]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/mempolicy.c
Linux 2.6.31-rc1
[mirror_ubuntu-jammy-kernel.git] / mm / mempolicy.c
CommitLineData
1da177e4
LT
1/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
8bccd85f 5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
1da177e4
LT
6 * 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.
8bccd85f 21 *
1da177e4
LT
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
8bccd85f
CL
24 * 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 *
1da177e4
LT
28 * 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.
8bccd85f 33 *
1da177e4
LT
34 * 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.
1da177e4
LT
66*/
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>
1da177e4
LT
74#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>
b488893a 80#include <linux/nsproxy.h>
1da177e4
LT
81#include <linux/interrupt.h>
82#include <linux/init.h>
83#include <linux/compat.h>
dc9aa5b9 84#include <linux/swap.h>
1a75a6c8
CL
85#include <linux/seq_file.h>
86#include <linux/proc_fs.h>
b20a3503 87#include <linux/migrate.h>
95a402c3 88#include <linux/rmap.h>
86c3a764 89#include <linux/security.h>
dbcb0f19 90#include <linux/syscalls.h>
095f1fc4 91#include <linux/ctype.h>
dc9aa5b9 92
1da177e4
LT
93#include <asm/tlbflush.h>
94#include <asm/uaccess.h>
95
62695a84
NP
96#include "internal.h"
97
38e35860 98/* Internal flags */
dc9aa5b9 99#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
38e35860 100#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
1a75a6c8 101#define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
dc9aa5b9 102
fcc234f8
PE
103static struct kmem_cache *policy_cache;
104static struct kmem_cache *sn_cache;
1da177e4 105
1da177e4
LT
106/* Highest zone. An specific allocation for a zone below that is not
107 policied. */
6267276f 108enum zone_type policy_zone = 0;
1da177e4 109
bea904d5
LS
110/*
111 * run-time system-wide default policy => local allocation
112 */
d42c6997 113struct mempolicy default_policy = {
1da177e4 114 .refcnt = ATOMIC_INIT(1), /* never free it */
bea904d5 115 .mode = MPOL_PREFERRED,
fc36b8d3 116 .flags = MPOL_F_LOCAL,
1da177e4
LT
117};
118
37012946
DR
119static const struct mempolicy_operations {
120 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
121 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
122} mpol_ops[MPOL_MAX];
123
19770b32 124/* Check that the nodemask contains at least one populated zone */
37012946 125static int is_valid_nodemask(const nodemask_t *nodemask)
1da177e4 126{
19770b32 127 int nd, k;
1da177e4 128
19770b32
MG
129 /* Check that there is something useful in this mask */
130 k = policy_zone;
131
132 for_each_node_mask(nd, *nodemask) {
133 struct zone *z;
134
135 for (k = 0; k <= policy_zone; k++) {
136 z = &NODE_DATA(nd)->node_zones[k];
137 if (z->present_pages > 0)
138 return 1;
dd942ae3 139 }
8af5e2eb 140 }
19770b32
MG
141
142 return 0;
1da177e4
LT
143}
144
f5b087b5
DR
145static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
146{
4c50bc01
DR
147 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
148}
149
150static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
151 const nodemask_t *rel)
152{
153 nodemask_t tmp;
154 nodes_fold(tmp, *orig, nodes_weight(*rel));
155 nodes_onto(*ret, tmp, *rel);
f5b087b5
DR
156}
157
37012946
DR
158static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
159{
160 if (nodes_empty(*nodes))
161 return -EINVAL;
162 pol->v.nodes = *nodes;
163 return 0;
164}
165
166static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
167{
168 if (!nodes)
fc36b8d3 169 pol->flags |= MPOL_F_LOCAL; /* local allocation */
37012946
DR
170 else if (nodes_empty(*nodes))
171 return -EINVAL; /* no allowed nodes */
172 else
173 pol->v.preferred_node = first_node(*nodes);
174 return 0;
175}
176
177static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
178{
179 if (!is_valid_nodemask(nodes))
180 return -EINVAL;
181 pol->v.nodes = *nodes;
182 return 0;
183}
184
58568d2a
MX
185/*
186 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
187 * any, for the new policy. mpol_new() has already validated the nodes
188 * parameter with respect to the policy mode and flags. But, we need to
189 * handle an empty nodemask with MPOL_PREFERRED here.
190 *
191 * Must be called holding task's alloc_lock to protect task's mems_allowed
192 * and mempolicy. May also be called holding the mmap_semaphore for write.
193 */
194static int mpol_set_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
195{
196 nodemask_t cpuset_context_nmask;
197 int ret;
198
199 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
200 if (pol == NULL)
201 return 0;
202
203 VM_BUG_ON(!nodes);
204 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
205 nodes = NULL; /* explicit local allocation */
206 else {
207 if (pol->flags & MPOL_F_RELATIVE_NODES)
208 mpol_relative_nodemask(&cpuset_context_nmask, nodes,
209 &cpuset_current_mems_allowed);
210 else
211 nodes_and(cpuset_context_nmask, *nodes,
212 cpuset_current_mems_allowed);
213 if (mpol_store_user_nodemask(pol))
214 pol->w.user_nodemask = *nodes;
215 else
216 pol->w.cpuset_mems_allowed =
217 cpuset_current_mems_allowed;
218 }
219
220 ret = mpol_ops[pol->mode].create(pol,
221 nodes ? &cpuset_context_nmask : NULL);
222 return ret;
223}
224
225/*
226 * This function just creates a new policy, does some check and simple
227 * initialization. You must invoke mpol_set_nodemask() to set nodes.
228 */
028fec41
DR
229static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
230 nodemask_t *nodes)
1da177e4
LT
231{
232 struct mempolicy *policy;
233
028fec41
DR
234 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
235 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
140d5a49 236
3e1f0645
DR
237 if (mode == MPOL_DEFAULT) {
238 if (nodes && !nodes_empty(*nodes))
37012946 239 return ERR_PTR(-EINVAL);
bea904d5 240 return NULL; /* simply delete any existing policy */
37012946 241 }
3e1f0645
DR
242 VM_BUG_ON(!nodes);
243
244 /*
245 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
246 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
247 * All other modes require a valid pointer to a non-empty nodemask.
248 */
249 if (mode == MPOL_PREFERRED) {
250 if (nodes_empty(*nodes)) {
251 if (((flags & MPOL_F_STATIC_NODES) ||
252 (flags & MPOL_F_RELATIVE_NODES)))
253 return ERR_PTR(-EINVAL);
3e1f0645
DR
254 }
255 } else if (nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
1da177e4
LT
257 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
258 if (!policy)
259 return ERR_PTR(-ENOMEM);
260 atomic_set(&policy->refcnt, 1);
45c4745a 261 policy->mode = mode;
3e1f0645 262 policy->flags = flags;
37012946 263
1da177e4 264 return policy;
37012946
DR
265}
266
52cd3b07
LS
267/* Slow path of a mpol destructor. */
268void __mpol_put(struct mempolicy *p)
269{
270 if (!atomic_dec_and_test(&p->refcnt))
271 return;
52cd3b07
LS
272 kmem_cache_free(policy_cache, p);
273}
274
37012946
DR
275static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
276{
277}
278
279static void mpol_rebind_nodemask(struct mempolicy *pol,
280 const nodemask_t *nodes)
281{
282 nodemask_t tmp;
283
284 if (pol->flags & MPOL_F_STATIC_NODES)
285 nodes_and(tmp, pol->w.user_nodemask, *nodes);
286 else if (pol->flags & MPOL_F_RELATIVE_NODES)
287 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
288 else {
289 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
290 *nodes);
291 pol->w.cpuset_mems_allowed = *nodes;
292 }
f5b087b5 293
37012946
DR
294 pol->v.nodes = tmp;
295 if (!node_isset(current->il_next, tmp)) {
296 current->il_next = next_node(current->il_next, tmp);
297 if (current->il_next >= MAX_NUMNODES)
298 current->il_next = first_node(tmp);
299 if (current->il_next >= MAX_NUMNODES)
300 current->il_next = numa_node_id();
301 }
302}
303
304static void mpol_rebind_preferred(struct mempolicy *pol,
305 const nodemask_t *nodes)
306{
307 nodemask_t tmp;
308
37012946
DR
309 if (pol->flags & MPOL_F_STATIC_NODES) {
310 int node = first_node(pol->w.user_nodemask);
311
fc36b8d3 312 if (node_isset(node, *nodes)) {
37012946 313 pol->v.preferred_node = node;
fc36b8d3
LS
314 pol->flags &= ~MPOL_F_LOCAL;
315 } else
316 pol->flags |= MPOL_F_LOCAL;
37012946
DR
317 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
318 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
319 pol->v.preferred_node = first_node(tmp);
fc36b8d3 320 } else if (!(pol->flags & MPOL_F_LOCAL)) {
37012946
DR
321 pol->v.preferred_node = node_remap(pol->v.preferred_node,
322 pol->w.cpuset_mems_allowed,
323 *nodes);
324 pol->w.cpuset_mems_allowed = *nodes;
325 }
1da177e4
LT
326}
327
1d0d2680
DR
328/* Migrate a policy to a different set of nodes */
329static void mpol_rebind_policy(struct mempolicy *pol,
330 const nodemask_t *newmask)
331{
1d0d2680
DR
332 if (!pol)
333 return;
1d0d2680
DR
334 if (!mpol_store_user_nodemask(pol) &&
335 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
336 return;
45c4745a 337 mpol_ops[pol->mode].rebind(pol, newmask);
1d0d2680
DR
338}
339
340/*
341 * Wrapper for mpol_rebind_policy() that just requires task
342 * pointer, and updates task mempolicy.
58568d2a
MX
343 *
344 * Called with task's alloc_lock held.
1d0d2680
DR
345 */
346
347void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
348{
349 mpol_rebind_policy(tsk->mempolicy, new);
350}
351
352/*
353 * Rebind each vma in mm to new nodemask.
354 *
355 * Call holding a reference to mm. Takes mm->mmap_sem during call.
356 */
357
358void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
359{
360 struct vm_area_struct *vma;
361
362 down_write(&mm->mmap_sem);
363 for (vma = mm->mmap; vma; vma = vma->vm_next)
364 mpol_rebind_policy(vma->vm_policy, new);
365 up_write(&mm->mmap_sem);
366}
367
37012946
DR
368static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
369 [MPOL_DEFAULT] = {
370 .rebind = mpol_rebind_default,
371 },
372 [MPOL_INTERLEAVE] = {
373 .create = mpol_new_interleave,
374 .rebind = mpol_rebind_nodemask,
375 },
376 [MPOL_PREFERRED] = {
377 .create = mpol_new_preferred,
378 .rebind = mpol_rebind_preferred,
379 },
380 [MPOL_BIND] = {
381 .create = mpol_new_bind,
382 .rebind = mpol_rebind_nodemask,
383 },
384};
385
397874df 386static void gather_stats(struct page *, void *, int pte_dirty);
fc301289
CL
387static void migrate_page_add(struct page *page, struct list_head *pagelist,
388 unsigned long flags);
1a75a6c8 389
38e35860 390/* Scan through pages checking if pages follow certain conditions. */
b5810039 391static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
dc9aa5b9
CL
392 unsigned long addr, unsigned long end,
393 const nodemask_t *nodes, unsigned long flags,
38e35860 394 void *private)
1da177e4 395{
91612e0d
HD
396 pte_t *orig_pte;
397 pte_t *pte;
705e87c0 398 spinlock_t *ptl;
941150a3 399
705e87c0 400 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
91612e0d 401 do {
6aab341e 402 struct page *page;
25ba77c1 403 int nid;
91612e0d
HD
404
405 if (!pte_present(*pte))
1da177e4 406 continue;
6aab341e
LT
407 page = vm_normal_page(vma, addr, *pte);
408 if (!page)
1da177e4 409 continue;
053837fc
NP
410 /*
411 * The check for PageReserved here is important to avoid
412 * handling zero pages and other pages that may have been
413 * marked special by the system.
414 *
415 * If the PageReserved would not be checked here then f.e.
416 * the location of the zero page could have an influence
417 * on MPOL_MF_STRICT, zero pages would be counted for
418 * the per node stats, and there would be useless attempts
419 * to put zero pages on the migration list.
420 */
f4598c8b
CL
421 if (PageReserved(page))
422 continue;
6aab341e 423 nid = page_to_nid(page);
38e35860
CL
424 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
425 continue;
426
1a75a6c8 427 if (flags & MPOL_MF_STATS)
397874df 428 gather_stats(page, private, pte_dirty(*pte));
053837fc 429 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
fc301289 430 migrate_page_add(page, private, flags);
38e35860
CL
431 else
432 break;
91612e0d 433 } while (pte++, addr += PAGE_SIZE, addr != end);
705e87c0 434 pte_unmap_unlock(orig_pte, ptl);
91612e0d
HD
435 return addr != end;
436}
437
b5810039 438static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
dc9aa5b9
CL
439 unsigned long addr, unsigned long end,
440 const nodemask_t *nodes, unsigned long flags,
38e35860 441 void *private)
91612e0d
HD
442{
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;
dc9aa5b9 451 if (check_pte_range(vma, pmd, addr, next, nodes,
38e35860 452 flags, private))
91612e0d
HD
453 return -EIO;
454 } while (pmd++, addr = next, addr != end);
455 return 0;
456}
457
b5810039 458static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
dc9aa5b9
CL
459 unsigned long addr, unsigned long end,
460 const nodemask_t *nodes, unsigned long flags,
38e35860 461 void *private)
91612e0d
HD
462{
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;
dc9aa5b9 471 if (check_pmd_range(vma, pud, addr, next, nodes,
38e35860 472 flags, private))
91612e0d
HD
473 return -EIO;
474 } while (pud++, addr = next, addr != end);
475 return 0;
476}
477
b5810039 478static inline int check_pgd_range(struct vm_area_struct *vma,
dc9aa5b9
CL
479 unsigned long addr, unsigned long end,
480 const nodemask_t *nodes, unsigned long flags,
38e35860 481 void *private)
91612e0d
HD
482{
483 pgd_t *pgd;
484 unsigned long next;
485
b5810039 486 pgd = pgd_offset(vma->vm_mm, addr);
91612e0d
HD
487 do {
488 next = pgd_addr_end(addr, end);
489 if (pgd_none_or_clear_bad(pgd))
490 continue;
dc9aa5b9 491 if (check_pud_range(vma, pgd, addr, next, nodes,
38e35860 492 flags, private))
91612e0d
HD
493 return -EIO;
494 } while (pgd++, addr = next, addr != end);
495 return 0;
1da177e4
LT
496}
497
dc9aa5b9
CL
498/*
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 */
1da177e4
LT
503static struct vm_area_struct *
504check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
38e35860 505 const nodemask_t *nodes, unsigned long flags, void *private)
1da177e4
LT
506{
507 int err;
508 struct vm_area_struct *first, *vma, *prev;
509
053837fc 510
1da177e4
LT
511 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) {
dc9aa5b9
CL
516 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)))) {
5b952b3c 526 unsigned long endvma = vma->vm_end;
dc9aa5b9 527
5b952b3c
AK
528 if (endvma > end)
529 endvma = end;
530 if (vma->vm_start > start)
531 start = vma->vm_start;
dc9aa5b9 532 err = check_pgd_range(vma, start, endvma, nodes,
38e35860 533 flags, private);
1da177e4
LT
534 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
140d5a49 550 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
1da177e4
LT
551 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;
f0be3d32 560 mpol_put(old);
1da177e4
LT
561 }
562 return err;
563}
564
565/* Step 2: apply policy to a range and do splits. */
566static int mbind_range(struct vm_area_struct *vma, unsigned long start,
567 unsigned long end, struct mempolicy *new)
568{
569 struct vm_area_struct *next;
570 int err;
571
572 err = 0;
573 for (; vma && vma->vm_start < end; vma = next) {
574 next = vma->vm_next;
575 if (vma->vm_start < start)
576 err = split_vma(vma->vm_mm, vma, start, 1);
577 if (!err && vma->vm_end > end)
578 err = split_vma(vma->vm_mm, vma, end, 0);
579 if (!err)
580 err = policy_vma(vma, new);
581 if (err)
582 break;
583 }
584 return err;
585}
586
c61afb18
PJ
587/*
588 * Update task->flags PF_MEMPOLICY bit: set iff non-default
589 * mempolicy. Allows more rapid checking of this (combined perhaps
590 * with other PF_* flag bits) on memory allocation hot code paths.
591 *
592 * If called from outside this file, the task 'p' should -only- be
593 * a newly forked child not yet visible on the task list, because
594 * manipulating the task flags of a visible task is not safe.
595 *
596 * The above limitation is why this routine has the funny name
597 * mpol_fix_fork_child_flag().
598 *
599 * It is also safe to call this with a task pointer of current,
600 * which the static wrapper mpol_set_task_struct_flag() does,
601 * for use within this file.
602 */
603
604void mpol_fix_fork_child_flag(struct task_struct *p)
605{
606 if (p->mempolicy)
607 p->flags |= PF_MEMPOLICY;
608 else
609 p->flags &= ~PF_MEMPOLICY;
610}
611
612static void mpol_set_task_struct_flag(void)
613{
614 mpol_fix_fork_child_flag(current);
615}
616
1da177e4 617/* Set the process memory policy */
028fec41
DR
618static long do_set_mempolicy(unsigned short mode, unsigned short flags,
619 nodemask_t *nodes)
1da177e4 620{
58568d2a 621 struct mempolicy *new, *old;
f4e53d91 622 struct mm_struct *mm = current->mm;
58568d2a 623 int ret;
1da177e4 624
028fec41 625 new = mpol_new(mode, flags, nodes);
1da177e4
LT
626 if (IS_ERR(new))
627 return PTR_ERR(new);
f4e53d91
LS
628
629 /*
630 * prevent changing our mempolicy while show_numa_maps()
631 * is using it.
632 * Note: do_set_mempolicy() can be called at init time
633 * with no 'mm'.
634 */
635 if (mm)
636 down_write(&mm->mmap_sem);
58568d2a
MX
637 task_lock(current);
638 ret = mpol_set_nodemask(new, nodes);
639 if (ret) {
640 task_unlock(current);
641 if (mm)
642 up_write(&mm->mmap_sem);
643 mpol_put(new);
644 return ret;
645 }
646 old = current->mempolicy;
1da177e4 647 current->mempolicy = new;
c61afb18 648 mpol_set_task_struct_flag();
45c4745a 649 if (new && new->mode == MPOL_INTERLEAVE &&
f5b087b5 650 nodes_weight(new->v.nodes))
dfcd3c0d 651 current->il_next = first_node(new->v.nodes);
58568d2a 652 task_unlock(current);
f4e53d91
LS
653 if (mm)
654 up_write(&mm->mmap_sem);
655
58568d2a 656 mpol_put(old);
1da177e4
LT
657 return 0;
658}
659
bea904d5
LS
660/*
661 * Return nodemask for policy for get_mempolicy() query
58568d2a
MX
662 *
663 * Called with task's alloc_lock held
bea904d5
LS
664 */
665static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
1da177e4 666{
dfcd3c0d 667 nodes_clear(*nodes);
bea904d5
LS
668 if (p == &default_policy)
669 return;
670
45c4745a 671 switch (p->mode) {
19770b32
MG
672 case MPOL_BIND:
673 /* Fall through */
1da177e4 674 case MPOL_INTERLEAVE:
dfcd3c0d 675 *nodes = p->v.nodes;
1da177e4
LT
676 break;
677 case MPOL_PREFERRED:
fc36b8d3 678 if (!(p->flags & MPOL_F_LOCAL))
dfcd3c0d 679 node_set(p->v.preferred_node, *nodes);
53f2556b 680 /* else return empty node mask for local allocation */
1da177e4
LT
681 break;
682 default:
683 BUG();
684 }
685}
686
687static int lookup_node(struct mm_struct *mm, unsigned long addr)
688{
689 struct page *p;
690 int err;
691
692 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
693 if (err >= 0) {
694 err = page_to_nid(p);
695 put_page(p);
696 }
697 return err;
698}
699
1da177e4 700/* Retrieve NUMA policy */
dbcb0f19
AB
701static long do_get_mempolicy(int *policy, nodemask_t *nmask,
702 unsigned long addr, unsigned long flags)
1da177e4 703{
8bccd85f 704 int err;
1da177e4
LT
705 struct mm_struct *mm = current->mm;
706 struct vm_area_struct *vma = NULL;
707 struct mempolicy *pol = current->mempolicy;
708
754af6f5
LS
709 if (flags &
710 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
1da177e4 711 return -EINVAL;
754af6f5
LS
712
713 if (flags & MPOL_F_MEMS_ALLOWED) {
714 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
715 return -EINVAL;
716 *policy = 0; /* just so it's initialized */
58568d2a 717 task_lock(current);
754af6f5 718 *nmask = cpuset_current_mems_allowed;
58568d2a 719 task_unlock(current);
754af6f5
LS
720 return 0;
721 }
722
1da177e4 723 if (flags & MPOL_F_ADDR) {
bea904d5
LS
724 /*
725 * Do NOT fall back to task policy if the
726 * vma/shared policy at addr is NULL. We
727 * want to return MPOL_DEFAULT in this case.
728 */
1da177e4
LT
729 down_read(&mm->mmap_sem);
730 vma = find_vma_intersection(mm, addr, addr+1);
731 if (!vma) {
732 up_read(&mm->mmap_sem);
733 return -EFAULT;
734 }
735 if (vma->vm_ops && vma->vm_ops->get_policy)
736 pol = vma->vm_ops->get_policy(vma, addr);
737 else
738 pol = vma->vm_policy;
739 } else if (addr)
740 return -EINVAL;
741
742 if (!pol)
bea904d5 743 pol = &default_policy; /* indicates default behavior */
1da177e4
LT
744
745 if (flags & MPOL_F_NODE) {
746 if (flags & MPOL_F_ADDR) {
747 err = lookup_node(mm, addr);
748 if (err < 0)
749 goto out;
8bccd85f 750 *policy = err;
1da177e4 751 } else if (pol == current->mempolicy &&
45c4745a 752 pol->mode == MPOL_INTERLEAVE) {
8bccd85f 753 *policy = current->il_next;
1da177e4
LT
754 } else {
755 err = -EINVAL;
756 goto out;
757 }
bea904d5
LS
758 } else {
759 *policy = pol == &default_policy ? MPOL_DEFAULT :
760 pol->mode;
d79df630
DR
761 /*
762 * Internal mempolicy flags must be masked off before exposing
763 * the policy to userspace.
764 */
765 *policy |= (pol->flags & MPOL_MODE_FLAGS);
bea904d5 766 }
1da177e4
LT
767
768 if (vma) {
769 up_read(&current->mm->mmap_sem);
770 vma = NULL;
771 }
772
1da177e4 773 err = 0;
58568d2a
MX
774 if (nmask) {
775 task_lock(current);
bea904d5 776 get_policy_nodemask(pol, nmask);
58568d2a
MX
777 task_unlock(current);
778 }
1da177e4
LT
779
780 out:
52cd3b07 781 mpol_cond_put(pol);
1da177e4
LT
782 if (vma)
783 up_read(&current->mm->mmap_sem);
784 return err;
785}
786
b20a3503 787#ifdef CONFIG_MIGRATION
6ce3c4c0
CL
788/*
789 * page migration
790 */
fc301289
CL
791static void migrate_page_add(struct page *page, struct list_head *pagelist,
792 unsigned long flags)
6ce3c4c0
CL
793{
794 /*
fc301289 795 * Avoid migrating a page that is shared with others.
6ce3c4c0 796 */
62695a84
NP
797 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
798 if (!isolate_lru_page(page)) {
799 list_add_tail(&page->lru, pagelist);
800 }
801 }
7e2ab150 802}
6ce3c4c0 803
742755a1 804static struct page *new_node_page(struct page *page, unsigned long node, int **x)
95a402c3 805{
6484eb3e 806 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
95a402c3
CL
807}
808
7e2ab150
CL
809/*
810 * Migrate pages from one node to a target node.
811 * Returns error or the number of pages not migrated.
812 */
dbcb0f19
AB
813static int migrate_to_node(struct mm_struct *mm, int source, int dest,
814 int flags)
7e2ab150
CL
815{
816 nodemask_t nmask;
817 LIST_HEAD(pagelist);
818 int err = 0;
819
820 nodes_clear(nmask);
821 node_set(source, nmask);
6ce3c4c0 822
7e2ab150
CL
823 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
824 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
825
aaa994b3 826 if (!list_empty(&pagelist))
95a402c3
CL
827 err = migrate_pages(&pagelist, new_node_page, dest);
828
7e2ab150 829 return err;
6ce3c4c0
CL
830}
831
39743889 832/*
7e2ab150
CL
833 * Move pages between the two nodesets so as to preserve the physical
834 * layout as much as possible.
39743889
CL
835 *
836 * Returns the number of page that could not be moved.
837 */
838int do_migrate_pages(struct mm_struct *mm,
839 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
840{
7e2ab150 841 int busy = 0;
0aedadf9 842 int err;
7e2ab150 843 nodemask_t tmp;
39743889 844
0aedadf9
CL
845 err = migrate_prep();
846 if (err)
847 return err;
848
53f2556b 849 down_read(&mm->mmap_sem);
39743889 850
7b2259b3
CL
851 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
852 if (err)
853 goto out;
854
7e2ab150
CL
855/*
856 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
857 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
858 * bit in 'tmp', and return that <source, dest> pair for migration.
859 * The pair of nodemasks 'to' and 'from' define the map.
860 *
861 * If no pair of bits is found that way, fallback to picking some
862 * pair of 'source' and 'dest' bits that are not the same. If the
863 * 'source' and 'dest' bits are the same, this represents a node
864 * that will be migrating to itself, so no pages need move.
865 *
866 * If no bits are left in 'tmp', or if all remaining bits left
867 * in 'tmp' correspond to the same bit in 'to', return false
868 * (nothing left to migrate).
869 *
870 * This lets us pick a pair of nodes to migrate between, such that
871 * if possible the dest node is not already occupied by some other
872 * source node, minimizing the risk of overloading the memory on a
873 * node that would happen if we migrated incoming memory to a node
874 * before migrating outgoing memory source that same node.
875 *
876 * A single scan of tmp is sufficient. As we go, we remember the
877 * most recent <s, d> pair that moved (s != d). If we find a pair
878 * that not only moved, but what's better, moved to an empty slot
879 * (d is not set in tmp), then we break out then, with that pair.
880 * Otherwise when we finish scannng from_tmp, we at least have the
881 * most recent <s, d> pair that moved. If we get all the way through
882 * the scan of tmp without finding any node that moved, much less
883 * moved to an empty node, then there is nothing left worth migrating.
884 */
d4984711 885
7e2ab150
CL
886 tmp = *from_nodes;
887 while (!nodes_empty(tmp)) {
888 int s,d;
889 int source = -1;
890 int dest = 0;
891
892 for_each_node_mask(s, tmp) {
893 d = node_remap(s, *from_nodes, *to_nodes);
894 if (s == d)
895 continue;
896
897 source = s; /* Node moved. Memorize */
898 dest = d;
899
900 /* dest not in remaining from nodes? */
901 if (!node_isset(dest, tmp))
902 break;
903 }
904 if (source == -1)
905 break;
906
907 node_clear(source, tmp);
908 err = migrate_to_node(mm, source, dest, flags);
909 if (err > 0)
910 busy += err;
911 if (err < 0)
912 break;
39743889 913 }
7b2259b3 914out:
39743889 915 up_read(&mm->mmap_sem);
7e2ab150
CL
916 if (err < 0)
917 return err;
918 return busy;
b20a3503
CL
919
920}
921
3ad33b24
LS
922/*
923 * Allocate a new page for page migration based on vma policy.
924 * Start assuming that page is mapped by vma pointed to by @private.
925 * Search forward from there, if not. N.B., this assumes that the
926 * list of pages handed to migrate_pages()--which is how we get here--
927 * is in virtual address order.
928 */
742755a1 929static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
95a402c3
CL
930{
931 struct vm_area_struct *vma = (struct vm_area_struct *)private;
3ad33b24 932 unsigned long uninitialized_var(address);
95a402c3 933
3ad33b24
LS
934 while (vma) {
935 address = page_address_in_vma(page, vma);
936 if (address != -EFAULT)
937 break;
938 vma = vma->vm_next;
939 }
940
941 /*
942 * if !vma, alloc_page_vma() will use task or system default policy
943 */
944 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
95a402c3 945}
b20a3503
CL
946#else
947
948static void migrate_page_add(struct page *page, struct list_head *pagelist,
949 unsigned long flags)
950{
39743889
CL
951}
952
b20a3503
CL
953int do_migrate_pages(struct mm_struct *mm,
954 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
955{
956 return -ENOSYS;
957}
95a402c3 958
69939749 959static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
95a402c3
CL
960{
961 return NULL;
962}
b20a3503
CL
963#endif
964
dbcb0f19 965static long do_mbind(unsigned long start, unsigned long len,
028fec41
DR
966 unsigned short mode, unsigned short mode_flags,
967 nodemask_t *nmask, unsigned long flags)
6ce3c4c0
CL
968{
969 struct vm_area_struct *vma;
970 struct mm_struct *mm = current->mm;
971 struct mempolicy *new;
972 unsigned long end;
973 int err;
974 LIST_HEAD(pagelist);
975
a3b51e01
DR
976 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
977 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
6ce3c4c0 978 return -EINVAL;
74c00241 979 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
6ce3c4c0
CL
980 return -EPERM;
981
982 if (start & ~PAGE_MASK)
983 return -EINVAL;
984
985 if (mode == MPOL_DEFAULT)
986 flags &= ~MPOL_MF_STRICT;
987
988 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
989 end = start + len;
990
991 if (end < start)
992 return -EINVAL;
993 if (end == start)
994 return 0;
995
028fec41 996 new = mpol_new(mode, mode_flags, nmask);
6ce3c4c0
CL
997 if (IS_ERR(new))
998 return PTR_ERR(new);
999
1000 /*
1001 * If we are using the default policy then operation
1002 * on discontinuous address spaces is okay after all
1003 */
1004 if (!new)
1005 flags |= MPOL_MF_DISCONTIG_OK;
1006
028fec41
DR
1007 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1008 start, start + len, mode, mode_flags,
1009 nmask ? nodes_addr(*nmask)[0] : -1);
6ce3c4c0 1010
0aedadf9
CL
1011 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1012
1013 err = migrate_prep();
1014 if (err)
1015 return err;
1016 }
6ce3c4c0 1017 down_write(&mm->mmap_sem);
58568d2a
MX
1018 task_lock(current);
1019 err = mpol_set_nodemask(new, nmask);
1020 task_unlock(current);
1021 if (err) {
1022 up_write(&mm->mmap_sem);
1023 mpol_put(new);
1024 return err;
1025 }
6ce3c4c0
CL
1026 vma = check_range(mm, start, end, nmask,
1027 flags | MPOL_MF_INVERT, &pagelist);
1028
1029 err = PTR_ERR(vma);
1030 if (!IS_ERR(vma)) {
1031 int nr_failed = 0;
1032
1033 err = mbind_range(vma, start, end, new);
7e2ab150 1034
6ce3c4c0 1035 if (!list_empty(&pagelist))
95a402c3
CL
1036 nr_failed = migrate_pages(&pagelist, new_vma_page,
1037 (unsigned long)vma);
6ce3c4c0
CL
1038
1039 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1040 err = -EIO;
1041 }
b20a3503 1042
6ce3c4c0 1043 up_write(&mm->mmap_sem);
f0be3d32 1044 mpol_put(new);
6ce3c4c0
CL
1045 return err;
1046}
1047
8bccd85f
CL
1048/*
1049 * User space interface with variable sized bitmaps for nodelists.
1050 */
1051
1052/* Copy a node mask from user space. */
39743889 1053static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
8bccd85f
CL
1054 unsigned long maxnode)
1055{
1056 unsigned long k;
1057 unsigned long nlongs;
1058 unsigned long endmask;
1059
1060 --maxnode;
1061 nodes_clear(*nodes);
1062 if (maxnode == 0 || !nmask)
1063 return 0;
a9c930ba 1064 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
636f13c1 1065 return -EINVAL;
8bccd85f
CL
1066
1067 nlongs = BITS_TO_LONGS(maxnode);
1068 if ((maxnode % BITS_PER_LONG) == 0)
1069 endmask = ~0UL;
1070 else
1071 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1072
1073 /* When the user specified more nodes than supported just check
1074 if the non supported part is all zero. */
1075 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1076 if (nlongs > PAGE_SIZE/sizeof(long))
1077 return -EINVAL;
1078 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1079 unsigned long t;
1080 if (get_user(t, nmask + k))
1081 return -EFAULT;
1082 if (k == nlongs - 1) {
1083 if (t & endmask)
1084 return -EINVAL;
1085 } else if (t)
1086 return -EINVAL;
1087 }
1088 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1089 endmask = ~0UL;
1090 }
1091
1092 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1093 return -EFAULT;
1094 nodes_addr(*nodes)[nlongs-1] &= endmask;
1095 return 0;
1096}
1097
1098/* Copy a kernel node mask to user space */
1099static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1100 nodemask_t *nodes)
1101{
1102 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1103 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1104
1105 if (copy > nbytes) {
1106 if (copy > PAGE_SIZE)
1107 return -EINVAL;
1108 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1109 return -EFAULT;
1110 copy = nbytes;
1111 }
1112 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1113}
1114
938bb9f5
HC
1115SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1116 unsigned long, mode, unsigned long __user *, nmask,
1117 unsigned long, maxnode, unsigned, flags)
8bccd85f
CL
1118{
1119 nodemask_t nodes;
1120 int err;
028fec41 1121 unsigned short mode_flags;
8bccd85f 1122
028fec41
DR
1123 mode_flags = mode & MPOL_MODE_FLAGS;
1124 mode &= ~MPOL_MODE_FLAGS;
a3b51e01
DR
1125 if (mode >= MPOL_MAX)
1126 return -EINVAL;
4c50bc01
DR
1127 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1128 (mode_flags & MPOL_F_RELATIVE_NODES))
1129 return -EINVAL;
8bccd85f
CL
1130 err = get_nodes(&nodes, nmask, maxnode);
1131 if (err)
1132 return err;
028fec41 1133 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
8bccd85f
CL
1134}
1135
1136/* Set the process memory policy */
938bb9f5
HC
1137SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1138 unsigned long, maxnode)
8bccd85f
CL
1139{
1140 int err;
1141 nodemask_t nodes;
028fec41 1142 unsigned short flags;
8bccd85f 1143
028fec41
DR
1144 flags = mode & MPOL_MODE_FLAGS;
1145 mode &= ~MPOL_MODE_FLAGS;
1146 if ((unsigned int)mode >= MPOL_MAX)
8bccd85f 1147 return -EINVAL;
4c50bc01
DR
1148 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1149 return -EINVAL;
8bccd85f
CL
1150 err = get_nodes(&nodes, nmask, maxnode);
1151 if (err)
1152 return err;
028fec41 1153 return do_set_mempolicy(mode, flags, &nodes);
8bccd85f
CL
1154}
1155
938bb9f5
HC
1156SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1157 const unsigned long __user *, old_nodes,
1158 const unsigned long __user *, new_nodes)
39743889 1159{
c69e8d9c 1160 const struct cred *cred = current_cred(), *tcred;
39743889
CL
1161 struct mm_struct *mm;
1162 struct task_struct *task;
1163 nodemask_t old;
1164 nodemask_t new;
1165 nodemask_t task_nodes;
1166 int err;
1167
1168 err = get_nodes(&old, old_nodes, maxnode);
1169 if (err)
1170 return err;
1171
1172 err = get_nodes(&new, new_nodes, maxnode);
1173 if (err)
1174 return err;
1175
1176 /* Find the mm_struct */
1177 read_lock(&tasklist_lock);
228ebcbe 1178 task = pid ? find_task_by_vpid(pid) : current;
39743889
CL
1179 if (!task) {
1180 read_unlock(&tasklist_lock);
1181 return -ESRCH;
1182 }
1183 mm = get_task_mm(task);
1184 read_unlock(&tasklist_lock);
1185
1186 if (!mm)
1187 return -EINVAL;
1188
1189 /*
1190 * Check if this process has the right to modify the specified
1191 * process. The right exists if the process has administrative
7f927fcc 1192 * capabilities, superuser privileges or the same
39743889
CL
1193 * userid as the target process.
1194 */
c69e8d9c
DH
1195 rcu_read_lock();
1196 tcred = __task_cred(task);
b6dff3ec
DH
1197 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1198 cred->uid != tcred->suid && cred->uid != tcred->uid &&
74c00241 1199 !capable(CAP_SYS_NICE)) {
c69e8d9c 1200 rcu_read_unlock();
39743889
CL
1201 err = -EPERM;
1202 goto out;
1203 }
c69e8d9c 1204 rcu_read_unlock();
39743889
CL
1205
1206 task_nodes = cpuset_mems_allowed(task);
1207 /* Is the user allowed to access the target nodes? */
74c00241 1208 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
39743889
CL
1209 err = -EPERM;
1210 goto out;
1211 }
1212
37b07e41 1213 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
3b42d28b
CL
1214 err = -EINVAL;
1215 goto out;
1216 }
1217
86c3a764
DQ
1218 err = security_task_movememory(task);
1219 if (err)
1220 goto out;
1221
511030bc 1222 err = do_migrate_pages(mm, &old, &new,
74c00241 1223 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
39743889
CL
1224out:
1225 mmput(mm);
1226 return err;
1227}
1228
1229
8bccd85f 1230/* Retrieve NUMA policy */
938bb9f5
HC
1231SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1232 unsigned long __user *, nmask, unsigned long, maxnode,
1233 unsigned long, addr, unsigned long, flags)
8bccd85f 1234{
dbcb0f19
AB
1235 int err;
1236 int uninitialized_var(pval);
8bccd85f
CL
1237 nodemask_t nodes;
1238
1239 if (nmask != NULL && maxnode < MAX_NUMNODES)
1240 return -EINVAL;
1241
1242 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1243
1244 if (err)
1245 return err;
1246
1247 if (policy && put_user(pval, policy))
1248 return -EFAULT;
1249
1250 if (nmask)
1251 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1252
1253 return err;
1254}
1255
1da177e4
LT
1256#ifdef CONFIG_COMPAT
1257
1258asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1259 compat_ulong_t __user *nmask,
1260 compat_ulong_t maxnode,
1261 compat_ulong_t addr, compat_ulong_t flags)
1262{
1263 long err;
1264 unsigned long __user *nm = NULL;
1265 unsigned long nr_bits, alloc_size;
1266 DECLARE_BITMAP(bm, MAX_NUMNODES);
1267
1268 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1269 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1270
1271 if (nmask)
1272 nm = compat_alloc_user_space(alloc_size);
1273
1274 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1275
1276 if (!err && nmask) {
1277 err = copy_from_user(bm, nm, alloc_size);
1278 /* ensure entire bitmap is zeroed */
1279 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1280 err |= compat_put_bitmap(nmask, bm, nr_bits);
1281 }
1282
1283 return err;
1284}
1285
1286asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1287 compat_ulong_t maxnode)
1288{
1289 long err = 0;
1290 unsigned long __user *nm = NULL;
1291 unsigned long nr_bits, alloc_size;
1292 DECLARE_BITMAP(bm, MAX_NUMNODES);
1293
1294 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1295 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1296
1297 if (nmask) {
1298 err = compat_get_bitmap(bm, nmask, nr_bits);
1299 nm = compat_alloc_user_space(alloc_size);
1300 err |= copy_to_user(nm, bm, alloc_size);
1301 }
1302
1303 if (err)
1304 return -EFAULT;
1305
1306 return sys_set_mempolicy(mode, nm, nr_bits+1);
1307}
1308
1309asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1310 compat_ulong_t mode, compat_ulong_t __user *nmask,
1311 compat_ulong_t maxnode, compat_ulong_t flags)
1312{
1313 long err = 0;
1314 unsigned long __user *nm = NULL;
1315 unsigned long nr_bits, alloc_size;
dfcd3c0d 1316 nodemask_t bm;
1da177e4
LT
1317
1318 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1319 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1320
1321 if (nmask) {
dfcd3c0d 1322 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1da177e4 1323 nm = compat_alloc_user_space(alloc_size);
dfcd3c0d 1324 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1da177e4
LT
1325 }
1326
1327 if (err)
1328 return -EFAULT;
1329
1330 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1331}
1332
1333#endif
1334
480eccf9
LS
1335/*
1336 * get_vma_policy(@task, @vma, @addr)
1337 * @task - task for fallback if vma policy == default
1338 * @vma - virtual memory area whose policy is sought
1339 * @addr - address in @vma for shared policy lookup
1340 *
1341 * Returns effective policy for a VMA at specified address.
1342 * Falls back to @task or system default policy, as necessary.
52cd3b07
LS
1343 * Current or other task's task mempolicy and non-shared vma policies
1344 * are protected by the task's mmap_sem, which must be held for read by
1345 * the caller.
1346 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1347 * count--added by the get_policy() vm_op, as appropriate--to protect against
1348 * freeing by another task. It is the caller's responsibility to free the
1349 * extra reference for shared policies.
480eccf9 1350 */
ae4d8c16 1351static struct mempolicy *get_vma_policy(struct task_struct *task,
48fce342 1352 struct vm_area_struct *vma, unsigned long addr)
1da177e4 1353{
6e21c8f1 1354 struct mempolicy *pol = task->mempolicy;
1da177e4
LT
1355
1356 if (vma) {
480eccf9 1357 if (vma->vm_ops && vma->vm_ops->get_policy) {
ae4d8c16
LS
1358 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1359 addr);
1360 if (vpol)
1361 pol = vpol;
bea904d5 1362 } else if (vma->vm_policy)
1da177e4
LT
1363 pol = vma->vm_policy;
1364 }
1365 if (!pol)
1366 pol = &default_policy;
1367 return pol;
1368}
1369
52cd3b07
LS
1370/*
1371 * Return a nodemask representing a mempolicy for filtering nodes for
1372 * page allocation
1373 */
1374static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
19770b32
MG
1375{
1376 /* Lower zones don't get a nodemask applied for MPOL_BIND */
45c4745a 1377 if (unlikely(policy->mode == MPOL_BIND) &&
19770b32
MG
1378 gfp_zone(gfp) >= policy_zone &&
1379 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1380 return &policy->v.nodes;
1381
1382 return NULL;
1383}
1384
52cd3b07
LS
1385/* Return a zonelist indicated by gfp for node representing a mempolicy */
1386static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
1da177e4 1387{
fc36b8d3 1388 int nd = numa_node_id();
1da177e4 1389
45c4745a 1390 switch (policy->mode) {
1da177e4 1391 case MPOL_PREFERRED:
fc36b8d3
LS
1392 if (!(policy->flags & MPOL_F_LOCAL))
1393 nd = policy->v.preferred_node;
1da177e4
LT
1394 break;
1395 case MPOL_BIND:
19770b32 1396 /*
52cd3b07
LS
1397 * Normally, MPOL_BIND allocations are node-local within the
1398 * allowed nodemask. However, if __GFP_THISNODE is set and the
1399 * current node is part of the mask, we use the zonelist for
1400 * the first node in the mask instead.
19770b32 1401 */
19770b32
MG
1402 if (unlikely(gfp & __GFP_THISNODE) &&
1403 unlikely(!node_isset(nd, policy->v.nodes)))
1404 nd = first_node(policy->v.nodes);
1405 break;
1da177e4 1406 case MPOL_INTERLEAVE: /* should not happen */
1da177e4
LT
1407 break;
1408 default:
1da177e4
LT
1409 BUG();
1410 }
0e88460d 1411 return node_zonelist(nd, gfp);
1da177e4
LT
1412}
1413
1414/* Do dynamic interleaving for a process */
1415static unsigned interleave_nodes(struct mempolicy *policy)
1416{
1417 unsigned nid, next;
1418 struct task_struct *me = current;
1419
1420 nid = me->il_next;
dfcd3c0d 1421 next = next_node(nid, policy->v.nodes);
1da177e4 1422 if (next >= MAX_NUMNODES)
dfcd3c0d 1423 next = first_node(policy->v.nodes);
f5b087b5
DR
1424 if (next < MAX_NUMNODES)
1425 me->il_next = next;
1da177e4
LT
1426 return nid;
1427}
1428
dc85da15
CL
1429/*
1430 * Depending on the memory policy provide a node from which to allocate the
1431 * next slab entry.
52cd3b07
LS
1432 * @policy must be protected by freeing by the caller. If @policy is
1433 * the current task's mempolicy, this protection is implicit, as only the
1434 * task can change it's policy. The system default policy requires no
1435 * such protection.
dc85da15
CL
1436 */
1437unsigned slab_node(struct mempolicy *policy)
1438{
fc36b8d3 1439 if (!policy || policy->flags & MPOL_F_LOCAL)
bea904d5
LS
1440 return numa_node_id();
1441
1442 switch (policy->mode) {
1443 case MPOL_PREFERRED:
fc36b8d3
LS
1444 /*
1445 * handled MPOL_F_LOCAL above
1446 */
1447 return policy->v.preferred_node;
765c4507 1448
dc85da15
CL
1449 case MPOL_INTERLEAVE:
1450 return interleave_nodes(policy);
1451
dd1a239f 1452 case MPOL_BIND: {
dc85da15
CL
1453 /*
1454 * Follow bind policy behavior and start allocation at the
1455 * first node.
1456 */
19770b32
MG
1457 struct zonelist *zonelist;
1458 struct zone *zone;
1459 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1460 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1461 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1462 &policy->v.nodes,
1463 &zone);
1464 return zone->node;
dd1a239f 1465 }
dc85da15 1466
dc85da15 1467 default:
bea904d5 1468 BUG();
dc85da15
CL
1469 }
1470}
1471
1da177e4
LT
1472/* Do static interleaving for a VMA with known offset. */
1473static unsigned offset_il_node(struct mempolicy *pol,
1474 struct vm_area_struct *vma, unsigned long off)
1475{
dfcd3c0d 1476 unsigned nnodes = nodes_weight(pol->v.nodes);
f5b087b5 1477 unsigned target;
1da177e4
LT
1478 int c;
1479 int nid = -1;
1480
f5b087b5
DR
1481 if (!nnodes)
1482 return numa_node_id();
1483 target = (unsigned int)off % nnodes;
1da177e4
LT
1484 c = 0;
1485 do {
dfcd3c0d 1486 nid = next_node(nid, pol->v.nodes);
1da177e4
LT
1487 c++;
1488 } while (c <= target);
1da177e4
LT
1489 return nid;
1490}
1491
5da7ca86
CL
1492/* Determine a node number for interleave */
1493static inline unsigned interleave_nid(struct mempolicy *pol,
1494 struct vm_area_struct *vma, unsigned long addr, int shift)
1495{
1496 if (vma) {
1497 unsigned long off;
1498
3b98b087
NA
1499 /*
1500 * for small pages, there is no difference between
1501 * shift and PAGE_SHIFT, so the bit-shift is safe.
1502 * for huge pages, since vm_pgoff is in units of small
1503 * pages, we need to shift off the always 0 bits to get
1504 * a useful offset.
1505 */
1506 BUG_ON(shift < PAGE_SHIFT);
1507 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
5da7ca86
CL
1508 off += (addr - vma->vm_start) >> shift;
1509 return offset_il_node(pol, vma, off);
1510 } else
1511 return interleave_nodes(pol);
1512}
1513
00ac59ad 1514#ifdef CONFIG_HUGETLBFS
480eccf9
LS
1515/*
1516 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1517 * @vma = virtual memory area whose policy is sought
1518 * @addr = address in @vma for shared policy lookup and interleave policy
1519 * @gfp_flags = for requested zone
19770b32
MG
1520 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1521 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
480eccf9 1522 *
52cd3b07
LS
1523 * Returns a zonelist suitable for a huge page allocation and a pointer
1524 * to the struct mempolicy for conditional unref after allocation.
1525 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1526 * @nodemask for filtering the zonelist.
480eccf9 1527 */
396faf03 1528struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
19770b32
MG
1529 gfp_t gfp_flags, struct mempolicy **mpol,
1530 nodemask_t **nodemask)
5da7ca86 1531{
480eccf9 1532 struct zonelist *zl;
5da7ca86 1533
52cd3b07 1534 *mpol = get_vma_policy(current, vma, addr);
19770b32 1535 *nodemask = NULL; /* assume !MPOL_BIND */
5da7ca86 1536
52cd3b07
LS
1537 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1538 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
a5516438 1539 huge_page_shift(hstate_vma(vma))), gfp_flags);
52cd3b07
LS
1540 } else {
1541 zl = policy_zonelist(gfp_flags, *mpol);
1542 if ((*mpol)->mode == MPOL_BIND)
1543 *nodemask = &(*mpol)->v.nodes;
480eccf9
LS
1544 }
1545 return zl;
5da7ca86 1546}
00ac59ad 1547#endif
5da7ca86 1548
1da177e4
LT
1549/* Allocate a page in interleaved policy.
1550 Own path because it needs to do special accounting. */
662f3a0b
AK
1551static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1552 unsigned nid)
1da177e4
LT
1553{
1554 struct zonelist *zl;
1555 struct page *page;
1556
0e88460d 1557 zl = node_zonelist(nid, gfp);
1da177e4 1558 page = __alloc_pages(gfp, order, zl);
dd1a239f 1559 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
ca889e6c 1560 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1da177e4
LT
1561 return page;
1562}
1563
1564/**
1565 * alloc_page_vma - Allocate a page for a VMA.
1566 *
1567 * @gfp:
1568 * %GFP_USER user allocation.
1569 * %GFP_KERNEL kernel allocations,
1570 * %GFP_HIGHMEM highmem/user allocations,
1571 * %GFP_FS allocation should not call back into a file system.
1572 * %GFP_ATOMIC don't sleep.
1573 *
1574 * @vma: Pointer to VMA or NULL if not available.
1575 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1576 *
1577 * This function allocates a page from the kernel page pool and applies
1578 * a NUMA policy associated with the VMA or the current process.
1579 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1580 * mm_struct of the VMA to prevent it from going away. Should be used for
1581 * all allocations for pages that will be mapped into
1582 * user space. Returns NULL when no page can be allocated.
1583 *
1584 * Should be called with the mm_sem of the vma hold.
1585 */
1586struct page *
dd0fc66f 1587alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1da177e4 1588{
6e21c8f1 1589 struct mempolicy *pol = get_vma_policy(current, vma, addr);
480eccf9 1590 struct zonelist *zl;
1da177e4 1591
45c4745a 1592 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1da177e4 1593 unsigned nid;
5da7ca86
CL
1594
1595 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
52cd3b07 1596 mpol_cond_put(pol);
1da177e4
LT
1597 return alloc_page_interleave(gfp, 0, nid);
1598 }
52cd3b07
LS
1599 zl = policy_zonelist(gfp, pol);
1600 if (unlikely(mpol_needs_cond_ref(pol))) {
480eccf9 1601 /*
52cd3b07 1602 * slow path: ref counted shared policy
480eccf9 1603 */
19770b32 1604 struct page *page = __alloc_pages_nodemask(gfp, 0,
52cd3b07 1605 zl, policy_nodemask(gfp, pol));
f0be3d32 1606 __mpol_put(pol);
480eccf9
LS
1607 return page;
1608 }
1609 /*
1610 * fast path: default or task policy
1611 */
52cd3b07 1612 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1da177e4
LT
1613}
1614
1615/**
1616 * alloc_pages_current - Allocate pages.
1617 *
1618 * @gfp:
1619 * %GFP_USER user allocation,
1620 * %GFP_KERNEL kernel allocation,
1621 * %GFP_HIGHMEM highmem allocation,
1622 * %GFP_FS don't call back into a file system.
1623 * %GFP_ATOMIC don't sleep.
1624 * @order: Power of two of allocation size in pages. 0 is a single page.
1625 *
1626 * Allocate a page from the kernel page pool. When not in
1627 * interrupt context and apply the current process NUMA policy.
1628 * Returns NULL when no page can be allocated.
1629 *
cf2a473c 1630 * Don't call cpuset_update_task_memory_state() unless
1da177e4
LT
1631 * 1) it's ok to take cpuset_sem (can WAIT), and
1632 * 2) allocating for current task (not interrupt).
1633 */
dd0fc66f 1634struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1da177e4
LT
1635{
1636 struct mempolicy *pol = current->mempolicy;
1637
9b819d20 1638 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1da177e4 1639 pol = &default_policy;
52cd3b07
LS
1640
1641 /*
1642 * No reference counting needed for current->mempolicy
1643 * nor system default_policy
1644 */
45c4745a 1645 if (pol->mode == MPOL_INTERLEAVE)
1da177e4 1646 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
19770b32 1647 return __alloc_pages_nodemask(gfp, order,
52cd3b07 1648 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
1da177e4
LT
1649}
1650EXPORT_SYMBOL(alloc_pages_current);
1651
4225399a 1652/*
846a16bf 1653 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
4225399a
PJ
1654 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1655 * with the mems_allowed returned by cpuset_mems_allowed(). This
1656 * keeps mempolicies cpuset relative after its cpuset moves. See
1657 * further kernel/cpuset.c update_nodemask().
1658 */
4225399a 1659
846a16bf
LS
1660/* Slow path of a mempolicy duplicate */
1661struct mempolicy *__mpol_dup(struct mempolicy *old)
1da177e4
LT
1662{
1663 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1664
1665 if (!new)
1666 return ERR_PTR(-ENOMEM);
4225399a
PJ
1667 if (current_cpuset_is_being_rebound()) {
1668 nodemask_t mems = cpuset_mems_allowed(current);
1669 mpol_rebind_policy(old, &mems);
1670 }
1da177e4
LT
1671 *new = *old;
1672 atomic_set(&new->refcnt, 1);
1da177e4
LT
1673 return new;
1674}
1675
52cd3b07
LS
1676/*
1677 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1678 * eliminate the * MPOL_F_* flags that require conditional ref and
1679 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1680 * after return. Use the returned value.
1681 *
1682 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1683 * policy lookup, even if the policy needs/has extra ref on lookup.
1684 * shmem_readahead needs this.
1685 */
1686struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1687 struct mempolicy *frompol)
1688{
1689 if (!mpol_needs_cond_ref(frompol))
1690 return frompol;
1691
1692 *tompol = *frompol;
1693 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1694 __mpol_put(frompol);
1695 return tompol;
1696}
1697
f5b087b5
DR
1698static int mpol_match_intent(const struct mempolicy *a,
1699 const struct mempolicy *b)
1700{
1701 if (a->flags != b->flags)
1702 return 0;
1703 if (!mpol_store_user_nodemask(a))
1704 return 1;
1705 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1706}
1707
1da177e4
LT
1708/* Slow path of a mempolicy comparison */
1709int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1710{
1711 if (!a || !b)
1712 return 0;
45c4745a 1713 if (a->mode != b->mode)
1da177e4 1714 return 0;
45c4745a 1715 if (a->mode != MPOL_DEFAULT && !mpol_match_intent(a, b))
f5b087b5 1716 return 0;
45c4745a 1717 switch (a->mode) {
19770b32
MG
1718 case MPOL_BIND:
1719 /* Fall through */
1da177e4 1720 case MPOL_INTERLEAVE:
dfcd3c0d 1721 return nodes_equal(a->v.nodes, b->v.nodes);
1da177e4 1722 case MPOL_PREFERRED:
fc36b8d3
LS
1723 return a->v.preferred_node == b->v.preferred_node &&
1724 a->flags == b->flags;
1da177e4
LT
1725 default:
1726 BUG();
1727 return 0;
1728 }
1729}
1730
1da177e4
LT
1731/*
1732 * Shared memory backing store policy support.
1733 *
1734 * Remember policies even when nobody has shared memory mapped.
1735 * The policies are kept in Red-Black tree linked from the inode.
1736 * They are protected by the sp->lock spinlock, which should be held
1737 * for any accesses to the tree.
1738 */
1739
1740/* lookup first element intersecting start-end */
1741/* Caller holds sp->lock */
1742static struct sp_node *
1743sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1744{
1745 struct rb_node *n = sp->root.rb_node;
1746
1747 while (n) {
1748 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1749
1750 if (start >= p->end)
1751 n = n->rb_right;
1752 else if (end <= p->start)
1753 n = n->rb_left;
1754 else
1755 break;
1756 }
1757 if (!n)
1758 return NULL;
1759 for (;;) {
1760 struct sp_node *w = NULL;
1761 struct rb_node *prev = rb_prev(n);
1762 if (!prev)
1763 break;
1764 w = rb_entry(prev, struct sp_node, nd);
1765 if (w->end <= start)
1766 break;
1767 n = prev;
1768 }
1769 return rb_entry(n, struct sp_node, nd);
1770}
1771
1772/* Insert a new shared policy into the list. */
1773/* Caller holds sp->lock */
1774static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1775{
1776 struct rb_node **p = &sp->root.rb_node;
1777 struct rb_node *parent = NULL;
1778 struct sp_node *nd;
1779
1780 while (*p) {
1781 parent = *p;
1782 nd = rb_entry(parent, struct sp_node, nd);
1783 if (new->start < nd->start)
1784 p = &(*p)->rb_left;
1785 else if (new->end > nd->end)
1786 p = &(*p)->rb_right;
1787 else
1788 BUG();
1789 }
1790 rb_link_node(&new->nd, parent, p);
1791 rb_insert_color(&new->nd, &sp->root);
140d5a49 1792 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
45c4745a 1793 new->policy ? new->policy->mode : 0);
1da177e4
LT
1794}
1795
1796/* Find shared policy intersecting idx */
1797struct mempolicy *
1798mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1799{
1800 struct mempolicy *pol = NULL;
1801 struct sp_node *sn;
1802
1803 if (!sp->root.rb_node)
1804 return NULL;
1805 spin_lock(&sp->lock);
1806 sn = sp_lookup(sp, idx, idx+1);
1807 if (sn) {
1808 mpol_get(sn->policy);
1809 pol = sn->policy;
1810 }
1811 spin_unlock(&sp->lock);
1812 return pol;
1813}
1814
1815static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1816{
140d5a49 1817 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1da177e4 1818 rb_erase(&n->nd, &sp->root);
f0be3d32 1819 mpol_put(n->policy);
1da177e4
LT
1820 kmem_cache_free(sn_cache, n);
1821}
1822
dbcb0f19
AB
1823static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1824 struct mempolicy *pol)
1da177e4
LT
1825{
1826 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1827
1828 if (!n)
1829 return NULL;
1830 n->start = start;
1831 n->end = end;
1832 mpol_get(pol);
aab0b102 1833 pol->flags |= MPOL_F_SHARED; /* for unref */
1da177e4
LT
1834 n->policy = pol;
1835 return n;
1836}
1837
1838/* Replace a policy range. */
1839static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1840 unsigned long end, struct sp_node *new)
1841{
1842 struct sp_node *n, *new2 = NULL;
1843
1844restart:
1845 spin_lock(&sp->lock);
1846 n = sp_lookup(sp, start, end);
1847 /* Take care of old policies in the same range. */
1848 while (n && n->start < end) {
1849 struct rb_node *next = rb_next(&n->nd);
1850 if (n->start >= start) {
1851 if (n->end <= end)
1852 sp_delete(sp, n);
1853 else
1854 n->start = end;
1855 } else {
1856 /* Old policy spanning whole new range. */
1857 if (n->end > end) {
1858 if (!new2) {
1859 spin_unlock(&sp->lock);
1860 new2 = sp_alloc(end, n->end, n->policy);
1861 if (!new2)
1862 return -ENOMEM;
1863 goto restart;
1864 }
1865 n->end = start;
1866 sp_insert(sp, new2);
1867 new2 = NULL;
1868 break;
1869 } else
1870 n->end = start;
1871 }
1872 if (!next)
1873 break;
1874 n = rb_entry(next, struct sp_node, nd);
1875 }
1876 if (new)
1877 sp_insert(sp, new);
1878 spin_unlock(&sp->lock);
1879 if (new2) {
f0be3d32 1880 mpol_put(new2->policy);
1da177e4
LT
1881 kmem_cache_free(sn_cache, new2);
1882 }
1883 return 0;
1884}
1885
71fe804b
LS
1886/**
1887 * mpol_shared_policy_init - initialize shared policy for inode
1888 * @sp: pointer to inode shared policy
1889 * @mpol: struct mempolicy to install
1890 *
1891 * Install non-NULL @mpol in inode's shared policy rb-tree.
1892 * On entry, the current task has a reference on a non-NULL @mpol.
1893 * This must be released on exit.
1894 */
1895void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
1896{
58568d2a
MX
1897 int ret;
1898
71fe804b
LS
1899 sp->root = RB_ROOT; /* empty tree == default mempolicy */
1900 spin_lock_init(&sp->lock);
1901
1902 if (mpol) {
1903 struct vm_area_struct pvma;
1904 struct mempolicy *new;
1905
1906 /* contextualize the tmpfs mount point mempolicy */
1907 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
58568d2a
MX
1908 if (IS_ERR(new)) {
1909 mpol_put(mpol); /* drop our ref on sb mpol */
71fe804b 1910 return; /* no valid nodemask intersection */
58568d2a
MX
1911 }
1912
1913 task_lock(current);
1914 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask);
1915 task_unlock(current);
1916 mpol_put(mpol); /* drop our ref on sb mpol */
1917 if (ret) {
1918 mpol_put(new);
1919 return;
1920 }
71fe804b
LS
1921
1922 /* Create pseudo-vma that contains just the policy */
1923 memset(&pvma, 0, sizeof(struct vm_area_struct));
1924 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
1925 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
1926 mpol_put(new); /* drop initial ref */
7339ff83
RH
1927 }
1928}
1929
1da177e4
LT
1930int mpol_set_shared_policy(struct shared_policy *info,
1931 struct vm_area_struct *vma, struct mempolicy *npol)
1932{
1933 int err;
1934 struct sp_node *new = NULL;
1935 unsigned long sz = vma_pages(vma);
1936
028fec41 1937 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1da177e4 1938 vma->vm_pgoff,
45c4745a 1939 sz, npol ? npol->mode : -1,
028fec41 1940 npol ? npol->flags : -1,
140d5a49 1941 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1da177e4
LT
1942
1943 if (npol) {
1944 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1945 if (!new)
1946 return -ENOMEM;
1947 }
1948 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1949 if (err && new)
1950 kmem_cache_free(sn_cache, new);
1951 return err;
1952}
1953
1954/* Free a backing policy store on inode delete. */
1955void mpol_free_shared_policy(struct shared_policy *p)
1956{
1957 struct sp_node *n;
1958 struct rb_node *next;
1959
1960 if (!p->root.rb_node)
1961 return;
1962 spin_lock(&p->lock);
1963 next = rb_first(&p->root);
1964 while (next) {
1965 n = rb_entry(next, struct sp_node, nd);
1966 next = rb_next(&n->nd);
90c5029e 1967 rb_erase(&n->nd, &p->root);
f0be3d32 1968 mpol_put(n->policy);
1da177e4
LT
1969 kmem_cache_free(sn_cache, n);
1970 }
1971 spin_unlock(&p->lock);
1da177e4
LT
1972}
1973
1974/* assumes fs == KERNEL_DS */
1975void __init numa_policy_init(void)
1976{
b71636e2
PM
1977 nodemask_t interleave_nodes;
1978 unsigned long largest = 0;
1979 int nid, prefer = 0;
1980
1da177e4
LT
1981 policy_cache = kmem_cache_create("numa_policy",
1982 sizeof(struct mempolicy),
20c2df83 1983 0, SLAB_PANIC, NULL);
1da177e4
LT
1984
1985 sn_cache = kmem_cache_create("shared_policy_node",
1986 sizeof(struct sp_node),
20c2df83 1987 0, SLAB_PANIC, NULL);
1da177e4 1988
b71636e2
PM
1989 /*
1990 * Set interleaving policy for system init. Interleaving is only
1991 * enabled across suitably sized nodes (default is >= 16MB), or
1992 * fall back to the largest node if they're all smaller.
1993 */
1994 nodes_clear(interleave_nodes);
56bbd65d 1995 for_each_node_state(nid, N_HIGH_MEMORY) {
b71636e2
PM
1996 unsigned long total_pages = node_present_pages(nid);
1997
1998 /* Preserve the largest node */
1999 if (largest < total_pages) {
2000 largest = total_pages;
2001 prefer = nid;
2002 }
2003
2004 /* Interleave this node? */
2005 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2006 node_set(nid, interleave_nodes);
2007 }
2008
2009 /* All too small, use the largest */
2010 if (unlikely(nodes_empty(interleave_nodes)))
2011 node_set(prefer, interleave_nodes);
1da177e4 2012
028fec41 2013 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
1da177e4
LT
2014 printk("numa_policy_init: interleaving failed\n");
2015}
2016
8bccd85f 2017/* Reset policy of current process to default */
1da177e4
LT
2018void numa_default_policy(void)
2019{
028fec41 2020 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
1da177e4 2021}
68860ec1 2022
095f1fc4
LS
2023/*
2024 * Parse and format mempolicy from/to strings
2025 */
2026
1a75a6c8 2027/*
fc36b8d3 2028 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
3f226aa1 2029 * Used only for mpol_parse_str() and mpol_to_str()
1a75a6c8 2030 */
53f2556b 2031#define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
15ad7cdc 2032static const char * const policy_types[] =
53f2556b 2033 { "default", "prefer", "bind", "interleave", "local" };
1a75a6c8 2034
095f1fc4
LS
2035
2036#ifdef CONFIG_TMPFS
2037/**
2038 * mpol_parse_str - parse string to mempolicy
2039 * @str: string containing mempolicy to parse
71fe804b
LS
2040 * @mpol: pointer to struct mempolicy pointer, returned on success.
2041 * @no_context: flag whether to "contextualize" the mempolicy
095f1fc4
LS
2042 *
2043 * Format of input:
2044 * <mode>[=<flags>][:<nodelist>]
2045 *
71fe804b
LS
2046 * if @no_context is true, save the input nodemask in w.user_nodemask in
2047 * the returned mempolicy. This will be used to "clone" the mempolicy in
2048 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2049 * mount option. Note that if 'static' or 'relative' mode flags were
2050 * specified, the input nodemask will already have been saved. Saving
2051 * it again is redundant, but safe.
2052 *
2053 * On success, returns 0, else 1
095f1fc4 2054 */
71fe804b 2055int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
095f1fc4 2056{
71fe804b
LS
2057 struct mempolicy *new = NULL;
2058 unsigned short uninitialized_var(mode);
2059 unsigned short uninitialized_var(mode_flags);
2060 nodemask_t nodes;
095f1fc4
LS
2061 char *nodelist = strchr(str, ':');
2062 char *flags = strchr(str, '=');
2063 int i;
2064 int err = 1;
2065
2066 if (nodelist) {
2067 /* NUL-terminate mode or flags string */
2068 *nodelist++ = '\0';
71fe804b 2069 if (nodelist_parse(nodelist, nodes))
095f1fc4 2070 goto out;
71fe804b 2071 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
095f1fc4 2072 goto out;
71fe804b
LS
2073 } else
2074 nodes_clear(nodes);
2075
095f1fc4
LS
2076 if (flags)
2077 *flags++ = '\0'; /* terminate mode string */
2078
3f226aa1 2079 for (i = 0; i <= MPOL_LOCAL; i++) {
095f1fc4 2080 if (!strcmp(str, policy_types[i])) {
71fe804b 2081 mode = i;
095f1fc4
LS
2082 break;
2083 }
2084 }
3f226aa1 2085 if (i > MPOL_LOCAL)
095f1fc4
LS
2086 goto out;
2087
71fe804b 2088 switch (mode) {
095f1fc4 2089 case MPOL_PREFERRED:
71fe804b
LS
2090 /*
2091 * Insist on a nodelist of one node only
2092 */
095f1fc4
LS
2093 if (nodelist) {
2094 char *rest = nodelist;
2095 while (isdigit(*rest))
2096 rest++;
2097 if (!*rest)
2098 err = 0;
2099 }
2100 break;
095f1fc4
LS
2101 case MPOL_INTERLEAVE:
2102 /*
2103 * Default to online nodes with memory if no nodelist
2104 */
2105 if (!nodelist)
71fe804b 2106 nodes = node_states[N_HIGH_MEMORY];
095f1fc4 2107 err = 0;
3f226aa1 2108 break;
71fe804b 2109 case MPOL_LOCAL:
3f226aa1 2110 /*
71fe804b 2111 * Don't allow a nodelist; mpol_new() checks flags
3f226aa1 2112 */
71fe804b 2113 if (nodelist)
3f226aa1 2114 goto out;
71fe804b 2115 mode = MPOL_PREFERRED;
3f226aa1 2116 break;
71fe804b
LS
2117
2118 /*
2119 * case MPOL_BIND: mpol_new() enforces non-empty nodemask.
2120 * case MPOL_DEFAULT: mpol_new() enforces empty nodemask, ignores flags.
2121 */
095f1fc4
LS
2122 }
2123
71fe804b 2124 mode_flags = 0;
095f1fc4
LS
2125 if (flags) {
2126 /*
2127 * Currently, we only support two mutually exclusive
2128 * mode flags.
2129 */
2130 if (!strcmp(flags, "static"))
71fe804b 2131 mode_flags |= MPOL_F_STATIC_NODES;
095f1fc4 2132 else if (!strcmp(flags, "relative"))
71fe804b 2133 mode_flags |= MPOL_F_RELATIVE_NODES;
095f1fc4
LS
2134 else
2135 err = 1;
2136 }
71fe804b
LS
2137
2138 new = mpol_new(mode, mode_flags, &nodes);
2139 if (IS_ERR(new))
2140 err = 1;
58568d2a
MX
2141 else {
2142 int ret;
2143
2144 task_lock(current);
2145 ret = mpol_set_nodemask(new, &nodes);
2146 task_unlock(current);
2147 if (ret)
2148 err = 1;
2149 else if (no_context) {
2150 /* save for contextualization */
2151 new->w.user_nodemask = nodes;
2152 }
2153 }
71fe804b 2154
095f1fc4
LS
2155out:
2156 /* Restore string for error message */
2157 if (nodelist)
2158 *--nodelist = ':';
2159 if (flags)
2160 *--flags = '=';
71fe804b
LS
2161 if (!err)
2162 *mpol = new;
095f1fc4
LS
2163 return err;
2164}
2165#endif /* CONFIG_TMPFS */
2166
71fe804b
LS
2167/**
2168 * mpol_to_str - format a mempolicy structure for printing
2169 * @buffer: to contain formatted mempolicy string
2170 * @maxlen: length of @buffer
2171 * @pol: pointer to mempolicy to be formatted
2172 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2173 *
1a75a6c8
CL
2174 * Convert a mempolicy into a string.
2175 * Returns the number of characters in buffer (if positive)
2176 * or an error (negative)
2177 */
71fe804b 2178int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
1a75a6c8
CL
2179{
2180 char *p = buffer;
2181 int l;
2182 nodemask_t nodes;
bea904d5 2183 unsigned short mode;
f5b087b5 2184 unsigned short flags = pol ? pol->flags : 0;
1a75a6c8 2185
2291990a
LS
2186 /*
2187 * Sanity check: room for longest mode, flag and some nodes
2188 */
2189 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2190
bea904d5
LS
2191 if (!pol || pol == &default_policy)
2192 mode = MPOL_DEFAULT;
2193 else
2194 mode = pol->mode;
2195
1a75a6c8
CL
2196 switch (mode) {
2197 case MPOL_DEFAULT:
2198 nodes_clear(nodes);
2199 break;
2200
2201 case MPOL_PREFERRED:
2202 nodes_clear(nodes);
fc36b8d3 2203 if (flags & MPOL_F_LOCAL)
53f2556b
LS
2204 mode = MPOL_LOCAL; /* pseudo-policy */
2205 else
fc36b8d3 2206 node_set(pol->v.preferred_node, nodes);
1a75a6c8
CL
2207 break;
2208
2209 case MPOL_BIND:
19770b32 2210 /* Fall through */
1a75a6c8 2211 case MPOL_INTERLEAVE:
71fe804b
LS
2212 if (no_context)
2213 nodes = pol->w.user_nodemask;
2214 else
2215 nodes = pol->v.nodes;
1a75a6c8
CL
2216 break;
2217
2218 default:
2219 BUG();
1a75a6c8
CL
2220 }
2221
2222 l = strlen(policy_types[mode]);
53f2556b
LS
2223 if (buffer + maxlen < p + l + 1)
2224 return -ENOSPC;
1a75a6c8
CL
2225
2226 strcpy(p, policy_types[mode]);
2227 p += l;
2228
fc36b8d3 2229 if (flags & MPOL_MODE_FLAGS) {
f5b087b5
DR
2230 if (buffer + maxlen < p + 2)
2231 return -ENOSPC;
2232 *p++ = '=';
2233
2291990a
LS
2234 /*
2235 * Currently, the only defined flags are mutually exclusive
2236 */
f5b087b5 2237 if (flags & MPOL_F_STATIC_NODES)
2291990a
LS
2238 p += snprintf(p, buffer + maxlen - p, "static");
2239 else if (flags & MPOL_F_RELATIVE_NODES)
2240 p += snprintf(p, buffer + maxlen - p, "relative");
f5b087b5
DR
2241 }
2242
1a75a6c8
CL
2243 if (!nodes_empty(nodes)) {
2244 if (buffer + maxlen < p + 2)
2245 return -ENOSPC;
095f1fc4 2246 *p++ = ':';
1a75a6c8
CL
2247 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2248 }
2249 return p - buffer;
2250}
2251
2252struct numa_maps {
2253 unsigned long pages;
2254 unsigned long anon;
397874df
CL
2255 unsigned long active;
2256 unsigned long writeback;
1a75a6c8 2257 unsigned long mapcount_max;
397874df
CL
2258 unsigned long dirty;
2259 unsigned long swapcache;
1a75a6c8
CL
2260 unsigned long node[MAX_NUMNODES];
2261};
2262
397874df 2263static void gather_stats(struct page *page, void *private, int pte_dirty)
1a75a6c8
CL
2264{
2265 struct numa_maps *md = private;
2266 int count = page_mapcount(page);
2267
397874df
CL
2268 md->pages++;
2269 if (pte_dirty || PageDirty(page))
2270 md->dirty++;
1a75a6c8 2271
397874df
CL
2272 if (PageSwapCache(page))
2273 md->swapcache++;
1a75a6c8 2274
894bc310 2275 if (PageActive(page) || PageUnevictable(page))
397874df
CL
2276 md->active++;
2277
2278 if (PageWriteback(page))
2279 md->writeback++;
1a75a6c8
CL
2280
2281 if (PageAnon(page))
2282 md->anon++;
2283
397874df
CL
2284 if (count > md->mapcount_max)
2285 md->mapcount_max = count;
2286
1a75a6c8 2287 md->node[page_to_nid(page)]++;
1a75a6c8
CL
2288}
2289
7f709ed0 2290#ifdef CONFIG_HUGETLB_PAGE
397874df
CL
2291static void check_huge_range(struct vm_area_struct *vma,
2292 unsigned long start, unsigned long end,
2293 struct numa_maps *md)
2294{
2295 unsigned long addr;
2296 struct page *page;
a5516438
AK
2297 struct hstate *h = hstate_vma(vma);
2298 unsigned long sz = huge_page_size(h);
397874df 2299
a5516438
AK
2300 for (addr = start; addr < end; addr += sz) {
2301 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2302 addr & huge_page_mask(h));
397874df
CL
2303 pte_t pte;
2304
2305 if (!ptep)
2306 continue;
2307
2308 pte = *ptep;
2309 if (pte_none(pte))
2310 continue;
2311
2312 page = pte_page(pte);
2313 if (!page)
2314 continue;
2315
2316 gather_stats(page, md, pte_dirty(*ptep));
2317 }
2318}
7f709ed0
AM
2319#else
2320static inline void check_huge_range(struct vm_area_struct *vma,
2321 unsigned long start, unsigned long end,
2322 struct numa_maps *md)
2323{
2324}
2325#endif
397874df 2326
53f2556b
LS
2327/*
2328 * Display pages allocated per node and memory policy via /proc.
2329 */
1a75a6c8
CL
2330int show_numa_map(struct seq_file *m, void *v)
2331{
99f89551 2332 struct proc_maps_private *priv = m->private;
1a75a6c8
CL
2333 struct vm_area_struct *vma = v;
2334 struct numa_maps *md;
397874df
CL
2335 struct file *file = vma->vm_file;
2336 struct mm_struct *mm = vma->vm_mm;
480eccf9 2337 struct mempolicy *pol;
1a75a6c8
CL
2338 int n;
2339 char buffer[50];
2340
397874df 2341 if (!mm)
1a75a6c8
CL
2342 return 0;
2343
2344 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2345 if (!md)
2346 return 0;
2347
480eccf9 2348 pol = get_vma_policy(priv->task, vma, vma->vm_start);
71fe804b 2349 mpol_to_str(buffer, sizeof(buffer), pol, 0);
52cd3b07 2350 mpol_cond_put(pol);
397874df
CL
2351
2352 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2353
2354 if (file) {
2355 seq_printf(m, " file=");
c32c2f63 2356 seq_path(m, &file->f_path, "\n\t= ");
397874df
CL
2357 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2358 seq_printf(m, " heap");
2359 } else if (vma->vm_start <= mm->start_stack &&
2360 vma->vm_end >= mm->start_stack) {
2361 seq_printf(m, " stack");
2362 }
2363
2364 if (is_vm_hugetlb_page(vma)) {
2365 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2366 seq_printf(m, " huge");
2367 } else {
a57ebfdb 2368 check_pgd_range(vma, vma->vm_start, vma->vm_end,
56bbd65d 2369 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
397874df
CL
2370 }
2371
2372 if (!md->pages)
2373 goto out;
1a75a6c8 2374
397874df
CL
2375 if (md->anon)
2376 seq_printf(m," anon=%lu",md->anon);
1a75a6c8 2377
397874df
CL
2378 if (md->dirty)
2379 seq_printf(m," dirty=%lu",md->dirty);
1a75a6c8 2380
397874df
CL
2381 if (md->pages != md->anon && md->pages != md->dirty)
2382 seq_printf(m, " mapped=%lu", md->pages);
1a75a6c8 2383
397874df
CL
2384 if (md->mapcount_max > 1)
2385 seq_printf(m, " mapmax=%lu", md->mapcount_max);
1a75a6c8 2386
397874df
CL
2387 if (md->swapcache)
2388 seq_printf(m," swapcache=%lu", md->swapcache);
2389
2390 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2391 seq_printf(m," active=%lu", md->active);
2392
2393 if (md->writeback)
2394 seq_printf(m," writeback=%lu", md->writeback);
2395
56bbd65d 2396 for_each_node_state(n, N_HIGH_MEMORY)
397874df
CL
2397 if (md->node[n])
2398 seq_printf(m, " N%d=%lu", n, md->node[n]);
2399out:
2400 seq_putc(m, '\n');
1a75a6c8
CL
2401 kfree(md);
2402
2403 if (m->count < m->size)
99f89551 2404 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;
1a75a6c8
CL
2405 return 0;
2406}