blob: 85e04268603143826e8cd80c533419feb221ad56 [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>
Mel Gormanf714f4f2013-12-18 17:08:33 -080039#include <linux/mmu_notifier.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080040
Michal Nazarewicz0d1836c2010-12-21 17:24:26 -080041#include <asm/tlbflush.h>
42
Mel Gorman7b2a2d42012-10-19 14:07:31 +010043#define CREATE_TRACE_POINTS
44#include <trace/events/migrate.h>
45
Christoph Lameterb20a3502006-03-22 00:09:12 -080046#include "internal.h"
47
Christoph Lameterb20a3502006-03-22 00:09:12 -080048/*
Christoph Lameter742755a2006-06-23 02:03:55 -070049 * migrate_prep() needs to be called before we start compiling a list of pages
Mel Gorman748446b2010-05-24 14:32:27 -070050 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
Christoph Lameterb20a3502006-03-22 00:09:12 -080052 */
53int migrate_prep(void)
54{
Christoph Lameterb20a3502006-03-22 00:09:12 -080055 /*
56 * Clear the LRU lists so pages can be isolated.
57 * Note that pages may be moved off the LRU after we have
58 * drained them. Those pages will fail to migrate like other
59 * pages that may be busy.
60 */
61 lru_add_drain_all();
62
63 return 0;
64}
65
Mel Gorman748446b2010-05-24 14:32:27 -070066/* Do the necessary work of migrate_prep but not if it involves other CPUs */
67int migrate_prep_local(void)
68{
69 lru_add_drain();
70
71 return 0;
72}
73
Christoph Lameterb20a3502006-03-22 00:09:12 -080074/*
Rafael Aquini5733c7d2012-12-11 16:02:47 -080075 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
77 *
Joonsoo Kim59c82b72014-01-21 15:51:17 -080078 * This function shall be used whenever the isolated pageset has been
79 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
80 * and isolate_huge_page().
Rafael Aquini5733c7d2012-12-11 16:02:47 -080081 */
82void putback_movable_pages(struct list_head *l)
83{
84 struct page *page;
85 struct page *page2;
86
87 list_for_each_entry_safe(page, page2, l, lru) {
Naoya Horiguchi31caf662013-09-11 14:21:59 -070088 if (unlikely(PageHuge(page))) {
89 putback_active_hugepage(page);
90 continue;
91 }
Rafael Aquini5733c7d2012-12-11 16:02:47 -080092 list_del(&page->lru);
93 dec_zone_page_state(page, NR_ISOLATED_ANON +
94 page_is_file_cache(page));
Rafael Aquini117aad12013-09-30 13:45:16 -070095 if (unlikely(isolated_balloon_page(page)))
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080096 balloon_page_putback(page);
97 else
98 putback_lru_page(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -080099 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800100}
101
Christoph Lameter06972122006-06-23 02:03:35 -0700102/*
103 * Restore a potential migration pte to a working pte entry
104 */
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800105static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
106 unsigned long addr, void *old)
Christoph Lameter06972122006-06-23 02:03:35 -0700107{
108 struct mm_struct *mm = vma->vm_mm;
109 swp_entry_t entry;
Christoph Lameter06972122006-06-23 02:03:35 -0700110 pmd_t *pmd;
111 pte_t *ptep, pte;
112 spinlock_t *ptl;
113
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900114 if (unlikely(PageHuge(new))) {
115 ptep = huge_pte_offset(mm, addr);
116 if (!ptep)
117 goto out;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800118 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900119 } else {
Bob Liu62190492012-12-11 16:00:37 -0800120 pmd = mm_find_pmd(mm, addr);
121 if (!pmd)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900122 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700123
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900124 ptep = pte_offset_map(pmd, addr);
Christoph Lameter06972122006-06-23 02:03:35 -0700125
Hugh Dickins486cf462011-10-19 12:50:35 -0700126 /*
127 * Peek to check is_swap_pte() before taking ptlock? No, we
128 * can race mremap's move_ptes(), which skips anon_vma lock.
129 */
Christoph Lameter06972122006-06-23 02:03:35 -0700130
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900131 ptl = pte_lockptr(mm, pmd);
132 }
133
Christoph Lameter06972122006-06-23 02:03:35 -0700134 spin_lock(ptl);
135 pte = *ptep;
136 if (!is_swap_pte(pte))
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800137 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700138
139 entry = pte_to_swp_entry(pte);
140
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800141 if (!is_migration_entry(entry) ||
142 migration_entry_to_page(entry) != old)
143 goto unlock;
Christoph Lameter06972122006-06-23 02:03:35 -0700144
Christoph Lameter06972122006-06-23 02:03:35 -0700145 get_page(new);
146 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
Cyrill Gorcunovc3d16e12013-10-16 13:46:51 -0700147 if (pte_swp_soft_dirty(*ptep))
148 pte = pte_mksoft_dirty(pte);
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100149
150 /* Recheck VMA as permissions can change since migration started */
Christoph Lameter06972122006-06-23 02:03:35 -0700151 if (is_write_migration_entry(entry))
Mel Gormand3cb8bf2014-10-02 19:47:41 +0100152 pte = maybe_mkwrite(pte, vma);
153
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200154#ifdef CONFIG_HUGETLB_PAGE
Tony Lube7517d2013-02-04 14:28:46 -0800155 if (PageHuge(new)) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900156 pte = pte_mkhuge(pte);
Tony Lube7517d2013-02-04 14:28:46 -0800157 pte = arch_make_huge_pte(pte, vma, new, 0);
158 }
Andi Kleen3ef8fd72010-10-11 16:03:21 +0200159#endif
Leonid Yegoshinc2cc4992013-05-24 15:55:18 -0700160 flush_dcache_page(new);
Christoph Lameter06972122006-06-23 02:03:35 -0700161 set_pte_at(mm, addr, ptep, pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700162
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900163 if (PageHuge(new)) {
164 if (PageAnon(new))
165 hugepage_add_anon_rmap(new, vma, addr);
166 else
167 page_dup_rmap(new);
168 } else if (PageAnon(new))
Christoph Lameter04e62a22006-06-23 02:03:38 -0700169 page_add_anon_rmap(new, vma, addr);
170 else
171 page_add_file_rmap(new);
172
173 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +0000174 update_mmu_cache(vma, addr, ptep);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800175unlock:
Christoph Lameter06972122006-06-23 02:03:35 -0700176 pte_unmap_unlock(ptep, ptl);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800177out:
178 return SWAP_AGAIN;
Christoph Lameter06972122006-06-23 02:03:35 -0700179}
180
181/*
Christoph Lameter04e62a22006-06-23 02:03:38 -0700182 * Get rid of all migration entries and replace them by
183 * references to the indicated page.
184 */
185static void remove_migration_ptes(struct page *old, struct page *new)
186{
Joonsoo Kim051ac832014-01-21 15:49:48 -0800187 struct rmap_walk_control rwc = {
188 .rmap_one = remove_migration_pte,
189 .arg = old,
190 };
191
192 rmap_walk(new, &rwc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700193}
194
195/*
Christoph Lameter06972122006-06-23 02:03:35 -0700196 * Something used the pte of a page under migration. We need to
197 * get to the page and wait until migration is finished.
198 * When we return from this function the fault will be retried.
Christoph Lameter06972122006-06-23 02:03:35 -0700199 */
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800200void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700201 spinlock_t *ptl)
Christoph Lameter06972122006-06-23 02:03:35 -0700202{
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700203 pte_t pte;
Christoph Lameter06972122006-06-23 02:03:35 -0700204 swp_entry_t entry;
205 struct page *page;
206
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700207 spin_lock(ptl);
Christoph Lameter06972122006-06-23 02:03:35 -0700208 pte = *ptep;
209 if (!is_swap_pte(pte))
210 goto out;
211
212 entry = pte_to_swp_entry(pte);
213 if (!is_migration_entry(entry))
214 goto out;
215
216 page = migration_entry_to_page(entry);
217
Nick Piggine2867812008-07-25 19:45:30 -0700218 /*
219 * Once radix-tree replacement of page migration started, page_count
220 * *must* be zero. And, we don't want to call wait_on_page_locked()
221 * against a page without get_page().
222 * So, we use get_page_unless_zero(), here. Even failed, page fault
223 * will occur again.
224 */
225 if (!get_page_unless_zero(page))
226 goto out;
Christoph Lameter06972122006-06-23 02:03:35 -0700227 pte_unmap_unlock(ptep, ptl);
228 wait_on_page_locked(page);
229 put_page(page);
230 return;
231out:
232 pte_unmap_unlock(ptep, ptl);
233}
234
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700235void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
236 unsigned long address)
237{
238 spinlock_t *ptl = pte_lockptr(mm, pmd);
239 pte_t *ptep = pte_offset_map(pmd, address);
240 __migration_entry_wait(mm, ptep, ptl);
241}
242
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800243void migration_entry_wait_huge(struct vm_area_struct *vma,
244 struct mm_struct *mm, pte_t *pte)
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700245{
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800246 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
Naoya Horiguchi30dad302013-06-12 14:05:04 -0700247 __migration_entry_wait(mm, pte, ptl);
248}
249
Mel Gormanb969c4a2012-01-12 17:19:34 -0800250#ifdef CONFIG_BLOCK
251/* Returns true if all buffers are successfully locked */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800252static bool buffer_migrate_lock_buffers(struct buffer_head *head,
253 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800254{
255 struct buffer_head *bh = head;
256
257 /* Simple case, sync compaction */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800258 if (mode != MIGRATE_ASYNC) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800259 do {
260 get_bh(bh);
261 lock_buffer(bh);
262 bh = bh->b_this_page;
263
264 } while (bh != head);
265
266 return true;
267 }
268
269 /* async case, we cannot block on lock_buffer so use trylock_buffer */
270 do {
271 get_bh(bh);
272 if (!trylock_buffer(bh)) {
273 /*
274 * We failed to lock the buffer and cannot stall in
275 * async migration. Release the taken locks
276 */
277 struct buffer_head *failed_bh = bh;
278 put_bh(failed_bh);
279 bh = head;
280 while (bh != failed_bh) {
281 unlock_buffer(bh);
282 put_bh(bh);
283 bh = bh->b_this_page;
284 }
285 return false;
286 }
287
288 bh = bh->b_this_page;
289 } while (bh != head);
290 return true;
291}
292#else
293static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800294 enum migrate_mode mode)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800295{
296 return true;
297}
298#endif /* CONFIG_BLOCK */
299
Christoph Lameterb20a3502006-03-22 00:09:12 -0800300/*
Christoph Lameterc3fcf8a2006-06-23 02:03:32 -0700301 * Replace the page in the mapping.
Christoph Lameter5b5c7122006-06-23 02:03:29 -0700302 *
303 * The number of remaining references must be:
304 * 1 for anonymous pages without a mapping
305 * 2 for pages with a mapping
David Howells266cf652009-04-03 16:42:36 +0100306 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800307 */
Gu Zheng36bc08c2013-07-16 17:56:16 +0800308int migrate_page_move_mapping(struct address_space *mapping,
Mel Gormanb969c4a2012-01-12 17:19:34 -0800309 struct page *newpage, struct page *page,
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500310 struct buffer_head *head, enum migrate_mode mode,
311 int extra_count)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800312{
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500313 int expected_count = 1 + extra_count;
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800314 void **pslot;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800315
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700316 if (!mapping) {
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700317 /* Anonymous page without mapping */
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500318 if (page_count(page) != expected_count)
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700319 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800320 return MIGRATEPAGE_SUCCESS;
Christoph Lameter6c5240a2006-06-23 02:03:37 -0700321 }
322
Nick Piggin19fd6232008-07-25 19:45:32 -0700323 spin_lock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800324
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800325 pslot = radix_tree_lookup_slot(&mapping->page_tree,
326 page_index(page));
Christoph Lameterb20a3502006-03-22 00:09:12 -0800327
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500328 expected_count += 1 + page_has_private(page);
Nick Piggine2867812008-07-25 19:45:30 -0700329 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800330 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700331 spin_unlock_irq(&mapping->tree_lock);
Christoph Lametere23ca002006-04-10 22:52:57 -0700332 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800333 }
334
Nick Piggine2867812008-07-25 19:45:30 -0700335 if (!page_freeze_refs(page, expected_count)) {
Nick Piggin19fd6232008-07-25 19:45:32 -0700336 spin_unlock_irq(&mapping->tree_lock);
Nick Piggine2867812008-07-25 19:45:30 -0700337 return -EAGAIN;
338 }
339
Christoph Lameterb20a3502006-03-22 00:09:12 -0800340 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800341 * In the async migration case of moving a page with buffers, lock the
342 * buffers using trylock before the mapping is moved. If the mapping
343 * was moved, we later failed to lock the buffers and could not move
344 * the mapping back due to an elevated page count, we would have to
345 * block waiting on other references to be dropped.
346 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800347 if (mode == MIGRATE_ASYNC && head &&
348 !buffer_migrate_lock_buffers(head, mode)) {
Mel Gormanb969c4a2012-01-12 17:19:34 -0800349 page_unfreeze_refs(page, expected_count);
350 spin_unlock_irq(&mapping->tree_lock);
351 return -EAGAIN;
352 }
353
354 /*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800355 * Now we know that no one else is looking at the page.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800356 */
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800357 get_page(newpage); /* add cache reference */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800358 if (PageSwapCache(page)) {
359 SetPageSwapCache(newpage);
360 set_page_private(newpage, page_private(page));
361 }
362
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800363 radix_tree_replace_slot(pslot, newpage);
364
365 /*
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800366 * Drop cache reference from old page by unfreezing
367 * to one less reference.
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800368 * We know this isn't the last reference.
369 */
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800370 page_unfreeze_refs(page, expected_count - 1);
Nick Piggin7cf9c2c2006-12-06 20:33:44 -0800371
Christoph Lameter0e8c7d02007-04-23 14:41:09 -0700372 /*
373 * If moved to a different zone then also account
374 * the page for that zone. Other VM counters will be
375 * taken care of when we establish references to the
376 * new page and drop references to the old page.
377 *
378 * Note that anonymous pages are accounted for
379 * via NR_FILE_PAGES and NR_ANON_PAGES if they
380 * are mapped to swap space.
381 */
382 __dec_zone_page_state(page, NR_FILE_PAGES);
383 __inc_zone_page_state(newpage, NR_FILE_PAGES);
Andrea Arcangeli99a15e22011-06-16 12:56:19 -0700384 if (!PageSwapCache(page) && PageSwapBacked(page)) {
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700385 __dec_zone_page_state(page, NR_SHMEM);
386 __inc_zone_page_state(newpage, NR_SHMEM);
387 }
Nick Piggin19fd6232008-07-25 19:45:32 -0700388 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800389
Rafael Aquini78bd5202012-12-11 16:02:31 -0800390 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800391}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800392
393/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900394 * The expected number of remaining references is the same as that
395 * of migrate_page_move_mapping().
396 */
397int migrate_huge_page_move_mapping(struct address_space *mapping,
398 struct page *newpage, struct page *page)
399{
400 int expected_count;
401 void **pslot;
402
403 if (!mapping) {
404 if (page_count(page) != 1)
405 return -EAGAIN;
Rafael Aquini78bd5202012-12-11 16:02:31 -0800406 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900407 }
408
409 spin_lock_irq(&mapping->tree_lock);
410
411 pslot = radix_tree_lookup_slot(&mapping->page_tree,
412 page_index(page));
413
414 expected_count = 2 + page_has_private(page);
415 if (page_count(page) != expected_count ||
Mel Gorman29c1f672011-01-13 15:47:21 -0800416 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900417 spin_unlock_irq(&mapping->tree_lock);
418 return -EAGAIN;
419 }
420
421 if (!page_freeze_refs(page, expected_count)) {
422 spin_unlock_irq(&mapping->tree_lock);
423 return -EAGAIN;
424 }
425
426 get_page(newpage);
427
428 radix_tree_replace_slot(pslot, newpage);
429
Jacobo Giralt937a94c2012-01-10 15:07:11 -0800430 page_unfreeze_refs(page, expected_count - 1);
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900431
432 spin_unlock_irq(&mapping->tree_lock);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800433 return MIGRATEPAGE_SUCCESS;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900434}
435
436/*
Dave Hansen30b0a102013-11-21 14:31:58 -0800437 * Gigantic pages are so large that we do not guarantee that page++ pointer
438 * arithmetic will work across the entire page. We need something more
439 * specialized.
440 */
441static void __copy_gigantic_page(struct page *dst, struct page *src,
442 int nr_pages)
443{
444 int i;
445 struct page *dst_base = dst;
446 struct page *src_base = src;
447
448 for (i = 0; i < nr_pages; ) {
449 cond_resched();
450 copy_highpage(dst, src);
451
452 i++;
453 dst = mem_map_next(dst, dst_base, i);
454 src = mem_map_next(src, src_base, i);
455 }
456}
457
458static void copy_huge_page(struct page *dst, struct page *src)
459{
460 int i;
461 int nr_pages;
462
463 if (PageHuge(src)) {
464 /* hugetlbfs page */
465 struct hstate *h = page_hstate(src);
466 nr_pages = pages_per_huge_page(h);
467
468 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
469 __copy_gigantic_page(dst, src, nr_pages);
470 return;
471 }
472 } else {
473 /* thp page */
474 BUG_ON(!PageTransHuge(src));
475 nr_pages = hpage_nr_pages(src);
476 }
477
478 for (i = 0; i < nr_pages; i++) {
479 cond_resched();
480 copy_highpage(dst + i, src + i);
481 }
482}
483
484/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800485 * Copy the page to its new location
486 */
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900487void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800488{
Rik van Riel7851a452013-10-07 11:29:23 +0100489 int cpupid;
490
Mel Gormanb32967f2012-11-19 12:35:47 +0000491 if (PageHuge(page) || PageTransHuge(page))
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900492 copy_huge_page(newpage, page);
493 else
494 copy_highpage(newpage, page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800495
496 if (PageError(page))
497 SetPageError(newpage);
498 if (PageReferenced(page))
499 SetPageReferenced(newpage);
500 if (PageUptodate(page))
501 SetPageUptodate(newpage);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700502 if (TestClearPageActive(page)) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800503 VM_BUG_ON_PAGE(PageUnevictable(page), page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800504 SetPageActive(newpage);
Lee Schermerhorn418b27e2009-12-14 17:59:54 -0800505 } else if (TestClearPageUnevictable(page))
506 SetPageUnevictable(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800507 if (PageChecked(page))
508 SetPageChecked(newpage);
509 if (PageMappedToDisk(page))
510 SetPageMappedToDisk(newpage);
511
512 if (PageDirty(page)) {
513 clear_page_dirty_for_io(page);
Nick Piggin3a902c52008-04-30 00:55:16 -0700514 /*
515 * Want to mark the page and the radix tree as dirty, and
516 * redo the accounting that clear_page_dirty_for_io undid,
517 * but we can't use set_page_dirty because that function
518 * is actually a signal that all of the page has become dirty.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300519 * Whereas only part of our page may be dirty.
Nick Piggin3a902c52008-04-30 00:55:16 -0700520 */
Hugh Dickins752dc182012-06-02 00:27:47 -0700521 if (PageSwapBacked(page))
522 SetPageDirty(newpage);
523 else
524 __set_page_dirty_nobuffers(newpage);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800525 }
526
Rik van Riel7851a452013-10-07 11:29:23 +0100527 /*
528 * Copy NUMA information to the new page, to prevent over-eager
529 * future migrations of this same page.
530 */
531 cpupid = page_cpupid_xchg_last(page, -1);
532 page_cpupid_xchg_last(newpage, cpupid);
533
Nick Pigginb291f002008-10-18 20:26:44 -0700534 mlock_migrate_page(newpage, page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -0800535 ksm_migrate_page(newpage, page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800536 /*
537 * Please do not reorder this without considering how mm/ksm.c's
538 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
539 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800540 ClearPageSwapCache(page);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800541 ClearPagePrivate(page);
542 set_page_private(page, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800543
544 /*
545 * If any waiters have accumulated on the new page then
546 * wake them up.
547 */
548 if (PageWriteback(newpage))
549 end_page_writeback(newpage);
550}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800551
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700552/************************************************************
553 * Migration functions
554 ***********************************************************/
555
Christoph Lameterb20a3502006-03-22 00:09:12 -0800556/*
557 * Common logic to directly migrate a single page suitable for
David Howells266cf652009-04-03 16:42:36 +0100558 * pages that do not use PagePrivate/PagePrivate2.
Christoph Lameterb20a3502006-03-22 00:09:12 -0800559 *
560 * Pages are locked upon entry and exit.
561 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700562int migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800563 struct page *newpage, struct page *page,
564 enum migrate_mode mode)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800565{
566 int rc;
567
568 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
569
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500570 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
Christoph Lameterb20a3502006-03-22 00:09:12 -0800571
Rafael Aquini78bd5202012-12-11 16:02:31 -0800572 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800573 return rc;
574
575 migrate_page_copy(newpage, page);
Rafael Aquini78bd5202012-12-11 16:02:31 -0800576 return MIGRATEPAGE_SUCCESS;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800577}
578EXPORT_SYMBOL(migrate_page);
579
David Howells93614012006-09-30 20:45:40 +0200580#ifdef CONFIG_BLOCK
Christoph Lameterb20a3502006-03-22 00:09:12 -0800581/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700582 * Migration function for pages with buffers. This function can only be used
583 * if the underlying filesystem guarantees that no other references to "page"
584 * exist.
585 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -0700586int buffer_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800587 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700588{
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700589 struct buffer_head *bh, *head;
590 int rc;
591
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700592 if (!page_has_buffers(page))
Mel Gormana6bc32b2012-01-12 17:19:43 -0800593 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700594
595 head = page_buffers(page);
596
Benjamin LaHaise8e321fe2013-12-21 17:56:08 -0500597 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700598
Rafael Aquini78bd5202012-12-11 16:02:31 -0800599 if (rc != MIGRATEPAGE_SUCCESS)
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700600 return rc;
601
Mel Gormanb969c4a2012-01-12 17:19:34 -0800602 /*
603 * In the async case, migrate_page_move_mapping locked the buffers
604 * with an IRQ-safe spinlock held. In the sync case, the buffers
605 * need to be locked now
606 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800607 if (mode != MIGRATE_ASYNC)
608 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700609
610 ClearPagePrivate(page);
611 set_page_private(newpage, page_private(page));
612 set_page_private(page, 0);
613 put_page(page);
614 get_page(newpage);
615
616 bh = head;
617 do {
618 set_bh_page(bh, newpage, bh_offset(bh));
619 bh = bh->b_this_page;
620
621 } while (bh != head);
622
623 SetPagePrivate(newpage);
624
625 migrate_page_copy(newpage, page);
626
627 bh = head;
628 do {
629 unlock_buffer(bh);
630 put_bh(bh);
631 bh = bh->b_this_page;
632
633 } while (bh != head);
634
Rafael Aquini78bd5202012-12-11 16:02:31 -0800635 return MIGRATEPAGE_SUCCESS;
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700636}
637EXPORT_SYMBOL(buffer_migrate_page);
David Howells93614012006-09-30 20:45:40 +0200638#endif
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700639
Christoph Lameter04e62a22006-06-23 02:03:38 -0700640/*
641 * Writeback a page to clean the dirty state
642 */
643static int writeout(struct address_space *mapping, struct page *page)
644{
645 struct writeback_control wbc = {
646 .sync_mode = WB_SYNC_NONE,
647 .nr_to_write = 1,
648 .range_start = 0,
649 .range_end = LLONG_MAX,
Christoph Lameter04e62a22006-06-23 02:03:38 -0700650 .for_reclaim = 1
651 };
652 int rc;
653
654 if (!mapping->a_ops->writepage)
655 /* No write method for the address space */
656 return -EINVAL;
657
658 if (!clear_page_dirty_for_io(page))
659 /* Someone else already triggered a write */
660 return -EAGAIN;
661
662 /*
663 * A dirty page may imply that the underlying filesystem has
664 * the page on some queue. So the page must be clean for
665 * migration. Writeout may mean we loose the lock and the
666 * page state is no longer what we checked for earlier.
667 * At this point we know that the migration attempt cannot
668 * be successful.
669 */
670 remove_migration_ptes(page, page);
671
672 rc = mapping->a_ops->writepage(page, &wbc);
Christoph Lameter04e62a22006-06-23 02:03:38 -0700673
674 if (rc != AOP_WRITEPAGE_ACTIVATE)
675 /* unlocked. Relock */
676 lock_page(page);
677
Hugh Dickinsbda85502008-11-19 15:36:36 -0800678 return (rc < 0) ? -EIO : -EAGAIN;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700679}
680
681/*
682 * Default handling if a filesystem does not provide a migration function.
683 */
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700684static int fallback_migrate_page(struct address_space *mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800685 struct page *newpage, struct page *page, enum migrate_mode mode)
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700686{
Mel Gormanb969c4a2012-01-12 17:19:34 -0800687 if (PageDirty(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800688 /* Only writeback pages in full synchronous migration */
689 if (mode != MIGRATE_SYNC)
Mel Gormanb969c4a2012-01-12 17:19:34 -0800690 return -EBUSY;
Christoph Lameter04e62a22006-06-23 02:03:38 -0700691 return writeout(mapping, page);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800692 }
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700693
694 /*
695 * Buffers may be managed in a filesystem specific way.
696 * We must have no buffers or drop them.
697 */
David Howells266cf652009-04-03 16:42:36 +0100698 if (page_has_private(page) &&
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700699 !try_to_release_page(page, GFP_KERNEL))
700 return -EAGAIN;
701
Mel Gormana6bc32b2012-01-12 17:19:43 -0800702 return migrate_page(mapping, newpage, page, mode);
Christoph Lameter8351a6e2006-06-23 02:03:33 -0700703}
704
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700705/*
Christoph Lametere24f0b82006-06-23 02:03:51 -0700706 * Move a page to a newly allocated page
707 * The page is locked and all ptes have been successfully removed.
708 *
709 * The new page will have replaced the old page if this function
710 * is successful.
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700711 *
712 * Return value:
713 * < 0 - error code
Rafael Aquini78bd5202012-12-11 16:02:31 -0800714 * MIGRATEPAGE_SUCCESS - success
Christoph Lametere24f0b82006-06-23 02:03:51 -0700715 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700716static int move_to_new_page(struct page *newpage, struct page *page,
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800717 int page_was_mapped, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700718{
719 struct address_space *mapping;
720 int rc;
721
722 /*
723 * Block others from accessing the page when we get around to
724 * establishing additional references. We are the only one
725 * holding a reference to the new page at this point.
726 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200727 if (!trylock_page(newpage))
Christoph Lametere24f0b82006-06-23 02:03:51 -0700728 BUG();
729
730 /* Prepare mapping for the new page.*/
731 newpage->index = page->index;
732 newpage->mapping = page->mapping;
Rik van Rielb2e18532008-10-18 20:26:30 -0700733 if (PageSwapBacked(page))
734 SetPageSwapBacked(newpage);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700735
736 mapping = page_mapping(page);
737 if (!mapping)
Mel Gormana6bc32b2012-01-12 17:19:43 -0800738 rc = migrate_page(mapping, newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800739 else if (mapping->a_ops->migratepage)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700740 /*
Mel Gormanb969c4a2012-01-12 17:19:34 -0800741 * Most pages have a mapping and most filesystems provide a
742 * migratepage callback. Anonymous pages are part of swap
743 * space which also has its own migratepage callback. This
744 * is the most common path for page migration.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700745 */
Mel Gormanb969c4a2012-01-12 17:19:34 -0800746 rc = mapping->a_ops->migratepage(mapping,
Mel Gormana6bc32b2012-01-12 17:19:43 -0800747 newpage, page, mode);
Mel Gormanb969c4a2012-01-12 17:19:34 -0800748 else
Mel Gormana6bc32b2012-01-12 17:19:43 -0800749 rc = fallback_migrate_page(mapping, newpage, page, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700750
Rafael Aquini78bd5202012-12-11 16:02:31 -0800751 if (rc != MIGRATEPAGE_SUCCESS) {
Christoph Lametere24f0b82006-06-23 02:03:51 -0700752 newpage->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700753 } else {
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700754 mem_cgroup_migrate(page, newpage, false);
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800755 if (page_was_mapped)
Mel Gorman3fe20112010-05-24 14:32:20 -0700756 remove_migration_ptes(page, newpage);
Konstantin Khlebnikov35512ec2012-02-03 15:37:13 -0800757 page->mapping = NULL;
Mel Gorman3fe20112010-05-24 14:32:20 -0700758 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700759
760 unlock_page(newpage);
761
762 return rc;
763}
764
Minchan Kim0dabec92011-10-31 17:06:57 -0700765static int __unmap_and_move(struct page *page, struct page *newpage,
Hugh Dickins9c620e22013-02-22 16:35:14 -0800766 int force, enum migrate_mode mode)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700767{
Minchan Kim0dabec92011-10-31 17:06:57 -0700768 int rc = -EAGAIN;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800769 int page_was_mapped = 0;
Mel Gorman3f6c8272010-05-24 14:32:17 -0700770 struct anon_vma *anon_vma = NULL;
Christoph Lameter95a402c2006-06-23 02:03:53 -0700771
Nick Piggin529ae9a2008-08-02 12:01:03 +0200772 if (!trylock_page(page)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -0800773 if (!force || mode == MIGRATE_ASYNC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700774 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800775
776 /*
777 * It's not safe for direct compaction to call lock_page.
778 * For example, during page readahead pages are added locked
779 * to the LRU. Later, when the IO completes the pages are
780 * marked uptodate and unlocked. However, the queueing
781 * could be merging multiple pages for one bio (e.g.
782 * mpage_readpages). If an allocation happens for the
783 * second or third page, the process can end up locking
784 * the same page twice and deadlocking. Rather than
785 * trying to be clever about what pages can be locked,
786 * avoid the use of lock_page for direct compaction
787 * altogether.
788 */
789 if (current->flags & PF_MEMALLOC)
Minchan Kim0dabec92011-10-31 17:06:57 -0700790 goto out;
Mel Gorman3e7d3442011-01-13 15:45:56 -0800791
Christoph Lametere24f0b82006-06-23 02:03:51 -0700792 lock_page(page);
793 }
794
795 if (PageWriteback(page)) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700796 /*
Jianguo Wufed5b642013-04-29 15:07:58 -0700797 * Only in the case of a full synchronous migration is it
Mel Gormana6bc32b2012-01-12 17:19:43 -0800798 * necessary to wait for PageWriteback. In the async case,
799 * the retry loop is too short and in the sync-light case,
800 * the overhead of stalling is too much
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700801 */
Mel Gormana6bc32b2012-01-12 17:19:43 -0800802 if (mode != MIGRATE_SYNC) {
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700803 rc = -EBUSY;
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700804 goto out_unlock;
Andrea Arcangeli11bc82d2011-03-22 16:33:11 -0700805 }
806 if (!force)
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700807 goto out_unlock;
Christoph Lametere24f0b82006-06-23 02:03:51 -0700808 wait_on_page_writeback(page);
809 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700810 /*
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700811 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
812 * we cannot notice that anon_vma is freed while we migrates a page.
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800813 * This get_anon_vma() delays freeing anon_vma pointer until the end
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700814 * of migration. File cache pages are no problem because of page_lock()
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700815 * File Caches may use write_page() or lock_page() in migration, then,
816 * just care Anon page here.
Christoph Lametere24f0b82006-06-23 02:03:51 -0700817 */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -0800818 if (PageAnon(page) && !PageKsm(page)) {
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800819 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000820 * Only page_lock_anon_vma_read() understands the subtleties of
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800821 * getting a hold on an anon_vma from outside one of its mms.
822 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700823 anon_vma = page_get_anon_vma(page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800824 if (anon_vma) {
825 /*
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700826 * Anon page
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800827 */
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800828 } else if (PageSwapCache(page)) {
Mel Gorman3fe20112010-05-24 14:32:20 -0700829 /*
830 * We cannot be sure that the anon_vma of an unmapped
831 * swapcache page is safe to use because we don't
832 * know in advance if the VMA that this page belonged
833 * to still exists. If the VMA and others sharing the
834 * data have been freed, then the anon_vma could
835 * already be invalid.
836 *
837 * To avoid this possibility, swapcache pages get
838 * migrated but are not remapped when migration
839 * completes
840 */
Mel Gorman3fe20112010-05-24 14:32:20 -0700841 } else {
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700842 goto out_unlock;
Mel Gorman3fe20112010-05-24 14:32:20 -0700843 }
KAMEZAWA Hiroyuki989f89c2007-08-30 23:56:21 -0700844 }
Shaohua Li62e1c552008-02-04 22:29:33 -0800845
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700846 if (unlikely(isolated_balloon_page(page))) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800847 /*
848 * A ballooned page does not need any special attention from
849 * physical to virtual reverse mapping procedures.
850 * Skip any attempt to unmap PTEs or to remap swap cache,
851 * in order to avoid burning cycles at rmap level, and perform
852 * the page migration right away (proteced by page lock).
853 */
854 rc = balloon_page_migrate(newpage, page, mode);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700855 goto out_unlock;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800856 }
857
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700858 /*
Shaohua Li62e1c552008-02-04 22:29:33 -0800859 * Corner case handling:
860 * 1. When a new swap-cache page is read into, it is added to the LRU
861 * and treated as swapcache but it has no rmap yet.
862 * Calling try_to_unmap() against a page->mapping==NULL page will
863 * trigger a BUG. So handle it here.
864 * 2. An orphaned page (see truncate_complete_page) might have
865 * fs-private metadata. The page can be picked up due to memory
866 * offlining. Everywhere else except page reclaim, the page is
867 * invisible to the vm, so the page can not be migrated. So try to
868 * free the metadata, so the page can be freed.
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700869 */
Shaohua Li62e1c552008-02-04 22:29:33 -0800870 if (!page->mapping) {
Sasha Levin309381fea2014-01-23 15:52:54 -0800871 VM_BUG_ON_PAGE(PageAnon(page), page);
Hugh Dickins1ce82b62011-01-13 15:47:30 -0800872 if (page_has_private(page)) {
Shaohua Li62e1c552008-02-04 22:29:33 -0800873 try_to_free_buffers(page);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700874 goto out_unlock;
Shaohua Li62e1c552008-02-04 22:29:33 -0800875 }
Shaohua Liabfc3482009-09-21 17:01:19 -0700876 goto skip_unmap;
Shaohua Li62e1c552008-02-04 22:29:33 -0800877 }
878
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700879 /* Establish migration ptes or remove ptes */
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800880 if (page_mapped(page)) {
881 try_to_unmap(page,
882 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
883 page_was_mapped = 1;
884 }
KAMEZAWA Hiroyukidc386d42007-07-26 10:41:07 -0700885
Shaohua Liabfc3482009-09-21 17:01:19 -0700886skip_unmap:
Christoph Lametere6a15302006-06-25 05:46:49 -0700887 if (!page_mapped(page))
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800888 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700889
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800890 if (rc && page_was_mapped)
Christoph Lametere24f0b82006-06-23 02:03:51 -0700891 remove_migration_ptes(page, page);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700892
893 /* Drop an anon_vma reference if we took one */
Rik van Riel76545062010-08-09 17:18:41 -0700894 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -0700895 put_anon_vma(anon_vma);
Mel Gorman3f6c8272010-05-24 14:32:17 -0700896
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700897out_unlock:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700898 unlock_page(page);
Minchan Kim0dabec92011-10-31 17:06:57 -0700899out:
900 return rc;
901}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700902
Minchan Kim0dabec92011-10-31 17:06:57 -0700903/*
904 * Obtain the lock on page, remove all ptes and migrate the page
905 * to the newly allocated page in newpage.
906 */
David Rientjes68711a72014-06-04 16:08:25 -0700907static int unmap_and_move(new_page_t get_new_page, free_page_t put_new_page,
908 unsigned long private, struct page *page, int force,
909 enum migrate_mode mode)
Minchan Kim0dabec92011-10-31 17:06:57 -0700910{
911 int rc = 0;
912 int *result = NULL;
913 struct page *newpage = get_new_page(page, private, &result);
914
915 if (!newpage)
916 return -ENOMEM;
917
918 if (page_count(page) == 1) {
919 /* page was freed from under us. So we are done. */
920 goto out;
921 }
922
923 if (unlikely(PageTransHuge(page)))
924 if (unlikely(split_huge_page(page)))
925 goto out;
926
Hugh Dickins9c620e22013-02-22 16:35:14 -0800927 rc = __unmap_and_move(page, newpage, force, mode);
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800928
Minchan Kim0dabec92011-10-31 17:06:57 -0700929out:
Christoph Lametere24f0b82006-06-23 02:03:51 -0700930 if (rc != -EAGAIN) {
Minchan Kim0dabec92011-10-31 17:06:57 -0700931 /*
932 * A page that has been migrated has all references
933 * removed and will be freed. A page that has not been
934 * migrated will have kepts its references and be
935 * restored.
936 */
937 list_del(&page->lru);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -0700938 dec_zone_page_state(page, NR_ISOLATED_ANON +
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700939 page_is_file_cache(page));
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700940 putback_lru_page(page);
Christoph Lametere24f0b82006-06-23 02:03:51 -0700941 }
David Rientjes68711a72014-06-04 16:08:25 -0700942
Christoph Lameter95a402c2006-06-23 02:03:53 -0700943 /*
David Rientjes68711a72014-06-04 16:08:25 -0700944 * If migration was not successful and there's a freeing callback, use
945 * it. Otherwise, putback_lru_page() will drop the reference grabbed
946 * during isolation.
Christoph Lameter95a402c2006-06-23 02:03:53 -0700947 */
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700948 if (rc != MIGRATEPAGE_SUCCESS && put_new_page) {
949 ClearPageSwapBacked(newpage);
David Rientjes68711a72014-06-04 16:08:25 -0700950 put_new_page(newpage, private);
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700951 } else if (unlikely(__is_movable_balloon_page(newpage))) {
952 /* drop our reference, page already in the balloon */
953 put_page(newpage);
Hugh Dickins8bdd6382014-07-26 12:58:23 -0700954 } else
David Rientjes68711a72014-06-04 16:08:25 -0700955 putback_lru_page(newpage);
956
Christoph Lameter742755a2006-06-23 02:03:55 -0700957 if (result) {
958 if (rc)
959 *result = rc;
960 else
961 *result = page_to_nid(newpage);
962 }
Christoph Lametere24f0b82006-06-23 02:03:51 -0700963 return rc;
964}
965
966/*
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900967 * Counterpart of unmap_and_move_page() for hugepage migration.
968 *
969 * This function doesn't wait the completion of hugepage I/O
970 * because there is no race between I/O and migration for hugepage.
971 * Note that currently hugepage I/O occurs only in direct I/O
972 * where no lock is held and PG_writeback is irrelevant,
973 * and writeback status of all subpages are counted in the reference
974 * count of the head page (i.e. if all subpages of a 2MB hugepage are
975 * under direct I/O, the reference of the head page is 512 and a bit more.)
976 * This means that when we try to migrate hugepage whose subpages are
977 * doing direct I/O, some references remain after try_to_unmap() and
978 * hugepage migration fails without data corruption.
979 *
980 * There is also no race when direct I/O is issued on the page under migration,
981 * because then pte is replaced with migration swap entry and direct I/O code
982 * will wait in the page fault for migration to complete.
983 */
984static int unmap_and_move_huge_page(new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -0700985 free_page_t put_new_page, unsigned long private,
986 struct page *hpage, int force,
987 enum migrate_mode mode)
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900988{
989 int rc = 0;
990 int *result = NULL;
Hugh Dickins2ebba6b2014-12-12 16:56:19 -0800991 int page_was_mapped = 0;
Joonsoo Kim32665f22014-01-21 15:51:15 -0800992 struct page *new_hpage;
Naoya Horiguchi290408d2010-09-08 10:19:35 +0900993 struct anon_vma *anon_vma = NULL;
994
Naoya Horiguchi83467ef2013-09-11 14:22:11 -0700995 /*
996 * Movability of hugepages depends on architectures and hugepage size.
997 * This check is necessary because some callers of hugepage migration
998 * like soft offline and memory hotremove don't walk through page
999 * tables or check whether the hugepage is pmd-based or not before
1000 * kicking migration.
1001 */
Naoya Horiguchi100873d2014-06-04 16:10:56 -07001002 if (!hugepage_migration_supported(page_hstate(hpage))) {
Joonsoo Kim32665f22014-01-21 15:51:15 -08001003 putback_active_hugepage(hpage);
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001004 return -ENOSYS;
Joonsoo Kim32665f22014-01-21 15:51:15 -08001005 }
Naoya Horiguchi83467ef2013-09-11 14:22:11 -07001006
Joonsoo Kim32665f22014-01-21 15:51:15 -08001007 new_hpage = get_new_page(hpage, private, &result);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001008 if (!new_hpage)
1009 return -ENOMEM;
1010
1011 rc = -EAGAIN;
1012
1013 if (!trylock_page(hpage)) {
Mel Gormana6bc32b2012-01-12 17:19:43 -08001014 if (!force || mode != MIGRATE_SYNC)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001015 goto out;
1016 lock_page(hpage);
1017 }
1018
Peter Zijlstra746b18d2011-05-24 17:12:10 -07001019 if (PageAnon(hpage))
1020 anon_vma = page_get_anon_vma(hpage);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001021
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001022 if (page_mapped(hpage)) {
1023 try_to_unmap(hpage,
1024 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1025 page_was_mapped = 1;
1026 }
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001027
1028 if (!page_mapped(hpage))
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001029 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001030
Hugh Dickins2ebba6b2014-12-12 16:56:19 -08001031 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001032 remove_migration_ptes(hpage, hpage);
1033
Hugh Dickinsfd4a4662011-01-13 15:47:31 -08001034 if (anon_vma)
Peter Zijlstra9e601092011-03-22 16:32:46 -07001035 put_anon_vma(anon_vma);
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001036
David Rientjes68711a72014-06-04 16:08:25 -07001037 if (rc == MIGRATEPAGE_SUCCESS)
Aneesh Kumar K.V8e6ac7f2012-07-31 16:42:27 -07001038 hugetlb_cgroup_migrate(hpage, new_hpage);
1039
Naoya Horiguchi290408d2010-09-08 10:19:35 +09001040 unlock_page(hpage);
Hillf Danton09761332011-12-08 14:34:20 -08001041out:
Naoya Horiguchib8ec1ce2013-09-11 14:22:01 -07001042 if (rc != -EAGAIN)
1043 putback_active_hugepage(hpage);
David Rientjes68711a72014-06-04 16:08:25 -07001044
1045 /*
1046 * If migration was not successful and there's a freeing callback, use
1047 * it. Otherwise, put_page() will drop the reference grabbed during
1048 * isolation.
1049 */
1050 if (rc != MIGRATEPAGE_SUCCESS && put_new_page)
1051 put_new_page(new_hpage, private);
1052 else
1053 put_page(new_hpage);
1054
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.
David Rientjes68711a72014-06-04 16:08:25 -07001071 * @put_new_page: The function used to free target pages if migration
1072 * fails, or NULL if no special handling is necessary.
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001073 * @private: Private data to be passed on to get_new_page()
1074 * @mode: The migration mode that specifies the constraints for
1075 * page migration, if any.
1076 * @reason: The reason for page migration.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001077 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001078 * The function returns after 10 attempts or if no pages are movable any more
1079 * because the list has become empty or no retryable pages exist any more.
1080 * The caller should call putback_lru_pages() to return pages to the LRU
Minchan Kim28bd6572011-01-25 15:07:26 -08001081 * or free list only if ret != 0.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001082 *
Srivatsa S. Bhatc73e5c92013-04-29 15:08:16 -07001083 * Returns the number of pages that were not migrated, or an error code.
Christoph Lameterb20a3502006-03-22 00:09:12 -08001084 */
Hugh Dickins9c620e22013-02-22 16:35:14 -08001085int migrate_pages(struct list_head *from, new_page_t get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001086 free_page_t put_new_page, unsigned long private,
1087 enum migrate_mode mode, int reason)
Christoph Lameterb20a3502006-03-22 00:09:12 -08001088{
Christoph Lametere24f0b82006-06-23 02:03:51 -07001089 int retry = 1;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001090 int nr_failed = 0;
Mel Gorman5647bc22012-10-19 10:46:20 +01001091 int nr_succeeded = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001092 int pass = 0;
1093 struct page *page;
1094 struct page *page2;
1095 int swapwrite = current->flags & PF_SWAPWRITE;
1096 int rc;
1097
1098 if (!swapwrite)
1099 current->flags |= PF_SWAPWRITE;
1100
Christoph Lametere24f0b82006-06-23 02:03:51 -07001101 for(pass = 0; pass < 10 && retry; pass++) {
1102 retry = 0;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001103
Christoph Lametere24f0b82006-06-23 02:03:51 -07001104 list_for_each_entry_safe(page, page2, from, lru) {
Christoph Lametere24f0b82006-06-23 02:03:51 -07001105 cond_resched();
Christoph Lameterb20a3502006-03-22 00:09:12 -08001106
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001107 if (PageHuge(page))
1108 rc = unmap_and_move_huge_page(get_new_page,
David Rientjes68711a72014-06-04 16:08:25 -07001109 put_new_page, private, page,
1110 pass > 2, mode);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07001111 else
David Rientjes68711a72014-06-04 16:08:25 -07001112 rc = unmap_and_move(get_new_page, put_new_page,
1113 private, page, pass > 2, mode);
Christoph Lameterb20a3502006-03-22 00:09:12 -08001114
Christoph Lametere24f0b82006-06-23 02:03:51 -07001115 switch(rc) {
Christoph Lameter95a402c2006-06-23 02:03:53 -07001116 case -ENOMEM:
1117 goto out;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001118 case -EAGAIN:
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001119 retry++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001120 break;
Rafael Aquini78bd5202012-12-11 16:02:31 -08001121 case MIGRATEPAGE_SUCCESS:
Mel Gorman5647bc22012-10-19 10:46:20 +01001122 nr_succeeded++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001123 break;
1124 default:
Naoya Horiguchi354a3362014-01-21 15:51:14 -08001125 /*
1126 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1127 * unlike -EAGAIN case, the failed page is
1128 * removed from migration page list and not
1129 * retried in the next outer loop.
1130 */
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001131 nr_failed++;
Christoph Lametere24f0b82006-06-23 02:03:51 -07001132 break;
Christoph Lameter2d1db3b2006-06-23 02:03:33 -07001133 }
Christoph Lameterb20a3502006-03-22 00:09:12 -08001134 }
1135 }
Rafael Aquini78bd5202012-12-11 16:02:31 -08001136 rc = nr_failed + retry;
Christoph Lameter95a402c2006-06-23 02:03:53 -07001137out:
Mel Gorman5647bc22012-10-19 10:46:20 +01001138 if (nr_succeeded)
1139 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1140 if (nr_failed)
1141 count_vm_events(PGMIGRATE_FAIL, nr_failed);
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001142 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1143
Christoph Lameterb20a3502006-03-22 00:09:12 -08001144 if (!swapwrite)
1145 current->flags &= ~PF_SWAPWRITE;
1146
Rafael Aquini78bd5202012-12-11 16:02:31 -08001147 return rc;
Christoph Lameterb20a3502006-03-22 00:09:12 -08001148}
1149
Christoph Lameter742755a2006-06-23 02:03:55 -07001150#ifdef CONFIG_NUMA
1151/*
1152 * Move a list of individual pages
1153 */
1154struct page_to_node {
1155 unsigned long addr;
1156 struct page *page;
1157 int node;
1158 int status;
1159};
1160
1161static struct page *new_page_node(struct page *p, unsigned long private,
1162 int **result)
1163{
1164 struct page_to_node *pm = (struct page_to_node *)private;
1165
1166 while (pm->node != MAX_NUMNODES && pm->page != p)
1167 pm++;
1168
1169 if (pm->node == MAX_NUMNODES)
1170 return NULL;
1171
1172 *result = &pm->status;
1173
Naoya Horiguchie632a932013-09-11 14:22:04 -07001174 if (PageHuge(p))
1175 return alloc_huge_page_node(page_hstate(compound_head(p)),
1176 pm->node);
1177 else
1178 return alloc_pages_exact_node(pm->node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001179 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
Christoph Lameter742755a2006-06-23 02:03:55 -07001180}
1181
1182/*
1183 * Move a set of pages as indicated in the pm array. The addr
1184 * field must be set to the virtual address of the page to be moved
1185 * and the node number must contain a valid target node.
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001186 * The pm array ends with node = MAX_NUMNODES.
Christoph Lameter742755a2006-06-23 02:03:55 -07001187 */
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001188static int do_move_page_to_node_array(struct mm_struct *mm,
1189 struct page_to_node *pm,
1190 int migrate_all)
Christoph Lameter742755a2006-06-23 02:03:55 -07001191{
1192 int err;
1193 struct page_to_node *pp;
1194 LIST_HEAD(pagelist);
1195
1196 down_read(&mm->mmap_sem);
1197
1198 /*
1199 * Build a list of pages to migrate
1200 */
Christoph Lameter742755a2006-06-23 02:03:55 -07001201 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1202 struct vm_area_struct *vma;
1203 struct page *page;
1204
Christoph Lameter742755a2006-06-23 02:03:55 -07001205 err = -EFAULT;
1206 vma = find_vma(mm, pp->addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001207 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
Christoph Lameter742755a2006-06-23 02:03:55 -07001208 goto set_status;
1209
Andrea Arcangeli500d65d2011-01-13 15:46:55 -08001210 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001211
1212 err = PTR_ERR(page);
1213 if (IS_ERR(page))
1214 goto set_status;
1215
Christoph Lameter742755a2006-06-23 02:03:55 -07001216 err = -ENOENT;
1217 if (!page)
1218 goto set_status;
1219
Hugh Dickins62b61f62009-12-14 17:59:33 -08001220 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001221 if (PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001222 goto put_and_set;
1223
1224 pp->page = page;
1225 err = page_to_nid(page);
1226
1227 if (err == pp->node)
1228 /*
1229 * Node already in the right place
1230 */
1231 goto put_and_set;
1232
1233 err = -EACCES;
1234 if (page_mapcount(page) > 1 &&
1235 !migrate_all)
1236 goto put_and_set;
1237
Naoya Horiguchie632a932013-09-11 14:22:04 -07001238 if (PageHuge(page)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08001239 if (PageHead(page))
1240 isolate_huge_page(page, &pagelist);
Naoya Horiguchie632a932013-09-11 14:22:04 -07001241 goto put_and_set;
1242 }
1243
Nick Piggin62695a82008-10-18 20:26:09 -07001244 err = isolate_lru_page(page);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001245 if (!err) {
Nick Piggin62695a82008-10-18 20:26:09 -07001246 list_add_tail(&page->lru, &pagelist);
KOSAKI Motohiro6d9c2852009-12-14 17:58:11 -08001247 inc_zone_page_state(page, NR_ISOLATED_ANON +
1248 page_is_file_cache(page));
1249 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001250put_and_set:
1251 /*
1252 * Either remove the duplicate refcount from
1253 * isolate_lru_page() or drop the page ref if it was
1254 * not isolated.
1255 */
1256 put_page(page);
1257set_status:
1258 pp->status = err;
1259 }
1260
Brice Gogline78bbfa2008-10-18 20:27:15 -07001261 err = 0;
Minchan Kimcf608ac2010-10-26 14:21:29 -07001262 if (!list_empty(&pagelist)) {
David Rientjes68711a72014-06-04 16:08:25 -07001263 err = migrate_pages(&pagelist, new_page_node, NULL,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001264 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001265 if (err)
Naoya Horiguchie632a932013-09-11 14:22:04 -07001266 putback_movable_pages(&pagelist);
Minchan Kimcf608ac2010-10-26 14:21:29 -07001267 }
Christoph Lameter742755a2006-06-23 02:03:55 -07001268
1269 up_read(&mm->mmap_sem);
1270 return err;
1271}
1272
1273/*
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001274 * Migrate an array of page address onto an array of nodes and fill
1275 * the corresponding array of status.
1276 */
Christoph Lameter3268c632012-03-21 16:34:06 -07001277static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001278 unsigned long nr_pages,
1279 const void __user * __user *pages,
1280 const int __user *nodes,
1281 int __user *status, int flags)
1282{
Brice Goglin3140a222009-01-06 14:38:57 -08001283 struct page_to_node *pm;
Brice Goglin3140a222009-01-06 14:38:57 -08001284 unsigned long chunk_nr_pages;
1285 unsigned long chunk_start;
1286 int err;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001287
Brice Goglin3140a222009-01-06 14:38:57 -08001288 err = -ENOMEM;
1289 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1290 if (!pm)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001291 goto out;
Brice Goglin35282a22009-06-16 15:32:43 -07001292
1293 migrate_prep();
1294
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001295 /*
Brice Goglin3140a222009-01-06 14:38:57 -08001296 * Store a chunk of page_to_node array in a page,
1297 * but keep the last one as a marker
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001298 */
Brice Goglin3140a222009-01-06 14:38:57 -08001299 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001300
Brice Goglin3140a222009-01-06 14:38:57 -08001301 for (chunk_start = 0;
1302 chunk_start < nr_pages;
1303 chunk_start += chunk_nr_pages) {
1304 int j;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001305
Brice Goglin3140a222009-01-06 14:38:57 -08001306 if (chunk_start + chunk_nr_pages > nr_pages)
1307 chunk_nr_pages = nr_pages - chunk_start;
1308
1309 /* fill the chunk pm with addrs and nodes from user-space */
1310 for (j = 0; j < chunk_nr_pages; j++) {
1311 const void __user *p;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001312 int node;
1313
Brice Goglin3140a222009-01-06 14:38:57 -08001314 err = -EFAULT;
1315 if (get_user(p, pages + j + chunk_start))
1316 goto out_pm;
1317 pm[j].addr = (unsigned long) p;
1318
1319 if (get_user(node, nodes + j + chunk_start))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001320 goto out_pm;
1321
1322 err = -ENODEV;
Linus Torvalds6f5a55f2010-02-05 16:16:50 -08001323 if (node < 0 || node >= MAX_NUMNODES)
1324 goto out_pm;
1325
Lai Jiangshan389162c2012-12-12 13:51:30 -08001326 if (!node_state(node, N_MEMORY))
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001327 goto out_pm;
1328
1329 err = -EACCES;
1330 if (!node_isset(node, task_nodes))
1331 goto out_pm;
1332
Brice Goglin3140a222009-01-06 14:38:57 -08001333 pm[j].node = node;
1334 }
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001335
Brice Goglin3140a222009-01-06 14:38:57 -08001336 /* End marker for this chunk */
1337 pm[chunk_nr_pages].node = MAX_NUMNODES;
1338
1339 /* Migrate this chunk */
1340 err = do_move_page_to_node_array(mm, pm,
1341 flags & MPOL_MF_MOVE_ALL);
1342 if (err < 0)
1343 goto out_pm;
1344
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001345 /* Return status information */
Brice Goglin3140a222009-01-06 14:38:57 -08001346 for (j = 0; j < chunk_nr_pages; j++)
1347 if (put_user(pm[j].status, status + j + chunk_start)) {
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001348 err = -EFAULT;
Brice Goglin3140a222009-01-06 14:38:57 -08001349 goto out_pm;
1350 }
1351 }
1352 err = 0;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001353
1354out_pm:
Brice Goglin3140a222009-01-06 14:38:57 -08001355 free_page((unsigned long)pm);
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001356out:
1357 return err;
1358}
1359
1360/*
Brice Goglin2f007e72008-10-18 20:27:16 -07001361 * Determine the nodes of an array of pages and store it in an array of status.
Christoph Lameter742755a2006-06-23 02:03:55 -07001362 */
Brice Goglin80bba122008-12-09 13:14:23 -08001363static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1364 const void __user **pages, int *status)
Christoph Lameter742755a2006-06-23 02:03:55 -07001365{
Brice Goglin2f007e72008-10-18 20:27:16 -07001366 unsigned long i;
Brice Goglin2f007e72008-10-18 20:27:16 -07001367
Christoph Lameter742755a2006-06-23 02:03:55 -07001368 down_read(&mm->mmap_sem);
1369
Brice Goglin2f007e72008-10-18 20:27:16 -07001370 for (i = 0; i < nr_pages; i++) {
Brice Goglin80bba122008-12-09 13:14:23 -08001371 unsigned long addr = (unsigned long)(*pages);
Christoph Lameter742755a2006-06-23 02:03:55 -07001372 struct vm_area_struct *vma;
1373 struct page *page;
KOSAKI Motohiroc095adb2008-12-16 16:06:43 +09001374 int err = -EFAULT;
Brice Goglin2f007e72008-10-18 20:27:16 -07001375
1376 vma = find_vma(mm, addr);
Gleb Natapov70384dc2010-10-26 14:22:07 -07001377 if (!vma || addr < vma->vm_start)
Christoph Lameter742755a2006-06-23 02:03:55 -07001378 goto set_status;
1379
Brice Goglin2f007e72008-10-18 20:27:16 -07001380 page = follow_page(vma, addr, 0);
Linus Torvalds89f5b7d2008-06-20 11:18:25 -07001381
1382 err = PTR_ERR(page);
1383 if (IS_ERR(page))
1384 goto set_status;
1385
Christoph Lameter742755a2006-06-23 02:03:55 -07001386 err = -ENOENT;
1387 /* Use PageReserved to check for zero page */
Hugh Dickinsb79bc0a2013-02-22 16:35:13 -08001388 if (!page || PageReserved(page))
Christoph Lameter742755a2006-06-23 02:03:55 -07001389 goto set_status;
1390
1391 err = page_to_nid(page);
1392set_status:
Brice Goglin80bba122008-12-09 13:14:23 -08001393 *status = err;
1394
1395 pages++;
1396 status++;
1397 }
1398
1399 up_read(&mm->mmap_sem);
1400}
1401
1402/*
1403 * Determine the nodes of a user array of pages and store it in
1404 * a user array of status.
1405 */
1406static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1407 const void __user * __user *pages,
1408 int __user *status)
1409{
1410#define DO_PAGES_STAT_CHUNK_NR 16
1411 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1412 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
Brice Goglin80bba122008-12-09 13:14:23 -08001413
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001414 while (nr_pages) {
1415 unsigned long chunk_nr;
Brice Goglin80bba122008-12-09 13:14:23 -08001416
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001417 chunk_nr = nr_pages;
1418 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1419 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1420
1421 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1422 break;
Brice Goglin80bba122008-12-09 13:14:23 -08001423
1424 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1425
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001426 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1427 break;
Christoph Lameter742755a2006-06-23 02:03:55 -07001428
H. Peter Anvin87b8d1a2010-02-18 16:13:40 -08001429 pages += chunk_nr;
1430 status += chunk_nr;
1431 nr_pages -= chunk_nr;
1432 }
1433 return nr_pages ? -EFAULT : 0;
Christoph Lameter742755a2006-06-23 02:03:55 -07001434}
1435
1436/*
1437 * Move a list of pages in the address space of the currently executing
1438 * process.
1439 */
Heiko Carstens938bb9f2009-01-14 14:14:30 +01001440SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1441 const void __user * __user *, pages,
1442 const int __user *, nodes,
1443 int __user *, status, int, flags)
Christoph Lameter742755a2006-06-23 02:03:55 -07001444{
David Howellsc69e8d92008-11-14 10:39:19 +11001445 const struct cred *cred = current_cred(), *tcred;
Christoph Lameter742755a2006-06-23 02:03:55 -07001446 struct task_struct *task;
Christoph Lameter742755a2006-06-23 02:03:55 -07001447 struct mm_struct *mm;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001448 int err;
Christoph Lameter3268c632012-03-21 16:34:06 -07001449 nodemask_t task_nodes;
Christoph Lameter742755a2006-06-23 02:03:55 -07001450
1451 /* Check flags */
1452 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1453 return -EINVAL;
1454
1455 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1456 return -EPERM;
1457
1458 /* Find the mm_struct */
Greg Thelena879bf52011-02-25 14:44:13 -08001459 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001460 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter742755a2006-06-23 02:03:55 -07001461 if (!task) {
Greg Thelena879bf52011-02-25 14:44:13 -08001462 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001463 return -ESRCH;
1464 }
Christoph Lameter3268c632012-03-21 16:34:06 -07001465 get_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001466
1467 /*
1468 * Check if this process has the right to modify the specified
1469 * process. The right exists if the process has administrative
1470 * capabilities, superuser privileges or the same
1471 * userid as the target process.
1472 */
David Howellsc69e8d92008-11-14 10:39:19 +11001473 tcred = __task_cred(task);
Eric W. Biedermanb38a86e2012-03-12 15:48:24 -07001474 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1475 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
Christoph Lameter742755a2006-06-23 02:03:55 -07001476 !capable(CAP_SYS_NICE)) {
David Howellsc69e8d92008-11-14 10:39:19 +11001477 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001478 err = -EPERM;
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001479 goto out;
Christoph Lameter742755a2006-06-23 02:03:55 -07001480 }
David Howellsc69e8d92008-11-14 10:39:19 +11001481 rcu_read_unlock();
Christoph Lameter742755a2006-06-23 02:03:55 -07001482
David Quigley86c3a762006-06-23 02:04:02 -07001483 err = security_task_movememory(task);
1484 if (err)
Brice Goglin5e9a0f02008-10-18 20:27:17 -07001485 goto out;
David Quigley86c3a762006-06-23 02:04:02 -07001486
Christoph Lameter3268c632012-03-21 16:34:06 -07001487 task_nodes = cpuset_mems_allowed(task);
1488 mm = get_task_mm(task);
1489 put_task_struct(task);
1490
Sasha Levin6e8b09e2012-04-25 16:01:53 -07001491 if (!mm)
1492 return -EINVAL;
1493
1494 if (nodes)
1495 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1496 nodes, status, flags);
1497 else
1498 err = do_pages_stat(mm, nr_pages, pages, status);
Christoph Lameter3268c632012-03-21 16:34:06 -07001499
1500 mmput(mm);
1501 return err;
David Quigley86c3a762006-06-23 02:04:02 -07001502
Christoph Lameter742755a2006-06-23 02:03:55 -07001503out:
Christoph Lameter3268c632012-03-21 16:34:06 -07001504 put_task_struct(task);
Christoph Lameter742755a2006-06-23 02:03:55 -07001505 return err;
1506}
Christoph Lameter742755a2006-06-23 02:03:55 -07001507
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001508#ifdef CONFIG_NUMA_BALANCING
1509/*
1510 * Returns true if this is a safe migration target node for misplaced NUMA
1511 * pages. Currently it only checks the watermarks which crude
1512 */
1513static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
Mel Gorman3abef4e2013-02-22 16:34:27 -08001514 unsigned long nr_migrate_pages)
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001515{
1516 int z;
1517 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1518 struct zone *zone = pgdat->node_zones + z;
1519
1520 if (!populated_zone(zone))
1521 continue;
1522
Lisa Du6e543d52013-09-11 14:22:36 -07001523 if (!zone_reclaimable(zone))
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001524 continue;
1525
1526 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1527 if (!zone_watermark_ok(zone, 0,
1528 high_wmark_pages(zone) +
1529 nr_migrate_pages,
1530 0, 0))
1531 continue;
1532 return true;
1533 }
1534 return false;
1535}
1536
1537static struct page *alloc_misplaced_dst_page(struct page *page,
1538 unsigned long data,
1539 int **result)
1540{
1541 int nid = (int) data;
1542 struct page *newpage;
1543
1544 newpage = alloc_pages_exact_node(nid,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001545 (GFP_HIGHUSER_MOVABLE |
1546 __GFP_THISNODE | __GFP_NOMEMALLOC |
1547 __GFP_NORETRY | __GFP_NOWARN) &
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001548 ~GFP_IOFS, 0);
Hillf Dantonbac03822012-11-27 14:46:24 +00001549
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001550 return newpage;
1551}
1552
1553/*
Mel Gormana8f60772012-11-14 21:41:46 +00001554 * page migration rate limiting control.
1555 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1556 * window of time. Default here says do not migrate more than 1280M per second.
Mel Gormane14808b2012-11-19 10:59:15 +00001557 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1558 * as it is faults that reset the window, pte updates will happen unconditionally
1559 * if there has not been a fault since @pteupdate_interval_millisecs after the
1560 * throttle window closed.
Mel Gormana8f60772012-11-14 21:41:46 +00001561 */
1562static unsigned int migrate_interval_millisecs __read_mostly = 100;
Mel Gormane14808b2012-11-19 10:59:15 +00001563static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
Mel Gormana8f60772012-11-14 21:41:46 +00001564static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1565
Mel Gormane14808b2012-11-19 10:59:15 +00001566/* Returns true if NUMA migration is currently rate limited */
1567bool migrate_ratelimited(int node)
1568{
1569 pg_data_t *pgdat = NODE_DATA(node);
1570
1571 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1572 msecs_to_jiffies(pteupdate_interval_millisecs)))
1573 return false;
1574
1575 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1576 return false;
1577
1578 return true;
1579}
1580
Mel Gormanb32967f2012-11-19 12:35:47 +00001581/* Returns true if the node is migrate rate-limited after the update */
Mel Gorman1c30e012014-01-21 15:50:58 -08001582static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1583 unsigned long nr_pages)
Mel Gormanb32967f2012-11-19 12:35:47 +00001584{
Mel Gormanb32967f2012-11-19 12:35:47 +00001585 /*
1586 * Rate-limit the amount of data that is being migrated to a node.
1587 * Optimal placement is no good if the memory bus is saturated and
1588 * all the time is being spent migrating!
1589 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001590 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001591 spin_lock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001592 pgdat->numabalancing_migrate_nr_pages = 0;
1593 pgdat->numabalancing_migrate_next_window = jiffies +
1594 msecs_to_jiffies(migrate_interval_millisecs);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001595 spin_unlock(&pgdat->numabalancing_migrate_lock);
Mel Gormanb32967f2012-11-19 12:35:47 +00001596 }
Mel Gormanaf1839d2014-01-21 15:51:01 -08001597 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1598 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1599 nr_pages);
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001600 return true;
Mel Gormanaf1839d2014-01-21 15:51:01 -08001601 }
Mel Gorman1c5e9c22014-01-21 15:50:59 -08001602
1603 /*
1604 * This is an unlocked non-atomic update so errors are possible.
1605 * The consequences are failing to migrate when we potentiall should
1606 * have which is not severe enough to warrant locking. If it is ever
1607 * a problem, it can be converted to a per-cpu counter.
1608 */
1609 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1610 return false;
Mel Gormanb32967f2012-11-19 12:35:47 +00001611}
1612
Mel Gorman1c30e012014-01-21 15:50:58 -08001613static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
Mel Gormanb32967f2012-11-19 12:35:47 +00001614{
Hugh Dickins340ef392013-02-22 16:34:33 -08001615 int page_lru;
Mel Gormanb32967f2012-11-19 12:35:47 +00001616
Sasha Levin309381fea2014-01-23 15:52:54 -08001617 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
Mel Gorman3abef4e2013-02-22 16:34:27 -08001618
Mel Gormanb32967f2012-11-19 12:35:47 +00001619 /* Avoid migrating to a node that is nearly full */
Hugh Dickins340ef392013-02-22 16:34:33 -08001620 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1621 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001622
Hugh Dickins340ef392013-02-22 16:34:33 -08001623 if (isolate_lru_page(page))
1624 return 0;
Mel Gormanb32967f2012-11-19 12:35:47 +00001625
1626 /*
Hugh Dickins340ef392013-02-22 16:34:33 -08001627 * migrate_misplaced_transhuge_page() skips page migration's usual
1628 * check on page_count(), so we must do it here, now that the page
1629 * has been isolated: a GUP pin, or any other pin, prevents migration.
1630 * The expected page count is 3: 1 for page's mapcount and 1 for the
1631 * caller's pin and 1 for the reference taken by isolate_lru_page().
1632 */
1633 if (PageTransHuge(page) && page_count(page) != 3) {
1634 putback_lru_page(page);
1635 return 0;
1636 }
1637
1638 page_lru = page_is_file_cache(page);
1639 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1640 hpage_nr_pages(page));
1641
1642 /*
1643 * Isolating the page has taken another reference, so the
1644 * caller's reference can be safely dropped without the page
1645 * disappearing underneath us during migration.
Mel Gormanb32967f2012-11-19 12:35:47 +00001646 */
1647 put_page(page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001648 return 1;
Mel Gormanb32967f2012-11-19 12:35:47 +00001649}
1650
Mel Gormande466bd2013-12-18 17:08:42 -08001651bool pmd_trans_migrating(pmd_t pmd)
1652{
1653 struct page *page = pmd_page(pmd);
1654 return PageLocked(page);
1655}
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,
David Rientjes68711a72014-06-04 16:08:25 -07001692 NULL, node, MIGRATE_ASYNC,
1693 MR_NUMA_MISPLACED);
Mel Gormanb32967f2012-11-19 12:35:47 +00001694 if (nr_remaining) {
Joonsoo Kim59c82b72014-01-21 15:51:17 -08001695 if (!list_empty(&migratepages)) {
1696 list_del(&page->lru);
1697 dec_zone_page_state(page, NR_ISOLATED_ANON +
1698 page_is_file_cache(page));
1699 putback_lru_page(page);
1700 }
Mel Gormanb32967f2012-11-19 12:35:47 +00001701 isolated = 0;
1702 } else
1703 count_vm_numa_event(NUMA_PAGE_MIGRATE);
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001704 BUG_ON(!list_empty(&migratepages));
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001705 return isolated;
Hugh Dickins340ef392013-02-22 16:34:33 -08001706
1707out:
1708 put_page(page);
1709 return 0;
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001710}
Mel Gorman220018d2012-12-05 09:32:56 +00001711#endif /* CONFIG_NUMA_BALANCING */
Mel Gormanb32967f2012-11-19 12:35:47 +00001712
Mel Gorman220018d2012-12-05 09:32:56 +00001713#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
Hugh Dickins340ef392013-02-22 16:34:33 -08001714/*
1715 * Migrates a THP to a given target node. page must be locked and is unlocked
1716 * before returning.
1717 */
Mel Gormanb32967f2012-11-19 12:35:47 +00001718int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1719 struct vm_area_struct *vma,
1720 pmd_t *pmd, pmd_t entry,
1721 unsigned long address,
1722 struct page *page, int node)
1723{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001724 spinlock_t *ptl;
Mel Gormanb32967f2012-11-19 12:35:47 +00001725 pg_data_t *pgdat = NODE_DATA(node);
1726 int isolated = 0;
1727 struct page *new_page = NULL;
Mel Gormanb32967f2012-11-19 12:35:47 +00001728 int page_lru = page_is_file_cache(page);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001729 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1730 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
Mel Gorman2b4847e2013-12-18 17:08:32 -08001731 pmd_t orig_entry;
Mel Gormanb32967f2012-11-19 12:35:47 +00001732
1733 /*
Mel Gormanb32967f2012-11-19 12:35:47 +00001734 * Rate-limit the amount of data that is being migrated to a node.
1735 * Optimal placement is no good if the memory bus is saturated and
1736 * all the time is being spent migrating!
1737 */
Mel Gormand28d43352012-11-29 09:24:36 +00001738 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
Mel Gormanb32967f2012-11-19 12:35:47 +00001739 goto out_dropref;
1740
1741 new_page = alloc_pages_node(node,
Johannes Weinere97ca8e2014-03-10 15:49:43 -07001742 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1743 HPAGE_PMD_ORDER);
Hugh Dickins340ef392013-02-22 16:34:33 -08001744 if (!new_page)
1745 goto out_fail;
1746
Mel Gormanb32967f2012-11-19 12:35:47 +00001747 isolated = numamigrate_isolate_page(pgdat, page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001748 if (!isolated) {
Mel Gormanb32967f2012-11-19 12:35:47 +00001749 put_page(new_page);
Hugh Dickins340ef392013-02-22 16:34:33 -08001750 goto out_fail;
Mel Gormanb32967f2012-11-19 12:35:47 +00001751 }
1752
Mel Gormanb0943d62013-12-18 17:08:46 -08001753 if (mm_tlb_flush_pending(mm))
1754 flush_tlb_range(vma, mmun_start, mmun_end);
1755
Mel Gormanb32967f2012-11-19 12:35:47 +00001756 /* Prepare a page as a migration target */
1757 __set_page_locked(new_page);
1758 SetPageSwapBacked(new_page);
1759
1760 /* anon mapping, we can simply copy page->mapping to the new page: */
1761 new_page->mapping = page->mapping;
1762 new_page->index = page->index;
1763 migrate_page_copy(new_page, page);
1764 WARN_ON(PageLRU(new_page));
1765
1766 /* Recheck the target PMD */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001767 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001768 ptl = pmd_lock(mm, pmd);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001769 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1770fail_putback:
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001771 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001772 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001773
1774 /* Reverse changes made by migrate_page_copy() */
1775 if (TestClearPageActive(new_page))
1776 SetPageActive(page);
1777 if (TestClearPageUnevictable(new_page))
1778 SetPageUnevictable(page);
1779 mlock_migrate_page(page, new_page);
1780
1781 unlock_page(new_page);
1782 put_page(new_page); /* Free it */
1783
Mel Gormana54a4072013-10-07 11:28:46 +01001784 /* Retake the callers reference and putback on LRU */
1785 get_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001786 putback_lru_page(page);
Mel Gormana54a4072013-10-07 11:28:46 +01001787 mod_zone_page_state(page_zone(page),
1788 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
Mel Gormaneb4489f62013-12-18 17:08:39 -08001789
1790 goto out_unlock;
Mel Gormanb32967f2012-11-19 12:35:47 +00001791 }
1792
Mel Gorman2b4847e2013-12-18 17:08:32 -08001793 orig_entry = *pmd;
Mel Gormanb32967f2012-11-19 12:35:47 +00001794 entry = mk_pmd(new_page, vma->vm_page_prot);
Mel Gormanb32967f2012-11-19 12:35:47 +00001795 entry = pmd_mkhuge(entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001796 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
Mel Gormanb32967f2012-11-19 12:35:47 +00001797
Mel Gorman2b4847e2013-12-18 17:08:32 -08001798 /*
1799 * Clear the old entry under pagetable lock and establish the new PTE.
1800 * Any parallel GUP will either observe the old page blocking on the
1801 * page lock, block on the page table lock or observe the new page.
1802 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1803 * guarantee the copy is visible before the pagetable update.
1804 */
Mel Gormanf714f4f2013-12-18 17:08:33 -08001805 flush_cache_range(vma, mmun_start, mmun_end);
Mel Gorman11de9922014-06-04 16:07:41 -07001806 page_add_anon_rmap(new_page, vma, mmun_start);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001807 pmdp_clear_flush_notify(vma, mmun_start, pmd);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001808 set_pmd_at(mm, mmun_start, pmd, entry);
1809 flush_tlb_range(vma, mmun_start, mmun_end);
Stephen Rothwellce4a9cc2012-12-10 19:50:57 +11001810 update_mmu_cache_pmd(vma, address, &entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001811
1812 if (page_count(page) != 2) {
Mel Gormanf714f4f2013-12-18 17:08:33 -08001813 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1814 flush_tlb_range(vma, mmun_start, mmun_end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11001815 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001816 update_mmu_cache_pmd(vma, address, &entry);
1817 page_remove_rmap(new_page);
1818 goto fail_putback;
1819 }
1820
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001821 mem_cgroup_migrate(page, new_page, false);
1822
Mel Gormanb32967f2012-11-19 12:35:47 +00001823 page_remove_rmap(page);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001824
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -08001825 spin_unlock(ptl);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001826 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Mel Gormanb32967f2012-11-19 12:35:47 +00001827
Mel Gorman11de9922014-06-04 16:07:41 -07001828 /* Take an "isolate" reference and put new page on the LRU. */
1829 get_page(new_page);
1830 putback_lru_page(new_page);
1831
Mel Gormanb32967f2012-11-19 12:35:47 +00001832 unlock_page(new_page);
1833 unlock_page(page);
1834 put_page(page); /* Drop the rmap reference */
1835 put_page(page); /* Drop the LRU isolation reference */
1836
1837 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1838 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1839
Mel Gormanb32967f2012-11-19 12:35:47 +00001840 mod_zone_page_state(page_zone(page),
1841 NR_ISOLATED_ANON + page_lru,
1842 -HPAGE_PMD_NR);
1843 return isolated;
1844
Hugh Dickins340ef392013-02-22 16:34:33 -08001845out_fail:
1846 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
Mel Gormanb32967f2012-11-19 12:35:47 +00001847out_dropref:
Mel Gorman2b4847e2013-12-18 17:08:32 -08001848 ptl = pmd_lock(mm, pmd);
1849 if (pmd_same(*pmd, entry)) {
Mel Gorman4d942462015-02-12 14:58:28 -08001850 entry = pmd_modify(entry, vma->vm_page_prot);
Mel Gormanf714f4f2013-12-18 17:08:33 -08001851 set_pmd_at(mm, mmun_start, pmd, entry);
Mel Gorman2b4847e2013-12-18 17:08:32 -08001852 update_mmu_cache_pmd(vma, address, &entry);
1853 }
1854 spin_unlock(ptl);
Mel Gormana54a4072013-10-07 11:28:46 +01001855
Mel Gormaneb4489f62013-12-18 17:08:39 -08001856out_unlock:
Hugh Dickins340ef392013-02-22 16:34:33 -08001857 unlock_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001858 put_page(page);
Mel Gormanb32967f2012-11-19 12:35:47 +00001859 return 0;
1860}
Peter Zijlstra7039e1d2012-10-25 14:16:34 +02001861#endif /* CONFIG_NUMA_BALANCING */
1862
1863#endif /* CONFIG_NUMA */