blob: a26bccd44ccb0a907662c08135399462bd816b9c [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));
164 if (is_write_migration_entry(entry))
165 pte = pte_mkwrite(pte);
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200166#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800167 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900168 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800169 pte = arch_make_huge_pte(pte, vma, new, 0);
170 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200171#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700172 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700173 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700174
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900175 if (PageHuge(new)) {
176 if (PageAnon(new))
177 hugepage_add_anon_rmap(new, vma, addr);
178 else
179 page_dup_rmap(new);
180 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700181 page_add_anon_rmap(new, vma, addr);
182 else
183 page_add_file_rmap(new);
184
185 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000186 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800187unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700188 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800189out:
190 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700191}
192
193/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700194 * Get rid of all migration entries and replace them by
195 * references to the indicated page.
196 */
197static void remove_migration_ptes(struct page *old, struct page *new)
198{
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800199 rmap_walk(new, remove_migration_pte, old);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700200}
201
202/*
Christoph Lameter06972122006-06-23 02:03:35 -0700203 * Something used the pte of a page under migration. We need to
204 * get to the page and wait until migration is finished.
205 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700206 */
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700207static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
208 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700209{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700210 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700211 swp_entry_t entry;
212 struct page *page;
213
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700214 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700215 pte = *ptep;
216 if (!is_swap_pte(pte))
217 goto out;
218
219 entry = pte_to_swp_entry(pte);
220 if (!is_migration_entry(entry))
221 goto out;
222
223 page = migration_entry_to_page(entry);
224
Nick Piggine2867812008-07-25 19:45:30 -0700225 /*
226 * Once radix-tree replacement of page migration started, page_count
227 * *must* be zero. And, we don't want to call wait_on_page_locked()
228 * against a page without get_page().
229 * So, we use get_page_unless_zero(), here. Even failed, page fault
230 * will occur again.
231 */
232 if (!get_page_unless_zero(page))
233 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700234 pte_unmap_unlock(ptep, ptl);
235 wait_on_page_locked(page);
236 put_page(page);
237 return;
238out:
239 pte_unmap_unlock(ptep, ptl);
240}
241
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700242void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
243 unsigned long address)
244{
245 spinlock_t *ptl = pte_lockptr(mm, pmd);
246 pte_t *ptep = pte_offset_map(pmd, address);
247 __migration_entry_wait(mm, ptep, ptl);
248}
249
250void migration_entry_wait_huge(struct mm_struct *mm, pte_t *pte)
251{
252 spinlock_t *ptl = &(mm)->page_table_lock;
253 __migration_entry_wait(mm, pte, ptl);
254}
255
Mel Gormanb969c4a2012-01-12 17:19:34 -0800256#ifdef CONFIG_BLOCK
257/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800258static bool buffer_migrate_lock_buffers(struct buffer_head *head,
259 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800260{
261 struct buffer_head *bh = head;
262
263 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800264 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800265 do {
266 get_bh(bh);
267 lock_buffer(bh);
268 bh = bh->b_this_page;
269
270 } while (bh != head);
271
272 return true;
273 }
274
275 /* async case, we cannot block on lock_buffer so use trylock_buffer */
276 do {
277 get_bh(bh);
278 if (!trylock_buffer(bh)) {
279 /*
280 * We failed to lock the buffer and cannot stall in
281 * async migration. Release the taken locks
282 */
283 struct buffer_head *failed_bh = bh;
284 put_bh(failed_bh);
285 bh = head;
286 while (bh != failed_bh) {
287 unlock_buffer(bh);
288 put_bh(bh);
289 bh = bh->b_this_page;
290 }
291 return false;
292 }
293
294 bh = bh->b_this_page;
295 } while (bh != head);
296 return true;
297}
298#else
299static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800300 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800301{
302 return true;
303}
304#endif /* CONFIG_BLOCK */
305
Christoph Lameterb20a3502006-03-22 00:09:12 -0800306/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700307 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700308 *
309 * The number of remaining references must be:
310 * 1 for anonymous pages without a mapping
311 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100312 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800313 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800314int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800315 struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800316 struct buffer_head *head, enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800317{
Peter Zijlstra7039e1d2012-10-25 14:16:34 +0200318 int expected_count = 0;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800319 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800320
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700321 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700322 /* Anonymous page without mapping */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700323 if (page_count(page) != 1)
324 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800325 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700326 }
327
Nick Piggin19fd6232008-07-25 19:45:32 -0700328 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800329
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800330 pslot = radix_tree_lookup_slot(&mapping->page_tree,
331 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800332
Johannes Weineredcf4742009-09-21 17:02:59 -0700333 expected_count = 2 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700334 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800335 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700336 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700337 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800338 }
339
Nick Piggine2867812008-07-25 19:45:30 -0700340 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700341 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700342 return -EAGAIN;
343 }
344
Christoph Lameterb20a3502006-03-22 00:09:12 -0800345 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800346 * In the async migration case of moving a page with buffers, lock the
347 * buffers using trylock before the mapping is moved. If the mapping
348 * was moved, we later failed to lock the buffers and could not move
349 * the mapping back due to an elevated page count, we would have to
350 * block waiting on other references to be dropped.
351 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800352 if (mode == MIGRATE_ASYNC && head &&
353 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800354 page_unfreeze_refs(page, expected_count);
355 spin_unlock_irq(&mapping->tree_lock);
356 return -EAGAIN;
357 }
358
359 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800360 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800361 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800362 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800363 if (PageSwapCache(page)) {
364 SetPageSwapCache(newpage);
365 set_page_private(newpage, page_private(page));
366 }
367
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800368 radix_tree_replace_slot(pslot, newpage);
369
370 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800371 * Drop cache reference from old page by unfreezing
372 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800373 * We know this isn't the last reference.
374 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800375 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800376
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700377 /*
378 * If moved to a different zone then also account
379 * the page for that zone. Other VM counters will be
380 * taken care of when we establish references to the
381 * new page and drop references to the old page.
382 *
383 * Note that anonymous pages are accounted for
384 * via NR_FILE_PAGES and NR_ANON_PAGES if they
385 * are mapped to swap space.
386 */
387 __dec_zone_page_state(page, NR_FILE_PAGES);
388 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700389 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700390 __dec_zone_page_state(page, NR_SHMEM);
391 __inc_zone_page_state(newpage, NR_SHMEM);
392 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700393 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800394
Rafael Aquini78bd5202012-12-11 16:02:31 -0800395 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800396}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800397
398/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900399 * The expected number of remaining references is the same as that
400 * of migrate_page_move_mapping().
401 */
402int migrate_huge_page_move_mapping(struct address_space *mapping,
403 struct page *newpage, struct page *page)
404{
405 int expected_count;
406 void **pslot;
407
408 if (!mapping) {
409 if (page_count(page) != 1)
410 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800411 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900412 }
413
414 spin_lock_irq(&mapping->tree_lock);
415
416 pslot = radix_tree_lookup_slot(&mapping->page_tree,
417 page_index(page));
418
419 expected_count = 2 + page_has_private(page);
420 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800421 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900422 spin_unlock_irq(&mapping->tree_lock);
423 return -EAGAIN;
424 }
425
426 if (!page_freeze_refs(page, expected_count)) {
427 spin_unlock_irq(&mapping->tree_lock);
428 return -EAGAIN;
429 }
430
431 get_page(newpage);
432
433 radix_tree_replace_slot(pslot, newpage);
434
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800435 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900436
437 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800438 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900439}
440
441/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800442 * Copy the page to its new location
443 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900444void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800445{
Mel Gormanb32967f2012-11-19 12:35:47 +0000446 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900447 copy_huge_page(newpage, page);
448 else
449 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800450
451 if (PageError(page))
452 SetPageError(newpage);
453 if (PageReferenced(page))
454 SetPageReferenced(newpage);
455 if (PageUptodate(page))
456 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700457 if (TestClearPageActive(page)) {
458 VM_BUG_ON(PageUnevictable(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800459 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800460 } else if (TestClearPageUnevictable(page))
461 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800462 if (PageChecked(page))
463 SetPageChecked(newpage);
464 if (PageMappedToDisk(page))
465 SetPageMappedToDisk(newpage);
466
467 if (PageDirty(page)) {
468 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700469 /*
470 * Want to mark the page and the radix tree as dirty, and
471 * redo the accounting that clear_page_dirty_for_io undid,
472 * but we can't use set_page_dirty because that function
473 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300474 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700475 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700476 if (PageSwapBacked(page))
477 SetPageDirty(newpage);
478 else
479 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800480 }
481
Nick Pigginb291f002008-10-18 20:26:44 -0700482 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800483 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800484 /*
485 * Please do not reorder this without considering how mm/ksm.c's
486 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
487 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800488 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800489 ClearPagePrivate(page);
490 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800491
492 /*
493 * If any waiters have accumulated on the new page then
494 * wake them up.
495 */
496 if (PageWriteback(newpage))
497 end_page_writeback(newpage);
498}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800499
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700500/************************************************************
501 * Migration functions
502 ***********************************************************/
503
504/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700505int fail_migrate_page(struct address_space *mapping,
506 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700507{
508 return -EIO;
509}
510EXPORT_SYMBOL(fail_migrate_page);
511
Christoph Lameterb20a3502006-03-22 00:09:12 -0800512/*
513 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100514 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800515 *
516 * Pages are locked upon entry and exit.
517 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700518int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800519 struct page *newpage, struct page *page,
520 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800521{
522 int rc;
523
524 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
525
Mel Gormana6bc32b2012-01-12 17:19:43 -0800526 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800527
Rafael Aquini78bd5202012-12-11 16:02:31 -0800528 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800529 return rc;
530
531 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800532 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800533}
534EXPORT_SYMBOL(migrate_page);
535
David Howells93614012006-09-30 20:45:40 +0200536#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800537/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700538 * Migration function for pages with buffers. This function can only be used
539 * if the underlying filesystem guarantees that no other references to "page"
540 * exist.
541 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700542int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800543 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700544{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700545 struct buffer_head *bh, *head;
546 int rc;
547
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700548 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800549 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700550
551 head = page_buffers(page);
552
Mel Gormana6bc32b2012-01-12 17:19:43 -0800553 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700554
Rafael Aquini78bd5202012-12-11 16:02:31 -0800555 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700556 return rc;
557
Mel Gormanb969c4a2012-01-12 17:19:34 -0800558 /*
559 * In the async case, migrate_page_move_mapping locked the buffers
560 * with an IRQ-safe spinlock held. In the sync case, the buffers
561 * need to be locked now
562 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800563 if (mode != MIGRATE_ASYNC)
564 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700565
566 ClearPagePrivate(page);
567 set_page_private(newpage, page_private(page));
568 set_page_private(page, 0);
569 put_page(page);
570 get_page(newpage);
571
572 bh = head;
573 do {
574 set_bh_page(bh, newpage, bh_offset(bh));
575 bh = bh->b_this_page;
576
577 } while (bh != head);
578
579 SetPagePrivate(newpage);
580
581 migrate_page_copy(newpage, page);
582
583 bh = head;
584 do {
585 unlock_buffer(bh);
586 put_bh(bh);
587 bh = bh->b_this_page;
588
589 } while (bh != head);
590
Rafael Aquini78bd5202012-12-11 16:02:31 -0800591 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700592}
593EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200594#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700595
Christoph Lameter04e62a22006-06-23 02:03:38 -0700596/*
597 * Writeback a page to clean the dirty state
598 */
599static int writeout(struct address_space *mapping, struct page *page)
600{
601 struct writeback_control wbc = {
602 .sync_mode = WB_SYNC_NONE,
603 .nr_to_write = 1,
604 .range_start = 0,
605 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700606 .for_reclaim = 1
607 };
608 int rc;
609
610 if (!mapping->a_ops->writepage)
611 /* No write method for the address space */
612 return -EINVAL;
613
614 if (!clear_page_dirty_for_io(page))
615 /* Someone else already triggered a write */
616 return -EAGAIN;
617
618 /*
619 * A dirty page may imply that the underlying filesystem has
620 * the page on some queue. So the page must be clean for
621 * migration. Writeout may mean we loose the lock and the
622 * page state is no longer what we checked for earlier.
623 * At this point we know that the migration attempt cannot
624 * be successful.
625 */
626 remove_migration_ptes(page, page);
627
628 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700629
630 if (rc != AOP_WRITEPAGE_ACTIVATE)
631 /* unlocked. Relock */
632 lock_page(page);
633
Hugh Dickinsbda85502008-11-19 15:36:36 -0800634 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700635}
636
637/*
638 * Default handling if a filesystem does not provide a migration function.
639 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700640static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800641 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700642{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800643 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800644 /* Only writeback pages in full synchronous migration */
645 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800646 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700647 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800648 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700649
650 /*
651 * Buffers may be managed in a filesystem specific way.
652 * We must have no buffers or drop them.
653 */
David Howells266cf652009-04-03 16:42:36 +0100654 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700655 !try_to_release_page(page, GFP_KERNEL))
656 return -EAGAIN;
657
Mel Gormana6bc32b2012-01-12 17:19:43 -0800658 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700659}
660
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700661/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700662 * Move a page to a newly allocated page
663 * The page is locked and all ptes have been successfully removed.
664 *
665 * The new page will have replaced the old page if this function
666 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700667 *
668 * Return value:
669 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800670 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700671 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700672static int move_to_new_page(struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800673 int remap_swapcache, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700674{
675 struct address_space *mapping;
676 int rc;
677
678 /*
679 * Block others from accessing the page when we get around to
680 * establishing additional references. We are the only one
681 * holding a reference to the new page at this point.
682 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200683 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700684 BUG();
685
686 /* Prepare mapping for the new page.*/
687 newpage->index = page->index;
688 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700689 if (PageSwapBacked(page))
690 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700691
692 mapping = page_mapping(page);
693 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800694 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800695 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700696 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800697 * Most pages have a mapping and most filesystems provide a
698 * migratepage callback. Anonymous pages are part of swap
699 * space which also has its own migratepage callback. This
700 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700701 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800702 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800703 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800704 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800705 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700706
Rafael Aquini78bd5202012-12-11 16:02:31 -0800707 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700708 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700709 } else {
710 if (remap_swapcache)
711 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800712 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700713 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700714
715 unlock_page(newpage);
716
717 return rc;
718}
719
Minchan Kim0dabec92011-10-31 17:06:57 -0700720static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800721 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700722{
Minchan Kim0dabec92011-10-31 17:06:57 -0700723 int rc = -EAGAIN;
Mel Gorman3fe20112010-05-24 14:32:20 -0700724 int remap_swapcache = 1;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -0700725 struct mem_cgroup *mem;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700726 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700727
Nick Piggin529ae9a2008-08-02 12:01:03 +0200728 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800729 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700730 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800731
732 /*
733 * It's not safe for direct compaction to call lock_page.
734 * For example, during page readahead pages are added locked
735 * to the LRU. Later, when the IO completes the pages are
736 * marked uptodate and unlocked. However, the queueing
737 * could be merging multiple pages for one bio (e.g.
738 * mpage_readpages). If an allocation happens for the
739 * second or third page, the process can end up locking
740 * the same page twice and deadlocking. Rather than
741 * trying to be clever about what pages can be locked,
742 * avoid the use of lock_page for direct compaction
743 * altogether.
744 */
745 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700746 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800747
Christoph Lametere24f0b82006-06-23 02:03:51 -0700748 lock_page(page);
749 }
750
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800751 /* charge against new page */
Johannes Weiner0030f532012-07-31 16:45:25 -0700752 mem_cgroup_prepare_migration(page, newpage, &mem);
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800753
Christoph Lametere24f0b82006-06-23 02:03:51 -0700754 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700755 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700756 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800757 * necessary to wait for PageWriteback. In the async case,
758 * the retry loop is too short and in the sync-light case,
759 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700760 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800761 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700762 rc = -EBUSY;
763 goto uncharge;
764 }
765 if (!force)
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800766 goto uncharge;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700767 wait_on_page_writeback(page);
768 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700769 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700770 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
771 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800772 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700773 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700774 * File Caches may use write_page() or lock_page() in migration, then,
775 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700776 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800777 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800778 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000779 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800780 * getting a hold on an anon_vma from outside one of its mms.
781 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700782 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800783 if (anon_vma) {
784 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700785 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800786 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800787 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700788 /*
789 * We cannot be sure that the anon_vma of an unmapped
790 * swapcache page is safe to use because we don't
791 * know in advance if the VMA that this page belonged
792 * to still exists. If the VMA and others sharing the
793 * data have been freed, then the anon_vma could
794 * already be invalid.
795 *
796 * To avoid this possibility, swapcache pages get
797 * migrated but are not remapped when migration
798 * completes
799 */
800 remap_swapcache = 0;
801 } else {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800802 goto uncharge;
Mel Gorman3fe20112010-05-24 14:32:20 -0700803 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700804 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800805
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800806 if (unlikely(balloon_page_movable(page))) {
807 /*
808 * A ballooned page does not need any special attention from
809 * physical to virtual reverse mapping procedures.
810 * Skip any attempt to unmap PTEs or to remap swap cache,
811 * in order to avoid burning cycles at rmap level, and perform
812 * the page migration right away (proteced by page lock).
813 */
814 rc = balloon_page_migrate(newpage, page, mode);
815 goto uncharge;
816 }
817
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700818 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800819 * Corner case handling:
820 * 1. When a new swap-cache page is read into, it is added to the LRU
821 * and treated as swapcache but it has no rmap yet.
822 * Calling try_to_unmap() against a page->mapping==NULL page will
823 * trigger a BUG. So handle it here.
824 * 2. An orphaned page (see truncate_complete_page) might have
825 * fs-private metadata. The page can be picked up due to memory
826 * offlining. Everywhere else except page reclaim, the page is
827 * invisible to the vm, so the page can not be migrated. So try to
828 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700829 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800830 if (!page->mapping) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800831 VM_BUG_ON(PageAnon(page));
832 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800833 try_to_free_buffers(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800834 goto uncharge;
Shaohua Li62e1c552008-02-04 22:29:33 -0800835 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700836 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800837 }
838
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700839 /* Establish migration ptes or remove ptes */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200840 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700841
Shaohua Liabfc3482009-09-21 17:01:19 -0700842skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700843 if (!page_mapped(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800844 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700845
Mel Gorman3fe20112010-05-24 14:32:20 -0700846 if (rc && remap_swapcache)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700847 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700848
849 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700850 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700851 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700852
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800853uncharge:
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800854 mem_cgroup_end_migration(mem, page, newpage,
855 (rc == MIGRATEPAGE_SUCCESS ||
856 rc == MIGRATEPAGE_BALLOON_SUCCESS));
Christoph Lametere24f0b82006-06-23 02:03:51 -0700857 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700858out:
859 return rc;
860}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700861
Minchan Kim0dabec92011-10-31 17:06:57 -0700862/*
863 * Obtain the lock on page, remove all ptes and migrate the page
864 * to the newly allocated page in newpage.
865 */
866static int unmap_and_move(new_page_t get_new_page, unsigned long private,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800867 struct page *page, int force, enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700868{
869 int rc = 0;
870 int *result = NULL;
871 struct page *newpage = get_new_page(page, private, &result);
872
873 if (!newpage)
874 return -ENOMEM;
875
876 if (page_count(page) == 1) {
877 /* page was freed from under us. So we are done. */
878 goto out;
879 }
880
881 if (unlikely(PageTransHuge(page)))
882 if (unlikely(split_huge_page(page)))
883 goto out;
884
Hugh Dickins9c620e22013-02-22 16:35:14 -0800885 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800886
887 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
888 /*
889 * A ballooned page has been migrated already.
890 * Now, it's the time to wrap-up counters,
891 * handle the page back to Buddy and return.
892 */
893 dec_zone_page_state(page, NR_ISOLATED_ANON +
894 page_is_file_cache(page));
895 balloon_page_free(page);
896 return MIGRATEPAGE_SUCCESS;
897 }
Minchan Kim0dabec92011-10-31 17:06:57 -0700898out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700899 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700900 /*
901 * A page that has been migrated has all references
902 * removed and will be freed. A page that has not been
903 * migrated will have kepts its references and be
904 * restored.
905 */
906 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700907 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700908 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700909 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700910 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700911 /*
912 * Move the new page to the LRU. If migration was not successful
913 * then this will free the page.
914 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700915 putback_lru_page(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700916 if (result) {
917 if (rc)
918 *result = rc;
919 else
920 *result = page_to_nid(newpage);
921 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700922 return rc;
923}
924
925/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900926 * Counterpart of unmap_and_move_page() for hugepage migration.
927 *
928 * This function doesn't wait the completion of hugepage I/O
929 * because there is no race between I/O and migration for hugepage.
930 * Note that currently hugepage I/O occurs only in direct I/O
931 * where no lock is held and PG_writeback is irrelevant,
932 * and writeback status of all subpages are counted in the reference
933 * count of the head page (i.e. if all subpages of a 2MB hugepage are
934 * under direct I/O, the reference of the head page is 512 and a bit more.)
935 * This means that when we try to migrate hugepage whose subpages are
936 * doing direct I/O, some references remain after try_to_unmap() and
937 * hugepage migration fails without data corruption.
938 *
939 * There is also no race when direct I/O is issued on the page under migration,
940 * because then pte is replaced with migration swap entry and direct I/O code
941 * will wait in the page fault for migration to complete.
942 */
943static int unmap_and_move_huge_page(new_page_t get_new_page,
944 unsigned long private, struct page *hpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800945 int force, enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900946{
947 int rc = 0;
948 int *result = NULL;
949 struct page *new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900950 struct anon_vma *anon_vma = NULL;
951
Naoya Horiguchi83467ef2013-09-11 14:22:11 -0700952 /*
953 * Movability of hugepages depends on architectures and hugepage size.
954 * This check is necessary because some callers of hugepage migration
955 * like soft offline and memory hotremove don't walk through page
956 * tables or check whether the hugepage is pmd-based or not before
957 * kicking migration.
958 */
959 if (!hugepage_migration_support(page_hstate(hpage)))
960 return -ENOSYS;
961
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900962 if (!new_hpage)
963 return -ENOMEM;
964
965 rc = -EAGAIN;
966
967 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800968 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900969 goto out;
970 lock_page(hpage);
971 }
972
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700973 if (PageAnon(hpage))
974 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900975
976 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
977
978 if (!page_mapped(hpage))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800979 rc = move_to_new_page(new_hpage, hpage, 1, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900980
981 if (rc)
982 remove_migration_ptes(hpage, hpage);
983
Hugh Dickinsfd4a4662011-01-13 15:47:31 -0800984 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700985 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -0700986
987 if (!rc)
988 hugetlb_cgroup_migrate(hpage, new_hpage);
989
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900990 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -0800991out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -0700992 if (rc != -EAGAIN)
993 putback_active_hugepage(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900994 put_page(new_hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900995 if (result) {
996 if (rc)
997 *result = rc;
998 else
999 *result = page_to_nid(new_hpage);
1000 }
1001 return rc;
1002}
1003
1004/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001005 * migrate_pages - migrate the pages specified in a list, to the free pages
1006 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001007 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001008 * @from: The list of pages to be migrated.
1009 * @get_new_page: The function used to allocate free pages to be used
1010 * as the target of the page migration.
1011 * @private: Private data to be passed on to get_new_page()
1012 * @mode: The migration mode that specifies the constraints for
1013 * page migration, if any.
1014 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001015 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001016 * The function returns after 10 attempts or if no pages are movable any more
1017 * because the list has become empty or no retryable pages exist any more.
1018 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001019 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001020 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001021 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001022 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001023int migrate_pages(struct list_head *from, new_page_t get_new_page,
1024 unsigned long private, enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001025{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001026 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001027 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001028 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001029 int pass = 0;
1030 struct page *page;
1031 struct page *page2;
1032 int swapwrite = current->flags & PF_SWAPWRITE;
1033 int rc;
1034
1035 if (!swapwrite)
1036 current->flags |= PF_SWAPWRITE;
1037
Christoph Lametere24f0b82006-06-23 02:03:51 -07001038 for(pass = 0; pass < 10 && retry; pass++) {
1039 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001040
Christoph Lametere24f0b82006-06-23 02:03:51 -07001041 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001042 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001043
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001044 if (PageHuge(page))
1045 rc = unmap_and_move_huge_page(get_new_page,
1046 private, page, pass > 2, mode);
1047 else
1048 rc = unmap_and_move(get_new_page, private,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001049 page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001050
Christoph Lametere24f0b82006-06-23 02:03:51 -07001051 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001052 case -ENOMEM:
1053 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001054 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001055 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001056 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001057 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001058 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001059 break;
1060 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001061 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001062 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001063 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001064 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001065 }
1066 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001067 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001068out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001069 if (nr_succeeded)
1070 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1071 if (nr_failed)
1072 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001073 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1074
Christoph Lameterb20a3502006-03-22 00:09:12 -08001075 if (!swapwrite)
1076 current->flags &= ~PF_SWAPWRITE;
1077
Rafael Aquini78bd5202012-12-11 16:02:31 -08001078 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001079}
1080
Christoph Lameter742755a2006-06-23 02:03:55 -07001081#ifdef CONFIG_NUMA
1082/*
1083 * Move a list of individual pages
1084 */
1085struct page_to_node {
1086 unsigned long addr;
1087 struct page *page;
1088 int node;
1089 int status;
1090};
1091
1092static struct page *new_page_node(struct page *p, unsigned long private,
1093 int **result)
1094{
1095 struct page_to_node *pm = (struct page_to_node *)private;
1096
1097 while (pm->node != MAX_NUMNODES && pm->page != p)
1098 pm++;
1099
1100 if (pm->node == MAX_NUMNODES)
1101 return NULL;
1102
1103 *result = &pm->status;
1104
Naoya Horiguchie632a932013-09-11 14:22:04 -07001105 if (PageHuge(p))
1106 return alloc_huge_page_node(page_hstate(compound_head(p)),
1107 pm->node);
1108 else
1109 return alloc_pages_exact_node(pm->node,
Mel Gorman769848c2007-07-17 04:03:05 -07001110 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001111}
1112
1113/*
1114 * Move a set of pages as indicated in the pm array. The addr
1115 * field must be set to the virtual address of the page to be moved
1116 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001117 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001118 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001119static int do_move_page_to_node_array(struct mm_struct *mm,
1120 struct page_to_node *pm,
1121 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001122{
1123 int err;
1124 struct page_to_node *pp;
1125 LIST_HEAD(pagelist);
1126
1127 down_read(&mm->mmap_sem);
1128
1129 /*
1130 * Build a list of pages to migrate
1131 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001132 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1133 struct vm_area_struct *vma;
1134 struct page *page;
1135
Christoph Lameter742755a2006-06-23 02:03:55 -07001136 err = -EFAULT;
1137 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001138 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001139 goto set_status;
1140
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001141 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001142
1143 err = PTR_ERR(page);
1144 if (IS_ERR(page))
1145 goto set_status;
1146
Christoph Lameter742755a2006-06-23 02:03:55 -07001147 err = -ENOENT;
1148 if (!page)
1149 goto set_status;
1150
Hugh Dickins62b61f62009-12-14 17:59:33 -08001151 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001152 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001153 goto put_and_set;
1154
1155 pp->page = page;
1156 err = page_to_nid(page);
1157
1158 if (err == pp->node)
1159 /*
1160 * Node already in the right place
1161 */
1162 goto put_and_set;
1163
1164 err = -EACCES;
1165 if (page_mapcount(page) > 1 &&
1166 !migrate_all)
1167 goto put_and_set;
1168
Naoya Horiguchie632a932013-09-11 14:22:04 -07001169 if (PageHuge(page)) {
1170 isolate_huge_page(page, &pagelist);
1171 goto put_and_set;
1172 }
1173
Nick Piggin62695a82008-10-18 20:26:09 -07001174 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001175 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001176 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001177 inc_zone_page_state(page, NR_ISOLATED_ANON +
1178 page_is_file_cache(page));
1179 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001180put_and_set:
1181 /*
1182 * Either remove the duplicate refcount from
1183 * isolate_lru_page() or drop the page ref if it was
1184 * not isolated.
1185 */
1186 put_page(page);
1187set_status:
1188 pp->status = err;
1189 }
1190
Brice Gogline78bbfa2008-10-18 20:27:15 -07001191 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001192 if (!list_empty(&pagelist)) {
Christoph Lameter742755a2006-06-23 02:03:55 -07001193 err = migrate_pages(&pagelist, new_page_node,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001194 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001195 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001196 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001197 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001198
1199 up_read(&mm->mmap_sem);
1200 return err;
1201}
1202
1203/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001204 * Migrate an array of page address onto an array of nodes and fill
1205 * the corresponding array of status.
1206 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001207static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001208 unsigned long nr_pages,
1209 const void __user * __user *pages,
1210 const int __user *nodes,
1211 int __user *status, int flags)
1212{
Brice Goglin3140a222009-01-06 14:38:57 -08001213 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001214 unsigned long chunk_nr_pages;
1215 unsigned long chunk_start;
1216 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001217
Brice Goglin3140a222009-01-06 14:38:57 -08001218 err = -ENOMEM;
1219 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1220 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001221 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001222
1223 migrate_prep();
1224
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001225 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001226 * Store a chunk of page_to_node array in a page,
1227 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001228 */
Brice Goglin3140a222009-01-06 14:38:57 -08001229 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001230
Brice Goglin3140a222009-01-06 14:38:57 -08001231 for (chunk_start = 0;
1232 chunk_start < nr_pages;
1233 chunk_start += chunk_nr_pages) {
1234 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001235
Brice Goglin3140a222009-01-06 14:38:57 -08001236 if (chunk_start + chunk_nr_pages > nr_pages)
1237 chunk_nr_pages = nr_pages - chunk_start;
1238
1239 /* fill the chunk pm with addrs and nodes from user-space */
1240 for (j = 0; j < chunk_nr_pages; j++) {
1241 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001242 int node;
1243
Brice Goglin3140a222009-01-06 14:38:57 -08001244 err = -EFAULT;
1245 if (get_user(p, pages + j + chunk_start))
1246 goto out_pm;
1247 pm[j].addr = (unsigned long) p;
1248
1249 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001250 goto out_pm;
1251
1252 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001253 if (node < 0 || node >= MAX_NUMNODES)
1254 goto out_pm;
1255
Lai Jiangshan389162c2012-12-12 13:51:30 -08001256 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001257 goto out_pm;
1258
1259 err = -EACCES;
1260 if (!node_isset(node, task_nodes))
1261 goto out_pm;
1262
Brice Goglin3140a222009-01-06 14:38:57 -08001263 pm[j].node = node;
1264 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001265
Brice Goglin3140a222009-01-06 14:38:57 -08001266 /* End marker for this chunk */
1267 pm[chunk_nr_pages].node = MAX_NUMNODES;
1268
1269 /* Migrate this chunk */
1270 err = do_move_page_to_node_array(mm, pm,
1271 flags & MPOL_MF_MOVE_ALL);
1272 if (err < 0)
1273 goto out_pm;
1274
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001275 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001276 for (j = 0; j < chunk_nr_pages; j++)
1277 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001278 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001279 goto out_pm;
1280 }
1281 }
1282 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001283
1284out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001285 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001286out:
1287 return err;
1288}
1289
1290/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001291 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001292 */
Brice Goglin80bba122008-12-09 13:14:23 -08001293static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1294 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001295{
Brice Goglin2f007e72008-10-18 20:27:16 -07001296 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001297
Christoph Lameter742755a2006-06-23 02:03:55 -07001298 down_read(&mm->mmap_sem);
1299
Brice Goglin2f007e72008-10-18 20:27:16 -07001300 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001301 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001302 struct vm_area_struct *vma;
1303 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001304 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001305
1306 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001307 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001308 goto set_status;
1309
Brice Goglin2f007e72008-10-18 20:27:16 -07001310 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001311
1312 err = PTR_ERR(page);
1313 if (IS_ERR(page))
1314 goto set_status;
1315
Christoph Lameter742755a2006-06-23 02:03:55 -07001316 err = -ENOENT;
1317 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001318 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001319 goto set_status;
1320
1321 err = page_to_nid(page);
1322set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001323 *status = err;
1324
1325 pages++;
1326 status++;
1327 }
1328
1329 up_read(&mm->mmap_sem);
1330}
1331
1332/*
1333 * Determine the nodes of a user array of pages and store it in
1334 * a user array of status.
1335 */
1336static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1337 const void __user * __user *pages,
1338 int __user *status)
1339{
1340#define DO_PAGES_STAT_CHUNK_NR 16
1341 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1342 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001343
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001344 while (nr_pages) {
1345 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001346
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001347 chunk_nr = nr_pages;
1348 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1349 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1350
1351 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1352 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001353
1354 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1355
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001356 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1357 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001358
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001359 pages += chunk_nr;
1360 status += chunk_nr;
1361 nr_pages -= chunk_nr;
1362 }
1363 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001364}
1365
1366/*
1367 * Move a list of pages in the address space of the currently executing
1368 * process.
1369 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001370SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1371 const void __user * __user *, pages,
1372 const int __user *, nodes,
1373 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001374{
David Howellsc69e8d92008-11-14 10:39:19 +11001375 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001376 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001377 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001378 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001379 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001380
1381 /* Check flags */
1382 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1383 return -EINVAL;
1384
1385 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1386 return -EPERM;
1387
1388 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001389 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001390 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001391 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001392 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001393 return -ESRCH;
1394 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001395 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001396
1397 /*
1398 * Check if this process has the right to modify the specified
1399 * process. The right exists if the process has administrative
1400 * capabilities, superuser privileges or the same
1401 * userid as the target process.
1402 */
David Howellsc69e8d92008-11-14 10:39:19 +11001403 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001404 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1405 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001406 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001407 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001408 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001409 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001410 }
David Howellsc69e8d92008-11-14 10:39:19 +11001411 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001412
David Quigley86c3a762006-06-23 02:04:02 -07001413 err = security_task_movememory(task);
1414 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001415 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001416
Christoph Lameter3268c632012-03-21 16:34:06 -07001417 task_nodes = cpuset_mems_allowed(task);
1418 mm = get_task_mm(task);
1419 put_task_struct(task);
1420
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001421 if (!mm)
1422 return -EINVAL;
1423
1424 if (nodes)
1425 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1426 nodes, status, flags);
1427 else
1428 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001429
1430 mmput(mm);
1431 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001432
Christoph Lameter742755a2006-06-23 02:03:55 -07001433out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001434 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001435 return err;
1436}
Christoph Lameter742755a2006-06-23 02:03:55 -07001437
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001438/*
1439 * Call migration functions in the vma_ops that may prepare
1440 * memory in a vm for migration. migration functions may perform
1441 * the migration for vmas that do not have an underlying page struct.
1442 */
1443int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1444 const nodemask_t *from, unsigned long flags)
1445{
1446 struct vm_area_struct *vma;
1447 int err = 0;
1448
Daisuke Nishimura1001c9f2009-02-11 13:04:18 -08001449 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001450 if (vma->vm_ops && vma->vm_ops->migrate) {
1451 err = vma->vm_ops->migrate(vma, to, from, flags);
1452 if (err)
1453 break;
1454 }
1455 }
1456 return err;
1457}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001458
1459#ifdef CONFIG_NUMA_BALANCING
1460/*
1461 * Returns true if this is a safe migration target node for misplaced NUMA
1462 * pages. Currently it only checks the watermarks which crude
1463 */
1464static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001465 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001466{
1467 int z;
1468 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1469 struct zone *zone = pgdat->node_zones + z;
1470
1471 if (!populated_zone(zone))
1472 continue;
1473
Lisa Du6e543d52013-09-11 14:22:36 -07001474 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001475 continue;
1476
1477 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1478 if (!zone_watermark_ok(zone, 0,
1479 high_wmark_pages(zone) +
1480 nr_migrate_pages,
1481 0, 0))
1482 continue;
1483 return true;
1484 }
1485 return false;
1486}
1487
1488static struct page *alloc_misplaced_dst_page(struct page *page,
1489 unsigned long data,
1490 int **result)
1491{
1492 int nid = (int) data;
1493 struct page *newpage;
1494
1495 newpage = alloc_pages_exact_node(nid,
1496 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1497 __GFP_NOMEMALLOC | __GFP_NORETRY |
1498 __GFP_NOWARN) &
1499 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001500 if (newpage)
Mel Gorman22b751c2013-02-22 16:34:59 -08001501 page_nid_xchg_last(newpage, page_nid_last(page));
Hillf Dantonbac03822012-11-27 14:46:24 +00001502
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001503 return newpage;
1504}
1505
1506/*
Mel Gormana8f60772012-11-14 21:41:46 +00001507 * page migration rate limiting control.
1508 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1509 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001510 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1511 * as it is faults that reset the window, pte updates will happen unconditionally
1512 * if there has not been a fault since @pteupdate_interval_millisecs after the
1513 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001514 */
1515static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001516static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001517static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1518
Mel Gormane14808b2012-11-19 10:59:15 +00001519/* Returns true if NUMA migration is currently rate limited */
1520bool migrate_ratelimited(int node)
1521{
1522 pg_data_t *pgdat = NODE_DATA(node);
1523
1524 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1525 msecs_to_jiffies(pteupdate_interval_millisecs)))
1526 return false;
1527
1528 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1529 return false;
1530
1531 return true;
1532}
1533
Mel Gormanb32967f2012-11-19 12:35:47 +00001534/* Returns true if the node is migrate rate-limited after the update */
Mel Gormand28d43352012-11-29 09:24:36 +00001535bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001536{
1537 bool rate_limited = false;
1538
1539 /*
1540 * Rate-limit the amount of data that is being migrated to a node.
1541 * Optimal placement is no good if the memory bus is saturated and
1542 * all the time is being spent migrating!
1543 */
1544 spin_lock(&pgdat->numabalancing_migrate_lock);
1545 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1546 pgdat->numabalancing_migrate_nr_pages = 0;
1547 pgdat->numabalancing_migrate_next_window = jiffies +
1548 msecs_to_jiffies(migrate_interval_millisecs);
1549 }
1550 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1551 rate_limited = true;
1552 else
Mel Gormand28d43352012-11-29 09:24:36 +00001553 pgdat->numabalancing_migrate_nr_pages += nr_pages;
Mel Gormanb32967f2012-11-19 12:35:47 +00001554 spin_unlock(&pgdat->numabalancing_migrate_lock);
1555
1556 return rate_limited;
1557}
1558
1559int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1560{
Hugh Dickins340ef392013-02-22 16:34:33 -08001561 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001562
Mel Gorman3abef4e2013-02-22 16:34:27 -08001563 VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1564
Mel Gormanb32967f2012-11-19 12:35:47 +00001565 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001566 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1567 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001568
Hugh Dickins340ef392013-02-22 16:34:33 -08001569 if (isolate_lru_page(page))
1570 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001571
1572 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001573 * migrate_misplaced_transhuge_page() skips page migration's usual
1574 * check on page_count(), so we must do it here, now that the page
1575 * has been isolated: a GUP pin, or any other pin, prevents migration.
1576 * The expected page count is 3: 1 for page's mapcount and 1 for the
1577 * caller's pin and 1 for the reference taken by isolate_lru_page().
1578 */
1579 if (PageTransHuge(page) && page_count(page) != 3) {
1580 putback_lru_page(page);
1581 return 0;
1582 }
1583
1584 page_lru = page_is_file_cache(page);
1585 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1586 hpage_nr_pages(page));
1587
1588 /*
1589 * Isolating the page has taken another reference, so the
1590 * caller's reference can be safely dropped without the page
1591 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001592 */
1593 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001594 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001595}
1596
Mel Gormana8f60772012-11-14 21:41:46 +00001597/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001598 * Attempt to migrate a misplaced page to the specified destination
1599 * node. Caller is expected to have an elevated reference count on
1600 * the page that will be dropped by this function before returning.
1601 */
1602int migrate_misplaced_page(struct page *page, int node)
1603{
Mel Gormana8f60772012-11-14 21:41:46 +00001604 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001605 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001606 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001607 LIST_HEAD(migratepages);
1608
1609 /*
1610 * Don't migrate pages that are mapped in multiple processes.
1611 * TODO: Handle false sharing detection instead of this hammer
1612 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001613 if (page_mapcount(page) != 1)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001614 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001615
Mel Gormana8f60772012-11-14 21:41:46 +00001616 /*
1617 * Rate-limit the amount of data that is being migrated to a node.
1618 * Optimal placement is no good if the memory bus is saturated and
1619 * all the time is being spent migrating!
1620 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001621 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001622 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001623
Mel Gormanb32967f2012-11-19 12:35:47 +00001624 isolated = numamigrate_isolate_page(pgdat, page);
1625 if (!isolated)
1626 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001627
Mel Gormanb32967f2012-11-19 12:35:47 +00001628 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001629 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1630 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001631 if (nr_remaining) {
1632 putback_lru_pages(&migratepages);
1633 isolated = 0;
1634 } else
1635 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001636 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001637 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001638
1639out:
1640 put_page(page);
1641 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001642}
Mel Gorman220018d2012-12-05 09:32:56 +00001643#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001644
Mel Gorman220018d2012-12-05 09:32:56 +00001645#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001646/*
1647 * Migrates a THP to a given target node. page must be locked and is unlocked
1648 * before returning.
1649 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001650int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1651 struct vm_area_struct *vma,
1652 pmd_t *pmd, pmd_t entry,
1653 unsigned long address,
1654 struct page *page, int node)
1655{
1656 unsigned long haddr = address & HPAGE_PMD_MASK;
1657 pg_data_t *pgdat = NODE_DATA(node);
1658 int isolated = 0;
1659 struct page *new_page = NULL;
1660 struct mem_cgroup *memcg = NULL;
1661 int page_lru = page_is_file_cache(page);
1662
1663 /*
1664 * Don't migrate pages that are mapped in multiple processes.
1665 * TODO: Handle false sharing detection instead of this hammer
1666 */
1667 if (page_mapcount(page) != 1)
1668 goto out_dropref;
1669
1670 /*
1671 * Rate-limit the amount of data that is being migrated to a node.
1672 * Optimal placement is no good if the memory bus is saturated and
1673 * all the time is being spent migrating!
1674 */
Mel Gormand28d43352012-11-29 09:24:36 +00001675 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001676 goto out_dropref;
1677
1678 new_page = alloc_pages_node(node,
1679 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001680 if (!new_page)
1681 goto out_fail;
1682
Mel Gorman22b751c2013-02-22 16:34:59 -08001683 page_nid_xchg_last(new_page, page_nid_last(page));
Mel Gormanb32967f2012-11-19 12:35:47 +00001684
1685 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001686 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001687 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001688 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001689 }
1690
1691 /* Prepare a page as a migration target */
1692 __set_page_locked(new_page);
1693 SetPageSwapBacked(new_page);
1694
1695 /* anon mapping, we can simply copy page->mapping to the new page: */
1696 new_page->mapping = page->mapping;
1697 new_page->index = page->index;
1698 migrate_page_copy(new_page, page);
1699 WARN_ON(PageLRU(new_page));
1700
1701 /* Recheck the target PMD */
1702 spin_lock(&mm->page_table_lock);
1703 if (unlikely(!pmd_same(*pmd, entry))) {
1704 spin_unlock(&mm->page_table_lock);
1705
1706 /* Reverse changes made by migrate_page_copy() */
1707 if (TestClearPageActive(new_page))
1708 SetPageActive(page);
1709 if (TestClearPageUnevictable(new_page))
1710 SetPageUnevictable(page);
1711 mlock_migrate_page(page, new_page);
1712
1713 unlock_page(new_page);
1714 put_page(new_page); /* Free it */
1715
1716 unlock_page(page);
1717 putback_lru_page(page);
1718
1719 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Hugh Dickins340ef392013-02-22 16:34:33 -08001720 isolated = 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001721 goto out;
1722 }
1723
1724 /*
1725 * Traditional migration needs to prepare the memcg charge
1726 * transaction early to prevent the old page from being
1727 * uncharged when installing migration entries. Here we can
1728 * save the potential rollback and start the charge transfer
1729 * only when migration is already known to end successfully.
1730 */
1731 mem_cgroup_prepare_migration(page, new_page, &memcg);
1732
1733 entry = mk_pmd(new_page, vma->vm_page_prot);
1734 entry = pmd_mknonnuma(entry);
1735 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1736 entry = pmd_mkhuge(entry);
1737
1738 page_add_new_anon_rmap(new_page, vma, haddr);
1739
1740 set_pmd_at(mm, haddr, pmd, entry);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001741 update_mmu_cache_pmd(vma, address, &entry);
Mel Gormanb32967f2012-11-19 12:35:47 +00001742 page_remove_rmap(page);
1743 /*
1744 * Finish the charge transaction under the page table lock to
1745 * prevent split_huge_page() from dividing up the charge
1746 * before it's fully transferred to the new page.
1747 */
1748 mem_cgroup_end_migration(memcg, page, new_page, true);
1749 spin_unlock(&mm->page_table_lock);
1750
1751 unlock_page(new_page);
1752 unlock_page(page);
1753 put_page(page); /* Drop the rmap reference */
1754 put_page(page); /* Drop the LRU isolation reference */
1755
1756 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1757 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1758
1759out:
1760 mod_zone_page_state(page_zone(page),
1761 NR_ISOLATED_ANON + page_lru,
1762 -HPAGE_PMD_NR);
1763 return isolated;
1764
Hugh Dickins340ef392013-02-22 16:34:33 -08001765out_fail:
1766 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001767out_dropref:
Hugh Dickins340ef392013-02-22 16:34:33 -08001768 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001769 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001770 return 0;
1771}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001772#endif /* CONFIG_NUMA_BALANCING */
1773
1774#endif /* CONFIG_NUMA */