]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
Revert "mm, page_alloc: only use per-cpu allocator for irq-safe requests"
[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>
bda807d4 34#include <linux/compaction.h>
4f5ca265 35#include <linux/syscalls.h>
290408d4 36#include <linux/hugetlb.h>
8e6ac7fa 37#include <linux/hugetlb_cgroup.h>
5a0e3ad6 38#include <linux/gfp.h>
bf6bddf1 39#include <linux/balloon_compaction.h>
f714f4f2 40#include <linux/mmu_notifier.h>
33c3fc71 41#include <linux/page_idle.h>
d435edca 42#include <linux/page_owner.h>
6e84f315 43#include <linux/sched/mm.h>
b20a3503 44
0d1836c3
MN
45#include <asm/tlbflush.h>
46
7b2a2d4a
MG
47#define CREATE_TRACE_POINTS
48#include <trace/events/migrate.h>
49
b20a3503
CL
50#include "internal.h"
51
b20a3503 52/*
742755a1 53 * migrate_prep() needs to be called before we start compiling a list of pages
748446bb
MG
54 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
55 * undesirable, use migrate_prep_local()
b20a3503
CL
56 */
57int migrate_prep(void)
58{
b20a3503
CL
59 /*
60 * Clear the LRU lists so pages can be isolated.
61 * Note that pages may be moved off the LRU after we have
62 * drained them. Those pages will fail to migrate like other
63 * pages that may be busy.
64 */
65 lru_add_drain_all();
66
67 return 0;
68}
69
748446bb
MG
70/* Do the necessary work of migrate_prep but not if it involves other CPUs */
71int migrate_prep_local(void)
72{
73 lru_add_drain();
74
75 return 0;
76}
77
9e5bcd61 78int isolate_movable_page(struct page *page, isolate_mode_t mode)
bda807d4
MK
79{
80 struct address_space *mapping;
81
82 /*
83 * Avoid burning cycles with pages that are yet under __free_pages(),
84 * or just got freed under us.
85 *
86 * In case we 'win' a race for a movable page being freed under us and
87 * raise its refcount preventing __free_pages() from doing its job
88 * the put_page() at the end of this block will take care of
89 * release this page, thus avoiding a nasty leakage.
90 */
91 if (unlikely(!get_page_unless_zero(page)))
92 goto out;
93
94 /*
95 * Check PageMovable before holding a PG_lock because page's owner
96 * assumes anybody doesn't touch PG_lock of newly allocated page
97 * so unconditionally grapping the lock ruins page's owner side.
98 */
99 if (unlikely(!__PageMovable(page)))
100 goto out_putpage;
101 /*
102 * As movable pages are not isolated from LRU lists, concurrent
103 * compaction threads can race against page migration functions
104 * as well as race against the releasing a page.
105 *
106 * In order to avoid having an already isolated movable page
107 * being (wrongly) re-isolated while it is under migration,
108 * or to avoid attempting to isolate pages being released,
109 * lets be sure we have the page lock
110 * before proceeding with the movable page isolation steps.
111 */
112 if (unlikely(!trylock_page(page)))
113 goto out_putpage;
114
115 if (!PageMovable(page) || PageIsolated(page))
116 goto out_no_isolated;
117
118 mapping = page_mapping(page);
119 VM_BUG_ON_PAGE(!mapping, page);
120
121 if (!mapping->a_ops->isolate_page(page, mode))
122 goto out_no_isolated;
123
124 /* Driver shouldn't use PG_isolated bit of page->flags */
125 WARN_ON_ONCE(PageIsolated(page));
126 __SetPageIsolated(page);
127 unlock_page(page);
128
9e5bcd61 129 return 0;
bda807d4
MK
130
131out_no_isolated:
132 unlock_page(page);
133out_putpage:
134 put_page(page);
135out:
9e5bcd61 136 return -EBUSY;
bda807d4
MK
137}
138
139/* It should be called on page which is PG_movable */
140void putback_movable_page(struct page *page)
141{
142 struct address_space *mapping;
143
144 VM_BUG_ON_PAGE(!PageLocked(page), page);
145 VM_BUG_ON_PAGE(!PageMovable(page), page);
146 VM_BUG_ON_PAGE(!PageIsolated(page), page);
147
148 mapping = page_mapping(page);
149 mapping->a_ops->putback_page(page);
150 __ClearPageIsolated(page);
151}
152
5733c7d1
RA
153/*
154 * Put previously isolated pages back onto the appropriate lists
155 * from where they were once taken off for compaction/migration.
156 *
59c82b70
JK
157 * This function shall be used whenever the isolated pageset has been
158 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
159 * and isolate_huge_page().
5733c7d1
RA
160 */
161void putback_movable_pages(struct list_head *l)
162{
163 struct page *page;
164 struct page *page2;
165
b20a3503 166 list_for_each_entry_safe(page, page2, l, lru) {
31caf665
NH
167 if (unlikely(PageHuge(page))) {
168 putback_active_hugepage(page);
169 continue;
170 }
e24f0b8f 171 list_del(&page->lru);
bda807d4
MK
172 /*
173 * We isolated non-lru movable page so here we can use
174 * __PageMovable because LRU page's mapping cannot have
175 * PAGE_MAPPING_MOVABLE.
176 */
b1123ea6 177 if (unlikely(__PageMovable(page))) {
bda807d4
MK
178 VM_BUG_ON_PAGE(!PageIsolated(page), page);
179 lock_page(page);
180 if (PageMovable(page))
181 putback_movable_page(page);
182 else
183 __ClearPageIsolated(page);
184 unlock_page(page);
185 put_page(page);
186 } else {
bf6bddf1 187 putback_lru_page(page);
6afcf8ef
ML
188 dec_node_page_state(page, NR_ISOLATED_ANON +
189 page_is_file_cache(page));
bda807d4 190 }
b20a3503 191 }
b20a3503
CL
192}
193
0697212a
CL
194/*
195 * Restore a potential migration pte to a working pte entry
196 */
3fe87967 197static int remove_migration_pte(struct page *page, struct vm_area_struct *vma,
e9995ef9 198 unsigned long addr, void *old)
0697212a 199{
3fe87967
KS
200 struct page_vma_mapped_walk pvmw = {
201 .page = old,
202 .vma = vma,
203 .address = addr,
204 .flags = PVMW_SYNC | PVMW_MIGRATION,
205 };
206 struct page *new;
207 pte_t pte;
0697212a 208 swp_entry_t entry;
0697212a 209
3fe87967
KS
210 VM_BUG_ON_PAGE(PageTail(page), page);
211 while (page_vma_mapped_walk(&pvmw)) {
4b0ece6f
NH
212 if (PageKsm(page))
213 new = page;
214 else
215 new = page - pvmw.page->index +
216 linear_page_index(vma, pvmw.address);
0697212a 217
3fe87967
KS
218 get_page(new);
219 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
220 if (pte_swp_soft_dirty(*pvmw.pte))
221 pte = pte_mksoft_dirty(pte);
0697212a 222
3fe87967
KS
223 /*
224 * Recheck VMA as permissions can change since migration started
225 */
226 entry = pte_to_swp_entry(*pvmw.pte);
227 if (is_write_migration_entry(entry))
228 pte = maybe_mkwrite(pte, vma);
d3cb8bf6 229
3ef8fd7f 230#ifdef CONFIG_HUGETLB_PAGE
3fe87967
KS
231 if (PageHuge(new)) {
232 pte = pte_mkhuge(pte);
233 pte = arch_make_huge_pte(pte, vma, new, 0);
234 }
3ef8fd7f 235#endif
3fe87967
KS
236 flush_dcache_page(new);
237 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
04e62a29 238
3fe87967
KS
239 if (PageHuge(new)) {
240 if (PageAnon(new))
241 hugepage_add_anon_rmap(new, vma, pvmw.address);
242 else
243 page_dup_rmap(new, true);
244 } else if (PageAnon(new))
245 page_add_anon_rmap(new, vma, pvmw.address, false);
290408d4 246 else
3fe87967 247 page_add_file_rmap(new, false);
04e62a29 248
3fe87967
KS
249 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
250 mlock_vma_page(new);
251
252 /* No need to invalidate - it was non-present before */
253 update_mmu_cache(vma, pvmw.address, pvmw.pte);
254 }
51afb12b 255
e9995ef9 256 return SWAP_AGAIN;
0697212a
CL
257}
258
04e62a29
CL
259/*
260 * Get rid of all migration entries and replace them by
261 * references to the indicated page.
262 */
e388466d 263void remove_migration_ptes(struct page *old, struct page *new, bool locked)
04e62a29 264{
051ac83a
JK
265 struct rmap_walk_control rwc = {
266 .rmap_one = remove_migration_pte,
267 .arg = old,
268 };
269
e388466d
KS
270 if (locked)
271 rmap_walk_locked(new, &rwc);
272 else
273 rmap_walk(new, &rwc);
04e62a29
CL
274}
275
0697212a
CL
276/*
277 * Something used the pte of a page under migration. We need to
278 * get to the page and wait until migration is finished.
279 * When we return from this function the fault will be retried.
0697212a 280 */
e66f17ff 281void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
30dad309 282 spinlock_t *ptl)
0697212a 283{
30dad309 284 pte_t pte;
0697212a
CL
285 swp_entry_t entry;
286 struct page *page;
287
30dad309 288 spin_lock(ptl);
0697212a
CL
289 pte = *ptep;
290 if (!is_swap_pte(pte))
291 goto out;
292
293 entry = pte_to_swp_entry(pte);
294 if (!is_migration_entry(entry))
295 goto out;
296
297 page = migration_entry_to_page(entry);
298
e286781d
NP
299 /*
300 * Once radix-tree replacement of page migration started, page_count
301 * *must* be zero. And, we don't want to call wait_on_page_locked()
302 * against a page without get_page().
303 * So, we use get_page_unless_zero(), here. Even failed, page fault
304 * will occur again.
305 */
306 if (!get_page_unless_zero(page))
307 goto out;
0697212a
CL
308 pte_unmap_unlock(ptep, ptl);
309 wait_on_page_locked(page);
310 put_page(page);
311 return;
312out:
313 pte_unmap_unlock(ptep, ptl);
314}
315
30dad309
NH
316void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
317 unsigned long address)
318{
319 spinlock_t *ptl = pte_lockptr(mm, pmd);
320 pte_t *ptep = pte_offset_map(pmd, address);
321 __migration_entry_wait(mm, ptep, ptl);
322}
323
cb900f41
KS
324void migration_entry_wait_huge(struct vm_area_struct *vma,
325 struct mm_struct *mm, pte_t *pte)
30dad309 326{
cb900f41 327 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
30dad309
NH
328 __migration_entry_wait(mm, pte, ptl);
329}
330
b969c4ab
MG
331#ifdef CONFIG_BLOCK
332/* Returns true if all buffers are successfully locked */
a6bc32b8
MG
333static bool buffer_migrate_lock_buffers(struct buffer_head *head,
334 enum migrate_mode mode)
b969c4ab
MG
335{
336 struct buffer_head *bh = head;
337
338 /* Simple case, sync compaction */
a6bc32b8 339 if (mode != MIGRATE_ASYNC) {
b969c4ab
MG
340 do {
341 get_bh(bh);
342 lock_buffer(bh);
343 bh = bh->b_this_page;
344
345 } while (bh != head);
346
347 return true;
348 }
349
350 /* async case, we cannot block on lock_buffer so use trylock_buffer */
351 do {
352 get_bh(bh);
353 if (!trylock_buffer(bh)) {
354 /*
355 * We failed to lock the buffer and cannot stall in
356 * async migration. Release the taken locks
357 */
358 struct buffer_head *failed_bh = bh;
359 put_bh(failed_bh);
360 bh = head;
361 while (bh != failed_bh) {
362 unlock_buffer(bh);
363 put_bh(bh);
364 bh = bh->b_this_page;
365 }
366 return false;
367 }
368
369 bh = bh->b_this_page;
370 } while (bh != head);
371 return true;
372}
373#else
374static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
a6bc32b8 375 enum migrate_mode mode)
b969c4ab
MG
376{
377 return true;
378}
379#endif /* CONFIG_BLOCK */
380
b20a3503 381/*
c3fcf8a5 382 * Replace the page in the mapping.
5b5c7120
CL
383 *
384 * The number of remaining references must be:
385 * 1 for anonymous pages without a mapping
386 * 2 for pages with a mapping
266cf658 387 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
b20a3503 388 */
36bc08cc 389int migrate_page_move_mapping(struct address_space *mapping,
b969c4ab 390 struct page *newpage, struct page *page,
8e321fef
BL
391 struct buffer_head *head, enum migrate_mode mode,
392 int extra_count)
b20a3503 393{
42cb14b1
HD
394 struct zone *oldzone, *newzone;
395 int dirty;
8e321fef 396 int expected_count = 1 + extra_count;
7cf9c2c7 397 void **pslot;
b20a3503 398
6c5240ae 399 if (!mapping) {
0e8c7d0f 400 /* Anonymous page without mapping */
8e321fef 401 if (page_count(page) != expected_count)
6c5240ae 402 return -EAGAIN;
cf4b769a
HD
403
404 /* No turning back from here */
cf4b769a
HD
405 newpage->index = page->index;
406 newpage->mapping = page->mapping;
407 if (PageSwapBacked(page))
fa9949da 408 __SetPageSwapBacked(newpage);
cf4b769a 409
78bd5209 410 return MIGRATEPAGE_SUCCESS;
6c5240ae
CL
411 }
412
42cb14b1
HD
413 oldzone = page_zone(page);
414 newzone = page_zone(newpage);
415
19fd6231 416 spin_lock_irq(&mapping->tree_lock);
b20a3503 417
7cf9c2c7
NP
418 pslot = radix_tree_lookup_slot(&mapping->page_tree,
419 page_index(page));
b20a3503 420
8e321fef 421 expected_count += 1 + page_has_private(page);
e286781d 422 if (page_count(page) != expected_count ||
29c1f677 423 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
19fd6231 424 spin_unlock_irq(&mapping->tree_lock);
e23ca00b 425 return -EAGAIN;
b20a3503
CL
426 }
427
fe896d18 428 if (!page_ref_freeze(page, expected_count)) {
19fd6231 429 spin_unlock_irq(&mapping->tree_lock);
e286781d
NP
430 return -EAGAIN;
431 }
432
b969c4ab
MG
433 /*
434 * In the async migration case of moving a page with buffers, lock the
435 * buffers using trylock before the mapping is moved. If the mapping
436 * was moved, we later failed to lock the buffers and could not move
437 * the mapping back due to an elevated page count, we would have to
438 * block waiting on other references to be dropped.
439 */
a6bc32b8
MG
440 if (mode == MIGRATE_ASYNC && head &&
441 !buffer_migrate_lock_buffers(head, mode)) {
fe896d18 442 page_ref_unfreeze(page, expected_count);
b969c4ab
MG
443 spin_unlock_irq(&mapping->tree_lock);
444 return -EAGAIN;
445 }
446
b20a3503 447 /*
cf4b769a
HD
448 * Now we know that no one else is looking at the page:
449 * no turning back from here.
b20a3503 450 */
cf4b769a
HD
451 newpage->index = page->index;
452 newpage->mapping = page->mapping;
7cf9c2c7 453 get_page(newpage); /* add cache reference */
6326fec1
NP
454 if (PageSwapBacked(page)) {
455 __SetPageSwapBacked(newpage);
456 if (PageSwapCache(page)) {
457 SetPageSwapCache(newpage);
458 set_page_private(newpage, page_private(page));
459 }
460 } else {
461 VM_BUG_ON_PAGE(PageSwapCache(page), page);
b20a3503
CL
462 }
463
42cb14b1
HD
464 /* Move dirty while page refs frozen and newpage not yet exposed */
465 dirty = PageDirty(page);
466 if (dirty) {
467 ClearPageDirty(page);
468 SetPageDirty(newpage);
469 }
470
6d75f366 471 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
7cf9c2c7
NP
472
473 /*
937a94c9
JG
474 * Drop cache reference from old page by unfreezing
475 * to one less reference.
7cf9c2c7
NP
476 * We know this isn't the last reference.
477 */
fe896d18 478 page_ref_unfreeze(page, expected_count - 1);
7cf9c2c7 479
42cb14b1
HD
480 spin_unlock(&mapping->tree_lock);
481 /* Leave irq disabled to prevent preemption while updating stats */
482
0e8c7d0f
CL
483 /*
484 * If moved to a different zone then also account
485 * the page for that zone. Other VM counters will be
486 * taken care of when we establish references to the
487 * new page and drop references to the old page.
488 *
489 * Note that anonymous pages are accounted for
4b9d0fab 490 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
0e8c7d0f
CL
491 * are mapped to swap space.
492 */
42cb14b1 493 if (newzone != oldzone) {
11fb9989
MG
494 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
495 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
42cb14b1 496 if (PageSwapBacked(page) && !PageSwapCache(page)) {
11fb9989
MG
497 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
498 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
42cb14b1
HD
499 }
500 if (dirty && mapping_cap_account_dirty(mapping)) {
11fb9989 501 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 502 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
11fb9989 503 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 504 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
42cb14b1 505 }
4b02108a 506 }
42cb14b1 507 local_irq_enable();
b20a3503 508
78bd5209 509 return MIGRATEPAGE_SUCCESS;
b20a3503 510}
1118dce7 511EXPORT_SYMBOL(migrate_page_move_mapping);
b20a3503 512
290408d4
NH
513/*
514 * The expected number of remaining references is the same as that
515 * of migrate_page_move_mapping().
516 */
517int migrate_huge_page_move_mapping(struct address_space *mapping,
518 struct page *newpage, struct page *page)
519{
520 int expected_count;
521 void **pslot;
522
290408d4
NH
523 spin_lock_irq(&mapping->tree_lock);
524
525 pslot = radix_tree_lookup_slot(&mapping->page_tree,
526 page_index(page));
527
528 expected_count = 2 + page_has_private(page);
529 if (page_count(page) != expected_count ||
29c1f677 530 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
290408d4
NH
531 spin_unlock_irq(&mapping->tree_lock);
532 return -EAGAIN;
533 }
534
fe896d18 535 if (!page_ref_freeze(page, expected_count)) {
290408d4
NH
536 spin_unlock_irq(&mapping->tree_lock);
537 return -EAGAIN;
538 }
539
cf4b769a
HD
540 newpage->index = page->index;
541 newpage->mapping = page->mapping;
6a93ca8f 542
290408d4
NH
543 get_page(newpage);
544
6d75f366 545 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
290408d4 546
fe896d18 547 page_ref_unfreeze(page, expected_count - 1);
290408d4
NH
548
549 spin_unlock_irq(&mapping->tree_lock);
6a93ca8f 550
78bd5209 551 return MIGRATEPAGE_SUCCESS;
290408d4
NH
552}
553
30b0a105
DH
554/*
555 * Gigantic pages are so large that we do not guarantee that page++ pointer
556 * arithmetic will work across the entire page. We need something more
557 * specialized.
558 */
559static void __copy_gigantic_page(struct page *dst, struct page *src,
560 int nr_pages)
561{
562 int i;
563 struct page *dst_base = dst;
564 struct page *src_base = src;
565
566 for (i = 0; i < nr_pages; ) {
567 cond_resched();
568 copy_highpage(dst, src);
569
570 i++;
571 dst = mem_map_next(dst, dst_base, i);
572 src = mem_map_next(src, src_base, i);
573 }
574}
575
576static void copy_huge_page(struct page *dst, struct page *src)
577{
578 int i;
579 int nr_pages;
580
581 if (PageHuge(src)) {
582 /* hugetlbfs page */
583 struct hstate *h = page_hstate(src);
584 nr_pages = pages_per_huge_page(h);
585
586 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
587 __copy_gigantic_page(dst, src, nr_pages);
588 return;
589 }
590 } else {
591 /* thp page */
592 BUG_ON(!PageTransHuge(src));
593 nr_pages = hpage_nr_pages(src);
594 }
595
596 for (i = 0; i < nr_pages; i++) {
597 cond_resched();
598 copy_highpage(dst + i, src + i);
599 }
600}
601
b20a3503
CL
602/*
603 * Copy the page to its new location
604 */
290408d4 605void migrate_page_copy(struct page *newpage, struct page *page)
b20a3503 606{
7851a45c
RR
607 int cpupid;
608
b32967ff 609 if (PageHuge(page) || PageTransHuge(page))
290408d4
NH
610 copy_huge_page(newpage, page);
611 else
612 copy_highpage(newpage, page);
b20a3503
CL
613
614 if (PageError(page))
615 SetPageError(newpage);
616 if (PageReferenced(page))
617 SetPageReferenced(newpage);
618 if (PageUptodate(page))
619 SetPageUptodate(newpage);
894bc310 620 if (TestClearPageActive(page)) {
309381fe 621 VM_BUG_ON_PAGE(PageUnevictable(page), page);
b20a3503 622 SetPageActive(newpage);
418b27ef
LS
623 } else if (TestClearPageUnevictable(page))
624 SetPageUnevictable(newpage);
b20a3503
CL
625 if (PageChecked(page))
626 SetPageChecked(newpage);
627 if (PageMappedToDisk(page))
628 SetPageMappedToDisk(newpage);
629
42cb14b1
HD
630 /* Move dirty on pages not done by migrate_page_move_mapping() */
631 if (PageDirty(page))
632 SetPageDirty(newpage);
b20a3503 633
33c3fc71
VD
634 if (page_is_young(page))
635 set_page_young(newpage);
636 if (page_is_idle(page))
637 set_page_idle(newpage);
638
7851a45c
RR
639 /*
640 * Copy NUMA information to the new page, to prevent over-eager
641 * future migrations of this same page.
642 */
643 cpupid = page_cpupid_xchg_last(page, -1);
644 page_cpupid_xchg_last(newpage, cpupid);
645
e9995ef9 646 ksm_migrate_page(newpage, page);
c8d6553b
HD
647 /*
648 * Please do not reorder this without considering how mm/ksm.c's
649 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
650 */
b3b3a99c
NH
651 if (PageSwapCache(page))
652 ClearPageSwapCache(page);
b20a3503
CL
653 ClearPagePrivate(page);
654 set_page_private(page, 0);
b20a3503
CL
655
656 /*
657 * If any waiters have accumulated on the new page then
658 * wake them up.
659 */
660 if (PageWriteback(newpage))
661 end_page_writeback(newpage);
d435edca
VB
662
663 copy_page_owner(page, newpage);
74485cf2
JW
664
665 mem_cgroup_migrate(page, newpage);
b20a3503 666}
1118dce7 667EXPORT_SYMBOL(migrate_page_copy);
b20a3503 668
1d8b85cc
CL
669/************************************************************
670 * Migration functions
671 ***********************************************************/
672
b20a3503 673/*
bda807d4 674 * Common logic to directly migrate a single LRU page suitable for
266cf658 675 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
676 *
677 * Pages are locked upon entry and exit.
678 */
2d1db3b1 679int migrate_page(struct address_space *mapping,
a6bc32b8
MG
680 struct page *newpage, struct page *page,
681 enum migrate_mode mode)
b20a3503
CL
682{
683 int rc;
684
685 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
686
8e321fef 687 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
b20a3503 688
78bd5209 689 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
690 return rc;
691
692 migrate_page_copy(newpage, page);
78bd5209 693 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
694}
695EXPORT_SYMBOL(migrate_page);
696
9361401e 697#ifdef CONFIG_BLOCK
1d8b85cc
CL
698/*
699 * Migration function for pages with buffers. This function can only be used
700 * if the underlying filesystem guarantees that no other references to "page"
701 * exist.
702 */
2d1db3b1 703int buffer_migrate_page(struct address_space *mapping,
a6bc32b8 704 struct page *newpage, struct page *page, enum migrate_mode mode)
1d8b85cc 705{
1d8b85cc
CL
706 struct buffer_head *bh, *head;
707 int rc;
708
1d8b85cc 709 if (!page_has_buffers(page))
a6bc32b8 710 return migrate_page(mapping, newpage, page, mode);
1d8b85cc
CL
711
712 head = page_buffers(page);
713
8e321fef 714 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
1d8b85cc 715
78bd5209 716 if (rc != MIGRATEPAGE_SUCCESS)
1d8b85cc
CL
717 return rc;
718
b969c4ab
MG
719 /*
720 * In the async case, migrate_page_move_mapping locked the buffers
721 * with an IRQ-safe spinlock held. In the sync case, the buffers
722 * need to be locked now
723 */
a6bc32b8
MG
724 if (mode != MIGRATE_ASYNC)
725 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
1d8b85cc
CL
726
727 ClearPagePrivate(page);
728 set_page_private(newpage, page_private(page));
729 set_page_private(page, 0);
730 put_page(page);
731 get_page(newpage);
732
733 bh = head;
734 do {
735 set_bh_page(bh, newpage, bh_offset(bh));
736 bh = bh->b_this_page;
737
738 } while (bh != head);
739
740 SetPagePrivate(newpage);
741
742 migrate_page_copy(newpage, page);
743
744 bh = head;
745 do {
746 unlock_buffer(bh);
747 put_bh(bh);
748 bh = bh->b_this_page;
749
750 } while (bh != head);
751
78bd5209 752 return MIGRATEPAGE_SUCCESS;
1d8b85cc
CL
753}
754EXPORT_SYMBOL(buffer_migrate_page);
9361401e 755#endif
1d8b85cc 756
04e62a29
CL
757/*
758 * Writeback a page to clean the dirty state
759 */
760static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 761{
04e62a29
CL
762 struct writeback_control wbc = {
763 .sync_mode = WB_SYNC_NONE,
764 .nr_to_write = 1,
765 .range_start = 0,
766 .range_end = LLONG_MAX,
04e62a29
CL
767 .for_reclaim = 1
768 };
769 int rc;
770
771 if (!mapping->a_ops->writepage)
772 /* No write method for the address space */
773 return -EINVAL;
774
775 if (!clear_page_dirty_for_io(page))
776 /* Someone else already triggered a write */
777 return -EAGAIN;
778
8351a6e4 779 /*
04e62a29
CL
780 * A dirty page may imply that the underlying filesystem has
781 * the page on some queue. So the page must be clean for
782 * migration. Writeout may mean we loose the lock and the
783 * page state is no longer what we checked for earlier.
784 * At this point we know that the migration attempt cannot
785 * be successful.
8351a6e4 786 */
e388466d 787 remove_migration_ptes(page, page, false);
8351a6e4 788
04e62a29 789 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 790
04e62a29
CL
791 if (rc != AOP_WRITEPAGE_ACTIVATE)
792 /* unlocked. Relock */
793 lock_page(page);
794
bda8550d 795 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
796}
797
798/*
799 * Default handling if a filesystem does not provide a migration function.
800 */
801static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 802 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 803{
b969c4ab 804 if (PageDirty(page)) {
a6bc32b8
MG
805 /* Only writeback pages in full synchronous migration */
806 if (mode != MIGRATE_SYNC)
b969c4ab 807 return -EBUSY;
04e62a29 808 return writeout(mapping, page);
b969c4ab 809 }
8351a6e4
CL
810
811 /*
812 * Buffers may be managed in a filesystem specific way.
813 * We must have no buffers or drop them.
814 */
266cf658 815 if (page_has_private(page) &&
8351a6e4
CL
816 !try_to_release_page(page, GFP_KERNEL))
817 return -EAGAIN;
818
a6bc32b8 819 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
820}
821
e24f0b8f
CL
822/*
823 * Move a page to a newly allocated page
824 * The page is locked and all ptes have been successfully removed.
825 *
826 * The new page will have replaced the old page if this function
827 * is successful.
894bc310
LS
828 *
829 * Return value:
830 * < 0 - error code
78bd5209 831 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 832 */
3fe2011f 833static int move_to_new_page(struct page *newpage, struct page *page,
5c3f9a67 834 enum migrate_mode mode)
e24f0b8f
CL
835{
836 struct address_space *mapping;
bda807d4
MK
837 int rc = -EAGAIN;
838 bool is_lru = !__PageMovable(page);
e24f0b8f 839
7db7671f
HD
840 VM_BUG_ON_PAGE(!PageLocked(page), page);
841 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
e24f0b8f 842
e24f0b8f 843 mapping = page_mapping(page);
bda807d4
MK
844
845 if (likely(is_lru)) {
846 if (!mapping)
847 rc = migrate_page(mapping, newpage, page, mode);
848 else if (mapping->a_ops->migratepage)
849 /*
850 * Most pages have a mapping and most filesystems
851 * provide a migratepage callback. Anonymous pages
852 * are part of swap space which also has its own
853 * migratepage callback. This is the most common path
854 * for page migration.
855 */
856 rc = mapping->a_ops->migratepage(mapping, newpage,
857 page, mode);
858 else
859 rc = fallback_migrate_page(mapping, newpage,
860 page, mode);
861 } else {
e24f0b8f 862 /*
bda807d4
MK
863 * In case of non-lru page, it could be released after
864 * isolation step. In that case, we shouldn't try migration.
e24f0b8f 865 */
bda807d4
MK
866 VM_BUG_ON_PAGE(!PageIsolated(page), page);
867 if (!PageMovable(page)) {
868 rc = MIGRATEPAGE_SUCCESS;
869 __ClearPageIsolated(page);
870 goto out;
871 }
872
873 rc = mapping->a_ops->migratepage(mapping, newpage,
874 page, mode);
875 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
876 !PageIsolated(page));
877 }
e24f0b8f 878
5c3f9a67
HD
879 /*
880 * When successful, old pagecache page->mapping must be cleared before
881 * page is freed; but stats require that PageAnon be left as PageAnon.
882 */
883 if (rc == MIGRATEPAGE_SUCCESS) {
bda807d4
MK
884 if (__PageMovable(page)) {
885 VM_BUG_ON_PAGE(!PageIsolated(page), page);
886
887 /*
888 * We clear PG_movable under page_lock so any compactor
889 * cannot try to migrate this page.
890 */
891 __ClearPageIsolated(page);
892 }
893
894 /*
895 * Anonymous and movable page->mapping will be cleard by
896 * free_pages_prepare so don't reset it here for keeping
897 * the type to work PageAnon, for example.
898 */
899 if (!PageMappingFlags(page))
5c3f9a67 900 page->mapping = NULL;
3fe2011f 901 }
bda807d4 902out:
e24f0b8f
CL
903 return rc;
904}
905
0dabec93 906static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 907 int force, enum migrate_mode mode)
e24f0b8f 908{
0dabec93 909 int rc = -EAGAIN;
2ebba6b7 910 int page_was_mapped = 0;
3f6c8272 911 struct anon_vma *anon_vma = NULL;
bda807d4 912 bool is_lru = !__PageMovable(page);
95a402c3 913
529ae9aa 914 if (!trylock_page(page)) {
a6bc32b8 915 if (!force || mode == MIGRATE_ASYNC)
0dabec93 916 goto out;
3e7d3449
MG
917
918 /*
919 * It's not safe for direct compaction to call lock_page.
920 * For example, during page readahead pages are added locked
921 * to the LRU. Later, when the IO completes the pages are
922 * marked uptodate and unlocked. However, the queueing
923 * could be merging multiple pages for one bio (e.g.
924 * mpage_readpages). If an allocation happens for the
925 * second or third page, the process can end up locking
926 * the same page twice and deadlocking. Rather than
927 * trying to be clever about what pages can be locked,
928 * avoid the use of lock_page for direct compaction
929 * altogether.
930 */
931 if (current->flags & PF_MEMALLOC)
0dabec93 932 goto out;
3e7d3449 933
e24f0b8f
CL
934 lock_page(page);
935 }
936
937 if (PageWriteback(page)) {
11bc82d6 938 /*
fed5b64a 939 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
940 * necessary to wait for PageWriteback. In the async case,
941 * the retry loop is too short and in the sync-light case,
942 * the overhead of stalling is too much
11bc82d6 943 */
a6bc32b8 944 if (mode != MIGRATE_SYNC) {
11bc82d6 945 rc = -EBUSY;
0a31bc97 946 goto out_unlock;
11bc82d6
AA
947 }
948 if (!force)
0a31bc97 949 goto out_unlock;
e24f0b8f
CL
950 wait_on_page_writeback(page);
951 }
03f15c86 952
e24f0b8f 953 /*
dc386d4d
KH
954 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
955 * we cannot notice that anon_vma is freed while we migrates a page.
1ce82b69 956 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 957 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
958 * File Caches may use write_page() or lock_page() in migration, then,
959 * just care Anon page here.
03f15c86
HD
960 *
961 * Only page_get_anon_vma() understands the subtleties of
962 * getting a hold on an anon_vma from outside one of its mms.
963 * But if we cannot get anon_vma, then we won't need it anyway,
964 * because that implies that the anon page is no longer mapped
965 * (and cannot be remapped so long as we hold the page lock).
dc386d4d 966 */
03f15c86 967 if (PageAnon(page) && !PageKsm(page))
746b18d4 968 anon_vma = page_get_anon_vma(page);
62e1c553 969
7db7671f
HD
970 /*
971 * Block others from accessing the new page when we get around to
972 * establishing additional references. We are usually the only one
973 * holding a reference to newpage at this point. We used to have a BUG
974 * here if trylock_page(newpage) fails, but would like to allow for
975 * cases where there might be a race with the previous use of newpage.
976 * This is much like races on refcount of oldpage: just don't BUG().
977 */
978 if (unlikely(!trylock_page(newpage)))
979 goto out_unlock;
980
bda807d4
MK
981 if (unlikely(!is_lru)) {
982 rc = move_to_new_page(newpage, page, mode);
983 goto out_unlock_both;
984 }
985
dc386d4d 986 /*
62e1c553
SL
987 * Corner case handling:
988 * 1. When a new swap-cache page is read into, it is added to the LRU
989 * and treated as swapcache but it has no rmap yet.
990 * Calling try_to_unmap() against a page->mapping==NULL page will
991 * trigger a BUG. So handle it here.
992 * 2. An orphaned page (see truncate_complete_page) might have
993 * fs-private metadata. The page can be picked up due to memory
994 * offlining. Everywhere else except page reclaim, the page is
995 * invisible to the vm, so the page can not be migrated. So try to
996 * free the metadata, so the page can be freed.
e24f0b8f 997 */
62e1c553 998 if (!page->mapping) {
309381fe 999 VM_BUG_ON_PAGE(PageAnon(page), page);
1ce82b69 1000 if (page_has_private(page)) {
62e1c553 1001 try_to_free_buffers(page);
7db7671f 1002 goto out_unlock_both;
62e1c553 1003 }
7db7671f
HD
1004 } else if (page_mapped(page)) {
1005 /* Establish migration ptes */
03f15c86
HD
1006 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1007 page);
2ebba6b7 1008 try_to_unmap(page,
da1b13cc 1009 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
2ebba6b7
HD
1010 page_was_mapped = 1;
1011 }
dc386d4d 1012
e6a1530d 1013 if (!page_mapped(page))
5c3f9a67 1014 rc = move_to_new_page(newpage, page, mode);
e24f0b8f 1015
5c3f9a67
HD
1016 if (page_was_mapped)
1017 remove_migration_ptes(page,
e388466d 1018 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
3f6c8272 1019
7db7671f
HD
1020out_unlock_both:
1021 unlock_page(newpage);
1022out_unlock:
3f6c8272 1023 /* Drop an anon_vma reference if we took one */
76545066 1024 if (anon_vma)
9e60109f 1025 put_anon_vma(anon_vma);
e24f0b8f 1026 unlock_page(page);
0dabec93 1027out:
c6c919eb
MK
1028 /*
1029 * If migration is successful, decrease refcount of the newpage
1030 * which will not free the page because new page owner increased
1031 * refcounter. As well, if it is LRU page, add the page to LRU
1032 * list in here.
1033 */
1034 if (rc == MIGRATEPAGE_SUCCESS) {
b1123ea6 1035 if (unlikely(__PageMovable(newpage)))
c6c919eb
MK
1036 put_page(newpage);
1037 else
1038 putback_lru_page(newpage);
1039 }
1040
0dabec93
MK
1041 return rc;
1042}
95a402c3 1043
ef2a5153
GU
1044/*
1045 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1046 * around it.
1047 */
1048#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1049#define ICE_noinline noinline
1050#else
1051#define ICE_noinline
1052#endif
1053
0dabec93
MK
1054/*
1055 * Obtain the lock on page, remove all ptes and migrate the page
1056 * to the newly allocated page in newpage.
1057 */
ef2a5153
GU
1058static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1059 free_page_t put_new_page,
1060 unsigned long private, struct page *page,
add05cec
NH
1061 int force, enum migrate_mode mode,
1062 enum migrate_reason reason)
0dabec93 1063{
2def7424 1064 int rc = MIGRATEPAGE_SUCCESS;
0dabec93 1065 int *result = NULL;
2def7424 1066 struct page *newpage;
0dabec93 1067
2def7424 1068 newpage = get_new_page(page, private, &result);
0dabec93
MK
1069 if (!newpage)
1070 return -ENOMEM;
1071
1072 if (page_count(page) == 1) {
1073 /* page was freed from under us. So we are done. */
c6c919eb
MK
1074 ClearPageActive(page);
1075 ClearPageUnevictable(page);
bda807d4
MK
1076 if (unlikely(__PageMovable(page))) {
1077 lock_page(page);
1078 if (!PageMovable(page))
1079 __ClearPageIsolated(page);
1080 unlock_page(page);
1081 }
c6c919eb
MK
1082 if (put_new_page)
1083 put_new_page(newpage, private);
1084 else
1085 put_page(newpage);
0dabec93
MK
1086 goto out;
1087 }
1088
4d2fa965
KS
1089 if (unlikely(PageTransHuge(page))) {
1090 lock_page(page);
1091 rc = split_huge_page(page);
1092 unlock_page(page);
1093 if (rc)
0dabec93 1094 goto out;
4d2fa965 1095 }
0dabec93 1096
9c620e2b 1097 rc = __unmap_and_move(page, newpage, force, mode);
c6c919eb 1098 if (rc == MIGRATEPAGE_SUCCESS)
7cd12b4a 1099 set_page_owner_migrate_reason(newpage, reason);
bf6bddf1 1100
0dabec93 1101out:
e24f0b8f 1102 if (rc != -EAGAIN) {
0dabec93
MK
1103 /*
1104 * A page that has been migrated has all references
1105 * removed and will be freed. A page that has not been
1106 * migrated will have kepts its references and be
1107 * restored.
1108 */
1109 list_del(&page->lru);
6afcf8ef
ML
1110
1111 /*
1112 * Compaction can migrate also non-LRU pages which are
1113 * not accounted to NR_ISOLATED_*. They can be recognized
1114 * as __PageMovable
1115 */
1116 if (likely(!__PageMovable(page)))
1117 dec_node_page_state(page, NR_ISOLATED_ANON +
1118 page_is_file_cache(page));
c6c919eb
MK
1119 }
1120
1121 /*
1122 * If migration is successful, releases reference grabbed during
1123 * isolation. Otherwise, restore the page to right list unless
1124 * we want to retry.
1125 */
1126 if (rc == MIGRATEPAGE_SUCCESS) {
1127 put_page(page);
1128 if (reason == MR_MEMORY_FAILURE) {
d7e69488 1129 /*
c6c919eb
MK
1130 * Set PG_HWPoison on just freed page
1131 * intentionally. Although it's rather weird,
1132 * it's how HWPoison flag works at the moment.
d7e69488 1133 */
da1b13cc
WL
1134 if (!test_set_page_hwpoison(page))
1135 num_poisoned_pages_inc();
c6c919eb
MK
1136 }
1137 } else {
bda807d4
MK
1138 if (rc != -EAGAIN) {
1139 if (likely(!__PageMovable(page))) {
1140 putback_lru_page(page);
1141 goto put_new;
1142 }
1143
1144 lock_page(page);
1145 if (PageMovable(page))
1146 putback_movable_page(page);
1147 else
1148 __ClearPageIsolated(page);
1149 unlock_page(page);
1150 put_page(page);
1151 }
1152put_new:
c6c919eb
MK
1153 if (put_new_page)
1154 put_new_page(newpage, private);
1155 else
1156 put_page(newpage);
e24f0b8f 1157 }
68711a74 1158
742755a1
CL
1159 if (result) {
1160 if (rc)
1161 *result = rc;
1162 else
1163 *result = page_to_nid(newpage);
1164 }
e24f0b8f
CL
1165 return rc;
1166}
1167
290408d4
NH
1168/*
1169 * Counterpart of unmap_and_move_page() for hugepage migration.
1170 *
1171 * This function doesn't wait the completion of hugepage I/O
1172 * because there is no race between I/O and migration for hugepage.
1173 * Note that currently hugepage I/O occurs only in direct I/O
1174 * where no lock is held and PG_writeback is irrelevant,
1175 * and writeback status of all subpages are counted in the reference
1176 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1177 * under direct I/O, the reference of the head page is 512 and a bit more.)
1178 * This means that when we try to migrate hugepage whose subpages are
1179 * doing direct I/O, some references remain after try_to_unmap() and
1180 * hugepage migration fails without data corruption.
1181 *
1182 * There is also no race when direct I/O is issued on the page under migration,
1183 * because then pte is replaced with migration swap entry and direct I/O code
1184 * will wait in the page fault for migration to complete.
1185 */
1186static int unmap_and_move_huge_page(new_page_t get_new_page,
68711a74
DR
1187 free_page_t put_new_page, unsigned long private,
1188 struct page *hpage, int force,
7cd12b4a 1189 enum migrate_mode mode, int reason)
290408d4 1190{
2def7424 1191 int rc = -EAGAIN;
290408d4 1192 int *result = NULL;
2ebba6b7 1193 int page_was_mapped = 0;
32665f2b 1194 struct page *new_hpage;
290408d4
NH
1195 struct anon_vma *anon_vma = NULL;
1196
83467efb
NH
1197 /*
1198 * Movability of hugepages depends on architectures and hugepage size.
1199 * This check is necessary because some callers of hugepage migration
1200 * like soft offline and memory hotremove don't walk through page
1201 * tables or check whether the hugepage is pmd-based or not before
1202 * kicking migration.
1203 */
100873d7 1204 if (!hugepage_migration_supported(page_hstate(hpage))) {
32665f2b 1205 putback_active_hugepage(hpage);
83467efb 1206 return -ENOSYS;
32665f2b 1207 }
83467efb 1208
32665f2b 1209 new_hpage = get_new_page(hpage, private, &result);
290408d4
NH
1210 if (!new_hpage)
1211 return -ENOMEM;
1212
290408d4 1213 if (!trylock_page(hpage)) {
a6bc32b8 1214 if (!force || mode != MIGRATE_SYNC)
290408d4
NH
1215 goto out;
1216 lock_page(hpage);
1217 }
1218
746b18d4
PZ
1219 if (PageAnon(hpage))
1220 anon_vma = page_get_anon_vma(hpage);
290408d4 1221
7db7671f
HD
1222 if (unlikely(!trylock_page(new_hpage)))
1223 goto put_anon;
1224
2ebba6b7
HD
1225 if (page_mapped(hpage)) {
1226 try_to_unmap(hpage,
1227 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1228 page_was_mapped = 1;
1229 }
290408d4
NH
1230
1231 if (!page_mapped(hpage))
5c3f9a67 1232 rc = move_to_new_page(new_hpage, hpage, mode);
290408d4 1233
5c3f9a67
HD
1234 if (page_was_mapped)
1235 remove_migration_ptes(hpage,
e388466d 1236 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
290408d4 1237
7db7671f
HD
1238 unlock_page(new_hpage);
1239
1240put_anon:
fd4a4663 1241 if (anon_vma)
9e60109f 1242 put_anon_vma(anon_vma);
8e6ac7fa 1243
2def7424 1244 if (rc == MIGRATEPAGE_SUCCESS) {
8e6ac7fa 1245 hugetlb_cgroup_migrate(hpage, new_hpage);
2def7424 1246 put_new_page = NULL;
7cd12b4a 1247 set_page_owner_migrate_reason(new_hpage, reason);
2def7424 1248 }
8e6ac7fa 1249
290408d4 1250 unlock_page(hpage);
09761333 1251out:
b8ec1cee
NH
1252 if (rc != -EAGAIN)
1253 putback_active_hugepage(hpage);
68711a74
DR
1254
1255 /*
1256 * If migration was not successful and there's a freeing callback, use
1257 * it. Otherwise, put_page() will drop the reference grabbed during
1258 * isolation.
1259 */
2def7424 1260 if (put_new_page)
68711a74
DR
1261 put_new_page(new_hpage, private);
1262 else
3aaa76e1 1263 putback_active_hugepage(new_hpage);
68711a74 1264
290408d4
NH
1265 if (result) {
1266 if (rc)
1267 *result = rc;
1268 else
1269 *result = page_to_nid(new_hpage);
1270 }
1271 return rc;
1272}
1273
b20a3503 1274/*
c73e5c9c
SB
1275 * migrate_pages - migrate the pages specified in a list, to the free pages
1276 * supplied as the target for the page migration
b20a3503 1277 *
c73e5c9c
SB
1278 * @from: The list of pages to be migrated.
1279 * @get_new_page: The function used to allocate free pages to be used
1280 * as the target of the page migration.
68711a74
DR
1281 * @put_new_page: The function used to free target pages if migration
1282 * fails, or NULL if no special handling is necessary.
c73e5c9c
SB
1283 * @private: Private data to be passed on to get_new_page()
1284 * @mode: The migration mode that specifies the constraints for
1285 * page migration, if any.
1286 * @reason: The reason for page migration.
b20a3503 1287 *
c73e5c9c
SB
1288 * The function returns after 10 attempts or if no pages are movable any more
1289 * because the list has become empty or no retryable pages exist any more.
14e0f9bc 1290 * The caller should call putback_movable_pages() to return pages to the LRU
28bd6578 1291 * or free list only if ret != 0.
b20a3503 1292 *
c73e5c9c 1293 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1294 */
9c620e2b 1295int migrate_pages(struct list_head *from, new_page_t get_new_page,
68711a74
DR
1296 free_page_t put_new_page, unsigned long private,
1297 enum migrate_mode mode, int reason)
b20a3503 1298{
e24f0b8f 1299 int retry = 1;
b20a3503 1300 int nr_failed = 0;
5647bc29 1301 int nr_succeeded = 0;
b20a3503
CL
1302 int pass = 0;
1303 struct page *page;
1304 struct page *page2;
1305 int swapwrite = current->flags & PF_SWAPWRITE;
1306 int rc;
1307
1308 if (!swapwrite)
1309 current->flags |= PF_SWAPWRITE;
1310
e24f0b8f
CL
1311 for(pass = 0; pass < 10 && retry; pass++) {
1312 retry = 0;
b20a3503 1313
e24f0b8f 1314 list_for_each_entry_safe(page, page2, from, lru) {
e24f0b8f 1315 cond_resched();
2d1db3b1 1316
31caf665
NH
1317 if (PageHuge(page))
1318 rc = unmap_and_move_huge_page(get_new_page,
68711a74 1319 put_new_page, private, page,
7cd12b4a 1320 pass > 2, mode, reason);
31caf665 1321 else
68711a74 1322 rc = unmap_and_move(get_new_page, put_new_page,
add05cec
NH
1323 private, page, pass > 2, mode,
1324 reason);
2d1db3b1 1325
e24f0b8f 1326 switch(rc) {
95a402c3 1327 case -ENOMEM:
dfef2ef4 1328 nr_failed++;
95a402c3 1329 goto out;
e24f0b8f 1330 case -EAGAIN:
2d1db3b1 1331 retry++;
e24f0b8f 1332 break;
78bd5209 1333 case MIGRATEPAGE_SUCCESS:
5647bc29 1334 nr_succeeded++;
e24f0b8f
CL
1335 break;
1336 default:
354a3363
NH
1337 /*
1338 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1339 * unlike -EAGAIN case, the failed page is
1340 * removed from migration page list and not
1341 * retried in the next outer loop.
1342 */
2d1db3b1 1343 nr_failed++;
e24f0b8f 1344 break;
2d1db3b1 1345 }
b20a3503
CL
1346 }
1347 }
f2f81fb2
VB
1348 nr_failed += retry;
1349 rc = nr_failed;
95a402c3 1350out:
5647bc29
MG
1351 if (nr_succeeded)
1352 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1353 if (nr_failed)
1354 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1355 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1356
b20a3503
CL
1357 if (!swapwrite)
1358 current->flags &= ~PF_SWAPWRITE;
1359
78bd5209 1360 return rc;
b20a3503 1361}
95a402c3 1362
742755a1
CL
1363#ifdef CONFIG_NUMA
1364/*
1365 * Move a list of individual pages
1366 */
1367struct page_to_node {
1368 unsigned long addr;
1369 struct page *page;
1370 int node;
1371 int status;
1372};
1373
1374static struct page *new_page_node(struct page *p, unsigned long private,
1375 int **result)
1376{
1377 struct page_to_node *pm = (struct page_to_node *)private;
1378
1379 while (pm->node != MAX_NUMNODES && pm->page != p)
1380 pm++;
1381
1382 if (pm->node == MAX_NUMNODES)
1383 return NULL;
1384
1385 *result = &pm->status;
1386
e632a938
NH
1387 if (PageHuge(p))
1388 return alloc_huge_page_node(page_hstate(compound_head(p)),
1389 pm->node);
1390 else
96db800f 1391 return __alloc_pages_node(pm->node,
e97ca8e5 1392 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
742755a1
CL
1393}
1394
1395/*
1396 * Move a set of pages as indicated in the pm array. The addr
1397 * field must be set to the virtual address of the page to be moved
1398 * and the node number must contain a valid target node.
5e9a0f02 1399 * The pm array ends with node = MAX_NUMNODES.
742755a1 1400 */
5e9a0f02
BG
1401static int do_move_page_to_node_array(struct mm_struct *mm,
1402 struct page_to_node *pm,
1403 int migrate_all)
742755a1
CL
1404{
1405 int err;
1406 struct page_to_node *pp;
1407 LIST_HEAD(pagelist);
1408
1409 down_read(&mm->mmap_sem);
1410
1411 /*
1412 * Build a list of pages to migrate
1413 */
742755a1
CL
1414 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1415 struct vm_area_struct *vma;
1416 struct page *page;
1417
742755a1
CL
1418 err = -EFAULT;
1419 vma = find_vma(mm, pp->addr);
70384dc6 1420 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
742755a1
CL
1421 goto set_status;
1422
d899844e
KS
1423 /* FOLL_DUMP to ignore special (like zero) pages */
1424 page = follow_page(vma, pp->addr,
1425 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
89f5b7da
LT
1426
1427 err = PTR_ERR(page);
1428 if (IS_ERR(page))
1429 goto set_status;
1430
742755a1
CL
1431 err = -ENOENT;
1432 if (!page)
1433 goto set_status;
1434
742755a1
CL
1435 pp->page = page;
1436 err = page_to_nid(page);
1437
1438 if (err == pp->node)
1439 /*
1440 * Node already in the right place
1441 */
1442 goto put_and_set;
1443
1444 err = -EACCES;
1445 if (page_mapcount(page) > 1 &&
1446 !migrate_all)
1447 goto put_and_set;
1448
e632a938 1449 if (PageHuge(page)) {
e66f17ff
NH
1450 if (PageHead(page))
1451 isolate_huge_page(page, &pagelist);
e632a938
NH
1452 goto put_and_set;
1453 }
1454
62695a84 1455 err = isolate_lru_page(page);
6d9c285a 1456 if (!err) {
62695a84 1457 list_add_tail(&page->lru, &pagelist);
599d0c95 1458 inc_node_page_state(page, NR_ISOLATED_ANON +
6d9c285a
KM
1459 page_is_file_cache(page));
1460 }
742755a1
CL
1461put_and_set:
1462 /*
1463 * Either remove the duplicate refcount from
1464 * isolate_lru_page() or drop the page ref if it was
1465 * not isolated.
1466 */
1467 put_page(page);
1468set_status:
1469 pp->status = err;
1470 }
1471
e78bbfa8 1472 err = 0;
cf608ac1 1473 if (!list_empty(&pagelist)) {
68711a74 1474 err = migrate_pages(&pagelist, new_page_node, NULL,
9c620e2b 1475 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
cf608ac1 1476 if (err)
e632a938 1477 putback_movable_pages(&pagelist);
cf608ac1 1478 }
742755a1
CL
1479
1480 up_read(&mm->mmap_sem);
1481 return err;
1482}
1483
5e9a0f02
BG
1484/*
1485 * Migrate an array of page address onto an array of nodes and fill
1486 * the corresponding array of status.
1487 */
3268c63e 1488static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1489 unsigned long nr_pages,
1490 const void __user * __user *pages,
1491 const int __user *nodes,
1492 int __user *status, int flags)
1493{
3140a227 1494 struct page_to_node *pm;
3140a227
BG
1495 unsigned long chunk_nr_pages;
1496 unsigned long chunk_start;
1497 int err;
5e9a0f02 1498
3140a227
BG
1499 err = -ENOMEM;
1500 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1501 if (!pm)
5e9a0f02 1502 goto out;
35282a2d
BG
1503
1504 migrate_prep();
1505
5e9a0f02 1506 /*
3140a227
BG
1507 * Store a chunk of page_to_node array in a page,
1508 * but keep the last one as a marker
5e9a0f02 1509 */
3140a227 1510 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
5e9a0f02 1511
3140a227
BG
1512 for (chunk_start = 0;
1513 chunk_start < nr_pages;
1514 chunk_start += chunk_nr_pages) {
1515 int j;
5e9a0f02 1516
3140a227
BG
1517 if (chunk_start + chunk_nr_pages > nr_pages)
1518 chunk_nr_pages = nr_pages - chunk_start;
1519
1520 /* fill the chunk pm with addrs and nodes from user-space */
1521 for (j = 0; j < chunk_nr_pages; j++) {
1522 const void __user *p;
5e9a0f02
BG
1523 int node;
1524
3140a227
BG
1525 err = -EFAULT;
1526 if (get_user(p, pages + j + chunk_start))
1527 goto out_pm;
1528 pm[j].addr = (unsigned long) p;
1529
1530 if (get_user(node, nodes + j + chunk_start))
5e9a0f02
BG
1531 goto out_pm;
1532
1533 err = -ENODEV;
6f5a55f1
LT
1534 if (node < 0 || node >= MAX_NUMNODES)
1535 goto out_pm;
1536
389162c2 1537 if (!node_state(node, N_MEMORY))
5e9a0f02
BG
1538 goto out_pm;
1539
1540 err = -EACCES;
1541 if (!node_isset(node, task_nodes))
1542 goto out_pm;
1543
3140a227
BG
1544 pm[j].node = node;
1545 }
1546
1547 /* End marker for this chunk */
1548 pm[chunk_nr_pages].node = MAX_NUMNODES;
1549
1550 /* Migrate this chunk */
1551 err = do_move_page_to_node_array(mm, pm,
1552 flags & MPOL_MF_MOVE_ALL);
1553 if (err < 0)
1554 goto out_pm;
5e9a0f02 1555
5e9a0f02 1556 /* Return status information */
3140a227
BG
1557 for (j = 0; j < chunk_nr_pages; j++)
1558 if (put_user(pm[j].status, status + j + chunk_start)) {
5e9a0f02 1559 err = -EFAULT;
3140a227
BG
1560 goto out_pm;
1561 }
1562 }
1563 err = 0;
5e9a0f02
BG
1564
1565out_pm:
3140a227 1566 free_page((unsigned long)pm);
5e9a0f02
BG
1567out:
1568 return err;
1569}
1570
742755a1 1571/*
2f007e74 1572 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1573 */
80bba129
BG
1574static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1575 const void __user **pages, int *status)
742755a1 1576{
2f007e74 1577 unsigned long i;
2f007e74 1578
742755a1
CL
1579 down_read(&mm->mmap_sem);
1580
2f007e74 1581 for (i = 0; i < nr_pages; i++) {
80bba129 1582 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1583 struct vm_area_struct *vma;
1584 struct page *page;
c095adbc 1585 int err = -EFAULT;
2f007e74
BG
1586
1587 vma = find_vma(mm, addr);
70384dc6 1588 if (!vma || addr < vma->vm_start)
742755a1
CL
1589 goto set_status;
1590
d899844e
KS
1591 /* FOLL_DUMP to ignore special (like zero) pages */
1592 page = follow_page(vma, addr, FOLL_DUMP);
89f5b7da
LT
1593
1594 err = PTR_ERR(page);
1595 if (IS_ERR(page))
1596 goto set_status;
1597
d899844e 1598 err = page ? page_to_nid(page) : -ENOENT;
742755a1 1599set_status:
80bba129
BG
1600 *status = err;
1601
1602 pages++;
1603 status++;
1604 }
1605
1606 up_read(&mm->mmap_sem);
1607}
1608
1609/*
1610 * Determine the nodes of a user array of pages and store it in
1611 * a user array of status.
1612 */
1613static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1614 const void __user * __user *pages,
1615 int __user *status)
1616{
1617#define DO_PAGES_STAT_CHUNK_NR 16
1618 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1619 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1620
87b8d1ad
PA
1621 while (nr_pages) {
1622 unsigned long chunk_nr;
80bba129 1623
87b8d1ad
PA
1624 chunk_nr = nr_pages;
1625 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1626 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1627
1628 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1629 break;
80bba129
BG
1630
1631 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1632
87b8d1ad
PA
1633 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1634 break;
742755a1 1635
87b8d1ad
PA
1636 pages += chunk_nr;
1637 status += chunk_nr;
1638 nr_pages -= chunk_nr;
1639 }
1640 return nr_pages ? -EFAULT : 0;
742755a1
CL
1641}
1642
1643/*
1644 * Move a list of pages in the address space of the currently executing
1645 * process.
1646 */
938bb9f5
HC
1647SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1648 const void __user * __user *, pages,
1649 const int __user *, nodes,
1650 int __user *, status, int, flags)
742755a1 1651{
c69e8d9c 1652 const struct cred *cred = current_cred(), *tcred;
742755a1 1653 struct task_struct *task;
742755a1 1654 struct mm_struct *mm;
5e9a0f02 1655 int err;
3268c63e 1656 nodemask_t task_nodes;
742755a1
CL
1657
1658 /* Check flags */
1659 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1660 return -EINVAL;
1661
1662 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1663 return -EPERM;
1664
1665 /* Find the mm_struct */
a879bf58 1666 rcu_read_lock();
228ebcbe 1667 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1668 if (!task) {
a879bf58 1669 rcu_read_unlock();
742755a1
CL
1670 return -ESRCH;
1671 }
3268c63e 1672 get_task_struct(task);
742755a1
CL
1673
1674 /*
1675 * Check if this process has the right to modify the specified
1676 * process. The right exists if the process has administrative
1677 * capabilities, superuser privileges or the same
1678 * userid as the target process.
1679 */
c69e8d9c 1680 tcred = __task_cred(task);
b38a86eb
EB
1681 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1682 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
742755a1 1683 !capable(CAP_SYS_NICE)) {
c69e8d9c 1684 rcu_read_unlock();
742755a1 1685 err = -EPERM;
5e9a0f02 1686 goto out;
742755a1 1687 }
c69e8d9c 1688 rcu_read_unlock();
742755a1 1689
86c3a764
DQ
1690 err = security_task_movememory(task);
1691 if (err)
5e9a0f02 1692 goto out;
86c3a764 1693
3268c63e
CL
1694 task_nodes = cpuset_mems_allowed(task);
1695 mm = get_task_mm(task);
1696 put_task_struct(task);
1697
6e8b09ea
SL
1698 if (!mm)
1699 return -EINVAL;
1700
1701 if (nodes)
1702 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1703 nodes, status, flags);
1704 else
1705 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1706
742755a1
CL
1707 mmput(mm);
1708 return err;
3268c63e
CL
1709
1710out:
1711 put_task_struct(task);
1712 return err;
742755a1 1713}
742755a1 1714
7039e1db
PZ
1715#ifdef CONFIG_NUMA_BALANCING
1716/*
1717 * Returns true if this is a safe migration target node for misplaced NUMA
1718 * pages. Currently it only checks the watermarks which crude
1719 */
1720static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1721 unsigned long nr_migrate_pages)
7039e1db
PZ
1722{
1723 int z;
599d0c95
MG
1724
1725 if (!pgdat_reclaimable(pgdat))
1726 return false;
1727
7039e1db
PZ
1728 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1729 struct zone *zone = pgdat->node_zones + z;
1730
1731 if (!populated_zone(zone))
1732 continue;
1733
7039e1db
PZ
1734 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1735 if (!zone_watermark_ok(zone, 0,
1736 high_wmark_pages(zone) +
1737 nr_migrate_pages,
1738 0, 0))
1739 continue;
1740 return true;
1741 }
1742 return false;
1743}
1744
1745static struct page *alloc_misplaced_dst_page(struct page *page,
1746 unsigned long data,
1747 int **result)
1748{
1749 int nid = (int) data;
1750 struct page *newpage;
1751
96db800f 1752 newpage = __alloc_pages_node(nid,
e97ca8e5
JW
1753 (GFP_HIGHUSER_MOVABLE |
1754 __GFP_THISNODE | __GFP_NOMEMALLOC |
1755 __GFP_NORETRY | __GFP_NOWARN) &
8479eba7 1756 ~__GFP_RECLAIM, 0);
bac0382c 1757
7039e1db
PZ
1758 return newpage;
1759}
1760
a8f60772
MG
1761/*
1762 * page migration rate limiting control.
1763 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1764 * window of time. Default here says do not migrate more than 1280M per second.
1765 */
1766static unsigned int migrate_interval_millisecs __read_mostly = 100;
1767static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1768
b32967ff 1769/* Returns true if the node is migrate rate-limited after the update */
1c30e017
MG
1770static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1771 unsigned long nr_pages)
7039e1db 1772{
a8f60772
MG
1773 /*
1774 * Rate-limit the amount of data that is being migrated to a node.
1775 * Optimal placement is no good if the memory bus is saturated and
1776 * all the time is being spent migrating!
1777 */
a8f60772 1778 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1c5e9c27 1779 spin_lock(&pgdat->numabalancing_migrate_lock);
a8f60772
MG
1780 pgdat->numabalancing_migrate_nr_pages = 0;
1781 pgdat->numabalancing_migrate_next_window = jiffies +
1782 msecs_to_jiffies(migrate_interval_millisecs);
1c5e9c27 1783 spin_unlock(&pgdat->numabalancing_migrate_lock);
a8f60772 1784 }
af1839d7
MG
1785 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1786 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1787 nr_pages);
1c5e9c27 1788 return true;
af1839d7 1789 }
1c5e9c27
MG
1790
1791 /*
1792 * This is an unlocked non-atomic update so errors are possible.
1793 * The consequences are failing to migrate when we potentiall should
1794 * have which is not severe enough to warrant locking. If it is ever
1795 * a problem, it can be converted to a per-cpu counter.
1796 */
1797 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1798 return false;
b32967ff
MG
1799}
1800
1c30e017 1801static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
b32967ff 1802{
340ef390 1803 int page_lru;
a8f60772 1804
309381fe 1805 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
3abef4e6 1806
7039e1db 1807 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1808 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1809 return 0;
7039e1db 1810
340ef390
HD
1811 if (isolate_lru_page(page))
1812 return 0;
7039e1db 1813
340ef390
HD
1814 /*
1815 * migrate_misplaced_transhuge_page() skips page migration's usual
1816 * check on page_count(), so we must do it here, now that the page
1817 * has been isolated: a GUP pin, or any other pin, prevents migration.
1818 * The expected page count is 3: 1 for page's mapcount and 1 for the
1819 * caller's pin and 1 for the reference taken by isolate_lru_page().
1820 */
1821 if (PageTransHuge(page) && page_count(page) != 3) {
1822 putback_lru_page(page);
1823 return 0;
7039e1db
PZ
1824 }
1825
340ef390 1826 page_lru = page_is_file_cache(page);
599d0c95 1827 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
340ef390
HD
1828 hpage_nr_pages(page));
1829
149c33e1 1830 /*
340ef390
HD
1831 * Isolating the page has taken another reference, so the
1832 * caller's reference can be safely dropped without the page
1833 * disappearing underneath us during migration.
149c33e1
MG
1834 */
1835 put_page(page);
340ef390 1836 return 1;
b32967ff
MG
1837}
1838
de466bd6
MG
1839bool pmd_trans_migrating(pmd_t pmd)
1840{
1841 struct page *page = pmd_page(pmd);
1842 return PageLocked(page);
1843}
1844
b32967ff
MG
1845/*
1846 * Attempt to migrate a misplaced page to the specified destination
1847 * node. Caller is expected to have an elevated reference count on
1848 * the page that will be dropped by this function before returning.
1849 */
1bc115d8
MG
1850int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1851 int node)
b32967ff
MG
1852{
1853 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1854 int isolated;
b32967ff
MG
1855 int nr_remaining;
1856 LIST_HEAD(migratepages);
1857
1858 /*
1bc115d8
MG
1859 * Don't migrate file pages that are mapped in multiple processes
1860 * with execute permissions as they are probably shared libraries.
b32967ff 1861 */
1bc115d8
MG
1862 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1863 (vma->vm_flags & VM_EXEC))
b32967ff 1864 goto out;
b32967ff
MG
1865
1866 /*
1867 * Rate-limit the amount of data that is being migrated to a node.
1868 * Optimal placement is no good if the memory bus is saturated and
1869 * all the time is being spent migrating!
1870 */
340ef390 1871 if (numamigrate_update_ratelimit(pgdat, 1))
b32967ff 1872 goto out;
b32967ff
MG
1873
1874 isolated = numamigrate_isolate_page(pgdat, page);
1875 if (!isolated)
1876 goto out;
1877
1878 list_add(&page->lru, &migratepages);
9c620e2b 1879 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
68711a74
DR
1880 NULL, node, MIGRATE_ASYNC,
1881 MR_NUMA_MISPLACED);
b32967ff 1882 if (nr_remaining) {
59c82b70
JK
1883 if (!list_empty(&migratepages)) {
1884 list_del(&page->lru);
599d0c95 1885 dec_node_page_state(page, NR_ISOLATED_ANON +
59c82b70
JK
1886 page_is_file_cache(page));
1887 putback_lru_page(page);
1888 }
b32967ff
MG
1889 isolated = 0;
1890 } else
1891 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1892 BUG_ON(!list_empty(&migratepages));
7039e1db 1893 return isolated;
340ef390
HD
1894
1895out:
1896 put_page(page);
1897 return 0;
7039e1db 1898}
220018d3 1899#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1900
220018d3 1901#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1902/*
1903 * Migrates a THP to a given target node. page must be locked and is unlocked
1904 * before returning.
1905 */
b32967ff
MG
1906int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1907 struct vm_area_struct *vma,
1908 pmd_t *pmd, pmd_t entry,
1909 unsigned long address,
1910 struct page *page, int node)
1911{
c4088ebd 1912 spinlock_t *ptl;
b32967ff
MG
1913 pg_data_t *pgdat = NODE_DATA(node);
1914 int isolated = 0;
1915 struct page *new_page = NULL;
b32967ff 1916 int page_lru = page_is_file_cache(page);
f714f4f2
MG
1917 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1918 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
2b4847e7 1919 pmd_t orig_entry;
b32967ff 1920
b32967ff
MG
1921 /*
1922 * Rate-limit the amount of data that is being migrated to a node.
1923 * Optimal placement is no good if the memory bus is saturated and
1924 * all the time is being spent migrating!
1925 */
d28d4335 1926 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
b32967ff
MG
1927 goto out_dropref;
1928
1929 new_page = alloc_pages_node(node,
25160354 1930 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
e97ca8e5 1931 HPAGE_PMD_ORDER);
340ef390
HD
1932 if (!new_page)
1933 goto out_fail;
9a982250 1934 prep_transhuge_page(new_page);
340ef390 1935
b32967ff 1936 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 1937 if (!isolated) {
b32967ff 1938 put_page(new_page);
340ef390 1939 goto out_fail;
b32967ff 1940 }
458aa76d
AK
1941 /*
1942 * We are not sure a pending tlb flush here is for a huge page
1943 * mapping or not. Hence use the tlb range variant
1944 */
b0943d61
MG
1945 if (mm_tlb_flush_pending(mm))
1946 flush_tlb_range(vma, mmun_start, mmun_end);
1947
b32967ff 1948 /* Prepare a page as a migration target */
48c935ad 1949 __SetPageLocked(new_page);
fa9949da 1950 __SetPageSwapBacked(new_page);
b32967ff
MG
1951
1952 /* anon mapping, we can simply copy page->mapping to the new page: */
1953 new_page->mapping = page->mapping;
1954 new_page->index = page->index;
1955 migrate_page_copy(new_page, page);
1956 WARN_ON(PageLRU(new_page));
1957
1958 /* Recheck the target PMD */
f714f4f2 1959 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 1960 ptl = pmd_lock(mm, pmd);
2b4847e7
MG
1961 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1962fail_putback:
c4088ebd 1963 spin_unlock(ptl);
f714f4f2 1964 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b32967ff
MG
1965
1966 /* Reverse changes made by migrate_page_copy() */
1967 if (TestClearPageActive(new_page))
1968 SetPageActive(page);
1969 if (TestClearPageUnevictable(new_page))
1970 SetPageUnevictable(page);
b32967ff
MG
1971
1972 unlock_page(new_page);
1973 put_page(new_page); /* Free it */
1974
a54a407f
MG
1975 /* Retake the callers reference and putback on LRU */
1976 get_page(page);
b32967ff 1977 putback_lru_page(page);
599d0c95 1978 mod_node_page_state(page_pgdat(page),
a54a407f 1979 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
eb4489f6
MG
1980
1981 goto out_unlock;
b32967ff
MG
1982 }
1983
2b4847e7 1984 orig_entry = *pmd;
10102459 1985 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2b4847e7 1986 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
b32967ff 1987
2b4847e7
MG
1988 /*
1989 * Clear the old entry under pagetable lock and establish the new PTE.
1990 * Any parallel GUP will either observe the old page blocking on the
1991 * page lock, block on the page table lock or observe the new page.
1992 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1993 * guarantee the copy is visible before the pagetable update.
1994 */
f714f4f2 1995 flush_cache_range(vma, mmun_start, mmun_end);
d281ee61 1996 page_add_anon_rmap(new_page, vma, mmun_start, true);
8809aa2d 1997 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
f714f4f2 1998 set_pmd_at(mm, mmun_start, pmd, entry);
ce4a9cc5 1999 update_mmu_cache_pmd(vma, address, &entry);
2b4847e7
MG
2000
2001 if (page_count(page) != 2) {
f714f4f2 2002 set_pmd_at(mm, mmun_start, pmd, orig_entry);
458aa76d 2003 flush_pmd_tlb_range(vma, mmun_start, mmun_end);
34ee645e 2004 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2b4847e7 2005 update_mmu_cache_pmd(vma, address, &entry);
d281ee61 2006 page_remove_rmap(new_page, true);
2b4847e7
MG
2007 goto fail_putback;
2008 }
2009
51afb12b 2010 mlock_migrate_page(new_page, page);
d281ee61 2011 page_remove_rmap(page, true);
7cd12b4a 2012 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2b4847e7 2013
c4088ebd 2014 spin_unlock(ptl);
f714f4f2 2015 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b32967ff 2016
11de9927
MG
2017 /* Take an "isolate" reference and put new page on the LRU. */
2018 get_page(new_page);
2019 putback_lru_page(new_page);
2020
b32967ff
MG
2021 unlock_page(new_page);
2022 unlock_page(page);
2023 put_page(page); /* Drop the rmap reference */
2024 put_page(page); /* Drop the LRU isolation reference */
2025
2026 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2027 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2028
599d0c95 2029 mod_node_page_state(page_pgdat(page),
b32967ff
MG
2030 NR_ISOLATED_ANON + page_lru,
2031 -HPAGE_PMD_NR);
2032 return isolated;
2033
340ef390
HD
2034out_fail:
2035 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
b32967ff 2036out_dropref:
2b4847e7
MG
2037 ptl = pmd_lock(mm, pmd);
2038 if (pmd_same(*pmd, entry)) {
4d942466 2039 entry = pmd_modify(entry, vma->vm_page_prot);
f714f4f2 2040 set_pmd_at(mm, mmun_start, pmd, entry);
2b4847e7
MG
2041 update_mmu_cache_pmd(vma, address, &entry);
2042 }
2043 spin_unlock(ptl);
a54a407f 2044
eb4489f6 2045out_unlock:
340ef390 2046 unlock_page(page);
b32967ff 2047 put_page(page);
b32967ff
MG
2048 return 0;
2049}
7039e1db
PZ
2050#endif /* CONFIG_NUMA_BALANCING */
2051
2052#endif /* CONFIG_NUMA */