mm: oom analysis: Show kernel stack usage in /proc/meminfo and OOM log output
[linux-2.6-block.git] / mm / migrate.c
CommitLineData
b20a3503
CL
1/*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
cde53535 12 * Christoph Lameter
b20a3503
CL
13 */
14
15#include <linux/migrate.h>
16#include <linux/module.h>
17#include <linux/swap.h>
0697212a 18#include <linux/swapops.h>
b20a3503 19#include <linux/pagemap.h>
e23ca00b 20#include <linux/buffer_head.h>
b20a3503 21#include <linux/mm_inline.h>
b488893a 22#include <linux/nsproxy.h>
b20a3503
CL
23#include <linux/pagevec.h>
24#include <linux/rmap.h>
25#include <linux/topology.h>
26#include <linux/cpu.h>
27#include <linux/cpuset.h>
04e62a29 28#include <linux/writeback.h>
742755a1
CL
29#include <linux/mempolicy.h>
30#include <linux/vmalloc.h>
86c3a764 31#include <linux/security.h>
8a9f3ccd 32#include <linux/memcontrol.h>
4f5ca265 33#include <linux/syscalls.h>
b20a3503
CL
34
35#include "internal.h"
36
b20a3503
CL
37#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
38
b20a3503 39/*
742755a1
CL
40 * migrate_prep() needs to be called before we start compiling a list of pages
41 * to be migrated using isolate_lru_page().
b20a3503
CL
42 */
43int migrate_prep(void)
44{
b20a3503
CL
45 /*
46 * Clear the LRU lists so pages can be isolated.
47 * Note that pages may be moved off the LRU after we have
48 * drained them. Those pages will fail to migrate like other
49 * pages that may be busy.
50 */
51 lru_add_drain_all();
52
53 return 0;
54}
55
b20a3503 56/*
894bc310
LS
57 * Add isolated pages on the list back to the LRU under page lock
58 * to avoid leaking evictable pages back onto unevictable list.
b20a3503
CL
59 *
60 * returns the number of pages put back.
61 */
62int putback_lru_pages(struct list_head *l)
63{
64 struct page *page;
65 struct page *page2;
66 int count = 0;
67
68 list_for_each_entry_safe(page, page2, l, lru) {
e24f0b8f 69 list_del(&page->lru);
894bc310 70 putback_lru_page(page);
b20a3503
CL
71 count++;
72 }
73 return count;
74}
75
0697212a
CL
76/*
77 * Restore a potential migration pte to a working pte entry
78 */
04e62a29 79static void remove_migration_pte(struct vm_area_struct *vma,
0697212a
CL
80 struct page *old, struct page *new)
81{
82 struct mm_struct *mm = vma->vm_mm;
83 swp_entry_t entry;
84 pgd_t *pgd;
85 pud_t *pud;
86 pmd_t *pmd;
87 pte_t *ptep, pte;
88 spinlock_t *ptl;
04e62a29
CL
89 unsigned long addr = page_address_in_vma(new, vma);
90
91 if (addr == -EFAULT)
92 return;
0697212a
CL
93
94 pgd = pgd_offset(mm, addr);
95 if (!pgd_present(*pgd))
96 return;
97
98 pud = pud_offset(pgd, addr);
99 if (!pud_present(*pud))
100 return;
101
102 pmd = pmd_offset(pud, addr);
103 if (!pmd_present(*pmd))
104 return;
105
106 ptep = pte_offset_map(pmd, addr);
107
108 if (!is_swap_pte(*ptep)) {
109 pte_unmap(ptep);
110 return;
111 }
112
113 ptl = pte_lockptr(mm, pmd);
114 spin_lock(ptl);
115 pte = *ptep;
116 if (!is_swap_pte(pte))
117 goto out;
118
119 entry = pte_to_swp_entry(pte);
120
121 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
122 goto out;
123
0697212a
CL
124 get_page(new);
125 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
126 if (is_write_migration_entry(entry))
127 pte = pte_mkwrite(pte);
97ee0524 128 flush_cache_page(vma, addr, pte_pfn(pte));
0697212a 129 set_pte_at(mm, addr, ptep, pte);
04e62a29
CL
130
131 if (PageAnon(new))
132 page_add_anon_rmap(new, vma, addr);
133 else
134 page_add_file_rmap(new);
135
136 /* No need to invalidate - it was non-present before */
137 update_mmu_cache(vma, addr, pte);
04e62a29 138
0697212a
CL
139out:
140 pte_unmap_unlock(ptep, ptl);
141}
142
143/*
04e62a29
CL
144 * Note that remove_file_migration_ptes will only work on regular mappings,
145 * Nonlinear mappings do not use migration entries.
146 */
147static void remove_file_migration_ptes(struct page *old, struct page *new)
148{
149 struct vm_area_struct *vma;
abfc3488 150 struct address_space *mapping = new->mapping;
04e62a29
CL
151 struct prio_tree_iter iter;
152 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
153
154 if (!mapping)
155 return;
156
157 spin_lock(&mapping->i_mmap_lock);
158
159 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
160 remove_migration_pte(vma, old, new);
161
162 spin_unlock(&mapping->i_mmap_lock);
163}
164
165/*
0697212a
CL
166 * Must hold mmap_sem lock on at least one of the vmas containing
167 * the page so that the anon_vma cannot vanish.
168 */
04e62a29 169static void remove_anon_migration_ptes(struct page *old, struct page *new)
0697212a
CL
170{
171 struct anon_vma *anon_vma;
172 struct vm_area_struct *vma;
173 unsigned long mapping;
174
175 mapping = (unsigned long)new->mapping;
176
177 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
178 return;
179
180 /*
181 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
182 */
183 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
184 spin_lock(&anon_vma->lock);
185
186 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
04e62a29 187 remove_migration_pte(vma, old, new);
0697212a
CL
188
189 spin_unlock(&anon_vma->lock);
190}
191
04e62a29
CL
192/*
193 * Get rid of all migration entries and replace them by
194 * references to the indicated page.
195 */
196static void remove_migration_ptes(struct page *old, struct page *new)
197{
198 if (PageAnon(new))
199 remove_anon_migration_ptes(old, new);
200 else
201 remove_file_migration_ptes(old, new);
202}
203
0697212a
CL
204/*
205 * Something used the pte of a page under migration. We need to
206 * get to the page and wait until migration is finished.
207 * When we return from this function the fault will be retried.
208 *
209 * This function is called from do_swap_page().
210 */
211void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
212 unsigned long address)
213{
214 pte_t *ptep, pte;
215 spinlock_t *ptl;
216 swp_entry_t entry;
217 struct page *page;
218
219 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
220 pte = *ptep;
221 if (!is_swap_pte(pte))
222 goto out;
223
224 entry = pte_to_swp_entry(pte);
225 if (!is_migration_entry(entry))
226 goto out;
227
228 page = migration_entry_to_page(entry);
229
e286781d
NP
230 /*
231 * Once radix-tree replacement of page migration started, page_count
232 * *must* be zero. And, we don't want to call wait_on_page_locked()
233 * against a page without get_page().
234 * So, we use get_page_unless_zero(), here. Even failed, page fault
235 * will occur again.
236 */
237 if (!get_page_unless_zero(page))
238 goto out;
0697212a
CL
239 pte_unmap_unlock(ptep, ptl);
240 wait_on_page_locked(page);
241 put_page(page);
242 return;
243out:
244 pte_unmap_unlock(ptep, ptl);
245}
246
b20a3503 247/*
c3fcf8a5 248 * Replace the page in the mapping.
5b5c7120
CL
249 *
250 * The number of remaining references must be:
251 * 1 for anonymous pages without a mapping
252 * 2 for pages with a mapping
266cf658 253 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
b20a3503 254 */
2d1db3b1
CL
255static int migrate_page_move_mapping(struct address_space *mapping,
256 struct page *newpage, struct page *page)
b20a3503 257{
e286781d 258 int expected_count;
7cf9c2c7 259 void **pslot;
b20a3503 260
6c5240ae 261 if (!mapping) {
0e8c7d0f 262 /* Anonymous page without mapping */
6c5240ae
CL
263 if (page_count(page) != 1)
264 return -EAGAIN;
265 return 0;
266 }
267
19fd6231 268 spin_lock_irq(&mapping->tree_lock);
b20a3503 269
7cf9c2c7
NP
270 pslot = radix_tree_lookup_slot(&mapping->page_tree,
271 page_index(page));
b20a3503 272
266cf658 273 expected_count = 2 + !!page_has_private(page);
e286781d 274 if (page_count(page) != expected_count ||
7cf9c2c7 275 (struct page *)radix_tree_deref_slot(pslot) != page) {
19fd6231 276 spin_unlock_irq(&mapping->tree_lock);
e23ca00b 277 return -EAGAIN;
b20a3503
CL
278 }
279
e286781d 280 if (!page_freeze_refs(page, expected_count)) {
19fd6231 281 spin_unlock_irq(&mapping->tree_lock);
e286781d
NP
282 return -EAGAIN;
283 }
284
b20a3503
CL
285 /*
286 * Now we know that no one else is looking at the page.
b20a3503 287 */
7cf9c2c7 288 get_page(newpage); /* add cache reference */
b20a3503
CL
289 if (PageSwapCache(page)) {
290 SetPageSwapCache(newpage);
291 set_page_private(newpage, page_private(page));
292 }
293
7cf9c2c7
NP
294 radix_tree_replace_slot(pslot, newpage);
295
e286781d 296 page_unfreeze_refs(page, expected_count);
7cf9c2c7
NP
297 /*
298 * Drop cache reference from old page.
299 * We know this isn't the last reference.
300 */
b20a3503 301 __put_page(page);
7cf9c2c7 302
0e8c7d0f
CL
303 /*
304 * If moved to a different zone then also account
305 * the page for that zone. Other VM counters will be
306 * taken care of when we establish references to the
307 * new page and drop references to the old page.
308 *
309 * Note that anonymous pages are accounted for
310 * via NR_FILE_PAGES and NR_ANON_PAGES if they
311 * are mapped to swap space.
312 */
313 __dec_zone_page_state(page, NR_FILE_PAGES);
314 __inc_zone_page_state(newpage, NR_FILE_PAGES);
315
19fd6231 316 spin_unlock_irq(&mapping->tree_lock);
b20a3503
CL
317
318 return 0;
319}
b20a3503
CL
320
321/*
322 * Copy the page to its new location
323 */
e7340f73 324static void migrate_page_copy(struct page *newpage, struct page *page)
b20a3503 325{
b7abea96
KH
326 int anon;
327
b20a3503
CL
328 copy_highpage(newpage, page);
329
330 if (PageError(page))
331 SetPageError(newpage);
332 if (PageReferenced(page))
333 SetPageReferenced(newpage);
334 if (PageUptodate(page))
335 SetPageUptodate(newpage);
894bc310
LS
336 if (TestClearPageActive(page)) {
337 VM_BUG_ON(PageUnevictable(page));
b20a3503 338 SetPageActive(newpage);
894bc310
LS
339 } else
340 unevictable_migrate_page(newpage, page);
b20a3503
CL
341 if (PageChecked(page))
342 SetPageChecked(newpage);
343 if (PageMappedToDisk(page))
344 SetPageMappedToDisk(newpage);
345
346 if (PageDirty(page)) {
347 clear_page_dirty_for_io(page);
3a902c5f
NP
348 /*
349 * Want to mark the page and the radix tree as dirty, and
350 * redo the accounting that clear_page_dirty_for_io undid,
351 * but we can't use set_page_dirty because that function
352 * is actually a signal that all of the page has become dirty.
353 * Wheras only part of our page may be dirty.
354 */
355 __set_page_dirty_nobuffers(newpage);
b20a3503
CL
356 }
357
b291f000
NP
358 mlock_migrate_page(newpage, page);
359
b20a3503 360 ClearPageSwapCache(page);
b20a3503
CL
361 ClearPagePrivate(page);
362 set_page_private(page, 0);
b7abea96
KH
363 /* page->mapping contains a flag for PageAnon() */
364 anon = PageAnon(page);
b20a3503
CL
365 page->mapping = NULL;
366
367 /*
368 * If any waiters have accumulated on the new page then
369 * wake them up.
370 */
371 if (PageWriteback(newpage))
372 end_page_writeback(newpage);
373}
b20a3503 374
1d8b85cc
CL
375/************************************************************
376 * Migration functions
377 ***********************************************************/
378
379/* Always fail migration. Used for mappings that are not movable */
2d1db3b1
CL
380int fail_migrate_page(struct address_space *mapping,
381 struct page *newpage, struct page *page)
1d8b85cc
CL
382{
383 return -EIO;
384}
385EXPORT_SYMBOL(fail_migrate_page);
386
b20a3503
CL
387/*
388 * Common logic to directly migrate a single page suitable for
266cf658 389 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
390 *
391 * Pages are locked upon entry and exit.
392 */
2d1db3b1
CL
393int migrate_page(struct address_space *mapping,
394 struct page *newpage, struct page *page)
b20a3503
CL
395{
396 int rc;
397
398 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
399
2d1db3b1 400 rc = migrate_page_move_mapping(mapping, newpage, page);
b20a3503
CL
401
402 if (rc)
403 return rc;
404
405 migrate_page_copy(newpage, page);
b20a3503
CL
406 return 0;
407}
408EXPORT_SYMBOL(migrate_page);
409
9361401e 410#ifdef CONFIG_BLOCK
1d8b85cc
CL
411/*
412 * Migration function for pages with buffers. This function can only be used
413 * if the underlying filesystem guarantees that no other references to "page"
414 * exist.
415 */
2d1db3b1
CL
416int buffer_migrate_page(struct address_space *mapping,
417 struct page *newpage, struct page *page)
1d8b85cc 418{
1d8b85cc
CL
419 struct buffer_head *bh, *head;
420 int rc;
421
1d8b85cc 422 if (!page_has_buffers(page))
2d1db3b1 423 return migrate_page(mapping, newpage, page);
1d8b85cc
CL
424
425 head = page_buffers(page);
426
2d1db3b1 427 rc = migrate_page_move_mapping(mapping, newpage, page);
1d8b85cc
CL
428
429 if (rc)
430 return rc;
431
432 bh = head;
433 do {
434 get_bh(bh);
435 lock_buffer(bh);
436 bh = bh->b_this_page;
437
438 } while (bh != head);
439
440 ClearPagePrivate(page);
441 set_page_private(newpage, page_private(page));
442 set_page_private(page, 0);
443 put_page(page);
444 get_page(newpage);
445
446 bh = head;
447 do {
448 set_bh_page(bh, newpage, bh_offset(bh));
449 bh = bh->b_this_page;
450
451 } while (bh != head);
452
453 SetPagePrivate(newpage);
454
455 migrate_page_copy(newpage, page);
456
457 bh = head;
458 do {
459 unlock_buffer(bh);
460 put_bh(bh);
461 bh = bh->b_this_page;
462
463 } while (bh != head);
464
465 return 0;
466}
467EXPORT_SYMBOL(buffer_migrate_page);
9361401e 468#endif
1d8b85cc 469
04e62a29
CL
470/*
471 * Writeback a page to clean the dirty state
472 */
473static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 474{
04e62a29
CL
475 struct writeback_control wbc = {
476 .sync_mode = WB_SYNC_NONE,
477 .nr_to_write = 1,
478 .range_start = 0,
479 .range_end = LLONG_MAX,
480 .nonblocking = 1,
481 .for_reclaim = 1
482 };
483 int rc;
484
485 if (!mapping->a_ops->writepage)
486 /* No write method for the address space */
487 return -EINVAL;
488
489 if (!clear_page_dirty_for_io(page))
490 /* Someone else already triggered a write */
491 return -EAGAIN;
492
8351a6e4 493 /*
04e62a29
CL
494 * A dirty page may imply that the underlying filesystem has
495 * the page on some queue. So the page must be clean for
496 * migration. Writeout may mean we loose the lock and the
497 * page state is no longer what we checked for earlier.
498 * At this point we know that the migration attempt cannot
499 * be successful.
8351a6e4 500 */
04e62a29 501 remove_migration_ptes(page, page);
8351a6e4 502
04e62a29 503 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 504
04e62a29
CL
505 if (rc != AOP_WRITEPAGE_ACTIVATE)
506 /* unlocked. Relock */
507 lock_page(page);
508
bda8550d 509 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
510}
511
512/*
513 * Default handling if a filesystem does not provide a migration function.
514 */
515static int fallback_migrate_page(struct address_space *mapping,
516 struct page *newpage, struct page *page)
517{
518 if (PageDirty(page))
519 return writeout(mapping, page);
8351a6e4
CL
520
521 /*
522 * Buffers may be managed in a filesystem specific way.
523 * We must have no buffers or drop them.
524 */
266cf658 525 if (page_has_private(page) &&
8351a6e4
CL
526 !try_to_release_page(page, GFP_KERNEL))
527 return -EAGAIN;
528
529 return migrate_page(mapping, newpage, page);
530}
531
e24f0b8f
CL
532/*
533 * Move a page to a newly allocated page
534 * The page is locked and all ptes have been successfully removed.
535 *
536 * The new page will have replaced the old page if this function
537 * is successful.
894bc310
LS
538 *
539 * Return value:
540 * < 0 - error code
541 * == 0 - success
e24f0b8f
CL
542 */
543static int move_to_new_page(struct page *newpage, struct page *page)
544{
545 struct address_space *mapping;
546 int rc;
547
548 /*
549 * Block others from accessing the page when we get around to
550 * establishing additional references. We are the only one
551 * holding a reference to the new page at this point.
552 */
529ae9aa 553 if (!trylock_page(newpage))
e24f0b8f
CL
554 BUG();
555
556 /* Prepare mapping for the new page.*/
557 newpage->index = page->index;
558 newpage->mapping = page->mapping;
b2e18538
RR
559 if (PageSwapBacked(page))
560 SetPageSwapBacked(newpage);
e24f0b8f
CL
561
562 mapping = page_mapping(page);
563 if (!mapping)
564 rc = migrate_page(mapping, newpage, page);
565 else if (mapping->a_ops->migratepage)
566 /*
567 * Most pages have a mapping and most filesystems
568 * should provide a migration function. Anonymous
569 * pages are part of swap space which also has its
570 * own migration function. This is the most common
571 * path for page migration.
572 */
573 rc = mapping->a_ops->migratepage(mapping,
574 newpage, page);
575 else
576 rc = fallback_migrate_page(mapping, newpage, page);
577
ae41be37 578 if (!rc) {
e24f0b8f 579 remove_migration_ptes(page, newpage);
ae41be37 580 } else
e24f0b8f
CL
581 newpage->mapping = NULL;
582
583 unlock_page(newpage);
584
585 return rc;
586}
587
588/*
589 * Obtain the lock on page, remove all ptes and migrate the page
590 * to the newly allocated page in newpage.
591 */
95a402c3
CL
592static int unmap_and_move(new_page_t get_new_page, unsigned long private,
593 struct page *page, int force)
e24f0b8f
CL
594{
595 int rc = 0;
742755a1
CL
596 int *result = NULL;
597 struct page *newpage = get_new_page(page, private, &result);
989f89c5 598 int rcu_locked = 0;
ae41be37 599 int charge = 0;
01b1ae63 600 struct mem_cgroup *mem;
95a402c3
CL
601
602 if (!newpage)
603 return -ENOMEM;
e24f0b8f 604
894bc310 605 if (page_count(page) == 1) {
e24f0b8f 606 /* page was freed from under us. So we are done. */
95a402c3 607 goto move_newpage;
894bc310 608 }
e24f0b8f 609
e8589cc1 610 /* prepare cgroup just returns 0 or -ENOMEM */
e24f0b8f 611 rc = -EAGAIN;
01b1ae63 612
529ae9aa 613 if (!trylock_page(page)) {
e24f0b8f 614 if (!force)
95a402c3 615 goto move_newpage;
e24f0b8f
CL
616 lock_page(page);
617 }
618
01b1ae63
KH
619 /* charge against new page */
620 charge = mem_cgroup_prepare_migration(page, &mem);
621 if (charge == -ENOMEM) {
622 rc = -ENOMEM;
623 goto unlock;
624 }
625 BUG_ON(charge);
626
e24f0b8f
CL
627 if (PageWriteback(page)) {
628 if (!force)
01b1ae63 629 goto uncharge;
e24f0b8f
CL
630 wait_on_page_writeback(page);
631 }
e24f0b8f 632 /*
dc386d4d
KH
633 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
634 * we cannot notice that anon_vma is freed while we migrates a page.
635 * This rcu_read_lock() delays freeing anon_vma pointer until the end
636 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
637 * File Caches may use write_page() or lock_page() in migration, then,
638 * just care Anon page here.
dc386d4d 639 */
989f89c5
KH
640 if (PageAnon(page)) {
641 rcu_read_lock();
642 rcu_locked = 1;
643 }
62e1c553 644
dc386d4d 645 /*
62e1c553
SL
646 * Corner case handling:
647 * 1. When a new swap-cache page is read into, it is added to the LRU
648 * and treated as swapcache but it has no rmap yet.
649 * Calling try_to_unmap() against a page->mapping==NULL page will
650 * trigger a BUG. So handle it here.
651 * 2. An orphaned page (see truncate_complete_page) might have
652 * fs-private metadata. The page can be picked up due to memory
653 * offlining. Everywhere else except page reclaim, the page is
654 * invisible to the vm, so the page can not be migrated. So try to
655 * free the metadata, so the page can be freed.
e24f0b8f 656 */
62e1c553 657 if (!page->mapping) {
266cf658 658 if (!PageAnon(page) && page_has_private(page)) {
62e1c553
SL
659 /*
660 * Go direct to try_to_free_buffers() here because
661 * a) that's what try_to_release_page() would do anyway
662 * b) we may be under rcu_read_lock() here, so we can't
663 * use GFP_KERNEL which is what try_to_release_page()
664 * needs to be effective.
665 */
666 try_to_free_buffers(page);
abfc3488 667 goto rcu_unlock;
62e1c553 668 }
abfc3488 669 goto skip_unmap;
62e1c553
SL
670 }
671
dc386d4d 672 /* Establish migration ptes or remove ptes */
e6a1530d 673 try_to_unmap(page, 1);
dc386d4d 674
abfc3488 675skip_unmap:
e6a1530d
CL
676 if (!page_mapped(page))
677 rc = move_to_new_page(newpage, page);
e24f0b8f 678
e8589cc1 679 if (rc)
e24f0b8f 680 remove_migration_ptes(page, page);
dc386d4d 681rcu_unlock:
989f89c5
KH
682 if (rcu_locked)
683 rcu_read_unlock();
01b1ae63
KH
684uncharge:
685 if (!charge)
686 mem_cgroup_end_migration(mem, page, newpage);
e24f0b8f
CL
687unlock:
688 unlock_page(page);
95a402c3 689
e24f0b8f 690 if (rc != -EAGAIN) {
aaa994b3
CL
691 /*
692 * A page that has been migrated has all references
693 * removed and will be freed. A page that has not been
694 * migrated will have kepts its references and be
695 * restored.
696 */
697 list_del(&page->lru);
894bc310 698 putback_lru_page(page);
e24f0b8f 699 }
95a402c3
CL
700
701move_newpage:
894bc310 702
95a402c3
CL
703 /*
704 * Move the new page to the LRU. If migration was not successful
705 * then this will free the page.
706 */
894bc310
LS
707 putback_lru_page(newpage);
708
742755a1
CL
709 if (result) {
710 if (rc)
711 *result = rc;
712 else
713 *result = page_to_nid(newpage);
714 }
e24f0b8f
CL
715 return rc;
716}
717
b20a3503
CL
718/*
719 * migrate_pages
720 *
95a402c3
CL
721 * The function takes one list of pages to migrate and a function
722 * that determines from the page to be migrated and the private data
723 * the target of the move and allocates the page.
b20a3503
CL
724 *
725 * The function returns after 10 attempts or if no pages
726 * are movable anymore because to has become empty
aaa994b3 727 * or no retryable pages exist anymore. All pages will be
e9534b3f 728 * returned to the LRU or freed.
b20a3503 729 *
95a402c3 730 * Return: Number of pages not migrated or error code.
b20a3503 731 */
95a402c3
CL
732int migrate_pages(struct list_head *from,
733 new_page_t get_new_page, unsigned long private)
b20a3503 734{
e24f0b8f 735 int retry = 1;
b20a3503
CL
736 int nr_failed = 0;
737 int pass = 0;
738 struct page *page;
739 struct page *page2;
740 int swapwrite = current->flags & PF_SWAPWRITE;
741 int rc;
742
743 if (!swapwrite)
744 current->flags |= PF_SWAPWRITE;
745
e24f0b8f
CL
746 for(pass = 0; pass < 10 && retry; pass++) {
747 retry = 0;
b20a3503 748
e24f0b8f 749 list_for_each_entry_safe(page, page2, from, lru) {
e24f0b8f 750 cond_resched();
2d1db3b1 751
95a402c3
CL
752 rc = unmap_and_move(get_new_page, private,
753 page, pass > 2);
2d1db3b1 754
e24f0b8f 755 switch(rc) {
95a402c3
CL
756 case -ENOMEM:
757 goto out;
e24f0b8f 758 case -EAGAIN:
2d1db3b1 759 retry++;
e24f0b8f
CL
760 break;
761 case 0:
e24f0b8f
CL
762 break;
763 default:
2d1db3b1 764 /* Permanent failure */
2d1db3b1 765 nr_failed++;
e24f0b8f 766 break;
2d1db3b1 767 }
b20a3503
CL
768 }
769 }
95a402c3
CL
770 rc = 0;
771out:
b20a3503
CL
772 if (!swapwrite)
773 current->flags &= ~PF_SWAPWRITE;
774
aaa994b3 775 putback_lru_pages(from);
b20a3503 776
95a402c3
CL
777 if (rc)
778 return rc;
b20a3503 779
95a402c3 780 return nr_failed + retry;
b20a3503 781}
95a402c3 782
742755a1
CL
783#ifdef CONFIG_NUMA
784/*
785 * Move a list of individual pages
786 */
787struct page_to_node {
788 unsigned long addr;
789 struct page *page;
790 int node;
791 int status;
792};
793
794static struct page *new_page_node(struct page *p, unsigned long private,
795 int **result)
796{
797 struct page_to_node *pm = (struct page_to_node *)private;
798
799 while (pm->node != MAX_NUMNODES && pm->page != p)
800 pm++;
801
802 if (pm->node == MAX_NUMNODES)
803 return NULL;
804
805 *result = &pm->status;
806
6484eb3e 807 return alloc_pages_exact_node(pm->node,
769848c0 808 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
742755a1
CL
809}
810
811/*
812 * Move a set of pages as indicated in the pm array. The addr
813 * field must be set to the virtual address of the page to be moved
814 * and the node number must contain a valid target node.
5e9a0f02 815 * The pm array ends with node = MAX_NUMNODES.
742755a1 816 */
5e9a0f02
BG
817static int do_move_page_to_node_array(struct mm_struct *mm,
818 struct page_to_node *pm,
819 int migrate_all)
742755a1
CL
820{
821 int err;
822 struct page_to_node *pp;
823 LIST_HEAD(pagelist);
824
825 down_read(&mm->mmap_sem);
826
827 /*
828 * Build a list of pages to migrate
829 */
742755a1
CL
830 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
831 struct vm_area_struct *vma;
832 struct page *page;
833
742755a1
CL
834 err = -EFAULT;
835 vma = find_vma(mm, pp->addr);
0dc952dc 836 if (!vma || !vma_migratable(vma))
742755a1
CL
837 goto set_status;
838
839 page = follow_page(vma, pp->addr, FOLL_GET);
89f5b7da
LT
840
841 err = PTR_ERR(page);
842 if (IS_ERR(page))
843 goto set_status;
844
742755a1
CL
845 err = -ENOENT;
846 if (!page)
847 goto set_status;
848
849 if (PageReserved(page)) /* Check for zero page */
850 goto put_and_set;
851
852 pp->page = page;
853 err = page_to_nid(page);
854
855 if (err == pp->node)
856 /*
857 * Node already in the right place
858 */
859 goto put_and_set;
860
861 err = -EACCES;
862 if (page_mapcount(page) > 1 &&
863 !migrate_all)
864 goto put_and_set;
865
62695a84
NP
866 err = isolate_lru_page(page);
867 if (!err)
868 list_add_tail(&page->lru, &pagelist);
742755a1
CL
869put_and_set:
870 /*
871 * Either remove the duplicate refcount from
872 * isolate_lru_page() or drop the page ref if it was
873 * not isolated.
874 */
875 put_page(page);
876set_status:
877 pp->status = err;
878 }
879
e78bbfa8 880 err = 0;
742755a1
CL
881 if (!list_empty(&pagelist))
882 err = migrate_pages(&pagelist, new_page_node,
883 (unsigned long)pm);
742755a1
CL
884
885 up_read(&mm->mmap_sem);
886 return err;
887}
888
5e9a0f02
BG
889/*
890 * Migrate an array of page address onto an array of nodes and fill
891 * the corresponding array of status.
892 */
893static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
894 unsigned long nr_pages,
895 const void __user * __user *pages,
896 const int __user *nodes,
897 int __user *status, int flags)
898{
3140a227 899 struct page_to_node *pm;
5e9a0f02 900 nodemask_t task_nodes;
3140a227
BG
901 unsigned long chunk_nr_pages;
902 unsigned long chunk_start;
903 int err;
5e9a0f02
BG
904
905 task_nodes = cpuset_mems_allowed(task);
906
3140a227
BG
907 err = -ENOMEM;
908 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
909 if (!pm)
5e9a0f02 910 goto out;
35282a2d
BG
911
912 migrate_prep();
913
5e9a0f02 914 /*
3140a227
BG
915 * Store a chunk of page_to_node array in a page,
916 * but keep the last one as a marker
5e9a0f02 917 */
3140a227 918 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
5e9a0f02 919
3140a227
BG
920 for (chunk_start = 0;
921 chunk_start < nr_pages;
922 chunk_start += chunk_nr_pages) {
923 int j;
5e9a0f02 924
3140a227
BG
925 if (chunk_start + chunk_nr_pages > nr_pages)
926 chunk_nr_pages = nr_pages - chunk_start;
927
928 /* fill the chunk pm with addrs and nodes from user-space */
929 for (j = 0; j < chunk_nr_pages; j++) {
930 const void __user *p;
5e9a0f02
BG
931 int node;
932
3140a227
BG
933 err = -EFAULT;
934 if (get_user(p, pages + j + chunk_start))
935 goto out_pm;
936 pm[j].addr = (unsigned long) p;
937
938 if (get_user(node, nodes + j + chunk_start))
5e9a0f02
BG
939 goto out_pm;
940
941 err = -ENODEV;
942 if (!node_state(node, N_HIGH_MEMORY))
943 goto out_pm;
944
945 err = -EACCES;
946 if (!node_isset(node, task_nodes))
947 goto out_pm;
948
3140a227
BG
949 pm[j].node = node;
950 }
951
952 /* End marker for this chunk */
953 pm[chunk_nr_pages].node = MAX_NUMNODES;
954
955 /* Migrate this chunk */
956 err = do_move_page_to_node_array(mm, pm,
957 flags & MPOL_MF_MOVE_ALL);
958 if (err < 0)
959 goto out_pm;
5e9a0f02 960
5e9a0f02 961 /* Return status information */
3140a227
BG
962 for (j = 0; j < chunk_nr_pages; j++)
963 if (put_user(pm[j].status, status + j + chunk_start)) {
5e9a0f02 964 err = -EFAULT;
3140a227
BG
965 goto out_pm;
966 }
967 }
968 err = 0;
5e9a0f02
BG
969
970out_pm:
3140a227 971 free_page((unsigned long)pm);
5e9a0f02
BG
972out:
973 return err;
974}
975
742755a1 976/*
2f007e74 977 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 978 */
80bba129
BG
979static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
980 const void __user **pages, int *status)
742755a1 981{
2f007e74 982 unsigned long i;
2f007e74 983
742755a1
CL
984 down_read(&mm->mmap_sem);
985
2f007e74 986 for (i = 0; i < nr_pages; i++) {
80bba129 987 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
988 struct vm_area_struct *vma;
989 struct page *page;
c095adbc 990 int err = -EFAULT;
2f007e74
BG
991
992 vma = find_vma(mm, addr);
742755a1
CL
993 if (!vma)
994 goto set_status;
995
2f007e74 996 page = follow_page(vma, addr, 0);
89f5b7da
LT
997
998 err = PTR_ERR(page);
999 if (IS_ERR(page))
1000 goto set_status;
1001
742755a1
CL
1002 err = -ENOENT;
1003 /* Use PageReserved to check for zero page */
1004 if (!page || PageReserved(page))
1005 goto set_status;
1006
1007 err = page_to_nid(page);
1008set_status:
80bba129
BG
1009 *status = err;
1010
1011 pages++;
1012 status++;
1013 }
1014
1015 up_read(&mm->mmap_sem);
1016}
1017
1018/*
1019 * Determine the nodes of a user array of pages and store it in
1020 * a user array of status.
1021 */
1022static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1023 const void __user * __user *pages,
1024 int __user *status)
1025{
1026#define DO_PAGES_STAT_CHUNK_NR 16
1027 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1028 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1029 unsigned long i, chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1030 int err;
1031
1032 for (i = 0; i < nr_pages; i += chunk_nr) {
1033 if (chunk_nr + i > nr_pages)
1034 chunk_nr = nr_pages - i;
1035
1036 err = copy_from_user(chunk_pages, &pages[i],
1037 chunk_nr * sizeof(*chunk_pages));
1038 if (err) {
1039 err = -EFAULT;
1040 goto out;
1041 }
1042
1043 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1044
1045 err = copy_to_user(&status[i], chunk_status,
1046 chunk_nr * sizeof(*chunk_status));
1047 if (err) {
1048 err = -EFAULT;
1049 goto out;
1050 }
742755a1 1051 }
2f007e74 1052 err = 0;
742755a1 1053
2f007e74 1054out:
2f007e74 1055 return err;
742755a1
CL
1056}
1057
1058/*
1059 * Move a list of pages in the address space of the currently executing
1060 * process.
1061 */
938bb9f5
HC
1062SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1063 const void __user * __user *, pages,
1064 const int __user *, nodes,
1065 int __user *, status, int, flags)
742755a1 1066{
c69e8d9c 1067 const struct cred *cred = current_cred(), *tcred;
742755a1 1068 struct task_struct *task;
742755a1 1069 struct mm_struct *mm;
5e9a0f02 1070 int err;
742755a1
CL
1071
1072 /* Check flags */
1073 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1074 return -EINVAL;
1075
1076 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1077 return -EPERM;
1078
1079 /* Find the mm_struct */
1080 read_lock(&tasklist_lock);
228ebcbe 1081 task = pid ? find_task_by_vpid(pid) : current;
742755a1
CL
1082 if (!task) {
1083 read_unlock(&tasklist_lock);
1084 return -ESRCH;
1085 }
1086 mm = get_task_mm(task);
1087 read_unlock(&tasklist_lock);
1088
1089 if (!mm)
1090 return -EINVAL;
1091
1092 /*
1093 * Check if this process has the right to modify the specified
1094 * process. The right exists if the process has administrative
1095 * capabilities, superuser privileges or the same
1096 * userid as the target process.
1097 */
c69e8d9c
DH
1098 rcu_read_lock();
1099 tcred = __task_cred(task);
b6dff3ec
DH
1100 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1101 cred->uid != tcred->suid && cred->uid != tcred->uid &&
742755a1 1102 !capable(CAP_SYS_NICE)) {
c69e8d9c 1103 rcu_read_unlock();
742755a1 1104 err = -EPERM;
5e9a0f02 1105 goto out;
742755a1 1106 }
c69e8d9c 1107 rcu_read_unlock();
742755a1 1108
86c3a764
DQ
1109 err = security_task_movememory(task);
1110 if (err)
5e9a0f02 1111 goto out;
86c3a764 1112
5e9a0f02
BG
1113 if (nodes) {
1114 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1115 flags);
1116 } else {
2f007e74 1117 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1
CL
1118 }
1119
742755a1 1120out:
742755a1
CL
1121 mmput(mm);
1122 return err;
1123}
742755a1 1124
7b2259b3
CL
1125/*
1126 * Call migration functions in the vma_ops that may prepare
1127 * memory in a vm for migration. migration functions may perform
1128 * the migration for vmas that do not have an underlying page struct.
1129 */
1130int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1131 const nodemask_t *from, unsigned long flags)
1132{
1133 struct vm_area_struct *vma;
1134 int err = 0;
1135
1001c9fb 1136 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
7b2259b3
CL
1137 if (vma->vm_ops && vma->vm_ops->migrate) {
1138 err = vma->vm_ops->migrate(vma, to, from, flags);
1139 if (err)
1140 break;
1141 }
1142 }
1143 return err;
1144}
83d1674a 1145#endif