]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
mm/hugetlb: pmd_huge() returns true for non-present hugepage
[mirror_ubuntu-jammy-kernel.git] / mm / migrate.c
CommitLineData
b20a3503
CL
1/*
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>
cde53535 12 * Christoph Lameter
b20a3503
CL
13 */
14
15#include <linux/migrate.h>
b95f1b31 16#include <linux/export.h>
b20a3503 17#include <linux/swap.h>
0697212a 18#include <linux/swapops.h>
b20a3503 19#include <linux/pagemap.h>
e23ca00b 20#include <linux/buffer_head.h>
b20a3503 21#include <linux/mm_inline.h>
b488893a 22#include <linux/nsproxy.h>
b20a3503 23#include <linux/pagevec.h>
e9995ef9 24#include <linux/ksm.h>
b20a3503
CL
25#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
04e62a29 29#include <linux/writeback.h>
742755a1
CL
30#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
86c3a764 32#include <linux/security.h>
8a9f3ccd 33#include <linux/memcontrol.h>
4f5ca265 34#include <linux/syscalls.h>
290408d4 35#include <linux/hugetlb.h>
8e6ac7fa 36#include <linux/hugetlb_cgroup.h>
5a0e3ad6 37#include <linux/gfp.h>
bf6bddf1 38#include <linux/balloon_compaction.h>
f714f4f2 39#include <linux/mmu_notifier.h>
b20a3503 40
0d1836c3
MN
41#include <asm/tlbflush.h>
42
7b2a2d4a
MG
43#define CREATE_TRACE_POINTS
44#include <trace/events/migrate.h>
45
b20a3503
CL
46#include "internal.h"
47
b20a3503 48/*
742755a1 49 * migrate_prep() needs to be called before we start compiling a list of pages
748446bb
MG
50 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
b20a3503
CL
52 */
53int migrate_prep(void)
54{
b20a3503
CL
55 /*
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
748446bb
MG
66/* 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
5733c7d1
RA
74/*
75 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
77 *
59c82b70
JK
78 * 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().
5733c7d1
RA
81 */
82void putback_movable_pages(struct list_head *l)
83{
84 struct page *page;
85 struct page *page2;
86
b20a3503 87 list_for_each_entry_safe(page, page2, l, lru) {
31caf665
NH
88 if (unlikely(PageHuge(page))) {
89 putback_active_hugepage(page);
90 continue;
91 }
e24f0b8f 92 list_del(&page->lru);
a731286d 93 dec_zone_page_state(page, NR_ISOLATED_ANON +
6c0b1351 94 page_is_file_cache(page));
117aad1e 95 if (unlikely(isolated_balloon_page(page)))
bf6bddf1
RA
96 balloon_page_putback(page);
97 else
98 putback_lru_page(page);
b20a3503 99 }
b20a3503
CL
100}
101
0697212a
CL
102/*
103 * Restore a potential migration pte to a working pte entry
104 */
e9995ef9
HD
105static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
106 unsigned long addr, void *old)
0697212a
CL
107{
108 struct mm_struct *mm = vma->vm_mm;
109 swp_entry_t entry;
0697212a
CL
110 pmd_t *pmd;
111 pte_t *ptep, pte;
112 spinlock_t *ptl;
113
290408d4
NH
114 if (unlikely(PageHuge(new))) {
115 ptep = huge_pte_offset(mm, addr);
116 if (!ptep)
117 goto out;
cb900f41 118 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
290408d4 119 } else {
6219049a
BL
120 pmd = mm_find_pmd(mm, addr);
121 if (!pmd)
290408d4 122 goto out;
0697212a 123
290408d4 124 ptep = pte_offset_map(pmd, addr);
0697212a 125
486cf46f
HD
126 /*
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 */
290408d4
NH
130
131 ptl = pte_lockptr(mm, pmd);
132 }
0697212a 133
0697212a
CL
134 spin_lock(ptl);
135 pte = *ptep;
136 if (!is_swap_pte(pte))
e9995ef9 137 goto unlock;
0697212a
CL
138
139 entry = pte_to_swp_entry(pte);
140
e9995ef9
HD
141 if (!is_migration_entry(entry) ||
142 migration_entry_to_page(entry) != old)
143 goto unlock;
0697212a 144
0697212a
CL
145 get_page(new);
146 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
c3d16e16
CG
147 if (pte_swp_soft_dirty(*ptep))
148 pte = pte_mksoft_dirty(pte);
d3cb8bf6
MG
149
150 /* Recheck VMA as permissions can change since migration started */
0697212a 151 if (is_write_migration_entry(entry))
d3cb8bf6
MG
152 pte = maybe_mkwrite(pte, vma);
153
3ef8fd7f 154#ifdef CONFIG_HUGETLB_PAGE
be7517d6 155 if (PageHuge(new)) {
290408d4 156 pte = pte_mkhuge(pte);
be7517d6
TL
157 pte = arch_make_huge_pte(pte, vma, new, 0);
158 }
3ef8fd7f 159#endif
c2cc499c 160 flush_dcache_page(new);
0697212a 161 set_pte_at(mm, addr, ptep, pte);
04e62a29 162
290408d4
NH
163 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))
04e62a29
CL
169 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 */
4b3073e1 174 update_mmu_cache(vma, addr, ptep);
e9995ef9 175unlock:
0697212a 176 pte_unmap_unlock(ptep, ptl);
e9995ef9
HD
177out:
178 return SWAP_AGAIN;
0697212a
CL
179}
180
04e62a29
CL
181/*
182 * 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{
051ac83a
JK
187 struct rmap_walk_control rwc = {
188 .rmap_one = remove_migration_pte,
189 .arg = old,
190 };
191
192 rmap_walk(new, &rwc);
04e62a29
CL
193}
194
0697212a
CL
195/*
196 * 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.
0697212a 199 */
30dad309
NH
200static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
201 spinlock_t *ptl)
0697212a 202{
30dad309 203 pte_t pte;
0697212a
CL
204 swp_entry_t entry;
205 struct page *page;
206
30dad309 207 spin_lock(ptl);
0697212a
CL
208 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
e286781d
NP
218 /*
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;
0697212a
CL
227 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
30dad309
NH
235void 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
cb900f41
KS
243void migration_entry_wait_huge(struct vm_area_struct *vma,
244 struct mm_struct *mm, pte_t *pte)
30dad309 245{
cb900f41 246 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
30dad309
NH
247 __migration_entry_wait(mm, pte, ptl);
248}
249
b969c4ab
MG
250#ifdef CONFIG_BLOCK
251/* Returns true if all buffers are successfully locked */
a6bc32b8
MG
252static bool buffer_migrate_lock_buffers(struct buffer_head *head,
253 enum migrate_mode mode)
b969c4ab
MG
254{
255 struct buffer_head *bh = head;
256
257 /* Simple case, sync compaction */
a6bc32b8 258 if (mode != MIGRATE_ASYNC) {
b969c4ab
MG
259 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,
a6bc32b8 294 enum migrate_mode mode)
b969c4ab
MG
295{
296 return true;
297}
298#endif /* CONFIG_BLOCK */
299
b20a3503 300/*
c3fcf8a5 301 * Replace the page in the mapping.
5b5c7120
CL
302 *
303 * The number of remaining references must be:
304 * 1 for anonymous pages without a mapping
305 * 2 for pages with a mapping
266cf658 306 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
b20a3503 307 */
36bc08cc 308int migrate_page_move_mapping(struct address_space *mapping,
b969c4ab 309 struct page *newpage, struct page *page,
8e321fef
BL
310 struct buffer_head *head, enum migrate_mode mode,
311 int extra_count)
b20a3503 312{
8e321fef 313 int expected_count = 1 + extra_count;
7cf9c2c7 314 void **pslot;
b20a3503 315
6c5240ae 316 if (!mapping) {
0e8c7d0f 317 /* Anonymous page without mapping */
8e321fef 318 if (page_count(page) != expected_count)
6c5240ae 319 return -EAGAIN;
78bd5209 320 return MIGRATEPAGE_SUCCESS;
6c5240ae
CL
321 }
322
19fd6231 323 spin_lock_irq(&mapping->tree_lock);
b20a3503 324
7cf9c2c7
NP
325 pslot = radix_tree_lookup_slot(&mapping->page_tree,
326 page_index(page));
b20a3503 327
8e321fef 328 expected_count += 1 + page_has_private(page);
e286781d 329 if (page_count(page) != expected_count ||
29c1f677 330 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
19fd6231 331 spin_unlock_irq(&mapping->tree_lock);
e23ca00b 332 return -EAGAIN;
b20a3503
CL
333 }
334
e286781d 335 if (!page_freeze_refs(page, expected_count)) {
19fd6231 336 spin_unlock_irq(&mapping->tree_lock);
e286781d
NP
337 return -EAGAIN;
338 }
339
b969c4ab
MG
340 /*
341 * 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 */
a6bc32b8
MG
347 if (mode == MIGRATE_ASYNC && head &&
348 !buffer_migrate_lock_buffers(head, mode)) {
b969c4ab
MG
349 page_unfreeze_refs(page, expected_count);
350 spin_unlock_irq(&mapping->tree_lock);
351 return -EAGAIN;
352 }
353
b20a3503
CL
354 /*
355 * Now we know that no one else is looking at the page.
b20a3503 356 */
7cf9c2c7 357 get_page(newpage); /* add cache reference */
b20a3503
CL
358 if (PageSwapCache(page)) {
359 SetPageSwapCache(newpage);
360 set_page_private(newpage, page_private(page));
361 }
362
7cf9c2c7
NP
363 radix_tree_replace_slot(pslot, newpage);
364
365 /*
937a94c9
JG
366 * Drop cache reference from old page by unfreezing
367 * to one less reference.
7cf9c2c7
NP
368 * We know this isn't the last reference.
369 */
937a94c9 370 page_unfreeze_refs(page, expected_count - 1);
7cf9c2c7 371
0e8c7d0f
CL
372 /*
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);
99a15e21 384 if (!PageSwapCache(page) && PageSwapBacked(page)) {
4b02108a
KM
385 __dec_zone_page_state(page, NR_SHMEM);
386 __inc_zone_page_state(newpage, NR_SHMEM);
387 }
19fd6231 388 spin_unlock_irq(&mapping->tree_lock);
b20a3503 389
78bd5209 390 return MIGRATEPAGE_SUCCESS;
b20a3503 391}
b20a3503 392
290408d4
NH
393/*
394 * 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;
78bd5209 406 return MIGRATEPAGE_SUCCESS;
290408d4
NH
407 }
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 ||
29c1f677 416 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
290408d4
NH
417 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
937a94c9 430 page_unfreeze_refs(page, expected_count - 1);
290408d4
NH
431
432 spin_unlock_irq(&mapping->tree_lock);
78bd5209 433 return MIGRATEPAGE_SUCCESS;
290408d4
NH
434}
435
30b0a105
DH
436/*
437 * 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
b20a3503
CL
484/*
485 * Copy the page to its new location
486 */
290408d4 487void migrate_page_copy(struct page *newpage, struct page *page)
b20a3503 488{
7851a45c
RR
489 int cpupid;
490
b32967ff 491 if (PageHuge(page) || PageTransHuge(page))
290408d4
NH
492 copy_huge_page(newpage, page);
493 else
494 copy_highpage(newpage, page);
b20a3503
CL
495
496 if (PageError(page))
497 SetPageError(newpage);
498 if (PageReferenced(page))
499 SetPageReferenced(newpage);
500 if (PageUptodate(page))
501 SetPageUptodate(newpage);
894bc310 502 if (TestClearPageActive(page)) {
309381fe 503 VM_BUG_ON_PAGE(PageUnevictable(page), page);
b20a3503 504 SetPageActive(newpage);
418b27ef
LS
505 } else if (TestClearPageUnevictable(page))
506 SetPageUnevictable(newpage);
b20a3503
CL
507 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);
3a902c5f
NP
514 /*
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.
25985edc 519 * Whereas only part of our page may be dirty.
3a902c5f 520 */
752dc185
HD
521 if (PageSwapBacked(page))
522 SetPageDirty(newpage);
523 else
524 __set_page_dirty_nobuffers(newpage);
b20a3503
CL
525 }
526
7851a45c
RR
527 /*
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
b291f000 534 mlock_migrate_page(newpage, page);
e9995ef9 535 ksm_migrate_page(newpage, page);
c8d6553b
HD
536 /*
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 */
b20a3503 540 ClearPageSwapCache(page);
b20a3503
CL
541 ClearPagePrivate(page);
542 set_page_private(page, 0);
b20a3503
CL
543
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}
b20a3503 551
1d8b85cc
CL
552/************************************************************
553 * Migration functions
554 ***********************************************************/
555
b20a3503
CL
556/*
557 * Common logic to directly migrate a single page suitable for
266cf658 558 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
559 *
560 * Pages are locked upon entry and exit.
561 */
2d1db3b1 562int migrate_page(struct address_space *mapping,
a6bc32b8
MG
563 struct page *newpage, struct page *page,
564 enum migrate_mode mode)
b20a3503
CL
565{
566 int rc;
567
568 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
569
8e321fef 570 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
b20a3503 571
78bd5209 572 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
573 return rc;
574
575 migrate_page_copy(newpage, page);
78bd5209 576 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
577}
578EXPORT_SYMBOL(migrate_page);
579
9361401e 580#ifdef CONFIG_BLOCK
1d8b85cc
CL
581/*
582 * 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 */
2d1db3b1 586int buffer_migrate_page(struct address_space *mapping,
a6bc32b8 587 struct page *newpage, struct page *page, enum migrate_mode mode)
1d8b85cc 588{
1d8b85cc
CL
589 struct buffer_head *bh, *head;
590 int rc;
591
1d8b85cc 592 if (!page_has_buffers(page))
a6bc32b8 593 return migrate_page(mapping, newpage, page, mode);
1d8b85cc
CL
594
595 head = page_buffers(page);
596
8e321fef 597 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
1d8b85cc 598
78bd5209 599 if (rc != MIGRATEPAGE_SUCCESS)
1d8b85cc
CL
600 return rc;
601
b969c4ab
MG
602 /*
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 */
a6bc32b8
MG
607 if (mode != MIGRATE_ASYNC)
608 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
1d8b85cc
CL
609
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
78bd5209 635 return MIGRATEPAGE_SUCCESS;
1d8b85cc
CL
636}
637EXPORT_SYMBOL(buffer_migrate_page);
9361401e 638#endif
1d8b85cc 639
04e62a29
CL
640/*
641 * Writeback a page to clean the dirty state
642 */
643static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 644{
04e62a29
CL
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,
04e62a29
CL
650 .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
8351a6e4 662 /*
04e62a29
CL
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.
8351a6e4 669 */
04e62a29 670 remove_migration_ptes(page, page);
8351a6e4 671
04e62a29 672 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 673
04e62a29
CL
674 if (rc != AOP_WRITEPAGE_ACTIVATE)
675 /* unlocked. Relock */
676 lock_page(page);
677
bda8550d 678 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
679}
680
681/*
682 * Default handling if a filesystem does not provide a migration function.
683 */
684static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 685 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 686{
b969c4ab 687 if (PageDirty(page)) {
a6bc32b8
MG
688 /* Only writeback pages in full synchronous migration */
689 if (mode != MIGRATE_SYNC)
b969c4ab 690 return -EBUSY;
04e62a29 691 return writeout(mapping, page);
b969c4ab 692 }
8351a6e4
CL
693
694 /*
695 * Buffers may be managed in a filesystem specific way.
696 * We must have no buffers or drop them.
697 */
266cf658 698 if (page_has_private(page) &&
8351a6e4
CL
699 !try_to_release_page(page, GFP_KERNEL))
700 return -EAGAIN;
701
a6bc32b8 702 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
703}
704
e24f0b8f
CL
705/*
706 * 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.
894bc310
LS
711 *
712 * Return value:
713 * < 0 - error code
78bd5209 714 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 715 */
3fe2011f 716static int move_to_new_page(struct page *newpage, struct page *page,
2ebba6b7 717 int page_was_mapped, enum migrate_mode mode)
e24f0b8f
CL
718{
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 */
529ae9aa 727 if (!trylock_page(newpage))
e24f0b8f
CL
728 BUG();
729
730 /* Prepare mapping for the new page.*/
731 newpage->index = page->index;
732 newpage->mapping = page->mapping;
b2e18538
RR
733 if (PageSwapBacked(page))
734 SetPageSwapBacked(newpage);
e24f0b8f
CL
735
736 mapping = page_mapping(page);
737 if (!mapping)
a6bc32b8 738 rc = migrate_page(mapping, newpage, page, mode);
b969c4ab 739 else if (mapping->a_ops->migratepage)
e24f0b8f 740 /*
b969c4ab
MG
741 * 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.
e24f0b8f 745 */
b969c4ab 746 rc = mapping->a_ops->migratepage(mapping,
a6bc32b8 747 newpage, page, mode);
b969c4ab 748 else
a6bc32b8 749 rc = fallback_migrate_page(mapping, newpage, page, mode);
e24f0b8f 750
78bd5209 751 if (rc != MIGRATEPAGE_SUCCESS) {
e24f0b8f 752 newpage->mapping = NULL;
3fe2011f 753 } else {
0a31bc97 754 mem_cgroup_migrate(page, newpage, false);
2ebba6b7 755 if (page_was_mapped)
3fe2011f 756 remove_migration_ptes(page, newpage);
35512eca 757 page->mapping = NULL;
3fe2011f 758 }
e24f0b8f
CL
759
760 unlock_page(newpage);
761
762 return rc;
763}
764
0dabec93 765static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 766 int force, enum migrate_mode mode)
e24f0b8f 767{
0dabec93 768 int rc = -EAGAIN;
2ebba6b7 769 int page_was_mapped = 0;
3f6c8272 770 struct anon_vma *anon_vma = NULL;
95a402c3 771
529ae9aa 772 if (!trylock_page(page)) {
a6bc32b8 773 if (!force || mode == MIGRATE_ASYNC)
0dabec93 774 goto out;
3e7d3449
MG
775
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)
0dabec93 790 goto out;
3e7d3449 791
e24f0b8f
CL
792 lock_page(page);
793 }
794
795 if (PageWriteback(page)) {
11bc82d6 796 /*
fed5b64a 797 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
798 * 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
11bc82d6 801 */
a6bc32b8 802 if (mode != MIGRATE_SYNC) {
11bc82d6 803 rc = -EBUSY;
0a31bc97 804 goto out_unlock;
11bc82d6
AA
805 }
806 if (!force)
0a31bc97 807 goto out_unlock;
e24f0b8f
CL
808 wait_on_page_writeback(page);
809 }
e24f0b8f 810 /*
dc386d4d
KH
811 * 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.
1ce82b69 813 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 814 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
815 * File Caches may use write_page() or lock_page() in migration, then,
816 * just care Anon page here.
dc386d4d 817 */
b79bc0a0 818 if (PageAnon(page) && !PageKsm(page)) {
1ce82b69 819 /*
4fc3f1d6 820 * Only page_lock_anon_vma_read() understands the subtleties of
1ce82b69
HD
821 * getting a hold on an anon_vma from outside one of its mms.
822 */
746b18d4 823 anon_vma = page_get_anon_vma(page);
1ce82b69
HD
824 if (anon_vma) {
825 /*
746b18d4 826 * Anon page
1ce82b69 827 */
1ce82b69 828 } else if (PageSwapCache(page)) {
3fe2011f
MG
829 /*
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 */
3fe2011f 841 } else {
0a31bc97 842 goto out_unlock;
3fe2011f 843 }
989f89c5 844 }
62e1c553 845
d6d86c0a 846 if (unlikely(isolated_balloon_page(page))) {
bf6bddf1
RA
847 /*
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);
0a31bc97 855 goto out_unlock;
bf6bddf1
RA
856 }
857
dc386d4d 858 /*
62e1c553
SL
859 * 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.
e24f0b8f 869 */
62e1c553 870 if (!page->mapping) {
309381fe 871 VM_BUG_ON_PAGE(PageAnon(page), page);
1ce82b69 872 if (page_has_private(page)) {
62e1c553 873 try_to_free_buffers(page);
0a31bc97 874 goto out_unlock;
62e1c553 875 }
abfc3488 876 goto skip_unmap;
62e1c553
SL
877 }
878
dc386d4d 879 /* Establish migration ptes or remove ptes */
2ebba6b7
HD
880 if (page_mapped(page)) {
881 try_to_unmap(page,
882 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
883 page_was_mapped = 1;
884 }
dc386d4d 885
abfc3488 886skip_unmap:
e6a1530d 887 if (!page_mapped(page))
2ebba6b7 888 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
e24f0b8f 889
2ebba6b7 890 if (rc && page_was_mapped)
e24f0b8f 891 remove_migration_ptes(page, page);
3f6c8272
MG
892
893 /* Drop an anon_vma reference if we took one */
76545066 894 if (anon_vma)
9e60109f 895 put_anon_vma(anon_vma);
3f6c8272 896
0a31bc97 897out_unlock:
e24f0b8f 898 unlock_page(page);
0dabec93
MK
899out:
900 return rc;
901}
95a402c3 902
0dabec93
MK
903/*
904 * Obtain the lock on page, remove all ptes and migrate the page
905 * to the newly allocated page in newpage.
906 */
68711a74
DR
907static 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)
0dabec93
MK
910{
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
9c620e2b 927 rc = __unmap_and_move(page, newpage, force, mode);
bf6bddf1 928
0dabec93 929out:
e24f0b8f 930 if (rc != -EAGAIN) {
0dabec93
MK
931 /*
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);
a731286d 938 dec_zone_page_state(page, NR_ISOLATED_ANON +
6c0b1351 939 page_is_file_cache(page));
894bc310 940 putback_lru_page(page);
e24f0b8f 941 }
68711a74 942
95a402c3 943 /*
68711a74
DR
944 * 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.
95a402c3 947 */
8bdd6380
HD
948 if (rc != MIGRATEPAGE_SUCCESS && put_new_page) {
949 ClearPageSwapBacked(newpage);
68711a74 950 put_new_page(newpage, private);
d6d86c0a
KK
951 } else if (unlikely(__is_movable_balloon_page(newpage))) {
952 /* drop our reference, page already in the balloon */
953 put_page(newpage);
8bdd6380 954 } else
68711a74
DR
955 putback_lru_page(newpage);
956
742755a1
CL
957 if (result) {
958 if (rc)
959 *result = rc;
960 else
961 *result = page_to_nid(newpage);
962 }
e24f0b8f
CL
963 return rc;
964}
965
290408d4
NH
966/*
967 * 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,
68711a74
DR
985 free_page_t put_new_page, unsigned long private,
986 struct page *hpage, int force,
987 enum migrate_mode mode)
290408d4
NH
988{
989 int rc = 0;
990 int *result = NULL;
2ebba6b7 991 int page_was_mapped = 0;
32665f2b 992 struct page *new_hpage;
290408d4
NH
993 struct anon_vma *anon_vma = NULL;
994
83467efb
NH
995 /*
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 */
100873d7 1002 if (!hugepage_migration_supported(page_hstate(hpage))) {
32665f2b 1003 putback_active_hugepage(hpage);
83467efb 1004 return -ENOSYS;
32665f2b 1005 }
83467efb 1006
32665f2b 1007 new_hpage = get_new_page(hpage, private, &result);
290408d4
NH
1008 if (!new_hpage)
1009 return -ENOMEM;
1010
1011 rc = -EAGAIN;
1012
1013 if (!trylock_page(hpage)) {
a6bc32b8 1014 if (!force || mode != MIGRATE_SYNC)
290408d4
NH
1015 goto out;
1016 lock_page(hpage);
1017 }
1018
746b18d4
PZ
1019 if (PageAnon(hpage))
1020 anon_vma = page_get_anon_vma(hpage);
290408d4 1021
2ebba6b7
HD
1022 if (page_mapped(hpage)) {
1023 try_to_unmap(hpage,
1024 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1025 page_was_mapped = 1;
1026 }
290408d4
NH
1027
1028 if (!page_mapped(hpage))
2ebba6b7 1029 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
290408d4 1030
2ebba6b7 1031 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
290408d4
NH
1032 remove_migration_ptes(hpage, hpage);
1033
fd4a4663 1034 if (anon_vma)
9e60109f 1035 put_anon_vma(anon_vma);
8e6ac7fa 1036
68711a74 1037 if (rc == MIGRATEPAGE_SUCCESS)
8e6ac7fa
AK
1038 hugetlb_cgroup_migrate(hpage, new_hpage);
1039
290408d4 1040 unlock_page(hpage);
09761333 1041out:
b8ec1cee
NH
1042 if (rc != -EAGAIN)
1043 putback_active_hugepage(hpage);
68711a74
DR
1044
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
290408d4
NH
1055 if (result) {
1056 if (rc)
1057 *result = rc;
1058 else
1059 *result = page_to_nid(new_hpage);
1060 }
1061 return rc;
1062}
1063
b20a3503 1064/*
c73e5c9c
SB
1065 * migrate_pages - migrate the pages specified in a list, to the free pages
1066 * supplied as the target for the page migration
b20a3503 1067 *
c73e5c9c
SB
1068 * @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.
68711a74
DR
1071 * @put_new_page: The function used to free target pages if migration
1072 * fails, or NULL if no special handling is necessary.
c73e5c9c
SB
1073 * @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.
b20a3503 1077 *
c73e5c9c
SB
1078 * 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
28bd6578 1081 * or free list only if ret != 0.
b20a3503 1082 *
c73e5c9c 1083 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1084 */
9c620e2b 1085int migrate_pages(struct list_head *from, new_page_t get_new_page,
68711a74
DR
1086 free_page_t put_new_page, unsigned long private,
1087 enum migrate_mode mode, int reason)
b20a3503 1088{
e24f0b8f 1089 int retry = 1;
b20a3503 1090 int nr_failed = 0;
5647bc29 1091 int nr_succeeded = 0;
b20a3503
CL
1092 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
e24f0b8f
CL
1101 for(pass = 0; pass < 10 && retry; pass++) {
1102 retry = 0;
b20a3503 1103
e24f0b8f 1104 list_for_each_entry_safe(page, page2, from, lru) {
e24f0b8f 1105 cond_resched();
2d1db3b1 1106
31caf665
NH
1107 if (PageHuge(page))
1108 rc = unmap_and_move_huge_page(get_new_page,
68711a74
DR
1109 put_new_page, private, page,
1110 pass > 2, mode);
31caf665 1111 else
68711a74
DR
1112 rc = unmap_and_move(get_new_page, put_new_page,
1113 private, page, pass > 2, mode);
2d1db3b1 1114
e24f0b8f 1115 switch(rc) {
95a402c3
CL
1116 case -ENOMEM:
1117 goto out;
e24f0b8f 1118 case -EAGAIN:
2d1db3b1 1119 retry++;
e24f0b8f 1120 break;
78bd5209 1121 case MIGRATEPAGE_SUCCESS:
5647bc29 1122 nr_succeeded++;
e24f0b8f
CL
1123 break;
1124 default:
354a3363
NH
1125 /*
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 */
2d1db3b1 1131 nr_failed++;
e24f0b8f 1132 break;
2d1db3b1 1133 }
b20a3503
CL
1134 }
1135 }
78bd5209 1136 rc = nr_failed + retry;
95a402c3 1137out:
5647bc29
MG
1138 if (nr_succeeded)
1139 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1140 if (nr_failed)
1141 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1142 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1143
b20a3503
CL
1144 if (!swapwrite)
1145 current->flags &= ~PF_SWAPWRITE;
1146
78bd5209 1147 return rc;
b20a3503 1148}
95a402c3 1149
742755a1
CL
1150#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
e632a938
NH
1174 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,
e97ca8e5 1179 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
742755a1
CL
1180}
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.
5e9a0f02 1186 * The pm array ends with node = MAX_NUMNODES.
742755a1 1187 */
5e9a0f02
BG
1188static int do_move_page_to_node_array(struct mm_struct *mm,
1189 struct page_to_node *pm,
1190 int migrate_all)
742755a1
CL
1191{
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 */
742755a1
CL
1201 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1202 struct vm_area_struct *vma;
1203 struct page *page;
1204
742755a1
CL
1205 err = -EFAULT;
1206 vma = find_vma(mm, pp->addr);
70384dc6 1207 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
742755a1
CL
1208 goto set_status;
1209
500d65d4 1210 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
89f5b7da
LT
1211
1212 err = PTR_ERR(page);
1213 if (IS_ERR(page))
1214 goto set_status;
1215
742755a1
CL
1216 err = -ENOENT;
1217 if (!page)
1218 goto set_status;
1219
62b61f61 1220 /* Use PageReserved to check for zero page */
b79bc0a0 1221 if (PageReserved(page))
742755a1
CL
1222 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
e632a938
NH
1238 if (PageHuge(page)) {
1239 isolate_huge_page(page, &pagelist);
1240 goto put_and_set;
1241 }
1242
62695a84 1243 err = isolate_lru_page(page);
6d9c285a 1244 if (!err) {
62695a84 1245 list_add_tail(&page->lru, &pagelist);
6d9c285a
KM
1246 inc_zone_page_state(page, NR_ISOLATED_ANON +
1247 page_is_file_cache(page));
1248 }
742755a1
CL
1249put_and_set:
1250 /*
1251 * Either remove the duplicate refcount from
1252 * isolate_lru_page() or drop the page ref if it was
1253 * not isolated.
1254 */
1255 put_page(page);
1256set_status:
1257 pp->status = err;
1258 }
1259
e78bbfa8 1260 err = 0;
cf608ac1 1261 if (!list_empty(&pagelist)) {
68711a74 1262 err = migrate_pages(&pagelist, new_page_node, NULL,
9c620e2b 1263 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
cf608ac1 1264 if (err)
e632a938 1265 putback_movable_pages(&pagelist);
cf608ac1 1266 }
742755a1
CL
1267
1268 up_read(&mm->mmap_sem);
1269 return err;
1270}
1271
5e9a0f02
BG
1272/*
1273 * Migrate an array of page address onto an array of nodes and fill
1274 * the corresponding array of status.
1275 */
3268c63e 1276static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1277 unsigned long nr_pages,
1278 const void __user * __user *pages,
1279 const int __user *nodes,
1280 int __user *status, int flags)
1281{
3140a227 1282 struct page_to_node *pm;
3140a227
BG
1283 unsigned long chunk_nr_pages;
1284 unsigned long chunk_start;
1285 int err;
5e9a0f02 1286
3140a227
BG
1287 err = -ENOMEM;
1288 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1289 if (!pm)
5e9a0f02 1290 goto out;
35282a2d
BG
1291
1292 migrate_prep();
1293
5e9a0f02 1294 /*
3140a227
BG
1295 * Store a chunk of page_to_node array in a page,
1296 * but keep the last one as a marker
5e9a0f02 1297 */
3140a227 1298 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
5e9a0f02 1299
3140a227
BG
1300 for (chunk_start = 0;
1301 chunk_start < nr_pages;
1302 chunk_start += chunk_nr_pages) {
1303 int j;
5e9a0f02 1304
3140a227
BG
1305 if (chunk_start + chunk_nr_pages > nr_pages)
1306 chunk_nr_pages = nr_pages - chunk_start;
1307
1308 /* fill the chunk pm with addrs and nodes from user-space */
1309 for (j = 0; j < chunk_nr_pages; j++) {
1310 const void __user *p;
5e9a0f02
BG
1311 int node;
1312
3140a227
BG
1313 err = -EFAULT;
1314 if (get_user(p, pages + j + chunk_start))
1315 goto out_pm;
1316 pm[j].addr = (unsigned long) p;
1317
1318 if (get_user(node, nodes + j + chunk_start))
5e9a0f02
BG
1319 goto out_pm;
1320
1321 err = -ENODEV;
6f5a55f1
LT
1322 if (node < 0 || node >= MAX_NUMNODES)
1323 goto out_pm;
1324
389162c2 1325 if (!node_state(node, N_MEMORY))
5e9a0f02
BG
1326 goto out_pm;
1327
1328 err = -EACCES;
1329 if (!node_isset(node, task_nodes))
1330 goto out_pm;
1331
3140a227
BG
1332 pm[j].node = node;
1333 }
1334
1335 /* End marker for this chunk */
1336 pm[chunk_nr_pages].node = MAX_NUMNODES;
1337
1338 /* Migrate this chunk */
1339 err = do_move_page_to_node_array(mm, pm,
1340 flags & MPOL_MF_MOVE_ALL);
1341 if (err < 0)
1342 goto out_pm;
5e9a0f02 1343
5e9a0f02 1344 /* Return status information */
3140a227
BG
1345 for (j = 0; j < chunk_nr_pages; j++)
1346 if (put_user(pm[j].status, status + j + chunk_start)) {
5e9a0f02 1347 err = -EFAULT;
3140a227
BG
1348 goto out_pm;
1349 }
1350 }
1351 err = 0;
5e9a0f02
BG
1352
1353out_pm:
3140a227 1354 free_page((unsigned long)pm);
5e9a0f02
BG
1355out:
1356 return err;
1357}
1358
742755a1 1359/*
2f007e74 1360 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1361 */
80bba129
BG
1362static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1363 const void __user **pages, int *status)
742755a1 1364{
2f007e74 1365 unsigned long i;
2f007e74 1366
742755a1
CL
1367 down_read(&mm->mmap_sem);
1368
2f007e74 1369 for (i = 0; i < nr_pages; i++) {
80bba129 1370 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1371 struct vm_area_struct *vma;
1372 struct page *page;
c095adbc 1373 int err = -EFAULT;
2f007e74
BG
1374
1375 vma = find_vma(mm, addr);
70384dc6 1376 if (!vma || addr < vma->vm_start)
742755a1
CL
1377 goto set_status;
1378
2f007e74 1379 page = follow_page(vma, addr, 0);
89f5b7da
LT
1380
1381 err = PTR_ERR(page);
1382 if (IS_ERR(page))
1383 goto set_status;
1384
742755a1
CL
1385 err = -ENOENT;
1386 /* Use PageReserved to check for zero page */
b79bc0a0 1387 if (!page || PageReserved(page))
742755a1
CL
1388 goto set_status;
1389
1390 err = page_to_nid(page);
1391set_status:
80bba129
BG
1392 *status = err;
1393
1394 pages++;
1395 status++;
1396 }
1397
1398 up_read(&mm->mmap_sem);
1399}
1400
1401/*
1402 * Determine the nodes of a user array of pages and store it in
1403 * a user array of status.
1404 */
1405static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1406 const void __user * __user *pages,
1407 int __user *status)
1408{
1409#define DO_PAGES_STAT_CHUNK_NR 16
1410 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1411 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1412
87b8d1ad
PA
1413 while (nr_pages) {
1414 unsigned long chunk_nr;
80bba129 1415
87b8d1ad
PA
1416 chunk_nr = nr_pages;
1417 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1418 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1419
1420 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1421 break;
80bba129
BG
1422
1423 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1424
87b8d1ad
PA
1425 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1426 break;
742755a1 1427
87b8d1ad
PA
1428 pages += chunk_nr;
1429 status += chunk_nr;
1430 nr_pages -= chunk_nr;
1431 }
1432 return nr_pages ? -EFAULT : 0;
742755a1
CL
1433}
1434
1435/*
1436 * Move a list of pages in the address space of the currently executing
1437 * process.
1438 */
938bb9f5
HC
1439SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1440 const void __user * __user *, pages,
1441 const int __user *, nodes,
1442 int __user *, status, int, flags)
742755a1 1443{
c69e8d9c 1444 const struct cred *cred = current_cred(), *tcred;
742755a1 1445 struct task_struct *task;
742755a1 1446 struct mm_struct *mm;
5e9a0f02 1447 int err;
3268c63e 1448 nodemask_t task_nodes;
742755a1
CL
1449
1450 /* Check flags */
1451 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1452 return -EINVAL;
1453
1454 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1455 return -EPERM;
1456
1457 /* Find the mm_struct */
a879bf58 1458 rcu_read_lock();
228ebcbe 1459 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1460 if (!task) {
a879bf58 1461 rcu_read_unlock();
742755a1
CL
1462 return -ESRCH;
1463 }
3268c63e 1464 get_task_struct(task);
742755a1
CL
1465
1466 /*
1467 * Check if this process has the right to modify the specified
1468 * process. The right exists if the process has administrative
1469 * capabilities, superuser privileges or the same
1470 * userid as the target process.
1471 */
c69e8d9c 1472 tcred = __task_cred(task);
b38a86eb
EB
1473 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1474 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
742755a1 1475 !capable(CAP_SYS_NICE)) {
c69e8d9c 1476 rcu_read_unlock();
742755a1 1477 err = -EPERM;
5e9a0f02 1478 goto out;
742755a1 1479 }
c69e8d9c 1480 rcu_read_unlock();
742755a1 1481
86c3a764
DQ
1482 err = security_task_movememory(task);
1483 if (err)
5e9a0f02 1484 goto out;
86c3a764 1485
3268c63e
CL
1486 task_nodes = cpuset_mems_allowed(task);
1487 mm = get_task_mm(task);
1488 put_task_struct(task);
1489
6e8b09ea
SL
1490 if (!mm)
1491 return -EINVAL;
1492
1493 if (nodes)
1494 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1495 nodes, status, flags);
1496 else
1497 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1498
742755a1
CL
1499 mmput(mm);
1500 return err;
3268c63e
CL
1501
1502out:
1503 put_task_struct(task);
1504 return err;
742755a1 1505}
742755a1 1506
7039e1db
PZ
1507#ifdef CONFIG_NUMA_BALANCING
1508/*
1509 * Returns true if this is a safe migration target node for misplaced NUMA
1510 * pages. Currently it only checks the watermarks which crude
1511 */
1512static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1513 unsigned long nr_migrate_pages)
7039e1db
PZ
1514{
1515 int z;
1516 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1517 struct zone *zone = pgdat->node_zones + z;
1518
1519 if (!populated_zone(zone))
1520 continue;
1521
6e543d57 1522 if (!zone_reclaimable(zone))
7039e1db
PZ
1523 continue;
1524
1525 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1526 if (!zone_watermark_ok(zone, 0,
1527 high_wmark_pages(zone) +
1528 nr_migrate_pages,
1529 0, 0))
1530 continue;
1531 return true;
1532 }
1533 return false;
1534}
1535
1536static struct page *alloc_misplaced_dst_page(struct page *page,
1537 unsigned long data,
1538 int **result)
1539{
1540 int nid = (int) data;
1541 struct page *newpage;
1542
1543 newpage = alloc_pages_exact_node(nid,
e97ca8e5
JW
1544 (GFP_HIGHUSER_MOVABLE |
1545 __GFP_THISNODE | __GFP_NOMEMALLOC |
1546 __GFP_NORETRY | __GFP_NOWARN) &
7039e1db 1547 ~GFP_IOFS, 0);
bac0382c 1548
7039e1db
PZ
1549 return newpage;
1550}
1551
a8f60772
MG
1552/*
1553 * page migration rate limiting control.
1554 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1555 * window of time. Default here says do not migrate more than 1280M per second.
e14808b4
MG
1556 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1557 * as it is faults that reset the window, pte updates will happen unconditionally
1558 * if there has not been a fault since @pteupdate_interval_millisecs after the
1559 * throttle window closed.
a8f60772
MG
1560 */
1561static unsigned int migrate_interval_millisecs __read_mostly = 100;
e14808b4 1562static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
a8f60772
MG
1563static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1564
e14808b4
MG
1565/* Returns true if NUMA migration is currently rate limited */
1566bool migrate_ratelimited(int node)
1567{
1568 pg_data_t *pgdat = NODE_DATA(node);
1569
1570 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1571 msecs_to_jiffies(pteupdate_interval_millisecs)))
1572 return false;
1573
1574 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1575 return false;
1576
1577 return true;
1578}
1579
b32967ff 1580/* Returns true if the node is migrate rate-limited after the update */
1c30e017
MG
1581static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1582 unsigned long nr_pages)
7039e1db 1583{
a8f60772
MG
1584 /*
1585 * Rate-limit the amount of data that is being migrated to a node.
1586 * Optimal placement is no good if the memory bus is saturated and
1587 * all the time is being spent migrating!
1588 */
a8f60772 1589 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1c5e9c27 1590 spin_lock(&pgdat->numabalancing_migrate_lock);
a8f60772
MG
1591 pgdat->numabalancing_migrate_nr_pages = 0;
1592 pgdat->numabalancing_migrate_next_window = jiffies +
1593 msecs_to_jiffies(migrate_interval_millisecs);
1c5e9c27 1594 spin_unlock(&pgdat->numabalancing_migrate_lock);
a8f60772 1595 }
af1839d7
MG
1596 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1597 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1598 nr_pages);
1c5e9c27 1599 return true;
af1839d7 1600 }
1c5e9c27
MG
1601
1602 /*
1603 * This is an unlocked non-atomic update so errors are possible.
1604 * The consequences are failing to migrate when we potentiall should
1605 * have which is not severe enough to warrant locking. If it is ever
1606 * a problem, it can be converted to a per-cpu counter.
1607 */
1608 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1609 return false;
b32967ff
MG
1610}
1611
1c30e017 1612static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
b32967ff 1613{
340ef390 1614 int page_lru;
a8f60772 1615
309381fe 1616 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
3abef4e6 1617
7039e1db 1618 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1619 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1620 return 0;
7039e1db 1621
340ef390
HD
1622 if (isolate_lru_page(page))
1623 return 0;
7039e1db 1624
340ef390
HD
1625 /*
1626 * migrate_misplaced_transhuge_page() skips page migration's usual
1627 * check on page_count(), so we must do it here, now that the page
1628 * has been isolated: a GUP pin, or any other pin, prevents migration.
1629 * The expected page count is 3: 1 for page's mapcount and 1 for the
1630 * caller's pin and 1 for the reference taken by isolate_lru_page().
1631 */
1632 if (PageTransHuge(page) && page_count(page) != 3) {
1633 putback_lru_page(page);
1634 return 0;
7039e1db
PZ
1635 }
1636
340ef390
HD
1637 page_lru = page_is_file_cache(page);
1638 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1639 hpage_nr_pages(page));
1640
149c33e1 1641 /*
340ef390
HD
1642 * Isolating the page has taken another reference, so the
1643 * caller's reference can be safely dropped without the page
1644 * disappearing underneath us during migration.
149c33e1
MG
1645 */
1646 put_page(page);
340ef390 1647 return 1;
b32967ff
MG
1648}
1649
de466bd6
MG
1650bool pmd_trans_migrating(pmd_t pmd)
1651{
1652 struct page *page = pmd_page(pmd);
1653 return PageLocked(page);
1654}
1655
1656void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1657{
1658 struct page *page = pmd_page(*pmd);
1659 wait_on_page_locked(page);
1660}
1661
b32967ff
MG
1662/*
1663 * Attempt to migrate a misplaced page to the specified destination
1664 * node. Caller is expected to have an elevated reference count on
1665 * the page that will be dropped by this function before returning.
1666 */
1bc115d8
MG
1667int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1668 int node)
b32967ff
MG
1669{
1670 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1671 int isolated;
b32967ff
MG
1672 int nr_remaining;
1673 LIST_HEAD(migratepages);
1674
1675 /*
1bc115d8
MG
1676 * Don't migrate file pages that are mapped in multiple processes
1677 * with execute permissions as they are probably shared libraries.
b32967ff 1678 */
1bc115d8
MG
1679 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1680 (vma->vm_flags & VM_EXEC))
b32967ff 1681 goto out;
b32967ff
MG
1682
1683 /*
1684 * Rate-limit the amount of data that is being migrated to a node.
1685 * Optimal placement is no good if the memory bus is saturated and
1686 * all the time is being spent migrating!
1687 */
340ef390 1688 if (numamigrate_update_ratelimit(pgdat, 1))
b32967ff 1689 goto out;
b32967ff
MG
1690
1691 isolated = numamigrate_isolate_page(pgdat, page);
1692 if (!isolated)
1693 goto out;
1694
1695 list_add(&page->lru, &migratepages);
9c620e2b 1696 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
68711a74
DR
1697 NULL, node, MIGRATE_ASYNC,
1698 MR_NUMA_MISPLACED);
b32967ff 1699 if (nr_remaining) {
59c82b70
JK
1700 if (!list_empty(&migratepages)) {
1701 list_del(&page->lru);
1702 dec_zone_page_state(page, NR_ISOLATED_ANON +
1703 page_is_file_cache(page));
1704 putback_lru_page(page);
1705 }
b32967ff
MG
1706 isolated = 0;
1707 } else
1708 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1709 BUG_ON(!list_empty(&migratepages));
7039e1db 1710 return isolated;
340ef390
HD
1711
1712out:
1713 put_page(page);
1714 return 0;
7039e1db 1715}
220018d3 1716#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1717
220018d3 1718#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1719/*
1720 * Migrates a THP to a given target node. page must be locked and is unlocked
1721 * before returning.
1722 */
b32967ff
MG
1723int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1724 struct vm_area_struct *vma,
1725 pmd_t *pmd, pmd_t entry,
1726 unsigned long address,
1727 struct page *page, int node)
1728{
c4088ebd 1729 spinlock_t *ptl;
b32967ff
MG
1730 pg_data_t *pgdat = NODE_DATA(node);
1731 int isolated = 0;
1732 struct page *new_page = NULL;
b32967ff 1733 int page_lru = page_is_file_cache(page);
f714f4f2
MG
1734 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1735 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
2b4847e7 1736 pmd_t orig_entry;
b32967ff 1737
b32967ff
MG
1738 /*
1739 * Rate-limit the amount of data that is being migrated to a node.
1740 * Optimal placement is no good if the memory bus is saturated and
1741 * all the time is being spent migrating!
1742 */
d28d4335 1743 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
b32967ff
MG
1744 goto out_dropref;
1745
1746 new_page = alloc_pages_node(node,
e97ca8e5
JW
1747 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1748 HPAGE_PMD_ORDER);
340ef390
HD
1749 if (!new_page)
1750 goto out_fail;
1751
b32967ff 1752 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 1753 if (!isolated) {
b32967ff 1754 put_page(new_page);
340ef390 1755 goto out_fail;
b32967ff
MG
1756 }
1757
b0943d61
MG
1758 if (mm_tlb_flush_pending(mm))
1759 flush_tlb_range(vma, mmun_start, mmun_end);
1760
b32967ff
MG
1761 /* Prepare a page as a migration target */
1762 __set_page_locked(new_page);
1763 SetPageSwapBacked(new_page);
1764
1765 /* anon mapping, we can simply copy page->mapping to the new page: */
1766 new_page->mapping = page->mapping;
1767 new_page->index = page->index;
1768 migrate_page_copy(new_page, page);
1769 WARN_ON(PageLRU(new_page));
1770
1771 /* Recheck the target PMD */
f714f4f2 1772 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 1773 ptl = pmd_lock(mm, pmd);
2b4847e7
MG
1774 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1775fail_putback:
c4088ebd 1776 spin_unlock(ptl);
f714f4f2 1777 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b32967ff
MG
1778
1779 /* Reverse changes made by migrate_page_copy() */
1780 if (TestClearPageActive(new_page))
1781 SetPageActive(page);
1782 if (TestClearPageUnevictable(new_page))
1783 SetPageUnevictable(page);
1784 mlock_migrate_page(page, new_page);
1785
1786 unlock_page(new_page);
1787 put_page(new_page); /* Free it */
1788
a54a407f
MG
1789 /* Retake the callers reference and putback on LRU */
1790 get_page(page);
b32967ff 1791 putback_lru_page(page);
a54a407f
MG
1792 mod_zone_page_state(page_zone(page),
1793 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
eb4489f6
MG
1794
1795 goto out_unlock;
b32967ff
MG
1796 }
1797
2b4847e7 1798 orig_entry = *pmd;
b32967ff 1799 entry = mk_pmd(new_page, vma->vm_page_prot);
b32967ff 1800 entry = pmd_mkhuge(entry);
2b4847e7 1801 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
b32967ff 1802
2b4847e7
MG
1803 /*
1804 * Clear the old entry under pagetable lock and establish the new PTE.
1805 * Any parallel GUP will either observe the old page blocking on the
1806 * page lock, block on the page table lock or observe the new page.
1807 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1808 * guarantee the copy is visible before the pagetable update.
1809 */
f714f4f2 1810 flush_cache_range(vma, mmun_start, mmun_end);
11de9927 1811 page_add_anon_rmap(new_page, vma, mmun_start);
34ee645e 1812 pmdp_clear_flush_notify(vma, mmun_start, pmd);
f714f4f2
MG
1813 set_pmd_at(mm, mmun_start, pmd, entry);
1814 flush_tlb_range(vma, mmun_start, mmun_end);
ce4a9cc5 1815 update_mmu_cache_pmd(vma, address, &entry);
2b4847e7
MG
1816
1817 if (page_count(page) != 2) {
f714f4f2
MG
1818 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1819 flush_tlb_range(vma, mmun_start, mmun_end);
34ee645e 1820 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2b4847e7
MG
1821 update_mmu_cache_pmd(vma, address, &entry);
1822 page_remove_rmap(new_page);
1823 goto fail_putback;
1824 }
1825
0a31bc97
JW
1826 mem_cgroup_migrate(page, new_page, false);
1827
b32967ff 1828 page_remove_rmap(page);
2b4847e7 1829
c4088ebd 1830 spin_unlock(ptl);
f714f4f2 1831 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b32967ff 1832
11de9927
MG
1833 /* Take an "isolate" reference and put new page on the LRU. */
1834 get_page(new_page);
1835 putback_lru_page(new_page);
1836
b32967ff
MG
1837 unlock_page(new_page);
1838 unlock_page(page);
1839 put_page(page); /* Drop the rmap reference */
1840 put_page(page); /* Drop the LRU isolation reference */
1841
1842 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1843 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1844
b32967ff
MG
1845 mod_zone_page_state(page_zone(page),
1846 NR_ISOLATED_ANON + page_lru,
1847 -HPAGE_PMD_NR);
1848 return isolated;
1849
340ef390
HD
1850out_fail:
1851 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
b32967ff 1852out_dropref:
2b4847e7
MG
1853 ptl = pmd_lock(mm, pmd);
1854 if (pmd_same(*pmd, entry)) {
1855 entry = pmd_mknonnuma(entry);
f714f4f2 1856 set_pmd_at(mm, mmun_start, pmd, entry);
2b4847e7
MG
1857 update_mmu_cache_pmd(vma, address, &entry);
1858 }
1859 spin_unlock(ptl);
a54a407f 1860
eb4489f6 1861out_unlock:
340ef390 1862 unlock_page(page);
b32967ff 1863 put_page(page);
b32967ff
MG
1864 return 0;
1865}
7039e1db
PZ
1866#endif /* CONFIG_NUMA_BALANCING */
1867
1868#endif /* CONFIG_NUMA */