blob: 7a7325ee1d089696a8073a84d2f748f326124805 [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>
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -070036#include <linux/hugetlb_cgroup.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090037#include <linux/gfp.h>
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080038#include <linux/balloon_compaction.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080039
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080040#include <asm/tlbflush.h>
41
Mel Gorman7b2a2d42012-10-19 14:07:31 +010042#define CREATE_TRACE_POINTS
43#include <trace/events/migrate.h>
44
Christoph Lameterb20a3502006-03-22 00:09:12 -080045#include "internal.h"
46
Christoph Lameterb20a3502006-03-22 00:09:12 -080047/*
Christoph Lameter742755a2006-06-23 02:03:55 -070048 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070049 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
50 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080051 */
52int migrate_prep(void)
53{
Christoph Lameterb20a3502006-03-22 00:09:12 -080054 /*
55 * Clear the LRU lists so pages can be isolated.
56 * Note that pages may be moved off the LRU after we have
57 * drained them. Those pages will fail to migrate like other
58 * pages that may be busy.
59 */
60 lru_add_drain_all();
61
62 return 0;
63}
64
Mel Gorman748446b2010-05-24 14:32:27 -070065/* Do the necessary work of migrate_prep but not if it involves other CPUs */
66int migrate_prep_local(void)
67{
68 lru_add_drain();
69
70 return 0;
71}
72
Christoph Lameterb20a3502006-03-22 00:09:12 -080073/*
Lee Schermerhorn894bc312008-10-18 20:26:39 -070074 * Add isolated pages on the list back to the LRU under page lock
75 * to avoid leaking evictable pages back onto unevictable list.
Christoph Lameterb20a3502006-03-22 00:09:12 -080076 */
Minchan Kime13861d2010-05-24 14:31:59 -070077void putback_lru_pages(struct list_head *l)
Christoph Lameterb20a3502006-03-22 00:09:12 -080078{
79 struct page *page;
80 struct page *page2;
Christoph Lameterb20a3502006-03-22 00:09:12 -080081
82 list_for_each_entry_safe(page, page2, l, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -070083 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -070084 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -070085 page_is_file_cache(page));
Rafael Aquini5733c7d2012-12-11 16:02:47 -080086 putback_lru_page(page);
87 }
88}
89
90/*
91 * Put previously isolated pages back onto the appropriate lists
92 * from where they were once taken off for compaction/migration.
93 *
94 * This function shall be used instead of putback_lru_pages(),
95 * whenever the isolated pageset has been built by isolate_migratepages_range()
96 */
97void putback_movable_pages(struct list_head *l)
98{
99 struct page *page;
100 struct page *page2;
101
102 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -0700103 if (unlikely(PageHuge(page))) {
104 putback_active_hugepage(page);
105 continue;
106 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -0800107 list_del(&page->lru);
108 dec_zone_page_state(page, NR_ISOLATED_ANON +
109 page_is_file_cache(page));
Rafael Aquini117aad12013-09-30 13:45:16 -0700110 if (unlikely(isolated_balloon_page(page)))
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800111 balloon_page_putback(page);
112 else
113 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800114 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800115}
116
Christoph Lameter06972122006-06-23 02:03:35 -0700117/*
118 * Restore a potential migration pte to a working pte entry
119 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800120static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
121 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700122{
123 struct mm_struct *mm = vma->vm_mm;
124 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700125 pmd_t *pmd;
126 pte_t *ptep, pte;
127 spinlock_t *ptl;
128
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900129 if (unlikely(PageHuge(new))) {
130 ptep = huge_pte_offset(mm, addr);
131 if (!ptep)
132 goto out;
133 ptl = &mm->page_table_lock;
134 } else {
Bob Liu62190492012-12-11 16:00:37 -0800135 pmd = mm_find_pmd(mm, addr);
136 if (!pmd)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900137 goto out;
Andrea Arcangeli500d65d2011-01-13 15:46:55 -0800138 if (pmd_trans_huge(*pmd))
139 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700140
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900141 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700142
Hugh Dickins486cf462011-10-19 12:50:35 -0700143 /*
144 * Peek to check is_swap_pte() before taking ptlock? No, we
145 * can race mremap's move_ptes(), which skips anon_vma lock.
146 */
Christoph Lameter06972122006-06-23 02:03:35 -0700147
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900148 ptl = pte_lockptr(mm, pmd);
149 }
150
Christoph Lameter06972122006-06-23 02:03:35 -0700151 spin_lock(ptl);
152 pte = *ptep;
153 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800154 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700155
156 entry = pte_to_swp_entry(pte);
157
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800158 if (!is_migration_entry(entry) ||
159 migration_entry_to_page(entry) != old)
160 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700161
Christoph Lameter06972122006-06-23 02:03:35 -0700162 get_page(new);
163 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700164 if (pte_swp_soft_dirty(*ptep))
165 pte = pte_mksoft_dirty(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700166 if (is_write_migration_entry(entry))
167 pte = pte_mkwrite(pte);
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200168#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800169 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900170 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800171 pte = arch_make_huge_pte(pte, vma, new, 0);
172 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200173#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700174 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700175 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700176
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900177 if (PageHuge(new)) {
178 if (PageAnon(new))
179 hugepage_add_anon_rmap(new, vma, addr);
180 else
181 page_dup_rmap(new);
182 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700183 page_add_anon_rmap(new, vma, addr);
184 else
185 page_add_file_rmap(new);
186
187 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000188 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800189unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700190 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800191out:
192 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700193}
194
195/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700196 * Get rid of all migration entries and replace them by
197 * references to the indicated page.
198 */
199static void remove_migration_ptes(struct page *old, struct page *new)
200{
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800201 rmap_walk(new, remove_migration_pte, old);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700202}
203
204/*
Christoph Lameter06972122006-06-23 02:03:35 -0700205 * Something used the pte of a page under migration. We need to
206 * get to the page and wait until migration is finished.
207 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700208 */
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700209static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
210 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700211{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700212 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700213 swp_entry_t entry;
214 struct page *page;
215
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700216 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700217 pte = *ptep;
218 if (!is_swap_pte(pte))
219 goto out;
220
221 entry = pte_to_swp_entry(pte);
222 if (!is_migration_entry(entry))
223 goto out;
224
225 page = migration_entry_to_page(entry);
226
Nick Piggine2867812008-07-25 19:45:30 -0700227 /*
228 * Once radix-tree replacement of page migration started, page_count
229 * *must* be zero. And, we don't want to call wait_on_page_locked()
230 * against a page without get_page().
231 * So, we use get_page_unless_zero(), here. Even failed, page fault
232 * will occur again.
233 */
234 if (!get_page_unless_zero(page))
235 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700236 pte_unmap_unlock(ptep, ptl);
237 wait_on_page_locked(page);
238 put_page(page);
239 return;
240out:
241 pte_unmap_unlock(ptep, ptl);
242}
243
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700244void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
245 unsigned long address)
246{
247 spinlock_t *ptl = pte_lockptr(mm, pmd);
248 pte_t *ptep = pte_offset_map(pmd, address);
249 __migration_entry_wait(mm, ptep, ptl);
250}
251
252void migration_entry_wait_huge(struct mm_struct *mm, pte_t *pte)
253{
254 spinlock_t *ptl = &(mm)->page_table_lock;
255 __migration_entry_wait(mm, pte, ptl);
256}
257
Mel Gormanb969c4a2012-01-12 17:19:34 -0800258#ifdef CONFIG_BLOCK
259/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800260static bool buffer_migrate_lock_buffers(struct buffer_head *head,
261 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800262{
263 struct buffer_head *bh = head;
264
265 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800266 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800267 do {
268 get_bh(bh);
269 lock_buffer(bh);
270 bh = bh->b_this_page;
271
272 } while (bh != head);
273
274 return true;
275 }
276
277 /* async case, we cannot block on lock_buffer so use trylock_buffer */
278 do {
279 get_bh(bh);
280 if (!trylock_buffer(bh)) {
281 /*
282 * We failed to lock the buffer and cannot stall in
283 * async migration. Release the taken locks
284 */
285 struct buffer_head *failed_bh = bh;
286 put_bh(failed_bh);
287 bh = head;
288 while (bh != failed_bh) {
289 unlock_buffer(bh);
290 put_bh(bh);
291 bh = bh->b_this_page;
292 }
293 return false;
294 }
295
296 bh = bh->b_this_page;
297 } while (bh != head);
298 return true;
299}
300#else
301static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800302 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800303{
304 return true;
305}
306#endif /* CONFIG_BLOCK */
307
Christoph Lameterb20a3502006-03-22 00:09:12 -0800308/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700309 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700310 *
311 * The number of remaining references must be:
312 * 1 for anonymous pages without a mapping
313 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100314 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800315 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800316int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800317 struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800318 struct buffer_head *head, enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800319{
Peter Zijlstra7039e1d2012-10-25 14:16:34 +0200320 int expected_count = 0;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800321 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800322
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700323 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700324 /* Anonymous page without mapping */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700325 if (page_count(page) != 1)
326 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800327 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700328 }
329
Nick Piggin19fd6232008-07-25 19:45:32 -0700330 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800331
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800332 pslot = radix_tree_lookup_slot(&mapping->page_tree,
333 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800334
Johannes Weineredcf4742009-09-21 17:02:59 -0700335 expected_count = 2 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700336 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800337 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700338 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700339 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800340 }
341
Nick Piggine2867812008-07-25 19:45:30 -0700342 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700343 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700344 return -EAGAIN;
345 }
346
Christoph Lameterb20a3502006-03-22 00:09:12 -0800347 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800348 * In the async migration case of moving a page with buffers, lock the
349 * buffers using trylock before the mapping is moved. If the mapping
350 * was moved, we later failed to lock the buffers and could not move
351 * the mapping back due to an elevated page count, we would have to
352 * block waiting on other references to be dropped.
353 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800354 if (mode == MIGRATE_ASYNC && head &&
355 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800356 page_unfreeze_refs(page, expected_count);
357 spin_unlock_irq(&mapping->tree_lock);
358 return -EAGAIN;
359 }
360
361 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800362 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800363 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800364 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800365 if (PageSwapCache(page)) {
366 SetPageSwapCache(newpage);
367 set_page_private(newpage, page_private(page));
368 }
369
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800370 radix_tree_replace_slot(pslot, newpage);
371
372 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800373 * Drop cache reference from old page by unfreezing
374 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800375 * We know this isn't the last reference.
376 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800377 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800378
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700379 /*
380 * If moved to a different zone then also account
381 * the page for that zone. Other VM counters will be
382 * taken care of when we establish references to the
383 * new page and drop references to the old page.
384 *
385 * Note that anonymous pages are accounted for
386 * via NR_FILE_PAGES and NR_ANON_PAGES if they
387 * are mapped to swap space.
388 */
389 __dec_zone_page_state(page, NR_FILE_PAGES);
390 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700391 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700392 __dec_zone_page_state(page, NR_SHMEM);
393 __inc_zone_page_state(newpage, NR_SHMEM);
394 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700395 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800396
Rafael Aquini78bd5202012-12-11 16:02:31 -0800397 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800398}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800399
400/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900401 * The expected number of remaining references is the same as that
402 * of migrate_page_move_mapping().
403 */
404int migrate_huge_page_move_mapping(struct address_space *mapping,
405 struct page *newpage, struct page *page)
406{
407 int expected_count;
408 void **pslot;
409
410 if (!mapping) {
411 if (page_count(page) != 1)
412 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800413 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900414 }
415
416 spin_lock_irq(&mapping->tree_lock);
417
418 pslot = radix_tree_lookup_slot(&mapping->page_tree,
419 page_index(page));
420
421 expected_count = 2 + page_has_private(page);
422 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800423 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900424 spin_unlock_irq(&mapping->tree_lock);
425 return -EAGAIN;
426 }
427
428 if (!page_freeze_refs(page, expected_count)) {
429 spin_unlock_irq(&mapping->tree_lock);
430 return -EAGAIN;
431 }
432
433 get_page(newpage);
434
435 radix_tree_replace_slot(pslot, newpage);
436
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800437 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900438
439 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800440 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900441}
442
443/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800444 * Copy the page to its new location
445 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900446void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800447{
Mel Gormanb32967f2012-11-19 12:35:47 +0000448 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900449 copy_huge_page(newpage, page);
450 else
451 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800452
453 if (PageError(page))
454 SetPageError(newpage);
455 if (PageReferenced(page))
456 SetPageReferenced(newpage);
457 if (PageUptodate(page))
458 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700459 if (TestClearPageActive(page)) {
460 VM_BUG_ON(PageUnevictable(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800461 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800462 } else if (TestClearPageUnevictable(page))
463 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800464 if (PageChecked(page))
465 SetPageChecked(newpage);
466 if (PageMappedToDisk(page))
467 SetPageMappedToDisk(newpage);
468
469 if (PageDirty(page)) {
470 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700471 /*
472 * Want to mark the page and the radix tree as dirty, and
473 * redo the accounting that clear_page_dirty_for_io undid,
474 * but we can't use set_page_dirty because that function
475 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300476 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700477 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700478 if (PageSwapBacked(page))
479 SetPageDirty(newpage);
480 else
481 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800482 }
483
Nick Pigginb291f002008-10-18 20:26:44 -0700484 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800485 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800486 /*
487 * Please do not reorder this without considering how mm/ksm.c's
488 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
489 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800490 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800491 ClearPagePrivate(page);
492 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800493
494 /*
495 * If any waiters have accumulated on the new page then
496 * wake them up.
497 */
498 if (PageWriteback(newpage))
499 end_page_writeback(newpage);
500}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800501
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700502/************************************************************
503 * Migration functions
504 ***********************************************************/
505
506/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700507int fail_migrate_page(struct address_space *mapping,
508 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700509{
510 return -EIO;
511}
512EXPORT_SYMBOL(fail_migrate_page);
513
Christoph Lameterb20a3502006-03-22 00:09:12 -0800514/*
515 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100516 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800517 *
518 * Pages are locked upon entry and exit.
519 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700520int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800521 struct page *newpage, struct page *page,
522 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800523{
524 int rc;
525
526 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
527
Mel Gormana6bc32b2012-01-12 17:19:43 -0800528 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800529
Rafael Aquini78bd5202012-12-11 16:02:31 -0800530 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800531 return rc;
532
533 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800534 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800535}
536EXPORT_SYMBOL(migrate_page);
537
David Howells93614012006-09-30 20:45:40 +0200538#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800539/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700540 * Migration function for pages with buffers. This function can only be used
541 * if the underlying filesystem guarantees that no other references to "page"
542 * exist.
543 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700544int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800545 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700546{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700547 struct buffer_head *bh, *head;
548 int rc;
549
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700550 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800551 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700552
553 head = page_buffers(page);
554
Mel Gormana6bc32b2012-01-12 17:19:43 -0800555 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700556
Rafael Aquini78bd5202012-12-11 16:02:31 -0800557 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700558 return rc;
559
Mel Gormanb969c4a2012-01-12 17:19:34 -0800560 /*
561 * In the async case, migrate_page_move_mapping locked the buffers
562 * with an IRQ-safe spinlock held. In the sync case, the buffers
563 * need to be locked now
564 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800565 if (mode != MIGRATE_ASYNC)
566 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700567
568 ClearPagePrivate(page);
569 set_page_private(newpage, page_private(page));
570 set_page_private(page, 0);
571 put_page(page);
572 get_page(newpage);
573
574 bh = head;
575 do {
576 set_bh_page(bh, newpage, bh_offset(bh));
577 bh = bh->b_this_page;
578
579 } while (bh != head);
580
581 SetPagePrivate(newpage);
582
583 migrate_page_copy(newpage, page);
584
585 bh = head;
586 do {
587 unlock_buffer(bh);
588 put_bh(bh);
589 bh = bh->b_this_page;
590
591 } while (bh != head);
592
Rafael Aquini78bd5202012-12-11 16:02:31 -0800593 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700594}
595EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200596#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700597
Christoph Lameter04e62a22006-06-23 02:03:38 -0700598/*
599 * Writeback a page to clean the dirty state
600 */
601static int writeout(struct address_space *mapping, struct page *page)
602{
603 struct writeback_control wbc = {
604 .sync_mode = WB_SYNC_NONE,
605 .nr_to_write = 1,
606 .range_start = 0,
607 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700608 .for_reclaim = 1
609 };
610 int rc;
611
612 if (!mapping->a_ops->writepage)
613 /* No write method for the address space */
614 return -EINVAL;
615
616 if (!clear_page_dirty_for_io(page))
617 /* Someone else already triggered a write */
618 return -EAGAIN;
619
620 /*
621 * A dirty page may imply that the underlying filesystem has
622 * the page on some queue. So the page must be clean for
623 * migration. Writeout may mean we loose the lock and the
624 * page state is no longer what we checked for earlier.
625 * At this point we know that the migration attempt cannot
626 * be successful.
627 */
628 remove_migration_ptes(page, page);
629
630 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700631
632 if (rc != AOP_WRITEPAGE_ACTIVATE)
633 /* unlocked. Relock */
634 lock_page(page);
635
Hugh Dickinsbda85502008-11-19 15:36:36 -0800636 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700637}
638
639/*
640 * Default handling if a filesystem does not provide a migration function.
641 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700642static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800643 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700644{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800645 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800646 /* Only writeback pages in full synchronous migration */
647 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800648 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700649 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800650 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700651
652 /*
653 * Buffers may be managed in a filesystem specific way.
654 * We must have no buffers or drop them.
655 */
David Howells266cf652009-04-03 16:42:36 +0100656 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700657 !try_to_release_page(page, GFP_KERNEL))
658 return -EAGAIN;
659
Mel Gormana6bc32b2012-01-12 17:19:43 -0800660 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700661}
662
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700663/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700664 * Move a page to a newly allocated page
665 * The page is locked and all ptes have been successfully removed.
666 *
667 * The new page will have replaced the old page if this function
668 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700669 *
670 * Return value:
671 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800672 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700673 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700674static int move_to_new_page(struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800675 int remap_swapcache, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700676{
677 struct address_space *mapping;
678 int rc;
679
680 /*
681 * Block others from accessing the page when we get around to
682 * establishing additional references. We are the only one
683 * holding a reference to the new page at this point.
684 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200685 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700686 BUG();
687
688 /* Prepare mapping for the new page.*/
689 newpage->index = page->index;
690 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700691 if (PageSwapBacked(page))
692 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700693
694 mapping = page_mapping(page);
695 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800696 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800697 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700698 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800699 * Most pages have a mapping and most filesystems provide a
700 * migratepage callback. Anonymous pages are part of swap
701 * space which also has its own migratepage callback. This
702 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700703 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800704 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800705 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800706 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800707 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700708
Rafael Aquini78bd5202012-12-11 16:02:31 -0800709 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700710 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700711 } else {
712 if (remap_swapcache)
713 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800714 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700715 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700716
717 unlock_page(newpage);
718
719 return rc;
720}
721
Minchan Kim0dabec92011-10-31 17:06:57 -0700722static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800723 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700724{
Minchan Kim0dabec92011-10-31 17:06:57 -0700725 int rc = -EAGAIN;
Mel Gorman3fe20112010-05-24 14:32:20 -0700726 int remap_swapcache = 1;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -0700727 struct mem_cgroup *mem;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700728 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700729
Nick Piggin529ae9a2008-08-02 12:01:03 +0200730 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800731 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700732 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800733
734 /*
735 * It's not safe for direct compaction to call lock_page.
736 * For example, during page readahead pages are added locked
737 * to the LRU. Later, when the IO completes the pages are
738 * marked uptodate and unlocked. However, the queueing
739 * could be merging multiple pages for one bio (e.g.
740 * mpage_readpages). If an allocation happens for the
741 * second or third page, the process can end up locking
742 * the same page twice and deadlocking. Rather than
743 * trying to be clever about what pages can be locked,
744 * avoid the use of lock_page for direct compaction
745 * altogether.
746 */
747 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700748 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800749
Christoph Lametere24f0b82006-06-23 02:03:51 -0700750 lock_page(page);
751 }
752
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800753 /* charge against new page */
Johannes Weiner0030f532012-07-31 16:45:25 -0700754 mem_cgroup_prepare_migration(page, newpage, &mem);
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800755
Christoph Lametere24f0b82006-06-23 02:03:51 -0700756 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700757 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700758 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800759 * necessary to wait for PageWriteback. In the async case,
760 * the retry loop is too short and in the sync-light case,
761 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700762 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800763 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700764 rc = -EBUSY;
765 goto uncharge;
766 }
767 if (!force)
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800768 goto uncharge;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700769 wait_on_page_writeback(page);
770 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700771 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700772 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
773 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800774 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700775 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700776 * File Caches may use write_page() or lock_page() in migration, then,
777 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700778 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800779 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800780 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000781 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800782 * getting a hold on an anon_vma from outside one of its mms.
783 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700784 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800785 if (anon_vma) {
786 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700787 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800788 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800789 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700790 /*
791 * We cannot be sure that the anon_vma of an unmapped
792 * swapcache page is safe to use because we don't
793 * know in advance if the VMA that this page belonged
794 * to still exists. If the VMA and others sharing the
795 * data have been freed, then the anon_vma could
796 * already be invalid.
797 *
798 * To avoid this possibility, swapcache pages get
799 * migrated but are not remapped when migration
800 * completes
801 */
802 remap_swapcache = 0;
803 } else {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800804 goto uncharge;
Mel Gorman3fe20112010-05-24 14:32:20 -0700805 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700806 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800807
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800808 if (unlikely(balloon_page_movable(page))) {
809 /*
810 * A ballooned page does not need any special attention from
811 * physical to virtual reverse mapping procedures.
812 * Skip any attempt to unmap PTEs or to remap swap cache,
813 * in order to avoid burning cycles at rmap level, and perform
814 * the page migration right away (proteced by page lock).
815 */
816 rc = balloon_page_migrate(newpage, page, mode);
817 goto uncharge;
818 }
819
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700820 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800821 * Corner case handling:
822 * 1. When a new swap-cache page is read into, it is added to the LRU
823 * and treated as swapcache but it has no rmap yet.
824 * Calling try_to_unmap() against a page->mapping==NULL page will
825 * trigger a BUG. So handle it here.
826 * 2. An orphaned page (see truncate_complete_page) might have
827 * fs-private metadata. The page can be picked up due to memory
828 * offlining. Everywhere else except page reclaim, the page is
829 * invisible to the vm, so the page can not be migrated. So try to
830 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700831 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800832 if (!page->mapping) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800833 VM_BUG_ON(PageAnon(page));
834 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800835 try_to_free_buffers(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800836 goto uncharge;
Shaohua Li62e1c552008-02-04 22:29:33 -0800837 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700838 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800839 }
840
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700841 /* Establish migration ptes or remove ptes */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200842 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700843
Shaohua Liabfc3482009-09-21 17:01:19 -0700844skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700845 if (!page_mapped(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800846 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700847
Mel Gorman3fe20112010-05-24 14:32:20 -0700848 if (rc && remap_swapcache)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700849 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700850
851 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700852 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700853 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700854
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800855uncharge:
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800856 mem_cgroup_end_migration(mem, page, newpage,
857 (rc == MIGRATEPAGE_SUCCESS ||
858 rc == MIGRATEPAGE_BALLOON_SUCCESS));
Christoph Lametere24f0b82006-06-23 02:03:51 -0700859 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700860out:
861 return rc;
862}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700863
Minchan Kim0dabec92011-10-31 17:06:57 -0700864/*
865 * Obtain the lock on page, remove all ptes and migrate the page
866 * to the newly allocated page in newpage.
867 */
868static int unmap_and_move(new_page_t get_new_page, unsigned long private,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800869 struct page *page, int force, enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700870{
871 int rc = 0;
872 int *result = NULL;
873 struct page *newpage = get_new_page(page, private, &result);
874
875 if (!newpage)
876 return -ENOMEM;
877
878 if (page_count(page) == 1) {
879 /* page was freed from under us. So we are done. */
880 goto out;
881 }
882
883 if (unlikely(PageTransHuge(page)))
884 if (unlikely(split_huge_page(page)))
885 goto out;
886
Hugh Dickins9c620e22013-02-22 16:35:14 -0800887 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800888
889 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
890 /*
891 * A ballooned page has been migrated already.
892 * Now, it's the time to wrap-up counters,
893 * handle the page back to Buddy and return.
894 */
895 dec_zone_page_state(page, NR_ISOLATED_ANON +
896 page_is_file_cache(page));
897 balloon_page_free(page);
898 return MIGRATEPAGE_SUCCESS;
899 }
Minchan Kim0dabec92011-10-31 17:06:57 -0700900out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700901 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700902 /*
903 * A page that has been migrated has all references
904 * removed and will be freed. A page that has not been
905 * migrated will have kepts its references and be
906 * restored.
907 */
908 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700909 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700910 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700911 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700912 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700913 /*
914 * Move the new page to the LRU. If migration was not successful
915 * then this will free the page.
916 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700917 putback_lru_page(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700918 if (result) {
919 if (rc)
920 *result = rc;
921 else
922 *result = page_to_nid(newpage);
923 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700924 return rc;
925}
926
927/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900928 * Counterpart of unmap_and_move_page() for hugepage migration.
929 *
930 * This function doesn't wait the completion of hugepage I/O
931 * because there is no race between I/O and migration for hugepage.
932 * Note that currently hugepage I/O occurs only in direct I/O
933 * where no lock is held and PG_writeback is irrelevant,
934 * and writeback status of all subpages are counted in the reference
935 * count of the head page (i.e. if all subpages of a 2MB hugepage are
936 * under direct I/O, the reference of the head page is 512 and a bit more.)
937 * This means that when we try to migrate hugepage whose subpages are
938 * doing direct I/O, some references remain after try_to_unmap() and
939 * hugepage migration fails without data corruption.
940 *
941 * There is also no race when direct I/O is issued on the page under migration,
942 * because then pte is replaced with migration swap entry and direct I/O code
943 * will wait in the page fault for migration to complete.
944 */
945static int unmap_and_move_huge_page(new_page_t get_new_page,
946 unsigned long private, struct page *hpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800947 int force, enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900948{
949 int rc = 0;
950 int *result = NULL;
951 struct page *new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900952 struct anon_vma *anon_vma = NULL;
953
Naoya Horiguchi83467ef2013-09-11 14:22:11 -0700954 /*
955 * Movability of hugepages depends on architectures and hugepage size.
956 * This check is necessary because some callers of hugepage migration
957 * like soft offline and memory hotremove don't walk through page
958 * tables or check whether the hugepage is pmd-based or not before
959 * kicking migration.
960 */
961 if (!hugepage_migration_support(page_hstate(hpage)))
962 return -ENOSYS;
963
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900964 if (!new_hpage)
965 return -ENOMEM;
966
967 rc = -EAGAIN;
968
969 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800970 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900971 goto out;
972 lock_page(hpage);
973 }
974
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700975 if (PageAnon(hpage))
976 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900977
978 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
979
980 if (!page_mapped(hpage))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800981 rc = move_to_new_page(new_hpage, hpage, 1, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900982
983 if (rc)
984 remove_migration_ptes(hpage, hpage);
985
Hugh Dickinsfd4a4662011-01-13 15:47:31 -0800986 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700987 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -0700988
989 if (!rc)
990 hugetlb_cgroup_migrate(hpage, new_hpage);
991
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900992 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -0800993out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -0700994 if (rc != -EAGAIN)
995 putback_active_hugepage(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900996 put_page(new_hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900997 if (result) {
998 if (rc)
999 *result = rc;
1000 else
1001 *result = page_to_nid(new_hpage);
1002 }
1003 return rc;
1004}
1005
1006/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001007 * migrate_pages - migrate the pages specified in a list, to the free pages
1008 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001009 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001010 * @from: The list of pages to be migrated.
1011 * @get_new_page: The function used to allocate free pages to be used
1012 * as the target of the page migration.
1013 * @private: Private data to be passed on to get_new_page()
1014 * @mode: The migration mode that specifies the constraints for
1015 * page migration, if any.
1016 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001017 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001018 * The function returns after 10 attempts or if no pages are movable any more
1019 * because the list has become empty or no retryable pages exist any more.
1020 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001021 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001022 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001023 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001024 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001025int migrate_pages(struct list_head *from, new_page_t get_new_page,
1026 unsigned long private, enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001027{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001028 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001029 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001030 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001031 int pass = 0;
1032 struct page *page;
1033 struct page *page2;
1034 int swapwrite = current->flags & PF_SWAPWRITE;
1035 int rc;
1036
1037 if (!swapwrite)
1038 current->flags |= PF_SWAPWRITE;
1039
Christoph Lametere24f0b82006-06-23 02:03:51 -07001040 for(pass = 0; pass < 10 && retry; pass++) {
1041 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001042
Christoph Lametere24f0b82006-06-23 02:03:51 -07001043 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001044 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001045
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001046 if (PageHuge(page))
1047 rc = unmap_and_move_huge_page(get_new_page,
1048 private, page, pass > 2, mode);
1049 else
1050 rc = unmap_and_move(get_new_page, private,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001051 page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001052
Christoph Lametere24f0b82006-06-23 02:03:51 -07001053 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001054 case -ENOMEM:
1055 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001056 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001057 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001058 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001059 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001060 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001061 break;
1062 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001063 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001064 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001065 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001066 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001067 }
1068 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001069 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001070out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001071 if (nr_succeeded)
1072 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1073 if (nr_failed)
1074 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001075 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1076
Christoph Lameterb20a3502006-03-22 00:09:12 -08001077 if (!swapwrite)
1078 current->flags &= ~PF_SWAPWRITE;
1079
Rafael Aquini78bd5202012-12-11 16:02:31 -08001080 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001081}
1082
Christoph Lameter742755a2006-06-23 02:03:55 -07001083#ifdef CONFIG_NUMA
1084/*
1085 * Move a list of individual pages
1086 */
1087struct page_to_node {
1088 unsigned long addr;
1089 struct page *page;
1090 int node;
1091 int status;
1092};
1093
1094static struct page *new_page_node(struct page *p, unsigned long private,
1095 int **result)
1096{
1097 struct page_to_node *pm = (struct page_to_node *)private;
1098
1099 while (pm->node != MAX_NUMNODES && pm->page != p)
1100 pm++;
1101
1102 if (pm->node == MAX_NUMNODES)
1103 return NULL;
1104
1105 *result = &pm->status;
1106
Naoya Horiguchie632a932013-09-11 14:22:04 -07001107 if (PageHuge(p))
1108 return alloc_huge_page_node(page_hstate(compound_head(p)),
1109 pm->node);
1110 else
1111 return alloc_pages_exact_node(pm->node,
Mel Gorman769848c2007-07-17 04:03:05 -07001112 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001113}
1114
1115/*
1116 * Move a set of pages as indicated in the pm array. The addr
1117 * field must be set to the virtual address of the page to be moved
1118 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001119 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001120 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001121static int do_move_page_to_node_array(struct mm_struct *mm,
1122 struct page_to_node *pm,
1123 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001124{
1125 int err;
1126 struct page_to_node *pp;
1127 LIST_HEAD(pagelist);
1128
1129 down_read(&mm->mmap_sem);
1130
1131 /*
1132 * Build a list of pages to migrate
1133 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001134 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1135 struct vm_area_struct *vma;
1136 struct page *page;
1137
Christoph Lameter742755a2006-06-23 02:03:55 -07001138 err = -EFAULT;
1139 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001140 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001141 goto set_status;
1142
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001143 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001144
1145 err = PTR_ERR(page);
1146 if (IS_ERR(page))
1147 goto set_status;
1148
Christoph Lameter742755a2006-06-23 02:03:55 -07001149 err = -ENOENT;
1150 if (!page)
1151 goto set_status;
1152
Hugh Dickins62b61f62009-12-14 17:59:33 -08001153 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001154 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001155 goto put_and_set;
1156
1157 pp->page = page;
1158 err = page_to_nid(page);
1159
1160 if (err == pp->node)
1161 /*
1162 * Node already in the right place
1163 */
1164 goto put_and_set;
1165
1166 err = -EACCES;
1167 if (page_mapcount(page) > 1 &&
1168 !migrate_all)
1169 goto put_and_set;
1170
Naoya Horiguchie632a932013-09-11 14:22:04 -07001171 if (PageHuge(page)) {
1172 isolate_huge_page(page, &pagelist);
1173 goto put_and_set;
1174 }
1175
Nick Piggin62695a82008-10-18 20:26:09 -07001176 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001177 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001178 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001179 inc_zone_page_state(page, NR_ISOLATED_ANON +
1180 page_is_file_cache(page));
1181 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001182put_and_set:
1183 /*
1184 * Either remove the duplicate refcount from
1185 * isolate_lru_page() or drop the page ref if it was
1186 * not isolated.
1187 */
1188 put_page(page);
1189set_status:
1190 pp->status = err;
1191 }
1192
Brice Gogline78bbfa2008-10-18 20:27:15 -07001193 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001194 if (!list_empty(&pagelist)) {
Christoph Lameter742755a2006-06-23 02:03:55 -07001195 err = migrate_pages(&pagelist, new_page_node,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001196 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001197 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001198 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001199 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001200
1201 up_read(&mm->mmap_sem);
1202 return err;
1203}
1204
1205/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001206 * Migrate an array of page address onto an array of nodes and fill
1207 * the corresponding array of status.
1208 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001209static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001210 unsigned long nr_pages,
1211 const void __user * __user *pages,
1212 const int __user *nodes,
1213 int __user *status, int flags)
1214{
Brice Goglin3140a222009-01-06 14:38:57 -08001215 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001216 unsigned long chunk_nr_pages;
1217 unsigned long chunk_start;
1218 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001219
Brice Goglin3140a222009-01-06 14:38:57 -08001220 err = -ENOMEM;
1221 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1222 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001223 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001224
1225 migrate_prep();
1226
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001227 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001228 * Store a chunk of page_to_node array in a page,
1229 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001230 */
Brice Goglin3140a222009-01-06 14:38:57 -08001231 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001232
Brice Goglin3140a222009-01-06 14:38:57 -08001233 for (chunk_start = 0;
1234 chunk_start < nr_pages;
1235 chunk_start += chunk_nr_pages) {
1236 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001237
Brice Goglin3140a222009-01-06 14:38:57 -08001238 if (chunk_start + chunk_nr_pages > nr_pages)
1239 chunk_nr_pages = nr_pages - chunk_start;
1240
1241 /* fill the chunk pm with addrs and nodes from user-space */
1242 for (j = 0; j < chunk_nr_pages; j++) {
1243 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001244 int node;
1245
Brice Goglin3140a222009-01-06 14:38:57 -08001246 err = -EFAULT;
1247 if (get_user(p, pages + j + chunk_start))
1248 goto out_pm;
1249 pm[j].addr = (unsigned long) p;
1250
1251 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001252 goto out_pm;
1253
1254 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001255 if (node < 0 || node >= MAX_NUMNODES)
1256 goto out_pm;
1257
Lai Jiangshan389162c2012-12-12 13:51:30 -08001258 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001259 goto out_pm;
1260
1261 err = -EACCES;
1262 if (!node_isset(node, task_nodes))
1263 goto out_pm;
1264
Brice Goglin3140a222009-01-06 14:38:57 -08001265 pm[j].node = node;
1266 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001267
Brice Goglin3140a222009-01-06 14:38:57 -08001268 /* End marker for this chunk */
1269 pm[chunk_nr_pages].node = MAX_NUMNODES;
1270
1271 /* Migrate this chunk */
1272 err = do_move_page_to_node_array(mm, pm,
1273 flags & MPOL_MF_MOVE_ALL);
1274 if (err < 0)
1275 goto out_pm;
1276
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001277 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001278 for (j = 0; j < chunk_nr_pages; j++)
1279 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001280 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001281 goto out_pm;
1282 }
1283 }
1284 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001285
1286out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001287 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001288out:
1289 return err;
1290}
1291
1292/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001293 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001294 */
Brice Goglin80bba122008-12-09 13:14:23 -08001295static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1296 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001297{
Brice Goglin2f007e72008-10-18 20:27:16 -07001298 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001299
Christoph Lameter742755a2006-06-23 02:03:55 -07001300 down_read(&mm->mmap_sem);
1301
Brice Goglin2f007e72008-10-18 20:27:16 -07001302 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001303 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001304 struct vm_area_struct *vma;
1305 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001306 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001307
1308 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001309 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001310 goto set_status;
1311
Brice Goglin2f007e72008-10-18 20:27:16 -07001312 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001313
1314 err = PTR_ERR(page);
1315 if (IS_ERR(page))
1316 goto set_status;
1317
Christoph Lameter742755a2006-06-23 02:03:55 -07001318 err = -ENOENT;
1319 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001320 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001321 goto set_status;
1322
1323 err = page_to_nid(page);
1324set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001325 *status = err;
1326
1327 pages++;
1328 status++;
1329 }
1330
1331 up_read(&mm->mmap_sem);
1332}
1333
1334/*
1335 * Determine the nodes of a user array of pages and store it in
1336 * a user array of status.
1337 */
1338static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1339 const void __user * __user *pages,
1340 int __user *status)
1341{
1342#define DO_PAGES_STAT_CHUNK_NR 16
1343 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1344 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001345
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001346 while (nr_pages) {
1347 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001348
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001349 chunk_nr = nr_pages;
1350 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1351 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1352
1353 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1354 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001355
1356 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1357
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001358 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1359 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001360
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001361 pages += chunk_nr;
1362 status += chunk_nr;
1363 nr_pages -= chunk_nr;
1364 }
1365 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001366}
1367
1368/*
1369 * Move a list of pages in the address space of the currently executing
1370 * process.
1371 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001372SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1373 const void __user * __user *, pages,
1374 const int __user *, nodes,
1375 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001376{
David Howellsc69e8d92008-11-14 10:39:19 +11001377 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001378 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001379 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001380 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001381 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001382
1383 /* Check flags */
1384 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1385 return -EINVAL;
1386
1387 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1388 return -EPERM;
1389
1390 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001391 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001392 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001393 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001394 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001395 return -ESRCH;
1396 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001397 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001398
1399 /*
1400 * Check if this process has the right to modify the specified
1401 * process. The right exists if the process has administrative
1402 * capabilities, superuser privileges or the same
1403 * userid as the target process.
1404 */
David Howellsc69e8d92008-11-14 10:39:19 +11001405 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001406 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1407 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001408 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001409 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001410 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001411 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001412 }
David Howellsc69e8d92008-11-14 10:39:19 +11001413 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001414
David Quigley86c3a762006-06-23 02:04:02 -07001415 err = security_task_movememory(task);
1416 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001417 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001418
Christoph Lameter3268c632012-03-21 16:34:06 -07001419 task_nodes = cpuset_mems_allowed(task);
1420 mm = get_task_mm(task);
1421 put_task_struct(task);
1422
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001423 if (!mm)
1424 return -EINVAL;
1425
1426 if (nodes)
1427 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1428 nodes, status, flags);
1429 else
1430 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001431
1432 mmput(mm);
1433 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001434
Christoph Lameter742755a2006-06-23 02:03:55 -07001435out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001436 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001437 return err;
1438}
Christoph Lameter742755a2006-06-23 02:03:55 -07001439
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001440/*
1441 * Call migration functions in the vma_ops that may prepare
1442 * memory in a vm for migration. migration functions may perform
1443 * the migration for vmas that do not have an underlying page struct.
1444 */
1445int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1446 const nodemask_t *from, unsigned long flags)
1447{
1448 struct vm_area_struct *vma;
1449 int err = 0;
1450
Daisuke Nishimura1001c9f2009-02-11 13:04:18 -08001451 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001452 if (vma->vm_ops && vma->vm_ops->migrate) {
1453 err = vma->vm_ops->migrate(vma, to, from, flags);
1454 if (err)
1455 break;
1456 }
1457 }
1458 return err;
1459}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001460
1461#ifdef CONFIG_NUMA_BALANCING
1462/*
1463 * Returns true if this is a safe migration target node for misplaced NUMA
1464 * pages. Currently it only checks the watermarks which crude
1465 */
1466static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001467 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001468{
1469 int z;
1470 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1471 struct zone *zone = pgdat->node_zones + z;
1472
1473 if (!populated_zone(zone))
1474 continue;
1475
Lisa Du6e543d52013-09-11 14:22:36 -07001476 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001477 continue;
1478
1479 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1480 if (!zone_watermark_ok(zone, 0,
1481 high_wmark_pages(zone) +
1482 nr_migrate_pages,
1483 0, 0))
1484 continue;
1485 return true;
1486 }
1487 return false;
1488}
1489
1490static struct page *alloc_misplaced_dst_page(struct page *page,
1491 unsigned long data,
1492 int **result)
1493{
1494 int nid = (int) data;
1495 struct page *newpage;
1496
1497 newpage = alloc_pages_exact_node(nid,
1498 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1499 __GFP_NOMEMALLOC | __GFP_NORETRY |
1500 __GFP_NOWARN) &
1501 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001502 if (newpage)
Mel Gorman22b751c2013-02-22 16:34:59 -08001503 page_nid_xchg_last(newpage, page_nid_last(page));
Hillf Dantonbac03822012-11-27 14:46:24 +00001504
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001505 return newpage;
1506}
1507
1508/*
Mel Gormana8f60772012-11-14 21:41:46 +00001509 * page migration rate limiting control.
1510 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1511 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001512 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1513 * as it is faults that reset the window, pte updates will happen unconditionally
1514 * if there has not been a fault since @pteupdate_interval_millisecs after the
1515 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001516 */
1517static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001518static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001519static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1520
Mel Gormane14808b2012-11-19 10:59:15 +00001521/* Returns true if NUMA migration is currently rate limited */
1522bool migrate_ratelimited(int node)
1523{
1524 pg_data_t *pgdat = NODE_DATA(node);
1525
1526 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1527 msecs_to_jiffies(pteupdate_interval_millisecs)))
1528 return false;
1529
1530 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1531 return false;
1532
1533 return true;
1534}
1535
Mel Gormanb32967f2012-11-19 12:35:47 +00001536/* Returns true if the node is migrate rate-limited after the update */
Mel Gormand28d43352012-11-29 09:24:36 +00001537bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001538{
1539 bool rate_limited = false;
1540
1541 /*
1542 * Rate-limit the amount of data that is being migrated to a node.
1543 * Optimal placement is no good if the memory bus is saturated and
1544 * all the time is being spent migrating!
1545 */
1546 spin_lock(&pgdat->numabalancing_migrate_lock);
1547 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1548 pgdat->numabalancing_migrate_nr_pages = 0;
1549 pgdat->numabalancing_migrate_next_window = jiffies +
1550 msecs_to_jiffies(migrate_interval_millisecs);
1551 }
1552 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1553 rate_limited = true;
1554 else
Mel Gormand28d43352012-11-29 09:24:36 +00001555 pgdat->numabalancing_migrate_nr_pages += nr_pages;
Mel Gormanb32967f2012-11-19 12:35:47 +00001556 spin_unlock(&pgdat->numabalancing_migrate_lock);
1557
1558 return rate_limited;
1559}
1560
1561int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1562{
Hugh Dickins340ef392013-02-22 16:34:33 -08001563 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001564
Mel Gorman3abef4e2013-02-22 16:34:27 -08001565 VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1566
Mel Gormanb32967f2012-11-19 12:35:47 +00001567 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001568 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1569 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001570
Hugh Dickins340ef392013-02-22 16:34:33 -08001571 if (isolate_lru_page(page))
1572 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001573
1574 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001575 * migrate_misplaced_transhuge_page() skips page migration's usual
1576 * check on page_count(), so we must do it here, now that the page
1577 * has been isolated: a GUP pin, or any other pin, prevents migration.
1578 * The expected page count is 3: 1 for page's mapcount and 1 for the
1579 * caller's pin and 1 for the reference taken by isolate_lru_page().
1580 */
1581 if (PageTransHuge(page) && page_count(page) != 3) {
1582 putback_lru_page(page);
1583 return 0;
1584 }
1585
1586 page_lru = page_is_file_cache(page);
1587 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1588 hpage_nr_pages(page));
1589
1590 /*
1591 * Isolating the page has taken another reference, so the
1592 * caller's reference can be safely dropped without the page
1593 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001594 */
1595 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001596 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001597}
1598
Mel Gormana8f60772012-11-14 21:41:46 +00001599/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001600 * Attempt to migrate a misplaced page to the specified destination
1601 * node. Caller is expected to have an elevated reference count on
1602 * the page that will be dropped by this function before returning.
1603 */
1604int migrate_misplaced_page(struct page *page, int node)
1605{
Mel Gormana8f60772012-11-14 21:41:46 +00001606 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001607 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001608 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001609 LIST_HEAD(migratepages);
1610
1611 /*
1612 * Don't migrate pages that are mapped in multiple processes.
1613 * TODO: Handle false sharing detection instead of this hammer
1614 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001615 if (page_mapcount(page) != 1)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001616 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001617
Mel Gormana8f60772012-11-14 21:41:46 +00001618 /*
1619 * Rate-limit the amount of data that is being migrated to a node.
1620 * Optimal placement is no good if the memory bus is saturated and
1621 * all the time is being spent migrating!
1622 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001623 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001624 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001625
Mel Gormanb32967f2012-11-19 12:35:47 +00001626 isolated = numamigrate_isolate_page(pgdat, page);
1627 if (!isolated)
1628 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001629
Mel Gormanb32967f2012-11-19 12:35:47 +00001630 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001631 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1632 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001633 if (nr_remaining) {
1634 putback_lru_pages(&migratepages);
1635 isolated = 0;
1636 } else
1637 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001638 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001639 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001640
1641out:
1642 put_page(page);
1643 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001644}
Mel Gorman220018d2012-12-05 09:32:56 +00001645#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001646
Mel Gorman220018d2012-12-05 09:32:56 +00001647#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001648/*
1649 * Migrates a THP to a given target node. page must be locked and is unlocked
1650 * before returning.
1651 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001652int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1653 struct vm_area_struct *vma,
1654 pmd_t *pmd, pmd_t entry,
1655 unsigned long address,
1656 struct page *page, int node)
1657{
1658 unsigned long haddr = address & HPAGE_PMD_MASK;
1659 pg_data_t *pgdat = NODE_DATA(node);
1660 int isolated = 0;
1661 struct page *new_page = NULL;
1662 struct mem_cgroup *memcg = NULL;
1663 int page_lru = page_is_file_cache(page);
1664
1665 /*
1666 * Don't migrate pages that are mapped in multiple processes.
1667 * TODO: Handle false sharing detection instead of this hammer
1668 */
1669 if (page_mapcount(page) != 1)
1670 goto out_dropref;
1671
1672 /*
1673 * Rate-limit the amount of data that is being migrated to a node.
1674 * Optimal placement is no good if the memory bus is saturated and
1675 * all the time is being spent migrating!
1676 */
Mel Gormand28d43352012-11-29 09:24:36 +00001677 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001678 goto out_dropref;
1679
1680 new_page = alloc_pages_node(node,
1681 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001682 if (!new_page)
1683 goto out_fail;
1684
Mel Gorman22b751c2013-02-22 16:34:59 -08001685 page_nid_xchg_last(new_page, page_nid_last(page));
Mel Gormanb32967f2012-11-19 12:35:47 +00001686
1687 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001688 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001689 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001690 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001691 }
1692
1693 /* Prepare a page as a migration target */
1694 __set_page_locked(new_page);
1695 SetPageSwapBacked(new_page);
1696
1697 /* anon mapping, we can simply copy page->mapping to the new page: */
1698 new_page->mapping = page->mapping;
1699 new_page->index = page->index;
1700 migrate_page_copy(new_page, page);
1701 WARN_ON(PageLRU(new_page));
1702
1703 /* Recheck the target PMD */
1704 spin_lock(&mm->page_table_lock);
1705 if (unlikely(!pmd_same(*pmd, entry))) {
1706 spin_unlock(&mm->page_table_lock);
1707
1708 /* Reverse changes made by migrate_page_copy() */
1709 if (TestClearPageActive(new_page))
1710 SetPageActive(page);
1711 if (TestClearPageUnevictable(new_page))
1712 SetPageUnevictable(page);
1713 mlock_migrate_page(page, new_page);
1714
1715 unlock_page(new_page);
1716 put_page(new_page); /* Free it */
1717
1718 unlock_page(page);
1719 putback_lru_page(page);
1720
1721 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Hugh Dickins340ef392013-02-22 16:34:33 -08001722 isolated = 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001723 goto out;
1724 }
1725
1726 /*
1727 * Traditional migration needs to prepare the memcg charge
1728 * transaction early to prevent the old page from being
1729 * uncharged when installing migration entries. Here we can
1730 * save the potential rollback and start the charge transfer
1731 * only when migration is already known to end successfully.
1732 */
1733 mem_cgroup_prepare_migration(page, new_page, &memcg);
1734
1735 entry = mk_pmd(new_page, vma->vm_page_prot);
1736 entry = pmd_mknonnuma(entry);
1737 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1738 entry = pmd_mkhuge(entry);
1739
1740 page_add_new_anon_rmap(new_page, vma, haddr);
1741
1742 set_pmd_at(mm, haddr, pmd, entry);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001743 update_mmu_cache_pmd(vma, address, &entry);
Mel Gormanb32967f2012-11-19 12:35:47 +00001744 page_remove_rmap(page);
1745 /*
1746 * Finish the charge transaction under the page table lock to
1747 * prevent split_huge_page() from dividing up the charge
1748 * before it's fully transferred to the new page.
1749 */
1750 mem_cgroup_end_migration(memcg, page, new_page, true);
1751 spin_unlock(&mm->page_table_lock);
1752
1753 unlock_page(new_page);
1754 unlock_page(page);
1755 put_page(page); /* Drop the rmap reference */
1756 put_page(page); /* Drop the LRU isolation reference */
1757
1758 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1759 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1760
1761out:
1762 mod_zone_page_state(page_zone(page),
1763 NR_ISOLATED_ANON + page_lru,
1764 -HPAGE_PMD_NR);
1765 return isolated;
1766
Hugh Dickins340ef392013-02-22 16:34:33 -08001767out_fail:
1768 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001769out_dropref:
Hugh Dickins340ef392013-02-22 16:34:33 -08001770 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001771 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001772 return 0;
1773}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001774#endif /* CONFIG_NUMA_BALANCING */
1775
1776#endif /* CONFIG_NUMA */