blob: 89ea0854332ec3fdf53123bbf7fe5f1de94f6599 [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>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040016#include <linux/export.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080017#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>
Hugh Dickinse9995ef2009-12-14 17:59:31 -080024#include <linux/ksm.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080025#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
Christoph Lameter04e62a22006-06-23 02:03:38 -070029#include <linux/writeback.h>
Christoph Lameter742755a2006-06-23 02:03:55 -070030#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
David Quigley86c3a762006-06-23 02:04:02 -070032#include <linux/security.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080033#include <linux/memcontrol.h>
Adrian Bunk4f5ca262008-07-23 21:27:02 -070034#include <linux/syscalls.h>
Naoya Horiguchi290408d2010-09-08 10:19:35 +090035#include <linux/hugetlb.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090036#include <linux/gfp.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080037
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080038#include <asm/tlbflush.h>
39
Christoph Lameterb20a3502006-03-22 00:09:12 -080040#include "internal.h"
41
Christoph Lameterb20a3502006-03-22 00:09:12 -080042/*
Christoph Lameter742755a2006-06-23 02:03:55 -070043 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070044 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
45 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080046 */
47int migrate_prep(void)
48{
Christoph Lameterb20a3502006-03-22 00:09:12 -080049 /*
50 * Clear the LRU lists so pages can be isolated.
51 * Note that pages may be moved off the LRU after we have
52 * drained them. Those pages will fail to migrate like other
53 * pages that may be busy.
54 */
55 lru_add_drain_all();
56
57 return 0;
58}
59
Mel Gorman748446b2010-05-24 14:32:27 -070060/* Do the necessary work of migrate_prep but not if it involves other CPUs */
61int migrate_prep_local(void)
62{
63 lru_add_drain();
64
65 return 0;
66}
67
Christoph Lameterb20a3502006-03-22 00:09:12 -080068/*
Lee Schermerhorn894bc312008-10-18 20:26:39 -070069 * Add isolated pages on the list back to the LRU under page lock
70 * to avoid leaking evictable pages back onto unevictable list.
Christoph Lameterb20a3502006-03-22 00:09:12 -080071 */
Minchan Kime13861d2010-05-24 14:31:59 -070072void putback_lru_pages(struct list_head *l)
Christoph Lameterb20a3502006-03-22 00:09:12 -080073{
74 struct page *page;
75 struct page *page2;
Christoph Lameterb20a3502006-03-22 00:09:12 -080076
77 list_for_each_entry_safe(page, page2, l, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -070078 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -070079 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -070080 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -070081 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -080082 }
Christoph Lameterb20a3502006-03-22 00:09:12 -080083}
84
Christoph Lameter06972122006-06-23 02:03:35 -070085/*
86 * Restore a potential migration pte to a working pte entry
87 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -080088static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
89 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -070090{
91 struct mm_struct *mm = vma->vm_mm;
92 swp_entry_t entry;
93 pgd_t *pgd;
94 pud_t *pud;
95 pmd_t *pmd;
96 pte_t *ptep, pte;
97 spinlock_t *ptl;
98
Naoya Horiguchi290408d2010-09-08 10:19:35 +090099 if (unlikely(PageHuge(new))) {
100 ptep = huge_pte_offset(mm, addr);
101 if (!ptep)
102 goto out;
103 ptl = &mm->page_table_lock;
104 } else {
105 pgd = pgd_offset(mm, addr);
106 if (!pgd_present(*pgd))
107 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700108
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900109 pud = pud_offset(pgd, addr);
110 if (!pud_present(*pud))
111 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700112
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900113 pmd = pmd_offset(pud, addr);
Andrea Arcangeli500d65d2011-01-13 15:46:55 -0800114 if (pmd_trans_huge(*pmd))
115 goto out;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900116 if (!pmd_present(*pmd))
117 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700118
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900119 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700120
Hugh Dickins486cf462011-10-19 12:50:35 -0700121 /*
122 * Peek to check is_swap_pte() before taking ptlock? No, we
123 * can race mremap's move_ptes(), which skips anon_vma lock.
124 */
Christoph Lameter06972122006-06-23 02:03:35 -0700125
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900126 ptl = pte_lockptr(mm, pmd);
127 }
128
Christoph Lameter06972122006-06-23 02:03:35 -0700129 spin_lock(ptl);
130 pte = *ptep;
131 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800132 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700133
134 entry = pte_to_swp_entry(pte);
135
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800136 if (!is_migration_entry(entry) ||
137 migration_entry_to_page(entry) != old)
138 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700139
Christoph Lameter06972122006-06-23 02:03:35 -0700140 get_page(new);
141 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
142 if (is_write_migration_entry(entry))
143 pte = pte_mkwrite(pte);
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200144#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900145 if (PageHuge(new))
146 pte = pte_mkhuge(pte);
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200147#endif
KAMEZAWA Hiroyuki97ee0522007-10-16 01:25:43 -0700148 flush_cache_page(vma, addr, pte_pfn(pte));
Christoph Lameter06972122006-06-23 02:03:35 -0700149 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700150
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900151 if (PageHuge(new)) {
152 if (PageAnon(new))
153 hugepage_add_anon_rmap(new, vma, addr);
154 else
155 page_dup_rmap(new);
156 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700157 page_add_anon_rmap(new, vma, addr);
158 else
159 page_add_file_rmap(new);
160
161 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000162 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800163unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700164 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800165out:
166 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700167}
168
169/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700170 * Get rid of all migration entries and replace them by
171 * references to the indicated page.
172 */
173static void remove_migration_ptes(struct page *old, struct page *new)
174{
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800175 rmap_walk(new, remove_migration_pte, old);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700176}
177
178/*
Christoph Lameter06972122006-06-23 02:03:35 -0700179 * Something used the pte of a page under migration. We need to
180 * get to the page and wait until migration is finished.
181 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700182 */
183void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
184 unsigned long address)
185{
186 pte_t *ptep, pte;
187 spinlock_t *ptl;
188 swp_entry_t entry;
189 struct page *page;
190
191 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
192 pte = *ptep;
193 if (!is_swap_pte(pte))
194 goto out;
195
196 entry = pte_to_swp_entry(pte);
197 if (!is_migration_entry(entry))
198 goto out;
199
200 page = migration_entry_to_page(entry);
201
Nick Piggine2867812008-07-25 19:45:30 -0700202 /*
203 * Once radix-tree replacement of page migration started, page_count
204 * *must* be zero. And, we don't want to call wait_on_page_locked()
205 * against a page without get_page().
206 * So, we use get_page_unless_zero(), here. Even failed, page fault
207 * will occur again.
208 */
209 if (!get_page_unless_zero(page))
210 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700211 pte_unmap_unlock(ptep, ptl);
212 wait_on_page_locked(page);
213 put_page(page);
214 return;
215out:
216 pte_unmap_unlock(ptep, ptl);
217}
218
Christoph Lameterb20a3502006-03-22 00:09:12 -0800219/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700220 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700221 *
222 * The number of remaining references must be:
223 * 1 for anonymous pages without a mapping
224 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100225 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800226 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700227static int migrate_page_move_mapping(struct address_space *mapping,
228 struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800229{
Nick Piggine2867812008-07-25 19:45:30 -0700230 int expected_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800231 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800232
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700233 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700234 /* Anonymous page without mapping */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700235 if (page_count(page) != 1)
236 return -EAGAIN;
237 return 0;
238 }
239
Nick Piggin19fd6232008-07-25 19:45:32 -0700240 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800241
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800242 pslot = radix_tree_lookup_slot(&mapping->page_tree,
243 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800244
Johannes Weineredcf4742009-09-21 17:02:59 -0700245 expected_count = 2 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700246 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800247 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700248 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700249 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800250 }
251
Nick Piggine2867812008-07-25 19:45:30 -0700252 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700253 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700254 return -EAGAIN;
255 }
256
Christoph Lameterb20a3502006-03-22 00:09:12 -0800257 /*
258 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800259 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800260 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800261 if (PageSwapCache(page)) {
262 SetPageSwapCache(newpage);
263 set_page_private(newpage, page_private(page));
264 }
265
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800266 radix_tree_replace_slot(pslot, newpage);
267
268 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800269 * Drop cache reference from old page by unfreezing
270 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800271 * We know this isn't the last reference.
272 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800273 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800274
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700275 /*
276 * If moved to a different zone then also account
277 * the page for that zone. Other VM counters will be
278 * taken care of when we establish references to the
279 * new page and drop references to the old page.
280 *
281 * Note that anonymous pages are accounted for
282 * via NR_FILE_PAGES and NR_ANON_PAGES if they
283 * are mapped to swap space.
284 */
285 __dec_zone_page_state(page, NR_FILE_PAGES);
286 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700287 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700288 __dec_zone_page_state(page, NR_SHMEM);
289 __inc_zone_page_state(newpage, NR_SHMEM);
290 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700291 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800292
293 return 0;
294}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800295
296/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900297 * The expected number of remaining references is the same as that
298 * of migrate_page_move_mapping().
299 */
300int migrate_huge_page_move_mapping(struct address_space *mapping,
301 struct page *newpage, struct page *page)
302{
303 int expected_count;
304 void **pslot;
305
306 if (!mapping) {
307 if (page_count(page) != 1)
308 return -EAGAIN;
309 return 0;
310 }
311
312 spin_lock_irq(&mapping->tree_lock);
313
314 pslot = radix_tree_lookup_slot(&mapping->page_tree,
315 page_index(page));
316
317 expected_count = 2 + page_has_private(page);
318 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800319 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900320 spin_unlock_irq(&mapping->tree_lock);
321 return -EAGAIN;
322 }
323
324 if (!page_freeze_refs(page, expected_count)) {
325 spin_unlock_irq(&mapping->tree_lock);
326 return -EAGAIN;
327 }
328
329 get_page(newpage);
330
331 radix_tree_replace_slot(pslot, newpage);
332
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800333 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900334
335 spin_unlock_irq(&mapping->tree_lock);
336 return 0;
337}
338
339/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800340 * Copy the page to its new location
341 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900342void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800343{
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900344 if (PageHuge(page))
345 copy_huge_page(newpage, page);
346 else
347 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800348
349 if (PageError(page))
350 SetPageError(newpage);
351 if (PageReferenced(page))
352 SetPageReferenced(newpage);
353 if (PageUptodate(page))
354 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700355 if (TestClearPageActive(page)) {
356 VM_BUG_ON(PageUnevictable(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800357 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800358 } else if (TestClearPageUnevictable(page))
359 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800360 if (PageChecked(page))
361 SetPageChecked(newpage);
362 if (PageMappedToDisk(page))
363 SetPageMappedToDisk(newpage);
364
365 if (PageDirty(page)) {
366 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700367 /*
368 * Want to mark the page and the radix tree as dirty, and
369 * redo the accounting that clear_page_dirty_for_io undid,
370 * but we can't use set_page_dirty because that function
371 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300372 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700373 */
374 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800375 }
376
Nick Pigginb291f002008-10-18 20:26:44 -0700377 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800378 ksm_migrate_page(newpage, page);
Nick Pigginb291f002008-10-18 20:26:44 -0700379
Christoph Lameterb20a3502006-03-22 00:09:12 -0800380 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800381 ClearPagePrivate(page);
382 set_page_private(page, 0);
383 page->mapping = NULL;
384
385 /*
386 * If any waiters have accumulated on the new page then
387 * wake them up.
388 */
389 if (PageWriteback(newpage))
390 end_page_writeback(newpage);
391}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800392
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700393/************************************************************
394 * Migration functions
395 ***********************************************************/
396
397/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700398int fail_migrate_page(struct address_space *mapping,
399 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700400{
401 return -EIO;
402}
403EXPORT_SYMBOL(fail_migrate_page);
404
Christoph Lameterb20a3502006-03-22 00:09:12 -0800405/*
406 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100407 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800408 *
409 * Pages are locked upon entry and exit.
410 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700411int migrate_page(struct address_space *mapping,
412 struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800413{
414 int rc;
415
416 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
417
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700418 rc = migrate_page_move_mapping(mapping, newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800419
420 if (rc)
421 return rc;
422
423 migrate_page_copy(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800424 return 0;
425}
426EXPORT_SYMBOL(migrate_page);
427
David Howells93614012006-09-30 20:45:40 +0200428#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800429/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700430 * Migration function for pages with buffers. This function can only be used
431 * if the underlying filesystem guarantees that no other references to "page"
432 * exist.
433 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700434int buffer_migrate_page(struct address_space *mapping,
435 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700436{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700437 struct buffer_head *bh, *head;
438 int rc;
439
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700440 if (!page_has_buffers(page))
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700441 return migrate_page(mapping, newpage, page);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700442
443 head = page_buffers(page);
444
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700445 rc = migrate_page_move_mapping(mapping, newpage, page);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700446
447 if (rc)
448 return rc;
449
450 bh = head;
451 do {
452 get_bh(bh);
453 lock_buffer(bh);
454 bh = bh->b_this_page;
455
456 } while (bh != head);
457
458 ClearPagePrivate(page);
459 set_page_private(newpage, page_private(page));
460 set_page_private(page, 0);
461 put_page(page);
462 get_page(newpage);
463
464 bh = head;
465 do {
466 set_bh_page(bh, newpage, bh_offset(bh));
467 bh = bh->b_this_page;
468
469 } while (bh != head);
470
471 SetPagePrivate(newpage);
472
473 migrate_page_copy(newpage, page);
474
475 bh = head;
476 do {
477 unlock_buffer(bh);
478 put_bh(bh);
479 bh = bh->b_this_page;
480
481 } while (bh != head);
482
483 return 0;
484}
485EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200486#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700487
Christoph Lameter04e62a22006-06-23 02:03:38 -0700488/*
489 * Writeback a page to clean the dirty state
490 */
491static int writeout(struct address_space *mapping, struct page *page)
492{
493 struct writeback_control wbc = {
494 .sync_mode = WB_SYNC_NONE,
495 .nr_to_write = 1,
496 .range_start = 0,
497 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700498 .for_reclaim = 1
499 };
500 int rc;
501
502 if (!mapping->a_ops->writepage)
503 /* No write method for the address space */
504 return -EINVAL;
505
506 if (!clear_page_dirty_for_io(page))
507 /* Someone else already triggered a write */
508 return -EAGAIN;
509
510 /*
511 * A dirty page may imply that the underlying filesystem has
512 * the page on some queue. So the page must be clean for
513 * migration. Writeout may mean we loose the lock and the
514 * page state is no longer what we checked for earlier.
515 * At this point we know that the migration attempt cannot
516 * be successful.
517 */
518 remove_migration_ptes(page, page);
519
520 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700521
522 if (rc != AOP_WRITEPAGE_ACTIVATE)
523 /* unlocked. Relock */
524 lock_page(page);
525
Hugh Dickinsbda85502008-11-19 15:36:36 -0800526 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700527}
528
529/*
530 * Default handling if a filesystem does not provide a migration function.
531 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700532static int fallback_migrate_page(struct address_space *mapping,
533 struct page *newpage, struct page *page)
534{
Christoph Lameter04e62a22006-06-23 02:03:38 -0700535 if (PageDirty(page))
536 return writeout(mapping, page);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700537
538 /*
539 * Buffers may be managed in a filesystem specific way.
540 * We must have no buffers or drop them.
541 */
David Howells266cf652009-04-03 16:42:36 +0100542 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700543 !try_to_release_page(page, GFP_KERNEL))
544 return -EAGAIN;
545
546 return migrate_page(mapping, newpage, page);
547}
548
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700549/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700550 * Move a page to a newly allocated page
551 * The page is locked and all ptes have been successfully removed.
552 *
553 * The new page will have replaced the old page if this function
554 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700555 *
556 * Return value:
557 * < 0 - error code
558 * == 0 - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700559 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700560static int move_to_new_page(struct page *newpage, struct page *page,
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700561 int remap_swapcache, bool sync)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700562{
563 struct address_space *mapping;
564 int rc;
565
566 /*
567 * Block others from accessing the page when we get around to
568 * establishing additional references. We are the only one
569 * holding a reference to the new page at this point.
570 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200571 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700572 BUG();
573
574 /* Prepare mapping for the new page.*/
575 newpage->index = page->index;
576 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700577 if (PageSwapBacked(page))
578 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700579
580 mapping = page_mapping(page);
581 if (!mapping)
582 rc = migrate_page(mapping, newpage, page);
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700583 else {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700584 /*
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700585 * Do not writeback pages if !sync and migratepage is
586 * not pointing to migrate_page() which is nonblocking
587 * (swapcache/tmpfs uses migratepage = migrate_page).
Christoph Lametere24f0b82006-06-23 02:03:51 -0700588 */
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700589 if (PageDirty(page) && !sync &&
590 mapping->a_ops->migratepage != migrate_page)
591 rc = -EBUSY;
592 else if (mapping->a_ops->migratepage)
593 /*
594 * Most pages have a mapping and most filesystems
595 * should provide a migration function. Anonymous
596 * pages are part of swap space which also has its
597 * own migration function. This is the most common
598 * path for page migration.
599 */
600 rc = mapping->a_ops->migratepage(mapping,
601 newpage, page);
602 else
603 rc = fallback_migrate_page(mapping, newpage, page);
604 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700605
Mel Gorman3fe20112010-05-24 14:32:20 -0700606 if (rc) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700607 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700608 } else {
609 if (remap_swapcache)
610 remove_migration_ptes(page, newpage);
611 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700612
613 unlock_page(newpage);
614
615 return rc;
616}
617
Minchan Kim0dabec92011-10-31 17:06:57 -0700618static int __unmap_and_move(struct page *page, struct page *newpage,
619 int force, bool offlining, bool sync)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700620{
Minchan Kim0dabec92011-10-31 17:06:57 -0700621 int rc = -EAGAIN;
Mel Gorman3fe20112010-05-24 14:32:20 -0700622 int remap_swapcache = 1;
KAMEZAWA Hiroyukiae41be32008-02-07 00:14:10 -0800623 int charge = 0;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -0700624 struct mem_cgroup *mem;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700625 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700626
Nick Piggin529ae9a2008-08-02 12:01:03 +0200627 if (!trylock_page(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700628 if (!force || !sync)
Minchan Kim0dabec92011-10-31 17:06:57 -0700629 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800630
631 /*
632 * It's not safe for direct compaction to call lock_page.
633 * For example, during page readahead pages are added locked
634 * to the LRU. Later, when the IO completes the pages are
635 * marked uptodate and unlocked. However, the queueing
636 * could be merging multiple pages for one bio (e.g.
637 * mpage_readpages). If an allocation happens for the
638 * second or third page, the process can end up locking
639 * the same page twice and deadlocking. Rather than
640 * trying to be clever about what pages can be locked,
641 * avoid the use of lock_page for direct compaction
642 * altogether.
643 */
644 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700645 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800646
Christoph Lametere24f0b82006-06-23 02:03:51 -0700647 lock_page(page);
648 }
649
Hugh Dickins62b61f62009-12-14 17:59:33 -0800650 /*
651 * Only memory hotplug's offline_pages() caller has locked out KSM,
652 * and can safely migrate a KSM page. The other cases have skipped
653 * PageKsm along with PageReserved - but it is only now when we have
654 * the page lock that we can be certain it will not go KSM beneath us
655 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
656 * its pagecount raised, but only here do we take the page lock which
657 * serializes that).
658 */
659 if (PageKsm(page) && !offlining) {
660 rc = -EBUSY;
661 goto unlock;
662 }
663
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800664 /* charge against new page */
Miklos Szeredief6a3c62011-03-22 16:30:52 -0700665 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800666 if (charge == -ENOMEM) {
667 rc = -ENOMEM;
668 goto unlock;
669 }
670 BUG_ON(charge);
671
Christoph Lametere24f0b82006-06-23 02:03:51 -0700672 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700673 /*
674 * For !sync, there is no point retrying as the retry loop
675 * is expected to be too short for PageWriteback to be cleared
676 */
677 if (!sync) {
678 rc = -EBUSY;
679 goto uncharge;
680 }
681 if (!force)
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800682 goto uncharge;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700683 wait_on_page_writeback(page);
684 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700685 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700686 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
687 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800688 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700689 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700690 * File Caches may use write_page() or lock_page() in migration, then,
691 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700692 */
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700693 if (PageAnon(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800694 /*
695 * Only page_lock_anon_vma() understands the subtleties of
696 * getting a hold on an anon_vma from outside one of its mms.
697 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700698 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800699 if (anon_vma) {
700 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700701 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800702 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800703 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700704 /*
705 * We cannot be sure that the anon_vma of an unmapped
706 * swapcache page is safe to use because we don't
707 * know in advance if the VMA that this page belonged
708 * to still exists. If the VMA and others sharing the
709 * data have been freed, then the anon_vma could
710 * already be invalid.
711 *
712 * To avoid this possibility, swapcache pages get
713 * migrated but are not remapped when migration
714 * completes
715 */
716 remap_swapcache = 0;
717 } else {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800718 goto uncharge;
Mel Gorman3fe20112010-05-24 14:32:20 -0700719 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700720 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800721
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700722 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800723 * Corner case handling:
724 * 1. When a new swap-cache page is read into, it is added to the LRU
725 * and treated as swapcache but it has no rmap yet.
726 * Calling try_to_unmap() against a page->mapping==NULL page will
727 * trigger a BUG. So handle it here.
728 * 2. An orphaned page (see truncate_complete_page) might have
729 * fs-private metadata. The page can be picked up due to memory
730 * offlining. Everywhere else except page reclaim, the page is
731 * invisible to the vm, so the page can not be migrated. So try to
732 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700733 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800734 if (!page->mapping) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800735 VM_BUG_ON(PageAnon(page));
736 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800737 try_to_free_buffers(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800738 goto uncharge;
Shaohua Li62e1c552008-02-04 22:29:33 -0800739 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700740 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800741 }
742
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700743 /* Establish migration ptes or remove ptes */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200744 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700745
Shaohua Liabfc3482009-09-21 17:01:19 -0700746skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700747 if (!page_mapped(page))
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700748 rc = move_to_new_page(newpage, page, remap_swapcache, sync);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700749
Mel Gorman3fe20112010-05-24 14:32:20 -0700750 if (rc && remap_swapcache)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700751 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700752
753 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700754 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700755 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700756
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800757uncharge:
758 if (!charge)
Daisuke Nishimura50de1dd2011-01-13 15:47:43 -0800759 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700760unlock:
761 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700762out:
763 return rc;
764}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700765
Minchan Kim0dabec92011-10-31 17:06:57 -0700766/*
767 * Obtain the lock on page, remove all ptes and migrate the page
768 * to the newly allocated page in newpage.
769 */
770static int unmap_and_move(new_page_t get_new_page, unsigned long private,
771 struct page *page, int force, bool offlining, bool sync)
772{
773 int rc = 0;
774 int *result = NULL;
775 struct page *newpage = get_new_page(page, private, &result);
776
777 if (!newpage)
778 return -ENOMEM;
779
780 if (page_count(page) == 1) {
781 /* page was freed from under us. So we are done. */
782 goto out;
783 }
784
785 if (unlikely(PageTransHuge(page)))
786 if (unlikely(split_huge_page(page)))
787 goto out;
788
789 rc = __unmap_and_move(page, newpage, force, offlining, sync);
790out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700791 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700792 /*
793 * A page that has been migrated has all references
794 * removed and will be freed. A page that has not been
795 * migrated will have kepts its references and be
796 * restored.
797 */
798 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700799 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700800 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700801 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700802 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700803 /*
804 * Move the new page to the LRU. If migration was not successful
805 * then this will free the page.
806 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700807 putback_lru_page(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700808 if (result) {
809 if (rc)
810 *result = rc;
811 else
812 *result = page_to_nid(newpage);
813 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700814 return rc;
815}
816
817/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900818 * Counterpart of unmap_and_move_page() for hugepage migration.
819 *
820 * This function doesn't wait the completion of hugepage I/O
821 * because there is no race between I/O and migration for hugepage.
822 * Note that currently hugepage I/O occurs only in direct I/O
823 * where no lock is held and PG_writeback is irrelevant,
824 * and writeback status of all subpages are counted in the reference
825 * count of the head page (i.e. if all subpages of a 2MB hugepage are
826 * under direct I/O, the reference of the head page is 512 and a bit more.)
827 * This means that when we try to migrate hugepage whose subpages are
828 * doing direct I/O, some references remain after try_to_unmap() and
829 * hugepage migration fails without data corruption.
830 *
831 * There is also no race when direct I/O is issued on the page under migration,
832 * because then pte is replaced with migration swap entry and direct I/O code
833 * will wait in the page fault for migration to complete.
834 */
835static int unmap_and_move_huge_page(new_page_t get_new_page,
836 unsigned long private, struct page *hpage,
Mel Gorman7f0f2492011-01-13 15:45:58 -0800837 int force, bool offlining, bool sync)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900838{
839 int rc = 0;
840 int *result = NULL;
841 struct page *new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900842 struct anon_vma *anon_vma = NULL;
843
844 if (!new_hpage)
845 return -ENOMEM;
846
847 rc = -EAGAIN;
848
849 if (!trylock_page(hpage)) {
Mel Gorman77f1fe62011-01-13 15:45:57 -0800850 if (!force || !sync)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900851 goto out;
852 lock_page(hpage);
853 }
854
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700855 if (PageAnon(hpage))
856 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900857
858 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
859
860 if (!page_mapped(hpage))
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700861 rc = move_to_new_page(new_hpage, hpage, 1, sync);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900862
863 if (rc)
864 remove_migration_ptes(hpage, hpage);
865
Hugh Dickinsfd4a4662011-01-13 15:47:31 -0800866 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700867 put_anon_vma(anon_vma);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900868 unlock_page(hpage);
869
Hillf Danton09761332011-12-08 14:34:20 -0800870out:
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900871 if (rc != -EAGAIN) {
872 list_del(&hpage->lru);
873 put_page(hpage);
874 }
875
876 put_page(new_hpage);
877
878 if (result) {
879 if (rc)
880 *result = rc;
881 else
882 *result = page_to_nid(new_hpage);
883 }
884 return rc;
885}
886
887/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800888 * migrate_pages
889 *
Christoph Lameter95a402c2006-06-23 02:03:53 -0700890 * The function takes one list of pages to migrate and a function
891 * that determines from the page to be migrated and the private data
892 * the target of the move and allocates the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800893 *
894 * The function returns after 10 attempts or if no pages
895 * are movable anymore because to has become empty
Minchan Kimcf608ac2010-10-26 14:21:29 -0700896 * or no retryable pages exist anymore.
897 * Caller should call putback_lru_pages to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -0800898 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800899 *
Christoph Lameter95a402c2006-06-23 02:03:53 -0700900 * Return: Number of pages not migrated or error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800901 */
Christoph Lameter95a402c2006-06-23 02:03:53 -0700902int migrate_pages(struct list_head *from,
Mel Gorman7f0f2492011-01-13 15:45:58 -0800903 new_page_t get_new_page, unsigned long private, bool offlining,
Mel Gorman77f1fe62011-01-13 15:45:57 -0800904 bool sync)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800905{
Christoph Lametere24f0b82006-06-23 02:03:51 -0700906 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800907 int nr_failed = 0;
908 int pass = 0;
909 struct page *page;
910 struct page *page2;
911 int swapwrite = current->flags & PF_SWAPWRITE;
912 int rc;
913
914 if (!swapwrite)
915 current->flags |= PF_SWAPWRITE;
916
Christoph Lametere24f0b82006-06-23 02:03:51 -0700917 for(pass = 0; pass < 10 && retry; pass++) {
918 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800919
Christoph Lametere24f0b82006-06-23 02:03:51 -0700920 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700921 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -0800922
Christoph Lameter95a402c2006-06-23 02:03:53 -0700923 rc = unmap_and_move(get_new_page, private,
Mel Gorman77f1fe62011-01-13 15:45:57 -0800924 page, pass > 2, offlining,
925 sync);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800926
Christoph Lametere24f0b82006-06-23 02:03:51 -0700927 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -0700928 case -ENOMEM:
929 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700930 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700931 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700932 break;
933 case 0:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700934 break;
935 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700936 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700937 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700938 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700939 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800940 }
941 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700942 rc = 0;
943out:
Christoph Lameterb20a3502006-03-22 00:09:12 -0800944 if (!swapwrite)
945 current->flags &= ~PF_SWAPWRITE;
946
Christoph Lameter95a402c2006-06-23 02:03:53 -0700947 if (rc)
948 return rc;
949
Christoph Lameterb20a3502006-03-22 00:09:12 -0800950 return nr_failed + retry;
951}
952
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900953int migrate_huge_pages(struct list_head *from,
Mel Gorman7f0f2492011-01-13 15:45:58 -0800954 new_page_t get_new_page, unsigned long private, bool offlining,
Mel Gorman77f1fe62011-01-13 15:45:57 -0800955 bool sync)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900956{
957 int retry = 1;
958 int nr_failed = 0;
959 int pass = 0;
960 struct page *page;
961 struct page *page2;
962 int rc;
963
964 for (pass = 0; pass < 10 && retry; pass++) {
965 retry = 0;
966
967 list_for_each_entry_safe(page, page2, from, lru) {
968 cond_resched();
969
970 rc = unmap_and_move_huge_page(get_new_page,
Mel Gorman77f1fe62011-01-13 15:45:57 -0800971 private, page, pass > 2, offlining,
972 sync);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900973
974 switch(rc) {
975 case -ENOMEM:
976 goto out;
977 case -EAGAIN:
978 retry++;
979 break;
980 case 0:
981 break;
982 default:
983 /* Permanent failure */
984 nr_failed++;
985 break;
986 }
987 }
988 }
989 rc = 0;
990out:
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900991 if (rc)
992 return rc;
993
994 return nr_failed + retry;
995}
996
Christoph Lameter742755a2006-06-23 02:03:55 -0700997#ifdef CONFIG_NUMA
998/*
999 * Move a list of individual pages
1000 */
1001struct page_to_node {
1002 unsigned long addr;
1003 struct page *page;
1004 int node;
1005 int status;
1006};
1007
1008static struct page *new_page_node(struct page *p, unsigned long private,
1009 int **result)
1010{
1011 struct page_to_node *pm = (struct page_to_node *)private;
1012
1013 while (pm->node != MAX_NUMNODES && pm->page != p)
1014 pm++;
1015
1016 if (pm->node == MAX_NUMNODES)
1017 return NULL;
1018
1019 *result = &pm->status;
1020
Mel Gorman6484eb32009-06-16 15:31:54 -07001021 return alloc_pages_exact_node(pm->node,
Mel Gorman769848c2007-07-17 04:03:05 -07001022 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001023}
1024
1025/*
1026 * Move a set of pages as indicated in the pm array. The addr
1027 * field must be set to the virtual address of the page to be moved
1028 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001029 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001030 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001031static int do_move_page_to_node_array(struct mm_struct *mm,
1032 struct page_to_node *pm,
1033 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001034{
1035 int err;
1036 struct page_to_node *pp;
1037 LIST_HEAD(pagelist);
1038
1039 down_read(&mm->mmap_sem);
1040
1041 /*
1042 * Build a list of pages to migrate
1043 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001044 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1045 struct vm_area_struct *vma;
1046 struct page *page;
1047
Christoph Lameter742755a2006-06-23 02:03:55 -07001048 err = -EFAULT;
1049 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001050 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001051 goto set_status;
1052
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001053 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001054
1055 err = PTR_ERR(page);
1056 if (IS_ERR(page))
1057 goto set_status;
1058
Christoph Lameter742755a2006-06-23 02:03:55 -07001059 err = -ENOENT;
1060 if (!page)
1061 goto set_status;
1062
Hugh Dickins62b61f62009-12-14 17:59:33 -08001063 /* Use PageReserved to check for zero page */
1064 if (PageReserved(page) || PageKsm(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001065 goto put_and_set;
1066
1067 pp->page = page;
1068 err = page_to_nid(page);
1069
1070 if (err == pp->node)
1071 /*
1072 * Node already in the right place
1073 */
1074 goto put_and_set;
1075
1076 err = -EACCES;
1077 if (page_mapcount(page) > 1 &&
1078 !migrate_all)
1079 goto put_and_set;
1080
Nick Piggin62695a82008-10-18 20:26:09 -07001081 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001082 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001083 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001084 inc_zone_page_state(page, NR_ISOLATED_ANON +
1085 page_is_file_cache(page));
1086 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001087put_and_set:
1088 /*
1089 * Either remove the duplicate refcount from
1090 * isolate_lru_page() or drop the page ref if it was
1091 * not isolated.
1092 */
1093 put_page(page);
1094set_status:
1095 pp->status = err;
1096 }
1097
Brice Gogline78bbfa2008-10-18 20:27:15 -07001098 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001099 if (!list_empty(&pagelist)) {
Christoph Lameter742755a2006-06-23 02:03:55 -07001100 err = migrate_pages(&pagelist, new_page_node,
Mel Gorman77f1fe62011-01-13 15:45:57 -08001101 (unsigned long)pm, 0, true);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001102 if (err)
1103 putback_lru_pages(&pagelist);
1104 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001105
1106 up_read(&mm->mmap_sem);
1107 return err;
1108}
1109
1110/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001111 * Migrate an array of page address onto an array of nodes and fill
1112 * the corresponding array of status.
1113 */
1114static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
1115 unsigned long nr_pages,
1116 const void __user * __user *pages,
1117 const int __user *nodes,
1118 int __user *status, int flags)
1119{
Brice Goglin3140a222009-01-06 14:38:57 -08001120 struct page_to_node *pm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001121 nodemask_t task_nodes;
Brice Goglin3140a222009-01-06 14:38:57 -08001122 unsigned long chunk_nr_pages;
1123 unsigned long chunk_start;
1124 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001125
1126 task_nodes = cpuset_mems_allowed(task);
1127
Brice Goglin3140a222009-01-06 14:38:57 -08001128 err = -ENOMEM;
1129 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1130 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001131 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001132
1133 migrate_prep();
1134
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001135 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001136 * Store a chunk of page_to_node array in a page,
1137 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001138 */
Brice Goglin3140a222009-01-06 14:38:57 -08001139 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001140
Brice Goglin3140a222009-01-06 14:38:57 -08001141 for (chunk_start = 0;
1142 chunk_start < nr_pages;
1143 chunk_start += chunk_nr_pages) {
1144 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001145
Brice Goglin3140a222009-01-06 14:38:57 -08001146 if (chunk_start + chunk_nr_pages > nr_pages)
1147 chunk_nr_pages = nr_pages - chunk_start;
1148
1149 /* fill the chunk pm with addrs and nodes from user-space */
1150 for (j = 0; j < chunk_nr_pages; j++) {
1151 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001152 int node;
1153
Brice Goglin3140a222009-01-06 14:38:57 -08001154 err = -EFAULT;
1155 if (get_user(p, pages + j + chunk_start))
1156 goto out_pm;
1157 pm[j].addr = (unsigned long) p;
1158
1159 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001160 goto out_pm;
1161
1162 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001163 if (node < 0 || node >= MAX_NUMNODES)
1164 goto out_pm;
1165
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001166 if (!node_state(node, N_HIGH_MEMORY))
1167 goto out_pm;
1168
1169 err = -EACCES;
1170 if (!node_isset(node, task_nodes))
1171 goto out_pm;
1172
Brice Goglin3140a222009-01-06 14:38:57 -08001173 pm[j].node = node;
1174 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001175
Brice Goglin3140a222009-01-06 14:38:57 -08001176 /* End marker for this chunk */
1177 pm[chunk_nr_pages].node = MAX_NUMNODES;
1178
1179 /* Migrate this chunk */
1180 err = do_move_page_to_node_array(mm, pm,
1181 flags & MPOL_MF_MOVE_ALL);
1182 if (err < 0)
1183 goto out_pm;
1184
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001185 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001186 for (j = 0; j < chunk_nr_pages; j++)
1187 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001188 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001189 goto out_pm;
1190 }
1191 }
1192 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001193
1194out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001195 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001196out:
1197 return err;
1198}
1199
1200/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001201 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001202 */
Brice Goglin80bba122008-12-09 13:14:23 -08001203static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1204 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001205{
Brice Goglin2f007e72008-10-18 20:27:16 -07001206 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001207
Christoph Lameter742755a2006-06-23 02:03:55 -07001208 down_read(&mm->mmap_sem);
1209
Brice Goglin2f007e72008-10-18 20:27:16 -07001210 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001211 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001212 struct vm_area_struct *vma;
1213 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001214 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001215
1216 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001217 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001218 goto set_status;
1219
Brice Goglin2f007e72008-10-18 20:27:16 -07001220 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001221
1222 err = PTR_ERR(page);
1223 if (IS_ERR(page))
1224 goto set_status;
1225
Christoph Lameter742755a2006-06-23 02:03:55 -07001226 err = -ENOENT;
1227 /* Use PageReserved to check for zero page */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001228 if (!page || PageReserved(page) || PageKsm(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001229 goto set_status;
1230
1231 err = page_to_nid(page);
1232set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001233 *status = err;
1234
1235 pages++;
1236 status++;
1237 }
1238
1239 up_read(&mm->mmap_sem);
1240}
1241
1242/*
1243 * Determine the nodes of a user array of pages and store it in
1244 * a user array of status.
1245 */
1246static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1247 const void __user * __user *pages,
1248 int __user *status)
1249{
1250#define DO_PAGES_STAT_CHUNK_NR 16
1251 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1252 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001253
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001254 while (nr_pages) {
1255 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001256
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001257 chunk_nr = nr_pages;
1258 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1259 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1260
1261 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1262 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001263
1264 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1265
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001266 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1267 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001268
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001269 pages += chunk_nr;
1270 status += chunk_nr;
1271 nr_pages -= chunk_nr;
1272 }
1273 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001274}
1275
1276/*
1277 * Move a list of pages in the address space of the currently executing
1278 * process.
1279 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001280SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1281 const void __user * __user *, pages,
1282 const int __user *, nodes,
1283 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001284{
David Howellsc69e8d92008-11-14 10:39:19 +11001285 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001286 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001287 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001288 int err;
Christoph Lameter742755a2006-06-23 02:03:55 -07001289
1290 /* Check flags */
1291 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1292 return -EINVAL;
1293
1294 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1295 return -EPERM;
1296
1297 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001298 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001299 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001300 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001301 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001302 return -ESRCH;
1303 }
1304 mm = get_task_mm(task);
Greg Thelena879bf52011-02-25 14:44:13 -08001305 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001306
1307 if (!mm)
1308 return -EINVAL;
1309
1310 /*
1311 * Check if this process has the right to modify the specified
1312 * process. The right exists if the process has administrative
1313 * capabilities, superuser privileges or the same
1314 * userid as the target process.
1315 */
David Howellsc69e8d92008-11-14 10:39:19 +11001316 rcu_read_lock();
1317 tcred = __task_cred(task);
David Howellsb6dff3e2008-11-14 10:39:16 +11001318 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1319 cred->uid != tcred->suid && cred->uid != tcred->uid &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001320 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001321 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001322 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001323 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001324 }
David Howellsc69e8d92008-11-14 10:39:19 +11001325 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001326
David Quigley86c3a762006-06-23 02:04:02 -07001327 err = security_task_movememory(task);
1328 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001329 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001330
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001331 if (nodes) {
1332 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1333 flags);
1334 } else {
Brice Goglin2f007e72008-10-18 20:27:16 -07001335 err = do_pages_stat(mm, nr_pages, pages, status);
Brice Goglin2f007e72008-10-18 20:27:16 -07001336 }
David Quigley86c3a762006-06-23 02:04:02 -07001337
Christoph Lameter742755a2006-06-23 02:03:55 -07001338out:
Christoph Lameter742755a2006-06-23 02:03:55 -07001339 mmput(mm);
1340 return err;
1341}
Christoph Lameter742755a2006-06-23 02:03:55 -07001342
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001343/*
1344 * Call migration functions in the vma_ops that may prepare
1345 * memory in a vm for migration. migration functions may perform
1346 * the migration for vmas that do not have an underlying page struct.
1347 */
1348int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1349 const nodemask_t *from, unsigned long flags)
1350{
1351 struct vm_area_struct *vma;
1352 int err = 0;
1353
Daisuke Nishimura1001c9f2009-02-11 13:04:18 -08001354 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001355 if (vma->vm_ops && vma->vm_ops->migrate) {
1356 err = vma->vm_ops->migrate(vma, to, from, flags);
1357 if (err)
1358 break;
1359 }
1360 }
1361 return err;
1362}
Gerald Schaefer83d16742008-07-23 21:28:22 -07001363#endif