]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
mm/kasan/quarantine.c: make quarantine_lock a raw_spinlock_t
[mirror_ubuntu-jammy-kernel.git] / mm / migrate.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
b20a3503 2/*
14e0f9bc 3 * Memory Migration functionality - linux/mm/migrate.c
b20a3503
CL
4 *
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 *
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
9 *
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
cde53535 13 * Christoph Lameter
b20a3503
CL
14 */
15
16#include <linux/migrate.h>
b95f1b31 17#include <linux/export.h>
b20a3503 18#include <linux/swap.h>
0697212a 19#include <linux/swapops.h>
b20a3503 20#include <linux/pagemap.h>
e23ca00b 21#include <linux/buffer_head.h>
b20a3503 22#include <linux/mm_inline.h>
b488893a 23#include <linux/nsproxy.h>
b20a3503 24#include <linux/pagevec.h>
e9995ef9 25#include <linux/ksm.h>
b20a3503
CL
26#include <linux/rmap.h>
27#include <linux/topology.h>
28#include <linux/cpu.h>
29#include <linux/cpuset.h>
04e62a29 30#include <linux/writeback.h>
742755a1
CL
31#include <linux/mempolicy.h>
32#include <linux/vmalloc.h>
86c3a764 33#include <linux/security.h>
42cb14b1 34#include <linux/backing-dev.h>
bda807d4 35#include <linux/compaction.h>
4f5ca265 36#include <linux/syscalls.h>
7addf443 37#include <linux/compat.h>
290408d4 38#include <linux/hugetlb.h>
8e6ac7fa 39#include <linux/hugetlb_cgroup.h>
5a0e3ad6 40#include <linux/gfp.h>
df6ad698 41#include <linux/pfn_t.h>
a5430dda 42#include <linux/memremap.h>
8315ada7 43#include <linux/userfaultfd_k.h>
bf6bddf1 44#include <linux/balloon_compaction.h>
f714f4f2 45#include <linux/mmu_notifier.h>
33c3fc71 46#include <linux/page_idle.h>
d435edca 47#include <linux/page_owner.h>
6e84f315 48#include <linux/sched/mm.h>
197e7e52 49#include <linux/ptrace.h>
b20a3503 50
0d1836c3
MN
51#include <asm/tlbflush.h>
52
7b2a2d4a
MG
53#define CREATE_TRACE_POINTS
54#include <trace/events/migrate.h>
55
b20a3503
CL
56#include "internal.h"
57
b20a3503 58/*
742755a1 59 * migrate_prep() needs to be called before we start compiling a list of pages
748446bb
MG
60 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
61 * undesirable, use migrate_prep_local()
b20a3503
CL
62 */
63int migrate_prep(void)
64{
b20a3503
CL
65 /*
66 * Clear the LRU lists so pages can be isolated.
67 * Note that pages may be moved off the LRU after we have
68 * drained them. Those pages will fail to migrate like other
69 * pages that may be busy.
70 */
71 lru_add_drain_all();
72
73 return 0;
74}
75
748446bb
MG
76/* Do the necessary work of migrate_prep but not if it involves other CPUs */
77int migrate_prep_local(void)
78{
79 lru_add_drain();
80
81 return 0;
82}
83
9e5bcd61 84int isolate_movable_page(struct page *page, isolate_mode_t mode)
bda807d4
MK
85{
86 struct address_space *mapping;
87
88 /*
89 * Avoid burning cycles with pages that are yet under __free_pages(),
90 * or just got freed under us.
91 *
92 * In case we 'win' a race for a movable page being freed under us and
93 * raise its refcount preventing __free_pages() from doing its job
94 * the put_page() at the end of this block will take care of
95 * release this page, thus avoiding a nasty leakage.
96 */
97 if (unlikely(!get_page_unless_zero(page)))
98 goto out;
99
100 /*
101 * Check PageMovable before holding a PG_lock because page's owner
102 * assumes anybody doesn't touch PG_lock of newly allocated page
103 * so unconditionally grapping the lock ruins page's owner side.
104 */
105 if (unlikely(!__PageMovable(page)))
106 goto out_putpage;
107 /*
108 * As movable pages are not isolated from LRU lists, concurrent
109 * compaction threads can race against page migration functions
110 * as well as race against the releasing a page.
111 *
112 * In order to avoid having an already isolated movable page
113 * being (wrongly) re-isolated while it is under migration,
114 * or to avoid attempting to isolate pages being released,
115 * lets be sure we have the page lock
116 * before proceeding with the movable page isolation steps.
117 */
118 if (unlikely(!trylock_page(page)))
119 goto out_putpage;
120
121 if (!PageMovable(page) || PageIsolated(page))
122 goto out_no_isolated;
123
124 mapping = page_mapping(page);
125 VM_BUG_ON_PAGE(!mapping, page);
126
127 if (!mapping->a_ops->isolate_page(page, mode))
128 goto out_no_isolated;
129
130 /* Driver shouldn't use PG_isolated bit of page->flags */
131 WARN_ON_ONCE(PageIsolated(page));
132 __SetPageIsolated(page);
133 unlock_page(page);
134
9e5bcd61 135 return 0;
bda807d4
MK
136
137out_no_isolated:
138 unlock_page(page);
139out_putpage:
140 put_page(page);
141out:
9e5bcd61 142 return -EBUSY;
bda807d4
MK
143}
144
145/* It should be called on page which is PG_movable */
146void putback_movable_page(struct page *page)
147{
148 struct address_space *mapping;
149
150 VM_BUG_ON_PAGE(!PageLocked(page), page);
151 VM_BUG_ON_PAGE(!PageMovable(page), page);
152 VM_BUG_ON_PAGE(!PageIsolated(page), page);
153
154 mapping = page_mapping(page);
155 mapping->a_ops->putback_page(page);
156 __ClearPageIsolated(page);
157}
158
5733c7d1
RA
159/*
160 * Put previously isolated pages back onto the appropriate lists
161 * from where they were once taken off for compaction/migration.
162 *
59c82b70
JK
163 * This function shall be used whenever the isolated pageset has been
164 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
165 * and isolate_huge_page().
5733c7d1
RA
166 */
167void putback_movable_pages(struct list_head *l)
168{
169 struct page *page;
170 struct page *page2;
171
b20a3503 172 list_for_each_entry_safe(page, page2, l, lru) {
31caf665
NH
173 if (unlikely(PageHuge(page))) {
174 putback_active_hugepage(page);
175 continue;
176 }
e24f0b8f 177 list_del(&page->lru);
bda807d4
MK
178 /*
179 * We isolated non-lru movable page so here we can use
180 * __PageMovable because LRU page's mapping cannot have
181 * PAGE_MAPPING_MOVABLE.
182 */
b1123ea6 183 if (unlikely(__PageMovable(page))) {
bda807d4
MK
184 VM_BUG_ON_PAGE(!PageIsolated(page), page);
185 lock_page(page);
186 if (PageMovable(page))
187 putback_movable_page(page);
188 else
189 __ClearPageIsolated(page);
190 unlock_page(page);
191 put_page(page);
192 } else {
e8db67eb
NH
193 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
194 page_is_file_cache(page), -hpage_nr_pages(page));
fc280fe8 195 putback_lru_page(page);
bda807d4 196 }
b20a3503 197 }
b20a3503
CL
198}
199
0697212a
CL
200/*
201 * Restore a potential migration pte to a working pte entry
202 */
e4b82222 203static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
e9995ef9 204 unsigned long addr, void *old)
0697212a 205{
3fe87967
KS
206 struct page_vma_mapped_walk pvmw = {
207 .page = old,
208 .vma = vma,
209 .address = addr,
210 .flags = PVMW_SYNC | PVMW_MIGRATION,
211 };
212 struct page *new;
213 pte_t pte;
0697212a 214 swp_entry_t entry;
0697212a 215
3fe87967
KS
216 VM_BUG_ON_PAGE(PageTail(page), page);
217 while (page_vma_mapped_walk(&pvmw)) {
4b0ece6f
NH
218 if (PageKsm(page))
219 new = page;
220 else
221 new = page - pvmw.page->index +
222 linear_page_index(vma, pvmw.address);
0697212a 223
616b8371
ZY
224#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
225 /* PMD-mapped THP migration entry */
226 if (!pvmw.pte) {
227 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
228 remove_migration_pmd(&pvmw, new);
229 continue;
230 }
231#endif
232
3fe87967
KS
233 get_page(new);
234 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
235 if (pte_swp_soft_dirty(*pvmw.pte))
236 pte = pte_mksoft_dirty(pte);
0697212a 237
3fe87967
KS
238 /*
239 * Recheck VMA as permissions can change since migration started
240 */
241 entry = pte_to_swp_entry(*pvmw.pte);
242 if (is_write_migration_entry(entry))
243 pte = maybe_mkwrite(pte, vma);
d3cb8bf6 244
df6ad698
JG
245 if (unlikely(is_zone_device_page(new))) {
246 if (is_device_private_page(new)) {
247 entry = make_device_private_entry(new, pte_write(pte));
248 pte = swp_entry_to_pte(entry);
249 } else if (is_device_public_page(new)) {
250 pte = pte_mkdevmap(pte);
251 flush_dcache_page(new);
252 }
a5430dda
JG
253 } else
254 flush_dcache_page(new);
255
3ef8fd7f 256#ifdef CONFIG_HUGETLB_PAGE
3fe87967
KS
257 if (PageHuge(new)) {
258 pte = pte_mkhuge(pte);
259 pte = arch_make_huge_pte(pte, vma, new, 0);
383321ab 260 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
3fe87967
KS
261 if (PageAnon(new))
262 hugepage_add_anon_rmap(new, vma, pvmw.address);
263 else
264 page_dup_rmap(new, true);
383321ab
AK
265 } else
266#endif
267 {
268 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
04e62a29 269
383321ab
AK
270 if (PageAnon(new))
271 page_add_anon_rmap(new, vma, pvmw.address, false);
272 else
273 page_add_file_rmap(new, false);
274 }
3fe87967
KS
275 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
276 mlock_vma_page(new);
277
e125fe40
KS
278 if (PageTransHuge(page) && PageMlocked(page))
279 clear_page_mlock(page);
280
3fe87967
KS
281 /* No need to invalidate - it was non-present before */
282 update_mmu_cache(vma, pvmw.address, pvmw.pte);
283 }
51afb12b 284
e4b82222 285 return true;
0697212a
CL
286}
287
04e62a29
CL
288/*
289 * Get rid of all migration entries and replace them by
290 * references to the indicated page.
291 */
e388466d 292void remove_migration_ptes(struct page *old, struct page *new, bool locked)
04e62a29 293{
051ac83a
JK
294 struct rmap_walk_control rwc = {
295 .rmap_one = remove_migration_pte,
296 .arg = old,
297 };
298
e388466d
KS
299 if (locked)
300 rmap_walk_locked(new, &rwc);
301 else
302 rmap_walk(new, &rwc);
04e62a29
CL
303}
304
0697212a
CL
305/*
306 * Something used the pte of a page under migration. We need to
307 * get to the page and wait until migration is finished.
308 * When we return from this function the fault will be retried.
0697212a 309 */
e66f17ff 310void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
30dad309 311 spinlock_t *ptl)
0697212a 312{
30dad309 313 pte_t pte;
0697212a
CL
314 swp_entry_t entry;
315 struct page *page;
316
30dad309 317 spin_lock(ptl);
0697212a
CL
318 pte = *ptep;
319 if (!is_swap_pte(pte))
320 goto out;
321
322 entry = pte_to_swp_entry(pte);
323 if (!is_migration_entry(entry))
324 goto out;
325
326 page = migration_entry_to_page(entry);
327
e286781d
NP
328 /*
329 * Once radix-tree replacement of page migration started, page_count
330 * *must* be zero. And, we don't want to call wait_on_page_locked()
331 * against a page without get_page().
332 * So, we use get_page_unless_zero(), here. Even failed, page fault
333 * will occur again.
334 */
335 if (!get_page_unless_zero(page))
336 goto out;
0697212a
CL
337 pte_unmap_unlock(ptep, ptl);
338 wait_on_page_locked(page);
339 put_page(page);
340 return;
341out:
342 pte_unmap_unlock(ptep, ptl);
343}
344
30dad309
NH
345void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
346 unsigned long address)
347{
348 spinlock_t *ptl = pte_lockptr(mm, pmd);
349 pte_t *ptep = pte_offset_map(pmd, address);
350 __migration_entry_wait(mm, ptep, ptl);
351}
352
cb900f41
KS
353void migration_entry_wait_huge(struct vm_area_struct *vma,
354 struct mm_struct *mm, pte_t *pte)
30dad309 355{
cb900f41 356 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
30dad309
NH
357 __migration_entry_wait(mm, pte, ptl);
358}
359
616b8371
ZY
360#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
361void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
362{
363 spinlock_t *ptl;
364 struct page *page;
365
366 ptl = pmd_lock(mm, pmd);
367 if (!is_pmd_migration_entry(*pmd))
368 goto unlock;
369 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
370 if (!get_page_unless_zero(page))
371 goto unlock;
372 spin_unlock(ptl);
373 wait_on_page_locked(page);
374 put_page(page);
375 return;
376unlock:
377 spin_unlock(ptl);
378}
379#endif
380
b969c4ab
MG
381#ifdef CONFIG_BLOCK
382/* Returns true if all buffers are successfully locked */
a6bc32b8
MG
383static bool buffer_migrate_lock_buffers(struct buffer_head *head,
384 enum migrate_mode mode)
b969c4ab
MG
385{
386 struct buffer_head *bh = head;
387
388 /* Simple case, sync compaction */
a6bc32b8 389 if (mode != MIGRATE_ASYNC) {
b969c4ab
MG
390 do {
391 get_bh(bh);
392 lock_buffer(bh);
393 bh = bh->b_this_page;
394
395 } while (bh != head);
396
397 return true;
398 }
399
400 /* async case, we cannot block on lock_buffer so use trylock_buffer */
401 do {
402 get_bh(bh);
403 if (!trylock_buffer(bh)) {
404 /*
405 * We failed to lock the buffer and cannot stall in
406 * async migration. Release the taken locks
407 */
408 struct buffer_head *failed_bh = bh;
409 put_bh(failed_bh);
410 bh = head;
411 while (bh != failed_bh) {
412 unlock_buffer(bh);
413 put_bh(bh);
414 bh = bh->b_this_page;
415 }
416 return false;
417 }
418
419 bh = bh->b_this_page;
420 } while (bh != head);
421 return true;
422}
423#else
424static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
a6bc32b8 425 enum migrate_mode mode)
b969c4ab
MG
426{
427 return true;
428}
429#endif /* CONFIG_BLOCK */
430
b20a3503 431/*
c3fcf8a5 432 * Replace the page in the mapping.
5b5c7120
CL
433 *
434 * The number of remaining references must be:
435 * 1 for anonymous pages without a mapping
436 * 2 for pages with a mapping
266cf658 437 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
b20a3503 438 */
36bc08cc 439int migrate_page_move_mapping(struct address_space *mapping,
b969c4ab 440 struct page *newpage, struct page *page,
8e321fef
BL
441 struct buffer_head *head, enum migrate_mode mode,
442 int extra_count)
b20a3503 443{
42cb14b1
HD
444 struct zone *oldzone, *newzone;
445 int dirty;
8e321fef 446 int expected_count = 1 + extra_count;
7cf9c2c7 447 void **pslot;
b20a3503 448
8763cb45 449 /*
df6ad698
JG
450 * Device public or private pages have an extra refcount as they are
451 * ZONE_DEVICE pages.
8763cb45 452 */
df6ad698
JG
453 expected_count += is_device_private_page(page);
454 expected_count += is_device_public_page(page);
8763cb45 455
6c5240ae 456 if (!mapping) {
0e8c7d0f 457 /* Anonymous page without mapping */
8e321fef 458 if (page_count(page) != expected_count)
6c5240ae 459 return -EAGAIN;
cf4b769a
HD
460
461 /* No turning back from here */
cf4b769a
HD
462 newpage->index = page->index;
463 newpage->mapping = page->mapping;
464 if (PageSwapBacked(page))
fa9949da 465 __SetPageSwapBacked(newpage);
cf4b769a 466
78bd5209 467 return MIGRATEPAGE_SUCCESS;
6c5240ae
CL
468 }
469
42cb14b1
HD
470 oldzone = page_zone(page);
471 newzone = page_zone(newpage);
472
b93b0163 473 xa_lock_irq(&mapping->i_pages);
b20a3503 474
b93b0163 475 pslot = radix_tree_lookup_slot(&mapping->i_pages,
7cf9c2c7 476 page_index(page));
b20a3503 477
e71769ae 478 expected_count += hpage_nr_pages(page) + page_has_private(page);
e286781d 479 if (page_count(page) != expected_count ||
b93b0163
MW
480 radix_tree_deref_slot_protected(pslot,
481 &mapping->i_pages.xa_lock) != page) {
482 xa_unlock_irq(&mapping->i_pages);
e23ca00b 483 return -EAGAIN;
b20a3503
CL
484 }
485
fe896d18 486 if (!page_ref_freeze(page, expected_count)) {
b93b0163 487 xa_unlock_irq(&mapping->i_pages);
e286781d
NP
488 return -EAGAIN;
489 }
490
b969c4ab
MG
491 /*
492 * In the async migration case of moving a page with buffers, lock the
493 * buffers using trylock before the mapping is moved. If the mapping
494 * was moved, we later failed to lock the buffers and could not move
495 * the mapping back due to an elevated page count, we would have to
496 * block waiting on other references to be dropped.
497 */
a6bc32b8
MG
498 if (mode == MIGRATE_ASYNC && head &&
499 !buffer_migrate_lock_buffers(head, mode)) {
fe896d18 500 page_ref_unfreeze(page, expected_count);
b93b0163 501 xa_unlock_irq(&mapping->i_pages);
b969c4ab
MG
502 return -EAGAIN;
503 }
504
b20a3503 505 /*
cf4b769a
HD
506 * Now we know that no one else is looking at the page:
507 * no turning back from here.
b20a3503 508 */
cf4b769a
HD
509 newpage->index = page->index;
510 newpage->mapping = page->mapping;
e71769ae 511 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
6326fec1
NP
512 if (PageSwapBacked(page)) {
513 __SetPageSwapBacked(newpage);
514 if (PageSwapCache(page)) {
515 SetPageSwapCache(newpage);
516 set_page_private(newpage, page_private(page));
517 }
518 } else {
519 VM_BUG_ON_PAGE(PageSwapCache(page), page);
b20a3503
CL
520 }
521
42cb14b1
HD
522 /* Move dirty while page refs frozen and newpage not yet exposed */
523 dirty = PageDirty(page);
524 if (dirty) {
525 ClearPageDirty(page);
526 SetPageDirty(newpage);
527 }
528
b93b0163 529 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
e71769ae
NH
530 if (PageTransHuge(page)) {
531 int i;
532 int index = page_index(page);
533
013567be 534 for (i = 1; i < HPAGE_PMD_NR; i++) {
e71769ae
NH
535 pslot = radix_tree_lookup_slot(&mapping->i_pages,
536 index + i);
537 radix_tree_replace_slot(&mapping->i_pages, pslot,
538 newpage + i);
539 }
e71769ae 540 }
7cf9c2c7
NP
541
542 /*
937a94c9
JG
543 * Drop cache reference from old page by unfreezing
544 * to one less reference.
7cf9c2c7
NP
545 * We know this isn't the last reference.
546 */
e71769ae 547 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
7cf9c2c7 548
b93b0163 549 xa_unlock(&mapping->i_pages);
42cb14b1
HD
550 /* Leave irq disabled to prevent preemption while updating stats */
551
0e8c7d0f
CL
552 /*
553 * If moved to a different zone then also account
554 * the page for that zone. Other VM counters will be
555 * taken care of when we establish references to the
556 * new page and drop references to the old page.
557 *
558 * Note that anonymous pages are accounted for
4b9d0fab 559 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
0e8c7d0f
CL
560 * are mapped to swap space.
561 */
42cb14b1 562 if (newzone != oldzone) {
11fb9989
MG
563 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
564 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
42cb14b1 565 if (PageSwapBacked(page) && !PageSwapCache(page)) {
11fb9989
MG
566 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
567 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
42cb14b1
HD
568 }
569 if (dirty && mapping_cap_account_dirty(mapping)) {
11fb9989 570 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 571 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
11fb9989 572 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 573 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
42cb14b1 574 }
4b02108a 575 }
42cb14b1 576 local_irq_enable();
b20a3503 577
78bd5209 578 return MIGRATEPAGE_SUCCESS;
b20a3503 579}
1118dce7 580EXPORT_SYMBOL(migrate_page_move_mapping);
b20a3503 581
290408d4
NH
582/*
583 * The expected number of remaining references is the same as that
584 * of migrate_page_move_mapping().
585 */
586int migrate_huge_page_move_mapping(struct address_space *mapping,
587 struct page *newpage, struct page *page)
588{
589 int expected_count;
590 void **pslot;
591
b93b0163 592 xa_lock_irq(&mapping->i_pages);
290408d4 593
b93b0163 594 pslot = radix_tree_lookup_slot(&mapping->i_pages, page_index(page));
290408d4
NH
595
596 expected_count = 2 + page_has_private(page);
597 if (page_count(page) != expected_count ||
b93b0163
MW
598 radix_tree_deref_slot_protected(pslot, &mapping->i_pages.xa_lock) != page) {
599 xa_unlock_irq(&mapping->i_pages);
290408d4
NH
600 return -EAGAIN;
601 }
602
fe896d18 603 if (!page_ref_freeze(page, expected_count)) {
b93b0163 604 xa_unlock_irq(&mapping->i_pages);
290408d4
NH
605 return -EAGAIN;
606 }
607
cf4b769a
HD
608 newpage->index = page->index;
609 newpage->mapping = page->mapping;
6a93ca8f 610
290408d4
NH
611 get_page(newpage);
612
b93b0163 613 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
290408d4 614
fe896d18 615 page_ref_unfreeze(page, expected_count - 1);
290408d4 616
b93b0163 617 xa_unlock_irq(&mapping->i_pages);
6a93ca8f 618
78bd5209 619 return MIGRATEPAGE_SUCCESS;
290408d4
NH
620}
621
30b0a105
DH
622/*
623 * Gigantic pages are so large that we do not guarantee that page++ pointer
624 * arithmetic will work across the entire page. We need something more
625 * specialized.
626 */
627static void __copy_gigantic_page(struct page *dst, struct page *src,
628 int nr_pages)
629{
630 int i;
631 struct page *dst_base = dst;
632 struct page *src_base = src;
633
634 for (i = 0; i < nr_pages; ) {
635 cond_resched();
636 copy_highpage(dst, src);
637
638 i++;
639 dst = mem_map_next(dst, dst_base, i);
640 src = mem_map_next(src, src_base, i);
641 }
642}
643
644static void copy_huge_page(struct page *dst, struct page *src)
645{
646 int i;
647 int nr_pages;
648
649 if (PageHuge(src)) {
650 /* hugetlbfs page */
651 struct hstate *h = page_hstate(src);
652 nr_pages = pages_per_huge_page(h);
653
654 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
655 __copy_gigantic_page(dst, src, nr_pages);
656 return;
657 }
658 } else {
659 /* thp page */
660 BUG_ON(!PageTransHuge(src));
661 nr_pages = hpage_nr_pages(src);
662 }
663
664 for (i = 0; i < nr_pages; i++) {
665 cond_resched();
666 copy_highpage(dst + i, src + i);
667 }
668}
669
b20a3503
CL
670/*
671 * Copy the page to its new location
672 */
2916ecc0 673void migrate_page_states(struct page *newpage, struct page *page)
b20a3503 674{
7851a45c
RR
675 int cpupid;
676
b20a3503
CL
677 if (PageError(page))
678 SetPageError(newpage);
679 if (PageReferenced(page))
680 SetPageReferenced(newpage);
681 if (PageUptodate(page))
682 SetPageUptodate(newpage);
894bc310 683 if (TestClearPageActive(page)) {
309381fe 684 VM_BUG_ON_PAGE(PageUnevictable(page), page);
b20a3503 685 SetPageActive(newpage);
418b27ef
LS
686 } else if (TestClearPageUnevictable(page))
687 SetPageUnevictable(newpage);
1899ad18
JW
688 if (PageWorkingset(page))
689 SetPageWorkingset(newpage);
b20a3503
CL
690 if (PageChecked(page))
691 SetPageChecked(newpage);
692 if (PageMappedToDisk(page))
693 SetPageMappedToDisk(newpage);
694
42cb14b1
HD
695 /* Move dirty on pages not done by migrate_page_move_mapping() */
696 if (PageDirty(page))
697 SetPageDirty(newpage);
b20a3503 698
33c3fc71
VD
699 if (page_is_young(page))
700 set_page_young(newpage);
701 if (page_is_idle(page))
702 set_page_idle(newpage);
703
7851a45c
RR
704 /*
705 * Copy NUMA information to the new page, to prevent over-eager
706 * future migrations of this same page.
707 */
708 cpupid = page_cpupid_xchg_last(page, -1);
709 page_cpupid_xchg_last(newpage, cpupid);
710
e9995ef9 711 ksm_migrate_page(newpage, page);
c8d6553b
HD
712 /*
713 * Please do not reorder this without considering how mm/ksm.c's
714 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
715 */
b3b3a99c
NH
716 if (PageSwapCache(page))
717 ClearPageSwapCache(page);
b20a3503
CL
718 ClearPagePrivate(page);
719 set_page_private(page, 0);
b20a3503
CL
720
721 /*
722 * If any waiters have accumulated on the new page then
723 * wake them up.
724 */
725 if (PageWriteback(newpage))
726 end_page_writeback(newpage);
d435edca
VB
727
728 copy_page_owner(page, newpage);
74485cf2
JW
729
730 mem_cgroup_migrate(page, newpage);
b20a3503 731}
2916ecc0
JG
732EXPORT_SYMBOL(migrate_page_states);
733
734void migrate_page_copy(struct page *newpage, struct page *page)
735{
736 if (PageHuge(page) || PageTransHuge(page))
737 copy_huge_page(newpage, page);
738 else
739 copy_highpage(newpage, page);
740
741 migrate_page_states(newpage, page);
742}
1118dce7 743EXPORT_SYMBOL(migrate_page_copy);
b20a3503 744
1d8b85cc
CL
745/************************************************************
746 * Migration functions
747 ***********************************************************/
748
b20a3503 749/*
bda807d4 750 * Common logic to directly migrate a single LRU page suitable for
266cf658 751 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
752 *
753 * Pages are locked upon entry and exit.
754 */
2d1db3b1 755int migrate_page(struct address_space *mapping,
a6bc32b8
MG
756 struct page *newpage, struct page *page,
757 enum migrate_mode mode)
b20a3503
CL
758{
759 int rc;
760
761 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
762
8e321fef 763 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
b20a3503 764
78bd5209 765 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
766 return rc;
767
2916ecc0
JG
768 if (mode != MIGRATE_SYNC_NO_COPY)
769 migrate_page_copy(newpage, page);
770 else
771 migrate_page_states(newpage, page);
78bd5209 772 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
773}
774EXPORT_SYMBOL(migrate_page);
775
9361401e 776#ifdef CONFIG_BLOCK
1d8b85cc
CL
777/*
778 * Migration function for pages with buffers. This function can only be used
779 * if the underlying filesystem guarantees that no other references to "page"
780 * exist.
781 */
2d1db3b1 782int buffer_migrate_page(struct address_space *mapping,
a6bc32b8 783 struct page *newpage, struct page *page, enum migrate_mode mode)
1d8b85cc 784{
1d8b85cc
CL
785 struct buffer_head *bh, *head;
786 int rc;
787
1d8b85cc 788 if (!page_has_buffers(page))
a6bc32b8 789 return migrate_page(mapping, newpage, page, mode);
1d8b85cc
CL
790
791 head = page_buffers(page);
792
8e321fef 793 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
1d8b85cc 794
78bd5209 795 if (rc != MIGRATEPAGE_SUCCESS)
1d8b85cc
CL
796 return rc;
797
b969c4ab
MG
798 /*
799 * In the async case, migrate_page_move_mapping locked the buffers
800 * with an IRQ-safe spinlock held. In the sync case, the buffers
801 * need to be locked now
802 */
a6bc32b8
MG
803 if (mode != MIGRATE_ASYNC)
804 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
1d8b85cc
CL
805
806 ClearPagePrivate(page);
807 set_page_private(newpage, page_private(page));
808 set_page_private(page, 0);
809 put_page(page);
810 get_page(newpage);
811
812 bh = head;
813 do {
814 set_bh_page(bh, newpage, bh_offset(bh));
815 bh = bh->b_this_page;
816
817 } while (bh != head);
818
819 SetPagePrivate(newpage);
820
2916ecc0
JG
821 if (mode != MIGRATE_SYNC_NO_COPY)
822 migrate_page_copy(newpage, page);
823 else
824 migrate_page_states(newpage, page);
1d8b85cc
CL
825
826 bh = head;
827 do {
828 unlock_buffer(bh);
2916ecc0 829 put_bh(bh);
1d8b85cc
CL
830 bh = bh->b_this_page;
831
832 } while (bh != head);
833
78bd5209 834 return MIGRATEPAGE_SUCCESS;
1d8b85cc
CL
835}
836EXPORT_SYMBOL(buffer_migrate_page);
9361401e 837#endif
1d8b85cc 838
04e62a29
CL
839/*
840 * Writeback a page to clean the dirty state
841 */
842static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 843{
04e62a29
CL
844 struct writeback_control wbc = {
845 .sync_mode = WB_SYNC_NONE,
846 .nr_to_write = 1,
847 .range_start = 0,
848 .range_end = LLONG_MAX,
04e62a29
CL
849 .for_reclaim = 1
850 };
851 int rc;
852
853 if (!mapping->a_ops->writepage)
854 /* No write method for the address space */
855 return -EINVAL;
856
857 if (!clear_page_dirty_for_io(page))
858 /* Someone else already triggered a write */
859 return -EAGAIN;
860
8351a6e4 861 /*
04e62a29
CL
862 * A dirty page may imply that the underlying filesystem has
863 * the page on some queue. So the page must be clean for
864 * migration. Writeout may mean we loose the lock and the
865 * page state is no longer what we checked for earlier.
866 * At this point we know that the migration attempt cannot
867 * be successful.
8351a6e4 868 */
e388466d 869 remove_migration_ptes(page, page, false);
8351a6e4 870
04e62a29 871 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 872
04e62a29
CL
873 if (rc != AOP_WRITEPAGE_ACTIVATE)
874 /* unlocked. Relock */
875 lock_page(page);
876
bda8550d 877 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
878}
879
880/*
881 * Default handling if a filesystem does not provide a migration function.
882 */
883static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 884 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 885{
b969c4ab 886 if (PageDirty(page)) {
a6bc32b8 887 /* Only writeback pages in full synchronous migration */
2916ecc0
JG
888 switch (mode) {
889 case MIGRATE_SYNC:
890 case MIGRATE_SYNC_NO_COPY:
891 break;
892 default:
b969c4ab 893 return -EBUSY;
2916ecc0 894 }
04e62a29 895 return writeout(mapping, page);
b969c4ab 896 }
8351a6e4
CL
897
898 /*
899 * Buffers may be managed in a filesystem specific way.
900 * We must have no buffers or drop them.
901 */
266cf658 902 if (page_has_private(page) &&
8351a6e4
CL
903 !try_to_release_page(page, GFP_KERNEL))
904 return -EAGAIN;
905
a6bc32b8 906 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
907}
908
e24f0b8f
CL
909/*
910 * Move a page to a newly allocated page
911 * The page is locked and all ptes have been successfully removed.
912 *
913 * The new page will have replaced the old page if this function
914 * is successful.
894bc310
LS
915 *
916 * Return value:
917 * < 0 - error code
78bd5209 918 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 919 */
3fe2011f 920static int move_to_new_page(struct page *newpage, struct page *page,
5c3f9a67 921 enum migrate_mode mode)
e24f0b8f
CL
922{
923 struct address_space *mapping;
bda807d4
MK
924 int rc = -EAGAIN;
925 bool is_lru = !__PageMovable(page);
e24f0b8f 926
7db7671f
HD
927 VM_BUG_ON_PAGE(!PageLocked(page), page);
928 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
e24f0b8f 929
e24f0b8f 930 mapping = page_mapping(page);
bda807d4
MK
931
932 if (likely(is_lru)) {
933 if (!mapping)
934 rc = migrate_page(mapping, newpage, page, mode);
935 else if (mapping->a_ops->migratepage)
936 /*
937 * Most pages have a mapping and most filesystems
938 * provide a migratepage callback. Anonymous pages
939 * are part of swap space which also has its own
940 * migratepage callback. This is the most common path
941 * for page migration.
942 */
943 rc = mapping->a_ops->migratepage(mapping, newpage,
944 page, mode);
945 else
946 rc = fallback_migrate_page(mapping, newpage,
947 page, mode);
948 } else {
e24f0b8f 949 /*
bda807d4
MK
950 * In case of non-lru page, it could be released after
951 * isolation step. In that case, we shouldn't try migration.
e24f0b8f 952 */
bda807d4
MK
953 VM_BUG_ON_PAGE(!PageIsolated(page), page);
954 if (!PageMovable(page)) {
955 rc = MIGRATEPAGE_SUCCESS;
956 __ClearPageIsolated(page);
957 goto out;
958 }
959
960 rc = mapping->a_ops->migratepage(mapping, newpage,
961 page, mode);
962 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
963 !PageIsolated(page));
964 }
e24f0b8f 965
5c3f9a67
HD
966 /*
967 * When successful, old pagecache page->mapping must be cleared before
968 * page is freed; but stats require that PageAnon be left as PageAnon.
969 */
970 if (rc == MIGRATEPAGE_SUCCESS) {
bda807d4
MK
971 if (__PageMovable(page)) {
972 VM_BUG_ON_PAGE(!PageIsolated(page), page);
973
974 /*
975 * We clear PG_movable under page_lock so any compactor
976 * cannot try to migrate this page.
977 */
978 __ClearPageIsolated(page);
979 }
980
981 /*
982 * Anonymous and movable page->mapping will be cleard by
983 * free_pages_prepare so don't reset it here for keeping
984 * the type to work PageAnon, for example.
985 */
986 if (!PageMappingFlags(page))
5c3f9a67 987 page->mapping = NULL;
3fe2011f 988 }
bda807d4 989out:
e24f0b8f
CL
990 return rc;
991}
992
0dabec93 993static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 994 int force, enum migrate_mode mode)
e24f0b8f 995{
0dabec93 996 int rc = -EAGAIN;
2ebba6b7 997 int page_was_mapped = 0;
3f6c8272 998 struct anon_vma *anon_vma = NULL;
bda807d4 999 bool is_lru = !__PageMovable(page);
95a402c3 1000
529ae9aa 1001 if (!trylock_page(page)) {
a6bc32b8 1002 if (!force || mode == MIGRATE_ASYNC)
0dabec93 1003 goto out;
3e7d3449
MG
1004
1005 /*
1006 * It's not safe for direct compaction to call lock_page.
1007 * For example, during page readahead pages are added locked
1008 * to the LRU. Later, when the IO completes the pages are
1009 * marked uptodate and unlocked. However, the queueing
1010 * could be merging multiple pages for one bio (e.g.
1011 * mpage_readpages). If an allocation happens for the
1012 * second or third page, the process can end up locking
1013 * the same page twice and deadlocking. Rather than
1014 * trying to be clever about what pages can be locked,
1015 * avoid the use of lock_page for direct compaction
1016 * altogether.
1017 */
1018 if (current->flags & PF_MEMALLOC)
0dabec93 1019 goto out;
3e7d3449 1020
e24f0b8f
CL
1021 lock_page(page);
1022 }
1023
1024 if (PageWriteback(page)) {
11bc82d6 1025 /*
fed5b64a 1026 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
1027 * necessary to wait for PageWriteback. In the async case,
1028 * the retry loop is too short and in the sync-light case,
1029 * the overhead of stalling is too much
11bc82d6 1030 */
2916ecc0
JG
1031 switch (mode) {
1032 case MIGRATE_SYNC:
1033 case MIGRATE_SYNC_NO_COPY:
1034 break;
1035 default:
11bc82d6 1036 rc = -EBUSY;
0a31bc97 1037 goto out_unlock;
11bc82d6
AA
1038 }
1039 if (!force)
0a31bc97 1040 goto out_unlock;
e24f0b8f
CL
1041 wait_on_page_writeback(page);
1042 }
03f15c86 1043
e24f0b8f 1044 /*
dc386d4d
KH
1045 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1046 * we cannot notice that anon_vma is freed while we migrates a page.
1ce82b69 1047 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 1048 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
1049 * File Caches may use write_page() or lock_page() in migration, then,
1050 * just care Anon page here.
03f15c86
HD
1051 *
1052 * Only page_get_anon_vma() understands the subtleties of
1053 * getting a hold on an anon_vma from outside one of its mms.
1054 * But if we cannot get anon_vma, then we won't need it anyway,
1055 * because that implies that the anon page is no longer mapped
1056 * (and cannot be remapped so long as we hold the page lock).
dc386d4d 1057 */
03f15c86 1058 if (PageAnon(page) && !PageKsm(page))
746b18d4 1059 anon_vma = page_get_anon_vma(page);
62e1c553 1060
7db7671f
HD
1061 /*
1062 * Block others from accessing the new page when we get around to
1063 * establishing additional references. We are usually the only one
1064 * holding a reference to newpage at this point. We used to have a BUG
1065 * here if trylock_page(newpage) fails, but would like to allow for
1066 * cases where there might be a race with the previous use of newpage.
1067 * This is much like races on refcount of oldpage: just don't BUG().
1068 */
1069 if (unlikely(!trylock_page(newpage)))
1070 goto out_unlock;
1071
bda807d4
MK
1072 if (unlikely(!is_lru)) {
1073 rc = move_to_new_page(newpage, page, mode);
1074 goto out_unlock_both;
1075 }
1076
dc386d4d 1077 /*
62e1c553
SL
1078 * Corner case handling:
1079 * 1. When a new swap-cache page is read into, it is added to the LRU
1080 * and treated as swapcache but it has no rmap yet.
1081 * Calling try_to_unmap() against a page->mapping==NULL page will
1082 * trigger a BUG. So handle it here.
1083 * 2. An orphaned page (see truncate_complete_page) might have
1084 * fs-private metadata. The page can be picked up due to memory
1085 * offlining. Everywhere else except page reclaim, the page is
1086 * invisible to the vm, so the page can not be migrated. So try to
1087 * free the metadata, so the page can be freed.
e24f0b8f 1088 */
62e1c553 1089 if (!page->mapping) {
309381fe 1090 VM_BUG_ON_PAGE(PageAnon(page), page);
1ce82b69 1091 if (page_has_private(page)) {
62e1c553 1092 try_to_free_buffers(page);
7db7671f 1093 goto out_unlock_both;
62e1c553 1094 }
7db7671f
HD
1095 } else if (page_mapped(page)) {
1096 /* Establish migration ptes */
03f15c86
HD
1097 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1098 page);
2ebba6b7 1099 try_to_unmap(page,
da1b13cc 1100 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
2ebba6b7
HD
1101 page_was_mapped = 1;
1102 }
dc386d4d 1103
e6a1530d 1104 if (!page_mapped(page))
5c3f9a67 1105 rc = move_to_new_page(newpage, page, mode);
e24f0b8f 1106
5c3f9a67
HD
1107 if (page_was_mapped)
1108 remove_migration_ptes(page,
e388466d 1109 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
3f6c8272 1110
7db7671f
HD
1111out_unlock_both:
1112 unlock_page(newpage);
1113out_unlock:
3f6c8272 1114 /* Drop an anon_vma reference if we took one */
76545066 1115 if (anon_vma)
9e60109f 1116 put_anon_vma(anon_vma);
e24f0b8f 1117 unlock_page(page);
0dabec93 1118out:
c6c919eb
MK
1119 /*
1120 * If migration is successful, decrease refcount of the newpage
1121 * which will not free the page because new page owner increased
1122 * refcounter. As well, if it is LRU page, add the page to LRU
1123 * list in here.
1124 */
1125 if (rc == MIGRATEPAGE_SUCCESS) {
b1123ea6 1126 if (unlikely(__PageMovable(newpage)))
c6c919eb
MK
1127 put_page(newpage);
1128 else
1129 putback_lru_page(newpage);
1130 }
1131
0dabec93
MK
1132 return rc;
1133}
95a402c3 1134
ef2a5153
GU
1135/*
1136 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1137 * around it.
1138 */
815f0ddb
ND
1139#if defined(CONFIG_ARM) && \
1140 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
ef2a5153
GU
1141#define ICE_noinline noinline
1142#else
1143#define ICE_noinline
1144#endif
1145
0dabec93
MK
1146/*
1147 * Obtain the lock on page, remove all ptes and migrate the page
1148 * to the newly allocated page in newpage.
1149 */
ef2a5153
GU
1150static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1151 free_page_t put_new_page,
1152 unsigned long private, struct page *page,
add05cec
NH
1153 int force, enum migrate_mode mode,
1154 enum migrate_reason reason)
0dabec93 1155{
2def7424 1156 int rc = MIGRATEPAGE_SUCCESS;
2def7424 1157 struct page *newpage;
0dabec93 1158
94723aaf
MH
1159 if (!thp_migration_supported() && PageTransHuge(page))
1160 return -ENOMEM;
1161
666feb21 1162 newpage = get_new_page(page, private);
0dabec93
MK
1163 if (!newpage)
1164 return -ENOMEM;
1165
1166 if (page_count(page) == 1) {
1167 /* page was freed from under us. So we are done. */
c6c919eb
MK
1168 ClearPageActive(page);
1169 ClearPageUnevictable(page);
bda807d4
MK
1170 if (unlikely(__PageMovable(page))) {
1171 lock_page(page);
1172 if (!PageMovable(page))
1173 __ClearPageIsolated(page);
1174 unlock_page(page);
1175 }
c6c919eb
MK
1176 if (put_new_page)
1177 put_new_page(newpage, private);
1178 else
1179 put_page(newpage);
0dabec93
MK
1180 goto out;
1181 }
1182
9c620e2b 1183 rc = __unmap_and_move(page, newpage, force, mode);
c6c919eb 1184 if (rc == MIGRATEPAGE_SUCCESS)
7cd12b4a 1185 set_page_owner_migrate_reason(newpage, reason);
bf6bddf1 1186
0dabec93 1187out:
e24f0b8f 1188 if (rc != -EAGAIN) {
0dabec93
MK
1189 /*
1190 * A page that has been migrated has all references
1191 * removed and will be freed. A page that has not been
1192 * migrated will have kepts its references and be
1193 * restored.
1194 */
1195 list_del(&page->lru);
6afcf8ef
ML
1196
1197 /*
1198 * Compaction can migrate also non-LRU pages which are
1199 * not accounted to NR_ISOLATED_*. They can be recognized
1200 * as __PageMovable
1201 */
1202 if (likely(!__PageMovable(page)))
e8db67eb
NH
1203 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1204 page_is_file_cache(page), -hpage_nr_pages(page));
c6c919eb
MK
1205 }
1206
1207 /*
1208 * If migration is successful, releases reference grabbed during
1209 * isolation. Otherwise, restore the page to right list unless
1210 * we want to retry.
1211 */
1212 if (rc == MIGRATEPAGE_SUCCESS) {
1213 put_page(page);
1214 if (reason == MR_MEMORY_FAILURE) {
d7e69488 1215 /*
c6c919eb
MK
1216 * Set PG_HWPoison on just freed page
1217 * intentionally. Although it's rather weird,
1218 * it's how HWPoison flag works at the moment.
d7e69488 1219 */
d4ae9916 1220 if (set_hwpoison_free_buddy_page(page))
da1b13cc 1221 num_poisoned_pages_inc();
c6c919eb
MK
1222 }
1223 } else {
bda807d4
MK
1224 if (rc != -EAGAIN) {
1225 if (likely(!__PageMovable(page))) {
1226 putback_lru_page(page);
1227 goto put_new;
1228 }
1229
1230 lock_page(page);
1231 if (PageMovable(page))
1232 putback_movable_page(page);
1233 else
1234 __ClearPageIsolated(page);
1235 unlock_page(page);
1236 put_page(page);
1237 }
1238put_new:
c6c919eb
MK
1239 if (put_new_page)
1240 put_new_page(newpage, private);
1241 else
1242 put_page(newpage);
e24f0b8f 1243 }
68711a74 1244
e24f0b8f
CL
1245 return rc;
1246}
1247
290408d4
NH
1248/*
1249 * Counterpart of unmap_and_move_page() for hugepage migration.
1250 *
1251 * This function doesn't wait the completion of hugepage I/O
1252 * because there is no race between I/O and migration for hugepage.
1253 * Note that currently hugepage I/O occurs only in direct I/O
1254 * where no lock is held and PG_writeback is irrelevant,
1255 * and writeback status of all subpages are counted in the reference
1256 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1257 * under direct I/O, the reference of the head page is 512 and a bit more.)
1258 * This means that when we try to migrate hugepage whose subpages are
1259 * doing direct I/O, some references remain after try_to_unmap() and
1260 * hugepage migration fails without data corruption.
1261 *
1262 * There is also no race when direct I/O is issued on the page under migration,
1263 * because then pte is replaced with migration swap entry and direct I/O code
1264 * will wait in the page fault for migration to complete.
1265 */
1266static int unmap_and_move_huge_page(new_page_t get_new_page,
68711a74
DR
1267 free_page_t put_new_page, unsigned long private,
1268 struct page *hpage, int force,
7cd12b4a 1269 enum migrate_mode mode, int reason)
290408d4 1270{
2def7424 1271 int rc = -EAGAIN;
2ebba6b7 1272 int page_was_mapped = 0;
32665f2b 1273 struct page *new_hpage;
290408d4
NH
1274 struct anon_vma *anon_vma = NULL;
1275
83467efb
NH
1276 /*
1277 * Movability of hugepages depends on architectures and hugepage size.
1278 * This check is necessary because some callers of hugepage migration
1279 * like soft offline and memory hotremove don't walk through page
1280 * tables or check whether the hugepage is pmd-based or not before
1281 * kicking migration.
1282 */
100873d7 1283 if (!hugepage_migration_supported(page_hstate(hpage))) {
32665f2b 1284 putback_active_hugepage(hpage);
83467efb 1285 return -ENOSYS;
32665f2b 1286 }
83467efb 1287
666feb21 1288 new_hpage = get_new_page(hpage, private);
290408d4
NH
1289 if (!new_hpage)
1290 return -ENOMEM;
1291
290408d4 1292 if (!trylock_page(hpage)) {
2916ecc0 1293 if (!force)
290408d4 1294 goto out;
2916ecc0
JG
1295 switch (mode) {
1296 case MIGRATE_SYNC:
1297 case MIGRATE_SYNC_NO_COPY:
1298 break;
1299 default:
1300 goto out;
1301 }
290408d4
NH
1302 lock_page(hpage);
1303 }
1304
746b18d4
PZ
1305 if (PageAnon(hpage))
1306 anon_vma = page_get_anon_vma(hpage);
290408d4 1307
7db7671f
HD
1308 if (unlikely(!trylock_page(new_hpage)))
1309 goto put_anon;
1310
2ebba6b7
HD
1311 if (page_mapped(hpage)) {
1312 try_to_unmap(hpage,
1313 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1314 page_was_mapped = 1;
1315 }
290408d4
NH
1316
1317 if (!page_mapped(hpage))
5c3f9a67 1318 rc = move_to_new_page(new_hpage, hpage, mode);
290408d4 1319
5c3f9a67
HD
1320 if (page_was_mapped)
1321 remove_migration_ptes(hpage,
e388466d 1322 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
290408d4 1323
7db7671f
HD
1324 unlock_page(new_hpage);
1325
1326put_anon:
fd4a4663 1327 if (anon_vma)
9e60109f 1328 put_anon_vma(anon_vma);
8e6ac7fa 1329
2def7424 1330 if (rc == MIGRATEPAGE_SUCCESS) {
ab5ac90a 1331 move_hugetlb_state(hpage, new_hpage, reason);
2def7424
HD
1332 put_new_page = NULL;
1333 }
8e6ac7fa 1334
290408d4 1335 unlock_page(hpage);
09761333 1336out:
b8ec1cee
NH
1337 if (rc != -EAGAIN)
1338 putback_active_hugepage(hpage);
68711a74
DR
1339
1340 /*
1341 * If migration was not successful and there's a freeing callback, use
1342 * it. Otherwise, put_page() will drop the reference grabbed during
1343 * isolation.
1344 */
2def7424 1345 if (put_new_page)
68711a74
DR
1346 put_new_page(new_hpage, private);
1347 else
3aaa76e1 1348 putback_active_hugepage(new_hpage);
68711a74 1349
290408d4
NH
1350 return rc;
1351}
1352
b20a3503 1353/*
c73e5c9c
SB
1354 * migrate_pages - migrate the pages specified in a list, to the free pages
1355 * supplied as the target for the page migration
b20a3503 1356 *
c73e5c9c
SB
1357 * @from: The list of pages to be migrated.
1358 * @get_new_page: The function used to allocate free pages to be used
1359 * as the target of the page migration.
68711a74
DR
1360 * @put_new_page: The function used to free target pages if migration
1361 * fails, or NULL if no special handling is necessary.
c73e5c9c
SB
1362 * @private: Private data to be passed on to get_new_page()
1363 * @mode: The migration mode that specifies the constraints for
1364 * page migration, if any.
1365 * @reason: The reason for page migration.
b20a3503 1366 *
c73e5c9c
SB
1367 * The function returns after 10 attempts or if no pages are movable any more
1368 * because the list has become empty or no retryable pages exist any more.
14e0f9bc 1369 * The caller should call putback_movable_pages() to return pages to the LRU
28bd6578 1370 * or free list only if ret != 0.
b20a3503 1371 *
c73e5c9c 1372 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1373 */
9c620e2b 1374int migrate_pages(struct list_head *from, new_page_t get_new_page,
68711a74
DR
1375 free_page_t put_new_page, unsigned long private,
1376 enum migrate_mode mode, int reason)
b20a3503 1377{
e24f0b8f 1378 int retry = 1;
b20a3503 1379 int nr_failed = 0;
5647bc29 1380 int nr_succeeded = 0;
b20a3503
CL
1381 int pass = 0;
1382 struct page *page;
1383 struct page *page2;
1384 int swapwrite = current->flags & PF_SWAPWRITE;
1385 int rc;
1386
1387 if (!swapwrite)
1388 current->flags |= PF_SWAPWRITE;
1389
e24f0b8f
CL
1390 for(pass = 0; pass < 10 && retry; pass++) {
1391 retry = 0;
b20a3503 1392
e24f0b8f 1393 list_for_each_entry_safe(page, page2, from, lru) {
94723aaf 1394retry:
e24f0b8f 1395 cond_resched();
2d1db3b1 1396
31caf665
NH
1397 if (PageHuge(page))
1398 rc = unmap_and_move_huge_page(get_new_page,
68711a74 1399 put_new_page, private, page,
7cd12b4a 1400 pass > 2, mode, reason);
31caf665 1401 else
68711a74 1402 rc = unmap_and_move(get_new_page, put_new_page,
add05cec
NH
1403 private, page, pass > 2, mode,
1404 reason);
2d1db3b1 1405
e24f0b8f 1406 switch(rc) {
95a402c3 1407 case -ENOMEM:
94723aaf
MH
1408 /*
1409 * THP migration might be unsupported or the
1410 * allocation could've failed so we should
1411 * retry on the same page with the THP split
1412 * to base pages.
1413 *
1414 * Head page is retried immediately and tail
1415 * pages are added to the tail of the list so
1416 * we encounter them after the rest of the list
1417 * is processed.
1418 */
e6112fc3 1419 if (PageTransHuge(page) && !PageHuge(page)) {
94723aaf
MH
1420 lock_page(page);
1421 rc = split_huge_page_to_list(page, from);
1422 unlock_page(page);
1423 if (!rc) {
1424 list_safe_reset_next(page, page2, lru);
1425 goto retry;
1426 }
1427 }
dfef2ef4 1428 nr_failed++;
95a402c3 1429 goto out;
e24f0b8f 1430 case -EAGAIN:
2d1db3b1 1431 retry++;
e24f0b8f 1432 break;
78bd5209 1433 case MIGRATEPAGE_SUCCESS:
5647bc29 1434 nr_succeeded++;
e24f0b8f
CL
1435 break;
1436 default:
354a3363
NH
1437 /*
1438 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1439 * unlike -EAGAIN case, the failed page is
1440 * removed from migration page list and not
1441 * retried in the next outer loop.
1442 */
2d1db3b1 1443 nr_failed++;
e24f0b8f 1444 break;
2d1db3b1 1445 }
b20a3503
CL
1446 }
1447 }
f2f81fb2
VB
1448 nr_failed += retry;
1449 rc = nr_failed;
95a402c3 1450out:
5647bc29
MG
1451 if (nr_succeeded)
1452 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1453 if (nr_failed)
1454 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1455 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1456
b20a3503
CL
1457 if (!swapwrite)
1458 current->flags &= ~PF_SWAPWRITE;
1459
78bd5209 1460 return rc;
b20a3503 1461}
95a402c3 1462
742755a1 1463#ifdef CONFIG_NUMA
742755a1 1464
a49bd4d7 1465static int store_status(int __user *status, int start, int value, int nr)
742755a1 1466{
a49bd4d7
MH
1467 while (nr-- > 0) {
1468 if (put_user(value, status + start))
1469 return -EFAULT;
1470 start++;
1471 }
1472
1473 return 0;
1474}
1475
1476static int do_move_pages_to_node(struct mm_struct *mm,
1477 struct list_head *pagelist, int node)
1478{
1479 int err;
1480
1481 if (list_empty(pagelist))
1482 return 0;
1483
1484 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1485 MIGRATE_SYNC, MR_SYSCALL);
1486 if (err)
1487 putback_movable_pages(pagelist);
1488 return err;
742755a1
CL
1489}
1490
1491/*
a49bd4d7
MH
1492 * Resolves the given address to a struct page, isolates it from the LRU and
1493 * puts it to the given pagelist.
1494 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1495 * queued or the page doesn't need to be migrated because it is already on
1496 * the target node
742755a1 1497 */
a49bd4d7
MH
1498static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1499 int node, struct list_head *pagelist, bool migrate_all)
742755a1 1500{
a49bd4d7
MH
1501 struct vm_area_struct *vma;
1502 struct page *page;
1503 unsigned int follflags;
742755a1 1504 int err;
742755a1
CL
1505
1506 down_read(&mm->mmap_sem);
a49bd4d7
MH
1507 err = -EFAULT;
1508 vma = find_vma(mm, addr);
1509 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1510 goto out;
742755a1 1511
a49bd4d7
MH
1512 /* FOLL_DUMP to ignore special (like zero) pages */
1513 follflags = FOLL_GET | FOLL_DUMP;
a49bd4d7 1514 page = follow_page(vma, addr, follflags);
89f5b7da 1515
a49bd4d7
MH
1516 err = PTR_ERR(page);
1517 if (IS_ERR(page))
1518 goto out;
89f5b7da 1519
a49bd4d7
MH
1520 err = -ENOENT;
1521 if (!page)
1522 goto out;
742755a1 1523
a49bd4d7
MH
1524 err = 0;
1525 if (page_to_nid(page) == node)
1526 goto out_putpage;
742755a1 1527
a49bd4d7
MH
1528 err = -EACCES;
1529 if (page_mapcount(page) > 1 && !migrate_all)
1530 goto out_putpage;
742755a1 1531
a49bd4d7
MH
1532 if (PageHuge(page)) {
1533 if (PageHead(page)) {
1534 isolate_huge_page(page, pagelist);
1535 err = 0;
e632a938 1536 }
a49bd4d7
MH
1537 } else {
1538 struct page *head;
e632a938 1539
e8db67eb
NH
1540 head = compound_head(page);
1541 err = isolate_lru_page(head);
cf608ac1 1542 if (err)
a49bd4d7 1543 goto out_putpage;
742755a1 1544
a49bd4d7
MH
1545 err = 0;
1546 list_add_tail(&head->lru, pagelist);
1547 mod_node_page_state(page_pgdat(head),
1548 NR_ISOLATED_ANON + page_is_file_cache(head),
1549 hpage_nr_pages(head));
1550 }
1551out_putpage:
1552 /*
1553 * Either remove the duplicate refcount from
1554 * isolate_lru_page() or drop the page ref if it was
1555 * not isolated.
1556 */
1557 put_page(page);
1558out:
742755a1
CL
1559 up_read(&mm->mmap_sem);
1560 return err;
1561}
1562
5e9a0f02
BG
1563/*
1564 * Migrate an array of page address onto an array of nodes and fill
1565 * the corresponding array of status.
1566 */
3268c63e 1567static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1568 unsigned long nr_pages,
1569 const void __user * __user *pages,
1570 const int __user *nodes,
1571 int __user *status, int flags)
1572{
a49bd4d7
MH
1573 int current_node = NUMA_NO_NODE;
1574 LIST_HEAD(pagelist);
1575 int start, i;
1576 int err = 0, err1;
35282a2d
BG
1577
1578 migrate_prep();
1579
a49bd4d7
MH
1580 for (i = start = 0; i < nr_pages; i++) {
1581 const void __user *p;
1582 unsigned long addr;
1583 int node;
3140a227 1584
a49bd4d7
MH
1585 err = -EFAULT;
1586 if (get_user(p, pages + i))
1587 goto out_flush;
1588 if (get_user(node, nodes + i))
1589 goto out_flush;
1590 addr = (unsigned long)p;
1591
1592 err = -ENODEV;
1593 if (node < 0 || node >= MAX_NUMNODES)
1594 goto out_flush;
1595 if (!node_state(node, N_MEMORY))
1596 goto out_flush;
5e9a0f02 1597
a49bd4d7
MH
1598 err = -EACCES;
1599 if (!node_isset(node, task_nodes))
1600 goto out_flush;
1601
1602 if (current_node == NUMA_NO_NODE) {
1603 current_node = node;
1604 start = i;
1605 } else if (node != current_node) {
1606 err = do_move_pages_to_node(mm, &pagelist, current_node);
1607 if (err)
1608 goto out;
1609 err = store_status(status, start, current_node, i - start);
1610 if (err)
1611 goto out;
1612 start = i;
1613 current_node = node;
3140a227
BG
1614 }
1615
a49bd4d7
MH
1616 /*
1617 * Errors in the page lookup or isolation are not fatal and we simply
1618 * report them via status
1619 */
1620 err = add_page_for_migration(mm, addr, current_node,
1621 &pagelist, flags & MPOL_MF_MOVE_ALL);
1622 if (!err)
1623 continue;
3140a227 1624
a49bd4d7
MH
1625 err = store_status(status, i, err, 1);
1626 if (err)
1627 goto out_flush;
5e9a0f02 1628
a49bd4d7
MH
1629 err = do_move_pages_to_node(mm, &pagelist, current_node);
1630 if (err)
1631 goto out;
1632 if (i > start) {
1633 err = store_status(status, start, current_node, i - start);
1634 if (err)
1635 goto out;
1636 }
1637 current_node = NUMA_NO_NODE;
3140a227 1638 }
a49bd4d7 1639out_flush:
8f175cf5
MH
1640 if (list_empty(&pagelist))
1641 return err;
1642
a49bd4d7
MH
1643 /* Make sure we do not overwrite the existing error */
1644 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1645 if (!err1)
1646 err1 = store_status(status, start, current_node, i - start);
1647 if (!err)
1648 err = err1;
5e9a0f02
BG
1649out:
1650 return err;
1651}
1652
742755a1 1653/*
2f007e74 1654 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1655 */
80bba129
BG
1656static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1657 const void __user **pages, int *status)
742755a1 1658{
2f007e74 1659 unsigned long i;
2f007e74 1660
742755a1
CL
1661 down_read(&mm->mmap_sem);
1662
2f007e74 1663 for (i = 0; i < nr_pages; i++) {
80bba129 1664 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1665 struct vm_area_struct *vma;
1666 struct page *page;
c095adbc 1667 int err = -EFAULT;
2f007e74
BG
1668
1669 vma = find_vma(mm, addr);
70384dc6 1670 if (!vma || addr < vma->vm_start)
742755a1
CL
1671 goto set_status;
1672
d899844e
KS
1673 /* FOLL_DUMP to ignore special (like zero) pages */
1674 page = follow_page(vma, addr, FOLL_DUMP);
89f5b7da
LT
1675
1676 err = PTR_ERR(page);
1677 if (IS_ERR(page))
1678 goto set_status;
1679
d899844e 1680 err = page ? page_to_nid(page) : -ENOENT;
742755a1 1681set_status:
80bba129
BG
1682 *status = err;
1683
1684 pages++;
1685 status++;
1686 }
1687
1688 up_read(&mm->mmap_sem);
1689}
1690
1691/*
1692 * Determine the nodes of a user array of pages and store it in
1693 * a user array of status.
1694 */
1695static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1696 const void __user * __user *pages,
1697 int __user *status)
1698{
1699#define DO_PAGES_STAT_CHUNK_NR 16
1700 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1701 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1702
87b8d1ad
PA
1703 while (nr_pages) {
1704 unsigned long chunk_nr;
80bba129 1705
87b8d1ad
PA
1706 chunk_nr = nr_pages;
1707 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1708 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1709
1710 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1711 break;
80bba129
BG
1712
1713 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1714
87b8d1ad
PA
1715 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1716 break;
742755a1 1717
87b8d1ad
PA
1718 pages += chunk_nr;
1719 status += chunk_nr;
1720 nr_pages -= chunk_nr;
1721 }
1722 return nr_pages ? -EFAULT : 0;
742755a1
CL
1723}
1724
1725/*
1726 * Move a list of pages in the address space of the currently executing
1727 * process.
1728 */
7addf443
DB
1729static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1730 const void __user * __user *pages,
1731 const int __user *nodes,
1732 int __user *status, int flags)
742755a1 1733{
742755a1 1734 struct task_struct *task;
742755a1 1735 struct mm_struct *mm;
5e9a0f02 1736 int err;
3268c63e 1737 nodemask_t task_nodes;
742755a1
CL
1738
1739 /* Check flags */
1740 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1741 return -EINVAL;
1742
1743 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1744 return -EPERM;
1745
1746 /* Find the mm_struct */
a879bf58 1747 rcu_read_lock();
228ebcbe 1748 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1749 if (!task) {
a879bf58 1750 rcu_read_unlock();
742755a1
CL
1751 return -ESRCH;
1752 }
3268c63e 1753 get_task_struct(task);
742755a1
CL
1754
1755 /*
1756 * Check if this process has the right to modify the specified
197e7e52 1757 * process. Use the regular "ptrace_may_access()" checks.
742755a1 1758 */
197e7e52 1759 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
c69e8d9c 1760 rcu_read_unlock();
742755a1 1761 err = -EPERM;
5e9a0f02 1762 goto out;
742755a1 1763 }
c69e8d9c 1764 rcu_read_unlock();
742755a1 1765
86c3a764
DQ
1766 err = security_task_movememory(task);
1767 if (err)
5e9a0f02 1768 goto out;
86c3a764 1769
3268c63e
CL
1770 task_nodes = cpuset_mems_allowed(task);
1771 mm = get_task_mm(task);
1772 put_task_struct(task);
1773
6e8b09ea
SL
1774 if (!mm)
1775 return -EINVAL;
1776
1777 if (nodes)
1778 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1779 nodes, status, flags);
1780 else
1781 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1782
742755a1
CL
1783 mmput(mm);
1784 return err;
3268c63e
CL
1785
1786out:
1787 put_task_struct(task);
1788 return err;
742755a1 1789}
742755a1 1790
7addf443
DB
1791SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1792 const void __user * __user *, pages,
1793 const int __user *, nodes,
1794 int __user *, status, int, flags)
1795{
1796 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1797}
1798
1799#ifdef CONFIG_COMPAT
1800COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1801 compat_uptr_t __user *, pages32,
1802 const int __user *, nodes,
1803 int __user *, status,
1804 int, flags)
1805{
1806 const void __user * __user *pages;
1807 int i;
1808
1809 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1810 for (i = 0; i < nr_pages; i++) {
1811 compat_uptr_t p;
1812
1813 if (get_user(p, pages32 + i) ||
1814 put_user(compat_ptr(p), pages + i))
1815 return -EFAULT;
1816 }
1817 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1818}
1819#endif /* CONFIG_COMPAT */
1820
7039e1db
PZ
1821#ifdef CONFIG_NUMA_BALANCING
1822/*
1823 * Returns true if this is a safe migration target node for misplaced NUMA
1824 * pages. Currently it only checks the watermarks which crude
1825 */
1826static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1827 unsigned long nr_migrate_pages)
7039e1db
PZ
1828{
1829 int z;
599d0c95 1830
7039e1db
PZ
1831 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1832 struct zone *zone = pgdat->node_zones + z;
1833
1834 if (!populated_zone(zone))
1835 continue;
1836
7039e1db
PZ
1837 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1838 if (!zone_watermark_ok(zone, 0,
1839 high_wmark_pages(zone) +
1840 nr_migrate_pages,
1841 0, 0))
1842 continue;
1843 return true;
1844 }
1845 return false;
1846}
1847
1848static struct page *alloc_misplaced_dst_page(struct page *page,
666feb21 1849 unsigned long data)
7039e1db
PZ
1850{
1851 int nid = (int) data;
1852 struct page *newpage;
1853
96db800f 1854 newpage = __alloc_pages_node(nid,
e97ca8e5
JW
1855 (GFP_HIGHUSER_MOVABLE |
1856 __GFP_THISNODE | __GFP_NOMEMALLOC |
1857 __GFP_NORETRY | __GFP_NOWARN) &
8479eba7 1858 ~__GFP_RECLAIM, 0);
bac0382c 1859
7039e1db
PZ
1860 return newpage;
1861}
1862
1c30e017 1863static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
b32967ff 1864{
340ef390 1865 int page_lru;
a8f60772 1866
309381fe 1867 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
3abef4e6 1868
7039e1db 1869 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1870 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1871 return 0;
7039e1db 1872
340ef390
HD
1873 if (isolate_lru_page(page))
1874 return 0;
7039e1db 1875
340ef390
HD
1876 /*
1877 * migrate_misplaced_transhuge_page() skips page migration's usual
1878 * check on page_count(), so we must do it here, now that the page
1879 * has been isolated: a GUP pin, or any other pin, prevents migration.
1880 * The expected page count is 3: 1 for page's mapcount and 1 for the
1881 * caller's pin and 1 for the reference taken by isolate_lru_page().
1882 */
1883 if (PageTransHuge(page) && page_count(page) != 3) {
1884 putback_lru_page(page);
1885 return 0;
7039e1db
PZ
1886 }
1887
340ef390 1888 page_lru = page_is_file_cache(page);
599d0c95 1889 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
340ef390
HD
1890 hpage_nr_pages(page));
1891
149c33e1 1892 /*
340ef390
HD
1893 * Isolating the page has taken another reference, so the
1894 * caller's reference can be safely dropped without the page
1895 * disappearing underneath us during migration.
149c33e1
MG
1896 */
1897 put_page(page);
340ef390 1898 return 1;
b32967ff
MG
1899}
1900
de466bd6
MG
1901bool pmd_trans_migrating(pmd_t pmd)
1902{
1903 struct page *page = pmd_page(pmd);
1904 return PageLocked(page);
1905}
1906
b32967ff
MG
1907/*
1908 * Attempt to migrate a misplaced page to the specified destination
1909 * node. Caller is expected to have an elevated reference count on
1910 * the page that will be dropped by this function before returning.
1911 */
1bc115d8
MG
1912int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1913 int node)
b32967ff
MG
1914{
1915 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1916 int isolated;
b32967ff
MG
1917 int nr_remaining;
1918 LIST_HEAD(migratepages);
1919
1920 /*
1bc115d8
MG
1921 * Don't migrate file pages that are mapped in multiple processes
1922 * with execute permissions as they are probably shared libraries.
b32967ff 1923 */
1bc115d8
MG
1924 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1925 (vma->vm_flags & VM_EXEC))
b32967ff 1926 goto out;
b32967ff 1927
09a913a7
MG
1928 /*
1929 * Also do not migrate dirty pages as not all filesystems can move
1930 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1931 */
1932 if (page_is_file_cache(page) && PageDirty(page))
1933 goto out;
1934
b32967ff
MG
1935 isolated = numamigrate_isolate_page(pgdat, page);
1936 if (!isolated)
1937 goto out;
1938
1939 list_add(&page->lru, &migratepages);
9c620e2b 1940 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
68711a74
DR
1941 NULL, node, MIGRATE_ASYNC,
1942 MR_NUMA_MISPLACED);
b32967ff 1943 if (nr_remaining) {
59c82b70
JK
1944 if (!list_empty(&migratepages)) {
1945 list_del(&page->lru);
599d0c95 1946 dec_node_page_state(page, NR_ISOLATED_ANON +
59c82b70
JK
1947 page_is_file_cache(page));
1948 putback_lru_page(page);
1949 }
b32967ff
MG
1950 isolated = 0;
1951 } else
1952 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1953 BUG_ON(!list_empty(&migratepages));
7039e1db 1954 return isolated;
340ef390
HD
1955
1956out:
1957 put_page(page);
1958 return 0;
7039e1db 1959}
220018d3 1960#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1961
220018d3 1962#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1963/*
1964 * Migrates a THP to a given target node. page must be locked and is unlocked
1965 * before returning.
1966 */
b32967ff
MG
1967int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1968 struct vm_area_struct *vma,
1969 pmd_t *pmd, pmd_t entry,
1970 unsigned long address,
1971 struct page *page, int node)
1972{
c4088ebd 1973 spinlock_t *ptl;
b32967ff
MG
1974 pg_data_t *pgdat = NODE_DATA(node);
1975 int isolated = 0;
1976 struct page *new_page = NULL;
b32967ff 1977 int page_lru = page_is_file_cache(page);
f714f4f2
MG
1978 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1979 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
b32967ff 1980
b32967ff 1981 new_page = alloc_pages_node(node,
25160354 1982 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
e97ca8e5 1983 HPAGE_PMD_ORDER);
340ef390
HD
1984 if (!new_page)
1985 goto out_fail;
9a982250 1986 prep_transhuge_page(new_page);
340ef390 1987
b32967ff 1988 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 1989 if (!isolated) {
b32967ff 1990 put_page(new_page);
340ef390 1991 goto out_fail;
b32967ff 1992 }
b0943d61 1993
b32967ff 1994 /* Prepare a page as a migration target */
48c935ad 1995 __SetPageLocked(new_page);
d44d363f
SL
1996 if (PageSwapBacked(page))
1997 __SetPageSwapBacked(new_page);
b32967ff
MG
1998
1999 /* anon mapping, we can simply copy page->mapping to the new page: */
2000 new_page->mapping = page->mapping;
2001 new_page->index = page->index;
2002 migrate_page_copy(new_page, page);
2003 WARN_ON(PageLRU(new_page));
2004
2005 /* Recheck the target PMD */
f714f4f2 2006 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2007 ptl = pmd_lock(mm, pmd);
f4e177d1 2008 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
c4088ebd 2009 spin_unlock(ptl);
f714f4f2 2010 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b32967ff
MG
2011
2012 /* Reverse changes made by migrate_page_copy() */
2013 if (TestClearPageActive(new_page))
2014 SetPageActive(page);
2015 if (TestClearPageUnevictable(new_page))
2016 SetPageUnevictable(page);
b32967ff
MG
2017
2018 unlock_page(new_page);
2019 put_page(new_page); /* Free it */
2020
a54a407f
MG
2021 /* Retake the callers reference and putback on LRU */
2022 get_page(page);
b32967ff 2023 putback_lru_page(page);
599d0c95 2024 mod_node_page_state(page_pgdat(page),
a54a407f 2025 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
eb4489f6
MG
2026
2027 goto out_unlock;
b32967ff
MG
2028 }
2029
10102459 2030 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
f55e1014 2031 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
b32967ff 2032
2b4847e7
MG
2033 /*
2034 * Clear the old entry under pagetable lock and establish the new PTE.
2035 * Any parallel GUP will either observe the old page blocking on the
2036 * page lock, block on the page table lock or observe the new page.
2037 * The SetPageUptodate on the new page and page_add_new_anon_rmap
2038 * guarantee the copy is visible before the pagetable update.
2039 */
f714f4f2 2040 flush_cache_range(vma, mmun_start, mmun_end);
d281ee61 2041 page_add_anon_rmap(new_page, vma, mmun_start, true);
8809aa2d 2042 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
f714f4f2 2043 set_pmd_at(mm, mmun_start, pmd, entry);
ce4a9cc5 2044 update_mmu_cache_pmd(vma, address, &entry);
2b4847e7 2045
f4e177d1 2046 page_ref_unfreeze(page, 2);
51afb12b 2047 mlock_migrate_page(new_page, page);
d281ee61 2048 page_remove_rmap(page, true);
7cd12b4a 2049 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2b4847e7 2050
c4088ebd 2051 spin_unlock(ptl);
4645b9fe
JG
2052 /*
2053 * No need to double call mmu_notifier->invalidate_range() callback as
2054 * the above pmdp_huge_clear_flush_notify() did already call it.
2055 */
2056 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
b32967ff 2057
11de9927
MG
2058 /* Take an "isolate" reference and put new page on the LRU. */
2059 get_page(new_page);
2060 putback_lru_page(new_page);
2061
b32967ff
MG
2062 unlock_page(new_page);
2063 unlock_page(page);
2064 put_page(page); /* Drop the rmap reference */
2065 put_page(page); /* Drop the LRU isolation reference */
2066
2067 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2068 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2069
599d0c95 2070 mod_node_page_state(page_pgdat(page),
b32967ff
MG
2071 NR_ISOLATED_ANON + page_lru,
2072 -HPAGE_PMD_NR);
2073 return isolated;
2074
340ef390
HD
2075out_fail:
2076 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2b4847e7
MG
2077 ptl = pmd_lock(mm, pmd);
2078 if (pmd_same(*pmd, entry)) {
4d942466 2079 entry = pmd_modify(entry, vma->vm_page_prot);
f714f4f2 2080 set_pmd_at(mm, mmun_start, pmd, entry);
2b4847e7
MG
2081 update_mmu_cache_pmd(vma, address, &entry);
2082 }
2083 spin_unlock(ptl);
a54a407f 2084
eb4489f6 2085out_unlock:
340ef390 2086 unlock_page(page);
b32967ff 2087 put_page(page);
b32967ff
MG
2088 return 0;
2089}
7039e1db
PZ
2090#endif /* CONFIG_NUMA_BALANCING */
2091
2092#endif /* CONFIG_NUMA */
8763cb45 2093
6b368cd4 2094#if defined(CONFIG_MIGRATE_VMA_HELPER)
8763cb45
JG
2095struct migrate_vma {
2096 struct vm_area_struct *vma;
2097 unsigned long *dst;
2098 unsigned long *src;
2099 unsigned long cpages;
2100 unsigned long npages;
2101 unsigned long start;
2102 unsigned long end;
2103};
2104
2105static int migrate_vma_collect_hole(unsigned long start,
2106 unsigned long end,
2107 struct mm_walk *walk)
2108{
2109 struct migrate_vma *migrate = walk->private;
2110 unsigned long addr;
2111
8315ada7 2112 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
e20d103b 2113 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
8315ada7 2114 migrate->dst[migrate->npages] = 0;
e20d103b 2115 migrate->npages++;
8315ada7
JG
2116 migrate->cpages++;
2117 }
2118
2119 return 0;
2120}
2121
2122static int migrate_vma_collect_skip(unsigned long start,
2123 unsigned long end,
2124 struct mm_walk *walk)
2125{
2126 struct migrate_vma *migrate = walk->private;
2127 unsigned long addr;
2128
8763cb45
JG
2129 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2130 migrate->dst[migrate->npages] = 0;
2131 migrate->src[migrate->npages++] = 0;
2132 }
2133
2134 return 0;
2135}
2136
2137static int migrate_vma_collect_pmd(pmd_t *pmdp,
2138 unsigned long start,
2139 unsigned long end,
2140 struct mm_walk *walk)
2141{
2142 struct migrate_vma *migrate = walk->private;
2143 struct vm_area_struct *vma = walk->vma;
2144 struct mm_struct *mm = vma->vm_mm;
8c3328f1 2145 unsigned long addr = start, unmapped = 0;
8763cb45
JG
2146 spinlock_t *ptl;
2147 pte_t *ptep;
2148
2149again:
2150 if (pmd_none(*pmdp))
2151 return migrate_vma_collect_hole(start, end, walk);
2152
2153 if (pmd_trans_huge(*pmdp)) {
2154 struct page *page;
2155
2156 ptl = pmd_lock(mm, pmdp);
2157 if (unlikely(!pmd_trans_huge(*pmdp))) {
2158 spin_unlock(ptl);
2159 goto again;
2160 }
2161
2162 page = pmd_page(*pmdp);
2163 if (is_huge_zero_page(page)) {
2164 spin_unlock(ptl);
2165 split_huge_pmd(vma, pmdp, addr);
2166 if (pmd_trans_unstable(pmdp))
8315ada7 2167 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2168 walk);
2169 } else {
2170 int ret;
2171
2172 get_page(page);
2173 spin_unlock(ptl);
2174 if (unlikely(!trylock_page(page)))
8315ada7 2175 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2176 walk);
2177 ret = split_huge_page(page);
2178 unlock_page(page);
2179 put_page(page);
8315ada7
JG
2180 if (ret)
2181 return migrate_vma_collect_skip(start, end,
2182 walk);
2183 if (pmd_none(*pmdp))
8763cb45
JG
2184 return migrate_vma_collect_hole(start, end,
2185 walk);
2186 }
2187 }
2188
2189 if (unlikely(pmd_bad(*pmdp)))
8315ada7 2190 return migrate_vma_collect_skip(start, end, walk);
8763cb45
JG
2191
2192 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
8c3328f1
JG
2193 arch_enter_lazy_mmu_mode();
2194
8763cb45
JG
2195 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2196 unsigned long mpfn, pfn;
2197 struct page *page;
8c3328f1 2198 swp_entry_t entry;
8763cb45
JG
2199 pte_t pte;
2200
2201 pte = *ptep;
2202 pfn = pte_pfn(pte);
2203
a5430dda 2204 if (pte_none(pte)) {
8315ada7
JG
2205 mpfn = MIGRATE_PFN_MIGRATE;
2206 migrate->cpages++;
2207 pfn = 0;
8763cb45
JG
2208 goto next;
2209 }
2210
a5430dda
JG
2211 if (!pte_present(pte)) {
2212 mpfn = pfn = 0;
2213
2214 /*
2215 * Only care about unaddressable device page special
2216 * page table entry. Other special swap entries are not
2217 * migratable, and we ignore regular swapped page.
2218 */
2219 entry = pte_to_swp_entry(pte);
2220 if (!is_device_private_entry(entry))
2221 goto next;
2222
2223 page = device_private_entry_to_page(entry);
2224 mpfn = migrate_pfn(page_to_pfn(page))|
2225 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2226 if (is_write_device_private_entry(entry))
2227 mpfn |= MIGRATE_PFN_WRITE;
2228 } else {
8315ada7
JG
2229 if (is_zero_pfn(pfn)) {
2230 mpfn = MIGRATE_PFN_MIGRATE;
2231 migrate->cpages++;
2232 pfn = 0;
2233 goto next;
2234 }
df6ad698 2235 page = _vm_normal_page(migrate->vma, addr, pte, true);
a5430dda
JG
2236 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2237 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2238 }
2239
8763cb45 2240 /* FIXME support THP */
8763cb45
JG
2241 if (!page || !page->mapping || PageTransCompound(page)) {
2242 mpfn = pfn = 0;
2243 goto next;
2244 }
a5430dda 2245 pfn = page_to_pfn(page);
8763cb45
JG
2246
2247 /*
2248 * By getting a reference on the page we pin it and that blocks
2249 * any kind of migration. Side effect is that it "freezes" the
2250 * pte.
2251 *
2252 * We drop this reference after isolating the page from the lru
2253 * for non device page (device page are not on the lru and thus
2254 * can't be dropped from it).
2255 */
2256 get_page(page);
2257 migrate->cpages++;
8763cb45 2258
8c3328f1
JG
2259 /*
2260 * Optimize for the common case where page is only mapped once
2261 * in one process. If we can lock the page, then we can safely
2262 * set up a special migration page table entry now.
2263 */
2264 if (trylock_page(page)) {
2265 pte_t swp_pte;
2266
2267 mpfn |= MIGRATE_PFN_LOCKED;
2268 ptep_get_and_clear(mm, addr, ptep);
2269
2270 /* Setup special migration page table entry */
07707125
RC
2271 entry = make_migration_entry(page, mpfn &
2272 MIGRATE_PFN_WRITE);
8c3328f1
JG
2273 swp_pte = swp_entry_to_pte(entry);
2274 if (pte_soft_dirty(pte))
2275 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2276 set_pte_at(mm, addr, ptep, swp_pte);
2277
2278 /*
2279 * This is like regular unmap: we remove the rmap and
2280 * drop page refcount. Page won't be freed, as we took
2281 * a reference just above.
2282 */
2283 page_remove_rmap(page, false);
2284 put_page(page);
a5430dda
JG
2285
2286 if (pte_present(pte))
2287 unmapped++;
8c3328f1
JG
2288 }
2289
8763cb45 2290next:
a5430dda 2291 migrate->dst[migrate->npages] = 0;
8763cb45
JG
2292 migrate->src[migrate->npages++] = mpfn;
2293 }
8c3328f1 2294 arch_leave_lazy_mmu_mode();
8763cb45
JG
2295 pte_unmap_unlock(ptep - 1, ptl);
2296
8c3328f1
JG
2297 /* Only flush the TLB if we actually modified any entries */
2298 if (unmapped)
2299 flush_tlb_range(walk->vma, start, end);
2300
8763cb45
JG
2301 return 0;
2302}
2303
2304/*
2305 * migrate_vma_collect() - collect pages over a range of virtual addresses
2306 * @migrate: migrate struct containing all migration information
2307 *
2308 * This will walk the CPU page table. For each virtual address backed by a
2309 * valid page, it updates the src array and takes a reference on the page, in
2310 * order to pin the page until we lock it and unmap it.
2311 */
2312static void migrate_vma_collect(struct migrate_vma *migrate)
2313{
2314 struct mm_walk mm_walk;
2315
2316 mm_walk.pmd_entry = migrate_vma_collect_pmd;
2317 mm_walk.pte_entry = NULL;
2318 mm_walk.pte_hole = migrate_vma_collect_hole;
2319 mm_walk.hugetlb_entry = NULL;
2320 mm_walk.test_walk = NULL;
2321 mm_walk.vma = migrate->vma;
2322 mm_walk.mm = migrate->vma->vm_mm;
2323 mm_walk.private = migrate;
2324
8c3328f1
JG
2325 mmu_notifier_invalidate_range_start(mm_walk.mm,
2326 migrate->start,
2327 migrate->end);
8763cb45 2328 walk_page_range(migrate->start, migrate->end, &mm_walk);
8c3328f1
JG
2329 mmu_notifier_invalidate_range_end(mm_walk.mm,
2330 migrate->start,
2331 migrate->end);
8763cb45
JG
2332
2333 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2334}
2335
2336/*
2337 * migrate_vma_check_page() - check if page is pinned or not
2338 * @page: struct page to check
2339 *
2340 * Pinned pages cannot be migrated. This is the same test as in
2341 * migrate_page_move_mapping(), except that here we allow migration of a
2342 * ZONE_DEVICE page.
2343 */
2344static bool migrate_vma_check_page(struct page *page)
2345{
2346 /*
2347 * One extra ref because caller holds an extra reference, either from
2348 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2349 * a device page.
2350 */
2351 int extra = 1;
2352
2353 /*
2354 * FIXME support THP (transparent huge page), it is bit more complex to
2355 * check them than regular pages, because they can be mapped with a pmd
2356 * or with a pte (split pte mapping).
2357 */
2358 if (PageCompound(page))
2359 return false;
2360
a5430dda
JG
2361 /* Page from ZONE_DEVICE have one extra reference */
2362 if (is_zone_device_page(page)) {
2363 /*
2364 * Private page can never be pin as they have no valid pte and
2365 * GUP will fail for those. Yet if there is a pending migration
2366 * a thread might try to wait on the pte migration entry and
2367 * will bump the page reference count. Sadly there is no way to
2368 * differentiate a regular pin from migration wait. Hence to
2369 * avoid 2 racing thread trying to migrate back to CPU to enter
2370 * infinite loop (one stoping migration because the other is
2371 * waiting on pte migration entry). We always return true here.
2372 *
2373 * FIXME proper solution is to rework migration_entry_wait() so
2374 * it does not need to take a reference on page.
2375 */
2376 if (is_device_private_page(page))
2377 return true;
2378
df6ad698
JG
2379 /*
2380 * Only allow device public page to be migrated and account for
2381 * the extra reference count imply by ZONE_DEVICE pages.
2382 */
2383 if (!is_device_public_page(page))
2384 return false;
2385 extra++;
a5430dda
JG
2386 }
2387
df6ad698
JG
2388 /* For file back page */
2389 if (page_mapping(page))
2390 extra += 1 + page_has_private(page);
2391
8763cb45
JG
2392 if ((page_count(page) - extra) > page_mapcount(page))
2393 return false;
2394
2395 return true;
2396}
2397
2398/*
2399 * migrate_vma_prepare() - lock pages and isolate them from the lru
2400 * @migrate: migrate struct containing all migration information
2401 *
2402 * This locks pages that have been collected by migrate_vma_collect(). Once each
2403 * page is locked it is isolated from the lru (for non-device pages). Finally,
2404 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2405 * migrated by concurrent kernel threads.
2406 */
2407static void migrate_vma_prepare(struct migrate_vma *migrate)
2408{
2409 const unsigned long npages = migrate->npages;
8c3328f1
JG
2410 const unsigned long start = migrate->start;
2411 unsigned long addr, i, restore = 0;
8763cb45 2412 bool allow_drain = true;
8763cb45
JG
2413
2414 lru_add_drain();
2415
2416 for (i = 0; (i < npages) && migrate->cpages; i++) {
2417 struct page *page = migrate_pfn_to_page(migrate->src[i]);
8c3328f1 2418 bool remap = true;
8763cb45
JG
2419
2420 if (!page)
2421 continue;
2422
8c3328f1
JG
2423 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2424 /*
2425 * Because we are migrating several pages there can be
2426 * a deadlock between 2 concurrent migration where each
2427 * are waiting on each other page lock.
2428 *
2429 * Make migrate_vma() a best effort thing and backoff
2430 * for any page we can not lock right away.
2431 */
2432 if (!trylock_page(page)) {
2433 migrate->src[i] = 0;
2434 migrate->cpages--;
2435 put_page(page);
2436 continue;
2437 }
2438 remap = false;
2439 migrate->src[i] |= MIGRATE_PFN_LOCKED;
8763cb45 2440 }
8763cb45 2441
a5430dda
JG
2442 /* ZONE_DEVICE pages are not on LRU */
2443 if (!is_zone_device_page(page)) {
2444 if (!PageLRU(page) && allow_drain) {
2445 /* Drain CPU's pagevec */
2446 lru_add_drain_all();
2447 allow_drain = false;
2448 }
8763cb45 2449
a5430dda
JG
2450 if (isolate_lru_page(page)) {
2451 if (remap) {
2452 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2453 migrate->cpages--;
2454 restore++;
2455 } else {
2456 migrate->src[i] = 0;
2457 unlock_page(page);
2458 migrate->cpages--;
2459 put_page(page);
2460 }
2461 continue;
8c3328f1 2462 }
a5430dda
JG
2463
2464 /* Drop the reference we took in collect */
2465 put_page(page);
8763cb45
JG
2466 }
2467
2468 if (!migrate_vma_check_page(page)) {
8c3328f1
JG
2469 if (remap) {
2470 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2471 migrate->cpages--;
2472 restore++;
8763cb45 2473
a5430dda
JG
2474 if (!is_zone_device_page(page)) {
2475 get_page(page);
2476 putback_lru_page(page);
2477 }
8c3328f1
JG
2478 } else {
2479 migrate->src[i] = 0;
2480 unlock_page(page);
2481 migrate->cpages--;
2482
a5430dda
JG
2483 if (!is_zone_device_page(page))
2484 putback_lru_page(page);
2485 else
2486 put_page(page);
8c3328f1 2487 }
8763cb45
JG
2488 }
2489 }
8c3328f1
JG
2490
2491 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2492 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2493
2494 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2495 continue;
2496
2497 remove_migration_pte(page, migrate->vma, addr, page);
2498
2499 migrate->src[i] = 0;
2500 unlock_page(page);
2501 put_page(page);
2502 restore--;
2503 }
8763cb45
JG
2504}
2505
2506/*
2507 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2508 * @migrate: migrate struct containing all migration information
2509 *
2510 * Replace page mapping (CPU page table pte) with a special migration pte entry
2511 * and check again if it has been pinned. Pinned pages are restored because we
2512 * cannot migrate them.
2513 *
2514 * This is the last step before we call the device driver callback to allocate
2515 * destination memory and copy contents of original page over to new page.
2516 */
2517static void migrate_vma_unmap(struct migrate_vma *migrate)
2518{
2519 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2520 const unsigned long npages = migrate->npages;
2521 const unsigned long start = migrate->start;
2522 unsigned long addr, i, restore = 0;
2523
2524 for (i = 0; i < npages; i++) {
2525 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2526
2527 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2528 continue;
2529
8c3328f1
JG
2530 if (page_mapped(page)) {
2531 try_to_unmap(page, flags);
2532 if (page_mapped(page))
2533 goto restore;
8763cb45 2534 }
8c3328f1
JG
2535
2536 if (migrate_vma_check_page(page))
2537 continue;
2538
2539restore:
2540 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2541 migrate->cpages--;
2542 restore++;
8763cb45
JG
2543 }
2544
2545 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2546 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2547
2548 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2549 continue;
2550
2551 remove_migration_ptes(page, page, false);
2552
2553 migrate->src[i] = 0;
2554 unlock_page(page);
2555 restore--;
2556
a5430dda
JG
2557 if (is_zone_device_page(page))
2558 put_page(page);
2559 else
2560 putback_lru_page(page);
8763cb45
JG
2561 }
2562}
2563
8315ada7
JG
2564static void migrate_vma_insert_page(struct migrate_vma *migrate,
2565 unsigned long addr,
2566 struct page *page,
2567 unsigned long *src,
2568 unsigned long *dst)
2569{
2570 struct vm_area_struct *vma = migrate->vma;
2571 struct mm_struct *mm = vma->vm_mm;
2572 struct mem_cgroup *memcg;
2573 bool flush = false;
2574 spinlock_t *ptl;
2575 pte_t entry;
2576 pgd_t *pgdp;
2577 p4d_t *p4dp;
2578 pud_t *pudp;
2579 pmd_t *pmdp;
2580 pte_t *ptep;
2581
2582 /* Only allow populating anonymous memory */
2583 if (!vma_is_anonymous(vma))
2584 goto abort;
2585
2586 pgdp = pgd_offset(mm, addr);
2587 p4dp = p4d_alloc(mm, pgdp, addr);
2588 if (!p4dp)
2589 goto abort;
2590 pudp = pud_alloc(mm, p4dp, addr);
2591 if (!pudp)
2592 goto abort;
2593 pmdp = pmd_alloc(mm, pudp, addr);
2594 if (!pmdp)
2595 goto abort;
2596
2597 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2598 goto abort;
2599
2600 /*
2601 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2602 * pte_offset_map() on pmds where a huge pmd might be created
2603 * from a different thread.
2604 *
2605 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2606 * parallel threads are excluded by other means.
2607 *
2608 * Here we only have down_read(mmap_sem).
2609 */
2610 if (pte_alloc(mm, pmdp, addr))
2611 goto abort;
2612
2613 /* See the comment in pte_alloc_one_map() */
2614 if (unlikely(pmd_trans_unstable(pmdp)))
2615 goto abort;
2616
2617 if (unlikely(anon_vma_prepare(vma)))
2618 goto abort;
2619 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2620 goto abort;
2621
2622 /*
2623 * The memory barrier inside __SetPageUptodate makes sure that
2624 * preceding stores to the page contents become visible before
2625 * the set_pte_at() write.
2626 */
2627 __SetPageUptodate(page);
2628
df6ad698
JG
2629 if (is_zone_device_page(page)) {
2630 if (is_device_private_page(page)) {
2631 swp_entry_t swp_entry;
2632
2633 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2634 entry = swp_entry_to_pte(swp_entry);
2635 } else if (is_device_public_page(page)) {
2636 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2637 if (vma->vm_flags & VM_WRITE)
2638 entry = pte_mkwrite(pte_mkdirty(entry));
2639 entry = pte_mkdevmap(entry);
2640 }
8315ada7
JG
2641 } else {
2642 entry = mk_pte(page, vma->vm_page_prot);
2643 if (vma->vm_flags & VM_WRITE)
2644 entry = pte_mkwrite(pte_mkdirty(entry));
2645 }
2646
2647 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2648
2649 if (pte_present(*ptep)) {
2650 unsigned long pfn = pte_pfn(*ptep);
2651
2652 if (!is_zero_pfn(pfn)) {
2653 pte_unmap_unlock(ptep, ptl);
2654 mem_cgroup_cancel_charge(page, memcg, false);
2655 goto abort;
2656 }
2657 flush = true;
2658 } else if (!pte_none(*ptep)) {
2659 pte_unmap_unlock(ptep, ptl);
2660 mem_cgroup_cancel_charge(page, memcg, false);
2661 goto abort;
2662 }
2663
2664 /*
2665 * Check for usefaultfd but do not deliver the fault. Instead,
2666 * just back off.
2667 */
2668 if (userfaultfd_missing(vma)) {
2669 pte_unmap_unlock(ptep, ptl);
2670 mem_cgroup_cancel_charge(page, memcg, false);
2671 goto abort;
2672 }
2673
2674 inc_mm_counter(mm, MM_ANONPAGES);
2675 page_add_new_anon_rmap(page, vma, addr, false);
2676 mem_cgroup_commit_charge(page, memcg, false, false);
2677 if (!is_zone_device_page(page))
2678 lru_cache_add_active_or_unevictable(page, vma);
2679 get_page(page);
2680
2681 if (flush) {
2682 flush_cache_page(vma, addr, pte_pfn(*ptep));
2683 ptep_clear_flush_notify(vma, addr, ptep);
2684 set_pte_at_notify(mm, addr, ptep, entry);
2685 update_mmu_cache(vma, addr, ptep);
2686 } else {
2687 /* No need to invalidate - it was non-present before */
2688 set_pte_at(mm, addr, ptep, entry);
2689 update_mmu_cache(vma, addr, ptep);
2690 }
2691
2692 pte_unmap_unlock(ptep, ptl);
2693 *src = MIGRATE_PFN_MIGRATE;
2694 return;
2695
2696abort:
2697 *src &= ~MIGRATE_PFN_MIGRATE;
2698}
2699
8763cb45
JG
2700/*
2701 * migrate_vma_pages() - migrate meta-data from src page to dst page
2702 * @migrate: migrate struct containing all migration information
2703 *
2704 * This migrates struct page meta-data from source struct page to destination
2705 * struct page. This effectively finishes the migration from source page to the
2706 * destination page.
2707 */
2708static void migrate_vma_pages(struct migrate_vma *migrate)
2709{
2710 const unsigned long npages = migrate->npages;
2711 const unsigned long start = migrate->start;
8315ada7
JG
2712 struct vm_area_struct *vma = migrate->vma;
2713 struct mm_struct *mm = vma->vm_mm;
2714 unsigned long addr, i, mmu_start;
2715 bool notified = false;
8763cb45
JG
2716
2717 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2718 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2719 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2720 struct address_space *mapping;
2721 int r;
2722
8315ada7
JG
2723 if (!newpage) {
2724 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
8763cb45 2725 continue;
8315ada7
JG
2726 }
2727
2728 if (!page) {
2729 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2730 continue;
2731 }
2732 if (!notified) {
2733 mmu_start = addr;
2734 notified = true;
2735 mmu_notifier_invalidate_range_start(mm,
2736 mmu_start,
2737 migrate->end);
2738 }
2739 migrate_vma_insert_page(migrate, addr, newpage,
2740 &migrate->src[i],
2741 &migrate->dst[i]);
8763cb45 2742 continue;
8315ada7 2743 }
8763cb45
JG
2744
2745 mapping = page_mapping(page);
2746
a5430dda
JG
2747 if (is_zone_device_page(newpage)) {
2748 if (is_device_private_page(newpage)) {
2749 /*
2750 * For now only support private anonymous when
2751 * migrating to un-addressable device memory.
2752 */
2753 if (mapping) {
2754 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2755 continue;
2756 }
df6ad698 2757 } else if (!is_device_public_page(newpage)) {
a5430dda
JG
2758 /*
2759 * Other types of ZONE_DEVICE page are not
2760 * supported.
2761 */
2762 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2763 continue;
2764 }
2765 }
2766
8763cb45
JG
2767 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2768 if (r != MIGRATEPAGE_SUCCESS)
2769 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2770 }
8315ada7 2771
4645b9fe
JG
2772 /*
2773 * No need to double call mmu_notifier->invalidate_range() callback as
2774 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2775 * did already call it.
2776 */
8315ada7 2777 if (notified)
4645b9fe
JG
2778 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2779 migrate->end);
8763cb45
JG
2780}
2781
2782/*
2783 * migrate_vma_finalize() - restore CPU page table entry
2784 * @migrate: migrate struct containing all migration information
2785 *
2786 * This replaces the special migration pte entry with either a mapping to the
2787 * new page if migration was successful for that page, or to the original page
2788 * otherwise.
2789 *
2790 * This also unlocks the pages and puts them back on the lru, or drops the extra
2791 * refcount, for device pages.
2792 */
2793static void migrate_vma_finalize(struct migrate_vma *migrate)
2794{
2795 const unsigned long npages = migrate->npages;
2796 unsigned long i;
2797
2798 for (i = 0; i < npages; i++) {
2799 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2800 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2801
8315ada7
JG
2802 if (!page) {
2803 if (newpage) {
2804 unlock_page(newpage);
2805 put_page(newpage);
2806 }
8763cb45 2807 continue;
8315ada7
JG
2808 }
2809
8763cb45
JG
2810 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2811 if (newpage) {
2812 unlock_page(newpage);
2813 put_page(newpage);
2814 }
2815 newpage = page;
2816 }
2817
2818 remove_migration_ptes(page, newpage, false);
2819 unlock_page(page);
2820 migrate->cpages--;
2821
a5430dda
JG
2822 if (is_zone_device_page(page))
2823 put_page(page);
2824 else
2825 putback_lru_page(page);
8763cb45
JG
2826
2827 if (newpage != page) {
2828 unlock_page(newpage);
a5430dda
JG
2829 if (is_zone_device_page(newpage))
2830 put_page(newpage);
2831 else
2832 putback_lru_page(newpage);
8763cb45
JG
2833 }
2834 }
2835}
2836
2837/*
2838 * migrate_vma() - migrate a range of memory inside vma
2839 *
2840 * @ops: migration callback for allocating destination memory and copying
2841 * @vma: virtual memory area containing the range to be migrated
2842 * @start: start address of the range to migrate (inclusive)
2843 * @end: end address of the range to migrate (exclusive)
2844 * @src: array of hmm_pfn_t containing source pfns
2845 * @dst: array of hmm_pfn_t containing destination pfns
2846 * @private: pointer passed back to each of the callback
2847 * Returns: 0 on success, error code otherwise
2848 *
2849 * This function tries to migrate a range of memory virtual address range, using
2850 * callbacks to allocate and copy memory from source to destination. First it
2851 * collects all the pages backing each virtual address in the range, saving this
2852 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2853 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2854 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2855 * in the corresponding src array entry. It then restores any pages that are
2856 * pinned, by remapping and unlocking those pages.
2857 *
2858 * At this point it calls the alloc_and_copy() callback. For documentation on
2859 * what is expected from that callback, see struct migrate_vma_ops comments in
2860 * include/linux/migrate.h
2861 *
2862 * After the alloc_and_copy() callback, this function goes over each entry in
2863 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2864 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2865 * then the function tries to migrate struct page information from the source
2866 * struct page to the destination struct page. If it fails to migrate the struct
2867 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2868 * array.
2869 *
2870 * At this point all successfully migrated pages have an entry in the src
2871 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2872 * array entry with MIGRATE_PFN_VALID flag set.
2873 *
2874 * It then calls the finalize_and_map() callback. See comments for "struct
2875 * migrate_vma_ops", in include/linux/migrate.h for details about
2876 * finalize_and_map() behavior.
2877 *
2878 * After the finalize_and_map() callback, for successfully migrated pages, this
2879 * function updates the CPU page table to point to new pages, otherwise it
2880 * restores the CPU page table to point to the original source pages.
2881 *
2882 * Function returns 0 after the above steps, even if no pages were migrated
2883 * (The function only returns an error if any of the arguments are invalid.)
2884 *
2885 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2886 * unsigned long entries.
2887 */
2888int migrate_vma(const struct migrate_vma_ops *ops,
2889 struct vm_area_struct *vma,
2890 unsigned long start,
2891 unsigned long end,
2892 unsigned long *src,
2893 unsigned long *dst,
2894 void *private)
2895{
2896 struct migrate_vma migrate;
2897
2898 /* Sanity check the arguments */
2899 start &= PAGE_MASK;
2900 end &= PAGE_MASK;
e1fb4a08
DJ
2901 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2902 vma_is_dax(vma))
8763cb45
JG
2903 return -EINVAL;
2904 if (start < vma->vm_start || start >= vma->vm_end)
2905 return -EINVAL;
2906 if (end <= vma->vm_start || end > vma->vm_end)
2907 return -EINVAL;
2908 if (!ops || !src || !dst || start >= end)
2909 return -EINVAL;
2910
2911 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2912 migrate.src = src;
2913 migrate.dst = dst;
2914 migrate.start = start;
2915 migrate.npages = 0;
2916 migrate.cpages = 0;
2917 migrate.end = end;
2918 migrate.vma = vma;
2919
2920 /* Collect, and try to unmap source pages */
2921 migrate_vma_collect(&migrate);
2922 if (!migrate.cpages)
2923 return 0;
2924
2925 /* Lock and isolate page */
2926 migrate_vma_prepare(&migrate);
2927 if (!migrate.cpages)
2928 return 0;
2929
2930 /* Unmap pages */
2931 migrate_vma_unmap(&migrate);
2932 if (!migrate.cpages)
2933 return 0;
2934
2935 /*
2936 * At this point pages are locked and unmapped, and thus they have
2937 * stable content and can safely be copied to destination memory that
2938 * is allocated by the callback.
2939 *
2940 * Note that migration can fail in migrate_vma_struct_page() for each
2941 * individual page.
2942 */
2943 ops->alloc_and_copy(vma, src, dst, start, end, private);
2944
2945 /* This does the real migration of struct page */
2946 migrate_vma_pages(&migrate);
2947
2948 ops->finalize_and_map(vma, src, dst, start, end, private);
2949
2950 /* Unlock and remap pages */
2951 migrate_vma_finalize(&migrate);
2952
2953 return 0;
2954}
2955EXPORT_SYMBOL(migrate_vma);
6b368cd4 2956#endif /* defined(MIGRATE_VMA_HELPER) */