]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
mm, oom: add oom victim's memcg to the oom context information
[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 328 /*
89eb946a 329 * Once page cache replacement of page migration started, page_count
e286781d
NP
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{
89eb946a 444 XA_STATE(xas, &mapping->i_pages, page_index(page));
42cb14b1
HD
445 struct zone *oldzone, *newzone;
446 int dirty;
8e321fef 447 int expected_count = 1 + extra_count;
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
89eb946a 473 xas_lock_irq(&xas);
b20a3503 474
e71769ae 475 expected_count += hpage_nr_pages(page) + page_has_private(page);
89eb946a
MW
476 if (page_count(page) != expected_count || xas_load(&xas) != page) {
477 xas_unlock_irq(&xas);
e23ca00b 478 return -EAGAIN;
b20a3503
CL
479 }
480
fe896d18 481 if (!page_ref_freeze(page, expected_count)) {
89eb946a 482 xas_unlock_irq(&xas);
e286781d
NP
483 return -EAGAIN;
484 }
485
b969c4ab
MG
486 /*
487 * In the async migration case of moving a page with buffers, lock the
488 * buffers using trylock before the mapping is moved. If the mapping
489 * was moved, we later failed to lock the buffers and could not move
490 * the mapping back due to an elevated page count, we would have to
491 * block waiting on other references to be dropped.
492 */
a6bc32b8
MG
493 if (mode == MIGRATE_ASYNC && head &&
494 !buffer_migrate_lock_buffers(head, mode)) {
fe896d18 495 page_ref_unfreeze(page, expected_count);
89eb946a 496 xas_unlock_irq(&xas);
b969c4ab
MG
497 return -EAGAIN;
498 }
499
b20a3503 500 /*
cf4b769a
HD
501 * Now we know that no one else is looking at the page:
502 * no turning back from here.
b20a3503 503 */
cf4b769a
HD
504 newpage->index = page->index;
505 newpage->mapping = page->mapping;
e71769ae 506 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
6326fec1
NP
507 if (PageSwapBacked(page)) {
508 __SetPageSwapBacked(newpage);
509 if (PageSwapCache(page)) {
510 SetPageSwapCache(newpage);
511 set_page_private(newpage, page_private(page));
512 }
513 } else {
514 VM_BUG_ON_PAGE(PageSwapCache(page), page);
b20a3503
CL
515 }
516
42cb14b1
HD
517 /* Move dirty while page refs frozen and newpage not yet exposed */
518 dirty = PageDirty(page);
519 if (dirty) {
520 ClearPageDirty(page);
521 SetPageDirty(newpage);
522 }
523
89eb946a 524 xas_store(&xas, newpage);
e71769ae
NH
525 if (PageTransHuge(page)) {
526 int i;
e71769ae 527
013567be 528 for (i = 1; i < HPAGE_PMD_NR; i++) {
89eb946a
MW
529 xas_next(&xas);
530 xas_store(&xas, newpage + i);
e71769ae 531 }
e71769ae 532 }
7cf9c2c7
NP
533
534 /*
937a94c9
JG
535 * Drop cache reference from old page by unfreezing
536 * to one less reference.
7cf9c2c7
NP
537 * We know this isn't the last reference.
538 */
e71769ae 539 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
7cf9c2c7 540
89eb946a 541 xas_unlock(&xas);
42cb14b1
HD
542 /* Leave irq disabled to prevent preemption while updating stats */
543
0e8c7d0f
CL
544 /*
545 * If moved to a different zone then also account
546 * the page for that zone. Other VM counters will be
547 * taken care of when we establish references to the
548 * new page and drop references to the old page.
549 *
550 * Note that anonymous pages are accounted for
4b9d0fab 551 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
0e8c7d0f
CL
552 * are mapped to swap space.
553 */
42cb14b1 554 if (newzone != oldzone) {
11fb9989
MG
555 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
556 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
42cb14b1 557 if (PageSwapBacked(page) && !PageSwapCache(page)) {
11fb9989
MG
558 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
559 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
42cb14b1
HD
560 }
561 if (dirty && mapping_cap_account_dirty(mapping)) {
11fb9989 562 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 563 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
11fb9989 564 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 565 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
42cb14b1 566 }
4b02108a 567 }
42cb14b1 568 local_irq_enable();
b20a3503 569
78bd5209 570 return MIGRATEPAGE_SUCCESS;
b20a3503 571}
1118dce7 572EXPORT_SYMBOL(migrate_page_move_mapping);
b20a3503 573
290408d4
NH
574/*
575 * The expected number of remaining references is the same as that
576 * of migrate_page_move_mapping().
577 */
578int migrate_huge_page_move_mapping(struct address_space *mapping,
579 struct page *newpage, struct page *page)
580{
89eb946a 581 XA_STATE(xas, &mapping->i_pages, page_index(page));
290408d4 582 int expected_count;
290408d4 583
89eb946a 584 xas_lock_irq(&xas);
290408d4 585 expected_count = 2 + page_has_private(page);
89eb946a
MW
586 if (page_count(page) != expected_count || xas_load(&xas) != page) {
587 xas_unlock_irq(&xas);
290408d4
NH
588 return -EAGAIN;
589 }
590
fe896d18 591 if (!page_ref_freeze(page, expected_count)) {
89eb946a 592 xas_unlock_irq(&xas);
290408d4
NH
593 return -EAGAIN;
594 }
595
cf4b769a
HD
596 newpage->index = page->index;
597 newpage->mapping = page->mapping;
6a93ca8f 598
290408d4
NH
599 get_page(newpage);
600
89eb946a 601 xas_store(&xas, newpage);
290408d4 602
fe896d18 603 page_ref_unfreeze(page, expected_count - 1);
290408d4 604
89eb946a 605 xas_unlock_irq(&xas);
6a93ca8f 606
78bd5209 607 return MIGRATEPAGE_SUCCESS;
290408d4
NH
608}
609
30b0a105
DH
610/*
611 * Gigantic pages are so large that we do not guarantee that page++ pointer
612 * arithmetic will work across the entire page. We need something more
613 * specialized.
614 */
615static void __copy_gigantic_page(struct page *dst, struct page *src,
616 int nr_pages)
617{
618 int i;
619 struct page *dst_base = dst;
620 struct page *src_base = src;
621
622 for (i = 0; i < nr_pages; ) {
623 cond_resched();
624 copy_highpage(dst, src);
625
626 i++;
627 dst = mem_map_next(dst, dst_base, i);
628 src = mem_map_next(src, src_base, i);
629 }
630}
631
632static void copy_huge_page(struct page *dst, struct page *src)
633{
634 int i;
635 int nr_pages;
636
637 if (PageHuge(src)) {
638 /* hugetlbfs page */
639 struct hstate *h = page_hstate(src);
640 nr_pages = pages_per_huge_page(h);
641
642 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
643 __copy_gigantic_page(dst, src, nr_pages);
644 return;
645 }
646 } else {
647 /* thp page */
648 BUG_ON(!PageTransHuge(src));
649 nr_pages = hpage_nr_pages(src);
650 }
651
652 for (i = 0; i < nr_pages; i++) {
653 cond_resched();
654 copy_highpage(dst + i, src + i);
655 }
656}
657
b20a3503
CL
658/*
659 * Copy the page to its new location
660 */
2916ecc0 661void migrate_page_states(struct page *newpage, struct page *page)
b20a3503 662{
7851a45c
RR
663 int cpupid;
664
b20a3503
CL
665 if (PageError(page))
666 SetPageError(newpage);
667 if (PageReferenced(page))
668 SetPageReferenced(newpage);
669 if (PageUptodate(page))
670 SetPageUptodate(newpage);
894bc310 671 if (TestClearPageActive(page)) {
309381fe 672 VM_BUG_ON_PAGE(PageUnevictable(page), page);
b20a3503 673 SetPageActive(newpage);
418b27ef
LS
674 } else if (TestClearPageUnevictable(page))
675 SetPageUnevictable(newpage);
1899ad18
JW
676 if (PageWorkingset(page))
677 SetPageWorkingset(newpage);
b20a3503
CL
678 if (PageChecked(page))
679 SetPageChecked(newpage);
680 if (PageMappedToDisk(page))
681 SetPageMappedToDisk(newpage);
682
42cb14b1
HD
683 /* Move dirty on pages not done by migrate_page_move_mapping() */
684 if (PageDirty(page))
685 SetPageDirty(newpage);
b20a3503 686
33c3fc71
VD
687 if (page_is_young(page))
688 set_page_young(newpage);
689 if (page_is_idle(page))
690 set_page_idle(newpage);
691
7851a45c
RR
692 /*
693 * Copy NUMA information to the new page, to prevent over-eager
694 * future migrations of this same page.
695 */
696 cpupid = page_cpupid_xchg_last(page, -1);
697 page_cpupid_xchg_last(newpage, cpupid);
698
e9995ef9 699 ksm_migrate_page(newpage, page);
c8d6553b
HD
700 /*
701 * Please do not reorder this without considering how mm/ksm.c's
702 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
703 */
b3b3a99c
NH
704 if (PageSwapCache(page))
705 ClearPageSwapCache(page);
b20a3503
CL
706 ClearPagePrivate(page);
707 set_page_private(page, 0);
b20a3503
CL
708
709 /*
710 * If any waiters have accumulated on the new page then
711 * wake them up.
712 */
713 if (PageWriteback(newpage))
714 end_page_writeback(newpage);
d435edca
VB
715
716 copy_page_owner(page, newpage);
74485cf2
JW
717
718 mem_cgroup_migrate(page, newpage);
b20a3503 719}
2916ecc0
JG
720EXPORT_SYMBOL(migrate_page_states);
721
722void migrate_page_copy(struct page *newpage, struct page *page)
723{
724 if (PageHuge(page) || PageTransHuge(page))
725 copy_huge_page(newpage, page);
726 else
727 copy_highpage(newpage, page);
728
729 migrate_page_states(newpage, page);
730}
1118dce7 731EXPORT_SYMBOL(migrate_page_copy);
b20a3503 732
1d8b85cc
CL
733/************************************************************
734 * Migration functions
735 ***********************************************************/
736
b20a3503 737/*
bda807d4 738 * Common logic to directly migrate a single LRU page suitable for
266cf658 739 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
740 *
741 * Pages are locked upon entry and exit.
742 */
2d1db3b1 743int migrate_page(struct address_space *mapping,
a6bc32b8
MG
744 struct page *newpage, struct page *page,
745 enum migrate_mode mode)
b20a3503
CL
746{
747 int rc;
748
749 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
750
8e321fef 751 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
b20a3503 752
78bd5209 753 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
754 return rc;
755
2916ecc0
JG
756 if (mode != MIGRATE_SYNC_NO_COPY)
757 migrate_page_copy(newpage, page);
758 else
759 migrate_page_states(newpage, page);
78bd5209 760 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
761}
762EXPORT_SYMBOL(migrate_page);
763
9361401e 764#ifdef CONFIG_BLOCK
1d8b85cc
CL
765/*
766 * Migration function for pages with buffers. This function can only be used
767 * if the underlying filesystem guarantees that no other references to "page"
768 * exist.
769 */
2d1db3b1 770int buffer_migrate_page(struct address_space *mapping,
a6bc32b8 771 struct page *newpage, struct page *page, enum migrate_mode mode)
1d8b85cc 772{
1d8b85cc
CL
773 struct buffer_head *bh, *head;
774 int rc;
775
1d8b85cc 776 if (!page_has_buffers(page))
a6bc32b8 777 return migrate_page(mapping, newpage, page, mode);
1d8b85cc
CL
778
779 head = page_buffers(page);
780
8e321fef 781 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
1d8b85cc 782
78bd5209 783 if (rc != MIGRATEPAGE_SUCCESS)
1d8b85cc
CL
784 return rc;
785
b969c4ab
MG
786 /*
787 * In the async case, migrate_page_move_mapping locked the buffers
788 * with an IRQ-safe spinlock held. In the sync case, the buffers
789 * need to be locked now
790 */
a6bc32b8
MG
791 if (mode != MIGRATE_ASYNC)
792 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
1d8b85cc
CL
793
794 ClearPagePrivate(page);
795 set_page_private(newpage, page_private(page));
796 set_page_private(page, 0);
797 put_page(page);
798 get_page(newpage);
799
800 bh = head;
801 do {
802 set_bh_page(bh, newpage, bh_offset(bh));
803 bh = bh->b_this_page;
804
805 } while (bh != head);
806
807 SetPagePrivate(newpage);
808
2916ecc0
JG
809 if (mode != MIGRATE_SYNC_NO_COPY)
810 migrate_page_copy(newpage, page);
811 else
812 migrate_page_states(newpage, page);
1d8b85cc
CL
813
814 bh = head;
815 do {
816 unlock_buffer(bh);
2916ecc0 817 put_bh(bh);
1d8b85cc
CL
818 bh = bh->b_this_page;
819
820 } while (bh != head);
821
78bd5209 822 return MIGRATEPAGE_SUCCESS;
1d8b85cc
CL
823}
824EXPORT_SYMBOL(buffer_migrate_page);
9361401e 825#endif
1d8b85cc 826
04e62a29
CL
827/*
828 * Writeback a page to clean the dirty state
829 */
830static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 831{
04e62a29
CL
832 struct writeback_control wbc = {
833 .sync_mode = WB_SYNC_NONE,
834 .nr_to_write = 1,
835 .range_start = 0,
836 .range_end = LLONG_MAX,
04e62a29
CL
837 .for_reclaim = 1
838 };
839 int rc;
840
841 if (!mapping->a_ops->writepage)
842 /* No write method for the address space */
843 return -EINVAL;
844
845 if (!clear_page_dirty_for_io(page))
846 /* Someone else already triggered a write */
847 return -EAGAIN;
848
8351a6e4 849 /*
04e62a29
CL
850 * A dirty page may imply that the underlying filesystem has
851 * the page on some queue. So the page must be clean for
852 * migration. Writeout may mean we loose the lock and the
853 * page state is no longer what we checked for earlier.
854 * At this point we know that the migration attempt cannot
855 * be successful.
8351a6e4 856 */
e388466d 857 remove_migration_ptes(page, page, false);
8351a6e4 858
04e62a29 859 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 860
04e62a29
CL
861 if (rc != AOP_WRITEPAGE_ACTIVATE)
862 /* unlocked. Relock */
863 lock_page(page);
864
bda8550d 865 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
866}
867
868/*
869 * Default handling if a filesystem does not provide a migration function.
870 */
871static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 872 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 873{
b969c4ab 874 if (PageDirty(page)) {
a6bc32b8 875 /* Only writeback pages in full synchronous migration */
2916ecc0
JG
876 switch (mode) {
877 case MIGRATE_SYNC:
878 case MIGRATE_SYNC_NO_COPY:
879 break;
880 default:
b969c4ab 881 return -EBUSY;
2916ecc0 882 }
04e62a29 883 return writeout(mapping, page);
b969c4ab 884 }
8351a6e4
CL
885
886 /*
887 * Buffers may be managed in a filesystem specific way.
888 * We must have no buffers or drop them.
889 */
266cf658 890 if (page_has_private(page) &&
8351a6e4
CL
891 !try_to_release_page(page, GFP_KERNEL))
892 return -EAGAIN;
893
a6bc32b8 894 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
895}
896
e24f0b8f
CL
897/*
898 * Move a page to a newly allocated page
899 * The page is locked and all ptes have been successfully removed.
900 *
901 * The new page will have replaced the old page if this function
902 * is successful.
894bc310
LS
903 *
904 * Return value:
905 * < 0 - error code
78bd5209 906 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 907 */
3fe2011f 908static int move_to_new_page(struct page *newpage, struct page *page,
5c3f9a67 909 enum migrate_mode mode)
e24f0b8f
CL
910{
911 struct address_space *mapping;
bda807d4
MK
912 int rc = -EAGAIN;
913 bool is_lru = !__PageMovable(page);
e24f0b8f 914
7db7671f
HD
915 VM_BUG_ON_PAGE(!PageLocked(page), page);
916 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
e24f0b8f 917
e24f0b8f 918 mapping = page_mapping(page);
bda807d4
MK
919
920 if (likely(is_lru)) {
921 if (!mapping)
922 rc = migrate_page(mapping, newpage, page, mode);
923 else if (mapping->a_ops->migratepage)
924 /*
925 * Most pages have a mapping and most filesystems
926 * provide a migratepage callback. Anonymous pages
927 * are part of swap space which also has its own
928 * migratepage callback. This is the most common path
929 * for page migration.
930 */
931 rc = mapping->a_ops->migratepage(mapping, newpage,
932 page, mode);
933 else
934 rc = fallback_migrate_page(mapping, newpage,
935 page, mode);
936 } else {
e24f0b8f 937 /*
bda807d4
MK
938 * In case of non-lru page, it could be released after
939 * isolation step. In that case, we shouldn't try migration.
e24f0b8f 940 */
bda807d4
MK
941 VM_BUG_ON_PAGE(!PageIsolated(page), page);
942 if (!PageMovable(page)) {
943 rc = MIGRATEPAGE_SUCCESS;
944 __ClearPageIsolated(page);
945 goto out;
946 }
947
948 rc = mapping->a_ops->migratepage(mapping, newpage,
949 page, mode);
950 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
951 !PageIsolated(page));
952 }
e24f0b8f 953
5c3f9a67
HD
954 /*
955 * When successful, old pagecache page->mapping must be cleared before
956 * page is freed; but stats require that PageAnon be left as PageAnon.
957 */
958 if (rc == MIGRATEPAGE_SUCCESS) {
bda807d4
MK
959 if (__PageMovable(page)) {
960 VM_BUG_ON_PAGE(!PageIsolated(page), page);
961
962 /*
963 * We clear PG_movable under page_lock so any compactor
964 * cannot try to migrate this page.
965 */
966 __ClearPageIsolated(page);
967 }
968
969 /*
970 * Anonymous and movable page->mapping will be cleard by
971 * free_pages_prepare so don't reset it here for keeping
972 * the type to work PageAnon, for example.
973 */
974 if (!PageMappingFlags(page))
5c3f9a67 975 page->mapping = NULL;
3fe2011f 976 }
bda807d4 977out:
e24f0b8f
CL
978 return rc;
979}
980
0dabec93 981static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 982 int force, enum migrate_mode mode)
e24f0b8f 983{
0dabec93 984 int rc = -EAGAIN;
2ebba6b7 985 int page_was_mapped = 0;
3f6c8272 986 struct anon_vma *anon_vma = NULL;
bda807d4 987 bool is_lru = !__PageMovable(page);
95a402c3 988
529ae9aa 989 if (!trylock_page(page)) {
a6bc32b8 990 if (!force || mode == MIGRATE_ASYNC)
0dabec93 991 goto out;
3e7d3449
MG
992
993 /*
994 * It's not safe for direct compaction to call lock_page.
995 * For example, during page readahead pages are added locked
996 * to the LRU. Later, when the IO completes the pages are
997 * marked uptodate and unlocked. However, the queueing
998 * could be merging multiple pages for one bio (e.g.
999 * mpage_readpages). If an allocation happens for the
1000 * second or third page, the process can end up locking
1001 * the same page twice and deadlocking. Rather than
1002 * trying to be clever about what pages can be locked,
1003 * avoid the use of lock_page for direct compaction
1004 * altogether.
1005 */
1006 if (current->flags & PF_MEMALLOC)
0dabec93 1007 goto out;
3e7d3449 1008
e24f0b8f
CL
1009 lock_page(page);
1010 }
1011
1012 if (PageWriteback(page)) {
11bc82d6 1013 /*
fed5b64a 1014 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
1015 * necessary to wait for PageWriteback. In the async case,
1016 * the retry loop is too short and in the sync-light case,
1017 * the overhead of stalling is too much
11bc82d6 1018 */
2916ecc0
JG
1019 switch (mode) {
1020 case MIGRATE_SYNC:
1021 case MIGRATE_SYNC_NO_COPY:
1022 break;
1023 default:
11bc82d6 1024 rc = -EBUSY;
0a31bc97 1025 goto out_unlock;
11bc82d6
AA
1026 }
1027 if (!force)
0a31bc97 1028 goto out_unlock;
e24f0b8f
CL
1029 wait_on_page_writeback(page);
1030 }
03f15c86 1031
e24f0b8f 1032 /*
dc386d4d
KH
1033 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1034 * we cannot notice that anon_vma is freed while we migrates a page.
1ce82b69 1035 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 1036 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
1037 * File Caches may use write_page() or lock_page() in migration, then,
1038 * just care Anon page here.
03f15c86
HD
1039 *
1040 * Only page_get_anon_vma() understands the subtleties of
1041 * getting a hold on an anon_vma from outside one of its mms.
1042 * But if we cannot get anon_vma, then we won't need it anyway,
1043 * because that implies that the anon page is no longer mapped
1044 * (and cannot be remapped so long as we hold the page lock).
dc386d4d 1045 */
03f15c86 1046 if (PageAnon(page) && !PageKsm(page))
746b18d4 1047 anon_vma = page_get_anon_vma(page);
62e1c553 1048
7db7671f
HD
1049 /*
1050 * Block others from accessing the new page when we get around to
1051 * establishing additional references. We are usually the only one
1052 * holding a reference to newpage at this point. We used to have a BUG
1053 * here if trylock_page(newpage) fails, but would like to allow for
1054 * cases where there might be a race with the previous use of newpage.
1055 * This is much like races on refcount of oldpage: just don't BUG().
1056 */
1057 if (unlikely(!trylock_page(newpage)))
1058 goto out_unlock;
1059
bda807d4
MK
1060 if (unlikely(!is_lru)) {
1061 rc = move_to_new_page(newpage, page, mode);
1062 goto out_unlock_both;
1063 }
1064
dc386d4d 1065 /*
62e1c553
SL
1066 * Corner case handling:
1067 * 1. When a new swap-cache page is read into, it is added to the LRU
1068 * and treated as swapcache but it has no rmap yet.
1069 * Calling try_to_unmap() against a page->mapping==NULL page will
1070 * trigger a BUG. So handle it here.
1071 * 2. An orphaned page (see truncate_complete_page) might have
1072 * fs-private metadata. The page can be picked up due to memory
1073 * offlining. Everywhere else except page reclaim, the page is
1074 * invisible to the vm, so the page can not be migrated. So try to
1075 * free the metadata, so the page can be freed.
e24f0b8f 1076 */
62e1c553 1077 if (!page->mapping) {
309381fe 1078 VM_BUG_ON_PAGE(PageAnon(page), page);
1ce82b69 1079 if (page_has_private(page)) {
62e1c553 1080 try_to_free_buffers(page);
7db7671f 1081 goto out_unlock_both;
62e1c553 1082 }
7db7671f
HD
1083 } else if (page_mapped(page)) {
1084 /* Establish migration ptes */
03f15c86
HD
1085 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1086 page);
2ebba6b7 1087 try_to_unmap(page,
da1b13cc 1088 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
2ebba6b7
HD
1089 page_was_mapped = 1;
1090 }
dc386d4d 1091
e6a1530d 1092 if (!page_mapped(page))
5c3f9a67 1093 rc = move_to_new_page(newpage, page, mode);
e24f0b8f 1094
5c3f9a67
HD
1095 if (page_was_mapped)
1096 remove_migration_ptes(page,
e388466d 1097 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
3f6c8272 1098
7db7671f
HD
1099out_unlock_both:
1100 unlock_page(newpage);
1101out_unlock:
3f6c8272 1102 /* Drop an anon_vma reference if we took one */
76545066 1103 if (anon_vma)
9e60109f 1104 put_anon_vma(anon_vma);
e24f0b8f 1105 unlock_page(page);
0dabec93 1106out:
c6c919eb
MK
1107 /*
1108 * If migration is successful, decrease refcount of the newpage
1109 * which will not free the page because new page owner increased
1110 * refcounter. As well, if it is LRU page, add the page to LRU
1111 * list in here.
1112 */
1113 if (rc == MIGRATEPAGE_SUCCESS) {
b1123ea6 1114 if (unlikely(__PageMovable(newpage)))
c6c919eb
MK
1115 put_page(newpage);
1116 else
1117 putback_lru_page(newpage);
1118 }
1119
0dabec93
MK
1120 return rc;
1121}
95a402c3 1122
ef2a5153
GU
1123/*
1124 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1125 * around it.
1126 */
815f0ddb
ND
1127#if defined(CONFIG_ARM) && \
1128 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
ef2a5153
GU
1129#define ICE_noinline noinline
1130#else
1131#define ICE_noinline
1132#endif
1133
0dabec93
MK
1134/*
1135 * Obtain the lock on page, remove all ptes and migrate the page
1136 * to the newly allocated page in newpage.
1137 */
ef2a5153
GU
1138static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1139 free_page_t put_new_page,
1140 unsigned long private, struct page *page,
add05cec
NH
1141 int force, enum migrate_mode mode,
1142 enum migrate_reason reason)
0dabec93 1143{
2def7424 1144 int rc = MIGRATEPAGE_SUCCESS;
2def7424 1145 struct page *newpage;
0dabec93 1146
94723aaf
MH
1147 if (!thp_migration_supported() && PageTransHuge(page))
1148 return -ENOMEM;
1149
666feb21 1150 newpage = get_new_page(page, private);
0dabec93
MK
1151 if (!newpage)
1152 return -ENOMEM;
1153
1154 if (page_count(page) == 1) {
1155 /* page was freed from under us. So we are done. */
c6c919eb
MK
1156 ClearPageActive(page);
1157 ClearPageUnevictable(page);
bda807d4
MK
1158 if (unlikely(__PageMovable(page))) {
1159 lock_page(page);
1160 if (!PageMovable(page))
1161 __ClearPageIsolated(page);
1162 unlock_page(page);
1163 }
c6c919eb
MK
1164 if (put_new_page)
1165 put_new_page(newpage, private);
1166 else
1167 put_page(newpage);
0dabec93
MK
1168 goto out;
1169 }
1170
9c620e2b 1171 rc = __unmap_and_move(page, newpage, force, mode);
c6c919eb 1172 if (rc == MIGRATEPAGE_SUCCESS)
7cd12b4a 1173 set_page_owner_migrate_reason(newpage, reason);
bf6bddf1 1174
0dabec93 1175out:
e24f0b8f 1176 if (rc != -EAGAIN) {
0dabec93
MK
1177 /*
1178 * A page that has been migrated has all references
1179 * removed and will be freed. A page that has not been
1180 * migrated will have kepts its references and be
1181 * restored.
1182 */
1183 list_del(&page->lru);
6afcf8ef
ML
1184
1185 /*
1186 * Compaction can migrate also non-LRU pages which are
1187 * not accounted to NR_ISOLATED_*. They can be recognized
1188 * as __PageMovable
1189 */
1190 if (likely(!__PageMovable(page)))
e8db67eb
NH
1191 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1192 page_is_file_cache(page), -hpage_nr_pages(page));
c6c919eb
MK
1193 }
1194
1195 /*
1196 * If migration is successful, releases reference grabbed during
1197 * isolation. Otherwise, restore the page to right list unless
1198 * we want to retry.
1199 */
1200 if (rc == MIGRATEPAGE_SUCCESS) {
1201 put_page(page);
1202 if (reason == MR_MEMORY_FAILURE) {
d7e69488 1203 /*
c6c919eb
MK
1204 * Set PG_HWPoison on just freed page
1205 * intentionally. Although it's rather weird,
1206 * it's how HWPoison flag works at the moment.
d7e69488 1207 */
d4ae9916 1208 if (set_hwpoison_free_buddy_page(page))
da1b13cc 1209 num_poisoned_pages_inc();
c6c919eb
MK
1210 }
1211 } else {
bda807d4
MK
1212 if (rc != -EAGAIN) {
1213 if (likely(!__PageMovable(page))) {
1214 putback_lru_page(page);
1215 goto put_new;
1216 }
1217
1218 lock_page(page);
1219 if (PageMovable(page))
1220 putback_movable_page(page);
1221 else
1222 __ClearPageIsolated(page);
1223 unlock_page(page);
1224 put_page(page);
1225 }
1226put_new:
c6c919eb
MK
1227 if (put_new_page)
1228 put_new_page(newpage, private);
1229 else
1230 put_page(newpage);
e24f0b8f 1231 }
68711a74 1232
e24f0b8f
CL
1233 return rc;
1234}
1235
290408d4
NH
1236/*
1237 * Counterpart of unmap_and_move_page() for hugepage migration.
1238 *
1239 * This function doesn't wait the completion of hugepage I/O
1240 * because there is no race between I/O and migration for hugepage.
1241 * Note that currently hugepage I/O occurs only in direct I/O
1242 * where no lock is held and PG_writeback is irrelevant,
1243 * and writeback status of all subpages are counted in the reference
1244 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1245 * under direct I/O, the reference of the head page is 512 and a bit more.)
1246 * This means that when we try to migrate hugepage whose subpages are
1247 * doing direct I/O, some references remain after try_to_unmap() and
1248 * hugepage migration fails without data corruption.
1249 *
1250 * There is also no race when direct I/O is issued on the page under migration,
1251 * because then pte is replaced with migration swap entry and direct I/O code
1252 * will wait in the page fault for migration to complete.
1253 */
1254static int unmap_and_move_huge_page(new_page_t get_new_page,
68711a74
DR
1255 free_page_t put_new_page, unsigned long private,
1256 struct page *hpage, int force,
7cd12b4a 1257 enum migrate_mode mode, int reason)
290408d4 1258{
2def7424 1259 int rc = -EAGAIN;
2ebba6b7 1260 int page_was_mapped = 0;
32665f2b 1261 struct page *new_hpage;
290408d4
NH
1262 struct anon_vma *anon_vma = NULL;
1263
83467efb
NH
1264 /*
1265 * Movability of hugepages depends on architectures and hugepage size.
1266 * This check is necessary because some callers of hugepage migration
1267 * like soft offline and memory hotremove don't walk through page
1268 * tables or check whether the hugepage is pmd-based or not before
1269 * kicking migration.
1270 */
100873d7 1271 if (!hugepage_migration_supported(page_hstate(hpage))) {
32665f2b 1272 putback_active_hugepage(hpage);
83467efb 1273 return -ENOSYS;
32665f2b 1274 }
83467efb 1275
666feb21 1276 new_hpage = get_new_page(hpage, private);
290408d4
NH
1277 if (!new_hpage)
1278 return -ENOMEM;
1279
290408d4 1280 if (!trylock_page(hpage)) {
2916ecc0 1281 if (!force)
290408d4 1282 goto out;
2916ecc0
JG
1283 switch (mode) {
1284 case MIGRATE_SYNC:
1285 case MIGRATE_SYNC_NO_COPY:
1286 break;
1287 default:
1288 goto out;
1289 }
290408d4
NH
1290 lock_page(hpage);
1291 }
1292
746b18d4
PZ
1293 if (PageAnon(hpage))
1294 anon_vma = page_get_anon_vma(hpage);
290408d4 1295
7db7671f
HD
1296 if (unlikely(!trylock_page(new_hpage)))
1297 goto put_anon;
1298
2ebba6b7
HD
1299 if (page_mapped(hpage)) {
1300 try_to_unmap(hpage,
1301 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1302 page_was_mapped = 1;
1303 }
290408d4
NH
1304
1305 if (!page_mapped(hpage))
5c3f9a67 1306 rc = move_to_new_page(new_hpage, hpage, mode);
290408d4 1307
5c3f9a67
HD
1308 if (page_was_mapped)
1309 remove_migration_ptes(hpage,
e388466d 1310 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
290408d4 1311
7db7671f
HD
1312 unlock_page(new_hpage);
1313
1314put_anon:
fd4a4663 1315 if (anon_vma)
9e60109f 1316 put_anon_vma(anon_vma);
8e6ac7fa 1317
2def7424 1318 if (rc == MIGRATEPAGE_SUCCESS) {
ab5ac90a 1319 move_hugetlb_state(hpage, new_hpage, reason);
2def7424
HD
1320 put_new_page = NULL;
1321 }
8e6ac7fa 1322
290408d4 1323 unlock_page(hpage);
09761333 1324out:
b8ec1cee
NH
1325 if (rc != -EAGAIN)
1326 putback_active_hugepage(hpage);
68711a74
DR
1327
1328 /*
1329 * If migration was not successful and there's a freeing callback, use
1330 * it. Otherwise, put_page() will drop the reference grabbed during
1331 * isolation.
1332 */
2def7424 1333 if (put_new_page)
68711a74
DR
1334 put_new_page(new_hpage, private);
1335 else
3aaa76e1 1336 putback_active_hugepage(new_hpage);
68711a74 1337
290408d4
NH
1338 return rc;
1339}
1340
b20a3503 1341/*
c73e5c9c
SB
1342 * migrate_pages - migrate the pages specified in a list, to the free pages
1343 * supplied as the target for the page migration
b20a3503 1344 *
c73e5c9c
SB
1345 * @from: The list of pages to be migrated.
1346 * @get_new_page: The function used to allocate free pages to be used
1347 * as the target of the page migration.
68711a74
DR
1348 * @put_new_page: The function used to free target pages if migration
1349 * fails, or NULL if no special handling is necessary.
c73e5c9c
SB
1350 * @private: Private data to be passed on to get_new_page()
1351 * @mode: The migration mode that specifies the constraints for
1352 * page migration, if any.
1353 * @reason: The reason for page migration.
b20a3503 1354 *
c73e5c9c
SB
1355 * The function returns after 10 attempts or if no pages are movable any more
1356 * because the list has become empty or no retryable pages exist any more.
14e0f9bc 1357 * The caller should call putback_movable_pages() to return pages to the LRU
28bd6578 1358 * or free list only if ret != 0.
b20a3503 1359 *
c73e5c9c 1360 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1361 */
9c620e2b 1362int migrate_pages(struct list_head *from, new_page_t get_new_page,
68711a74
DR
1363 free_page_t put_new_page, unsigned long private,
1364 enum migrate_mode mode, int reason)
b20a3503 1365{
e24f0b8f 1366 int retry = 1;
b20a3503 1367 int nr_failed = 0;
5647bc29 1368 int nr_succeeded = 0;
b20a3503
CL
1369 int pass = 0;
1370 struct page *page;
1371 struct page *page2;
1372 int swapwrite = current->flags & PF_SWAPWRITE;
1373 int rc;
1374
1375 if (!swapwrite)
1376 current->flags |= PF_SWAPWRITE;
1377
e24f0b8f
CL
1378 for(pass = 0; pass < 10 && retry; pass++) {
1379 retry = 0;
b20a3503 1380
e24f0b8f 1381 list_for_each_entry_safe(page, page2, from, lru) {
94723aaf 1382retry:
e24f0b8f 1383 cond_resched();
2d1db3b1 1384
31caf665
NH
1385 if (PageHuge(page))
1386 rc = unmap_and_move_huge_page(get_new_page,
68711a74 1387 put_new_page, private, page,
7cd12b4a 1388 pass > 2, mode, reason);
31caf665 1389 else
68711a74 1390 rc = unmap_and_move(get_new_page, put_new_page,
add05cec
NH
1391 private, page, pass > 2, mode,
1392 reason);
2d1db3b1 1393
e24f0b8f 1394 switch(rc) {
95a402c3 1395 case -ENOMEM:
94723aaf
MH
1396 /*
1397 * THP migration might be unsupported or the
1398 * allocation could've failed so we should
1399 * retry on the same page with the THP split
1400 * to base pages.
1401 *
1402 * Head page is retried immediately and tail
1403 * pages are added to the tail of the list so
1404 * we encounter them after the rest of the list
1405 * is processed.
1406 */
e6112fc3 1407 if (PageTransHuge(page) && !PageHuge(page)) {
94723aaf
MH
1408 lock_page(page);
1409 rc = split_huge_page_to_list(page, from);
1410 unlock_page(page);
1411 if (!rc) {
1412 list_safe_reset_next(page, page2, lru);
1413 goto retry;
1414 }
1415 }
dfef2ef4 1416 nr_failed++;
95a402c3 1417 goto out;
e24f0b8f 1418 case -EAGAIN:
2d1db3b1 1419 retry++;
e24f0b8f 1420 break;
78bd5209 1421 case MIGRATEPAGE_SUCCESS:
5647bc29 1422 nr_succeeded++;
e24f0b8f
CL
1423 break;
1424 default:
354a3363
NH
1425 /*
1426 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1427 * unlike -EAGAIN case, the failed page is
1428 * removed from migration page list and not
1429 * retried in the next outer loop.
1430 */
2d1db3b1 1431 nr_failed++;
e24f0b8f 1432 break;
2d1db3b1 1433 }
b20a3503
CL
1434 }
1435 }
f2f81fb2
VB
1436 nr_failed += retry;
1437 rc = nr_failed;
95a402c3 1438out:
5647bc29
MG
1439 if (nr_succeeded)
1440 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1441 if (nr_failed)
1442 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1443 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1444
b20a3503
CL
1445 if (!swapwrite)
1446 current->flags &= ~PF_SWAPWRITE;
1447
78bd5209 1448 return rc;
b20a3503 1449}
95a402c3 1450
742755a1 1451#ifdef CONFIG_NUMA
742755a1 1452
a49bd4d7 1453static int store_status(int __user *status, int start, int value, int nr)
742755a1 1454{
a49bd4d7
MH
1455 while (nr-- > 0) {
1456 if (put_user(value, status + start))
1457 return -EFAULT;
1458 start++;
1459 }
1460
1461 return 0;
1462}
1463
1464static int do_move_pages_to_node(struct mm_struct *mm,
1465 struct list_head *pagelist, int node)
1466{
1467 int err;
1468
1469 if (list_empty(pagelist))
1470 return 0;
1471
1472 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1473 MIGRATE_SYNC, MR_SYSCALL);
1474 if (err)
1475 putback_movable_pages(pagelist);
1476 return err;
742755a1
CL
1477}
1478
1479/*
a49bd4d7
MH
1480 * Resolves the given address to a struct page, isolates it from the LRU and
1481 * puts it to the given pagelist.
1482 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1483 * queued or the page doesn't need to be migrated because it is already on
1484 * the target node
742755a1 1485 */
a49bd4d7
MH
1486static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1487 int node, struct list_head *pagelist, bool migrate_all)
742755a1 1488{
a49bd4d7
MH
1489 struct vm_area_struct *vma;
1490 struct page *page;
1491 unsigned int follflags;
742755a1 1492 int err;
742755a1
CL
1493
1494 down_read(&mm->mmap_sem);
a49bd4d7
MH
1495 err = -EFAULT;
1496 vma = find_vma(mm, addr);
1497 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1498 goto out;
742755a1 1499
a49bd4d7
MH
1500 /* FOLL_DUMP to ignore special (like zero) pages */
1501 follflags = FOLL_GET | FOLL_DUMP;
a49bd4d7 1502 page = follow_page(vma, addr, follflags);
89f5b7da 1503
a49bd4d7
MH
1504 err = PTR_ERR(page);
1505 if (IS_ERR(page))
1506 goto out;
89f5b7da 1507
a49bd4d7
MH
1508 err = -ENOENT;
1509 if (!page)
1510 goto out;
742755a1 1511
a49bd4d7
MH
1512 err = 0;
1513 if (page_to_nid(page) == node)
1514 goto out_putpage;
742755a1 1515
a49bd4d7
MH
1516 err = -EACCES;
1517 if (page_mapcount(page) > 1 && !migrate_all)
1518 goto out_putpage;
742755a1 1519
a49bd4d7
MH
1520 if (PageHuge(page)) {
1521 if (PageHead(page)) {
1522 isolate_huge_page(page, pagelist);
1523 err = 0;
e632a938 1524 }
a49bd4d7
MH
1525 } else {
1526 struct page *head;
e632a938 1527
e8db67eb
NH
1528 head = compound_head(page);
1529 err = isolate_lru_page(head);
cf608ac1 1530 if (err)
a49bd4d7 1531 goto out_putpage;
742755a1 1532
a49bd4d7
MH
1533 err = 0;
1534 list_add_tail(&head->lru, pagelist);
1535 mod_node_page_state(page_pgdat(head),
1536 NR_ISOLATED_ANON + page_is_file_cache(head),
1537 hpage_nr_pages(head));
1538 }
1539out_putpage:
1540 /*
1541 * Either remove the duplicate refcount from
1542 * isolate_lru_page() or drop the page ref if it was
1543 * not isolated.
1544 */
1545 put_page(page);
1546out:
742755a1
CL
1547 up_read(&mm->mmap_sem);
1548 return err;
1549}
1550
5e9a0f02
BG
1551/*
1552 * Migrate an array of page address onto an array of nodes and fill
1553 * the corresponding array of status.
1554 */
3268c63e 1555static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1556 unsigned long nr_pages,
1557 const void __user * __user *pages,
1558 const int __user *nodes,
1559 int __user *status, int flags)
1560{
a49bd4d7
MH
1561 int current_node = NUMA_NO_NODE;
1562 LIST_HEAD(pagelist);
1563 int start, i;
1564 int err = 0, err1;
35282a2d
BG
1565
1566 migrate_prep();
1567
a49bd4d7
MH
1568 for (i = start = 0; i < nr_pages; i++) {
1569 const void __user *p;
1570 unsigned long addr;
1571 int node;
3140a227 1572
a49bd4d7
MH
1573 err = -EFAULT;
1574 if (get_user(p, pages + i))
1575 goto out_flush;
1576 if (get_user(node, nodes + i))
1577 goto out_flush;
1578 addr = (unsigned long)p;
1579
1580 err = -ENODEV;
1581 if (node < 0 || node >= MAX_NUMNODES)
1582 goto out_flush;
1583 if (!node_state(node, N_MEMORY))
1584 goto out_flush;
5e9a0f02 1585
a49bd4d7
MH
1586 err = -EACCES;
1587 if (!node_isset(node, task_nodes))
1588 goto out_flush;
1589
1590 if (current_node == NUMA_NO_NODE) {
1591 current_node = node;
1592 start = i;
1593 } else if (node != current_node) {
1594 err = do_move_pages_to_node(mm, &pagelist, current_node);
1595 if (err)
1596 goto out;
1597 err = store_status(status, start, current_node, i - start);
1598 if (err)
1599 goto out;
1600 start = i;
1601 current_node = node;
3140a227
BG
1602 }
1603
a49bd4d7
MH
1604 /*
1605 * Errors in the page lookup or isolation are not fatal and we simply
1606 * report them via status
1607 */
1608 err = add_page_for_migration(mm, addr, current_node,
1609 &pagelist, flags & MPOL_MF_MOVE_ALL);
1610 if (!err)
1611 continue;
3140a227 1612
a49bd4d7
MH
1613 err = store_status(status, i, err, 1);
1614 if (err)
1615 goto out_flush;
5e9a0f02 1616
a49bd4d7
MH
1617 err = do_move_pages_to_node(mm, &pagelist, current_node);
1618 if (err)
1619 goto out;
1620 if (i > start) {
1621 err = store_status(status, start, current_node, i - start);
1622 if (err)
1623 goto out;
1624 }
1625 current_node = NUMA_NO_NODE;
3140a227 1626 }
a49bd4d7 1627out_flush:
8f175cf5
MH
1628 if (list_empty(&pagelist))
1629 return err;
1630
a49bd4d7
MH
1631 /* Make sure we do not overwrite the existing error */
1632 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1633 if (!err1)
1634 err1 = store_status(status, start, current_node, i - start);
1635 if (!err)
1636 err = err1;
5e9a0f02
BG
1637out:
1638 return err;
1639}
1640
742755a1 1641/*
2f007e74 1642 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1643 */
80bba129
BG
1644static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1645 const void __user **pages, int *status)
742755a1 1646{
2f007e74 1647 unsigned long i;
2f007e74 1648
742755a1
CL
1649 down_read(&mm->mmap_sem);
1650
2f007e74 1651 for (i = 0; i < nr_pages; i++) {
80bba129 1652 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1653 struct vm_area_struct *vma;
1654 struct page *page;
c095adbc 1655 int err = -EFAULT;
2f007e74
BG
1656
1657 vma = find_vma(mm, addr);
70384dc6 1658 if (!vma || addr < vma->vm_start)
742755a1
CL
1659 goto set_status;
1660
d899844e
KS
1661 /* FOLL_DUMP to ignore special (like zero) pages */
1662 page = follow_page(vma, addr, FOLL_DUMP);
89f5b7da
LT
1663
1664 err = PTR_ERR(page);
1665 if (IS_ERR(page))
1666 goto set_status;
1667
d899844e 1668 err = page ? page_to_nid(page) : -ENOENT;
742755a1 1669set_status:
80bba129
BG
1670 *status = err;
1671
1672 pages++;
1673 status++;
1674 }
1675
1676 up_read(&mm->mmap_sem);
1677}
1678
1679/*
1680 * Determine the nodes of a user array of pages and store it in
1681 * a user array of status.
1682 */
1683static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1684 const void __user * __user *pages,
1685 int __user *status)
1686{
1687#define DO_PAGES_STAT_CHUNK_NR 16
1688 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1689 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1690
87b8d1ad
PA
1691 while (nr_pages) {
1692 unsigned long chunk_nr;
80bba129 1693
87b8d1ad
PA
1694 chunk_nr = nr_pages;
1695 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1696 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1697
1698 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1699 break;
80bba129
BG
1700
1701 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1702
87b8d1ad
PA
1703 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1704 break;
742755a1 1705
87b8d1ad
PA
1706 pages += chunk_nr;
1707 status += chunk_nr;
1708 nr_pages -= chunk_nr;
1709 }
1710 return nr_pages ? -EFAULT : 0;
742755a1
CL
1711}
1712
1713/*
1714 * Move a list of pages in the address space of the currently executing
1715 * process.
1716 */
7addf443
DB
1717static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1718 const void __user * __user *pages,
1719 const int __user *nodes,
1720 int __user *status, int flags)
742755a1 1721{
742755a1 1722 struct task_struct *task;
742755a1 1723 struct mm_struct *mm;
5e9a0f02 1724 int err;
3268c63e 1725 nodemask_t task_nodes;
742755a1
CL
1726
1727 /* Check flags */
1728 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1729 return -EINVAL;
1730
1731 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1732 return -EPERM;
1733
1734 /* Find the mm_struct */
a879bf58 1735 rcu_read_lock();
228ebcbe 1736 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1737 if (!task) {
a879bf58 1738 rcu_read_unlock();
742755a1
CL
1739 return -ESRCH;
1740 }
3268c63e 1741 get_task_struct(task);
742755a1
CL
1742
1743 /*
1744 * Check if this process has the right to modify the specified
197e7e52 1745 * process. Use the regular "ptrace_may_access()" checks.
742755a1 1746 */
197e7e52 1747 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
c69e8d9c 1748 rcu_read_unlock();
742755a1 1749 err = -EPERM;
5e9a0f02 1750 goto out;
742755a1 1751 }
c69e8d9c 1752 rcu_read_unlock();
742755a1 1753
86c3a764
DQ
1754 err = security_task_movememory(task);
1755 if (err)
5e9a0f02 1756 goto out;
86c3a764 1757
3268c63e
CL
1758 task_nodes = cpuset_mems_allowed(task);
1759 mm = get_task_mm(task);
1760 put_task_struct(task);
1761
6e8b09ea
SL
1762 if (!mm)
1763 return -EINVAL;
1764
1765 if (nodes)
1766 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1767 nodes, status, flags);
1768 else
1769 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1770
742755a1
CL
1771 mmput(mm);
1772 return err;
3268c63e
CL
1773
1774out:
1775 put_task_struct(task);
1776 return err;
742755a1 1777}
742755a1 1778
7addf443
DB
1779SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1780 const void __user * __user *, pages,
1781 const int __user *, nodes,
1782 int __user *, status, int, flags)
1783{
1784 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1785}
1786
1787#ifdef CONFIG_COMPAT
1788COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1789 compat_uptr_t __user *, pages32,
1790 const int __user *, nodes,
1791 int __user *, status,
1792 int, flags)
1793{
1794 const void __user * __user *pages;
1795 int i;
1796
1797 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1798 for (i = 0; i < nr_pages; i++) {
1799 compat_uptr_t p;
1800
1801 if (get_user(p, pages32 + i) ||
1802 put_user(compat_ptr(p), pages + i))
1803 return -EFAULT;
1804 }
1805 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1806}
1807#endif /* CONFIG_COMPAT */
1808
7039e1db
PZ
1809#ifdef CONFIG_NUMA_BALANCING
1810/*
1811 * Returns true if this is a safe migration target node for misplaced NUMA
1812 * pages. Currently it only checks the watermarks which crude
1813 */
1814static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1815 unsigned long nr_migrate_pages)
7039e1db
PZ
1816{
1817 int z;
599d0c95 1818
7039e1db
PZ
1819 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1820 struct zone *zone = pgdat->node_zones + z;
1821
1822 if (!populated_zone(zone))
1823 continue;
1824
7039e1db
PZ
1825 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1826 if (!zone_watermark_ok(zone, 0,
1827 high_wmark_pages(zone) +
1828 nr_migrate_pages,
1829 0, 0))
1830 continue;
1831 return true;
1832 }
1833 return false;
1834}
1835
1836static struct page *alloc_misplaced_dst_page(struct page *page,
666feb21 1837 unsigned long data)
7039e1db
PZ
1838{
1839 int nid = (int) data;
1840 struct page *newpage;
1841
96db800f 1842 newpage = __alloc_pages_node(nid,
e97ca8e5
JW
1843 (GFP_HIGHUSER_MOVABLE |
1844 __GFP_THISNODE | __GFP_NOMEMALLOC |
1845 __GFP_NORETRY | __GFP_NOWARN) &
8479eba7 1846 ~__GFP_RECLAIM, 0);
bac0382c 1847
7039e1db
PZ
1848 return newpage;
1849}
1850
1c30e017 1851static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
b32967ff 1852{
340ef390 1853 int page_lru;
a8f60772 1854
309381fe 1855 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
3abef4e6 1856
7039e1db 1857 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1858 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1859 return 0;
7039e1db 1860
340ef390
HD
1861 if (isolate_lru_page(page))
1862 return 0;
7039e1db 1863
340ef390
HD
1864 /*
1865 * migrate_misplaced_transhuge_page() skips page migration's usual
1866 * check on page_count(), so we must do it here, now that the page
1867 * has been isolated: a GUP pin, or any other pin, prevents migration.
1868 * The expected page count is 3: 1 for page's mapcount and 1 for the
1869 * caller's pin and 1 for the reference taken by isolate_lru_page().
1870 */
1871 if (PageTransHuge(page) && page_count(page) != 3) {
1872 putback_lru_page(page);
1873 return 0;
7039e1db
PZ
1874 }
1875
340ef390 1876 page_lru = page_is_file_cache(page);
599d0c95 1877 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
340ef390
HD
1878 hpage_nr_pages(page));
1879
149c33e1 1880 /*
340ef390
HD
1881 * Isolating the page has taken another reference, so the
1882 * caller's reference can be safely dropped without the page
1883 * disappearing underneath us during migration.
149c33e1
MG
1884 */
1885 put_page(page);
340ef390 1886 return 1;
b32967ff
MG
1887}
1888
de466bd6
MG
1889bool pmd_trans_migrating(pmd_t pmd)
1890{
1891 struct page *page = pmd_page(pmd);
1892 return PageLocked(page);
1893}
1894
b32967ff
MG
1895/*
1896 * Attempt to migrate a misplaced page to the specified destination
1897 * node. Caller is expected to have an elevated reference count on
1898 * the page that will be dropped by this function before returning.
1899 */
1bc115d8
MG
1900int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1901 int node)
b32967ff
MG
1902{
1903 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1904 int isolated;
b32967ff
MG
1905 int nr_remaining;
1906 LIST_HEAD(migratepages);
1907
1908 /*
1bc115d8
MG
1909 * Don't migrate file pages that are mapped in multiple processes
1910 * with execute permissions as they are probably shared libraries.
b32967ff 1911 */
1bc115d8
MG
1912 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1913 (vma->vm_flags & VM_EXEC))
b32967ff 1914 goto out;
b32967ff 1915
09a913a7
MG
1916 /*
1917 * Also do not migrate dirty pages as not all filesystems can move
1918 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1919 */
1920 if (page_is_file_cache(page) && PageDirty(page))
1921 goto out;
1922
b32967ff
MG
1923 isolated = numamigrate_isolate_page(pgdat, page);
1924 if (!isolated)
1925 goto out;
1926
1927 list_add(&page->lru, &migratepages);
9c620e2b 1928 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
68711a74
DR
1929 NULL, node, MIGRATE_ASYNC,
1930 MR_NUMA_MISPLACED);
b32967ff 1931 if (nr_remaining) {
59c82b70
JK
1932 if (!list_empty(&migratepages)) {
1933 list_del(&page->lru);
599d0c95 1934 dec_node_page_state(page, NR_ISOLATED_ANON +
59c82b70
JK
1935 page_is_file_cache(page));
1936 putback_lru_page(page);
1937 }
b32967ff
MG
1938 isolated = 0;
1939 } else
1940 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1941 BUG_ON(!list_empty(&migratepages));
7039e1db 1942 return isolated;
340ef390
HD
1943
1944out:
1945 put_page(page);
1946 return 0;
7039e1db 1947}
220018d3 1948#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1949
220018d3 1950#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1951/*
1952 * Migrates a THP to a given target node. page must be locked and is unlocked
1953 * before returning.
1954 */
b32967ff
MG
1955int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1956 struct vm_area_struct *vma,
1957 pmd_t *pmd, pmd_t entry,
1958 unsigned long address,
1959 struct page *page, int node)
1960{
c4088ebd 1961 spinlock_t *ptl;
b32967ff
MG
1962 pg_data_t *pgdat = NODE_DATA(node);
1963 int isolated = 0;
1964 struct page *new_page = NULL;
b32967ff 1965 int page_lru = page_is_file_cache(page);
7066f0f9 1966 unsigned long start = address & HPAGE_PMD_MASK;
b32967ff 1967
b32967ff 1968 new_page = alloc_pages_node(node,
25160354 1969 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
e97ca8e5 1970 HPAGE_PMD_ORDER);
340ef390
HD
1971 if (!new_page)
1972 goto out_fail;
9a982250 1973 prep_transhuge_page(new_page);
340ef390 1974
b32967ff 1975 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 1976 if (!isolated) {
b32967ff 1977 put_page(new_page);
340ef390 1978 goto out_fail;
b32967ff 1979 }
b0943d61 1980
b32967ff 1981 /* Prepare a page as a migration target */
48c935ad 1982 __SetPageLocked(new_page);
d44d363f
SL
1983 if (PageSwapBacked(page))
1984 __SetPageSwapBacked(new_page);
b32967ff
MG
1985
1986 /* anon mapping, we can simply copy page->mapping to the new page: */
1987 new_page->mapping = page->mapping;
1988 new_page->index = page->index;
7eef5f97
AA
1989 /* flush the cache before copying using the kernel virtual address */
1990 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
b32967ff
MG
1991 migrate_page_copy(new_page, page);
1992 WARN_ON(PageLRU(new_page));
1993
1994 /* Recheck the target PMD */
c4088ebd 1995 ptl = pmd_lock(mm, pmd);
f4e177d1 1996 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
c4088ebd 1997 spin_unlock(ptl);
b32967ff
MG
1998
1999 /* Reverse changes made by migrate_page_copy() */
2000 if (TestClearPageActive(new_page))
2001 SetPageActive(page);
2002 if (TestClearPageUnevictable(new_page))
2003 SetPageUnevictable(page);
b32967ff
MG
2004
2005 unlock_page(new_page);
2006 put_page(new_page); /* Free it */
2007
a54a407f
MG
2008 /* Retake the callers reference and putback on LRU */
2009 get_page(page);
b32967ff 2010 putback_lru_page(page);
599d0c95 2011 mod_node_page_state(page_pgdat(page),
a54a407f 2012 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
eb4489f6
MG
2013
2014 goto out_unlock;
b32967ff
MG
2015 }
2016
10102459 2017 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
f55e1014 2018 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
b32967ff 2019
2b4847e7 2020 /*
d7c33934
AA
2021 * Overwrite the old entry under pagetable lock and establish
2022 * the new PTE. Any parallel GUP will either observe the old
2023 * page blocking on the page lock, block on the page table
2024 * lock or observe the new page. The SetPageUptodate on the
2025 * new page and page_add_new_anon_rmap guarantee the copy is
2026 * visible before the pagetable update.
2b4847e7 2027 */
7066f0f9 2028 page_add_anon_rmap(new_page, vma, start, true);
d7c33934
AA
2029 /*
2030 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2031 * has already been flushed globally. So no TLB can be currently
2032 * caching this non present pmd mapping. There's no need to clear the
2033 * pmd before doing set_pmd_at(), nor to flush the TLB after
2034 * set_pmd_at(). Clearing the pmd here would introduce a race
2035 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2036 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2037 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2038 * pmd.
2039 */
7066f0f9 2040 set_pmd_at(mm, start, pmd, entry);
ce4a9cc5 2041 update_mmu_cache_pmd(vma, address, &entry);
2b4847e7 2042
f4e177d1 2043 page_ref_unfreeze(page, 2);
51afb12b 2044 mlock_migrate_page(new_page, page);
d281ee61 2045 page_remove_rmap(page, true);
7cd12b4a 2046 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2b4847e7 2047
c4088ebd 2048 spin_unlock(ptl);
b32967ff 2049
11de9927
MG
2050 /* Take an "isolate" reference and put new page on the LRU. */
2051 get_page(new_page);
2052 putback_lru_page(new_page);
2053
b32967ff
MG
2054 unlock_page(new_page);
2055 unlock_page(page);
2056 put_page(page); /* Drop the rmap reference */
2057 put_page(page); /* Drop the LRU isolation reference */
2058
2059 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2060 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2061
599d0c95 2062 mod_node_page_state(page_pgdat(page),
b32967ff
MG
2063 NR_ISOLATED_ANON + page_lru,
2064 -HPAGE_PMD_NR);
2065 return isolated;
2066
340ef390
HD
2067out_fail:
2068 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2b4847e7
MG
2069 ptl = pmd_lock(mm, pmd);
2070 if (pmd_same(*pmd, entry)) {
4d942466 2071 entry = pmd_modify(entry, vma->vm_page_prot);
7066f0f9 2072 set_pmd_at(mm, start, pmd, entry);
2b4847e7
MG
2073 update_mmu_cache_pmd(vma, address, &entry);
2074 }
2075 spin_unlock(ptl);
a54a407f 2076
eb4489f6 2077out_unlock:
340ef390 2078 unlock_page(page);
b32967ff 2079 put_page(page);
b32967ff
MG
2080 return 0;
2081}
7039e1db
PZ
2082#endif /* CONFIG_NUMA_BALANCING */
2083
2084#endif /* CONFIG_NUMA */
8763cb45 2085
6b368cd4 2086#if defined(CONFIG_MIGRATE_VMA_HELPER)
8763cb45
JG
2087struct migrate_vma {
2088 struct vm_area_struct *vma;
2089 unsigned long *dst;
2090 unsigned long *src;
2091 unsigned long cpages;
2092 unsigned long npages;
2093 unsigned long start;
2094 unsigned long end;
2095};
2096
2097static int migrate_vma_collect_hole(unsigned long start,
2098 unsigned long end,
2099 struct mm_walk *walk)
2100{
2101 struct migrate_vma *migrate = walk->private;
2102 unsigned long addr;
2103
8315ada7 2104 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
e20d103b 2105 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
8315ada7 2106 migrate->dst[migrate->npages] = 0;
e20d103b 2107 migrate->npages++;
8315ada7
JG
2108 migrate->cpages++;
2109 }
2110
2111 return 0;
2112}
2113
2114static int migrate_vma_collect_skip(unsigned long start,
2115 unsigned long end,
2116 struct mm_walk *walk)
2117{
2118 struct migrate_vma *migrate = walk->private;
2119 unsigned long addr;
2120
8763cb45
JG
2121 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2122 migrate->dst[migrate->npages] = 0;
2123 migrate->src[migrate->npages++] = 0;
2124 }
2125
2126 return 0;
2127}
2128
2129static int migrate_vma_collect_pmd(pmd_t *pmdp,
2130 unsigned long start,
2131 unsigned long end,
2132 struct mm_walk *walk)
2133{
2134 struct migrate_vma *migrate = walk->private;
2135 struct vm_area_struct *vma = walk->vma;
2136 struct mm_struct *mm = vma->vm_mm;
8c3328f1 2137 unsigned long addr = start, unmapped = 0;
8763cb45
JG
2138 spinlock_t *ptl;
2139 pte_t *ptep;
2140
2141again:
2142 if (pmd_none(*pmdp))
2143 return migrate_vma_collect_hole(start, end, walk);
2144
2145 if (pmd_trans_huge(*pmdp)) {
2146 struct page *page;
2147
2148 ptl = pmd_lock(mm, pmdp);
2149 if (unlikely(!pmd_trans_huge(*pmdp))) {
2150 spin_unlock(ptl);
2151 goto again;
2152 }
2153
2154 page = pmd_page(*pmdp);
2155 if (is_huge_zero_page(page)) {
2156 spin_unlock(ptl);
2157 split_huge_pmd(vma, pmdp, addr);
2158 if (pmd_trans_unstable(pmdp))
8315ada7 2159 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2160 walk);
2161 } else {
2162 int ret;
2163
2164 get_page(page);
2165 spin_unlock(ptl);
2166 if (unlikely(!trylock_page(page)))
8315ada7 2167 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2168 walk);
2169 ret = split_huge_page(page);
2170 unlock_page(page);
2171 put_page(page);
8315ada7
JG
2172 if (ret)
2173 return migrate_vma_collect_skip(start, end,
2174 walk);
2175 if (pmd_none(*pmdp))
8763cb45
JG
2176 return migrate_vma_collect_hole(start, end,
2177 walk);
2178 }
2179 }
2180
2181 if (unlikely(pmd_bad(*pmdp)))
8315ada7 2182 return migrate_vma_collect_skip(start, end, walk);
8763cb45
JG
2183
2184 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
8c3328f1
JG
2185 arch_enter_lazy_mmu_mode();
2186
8763cb45
JG
2187 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2188 unsigned long mpfn, pfn;
2189 struct page *page;
8c3328f1 2190 swp_entry_t entry;
8763cb45
JG
2191 pte_t pte;
2192
2193 pte = *ptep;
2194 pfn = pte_pfn(pte);
2195
a5430dda 2196 if (pte_none(pte)) {
8315ada7
JG
2197 mpfn = MIGRATE_PFN_MIGRATE;
2198 migrate->cpages++;
2199 pfn = 0;
8763cb45
JG
2200 goto next;
2201 }
2202
a5430dda
JG
2203 if (!pte_present(pte)) {
2204 mpfn = pfn = 0;
2205
2206 /*
2207 * Only care about unaddressable device page special
2208 * page table entry. Other special swap entries are not
2209 * migratable, and we ignore regular swapped page.
2210 */
2211 entry = pte_to_swp_entry(pte);
2212 if (!is_device_private_entry(entry))
2213 goto next;
2214
2215 page = device_private_entry_to_page(entry);
2216 mpfn = migrate_pfn(page_to_pfn(page))|
2217 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2218 if (is_write_device_private_entry(entry))
2219 mpfn |= MIGRATE_PFN_WRITE;
2220 } else {
8315ada7
JG
2221 if (is_zero_pfn(pfn)) {
2222 mpfn = MIGRATE_PFN_MIGRATE;
2223 migrate->cpages++;
2224 pfn = 0;
2225 goto next;
2226 }
df6ad698 2227 page = _vm_normal_page(migrate->vma, addr, pte, true);
a5430dda
JG
2228 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2229 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2230 }
2231
8763cb45 2232 /* FIXME support THP */
8763cb45
JG
2233 if (!page || !page->mapping || PageTransCompound(page)) {
2234 mpfn = pfn = 0;
2235 goto next;
2236 }
a5430dda 2237 pfn = page_to_pfn(page);
8763cb45
JG
2238
2239 /*
2240 * By getting a reference on the page we pin it and that blocks
2241 * any kind of migration. Side effect is that it "freezes" the
2242 * pte.
2243 *
2244 * We drop this reference after isolating the page from the lru
2245 * for non device page (device page are not on the lru and thus
2246 * can't be dropped from it).
2247 */
2248 get_page(page);
2249 migrate->cpages++;
8763cb45 2250
8c3328f1
JG
2251 /*
2252 * Optimize for the common case where page is only mapped once
2253 * in one process. If we can lock the page, then we can safely
2254 * set up a special migration page table entry now.
2255 */
2256 if (trylock_page(page)) {
2257 pte_t swp_pte;
2258
2259 mpfn |= MIGRATE_PFN_LOCKED;
2260 ptep_get_and_clear(mm, addr, ptep);
2261
2262 /* Setup special migration page table entry */
07707125
RC
2263 entry = make_migration_entry(page, mpfn &
2264 MIGRATE_PFN_WRITE);
8c3328f1
JG
2265 swp_pte = swp_entry_to_pte(entry);
2266 if (pte_soft_dirty(pte))
2267 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2268 set_pte_at(mm, addr, ptep, swp_pte);
2269
2270 /*
2271 * This is like regular unmap: we remove the rmap and
2272 * drop page refcount. Page won't be freed, as we took
2273 * a reference just above.
2274 */
2275 page_remove_rmap(page, false);
2276 put_page(page);
a5430dda
JG
2277
2278 if (pte_present(pte))
2279 unmapped++;
8c3328f1
JG
2280 }
2281
8763cb45 2282next:
a5430dda 2283 migrate->dst[migrate->npages] = 0;
8763cb45
JG
2284 migrate->src[migrate->npages++] = mpfn;
2285 }
8c3328f1 2286 arch_leave_lazy_mmu_mode();
8763cb45
JG
2287 pte_unmap_unlock(ptep - 1, ptl);
2288
8c3328f1
JG
2289 /* Only flush the TLB if we actually modified any entries */
2290 if (unmapped)
2291 flush_tlb_range(walk->vma, start, end);
2292
8763cb45
JG
2293 return 0;
2294}
2295
2296/*
2297 * migrate_vma_collect() - collect pages over a range of virtual addresses
2298 * @migrate: migrate struct containing all migration information
2299 *
2300 * This will walk the CPU page table. For each virtual address backed by a
2301 * valid page, it updates the src array and takes a reference on the page, in
2302 * order to pin the page until we lock it and unmap it.
2303 */
2304static void migrate_vma_collect(struct migrate_vma *migrate)
2305{
2306 struct mm_walk mm_walk;
2307
2308 mm_walk.pmd_entry = migrate_vma_collect_pmd;
2309 mm_walk.pte_entry = NULL;
2310 mm_walk.pte_hole = migrate_vma_collect_hole;
2311 mm_walk.hugetlb_entry = NULL;
2312 mm_walk.test_walk = NULL;
2313 mm_walk.vma = migrate->vma;
2314 mm_walk.mm = migrate->vma->vm_mm;
2315 mm_walk.private = migrate;
2316
8c3328f1
JG
2317 mmu_notifier_invalidate_range_start(mm_walk.mm,
2318 migrate->start,
2319 migrate->end);
8763cb45 2320 walk_page_range(migrate->start, migrate->end, &mm_walk);
8c3328f1
JG
2321 mmu_notifier_invalidate_range_end(mm_walk.mm,
2322 migrate->start,
2323 migrate->end);
8763cb45
JG
2324
2325 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2326}
2327
2328/*
2329 * migrate_vma_check_page() - check if page is pinned or not
2330 * @page: struct page to check
2331 *
2332 * Pinned pages cannot be migrated. This is the same test as in
2333 * migrate_page_move_mapping(), except that here we allow migration of a
2334 * ZONE_DEVICE page.
2335 */
2336static bool migrate_vma_check_page(struct page *page)
2337{
2338 /*
2339 * One extra ref because caller holds an extra reference, either from
2340 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2341 * a device page.
2342 */
2343 int extra = 1;
2344
2345 /*
2346 * FIXME support THP (transparent huge page), it is bit more complex to
2347 * check them than regular pages, because they can be mapped with a pmd
2348 * or with a pte (split pte mapping).
2349 */
2350 if (PageCompound(page))
2351 return false;
2352
a5430dda
JG
2353 /* Page from ZONE_DEVICE have one extra reference */
2354 if (is_zone_device_page(page)) {
2355 /*
2356 * Private page can never be pin as they have no valid pte and
2357 * GUP will fail for those. Yet if there is a pending migration
2358 * a thread might try to wait on the pte migration entry and
2359 * will bump the page reference count. Sadly there is no way to
2360 * differentiate a regular pin from migration wait. Hence to
2361 * avoid 2 racing thread trying to migrate back to CPU to enter
2362 * infinite loop (one stoping migration because the other is
2363 * waiting on pte migration entry). We always return true here.
2364 *
2365 * FIXME proper solution is to rework migration_entry_wait() so
2366 * it does not need to take a reference on page.
2367 */
2368 if (is_device_private_page(page))
2369 return true;
2370
df6ad698
JG
2371 /*
2372 * Only allow device public page to be migrated and account for
2373 * the extra reference count imply by ZONE_DEVICE pages.
2374 */
2375 if (!is_device_public_page(page))
2376 return false;
2377 extra++;
a5430dda
JG
2378 }
2379
df6ad698
JG
2380 /* For file back page */
2381 if (page_mapping(page))
2382 extra += 1 + page_has_private(page);
2383
8763cb45
JG
2384 if ((page_count(page) - extra) > page_mapcount(page))
2385 return false;
2386
2387 return true;
2388}
2389
2390/*
2391 * migrate_vma_prepare() - lock pages and isolate them from the lru
2392 * @migrate: migrate struct containing all migration information
2393 *
2394 * This locks pages that have been collected by migrate_vma_collect(). Once each
2395 * page is locked it is isolated from the lru (for non-device pages). Finally,
2396 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2397 * migrated by concurrent kernel threads.
2398 */
2399static void migrate_vma_prepare(struct migrate_vma *migrate)
2400{
2401 const unsigned long npages = migrate->npages;
8c3328f1
JG
2402 const unsigned long start = migrate->start;
2403 unsigned long addr, i, restore = 0;
8763cb45 2404 bool allow_drain = true;
8763cb45
JG
2405
2406 lru_add_drain();
2407
2408 for (i = 0; (i < npages) && migrate->cpages; i++) {
2409 struct page *page = migrate_pfn_to_page(migrate->src[i]);
8c3328f1 2410 bool remap = true;
8763cb45
JG
2411
2412 if (!page)
2413 continue;
2414
8c3328f1
JG
2415 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2416 /*
2417 * Because we are migrating several pages there can be
2418 * a deadlock between 2 concurrent migration where each
2419 * are waiting on each other page lock.
2420 *
2421 * Make migrate_vma() a best effort thing and backoff
2422 * for any page we can not lock right away.
2423 */
2424 if (!trylock_page(page)) {
2425 migrate->src[i] = 0;
2426 migrate->cpages--;
2427 put_page(page);
2428 continue;
2429 }
2430 remap = false;
2431 migrate->src[i] |= MIGRATE_PFN_LOCKED;
8763cb45 2432 }
8763cb45 2433
a5430dda
JG
2434 /* ZONE_DEVICE pages are not on LRU */
2435 if (!is_zone_device_page(page)) {
2436 if (!PageLRU(page) && allow_drain) {
2437 /* Drain CPU's pagevec */
2438 lru_add_drain_all();
2439 allow_drain = false;
2440 }
8763cb45 2441
a5430dda
JG
2442 if (isolate_lru_page(page)) {
2443 if (remap) {
2444 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2445 migrate->cpages--;
2446 restore++;
2447 } else {
2448 migrate->src[i] = 0;
2449 unlock_page(page);
2450 migrate->cpages--;
2451 put_page(page);
2452 }
2453 continue;
8c3328f1 2454 }
a5430dda
JG
2455
2456 /* Drop the reference we took in collect */
2457 put_page(page);
8763cb45
JG
2458 }
2459
2460 if (!migrate_vma_check_page(page)) {
8c3328f1
JG
2461 if (remap) {
2462 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2463 migrate->cpages--;
2464 restore++;
8763cb45 2465
a5430dda
JG
2466 if (!is_zone_device_page(page)) {
2467 get_page(page);
2468 putback_lru_page(page);
2469 }
8c3328f1
JG
2470 } else {
2471 migrate->src[i] = 0;
2472 unlock_page(page);
2473 migrate->cpages--;
2474
a5430dda
JG
2475 if (!is_zone_device_page(page))
2476 putback_lru_page(page);
2477 else
2478 put_page(page);
8c3328f1 2479 }
8763cb45
JG
2480 }
2481 }
8c3328f1
JG
2482
2483 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2484 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2485
2486 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2487 continue;
2488
2489 remove_migration_pte(page, migrate->vma, addr, page);
2490
2491 migrate->src[i] = 0;
2492 unlock_page(page);
2493 put_page(page);
2494 restore--;
2495 }
8763cb45
JG
2496}
2497
2498/*
2499 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2500 * @migrate: migrate struct containing all migration information
2501 *
2502 * Replace page mapping (CPU page table pte) with a special migration pte entry
2503 * and check again if it has been pinned. Pinned pages are restored because we
2504 * cannot migrate them.
2505 *
2506 * This is the last step before we call the device driver callback to allocate
2507 * destination memory and copy contents of original page over to new page.
2508 */
2509static void migrate_vma_unmap(struct migrate_vma *migrate)
2510{
2511 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2512 const unsigned long npages = migrate->npages;
2513 const unsigned long start = migrate->start;
2514 unsigned long addr, i, restore = 0;
2515
2516 for (i = 0; i < npages; i++) {
2517 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2518
2519 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2520 continue;
2521
8c3328f1
JG
2522 if (page_mapped(page)) {
2523 try_to_unmap(page, flags);
2524 if (page_mapped(page))
2525 goto restore;
8763cb45 2526 }
8c3328f1
JG
2527
2528 if (migrate_vma_check_page(page))
2529 continue;
2530
2531restore:
2532 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2533 migrate->cpages--;
2534 restore++;
8763cb45
JG
2535 }
2536
2537 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2538 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2539
2540 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2541 continue;
2542
2543 remove_migration_ptes(page, page, false);
2544
2545 migrate->src[i] = 0;
2546 unlock_page(page);
2547 restore--;
2548
a5430dda
JG
2549 if (is_zone_device_page(page))
2550 put_page(page);
2551 else
2552 putback_lru_page(page);
8763cb45
JG
2553 }
2554}
2555
8315ada7
JG
2556static void migrate_vma_insert_page(struct migrate_vma *migrate,
2557 unsigned long addr,
2558 struct page *page,
2559 unsigned long *src,
2560 unsigned long *dst)
2561{
2562 struct vm_area_struct *vma = migrate->vma;
2563 struct mm_struct *mm = vma->vm_mm;
2564 struct mem_cgroup *memcg;
2565 bool flush = false;
2566 spinlock_t *ptl;
2567 pte_t entry;
2568 pgd_t *pgdp;
2569 p4d_t *p4dp;
2570 pud_t *pudp;
2571 pmd_t *pmdp;
2572 pte_t *ptep;
2573
2574 /* Only allow populating anonymous memory */
2575 if (!vma_is_anonymous(vma))
2576 goto abort;
2577
2578 pgdp = pgd_offset(mm, addr);
2579 p4dp = p4d_alloc(mm, pgdp, addr);
2580 if (!p4dp)
2581 goto abort;
2582 pudp = pud_alloc(mm, p4dp, addr);
2583 if (!pudp)
2584 goto abort;
2585 pmdp = pmd_alloc(mm, pudp, addr);
2586 if (!pmdp)
2587 goto abort;
2588
2589 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2590 goto abort;
2591
2592 /*
2593 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2594 * pte_offset_map() on pmds where a huge pmd might be created
2595 * from a different thread.
2596 *
2597 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2598 * parallel threads are excluded by other means.
2599 *
2600 * Here we only have down_read(mmap_sem).
2601 */
2602 if (pte_alloc(mm, pmdp, addr))
2603 goto abort;
2604
2605 /* See the comment in pte_alloc_one_map() */
2606 if (unlikely(pmd_trans_unstable(pmdp)))
2607 goto abort;
2608
2609 if (unlikely(anon_vma_prepare(vma)))
2610 goto abort;
2611 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2612 goto abort;
2613
2614 /*
2615 * The memory barrier inside __SetPageUptodate makes sure that
2616 * preceding stores to the page contents become visible before
2617 * the set_pte_at() write.
2618 */
2619 __SetPageUptodate(page);
2620
df6ad698
JG
2621 if (is_zone_device_page(page)) {
2622 if (is_device_private_page(page)) {
2623 swp_entry_t swp_entry;
2624
2625 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2626 entry = swp_entry_to_pte(swp_entry);
2627 } else if (is_device_public_page(page)) {
2628 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2629 if (vma->vm_flags & VM_WRITE)
2630 entry = pte_mkwrite(pte_mkdirty(entry));
2631 entry = pte_mkdevmap(entry);
2632 }
8315ada7
JG
2633 } else {
2634 entry = mk_pte(page, vma->vm_page_prot);
2635 if (vma->vm_flags & VM_WRITE)
2636 entry = pte_mkwrite(pte_mkdirty(entry));
2637 }
2638
2639 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2640
2641 if (pte_present(*ptep)) {
2642 unsigned long pfn = pte_pfn(*ptep);
2643
2644 if (!is_zero_pfn(pfn)) {
2645 pte_unmap_unlock(ptep, ptl);
2646 mem_cgroup_cancel_charge(page, memcg, false);
2647 goto abort;
2648 }
2649 flush = true;
2650 } else if (!pte_none(*ptep)) {
2651 pte_unmap_unlock(ptep, ptl);
2652 mem_cgroup_cancel_charge(page, memcg, false);
2653 goto abort;
2654 }
2655
2656 /*
2657 * Check for usefaultfd but do not deliver the fault. Instead,
2658 * just back off.
2659 */
2660 if (userfaultfd_missing(vma)) {
2661 pte_unmap_unlock(ptep, ptl);
2662 mem_cgroup_cancel_charge(page, memcg, false);
2663 goto abort;
2664 }
2665
2666 inc_mm_counter(mm, MM_ANONPAGES);
2667 page_add_new_anon_rmap(page, vma, addr, false);
2668 mem_cgroup_commit_charge(page, memcg, false, false);
2669 if (!is_zone_device_page(page))
2670 lru_cache_add_active_or_unevictable(page, vma);
2671 get_page(page);
2672
2673 if (flush) {
2674 flush_cache_page(vma, addr, pte_pfn(*ptep));
2675 ptep_clear_flush_notify(vma, addr, ptep);
2676 set_pte_at_notify(mm, addr, ptep, entry);
2677 update_mmu_cache(vma, addr, ptep);
2678 } else {
2679 /* No need to invalidate - it was non-present before */
2680 set_pte_at(mm, addr, ptep, entry);
2681 update_mmu_cache(vma, addr, ptep);
2682 }
2683
2684 pte_unmap_unlock(ptep, ptl);
2685 *src = MIGRATE_PFN_MIGRATE;
2686 return;
2687
2688abort:
2689 *src &= ~MIGRATE_PFN_MIGRATE;
2690}
2691
8763cb45
JG
2692/*
2693 * migrate_vma_pages() - migrate meta-data from src page to dst page
2694 * @migrate: migrate struct containing all migration information
2695 *
2696 * This migrates struct page meta-data from source struct page to destination
2697 * struct page. This effectively finishes the migration from source page to the
2698 * destination page.
2699 */
2700static void migrate_vma_pages(struct migrate_vma *migrate)
2701{
2702 const unsigned long npages = migrate->npages;
2703 const unsigned long start = migrate->start;
8315ada7
JG
2704 struct vm_area_struct *vma = migrate->vma;
2705 struct mm_struct *mm = vma->vm_mm;
2706 unsigned long addr, i, mmu_start;
2707 bool notified = false;
8763cb45
JG
2708
2709 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2710 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2711 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2712 struct address_space *mapping;
2713 int r;
2714
8315ada7
JG
2715 if (!newpage) {
2716 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
8763cb45 2717 continue;
8315ada7
JG
2718 }
2719
2720 if (!page) {
2721 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2722 continue;
2723 }
2724 if (!notified) {
2725 mmu_start = addr;
2726 notified = true;
2727 mmu_notifier_invalidate_range_start(mm,
2728 mmu_start,
2729 migrate->end);
2730 }
2731 migrate_vma_insert_page(migrate, addr, newpage,
2732 &migrate->src[i],
2733 &migrate->dst[i]);
8763cb45 2734 continue;
8315ada7 2735 }
8763cb45
JG
2736
2737 mapping = page_mapping(page);
2738
a5430dda
JG
2739 if (is_zone_device_page(newpage)) {
2740 if (is_device_private_page(newpage)) {
2741 /*
2742 * For now only support private anonymous when
2743 * migrating to un-addressable device memory.
2744 */
2745 if (mapping) {
2746 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2747 continue;
2748 }
df6ad698 2749 } else if (!is_device_public_page(newpage)) {
a5430dda
JG
2750 /*
2751 * Other types of ZONE_DEVICE page are not
2752 * supported.
2753 */
2754 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2755 continue;
2756 }
2757 }
2758
8763cb45
JG
2759 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2760 if (r != MIGRATEPAGE_SUCCESS)
2761 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2762 }
8315ada7 2763
4645b9fe
JG
2764 /*
2765 * No need to double call mmu_notifier->invalidate_range() callback as
2766 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2767 * did already call it.
2768 */
8315ada7 2769 if (notified)
4645b9fe
JG
2770 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2771 migrate->end);
8763cb45
JG
2772}
2773
2774/*
2775 * migrate_vma_finalize() - restore CPU page table entry
2776 * @migrate: migrate struct containing all migration information
2777 *
2778 * This replaces the special migration pte entry with either a mapping to the
2779 * new page if migration was successful for that page, or to the original page
2780 * otherwise.
2781 *
2782 * This also unlocks the pages and puts them back on the lru, or drops the extra
2783 * refcount, for device pages.
2784 */
2785static void migrate_vma_finalize(struct migrate_vma *migrate)
2786{
2787 const unsigned long npages = migrate->npages;
2788 unsigned long i;
2789
2790 for (i = 0; i < npages; i++) {
2791 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2792 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2793
8315ada7
JG
2794 if (!page) {
2795 if (newpage) {
2796 unlock_page(newpage);
2797 put_page(newpage);
2798 }
8763cb45 2799 continue;
8315ada7
JG
2800 }
2801
8763cb45
JG
2802 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2803 if (newpage) {
2804 unlock_page(newpage);
2805 put_page(newpage);
2806 }
2807 newpage = page;
2808 }
2809
2810 remove_migration_ptes(page, newpage, false);
2811 unlock_page(page);
2812 migrate->cpages--;
2813
a5430dda
JG
2814 if (is_zone_device_page(page))
2815 put_page(page);
2816 else
2817 putback_lru_page(page);
8763cb45
JG
2818
2819 if (newpage != page) {
2820 unlock_page(newpage);
a5430dda
JG
2821 if (is_zone_device_page(newpage))
2822 put_page(newpage);
2823 else
2824 putback_lru_page(newpage);
8763cb45
JG
2825 }
2826 }
2827}
2828
2829/*
2830 * migrate_vma() - migrate a range of memory inside vma
2831 *
2832 * @ops: migration callback for allocating destination memory and copying
2833 * @vma: virtual memory area containing the range to be migrated
2834 * @start: start address of the range to migrate (inclusive)
2835 * @end: end address of the range to migrate (exclusive)
2836 * @src: array of hmm_pfn_t containing source pfns
2837 * @dst: array of hmm_pfn_t containing destination pfns
2838 * @private: pointer passed back to each of the callback
2839 * Returns: 0 on success, error code otherwise
2840 *
2841 * This function tries to migrate a range of memory virtual address range, using
2842 * callbacks to allocate and copy memory from source to destination. First it
2843 * collects all the pages backing each virtual address in the range, saving this
2844 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2845 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2846 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2847 * in the corresponding src array entry. It then restores any pages that are
2848 * pinned, by remapping and unlocking those pages.
2849 *
2850 * At this point it calls the alloc_and_copy() callback. For documentation on
2851 * what is expected from that callback, see struct migrate_vma_ops comments in
2852 * include/linux/migrate.h
2853 *
2854 * After the alloc_and_copy() callback, this function goes over each entry in
2855 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2856 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2857 * then the function tries to migrate struct page information from the source
2858 * struct page to the destination struct page. If it fails to migrate the struct
2859 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2860 * array.
2861 *
2862 * At this point all successfully migrated pages have an entry in the src
2863 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2864 * array entry with MIGRATE_PFN_VALID flag set.
2865 *
2866 * It then calls the finalize_and_map() callback. See comments for "struct
2867 * migrate_vma_ops", in include/linux/migrate.h for details about
2868 * finalize_and_map() behavior.
2869 *
2870 * After the finalize_and_map() callback, for successfully migrated pages, this
2871 * function updates the CPU page table to point to new pages, otherwise it
2872 * restores the CPU page table to point to the original source pages.
2873 *
2874 * Function returns 0 after the above steps, even if no pages were migrated
2875 * (The function only returns an error if any of the arguments are invalid.)
2876 *
2877 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2878 * unsigned long entries.
2879 */
2880int migrate_vma(const struct migrate_vma_ops *ops,
2881 struct vm_area_struct *vma,
2882 unsigned long start,
2883 unsigned long end,
2884 unsigned long *src,
2885 unsigned long *dst,
2886 void *private)
2887{
2888 struct migrate_vma migrate;
2889
2890 /* Sanity check the arguments */
2891 start &= PAGE_MASK;
2892 end &= PAGE_MASK;
e1fb4a08
DJ
2893 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2894 vma_is_dax(vma))
8763cb45
JG
2895 return -EINVAL;
2896 if (start < vma->vm_start || start >= vma->vm_end)
2897 return -EINVAL;
2898 if (end <= vma->vm_start || end > vma->vm_end)
2899 return -EINVAL;
2900 if (!ops || !src || !dst || start >= end)
2901 return -EINVAL;
2902
2903 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2904 migrate.src = src;
2905 migrate.dst = dst;
2906 migrate.start = start;
2907 migrate.npages = 0;
2908 migrate.cpages = 0;
2909 migrate.end = end;
2910 migrate.vma = vma;
2911
2912 /* Collect, and try to unmap source pages */
2913 migrate_vma_collect(&migrate);
2914 if (!migrate.cpages)
2915 return 0;
2916
2917 /* Lock and isolate page */
2918 migrate_vma_prepare(&migrate);
2919 if (!migrate.cpages)
2920 return 0;
2921
2922 /* Unmap pages */
2923 migrate_vma_unmap(&migrate);
2924 if (!migrate.cpages)
2925 return 0;
2926
2927 /*
2928 * At this point pages are locked and unmapped, and thus they have
2929 * stable content and can safely be copied to destination memory that
2930 * is allocated by the callback.
2931 *
2932 * Note that migration can fail in migrate_vma_struct_page() for each
2933 * individual page.
2934 */
2935 ops->alloc_and_copy(vma, src, dst, start, end, private);
2936
2937 /* This does the real migration of struct page */
2938 migrate_vma_pages(&migrate);
2939
2940 ops->finalize_and_map(vma, src, dst, start, end, private);
2941
2942 /* Unlock and remap pages */
2943 migrate_vma_finalize(&migrate);
2944
2945 return 0;
2946}
2947EXPORT_SYMBOL(migrate_vma);
6b368cd4 2948#endif /* defined(MIGRATE_VMA_HELPER) */