]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
rmap: extend try_to_unmap() to be usable by split_huge_page()
[mirror_ubuntu-jammy-kernel.git] / mm / migrate.c
CommitLineData
b20a3503 1/*
14e0f9bc 2 * Memory Migration functionality - linux/mm/migrate.c
b20a3503
CL
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>
42cb14b1 33#include <linux/backing-dev.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>
33c3fc71 40#include <linux/page_idle.h>
d435edca 41#include <linux/page_owner.h>
b20a3503 42
0d1836c3
MN
43#include <asm/tlbflush.h>
44
7b2a2d4a
MG
45#define CREATE_TRACE_POINTS
46#include <trace/events/migrate.h>
47
b20a3503
CL
48#include "internal.h"
49
b20a3503 50/*
742755a1 51 * migrate_prep() needs to be called before we start compiling a list of pages
748446bb
MG
52 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
53 * undesirable, use migrate_prep_local()
b20a3503
CL
54 */
55int migrate_prep(void)
56{
b20a3503
CL
57 /*
58 * Clear the LRU lists so pages can be isolated.
59 * Note that pages may be moved off the LRU after we have
60 * drained them. Those pages will fail to migrate like other
61 * pages that may be busy.
62 */
63 lru_add_drain_all();
64
65 return 0;
66}
67
748446bb
MG
68/* Do the necessary work of migrate_prep but not if it involves other CPUs */
69int migrate_prep_local(void)
70{
71 lru_add_drain();
72
73 return 0;
74}
75
5733c7d1
RA
76/*
77 * Put previously isolated pages back onto the appropriate lists
78 * from where they were once taken off for compaction/migration.
79 *
59c82b70
JK
80 * This function shall be used whenever the isolated pageset has been
81 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
82 * and isolate_huge_page().
5733c7d1
RA
83 */
84void putback_movable_pages(struct list_head *l)
85{
86 struct page *page;
87 struct page *page2;
88
b20a3503 89 list_for_each_entry_safe(page, page2, l, lru) {
31caf665
NH
90 if (unlikely(PageHuge(page))) {
91 putback_active_hugepage(page);
92 continue;
93 }
e24f0b8f 94 list_del(&page->lru);
a731286d 95 dec_zone_page_state(page, NR_ISOLATED_ANON +
6c0b1351 96 page_is_file_cache(page));
117aad1e 97 if (unlikely(isolated_balloon_page(page)))
bf6bddf1
RA
98 balloon_page_putback(page);
99 else
100 putback_lru_page(page);
b20a3503 101 }
b20a3503
CL
102}
103
0697212a
CL
104/*
105 * Restore a potential migration pte to a working pte entry
106 */
e9995ef9
HD
107static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
108 unsigned long addr, void *old)
0697212a
CL
109{
110 struct mm_struct *mm = vma->vm_mm;
111 swp_entry_t entry;
0697212a
CL
112 pmd_t *pmd;
113 pte_t *ptep, pte;
114 spinlock_t *ptl;
115
290408d4
NH
116 if (unlikely(PageHuge(new))) {
117 ptep = huge_pte_offset(mm, addr);
118 if (!ptep)
119 goto out;
cb900f41 120 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
290408d4 121 } else {
6219049a
BL
122 pmd = mm_find_pmd(mm, addr);
123 if (!pmd)
290408d4 124 goto out;
0697212a 125
290408d4 126 ptep = pte_offset_map(pmd, addr);
0697212a 127
486cf46f
HD
128 /*
129 * Peek to check is_swap_pte() before taking ptlock? No, we
130 * can race mremap's move_ptes(), which skips anon_vma lock.
131 */
290408d4
NH
132
133 ptl = pte_lockptr(mm, pmd);
134 }
0697212a 135
0697212a
CL
136 spin_lock(ptl);
137 pte = *ptep;
138 if (!is_swap_pte(pte))
e9995ef9 139 goto unlock;
0697212a
CL
140
141 entry = pte_to_swp_entry(pte);
142
e9995ef9
HD
143 if (!is_migration_entry(entry) ||
144 migration_entry_to_page(entry) != old)
145 goto unlock;
0697212a 146
0697212a
CL
147 get_page(new);
148 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
c3d16e16
CG
149 if (pte_swp_soft_dirty(*ptep))
150 pte = pte_mksoft_dirty(pte);
d3cb8bf6
MG
151
152 /* Recheck VMA as permissions can change since migration started */
0697212a 153 if (is_write_migration_entry(entry))
d3cb8bf6
MG
154 pte = maybe_mkwrite(pte, vma);
155
3ef8fd7f 156#ifdef CONFIG_HUGETLB_PAGE
be7517d6 157 if (PageHuge(new)) {
290408d4 158 pte = pte_mkhuge(pte);
be7517d6
TL
159 pte = arch_make_huge_pte(pte, vma, new, 0);
160 }
3ef8fd7f 161#endif
c2cc499c 162 flush_dcache_page(new);
0697212a 163 set_pte_at(mm, addr, ptep, pte);
04e62a29 164
290408d4
NH
165 if (PageHuge(new)) {
166 if (PageAnon(new))
167 hugepage_add_anon_rmap(new, vma, addr);
168 else
53f9263b 169 page_dup_rmap(new, true);
290408d4 170 } else if (PageAnon(new))
d281ee61 171 page_add_anon_rmap(new, vma, addr, false);
04e62a29
CL
172 else
173 page_add_file_rmap(new);
174
51afb12b
HD
175 if (vma->vm_flags & VM_LOCKED)
176 mlock_vma_page(new);
177
04e62a29 178 /* No need to invalidate - it was non-present before */
4b3073e1 179 update_mmu_cache(vma, addr, ptep);
e9995ef9 180unlock:
0697212a 181 pte_unmap_unlock(ptep, ptl);
e9995ef9
HD
182out:
183 return SWAP_AGAIN;
0697212a
CL
184}
185
04e62a29
CL
186/*
187 * Get rid of all migration entries and replace them by
188 * references to the indicated page.
189 */
190static void remove_migration_ptes(struct page *old, struct page *new)
191{
051ac83a
JK
192 struct rmap_walk_control rwc = {
193 .rmap_one = remove_migration_pte,
194 .arg = old,
195 };
196
197 rmap_walk(new, &rwc);
04e62a29
CL
198}
199
0697212a
CL
200/*
201 * Something used the pte of a page under migration. We need to
202 * get to the page and wait until migration is finished.
203 * When we return from this function the fault will be retried.
0697212a 204 */
e66f17ff 205void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
30dad309 206 spinlock_t *ptl)
0697212a 207{
30dad309 208 pte_t pte;
0697212a
CL
209 swp_entry_t entry;
210 struct page *page;
211
30dad309 212 spin_lock(ptl);
0697212a
CL
213 pte = *ptep;
214 if (!is_swap_pte(pte))
215 goto out;
216
217 entry = pte_to_swp_entry(pte);
218 if (!is_migration_entry(entry))
219 goto out;
220
221 page = migration_entry_to_page(entry);
222
e286781d
NP
223 /*
224 * Once radix-tree replacement of page migration started, page_count
225 * *must* be zero. And, we don't want to call wait_on_page_locked()
226 * against a page without get_page().
227 * So, we use get_page_unless_zero(), here. Even failed, page fault
228 * will occur again.
229 */
230 if (!get_page_unless_zero(page))
231 goto out;
0697212a
CL
232 pte_unmap_unlock(ptep, ptl);
233 wait_on_page_locked(page);
234 put_page(page);
235 return;
236out:
237 pte_unmap_unlock(ptep, ptl);
238}
239
30dad309
NH
240void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
241 unsigned long address)
242{
243 spinlock_t *ptl = pte_lockptr(mm, pmd);
244 pte_t *ptep = pte_offset_map(pmd, address);
245 __migration_entry_wait(mm, ptep, ptl);
246}
247
cb900f41
KS
248void migration_entry_wait_huge(struct vm_area_struct *vma,
249 struct mm_struct *mm, pte_t *pte)
30dad309 250{
cb900f41 251 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
30dad309
NH
252 __migration_entry_wait(mm, pte, ptl);
253}
254
b969c4ab
MG
255#ifdef CONFIG_BLOCK
256/* Returns true if all buffers are successfully locked */
a6bc32b8
MG
257static bool buffer_migrate_lock_buffers(struct buffer_head *head,
258 enum migrate_mode mode)
b969c4ab
MG
259{
260 struct buffer_head *bh = head;
261
262 /* Simple case, sync compaction */
a6bc32b8 263 if (mode != MIGRATE_ASYNC) {
b969c4ab
MG
264 do {
265 get_bh(bh);
266 lock_buffer(bh);
267 bh = bh->b_this_page;
268
269 } while (bh != head);
270
271 return true;
272 }
273
274 /* async case, we cannot block on lock_buffer so use trylock_buffer */
275 do {
276 get_bh(bh);
277 if (!trylock_buffer(bh)) {
278 /*
279 * We failed to lock the buffer and cannot stall in
280 * async migration. Release the taken locks
281 */
282 struct buffer_head *failed_bh = bh;
283 put_bh(failed_bh);
284 bh = head;
285 while (bh != failed_bh) {
286 unlock_buffer(bh);
287 put_bh(bh);
288 bh = bh->b_this_page;
289 }
290 return false;
291 }
292
293 bh = bh->b_this_page;
294 } while (bh != head);
295 return true;
296}
297#else
298static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
a6bc32b8 299 enum migrate_mode mode)
b969c4ab
MG
300{
301 return true;
302}
303#endif /* CONFIG_BLOCK */
304
b20a3503 305/*
c3fcf8a5 306 * Replace the page in the mapping.
5b5c7120
CL
307 *
308 * The number of remaining references must be:
309 * 1 for anonymous pages without a mapping
310 * 2 for pages with a mapping
266cf658 311 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
b20a3503 312 */
36bc08cc 313int migrate_page_move_mapping(struct address_space *mapping,
b969c4ab 314 struct page *newpage, struct page *page,
8e321fef
BL
315 struct buffer_head *head, enum migrate_mode mode,
316 int extra_count)
b20a3503 317{
42cb14b1
HD
318 struct zone *oldzone, *newzone;
319 int dirty;
8e321fef 320 int expected_count = 1 + extra_count;
7cf9c2c7 321 void **pslot;
b20a3503 322
6c5240ae 323 if (!mapping) {
0e8c7d0f 324 /* Anonymous page without mapping */
8e321fef 325 if (page_count(page) != expected_count)
6c5240ae 326 return -EAGAIN;
cf4b769a
HD
327
328 /* No turning back from here */
cf4b769a
HD
329 newpage->index = page->index;
330 newpage->mapping = page->mapping;
331 if (PageSwapBacked(page))
332 SetPageSwapBacked(newpage);
333
78bd5209 334 return MIGRATEPAGE_SUCCESS;
6c5240ae
CL
335 }
336
42cb14b1
HD
337 oldzone = page_zone(page);
338 newzone = page_zone(newpage);
339
19fd6231 340 spin_lock_irq(&mapping->tree_lock);
b20a3503 341
7cf9c2c7
NP
342 pslot = radix_tree_lookup_slot(&mapping->page_tree,
343 page_index(page));
b20a3503 344
8e321fef 345 expected_count += 1 + page_has_private(page);
e286781d 346 if (page_count(page) != expected_count ||
29c1f677 347 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
19fd6231 348 spin_unlock_irq(&mapping->tree_lock);
e23ca00b 349 return -EAGAIN;
b20a3503
CL
350 }
351
fe896d18 352 if (!page_ref_freeze(page, expected_count)) {
19fd6231 353 spin_unlock_irq(&mapping->tree_lock);
e286781d
NP
354 return -EAGAIN;
355 }
356
b969c4ab
MG
357 /*
358 * In the async migration case of moving a page with buffers, lock the
359 * buffers using trylock before the mapping is moved. If the mapping
360 * was moved, we later failed to lock the buffers and could not move
361 * the mapping back due to an elevated page count, we would have to
362 * block waiting on other references to be dropped.
363 */
a6bc32b8
MG
364 if (mode == MIGRATE_ASYNC && head &&
365 !buffer_migrate_lock_buffers(head, mode)) {
fe896d18 366 page_ref_unfreeze(page, expected_count);
b969c4ab
MG
367 spin_unlock_irq(&mapping->tree_lock);
368 return -EAGAIN;
369 }
370
b20a3503 371 /*
cf4b769a
HD
372 * Now we know that no one else is looking at the page:
373 * no turning back from here.
b20a3503 374 */
cf4b769a
HD
375 newpage->index = page->index;
376 newpage->mapping = page->mapping;
377 if (PageSwapBacked(page))
378 SetPageSwapBacked(newpage);
379
7cf9c2c7 380 get_page(newpage); /* add cache reference */
b20a3503
CL
381 if (PageSwapCache(page)) {
382 SetPageSwapCache(newpage);
383 set_page_private(newpage, page_private(page));
384 }
385
42cb14b1
HD
386 /* Move dirty while page refs frozen and newpage not yet exposed */
387 dirty = PageDirty(page);
388 if (dirty) {
389 ClearPageDirty(page);
390 SetPageDirty(newpage);
391 }
392
7cf9c2c7
NP
393 radix_tree_replace_slot(pslot, newpage);
394
395 /*
937a94c9
JG
396 * Drop cache reference from old page by unfreezing
397 * to one less reference.
7cf9c2c7
NP
398 * We know this isn't the last reference.
399 */
fe896d18 400 page_ref_unfreeze(page, expected_count - 1);
7cf9c2c7 401
42cb14b1
HD
402 spin_unlock(&mapping->tree_lock);
403 /* Leave irq disabled to prevent preemption while updating stats */
404
0e8c7d0f
CL
405 /*
406 * If moved to a different zone then also account
407 * the page for that zone. Other VM counters will be
408 * taken care of when we establish references to the
409 * new page and drop references to the old page.
410 *
411 * Note that anonymous pages are accounted for
412 * via NR_FILE_PAGES and NR_ANON_PAGES if they
413 * are mapped to swap space.
414 */
42cb14b1
HD
415 if (newzone != oldzone) {
416 __dec_zone_state(oldzone, NR_FILE_PAGES);
417 __inc_zone_state(newzone, NR_FILE_PAGES);
418 if (PageSwapBacked(page) && !PageSwapCache(page)) {
419 __dec_zone_state(oldzone, NR_SHMEM);
420 __inc_zone_state(newzone, NR_SHMEM);
421 }
422 if (dirty && mapping_cap_account_dirty(mapping)) {
423 __dec_zone_state(oldzone, NR_FILE_DIRTY);
424 __inc_zone_state(newzone, NR_FILE_DIRTY);
425 }
4b02108a 426 }
42cb14b1 427 local_irq_enable();
b20a3503 428
78bd5209 429 return MIGRATEPAGE_SUCCESS;
b20a3503 430}
b20a3503 431
290408d4
NH
432/*
433 * The expected number of remaining references is the same as that
434 * of migrate_page_move_mapping().
435 */
436int migrate_huge_page_move_mapping(struct address_space *mapping,
437 struct page *newpage, struct page *page)
438{
439 int expected_count;
440 void **pslot;
441
290408d4
NH
442 spin_lock_irq(&mapping->tree_lock);
443
444 pslot = radix_tree_lookup_slot(&mapping->page_tree,
445 page_index(page));
446
447 expected_count = 2 + page_has_private(page);
448 if (page_count(page) != expected_count ||
29c1f677 449 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
290408d4
NH
450 spin_unlock_irq(&mapping->tree_lock);
451 return -EAGAIN;
452 }
453
fe896d18 454 if (!page_ref_freeze(page, expected_count)) {
290408d4
NH
455 spin_unlock_irq(&mapping->tree_lock);
456 return -EAGAIN;
457 }
458
cf4b769a
HD
459 newpage->index = page->index;
460 newpage->mapping = page->mapping;
6a93ca8f 461
290408d4
NH
462 get_page(newpage);
463
464 radix_tree_replace_slot(pslot, newpage);
465
fe896d18 466 page_ref_unfreeze(page, expected_count - 1);
290408d4
NH
467
468 spin_unlock_irq(&mapping->tree_lock);
6a93ca8f 469
78bd5209 470 return MIGRATEPAGE_SUCCESS;
290408d4
NH
471}
472
30b0a105
DH
473/*
474 * Gigantic pages are so large that we do not guarantee that page++ pointer
475 * arithmetic will work across the entire page. We need something more
476 * specialized.
477 */
478static void __copy_gigantic_page(struct page *dst, struct page *src,
479 int nr_pages)
480{
481 int i;
482 struct page *dst_base = dst;
483 struct page *src_base = src;
484
485 for (i = 0; i < nr_pages; ) {
486 cond_resched();
487 copy_highpage(dst, src);
488
489 i++;
490 dst = mem_map_next(dst, dst_base, i);
491 src = mem_map_next(src, src_base, i);
492 }
493}
494
495static void copy_huge_page(struct page *dst, struct page *src)
496{
497 int i;
498 int nr_pages;
499
500 if (PageHuge(src)) {
501 /* hugetlbfs page */
502 struct hstate *h = page_hstate(src);
503 nr_pages = pages_per_huge_page(h);
504
505 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
506 __copy_gigantic_page(dst, src, nr_pages);
507 return;
508 }
509 } else {
510 /* thp page */
511 BUG_ON(!PageTransHuge(src));
512 nr_pages = hpage_nr_pages(src);
513 }
514
515 for (i = 0; i < nr_pages; i++) {
516 cond_resched();
517 copy_highpage(dst + i, src + i);
518 }
519}
520
b20a3503
CL
521/*
522 * Copy the page to its new location
523 */
290408d4 524void migrate_page_copy(struct page *newpage, struct page *page)
b20a3503 525{
7851a45c
RR
526 int cpupid;
527
b32967ff 528 if (PageHuge(page) || PageTransHuge(page))
290408d4
NH
529 copy_huge_page(newpage, page);
530 else
531 copy_highpage(newpage, page);
b20a3503
CL
532
533 if (PageError(page))
534 SetPageError(newpage);
535 if (PageReferenced(page))
536 SetPageReferenced(newpage);
537 if (PageUptodate(page))
538 SetPageUptodate(newpage);
894bc310 539 if (TestClearPageActive(page)) {
309381fe 540 VM_BUG_ON_PAGE(PageUnevictable(page), page);
b20a3503 541 SetPageActive(newpage);
418b27ef
LS
542 } else if (TestClearPageUnevictable(page))
543 SetPageUnevictable(newpage);
b20a3503
CL
544 if (PageChecked(page))
545 SetPageChecked(newpage);
546 if (PageMappedToDisk(page))
547 SetPageMappedToDisk(newpage);
548
42cb14b1
HD
549 /* Move dirty on pages not done by migrate_page_move_mapping() */
550 if (PageDirty(page))
551 SetPageDirty(newpage);
b20a3503 552
33c3fc71
VD
553 if (page_is_young(page))
554 set_page_young(newpage);
555 if (page_is_idle(page))
556 set_page_idle(newpage);
557
7851a45c
RR
558 /*
559 * Copy NUMA information to the new page, to prevent over-eager
560 * future migrations of this same page.
561 */
562 cpupid = page_cpupid_xchg_last(page, -1);
563 page_cpupid_xchg_last(newpage, cpupid);
564
e9995ef9 565 ksm_migrate_page(newpage, page);
c8d6553b
HD
566 /*
567 * Please do not reorder this without considering how mm/ksm.c's
568 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
569 */
b3b3a99c
NH
570 if (PageSwapCache(page))
571 ClearPageSwapCache(page);
b20a3503
CL
572 ClearPagePrivate(page);
573 set_page_private(page, 0);
b20a3503
CL
574
575 /*
576 * If any waiters have accumulated on the new page then
577 * wake them up.
578 */
579 if (PageWriteback(newpage))
580 end_page_writeback(newpage);
d435edca
VB
581
582 copy_page_owner(page, newpage);
74485cf2
JW
583
584 mem_cgroup_migrate(page, newpage);
b20a3503 585}
b20a3503 586
1d8b85cc
CL
587/************************************************************
588 * Migration functions
589 ***********************************************************/
590
b20a3503
CL
591/*
592 * Common logic to directly migrate a single page suitable for
266cf658 593 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
594 *
595 * Pages are locked upon entry and exit.
596 */
2d1db3b1 597int migrate_page(struct address_space *mapping,
a6bc32b8
MG
598 struct page *newpage, struct page *page,
599 enum migrate_mode mode)
b20a3503
CL
600{
601 int rc;
602
603 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
604
8e321fef 605 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
b20a3503 606
78bd5209 607 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
608 return rc;
609
610 migrate_page_copy(newpage, page);
78bd5209 611 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
612}
613EXPORT_SYMBOL(migrate_page);
614
9361401e 615#ifdef CONFIG_BLOCK
1d8b85cc
CL
616/*
617 * Migration function for pages with buffers. This function can only be used
618 * if the underlying filesystem guarantees that no other references to "page"
619 * exist.
620 */
2d1db3b1 621int buffer_migrate_page(struct address_space *mapping,
a6bc32b8 622 struct page *newpage, struct page *page, enum migrate_mode mode)
1d8b85cc 623{
1d8b85cc
CL
624 struct buffer_head *bh, *head;
625 int rc;
626
1d8b85cc 627 if (!page_has_buffers(page))
a6bc32b8 628 return migrate_page(mapping, newpage, page, mode);
1d8b85cc
CL
629
630 head = page_buffers(page);
631
8e321fef 632 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
1d8b85cc 633
78bd5209 634 if (rc != MIGRATEPAGE_SUCCESS)
1d8b85cc
CL
635 return rc;
636
b969c4ab
MG
637 /*
638 * In the async case, migrate_page_move_mapping locked the buffers
639 * with an IRQ-safe spinlock held. In the sync case, the buffers
640 * need to be locked now
641 */
a6bc32b8
MG
642 if (mode != MIGRATE_ASYNC)
643 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
1d8b85cc
CL
644
645 ClearPagePrivate(page);
646 set_page_private(newpage, page_private(page));
647 set_page_private(page, 0);
648 put_page(page);
649 get_page(newpage);
650
651 bh = head;
652 do {
653 set_bh_page(bh, newpage, bh_offset(bh));
654 bh = bh->b_this_page;
655
656 } while (bh != head);
657
658 SetPagePrivate(newpage);
659
660 migrate_page_copy(newpage, page);
661
662 bh = head;
663 do {
664 unlock_buffer(bh);
665 put_bh(bh);
666 bh = bh->b_this_page;
667
668 } while (bh != head);
669
78bd5209 670 return MIGRATEPAGE_SUCCESS;
1d8b85cc
CL
671}
672EXPORT_SYMBOL(buffer_migrate_page);
9361401e 673#endif
1d8b85cc 674
04e62a29
CL
675/*
676 * Writeback a page to clean the dirty state
677 */
678static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 679{
04e62a29
CL
680 struct writeback_control wbc = {
681 .sync_mode = WB_SYNC_NONE,
682 .nr_to_write = 1,
683 .range_start = 0,
684 .range_end = LLONG_MAX,
04e62a29
CL
685 .for_reclaim = 1
686 };
687 int rc;
688
689 if (!mapping->a_ops->writepage)
690 /* No write method for the address space */
691 return -EINVAL;
692
693 if (!clear_page_dirty_for_io(page))
694 /* Someone else already triggered a write */
695 return -EAGAIN;
696
8351a6e4 697 /*
04e62a29
CL
698 * A dirty page may imply that the underlying filesystem has
699 * the page on some queue. So the page must be clean for
700 * migration. Writeout may mean we loose the lock and the
701 * page state is no longer what we checked for earlier.
702 * At this point we know that the migration attempt cannot
703 * be successful.
8351a6e4 704 */
04e62a29 705 remove_migration_ptes(page, page);
8351a6e4 706
04e62a29 707 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 708
04e62a29
CL
709 if (rc != AOP_WRITEPAGE_ACTIVATE)
710 /* unlocked. Relock */
711 lock_page(page);
712
bda8550d 713 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
714}
715
716/*
717 * Default handling if a filesystem does not provide a migration function.
718 */
719static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 720 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 721{
b969c4ab 722 if (PageDirty(page)) {
a6bc32b8
MG
723 /* Only writeback pages in full synchronous migration */
724 if (mode != MIGRATE_SYNC)
b969c4ab 725 return -EBUSY;
04e62a29 726 return writeout(mapping, page);
b969c4ab 727 }
8351a6e4
CL
728
729 /*
730 * Buffers may be managed in a filesystem specific way.
731 * We must have no buffers or drop them.
732 */
266cf658 733 if (page_has_private(page) &&
8351a6e4
CL
734 !try_to_release_page(page, GFP_KERNEL))
735 return -EAGAIN;
736
a6bc32b8 737 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
738}
739
e24f0b8f
CL
740/*
741 * Move a page to a newly allocated page
742 * The page is locked and all ptes have been successfully removed.
743 *
744 * The new page will have replaced the old page if this function
745 * is successful.
894bc310
LS
746 *
747 * Return value:
748 * < 0 - error code
78bd5209 749 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 750 */
3fe2011f 751static int move_to_new_page(struct page *newpage, struct page *page,
5c3f9a67 752 enum migrate_mode mode)
e24f0b8f
CL
753{
754 struct address_space *mapping;
755 int rc;
756
7db7671f
HD
757 VM_BUG_ON_PAGE(!PageLocked(page), page);
758 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
e24f0b8f 759
e24f0b8f
CL
760 mapping = page_mapping(page);
761 if (!mapping)
a6bc32b8 762 rc = migrate_page(mapping, newpage, page, mode);
b969c4ab 763 else if (mapping->a_ops->migratepage)
e24f0b8f 764 /*
b969c4ab
MG
765 * Most pages have a mapping and most filesystems provide a
766 * migratepage callback. Anonymous pages are part of swap
767 * space which also has its own migratepage callback. This
768 * is the most common path for page migration.
e24f0b8f 769 */
5c3f9a67 770 rc = mapping->a_ops->migratepage(mapping, newpage, page, mode);
b969c4ab 771 else
a6bc32b8 772 rc = fallback_migrate_page(mapping, newpage, page, mode);
e24f0b8f 773
5c3f9a67
HD
774 /*
775 * When successful, old pagecache page->mapping must be cleared before
776 * page is freed; but stats require that PageAnon be left as PageAnon.
777 */
778 if (rc == MIGRATEPAGE_SUCCESS) {
5c3f9a67
HD
779 if (!PageAnon(page))
780 page->mapping = NULL;
3fe2011f 781 }
e24f0b8f
CL
782 return rc;
783}
784
0dabec93 785static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 786 int force, enum migrate_mode mode)
e24f0b8f 787{
0dabec93 788 int rc = -EAGAIN;
2ebba6b7 789 int page_was_mapped = 0;
3f6c8272 790 struct anon_vma *anon_vma = NULL;
95a402c3 791
529ae9aa 792 if (!trylock_page(page)) {
a6bc32b8 793 if (!force || mode == MIGRATE_ASYNC)
0dabec93 794 goto out;
3e7d3449
MG
795
796 /*
797 * It's not safe for direct compaction to call lock_page.
798 * For example, during page readahead pages are added locked
799 * to the LRU. Later, when the IO completes the pages are
800 * marked uptodate and unlocked. However, the queueing
801 * could be merging multiple pages for one bio (e.g.
802 * mpage_readpages). If an allocation happens for the
803 * second or third page, the process can end up locking
804 * the same page twice and deadlocking. Rather than
805 * trying to be clever about what pages can be locked,
806 * avoid the use of lock_page for direct compaction
807 * altogether.
808 */
809 if (current->flags & PF_MEMALLOC)
0dabec93 810 goto out;
3e7d3449 811
e24f0b8f
CL
812 lock_page(page);
813 }
814
815 if (PageWriteback(page)) {
11bc82d6 816 /*
fed5b64a 817 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
818 * necessary to wait for PageWriteback. In the async case,
819 * the retry loop is too short and in the sync-light case,
820 * the overhead of stalling is too much
11bc82d6 821 */
a6bc32b8 822 if (mode != MIGRATE_SYNC) {
11bc82d6 823 rc = -EBUSY;
0a31bc97 824 goto out_unlock;
11bc82d6
AA
825 }
826 if (!force)
0a31bc97 827 goto out_unlock;
e24f0b8f
CL
828 wait_on_page_writeback(page);
829 }
03f15c86 830
e24f0b8f 831 /*
dc386d4d
KH
832 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
833 * we cannot notice that anon_vma is freed while we migrates a page.
1ce82b69 834 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 835 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
836 * File Caches may use write_page() or lock_page() in migration, then,
837 * just care Anon page here.
03f15c86
HD
838 *
839 * Only page_get_anon_vma() understands the subtleties of
840 * getting a hold on an anon_vma from outside one of its mms.
841 * But if we cannot get anon_vma, then we won't need it anyway,
842 * because that implies that the anon page is no longer mapped
843 * (and cannot be remapped so long as we hold the page lock).
dc386d4d 844 */
03f15c86 845 if (PageAnon(page) && !PageKsm(page))
746b18d4 846 anon_vma = page_get_anon_vma(page);
62e1c553 847
7db7671f
HD
848 /*
849 * Block others from accessing the new page when we get around to
850 * establishing additional references. We are usually the only one
851 * holding a reference to newpage at this point. We used to have a BUG
852 * here if trylock_page(newpage) fails, but would like to allow for
853 * cases where there might be a race with the previous use of newpage.
854 * This is much like races on refcount of oldpage: just don't BUG().
855 */
856 if (unlikely(!trylock_page(newpage)))
857 goto out_unlock;
858
d6d86c0a 859 if (unlikely(isolated_balloon_page(page))) {
bf6bddf1
RA
860 /*
861 * A ballooned page does not need any special attention from
862 * physical to virtual reverse mapping procedures.
863 * Skip any attempt to unmap PTEs or to remap swap cache,
864 * in order to avoid burning cycles at rmap level, and perform
865 * the page migration right away (proteced by page lock).
866 */
867 rc = balloon_page_migrate(newpage, page, mode);
7db7671f 868 goto out_unlock_both;
bf6bddf1
RA
869 }
870
dc386d4d 871 /*
62e1c553
SL
872 * Corner case handling:
873 * 1. When a new swap-cache page is read into, it is added to the LRU
874 * and treated as swapcache but it has no rmap yet.
875 * Calling try_to_unmap() against a page->mapping==NULL page will
876 * trigger a BUG. So handle it here.
877 * 2. An orphaned page (see truncate_complete_page) might have
878 * fs-private metadata. The page can be picked up due to memory
879 * offlining. Everywhere else except page reclaim, the page is
880 * invisible to the vm, so the page can not be migrated. So try to
881 * free the metadata, so the page can be freed.
e24f0b8f 882 */
62e1c553 883 if (!page->mapping) {
309381fe 884 VM_BUG_ON_PAGE(PageAnon(page), page);
1ce82b69 885 if (page_has_private(page)) {
62e1c553 886 try_to_free_buffers(page);
7db7671f 887 goto out_unlock_both;
62e1c553 888 }
7db7671f
HD
889 } else if (page_mapped(page)) {
890 /* Establish migration ptes */
03f15c86
HD
891 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
892 page);
2ebba6b7 893 try_to_unmap(page,
da1b13cc 894 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
2ebba6b7
HD
895 page_was_mapped = 1;
896 }
dc386d4d 897
e6a1530d 898 if (!page_mapped(page))
5c3f9a67 899 rc = move_to_new_page(newpage, page, mode);
e24f0b8f 900
5c3f9a67
HD
901 if (page_was_mapped)
902 remove_migration_ptes(page,
903 rc == MIGRATEPAGE_SUCCESS ? newpage : page);
3f6c8272 904
7db7671f
HD
905out_unlock_both:
906 unlock_page(newpage);
907out_unlock:
3f6c8272 908 /* Drop an anon_vma reference if we took one */
76545066 909 if (anon_vma)
9e60109f 910 put_anon_vma(anon_vma);
e24f0b8f 911 unlock_page(page);
0dabec93
MK
912out:
913 return rc;
914}
95a402c3 915
ef2a5153
GU
916/*
917 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
918 * around it.
919 */
920#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
921#define ICE_noinline noinline
922#else
923#define ICE_noinline
924#endif
925
0dabec93
MK
926/*
927 * Obtain the lock on page, remove all ptes and migrate the page
928 * to the newly allocated page in newpage.
929 */
ef2a5153
GU
930static ICE_noinline int unmap_and_move(new_page_t get_new_page,
931 free_page_t put_new_page,
932 unsigned long private, struct page *page,
add05cec
NH
933 int force, enum migrate_mode mode,
934 enum migrate_reason reason)
0dabec93 935{
2def7424 936 int rc = MIGRATEPAGE_SUCCESS;
0dabec93 937 int *result = NULL;
2def7424 938 struct page *newpage;
0dabec93 939
2def7424 940 newpage = get_new_page(page, private, &result);
0dabec93
MK
941 if (!newpage)
942 return -ENOMEM;
943
944 if (page_count(page) == 1) {
945 /* page was freed from under us. So we are done. */
946 goto out;
947 }
948
4d2fa965
KS
949 if (unlikely(PageTransHuge(page))) {
950 lock_page(page);
951 rc = split_huge_page(page);
952 unlock_page(page);
953 if (rc)
0dabec93 954 goto out;
4d2fa965 955 }
0dabec93 956
9c620e2b 957 rc = __unmap_and_move(page, newpage, force, mode);
7cd12b4a 958 if (rc == MIGRATEPAGE_SUCCESS) {
2def7424 959 put_new_page = NULL;
7cd12b4a
VB
960 set_page_owner_migrate_reason(newpage, reason);
961 }
bf6bddf1 962
0dabec93 963out:
e24f0b8f 964 if (rc != -EAGAIN) {
0dabec93
MK
965 /*
966 * A page that has been migrated has all references
967 * removed and will be freed. A page that has not been
968 * migrated will have kepts its references and be
969 * restored.
970 */
971 list_del(&page->lru);
a731286d 972 dec_zone_page_state(page, NR_ISOLATED_ANON +
6c0b1351 973 page_is_file_cache(page));
f4c18e6f 974 /* Soft-offlined page shouldn't go through lru cache list */
da1b13cc 975 if (reason == MR_MEMORY_FAILURE) {
f4c18e6f 976 put_page(page);
da1b13cc
WL
977 if (!test_set_page_hwpoison(page))
978 num_poisoned_pages_inc();
979 } else
add05cec 980 putback_lru_page(page);
e24f0b8f 981 }
68711a74 982
95a402c3 983 /*
68711a74
DR
984 * If migration was not successful and there's a freeing callback, use
985 * it. Otherwise, putback_lru_page() will drop the reference grabbed
986 * during isolation.
95a402c3 987 */
cf4b769a 988 if (put_new_page)
68711a74 989 put_new_page(newpage, private);
cf4b769a 990 else if (unlikely(__is_movable_balloon_page(newpage))) {
d6d86c0a
KK
991 /* drop our reference, page already in the balloon */
992 put_page(newpage);
8bdd6380 993 } else
68711a74
DR
994 putback_lru_page(newpage);
995
742755a1
CL
996 if (result) {
997 if (rc)
998 *result = rc;
999 else
1000 *result = page_to_nid(newpage);
1001 }
e24f0b8f
CL
1002 return rc;
1003}
1004
290408d4
NH
1005/*
1006 * Counterpart of unmap_and_move_page() for hugepage migration.
1007 *
1008 * This function doesn't wait the completion of hugepage I/O
1009 * because there is no race between I/O and migration for hugepage.
1010 * Note that currently hugepage I/O occurs only in direct I/O
1011 * where no lock is held and PG_writeback is irrelevant,
1012 * and writeback status of all subpages are counted in the reference
1013 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1014 * under direct I/O, the reference of the head page is 512 and a bit more.)
1015 * This means that when we try to migrate hugepage whose subpages are
1016 * doing direct I/O, some references remain after try_to_unmap() and
1017 * hugepage migration fails without data corruption.
1018 *
1019 * There is also no race when direct I/O is issued on the page under migration,
1020 * because then pte is replaced with migration swap entry and direct I/O code
1021 * will wait in the page fault for migration to complete.
1022 */
1023static int unmap_and_move_huge_page(new_page_t get_new_page,
68711a74
DR
1024 free_page_t put_new_page, unsigned long private,
1025 struct page *hpage, int force,
7cd12b4a 1026 enum migrate_mode mode, int reason)
290408d4 1027{
2def7424 1028 int rc = -EAGAIN;
290408d4 1029 int *result = NULL;
2ebba6b7 1030 int page_was_mapped = 0;
32665f2b 1031 struct page *new_hpage;
290408d4
NH
1032 struct anon_vma *anon_vma = NULL;
1033
83467efb
NH
1034 /*
1035 * Movability of hugepages depends on architectures and hugepage size.
1036 * This check is necessary because some callers of hugepage migration
1037 * like soft offline and memory hotremove don't walk through page
1038 * tables or check whether the hugepage is pmd-based or not before
1039 * kicking migration.
1040 */
100873d7 1041 if (!hugepage_migration_supported(page_hstate(hpage))) {
32665f2b 1042 putback_active_hugepage(hpage);
83467efb 1043 return -ENOSYS;
32665f2b 1044 }
83467efb 1045
32665f2b 1046 new_hpage = get_new_page(hpage, private, &result);
290408d4
NH
1047 if (!new_hpage)
1048 return -ENOMEM;
1049
290408d4 1050 if (!trylock_page(hpage)) {
a6bc32b8 1051 if (!force || mode != MIGRATE_SYNC)
290408d4
NH
1052 goto out;
1053 lock_page(hpage);
1054 }
1055
746b18d4
PZ
1056 if (PageAnon(hpage))
1057 anon_vma = page_get_anon_vma(hpage);
290408d4 1058
7db7671f
HD
1059 if (unlikely(!trylock_page(new_hpage)))
1060 goto put_anon;
1061
2ebba6b7
HD
1062 if (page_mapped(hpage)) {
1063 try_to_unmap(hpage,
1064 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1065 page_was_mapped = 1;
1066 }
290408d4
NH
1067
1068 if (!page_mapped(hpage))
5c3f9a67 1069 rc = move_to_new_page(new_hpage, hpage, mode);
290408d4 1070
5c3f9a67
HD
1071 if (page_was_mapped)
1072 remove_migration_ptes(hpage,
1073 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage);
290408d4 1074
7db7671f
HD
1075 unlock_page(new_hpage);
1076
1077put_anon:
fd4a4663 1078 if (anon_vma)
9e60109f 1079 put_anon_vma(anon_vma);
8e6ac7fa 1080
2def7424 1081 if (rc == MIGRATEPAGE_SUCCESS) {
8e6ac7fa 1082 hugetlb_cgroup_migrate(hpage, new_hpage);
2def7424 1083 put_new_page = NULL;
7cd12b4a 1084 set_page_owner_migrate_reason(new_hpage, reason);
2def7424 1085 }
8e6ac7fa 1086
290408d4 1087 unlock_page(hpage);
09761333 1088out:
b8ec1cee
NH
1089 if (rc != -EAGAIN)
1090 putback_active_hugepage(hpage);
68711a74
DR
1091
1092 /*
1093 * If migration was not successful and there's a freeing callback, use
1094 * it. Otherwise, put_page() will drop the reference grabbed during
1095 * isolation.
1096 */
2def7424 1097 if (put_new_page)
68711a74
DR
1098 put_new_page(new_hpage, private);
1099 else
3aaa76e1 1100 putback_active_hugepage(new_hpage);
68711a74 1101
290408d4
NH
1102 if (result) {
1103 if (rc)
1104 *result = rc;
1105 else
1106 *result = page_to_nid(new_hpage);
1107 }
1108 return rc;
1109}
1110
b20a3503 1111/*
c73e5c9c
SB
1112 * migrate_pages - migrate the pages specified in a list, to the free pages
1113 * supplied as the target for the page migration
b20a3503 1114 *
c73e5c9c
SB
1115 * @from: The list of pages to be migrated.
1116 * @get_new_page: The function used to allocate free pages to be used
1117 * as the target of the page migration.
68711a74
DR
1118 * @put_new_page: The function used to free target pages if migration
1119 * fails, or NULL if no special handling is necessary.
c73e5c9c
SB
1120 * @private: Private data to be passed on to get_new_page()
1121 * @mode: The migration mode that specifies the constraints for
1122 * page migration, if any.
1123 * @reason: The reason for page migration.
b20a3503 1124 *
c73e5c9c
SB
1125 * The function returns after 10 attempts or if no pages are movable any more
1126 * because the list has become empty or no retryable pages exist any more.
14e0f9bc 1127 * The caller should call putback_movable_pages() to return pages to the LRU
28bd6578 1128 * or free list only if ret != 0.
b20a3503 1129 *
c73e5c9c 1130 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1131 */
9c620e2b 1132int migrate_pages(struct list_head *from, new_page_t get_new_page,
68711a74
DR
1133 free_page_t put_new_page, unsigned long private,
1134 enum migrate_mode mode, int reason)
b20a3503 1135{
e24f0b8f 1136 int retry = 1;
b20a3503 1137 int nr_failed = 0;
5647bc29 1138 int nr_succeeded = 0;
b20a3503
CL
1139 int pass = 0;
1140 struct page *page;
1141 struct page *page2;
1142 int swapwrite = current->flags & PF_SWAPWRITE;
1143 int rc;
1144
1145 if (!swapwrite)
1146 current->flags |= PF_SWAPWRITE;
1147
e24f0b8f
CL
1148 for(pass = 0; pass < 10 && retry; pass++) {
1149 retry = 0;
b20a3503 1150
e24f0b8f 1151 list_for_each_entry_safe(page, page2, from, lru) {
e24f0b8f 1152 cond_resched();
2d1db3b1 1153
31caf665
NH
1154 if (PageHuge(page))
1155 rc = unmap_and_move_huge_page(get_new_page,
68711a74 1156 put_new_page, private, page,
7cd12b4a 1157 pass > 2, mode, reason);
31caf665 1158 else
68711a74 1159 rc = unmap_and_move(get_new_page, put_new_page,
add05cec
NH
1160 private, page, pass > 2, mode,
1161 reason);
2d1db3b1 1162
e24f0b8f 1163 switch(rc) {
95a402c3
CL
1164 case -ENOMEM:
1165 goto out;
e24f0b8f 1166 case -EAGAIN:
2d1db3b1 1167 retry++;
e24f0b8f 1168 break;
78bd5209 1169 case MIGRATEPAGE_SUCCESS:
5647bc29 1170 nr_succeeded++;
e24f0b8f
CL
1171 break;
1172 default:
354a3363
NH
1173 /*
1174 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1175 * unlike -EAGAIN case, the failed page is
1176 * removed from migration page list and not
1177 * retried in the next outer loop.
1178 */
2d1db3b1 1179 nr_failed++;
e24f0b8f 1180 break;
2d1db3b1 1181 }
b20a3503
CL
1182 }
1183 }
f2f81fb2
VB
1184 nr_failed += retry;
1185 rc = nr_failed;
95a402c3 1186out:
5647bc29
MG
1187 if (nr_succeeded)
1188 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1189 if (nr_failed)
1190 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1191 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1192
b20a3503
CL
1193 if (!swapwrite)
1194 current->flags &= ~PF_SWAPWRITE;
1195
78bd5209 1196 return rc;
b20a3503 1197}
95a402c3 1198
742755a1
CL
1199#ifdef CONFIG_NUMA
1200/*
1201 * Move a list of individual pages
1202 */
1203struct page_to_node {
1204 unsigned long addr;
1205 struct page *page;
1206 int node;
1207 int status;
1208};
1209
1210static struct page *new_page_node(struct page *p, unsigned long private,
1211 int **result)
1212{
1213 struct page_to_node *pm = (struct page_to_node *)private;
1214
1215 while (pm->node != MAX_NUMNODES && pm->page != p)
1216 pm++;
1217
1218 if (pm->node == MAX_NUMNODES)
1219 return NULL;
1220
1221 *result = &pm->status;
1222
e632a938
NH
1223 if (PageHuge(p))
1224 return alloc_huge_page_node(page_hstate(compound_head(p)),
1225 pm->node);
1226 else
96db800f 1227 return __alloc_pages_node(pm->node,
e97ca8e5 1228 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
742755a1
CL
1229}
1230
1231/*
1232 * Move a set of pages as indicated in the pm array. The addr
1233 * field must be set to the virtual address of the page to be moved
1234 * and the node number must contain a valid target node.
5e9a0f02 1235 * The pm array ends with node = MAX_NUMNODES.
742755a1 1236 */
5e9a0f02
BG
1237static int do_move_page_to_node_array(struct mm_struct *mm,
1238 struct page_to_node *pm,
1239 int migrate_all)
742755a1
CL
1240{
1241 int err;
1242 struct page_to_node *pp;
1243 LIST_HEAD(pagelist);
1244
1245 down_read(&mm->mmap_sem);
1246
1247 /*
1248 * Build a list of pages to migrate
1249 */
742755a1
CL
1250 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1251 struct vm_area_struct *vma;
1252 struct page *page;
1253
742755a1
CL
1254 err = -EFAULT;
1255 vma = find_vma(mm, pp->addr);
70384dc6 1256 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
742755a1
CL
1257 goto set_status;
1258
d899844e
KS
1259 /* FOLL_DUMP to ignore special (like zero) pages */
1260 page = follow_page(vma, pp->addr,
1261 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
89f5b7da
LT
1262
1263 err = PTR_ERR(page);
1264 if (IS_ERR(page))
1265 goto set_status;
1266
742755a1
CL
1267 err = -ENOENT;
1268 if (!page)
1269 goto set_status;
1270
742755a1
CL
1271 pp->page = page;
1272 err = page_to_nid(page);
1273
1274 if (err == pp->node)
1275 /*
1276 * Node already in the right place
1277 */
1278 goto put_and_set;
1279
1280 err = -EACCES;
1281 if (page_mapcount(page) > 1 &&
1282 !migrate_all)
1283 goto put_and_set;
1284
e632a938 1285 if (PageHuge(page)) {
e66f17ff
NH
1286 if (PageHead(page))
1287 isolate_huge_page(page, &pagelist);
e632a938
NH
1288 goto put_and_set;
1289 }
1290
62695a84 1291 err = isolate_lru_page(page);
6d9c285a 1292 if (!err) {
62695a84 1293 list_add_tail(&page->lru, &pagelist);
6d9c285a
KM
1294 inc_zone_page_state(page, NR_ISOLATED_ANON +
1295 page_is_file_cache(page));
1296 }
742755a1
CL
1297put_and_set:
1298 /*
1299 * Either remove the duplicate refcount from
1300 * isolate_lru_page() or drop the page ref if it was
1301 * not isolated.
1302 */
1303 put_page(page);
1304set_status:
1305 pp->status = err;
1306 }
1307
e78bbfa8 1308 err = 0;
cf608ac1 1309 if (!list_empty(&pagelist)) {
68711a74 1310 err = migrate_pages(&pagelist, new_page_node, NULL,
9c620e2b 1311 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
cf608ac1 1312 if (err)
e632a938 1313 putback_movable_pages(&pagelist);
cf608ac1 1314 }
742755a1
CL
1315
1316 up_read(&mm->mmap_sem);
1317 return err;
1318}
1319
5e9a0f02
BG
1320/*
1321 * Migrate an array of page address onto an array of nodes and fill
1322 * the corresponding array of status.
1323 */
3268c63e 1324static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1325 unsigned long nr_pages,
1326 const void __user * __user *pages,
1327 const int __user *nodes,
1328 int __user *status, int flags)
1329{
3140a227 1330 struct page_to_node *pm;
3140a227
BG
1331 unsigned long chunk_nr_pages;
1332 unsigned long chunk_start;
1333 int err;
5e9a0f02 1334
3140a227
BG
1335 err = -ENOMEM;
1336 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1337 if (!pm)
5e9a0f02 1338 goto out;
35282a2d
BG
1339
1340 migrate_prep();
1341
5e9a0f02 1342 /*
3140a227
BG
1343 * Store a chunk of page_to_node array in a page,
1344 * but keep the last one as a marker
5e9a0f02 1345 */
3140a227 1346 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
5e9a0f02 1347
3140a227
BG
1348 for (chunk_start = 0;
1349 chunk_start < nr_pages;
1350 chunk_start += chunk_nr_pages) {
1351 int j;
5e9a0f02 1352
3140a227
BG
1353 if (chunk_start + chunk_nr_pages > nr_pages)
1354 chunk_nr_pages = nr_pages - chunk_start;
1355
1356 /* fill the chunk pm with addrs and nodes from user-space */
1357 for (j = 0; j < chunk_nr_pages; j++) {
1358 const void __user *p;
5e9a0f02
BG
1359 int node;
1360
3140a227
BG
1361 err = -EFAULT;
1362 if (get_user(p, pages + j + chunk_start))
1363 goto out_pm;
1364 pm[j].addr = (unsigned long) p;
1365
1366 if (get_user(node, nodes + j + chunk_start))
5e9a0f02
BG
1367 goto out_pm;
1368
1369 err = -ENODEV;
6f5a55f1
LT
1370 if (node < 0 || node >= MAX_NUMNODES)
1371 goto out_pm;
1372
389162c2 1373 if (!node_state(node, N_MEMORY))
5e9a0f02
BG
1374 goto out_pm;
1375
1376 err = -EACCES;
1377 if (!node_isset(node, task_nodes))
1378 goto out_pm;
1379
3140a227
BG
1380 pm[j].node = node;
1381 }
1382
1383 /* End marker for this chunk */
1384 pm[chunk_nr_pages].node = MAX_NUMNODES;
1385
1386 /* Migrate this chunk */
1387 err = do_move_page_to_node_array(mm, pm,
1388 flags & MPOL_MF_MOVE_ALL);
1389 if (err < 0)
1390 goto out_pm;
5e9a0f02 1391
5e9a0f02 1392 /* Return status information */
3140a227
BG
1393 for (j = 0; j < chunk_nr_pages; j++)
1394 if (put_user(pm[j].status, status + j + chunk_start)) {
5e9a0f02 1395 err = -EFAULT;
3140a227
BG
1396 goto out_pm;
1397 }
1398 }
1399 err = 0;
5e9a0f02
BG
1400
1401out_pm:
3140a227 1402 free_page((unsigned long)pm);
5e9a0f02
BG
1403out:
1404 return err;
1405}
1406
742755a1 1407/*
2f007e74 1408 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1409 */
80bba129
BG
1410static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1411 const void __user **pages, int *status)
742755a1 1412{
2f007e74 1413 unsigned long i;
2f007e74 1414
742755a1
CL
1415 down_read(&mm->mmap_sem);
1416
2f007e74 1417 for (i = 0; i < nr_pages; i++) {
80bba129 1418 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1419 struct vm_area_struct *vma;
1420 struct page *page;
c095adbc 1421 int err = -EFAULT;
2f007e74
BG
1422
1423 vma = find_vma(mm, addr);
70384dc6 1424 if (!vma || addr < vma->vm_start)
742755a1
CL
1425 goto set_status;
1426
d899844e
KS
1427 /* FOLL_DUMP to ignore special (like zero) pages */
1428 page = follow_page(vma, addr, FOLL_DUMP);
89f5b7da
LT
1429
1430 err = PTR_ERR(page);
1431 if (IS_ERR(page))
1432 goto set_status;
1433
d899844e 1434 err = page ? page_to_nid(page) : -ENOENT;
742755a1 1435set_status:
80bba129
BG
1436 *status = err;
1437
1438 pages++;
1439 status++;
1440 }
1441
1442 up_read(&mm->mmap_sem);
1443}
1444
1445/*
1446 * Determine the nodes of a user array of pages and store it in
1447 * a user array of status.
1448 */
1449static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1450 const void __user * __user *pages,
1451 int __user *status)
1452{
1453#define DO_PAGES_STAT_CHUNK_NR 16
1454 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1455 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1456
87b8d1ad
PA
1457 while (nr_pages) {
1458 unsigned long chunk_nr;
80bba129 1459
87b8d1ad
PA
1460 chunk_nr = nr_pages;
1461 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1462 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1463
1464 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1465 break;
80bba129
BG
1466
1467 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1468
87b8d1ad
PA
1469 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1470 break;
742755a1 1471
87b8d1ad
PA
1472 pages += chunk_nr;
1473 status += chunk_nr;
1474 nr_pages -= chunk_nr;
1475 }
1476 return nr_pages ? -EFAULT : 0;
742755a1
CL
1477}
1478
1479/*
1480 * Move a list of pages in the address space of the currently executing
1481 * process.
1482 */
938bb9f5
HC
1483SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1484 const void __user * __user *, pages,
1485 const int __user *, nodes,
1486 int __user *, status, int, flags)
742755a1 1487{
c69e8d9c 1488 const struct cred *cred = current_cred(), *tcred;
742755a1 1489 struct task_struct *task;
742755a1 1490 struct mm_struct *mm;
5e9a0f02 1491 int err;
3268c63e 1492 nodemask_t task_nodes;
742755a1
CL
1493
1494 /* Check flags */
1495 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1496 return -EINVAL;
1497
1498 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1499 return -EPERM;
1500
1501 /* Find the mm_struct */
a879bf58 1502 rcu_read_lock();
228ebcbe 1503 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1504 if (!task) {
a879bf58 1505 rcu_read_unlock();
742755a1
CL
1506 return -ESRCH;
1507 }
3268c63e 1508 get_task_struct(task);
742755a1
CL
1509
1510 /*
1511 * Check if this process has the right to modify the specified
1512 * process. The right exists if the process has administrative
1513 * capabilities, superuser privileges or the same
1514 * userid as the target process.
1515 */
c69e8d9c 1516 tcred = __task_cred(task);
b38a86eb
EB
1517 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1518 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
742755a1 1519 !capable(CAP_SYS_NICE)) {
c69e8d9c 1520 rcu_read_unlock();
742755a1 1521 err = -EPERM;
5e9a0f02 1522 goto out;
742755a1 1523 }
c69e8d9c 1524 rcu_read_unlock();
742755a1 1525
86c3a764
DQ
1526 err = security_task_movememory(task);
1527 if (err)
5e9a0f02 1528 goto out;
86c3a764 1529
3268c63e
CL
1530 task_nodes = cpuset_mems_allowed(task);
1531 mm = get_task_mm(task);
1532 put_task_struct(task);
1533
6e8b09ea
SL
1534 if (!mm)
1535 return -EINVAL;
1536
1537 if (nodes)
1538 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1539 nodes, status, flags);
1540 else
1541 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1542
742755a1
CL
1543 mmput(mm);
1544 return err;
3268c63e
CL
1545
1546out:
1547 put_task_struct(task);
1548 return err;
742755a1 1549}
742755a1 1550
7039e1db
PZ
1551#ifdef CONFIG_NUMA_BALANCING
1552/*
1553 * Returns true if this is a safe migration target node for misplaced NUMA
1554 * pages. Currently it only checks the watermarks which crude
1555 */
1556static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1557 unsigned long nr_migrate_pages)
7039e1db
PZ
1558{
1559 int z;
1560 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1561 struct zone *zone = pgdat->node_zones + z;
1562
1563 if (!populated_zone(zone))
1564 continue;
1565
6e543d57 1566 if (!zone_reclaimable(zone))
7039e1db
PZ
1567 continue;
1568
1569 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1570 if (!zone_watermark_ok(zone, 0,
1571 high_wmark_pages(zone) +
1572 nr_migrate_pages,
1573 0, 0))
1574 continue;
1575 return true;
1576 }
1577 return false;
1578}
1579
1580static struct page *alloc_misplaced_dst_page(struct page *page,
1581 unsigned long data,
1582 int **result)
1583{
1584 int nid = (int) data;
1585 struct page *newpage;
1586
96db800f 1587 newpage = __alloc_pages_node(nid,
e97ca8e5
JW
1588 (GFP_HIGHUSER_MOVABLE |
1589 __GFP_THISNODE | __GFP_NOMEMALLOC |
1590 __GFP_NORETRY | __GFP_NOWARN) &
8479eba7 1591 ~__GFP_RECLAIM, 0);
bac0382c 1592
7039e1db
PZ
1593 return newpage;
1594}
1595
a8f60772
MG
1596/*
1597 * page migration rate limiting control.
1598 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1599 * window of time. Default here says do not migrate more than 1280M per second.
1600 */
1601static unsigned int migrate_interval_millisecs __read_mostly = 100;
1602static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1603
b32967ff 1604/* Returns true if the node is migrate rate-limited after the update */
1c30e017
MG
1605static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1606 unsigned long nr_pages)
7039e1db 1607{
a8f60772
MG
1608 /*
1609 * Rate-limit the amount of data that is being migrated to a node.
1610 * Optimal placement is no good if the memory bus is saturated and
1611 * all the time is being spent migrating!
1612 */
a8f60772 1613 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1c5e9c27 1614 spin_lock(&pgdat->numabalancing_migrate_lock);
a8f60772
MG
1615 pgdat->numabalancing_migrate_nr_pages = 0;
1616 pgdat->numabalancing_migrate_next_window = jiffies +
1617 msecs_to_jiffies(migrate_interval_millisecs);
1c5e9c27 1618 spin_unlock(&pgdat->numabalancing_migrate_lock);
a8f60772 1619 }
af1839d7
MG
1620 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1621 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1622 nr_pages);
1c5e9c27 1623 return true;
af1839d7 1624 }
1c5e9c27
MG
1625
1626 /*
1627 * This is an unlocked non-atomic update so errors are possible.
1628 * The consequences are failing to migrate when we potentiall should
1629 * have which is not severe enough to warrant locking. If it is ever
1630 * a problem, it can be converted to a per-cpu counter.
1631 */
1632 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1633 return false;
b32967ff
MG
1634}
1635
1c30e017 1636static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
b32967ff 1637{
340ef390 1638 int page_lru;
a8f60772 1639
309381fe 1640 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
3abef4e6 1641
7039e1db 1642 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1643 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1644 return 0;
7039e1db 1645
340ef390
HD
1646 if (isolate_lru_page(page))
1647 return 0;
7039e1db 1648
340ef390
HD
1649 /*
1650 * migrate_misplaced_transhuge_page() skips page migration's usual
1651 * check on page_count(), so we must do it here, now that the page
1652 * has been isolated: a GUP pin, or any other pin, prevents migration.
1653 * The expected page count is 3: 1 for page's mapcount and 1 for the
1654 * caller's pin and 1 for the reference taken by isolate_lru_page().
1655 */
1656 if (PageTransHuge(page) && page_count(page) != 3) {
1657 putback_lru_page(page);
1658 return 0;
7039e1db
PZ
1659 }
1660
340ef390
HD
1661 page_lru = page_is_file_cache(page);
1662 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1663 hpage_nr_pages(page));
1664
149c33e1 1665 /*
340ef390
HD
1666 * Isolating the page has taken another reference, so the
1667 * caller's reference can be safely dropped without the page
1668 * disappearing underneath us during migration.
149c33e1
MG
1669 */
1670 put_page(page);
340ef390 1671 return 1;
b32967ff
MG
1672}
1673
de466bd6
MG
1674bool pmd_trans_migrating(pmd_t pmd)
1675{
1676 struct page *page = pmd_page(pmd);
1677 return PageLocked(page);
1678}
1679
b32967ff
MG
1680/*
1681 * Attempt to migrate a misplaced page to the specified destination
1682 * node. Caller is expected to have an elevated reference count on
1683 * the page that will be dropped by this function before returning.
1684 */
1bc115d8
MG
1685int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1686 int node)
b32967ff
MG
1687{
1688 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1689 int isolated;
b32967ff
MG
1690 int nr_remaining;
1691 LIST_HEAD(migratepages);
1692
1693 /*
1bc115d8
MG
1694 * Don't migrate file pages that are mapped in multiple processes
1695 * with execute permissions as they are probably shared libraries.
b32967ff 1696 */
1bc115d8
MG
1697 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1698 (vma->vm_flags & VM_EXEC))
b32967ff 1699 goto out;
b32967ff
MG
1700
1701 /*
1702 * Rate-limit the amount of data that is being migrated to a node.
1703 * Optimal placement is no good if the memory bus is saturated and
1704 * all the time is being spent migrating!
1705 */
340ef390 1706 if (numamigrate_update_ratelimit(pgdat, 1))
b32967ff 1707 goto out;
b32967ff
MG
1708
1709 isolated = numamigrate_isolate_page(pgdat, page);
1710 if (!isolated)
1711 goto out;
1712
1713 list_add(&page->lru, &migratepages);
9c620e2b 1714 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
68711a74
DR
1715 NULL, node, MIGRATE_ASYNC,
1716 MR_NUMA_MISPLACED);
b32967ff 1717 if (nr_remaining) {
59c82b70
JK
1718 if (!list_empty(&migratepages)) {
1719 list_del(&page->lru);
1720 dec_zone_page_state(page, NR_ISOLATED_ANON +
1721 page_is_file_cache(page));
1722 putback_lru_page(page);
1723 }
b32967ff
MG
1724 isolated = 0;
1725 } else
1726 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1727 BUG_ON(!list_empty(&migratepages));
7039e1db 1728 return isolated;
340ef390
HD
1729
1730out:
1731 put_page(page);
1732 return 0;
7039e1db 1733}
220018d3 1734#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1735
220018d3 1736#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1737/*
1738 * Migrates a THP to a given target node. page must be locked and is unlocked
1739 * before returning.
1740 */
b32967ff
MG
1741int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1742 struct vm_area_struct *vma,
1743 pmd_t *pmd, pmd_t entry,
1744 unsigned long address,
1745 struct page *page, int node)
1746{
c4088ebd 1747 spinlock_t *ptl;
b32967ff
MG
1748 pg_data_t *pgdat = NODE_DATA(node);
1749 int isolated = 0;
1750 struct page *new_page = NULL;
b32967ff 1751 int page_lru = page_is_file_cache(page);
f714f4f2
MG
1752 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1753 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
2b4847e7 1754 pmd_t orig_entry;
b32967ff 1755
b32967ff
MG
1756 /*
1757 * Rate-limit the amount of data that is being migrated to a node.
1758 * Optimal placement is no good if the memory bus is saturated and
1759 * all the time is being spent migrating!
1760 */
d28d4335 1761 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
b32967ff
MG
1762 goto out_dropref;
1763
1764 new_page = alloc_pages_node(node,
71baba4b 1765 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
e97ca8e5 1766 HPAGE_PMD_ORDER);
340ef390
HD
1767 if (!new_page)
1768 goto out_fail;
9a982250 1769 prep_transhuge_page(new_page);
340ef390 1770
b32967ff 1771 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 1772 if (!isolated) {
b32967ff 1773 put_page(new_page);
340ef390 1774 goto out_fail;
b32967ff 1775 }
458aa76d
AK
1776 /*
1777 * We are not sure a pending tlb flush here is for a huge page
1778 * mapping or not. Hence use the tlb range variant
1779 */
b0943d61
MG
1780 if (mm_tlb_flush_pending(mm))
1781 flush_tlb_range(vma, mmun_start, mmun_end);
1782
b32967ff 1783 /* Prepare a page as a migration target */
48c935ad 1784 __SetPageLocked(new_page);
b32967ff
MG
1785 SetPageSwapBacked(new_page);
1786
1787 /* anon mapping, we can simply copy page->mapping to the new page: */
1788 new_page->mapping = page->mapping;
1789 new_page->index = page->index;
1790 migrate_page_copy(new_page, page);
1791 WARN_ON(PageLRU(new_page));
1792
1793 /* Recheck the target PMD */
f714f4f2 1794 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 1795 ptl = pmd_lock(mm, pmd);
2b4847e7
MG
1796 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1797fail_putback:
c4088ebd 1798 spin_unlock(ptl);
f714f4f2 1799 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b32967ff
MG
1800
1801 /* Reverse changes made by migrate_page_copy() */
1802 if (TestClearPageActive(new_page))
1803 SetPageActive(page);
1804 if (TestClearPageUnevictable(new_page))
1805 SetPageUnevictable(page);
b32967ff
MG
1806
1807 unlock_page(new_page);
1808 put_page(new_page); /* Free it */
1809
a54a407f
MG
1810 /* Retake the callers reference and putback on LRU */
1811 get_page(page);
b32967ff 1812 putback_lru_page(page);
a54a407f
MG
1813 mod_zone_page_state(page_zone(page),
1814 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
eb4489f6
MG
1815
1816 goto out_unlock;
b32967ff
MG
1817 }
1818
2b4847e7 1819 orig_entry = *pmd;
b32967ff 1820 entry = mk_pmd(new_page, vma->vm_page_prot);
b32967ff 1821 entry = pmd_mkhuge(entry);
2b4847e7 1822 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
b32967ff 1823
2b4847e7
MG
1824 /*
1825 * Clear the old entry under pagetable lock and establish the new PTE.
1826 * Any parallel GUP will either observe the old page blocking on the
1827 * page lock, block on the page table lock or observe the new page.
1828 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1829 * guarantee the copy is visible before the pagetable update.
1830 */
f714f4f2 1831 flush_cache_range(vma, mmun_start, mmun_end);
d281ee61 1832 page_add_anon_rmap(new_page, vma, mmun_start, true);
8809aa2d 1833 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
f714f4f2 1834 set_pmd_at(mm, mmun_start, pmd, entry);
ce4a9cc5 1835 update_mmu_cache_pmd(vma, address, &entry);
2b4847e7
MG
1836
1837 if (page_count(page) != 2) {
f714f4f2 1838 set_pmd_at(mm, mmun_start, pmd, orig_entry);
458aa76d 1839 flush_pmd_tlb_range(vma, mmun_start, mmun_end);
34ee645e 1840 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2b4847e7 1841 update_mmu_cache_pmd(vma, address, &entry);
d281ee61 1842 page_remove_rmap(new_page, true);
2b4847e7
MG
1843 goto fail_putback;
1844 }
1845
51afb12b 1846 mlock_migrate_page(new_page, page);
d281ee61 1847 page_remove_rmap(page, true);
7cd12b4a 1848 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2b4847e7 1849
c4088ebd 1850 spin_unlock(ptl);
f714f4f2 1851 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b32967ff 1852
11de9927
MG
1853 /* Take an "isolate" reference and put new page on the LRU. */
1854 get_page(new_page);
1855 putback_lru_page(new_page);
1856
b32967ff
MG
1857 unlock_page(new_page);
1858 unlock_page(page);
1859 put_page(page); /* Drop the rmap reference */
1860 put_page(page); /* Drop the LRU isolation reference */
1861
1862 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1863 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1864
b32967ff
MG
1865 mod_zone_page_state(page_zone(page),
1866 NR_ISOLATED_ANON + page_lru,
1867 -HPAGE_PMD_NR);
1868 return isolated;
1869
340ef390
HD
1870out_fail:
1871 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
b32967ff 1872out_dropref:
2b4847e7
MG
1873 ptl = pmd_lock(mm, pmd);
1874 if (pmd_same(*pmd, entry)) {
4d942466 1875 entry = pmd_modify(entry, vma->vm_page_prot);
f714f4f2 1876 set_pmd_at(mm, mmun_start, pmd, entry);
2b4847e7
MG
1877 update_mmu_cache_pmd(vma, address, &entry);
1878 }
1879 spin_unlock(ptl);
a54a407f 1880
eb4489f6 1881out_unlock:
340ef390 1882 unlock_page(page);
b32967ff 1883 put_page(page);
b32967ff
MG
1884 return 0;
1885}
7039e1db
PZ
1886#endif /* CONFIG_NUMA_BALANCING */
1887
1888#endif /* CONFIG_NUMA */