}
159
David Rientjes37012942008-04-28 02:12:33 -0700160static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
161{
162 if (nodes_empty(*nodes))
163 return -EINVAL;
164 pol->v.nodes = *nodes;
165 return 0;
166}
167
168static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
169{
170 if (!nodes)
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700171 pol->flags |= MPOL_F_LOCAL; /* local allocation */
David Rientjes37012942008-04-28 02:12:33 -0700172 else if (nodes_empty(*nodes))
173 return -EINVAL; /* no allowed nodes */
174 else
175 pol->v.preferred_node = first_node(*nodes);
176 return 0;
177}
178
179static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
180{
181 if (!is_valid_nodemask(nodes))
182 return -EINVAL;
183 pol->v.nodes = *nodes;
184 return 0;
185}
186
Miao Xie58568d22009-06-16 15:31:49 -0700187/*
188 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
189 * any, for the new policy. mpol_new() has already validated the nodes
190 * parameter with respect to the policy mode and flags. But, we need to
191 * handle an empty nodemask with MPOL_PREFERRED here.
192 *
193 * Must be called holding task's alloc_lock to protect task's mems_allowed
194 * and mempolicy. May also be called holding the mmap_semaphore for write.
195 */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700196static int mpol_set_nodemask(struct mempolicy *pol,
197 const nodemask_t *nodes, struct nodemask_scratch *nsc)
Miao Xie58568d22009-06-16 15:31:49 -0700198{
Miao Xie58568d22009-06-16 15:31:49 -0700199 int ret;
200
201 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
202 if (pol == NULL)
203 return 0;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700204 /* Check N_HIGH_MEMORY */
205 nodes_and(nsc->mask1,
206 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
Miao Xie58568d22009-06-16 15:31:49 -0700207
208 VM_BUG_ON(!nodes);
209 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
210 nodes = NULL; /* explicit local allocation */
211 else {
212 if (pol->flags & MPOL_F_RELATIVE_NODES)
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700213 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
Miao Xie58568d22009-06-16 15:31:49 -0700214 else
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700215 nodes_and(nsc->mask2, *nodes, nsc->mask1);
216
Miao Xie58568d22009-06-16 15:31:49 -0700217 if (mpol_store_user_nodemask(pol))
218 pol->w.user_nodemask = *nodes;
219 else
220 pol->w.cpuset_mems_allowed =
221 cpuset_current_mems_allowed;
222 }
223
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700224 if (nodes)
225 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
226 else
227 ret = mpol_ops[pol->mode].create(pol, NULL);
Miao Xie58568d22009-06-16 15:31:49 -0700228 return ret;
229}
230
231/*
232 * This function just creates a new policy, does some check and simple
233 * initialization. You must invoke mpol_set_nodemask() to set nodes.
234 */
David Rientjes028fec42008-04-28 02:12:25 -0700235static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
236 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237{
238 struct mempolicy *policy;
239
David Rientjes028fec42008-04-28 02:12:25 -0700240 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
241 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
Paul Mundt140d5a42007-07-15 23:38:16 -0700242
David Rientjes3e1f0642008-04-28 02:12:34 -0700243 if (mode == MPOL_DEFAULT) {
244 if (nodes && !nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700245 return ERR_PTR(-EINVAL);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700246 return NULL; /* simply delete any existing policy */
David Rientjes37012942008-04-28 02:12:33 -0700247 }
David Rientjes3e1f0642008-04-28 02:12:34 -0700248 VM_BUG_ON(!nodes);
249
250 /*
251 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
252 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
253 * All other modes require a valid pointer to a non-empty nodemask.
254 */
255 if (mode == MPOL_PREFERRED) {
256 if (nodes_empty(*nodes)) {
257 if (((flags & MPOL_F_STATIC_NODES) ||
258 (flags & MPOL_F_RELATIVE_NODES)))
259 return ERR_PTR(-EINVAL);
David Rientjes3e1f0642008-04-28 02:12:34 -0700260 }
261 } else if (nodes_empty(*nodes))
262 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
264 if (!policy)
265 return ERR_PTR(-ENOMEM);
266 atomic_set(&policy->refcnt, 1);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700267 policy->mode = mode;
David Rientjes3e1f0642008-04-28 02:12:34 -0700268 policy->flags = flags;
David Rientjesf5b087b2008-04-28 02:12:27 -0700269
David Rientjes37012942008-04-28 02:12:33 -0700270 return policy;
271}
272
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700273/* Slow path of a mpol destructor. */
274void __mpol_put(struct mempolicy *p)
275{
276 if (!atomic_dec_and_test(&p->refcnt))
277 return;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700278 kmem_cache_free(policy_cache, p);
279}
280
David Rientjes37012942008-04-28 02:12:33 -0700281static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
282{
283}
284
285static void mpol_rebind_nodemask(struct mempolicy *pol,
286 const nodemask_t *nodes)
287{
288 nodemask_t tmp;
289
290 if (pol->flags & MPOL_F_STATIC_NODES)
291 nodes_and(tmp, pol->w.user_nodemask, *nodes);
292 else if (pol->flags & MPOL_F_RELATIVE_NODES)
293 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
294 else {
295 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
296 *nodes);
297 pol->w.cpuset_mems_allowed = *nodes;
298 }
299
300 pol->v.nodes = tmp;
301 if (!node_isset(current->il_next, tmp)) {
302 current->il_next = next_node(current->il_next, tmp);
303 if (current->il_next >= MAX_NUMNODES)
304 current->il_next = first_node(tmp);
305 if (current->il_next >= MAX_NUMNODES)
306 current->il_next = numa_node_id();
307 }
308}
309
310static void mpol_rebind_preferred(struct mempolicy *pol,
311 const nodemask_t *nodes)
312{
313 nodemask_t tmp;
314
David Rientjes37012942008-04-28 02:12:33 -0700315 if (pol->flags & MPOL_F_STATIC_NODES) {
316 int node = first_node(pol->w.user_nodemask);
317
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700318 if (node_isset(node, *nodes)) {
David Rientjes37012942008-04-28 02:12:33 -0700319 pol->v.preferred_node = node;
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700320 pol->flags &= ~MPOL_F_LOCAL;
321 } else
322 pol->flags |= MPOL_F_LOCAL;
David Rientjes37012942008-04-28 02:12:33 -0700323 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
324 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
325 pol->v.preferred_node = first_node(tmp);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700326 } else if (!(pol->flags & MPOL_F_LOCAL)) {
David Rientjes37012942008-04-28 02:12:33 -0700327 pol->v.preferred_node = node_remap(pol->v.preferred_node,
328 pol->w.cpuset_mems_allowed,
329 *nodes);
330 pol->w.cpuset_mems_allowed = *nodes;
331 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332}
333
David Rientjes1d0d2682008-04-28 02:12:32 -0700334/* Migrate a policy to a different set of nodes */
335static void mpol_rebind_policy(struct mempolicy *pol,
336 const nodemask_t *newmask)
337{
David Rientjes1d0d2682008-04-28 02:12:32 -0700338 if (!pol)
339 return;
David Rientjes1d0d2682008-04-28 02:12:32 -0700340 if (!mpol_store_user_nodemask(pol) &&
341 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
342 return;
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700343 mpol_ops[pol->mode].rebind(pol, newmask);
David Rientjes1d0d2682008-04-28 02:12:32 -0700344}
345
346/*
347 * Wrapper for mpol_rebind_policy() that just requires task
348 * pointer, and updates task mempolicy.
Miao Xie58568d22009-06-16 15:31:49 -0700349 *
350 * Called with task's alloc_lock held.
David Rientjes1d0d2682008-04-28 02:12:32 -0700351 */
352
353void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
354{
355 mpol_rebind_policy(tsk->mempolicy, new);
356}
357
358/*
359 * Rebind each vma in mm to new nodemask.
360 *
361 * Call holding a reference to mm. Takes mm->mmap_sem during call.
362 */
363
364void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
365{
366 struct vm_area_struct *vma;
367
368 down_write(&mm->mmap_sem);
369 for (vma = mm->mmap; vma; vma = vma->vm_next)
370 mpol_rebind_policy(vma->vm_policy, new);
371 up_write(&mm->mmap_sem);
372}
373
David Rientjes37012942008-04-28 02:12:33 -0700374static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
375 [MPOL_DEFAULT] = {
376 .rebind = mpol_rebind_default,
377 },
378 [MPOL_INTERLEAVE] = {
379 .create = mpol_new_interleave,
380 .rebind = mpol_rebind_nodemask,
381 },
382 [MPOL_PREFERRED] = {
383 .create = mpol_new_preferred,
384 .rebind = mpol_rebind_preferred,
385 },
386 [MPOL_BIND] = {
387 .create = mpol_new_bind,
388 .rebind = mpol_rebind_nodemask,
389 },
390};
391
Christoph Lameter397874d2006-03-06 15:42:53 -0800392static void gather_stats(struct page *, void *, int pte_dirty);
Christoph Lameterfc301282006-01-18 17:42:29 -0800393static void migrate_page_add(struct page *page, struct list_head *pagelist,
394 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800395
Christoph Lameter38e35862006-01-08 01:01:01 -0800396/* Scan through pages checking if pages follow certain conditions. */
Nick Pigginb5810032005-10-29 18:16:12 -0700397static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800398 unsigned long addr, unsigned long end,
399 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800400 void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401{
Hugh Dickins91612e02005-06-21 17:15:07 -0700402 pte_t *orig_pte;
403 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700404 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700405
Hugh Dickins705e87c2005-10-29 18:16:27 -0700406 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700407 do {
Linus Torvalds6aab3412005-11-28 14:34:23 -0800408 struct page *page;
Andy Whitcroft25ba77c2006-12-06 20:33:03 -0800409 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700410
411 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800413 page = vm_normal_page(vma, addr, *pte);
414 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800416 /*
Hugh Dickins62b61f62009-12-14 17:59:33 -0800417 * vm_normal_page() filters out zero pages, but there might
418 * still be PageReserved pages to skip, perhaps in a VDSO.
419 * And we cannot move PageKsm pages sensibly or safely yet.
Nick Piggin053837f2006-01-18 17:42:27 -0800420 */
Hugh Dickins62b61f62009-12-14 17:59:33 -0800421 if (PageReserved(page) || PageKsm(page))
Christoph Lameterf4598c82006-01-12 01:05:20 -0800422 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800423 nid = page_to_nid(page);
Christoph Lameter38e35862006-01-08 01:01:01 -0800424 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
425 continue;
426
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800427 if (flags & MPOL_MF_STATS)
Christoph Lameter397874d2006-03-06 15:42:53 -0800428 gather_stats(page, private, pte_dirty(*pte));
Nick Piggin053837f2006-01-18 17:42:27 -0800429 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameterfc301282006-01-18 17:42:29 -0800430 migrate_page_add(page, private, flags);
Christoph Lameter38e35862006-01-08 01:01:01 -0800431 else
432 break;
Hugh Dickins91612e02005-06-21 17:15:07 -0700433 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickins705e87c2005-10-29 18:16:27 -0700434 pte_unmap_unlock(orig_pte, ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700435 return addr != end;
436}
437
Nick Pigginb5810032005-10-29 18:16:12 -0700438static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800439 unsigned long addr, unsigned long end,
440 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800441 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700442{
443 pmd_t *pmd;
444 unsigned long next;
445
446 pmd = pmd_offset(pud, addr);
447 do {
448 next = pmd_addr_end(addr, end);
449 if (pmd_none_or_clear_bad(pmd))
450 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800451 if (check_pte_range(vma, pmd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800452 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700453 return -EIO;
454 } while (pmd++, addr = next, addr != end);
455 return 0;
456}
457
Nick Pigginb5810032005-10-29 18:16:12 -0700458static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800459 unsigned long addr, unsigned long end,
460 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800461 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700462{
463 pud_t *pud;
464 unsigned long next;
465
466 pud = pud_offset(pgd, addr);
467 do {
468 next = pud_addr_end(addr, end);
469 if (pud_none_or_clear_bad(pud))
470 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800471 if (check_pmd_range(vma, pud, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800472 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700473 return -EIO;
474 } while (pud++, addr = next, addr != end);
475 return 0;
476}
477
Nick Pigginb5810032005-10-29 18:16:12 -0700478static inline int check_pgd_range(struct vm_area_struct *vma,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800479 unsigned long addr, unsigned long end,
480 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800481 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700482{
483 pgd_t *pgd;
484 unsigned long next;
485
Nick Pigginb5810032005-10-29 18:16:12 -0700486 pgd = pgd_offset(vma->vm_mm, addr);
Hugh Dickins91612e02005-06-21 17:15:07 -0700487 do {
488 next = pgd_addr_end(addr, end);
489 if (pgd_none_or_clear_bad(pgd))
490 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800491 if (check_pud_range(vma, pgd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800492 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700493 return -EIO;
494 } while (pgd++, addr = next, addr != end);
495 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496}
497
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800498/*
499 * Check if all pages in a range are on a set of nodes.
500 * If pagelist != NULL then isolate pages from the LRU and
501 * put them on the pagelist.
502 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503static struct vm_area_struct *
504check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Christoph Lameter38e35862006-01-08 01:01:01 -0800505 const nodemask_t *nodes, unsigned long flags, void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506{
507 int err;
508 struct vm_area_struct *first, *vma, *prev;
509
Nick Piggin053837f2006-01-18 17:42:27 -0800510
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 first = find_vma(mm, start);
512 if (!first)
513 return ERR_PTR(-EFAULT);
514 prev = NULL;
515 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800516 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
517 if (!vma->vm_next && vma->vm_end < end)
518 return ERR_PTR(-EFAULT);
519 if (prev && prev->vm_end < vma->vm_start)
520 return ERR_PTR(-EFAULT);
521 }
522 if (!is_vm_hugetlb_page(vma) &&
523 ((flags & MPOL_MF_STRICT) ||
524 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
525 vma_migratable(vma)))) {
Andi Kleen5b952b32005-09-13 01:25:08 -0700526 unsigned long endvma = vma->vm_end;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800527
Andi Kleen5b952b32005-09-13 01:25:08 -0700528 if (endvma > end)
529 endvma = end;
530 if (vma->vm_start > start)
531 start = vma->vm_start;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800532 err = check_pgd_range(vma, start, endvma, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800533 flags, private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 if (err) {
535 first = ERR_PTR(err);
536 break;
537 }
538 }
539 prev = vma;
540 }
541 return first;
542}
543
544/* Apply policy to a single VMA */
545static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
546{
547 int err = 0;
548 struct mempolicy *old = vma->vm_policy;
549
Paul Mundt140d5a42007-07-15 23:38:16 -0700550 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 vma->vm_start, vma->vm_end, vma->vm_pgoff,
552 vma->vm_ops, vma->vm_file,
553 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
554
555 if (vma->vm_ops && vma->vm_ops->set_policy)
556 err = vma->vm_ops->set_policy(vma, new);
557 if (!err) {
558 mpol_get(new);
559 vma->vm_policy = new;
Lee Schermerhornf0be3d32008-04-28 02:13:08 -0700560 mpol_put(old);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 }
562 return err;
563}
564
565/* Step 2: apply policy to a range and do splits. */
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800566static int mbind_range(struct mm_struct *mm, unsigned long start,
567 unsigned long end, struct mempolicy *new_pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568{
569 struct vm_area_struct *next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800570 struct vm_area_struct *prev;
571 struct vm_area_struct *vma;
572 int err = 0;
573 pgoff_t pgoff;
574 unsigned long vmstart;
575 unsigned long vmend;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800577 vma = find_vma_prev(mm, start, &prev);
578 if (!vma || vma->vm_start > start)
579 return -EFAULT;
580
581 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 next = vma->vm_next;
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800583 vmstart = max(start, vma->vm_start);
584 vmend = min(end, vma->vm_end);
585
586 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
587 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
588 vma->anon_vma, vma->vm_file, pgoff, new_pol);
589 if (prev) {
590 vma = prev;
591 next = vma->vm_next;
592 continue;
593 }
594 if (vma->vm_start != vmstart) {
595 err = split_vma(vma->vm_mm, vma, vmstart, 1);
596 if (err)
597 goto out;
598 }
599 if (vma->vm_end != vmend) {
600 err = split_vma(vma->vm_mm, vma, vmend, 0);
601 if (err)
602 goto out;
603 }
604 err = policy_vma(vma, new_pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 if (err)
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800606 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 }
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -0800608
609 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 return err;
611}
612
Paul Jacksonc61afb12006-03-24 03:16:08 -0800613/*
614 * Update task->flags PF_MEMPOLICY bit: set iff non-default
615 * mempolicy. Allows more rapid checking of this (combined perhaps
616 * with other PF_* flag bits) on memory allocation hot code paths.
617 *
618 * If called from outside this file, the task 'p' should -only- be
619 * a newly forked child not yet visible on the task list, because
620 * manipulating the task flags of a visible task is not safe.
621 *
622 * The above limitation is why this routine has the funny name
623 * mpol_fix_fork_child_flag().
624 *
625 * It is also safe to call this with a task pointer of current,
626 * which the static wrapper mpol_set_task_struct_flag() does,
627 * for use within this file.
628 */
629
630void mpol_fix_fork_child_flag(struct task_struct *p)
631{
632 if (p->mempolicy)
633 p->flags |= PF_MEMPOLICY;
634 else
635 p->flags &= ~PF_MEMPOLICY;
636}
637
638static void mpol_set_task_struct_flag(void)
639{
640 mpol_fix_fork_child_flag(current);
641}
642
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700644static long do_set_mempolicy(unsigned short mode, unsigned short flags,
645 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646{
Miao Xie58568d22009-06-16 15:31:49 -0700647 struct mempolicy *new, *old;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700648 struct mm_struct *mm = current->mm;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700649 NODEMASK_SCRATCH(scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700650 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700652 if (!scratch)
653 return -ENOMEM;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700654
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700655 new = mpol_new(mode, flags, nodes);
656 if (IS_ERR(new)) {
657 ret = PTR_ERR(new);
658 goto out;
659 }
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700660 /*
661 * prevent changing our mempolicy while show_numa_maps()
662 * is using it.
663 * Note: do_set_mempolicy() can be called at init time
664 * with no 'mm'.
665 */
666 if (mm)
667 down_write(&mm->mmap_sem);
Miao Xie58568d22009-06-16 15:31:49 -0700668 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700669 ret = mpol_set_nodemask(new, nodes, scratch);
Miao Xie58568d22009-06-16 15:31:49 -0700670 if (ret) {
671 task_unlock(current);
672 if (mm)
673 up_write(&mm->mmap_sem);
674 mpol_put(new);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700675 goto out;
Miao Xie58568d22009-06-16 15:31:49 -0700676 }
677 old = current->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 current->mempolicy = new;
Paul Jacksonc61afb12006-03-24 03:16:08 -0800679 mpol_set_task_struct_flag();
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700680 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700681 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700682 current->il_next = first_node(new->v.nodes);
Miao Xie58568d22009-06-16 15:31:49 -0700683 task_unlock(current);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700684 if (mm)
685 up_write(&mm->mmap_sem);
686
Miao Xie58568d22009-06-16 15:31:49 -0700687 mpol_put(old);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -0700688 ret = 0;
689out:
690 NODEMASK_SCRATCH_FREE(scratch);
691 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692}
693
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700694/*
695 * Return nodemask for policy for get_mempolicy() query
Miao Xie58568d22009-06-16 15:31:49 -0700696 *
697 * Called with task's alloc_lock held
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700698 */
699static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700701 nodes_clear(*nodes);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700702 if (p == &default_policy)
703 return;
704
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700705 switch (p->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -0700706 case MPOL_BIND:
707 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700709 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 break;
711 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -0700712 if (!(p->flags & MPOL_F_LOCAL))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700713 node_set(p->v.preferred_node, *nodes);
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700714 /* else return empty node mask for local allocation */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 break;
716 default:
717 BUG();
718 }
719}
720
721static int lookup_node(struct mm_struct *mm, unsigned long addr)
722{
723 struct page *p;
724 int err;
725
726 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
727 if (err >= 0) {
728 err = page_to_nid(p);
729 put_page(p);
730 }
731 return err;
732}
733
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700735static long do_get_mempolicy(int *policy, nodemask_t *nmask,
736 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700738 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 struct mm_struct *mm = current->mm;
740 struct vm_area_struct *vma = NULL;
741 struct mempolicy *pol = current->mempolicy;
742
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700743 if (flags &
744 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700746
747 if (flags & MPOL_F_MEMS_ALLOWED) {
748 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
749 return -EINVAL;
750 *policy = 0; /* just so it's initialized */
Miao Xie58568d22009-06-16 15:31:49 -0700751 task_lock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700752 *nmask = cpuset_current_mems_allowed;
Miao Xie58568d22009-06-16 15:31:49 -0700753 task_unlock(current);
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700754 return 0;
755 }
756
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 if (flags & MPOL_F_ADDR) {
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700758 /*
759 * Do NOT fall back to task policy if the
760 * vma/shared policy at addr is NULL. We
761 * want to return MPOL_DEFAULT in this case.
762 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 down_read(&mm->mmap_sem);
764 vma = find_vma_intersection(mm, addr, addr+1);
765 if (!vma) {
766 up_read(&mm->mmap_sem);
767 return -EFAULT;
768 }
769 if (vma->vm_ops && vma->vm_ops->get_policy)
770 pol = vma->vm_ops->get_policy(vma, addr);
771 else
772 pol = vma->vm_policy;
773 } else if (addr)
774 return -EINVAL;
775
776 if (!pol)
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700777 pol = &default_policy; /* indicates default behavior */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778
779 if (flags & MPOL_F_NODE) {
780 if (flags & MPOL_F_ADDR) {
781 err = lookup_node(mm, addr);
782 if (err < 0)
783 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700784 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700786 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700787 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 } else {
789 err = -EINVAL;
790 goto out;
791 }
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700792 } else {
793 *policy = pol == &default_policy ? MPOL_DEFAULT :
794 pol->mode;
David Rientjesd79df632008-07-04 12:24:13 -0700795 /*
796 * Internal mempolicy flags must be masked off before exposing
797 * the policy to userspace.
798 */
799 *policy |= (pol->flags & MPOL_MODE_FLAGS);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700800 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801
802 if (vma) {
803 up_read(&current->mm->mmap_sem);
804 vma = NULL;
805 }
806
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 err = 0;
Miao Xie58568d22009-06-16 15:31:49 -0700808 if (nmask) {
809 task_lock(current);
Lee Schermerhornbea904d2008-04-28 02:13:18 -0700810 get_policy_nodemask(pol, nmask);
Miao Xie58568d22009-06-16 15:31:49 -0700811 task_unlock(current);
812 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813
814 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700815 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 if (vma)
817 up_read(&current->mm->mmap_sem);
818 return err;
819}
820
Christoph Lameterb20a3502006-03-22 00:09:12 -0800821#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700822/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800823 * page migration
824 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800825static void migrate_page_add(struct page *page, struct list_head *pagelist,
826 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800827{
828 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800829 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800830 */
Nick Piggin62695a82008-10-18 20:26:09 -0700831 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
832 if (!isolate_lru_page(page)) {
833 list_add_tail(&page->lru, pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -0800834 inc_zone_page_state(page, NR_ISOLATED_ANON +
835 page_is_file_cache(page));
Nick Piggin62695a82008-10-18 20:26:09 -0700836 }
837 }
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800838}
839
Christoph Lameter742755a2006-06-23 02:03:55 -0700840static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700841{
Mel Gorman6484eb32009-06-16 15:31:54 -0700842 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700843}
844
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800845/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800846 * Migrate pages from one node to a target node.
847 * Returns error or the number of pages not migrated.
848 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700849static int migrate_to_node(struct mm_struct *mm, int source, int dest,
850 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800851{
852 nodemask_t nmask;
853 LIST_HEAD(pagelist);
854 int err = 0;
855
856 nodes_clear(nmask);
857 node_set(source, nmask);
858
859 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
860 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
861
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700862 if (!list_empty(&pagelist))
Hugh Dickins62b61f62009-12-14 17:59:33 -0800863 err = migrate_pages(&pagelist, new_node_page, dest, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700864
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800865 return err;
866}
867
868/*
869 * Move pages between the two nodesets so as to preserve the physical
870 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -0800871 *
872 * Returns the number of page that could not be moved.
873 */
874int do_migrate_pages(struct mm_struct *mm,
875 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
876{
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800877 int busy = 0;
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800878 int err;
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800879 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -0800880
Christoph Lameter0aedadf2008-11-06 12:53:30 -0800881 err = migrate_prep();
882 if (err)
883 return err;
884
Lee Schermerhorn53f25562008-04-28 02:13:20 -0700885 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -0800886
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700887 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
888 if (err)
889 goto out;
890
KOSAKI Motohiroda0aa132010-03-05 13:41:59 -0800891 /*
892 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
893 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
894 * bit in 'tmp', and return that <source, dest> pair for migration.
895 * The pair of nodemasks 'to' and 'from' define the map.
896 *
897 * If no pair of bits is found that way, fallback to picking some
898 * pair of 'source' and 'dest' bits that are not the same. If the
899 * 'source' and 'dest' bits are the same, this represents a node
900 * that will be migrating to itself, so no pages need move.
901 *
902 * If no bits are left in 'tmp', or if all remaining bits left
903 * in 'tmp' correspond to the same bit in 'to', return false
904 * (nothing left to migrate).
905 *
906 * This lets us pick a pair of nodes to migrate between, such that
907 * if possible the dest node is not already occupied by some other
908 * source node, minimizing the risk of overloading the memory on a
909 * node that would happen if we migrated incoming memory to a node
910 * before migrating outgoing memory source that same node.
911 *
912 * A single scan of tmp is sufficient. As we go, we remember the
913 * most recent <s, d> pair that moved (s != d). If we find a pair
914 * that not only moved, but what's better, moved to an empty slot
915 * (d is not set in tmp), then we break out then, with that pair.
916 * Otherwise when we finish scannng from_tmp, we at least have the
917 * most recent <s, d> pair that moved. If we get all the way through
918 * the scan of tmp without finding any node that moved, much less
919 * moved to an empty node, then there is nothing left worth migrating.
920 */
Christoph Lameterd4984712006-01-08 01:00:55 -0800921
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800922 tmp = *from_nodes;
923 while (!nodes_empty(tmp)) {
924 int s,d;
925 int source = -1;
926 int dest = 0;
927
928 for_each_node_mask(s, tmp) {
929 d = node_remap(s, *from_nodes, *to_nodes);
930 if (s == d)
931 continue;
932
933 source = s; /* Node moved. Memorize */
934 dest = d;
935
936 /* dest not in remaining from nodes? */
937 if (!node_isset(dest, tmp))
938 break;
939 }
940 if (source == -1)
941 break;
942
943 node_clear(source, tmp);
944 err = migrate_to_node(mm, source, dest, flags);
945 if (err > 0)
946 busy += err;
947 if (err < 0)
948 break;
Christoph Lameter39743882006-01-08 01:00:51 -0800949 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700950out:
Christoph Lameter39743882006-01-08 01:00:51 -0800951 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800952 if (err < 0)
953 return err;
954 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800955
Christoph Lameter39743882006-01-08 01:00:51 -0800956}
957
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800958/*
959 * Allocate a new page for page migration based on vma policy.
960 * Start assuming that page is mapped by vma pointed to by @private.
961 * Search forward from there, if not. N.B., this assumes that the
962 * list of pages handed to migrate_pages()--which is how we get here--
963 * is in virtual address order.
964 */
Christoph Lameter742755a2006-06-23 02:03:55 -0700965static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700966{
967 struct vm_area_struct *vma = (struct vm_area_struct *)private;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800968 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700969
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800970 while (vma) {
971 address = page_address_in_vma(page, vma);
972 if (address != -EFAULT)
973 break;
974 vma = vma->vm_next;
975 }
976
977 /*
978 * if !vma, alloc_page_vma() will use task or system default policy
979 */
980 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700981}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800982#else
983
984static void migrate_page_add(struct page *page, struct list_head *pagelist,
985 unsigned long flags)
986{
987}
988
989int do_migrate_pages(struct mm_struct *mm,
990 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
991{
992 return -ENOSYS;
993}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700994
Keith Owens69939742006-10-11 01:21:28 -0700995static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700996{
997 return NULL;
998}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800999#endif
1000
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001001static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -07001002 unsigned short mode, unsigned short mode_flags,
1003 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001004{
1005 struct vm_area_struct *vma;
1006 struct mm_struct *mm = current->mm;
1007 struct mempolicy *new;
1008 unsigned long end;
1009 int err;
1010 LIST_HEAD(pagelist);
1011
David Rientjesa3b51e02008-04-28 02:12:23 -07001012 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1013 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001014 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -08001015 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001016 return -EPERM;
1017
1018 if (start & ~PAGE_MASK)
1019 return -EINVAL;
1020
1021 if (mode == MPOL_DEFAULT)
1022 flags &= ~MPOL_MF_STRICT;
1023
1024 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1025 end = start + len;
1026
1027 if (end < start)
1028 return -EINVAL;
1029 if (end == start)
1030 return 0;
1031
David Rientjes028fec42008-04-28 02:12:25 -07001032 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001033 if (IS_ERR(new))
1034 return PTR_ERR(new);
1035
1036 /*
1037 * If we are using the default policy then operation
1038 * on discontinuous address spaces is okay after all
1039 */
1040 if (!new)
1041 flags |= MPOL_MF_DISCONTIG_OK;
1042
David Rientjes028fec42008-04-28 02:12:25 -07001043 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1044 start, start + len, mode, mode_flags,
1045 nmask ? nodes_addr(*nmask)[0] : -1);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001046
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001047 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1048
1049 err = migrate_prep();
1050 if (err)
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001051 goto mpol_out;
Christoph Lameter0aedadf2008-11-06 12:53:30 -08001052 }
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001053 {
1054 NODEMASK_SCRATCH(scratch);
1055 if (scratch) {
1056 down_write(&mm->mmap_sem);
1057 task_lock(current);
1058 err = mpol_set_nodemask(new, nmask, scratch);
1059 task_unlock(current);
1060 if (err)
1061 up_write(&mm->mmap_sem);
1062 } else
1063 err = -ENOMEM;
1064 NODEMASK_SCRATCH_FREE(scratch);
1065 }
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001066 if (err)
1067 goto mpol_out;
1068
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001069 vma = check_range(mm, start, end, nmask,
1070 flags | MPOL_MF_INVERT, &pagelist);
1071
1072 err = PTR_ERR(vma);
1073 if (!IS_ERR(vma)) {
1074 int nr_failed = 0;
1075
KOSAKI Motohiro9d8cebd2010-03-05 13:41:57 -08001076 err = mbind_range(mm, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -08001077
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001078 if (!list_empty(&pagelist))
Christoph Lameter95a402c2006-06-23 02:03:53 -07001079 nr_failed = migrate_pages(&pagelist, new_vma_page,
Hugh Dickins62b61f62009-12-14 17:59:33 -08001080 (unsigned long)vma, 0);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001081
1082 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1083 err = -EIO;
KOSAKI Motohiroab8a3e12009-10-26 16:49:58 -07001084 } else
1085 putback_lru_pages(&pagelist);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001086
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001087 up_write(&mm->mmap_sem);
KOSAKI Motohirob05ca732009-10-26 16:49:59 -07001088 mpol_out:
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001089 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -08001090 return err;
1091}
1092
Christoph Lameter39743882006-01-08 01:00:51 -08001093/*
Christoph Lameter8bccd852005-10-29 18:16:59 -07001094 * User space interface with variable sized bitmaps for nodelists.
1095 */
1096
1097/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -08001098static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -07001099 unsigned long maxnode)
1100{
1101 unsigned long k;
1102 unsigned long nlongs;
1103 unsigned long endmask;
1104
1105 --maxnode;
1106 nodes_clear(*nodes);
1107 if (maxnode == 0 || !nmask)
1108 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -08001109 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -08001110 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001111
1112 nlongs = BITS_TO_LONGS(maxnode);
1113 if ((maxnode % BITS_PER_LONG) == 0)
1114 endmask = ~0UL;
1115 else
1116 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1117
1118 /* When the user specified more nodes than supported just check
1119 if the non supported part is all zero. */
1120 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1121 if (nlongs > PAGE_SIZE/sizeof(long))
1122 return -EINVAL;
1123 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1124 unsigned long t;
1125 if (get_user(t, nmask + k))
1126 return -EFAULT;
1127 if (k == nlongs - 1) {
1128 if (t & endmask)
1129 return -EINVAL;
1130 } else if (t)
1131 return -EINVAL;
1132 }
1133 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1134 endmask = ~0UL;
1135 }
1136
1137 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1138 return -EFAULT;
1139 nodes_addr(*nodes)[nlongs-1] &= endmask;
1140 return 0;
1141}
1142
1143/* Copy a kernel node mask to user space */
1144static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1145 nodemask_t *nodes)
1146{
1147 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1148 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1149
1150 if (copy > nbytes) {
1151 if (copy > PAGE_SIZE)
1152 return -EINVAL;
1153 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1154 return -EFAULT;
1155 copy = nbytes;
1156 }
1157 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1158}
1159
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001160SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1161 unsigned long, mode, unsigned long __user *, nmask,
1162 unsigned long, maxnode, unsigned, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001163{
1164 nodemask_t nodes;
1165 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001166 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001167
David Rientjes028fec42008-04-28 02:12:25 -07001168 mode_flags = mode & MPOL_MODE_FLAGS;
1169 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001170 if (mode >= MPOL_MAX)
1171 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001172 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1173 (mode_flags & MPOL_F_RELATIVE_NODES))
1174 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001175 err = get_nodes(&nodes, nmask, maxnode);
1176 if (err)
1177 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001178 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001179}
1180
1181/* Set the process memory policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001182SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1183 unsigned long, maxnode)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001184{
1185 int err;
1186 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001187 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001188
David Rientjes028fec42008-04-28 02:12:25 -07001189 flags = mode & MPOL_MODE_FLAGS;
1190 mode &= ~MPOL_MODE_FLAGS;
1191 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001192 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001193 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1194 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001195 err = get_nodes(&nodes, nmask, maxnode);
1196 if (err)
1197 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001198 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001199}
1200
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001201SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1202 const unsigned long __user *, old_nodes,
1203 const unsigned long __user *, new_nodes)
Christoph Lameter39743882006-01-08 01:00:51 -08001204{
David Howellsc69e8d92008-11-14 10:39:19 +11001205 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter39743882006-01-08 01:00:51 -08001206 struct mm_struct *mm;
1207 struct task_struct *task;
1208 nodemask_t old;
1209 nodemask_t new;
1210 nodemask_t task_nodes;
1211 int err;
1212
1213 err = get_nodes(&old, old_nodes, maxnode);
1214 if (err)
1215 return err;
1216
1217 err = get_nodes(&new, new_nodes, maxnode);
1218 if (err)
1219 return err;
1220
1221 /* Find the mm_struct */
1222 read_lock(&tasklist_lock);
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001223 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001224 if (!task) {
1225 read_unlock(&tasklist_lock);
1226 return -ESRCH;
1227 }
1228 mm = get_task_mm(task);
1229 read_unlock(&tasklist_lock);
1230
1231 if (!mm)
1232 return -EINVAL;
1233
1234 /*
1235 * Check if this process has the right to modify the specified
1236 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001237 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001238 * userid as the target process.
1239 */
David Howellsc69e8d92008-11-14 10:39:19 +11001240 rcu_read_lock();
1241 tcred = __task_cred(task);
David Howellsb6dff3e2008-11-14 10:39:16 +11001242 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1243 cred->uid != tcred->suid && cred->uid != tcred->uid &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001244 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001245 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001246 err = -EPERM;
1247 goto out;
1248 }
David Howellsc69e8d92008-11-14 10:39:19 +11001249 rcu_read_unlock();
Christoph Lameter39743882006-01-08 01:00:51 -08001250
1251 task_nodes = cpuset_mems_allowed(task);
1252 /* Is the user allowed to access the target nodes? */
Christoph Lameter74c00242006-03-14 19:50:21 -08001253 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001254 err = -EPERM;
1255 goto out;
1256 }
1257
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001258 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001259 err = -EINVAL;
1260 goto out;
1261 }
1262
David Quigley86c3a762006-06-23 02:04:02 -07001263 err = security_task_movememory(task);
1264 if (err)
1265 goto out;
1266
Christoph Lameter511030b2006-02-28 16:58:57 -08001267 err = do_migrate_pages(mm, &old, &new,
Christoph Lameter74c00242006-03-14 19:50:21 -08001268 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter39743882006-01-08 01:00:51 -08001269out:
1270 mmput(mm);
1271 return err;
1272}
1273
1274
Christoph Lameter8bccd852005-10-29 18:16:59 -07001275/* Retrieve NUMA policy */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001276SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1277 unsigned long __user *, nmask, unsigned long, maxnode,
1278 unsigned long, addr, unsigned long, flags)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001279{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001280 int err;
1281 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001282 nodemask_t nodes;
1283
1284 if (nmask != NULL && maxnode < MAX_NUMNODES)
1285 return -EINVAL;
1286
1287 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1288
1289 if (err)
1290 return err;
1291
1292 if (policy && put_user(pval, policy))
1293 return -EFAULT;
1294
1295 if (nmask)
1296 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1297
1298 return err;
1299}
1300
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301#ifdef CONFIG_COMPAT
1302
1303asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1304 compat_ulong_t __user *nmask,
1305 compat_ulong_t maxnode,
1306 compat_ulong_t addr, compat_ulong_t flags)
1307{
1308 long err;
1309 unsigned long __user *nm = NULL;
1310 unsigned long nr_bits, alloc_size;
1311 DECLARE_BITMAP(bm, MAX_NUMNODES);
1312
1313 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1314 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1315
1316 if (nmask)
1317 nm = compat_alloc_user_space(alloc_size);
1318
1319 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1320
1321 if (!err && nmask) {
1322 err = copy_from_user(bm, nm, alloc_size);
1323 /* ensure entire bitmap is zeroed */
1324 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1325 err |= compat_put_bitmap(nmask, bm, nr_bits);
1326 }
1327
1328 return err;
1329}
1330
1331asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1332 compat_ulong_t maxnode)
1333{
1334 long err = 0;
1335 unsigned long __user *nm = NULL;
1336 unsigned long nr_bits, alloc_size;
1337 DECLARE_BITMAP(bm, MAX_NUMNODES);
1338
1339 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1340 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1341
1342 if (nmask) {
1343 err = compat_get_bitmap(bm, nmask, nr_bits);
1344 nm = compat_alloc_user_space(alloc_size);
1345 err |= copy_to_user(nm, bm, alloc_size);
1346 }
1347
1348 if (err)
1349 return -EFAULT;
1350
1351 return sys_set_mempolicy(mode, nm, nr_bits+1);
1352}
1353
1354asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1355 compat_ulong_t mode, compat_ulong_t __user *nmask,
1356 compat_ulong_t maxnode, compat_ulong_t flags)
1357{
1358 long err = 0;
1359 unsigned long __user *nm = NULL;
1360 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001361 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
1363 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1364 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1365
1366 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001367 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001369 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 }
1371
1372 if (err)
1373 return -EFAULT;
1374
1375 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1376}
1377
1378#endif
1379
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001380/*
1381 * get_vma_policy(@task, @vma, @addr)
1382 * @task - task for fallback if vma policy == default
1383 * @vma - virtual memory area whose policy is sought
1384 * @addr - address in @vma for shared policy lookup
1385 *
1386 * Returns effective policy for a VMA at specified address.
1387 * Falls back to @task or system default policy, as necessary.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001388 * Current or other task's task mempolicy and non-shared vma policies
1389 * are protected by the task's mmap_sem, which must be held for read by
1390 * the caller.
1391 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1392 * count--added by the get_policy() vm_op, as appropriate--to protect against
1393 * freeing by another task. It is the caller's responsibility to free the
1394 * extra reference for shared policies.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001395 */
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001396static struct mempolicy *get_vma_policy(struct task_struct *task,
Christoph Lameter48fce342006-01-08 01:01:03 -08001397 struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001399 struct mempolicy *pol = task->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400
1401 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001402 if (vma->vm_ops && vma->vm_ops->get_policy) {
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001403 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1404 addr);
1405 if (vpol)
1406 pol = vpol;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001407 } else if (vma->vm_policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 pol = vma->vm_policy;
1409 }
1410 if (!pol)
1411 pol = &default_policy;
1412 return pol;
1413}
1414
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001415/*
1416 * Return a nodemask representing a mempolicy for filtering nodes for
1417 * page allocation
1418 */
1419static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001420{
1421 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001422 if (unlikely(policy->mode == MPOL_BIND) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001423 gfp_zone(gfp) >= policy_zone &&
1424 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1425 return &policy->v.nodes;
1426
1427 return NULL;
1428}
1429
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001430/* Return a zonelist indicated by gfp for node representing a mempolicy */
1431static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432{
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001433 int nd = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001435 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001437 if (!(policy->flags & MPOL_F_LOCAL))
1438 nd = policy->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 break;
1440 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001441 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001442 * Normally, MPOL_BIND allocations are node-local within the
1443 * allowed nodemask. However, if __GFP_THISNODE is set and the
1444 * current node is part of the mask, we use the zonelist for
1445 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001446 */
Mel Gorman19770b32008-04-28 02:12:18 -07001447 if (unlikely(gfp & __GFP_THISNODE) &&
1448 unlikely(!node_isset(nd, policy->v.nodes)))
1449 nd = first_node(policy->v.nodes);
1450 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 case MPOL_INTERLEAVE: /* should not happen */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 break;
1453 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 BUG();
1455 }
Mel Gorman0e884602008-04-28 02:12:14 -07001456 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457}
1458
1459/* Do dynamic interleaving for a process */
1460static unsigned interleave_nodes(struct mempolicy *policy)
1461{
1462 unsigned nid, next;
1463 struct task_struct *me = current;
1464
1465 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001466 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001468 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001469 if (next < MAX_NUMNODES)
1470 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 return nid;
1472}
1473
Christoph Lameterdc85da12006-01-18 17:42:36 -08001474/*
1475 * Depending on the memory policy provide a node from which to allocate the
1476 * next slab entry.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001477 * @policy must be protected by freeing by the caller. If @policy is
1478 * the current task's mempolicy, this protection is implicit, as only the
1479 * task can change it's policy. The system default policy requires no
1480 * such protection.
Christoph Lameterdc85da12006-01-18 17:42:36 -08001481 */
1482unsigned slab_node(struct mempolicy *policy)
1483{
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001484 if (!policy || policy->flags & MPOL_F_LOCAL)
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001485 return numa_node_id();
Christoph Lameter765c4502006-09-27 01:50:08 -07001486
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001487 switch (policy->mode) {
1488 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001489 /*
1490 * handled MPOL_F_LOCAL above
1491 */
1492 return policy->v.preferred_node;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001493
Christoph Lameterdc85da12006-01-18 17:42:36 -08001494 case MPOL_INTERLEAVE:
1495 return interleave_nodes(policy);
1496
Mel Gormandd1a2392008-04-28 02:12:17 -07001497 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001498 /*
1499 * Follow bind policy behavior and start allocation at the
1500 * first node.
1501 */
Mel Gorman19770b32008-04-28 02:12:18 -07001502 struct zonelist *zonelist;
1503 struct zone *zone;
1504 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1505 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1506 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1507 &policy->v.nodes,
1508 &zone);
1509 return zone->node;
Mel Gormandd1a2392008-04-28 02:12:17 -07001510 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001511
Christoph Lameterdc85da12006-01-18 17:42:36 -08001512 default:
Lee Schermerhornbea904d2008-04-28 02:13:18 -07001513 BUG();
Christoph Lameterdc85da12006-01-18 17:42:36 -08001514 }
1515}
1516
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517/* Do static interleaving for a VMA with known offset. */
1518static unsigned offset_il_node(struct mempolicy *pol,
1519 struct vm_area_struct *vma, unsigned long off)
1520{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001521 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001522 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 int c;
1524 int nid = -1;
1525
David Rientjesf5b087b2008-04-28 02:12:27 -07001526 if (!nnodes)
1527 return numa_node_id();
1528 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 c = 0;
1530 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001531 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 c++;
1533 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 return nid;
1535}
1536
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001537/* Determine a node number for interleave */
1538static inline unsigned interleave_nid(struct mempolicy *pol,
1539 struct vm_area_struct *vma, unsigned long addr, int shift)
1540{
1541 if (vma) {
1542 unsigned long off;
1543
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001544 /*
1545 * for small pages, there is no difference between
1546 * shift and PAGE_SHIFT, so the bit-shift is safe.
1547 * for huge pages, since vm_pgoff is in units of small
1548 * pages, we need to shift off the always 0 bits to get
1549 * a useful offset.
1550 */
1551 BUG_ON(shift < PAGE_SHIFT);
1552 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001553 off += (addr - vma->vm_start) >> shift;
1554 return offset_il_node(pol, vma, off);
1555 } else
1556 return interleave_nodes(pol);
1557}
1558
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001559#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001560/*
1561 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1562 * @vma = virtual memory area whose policy is sought
1563 * @addr = address in @vma for shared policy lookup and interleave policy
1564 * @gfp_flags = for requested zone
Mel Gorman19770b32008-04-28 02:12:18 -07001565 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1566 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001567 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001568 * Returns a zonelist suitable for a huge page allocation and a pointer
1569 * to the struct mempolicy for conditional unref after allocation.
1570 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1571 * @nodemask for filtering the zonelist.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001572 */
Mel Gorman396faf02007-07-17 04:03:13 -07001573struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001574 gfp_t gfp_flags, struct mempolicy **mpol,
1575 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001576{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001577 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001578
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001579 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001580 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001581
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001582 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1583 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
Andi Kleena5516432008-07-23 21:27:41 -07001584 huge_page_shift(hstate_vma(vma))), gfp_flags);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001585 } else {
1586 zl = policy_zonelist(gfp_flags, *mpol);
1587 if ((*mpol)->mode == MPOL_BIND)
1588 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001589 }
1590 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001591}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001592
1593/*
1594 * init_nodemask_of_mempolicy
1595 *
1596 * If the current task's mempolicy is "default" [NULL], return 'false'
1597 * to indicate default policy. Otherwise, extract the policy nodemask
1598 * for 'bind' or 'interleave' policy into the argument nodemask, or
1599 * initialize the argument nodemask to contain the single node for
1600 * 'preferred' or 'local' policy and return 'true' to indicate presence
1601 * of non-default mempolicy.
1602 *
1603 * We don't bother with reference counting the mempolicy [mpol_get/put]
1604 * because the current task is examining it's own mempolicy and a task's
1605 * mempolicy is only ever changed by the task itself.
1606 *
1607 * N.B., it is the caller's responsibility to free a returned nodemask.
1608 */
1609bool init_nodemask_of_mempolicy(nodemask_t *mask)
1610{
1611 struct mempolicy *mempolicy;
1612 int nid;
1613
1614 if (!(mask && current->mempolicy))
1615 return false;
1616
1617 mempolicy = current->mempolicy;
1618 switch (mempolicy->mode) {
1619 case MPOL_PREFERRED:
1620 if (mempolicy->flags & MPOL_F_LOCAL)
1621 nid = numa_node_id();
1622 else
1623 nid = mempolicy->v.preferred_node;
1624 init_nodemask_of_node(mask, nid);
1625 break;
1626
1627 case MPOL_BIND:
1628 /* Fall through */
1629 case MPOL_INTERLEAVE:
1630 *mask = mempolicy->v.nodes;
1631 break;
1632
1633 default:
1634 BUG();
1635 }
1636
1637 return true;
1638}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001639#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001640
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641/* Allocate a page in interleaved policy.
1642 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001643static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1644 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645{
1646 struct zonelist *zl;
1647 struct page *page;
1648
Mel Gorman0e884602008-04-28 02:12:14 -07001649 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001651 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001652 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 return page;
1654}
1655
1656/**
1657 * alloc_page_vma - Allocate a page for a VMA.
1658 *
1659 * @gfp:
1660 * %GFP_USER user allocation.
1661 * %GFP_KERNEL kernel allocations,
1662 * %GFP_HIGHMEM highmem/user allocations,
1663 * %GFP_FS allocation should not call back into a file system.
1664 * %GFP_ATOMIC don't sleep.
1665 *
1666 * @vma: Pointer to VMA or NULL if not available.
1667 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1668 *
1669 * This function allocates a page from the kernel page pool and applies
1670 * a NUMA policy associated with the VMA or the current process.
1671 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1672 * mm_struct of the VMA to prevent it from going away. Should be used for
1673 * all allocations for pages that will be mapped into
1674 * user space. Returns NULL when no page can be allocated.
1675 *
1676 * Should be called with the mm_sem of the vma hold.
1677 */
1678struct page *
Al Virodd0fc662005-10-07 07:46:04 +01001679alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001681 struct mempolicy *pol = get_vma_policy(current, vma, addr);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001682 struct zonelist *zl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001684 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001686
1687 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001688 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 return alloc_page_interleave(gfp, 0, nid);
1690 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001691 zl = policy_zonelist(gfp, pol);
1692 if (unlikely(mpol_needs_cond_ref(pol))) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001693 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001694 * slow path: ref counted shared policy
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001695 */
Mel Gorman19770b32008-04-28 02:12:18 -07001696 struct page *page = __alloc_pages_nodemask(gfp, 0,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001697 zl, policy_nodemask(gfp, pol));
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001698 __mpol_put(pol);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001699 return page;
1700 }
1701 /*
1702 * fast path: default or task policy
1703 */
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001704 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705}
1706
1707/**
1708 * alloc_pages_current - Allocate pages.
1709 *
1710 * @gfp:
1711 * %GFP_USER user allocation,
1712 * %GFP_KERNEL kernel allocation,
1713 * %GFP_HIGHMEM highmem allocation,
1714 * %GFP_FS don't call back into a file system.
1715 * %GFP_ATOMIC don't sleep.
1716 * @order: Power of two of allocation size in pages. 0 is a single page.
1717 *
1718 * Allocate a page from the kernel page pool. When not in
1719 * interrupt context and apply the current process NUMA policy.
1720 * Returns NULL when no page can be allocated.
1721 *
Paul Jacksoncf2a4732006-01-08 01:01:54 -08001722 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 * 1) it's ok to take cpuset_sem (can WAIT), and
1724 * 2) allocating for current task (not interrupt).
1725 */
Al Virodd0fc662005-10-07 07:46:04 +01001726struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727{
1728 struct mempolicy *pol = current->mempolicy;
1729
Christoph Lameter9b819d22006-09-25 23:31:40 -07001730 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 pol = &default_policy;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001732
1733 /*
1734 * No reference counting needed for current->mempolicy
1735 * nor system default_policy
1736 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001737 if (pol->mode == MPOL_INTERLEAVE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
Mel Gorman19770b32008-04-28 02:12:18 -07001739 return __alloc_pages_nodemask(gfp, order,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001740 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741}
1742EXPORT_SYMBOL(alloc_pages_current);
1743
Paul Jackson42253992006-01-08 01:01:59 -08001744/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001745 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08001746 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1747 * with the mems_allowed returned by cpuset_mems_allowed(). This
1748 * keeps mempolicies cpuset relative after its cpuset moves. See
1749 * further kernel/cpuset.c update_nodemask().
1750 */
Paul Jackson42253992006-01-08 01:01:59 -08001751
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001752/* Slow path of a mempolicy duplicate */
1753struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754{
1755 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1756
1757 if (!new)
1758 return ERR_PTR(-ENOMEM);
Paul Jackson42253992006-01-08 01:01:59 -08001759 if (current_cpuset_is_being_rebound()) {
1760 nodemask_t mems = cpuset_mems_allowed(current);
1761 mpol_rebind_policy(old, &mems);
1762 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 *new = *old;
1764 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 return new;
1766}
1767
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001768/*
1769 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1770 * eliminate the * MPOL_F_* flags that require conditional ref and
1771 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1772 * after return. Use the returned value.
1773 *
1774 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1775 * policy lookup, even if the policy needs/has extra ref on lookup.
1776 * shmem_readahead needs this.
1777 */
1778struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1779 struct mempolicy *frompol)
1780{
1781 if (!mpol_needs_cond_ref(frompol))
1782 return frompol;
1783
1784 *tompol = *frompol;
1785 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1786 __mpol_put(frompol);
1787 return tompol;
1788}
1789
David Rientjesf5b087b2008-04-28 02:12:27 -07001790static int mpol_match_intent(const struct mempolicy *a,
1791 const struct mempolicy *b)
1792{
1793 if (a->flags != b->flags)
1794 return 0;
1795 if (!mpol_store_user_nodemask(a))
1796 return 1;
1797 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1798}
1799
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800/* Slow path of a mempolicy comparison */
1801int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1802{
1803 if (!a || !b)
1804 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001805 if (a->mode != b->mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001807 if (a->mode != MPOL_DEFAULT && !mpol_match_intent(a, b))
David Rientjesf5b087b2008-04-28 02:12:27 -07001808 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001809 switch (a->mode) {
Mel Gorman19770b32008-04-28 02:12:18 -07001810 case MPOL_BIND:
1811 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -07001813 return nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814 case MPOL_PREFERRED:
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07001815 return a->v.preferred_node == b->v.preferred_node &&
1816 a->flags == b->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 default:
1818 BUG();
1819 return 0;
1820 }
1821}
1822
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 * Shared memory backing store policy support.
1825 *
1826 * Remember policies even when nobody has shared memory mapped.
1827 * The policies are kept in Red-Black tree linked from the inode.
1828 * They are protected by the sp->lock spinlock, which should be held
1829 * for any accesses to the tree.
1830 */
1831
1832/* lookup first element intersecting start-end */
1833/* Caller holds sp->lock */
1834static struct sp_node *
1835sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1836{
1837 struct rb_node *n = sp->root.rb_node;
1838
1839 while (n) {
1840 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1841
1842 if (start >= p->end)
1843 n = n->rb_right;
1844 else if (end <= p->start)
1845 n = n->rb_left;
1846 else
1847 break;
1848 }
1849 if (!n)
1850 return NULL;
1851 for (;;) {
1852 struct sp_node *w = NULL;
1853 struct rb_node *prev = rb_prev(n);
1854 if (!prev)
1855 break;
1856 w = rb_entry(prev, struct sp_node, nd);
1857 if (w->end <= start)
1858 break;
1859 n = prev;
1860 }
1861 return rb_entry(n, struct sp_node, nd);
1862}
1863
1864/* Insert a new shared policy into the list. */
1865/* Caller holds sp->lock */
1866static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1867{
1868 struct rb_node **p = &sp->root.rb_node;
1869 struct rb_node *parent = NULL;
1870 struct sp_node *nd;
1871
1872 while (*p) {
1873 parent = *p;
1874 nd = rb_entry(parent, struct sp_node, nd);
1875 if (new->start < nd->start)
1876 p = &(*p)->rb_left;
1877 else if (new->end > nd->end)
1878 p = &(*p)->rb_right;
1879 else
1880 BUG();
1881 }
1882 rb_link_node(&new->nd, parent, p);
1883 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07001884 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001885 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886}
1887
1888/* Find shared policy intersecting idx */
1889struct mempolicy *
1890mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1891{
1892 struct mempolicy *pol = NULL;
1893 struct sp_node *sn;
1894
1895 if (!sp->root.rb_node)
1896 return NULL;
1897 spin_lock(&sp->lock);
1898 sn = sp_lookup(sp, idx, idx+1);
1899 if (sn) {
1900 mpol_get(sn->policy);
1901 pol = sn->policy;
1902 }
1903 spin_unlock(&sp->lock);
1904 return pol;
1905}
1906
1907static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1908{
Paul Mundt140d5a42007-07-15 23:38:16 -07001909 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 rb_erase(&n->nd, &sp->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001911 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 kmem_cache_free(sn_cache, n);
1913}
1914
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001915static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1916 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917{
1918 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1919
1920 if (!n)
1921 return NULL;
1922 n->start = start;
1923 n->end = end;
1924 mpol_get(pol);
Lee Schermerhornaab0b102008-04-28 02:13:13 -07001925 pol->flags |= MPOL_F_SHARED; /* for unref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 n->policy = pol;
1927 return n;
1928}
1929
1930/* Replace a policy range. */
1931static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1932 unsigned long end, struct sp_node *new)
1933{
1934 struct sp_node *n, *new2 = NULL;
1935
1936restart:
1937 spin_lock(&sp->lock);
1938 n = sp_lookup(sp, start, end);
1939 /* Take care of old policies in the same range. */
1940 while (n && n->start < end) {
1941 struct rb_node *next = rb_next(&n->nd);
1942 if (n->start >= start) {
1943 if (n->end <= end)
1944 sp_delete(sp, n);
1945 else
1946 n->start = end;
1947 } else {
1948 /* Old policy spanning whole new range. */
1949 if (n->end > end) {
1950 if (!new2) {
1951 spin_unlock(&sp->lock);
1952 new2 = sp_alloc(end, n->end, n->policy);
1953 if (!new2)
1954 return -ENOMEM;
1955 goto restart;
1956 }
1957 n->end = start;
1958 sp_insert(sp, new2);
1959 new2 = NULL;
1960 break;
1961 } else
1962 n->end = start;
1963 }
1964 if (!next)
1965 break;
1966 n = rb_entry(next, struct sp_node, nd);
1967 }
1968 if (new)
1969 sp_insert(sp, new);
1970 spin_unlock(&sp->lock);
1971 if (new2) {
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001972 mpol_put(new2->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 kmem_cache_free(sn_cache, new2);
1974 }
1975 return 0;
1976}
1977
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07001978/**
1979 * mpol_shared_policy_init - initialize shared policy for inode
1980 * @sp: pointer to inode shared policy
1981 * @mpol: struct mempolicy to install
1982 *
1983 * Install non-NULL @mpol in inode's shared policy rb-tree.
1984 * On entry, the current task has a reference on a non-NULL @mpol.
1985 * This must be released on exit.
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001986 * This is called at get_inode() calls and we can use GFP_KERNEL.
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07001987 */
1988void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
Robin Holt7339ff82006-01-14 13:20:48 -08001989{
Miao Xie58568d22009-06-16 15:31:49 -07001990 int ret;
1991
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07001992 sp->root = RB_ROOT; /* empty tree == default mempolicy */
1993 spin_lock_init(&sp->lock);
Robin Holt7339ff82006-01-14 13:20:48 -08001994
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07001995 if (mpol) {
1996 struct vm_area_struct pvma;
1997 struct mempolicy *new;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07001998 NODEMASK_SCRATCH(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08001999
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002000 if (!scratch)
2001 return;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002002 /* contextualize the tmpfs mount point mempolicy */
2003 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
Miao Xie58568d22009-06-16 15:31:49 -07002004 if (IS_ERR(new)) {
2005 mpol_put(mpol); /* drop our ref on sb mpol */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002006 NODEMASK_SCRATCH_FREE(scratch);
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002007 return; /* no valid nodemask intersection */
Miao Xie58568d22009-06-16 15:31:49 -07002008 }
2009
2010 task_lock(current);
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002011 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002012 task_unlock(current);
2013 mpol_put(mpol); /* drop our ref on sb mpol */
2014 if (ret) {
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002015 NODEMASK_SCRATCH_FREE(scratch);
Miao Xie58568d22009-06-16 15:31:49 -07002016 mpol_put(new);
2017 return;
2018 }
Robin Holt7339ff82006-01-14 13:20:48 -08002019
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002020 /* Create pseudo-vma that contains just the policy */
2021 memset(&pvma, 0, sizeof(struct vm_area_struct));
2022 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2023 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2024 mpol_put(new); /* drop initial ref */
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002025 NODEMASK_SCRATCH_FREE(scratch);
Robin Holt7339ff82006-01-14 13:20:48 -08002026 }
2027}
2028
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029int mpol_set_shared_policy(struct shared_policy *info,
2030 struct vm_area_struct *vma, struct mempolicy *npol)
2031{
2032 int err;
2033 struct sp_node *new = NULL;
2034 unsigned long sz = vma_pages(vma);
2035
David Rientjes028fec42008-04-28 02:12:25 -07002036 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07002038 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07002039 npol ? npol->flags : -1,
Paul Mundt140d5a42007-07-15 23:38:16 -07002040 npol ? nodes_addr(npol->v.nodes)[0] : -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041
2042 if (npol) {
2043 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2044 if (!new)
2045 return -ENOMEM;
2046 }
2047 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2048 if (err && new)
2049 kmem_cache_free(sn_cache, new);
2050 return err;
2051}
2052
2053/* Free a backing policy store on inode delete. */
2054void mpol_free_shared_policy(struct shared_policy *p)
2055{
2056 struct sp_node *n;
2057 struct rb_node *next;
2058
2059 if (!p->root.rb_node)
2060 return;
2061 spin_lock(&p->lock);
2062 next = rb_first(&p->root);
2063 while (next) {
2064 n = rb_entry(next, struct sp_node, nd);
2065 next = rb_next(&n->nd);
Andi Kleen90c50292005-07-27 11:43:50 -07002066 rb_erase(&n->nd, &p->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07002067 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 kmem_cache_free(sn_cache, n);
2069 }
2070 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071}
2072
2073/* assumes fs == KERNEL_DS */
2074void __init numa_policy_init(void)
2075{
Paul Mundtb71636e2007-07-15 23:38:15 -07002076 nodemask_t interleave_nodes;
2077 unsigned long largest = 0;
2078 int nid, prefer = 0;
2079
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 policy_cache = kmem_cache_create("numa_policy",
2081 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09002082 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083
2084 sn_cache = kmem_cache_create("shared_policy_node",
2085 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09002086 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087
Paul Mundtb71636e2007-07-15 23:38:15 -07002088 /*
2089 * Set interleaving policy for system init. Interleaving is only
2090 * enabled across suitably sized nodes (default is >= 16MB), or
2091 * fall back to the largest node if they're all smaller.
2092 */
2093 nodes_clear(interleave_nodes);
Christoph Lameter56bbd652007-10-16 01:25:35 -07002094 for_each_node_state(nid, N_HIGH_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07002095 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096
Paul Mundtb71636e2007-07-15 23:38:15 -07002097 /* Preserve the largest node */
2098 if (largest < total_pages) {
2099 largest = total_pages;
2100 prefer = nid;
2101 }
2102
2103 /* Interleave this node? */
2104 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2105 node_set(nid, interleave_nodes);
2106 }
2107
2108 /* All too small, use the largest */
2109 if (unlikely(nodes_empty(interleave_nodes)))
2110 node_set(prefer, interleave_nodes);
2111
David Rientjes028fec42008-04-28 02:12:25 -07002112 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 printk("numa_policy_init: interleaving failed\n");
2114}
2115
Christoph Lameter8bccd852005-10-29 18:16:59 -07002116/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117void numa_default_policy(void)
2118{
David Rientjes028fec42008-04-28 02:12:25 -07002119 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120}
Paul Jackson68860ec2005-10-30 15:02:36 -08002121
Paul Jackson42253992006-01-08 01:01:59 -08002122/*
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002123 * Parse and format mempolicy from/to strings
2124 */
2125
2126/*
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002127 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002128 * Used only for mpol_parse_str() and mpol_to_str()
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002129 */
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002130#define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
Helge Deller15ad7cd2006-12-06 20:40:36 -08002131static const char * const policy_types[] =
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002132 { "default", "prefer", "bind", "interleave", "local" };
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002133
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002134
2135#ifdef CONFIG_TMPFS
2136/**
2137 * mpol_parse_str - parse string to mempolicy
2138 * @str: string containing mempolicy to parse
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002139 * @mpol: pointer to struct mempolicy pointer, returned on success.
2140 * @no_context: flag whether to "contextualize" the mempolicy
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002141 *
2142 * Format of input:
2143 * <mode>[=<flags>][:<nodelist>]
2144 *
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002145 * if @no_context is true, save the input nodemask in w.user_nodemask in
2146 * the returned mempolicy. This will be used to "clone" the mempolicy in
2147 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2148 * mount option. Note that if 'static' or 'relative' mode flags were
2149 * specified, the input nodemask will already have been saved. Saving
2150 * it again is redundant, but safe.
2151 *
2152 * On success, returns 0, else 1
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002153 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002154int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002155{
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002156 struct mempolicy *new = NULL;
2157 unsigned short uninitialized_var(mode);
2158 unsigned short uninitialized_var(mode_flags);
2159 nodemask_t nodes;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002160 char *nodelist = strchr(str, ':');
2161 char *flags = strchr(str, '=');
2162 int i;
2163 int err = 1;
2164
2165 if (nodelist) {
2166 /* NUL-terminate mode or flags string */
2167 *nodelist++ = '\0';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002168 if (nodelist_parse(nodelist, nodes))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002169 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002170 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002171 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002172 } else
2173 nodes_clear(nodes);
2174
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002175 if (flags)
2176 *flags++ = '\0'; /* terminate mode string */
2177
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002178 for (i = 0; i <= MPOL_LOCAL; i++) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002179 if (!strcmp(str, policy_types[i])) {
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002180 mode = i;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002181 break;
2182 }
2183 }
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002184 if (i > MPOL_LOCAL)
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002185 goto out;
2186
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002187 switch (mode) {
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002188 case MPOL_PREFERRED:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002189 /*
2190 * Insist on a nodelist of one node only
2191 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002192 if (nodelist) {
2193 char *rest = nodelist;
2194 while (isdigit(*rest))
2195 rest++;
2196 if (!*rest)
2197 err = 0;
2198 }
2199 break;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002200 case MPOL_INTERLEAVE:
2201 /*
2202 * Default to online nodes with memory if no nodelist
2203 */
2204 if (!nodelist)
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002205 nodes = node_states[N_HIGH_MEMORY];
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002206 err = 0;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002207 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002208 case MPOL_LOCAL:
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002209 /*
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002210 * Don't allow a nodelist; mpol_new() checks flags
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002211 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002212 if (nodelist)
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002213 goto out;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002214 mode = MPOL_PREFERRED;
Lee Schermerhorn3f226aa2008-04-28 02:13:24 -07002215 break;
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002216
2217 /*
2218 * case MPOL_BIND: mpol_new() enforces non-empty nodemask.
2219 * case MPOL_DEFAULT: mpol_new() enforces empty nodemask, ignores flags.
2220 */
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002221 }
2222
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002223 mode_flags = 0;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002224 if (flags) {
2225 /*
2226 * Currently, we only support two mutually exclusive
2227 * mode flags.
2228 */
2229 if (!strcmp(flags, "static"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002230 mode_flags |= MPOL_F_STATIC_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002231 else if (!strcmp(flags, "relative"))
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002232 mode_flags |= MPOL_F_RELATIVE_NODES;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002233 else
2234 err = 1;
2235 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002236
2237 new = mpol_new(mode, mode_flags, &nodes);
2238 if (IS_ERR(new))
2239 err = 1;
Miao Xie58568d22009-06-16 15:31:49 -07002240 else {
2241 int ret;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002242 NODEMASK_SCRATCH(scratch);
2243 if (scratch) {
2244 task_lock(current);
2245 ret = mpol_set_nodemask(new, &nodes, scratch);
2246 task_unlock(current);
2247 } else
2248 ret = -ENOMEM;
2249 NODEMASK_SCRATCH_FREE(scratch);
2250 if (ret) {
Miao Xie58568d22009-06-16 15:31:49 -07002251 err = 1;
KAMEZAWA Hiroyuki4bfc4492009-08-06 15:07:33 -07002252 mpol_put(new);
2253 } else if (no_context) {
Miao Xie58568d22009-06-16 15:31:49 -07002254 /* save for contextualization */
2255 new->w.user_nodemask = nodes;
2256 }
2257 }
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002258
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002259out:
2260 /* Restore string for error message */
2261 if (nodelist)
2262 *--nodelist = ':';
2263 if (flags)
2264 *--flags = '=';
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002265 if (!err)
2266 *mpol = new;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002267 return err;
2268}
2269#endif /* CONFIG_TMPFS */
2270
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002271/**
2272 * mpol_to_str - format a mempolicy structure for printing
2273 * @buffer: to contain formatted mempolicy string
2274 * @maxlen: length of @buffer
2275 * @pol: pointer to mempolicy to be formatted
2276 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2277 *
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002278 * Convert a mempolicy into a string.
2279 * Returns the number of characters in buffer (if positive)
2280 * or an error (negative)
2281 */
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002282int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002283{
2284 char *p = buffer;
2285 int l;
2286 nodemask_t nodes;
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002287 unsigned short mode;
David Rientjesf5b087b2008-04-28 02:12:27 -07002288 unsigned short flags = pol ? pol->flags : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002289
Lee Schermerhorn22919902008-04-28 02:13:22 -07002290 /*
2291 * Sanity check: room for longest mode, flag and some nodes
2292 */
2293 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2294
Lee Schermerhornbea904d2008-04-28 02:13:18 -07002295 if (!pol || pol == &default_policy)
2296 mode = MPOL_DEFAULT;
2297 else
2298 mode = pol->mode;
2299
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002300 switch (mode) {
2301 case MPOL_DEFAULT:
2302 nodes_clear(nodes);
2303 break;
2304
2305 case MPOL_PREFERRED:
2306 nodes_clear(nodes);
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002307 if (flags & MPOL_F_LOCAL)
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002308 mode = MPOL_LOCAL; /* pseudo-policy */
2309 else
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002310 node_set(pol->v.preferred_node, nodes);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002311 break;
2312
2313 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07002314 /* Fall through */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002315 case MPOL_INTERLEAVE:
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002316 if (no_context)
2317 nodes = pol->w.user_nodemask;
2318 else
2319 nodes = pol->v.nodes;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002320 break;
2321
2322 default:
2323 BUG();
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002324 }
2325
2326 l = strlen(policy_types[mode]);
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002327 if (buffer + maxlen < p + l + 1)
2328 return -ENOSPC;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002329
2330 strcpy(p, policy_types[mode]);
2331 p += l;
2332
Lee Schermerhornfc36b8d2008-04-28 02:13:21 -07002333 if (flags & MPOL_MODE_FLAGS) {
David Rientjesf5b087b2008-04-28 02:12:27 -07002334 if (buffer + maxlen < p + 2)
2335 return -ENOSPC;
2336 *p++ = '=';
2337
Lee Schermerhorn22919902008-04-28 02:13:22 -07002338 /*
2339 * Currently, the only defined flags are mutually exclusive
2340 */
David Rientjesf5b087b2008-04-28 02:12:27 -07002341 if (flags & MPOL_F_STATIC_NODES)
Lee Schermerhorn22919902008-04-28 02:13:22 -07002342 p += snprintf(p, buffer + maxlen - p, "static");
2343 else if (flags & MPOL_F_RELATIVE_NODES)
2344 p += snprintf(p, buffer + maxlen - p, "relative");
David Rientjesf5b087b2008-04-28 02:12:27 -07002345 }
2346
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002347 if (!nodes_empty(nodes)) {
2348 if (buffer + maxlen < p + 2)
2349 return -ENOSPC;
Lee Schermerhorn095f1fc2008-04-28 02:13:23 -07002350 *p++ = ':';
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002351 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2352 }
2353 return p - buffer;
2354}
2355
2356struct numa_maps {
2357 unsigned long pages;
2358 unsigned long anon;
Christoph Lameter397874d2006-03-06 15:42:53 -08002359 unsigned long active;
2360 unsigned long writeback;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002361 unsigned long mapcount_max;
Christoph Lameter397874d2006-03-06 15:42:53 -08002362 unsigned long dirty;
2363 unsigned long swapcache;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002364 unsigned long node[MAX_NUMNODES];
2365};
2366
Christoph Lameter397874d2006-03-06 15:42:53 -08002367static void gather_stats(struct page *page, void *private, int pte_dirty)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002368{
2369 struct numa_maps *md = private;
2370 int count = page_mapcount(page);
2371
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002372 md->pages++;
Christoph Lameter397874d2006-03-06 15:42:53 -08002373 if (pte_dirty || PageDirty(page))
2374 md->dirty++;
2375
2376 if (PageSwapCache(page))
2377 md->swapcache++;
2378
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002379 if (PageActive(page) || PageUnevictable(page))
Christoph Lameter397874d2006-03-06 15:42:53 -08002380 md->active++;
2381
2382 if (PageWriteback(page))
2383 md->writeback++;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002384
2385 if (PageAnon(page))
2386 md->anon++;
2387
Christoph Lameter397874d2006-03-06 15:42:53 -08002388 if (count > md->mapcount_max)
2389 md->mapcount_max = count;
2390
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002391 md->node[page_to_nid(page)]++;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002392}
2393
Andrew Morton7f709ed2006-03-07 21:55:22 -08002394#ifdef CONFIG_HUGETLB_PAGE
Christoph Lameter397874d2006-03-06 15:42:53 -08002395static void check_huge_range(struct vm_area_struct *vma,
2396 unsigned long start, unsigned long end,
2397 struct numa_maps *md)
2398{
2399 unsigned long addr;
2400 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002401 struct hstate *h = hstate_vma(vma);
2402 unsigned long sz = huge_page_size(h);
Christoph Lameter397874d2006-03-06 15:42:53 -08002403
Andi Kleena5516432008-07-23 21:27:41 -07002404 for (addr = start; addr < end; addr += sz) {
2405 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2406 addr & huge_page_mask(h));
Christoph Lameter397874d2006-03-06 15:42:53 -08002407 pte_t pte;
2408
2409 if (!ptep)
2410 continue;
2411
2412 pte = *ptep;
2413 if (pte_none(pte))
2414 continue;
2415
2416 page = pte_page(pte);
2417 if (!page)
2418 continue;
2419
2420 gather_stats(page, md, pte_dirty(*ptep));
2421 }
2422}
Andrew Morton7f709ed2006-03-07 21:55:22 -08002423#else
2424static inline void check_huge_range(struct vm_area_struct *vma,
2425 unsigned long start, unsigned long end,
2426 struct numa_maps *md)
2427{
2428}
2429#endif
Christoph Lameter397874d2006-03-06 15:42:53 -08002430
Lee Schermerhorn53f25562008-04-28 02:13:20 -07002431/*
2432 * Display pages allocated per node and memory policy via /proc.
2433 */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002434int show_numa_map(struct seq_file *m, void *v)
2435{
Eric W. Biederman99f89552006-06-26 00:25:55 -07002436 struct proc_maps_private *priv = m->private;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002437 struct vm_area_struct *vma = v;
2438 struct numa_maps *md;
Christoph Lameter397874d2006-03-06 15:42:53 -08002439 struct file *file = vma->vm_file;
2440 struct mm_struct *mm = vma->vm_mm;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07002441 struct mempolicy *pol;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002442 int n;
2443 char buffer[50];
2444
Christoph Lameter397874d2006-03-06 15:42:53 -08002445 if (!mm)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002446 return 0;
2447
2448 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2449 if (!md)
2450 return 0;
2451
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07002452 pol = get_vma_policy(priv->task, vma, vma->vm_start);
Lee Schermerhorn71fe8042008-04-28 02:13:26 -07002453 mpol_to_str(buffer, sizeof(buffer), pol, 0);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002454 mpol_cond_put(pol);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002455
Christoph Lameter397874d2006-03-06 15:42:53 -08002456 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002457
Christoph Lameter397874d2006-03-06 15:42:53 -08002458 if (file) {
2459 seq_printf(m, " file=");
Jan Blunckc32c2f62008-02-14 19:38:43 -08002460 seq_path(m, &file->f_path, "\n\t= ");
Christoph Lameter397874d2006-03-06 15:42:53 -08002461 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2462 seq_printf(m, " heap");
2463 } else if (vma->vm_start <= mm->start_stack &&
2464 vma->vm_end >= mm->start_stack) {
2465 seq_printf(m, " stack");
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002466 }
Christoph Lameter397874d2006-03-06 15:42:53 -08002467
2468 if (is_vm_hugetlb_page(vma)) {
2469 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2470 seq_printf(m, " huge");
2471 } else {
2472 check_pgd_range(vma, vma->vm_start, vma->vm_end,
Christoph Lameter56bbd652007-10-16 01:25:35 -07002473 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
Christoph Lameter397874d2006-03-06 15:42:53 -08002474 }
2475
2476 if (!md->pages)
2477 goto out;
2478
2479 if (md->anon)
2480 seq_printf(m," anon=%lu",md->anon);
2481
2482 if (md->dirty)
2483 seq_printf(m," dirty=%lu",md->dirty);
2484
2485 if (md->pages != md->anon && md->pages != md->dirty)
2486 seq_printf(m, " mapped=%lu", md->pages);
2487
2488 if (md->mapcount_max > 1)
2489 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2490
2491 if (md->swapcache)
2492 seq_printf(m," swapcache=%lu", md->swapcache);
2493
2494 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2495 seq_printf(m," active=%lu", md->active);
2496
2497 if (md->writeback)
2498 seq_printf(m," writeback=%lu", md->writeback);
2499
Christoph Lameter56bbd652007-10-16 01:25:35 -07002500 for_each_node_state(n, N_HIGH_MEMORY)
Christoph Lameter397874d2006-03-06 15:42:53 -08002501 if (md->node[n])
2502 seq_printf(m, " N%d=%lu", n, md->node[n]);
2503out:
2504 seq_putc(m, '\n');
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002505 kfree(md);
2506
2507 if (m->count < m->size)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002508 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002509 return 0;
2510}