blob: f6d7f8efd1a84b5fb1729013890ecaa3d6caa609 [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>
Christoph Lametercde53532008-07-04 09:59:22 -070012 * Christoph Lameter
Christoph Lameterb20a3502006-03-22 00:09:12 -080013 */
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>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080032#include <linux/memcontrol.h>
Adrian Bunk4f5ca262008-07-23 21:27:02 -070033#include <linux/syscalls.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080034
35#include "internal.h"
36
Christoph Lameterb20a3502006-03-22 00:09:12 -080037#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
38
39/*
40 * Isolate one page from the LRU lists. If successful put it onto
41 * the indicated list with elevated page count.
42 *
43 * Result:
44 * -EBUSY: page not on LRU list
45 * 0: page removed from LRU list and added to the specified list.
46 */
47int isolate_lru_page(struct page *page, struct list_head *pagelist)
48{
49 int ret = -EBUSY;
50
51 if (PageLRU(page)) {
52 struct zone *zone = page_zone(page);
53
54 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyuki3dd9fe82007-07-26 10:41:08 -070055 if (PageLRU(page) && get_page_unless_zero(page)) {
Christoph Lameterb20a3502006-03-22 00:09:12 -080056 ret = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -080057 ClearPageLRU(page);
58 if (PageActive(page))
59 del_page_from_active_list(zone, page);
60 else
61 del_page_from_inactive_list(zone, page);
62 list_add_tail(&page->lru, pagelist);
63 }
64 spin_unlock_irq(&zone->lru_lock);
65 }
66 return ret;
67}
68
69/*
Christoph Lameter742755a2006-06-23 02:03:55 -070070 * migrate_prep() needs to be called before we start compiling a list of pages
71 * to be migrated using isolate_lru_page().
Christoph Lameterb20a3502006-03-22 00:09:12 -080072 */
73int migrate_prep(void)
74{
Christoph Lameterb20a3502006-03-22 00:09:12 -080075 /*
76 * Clear the LRU lists so pages can be isolated.
77 * Note that pages may be moved off the LRU after we have
78 * drained them. Those pages will fail to migrate like other
79 * pages that may be busy.
80 */
81 lru_add_drain_all();
82
83 return 0;
84}
85
86static inline void move_to_lru(struct page *page)
87{
Christoph Lameterb20a3502006-03-22 00:09:12 -080088 if (PageActive(page)) {
89 /*
90 * lru_cache_add_active checks that
91 * the PG_active bit is off.
92 */
93 ClearPageActive(page);
94 lru_cache_add_active(page);
95 } else {
96 lru_cache_add(page);
97 }
98 put_page(page);
99}
100
101/*
102 * Add isolated pages on the list back to the LRU.
103 *
104 * returns the number of pages put back.
105 */
106int putback_lru_pages(struct list_head *l)
107{
108 struct page *page;
109 struct page *page2;
110 int count = 0;
111
112 list_for_each_entry_safe(page, page2, l, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700113 list_del(&page->lru);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800114 move_to_lru(page);
115 count++;
116 }
117 return count;
118}
119
Christoph Lameter06972122006-06-23 02:03:35 -0700120/*
121 * Restore a potential migration pte to a working pte entry
122 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700123static void remove_migration_pte(struct vm_area_struct *vma,
Christoph Lameter06972122006-06-23 02:03:35 -0700124 struct page *old, struct page *new)
125{
126 struct mm_struct *mm = vma->vm_mm;
127 swp_entry_t entry;
128 pgd_t *pgd;
129 pud_t *pud;
130 pmd_t *pmd;
131 pte_t *ptep, pte;
132 spinlock_t *ptl;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700133 unsigned long addr = page_address_in_vma(new, vma);
134
135 if (addr == -EFAULT)
136 return;
Christoph Lameter06972122006-06-23 02:03:35 -0700137
138 pgd = pgd_offset(mm, addr);
139 if (!pgd_present(*pgd))
140 return;
141
142 pud = pud_offset(pgd, addr);
143 if (!pud_present(*pud))
144 return;
145
146 pmd = pmd_offset(pud, addr);
147 if (!pmd_present(*pmd))
148 return;
149
150 ptep = pte_offset_map(pmd, addr);
151
152 if (!is_swap_pte(*ptep)) {
153 pte_unmap(ptep);
154 return;
155 }
156
157 ptl = pte_lockptr(mm, pmd);
158 spin_lock(ptl);
159 pte = *ptep;
160 if (!is_swap_pte(pte))
161 goto out;
162
163 entry = pte_to_swp_entry(pte);
164
165 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
166 goto out;
167
Hugh Dickins98837c72008-03-04 14:29:06 -0800168 /*
169 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
170 * Failure is not an option here: we're now expected to remove every
171 * migration pte, and will cause crashes otherwise. Normally this
172 * is not an issue: mem_cgroup_prepare_migration bumped up the old
173 * page_cgroup count for safety, that's now attached to the new page,
174 * so this charge should just be another incrementation of the count,
175 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
176 * there's been a force_empty, those reference counts may no longer
177 * be reliable, and this charge can actually fail: oh well, we don't
178 * make the situation any worse by proceeding as if it had succeeded.
179 */
180 mem_cgroup_charge(new, mm, GFP_ATOMIC);
181
Christoph Lameter06972122006-06-23 02:03:35 -0700182 get_page(new);
183 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
184 if (is_write_migration_entry(entry))
185 pte = pte_mkwrite(pte);
KAMEZAWA Hiroyuki97ee0522007-10-16 01:25:43 -0700186 flush_cache_page(vma, addr, pte_pfn(pte));
Christoph Lameter06972122006-06-23 02:03:35 -0700187 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700188
189 if (PageAnon(new))
190 page_add_anon_rmap(new, vma, addr);
191 else
192 page_add_file_rmap(new);
193
194 /* No need to invalidate - it was non-present before */
195 update_mmu_cache(vma, addr, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700196
Christoph Lameter06972122006-06-23 02:03:35 -0700197out:
198 pte_unmap_unlock(ptep, ptl);
199}
200
201/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700202 * Note that remove_file_migration_ptes will only work on regular mappings,
203 * Nonlinear mappings do not use migration entries.
204 */
205static void remove_file_migration_ptes(struct page *old, struct page *new)
206{
207 struct vm_area_struct *vma;
208 struct address_space *mapping = page_mapping(new);
209 struct prio_tree_iter iter;
210 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
211
212 if (!mapping)
213 return;
214
215 spin_lock(&mapping->i_mmap_lock);
216
217 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
218 remove_migration_pte(vma, old, new);
219
220 spin_unlock(&mapping->i_mmap_lock);
221}
222
223/*
Christoph Lameter06972122006-06-23 02:03:35 -0700224 * Must hold mmap_sem lock on at least one of the vmas containing
225 * the page so that the anon_vma cannot vanish.
226 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700227static void remove_anon_migration_ptes(struct page *old, struct page *new)
Christoph Lameter06972122006-06-23 02:03:35 -0700228{
229 struct anon_vma *anon_vma;
230 struct vm_area_struct *vma;
231 unsigned long mapping;
232
233 mapping = (unsigned long)new->mapping;
234
235 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
236 return;
237
238 /*
239 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
240 */
241 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
242 spin_lock(&anon_vma->lock);
243
244 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
Christoph Lameter04e62a22006-06-23 02:03:38 -0700245 remove_migration_pte(vma, old, new);
Christoph Lameter06972122006-06-23 02:03:35 -0700246
247 spin_unlock(&anon_vma->lock);
248}
249
250/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700251 * Get rid of all migration entries and replace them by
252 * references to the indicated page.
253 */
254static void remove_migration_ptes(struct page *old, struct page *new)
255{
256 if (PageAnon(new))
257 remove_anon_migration_ptes(old, new);
258 else
259 remove_file_migration_ptes(old, new);
260}
261
262/*
Christoph Lameter06972122006-06-23 02:03:35 -0700263 * Something used the pte of a page under migration. We need to
264 * get to the page and wait until migration is finished.
265 * When we return from this function the fault will be retried.
266 *
267 * This function is called from do_swap_page().
268 */
269void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
270 unsigned long address)
271{
272 pte_t *ptep, pte;
273 spinlock_t *ptl;
274 swp_entry_t entry;
275 struct page *page;
276
277 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
278 pte = *ptep;
279 if (!is_swap_pte(pte))
280 goto out;
281
282 entry = pte_to_swp_entry(pte);
283 if (!is_migration_entry(entry))
284 goto out;
285
286 page = migration_entry_to_page(entry);
287
288 get_page(page);
289 pte_unmap_unlock(ptep, ptl);
290 wait_on_page_locked(page);
291 put_page(page);
292 return;
293out:
294 pte_unmap_unlock(ptep, ptl);
295}
296
Christoph Lameterb20a3502006-03-22 00:09:12 -0800297/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700298 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700299 *
300 * The number of remaining references must be:
301 * 1 for anonymous pages without a mapping
302 * 2 for pages with a mapping
303 * 3 for pages with a mapping and PagePrivate set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800304 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700305static int migrate_page_move_mapping(struct address_space *mapping,
306 struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800307{
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800308 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800309
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700310 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700311 /* Anonymous page without mapping */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700312 if (page_count(page) != 1)
313 return -EAGAIN;
314 return 0;
315 }
316
Christoph Lameterb20a3502006-03-22 00:09:12 -0800317 write_lock_irq(&mapping->tree_lock);
318
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800319 pslot = radix_tree_lookup_slot(&mapping->page_tree,
320 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800321
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700322 if (page_count(page) != 2 + !!PagePrivate(page) ||
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800323 (struct page *)radix_tree_deref_slot(pslot) != page) {
Christoph Lameterb20a3502006-03-22 00:09:12 -0800324 write_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700325 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800326 }
327
328 /*
329 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800330 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800331 get_page(newpage); /* add cache reference */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700332#ifdef CONFIG_SWAP
Christoph Lameterb20a3502006-03-22 00:09:12 -0800333 if (PageSwapCache(page)) {
334 SetPageSwapCache(newpage);
335 set_page_private(newpage, page_private(page));
336 }
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700337#endif
Christoph Lameterb20a3502006-03-22 00:09:12 -0800338
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800339 radix_tree_replace_slot(pslot, newpage);
340
341 /*
342 * Drop cache reference from old page.
343 * We know this isn't the last reference.
344 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800345 __put_page(page);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800346
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700347 /*
348 * If moved to a different zone then also account
349 * the page for that zone. Other VM counters will be
350 * taken care of when we establish references to the
351 * new page and drop references to the old page.
352 *
353 * Note that anonymous pages are accounted for
354 * via NR_FILE_PAGES and NR_ANON_PAGES if they
355 * are mapped to swap space.
356 */
357 __dec_zone_page_state(page, NR_FILE_PAGES);
358 __inc_zone_page_state(newpage, NR_FILE_PAGES);
359
Christoph Lameterb20a3502006-03-22 00:09:12 -0800360 write_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukie8589cc2008-07-25 01:47:10 -0700361 if (!PageSwapCache(newpage)) {
362 mem_cgroup_uncharge_page(page);
363 mem_cgroup_getref(newpage);
364 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800365
366 return 0;
367}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800368
369/*
370 * Copy the page to its new location
371 */
Christoph Lametere7340f72006-06-23 02:03:29 -0700372static void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800373{
374 copy_highpage(newpage, page);
375
376 if (PageError(page))
377 SetPageError(newpage);
378 if (PageReferenced(page))
379 SetPageReferenced(newpage);
380 if (PageUptodate(page))
381 SetPageUptodate(newpage);
382 if (PageActive(page))
383 SetPageActive(newpage);
384 if (PageChecked(page))
385 SetPageChecked(newpage);
386 if (PageMappedToDisk(page))
387 SetPageMappedToDisk(newpage);
388
389 if (PageDirty(page)) {
390 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700391 /*
392 * Want to mark the page and the radix tree as dirty, and
393 * redo the accounting that clear_page_dirty_for_io undid,
394 * but we can't use set_page_dirty because that function
395 * is actually a signal that all of the page has become dirty.
396 * Wheras only part of our page may be dirty.
397 */
398 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800399 }
400
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700401#ifdef CONFIG_SWAP
Christoph Lameterb20a3502006-03-22 00:09:12 -0800402 ClearPageSwapCache(page);
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700403#endif
Christoph Lameterb20a3502006-03-22 00:09:12 -0800404 ClearPageActive(page);
405 ClearPagePrivate(page);
406 set_page_private(page, 0);
407 page->mapping = NULL;
408
409 /*
410 * If any waiters have accumulated on the new page then
411 * wake them up.
412 */
413 if (PageWriteback(newpage))
414 end_page_writeback(newpage);
415}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800416
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700417/************************************************************
418 * Migration functions
419 ***********************************************************/
420
421/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700422int fail_migrate_page(struct address_space *mapping,
423 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700424{
425 return -EIO;
426}
427EXPORT_SYMBOL(fail_migrate_page);
428
Christoph Lameterb20a3502006-03-22 00:09:12 -0800429/*
430 * Common logic to directly migrate a single page suitable for
431 * pages that do not use PagePrivate.
432 *
433 * Pages are locked upon entry and exit.
434 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700435int migrate_page(struct address_space *mapping,
436 struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800437{
438 int rc;
439
440 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
441
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700442 rc = migrate_page_move_mapping(mapping, newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800443
444 if (rc)
445 return rc;
446
447 migrate_page_copy(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800448 return 0;
449}
450EXPORT_SYMBOL(migrate_page);
451
David Howells93614012006-09-30 20:45:40 +0200452#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800453/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700454 * Migration function for pages with buffers. This function can only be used
455 * if the underlying filesystem guarantees that no other references to "page"
456 * exist.
457 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700458int buffer_migrate_page(struct address_space *mapping,
459 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700460{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700461 struct buffer_head *bh, *head;
462 int rc;
463
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700464 if (!page_has_buffers(page))
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700465 return migrate_page(mapping, newpage, page);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700466
467 head = page_buffers(page);
468
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700469 rc = migrate_page_move_mapping(mapping, newpage, page);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700470
471 if (rc)
472 return rc;
473
474 bh = head;
475 do {
476 get_bh(bh);
477 lock_buffer(bh);
478 bh = bh->b_this_page;
479
480 } while (bh != head);
481
482 ClearPagePrivate(page);
483 set_page_private(newpage, page_private(page));
484 set_page_private(page, 0);
485 put_page(page);
486 get_page(newpage);
487
488 bh = head;
489 do {
490 set_bh_page(bh, newpage, bh_offset(bh));
491 bh = bh->b_this_page;
492
493 } while (bh != head);
494
495 SetPagePrivate(newpage);
496
497 migrate_page_copy(newpage, page);
498
499 bh = head;
500 do {
501 unlock_buffer(bh);
502 put_bh(bh);
503 bh = bh->b_this_page;
504
505 } while (bh != head);
506
507 return 0;
508}
509EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200510#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700511
Christoph Lameter04e62a22006-06-23 02:03:38 -0700512/*
513 * Writeback a page to clean the dirty state
514 */
515static int writeout(struct address_space *mapping, struct page *page)
516{
517 struct writeback_control wbc = {
518 .sync_mode = WB_SYNC_NONE,
519 .nr_to_write = 1,
520 .range_start = 0,
521 .range_end = LLONG_MAX,
522 .nonblocking = 1,
523 .for_reclaim = 1
524 };
525 int rc;
526
527 if (!mapping->a_ops->writepage)
528 /* No write method for the address space */
529 return -EINVAL;
530
531 if (!clear_page_dirty_for_io(page))
532 /* Someone else already triggered a write */
533 return -EAGAIN;
534
535 /*
536 * A dirty page may imply that the underlying filesystem has
537 * the page on some queue. So the page must be clean for
538 * migration. Writeout may mean we loose the lock and the
539 * page state is no longer what we checked for earlier.
540 * At this point we know that the migration attempt cannot
541 * be successful.
542 */
543 remove_migration_ptes(page, page);
544
545 rc = mapping->a_ops->writepage(page, &wbc);
546 if (rc < 0)
547 /* I/O Error writing */
548 return -EIO;
549
550 if (rc != AOP_WRITEPAGE_ACTIVATE)
551 /* unlocked. Relock */
552 lock_page(page);
553
554 return -EAGAIN;
555}
556
557/*
558 * Default handling if a filesystem does not provide a migration function.
559 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700560static int fallback_migrate_page(struct address_space *mapping,
561 struct page *newpage, struct page *page)
562{
Christoph Lameter04e62a22006-06-23 02:03:38 -0700563 if (PageDirty(page))
564 return writeout(mapping, page);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700565
566 /*
567 * Buffers may be managed in a filesystem specific way.
568 * We must have no buffers or drop them.
569 */
David Howellsb398f6b2006-08-29 19:05:58 +0100570 if (PagePrivate(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700571 !try_to_release_page(page, GFP_KERNEL))
572 return -EAGAIN;
573
574 return migrate_page(mapping, newpage, page);
575}
576
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700577/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700578 * Move a page to a newly allocated page
579 * The page is locked and all ptes have been successfully removed.
580 *
581 * The new page will have replaced the old page if this function
582 * is successful.
583 */
584static int move_to_new_page(struct page *newpage, struct page *page)
585{
586 struct address_space *mapping;
587 int rc;
588
589 /*
590 * Block others from accessing the page when we get around to
591 * establishing additional references. We are the only one
592 * holding a reference to the new page at this point.
593 */
594 if (TestSetPageLocked(newpage))
595 BUG();
596
597 /* Prepare mapping for the new page.*/
598 newpage->index = page->index;
599 newpage->mapping = page->mapping;
600
601 mapping = page_mapping(page);
602 if (!mapping)
603 rc = migrate_page(mapping, newpage, page);
604 else if (mapping->a_ops->migratepage)
605 /*
606 * Most pages have a mapping and most filesystems
607 * should provide a migration function. Anonymous
608 * pages are part of swap space which also has its
609 * own migration function. This is the most common
610 * path for page migration.
611 */
612 rc = mapping->a_ops->migratepage(mapping,
613 newpage, page);
614 else
615 rc = fallback_migrate_page(mapping, newpage, page);
616
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800617 if (!rc) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700618 remove_migration_ptes(page, newpage);
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800619 } else
Christoph Lametere24f0b82006-06-23 02:03:51 -0700620 newpage->mapping = NULL;
621
622 unlock_page(newpage);
623
624 return rc;
625}
626
627/*
628 * Obtain the lock on page, remove all ptes and migrate the page
629 * to the newly allocated page in newpage.
630 */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700631static int unmap_and_move(new_page_t get_new_page, unsigned long private,
632 struct page *page, int force)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700633{
634 int rc = 0;
Christoph Lameter742755a2006-06-23 02:03:55 -0700635 int *result = NULL;
636 struct page *newpage = get_new_page(page, private, &result);
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700637 int rcu_locked = 0;
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800638 int charge = 0;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700639
640 if (!newpage)
641 return -ENOMEM;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700642
643 if (page_count(page) == 1)
644 /* page was freed from under us. So we are done. */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700645 goto move_newpage;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700646
KAMEZAWA Hiroyukie8589cc2008-07-25 01:47:10 -0700647 charge = mem_cgroup_prepare_migration(page, newpage);
648 if (charge == -ENOMEM) {
649 rc = -ENOMEM;
650 goto move_newpage;
651 }
652 /* prepare cgroup just returns 0 or -ENOMEM */
653 BUG_ON(charge);
654
Christoph Lametere24f0b82006-06-23 02:03:51 -0700655 rc = -EAGAIN;
656 if (TestSetPageLocked(page)) {
657 if (!force)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700658 goto move_newpage;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700659 lock_page(page);
660 }
661
662 if (PageWriteback(page)) {
663 if (!force)
664 goto unlock;
665 wait_on_page_writeback(page);
666 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700667 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700668 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
669 * we cannot notice that anon_vma is freed while we migrates a page.
670 * This rcu_read_lock() delays freeing anon_vma pointer until the end
671 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700672 * File Caches may use write_page() or lock_page() in migration, then,
673 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700674 */
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700675 if (PageAnon(page)) {
676 rcu_read_lock();
677 rcu_locked = 1;
678 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800679
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700680 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800681 * Corner case handling:
682 * 1. When a new swap-cache page is read into, it is added to the LRU
683 * and treated as swapcache but it has no rmap yet.
684 * Calling try_to_unmap() against a page->mapping==NULL page will
685 * trigger a BUG. So handle it here.
686 * 2. An orphaned page (see truncate_complete_page) might have
687 * fs-private metadata. The page can be picked up due to memory
688 * offlining. Everywhere else except page reclaim, the page is
689 * invisible to the vm, so the page can not be migrated. So try to
690 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700691 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800692 if (!page->mapping) {
693 if (!PageAnon(page) && PagePrivate(page)) {
694 /*
695 * Go direct to try_to_free_buffers() here because
696 * a) that's what try_to_release_page() would do anyway
697 * b) we may be under rcu_read_lock() here, so we can't
698 * use GFP_KERNEL which is what try_to_release_page()
699 * needs to be effective.
700 */
701 try_to_free_buffers(page);
702 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700703 goto rcu_unlock;
Shaohua Li62e1c552008-02-04 22:29:33 -0800704 }
705
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700706 /* Establish migration ptes or remove ptes */
Christoph Lametere6a15302006-06-25 05:46:49 -0700707 try_to_unmap(page, 1);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700708
Christoph Lametere6a15302006-06-25 05:46:49 -0700709 if (!page_mapped(page))
710 rc = move_to_new_page(newpage, page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700711
KAMEZAWA Hiroyukie8589cc2008-07-25 01:47:10 -0700712 if (rc)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700713 remove_migration_ptes(page, page);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700714rcu_unlock:
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700715 if (rcu_locked)
716 rcu_read_unlock();
Christoph Lametere6a15302006-06-25 05:46:49 -0700717
Christoph Lametere24f0b82006-06-23 02:03:51 -0700718unlock:
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700719
Christoph Lametere24f0b82006-06-23 02:03:51 -0700720 unlock_page(page);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700721
Christoph Lametere24f0b82006-06-23 02:03:51 -0700722 if (rc != -EAGAIN) {
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700723 /*
724 * A page that has been migrated has all references
725 * removed and will be freed. A page that has not been
726 * migrated will have kepts its references and be
727 * restored.
728 */
729 list_del(&page->lru);
730 move_to_lru(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700731 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700732
733move_newpage:
KAMEZAWA Hiroyukie8589cc2008-07-25 01:47:10 -0700734 if (!charge)
735 mem_cgroup_end_migration(newpage);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700736 /*
737 * Move the new page to the LRU. If migration was not successful
738 * then this will free the page.
739 */
740 move_to_lru(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700741 if (result) {
742 if (rc)
743 *result = rc;
744 else
745 *result = page_to_nid(newpage);
746 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700747 return rc;
748}
749
750/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800751 * migrate_pages
752 *
Christoph Lameter95a402c2006-06-23 02:03:53 -0700753 * The function takes one list of pages to migrate and a function
754 * that determines from the page to be migrated and the private data
755 * the target of the move and allocates the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800756 *
757 * The function returns after 10 attempts or if no pages
758 * are movable anymore because to has become empty
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700759 * or no retryable pages exist anymore. All pages will be
Gabriel Craciunescue9534b32007-10-20 02:13:26 +0200760 * returned to the LRU or freed.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800761 *
Christoph Lameter95a402c2006-06-23 02:03:53 -0700762 * Return: Number of pages not migrated or error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800763 */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700764int migrate_pages(struct list_head *from,
765 new_page_t get_new_page, unsigned long private)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800766{
Christoph Lametere24f0b82006-06-23 02:03:51 -0700767 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800768 int nr_failed = 0;
769 int pass = 0;
770 struct page *page;
771 struct page *page2;
772 int swapwrite = current->flags & PF_SWAPWRITE;
773 int rc;
774
775 if (!swapwrite)
776 current->flags |= PF_SWAPWRITE;
777
Christoph Lametere24f0b82006-06-23 02:03:51 -0700778 for(pass = 0; pass < 10 && retry; pass++) {
779 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800780
Christoph Lametere24f0b82006-06-23 02:03:51 -0700781 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700782 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -0800783
Christoph Lameter95a402c2006-06-23 02:03:53 -0700784 rc = unmap_and_move(get_new_page, private,
785 page, pass > 2);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800786
Christoph Lametere24f0b82006-06-23 02:03:51 -0700787 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -0700788 case -ENOMEM:
789 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700790 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700791 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700792 break;
793 case 0:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700794 break;
795 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700796 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700797 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700798 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700799 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800800 }
801 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700802 rc = 0;
803out:
Christoph Lameterb20a3502006-03-22 00:09:12 -0800804 if (!swapwrite)
805 current->flags &= ~PF_SWAPWRITE;
806
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700807 putback_lru_pages(from);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700808
809 if (rc)
810 return rc;
811
Christoph Lameterb20a3502006-03-22 00:09:12 -0800812 return nr_failed + retry;
813}
814
Christoph Lameter742755a2006-06-23 02:03:55 -0700815#ifdef CONFIG_NUMA
816/*
817 * Move a list of individual pages
818 */
819struct page_to_node {
820 unsigned long addr;
821 struct page *page;
822 int node;
823 int status;
824};
825
826static struct page *new_page_node(struct page *p, unsigned long private,
827 int **result)
828{
829 struct page_to_node *pm = (struct page_to_node *)private;
830
831 while (pm->node != MAX_NUMNODES && pm->page != p)
832 pm++;
833
834 if (pm->node == MAX_NUMNODES)
835 return NULL;
836
837 *result = &pm->status;
838
Mel Gorman769848c2007-07-17 04:03:05 -0700839 return alloc_pages_node(pm->node,
840 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -0700841}
842
843/*
844 * Move a set of pages as indicated in the pm array. The addr
845 * field must be set to the virtual address of the page to be moved
846 * and the node number must contain a valid target node.
847 */
848static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
849 int migrate_all)
850{
851 int err;
852 struct page_to_node *pp;
853 LIST_HEAD(pagelist);
854
855 down_read(&mm->mmap_sem);
856
857 /*
858 * Build a list of pages to migrate
859 */
860 migrate_prep();
861 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
862 struct vm_area_struct *vma;
863 struct page *page;
864
865 /*
866 * A valid page pointer that will not match any of the
867 * pages that will be moved.
868 */
869 pp->page = ZERO_PAGE(0);
870
871 err = -EFAULT;
872 vma = find_vma(mm, pp->addr);
Christoph Lameter0dc952d2007-03-05 00:30:33 -0800873 if (!vma || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -0700874 goto set_status;
875
876 page = follow_page(vma, pp->addr, FOLL_GET);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -0700877
878 err = PTR_ERR(page);
879 if (IS_ERR(page))
880 goto set_status;
881
Christoph Lameter742755a2006-06-23 02:03:55 -0700882 err = -ENOENT;
883 if (!page)
884 goto set_status;
885
886 if (PageReserved(page)) /* Check for zero page */
887 goto put_and_set;
888
889 pp->page = page;
890 err = page_to_nid(page);
891
892 if (err == pp->node)
893 /*
894 * Node already in the right place
895 */
896 goto put_and_set;
897
898 err = -EACCES;
899 if (page_mapcount(page) > 1 &&
900 !migrate_all)
901 goto put_and_set;
902
903 err = isolate_lru_page(page, &pagelist);
904put_and_set:
905 /*
906 * Either remove the duplicate refcount from
907 * isolate_lru_page() or drop the page ref if it was
908 * not isolated.
909 */
910 put_page(page);
911set_status:
912 pp->status = err;
913 }
914
915 if (!list_empty(&pagelist))
916 err = migrate_pages(&pagelist, new_page_node,
917 (unsigned long)pm);
918 else
919 err = -ENOENT;
920
921 up_read(&mm->mmap_sem);
922 return err;
923}
924
925/*
926 * Determine the nodes of a list of pages. The addr in the pm array
927 * must have been set to the virtual address of which we want to determine
928 * the node number.
929 */
930static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
931{
932 down_read(&mm->mmap_sem);
933
934 for ( ; pm->node != MAX_NUMNODES; pm++) {
935 struct vm_area_struct *vma;
936 struct page *page;
937 int err;
938
939 err = -EFAULT;
940 vma = find_vma(mm, pm->addr);
941 if (!vma)
942 goto set_status;
943
944 page = follow_page(vma, pm->addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -0700945
946 err = PTR_ERR(page);
947 if (IS_ERR(page))
948 goto set_status;
949
Christoph Lameter742755a2006-06-23 02:03:55 -0700950 err = -ENOENT;
951 /* Use PageReserved to check for zero page */
952 if (!page || PageReserved(page))
953 goto set_status;
954
955 err = page_to_nid(page);
956set_status:
957 pm->status = err;
958 }
959
960 up_read(&mm->mmap_sem);
961 return 0;
962}
963
964/*
965 * Move a list of pages in the address space of the currently executing
966 * process.
967 */
968asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
969 const void __user * __user *pages,
970 const int __user *nodes,
971 int __user *status, int flags)
972{
973 int err = 0;
974 int i;
975 struct task_struct *task;
976 nodemask_t task_nodes;
977 struct mm_struct *mm;
978 struct page_to_node *pm = NULL;
979
980 /* Check flags */
981 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
982 return -EINVAL;
983
984 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
985 return -EPERM;
986
987 /* Find the mm_struct */
988 read_lock(&tasklist_lock);
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -0700989 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -0700990 if (!task) {
991 read_unlock(&tasklist_lock);
992 return -ESRCH;
993 }
994 mm = get_task_mm(task);
995 read_unlock(&tasklist_lock);
996
997 if (!mm)
998 return -EINVAL;
999
1000 /*
1001 * Check if this process has the right to modify the specified
1002 * process. The right exists if the process has administrative
1003 * capabilities, superuser privileges or the same
1004 * userid as the target process.
1005 */
1006 if ((current->euid != task->suid) && (current->euid != task->uid) &&
1007 (current->uid != task->suid) && (current->uid != task->uid) &&
1008 !capable(CAP_SYS_NICE)) {
1009 err = -EPERM;
1010 goto out2;
1011 }
1012
David Quigley86c3a762006-06-23 02:04:02 -07001013 err = security_task_movememory(task);
1014 if (err)
1015 goto out2;
1016
1017
Christoph Lameter742755a2006-06-23 02:03:55 -07001018 task_nodes = cpuset_mems_allowed(task);
1019
1020 /* Limit nr_pages so that the multiplication may not overflow */
1021 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1022 err = -E2BIG;
1023 goto out2;
1024 }
1025
1026 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1027 if (!pm) {
1028 err = -ENOMEM;
1029 goto out2;
1030 }
1031
1032 /*
1033 * Get parameters from user space and initialize the pm
1034 * array. Return various errors if the user did something wrong.
1035 */
1036 for (i = 0; i < nr_pages; i++) {
Al Viro9d966d42007-10-14 19:34:10 +01001037 const void __user *p;
Christoph Lameter742755a2006-06-23 02:03:55 -07001038
1039 err = -EFAULT;
1040 if (get_user(p, pages + i))
1041 goto out;
1042
1043 pm[i].addr = (unsigned long)p;
1044 if (nodes) {
1045 int node;
1046
1047 if (get_user(node, nodes + i))
1048 goto out;
1049
1050 err = -ENODEV;
Christoph Lameter56bbd652007-10-16 01:25:35 -07001051 if (!node_state(node, N_HIGH_MEMORY))
Christoph Lameter742755a2006-06-23 02:03:55 -07001052 goto out;
1053
1054 err = -EACCES;
1055 if (!node_isset(node, task_nodes))
1056 goto out;
1057
1058 pm[i].node = node;
Stephen Rothwell8ce08462006-11-02 22:07:28 -08001059 } else
1060 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
Christoph Lameter742755a2006-06-23 02:03:55 -07001061 }
1062 /* End marker */
1063 pm[nr_pages].node = MAX_NUMNODES;
1064
1065 if (nodes)
1066 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1067 else
1068 err = do_pages_stat(mm, pm);
1069
1070 if (err >= 0)
1071 /* Return status information */
1072 for (i = 0; i < nr_pages; i++)
1073 if (put_user(pm[i].status, status + i))
1074 err = -EFAULT;
1075
1076out:
1077 vfree(pm);
1078out2:
1079 mmput(mm);
1080 return err;
1081}
Christoph Lameter742755a2006-06-23 02:03:55 -07001082
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001083/*
1084 * Call migration functions in the vma_ops that may prepare
1085 * memory in a vm for migration. migration functions may perform
1086 * the migration for vmas that do not have an underlying page struct.
1087 */
1088int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1089 const nodemask_t *from, unsigned long flags)
1090{
1091 struct vm_area_struct *vma;
1092 int err = 0;
1093
1094 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1095 if (vma->vm_ops && vma->vm_ops->migrate) {
1096 err = vma->vm_ops->migrate(vma, to, from, flags);
1097 if (err)
1098 break;
1099 }
1100 }
1101 return err;
1102}
Gerald Schaefer83d16742008-07-23 21:28:22 -07001103#endif