blob: 112bcaeaa1046076af8618b7f06b2cfbc32faf8a [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>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080032#include <linux/memcontrol.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080033
34#include "internal.h"
35
Christoph Lameterb20a3502006-03-22 00:09:12 -080036#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
37
38/*
39 * Isolate one page from the LRU lists. If successful put it onto
40 * the indicated list with elevated page count.
41 *
42 * Result:
43 * -EBUSY: page not on LRU list
44 * 0: page removed from LRU list and added to the specified list.
45 */
46int isolate_lru_page(struct page *page, struct list_head *pagelist)
47{
48 int ret = -EBUSY;
49
50 if (PageLRU(page)) {
51 struct zone *zone = page_zone(page);
52
53 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyuki3dd9fe82007-07-26 10:41:08 -070054 if (PageLRU(page) && get_page_unless_zero(page)) {
Christoph Lameterb20a3502006-03-22 00:09:12 -080055 ret = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -080056 ClearPageLRU(page);
57 if (PageActive(page))
58 del_page_from_active_list(zone, page);
59 else
60 del_page_from_inactive_list(zone, page);
61 list_add_tail(&page->lru, pagelist);
62 }
63 spin_unlock_irq(&zone->lru_lock);
64 }
65 return ret;
66}
67
68/*
Christoph Lameter742755a2006-06-23 02:03:55 -070069 * migrate_prep() needs to be called before we start compiling a list of pages
70 * to be migrated using isolate_lru_page().
Christoph Lameterb20a3502006-03-22 00:09:12 -080071 */
72int migrate_prep(void)
73{
Christoph Lameterb20a3502006-03-22 00:09:12 -080074 /*
75 * Clear the LRU lists so pages can be isolated.
76 * Note that pages may be moved off the LRU after we have
77 * drained them. Those pages will fail to migrate like other
78 * pages that may be busy.
79 */
80 lru_add_drain_all();
81
82 return 0;
83}
84
85static inline void move_to_lru(struct page *page)
86{
Christoph Lameterb20a3502006-03-22 00:09:12 -080087 if (PageActive(page)) {
88 /*
89 * lru_cache_add_active checks that
90 * the PG_active bit is off.
91 */
92 ClearPageActive(page);
93 lru_cache_add_active(page);
94 } else {
95 lru_cache_add(page);
96 }
97 put_page(page);
98}
99
100/*
101 * Add isolated pages on the list back to the LRU.
102 *
103 * returns the number of pages put back.
104 */
105int putback_lru_pages(struct list_head *l)
106{
107 struct page *page;
108 struct page *page2;
109 int count = 0;
110
111 list_for_each_entry_safe(page, page2, l, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700112 list_del(&page->lru);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800113 move_to_lru(page);
114 count++;
115 }
116 return count;
117}
118
Christoph Lameter06972122006-06-23 02:03:35 -0700119/*
120 * Restore a potential migration pte to a working pte entry
121 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700122static void remove_migration_pte(struct vm_area_struct *vma,
Christoph Lameter06972122006-06-23 02:03:35 -0700123 struct page *old, struct page *new)
124{
125 struct mm_struct *mm = vma->vm_mm;
126 swp_entry_t entry;
127 pgd_t *pgd;
128 pud_t *pud;
129 pmd_t *pmd;
130 pte_t *ptep, pte;
131 spinlock_t *ptl;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700132 unsigned long addr = page_address_in_vma(new, vma);
133
134 if (addr == -EFAULT)
135 return;
Christoph Lameter06972122006-06-23 02:03:35 -0700136
137 pgd = pgd_offset(mm, addr);
138 if (!pgd_present(*pgd))
139 return;
140
141 pud = pud_offset(pgd, addr);
142 if (!pud_present(*pud))
143 return;
144
145 pmd = pmd_offset(pud, addr);
146 if (!pmd_present(*pmd))
147 return;
148
149 ptep = pte_offset_map(pmd, addr);
150
151 if (!is_swap_pte(*ptep)) {
152 pte_unmap(ptep);
153 return;
154 }
155
156 ptl = pte_lockptr(mm, pmd);
157 spin_lock(ptl);
158 pte = *ptep;
159 if (!is_swap_pte(pte))
160 goto out;
161
162 entry = pte_to_swp_entry(pte);
163
164 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
165 goto out;
166
Hugh Dickins98837c72008-03-04 14:29:06 -0800167 /*
168 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
169 * Failure is not an option here: we're now expected to remove every
170 * migration pte, and will cause crashes otherwise. Normally this
171 * is not an issue: mem_cgroup_prepare_migration bumped up the old
172 * page_cgroup count for safety, that's now attached to the new page,
173 * so this charge should just be another incrementation of the count,
174 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
175 * there's been a force_empty, those reference counts may no longer
176 * be reliable, and this charge can actually fail: oh well, we don't
177 * make the situation any worse by proceeding as if it had succeeded.
178 */
179 mem_cgroup_charge(new, mm, GFP_ATOMIC);
180
Christoph Lameter06972122006-06-23 02:03:35 -0700181 get_page(new);
182 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
183 if (is_write_migration_entry(entry))
184 pte = pte_mkwrite(pte);
KAMEZAWA Hiroyuki97ee0522007-10-16 01:25:43 -0700185 flush_cache_page(vma, addr, pte_pfn(pte));
Christoph Lameter06972122006-06-23 02:03:35 -0700186 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700187
188 if (PageAnon(new))
189 page_add_anon_rmap(new, vma, addr);
190 else
191 page_add_file_rmap(new);
192
193 /* No need to invalidate - it was non-present before */
194 update_mmu_cache(vma, addr, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700195
Christoph Lameter06972122006-06-23 02:03:35 -0700196out:
197 pte_unmap_unlock(ptep, ptl);
198}
199
200/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700201 * Note that remove_file_migration_ptes will only work on regular mappings,
202 * Nonlinear mappings do not use migration entries.
203 */
204static void remove_file_migration_ptes(struct page *old, struct page *new)
205{
206 struct vm_area_struct *vma;
207 struct address_space *mapping = page_mapping(new);
208 struct prio_tree_iter iter;
209 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
210
211 if (!mapping)
212 return;
213
214 spin_lock(&mapping->i_mmap_lock);
215
216 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
217 remove_migration_pte(vma, old, new);
218
219 spin_unlock(&mapping->i_mmap_lock);
220}
221
222/*
Christoph Lameter06972122006-06-23 02:03:35 -0700223 * Must hold mmap_sem lock on at least one of the vmas containing
224 * the page so that the anon_vma cannot vanish.
225 */
Christoph Lameter04e62a22006-06-23 02:03:38 -0700226static void remove_anon_migration_ptes(struct page *old, struct page *new)
Christoph Lameter06972122006-06-23 02:03:35 -0700227{
228 struct anon_vma *anon_vma;
229 struct vm_area_struct *vma;
230 unsigned long mapping;
231
232 mapping = (unsigned long)new->mapping;
233
234 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
235 return;
236
237 /*
238 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
239 */
240 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
241 spin_lock(&anon_vma->lock);
242
243 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
Christoph Lameter04e62a22006-06-23 02:03:38 -0700244 remove_migration_pte(vma, old, new);
Christoph Lameter06972122006-06-23 02:03:35 -0700245
246 spin_unlock(&anon_vma->lock);
247}
248
249/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700250 * Get rid of all migration entries and replace them by
251 * references to the indicated page.
252 */
253static void remove_migration_ptes(struct page *old, struct page *new)
254{
255 if (PageAnon(new))
256 remove_anon_migration_ptes(old, new);
257 else
258 remove_file_migration_ptes(old, new);
259}
260
261/*
Christoph Lameter06972122006-06-23 02:03:35 -0700262 * Something used the pte of a page under migration. We need to
263 * get to the page and wait until migration is finished.
264 * When we return from this function the fault will be retried.
265 *
266 * This function is called from do_swap_page().
267 */
268void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
269 unsigned long address)
270{
271 pte_t *ptep, pte;
272 spinlock_t *ptl;
273 swp_entry_t entry;
274 struct page *page;
275
276 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
277 pte = *ptep;
278 if (!is_swap_pte(pte))
279 goto out;
280
281 entry = pte_to_swp_entry(pte);
282 if (!is_migration_entry(entry))
283 goto out;
284
285 page = migration_entry_to_page(entry);
286
287 get_page(page);
288 pte_unmap_unlock(ptep, ptl);
289 wait_on_page_locked(page);
290 put_page(page);
291 return;
292out:
293 pte_unmap_unlock(ptep, ptl);
294}
295
Christoph Lameterb20a3502006-03-22 00:09:12 -0800296/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700297 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700298 *
299 * The number of remaining references must be:
300 * 1 for anonymous pages without a mapping
301 * 2 for pages with a mapping
302 * 3 for pages with a mapping and PagePrivate set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800303 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700304static int migrate_page_move_mapping(struct address_space *mapping,
305 struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800306{
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800307 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800308
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700309 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700310 /* Anonymous page without mapping */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700311 if (page_count(page) != 1)
312 return -EAGAIN;
313 return 0;
314 }
315
Christoph Lameterb20a3502006-03-22 00:09:12 -0800316 write_lock_irq(&mapping->tree_lock);
317
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800318 pslot = radix_tree_lookup_slot(&mapping->page_tree,
319 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800320
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700321 if (page_count(page) != 2 + !!PagePrivate(page) ||
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800322 (struct page *)radix_tree_deref_slot(pslot) != page) {
Christoph Lameterb20a3502006-03-22 00:09:12 -0800323 write_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700324 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800325 }
326
327 /*
328 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800329 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800330 get_page(newpage); /* add cache reference */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700331#ifdef CONFIG_SWAP
Christoph Lameterb20a3502006-03-22 00:09:12 -0800332 if (PageSwapCache(page)) {
333 SetPageSwapCache(newpage);
334 set_page_private(newpage, page_private(page));
335 }
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700336#endif
Christoph Lameterb20a3502006-03-22 00:09:12 -0800337
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800338 radix_tree_replace_slot(pslot, newpage);
339
340 /*
341 * Drop cache reference from old page.
342 * We know this isn't the last reference.
343 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800344 __put_page(page);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800345
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700346 /*
347 * If moved to a different zone then also account
348 * the page for that zone. Other VM counters will be
349 * taken care of when we establish references to the
350 * new page and drop references to the old page.
351 *
352 * Note that anonymous pages are accounted for
353 * via NR_FILE_PAGES and NR_ANON_PAGES if they
354 * are mapped to swap space.
355 */
356 __dec_zone_page_state(page, NR_FILE_PAGES);
357 __inc_zone_page_state(newpage, NR_FILE_PAGES);
358
Christoph Lameterb20a3502006-03-22 00:09:12 -0800359 write_unlock_irq(&mapping->tree_lock);
360
361 return 0;
362}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800363
364/*
365 * Copy the page to its new location
366 */
Christoph Lametere7340f72006-06-23 02:03:29 -0700367static void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800368{
369 copy_highpage(newpage, page);
370
371 if (PageError(page))
372 SetPageError(newpage);
373 if (PageReferenced(page))
374 SetPageReferenced(newpage);
375 if (PageUptodate(page))
376 SetPageUptodate(newpage);
377 if (PageActive(page))
378 SetPageActive(newpage);
379 if (PageChecked(page))
380 SetPageChecked(newpage);
381 if (PageMappedToDisk(page))
382 SetPageMappedToDisk(newpage);
383
384 if (PageDirty(page)) {
385 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700386 /*
387 * Want to mark the page and the radix tree as dirty, and
388 * redo the accounting that clear_page_dirty_for_io undid,
389 * but we can't use set_page_dirty because that function
390 * is actually a signal that all of the page has become dirty.
391 * Wheras only part of our page may be dirty.
392 */
393 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800394 }
395
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700396#ifdef CONFIG_SWAP
Christoph Lameterb20a3502006-03-22 00:09:12 -0800397 ClearPageSwapCache(page);
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700398#endif
Christoph Lameterb20a3502006-03-22 00:09:12 -0800399 ClearPageActive(page);
400 ClearPagePrivate(page);
401 set_page_private(page, 0);
402 page->mapping = NULL;
403
404 /*
405 * If any waiters have accumulated on the new page then
406 * wake them up.
407 */
408 if (PageWriteback(newpage))
409 end_page_writeback(newpage);
410}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800411
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700412/************************************************************
413 * Migration functions
414 ***********************************************************/
415
416/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700417int fail_migrate_page(struct address_space *mapping,
418 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700419{
420 return -EIO;
421}
422EXPORT_SYMBOL(fail_migrate_page);
423
Christoph Lameterb20a3502006-03-22 00:09:12 -0800424/*
425 * Common logic to directly migrate a single page suitable for
426 * pages that do not use PagePrivate.
427 *
428 * Pages are locked upon entry and exit.
429 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700430int migrate_page(struct address_space *mapping,
431 struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800432{
433 int rc;
434
435 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
436
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700437 rc = migrate_page_move_mapping(mapping, newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800438
439 if (rc)
440 return rc;
441
442 migrate_page_copy(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800443 return 0;
444}
445EXPORT_SYMBOL(migrate_page);
446
David Howells93614012006-09-30 20:45:40 +0200447#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800448/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700449 * Migration function for pages with buffers. This function can only be used
450 * if the underlying filesystem guarantees that no other references to "page"
451 * exist.
452 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700453int buffer_migrate_page(struct address_space *mapping,
454 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700455{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700456 struct buffer_head *bh, *head;
457 int rc;
458
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700459 if (!page_has_buffers(page))
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700460 return migrate_page(mapping, newpage, page);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700461
462 head = page_buffers(page);
463
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700464 rc = migrate_page_move_mapping(mapping, newpage, page);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700465
466 if (rc)
467 return rc;
468
469 bh = head;
470 do {
471 get_bh(bh);
472 lock_buffer(bh);
473 bh = bh->b_this_page;
474
475 } while (bh != head);
476
477 ClearPagePrivate(page);
478 set_page_private(newpage, page_private(page));
479 set_page_private(page, 0);
480 put_page(page);
481 get_page(newpage);
482
483 bh = head;
484 do {
485 set_bh_page(bh, newpage, bh_offset(bh));
486 bh = bh->b_this_page;
487
488 } while (bh != head);
489
490 SetPagePrivate(newpage);
491
492 migrate_page_copy(newpage, page);
493
494 bh = head;
495 do {
496 unlock_buffer(bh);
497 put_bh(bh);
498 bh = bh->b_this_page;
499
500 } while (bh != head);
501
502 return 0;
503}
504EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200505#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700506
Christoph Lameter04e62a22006-06-23 02:03:38 -0700507/*
508 * Writeback a page to clean the dirty state
509 */
510static int writeout(struct address_space *mapping, struct page *page)
511{
512 struct writeback_control wbc = {
513 .sync_mode = WB_SYNC_NONE,
514 .nr_to_write = 1,
515 .range_start = 0,
516 .range_end = LLONG_MAX,
517 .nonblocking = 1,
518 .for_reclaim = 1
519 };
520 int rc;
521
522 if (!mapping->a_ops->writepage)
523 /* No write method for the address space */
524 return -EINVAL;
525
526 if (!clear_page_dirty_for_io(page))
527 /* Someone else already triggered a write */
528 return -EAGAIN;
529
530 /*
531 * A dirty page may imply that the underlying filesystem has
532 * the page on some queue. So the page must be clean for
533 * migration. Writeout may mean we loose the lock and the
534 * page state is no longer what we checked for earlier.
535 * At this point we know that the migration attempt cannot
536 * be successful.
537 */
538 remove_migration_ptes(page, page);
539
540 rc = mapping->a_ops->writepage(page, &wbc);
541 if (rc < 0)
542 /* I/O Error writing */
543 return -EIO;
544
545 if (rc != AOP_WRITEPAGE_ACTIVATE)
546 /* unlocked. Relock */
547 lock_page(page);
548
549 return -EAGAIN;
550}
551
552/*
553 * Default handling if a filesystem does not provide a migration function.
554 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700555static int fallback_migrate_page(struct address_space *mapping,
556 struct page *newpage, struct page *page)
557{
Christoph Lameter04e62a22006-06-23 02:03:38 -0700558 if (PageDirty(page))
559 return writeout(mapping, page);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700560
561 /*
562 * Buffers may be managed in a filesystem specific way.
563 * We must have no buffers or drop them.
564 */
David Howellsb398f6b2006-08-29 19:05:58 +0100565 if (PagePrivate(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700566 !try_to_release_page(page, GFP_KERNEL))
567 return -EAGAIN;
568
569 return migrate_page(mapping, newpage, page);
570}
571
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700572/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700573 * Move a page to a newly allocated page
574 * The page is locked and all ptes have been successfully removed.
575 *
576 * The new page will have replaced the old page if this function
577 * is successful.
578 */
579static int move_to_new_page(struct page *newpage, struct page *page)
580{
581 struct address_space *mapping;
582 int rc;
583
584 /*
585 * Block others from accessing the page when we get around to
586 * establishing additional references. We are the only one
587 * holding a reference to the new page at this point.
588 */
589 if (TestSetPageLocked(newpage))
590 BUG();
591
592 /* Prepare mapping for the new page.*/
593 newpage->index = page->index;
594 newpage->mapping = page->mapping;
595
596 mapping = page_mapping(page);
597 if (!mapping)
598 rc = migrate_page(mapping, newpage, page);
599 else if (mapping->a_ops->migratepage)
600 /*
601 * Most pages have a mapping and most filesystems
602 * should provide a migration function. Anonymous
603 * pages are part of swap space which also has its
604 * own migration function. This is the most common
605 * path for page migration.
606 */
607 rc = mapping->a_ops->migratepage(mapping,
608 newpage, page);
609 else
610 rc = fallback_migrate_page(mapping, newpage, page);
611
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800612 if (!rc) {
613 mem_cgroup_page_migration(page, newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700614 remove_migration_ptes(page, newpage);
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800615 } else
Christoph Lametere24f0b82006-06-23 02:03:51 -0700616 newpage->mapping = NULL;
617
618 unlock_page(newpage);
619
620 return rc;
621}
622
623/*
624 * Obtain the lock on page, remove all ptes and migrate the page
625 * to the newly allocated page in newpage.
626 */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700627static int unmap_and_move(new_page_t get_new_page, unsigned long private,
628 struct page *page, int force)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700629{
630 int rc = 0;
Christoph Lameter742755a2006-06-23 02:03:55 -0700631 int *result = NULL;
632 struct page *newpage = get_new_page(page, private, &result);
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700633 int rcu_locked = 0;
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800634 int charge = 0;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700635
636 if (!newpage)
637 return -ENOMEM;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700638
639 if (page_count(page) == 1)
640 /* page was freed from under us. So we are done. */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700641 goto move_newpage;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700642
643 rc = -EAGAIN;
644 if (TestSetPageLocked(page)) {
645 if (!force)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700646 goto move_newpage;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700647 lock_page(page);
648 }
649
650 if (PageWriteback(page)) {
651 if (!force)
652 goto unlock;
653 wait_on_page_writeback(page);
654 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700655 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700656 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
657 * we cannot notice that anon_vma is freed while we migrates a page.
658 * This rcu_read_lock() delays freeing anon_vma pointer until the end
659 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700660 * File Caches may use write_page() or lock_page() in migration, then,
661 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700662 */
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700663 if (PageAnon(page)) {
664 rcu_read_lock();
665 rcu_locked = 1;
666 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800667
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700668 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800669 * Corner case handling:
670 * 1. When a new swap-cache page is read into, it is added to the LRU
671 * and treated as swapcache but it has no rmap yet.
672 * Calling try_to_unmap() against a page->mapping==NULL page will
673 * trigger a BUG. So handle it here.
674 * 2. An orphaned page (see truncate_complete_page) might have
675 * fs-private metadata. The page can be picked up due to memory
676 * offlining. Everywhere else except page reclaim, the page is
677 * invisible to the vm, so the page can not be migrated. So try to
678 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700679 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800680 if (!page->mapping) {
681 if (!PageAnon(page) && PagePrivate(page)) {
682 /*
683 * Go direct to try_to_free_buffers() here because
684 * a) that's what try_to_release_page() would do anyway
685 * b) we may be under rcu_read_lock() here, so we can't
686 * use GFP_KERNEL which is what try_to_release_page()
687 * needs to be effective.
688 */
689 try_to_free_buffers(page);
690 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700691 goto rcu_unlock;
Shaohua Li62e1c552008-02-04 22:29:33 -0800692 }
693
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800694 charge = mem_cgroup_prepare_migration(page);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700695 /* Establish migration ptes or remove ptes */
Christoph Lametere6a15302006-06-25 05:46:49 -0700696 try_to_unmap(page, 1);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700697
Christoph Lametere6a15302006-06-25 05:46:49 -0700698 if (!page_mapped(page))
699 rc = move_to_new_page(newpage, page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700700
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800701 if (rc) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700702 remove_migration_ptes(page, page);
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800703 if (charge)
704 mem_cgroup_end_migration(page);
705 } else if (charge)
706 mem_cgroup_end_migration(newpage);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700707rcu_unlock:
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700708 if (rcu_locked)
709 rcu_read_unlock();
Christoph Lametere6a15302006-06-25 05:46:49 -0700710
Christoph Lametere24f0b82006-06-23 02:03:51 -0700711unlock:
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700712
Christoph Lametere24f0b82006-06-23 02:03:51 -0700713 unlock_page(page);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700714
Christoph Lametere24f0b82006-06-23 02:03:51 -0700715 if (rc != -EAGAIN) {
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700716 /*
717 * A page that has been migrated has all references
718 * removed and will be freed. A page that has not been
719 * migrated will have kepts its references and be
720 * restored.
721 */
722 list_del(&page->lru);
723 move_to_lru(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700724 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700725
726move_newpage:
727 /*
728 * Move the new page to the LRU. If migration was not successful
729 * then this will free the page.
730 */
731 move_to_lru(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700732 if (result) {
733 if (rc)
734 *result = rc;
735 else
736 *result = page_to_nid(newpage);
737 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700738 return rc;
739}
740
741/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800742 * migrate_pages
743 *
Christoph Lameter95a402c2006-06-23 02:03:53 -0700744 * The function takes one list of pages to migrate and a function
745 * that determines from the page to be migrated and the private data
746 * the target of the move and allocates the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800747 *
748 * The function returns after 10 attempts or if no pages
749 * are movable anymore because to has become empty
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700750 * or no retryable pages exist anymore. All pages will be
Gabriel Craciunescue9534b32007-10-20 02:13:26 +0200751 * returned to the LRU or freed.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800752 *
Christoph Lameter95a402c2006-06-23 02:03:53 -0700753 * Return: Number of pages not migrated or error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800754 */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700755int migrate_pages(struct list_head *from,
756 new_page_t get_new_page, unsigned long private)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800757{
Christoph Lametere24f0b82006-06-23 02:03:51 -0700758 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800759 int nr_failed = 0;
760 int pass = 0;
761 struct page *page;
762 struct page *page2;
763 int swapwrite = current->flags & PF_SWAPWRITE;
764 int rc;
765
766 if (!swapwrite)
767 current->flags |= PF_SWAPWRITE;
768
Christoph Lametere24f0b82006-06-23 02:03:51 -0700769 for(pass = 0; pass < 10 && retry; pass++) {
770 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800771
Christoph Lametere24f0b82006-06-23 02:03:51 -0700772 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700773 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -0800774
Christoph Lameter95a402c2006-06-23 02:03:53 -0700775 rc = unmap_and_move(get_new_page, private,
776 page, pass > 2);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800777
Christoph Lametere24f0b82006-06-23 02:03:51 -0700778 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -0700779 case -ENOMEM:
780 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700781 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700782 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700783 break;
784 case 0:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700785 break;
786 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700787 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700788 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700789 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700790 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800791 }
792 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700793 rc = 0;
794out:
Christoph Lameterb20a3502006-03-22 00:09:12 -0800795 if (!swapwrite)
796 current->flags &= ~PF_SWAPWRITE;
797
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700798 putback_lru_pages(from);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700799
800 if (rc)
801 return rc;
802
Christoph Lameterb20a3502006-03-22 00:09:12 -0800803 return nr_failed + retry;
804}
805
Christoph Lameter742755a2006-06-23 02:03:55 -0700806#ifdef CONFIG_NUMA
807/*
808 * Move a list of individual pages
809 */
810struct page_to_node {
811 unsigned long addr;
812 struct page *page;
813 int node;
814 int status;
815};
816
817static struct page *new_page_node(struct page *p, unsigned long private,
818 int **result)
819{
820 struct page_to_node *pm = (struct page_to_node *)private;
821
822 while (pm->node != MAX_NUMNODES && pm->page != p)
823 pm++;
824
825 if (pm->node == MAX_NUMNODES)
826 return NULL;
827
828 *result = &pm->status;
829
Mel Gorman769848c2007-07-17 04:03:05 -0700830 return alloc_pages_node(pm->node,
831 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -0700832}
833
834/*
835 * Move a set of pages as indicated in the pm array. The addr
836 * field must be set to the virtual address of the page to be moved
837 * and the node number must contain a valid target node.
838 */
839static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
840 int migrate_all)
841{
842 int err;
843 struct page_to_node *pp;
844 LIST_HEAD(pagelist);
845
846 down_read(&mm->mmap_sem);
847
848 /*
849 * Build a list of pages to migrate
850 */
851 migrate_prep();
852 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
853 struct vm_area_struct *vma;
854 struct page *page;
855
856 /*
857 * A valid page pointer that will not match any of the
858 * pages that will be moved.
859 */
860 pp->page = ZERO_PAGE(0);
861
862 err = -EFAULT;
863 vma = find_vma(mm, pp->addr);
Christoph Lameter0dc952d2007-03-05 00:30:33 -0800864 if (!vma || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -0700865 goto set_status;
866
867 page = follow_page(vma, pp->addr, FOLL_GET);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -0700868
869 err = PTR_ERR(page);
870 if (IS_ERR(page))
871 goto set_status;
872
Christoph Lameter742755a2006-06-23 02:03:55 -0700873 err = -ENOENT;
874 if (!page)
875 goto set_status;
876
877 if (PageReserved(page)) /* Check for zero page */
878 goto put_and_set;
879
880 pp->page = page;
881 err = page_to_nid(page);
882
883 if (err == pp->node)
884 /*
885 * Node already in the right place
886 */
887 goto put_and_set;
888
889 err = -EACCES;
890 if (page_mapcount(page) > 1 &&
891 !migrate_all)
892 goto put_and_set;
893
894 err = isolate_lru_page(page, &pagelist);
895put_and_set:
896 /*
897 * Either remove the duplicate refcount from
898 * isolate_lru_page() or drop the page ref if it was
899 * not isolated.
900 */
901 put_page(page);
902set_status:
903 pp->status = err;
904 }
905
906 if (!list_empty(&pagelist))
907 err = migrate_pages(&pagelist, new_page_node,
908 (unsigned long)pm);
909 else
910 err = -ENOENT;
911
912 up_read(&mm->mmap_sem);
913 return err;
914}
915
916/*
917 * Determine the nodes of a list of pages. The addr in the pm array
918 * must have been set to the virtual address of which we want to determine
919 * the node number.
920 */
921static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
922{
923 down_read(&mm->mmap_sem);
924
925 for ( ; pm->node != MAX_NUMNODES; pm++) {
926 struct vm_area_struct *vma;
927 struct page *page;
928 int err;
929
930 err = -EFAULT;
931 vma = find_vma(mm, pm->addr);
932 if (!vma)
933 goto set_status;
934
935 page = follow_page(vma, pm->addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -0700936
937 err = PTR_ERR(page);
938 if (IS_ERR(page))
939 goto set_status;
940
Christoph Lameter742755a2006-06-23 02:03:55 -0700941 err = -ENOENT;
942 /* Use PageReserved to check for zero page */
943 if (!page || PageReserved(page))
944 goto set_status;
945
946 err = page_to_nid(page);
947set_status:
948 pm->status = err;
949 }
950
951 up_read(&mm->mmap_sem);
952 return 0;
953}
954
955/*
956 * Move a list of pages in the address space of the currently executing
957 * process.
958 */
959asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
960 const void __user * __user *pages,
961 const int __user *nodes,
962 int __user *status, int flags)
963{
964 int err = 0;
965 int i;
966 struct task_struct *task;
967 nodemask_t task_nodes;
968 struct mm_struct *mm;
969 struct page_to_node *pm = NULL;
970
971 /* Check flags */
972 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
973 return -EINVAL;
974
975 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
976 return -EPERM;
977
978 /* Find the mm_struct */
979 read_lock(&tasklist_lock);
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -0700980 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -0700981 if (!task) {
982 read_unlock(&tasklist_lock);
983 return -ESRCH;
984 }
985 mm = get_task_mm(task);
986 read_unlock(&tasklist_lock);
987
988 if (!mm)
989 return -EINVAL;
990
991 /*
992 * Check if this process has the right to modify the specified
993 * process. The right exists if the process has administrative
994 * capabilities, superuser privileges or the same
995 * userid as the target process.
996 */
997 if ((current->euid != task->suid) && (current->euid != task->uid) &&
998 (current->uid != task->suid) && (current->uid != task->uid) &&
999 !capable(CAP_SYS_NICE)) {
1000 err = -EPERM;
1001 goto out2;
1002 }
1003
David Quigley86c3a762006-06-23 02:04:02 -07001004 err = security_task_movememory(task);
1005 if (err)
1006 goto out2;
1007
1008
Christoph Lameter742755a2006-06-23 02:03:55 -07001009 task_nodes = cpuset_mems_allowed(task);
1010
1011 /* Limit nr_pages so that the multiplication may not overflow */
1012 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1013 err = -E2BIG;
1014 goto out2;
1015 }
1016
1017 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1018 if (!pm) {
1019 err = -ENOMEM;
1020 goto out2;
1021 }
1022
1023 /*
1024 * Get parameters from user space and initialize the pm
1025 * array. Return various errors if the user did something wrong.
1026 */
1027 for (i = 0; i < nr_pages; i++) {
Al Viro9d966d42007-10-14 19:34:10 +01001028 const void __user *p;
Christoph Lameter742755a2006-06-23 02:03:55 -07001029
1030 err = -EFAULT;
1031 if (get_user(p, pages + i))
1032 goto out;
1033
1034 pm[i].addr = (unsigned long)p;
1035 if (nodes) {
1036 int node;
1037
1038 if (get_user(node, nodes + i))
1039 goto out;
1040
1041 err = -ENODEV;
Christoph Lameter56bbd652007-10-16 01:25:35 -07001042 if (!node_state(node, N_HIGH_MEMORY))
Christoph Lameter742755a2006-06-23 02:03:55 -07001043 goto out;
1044
1045 err = -EACCES;
1046 if (!node_isset(node, task_nodes))
1047 goto out;
1048
1049 pm[i].node = node;
Stephen Rothwell8ce08462006-11-02 22:07:28 -08001050 } else
1051 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
Christoph Lameter742755a2006-06-23 02:03:55 -07001052 }
1053 /* End marker */
1054 pm[nr_pages].node = MAX_NUMNODES;
1055
1056 if (nodes)
1057 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1058 else
1059 err = do_pages_stat(mm, pm);
1060
1061 if (err >= 0)
1062 /* Return status information */
1063 for (i = 0; i < nr_pages; i++)
1064 if (put_user(pm[i].status, status + i))
1065 err = -EFAULT;
1066
1067out:
1068 vfree(pm);
1069out2:
1070 mmput(mm);
1071 return err;
1072}
1073#endif
1074
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001075/*
1076 * Call migration functions in the vma_ops that may prepare
1077 * memory in a vm for migration. migration functions may perform
1078 * the migration for vmas that do not have an underlying page struct.
1079 */
1080int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1081 const nodemask_t *from, unsigned long flags)
1082{
1083 struct vm_area_struct *vma;
1084 int err = 0;
1085
1086 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1087 if (vma->vm_ops && vma->vm_ops->migrate) {
1088 err = vma->vm_ops->migrate(vma, to, from, flags);
1089 if (err)
1090 break;
1091 }
1092 }
1093 return err;
1094}