]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
mm: migrate: make core migration code aware of hugepage
[mirror_ubuntu-jammy-kernel.git] / mm / migrate.c
CommitLineData
b20a3503
CL
1/*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
cde53535 12 * Christoph Lameter
b20a3503
CL
13 */
14
15#include <linux/migrate.h>
b95f1b31 16#include <linux/export.h>
b20a3503 17#include <linux/swap.h>
0697212a 18#include <linux/swapops.h>
b20a3503 19#include <linux/pagemap.h>
e23ca00b 20#include <linux/buffer_head.h>
b20a3503 21#include <linux/mm_inline.h>
b488893a 22#include <linux/nsproxy.h>
b20a3503 23#include <linux/pagevec.h>
e9995ef9 24#include <linux/ksm.h>
b20a3503
CL
25#include <linux/rmap.h>
26#include <linux/topology.h>
27#include <linux/cpu.h>
28#include <linux/cpuset.h>
04e62a29 29#include <linux/writeback.h>
742755a1
CL
30#include <linux/mempolicy.h>
31#include <linux/vmalloc.h>
86c3a764 32#include <linux/security.h>
8a9f3ccd 33#include <linux/memcontrol.h>
4f5ca265 34#include <linux/syscalls.h>
290408d4 35#include <linux/hugetlb.h>
8e6ac7fa 36#include <linux/hugetlb_cgroup.h>
5a0e3ad6 37#include <linux/gfp.h>
bf6bddf1 38#include <linux/balloon_compaction.h>
b20a3503 39
0d1836c3
MN
40#include <asm/tlbflush.h>
41
7b2a2d4a
MG
42#define CREATE_TRACE_POINTS
43#include <trace/events/migrate.h>
44
b20a3503
CL
45#include "internal.h"
46
b20a3503 47/*
742755a1 48 * migrate_prep() needs to be called before we start compiling a list of pages
748446bb
MG
49 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
50 * undesirable, use migrate_prep_local()
b20a3503
CL
51 */
52int migrate_prep(void)
53{
b20a3503
CL
54 /*
55 * Clear the LRU lists so pages can be isolated.
56 * Note that pages may be moved off the LRU after we have
57 * drained them. Those pages will fail to migrate like other
58 * pages that may be busy.
59 */
60 lru_add_drain_all();
61
62 return 0;
63}
64
748446bb
MG
65/* Do the necessary work of migrate_prep but not if it involves other CPUs */
66int migrate_prep_local(void)
67{
68 lru_add_drain();
69
70 return 0;
71}
72
b20a3503 73/*
894bc310
LS
74 * Add isolated pages on the list back to the LRU under page lock
75 * to avoid leaking evictable pages back onto unevictable list.
b20a3503 76 */
e13861d8 77void putback_lru_pages(struct list_head *l)
b20a3503
CL
78{
79 struct page *page;
80 struct page *page2;
b20a3503 81
5733c7d1
RA
82 list_for_each_entry_safe(page, page2, l, lru) {
83 list_del(&page->lru);
84 dec_zone_page_state(page, NR_ISOLATED_ANON +
85 page_is_file_cache(page));
86 putback_lru_page(page);
87 }
88}
89
90/*
91 * Put previously isolated pages back onto the appropriate lists
92 * from where they were once taken off for compaction/migration.
93 *
94 * This function shall be used instead of putback_lru_pages(),
95 * whenever the isolated pageset has been built by isolate_migratepages_range()
96 */
97void putback_movable_pages(struct list_head *l)
98{
99 struct page *page;
100 struct page *page2;
101
b20a3503 102 list_for_each_entry_safe(page, page2, l, lru) {
31caf665
NH
103 if (unlikely(PageHuge(page))) {
104 putback_active_hugepage(page);
105 continue;
106 }
e24f0b8f 107 list_del(&page->lru);
a731286d 108 dec_zone_page_state(page, NR_ISOLATED_ANON +
6c0b1351 109 page_is_file_cache(page));
bf6bddf1
RA
110 if (unlikely(balloon_page_movable(page)))
111 balloon_page_putback(page);
112 else
113 putback_lru_page(page);
b20a3503 114 }
b20a3503
CL
115}
116
0697212a
CL
117/*
118 * Restore a potential migration pte to a working pte entry
119 */
e9995ef9
HD
120static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
121 unsigned long addr, void *old)
0697212a
CL
122{
123 struct mm_struct *mm = vma->vm_mm;
124 swp_entry_t entry;
0697212a
CL
125 pmd_t *pmd;
126 pte_t *ptep, pte;
127 spinlock_t *ptl;
128
290408d4
NH
129 if (unlikely(PageHuge(new))) {
130 ptep = huge_pte_offset(mm, addr);
131 if (!ptep)
132 goto out;
133 ptl = &mm->page_table_lock;
134 } else {
6219049a
BL
135 pmd = mm_find_pmd(mm, addr);
136 if (!pmd)
290408d4 137 goto out;
500d65d4
AA
138 if (pmd_trans_huge(*pmd))
139 goto out;
0697212a 140
290408d4 141 ptep = pte_offset_map(pmd, addr);
0697212a 142
486cf46f
HD
143 /*
144 * Peek to check is_swap_pte() before taking ptlock? No, we
145 * can race mremap's move_ptes(), which skips anon_vma lock.
146 */
290408d4
NH
147
148 ptl = pte_lockptr(mm, pmd);
149 }
0697212a 150
0697212a
CL
151 spin_lock(ptl);
152 pte = *ptep;
153 if (!is_swap_pte(pte))
e9995ef9 154 goto unlock;
0697212a
CL
155
156 entry = pte_to_swp_entry(pte);
157
e9995ef9
HD
158 if (!is_migration_entry(entry) ||
159 migration_entry_to_page(entry) != old)
160 goto unlock;
0697212a 161
0697212a
CL
162 get_page(new);
163 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
164 if (is_write_migration_entry(entry))
165 pte = pte_mkwrite(pte);
3ef8fd7f 166#ifdef CONFIG_HUGETLB_PAGE
be7517d6 167 if (PageHuge(new)) {
290408d4 168 pte = pte_mkhuge(pte);
be7517d6
TL
169 pte = arch_make_huge_pte(pte, vma, new, 0);
170 }
3ef8fd7f 171#endif
c2cc499c 172 flush_dcache_page(new);
0697212a 173 set_pte_at(mm, addr, ptep, pte);
04e62a29 174
290408d4
NH
175 if (PageHuge(new)) {
176 if (PageAnon(new))
177 hugepage_add_anon_rmap(new, vma, addr);
178 else
179 page_dup_rmap(new);
180 } else if (PageAnon(new))
04e62a29
CL
181 page_add_anon_rmap(new, vma, addr);
182 else
183 page_add_file_rmap(new);
184
185 /* No need to invalidate - it was non-present before */
4b3073e1 186 update_mmu_cache(vma, addr, ptep);
e9995ef9 187unlock:
0697212a 188 pte_unmap_unlock(ptep, ptl);
e9995ef9
HD
189out:
190 return SWAP_AGAIN;
0697212a
CL
191}
192
04e62a29
CL
193/*
194 * Get rid of all migration entries and replace them by
195 * references to the indicated page.
196 */
197static void remove_migration_ptes(struct page *old, struct page *new)
198{
e9995ef9 199 rmap_walk(new, remove_migration_pte, old);
04e62a29
CL
200}
201
0697212a
CL
202/*
203 * Something used the pte of a page under migration. We need to
204 * get to the page and wait until migration is finished.
205 * When we return from this function the fault will be retried.
0697212a 206 */
30dad309
NH
207static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
208 spinlock_t *ptl)
0697212a 209{
30dad309 210 pte_t pte;
0697212a
CL
211 swp_entry_t entry;
212 struct page *page;
213
30dad309 214 spin_lock(ptl);
0697212a
CL
215 pte = *ptep;
216 if (!is_swap_pte(pte))
217 goto out;
218
219 entry = pte_to_swp_entry(pte);
220 if (!is_migration_entry(entry))
221 goto out;
222
223 page = migration_entry_to_page(entry);
224
e286781d
NP
225 /*
226 * Once radix-tree replacement of page migration started, page_count
227 * *must* be zero. And, we don't want to call wait_on_page_locked()
228 * against a page without get_page().
229 * So, we use get_page_unless_zero(), here. Even failed, page fault
230 * will occur again.
231 */
232 if (!get_page_unless_zero(page))
233 goto out;
0697212a
CL
234 pte_unmap_unlock(ptep, ptl);
235 wait_on_page_locked(page);
236 put_page(page);
237 return;
238out:
239 pte_unmap_unlock(ptep, ptl);
240}
241
30dad309
NH
242void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
243 unsigned long address)
244{
245 spinlock_t *ptl = pte_lockptr(mm, pmd);
246 pte_t *ptep = pte_offset_map(pmd, address);
247 __migration_entry_wait(mm, ptep, ptl);
248}
249
250void migration_entry_wait_huge(struct mm_struct *mm, pte_t *pte)
251{
252 spinlock_t *ptl = &(mm)->page_table_lock;
253 __migration_entry_wait(mm, pte, ptl);
254}
255
b969c4ab
MG
256#ifdef CONFIG_BLOCK
257/* Returns true if all buffers are successfully locked */
a6bc32b8
MG
258static bool buffer_migrate_lock_buffers(struct buffer_head *head,
259 enum migrate_mode mode)
b969c4ab
MG
260{
261 struct buffer_head *bh = head;
262
263 /* Simple case, sync compaction */
a6bc32b8 264 if (mode != MIGRATE_ASYNC) {
b969c4ab
MG
265 do {
266 get_bh(bh);
267 lock_buffer(bh);
268 bh = bh->b_this_page;
269
270 } while (bh != head);
271
272 return true;
273 }
274
275 /* async case, we cannot block on lock_buffer so use trylock_buffer */
276 do {
277 get_bh(bh);
278 if (!trylock_buffer(bh)) {
279 /*
280 * We failed to lock the buffer and cannot stall in
281 * async migration. Release the taken locks
282 */
283 struct buffer_head *failed_bh = bh;
284 put_bh(failed_bh);
285 bh = head;
286 while (bh != failed_bh) {
287 unlock_buffer(bh);
288 put_bh(bh);
289 bh = bh->b_this_page;
290 }
291 return false;
292 }
293
294 bh = bh->b_this_page;
295 } while (bh != head);
296 return true;
297}
298#else
299static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
a6bc32b8 300 enum migrate_mode mode)
b969c4ab
MG
301{
302 return true;
303}
304#endif /* CONFIG_BLOCK */
305
b20a3503 306/*
c3fcf8a5 307 * Replace the page in the mapping.
5b5c7120
CL
308 *
309 * The number of remaining references must be:
310 * 1 for anonymous pages without a mapping
311 * 2 for pages with a mapping
266cf658 312 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
b20a3503 313 */
2d1db3b1 314static int migrate_page_move_mapping(struct address_space *mapping,
b969c4ab 315 struct page *newpage, struct page *page,
a6bc32b8 316 struct buffer_head *head, enum migrate_mode mode)
b20a3503 317{
7039e1db 318 int expected_count = 0;
7cf9c2c7 319 void **pslot;
b20a3503 320
6c5240ae 321 if (!mapping) {
0e8c7d0f 322 /* Anonymous page without mapping */
6c5240ae
CL
323 if (page_count(page) != 1)
324 return -EAGAIN;
78bd5209 325 return MIGRATEPAGE_SUCCESS;
6c5240ae
CL
326 }
327
19fd6231 328 spin_lock_irq(&mapping->tree_lock);
b20a3503 329
7cf9c2c7
NP
330 pslot = radix_tree_lookup_slot(&mapping->page_tree,
331 page_index(page));
b20a3503 332
edcf4748 333 expected_count = 2 + page_has_private(page);
e286781d 334 if (page_count(page) != expected_count ||
29c1f677 335 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
19fd6231 336 spin_unlock_irq(&mapping->tree_lock);
e23ca00b 337 return -EAGAIN;
b20a3503
CL
338 }
339
e286781d 340 if (!page_freeze_refs(page, expected_count)) {
19fd6231 341 spin_unlock_irq(&mapping->tree_lock);
e286781d
NP
342 return -EAGAIN;
343 }
344
b969c4ab
MG
345 /*
346 * In the async migration case of moving a page with buffers, lock the
347 * buffers using trylock before the mapping is moved. If the mapping
348 * was moved, we later failed to lock the buffers and could not move
349 * the mapping back due to an elevated page count, we would have to
350 * block waiting on other references to be dropped.
351 */
a6bc32b8
MG
352 if (mode == MIGRATE_ASYNC && head &&
353 !buffer_migrate_lock_buffers(head, mode)) {
b969c4ab
MG
354 page_unfreeze_refs(page, expected_count);
355 spin_unlock_irq(&mapping->tree_lock);
356 return -EAGAIN;
357 }
358
b20a3503
CL
359 /*
360 * Now we know that no one else is looking at the page.
b20a3503 361 */
7cf9c2c7 362 get_page(newpage); /* add cache reference */
b20a3503
CL
363 if (PageSwapCache(page)) {
364 SetPageSwapCache(newpage);
365 set_page_private(newpage, page_private(page));
366 }
367
7cf9c2c7
NP
368 radix_tree_replace_slot(pslot, newpage);
369
370 /*
937a94c9
JG
371 * Drop cache reference from old page by unfreezing
372 * to one less reference.
7cf9c2c7
NP
373 * We know this isn't the last reference.
374 */
937a94c9 375 page_unfreeze_refs(page, expected_count - 1);
7cf9c2c7 376
0e8c7d0f
CL
377 /*
378 * If moved to a different zone then also account
379 * the page for that zone. Other VM counters will be
380 * taken care of when we establish references to the
381 * new page and drop references to the old page.
382 *
383 * Note that anonymous pages are accounted for
384 * via NR_FILE_PAGES and NR_ANON_PAGES if they
385 * are mapped to swap space.
386 */
387 __dec_zone_page_state(page, NR_FILE_PAGES);
388 __inc_zone_page_state(newpage, NR_FILE_PAGES);
99a15e21 389 if (!PageSwapCache(page) && PageSwapBacked(page)) {
4b02108a
KM
390 __dec_zone_page_state(page, NR_SHMEM);
391 __inc_zone_page_state(newpage, NR_SHMEM);
392 }
19fd6231 393 spin_unlock_irq(&mapping->tree_lock);
b20a3503 394
78bd5209 395 return MIGRATEPAGE_SUCCESS;
b20a3503 396}
b20a3503 397
290408d4
NH
398/*
399 * The expected number of remaining references is the same as that
400 * of migrate_page_move_mapping().
401 */
402int migrate_huge_page_move_mapping(struct address_space *mapping,
403 struct page *newpage, struct page *page)
404{
405 int expected_count;
406 void **pslot;
407
408 if (!mapping) {
409 if (page_count(page) != 1)
410 return -EAGAIN;
78bd5209 411 return MIGRATEPAGE_SUCCESS;
290408d4
NH
412 }
413
414 spin_lock_irq(&mapping->tree_lock);
415
416 pslot = radix_tree_lookup_slot(&mapping->page_tree,
417 page_index(page));
418
419 expected_count = 2 + page_has_private(page);
420 if (page_count(page) != expected_count ||
29c1f677 421 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
290408d4
NH
422 spin_unlock_irq(&mapping->tree_lock);
423 return -EAGAIN;
424 }
425
426 if (!page_freeze_refs(page, expected_count)) {
427 spin_unlock_irq(&mapping->tree_lock);
428 return -EAGAIN;
429 }
430
431 get_page(newpage);
432
433 radix_tree_replace_slot(pslot, newpage);
434
937a94c9 435 page_unfreeze_refs(page, expected_count - 1);
290408d4
NH
436
437 spin_unlock_irq(&mapping->tree_lock);
78bd5209 438 return MIGRATEPAGE_SUCCESS;
290408d4
NH
439}
440
b20a3503
CL
441/*
442 * Copy the page to its new location
443 */
290408d4 444void migrate_page_copy(struct page *newpage, struct page *page)
b20a3503 445{
b32967ff 446 if (PageHuge(page) || PageTransHuge(page))
290408d4
NH
447 copy_huge_page(newpage, page);
448 else
449 copy_highpage(newpage, page);
b20a3503
CL
450
451 if (PageError(page))
452 SetPageError(newpage);
453 if (PageReferenced(page))
454 SetPageReferenced(newpage);
455 if (PageUptodate(page))
456 SetPageUptodate(newpage);
894bc310
LS
457 if (TestClearPageActive(page)) {
458 VM_BUG_ON(PageUnevictable(page));
b20a3503 459 SetPageActive(newpage);
418b27ef
LS
460 } else if (TestClearPageUnevictable(page))
461 SetPageUnevictable(newpage);
b20a3503
CL
462 if (PageChecked(page))
463 SetPageChecked(newpage);
464 if (PageMappedToDisk(page))
465 SetPageMappedToDisk(newpage);
466
467 if (PageDirty(page)) {
468 clear_page_dirty_for_io(page);
3a902c5f
NP
469 /*
470 * Want to mark the page and the radix tree as dirty, and
471 * redo the accounting that clear_page_dirty_for_io undid,
472 * but we can't use set_page_dirty because that function
473 * is actually a signal that all of the page has become dirty.
25985edc 474 * Whereas only part of our page may be dirty.
3a902c5f 475 */
752dc185
HD
476 if (PageSwapBacked(page))
477 SetPageDirty(newpage);
478 else
479 __set_page_dirty_nobuffers(newpage);
b20a3503
CL
480 }
481
b291f000 482 mlock_migrate_page(newpage, page);
e9995ef9 483 ksm_migrate_page(newpage, page);
c8d6553b
HD
484 /*
485 * Please do not reorder this without considering how mm/ksm.c's
486 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
487 */
b20a3503 488 ClearPageSwapCache(page);
b20a3503
CL
489 ClearPagePrivate(page);
490 set_page_private(page, 0);
b20a3503
CL
491
492 /*
493 * If any waiters have accumulated on the new page then
494 * wake them up.
495 */
496 if (PageWriteback(newpage))
497 end_page_writeback(newpage);
498}
b20a3503 499
1d8b85cc
CL
500/************************************************************
501 * Migration functions
502 ***********************************************************/
503
504/* Always fail migration. Used for mappings that are not movable */
2d1db3b1
CL
505int fail_migrate_page(struct address_space *mapping,
506 struct page *newpage, struct page *page)
1d8b85cc
CL
507{
508 return -EIO;
509}
510EXPORT_SYMBOL(fail_migrate_page);
511
b20a3503
CL
512/*
513 * Common logic to directly migrate a single page suitable for
266cf658 514 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
515 *
516 * Pages are locked upon entry and exit.
517 */
2d1db3b1 518int migrate_page(struct address_space *mapping,
a6bc32b8
MG
519 struct page *newpage, struct page *page,
520 enum migrate_mode mode)
b20a3503
CL
521{
522 int rc;
523
524 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
525
a6bc32b8 526 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
b20a3503 527
78bd5209 528 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
529 return rc;
530
531 migrate_page_copy(newpage, page);
78bd5209 532 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
533}
534EXPORT_SYMBOL(migrate_page);
535
9361401e 536#ifdef CONFIG_BLOCK
1d8b85cc
CL
537/*
538 * Migration function for pages with buffers. This function can only be used
539 * if the underlying filesystem guarantees that no other references to "page"
540 * exist.
541 */
2d1db3b1 542int buffer_migrate_page(struct address_space *mapping,
a6bc32b8 543 struct page *newpage, struct page *page, enum migrate_mode mode)
1d8b85cc 544{
1d8b85cc
CL
545 struct buffer_head *bh, *head;
546 int rc;
547
1d8b85cc 548 if (!page_has_buffers(page))
a6bc32b8 549 return migrate_page(mapping, newpage, page, mode);
1d8b85cc
CL
550
551 head = page_buffers(page);
552
a6bc32b8 553 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
1d8b85cc 554
78bd5209 555 if (rc != MIGRATEPAGE_SUCCESS)
1d8b85cc
CL
556 return rc;
557
b969c4ab
MG
558 /*
559 * In the async case, migrate_page_move_mapping locked the buffers
560 * with an IRQ-safe spinlock held. In the sync case, the buffers
561 * need to be locked now
562 */
a6bc32b8
MG
563 if (mode != MIGRATE_ASYNC)
564 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
1d8b85cc
CL
565
566 ClearPagePrivate(page);
567 set_page_private(newpage, page_private(page));
568 set_page_private(page, 0);
569 put_page(page);
570 get_page(newpage);
571
572 bh = head;
573 do {
574 set_bh_page(bh, newpage, bh_offset(bh));
575 bh = bh->b_this_page;
576
577 } while (bh != head);
578
579 SetPagePrivate(newpage);
580
581 migrate_page_copy(newpage, page);
582
583 bh = head;
584 do {
585 unlock_buffer(bh);
586 put_bh(bh);
587 bh = bh->b_this_page;
588
589 } while (bh != head);
590
78bd5209 591 return MIGRATEPAGE_SUCCESS;
1d8b85cc
CL
592}
593EXPORT_SYMBOL(buffer_migrate_page);
9361401e 594#endif
1d8b85cc 595
04e62a29
CL
596/*
597 * Writeback a page to clean the dirty state
598 */
599static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 600{
04e62a29
CL
601 struct writeback_control wbc = {
602 .sync_mode = WB_SYNC_NONE,
603 .nr_to_write = 1,
604 .range_start = 0,
605 .range_end = LLONG_MAX,
04e62a29
CL
606 .for_reclaim = 1
607 };
608 int rc;
609
610 if (!mapping->a_ops->writepage)
611 /* No write method for the address space */
612 return -EINVAL;
613
614 if (!clear_page_dirty_for_io(page))
615 /* Someone else already triggered a write */
616 return -EAGAIN;
617
8351a6e4 618 /*
04e62a29
CL
619 * A dirty page may imply that the underlying filesystem has
620 * the page on some queue. So the page must be clean for
621 * migration. Writeout may mean we loose the lock and the
622 * page state is no longer what we checked for earlier.
623 * At this point we know that the migration attempt cannot
624 * be successful.
8351a6e4 625 */
04e62a29 626 remove_migration_ptes(page, page);
8351a6e4 627
04e62a29 628 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 629
04e62a29
CL
630 if (rc != AOP_WRITEPAGE_ACTIVATE)
631 /* unlocked. Relock */
632 lock_page(page);
633
bda8550d 634 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
635}
636
637/*
638 * Default handling if a filesystem does not provide a migration function.
639 */
640static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 641 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 642{
b969c4ab 643 if (PageDirty(page)) {
a6bc32b8
MG
644 /* Only writeback pages in full synchronous migration */
645 if (mode != MIGRATE_SYNC)
b969c4ab 646 return -EBUSY;
04e62a29 647 return writeout(mapping, page);
b969c4ab 648 }
8351a6e4
CL
649
650 /*
651 * Buffers may be managed in a filesystem specific way.
652 * We must have no buffers or drop them.
653 */
266cf658 654 if (page_has_private(page) &&
8351a6e4
CL
655 !try_to_release_page(page, GFP_KERNEL))
656 return -EAGAIN;
657
a6bc32b8 658 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
659}
660
e24f0b8f
CL
661/*
662 * Move a page to a newly allocated page
663 * The page is locked and all ptes have been successfully removed.
664 *
665 * The new page will have replaced the old page if this function
666 * is successful.
894bc310
LS
667 *
668 * Return value:
669 * < 0 - error code
78bd5209 670 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 671 */
3fe2011f 672static int move_to_new_page(struct page *newpage, struct page *page,
a6bc32b8 673 int remap_swapcache, enum migrate_mode mode)
e24f0b8f
CL
674{
675 struct address_space *mapping;
676 int rc;
677
678 /*
679 * Block others from accessing the page when we get around to
680 * establishing additional references. We are the only one
681 * holding a reference to the new page at this point.
682 */
529ae9aa 683 if (!trylock_page(newpage))
e24f0b8f
CL
684 BUG();
685
686 /* Prepare mapping for the new page.*/
687 newpage->index = page->index;
688 newpage->mapping = page->mapping;
b2e18538
RR
689 if (PageSwapBacked(page))
690 SetPageSwapBacked(newpage);
e24f0b8f
CL
691
692 mapping = page_mapping(page);
693 if (!mapping)
a6bc32b8 694 rc = migrate_page(mapping, newpage, page, mode);
b969c4ab 695 else if (mapping->a_ops->migratepage)
e24f0b8f 696 /*
b969c4ab
MG
697 * Most pages have a mapping and most filesystems provide a
698 * migratepage callback. Anonymous pages are part of swap
699 * space which also has its own migratepage callback. This
700 * is the most common path for page migration.
e24f0b8f 701 */
b969c4ab 702 rc = mapping->a_ops->migratepage(mapping,
a6bc32b8 703 newpage, page, mode);
b969c4ab 704 else
a6bc32b8 705 rc = fallback_migrate_page(mapping, newpage, page, mode);
e24f0b8f 706
78bd5209 707 if (rc != MIGRATEPAGE_SUCCESS) {
e24f0b8f 708 newpage->mapping = NULL;
3fe2011f
MG
709 } else {
710 if (remap_swapcache)
711 remove_migration_ptes(page, newpage);
35512eca 712 page->mapping = NULL;
3fe2011f 713 }
e24f0b8f
CL
714
715 unlock_page(newpage);
716
717 return rc;
718}
719
0dabec93 720static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 721 int force, enum migrate_mode mode)
e24f0b8f 722{
0dabec93 723 int rc = -EAGAIN;
3fe2011f 724 int remap_swapcache = 1;
56039efa 725 struct mem_cgroup *mem;
3f6c8272 726 struct anon_vma *anon_vma = NULL;
95a402c3 727
529ae9aa 728 if (!trylock_page(page)) {
a6bc32b8 729 if (!force || mode == MIGRATE_ASYNC)
0dabec93 730 goto out;
3e7d3449
MG
731
732 /*
733 * It's not safe for direct compaction to call lock_page.
734 * For example, during page readahead pages are added locked
735 * to the LRU. Later, when the IO completes the pages are
736 * marked uptodate and unlocked. However, the queueing
737 * could be merging multiple pages for one bio (e.g.
738 * mpage_readpages). If an allocation happens for the
739 * second or third page, the process can end up locking
740 * the same page twice and deadlocking. Rather than
741 * trying to be clever about what pages can be locked,
742 * avoid the use of lock_page for direct compaction
743 * altogether.
744 */
745 if (current->flags & PF_MEMALLOC)
0dabec93 746 goto out;
3e7d3449 747
e24f0b8f
CL
748 lock_page(page);
749 }
750
01b1ae63 751 /* charge against new page */
0030f535 752 mem_cgroup_prepare_migration(page, newpage, &mem);
01b1ae63 753
e24f0b8f 754 if (PageWriteback(page)) {
11bc82d6 755 /*
fed5b64a 756 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
757 * necessary to wait for PageWriteback. In the async case,
758 * the retry loop is too short and in the sync-light case,
759 * the overhead of stalling is too much
11bc82d6 760 */
a6bc32b8 761 if (mode != MIGRATE_SYNC) {
11bc82d6
AA
762 rc = -EBUSY;
763 goto uncharge;
764 }
765 if (!force)
01b1ae63 766 goto uncharge;
e24f0b8f
CL
767 wait_on_page_writeback(page);
768 }
e24f0b8f 769 /*
dc386d4d
KH
770 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
771 * we cannot notice that anon_vma is freed while we migrates a page.
1ce82b69 772 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 773 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
774 * File Caches may use write_page() or lock_page() in migration, then,
775 * just care Anon page here.
dc386d4d 776 */
b79bc0a0 777 if (PageAnon(page) && !PageKsm(page)) {
1ce82b69 778 /*
4fc3f1d6 779 * Only page_lock_anon_vma_read() understands the subtleties of
1ce82b69
HD
780 * getting a hold on an anon_vma from outside one of its mms.
781 */
746b18d4 782 anon_vma = page_get_anon_vma(page);
1ce82b69
HD
783 if (anon_vma) {
784 /*
746b18d4 785 * Anon page
1ce82b69 786 */
1ce82b69 787 } else if (PageSwapCache(page)) {
3fe2011f
MG
788 /*
789 * We cannot be sure that the anon_vma of an unmapped
790 * swapcache page is safe to use because we don't
791 * know in advance if the VMA that this page belonged
792 * to still exists. If the VMA and others sharing the
793 * data have been freed, then the anon_vma could
794 * already be invalid.
795 *
796 * To avoid this possibility, swapcache pages get
797 * migrated but are not remapped when migration
798 * completes
799 */
800 remap_swapcache = 0;
801 } else {
1ce82b69 802 goto uncharge;
3fe2011f 803 }
989f89c5 804 }
62e1c553 805
bf6bddf1
RA
806 if (unlikely(balloon_page_movable(page))) {
807 /*
808 * A ballooned page does not need any special attention from
809 * physical to virtual reverse mapping procedures.
810 * Skip any attempt to unmap PTEs or to remap swap cache,
811 * in order to avoid burning cycles at rmap level, and perform
812 * the page migration right away (proteced by page lock).
813 */
814 rc = balloon_page_migrate(newpage, page, mode);
815 goto uncharge;
816 }
817
dc386d4d 818 /*
62e1c553
SL
819 * Corner case handling:
820 * 1. When a new swap-cache page is read into, it is added to the LRU
821 * and treated as swapcache but it has no rmap yet.
822 * Calling try_to_unmap() against a page->mapping==NULL page will
823 * trigger a BUG. So handle it here.
824 * 2. An orphaned page (see truncate_complete_page) might have
825 * fs-private metadata. The page can be picked up due to memory
826 * offlining. Everywhere else except page reclaim, the page is
827 * invisible to the vm, so the page can not be migrated. So try to
828 * free the metadata, so the page can be freed.
e24f0b8f 829 */
62e1c553 830 if (!page->mapping) {
1ce82b69
HD
831 VM_BUG_ON(PageAnon(page));
832 if (page_has_private(page)) {
62e1c553 833 try_to_free_buffers(page);
1ce82b69 834 goto uncharge;
62e1c553 835 }
abfc3488 836 goto skip_unmap;
62e1c553
SL
837 }
838
dc386d4d 839 /* Establish migration ptes or remove ptes */
14fa31b8 840 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
dc386d4d 841
abfc3488 842skip_unmap:
e6a1530d 843 if (!page_mapped(page))
a6bc32b8 844 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
e24f0b8f 845
3fe2011f 846 if (rc && remap_swapcache)
e24f0b8f 847 remove_migration_ptes(page, page);
3f6c8272
MG
848
849 /* Drop an anon_vma reference if we took one */
76545066 850 if (anon_vma)
9e60109f 851 put_anon_vma(anon_vma);
3f6c8272 852
01b1ae63 853uncharge:
bf6bddf1
RA
854 mem_cgroup_end_migration(mem, page, newpage,
855 (rc == MIGRATEPAGE_SUCCESS ||
856 rc == MIGRATEPAGE_BALLOON_SUCCESS));
e24f0b8f 857 unlock_page(page);
0dabec93
MK
858out:
859 return rc;
860}
95a402c3 861
0dabec93
MK
862/*
863 * Obtain the lock on page, remove all ptes and migrate the page
864 * to the newly allocated page in newpage.
865 */
866static int unmap_and_move(new_page_t get_new_page, unsigned long private,
9c620e2b 867 struct page *page, int force, enum migrate_mode mode)
0dabec93
MK
868{
869 int rc = 0;
870 int *result = NULL;
871 struct page *newpage = get_new_page(page, private, &result);
872
873 if (!newpage)
874 return -ENOMEM;
875
876 if (page_count(page) == 1) {
877 /* page was freed from under us. So we are done. */
878 goto out;
879 }
880
881 if (unlikely(PageTransHuge(page)))
882 if (unlikely(split_huge_page(page)))
883 goto out;
884
9c620e2b 885 rc = __unmap_and_move(page, newpage, force, mode);
bf6bddf1
RA
886
887 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
888 /*
889 * A ballooned page has been migrated already.
890 * Now, it's the time to wrap-up counters,
891 * handle the page back to Buddy and return.
892 */
893 dec_zone_page_state(page, NR_ISOLATED_ANON +
894 page_is_file_cache(page));
895 balloon_page_free(page);
896 return MIGRATEPAGE_SUCCESS;
897 }
0dabec93 898out:
e24f0b8f 899 if (rc != -EAGAIN) {
0dabec93
MK
900 /*
901 * A page that has been migrated has all references
902 * removed and will be freed. A page that has not been
903 * migrated will have kepts its references and be
904 * restored.
905 */
906 list_del(&page->lru);
a731286d 907 dec_zone_page_state(page, NR_ISOLATED_ANON +
6c0b1351 908 page_is_file_cache(page));
894bc310 909 putback_lru_page(page);
e24f0b8f 910 }
95a402c3
CL
911 /*
912 * Move the new page to the LRU. If migration was not successful
913 * then this will free the page.
914 */
894bc310 915 putback_lru_page(newpage);
742755a1
CL
916 if (result) {
917 if (rc)
918 *result = rc;
919 else
920 *result = page_to_nid(newpage);
921 }
e24f0b8f
CL
922 return rc;
923}
924
290408d4
NH
925/*
926 * Counterpart of unmap_and_move_page() for hugepage migration.
927 *
928 * This function doesn't wait the completion of hugepage I/O
929 * because there is no race between I/O and migration for hugepage.
930 * Note that currently hugepage I/O occurs only in direct I/O
931 * where no lock is held and PG_writeback is irrelevant,
932 * and writeback status of all subpages are counted in the reference
933 * count of the head page (i.e. if all subpages of a 2MB hugepage are
934 * under direct I/O, the reference of the head page is 512 and a bit more.)
935 * This means that when we try to migrate hugepage whose subpages are
936 * doing direct I/O, some references remain after try_to_unmap() and
937 * hugepage migration fails without data corruption.
938 *
939 * There is also no race when direct I/O is issued on the page under migration,
940 * because then pte is replaced with migration swap entry and direct I/O code
941 * will wait in the page fault for migration to complete.
942 */
943static int unmap_and_move_huge_page(new_page_t get_new_page,
944 unsigned long private, struct page *hpage,
9c620e2b 945 int force, enum migrate_mode mode)
290408d4
NH
946{
947 int rc = 0;
948 int *result = NULL;
949 struct page *new_hpage = get_new_page(hpage, private, &result);
290408d4
NH
950 struct anon_vma *anon_vma = NULL;
951
952 if (!new_hpage)
953 return -ENOMEM;
954
955 rc = -EAGAIN;
956
957 if (!trylock_page(hpage)) {
a6bc32b8 958 if (!force || mode != MIGRATE_SYNC)
290408d4
NH
959 goto out;
960 lock_page(hpage);
961 }
962
746b18d4
PZ
963 if (PageAnon(hpage))
964 anon_vma = page_get_anon_vma(hpage);
290408d4
NH
965
966 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
967
968 if (!page_mapped(hpage))
a6bc32b8 969 rc = move_to_new_page(new_hpage, hpage, 1, mode);
290408d4
NH
970
971 if (rc)
972 remove_migration_ptes(hpage, hpage);
973
fd4a4663 974 if (anon_vma)
9e60109f 975 put_anon_vma(anon_vma);
8e6ac7fa
AK
976
977 if (!rc)
978 hugetlb_cgroup_migrate(hpage, new_hpage);
979
290408d4 980 unlock_page(hpage);
09761333 981out:
290408d4 982 put_page(new_hpage);
290408d4
NH
983 if (result) {
984 if (rc)
985 *result = rc;
986 else
987 *result = page_to_nid(new_hpage);
988 }
989 return rc;
990}
991
b20a3503 992/*
c73e5c9c
SB
993 * migrate_pages - migrate the pages specified in a list, to the free pages
994 * supplied as the target for the page migration
b20a3503 995 *
c73e5c9c
SB
996 * @from: The list of pages to be migrated.
997 * @get_new_page: The function used to allocate free pages to be used
998 * as the target of the page migration.
999 * @private: Private data to be passed on to get_new_page()
1000 * @mode: The migration mode that specifies the constraints for
1001 * page migration, if any.
1002 * @reason: The reason for page migration.
b20a3503 1003 *
c73e5c9c
SB
1004 * The function returns after 10 attempts or if no pages are movable any more
1005 * because the list has become empty or no retryable pages exist any more.
1006 * The caller should call putback_lru_pages() to return pages to the LRU
28bd6578 1007 * or free list only if ret != 0.
b20a3503 1008 *
c73e5c9c 1009 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1010 */
9c620e2b
HD
1011int migrate_pages(struct list_head *from, new_page_t get_new_page,
1012 unsigned long private, enum migrate_mode mode, int reason)
b20a3503 1013{
e24f0b8f 1014 int retry = 1;
b20a3503 1015 int nr_failed = 0;
5647bc29 1016 int nr_succeeded = 0;
b20a3503
CL
1017 int pass = 0;
1018 struct page *page;
1019 struct page *page2;
1020 int swapwrite = current->flags & PF_SWAPWRITE;
1021 int rc;
1022
1023 if (!swapwrite)
1024 current->flags |= PF_SWAPWRITE;
1025
e24f0b8f
CL
1026 for(pass = 0; pass < 10 && retry; pass++) {
1027 retry = 0;
b20a3503 1028
e24f0b8f 1029 list_for_each_entry_safe(page, page2, from, lru) {
e24f0b8f 1030 cond_resched();
2d1db3b1 1031
31caf665
NH
1032 if (PageHuge(page))
1033 rc = unmap_and_move_huge_page(get_new_page,
1034 private, page, pass > 2, mode);
1035 else
1036 rc = unmap_and_move(get_new_page, private,
9c620e2b 1037 page, pass > 2, mode);
2d1db3b1 1038
e24f0b8f 1039 switch(rc) {
95a402c3
CL
1040 case -ENOMEM:
1041 goto out;
e24f0b8f 1042 case -EAGAIN:
2d1db3b1 1043 retry++;
e24f0b8f 1044 break;
78bd5209 1045 case MIGRATEPAGE_SUCCESS:
5647bc29 1046 nr_succeeded++;
e24f0b8f
CL
1047 break;
1048 default:
2d1db3b1 1049 /* Permanent failure */
2d1db3b1 1050 nr_failed++;
e24f0b8f 1051 break;
2d1db3b1 1052 }
b20a3503
CL
1053 }
1054 }
78bd5209 1055 rc = nr_failed + retry;
95a402c3 1056out:
5647bc29
MG
1057 if (nr_succeeded)
1058 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1059 if (nr_failed)
1060 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1061 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1062
b20a3503
CL
1063 if (!swapwrite)
1064 current->flags &= ~PF_SWAPWRITE;
1065
78bd5209 1066 return rc;
b20a3503 1067}
95a402c3 1068
189ebff2 1069int migrate_huge_page(struct page *hpage, new_page_t get_new_page,
9c620e2b 1070 unsigned long private, enum migrate_mode mode)
290408d4 1071{
189ebff2
AK
1072 int pass, rc;
1073
1074 for (pass = 0; pass < 10; pass++) {
9c620e2b
HD
1075 rc = unmap_and_move_huge_page(get_new_page, private,
1076 hpage, pass > 2, mode);
189ebff2
AK
1077 switch (rc) {
1078 case -ENOMEM:
1079 goto out;
1080 case -EAGAIN:
1081 /* try again */
290408d4 1082 cond_resched();
189ebff2 1083 break;
78bd5209 1084 case MIGRATEPAGE_SUCCESS:
189ebff2
AK
1085 goto out;
1086 default:
1087 rc = -EIO;
1088 goto out;
290408d4
NH
1089 }
1090 }
290408d4 1091out:
189ebff2 1092 return rc;
290408d4
NH
1093}
1094
742755a1
CL
1095#ifdef CONFIG_NUMA
1096/*
1097 * Move a list of individual pages
1098 */
1099struct page_to_node {
1100 unsigned long addr;
1101 struct page *page;
1102 int node;
1103 int status;
1104};
1105
1106static struct page *new_page_node(struct page *p, unsigned long private,
1107 int **result)
1108{
1109 struct page_to_node *pm = (struct page_to_node *)private;
1110
1111 while (pm->node != MAX_NUMNODES && pm->page != p)
1112 pm++;
1113
1114 if (pm->node == MAX_NUMNODES)
1115 return NULL;
1116
1117 *result = &pm->status;
1118
6484eb3e 1119 return alloc_pages_exact_node(pm->node,
769848c0 1120 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
742755a1
CL
1121}
1122
1123/*
1124 * Move a set of pages as indicated in the pm array. The addr
1125 * field must be set to the virtual address of the page to be moved
1126 * and the node number must contain a valid target node.
5e9a0f02 1127 * The pm array ends with node = MAX_NUMNODES.
742755a1 1128 */
5e9a0f02
BG
1129static int do_move_page_to_node_array(struct mm_struct *mm,
1130 struct page_to_node *pm,
1131 int migrate_all)
742755a1
CL
1132{
1133 int err;
1134 struct page_to_node *pp;
1135 LIST_HEAD(pagelist);
1136
1137 down_read(&mm->mmap_sem);
1138
1139 /*
1140 * Build a list of pages to migrate
1141 */
742755a1
CL
1142 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1143 struct vm_area_struct *vma;
1144 struct page *page;
1145
742755a1
CL
1146 err = -EFAULT;
1147 vma = find_vma(mm, pp->addr);
70384dc6 1148 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
742755a1
CL
1149 goto set_status;
1150
500d65d4 1151 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
89f5b7da
LT
1152
1153 err = PTR_ERR(page);
1154 if (IS_ERR(page))
1155 goto set_status;
1156
742755a1
CL
1157 err = -ENOENT;
1158 if (!page)
1159 goto set_status;
1160
62b61f61 1161 /* Use PageReserved to check for zero page */
b79bc0a0 1162 if (PageReserved(page))
742755a1
CL
1163 goto put_and_set;
1164
1165 pp->page = page;
1166 err = page_to_nid(page);
1167
1168 if (err == pp->node)
1169 /*
1170 * Node already in the right place
1171 */
1172 goto put_and_set;
1173
1174 err = -EACCES;
1175 if (page_mapcount(page) > 1 &&
1176 !migrate_all)
1177 goto put_and_set;
1178
62695a84 1179 err = isolate_lru_page(page);
6d9c285a 1180 if (!err) {
62695a84 1181 list_add_tail(&page->lru, &pagelist);
6d9c285a
KM
1182 inc_zone_page_state(page, NR_ISOLATED_ANON +
1183 page_is_file_cache(page));
1184 }
742755a1
CL
1185put_and_set:
1186 /*
1187 * Either remove the duplicate refcount from
1188 * isolate_lru_page() or drop the page ref if it was
1189 * not isolated.
1190 */
1191 put_page(page);
1192set_status:
1193 pp->status = err;
1194 }
1195
e78bbfa8 1196 err = 0;
cf608ac1 1197 if (!list_empty(&pagelist)) {
742755a1 1198 err = migrate_pages(&pagelist, new_page_node,
9c620e2b 1199 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
cf608ac1
MK
1200 if (err)
1201 putback_lru_pages(&pagelist);
1202 }
742755a1
CL
1203
1204 up_read(&mm->mmap_sem);
1205 return err;
1206}
1207
5e9a0f02
BG
1208/*
1209 * Migrate an array of page address onto an array of nodes and fill
1210 * the corresponding array of status.
1211 */
3268c63e 1212static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1213 unsigned long nr_pages,
1214 const void __user * __user *pages,
1215 const int __user *nodes,
1216 int __user *status, int flags)
1217{
3140a227 1218 struct page_to_node *pm;
3140a227
BG
1219 unsigned long chunk_nr_pages;
1220 unsigned long chunk_start;
1221 int err;
5e9a0f02 1222
3140a227
BG
1223 err = -ENOMEM;
1224 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1225 if (!pm)
5e9a0f02 1226 goto out;
35282a2d
BG
1227
1228 migrate_prep();
1229
5e9a0f02 1230 /*
3140a227
BG
1231 * Store a chunk of page_to_node array in a page,
1232 * but keep the last one as a marker
5e9a0f02 1233 */
3140a227 1234 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
5e9a0f02 1235
3140a227
BG
1236 for (chunk_start = 0;
1237 chunk_start < nr_pages;
1238 chunk_start += chunk_nr_pages) {
1239 int j;
5e9a0f02 1240
3140a227
BG
1241 if (chunk_start + chunk_nr_pages > nr_pages)
1242 chunk_nr_pages = nr_pages - chunk_start;
1243
1244 /* fill the chunk pm with addrs and nodes from user-space */
1245 for (j = 0; j < chunk_nr_pages; j++) {
1246 const void __user *p;
5e9a0f02
BG
1247 int node;
1248
3140a227
BG
1249 err = -EFAULT;
1250 if (get_user(p, pages + j + chunk_start))
1251 goto out_pm;
1252 pm[j].addr = (unsigned long) p;
1253
1254 if (get_user(node, nodes + j + chunk_start))
5e9a0f02
BG
1255 goto out_pm;
1256
1257 err = -ENODEV;
6f5a55f1
LT
1258 if (node < 0 || node >= MAX_NUMNODES)
1259 goto out_pm;
1260
389162c2 1261 if (!node_state(node, N_MEMORY))
5e9a0f02
BG
1262 goto out_pm;
1263
1264 err = -EACCES;
1265 if (!node_isset(node, task_nodes))
1266 goto out_pm;
1267
3140a227
BG
1268 pm[j].node = node;
1269 }
1270
1271 /* End marker for this chunk */
1272 pm[chunk_nr_pages].node = MAX_NUMNODES;
1273
1274 /* Migrate this chunk */
1275 err = do_move_page_to_node_array(mm, pm,
1276 flags & MPOL_MF_MOVE_ALL);
1277 if (err < 0)
1278 goto out_pm;
5e9a0f02 1279
5e9a0f02 1280 /* Return status information */
3140a227
BG
1281 for (j = 0; j < chunk_nr_pages; j++)
1282 if (put_user(pm[j].status, status + j + chunk_start)) {
5e9a0f02 1283 err = -EFAULT;
3140a227
BG
1284 goto out_pm;
1285 }
1286 }
1287 err = 0;
5e9a0f02
BG
1288
1289out_pm:
3140a227 1290 free_page((unsigned long)pm);
5e9a0f02
BG
1291out:
1292 return err;
1293}
1294
742755a1 1295/*
2f007e74 1296 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1297 */
80bba129
BG
1298static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1299 const void __user **pages, int *status)
742755a1 1300{
2f007e74 1301 unsigned long i;
2f007e74 1302
742755a1
CL
1303 down_read(&mm->mmap_sem);
1304
2f007e74 1305 for (i = 0; i < nr_pages; i++) {
80bba129 1306 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1307 struct vm_area_struct *vma;
1308 struct page *page;
c095adbc 1309 int err = -EFAULT;
2f007e74
BG
1310
1311 vma = find_vma(mm, addr);
70384dc6 1312 if (!vma || addr < vma->vm_start)
742755a1
CL
1313 goto set_status;
1314
2f007e74 1315 page = follow_page(vma, addr, 0);
89f5b7da
LT
1316
1317 err = PTR_ERR(page);
1318 if (IS_ERR(page))
1319 goto set_status;
1320
742755a1
CL
1321 err = -ENOENT;
1322 /* Use PageReserved to check for zero page */
b79bc0a0 1323 if (!page || PageReserved(page))
742755a1
CL
1324 goto set_status;
1325
1326 err = page_to_nid(page);
1327set_status:
80bba129
BG
1328 *status = err;
1329
1330 pages++;
1331 status++;
1332 }
1333
1334 up_read(&mm->mmap_sem);
1335}
1336
1337/*
1338 * Determine the nodes of a user array of pages and store it in
1339 * a user array of status.
1340 */
1341static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1342 const void __user * __user *pages,
1343 int __user *status)
1344{
1345#define DO_PAGES_STAT_CHUNK_NR 16
1346 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1347 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1348
87b8d1ad
PA
1349 while (nr_pages) {
1350 unsigned long chunk_nr;
80bba129 1351
87b8d1ad
PA
1352 chunk_nr = nr_pages;
1353 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1354 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1355
1356 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1357 break;
80bba129
BG
1358
1359 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1360
87b8d1ad
PA
1361 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1362 break;
742755a1 1363
87b8d1ad
PA
1364 pages += chunk_nr;
1365 status += chunk_nr;
1366 nr_pages -= chunk_nr;
1367 }
1368 return nr_pages ? -EFAULT : 0;
742755a1
CL
1369}
1370
1371/*
1372 * Move a list of pages in the address space of the currently executing
1373 * process.
1374 */
938bb9f5
HC
1375SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1376 const void __user * __user *, pages,
1377 const int __user *, nodes,
1378 int __user *, status, int, flags)
742755a1 1379{
c69e8d9c 1380 const struct cred *cred = current_cred(), *tcred;
742755a1 1381 struct task_struct *task;
742755a1 1382 struct mm_struct *mm;
5e9a0f02 1383 int err;
3268c63e 1384 nodemask_t task_nodes;
742755a1
CL
1385
1386 /* Check flags */
1387 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1388 return -EINVAL;
1389
1390 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1391 return -EPERM;
1392
1393 /* Find the mm_struct */
a879bf58 1394 rcu_read_lock();
228ebcbe 1395 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1396 if (!task) {
a879bf58 1397 rcu_read_unlock();
742755a1
CL
1398 return -ESRCH;
1399 }
3268c63e 1400 get_task_struct(task);
742755a1
CL
1401
1402 /*
1403 * Check if this process has the right to modify the specified
1404 * process. The right exists if the process has administrative
1405 * capabilities, superuser privileges or the same
1406 * userid as the target process.
1407 */
c69e8d9c 1408 tcred = __task_cred(task);
b38a86eb
EB
1409 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1410 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
742755a1 1411 !capable(CAP_SYS_NICE)) {
c69e8d9c 1412 rcu_read_unlock();
742755a1 1413 err = -EPERM;
5e9a0f02 1414 goto out;
742755a1 1415 }
c69e8d9c 1416 rcu_read_unlock();
742755a1 1417
86c3a764
DQ
1418 err = security_task_movememory(task);
1419 if (err)
5e9a0f02 1420 goto out;
86c3a764 1421
3268c63e
CL
1422 task_nodes = cpuset_mems_allowed(task);
1423 mm = get_task_mm(task);
1424 put_task_struct(task);
1425
6e8b09ea
SL
1426 if (!mm)
1427 return -EINVAL;
1428
1429 if (nodes)
1430 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1431 nodes, status, flags);
1432 else
1433 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1434
742755a1
CL
1435 mmput(mm);
1436 return err;
3268c63e
CL
1437
1438out:
1439 put_task_struct(task);
1440 return err;
742755a1 1441}
742755a1 1442
7b2259b3
CL
1443/*
1444 * Call migration functions in the vma_ops that may prepare
1445 * memory in a vm for migration. migration functions may perform
1446 * the migration for vmas that do not have an underlying page struct.
1447 */
1448int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1449 const nodemask_t *from, unsigned long flags)
1450{
1451 struct vm_area_struct *vma;
1452 int err = 0;
1453
1001c9fb 1454 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
7b2259b3
CL
1455 if (vma->vm_ops && vma->vm_ops->migrate) {
1456 err = vma->vm_ops->migrate(vma, to, from, flags);
1457 if (err)
1458 break;
1459 }
1460 }
1461 return err;
1462}
7039e1db
PZ
1463
1464#ifdef CONFIG_NUMA_BALANCING
1465/*
1466 * Returns true if this is a safe migration target node for misplaced NUMA
1467 * pages. Currently it only checks the watermarks which crude
1468 */
1469static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1470 unsigned long nr_migrate_pages)
7039e1db
PZ
1471{
1472 int z;
1473 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1474 struct zone *zone = pgdat->node_zones + z;
1475
1476 if (!populated_zone(zone))
1477 continue;
1478
1479 if (zone->all_unreclaimable)
1480 continue;
1481
1482 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1483 if (!zone_watermark_ok(zone, 0,
1484 high_wmark_pages(zone) +
1485 nr_migrate_pages,
1486 0, 0))
1487 continue;
1488 return true;
1489 }
1490 return false;
1491}
1492
1493static struct page *alloc_misplaced_dst_page(struct page *page,
1494 unsigned long data,
1495 int **result)
1496{
1497 int nid = (int) data;
1498 struct page *newpage;
1499
1500 newpage = alloc_pages_exact_node(nid,
1501 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1502 __GFP_NOMEMALLOC | __GFP_NORETRY |
1503 __GFP_NOWARN) &
1504 ~GFP_IOFS, 0);
bac0382c 1505 if (newpage)
22b751c3 1506 page_nid_xchg_last(newpage, page_nid_last(page));
bac0382c 1507
7039e1db
PZ
1508 return newpage;
1509}
1510
a8f60772
MG
1511/*
1512 * page migration rate limiting control.
1513 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1514 * window of time. Default here says do not migrate more than 1280M per second.
e14808b4
MG
1515 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1516 * as it is faults that reset the window, pte updates will happen unconditionally
1517 * if there has not been a fault since @pteupdate_interval_millisecs after the
1518 * throttle window closed.
a8f60772
MG
1519 */
1520static unsigned int migrate_interval_millisecs __read_mostly = 100;
e14808b4 1521static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
a8f60772
MG
1522static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1523
e14808b4
MG
1524/* Returns true if NUMA migration is currently rate limited */
1525bool migrate_ratelimited(int node)
1526{
1527 pg_data_t *pgdat = NODE_DATA(node);
1528
1529 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1530 msecs_to_jiffies(pteupdate_interval_millisecs)))
1531 return false;
1532
1533 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1534 return false;
1535
1536 return true;
1537}
1538
b32967ff 1539/* Returns true if the node is migrate rate-limited after the update */
d28d4335 1540bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
7039e1db 1541{
b32967ff 1542 bool rate_limited = false;
7039e1db 1543
a8f60772
MG
1544 /*
1545 * Rate-limit the amount of data that is being migrated to a node.
1546 * Optimal placement is no good if the memory bus is saturated and
1547 * all the time is being spent migrating!
1548 */
1549 spin_lock(&pgdat->numabalancing_migrate_lock);
1550 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1551 pgdat->numabalancing_migrate_nr_pages = 0;
1552 pgdat->numabalancing_migrate_next_window = jiffies +
1553 msecs_to_jiffies(migrate_interval_millisecs);
1554 }
b32967ff
MG
1555 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1556 rate_limited = true;
1557 else
d28d4335 1558 pgdat->numabalancing_migrate_nr_pages += nr_pages;
a8f60772 1559 spin_unlock(&pgdat->numabalancing_migrate_lock);
b32967ff
MG
1560
1561 return rate_limited;
1562}
1563
1564int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1565{
340ef390 1566 int page_lru;
a8f60772 1567
3abef4e6
MG
1568 VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1569
7039e1db 1570 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1571 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1572 return 0;
7039e1db 1573
340ef390
HD
1574 if (isolate_lru_page(page))
1575 return 0;
7039e1db 1576
340ef390
HD
1577 /*
1578 * migrate_misplaced_transhuge_page() skips page migration's usual
1579 * check on page_count(), so we must do it here, now that the page
1580 * has been isolated: a GUP pin, or any other pin, prevents migration.
1581 * The expected page count is 3: 1 for page's mapcount and 1 for the
1582 * caller's pin and 1 for the reference taken by isolate_lru_page().
1583 */
1584 if (PageTransHuge(page) && page_count(page) != 3) {
1585 putback_lru_page(page);
1586 return 0;
7039e1db
PZ
1587 }
1588
340ef390
HD
1589 page_lru = page_is_file_cache(page);
1590 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1591 hpage_nr_pages(page));
1592
149c33e1 1593 /*
340ef390
HD
1594 * Isolating the page has taken another reference, so the
1595 * caller's reference can be safely dropped without the page
1596 * disappearing underneath us during migration.
149c33e1
MG
1597 */
1598 put_page(page);
340ef390 1599 return 1;
b32967ff
MG
1600}
1601
1602/*
1603 * Attempt to migrate a misplaced page to the specified destination
1604 * node. Caller is expected to have an elevated reference count on
1605 * the page that will be dropped by this function before returning.
1606 */
1607int migrate_misplaced_page(struct page *page, int node)
1608{
1609 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1610 int isolated;
b32967ff
MG
1611 int nr_remaining;
1612 LIST_HEAD(migratepages);
1613
1614 /*
1615 * Don't migrate pages that are mapped in multiple processes.
1616 * TODO: Handle false sharing detection instead of this hammer
1617 */
340ef390 1618 if (page_mapcount(page) != 1)
b32967ff 1619 goto out;
b32967ff
MG
1620
1621 /*
1622 * Rate-limit the amount of data that is being migrated to a node.
1623 * Optimal placement is no good if the memory bus is saturated and
1624 * all the time is being spent migrating!
1625 */
340ef390 1626 if (numamigrate_update_ratelimit(pgdat, 1))
b32967ff 1627 goto out;
b32967ff
MG
1628
1629 isolated = numamigrate_isolate_page(pgdat, page);
1630 if (!isolated)
1631 goto out;
1632
1633 list_add(&page->lru, &migratepages);
9c620e2b
HD
1634 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1635 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
b32967ff
MG
1636 if (nr_remaining) {
1637 putback_lru_pages(&migratepages);
1638 isolated = 0;
1639 } else
1640 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1641 BUG_ON(!list_empty(&migratepages));
7039e1db 1642 return isolated;
340ef390
HD
1643
1644out:
1645 put_page(page);
1646 return 0;
7039e1db 1647}
220018d3 1648#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1649
220018d3 1650#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1651/*
1652 * Migrates a THP to a given target node. page must be locked and is unlocked
1653 * before returning.
1654 */
b32967ff
MG
1655int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1656 struct vm_area_struct *vma,
1657 pmd_t *pmd, pmd_t entry,
1658 unsigned long address,
1659 struct page *page, int node)
1660{
1661 unsigned long haddr = address & HPAGE_PMD_MASK;
1662 pg_data_t *pgdat = NODE_DATA(node);
1663 int isolated = 0;
1664 struct page *new_page = NULL;
1665 struct mem_cgroup *memcg = NULL;
1666 int page_lru = page_is_file_cache(page);
1667
1668 /*
1669 * Don't migrate pages that are mapped in multiple processes.
1670 * TODO: Handle false sharing detection instead of this hammer
1671 */
1672 if (page_mapcount(page) != 1)
1673 goto out_dropref;
1674
1675 /*
1676 * Rate-limit the amount of data that is being migrated to a node.
1677 * Optimal placement is no good if the memory bus is saturated and
1678 * all the time is being spent migrating!
1679 */
d28d4335 1680 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
b32967ff
MG
1681 goto out_dropref;
1682
1683 new_page = alloc_pages_node(node,
1684 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
340ef390
HD
1685 if (!new_page)
1686 goto out_fail;
1687
22b751c3 1688 page_nid_xchg_last(new_page, page_nid_last(page));
b32967ff
MG
1689
1690 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 1691 if (!isolated) {
b32967ff 1692 put_page(new_page);
340ef390 1693 goto out_fail;
b32967ff
MG
1694 }
1695
1696 /* Prepare a page as a migration target */
1697 __set_page_locked(new_page);
1698 SetPageSwapBacked(new_page);
1699
1700 /* anon mapping, we can simply copy page->mapping to the new page: */
1701 new_page->mapping = page->mapping;
1702 new_page->index = page->index;
1703 migrate_page_copy(new_page, page);
1704 WARN_ON(PageLRU(new_page));
1705
1706 /* Recheck the target PMD */
1707 spin_lock(&mm->page_table_lock);
1708 if (unlikely(!pmd_same(*pmd, entry))) {
1709 spin_unlock(&mm->page_table_lock);
1710
1711 /* Reverse changes made by migrate_page_copy() */
1712 if (TestClearPageActive(new_page))
1713 SetPageActive(page);
1714 if (TestClearPageUnevictable(new_page))
1715 SetPageUnevictable(page);
1716 mlock_migrate_page(page, new_page);
1717
1718 unlock_page(new_page);
1719 put_page(new_page); /* Free it */
1720
1721 unlock_page(page);
1722 putback_lru_page(page);
1723
1724 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
340ef390 1725 isolated = 0;
b32967ff
MG
1726 goto out;
1727 }
1728
1729 /*
1730 * Traditional migration needs to prepare the memcg charge
1731 * transaction early to prevent the old page from being
1732 * uncharged when installing migration entries. Here we can
1733 * save the potential rollback and start the charge transfer
1734 * only when migration is already known to end successfully.
1735 */
1736 mem_cgroup_prepare_migration(page, new_page, &memcg);
1737
1738 entry = mk_pmd(new_page, vma->vm_page_prot);
1739 entry = pmd_mknonnuma(entry);
1740 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1741 entry = pmd_mkhuge(entry);
1742
1743 page_add_new_anon_rmap(new_page, vma, haddr);
1744
1745 set_pmd_at(mm, haddr, pmd, entry);
ce4a9cc5 1746 update_mmu_cache_pmd(vma, address, &entry);
b32967ff
MG
1747 page_remove_rmap(page);
1748 /*
1749 * Finish the charge transaction under the page table lock to
1750 * prevent split_huge_page() from dividing up the charge
1751 * before it's fully transferred to the new page.
1752 */
1753 mem_cgroup_end_migration(memcg, page, new_page, true);
1754 spin_unlock(&mm->page_table_lock);
1755
1756 unlock_page(new_page);
1757 unlock_page(page);
1758 put_page(page); /* Drop the rmap reference */
1759 put_page(page); /* Drop the LRU isolation reference */
1760
1761 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1762 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1763
1764out:
1765 mod_zone_page_state(page_zone(page),
1766 NR_ISOLATED_ANON + page_lru,
1767 -HPAGE_PMD_NR);
1768 return isolated;
1769
340ef390
HD
1770out_fail:
1771 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
b32967ff 1772out_dropref:
340ef390 1773 unlock_page(page);
b32967ff 1774 put_page(page);
b32967ff
MG
1775 return 0;
1776}
7039e1db
PZ
1777#endif /* CONFIG_NUMA_BALANCING */
1778
1779#endif /* CONFIG_NUMA */