blob: 316e720a2023c260b0205cdd2cd6be7d257986ac [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;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800133 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900134 } 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
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800252void migration_entry_wait_huge(struct vm_area_struct *vma,
253 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700254{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800255 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700256 __migration_entry_wait(mm, pte, ptl);
257}
258
Mel Gormanb969c4a2012-01-12 17:19:34 -0800259#ifdef CONFIG_BLOCK
260/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800261static bool buffer_migrate_lock_buffers(struct buffer_head *head,
262 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800263{
264 struct buffer_head *bh = head;
265
266 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800267 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800268 do {
269 get_bh(bh);
270 lock_buffer(bh);
271 bh = bh->b_this_page;
272
273 } while (bh != head);
274
275 return true;
276 }
277
278 /* async case, we cannot block on lock_buffer so use trylock_buffer */
279 do {
280 get_bh(bh);
281 if (!trylock_buffer(bh)) {
282 /*
283 * We failed to lock the buffer and cannot stall in
284 * async migration. Release the taken locks
285 */
286 struct buffer_head *failed_bh = bh;
287 put_bh(failed_bh);
288 bh = head;
289 while (bh != failed_bh) {
290 unlock_buffer(bh);
291 put_bh(bh);
292 bh = bh->b_this_page;
293 }
294 return false;
295 }
296
297 bh = bh->b_this_page;
298 } while (bh != head);
299 return true;
300}
301#else
302static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800303 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800304{
305 return true;
306}
307#endif /* CONFIG_BLOCK */
308
Christoph Lameterb20a3502006-03-22 00:09:12 -0800309/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700310 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700311 *
312 * The number of remaining references must be:
313 * 1 for anonymous pages without a mapping
314 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100315 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800316 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800317int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800318 struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800319 struct buffer_head *head, enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800320{
Peter Zijlstra7039e1d2012-10-25 14:16:34 +0200321 int expected_count = 0;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800322 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800323
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700324 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700325 /* Anonymous page without mapping */
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700326 if (page_count(page) != 1)
327 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800328 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700329 }
330
Nick Piggin19fd6232008-07-25 19:45:32 -0700331 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800332
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800333 pslot = radix_tree_lookup_slot(&mapping->page_tree,
334 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800335
Johannes Weineredcf4742009-09-21 17:02:59 -0700336 expected_count = 2 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700337 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800338 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700339 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700340 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800341 }
342
Nick Piggine2867812008-07-25 19:45:30 -0700343 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700344 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700345 return -EAGAIN;
346 }
347
Christoph Lameterb20a3502006-03-22 00:09:12 -0800348 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800349 * In the async migration case of moving a page with buffers, lock the
350 * buffers using trylock before the mapping is moved. If the mapping
351 * was moved, we later failed to lock the buffers and could not move
352 * the mapping back due to an elevated page count, we would have to
353 * block waiting on other references to be dropped.
354 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800355 if (mode == MIGRATE_ASYNC && head &&
356 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800357 page_unfreeze_refs(page, expected_count);
358 spin_unlock_irq(&mapping->tree_lock);
359 return -EAGAIN;
360 }
361
362 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800363 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800364 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800365 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800366 if (PageSwapCache(page)) {
367 SetPageSwapCache(newpage);
368 set_page_private(newpage, page_private(page));
369 }
370
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800371 radix_tree_replace_slot(pslot, newpage);
372
373 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800374 * Drop cache reference from old page by unfreezing
375 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800376 * We know this isn't the last reference.
377 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800378 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800379
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700380 /*
381 * If moved to a different zone then also account
382 * the page for that zone. Other VM counters will be
383 * taken care of when we establish references to the
384 * new page and drop references to the old page.
385 *
386 * Note that anonymous pages are accounted for
387 * via NR_FILE_PAGES and NR_ANON_PAGES if they
388 * are mapped to swap space.
389 */
390 __dec_zone_page_state(page, NR_FILE_PAGES);
391 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700392 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700393 __dec_zone_page_state(page, NR_SHMEM);
394 __inc_zone_page_state(newpage, NR_SHMEM);
395 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700396 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800397
Rafael Aquini78bd5202012-12-11 16:02:31 -0800398 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800399}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800400
401/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900402 * The expected number of remaining references is the same as that
403 * of migrate_page_move_mapping().
404 */
405int migrate_huge_page_move_mapping(struct address_space *mapping,
406 struct page *newpage, struct page *page)
407{
408 int expected_count;
409 void **pslot;
410
411 if (!mapping) {
412 if (page_count(page) != 1)
413 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800414 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900415 }
416
417 spin_lock_irq(&mapping->tree_lock);
418
419 pslot = radix_tree_lookup_slot(&mapping->page_tree,
420 page_index(page));
421
422 expected_count = 2 + page_has_private(page);
423 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800424 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900425 spin_unlock_irq(&mapping->tree_lock);
426 return -EAGAIN;
427 }
428
429 if (!page_freeze_refs(page, expected_count)) {
430 spin_unlock_irq(&mapping->tree_lock);
431 return -EAGAIN;
432 }
433
434 get_page(newpage);
435
436 radix_tree_replace_slot(pslot, newpage);
437
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800438 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900439
440 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800441 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900442}
443
444/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800445 * Copy the page to its new location
446 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900447void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800448{
Rik van Riel7851a452013-10-07 11:29:23 +0100449 int cpupid;
450
Mel Gormanb32967f2012-11-19 12:35:47 +0000451 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900452 copy_huge_page(newpage, page);
453 else
454 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800455
456 if (PageError(page))
457 SetPageError(newpage);
458 if (PageReferenced(page))
459 SetPageReferenced(newpage);
460 if (PageUptodate(page))
461 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700462 if (TestClearPageActive(page)) {
463 VM_BUG_ON(PageUnevictable(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800464 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800465 } else if (TestClearPageUnevictable(page))
466 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800467 if (PageChecked(page))
468 SetPageChecked(newpage);
469 if (PageMappedToDisk(page))
470 SetPageMappedToDisk(newpage);
471
472 if (PageDirty(page)) {
473 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700474 /*
475 * Want to mark the page and the radix tree as dirty, and
476 * redo the accounting that clear_page_dirty_for_io undid,
477 * but we can't use set_page_dirty because that function
478 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300479 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700480 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700481 if (PageSwapBacked(page))
482 SetPageDirty(newpage);
483 else
484 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800485 }
486
Rik van Riel7851a452013-10-07 11:29:23 +0100487 /*
488 * Copy NUMA information to the new page, to prevent over-eager
489 * future migrations of this same page.
490 */
491 cpupid = page_cpupid_xchg_last(page, -1);
492 page_cpupid_xchg_last(newpage, cpupid);
493
Nick Pigginb291f002008-10-18 20:26:44 -0700494 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800495 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800496 /*
497 * Please do not reorder this without considering how mm/ksm.c's
498 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
499 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800500 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800501 ClearPagePrivate(page);
502 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800503
504 /*
505 * If any waiters have accumulated on the new page then
506 * wake them up.
507 */
508 if (PageWriteback(newpage))
509 end_page_writeback(newpage);
510}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800511
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700512/************************************************************
513 * Migration functions
514 ***********************************************************/
515
516/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700517int fail_migrate_page(struct address_space *mapping,
518 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700519{
520 return -EIO;
521}
522EXPORT_SYMBOL(fail_migrate_page);
523
Christoph Lameterb20a3502006-03-22 00:09:12 -0800524/*
525 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100526 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800527 *
528 * Pages are locked upon entry and exit.
529 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700530int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800531 struct page *newpage, struct page *page,
532 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800533{
534 int rc;
535
536 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
537
Mel Gormana6bc32b2012-01-12 17:19:43 -0800538 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800539
Rafael Aquini78bd5202012-12-11 16:02:31 -0800540 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800541 return rc;
542
543 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800544 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800545}
546EXPORT_SYMBOL(migrate_page);
547
David Howells93614012006-09-30 20:45:40 +0200548#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800549/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700550 * Migration function for pages with buffers. This function can only be used
551 * if the underlying filesystem guarantees that no other references to "page"
552 * exist.
553 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700554int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800555 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700556{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700557 struct buffer_head *bh, *head;
558 int rc;
559
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700560 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800561 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700562
563 head = page_buffers(page);
564
Mel Gormana6bc32b2012-01-12 17:19:43 -0800565 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700566
Rafael Aquini78bd5202012-12-11 16:02:31 -0800567 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700568 return rc;
569
Mel Gormanb969c4a2012-01-12 17:19:34 -0800570 /*
571 * In the async case, migrate_page_move_mapping locked the buffers
572 * with an IRQ-safe spinlock held. In the sync case, the buffers
573 * need to be locked now
574 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800575 if (mode != MIGRATE_ASYNC)
576 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700577
578 ClearPagePrivate(page);
579 set_page_private(newpage, page_private(page));
580 set_page_private(page, 0);
581 put_page(page);
582 get_page(newpage);
583
584 bh = head;
585 do {
586 set_bh_page(bh, newpage, bh_offset(bh));
587 bh = bh->b_this_page;
588
589 } while (bh != head);
590
591 SetPagePrivate(newpage);
592
593 migrate_page_copy(newpage, page);
594
595 bh = head;
596 do {
597 unlock_buffer(bh);
598 put_bh(bh);
599 bh = bh->b_this_page;
600
601 } while (bh != head);
602
Rafael Aquini78bd5202012-12-11 16:02:31 -0800603 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700604}
605EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200606#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700607
Christoph Lameter04e62a22006-06-23 02:03:38 -0700608/*
609 * Writeback a page to clean the dirty state
610 */
611static int writeout(struct address_space *mapping, struct page *page)
612{
613 struct writeback_control wbc = {
614 .sync_mode = WB_SYNC_NONE,
615 .nr_to_write = 1,
616 .range_start = 0,
617 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700618 .for_reclaim = 1
619 };
620 int rc;
621
622 if (!mapping->a_ops->writepage)
623 /* No write method for the address space */
624 return -EINVAL;
625
626 if (!clear_page_dirty_for_io(page))
627 /* Someone else already triggered a write */
628 return -EAGAIN;
629
630 /*
631 * A dirty page may imply that the underlying filesystem has
632 * the page on some queue. So the page must be clean for
633 * migration. Writeout may mean we loose the lock and the
634 * page state is no longer what we checked for earlier.
635 * At this point we know that the migration attempt cannot
636 * be successful.
637 */
638 remove_migration_ptes(page, page);
639
640 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700641
642 if (rc != AOP_WRITEPAGE_ACTIVATE)
643 /* unlocked. Relock */
644 lock_page(page);
645
Hugh Dickinsbda85502008-11-19 15:36:36 -0800646 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700647}
648
649/*
650 * Default handling if a filesystem does not provide a migration function.
651 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700652static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800653 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700654{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800655 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800656 /* Only writeback pages in full synchronous migration */
657 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800658 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700659 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800660 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700661
662 /*
663 * Buffers may be managed in a filesystem specific way.
664 * We must have no buffers or drop them.
665 */
David Howells266cf652009-04-03 16:42:36 +0100666 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700667 !try_to_release_page(page, GFP_KERNEL))
668 return -EAGAIN;
669
Mel Gormana6bc32b2012-01-12 17:19:43 -0800670 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700671}
672
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700673/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700674 * Move a page to a newly allocated page
675 * The page is locked and all ptes have been successfully removed.
676 *
677 * The new page will have replaced the old page if this function
678 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700679 *
680 * Return value:
681 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800682 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700683 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700684static int move_to_new_page(struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800685 int remap_swapcache, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700686{
687 struct address_space *mapping;
688 int rc;
689
690 /*
691 * Block others from accessing the page when we get around to
692 * establishing additional references. We are the only one
693 * holding a reference to the new page at this point.
694 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200695 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700696 BUG();
697
698 /* Prepare mapping for the new page.*/
699 newpage->index = page->index;
700 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700701 if (PageSwapBacked(page))
702 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700703
704 mapping = page_mapping(page);
705 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800706 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800707 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700708 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800709 * Most pages have a mapping and most filesystems provide a
710 * migratepage callback. Anonymous pages are part of swap
711 * space which also has its own migratepage callback. This
712 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700713 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800714 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800715 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800716 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800717 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700718
Rafael Aquini78bd5202012-12-11 16:02:31 -0800719 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700720 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700721 } else {
722 if (remap_swapcache)
723 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800724 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700725 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700726
727 unlock_page(newpage);
728
729 return rc;
730}
731
Minchan Kim0dabec92011-10-31 17:06:57 -0700732static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800733 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700734{
Minchan Kim0dabec92011-10-31 17:06:57 -0700735 int rc = -EAGAIN;
Mel Gorman3fe20112010-05-24 14:32:20 -0700736 int remap_swapcache = 1;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -0700737 struct mem_cgroup *mem;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700738 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700739
Nick Piggin529ae9a2008-08-02 12:01:03 +0200740 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800741 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700742 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800743
744 /*
745 * It's not safe for direct compaction to call lock_page.
746 * For example, during page readahead pages are added locked
747 * to the LRU. Later, when the IO completes the pages are
748 * marked uptodate and unlocked. However, the queueing
749 * could be merging multiple pages for one bio (e.g.
750 * mpage_readpages). If an allocation happens for the
751 * second or third page, the process can end up locking
752 * the same page twice and deadlocking. Rather than
753 * trying to be clever about what pages can be locked,
754 * avoid the use of lock_page for direct compaction
755 * altogether.
756 */
757 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700758 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800759
Christoph Lametere24f0b82006-06-23 02:03:51 -0700760 lock_page(page);
761 }
762
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800763 /* charge against new page */
Johannes Weiner0030f532012-07-31 16:45:25 -0700764 mem_cgroup_prepare_migration(page, newpage, &mem);
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800765
Christoph Lametere24f0b82006-06-23 02:03:51 -0700766 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700767 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700768 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800769 * necessary to wait for PageWriteback. In the async case,
770 * the retry loop is too short and in the sync-light case,
771 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700772 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800773 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700774 rc = -EBUSY;
775 goto uncharge;
776 }
777 if (!force)
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800778 goto uncharge;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700779 wait_on_page_writeback(page);
780 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700781 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700782 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
783 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800784 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700785 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700786 * File Caches may use write_page() or lock_page() in migration, then,
787 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700788 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800789 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800790 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000791 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800792 * getting a hold on an anon_vma from outside one of its mms.
793 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700794 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800795 if (anon_vma) {
796 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700797 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800798 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800799 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700800 /*
801 * We cannot be sure that the anon_vma of an unmapped
802 * swapcache page is safe to use because we don't
803 * know in advance if the VMA that this page belonged
804 * to still exists. If the VMA and others sharing the
805 * data have been freed, then the anon_vma could
806 * already be invalid.
807 *
808 * To avoid this possibility, swapcache pages get
809 * migrated but are not remapped when migration
810 * completes
811 */
812 remap_swapcache = 0;
813 } else {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800814 goto uncharge;
Mel Gorman3fe20112010-05-24 14:32:20 -0700815 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700816 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800817
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800818 if (unlikely(balloon_page_movable(page))) {
819 /*
820 * A ballooned page does not need any special attention from
821 * physical to virtual reverse mapping procedures.
822 * Skip any attempt to unmap PTEs or to remap swap cache,
823 * in order to avoid burning cycles at rmap level, and perform
824 * the page migration right away (proteced by page lock).
825 */
826 rc = balloon_page_migrate(newpage, page, mode);
827 goto uncharge;
828 }
829
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700830 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800831 * Corner case handling:
832 * 1. When a new swap-cache page is read into, it is added to the LRU
833 * and treated as swapcache but it has no rmap yet.
834 * Calling try_to_unmap() against a page->mapping==NULL page will
835 * trigger a BUG. So handle it here.
836 * 2. An orphaned page (see truncate_complete_page) might have
837 * fs-private metadata. The page can be picked up due to memory
838 * offlining. Everywhere else except page reclaim, the page is
839 * invisible to the vm, so the page can not be migrated. So try to
840 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700841 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800842 if (!page->mapping) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800843 VM_BUG_ON(PageAnon(page));
844 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800845 try_to_free_buffers(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800846 goto uncharge;
Shaohua Li62e1c552008-02-04 22:29:33 -0800847 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700848 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800849 }
850
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700851 /* Establish migration ptes or remove ptes */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200852 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700853
Shaohua Liabfc3482009-09-21 17:01:19 -0700854skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700855 if (!page_mapped(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800856 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700857
Mel Gorman3fe20112010-05-24 14:32:20 -0700858 if (rc && remap_swapcache)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700859 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700860
861 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700862 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700863 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700864
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800865uncharge:
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800866 mem_cgroup_end_migration(mem, page, newpage,
867 (rc == MIGRATEPAGE_SUCCESS ||
868 rc == MIGRATEPAGE_BALLOON_SUCCESS));
Christoph Lametere24f0b82006-06-23 02:03:51 -0700869 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700870out:
871 return rc;
872}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700873
Minchan Kim0dabec92011-10-31 17:06:57 -0700874/*
875 * Obtain the lock on page, remove all ptes and migrate the page
876 * to the newly allocated page in newpage.
877 */
878static int unmap_and_move(new_page_t get_new_page, unsigned long private,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800879 struct page *page, int force, enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700880{
881 int rc = 0;
882 int *result = NULL;
883 struct page *newpage = get_new_page(page, private, &result);
884
885 if (!newpage)
886 return -ENOMEM;
887
888 if (page_count(page) == 1) {
889 /* page was freed from under us. So we are done. */
890 goto out;
891 }
892
893 if (unlikely(PageTransHuge(page)))
894 if (unlikely(split_huge_page(page)))
895 goto out;
896
Hugh Dickins9c620e22013-02-22 16:35:14 -0800897 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800898
899 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
900 /*
901 * A ballooned page has been migrated already.
902 * Now, it's the time to wrap-up counters,
903 * handle the page back to Buddy and return.
904 */
905 dec_zone_page_state(page, NR_ISOLATED_ANON +
906 page_is_file_cache(page));
907 balloon_page_free(page);
908 return MIGRATEPAGE_SUCCESS;
909 }
Minchan Kim0dabec92011-10-31 17:06:57 -0700910out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700911 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700912 /*
913 * A page that has been migrated has all references
914 * removed and will be freed. A page that has not been
915 * migrated will have kepts its references and be
916 * restored.
917 */
918 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700919 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700920 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700921 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700922 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700923 /*
924 * Move the new page to the LRU. If migration was not successful
925 * then this will free the page.
926 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700927 putback_lru_page(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700928 if (result) {
929 if (rc)
930 *result = rc;
931 else
932 *result = page_to_nid(newpage);
933 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700934 return rc;
935}
936
937/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900938 * Counterpart of unmap_and_move_page() for hugepage migration.
939 *
940 * This function doesn't wait the completion of hugepage I/O
941 * because there is no race between I/O and migration for hugepage.
942 * Note that currently hugepage I/O occurs only in direct I/O
943 * where no lock is held and PG_writeback is irrelevant,
944 * and writeback status of all subpages are counted in the reference
945 * count of the head page (i.e. if all subpages of a 2MB hugepage are
946 * under direct I/O, the reference of the head page is 512 and a bit more.)
947 * This means that when we try to migrate hugepage whose subpages are
948 * doing direct I/O, some references remain after try_to_unmap() and
949 * hugepage migration fails without data corruption.
950 *
951 * There is also no race when direct I/O is issued on the page under migration,
952 * because then pte is replaced with migration swap entry and direct I/O code
953 * will wait in the page fault for migration to complete.
954 */
955static int unmap_and_move_huge_page(new_page_t get_new_page,
956 unsigned long private, struct page *hpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800957 int force, enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900958{
959 int rc = 0;
960 int *result = NULL;
961 struct page *new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900962 struct anon_vma *anon_vma = NULL;
963
Naoya Horiguchi83467ef2013-09-11 14:22:11 -0700964 /*
965 * Movability of hugepages depends on architectures and hugepage size.
966 * This check is necessary because some callers of hugepage migration
967 * like soft offline and memory hotremove don't walk through page
968 * tables or check whether the hugepage is pmd-based or not before
969 * kicking migration.
970 */
971 if (!hugepage_migration_support(page_hstate(hpage)))
972 return -ENOSYS;
973
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900974 if (!new_hpage)
975 return -ENOMEM;
976
977 rc = -EAGAIN;
978
979 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800980 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900981 goto out;
982 lock_page(hpage);
983 }
984
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700985 if (PageAnon(hpage))
986 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900987
988 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
989
990 if (!page_mapped(hpage))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800991 rc = move_to_new_page(new_hpage, hpage, 1, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900992
993 if (rc)
994 remove_migration_ptes(hpage, hpage);
995
Hugh Dickinsfd4a4662011-01-13 15:47:31 -0800996 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700997 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -0700998
999 if (!rc)
1000 hugetlb_cgroup_migrate(hpage, new_hpage);
1001
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001002 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001003out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001004 if (rc != -EAGAIN)
1005 putback_active_hugepage(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001006 put_page(new_hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001007 if (result) {
1008 if (rc)
1009 *result = rc;
1010 else
1011 *result = page_to_nid(new_hpage);
1012 }
1013 return rc;
1014}
1015
1016/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001017 * migrate_pages - migrate the pages specified in a list, to the free pages
1018 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001019 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001020 * @from: The list of pages to be migrated.
1021 * @get_new_page: The function used to allocate free pages to be used
1022 * as the target of the page migration.
1023 * @private: Private data to be passed on to get_new_page()
1024 * @mode: The migration mode that specifies the constraints for
1025 * page migration, if any.
1026 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001027 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001028 * The function returns after 10 attempts or if no pages are movable any more
1029 * because the list has become empty or no retryable pages exist any more.
1030 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001031 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001032 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001033 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001034 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001035int migrate_pages(struct list_head *from, new_page_t get_new_page,
1036 unsigned long private, enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001037{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001038 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001039 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001040 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001041 int pass = 0;
1042 struct page *page;
1043 struct page *page2;
1044 int swapwrite = current->flags & PF_SWAPWRITE;
1045 int rc;
1046
1047 if (!swapwrite)
1048 current->flags |= PF_SWAPWRITE;
1049
Christoph Lametere24f0b82006-06-23 02:03:51 -07001050 for(pass = 0; pass < 10 && retry; pass++) {
1051 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001052
Christoph Lametere24f0b82006-06-23 02:03:51 -07001053 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001054 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001055
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001056 if (PageHuge(page))
1057 rc = unmap_and_move_huge_page(get_new_page,
1058 private, page, pass > 2, mode);
1059 else
1060 rc = unmap_and_move(get_new_page, private,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001061 page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001062
Christoph Lametere24f0b82006-06-23 02:03:51 -07001063 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001064 case -ENOMEM:
1065 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001066 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001067 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001068 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001069 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001070 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001071 break;
1072 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001073 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001074 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001075 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001076 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001077 }
1078 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001079 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001080out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001081 if (nr_succeeded)
1082 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1083 if (nr_failed)
1084 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001085 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1086
Christoph Lameterb20a3502006-03-22 00:09:12 -08001087 if (!swapwrite)
1088 current->flags &= ~PF_SWAPWRITE;
1089
Rafael Aquini78bd5202012-12-11 16:02:31 -08001090 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001091}
1092
Christoph Lameter742755a2006-06-23 02:03:55 -07001093#ifdef CONFIG_NUMA
1094/*
1095 * Move a list of individual pages
1096 */
1097struct page_to_node {
1098 unsigned long addr;
1099 struct page *page;
1100 int node;
1101 int status;
1102};
1103
1104static struct page *new_page_node(struct page *p, unsigned long private,
1105 int **result)
1106{
1107 struct page_to_node *pm = (struct page_to_node *)private;
1108
1109 while (pm->node != MAX_NUMNODES && pm->page != p)
1110 pm++;
1111
1112 if (pm->node == MAX_NUMNODES)
1113 return NULL;
1114
1115 *result = &pm->status;
1116
Naoya Horiguchie632a932013-09-11 14:22:04 -07001117 if (PageHuge(p))
1118 return alloc_huge_page_node(page_hstate(compound_head(p)),
1119 pm->node);
1120 else
1121 return alloc_pages_exact_node(pm->node,
Mel Gorman769848c2007-07-17 04:03:05 -07001122 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001123}
1124
1125/*
1126 * Move a set of pages as indicated in the pm array. The addr
1127 * field must be set to the virtual address of the page to be moved
1128 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001129 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001130 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001131static int do_move_page_to_node_array(struct mm_struct *mm,
1132 struct page_to_node *pm,
1133 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001134{
1135 int err;
1136 struct page_to_node *pp;
1137 LIST_HEAD(pagelist);
1138
1139 down_read(&mm->mmap_sem);
1140
1141 /*
1142 * Build a list of pages to migrate
1143 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001144 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1145 struct vm_area_struct *vma;
1146 struct page *page;
1147
Christoph Lameter742755a2006-06-23 02:03:55 -07001148 err = -EFAULT;
1149 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001150 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001151 goto set_status;
1152
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001153 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001154
1155 err = PTR_ERR(page);
1156 if (IS_ERR(page))
1157 goto set_status;
1158
Christoph Lameter742755a2006-06-23 02:03:55 -07001159 err = -ENOENT;
1160 if (!page)
1161 goto set_status;
1162
Hugh Dickins62b61f62009-12-14 17:59:33 -08001163 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001164 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001165 goto put_and_set;
1166
1167 pp->page = page;
1168 err = page_to_nid(page);
1169
1170 if (err == pp->node)
1171 /*
1172 * Node already in the right place
1173 */
1174 goto put_and_set;
1175
1176 err = -EACCES;
1177 if (page_mapcount(page) > 1 &&
1178 !migrate_all)
1179 goto put_and_set;
1180
Naoya Horiguchie632a932013-09-11 14:22:04 -07001181 if (PageHuge(page)) {
1182 isolate_huge_page(page, &pagelist);
1183 goto put_and_set;
1184 }
1185
Nick Piggin62695a82008-10-18 20:26:09 -07001186 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001187 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001188 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001189 inc_zone_page_state(page, NR_ISOLATED_ANON +
1190 page_is_file_cache(page));
1191 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001192put_and_set:
1193 /*
1194 * Either remove the duplicate refcount from
1195 * isolate_lru_page() or drop the page ref if it was
1196 * not isolated.
1197 */
1198 put_page(page);
1199set_status:
1200 pp->status = err;
1201 }
1202
Brice Gogline78bbfa2008-10-18 20:27:15 -07001203 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001204 if (!list_empty(&pagelist)) {
Christoph Lameter742755a2006-06-23 02:03:55 -07001205 err = migrate_pages(&pagelist, new_page_node,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001206 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001207 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001208 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001209 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001210
1211 up_read(&mm->mmap_sem);
1212 return err;
1213}
1214
1215/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001216 * Migrate an array of page address onto an array of nodes and fill
1217 * the corresponding array of status.
1218 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001219static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001220 unsigned long nr_pages,
1221 const void __user * __user *pages,
1222 const int __user *nodes,
1223 int __user *status, int flags)
1224{
Brice Goglin3140a222009-01-06 14:38:57 -08001225 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001226 unsigned long chunk_nr_pages;
1227 unsigned long chunk_start;
1228 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001229
Brice Goglin3140a222009-01-06 14:38:57 -08001230 err = -ENOMEM;
1231 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1232 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001233 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001234
1235 migrate_prep();
1236
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001237 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001238 * Store a chunk of page_to_node array in a page,
1239 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001240 */
Brice Goglin3140a222009-01-06 14:38:57 -08001241 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001242
Brice Goglin3140a222009-01-06 14:38:57 -08001243 for (chunk_start = 0;
1244 chunk_start < nr_pages;
1245 chunk_start += chunk_nr_pages) {
1246 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001247
Brice Goglin3140a222009-01-06 14:38:57 -08001248 if (chunk_start + chunk_nr_pages > nr_pages)
1249 chunk_nr_pages = nr_pages - chunk_start;
1250
1251 /* fill the chunk pm with addrs and nodes from user-space */
1252 for (j = 0; j < chunk_nr_pages; j++) {
1253 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001254 int node;
1255
Brice Goglin3140a222009-01-06 14:38:57 -08001256 err = -EFAULT;
1257 if (get_user(p, pages + j + chunk_start))
1258 goto out_pm;
1259 pm[j].addr = (unsigned long) p;
1260
1261 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001262 goto out_pm;
1263
1264 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001265 if (node < 0 || node >= MAX_NUMNODES)
1266 goto out_pm;
1267
Lai Jiangshan389162c2012-12-12 13:51:30 -08001268 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001269 goto out_pm;
1270
1271 err = -EACCES;
1272 if (!node_isset(node, task_nodes))
1273 goto out_pm;
1274
Brice Goglin3140a222009-01-06 14:38:57 -08001275 pm[j].node = node;
1276 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001277
Brice Goglin3140a222009-01-06 14:38:57 -08001278 /* End marker for this chunk */
1279 pm[chunk_nr_pages].node = MAX_NUMNODES;
1280
1281 /* Migrate this chunk */
1282 err = do_move_page_to_node_array(mm, pm,
1283 flags & MPOL_MF_MOVE_ALL);
1284 if (err < 0)
1285 goto out_pm;
1286
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001287 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001288 for (j = 0; j < chunk_nr_pages; j++)
1289 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001290 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001291 goto out_pm;
1292 }
1293 }
1294 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001295
1296out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001297 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001298out:
1299 return err;
1300}
1301
1302/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001303 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001304 */
Brice Goglin80bba122008-12-09 13:14:23 -08001305static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1306 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001307{
Brice Goglin2f007e72008-10-18 20:27:16 -07001308 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001309
Christoph Lameter742755a2006-06-23 02:03:55 -07001310 down_read(&mm->mmap_sem);
1311
Brice Goglin2f007e72008-10-18 20:27:16 -07001312 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001313 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001314 struct vm_area_struct *vma;
1315 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001316 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001317
1318 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001319 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001320 goto set_status;
1321
Brice Goglin2f007e72008-10-18 20:27:16 -07001322 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001323
1324 err = PTR_ERR(page);
1325 if (IS_ERR(page))
1326 goto set_status;
1327
Christoph Lameter742755a2006-06-23 02:03:55 -07001328 err = -ENOENT;
1329 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001330 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001331 goto set_status;
1332
1333 err = page_to_nid(page);
1334set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001335 *status = err;
1336
1337 pages++;
1338 status++;
1339 }
1340
1341 up_read(&mm->mmap_sem);
1342}
1343
1344/*
1345 * Determine the nodes of a user array of pages and store it in
1346 * a user array of status.
1347 */
1348static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1349 const void __user * __user *pages,
1350 int __user *status)
1351{
1352#define DO_PAGES_STAT_CHUNK_NR 16
1353 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1354 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001355
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001356 while (nr_pages) {
1357 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001358
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001359 chunk_nr = nr_pages;
1360 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1361 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1362
1363 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1364 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001365
1366 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1367
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001368 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1369 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001370
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001371 pages += chunk_nr;
1372 status += chunk_nr;
1373 nr_pages -= chunk_nr;
1374 }
1375 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001376}
1377
1378/*
1379 * Move a list of pages in the address space of the currently executing
1380 * process.
1381 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001382SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1383 const void __user * __user *, pages,
1384 const int __user *, nodes,
1385 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001386{
David Howellsc69e8d92008-11-14 10:39:19 +11001387 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001388 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001389 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001390 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001391 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001392
1393 /* Check flags */
1394 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1395 return -EINVAL;
1396
1397 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1398 return -EPERM;
1399
1400 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001401 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001402 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001403 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001404 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001405 return -ESRCH;
1406 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001407 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001408
1409 /*
1410 * Check if this process has the right to modify the specified
1411 * process. The right exists if the process has administrative
1412 * capabilities, superuser privileges or the same
1413 * userid as the target process.
1414 */
David Howellsc69e8d92008-11-14 10:39:19 +11001415 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001416 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1417 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001418 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001419 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001420 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001421 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001422 }
David Howellsc69e8d92008-11-14 10:39:19 +11001423 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001424
David Quigley86c3a762006-06-23 02:04:02 -07001425 err = security_task_movememory(task);
1426 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001427 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001428
Christoph Lameter3268c632012-03-21 16:34:06 -07001429 task_nodes = cpuset_mems_allowed(task);
1430 mm = get_task_mm(task);
1431 put_task_struct(task);
1432
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001433 if (!mm)
1434 return -EINVAL;
1435
1436 if (nodes)
1437 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1438 nodes, status, flags);
1439 else
1440 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001441
1442 mmput(mm);
1443 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001444
Christoph Lameter742755a2006-06-23 02:03:55 -07001445out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001446 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001447 return err;
1448}
Christoph Lameter742755a2006-06-23 02:03:55 -07001449
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001450/*
1451 * Call migration functions in the vma_ops that may prepare
1452 * memory in a vm for migration. migration functions may perform
1453 * the migration for vmas that do not have an underlying page struct.
1454 */
1455int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1456 const nodemask_t *from, unsigned long flags)
1457{
1458 struct vm_area_struct *vma;
1459 int err = 0;
1460
Daisuke Nishimura1001c9f2009-02-11 13:04:18 -08001461 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001462 if (vma->vm_ops && vma->vm_ops->migrate) {
1463 err = vma->vm_ops->migrate(vma, to, from, flags);
1464 if (err)
1465 break;
1466 }
1467 }
1468 return err;
1469}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001470
1471#ifdef CONFIG_NUMA_BALANCING
1472/*
1473 * Returns true if this is a safe migration target node for misplaced NUMA
1474 * pages. Currently it only checks the watermarks which crude
1475 */
1476static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001477 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001478{
1479 int z;
1480 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1481 struct zone *zone = pgdat->node_zones + z;
1482
1483 if (!populated_zone(zone))
1484 continue;
1485
Lisa Du6e543d52013-09-11 14:22:36 -07001486 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001487 continue;
1488
1489 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1490 if (!zone_watermark_ok(zone, 0,
1491 high_wmark_pages(zone) +
1492 nr_migrate_pages,
1493 0, 0))
1494 continue;
1495 return true;
1496 }
1497 return false;
1498}
1499
1500static struct page *alloc_misplaced_dst_page(struct page *page,
1501 unsigned long data,
1502 int **result)
1503{
1504 int nid = (int) data;
1505 struct page *newpage;
1506
1507 newpage = alloc_pages_exact_node(nid,
1508 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1509 __GFP_NOMEMALLOC | __GFP_NORETRY |
1510 __GFP_NOWARN) &
1511 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001512 if (newpage)
Peter Zijlstra90572892013-10-07 11:29:20 +01001513 page_cpupid_xchg_last(newpage, page_cpupid_last(page));
Hillf Dantonbac03822012-11-27 14:46:24 +00001514
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001515 return newpage;
1516}
1517
1518/*
Mel Gormana8f60772012-11-14 21:41:46 +00001519 * page migration rate limiting control.
1520 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1521 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001522 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1523 * as it is faults that reset the window, pte updates will happen unconditionally
1524 * if there has not been a fault since @pteupdate_interval_millisecs after the
1525 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001526 */
1527static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001528static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001529static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1530
Mel Gormane14808b2012-11-19 10:59:15 +00001531/* Returns true if NUMA migration is currently rate limited */
1532bool migrate_ratelimited(int node)
1533{
1534 pg_data_t *pgdat = NODE_DATA(node);
1535
1536 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1537 msecs_to_jiffies(pteupdate_interval_millisecs)))
1538 return false;
1539
1540 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1541 return false;
1542
1543 return true;
1544}
1545
Mel Gormanb32967f2012-11-19 12:35:47 +00001546/* Returns true if the node is migrate rate-limited after the update */
Mel Gormand28d43352012-11-29 09:24:36 +00001547bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001548{
1549 bool rate_limited = false;
1550
1551 /*
1552 * Rate-limit the amount of data that is being migrated to a node.
1553 * Optimal placement is no good if the memory bus is saturated and
1554 * all the time is being spent migrating!
1555 */
1556 spin_lock(&pgdat->numabalancing_migrate_lock);
1557 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1558 pgdat->numabalancing_migrate_nr_pages = 0;
1559 pgdat->numabalancing_migrate_next_window = jiffies +
1560 msecs_to_jiffies(migrate_interval_millisecs);
1561 }
1562 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1563 rate_limited = true;
1564 else
Mel Gormand28d43352012-11-29 09:24:36 +00001565 pgdat->numabalancing_migrate_nr_pages += nr_pages;
Mel Gormanb32967f2012-11-19 12:35:47 +00001566 spin_unlock(&pgdat->numabalancing_migrate_lock);
1567
1568 return rate_limited;
1569}
1570
1571int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1572{
Hugh Dickins340ef392013-02-22 16:34:33 -08001573 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001574
Mel Gorman3abef4e2013-02-22 16:34:27 -08001575 VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1576
Mel Gormanb32967f2012-11-19 12:35:47 +00001577 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001578 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1579 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001580
Hugh Dickins340ef392013-02-22 16:34:33 -08001581 if (isolate_lru_page(page))
1582 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001583
1584 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001585 * migrate_misplaced_transhuge_page() skips page migration's usual
1586 * check on page_count(), so we must do it here, now that the page
1587 * has been isolated: a GUP pin, or any other pin, prevents migration.
1588 * The expected page count is 3: 1 for page's mapcount and 1 for the
1589 * caller's pin and 1 for the reference taken by isolate_lru_page().
1590 */
1591 if (PageTransHuge(page) && page_count(page) != 3) {
1592 putback_lru_page(page);
1593 return 0;
1594 }
1595
1596 page_lru = page_is_file_cache(page);
1597 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1598 hpage_nr_pages(page));
1599
1600 /*
1601 * Isolating the page has taken another reference, so the
1602 * caller's reference can be safely dropped without the page
1603 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001604 */
1605 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001606 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001607}
1608
Mel Gormana8f60772012-11-14 21:41:46 +00001609/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001610 * Attempt to migrate a misplaced page to the specified destination
1611 * node. Caller is expected to have an elevated reference count on
1612 * the page that will be dropped by this function before returning.
1613 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001614int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1615 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001616{
Mel Gormana8f60772012-11-14 21:41:46 +00001617 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001618 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001619 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001620 LIST_HEAD(migratepages);
1621
1622 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001623 * Don't migrate file pages that are mapped in multiple processes
1624 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001625 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001626 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1627 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001628 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001629
Mel Gormana8f60772012-11-14 21:41:46 +00001630 /*
1631 * Rate-limit the amount of data that is being migrated to a node.
1632 * Optimal placement is no good if the memory bus is saturated and
1633 * all the time is being spent migrating!
1634 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001635 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001636 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001637
Mel Gormanb32967f2012-11-19 12:35:47 +00001638 isolated = numamigrate_isolate_page(pgdat, page);
1639 if (!isolated)
1640 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001641
Mel Gormanb32967f2012-11-19 12:35:47 +00001642 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001643 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1644 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001645 if (nr_remaining) {
1646 putback_lru_pages(&migratepages);
1647 isolated = 0;
1648 } else
1649 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001650 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001651 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001652
1653out:
1654 put_page(page);
1655 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001656}
Mel Gorman220018d2012-12-05 09:32:56 +00001657#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001658
Mel Gorman220018d2012-12-05 09:32:56 +00001659#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001660/*
1661 * Migrates a THP to a given target node. page must be locked and is unlocked
1662 * before returning.
1663 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001664int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1665 struct vm_area_struct *vma,
1666 pmd_t *pmd, pmd_t entry,
1667 unsigned long address,
1668 struct page *page, int node)
1669{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001670 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001671 unsigned long haddr = address & HPAGE_PMD_MASK;
1672 pg_data_t *pgdat = NODE_DATA(node);
1673 int isolated = 0;
1674 struct page *new_page = NULL;
1675 struct mem_cgroup *memcg = NULL;
1676 int page_lru = page_is_file_cache(page);
1677
1678 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001679 * Rate-limit the amount of data that is being migrated to a node.
1680 * Optimal placement is no good if the memory bus is saturated and
1681 * all the time is being spent migrating!
1682 */
Mel Gormand28d43352012-11-29 09:24:36 +00001683 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001684 goto out_dropref;
1685
1686 new_page = alloc_pages_node(node,
1687 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001688 if (!new_page)
1689 goto out_fail;
1690
Peter Zijlstra90572892013-10-07 11:29:20 +01001691 page_cpupid_xchg_last(new_page, page_cpupid_last(page));
Mel Gormanb32967f2012-11-19 12:35:47 +00001692
1693 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001694 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001695 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001696 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001697 }
1698
1699 /* Prepare a page as a migration target */
1700 __set_page_locked(new_page);
1701 SetPageSwapBacked(new_page);
1702
1703 /* anon mapping, we can simply copy page->mapping to the new page: */
1704 new_page->mapping = page->mapping;
1705 new_page->index = page->index;
1706 migrate_page_copy(new_page, page);
1707 WARN_ON(PageLRU(new_page));
1708
1709 /* Recheck the target PMD */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001710 ptl = pmd_lock(mm, pmd);
Mel Gormanb32967f2012-11-19 12:35:47 +00001711 if (unlikely(!pmd_same(*pmd, entry))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001712 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001713
1714 /* Reverse changes made by migrate_page_copy() */
1715 if (TestClearPageActive(new_page))
1716 SetPageActive(page);
1717 if (TestClearPageUnevictable(new_page))
1718 SetPageUnevictable(page);
1719 mlock_migrate_page(page, new_page);
1720
1721 unlock_page(new_page);
1722 put_page(new_page); /* Free it */
1723
Mel Gormana54a4072013-10-07 11:28:46 +01001724 /* Retake the callers reference and putback on LRU */
1725 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001726 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001727 mod_zone_page_state(page_zone(page),
1728 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1729 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001730 }
1731
1732 /*
1733 * Traditional migration needs to prepare the memcg charge
1734 * transaction early to prevent the old page from being
1735 * uncharged when installing migration entries. Here we can
1736 * save the potential rollback and start the charge transfer
1737 * only when migration is already known to end successfully.
1738 */
1739 mem_cgroup_prepare_migration(page, new_page, &memcg);
1740
1741 entry = mk_pmd(new_page, vma->vm_page_prot);
1742 entry = pmd_mknonnuma(entry);
1743 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1744 entry = pmd_mkhuge(entry);
1745
Mel Gormana54a4072013-10-07 11:28:46 +01001746 pmdp_clear_flush(vma, haddr, pmd);
Mel Gormanb32967f2012-11-19 12:35:47 +00001747 set_pmd_at(mm, haddr, pmd, entry);
Mel Gormana54a4072013-10-07 11:28:46 +01001748 page_add_new_anon_rmap(new_page, vma, haddr);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001749 update_mmu_cache_pmd(vma, address, &entry);
Mel Gormanb32967f2012-11-19 12:35:47 +00001750 page_remove_rmap(page);
1751 /*
1752 * Finish the charge transaction under the page table lock to
1753 * prevent split_huge_page() from dividing up the charge
1754 * before it's fully transferred to the new page.
1755 */
1756 mem_cgroup_end_migration(memcg, page, new_page, true);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001757 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001758
1759 unlock_page(new_page);
1760 unlock_page(page);
1761 put_page(page); /* Drop the rmap reference */
1762 put_page(page); /* Drop the LRU isolation reference */
1763
1764 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1765 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1766
Mel Gormanb32967f2012-11-19 12:35:47 +00001767 mod_zone_page_state(page_zone(page),
1768 NR_ISOLATED_ANON + page_lru,
1769 -HPAGE_PMD_NR);
1770 return isolated;
1771
Hugh Dickins340ef392013-02-22 16:34:33 -08001772out_fail:
1773 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001774out_dropref:
Mel Gormana54a4072013-10-07 11:28:46 +01001775 entry = pmd_mknonnuma(entry);
1776 set_pmd_at(mm, haddr, pmd, entry);
1777 update_mmu_cache_pmd(vma, address, &entry);
1778
Hugh Dickins340ef392013-02-22 16:34:33 -08001779 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001780 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001781 return 0;
1782}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001783#endif /* CONFIG_NUMA_BALANCING */
1784
1785#endif /* CONFIG_NUMA */