blob: 4d6ee03db94608e3bc49e5bffb0d9fdd697ec449 [file] [log] [blame]
Christoph Lameterb20a3502006-03-22 00:09:12 -08001/*
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>
12 * Christoph Lameter <clameter@sgi.com>
13 */
14
15#include <linux/migrate.h>
16#include <linux/module.h>
17#include <linux/swap.h>
Christoph Lameter06972122006-06-23 02:03:35 -070018#include <linux/swapops.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080019#include <linux/pagemap.h>
Christoph Lametere23ca002006-04-10 22:52:57 -070020#include <linux/buffer_head.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080021#include <linux/mm_inline.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070022#include <linux/nsproxy.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080023#include <linux/pagevec.h>
24#include <linux/rmap.h>
25#include <linux/topology.h>
26#include <linux/cpu.h>
27#include <linux/cpuset.h>
Christoph Lameter04e62a22006-06-23 02:03:38 -070028#include <linux/writeback.h>
Christoph Lameter742755a2006-06-23 02:03:55 -070029#include <linux/mempolicy.h>
30#include <linux/vmalloc.h>
David Quigley86c3a762006-06-23 02:04:02 -070031#include <linux/security.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080032
33#include "internal.h"
34
Christoph Lameterb20a3502006-03-22 00:09:12 -080035#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
36
37/*
38 * Isolate one page from the LRU lists. If successful put it onto
39 * the indicated list with elevated page count.
40 *
41 * Result:
42 * -EBUSY: page not on LRU list
43 * 0: page removed from LRU list and added to the specified list.
44 */
45int isolate_lru_page(struct page *page, struct list_head *pagelist)
46{
47 int ret = -EBUSY;
48
49 if (PageLRU(page)) {
50 struct zone *zone = page_zone(page);
51
52 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyuki3dd9fe82007-07-26 10:41:08 -070053 if (PageLRU(page) && get_page_unless_zero(page)) {
Christoph Lameterb20a3502006-03-22 00:09:12 -080054 ret = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -080055 ClearPageLRU(page);
56 if (PageActive(page))
57 del_page_from_active_list(zone, page);
58 else
59 del_page_from_inactive_list(zone, page);
60 list_add_tail(&page->lru, pagelist);
61 }
62 spin_unlock_irq(&zone->lru_lock);
63 }
64 return ret;
65}
66
67/*
Christoph Lameter742755a2006-06-23 02:03:55 -070068 * migrate_prep() needs to be called before we start compiling a list of pages
69 * to be migrated using isolate_lru_page().
Christoph Lameterb20a3502006-03-22 00:09:12 -080070 */
71int migrate_prep(void)
72{
Christoph Lameterb20a3502006-03-22 00:09:12 -080073 /*
74 * Clear the LRU lists so pages can be isolated.
75 * Note that pages may be moved off the LRU after we have
76 * drained them. Those pages will fail to migrate like other
77 * pages that may be busy.
78 */
79 lru_add_drain_all();
80
81 return 0;
82}
83
84static inline void move_to_lru(struct page *page)
85{
Christoph Lameterb20a3502006-03-22 00:09:12 -080086 if (PageActive(page)) {
87 /*
88 * lru_cache_add_active checks that
89 * the PG_active bit is off.
90 */
91 ClearPageActive(page);
92 lru_cache_add_active(page);
93 } else {
94 lru_cache_add(page);
95 }
96 put_page(page);
97}
98
99/*
100 * Add isolated pages on the list back to the LRU.
101 *
102 * returns the number of pages put back.
103 */
104int putback_lru_pages(struct list_head *l)
105{
106 struct page *page;
107 struct page *page2;
108 int count = 0;
109
110 list_for_each_entry_safe(page, page2, l, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700111 list_del(&page->lru);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800112 move_to_lru(page);
113 count++;
114 }
115 return count;
116}
117
Christoph Lameter06972122006-06-23 02:03:35 -0700118static inline int is_swap_pte(pte_t pte)
119{
120 return !pte_none(pte) && !pte_present(pte) && !pte_file(pte);
121}
122
123/*
124 * Restore a potential migration pte to a working pte entry
125 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700126static void remove_migration_pte(struct vm_area_struct *vma,
Christoph Lameter06972122006-06-23 02:03:35 -0700127 struct page *old, struct page *new)
128{
129 struct mm_struct *mm = vma->vm_mm;
130 swp_entry_t entry;
131 pgd_t *pgd;
132 pud_t *pud;
133 pmd_t *pmd;
134 pte_t *ptep, pte;
135 spinlock_t *ptl;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700136 unsigned long addr = page_address_in_vma(new, vma);
137
138 if (addr == -EFAULT)
139 return;
Christoph Lameter06972122006-06-23 02:03:35 -0700140
141 pgd = pgd_offset(mm, addr);
142 if (!pgd_present(*pgd))
143 return;
144
145 pud = pud_offset(pgd, addr);
146 if (!pud_present(*pud))
147 return;
148
149 pmd = pmd_offset(pud, addr);
150 if (!pmd_present(*pmd))
151 return;
152
153 ptep = pte_offset_map(pmd, addr);
154
155 if (!is_swap_pte(*ptep)) {
156 pte_unmap(ptep);
157 return;
158 }
159
160 ptl = pte_lockptr(mm, pmd);
161 spin_lock(ptl);
162 pte = *ptep;
163 if (!is_swap_pte(pte))
164 goto out;
165
166 entry = pte_to_swp_entry(pte);
167
168 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
169 goto out;
170
Christoph Lameter06972122006-06-23 02:03:35 -0700171 get_page(new);
172 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
173 if (is_write_migration_entry(entry))
174 pte = pte_mkwrite(pte);
KAMEZAWA Hiroyuki97ee0522007-10-16 01:25:43 -0700175 flush_cache_page(vma, addr, pte_pfn(pte));
Christoph Lameter06972122006-06-23 02:03:35 -0700176 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700177
178 if (PageAnon(new))
179 page_add_anon_rmap(new, vma, addr);
180 else
181 page_add_file_rmap(new);
182
183 /* No need to invalidate - it was non-present before */
184 update_mmu_cache(vma, addr, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700185
Christoph Lameter06972122006-06-23 02:03:35 -0700186out:
187 pte_unmap_unlock(ptep, ptl);
188}
189
190/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700191 * Note that remove_file_migration_ptes will only work on regular mappings,
192 * Nonlinear mappings do not use migration entries.
193 */
194static void remove_file_migration_ptes(struct page *old, struct page *new)
195{
196 struct vm_area_struct *vma;
197 struct address_space *mapping = page_mapping(new);
198 struct prio_tree_iter iter;
199 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
200
201 if (!mapping)
202 return;
203
204 spin_lock(&mapping->i_mmap_lock);
205
206 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
207 remove_migration_pte(vma, old, new);
208
209 spin_unlock(&mapping->i_mmap_lock);
210}
211
212/*
Christoph Lameter06972122006-06-23 02:03:35 -0700213 * Must hold mmap_sem lock on at least one of the vmas containing
214 * the page so that the anon_vma cannot vanish.
215 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700216static void remove_anon_migration_ptes(struct page *old, struct page *new)
Christoph Lameter06972122006-06-23 02:03:35 -0700217{
218 struct anon_vma *anon_vma;
219 struct vm_area_struct *vma;
220 unsigned long mapping;
221
222 mapping = (unsigned long)new->mapping;
223
224 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
225 return;
226
227 /*
228 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
229 */
230 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
231 spin_lock(&anon_vma->lock);
232
233 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
Christoph Lameter04e62a22006-06-23 02:03:38 -0700234 remove_migration_pte(vma, old, new);
Christoph Lameter06972122006-06-23 02:03:35 -0700235
236 spin_unlock(&anon_vma->lock);
237}
238
239/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700240 * Get rid of all migration entries and replace them by
241 * references to the indicated page.
242 */
243static void remove_migration_ptes(struct page *old, struct page *new)
244{
245 if (PageAnon(new))
246 remove_anon_migration_ptes(old, new);
247 else
248 remove_file_migration_ptes(old, new);
249}
250
251/*
Christoph Lameter06972122006-06-23 02:03:35 -0700252 * Something used the pte of a page under migration. We need to
253 * get to the page and wait until migration is finished.
254 * When we return from this function the fault will be retried.
255 *
256 * This function is called from do_swap_page().
257 */
258void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
259 unsigned long address)
260{
261 pte_t *ptep, pte;
262 spinlock_t *ptl;
263 swp_entry_t entry;
264 struct page *page;
265
266 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
267 pte = *ptep;
268 if (!is_swap_pte(pte))
269 goto out;
270
271 entry = pte_to_swp_entry(pte);
272 if (!is_migration_entry(entry))
273 goto out;
274
275 page = migration_entry_to_page(entry);
276
277 get_page(page);
278 pte_unmap_unlock(ptep, ptl);
279 wait_on_page_locked(page);
280 put_page(page);
281 return;
282out:
283 pte_unmap_unlock(ptep, ptl);
284}
285
Christoph Lameterb20a3502006-03-22 00:09:12 -0800286/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700287 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700288 *
289 * The number of remaining references must be:
290 * 1 for anonymous pages without a mapping
291 * 2 for pages with a mapping
292 * 3 for pages with a mapping and PagePrivate set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800293 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700294static int migrate_page_move_mapping(struct address_space *mapping,
295 struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800296{
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800297 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800298
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700299 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700300 /* Anonymous page without mapping */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700301 if (page_count(page) != 1)
302 return -EAGAIN;
303 return 0;
304 }
305
Christoph Lameterb20a3502006-03-22 00:09:12 -0800306 write_lock_irq(&mapping->tree_lock);
307
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800308 pslot = radix_tree_lookup_slot(&mapping->page_tree,
309 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800310
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700311 if (page_count(page) != 2 + !!PagePrivate(page) ||
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800312 (struct page *)radix_tree_deref_slot(pslot) != page) {
Christoph Lameterb20a3502006-03-22 00:09:12 -0800313 write_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700314 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800315 }
316
317 /*
318 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800319 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800320 get_page(newpage); /* add cache reference */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700321#ifdef CONFIG_SWAP
Christoph Lameterb20a3502006-03-22 00:09:12 -0800322 if (PageSwapCache(page)) {
323 SetPageSwapCache(newpage);
324 set_page_private(newpage, page_private(page));
325 }
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700326#endif
Christoph Lameterb20a3502006-03-22 00:09:12 -0800327
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800328 radix_tree_replace_slot(pslot, newpage);
329
330 /*
331 * Drop cache reference from old page.
332 * We know this isn't the last reference.
333 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800334 __put_page(page);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800335
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700336 /*
337 * If moved to a different zone then also account
338 * the page for that zone. Other VM counters will be
339 * taken care of when we establish references to the
340 * new page and drop references to the old page.
341 *
342 * Note that anonymous pages are accounted for
343 * via NR_FILE_PAGES and NR_ANON_PAGES if they
344 * are mapped to swap space.
345 */
346 __dec_zone_page_state(page, NR_FILE_PAGES);
347 __inc_zone_page_state(newpage, NR_FILE_PAGES);
348
Christoph Lameterb20a3502006-03-22 00:09:12 -0800349 write_unlock_irq(&mapping->tree_lock);
350
351 return 0;
352}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800353
354/*
355 * Copy the page to its new location
356 */
Christoph Lametere7340f72006-06-23 02:03:29 -0700357static void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800358{
359 copy_highpage(newpage, page);
360
361 if (PageError(page))
362 SetPageError(newpage);
363 if (PageReferenced(page))
364 SetPageReferenced(newpage);
365 if (PageUptodate(page))
366 SetPageUptodate(newpage);
367 if (PageActive(page))
368 SetPageActive(newpage);
369 if (PageChecked(page))
370 SetPageChecked(newpage);
371 if (PageMappedToDisk(page))
372 SetPageMappedToDisk(newpage);
373
374 if (PageDirty(page)) {
375 clear_page_dirty_for_io(page);
376 set_page_dirty(newpage);
377 }
378
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700379#ifdef CONFIG_SWAP
Christoph Lameterb20a3502006-03-22 00:09:12 -0800380 ClearPageSwapCache(page);
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700381#endif
Christoph Lameterb20a3502006-03-22 00:09:12 -0800382 ClearPageActive(page);
383 ClearPagePrivate(page);
384 set_page_private(page, 0);
385 page->mapping = NULL;
386
387 /*
388 * If any waiters have accumulated on the new page then
389 * wake them up.
390 */
391 if (PageWriteback(newpage))
392 end_page_writeback(newpage);
393}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800394
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700395/************************************************************
396 * Migration functions
397 ***********************************************************/
398
399/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700400int fail_migrate_page(struct address_space *mapping,
401 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700402{
403 return -EIO;
404}
405EXPORT_SYMBOL(fail_migrate_page);
406
Christoph Lameterb20a3502006-03-22 00:09:12 -0800407/*
408 * Common logic to directly migrate a single page suitable for
409 * pages that do not use PagePrivate.
410 *
411 * Pages are locked upon entry and exit.
412 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700413int migrate_page(struct address_space *mapping,
414 struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800415{
416 int rc;
417
418 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
419
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700420 rc = migrate_page_move_mapping(mapping, newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800421
422 if (rc)
423 return rc;
424
425 migrate_page_copy(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800426 return 0;
427}
428EXPORT_SYMBOL(migrate_page);
429
David Howells93614012006-09-30 20:45:40 +0200430#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800431/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700432 * Migration function for pages with buffers. This function can only be used
433 * if the underlying filesystem guarantees that no other references to "page"
434 * exist.
435 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700436int buffer_migrate_page(struct address_space *mapping,
437 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700438{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700439 struct buffer_head *bh, *head;
440 int rc;
441
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700442 if (!page_has_buffers(page))
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700443 return migrate_page(mapping, newpage, page);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700444
445 head = page_buffers(page);
446
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700447 rc = migrate_page_move_mapping(mapping, newpage, page);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700448
449 if (rc)
450 return rc;
451
452 bh = head;
453 do {
454 get_bh(bh);
455 lock_buffer(bh);
456 bh = bh->b_this_page;
457
458 } while (bh != head);
459
460 ClearPagePrivate(page);
461 set_page_private(newpage, page_private(page));
462 set_page_private(page, 0);
463 put_page(page);
464 get_page(newpage);
465
466 bh = head;
467 do {
468 set_bh_page(bh, newpage, bh_offset(bh));
469 bh = bh->b_this_page;
470
471 } while (bh != head);
472
473 SetPagePrivate(newpage);
474
475 migrate_page_copy(newpage, page);
476
477 bh = head;
478 do {
479 unlock_buffer(bh);
480 put_bh(bh);
481 bh = bh->b_this_page;
482
483 } while (bh != head);
484
485 return 0;
486}
487EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200488#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700489
Christoph Lameter04e62a22006-06-23 02:03:38 -0700490/*
491 * Writeback a page to clean the dirty state
492 */
493static int writeout(struct address_space *mapping, struct page *page)
494{
495 struct writeback_control wbc = {
496 .sync_mode = WB_SYNC_NONE,
497 .nr_to_write = 1,
498 .range_start = 0,
499 .range_end = LLONG_MAX,
500 .nonblocking = 1,
501 .for_reclaim = 1
502 };
503 int rc;
504
505 if (!mapping->a_ops->writepage)
506 /* No write method for the address space */
507 return -EINVAL;
508
509 if (!clear_page_dirty_for_io(page))
510 /* Someone else already triggered a write */
511 return -EAGAIN;
512
513 /*
514 * A dirty page may imply that the underlying filesystem has
515 * the page on some queue. So the page must be clean for
516 * migration. Writeout may mean we loose the lock and the
517 * page state is no longer what we checked for earlier.
518 * At this point we know that the migration attempt cannot
519 * be successful.
520 */
521 remove_migration_ptes(page, page);
522
523 rc = mapping->a_ops->writepage(page, &wbc);
524 if (rc < 0)
525 /* I/O Error writing */
526 return -EIO;
527
528 if (rc != AOP_WRITEPAGE_ACTIVATE)
529 /* unlocked. Relock */
530 lock_page(page);
531
532 return -EAGAIN;
533}
534
535/*
536 * Default handling if a filesystem does not provide a migration function.
537 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700538static int fallback_migrate_page(struct address_space *mapping,
539 struct page *newpage, struct page *page)
540{
Christoph Lameter04e62a22006-06-23 02:03:38 -0700541 if (PageDirty(page))
542 return writeout(mapping, page);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700543
544 /*
545 * Buffers may be managed in a filesystem specific way.
546 * We must have no buffers or drop them.
547 */
David Howellsb398f6b2006-08-29 19:05:58 +0100548 if (PagePrivate(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700549 !try_to_release_page(page, GFP_KERNEL))
550 return -EAGAIN;
551
552 return migrate_page(mapping, newpage, page);
553}
554
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700555/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700556 * Move a page to a newly allocated page
557 * The page is locked and all ptes have been successfully removed.
558 *
559 * The new page will have replaced the old page if this function
560 * is successful.
561 */
562static int move_to_new_page(struct page *newpage, struct page *page)
563{
564 struct address_space *mapping;
565 int rc;
566
567 /*
568 * Block others from accessing the page when we get around to
569 * establishing additional references. We are the only one
570 * holding a reference to the new page at this point.
571 */
572 if (TestSetPageLocked(newpage))
573 BUG();
574
575 /* Prepare mapping for the new page.*/
576 newpage->index = page->index;
577 newpage->mapping = page->mapping;
578
579 mapping = page_mapping(page);
580 if (!mapping)
581 rc = migrate_page(mapping, newpage, page);
582 else if (mapping->a_ops->migratepage)
583 /*
584 * Most pages have a mapping and most filesystems
585 * should provide a migration function. Anonymous
586 * pages are part of swap space which also has its
587 * own migration function. This is the most common
588 * path for page migration.
589 */
590 rc = mapping->a_ops->migratepage(mapping,
591 newpage, page);
592 else
593 rc = fallback_migrate_page(mapping, newpage, page);
594
595 if (!rc)
596 remove_migration_ptes(page, newpage);
597 else
598 newpage->mapping = NULL;
599
600 unlock_page(newpage);
601
602 return rc;
603}
604
605/*
606 * Obtain the lock on page, remove all ptes and migrate the page
607 * to the newly allocated page in newpage.
608 */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700609static int unmap_and_move(new_page_t get_new_page, unsigned long private,
610 struct page *page, int force)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700611{
612 int rc = 0;
Christoph Lameter742755a2006-06-23 02:03:55 -0700613 int *result = NULL;
614 struct page *newpage = get_new_page(page, private, &result);
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700615 int rcu_locked = 0;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700616
617 if (!newpage)
618 return -ENOMEM;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700619
620 if (page_count(page) == 1)
621 /* page was freed from under us. So we are done. */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700622 goto move_newpage;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700623
624 rc = -EAGAIN;
625 if (TestSetPageLocked(page)) {
626 if (!force)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700627 goto move_newpage;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700628 lock_page(page);
629 }
630
631 if (PageWriteback(page)) {
632 if (!force)
633 goto unlock;
634 wait_on_page_writeback(page);
635 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700636 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700637 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
638 * we cannot notice that anon_vma is freed while we migrates a page.
639 * This rcu_read_lock() delays freeing anon_vma pointer until the end
640 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700641 * File Caches may use write_page() or lock_page() in migration, then,
642 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700643 */
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700644 if (PageAnon(page)) {
645 rcu_read_lock();
646 rcu_locked = 1;
647 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700648 /*
649 * This is a corner case handling.
650 * When a new swap-cache is read into, it is linked to LRU
651 * and treated as swapcache but has no rmap yet.
652 * Calling try_to_unmap() against a page->mapping==NULL page is
653 * BUG. So handle it here.
654 */
655 if (!page->mapping)
656 goto rcu_unlock;
657 /* Establish migration ptes or remove ptes */
Christoph Lametere6a15302006-06-25 05:46:49 -0700658 try_to_unmap(page, 1);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700659
Christoph Lametere6a15302006-06-25 05:46:49 -0700660 if (!page_mapped(page))
661 rc = move_to_new_page(newpage, page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700662
663 if (rc)
664 remove_migration_ptes(page, page);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700665rcu_unlock:
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700666 if (rcu_locked)
667 rcu_read_unlock();
Christoph Lametere6a15302006-06-25 05:46:49 -0700668
Christoph Lametere24f0b82006-06-23 02:03:51 -0700669unlock:
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700670
Christoph Lametere24f0b82006-06-23 02:03:51 -0700671 unlock_page(page);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700672
Christoph Lametere24f0b82006-06-23 02:03:51 -0700673 if (rc != -EAGAIN) {
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700674 /*
675 * A page that has been migrated has all references
676 * removed and will be freed. A page that has not been
677 * migrated will have kepts its references and be
678 * restored.
679 */
680 list_del(&page->lru);
681 move_to_lru(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700682 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700683
684move_newpage:
685 /*
686 * Move the new page to the LRU. If migration was not successful
687 * then this will free the page.
688 */
689 move_to_lru(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700690 if (result) {
691 if (rc)
692 *result = rc;
693 else
694 *result = page_to_nid(newpage);
695 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700696 return rc;
697}
698
699/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800700 * migrate_pages
701 *
Christoph Lameter95a402c2006-06-23 02:03:53 -0700702 * The function takes one list of pages to migrate and a function
703 * that determines from the page to be migrated and the private data
704 * the target of the move and allocates the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800705 *
706 * The function returns after 10 attempts or if no pages
707 * are movable anymore because to has become empty
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700708 * or no retryable pages exist anymore. All pages will be
709 * retruned to the LRU or freed.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800710 *
Christoph Lameter95a402c2006-06-23 02:03:53 -0700711 * Return: Number of pages not migrated or error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800712 */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700713int migrate_pages(struct list_head *from,
714 new_page_t get_new_page, unsigned long private)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800715{
Christoph Lametere24f0b82006-06-23 02:03:51 -0700716 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800717 int nr_failed = 0;
718 int pass = 0;
719 struct page *page;
720 struct page *page2;
721 int swapwrite = current->flags & PF_SWAPWRITE;
722 int rc;
723
724 if (!swapwrite)
725 current->flags |= PF_SWAPWRITE;
726
Christoph Lametere24f0b82006-06-23 02:03:51 -0700727 for(pass = 0; pass < 10 && retry; pass++) {
728 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800729
Christoph Lametere24f0b82006-06-23 02:03:51 -0700730 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700731 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -0800732
Christoph Lameter95a402c2006-06-23 02:03:53 -0700733 rc = unmap_and_move(get_new_page, private,
734 page, pass > 2);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800735
Christoph Lametere24f0b82006-06-23 02:03:51 -0700736 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -0700737 case -ENOMEM:
738 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700739 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700740 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700741 break;
742 case 0:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700743 break;
744 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700745 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700746 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700747 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700748 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800749 }
750 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700751 rc = 0;
752out:
Christoph Lameterb20a3502006-03-22 00:09:12 -0800753 if (!swapwrite)
754 current->flags &= ~PF_SWAPWRITE;
755
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700756 putback_lru_pages(from);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700757
758 if (rc)
759 return rc;
760
Christoph Lameterb20a3502006-03-22 00:09:12 -0800761 return nr_failed + retry;
762}
763
Christoph Lameter742755a2006-06-23 02:03:55 -0700764#ifdef CONFIG_NUMA
765/*
766 * Move a list of individual pages
767 */
768struct page_to_node {
769 unsigned long addr;
770 struct page *page;
771 int node;
772 int status;
773};
774
775static struct page *new_page_node(struct page *p, unsigned long private,
776 int **result)
777{
778 struct page_to_node *pm = (struct page_to_node *)private;
779
780 while (pm->node != MAX_NUMNODES && pm->page != p)
781 pm++;
782
783 if (pm->node == MAX_NUMNODES)
784 return NULL;
785
786 *result = &pm->status;
787
Mel Gorman769848c2007-07-17 04:03:05 -0700788 return alloc_pages_node(pm->node,
789 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -0700790}
791
792/*
793 * Move a set of pages as indicated in the pm array. The addr
794 * field must be set to the virtual address of the page to be moved
795 * and the node number must contain a valid target node.
796 */
797static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
798 int migrate_all)
799{
800 int err;
801 struct page_to_node *pp;
802 LIST_HEAD(pagelist);
803
804 down_read(&mm->mmap_sem);
805
806 /*
807 * Build a list of pages to migrate
808 */
809 migrate_prep();
810 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
811 struct vm_area_struct *vma;
812 struct page *page;
813
814 /*
815 * A valid page pointer that will not match any of the
816 * pages that will be moved.
817 */
818 pp->page = ZERO_PAGE(0);
819
820 err = -EFAULT;
821 vma = find_vma(mm, pp->addr);
Christoph Lameter0dc952d2007-03-05 00:30:33 -0800822 if (!vma || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -0700823 goto set_status;
824
825 page = follow_page(vma, pp->addr, FOLL_GET);
826 err = -ENOENT;
827 if (!page)
828 goto set_status;
829
830 if (PageReserved(page)) /* Check for zero page */
831 goto put_and_set;
832
833 pp->page = page;
834 err = page_to_nid(page);
835
836 if (err == pp->node)
837 /*
838 * Node already in the right place
839 */
840 goto put_and_set;
841
842 err = -EACCES;
843 if (page_mapcount(page) > 1 &&
844 !migrate_all)
845 goto put_and_set;
846
847 err = isolate_lru_page(page, &pagelist);
848put_and_set:
849 /*
850 * Either remove the duplicate refcount from
851 * isolate_lru_page() or drop the page ref if it was
852 * not isolated.
853 */
854 put_page(page);
855set_status:
856 pp->status = err;
857 }
858
859 if (!list_empty(&pagelist))
860 err = migrate_pages(&pagelist, new_page_node,
861 (unsigned long)pm);
862 else
863 err = -ENOENT;
864
865 up_read(&mm->mmap_sem);
866 return err;
867}
868
869/*
870 * Determine the nodes of a list of pages. The addr in the pm array
871 * must have been set to the virtual address of which we want to determine
872 * the node number.
873 */
874static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
875{
876 down_read(&mm->mmap_sem);
877
878 for ( ; pm->node != MAX_NUMNODES; pm++) {
879 struct vm_area_struct *vma;
880 struct page *page;
881 int err;
882
883 err = -EFAULT;
884 vma = find_vma(mm, pm->addr);
885 if (!vma)
886 goto set_status;
887
888 page = follow_page(vma, pm->addr, 0);
889 err = -ENOENT;
890 /* Use PageReserved to check for zero page */
891 if (!page || PageReserved(page))
892 goto set_status;
893
894 err = page_to_nid(page);
895set_status:
896 pm->status = err;
897 }
898
899 up_read(&mm->mmap_sem);
900 return 0;
901}
902
903/*
904 * Move a list of pages in the address space of the currently executing
905 * process.
906 */
907asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
908 const void __user * __user *pages,
909 const int __user *nodes,
910 int __user *status, int flags)
911{
912 int err = 0;
913 int i;
914 struct task_struct *task;
915 nodemask_t task_nodes;
916 struct mm_struct *mm;
917 struct page_to_node *pm = NULL;
918
919 /* Check flags */
920 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
921 return -EINVAL;
922
923 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
924 return -EPERM;
925
926 /* Find the mm_struct */
927 read_lock(&tasklist_lock);
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -0700928 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -0700929 if (!task) {
930 read_unlock(&tasklist_lock);
931 return -ESRCH;
932 }
933 mm = get_task_mm(task);
934 read_unlock(&tasklist_lock);
935
936 if (!mm)
937 return -EINVAL;
938
939 /*
940 * Check if this process has the right to modify the specified
941 * process. The right exists if the process has administrative
942 * capabilities, superuser privileges or the same
943 * userid as the target process.
944 */
945 if ((current->euid != task->suid) && (current->euid != task->uid) &&
946 (current->uid != task->suid) && (current->uid != task->uid) &&
947 !capable(CAP_SYS_NICE)) {
948 err = -EPERM;
949 goto out2;
950 }
951
David Quigley86c3a762006-06-23 02:04:02 -0700952 err = security_task_movememory(task);
953 if (err)
954 goto out2;
955
956
Christoph Lameter742755a2006-06-23 02:03:55 -0700957 task_nodes = cpuset_mems_allowed(task);
958
959 /* Limit nr_pages so that the multiplication may not overflow */
960 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
961 err = -E2BIG;
962 goto out2;
963 }
964
965 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
966 if (!pm) {
967 err = -ENOMEM;
968 goto out2;
969 }
970
971 /*
972 * Get parameters from user space and initialize the pm
973 * array. Return various errors if the user did something wrong.
974 */
975 for (i = 0; i < nr_pages; i++) {
Al Viro9d966d42007-10-14 19:34:10 +0100976 const void __user *p;
Christoph Lameter742755a2006-06-23 02:03:55 -0700977
978 err = -EFAULT;
979 if (get_user(p, pages + i))
980 goto out;
981
982 pm[i].addr = (unsigned long)p;
983 if (nodes) {
984 int node;
985
986 if (get_user(node, nodes + i))
987 goto out;
988
989 err = -ENODEV;
Christoph Lameter56bbd652007-10-16 01:25:35 -0700990 if (!node_state(node, N_HIGH_MEMORY))
Christoph Lameter742755a2006-06-23 02:03:55 -0700991 goto out;
992
993 err = -EACCES;
994 if (!node_isset(node, task_nodes))
995 goto out;
996
997 pm[i].node = node;
Stephen Rothwell8ce08462006-11-02 22:07:28 -0800998 } else
999 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
Christoph Lameter742755a2006-06-23 02:03:55 -07001000 }
1001 /* End marker */
1002 pm[nr_pages].node = MAX_NUMNODES;
1003
1004 if (nodes)
1005 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1006 else
1007 err = do_pages_stat(mm, pm);
1008
1009 if (err >= 0)
1010 /* Return status information */
1011 for (i = 0; i < nr_pages; i++)
1012 if (put_user(pm[i].status, status + i))
1013 err = -EFAULT;
1014
1015out:
1016 vfree(pm);
1017out2:
1018 mmput(mm);
1019 return err;
1020}
1021#endif
1022
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001023/*
1024 * Call migration functions in the vma_ops that may prepare
1025 * memory in a vm for migration. migration functions may perform
1026 * the migration for vmas that do not have an underlying page struct.
1027 */
1028int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1029 const nodemask_t *from, unsigned long flags)
1030{
1031 struct vm_area_struct *vma;
1032 int err = 0;
1033
1034 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1035 if (vma->vm_ops && vma->vm_ops->migrate) {
1036 err = vma->vm_ops->migrate(vma, to, from, flags);
1037 if (err)
1038 break;
1039 }
1040 }
1041 return err;
1042}