blob: bb940045fe8595842ed58f2e32f87b83d40485e1 [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/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800445 * Gigantic pages are so large that we do not guarantee that page++ pointer
446 * arithmetic will work across the entire page. We need something more
447 * specialized.
448 */
449static void __copy_gigantic_page(struct page *dst, struct page *src,
450 int nr_pages)
451{
452 int i;
453 struct page *dst_base = dst;
454 struct page *src_base = src;
455
456 for (i = 0; i < nr_pages; ) {
457 cond_resched();
458 copy_highpage(dst, src);
459
460 i++;
461 dst = mem_map_next(dst, dst_base, i);
462 src = mem_map_next(src, src_base, i);
463 }
464}
465
466static void copy_huge_page(struct page *dst, struct page *src)
467{
468 int i;
469 int nr_pages;
470
471 if (PageHuge(src)) {
472 /* hugetlbfs page */
473 struct hstate *h = page_hstate(src);
474 nr_pages = pages_per_huge_page(h);
475
476 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
477 __copy_gigantic_page(dst, src, nr_pages);
478 return;
479 }
480 } else {
481 /* thp page */
482 BUG_ON(!PageTransHuge(src));
483 nr_pages = hpage_nr_pages(src);
484 }
485
486 for (i = 0; i < nr_pages; i++) {
487 cond_resched();
488 copy_highpage(dst + i, src + i);
489 }
490}
491
492/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800493 * Copy the page to its new location
494 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900495void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800496{
Rik van Riel7851a452013-10-07 11:29:23 +0100497 int cpupid;
498
Mel Gormanb32967f2012-11-19 12:35:47 +0000499 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900500 copy_huge_page(newpage, page);
501 else
502 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800503
504 if (PageError(page))
505 SetPageError(newpage);
506 if (PageReferenced(page))
507 SetPageReferenced(newpage);
508 if (PageUptodate(page))
509 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700510 if (TestClearPageActive(page)) {
511 VM_BUG_ON(PageUnevictable(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800512 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800513 } else if (TestClearPageUnevictable(page))
514 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800515 if (PageChecked(page))
516 SetPageChecked(newpage);
517 if (PageMappedToDisk(page))
518 SetPageMappedToDisk(newpage);
519
520 if (PageDirty(page)) {
521 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700522 /*
523 * Want to mark the page and the radix tree as dirty, and
524 * redo the accounting that clear_page_dirty_for_io undid,
525 * but we can't use set_page_dirty because that function
526 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300527 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700528 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700529 if (PageSwapBacked(page))
530 SetPageDirty(newpage);
531 else
532 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800533 }
534
Rik van Riel7851a452013-10-07 11:29:23 +0100535 /*
536 * Copy NUMA information to the new page, to prevent over-eager
537 * future migrations of this same page.
538 */
539 cpupid = page_cpupid_xchg_last(page, -1);
540 page_cpupid_xchg_last(newpage, cpupid);
541
Nick Pigginb291f002008-10-18 20:26:44 -0700542 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800543 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800544 /*
545 * Please do not reorder this without considering how mm/ksm.c's
546 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
547 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800548 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800549 ClearPagePrivate(page);
550 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800551
552 /*
553 * If any waiters have accumulated on the new page then
554 * wake them up.
555 */
556 if (PageWriteback(newpage))
557 end_page_writeback(newpage);
558}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800559
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700560/************************************************************
561 * Migration functions
562 ***********************************************************/
563
564/* Always fail migration. Used for mappings that are not movable */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700565int fail_migrate_page(struct address_space *mapping,
566 struct page *newpage, struct page *page)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700567{
568 return -EIO;
569}
570EXPORT_SYMBOL(fail_migrate_page);
571
Christoph Lameterb20a3502006-03-22 00:09:12 -0800572/*
573 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100574 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800575 *
576 * Pages are locked upon entry and exit.
577 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700578int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800579 struct page *newpage, struct page *page,
580 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800581{
582 int rc;
583
584 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
585
Mel Gormana6bc32b2012-01-12 17:19:43 -0800586 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800587
Rafael Aquini78bd5202012-12-11 16:02:31 -0800588 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800589 return rc;
590
591 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800592 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800593}
594EXPORT_SYMBOL(migrate_page);
595
David Howells93614012006-09-30 20:45:40 +0200596#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800597/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700598 * Migration function for pages with buffers. This function can only be used
599 * if the underlying filesystem guarantees that no other references to "page"
600 * exist.
601 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700602int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800603 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700604{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700605 struct buffer_head *bh, *head;
606 int rc;
607
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700608 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800609 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700610
611 head = page_buffers(page);
612
Mel Gormana6bc32b2012-01-12 17:19:43 -0800613 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700614
Rafael Aquini78bd5202012-12-11 16:02:31 -0800615 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700616 return rc;
617
Mel Gormanb969c4a2012-01-12 17:19:34 -0800618 /*
619 * In the async case, migrate_page_move_mapping locked the buffers
620 * with an IRQ-safe spinlock held. In the sync case, the buffers
621 * need to be locked now
622 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800623 if (mode != MIGRATE_ASYNC)
624 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700625
626 ClearPagePrivate(page);
627 set_page_private(newpage, page_private(page));
628 set_page_private(page, 0);
629 put_page(page);
630 get_page(newpage);
631
632 bh = head;
633 do {
634 set_bh_page(bh, newpage, bh_offset(bh));
635 bh = bh->b_this_page;
636
637 } while (bh != head);
638
639 SetPagePrivate(newpage);
640
641 migrate_page_copy(newpage, page);
642
643 bh = head;
644 do {
645 unlock_buffer(bh);
646 put_bh(bh);
647 bh = bh->b_this_page;
648
649 } while (bh != head);
650
Rafael Aquini78bd5202012-12-11 16:02:31 -0800651 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700652}
653EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200654#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700655
Christoph Lameter04e62a22006-06-23 02:03:38 -0700656/*
657 * Writeback a page to clean the dirty state
658 */
659static int writeout(struct address_space *mapping, struct page *page)
660{
661 struct writeback_control wbc = {
662 .sync_mode = WB_SYNC_NONE,
663 .nr_to_write = 1,
664 .range_start = 0,
665 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700666 .for_reclaim = 1
667 };
668 int rc;
669
670 if (!mapping->a_ops->writepage)
671 /* No write method for the address space */
672 return -EINVAL;
673
674 if (!clear_page_dirty_for_io(page))
675 /* Someone else already triggered a write */
676 return -EAGAIN;
677
678 /*
679 * A dirty page may imply that the underlying filesystem has
680 * the page on some queue. So the page must be clean for
681 * migration. Writeout may mean we loose the lock and the
682 * page state is no longer what we checked for earlier.
683 * At this point we know that the migration attempt cannot
684 * be successful.
685 */
686 remove_migration_ptes(page, page);
687
688 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700689
690 if (rc != AOP_WRITEPAGE_ACTIVATE)
691 /* unlocked. Relock */
692 lock_page(page);
693
Hugh Dickinsbda85502008-11-19 15:36:36 -0800694 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700695}
696
697/*
698 * Default handling if a filesystem does not provide a migration function.
699 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700700static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800701 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700702{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800703 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800704 /* Only writeback pages in full synchronous migration */
705 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800706 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700707 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800708 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700709
710 /*
711 * Buffers may be managed in a filesystem specific way.
712 * We must have no buffers or drop them.
713 */
David Howells266cf652009-04-03 16:42:36 +0100714 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700715 !try_to_release_page(page, GFP_KERNEL))
716 return -EAGAIN;
717
Mel Gormana6bc32b2012-01-12 17:19:43 -0800718 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700719}
720
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700721/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700722 * Move a page to a newly allocated page
723 * The page is locked and all ptes have been successfully removed.
724 *
725 * The new page will have replaced the old page if this function
726 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700727 *
728 * Return value:
729 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800730 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700731 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700732static int move_to_new_page(struct page *newpage, struct page *page,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800733 int remap_swapcache, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700734{
735 struct address_space *mapping;
736 int rc;
737
738 /*
739 * Block others from accessing the page when we get around to
740 * establishing additional references. We are the only one
741 * holding a reference to the new page at this point.
742 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200743 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700744 BUG();
745
746 /* Prepare mapping for the new page.*/
747 newpage->index = page->index;
748 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700749 if (PageSwapBacked(page))
750 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700751
752 mapping = page_mapping(page);
753 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800754 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800755 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700756 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800757 * Most pages have a mapping and most filesystems provide a
758 * migratepage callback. Anonymous pages are part of swap
759 * space which also has its own migratepage callback. This
760 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700761 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800762 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800763 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800764 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800765 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700766
Rafael Aquini78bd5202012-12-11 16:02:31 -0800767 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700768 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700769 } else {
770 if (remap_swapcache)
771 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800772 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700773 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700774
775 unlock_page(newpage);
776
777 return rc;
778}
779
Minchan Kim0dabec92011-10-31 17:06:57 -0700780static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800781 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700782{
Minchan Kim0dabec92011-10-31 17:06:57 -0700783 int rc = -EAGAIN;
Mel Gorman3fe20112010-05-24 14:32:20 -0700784 int remap_swapcache = 1;
KAMEZAWA Hiroyuki56039ef2011-03-23 16:42:19 -0700785 struct mem_cgroup *mem;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700786 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700787
Nick Piggin529ae9a2008-08-02 12:01:03 +0200788 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800789 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700790 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800791
792 /*
793 * It's not safe for direct compaction to call lock_page.
794 * For example, during page readahead pages are added locked
795 * to the LRU. Later, when the IO completes the pages are
796 * marked uptodate and unlocked. However, the queueing
797 * could be merging multiple pages for one bio (e.g.
798 * mpage_readpages). If an allocation happens for the
799 * second or third page, the process can end up locking
800 * the same page twice and deadlocking. Rather than
801 * trying to be clever about what pages can be locked,
802 * avoid the use of lock_page for direct compaction
803 * altogether.
804 */
805 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700806 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800807
Christoph Lametere24f0b82006-06-23 02:03:51 -0700808 lock_page(page);
809 }
810
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800811 /* charge against new page */
Johannes Weiner0030f532012-07-31 16:45:25 -0700812 mem_cgroup_prepare_migration(page, newpage, &mem);
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800813
Christoph Lametere24f0b82006-06-23 02:03:51 -0700814 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700815 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700816 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800817 * necessary to wait for PageWriteback. In the async case,
818 * the retry loop is too short and in the sync-light case,
819 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700820 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800821 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700822 rc = -EBUSY;
823 goto uncharge;
824 }
825 if (!force)
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800826 goto uncharge;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700827 wait_on_page_writeback(page);
828 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700829 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700830 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
831 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800832 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700833 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700834 * File Caches may use write_page() or lock_page() in migration, then,
835 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700836 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800837 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800838 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000839 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800840 * getting a hold on an anon_vma from outside one of its mms.
841 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700842 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800843 if (anon_vma) {
844 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700845 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800846 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800847 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700848 /*
849 * We cannot be sure that the anon_vma of an unmapped
850 * swapcache page is safe to use because we don't
851 * know in advance if the VMA that this page belonged
852 * to still exists. If the VMA and others sharing the
853 * data have been freed, then the anon_vma could
854 * already be invalid.
855 *
856 * To avoid this possibility, swapcache pages get
857 * migrated but are not remapped when migration
858 * completes
859 */
860 remap_swapcache = 0;
861 } else {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800862 goto uncharge;
Mel Gorman3fe20112010-05-24 14:32:20 -0700863 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700864 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800865
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800866 if (unlikely(balloon_page_movable(page))) {
867 /*
868 * A ballooned page does not need any special attention from
869 * physical to virtual reverse mapping procedures.
870 * Skip any attempt to unmap PTEs or to remap swap cache,
871 * in order to avoid burning cycles at rmap level, and perform
872 * the page migration right away (proteced by page lock).
873 */
874 rc = balloon_page_migrate(newpage, page, mode);
875 goto uncharge;
876 }
877
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700878 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800879 * Corner case handling:
880 * 1. When a new swap-cache page is read into, it is added to the LRU
881 * and treated as swapcache but it has no rmap yet.
882 * Calling try_to_unmap() against a page->mapping==NULL page will
883 * trigger a BUG. So handle it here.
884 * 2. An orphaned page (see truncate_complete_page) might have
885 * fs-private metadata. The page can be picked up due to memory
886 * offlining. Everywhere else except page reclaim, the page is
887 * invisible to the vm, so the page can not be migrated. So try to
888 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700889 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800890 if (!page->mapping) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800891 VM_BUG_ON(PageAnon(page));
892 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800893 try_to_free_buffers(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800894 goto uncharge;
Shaohua Li62e1c552008-02-04 22:29:33 -0800895 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700896 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800897 }
898
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700899 /* Establish migration ptes or remove ptes */
Andi Kleen14fa31b2009-09-16 11:50:10 +0200900 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700901
Shaohua Liabfc3482009-09-21 17:01:19 -0700902skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700903 if (!page_mapped(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800904 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700905
Mel Gorman3fe20112010-05-24 14:32:20 -0700906 if (rc && remap_swapcache)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700907 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700908
909 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700910 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700911 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700912
KAMEZAWA Hiroyuki01b1ae62009-01-07 18:07:50 -0800913uncharge:
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800914 mem_cgroup_end_migration(mem, page, newpage,
915 (rc == MIGRATEPAGE_SUCCESS ||
916 rc == MIGRATEPAGE_BALLOON_SUCCESS));
Christoph Lametere24f0b82006-06-23 02:03:51 -0700917 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700918out:
919 return rc;
920}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700921
Minchan Kim0dabec92011-10-31 17:06:57 -0700922/*
923 * Obtain the lock on page, remove all ptes and migrate the page
924 * to the newly allocated page in newpage.
925 */
926static int unmap_and_move(new_page_t get_new_page, unsigned long private,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800927 struct page *page, int force, enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700928{
929 int rc = 0;
930 int *result = NULL;
931 struct page *newpage = get_new_page(page, private, &result);
932
933 if (!newpage)
934 return -ENOMEM;
935
936 if (page_count(page) == 1) {
937 /* page was freed from under us. So we are done. */
938 goto out;
939 }
940
941 if (unlikely(PageTransHuge(page)))
942 if (unlikely(split_huge_page(page)))
943 goto out;
944
Hugh Dickins9c620e22013-02-22 16:35:14 -0800945 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800946
947 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
948 /*
949 * A ballooned page has been migrated already.
950 * Now, it's the time to wrap-up counters,
951 * handle the page back to Buddy and return.
952 */
953 dec_zone_page_state(page, NR_ISOLATED_ANON +
954 page_is_file_cache(page));
955 balloon_page_free(page);
956 return MIGRATEPAGE_SUCCESS;
957 }
Minchan Kim0dabec92011-10-31 17:06:57 -0700958out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700959 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700960 /*
961 * A page that has been migrated has all references
962 * removed and will be freed. A page that has not been
963 * migrated will have kepts its references and be
964 * restored.
965 */
966 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700967 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700968 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700969 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700970 }
Christoph Lameter95a402c2006-06-23 02:03:53 -0700971 /*
972 * Move the new page to the LRU. If migration was not successful
973 * then this will free the page.
974 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700975 putback_lru_page(newpage);
Christoph Lameter742755a2006-06-23 02:03:55 -0700976 if (result) {
977 if (rc)
978 *result = rc;
979 else
980 *result = page_to_nid(newpage);
981 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700982 return rc;
983}
984
985/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900986 * Counterpart of unmap_and_move_page() for hugepage migration.
987 *
988 * This function doesn't wait the completion of hugepage I/O
989 * because there is no race between I/O and migration for hugepage.
990 * Note that currently hugepage I/O occurs only in direct I/O
991 * where no lock is held and PG_writeback is irrelevant,
992 * and writeback status of all subpages are counted in the reference
993 * count of the head page (i.e. if all subpages of a 2MB hugepage are
994 * under direct I/O, the reference of the head page is 512 and a bit more.)
995 * This means that when we try to migrate hugepage whose subpages are
996 * doing direct I/O, some references remain after try_to_unmap() and
997 * hugepage migration fails without data corruption.
998 *
999 * There is also no race when direct I/O is issued on the page under migration,
1000 * because then pte is replaced with migration swap entry and direct I/O code
1001 * will wait in the page fault for migration to complete.
1002 */
1003static int unmap_and_move_huge_page(new_page_t get_new_page,
1004 unsigned long private, struct page *hpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001005 int force, enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001006{
1007 int rc = 0;
1008 int *result = NULL;
1009 struct page *new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001010 struct anon_vma *anon_vma = NULL;
1011
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001012 /*
1013 * Movability of hugepages depends on architectures and hugepage size.
1014 * This check is necessary because some callers of hugepage migration
1015 * like soft offline and memory hotremove don't walk through page
1016 * tables or check whether the hugepage is pmd-based or not before
1017 * kicking migration.
1018 */
1019 if (!hugepage_migration_support(page_hstate(hpage)))
1020 return -ENOSYS;
1021
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001022 if (!new_hpage)
1023 return -ENOMEM;
1024
1025 rc = -EAGAIN;
1026
1027 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001028 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001029 goto out;
1030 lock_page(hpage);
1031 }
1032
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001033 if (PageAnon(hpage))
1034 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001035
1036 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1037
1038 if (!page_mapped(hpage))
Mel Gormana6bc32b2012-01-12 17:19:43 -08001039 rc = move_to_new_page(new_hpage, hpage, 1, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001040
1041 if (rc)
1042 remove_migration_ptes(hpage, hpage);
1043
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001044 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001045 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001046
1047 if (!rc)
1048 hugetlb_cgroup_migrate(hpage, new_hpage);
1049
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001050 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001051out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001052 if (rc != -EAGAIN)
1053 putback_active_hugepage(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001054 put_page(new_hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001055 if (result) {
1056 if (rc)
1057 *result = rc;
1058 else
1059 *result = page_to_nid(new_hpage);
1060 }
1061 return rc;
1062}
1063
1064/*
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001065 * migrate_pages - migrate the pages specified in a list, to the free pages
1066 * supplied as the target for the page migration
Christoph Lameterb20a3502006-03-22 00:09:12 -08001067 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001068 * @from: The list of pages to be migrated.
1069 * @get_new_page: The function used to allocate free pages to be used
1070 * as the target of the page migration.
1071 * @private: Private data to be passed on to get_new_page()
1072 * @mode: The migration mode that specifies the constraints for
1073 * page migration, if any.
1074 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001075 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001076 * The function returns after 10 attempts or if no pages are movable any more
1077 * because the list has become empty or no retryable pages exist any more.
1078 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001079 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001080 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001081 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001082 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001083int migrate_pages(struct list_head *from, new_page_t get_new_page,
1084 unsigned long private, enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001085{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001086 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001087 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001088 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001089 int pass = 0;
1090 struct page *page;
1091 struct page *page2;
1092 int swapwrite = current->flags & PF_SWAPWRITE;
1093 int rc;
1094
1095 if (!swapwrite)
1096 current->flags |= PF_SWAPWRITE;
1097
Christoph Lametere24f0b82006-06-23 02:03:51 -07001098 for(pass = 0; pass < 10 && retry; pass++) {
1099 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001100
Christoph Lametere24f0b82006-06-23 02:03:51 -07001101 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001102 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001103
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001104 if (PageHuge(page))
1105 rc = unmap_and_move_huge_page(get_new_page,
1106 private, page, pass > 2, mode);
1107 else
1108 rc = unmap_and_move(get_new_page, private,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001109 page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001110
Christoph Lametere24f0b82006-06-23 02:03:51 -07001111 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001112 case -ENOMEM:
1113 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001114 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001115 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001116 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001117 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001118 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001119 break;
1120 default:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001121 /* Permanent failure */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001122 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001123 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001124 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001125 }
1126 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001127 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001128out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001129 if (nr_succeeded)
1130 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1131 if (nr_failed)
1132 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001133 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1134
Christoph Lameterb20a3502006-03-22 00:09:12 -08001135 if (!swapwrite)
1136 current->flags &= ~PF_SWAPWRITE;
1137
Rafael Aquini78bd5202012-12-11 16:02:31 -08001138 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001139}
1140
Christoph Lameter742755a2006-06-23 02:03:55 -07001141#ifdef CONFIG_NUMA
1142/*
1143 * Move a list of individual pages
1144 */
1145struct page_to_node {
1146 unsigned long addr;
1147 struct page *page;
1148 int node;
1149 int status;
1150};
1151
1152static struct page *new_page_node(struct page *p, unsigned long private,
1153 int **result)
1154{
1155 struct page_to_node *pm = (struct page_to_node *)private;
1156
1157 while (pm->node != MAX_NUMNODES && pm->page != p)
1158 pm++;
1159
1160 if (pm->node == MAX_NUMNODES)
1161 return NULL;
1162
1163 *result = &pm->status;
1164
Naoya Horiguchie632a932013-09-11 14:22:04 -07001165 if (PageHuge(p))
1166 return alloc_huge_page_node(page_hstate(compound_head(p)),
1167 pm->node);
1168 else
1169 return alloc_pages_exact_node(pm->node,
Mel Gorman769848c2007-07-17 04:03:05 -07001170 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001171}
1172
1173/*
1174 * Move a set of pages as indicated in the pm array. The addr
1175 * field must be set to the virtual address of the page to be moved
1176 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001177 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001178 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001179static int do_move_page_to_node_array(struct mm_struct *mm,
1180 struct page_to_node *pm,
1181 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001182{
1183 int err;
1184 struct page_to_node *pp;
1185 LIST_HEAD(pagelist);
1186
1187 down_read(&mm->mmap_sem);
1188
1189 /*
1190 * Build a list of pages to migrate
1191 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001192 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1193 struct vm_area_struct *vma;
1194 struct page *page;
1195
Christoph Lameter742755a2006-06-23 02:03:55 -07001196 err = -EFAULT;
1197 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001198 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001199 goto set_status;
1200
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001201 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001202
1203 err = PTR_ERR(page);
1204 if (IS_ERR(page))
1205 goto set_status;
1206
Christoph Lameter742755a2006-06-23 02:03:55 -07001207 err = -ENOENT;
1208 if (!page)
1209 goto set_status;
1210
Hugh Dickins62b61f62009-12-14 17:59:33 -08001211 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001212 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001213 goto put_and_set;
1214
1215 pp->page = page;
1216 err = page_to_nid(page);
1217
1218 if (err == pp->node)
1219 /*
1220 * Node already in the right place
1221 */
1222 goto put_and_set;
1223
1224 err = -EACCES;
1225 if (page_mapcount(page) > 1 &&
1226 !migrate_all)
1227 goto put_and_set;
1228
Naoya Horiguchie632a932013-09-11 14:22:04 -07001229 if (PageHuge(page)) {
1230 isolate_huge_page(page, &pagelist);
1231 goto put_and_set;
1232 }
1233
Nick Piggin62695a82008-10-18 20:26:09 -07001234 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001235 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001236 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001237 inc_zone_page_state(page, NR_ISOLATED_ANON +
1238 page_is_file_cache(page));
1239 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001240put_and_set:
1241 /*
1242 * Either remove the duplicate refcount from
1243 * isolate_lru_page() or drop the page ref if it was
1244 * not isolated.
1245 */
1246 put_page(page);
1247set_status:
1248 pp->status = err;
1249 }
1250
Brice Gogline78bbfa2008-10-18 20:27:15 -07001251 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001252 if (!list_empty(&pagelist)) {
Christoph Lameter742755a2006-06-23 02:03:55 -07001253 err = migrate_pages(&pagelist, new_page_node,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001254 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001255 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001256 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001257 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001258
1259 up_read(&mm->mmap_sem);
1260 return err;
1261}
1262
1263/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001264 * Migrate an array of page address onto an array of nodes and fill
1265 * the corresponding array of status.
1266 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001267static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001268 unsigned long nr_pages,
1269 const void __user * __user *pages,
1270 const int __user *nodes,
1271 int __user *status, int flags)
1272{
Brice Goglin3140a222009-01-06 14:38:57 -08001273 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001274 unsigned long chunk_nr_pages;
1275 unsigned long chunk_start;
1276 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001277
Brice Goglin3140a222009-01-06 14:38:57 -08001278 err = -ENOMEM;
1279 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1280 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001281 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001282
1283 migrate_prep();
1284
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001285 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001286 * Store a chunk of page_to_node array in a page,
1287 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001288 */
Brice Goglin3140a222009-01-06 14:38:57 -08001289 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001290
Brice Goglin3140a222009-01-06 14:38:57 -08001291 for (chunk_start = 0;
1292 chunk_start < nr_pages;
1293 chunk_start += chunk_nr_pages) {
1294 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001295
Brice Goglin3140a222009-01-06 14:38:57 -08001296 if (chunk_start + chunk_nr_pages > nr_pages)
1297 chunk_nr_pages = nr_pages - chunk_start;
1298
1299 /* fill the chunk pm with addrs and nodes from user-space */
1300 for (j = 0; j < chunk_nr_pages; j++) {
1301 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001302 int node;
1303
Brice Goglin3140a222009-01-06 14:38:57 -08001304 err = -EFAULT;
1305 if (get_user(p, pages + j + chunk_start))
1306 goto out_pm;
1307 pm[j].addr = (unsigned long) p;
1308
1309 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001310 goto out_pm;
1311
1312 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001313 if (node < 0 || node >= MAX_NUMNODES)
1314 goto out_pm;
1315
Lai Jiangshan389162c2012-12-12 13:51:30 -08001316 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001317 goto out_pm;
1318
1319 err = -EACCES;
1320 if (!node_isset(node, task_nodes))
1321 goto out_pm;
1322
Brice Goglin3140a222009-01-06 14:38:57 -08001323 pm[j].node = node;
1324 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001325
Brice Goglin3140a222009-01-06 14:38:57 -08001326 /* End marker for this chunk */
1327 pm[chunk_nr_pages].node = MAX_NUMNODES;
1328
1329 /* Migrate this chunk */
1330 err = do_move_page_to_node_array(mm, pm,
1331 flags & MPOL_MF_MOVE_ALL);
1332 if (err < 0)
1333 goto out_pm;
1334
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001335 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001336 for (j = 0; j < chunk_nr_pages; j++)
1337 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001338 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001339 goto out_pm;
1340 }
1341 }
1342 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001343
1344out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001345 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001346out:
1347 return err;
1348}
1349
1350/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001351 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001352 */
Brice Goglin80bba122008-12-09 13:14:23 -08001353static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1354 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001355{
Brice Goglin2f007e72008-10-18 20:27:16 -07001356 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001357
Christoph Lameter742755a2006-06-23 02:03:55 -07001358 down_read(&mm->mmap_sem);
1359
Brice Goglin2f007e72008-10-18 20:27:16 -07001360 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001361 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001362 struct vm_area_struct *vma;
1363 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001364 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001365
1366 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001367 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001368 goto set_status;
1369
Brice Goglin2f007e72008-10-18 20:27:16 -07001370 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001371
1372 err = PTR_ERR(page);
1373 if (IS_ERR(page))
1374 goto set_status;
1375
Christoph Lameter742755a2006-06-23 02:03:55 -07001376 err = -ENOENT;
1377 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001378 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001379 goto set_status;
1380
1381 err = page_to_nid(page);
1382set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001383 *status = err;
1384
1385 pages++;
1386 status++;
1387 }
1388
1389 up_read(&mm->mmap_sem);
1390}
1391
1392/*
1393 * Determine the nodes of a user array of pages and store it in
1394 * a user array of status.
1395 */
1396static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1397 const void __user * __user *pages,
1398 int __user *status)
1399{
1400#define DO_PAGES_STAT_CHUNK_NR 16
1401 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1402 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001403
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001404 while (nr_pages) {
1405 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001406
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001407 chunk_nr = nr_pages;
1408 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1409 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1410
1411 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1412 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001413
1414 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1415
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001416 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1417 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001418
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001419 pages += chunk_nr;
1420 status += chunk_nr;
1421 nr_pages -= chunk_nr;
1422 }
1423 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001424}
1425
1426/*
1427 * Move a list of pages in the address space of the currently executing
1428 * process.
1429 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001430SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1431 const void __user * __user *, pages,
1432 const int __user *, nodes,
1433 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001434{
David Howellsc69e8d92008-11-14 10:39:19 +11001435 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001436 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001437 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001438 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001439 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001440
1441 /* Check flags */
1442 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1443 return -EINVAL;
1444
1445 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1446 return -EPERM;
1447
1448 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001449 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001450 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001451 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001452 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001453 return -ESRCH;
1454 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001455 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001456
1457 /*
1458 * Check if this process has the right to modify the specified
1459 * process. The right exists if the process has administrative
1460 * capabilities, superuser privileges or the same
1461 * userid as the target process.
1462 */
David Howellsc69e8d92008-11-14 10:39:19 +11001463 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001464 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1465 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001466 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001467 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001468 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001469 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001470 }
David Howellsc69e8d92008-11-14 10:39:19 +11001471 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001472
David Quigley86c3a762006-06-23 02:04:02 -07001473 err = security_task_movememory(task);
1474 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001475 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001476
Christoph Lameter3268c632012-03-21 16:34:06 -07001477 task_nodes = cpuset_mems_allowed(task);
1478 mm = get_task_mm(task);
1479 put_task_struct(task);
1480
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001481 if (!mm)
1482 return -EINVAL;
1483
1484 if (nodes)
1485 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1486 nodes, status, flags);
1487 else
1488 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001489
1490 mmput(mm);
1491 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001492
Christoph Lameter742755a2006-06-23 02:03:55 -07001493out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001494 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001495 return err;
1496}
Christoph Lameter742755a2006-06-23 02:03:55 -07001497
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001498/*
1499 * Call migration functions in the vma_ops that may prepare
1500 * memory in a vm for migration. migration functions may perform
1501 * the migration for vmas that do not have an underlying page struct.
1502 */
1503int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1504 const nodemask_t *from, unsigned long flags)
1505{
1506 struct vm_area_struct *vma;
1507 int err = 0;
1508
Daisuke Nishimura1001c9f2009-02-11 13:04:18 -08001509 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
Christoph Lameter7b2259b2006-06-25 05:46:48 -07001510 if (vma->vm_ops && vma->vm_ops->migrate) {
1511 err = vma->vm_ops->migrate(vma, to, from, flags);
1512 if (err)
1513 break;
1514 }
1515 }
1516 return err;
1517}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001518
1519#ifdef CONFIG_NUMA_BALANCING
1520/*
1521 * Returns true if this is a safe migration target node for misplaced NUMA
1522 * pages. Currently it only checks the watermarks which crude
1523 */
1524static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001525 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001526{
1527 int z;
1528 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1529 struct zone *zone = pgdat->node_zones + z;
1530
1531 if (!populated_zone(zone))
1532 continue;
1533
Lisa Du6e543d52013-09-11 14:22:36 -07001534 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001535 continue;
1536
1537 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1538 if (!zone_watermark_ok(zone, 0,
1539 high_wmark_pages(zone) +
1540 nr_migrate_pages,
1541 0, 0))
1542 continue;
1543 return true;
1544 }
1545 return false;
1546}
1547
1548static struct page *alloc_misplaced_dst_page(struct page *page,
1549 unsigned long data,
1550 int **result)
1551{
1552 int nid = (int) data;
1553 struct page *newpage;
1554
1555 newpage = alloc_pages_exact_node(nid,
1556 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1557 __GFP_NOMEMALLOC | __GFP_NORETRY |
1558 __GFP_NOWARN) &
1559 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001560 if (newpage)
Peter Zijlstra90572892013-10-07 11:29:20 +01001561 page_cpupid_xchg_last(newpage, page_cpupid_last(page));
Hillf Dantonbac03822012-11-27 14:46:24 +00001562
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001563 return newpage;
1564}
1565
1566/*
Mel Gormana8f60772012-11-14 21:41:46 +00001567 * page migration rate limiting control.
1568 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1569 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001570 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1571 * as it is faults that reset the window, pte updates will happen unconditionally
1572 * if there has not been a fault since @pteupdate_interval_millisecs after the
1573 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001574 */
1575static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001576static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001577static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1578
Mel Gormane14808b2012-11-19 10:59:15 +00001579/* Returns true if NUMA migration is currently rate limited */
1580bool migrate_ratelimited(int node)
1581{
1582 pg_data_t *pgdat = NODE_DATA(node);
1583
1584 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1585 msecs_to_jiffies(pteupdate_interval_millisecs)))
1586 return false;
1587
1588 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1589 return false;
1590
1591 return true;
1592}
1593
Mel Gormanb32967f2012-11-19 12:35:47 +00001594/* Returns true if the node is migrate rate-limited after the update */
Mel Gormand28d43352012-11-29 09:24:36 +00001595bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001596{
1597 bool rate_limited = false;
1598
1599 /*
1600 * Rate-limit the amount of data that is being migrated to a node.
1601 * Optimal placement is no good if the memory bus is saturated and
1602 * all the time is being spent migrating!
1603 */
1604 spin_lock(&pgdat->numabalancing_migrate_lock);
1605 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1606 pgdat->numabalancing_migrate_nr_pages = 0;
1607 pgdat->numabalancing_migrate_next_window = jiffies +
1608 msecs_to_jiffies(migrate_interval_millisecs);
1609 }
1610 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1611 rate_limited = true;
1612 else
Mel Gormand28d43352012-11-29 09:24:36 +00001613 pgdat->numabalancing_migrate_nr_pages += nr_pages;
Mel Gormanb32967f2012-11-19 12:35:47 +00001614 spin_unlock(&pgdat->numabalancing_migrate_lock);
1615
1616 return rate_limited;
1617}
1618
1619int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1620{
Hugh Dickins340ef392013-02-22 16:34:33 -08001621 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001622
Mel Gorman3abef4e2013-02-22 16:34:27 -08001623 VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1624
Mel Gormanb32967f2012-11-19 12:35:47 +00001625 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001626 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1627 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001628
Hugh Dickins340ef392013-02-22 16:34:33 -08001629 if (isolate_lru_page(page))
1630 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001631
1632 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001633 * migrate_misplaced_transhuge_page() skips page migration's usual
1634 * check on page_count(), so we must do it here, now that the page
1635 * has been isolated: a GUP pin, or any other pin, prevents migration.
1636 * The expected page count is 3: 1 for page's mapcount and 1 for the
1637 * caller's pin and 1 for the reference taken by isolate_lru_page().
1638 */
1639 if (PageTransHuge(page) && page_count(page) != 3) {
1640 putback_lru_page(page);
1641 return 0;
1642 }
1643
1644 page_lru = page_is_file_cache(page);
1645 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1646 hpage_nr_pages(page));
1647
1648 /*
1649 * Isolating the page has taken another reference, so the
1650 * caller's reference can be safely dropped without the page
1651 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001652 */
1653 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001654 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001655}
1656
Mel Gormana8f60772012-11-14 21:41:46 +00001657/*
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001658 * Attempt to migrate a misplaced page to the specified destination
1659 * node. Caller is expected to have an elevated reference count on
1660 * the page that will be dropped by this function before returning.
1661 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001662int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1663 int node)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001664{
Mel Gormana8f60772012-11-14 21:41:46 +00001665 pg_data_t *pgdat = NODE_DATA(node);
Hugh Dickins340ef392013-02-22 16:34:33 -08001666 int isolated;
Mel Gormanb32967f2012-11-19 12:35:47 +00001667 int nr_remaining;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001668 LIST_HEAD(migratepages);
1669
1670 /*
Mel Gorman1bc115d2013-10-07 11:29:05 +01001671 * Don't migrate file pages that are mapped in multiple processes
1672 * with execute permissions as they are probably shared libraries.
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001673 */
Mel Gorman1bc115d2013-10-07 11:29:05 +01001674 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1675 (vma->vm_flags & VM_EXEC))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001676 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001677
Mel Gormana8f60772012-11-14 21:41:46 +00001678 /*
1679 * 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 */
Hugh Dickins340ef392013-02-22 16:34:33 -08001683 if (numamigrate_update_ratelimit(pgdat, 1))
Mel Gormana8f60772012-11-14 21:41:46 +00001684 goto out;
Mel Gormana8f60772012-11-14 21:41:46 +00001685
Mel Gormanb32967f2012-11-19 12:35:47 +00001686 isolated = numamigrate_isolate_page(pgdat, page);
1687 if (!isolated)
1688 goto out;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001689
Mel Gormanb32967f2012-11-19 12:35:47 +00001690 list_add(&page->lru, &migratepages);
Hugh Dickins9c620e22013-02-22 16:35:14 -08001691 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1692 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001693 if (nr_remaining) {
1694 putback_lru_pages(&migratepages);
1695 isolated = 0;
1696 } else
1697 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001698 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001699 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001700
1701out:
1702 put_page(page);
1703 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001704}
Mel Gorman220018d2012-12-05 09:32:56 +00001705#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001706
Mel Gorman220018d2012-12-05 09:32:56 +00001707#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001708/*
1709 * Migrates a THP to a given target node. page must be locked and is unlocked
1710 * before returning.
1711 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001712int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1713 struct vm_area_struct *vma,
1714 pmd_t *pmd, pmd_t entry,
1715 unsigned long address,
1716 struct page *page, int node)
1717{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001718 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001719 unsigned long haddr = address & HPAGE_PMD_MASK;
1720 pg_data_t *pgdat = NODE_DATA(node);
1721 int isolated = 0;
1722 struct page *new_page = NULL;
1723 struct mem_cgroup *memcg = NULL;
1724 int page_lru = page_is_file_cache(page);
1725
1726 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001727 * Rate-limit the amount of data that is being migrated to a node.
1728 * Optimal placement is no good if the memory bus is saturated and
1729 * all the time is being spent migrating!
1730 */
Mel Gormand28d43352012-11-29 09:24:36 +00001731 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001732 goto out_dropref;
1733
1734 new_page = alloc_pages_node(node,
1735 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001736 if (!new_page)
1737 goto out_fail;
1738
Peter Zijlstra90572892013-10-07 11:29:20 +01001739 page_cpupid_xchg_last(new_page, page_cpupid_last(page));
Mel Gormanb32967f2012-11-19 12:35:47 +00001740
1741 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001742 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001743 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001744 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001745 }
1746
1747 /* Prepare a page as a migration target */
1748 __set_page_locked(new_page);
1749 SetPageSwapBacked(new_page);
1750
1751 /* anon mapping, we can simply copy page->mapping to the new page: */
1752 new_page->mapping = page->mapping;
1753 new_page->index = page->index;
1754 migrate_page_copy(new_page, page);
1755 WARN_ON(PageLRU(new_page));
1756
1757 /* Recheck the target PMD */
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001758 ptl = pmd_lock(mm, pmd);
Mel Gormanb32967f2012-11-19 12:35:47 +00001759 if (unlikely(!pmd_same(*pmd, entry))) {
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001760 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001761
1762 /* Reverse changes made by migrate_page_copy() */
1763 if (TestClearPageActive(new_page))
1764 SetPageActive(page);
1765 if (TestClearPageUnevictable(new_page))
1766 SetPageUnevictable(page);
1767 mlock_migrate_page(page, new_page);
1768
1769 unlock_page(new_page);
1770 put_page(new_page); /* Free it */
1771
Mel Gormana54a4072013-10-07 11:28:46 +01001772 /* Retake the callers reference and putback on LRU */
1773 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001774 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001775 mod_zone_page_state(page_zone(page),
1776 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1777 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001778 }
1779
1780 /*
1781 * Traditional migration needs to prepare the memcg charge
1782 * transaction early to prevent the old page from being
1783 * uncharged when installing migration entries. Here we can
1784 * save the potential rollback and start the charge transfer
1785 * only when migration is already known to end successfully.
1786 */
1787 mem_cgroup_prepare_migration(page, new_page, &memcg);
1788
1789 entry = mk_pmd(new_page, vma->vm_page_prot);
1790 entry = pmd_mknonnuma(entry);
1791 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1792 entry = pmd_mkhuge(entry);
1793
Mel Gormana54a4072013-10-07 11:28:46 +01001794 pmdp_clear_flush(vma, haddr, pmd);
Mel Gormanb32967f2012-11-19 12:35:47 +00001795 set_pmd_at(mm, haddr, pmd, entry);
Mel Gormana54a4072013-10-07 11:28:46 +01001796 page_add_new_anon_rmap(new_page, vma, haddr);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001797 update_mmu_cache_pmd(vma, address, &entry);
Mel Gormanb32967f2012-11-19 12:35:47 +00001798 page_remove_rmap(page);
1799 /*
1800 * Finish the charge transaction under the page table lock to
1801 * prevent split_huge_page() from dividing up the charge
1802 * before it's fully transferred to the new page.
1803 */
1804 mem_cgroup_end_migration(memcg, page, new_page, true);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001805 spin_unlock(ptl);
Mel Gormanb32967f2012-11-19 12:35:47 +00001806
1807 unlock_page(new_page);
1808 unlock_page(page);
1809 put_page(page); /* Drop the rmap reference */
1810 put_page(page); /* Drop the LRU isolation reference */
1811
1812 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1813 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1814
Mel Gormanb32967f2012-11-19 12:35:47 +00001815 mod_zone_page_state(page_zone(page),
1816 NR_ISOLATED_ANON + page_lru,
1817 -HPAGE_PMD_NR);
1818 return isolated;
1819
Hugh Dickins340ef392013-02-22 16:34:33 -08001820out_fail:
1821 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001822out_dropref:
Mel Gormana54a4072013-10-07 11:28:46 +01001823 entry = pmd_mknonnuma(entry);
1824 set_pmd_at(mm, haddr, pmd, entry);
1825 update_mmu_cache_pmd(vma, address, &entry);
1826
Hugh Dickins340ef392013-02-22 16:34:33 -08001827 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001828 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001829 return 0;
1830}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001831#endif /* CONFIG_NUMA_BALANCING */
1832
1833#endif /* CONFIG_NUMA */