]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
mm/mmu_notifiers: annotate with might_sleep()
[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
8bb4e7a2 103 * so unconditionally grabbing the lock ruins page's owner side.
bda807d4
MK
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);
df6ad698 249 }
d2b2c6dd 250 }
a5430dda 251
3ef8fd7f 252#ifdef CONFIG_HUGETLB_PAGE
3fe87967
KS
253 if (PageHuge(new)) {
254 pte = pte_mkhuge(pte);
255 pte = arch_make_huge_pte(pte, vma, new, 0);
383321ab 256 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
3fe87967
KS
257 if (PageAnon(new))
258 hugepage_add_anon_rmap(new, vma, pvmw.address);
259 else
260 page_dup_rmap(new, true);
383321ab
AK
261 } else
262#endif
263 {
264 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
04e62a29 265
383321ab
AK
266 if (PageAnon(new))
267 page_add_anon_rmap(new, vma, pvmw.address, false);
268 else
269 page_add_file_rmap(new, false);
270 }
3fe87967
KS
271 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
272 mlock_vma_page(new);
273
e125fe40
KS
274 if (PageTransHuge(page) && PageMlocked(page))
275 clear_page_mlock(page);
276
3fe87967
KS
277 /* No need to invalidate - it was non-present before */
278 update_mmu_cache(vma, pvmw.address, pvmw.pte);
279 }
51afb12b 280
e4b82222 281 return true;
0697212a
CL
282}
283
04e62a29
CL
284/*
285 * Get rid of all migration entries and replace them by
286 * references to the indicated page.
287 */
e388466d 288void remove_migration_ptes(struct page *old, struct page *new, bool locked)
04e62a29 289{
051ac83a
JK
290 struct rmap_walk_control rwc = {
291 .rmap_one = remove_migration_pte,
292 .arg = old,
293 };
294
e388466d
KS
295 if (locked)
296 rmap_walk_locked(new, &rwc);
297 else
298 rmap_walk(new, &rwc);
04e62a29
CL
299}
300
0697212a
CL
301/*
302 * Something used the pte of a page under migration. We need to
303 * get to the page and wait until migration is finished.
304 * When we return from this function the fault will be retried.
0697212a 305 */
e66f17ff 306void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
30dad309 307 spinlock_t *ptl)
0697212a 308{
30dad309 309 pte_t pte;
0697212a
CL
310 swp_entry_t entry;
311 struct page *page;
312
30dad309 313 spin_lock(ptl);
0697212a
CL
314 pte = *ptep;
315 if (!is_swap_pte(pte))
316 goto out;
317
318 entry = pte_to_swp_entry(pte);
319 if (!is_migration_entry(entry))
320 goto out;
321
322 page = migration_entry_to_page(entry);
323
e286781d 324 /*
89eb946a 325 * Once page cache replacement of page migration started, page_count
9a1ea439
HD
326 * is zero; but we must not call put_and_wait_on_page_locked() without
327 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
e286781d
NP
328 */
329 if (!get_page_unless_zero(page))
330 goto out;
0697212a 331 pte_unmap_unlock(ptep, ptl);
9a1ea439 332 put_and_wait_on_page_locked(page);
0697212a
CL
333 return;
334out:
335 pte_unmap_unlock(ptep, ptl);
336}
337
30dad309
NH
338void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
339 unsigned long address)
340{
341 spinlock_t *ptl = pte_lockptr(mm, pmd);
342 pte_t *ptep = pte_offset_map(pmd, address);
343 __migration_entry_wait(mm, ptep, ptl);
344}
345
cb900f41
KS
346void migration_entry_wait_huge(struct vm_area_struct *vma,
347 struct mm_struct *mm, pte_t *pte)
30dad309 348{
cb900f41 349 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
30dad309
NH
350 __migration_entry_wait(mm, pte, ptl);
351}
352
616b8371
ZY
353#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
354void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
355{
356 spinlock_t *ptl;
357 struct page *page;
358
359 ptl = pmd_lock(mm, pmd);
360 if (!is_pmd_migration_entry(*pmd))
361 goto unlock;
362 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
363 if (!get_page_unless_zero(page))
364 goto unlock;
365 spin_unlock(ptl);
9a1ea439 366 put_and_wait_on_page_locked(page);
616b8371
ZY
367 return;
368unlock:
369 spin_unlock(ptl);
370}
371#endif
372
f900482d 373static int expected_page_refs(struct address_space *mapping, struct page *page)
0b3901b3
JK
374{
375 int expected_count = 1;
376
377 /*
378 * Device public or private pages have an extra refcount as they are
379 * ZONE_DEVICE pages.
380 */
381 expected_count += is_device_private_page(page);
f900482d 382 if (mapping)
0b3901b3
JK
383 expected_count += hpage_nr_pages(page) + page_has_private(page);
384
385 return expected_count;
386}
387
b20a3503 388/*
c3fcf8a5 389 * Replace the page in the mapping.
5b5c7120
CL
390 *
391 * The number of remaining references must be:
392 * 1 for anonymous pages without a mapping
393 * 2 for pages with a mapping
266cf658 394 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
b20a3503 395 */
36bc08cc 396int migrate_page_move_mapping(struct address_space *mapping,
37109694 397 struct page *newpage, struct page *page, int extra_count)
b20a3503 398{
89eb946a 399 XA_STATE(xas, &mapping->i_pages, page_index(page));
42cb14b1
HD
400 struct zone *oldzone, *newzone;
401 int dirty;
f900482d 402 int expected_count = expected_page_refs(mapping, page) + extra_count;
8763cb45 403
6c5240ae 404 if (!mapping) {
0e8c7d0f 405 /* Anonymous page without mapping */
8e321fef 406 if (page_count(page) != expected_count)
6c5240ae 407 return -EAGAIN;
cf4b769a
HD
408
409 /* No turning back from here */
cf4b769a
HD
410 newpage->index = page->index;
411 newpage->mapping = page->mapping;
412 if (PageSwapBacked(page))
fa9949da 413 __SetPageSwapBacked(newpage);
cf4b769a 414
78bd5209 415 return MIGRATEPAGE_SUCCESS;
6c5240ae
CL
416 }
417
42cb14b1
HD
418 oldzone = page_zone(page);
419 newzone = page_zone(newpage);
420
89eb946a 421 xas_lock_irq(&xas);
89eb946a
MW
422 if (page_count(page) != expected_count || xas_load(&xas) != page) {
423 xas_unlock_irq(&xas);
e23ca00b 424 return -EAGAIN;
b20a3503
CL
425 }
426
fe896d18 427 if (!page_ref_freeze(page, expected_count)) {
89eb946a 428 xas_unlock_irq(&xas);
e286781d
NP
429 return -EAGAIN;
430 }
431
b20a3503 432 /*
cf4b769a
HD
433 * Now we know that no one else is looking at the page:
434 * no turning back from here.
b20a3503 435 */
cf4b769a
HD
436 newpage->index = page->index;
437 newpage->mapping = page->mapping;
e71769ae 438 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
6326fec1
NP
439 if (PageSwapBacked(page)) {
440 __SetPageSwapBacked(newpage);
441 if (PageSwapCache(page)) {
442 SetPageSwapCache(newpage);
443 set_page_private(newpage, page_private(page));
444 }
445 } else {
446 VM_BUG_ON_PAGE(PageSwapCache(page), page);
b20a3503
CL
447 }
448
42cb14b1
HD
449 /* Move dirty while page refs frozen and newpage not yet exposed */
450 dirty = PageDirty(page);
451 if (dirty) {
452 ClearPageDirty(page);
453 SetPageDirty(newpage);
454 }
455
89eb946a 456 xas_store(&xas, newpage);
e71769ae
NH
457 if (PageTransHuge(page)) {
458 int i;
e71769ae 459
013567be 460 for (i = 1; i < HPAGE_PMD_NR; i++) {
89eb946a 461 xas_next(&xas);
69bf4b6b 462 xas_store(&xas, newpage + i);
e71769ae 463 }
e71769ae 464 }
7cf9c2c7
NP
465
466 /*
937a94c9
JG
467 * Drop cache reference from old page by unfreezing
468 * to one less reference.
7cf9c2c7
NP
469 * We know this isn't the last reference.
470 */
e71769ae 471 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
7cf9c2c7 472
89eb946a 473 xas_unlock(&xas);
42cb14b1
HD
474 /* Leave irq disabled to prevent preemption while updating stats */
475
0e8c7d0f
CL
476 /*
477 * If moved to a different zone then also account
478 * the page for that zone. Other VM counters will be
479 * taken care of when we establish references to the
480 * new page and drop references to the old page.
481 *
482 * Note that anonymous pages are accounted for
4b9d0fab 483 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
0e8c7d0f
CL
484 * are mapped to swap space.
485 */
42cb14b1 486 if (newzone != oldzone) {
11fb9989
MG
487 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
488 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
42cb14b1 489 if (PageSwapBacked(page) && !PageSwapCache(page)) {
11fb9989
MG
490 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
491 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
42cb14b1
HD
492 }
493 if (dirty && mapping_cap_account_dirty(mapping)) {
11fb9989 494 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 495 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
11fb9989 496 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 497 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
42cb14b1 498 }
4b02108a 499 }
42cb14b1 500 local_irq_enable();
b20a3503 501
78bd5209 502 return MIGRATEPAGE_SUCCESS;
b20a3503 503}
1118dce7 504EXPORT_SYMBOL(migrate_page_move_mapping);
b20a3503 505
290408d4
NH
506/*
507 * The expected number of remaining references is the same as that
508 * of migrate_page_move_mapping().
509 */
510int migrate_huge_page_move_mapping(struct address_space *mapping,
511 struct page *newpage, struct page *page)
512{
89eb946a 513 XA_STATE(xas, &mapping->i_pages, page_index(page));
290408d4 514 int expected_count;
290408d4 515
89eb946a 516 xas_lock_irq(&xas);
290408d4 517 expected_count = 2 + page_has_private(page);
89eb946a
MW
518 if (page_count(page) != expected_count || xas_load(&xas) != page) {
519 xas_unlock_irq(&xas);
290408d4
NH
520 return -EAGAIN;
521 }
522
fe896d18 523 if (!page_ref_freeze(page, expected_count)) {
89eb946a 524 xas_unlock_irq(&xas);
290408d4
NH
525 return -EAGAIN;
526 }
527
cf4b769a
HD
528 newpage->index = page->index;
529 newpage->mapping = page->mapping;
6a93ca8f 530
290408d4
NH
531 get_page(newpage);
532
89eb946a 533 xas_store(&xas, newpage);
290408d4 534
fe896d18 535 page_ref_unfreeze(page, expected_count - 1);
290408d4 536
89eb946a 537 xas_unlock_irq(&xas);
6a93ca8f 538
78bd5209 539 return MIGRATEPAGE_SUCCESS;
290408d4
NH
540}
541
30b0a105
DH
542/*
543 * Gigantic pages are so large that we do not guarantee that page++ pointer
544 * arithmetic will work across the entire page. We need something more
545 * specialized.
546 */
547static void __copy_gigantic_page(struct page *dst, struct page *src,
548 int nr_pages)
549{
550 int i;
551 struct page *dst_base = dst;
552 struct page *src_base = src;
553
554 for (i = 0; i < nr_pages; ) {
555 cond_resched();
556 copy_highpage(dst, src);
557
558 i++;
559 dst = mem_map_next(dst, dst_base, i);
560 src = mem_map_next(src, src_base, i);
561 }
562}
563
564static void copy_huge_page(struct page *dst, struct page *src)
565{
566 int i;
567 int nr_pages;
568
569 if (PageHuge(src)) {
570 /* hugetlbfs page */
571 struct hstate *h = page_hstate(src);
572 nr_pages = pages_per_huge_page(h);
573
574 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
575 __copy_gigantic_page(dst, src, nr_pages);
576 return;
577 }
578 } else {
579 /* thp page */
580 BUG_ON(!PageTransHuge(src));
581 nr_pages = hpage_nr_pages(src);
582 }
583
584 for (i = 0; i < nr_pages; i++) {
585 cond_resched();
586 copy_highpage(dst + i, src + i);
587 }
588}
589
b20a3503
CL
590/*
591 * Copy the page to its new location
592 */
2916ecc0 593void migrate_page_states(struct page *newpage, struct page *page)
b20a3503 594{
7851a45c
RR
595 int cpupid;
596
b20a3503
CL
597 if (PageError(page))
598 SetPageError(newpage);
599 if (PageReferenced(page))
600 SetPageReferenced(newpage);
601 if (PageUptodate(page))
602 SetPageUptodate(newpage);
894bc310 603 if (TestClearPageActive(page)) {
309381fe 604 VM_BUG_ON_PAGE(PageUnevictable(page), page);
b20a3503 605 SetPageActive(newpage);
418b27ef
LS
606 } else if (TestClearPageUnevictable(page))
607 SetPageUnevictable(newpage);
1899ad18
JW
608 if (PageWorkingset(page))
609 SetPageWorkingset(newpage);
b20a3503
CL
610 if (PageChecked(page))
611 SetPageChecked(newpage);
612 if (PageMappedToDisk(page))
613 SetPageMappedToDisk(newpage);
614
42cb14b1
HD
615 /* Move dirty on pages not done by migrate_page_move_mapping() */
616 if (PageDirty(page))
617 SetPageDirty(newpage);
b20a3503 618
33c3fc71
VD
619 if (page_is_young(page))
620 set_page_young(newpage);
621 if (page_is_idle(page))
622 set_page_idle(newpage);
623
7851a45c
RR
624 /*
625 * Copy NUMA information to the new page, to prevent over-eager
626 * future migrations of this same page.
627 */
628 cpupid = page_cpupid_xchg_last(page, -1);
629 page_cpupid_xchg_last(newpage, cpupid);
630
e9995ef9 631 ksm_migrate_page(newpage, page);
c8d6553b
HD
632 /*
633 * Please do not reorder this without considering how mm/ksm.c's
634 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
635 */
b3b3a99c
NH
636 if (PageSwapCache(page))
637 ClearPageSwapCache(page);
b20a3503
CL
638 ClearPagePrivate(page);
639 set_page_private(page, 0);
b20a3503
CL
640
641 /*
642 * If any waiters have accumulated on the new page then
643 * wake them up.
644 */
645 if (PageWriteback(newpage))
646 end_page_writeback(newpage);
d435edca
VB
647
648 copy_page_owner(page, newpage);
74485cf2
JW
649
650 mem_cgroup_migrate(page, newpage);
b20a3503 651}
2916ecc0
JG
652EXPORT_SYMBOL(migrate_page_states);
653
654void migrate_page_copy(struct page *newpage, struct page *page)
655{
656 if (PageHuge(page) || PageTransHuge(page))
657 copy_huge_page(newpage, page);
658 else
659 copy_highpage(newpage, page);
660
661 migrate_page_states(newpage, page);
662}
1118dce7 663EXPORT_SYMBOL(migrate_page_copy);
b20a3503 664
1d8b85cc
CL
665/************************************************************
666 * Migration functions
667 ***********************************************************/
668
b20a3503 669/*
bda807d4 670 * Common logic to directly migrate a single LRU page suitable for
266cf658 671 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
672 *
673 * Pages are locked upon entry and exit.
674 */
2d1db3b1 675int migrate_page(struct address_space *mapping,
a6bc32b8
MG
676 struct page *newpage, struct page *page,
677 enum migrate_mode mode)
b20a3503
CL
678{
679 int rc;
680
681 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
682
37109694 683 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
b20a3503 684
78bd5209 685 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
686 return rc;
687
2916ecc0
JG
688 if (mode != MIGRATE_SYNC_NO_COPY)
689 migrate_page_copy(newpage, page);
690 else
691 migrate_page_states(newpage, page);
78bd5209 692 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
693}
694EXPORT_SYMBOL(migrate_page);
695
9361401e 696#ifdef CONFIG_BLOCK
84ade7c1
JK
697/* Returns true if all buffers are successfully locked */
698static bool buffer_migrate_lock_buffers(struct buffer_head *head,
699 enum migrate_mode mode)
700{
701 struct buffer_head *bh = head;
702
703 /* Simple case, sync compaction */
704 if (mode != MIGRATE_ASYNC) {
705 do {
84ade7c1
JK
706 lock_buffer(bh);
707 bh = bh->b_this_page;
708
709 } while (bh != head);
710
711 return true;
712 }
713
714 /* async case, we cannot block on lock_buffer so use trylock_buffer */
715 do {
84ade7c1
JK
716 if (!trylock_buffer(bh)) {
717 /*
718 * We failed to lock the buffer and cannot stall in
719 * async migration. Release the taken locks
720 */
721 struct buffer_head *failed_bh = bh;
84ade7c1
JK
722 bh = head;
723 while (bh != failed_bh) {
724 unlock_buffer(bh);
84ade7c1
JK
725 bh = bh->b_this_page;
726 }
727 return false;
728 }
729
730 bh = bh->b_this_page;
731 } while (bh != head);
732 return true;
733}
734
89cb0888
JK
735static int __buffer_migrate_page(struct address_space *mapping,
736 struct page *newpage, struct page *page, enum migrate_mode mode,
737 bool check_refs)
1d8b85cc 738{
1d8b85cc
CL
739 struct buffer_head *bh, *head;
740 int rc;
cc4f11e6 741 int expected_count;
1d8b85cc 742
1d8b85cc 743 if (!page_has_buffers(page))
a6bc32b8 744 return migrate_page(mapping, newpage, page, mode);
1d8b85cc 745
cc4f11e6 746 /* Check whether page does not have extra refs before we do more work */
f900482d 747 expected_count = expected_page_refs(mapping, page);
cc4f11e6
JK
748 if (page_count(page) != expected_count)
749 return -EAGAIN;
1d8b85cc 750
cc4f11e6
JK
751 head = page_buffers(page);
752 if (!buffer_migrate_lock_buffers(head, mode))
753 return -EAGAIN;
1d8b85cc 754
89cb0888
JK
755 if (check_refs) {
756 bool busy;
757 bool invalidated = false;
758
759recheck_buffers:
760 busy = false;
761 spin_lock(&mapping->private_lock);
762 bh = head;
763 do {
764 if (atomic_read(&bh->b_count)) {
765 busy = true;
766 break;
767 }
768 bh = bh->b_this_page;
769 } while (bh != head);
89cb0888
JK
770 if (busy) {
771 if (invalidated) {
772 rc = -EAGAIN;
773 goto unlock_buffers;
774 }
ebdf4de5 775 spin_unlock(&mapping->private_lock);
89cb0888
JK
776 invalidate_bh_lrus();
777 invalidated = true;
778 goto recheck_buffers;
779 }
780 }
781
37109694 782 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
78bd5209 783 if (rc != MIGRATEPAGE_SUCCESS)
cc4f11e6 784 goto unlock_buffers;
1d8b85cc
CL
785
786 ClearPagePrivate(page);
787 set_page_private(newpage, page_private(page));
788 set_page_private(page, 0);
789 put_page(page);
790 get_page(newpage);
791
792 bh = head;
793 do {
794 set_bh_page(bh, newpage, bh_offset(bh));
795 bh = bh->b_this_page;
796
797 } while (bh != head);
798
799 SetPagePrivate(newpage);
800
2916ecc0
JG
801 if (mode != MIGRATE_SYNC_NO_COPY)
802 migrate_page_copy(newpage, page);
803 else
804 migrate_page_states(newpage, page);
1d8b85cc 805
cc4f11e6
JK
806 rc = MIGRATEPAGE_SUCCESS;
807unlock_buffers:
ebdf4de5
JK
808 if (check_refs)
809 spin_unlock(&mapping->private_lock);
1d8b85cc
CL
810 bh = head;
811 do {
812 unlock_buffer(bh);
1d8b85cc
CL
813 bh = bh->b_this_page;
814
815 } while (bh != head);
816
cc4f11e6 817 return rc;
1d8b85cc 818}
89cb0888
JK
819
820/*
821 * Migration function for pages with buffers. This function can only be used
822 * if the underlying filesystem guarantees that no other references to "page"
823 * exist. For example attached buffer heads are accessed only under page lock.
824 */
825int buffer_migrate_page(struct address_space *mapping,
826 struct page *newpage, struct page *page, enum migrate_mode mode)
827{
828 return __buffer_migrate_page(mapping, newpage, page, mode, false);
829}
1d8b85cc 830EXPORT_SYMBOL(buffer_migrate_page);
89cb0888
JK
831
832/*
833 * Same as above except that this variant is more careful and checks that there
834 * are also no buffer head references. This function is the right one for
835 * mappings where buffer heads are directly looked up and referenced (such as
836 * block device mappings).
837 */
838int buffer_migrate_page_norefs(struct address_space *mapping,
839 struct page *newpage, struct page *page, enum migrate_mode mode)
840{
841 return __buffer_migrate_page(mapping, newpage, page, mode, true);
842}
9361401e 843#endif
1d8b85cc 844
04e62a29
CL
845/*
846 * Writeback a page to clean the dirty state
847 */
848static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 849{
04e62a29
CL
850 struct writeback_control wbc = {
851 .sync_mode = WB_SYNC_NONE,
852 .nr_to_write = 1,
853 .range_start = 0,
854 .range_end = LLONG_MAX,
04e62a29
CL
855 .for_reclaim = 1
856 };
857 int rc;
858
859 if (!mapping->a_ops->writepage)
860 /* No write method for the address space */
861 return -EINVAL;
862
863 if (!clear_page_dirty_for_io(page))
864 /* Someone else already triggered a write */
865 return -EAGAIN;
866
8351a6e4 867 /*
04e62a29
CL
868 * A dirty page may imply that the underlying filesystem has
869 * the page on some queue. So the page must be clean for
870 * migration. Writeout may mean we loose the lock and the
871 * page state is no longer what we checked for earlier.
872 * At this point we know that the migration attempt cannot
873 * be successful.
8351a6e4 874 */
e388466d 875 remove_migration_ptes(page, page, false);
8351a6e4 876
04e62a29 877 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 878
04e62a29
CL
879 if (rc != AOP_WRITEPAGE_ACTIVATE)
880 /* unlocked. Relock */
881 lock_page(page);
882
bda8550d 883 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
884}
885
886/*
887 * Default handling if a filesystem does not provide a migration function.
888 */
889static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 890 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 891{
b969c4ab 892 if (PageDirty(page)) {
a6bc32b8 893 /* Only writeback pages in full synchronous migration */
2916ecc0
JG
894 switch (mode) {
895 case MIGRATE_SYNC:
896 case MIGRATE_SYNC_NO_COPY:
897 break;
898 default:
b969c4ab 899 return -EBUSY;
2916ecc0 900 }
04e62a29 901 return writeout(mapping, page);
b969c4ab 902 }
8351a6e4
CL
903
904 /*
905 * Buffers may be managed in a filesystem specific way.
906 * We must have no buffers or drop them.
907 */
266cf658 908 if (page_has_private(page) &&
8351a6e4 909 !try_to_release_page(page, GFP_KERNEL))
806031bb 910 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
8351a6e4 911
a6bc32b8 912 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
913}
914
e24f0b8f
CL
915/*
916 * Move a page to a newly allocated page
917 * The page is locked and all ptes have been successfully removed.
918 *
919 * The new page will have replaced the old page if this function
920 * is successful.
894bc310
LS
921 *
922 * Return value:
923 * < 0 - error code
78bd5209 924 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 925 */
3fe2011f 926static int move_to_new_page(struct page *newpage, struct page *page,
5c3f9a67 927 enum migrate_mode mode)
e24f0b8f
CL
928{
929 struct address_space *mapping;
bda807d4
MK
930 int rc = -EAGAIN;
931 bool is_lru = !__PageMovable(page);
e24f0b8f 932
7db7671f
HD
933 VM_BUG_ON_PAGE(!PageLocked(page), page);
934 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
e24f0b8f 935
e24f0b8f 936 mapping = page_mapping(page);
bda807d4
MK
937
938 if (likely(is_lru)) {
939 if (!mapping)
940 rc = migrate_page(mapping, newpage, page, mode);
941 else if (mapping->a_ops->migratepage)
942 /*
943 * Most pages have a mapping and most filesystems
944 * provide a migratepage callback. Anonymous pages
945 * are part of swap space which also has its own
946 * migratepage callback. This is the most common path
947 * for page migration.
948 */
949 rc = mapping->a_ops->migratepage(mapping, newpage,
950 page, mode);
951 else
952 rc = fallback_migrate_page(mapping, newpage,
953 page, mode);
954 } else {
e24f0b8f 955 /*
bda807d4
MK
956 * In case of non-lru page, it could be released after
957 * isolation step. In that case, we shouldn't try migration.
e24f0b8f 958 */
bda807d4
MK
959 VM_BUG_ON_PAGE(!PageIsolated(page), page);
960 if (!PageMovable(page)) {
961 rc = MIGRATEPAGE_SUCCESS;
962 __ClearPageIsolated(page);
963 goto out;
964 }
965
966 rc = mapping->a_ops->migratepage(mapping, newpage,
967 page, mode);
968 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
969 !PageIsolated(page));
970 }
e24f0b8f 971
5c3f9a67
HD
972 /*
973 * When successful, old pagecache page->mapping must be cleared before
974 * page is freed; but stats require that PageAnon be left as PageAnon.
975 */
976 if (rc == MIGRATEPAGE_SUCCESS) {
bda807d4
MK
977 if (__PageMovable(page)) {
978 VM_BUG_ON_PAGE(!PageIsolated(page), page);
979
980 /*
981 * We clear PG_movable under page_lock so any compactor
982 * cannot try to migrate this page.
983 */
984 __ClearPageIsolated(page);
985 }
986
987 /*
988 * Anonymous and movable page->mapping will be cleard by
989 * free_pages_prepare so don't reset it here for keeping
990 * the type to work PageAnon, for example.
991 */
992 if (!PageMappingFlags(page))
5c3f9a67 993 page->mapping = NULL;
d2b2c6dd 994
25b2995a 995 if (likely(!is_zone_device_page(newpage)))
d2b2c6dd
LP
996 flush_dcache_page(newpage);
997
3fe2011f 998 }
bda807d4 999out:
e24f0b8f
CL
1000 return rc;
1001}
1002
0dabec93 1003static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 1004 int force, enum migrate_mode mode)
e24f0b8f 1005{
0dabec93 1006 int rc = -EAGAIN;
2ebba6b7 1007 int page_was_mapped = 0;
3f6c8272 1008 struct anon_vma *anon_vma = NULL;
bda807d4 1009 bool is_lru = !__PageMovable(page);
95a402c3 1010
529ae9aa 1011 if (!trylock_page(page)) {
a6bc32b8 1012 if (!force || mode == MIGRATE_ASYNC)
0dabec93 1013 goto out;
3e7d3449
MG
1014
1015 /*
1016 * It's not safe for direct compaction to call lock_page.
1017 * For example, during page readahead pages are added locked
1018 * to the LRU. Later, when the IO completes the pages are
1019 * marked uptodate and unlocked. However, the queueing
1020 * could be merging multiple pages for one bio (e.g.
1021 * mpage_readpages). If an allocation happens for the
1022 * second or third page, the process can end up locking
1023 * the same page twice and deadlocking. Rather than
1024 * trying to be clever about what pages can be locked,
1025 * avoid the use of lock_page for direct compaction
1026 * altogether.
1027 */
1028 if (current->flags & PF_MEMALLOC)
0dabec93 1029 goto out;
3e7d3449 1030
e24f0b8f
CL
1031 lock_page(page);
1032 }
1033
1034 if (PageWriteback(page)) {
11bc82d6 1035 /*
fed5b64a 1036 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
1037 * necessary to wait for PageWriteback. In the async case,
1038 * the retry loop is too short and in the sync-light case,
1039 * the overhead of stalling is too much
11bc82d6 1040 */
2916ecc0
JG
1041 switch (mode) {
1042 case MIGRATE_SYNC:
1043 case MIGRATE_SYNC_NO_COPY:
1044 break;
1045 default:
11bc82d6 1046 rc = -EBUSY;
0a31bc97 1047 goto out_unlock;
11bc82d6
AA
1048 }
1049 if (!force)
0a31bc97 1050 goto out_unlock;
e24f0b8f
CL
1051 wait_on_page_writeback(page);
1052 }
03f15c86 1053
e24f0b8f 1054 /*
dc386d4d
KH
1055 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1056 * we cannot notice that anon_vma is freed while we migrates a page.
1ce82b69 1057 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 1058 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
1059 * File Caches may use write_page() or lock_page() in migration, then,
1060 * just care Anon page here.
03f15c86
HD
1061 *
1062 * Only page_get_anon_vma() understands the subtleties of
1063 * getting a hold on an anon_vma from outside one of its mms.
1064 * But if we cannot get anon_vma, then we won't need it anyway,
1065 * because that implies that the anon page is no longer mapped
1066 * (and cannot be remapped so long as we hold the page lock).
dc386d4d 1067 */
03f15c86 1068 if (PageAnon(page) && !PageKsm(page))
746b18d4 1069 anon_vma = page_get_anon_vma(page);
62e1c553 1070
7db7671f
HD
1071 /*
1072 * Block others from accessing the new page when we get around to
1073 * establishing additional references. We are usually the only one
1074 * holding a reference to newpage at this point. We used to have a BUG
1075 * here if trylock_page(newpage) fails, but would like to allow for
1076 * cases where there might be a race with the previous use of newpage.
1077 * This is much like races on refcount of oldpage: just don't BUG().
1078 */
1079 if (unlikely(!trylock_page(newpage)))
1080 goto out_unlock;
1081
bda807d4
MK
1082 if (unlikely(!is_lru)) {
1083 rc = move_to_new_page(newpage, page, mode);
1084 goto out_unlock_both;
1085 }
1086
dc386d4d 1087 /*
62e1c553
SL
1088 * Corner case handling:
1089 * 1. When a new swap-cache page is read into, it is added to the LRU
1090 * and treated as swapcache but it has no rmap yet.
1091 * Calling try_to_unmap() against a page->mapping==NULL page will
1092 * trigger a BUG. So handle it here.
1093 * 2. An orphaned page (see truncate_complete_page) might have
1094 * fs-private metadata. The page can be picked up due to memory
1095 * offlining. Everywhere else except page reclaim, the page is
1096 * invisible to the vm, so the page can not be migrated. So try to
1097 * free the metadata, so the page can be freed.
e24f0b8f 1098 */
62e1c553 1099 if (!page->mapping) {
309381fe 1100 VM_BUG_ON_PAGE(PageAnon(page), page);
1ce82b69 1101 if (page_has_private(page)) {
62e1c553 1102 try_to_free_buffers(page);
7db7671f 1103 goto out_unlock_both;
62e1c553 1104 }
7db7671f
HD
1105 } else if (page_mapped(page)) {
1106 /* Establish migration ptes */
03f15c86
HD
1107 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1108 page);
2ebba6b7 1109 try_to_unmap(page,
da1b13cc 1110 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
2ebba6b7
HD
1111 page_was_mapped = 1;
1112 }
dc386d4d 1113
e6a1530d 1114 if (!page_mapped(page))
5c3f9a67 1115 rc = move_to_new_page(newpage, page, mode);
e24f0b8f 1116
5c3f9a67
HD
1117 if (page_was_mapped)
1118 remove_migration_ptes(page,
e388466d 1119 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
3f6c8272 1120
7db7671f
HD
1121out_unlock_both:
1122 unlock_page(newpage);
1123out_unlock:
3f6c8272 1124 /* Drop an anon_vma reference if we took one */
76545066 1125 if (anon_vma)
9e60109f 1126 put_anon_vma(anon_vma);
e24f0b8f 1127 unlock_page(page);
0dabec93 1128out:
c6c919eb
MK
1129 /*
1130 * If migration is successful, decrease refcount of the newpage
1131 * which will not free the page because new page owner increased
1132 * refcounter. As well, if it is LRU page, add the page to LRU
e0a352fa
DH
1133 * list in here. Use the old state of the isolated source page to
1134 * determine if we migrated a LRU page. newpage was already unlocked
1135 * and possibly modified by its owner - don't rely on the page
1136 * state.
c6c919eb
MK
1137 */
1138 if (rc == MIGRATEPAGE_SUCCESS) {
e0a352fa 1139 if (unlikely(!is_lru))
c6c919eb
MK
1140 put_page(newpage);
1141 else
1142 putback_lru_page(newpage);
1143 }
1144
0dabec93
MK
1145 return rc;
1146}
95a402c3 1147
ef2a5153
GU
1148/*
1149 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1150 * around it.
1151 */
815f0ddb
ND
1152#if defined(CONFIG_ARM) && \
1153 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
ef2a5153
GU
1154#define ICE_noinline noinline
1155#else
1156#define ICE_noinline
1157#endif
1158
0dabec93
MK
1159/*
1160 * Obtain the lock on page, remove all ptes and migrate the page
1161 * to the newly allocated page in newpage.
1162 */
ef2a5153
GU
1163static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1164 free_page_t put_new_page,
1165 unsigned long private, struct page *page,
add05cec
NH
1166 int force, enum migrate_mode mode,
1167 enum migrate_reason reason)
0dabec93 1168{
2def7424 1169 int rc = MIGRATEPAGE_SUCCESS;
2def7424 1170 struct page *newpage;
0dabec93 1171
94723aaf
MH
1172 if (!thp_migration_supported() && PageTransHuge(page))
1173 return -ENOMEM;
1174
666feb21 1175 newpage = get_new_page(page, private);
0dabec93
MK
1176 if (!newpage)
1177 return -ENOMEM;
1178
1179 if (page_count(page) == 1) {
1180 /* page was freed from under us. So we are done. */
c6c919eb
MK
1181 ClearPageActive(page);
1182 ClearPageUnevictable(page);
bda807d4
MK
1183 if (unlikely(__PageMovable(page))) {
1184 lock_page(page);
1185 if (!PageMovable(page))
1186 __ClearPageIsolated(page);
1187 unlock_page(page);
1188 }
c6c919eb
MK
1189 if (put_new_page)
1190 put_new_page(newpage, private);
1191 else
1192 put_page(newpage);
0dabec93
MK
1193 goto out;
1194 }
1195
9c620e2b 1196 rc = __unmap_and_move(page, newpage, force, mode);
c6c919eb 1197 if (rc == MIGRATEPAGE_SUCCESS)
7cd12b4a 1198 set_page_owner_migrate_reason(newpage, reason);
bf6bddf1 1199
0dabec93 1200out:
e24f0b8f 1201 if (rc != -EAGAIN) {
0dabec93
MK
1202 /*
1203 * A page that has been migrated has all references
1204 * removed and will be freed. A page that has not been
1205 * migrated will have kepts its references and be
1206 * restored.
1207 */
1208 list_del(&page->lru);
6afcf8ef
ML
1209
1210 /*
1211 * Compaction can migrate also non-LRU pages which are
1212 * not accounted to NR_ISOLATED_*. They can be recognized
1213 * as __PageMovable
1214 */
1215 if (likely(!__PageMovable(page)))
e8db67eb
NH
1216 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1217 page_is_file_cache(page), -hpage_nr_pages(page));
c6c919eb
MK
1218 }
1219
1220 /*
1221 * If migration is successful, releases reference grabbed during
1222 * isolation. Otherwise, restore the page to right list unless
1223 * we want to retry.
1224 */
1225 if (rc == MIGRATEPAGE_SUCCESS) {
1226 put_page(page);
1227 if (reason == MR_MEMORY_FAILURE) {
d7e69488 1228 /*
c6c919eb
MK
1229 * Set PG_HWPoison on just freed page
1230 * intentionally. Although it's rather weird,
1231 * it's how HWPoison flag works at the moment.
d7e69488 1232 */
d4ae9916 1233 if (set_hwpoison_free_buddy_page(page))
da1b13cc 1234 num_poisoned_pages_inc();
c6c919eb
MK
1235 }
1236 } else {
bda807d4
MK
1237 if (rc != -EAGAIN) {
1238 if (likely(!__PageMovable(page))) {
1239 putback_lru_page(page);
1240 goto put_new;
1241 }
1242
1243 lock_page(page);
1244 if (PageMovable(page))
1245 putback_movable_page(page);
1246 else
1247 __ClearPageIsolated(page);
1248 unlock_page(page);
1249 put_page(page);
1250 }
1251put_new:
c6c919eb
MK
1252 if (put_new_page)
1253 put_new_page(newpage, private);
1254 else
1255 put_page(newpage);
e24f0b8f 1256 }
68711a74 1257
e24f0b8f
CL
1258 return rc;
1259}
1260
290408d4
NH
1261/*
1262 * Counterpart of unmap_and_move_page() for hugepage migration.
1263 *
1264 * This function doesn't wait the completion of hugepage I/O
1265 * because there is no race between I/O and migration for hugepage.
1266 * Note that currently hugepage I/O occurs only in direct I/O
1267 * where no lock is held and PG_writeback is irrelevant,
1268 * and writeback status of all subpages are counted in the reference
1269 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1270 * under direct I/O, the reference of the head page is 512 and a bit more.)
1271 * This means that when we try to migrate hugepage whose subpages are
1272 * doing direct I/O, some references remain after try_to_unmap() and
1273 * hugepage migration fails without data corruption.
1274 *
1275 * There is also no race when direct I/O is issued on the page under migration,
1276 * because then pte is replaced with migration swap entry and direct I/O code
1277 * will wait in the page fault for migration to complete.
1278 */
1279static int unmap_and_move_huge_page(new_page_t get_new_page,
68711a74
DR
1280 free_page_t put_new_page, unsigned long private,
1281 struct page *hpage, int force,
7cd12b4a 1282 enum migrate_mode mode, int reason)
290408d4 1283{
2def7424 1284 int rc = -EAGAIN;
2ebba6b7 1285 int page_was_mapped = 0;
32665f2b 1286 struct page *new_hpage;
290408d4
NH
1287 struct anon_vma *anon_vma = NULL;
1288
83467efb 1289 /*
7ed2c31d 1290 * Migratability of hugepages depends on architectures and their size.
83467efb
NH
1291 * This check is necessary because some callers of hugepage migration
1292 * like soft offline and memory hotremove don't walk through page
1293 * tables or check whether the hugepage is pmd-based or not before
1294 * kicking migration.
1295 */
100873d7 1296 if (!hugepage_migration_supported(page_hstate(hpage))) {
32665f2b 1297 putback_active_hugepage(hpage);
83467efb 1298 return -ENOSYS;
32665f2b 1299 }
83467efb 1300
666feb21 1301 new_hpage = get_new_page(hpage, private);
290408d4
NH
1302 if (!new_hpage)
1303 return -ENOMEM;
1304
290408d4 1305 if (!trylock_page(hpage)) {
2916ecc0 1306 if (!force)
290408d4 1307 goto out;
2916ecc0
JG
1308 switch (mode) {
1309 case MIGRATE_SYNC:
1310 case MIGRATE_SYNC_NO_COPY:
1311 break;
1312 default:
1313 goto out;
1314 }
290408d4
NH
1315 lock_page(hpage);
1316 }
1317
cb6acd01
MK
1318 /*
1319 * Check for pages which are in the process of being freed. Without
1320 * page_mapping() set, hugetlbfs specific move page routine will not
1321 * be called and we could leak usage counts for subpools.
1322 */
1323 if (page_private(hpage) && !page_mapping(hpage)) {
1324 rc = -EBUSY;
1325 goto out_unlock;
1326 }
1327
746b18d4
PZ
1328 if (PageAnon(hpage))
1329 anon_vma = page_get_anon_vma(hpage);
290408d4 1330
7db7671f
HD
1331 if (unlikely(!trylock_page(new_hpage)))
1332 goto put_anon;
1333
2ebba6b7
HD
1334 if (page_mapped(hpage)) {
1335 try_to_unmap(hpage,
ddeaab32 1336 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
2ebba6b7
HD
1337 page_was_mapped = 1;
1338 }
290408d4
NH
1339
1340 if (!page_mapped(hpage))
5c3f9a67 1341 rc = move_to_new_page(new_hpage, hpage, mode);
290408d4 1342
5c3f9a67
HD
1343 if (page_was_mapped)
1344 remove_migration_ptes(hpage,
e388466d 1345 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
290408d4 1346
7db7671f
HD
1347 unlock_page(new_hpage);
1348
1349put_anon:
fd4a4663 1350 if (anon_vma)
9e60109f 1351 put_anon_vma(anon_vma);
8e6ac7fa 1352
2def7424 1353 if (rc == MIGRATEPAGE_SUCCESS) {
ab5ac90a 1354 move_hugetlb_state(hpage, new_hpage, reason);
2def7424
HD
1355 put_new_page = NULL;
1356 }
8e6ac7fa 1357
cb6acd01 1358out_unlock:
290408d4 1359 unlock_page(hpage);
09761333 1360out:
b8ec1cee
NH
1361 if (rc != -EAGAIN)
1362 putback_active_hugepage(hpage);
68711a74
DR
1363
1364 /*
1365 * If migration was not successful and there's a freeing callback, use
1366 * it. Otherwise, put_page() will drop the reference grabbed during
1367 * isolation.
1368 */
2def7424 1369 if (put_new_page)
68711a74
DR
1370 put_new_page(new_hpage, private);
1371 else
3aaa76e1 1372 putback_active_hugepage(new_hpage);
68711a74 1373
290408d4
NH
1374 return rc;
1375}
1376
b20a3503 1377/*
c73e5c9c
SB
1378 * migrate_pages - migrate the pages specified in a list, to the free pages
1379 * supplied as the target for the page migration
b20a3503 1380 *
c73e5c9c
SB
1381 * @from: The list of pages to be migrated.
1382 * @get_new_page: The function used to allocate free pages to be used
1383 * as the target of the page migration.
68711a74
DR
1384 * @put_new_page: The function used to free target pages if migration
1385 * fails, or NULL if no special handling is necessary.
c73e5c9c
SB
1386 * @private: Private data to be passed on to get_new_page()
1387 * @mode: The migration mode that specifies the constraints for
1388 * page migration, if any.
1389 * @reason: The reason for page migration.
b20a3503 1390 *
c73e5c9c
SB
1391 * The function returns after 10 attempts or if no pages are movable any more
1392 * because the list has become empty or no retryable pages exist any more.
14e0f9bc 1393 * The caller should call putback_movable_pages() to return pages to the LRU
28bd6578 1394 * or free list only if ret != 0.
b20a3503 1395 *
c73e5c9c 1396 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1397 */
9c620e2b 1398int migrate_pages(struct list_head *from, new_page_t get_new_page,
68711a74
DR
1399 free_page_t put_new_page, unsigned long private,
1400 enum migrate_mode mode, int reason)
b20a3503 1401{
e24f0b8f 1402 int retry = 1;
b20a3503 1403 int nr_failed = 0;
5647bc29 1404 int nr_succeeded = 0;
b20a3503
CL
1405 int pass = 0;
1406 struct page *page;
1407 struct page *page2;
1408 int swapwrite = current->flags & PF_SWAPWRITE;
1409 int rc;
1410
1411 if (!swapwrite)
1412 current->flags |= PF_SWAPWRITE;
1413
e24f0b8f
CL
1414 for(pass = 0; pass < 10 && retry; pass++) {
1415 retry = 0;
b20a3503 1416
e24f0b8f 1417 list_for_each_entry_safe(page, page2, from, lru) {
94723aaf 1418retry:
e24f0b8f 1419 cond_resched();
2d1db3b1 1420
31caf665
NH
1421 if (PageHuge(page))
1422 rc = unmap_and_move_huge_page(get_new_page,
68711a74 1423 put_new_page, private, page,
7cd12b4a 1424 pass > 2, mode, reason);
31caf665 1425 else
68711a74 1426 rc = unmap_and_move(get_new_page, put_new_page,
add05cec
NH
1427 private, page, pass > 2, mode,
1428 reason);
2d1db3b1 1429
e24f0b8f 1430 switch(rc) {
95a402c3 1431 case -ENOMEM:
94723aaf
MH
1432 /*
1433 * THP migration might be unsupported or the
1434 * allocation could've failed so we should
1435 * retry on the same page with the THP split
1436 * to base pages.
1437 *
1438 * Head page is retried immediately and tail
1439 * pages are added to the tail of the list so
1440 * we encounter them after the rest of the list
1441 * is processed.
1442 */
e6112fc3 1443 if (PageTransHuge(page) && !PageHuge(page)) {
94723aaf
MH
1444 lock_page(page);
1445 rc = split_huge_page_to_list(page, from);
1446 unlock_page(page);
1447 if (!rc) {
1448 list_safe_reset_next(page, page2, lru);
1449 goto retry;
1450 }
1451 }
dfef2ef4 1452 nr_failed++;
95a402c3 1453 goto out;
e24f0b8f 1454 case -EAGAIN:
2d1db3b1 1455 retry++;
e24f0b8f 1456 break;
78bd5209 1457 case MIGRATEPAGE_SUCCESS:
5647bc29 1458 nr_succeeded++;
e24f0b8f
CL
1459 break;
1460 default:
354a3363
NH
1461 /*
1462 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1463 * unlike -EAGAIN case, the failed page is
1464 * removed from migration page list and not
1465 * retried in the next outer loop.
1466 */
2d1db3b1 1467 nr_failed++;
e24f0b8f 1468 break;
2d1db3b1 1469 }
b20a3503
CL
1470 }
1471 }
f2f81fb2
VB
1472 nr_failed += retry;
1473 rc = nr_failed;
95a402c3 1474out:
5647bc29
MG
1475 if (nr_succeeded)
1476 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1477 if (nr_failed)
1478 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1479 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1480
b20a3503
CL
1481 if (!swapwrite)
1482 current->flags &= ~PF_SWAPWRITE;
1483
78bd5209 1484 return rc;
b20a3503 1485}
95a402c3 1486
742755a1 1487#ifdef CONFIG_NUMA
742755a1 1488
a49bd4d7 1489static int store_status(int __user *status, int start, int value, int nr)
742755a1 1490{
a49bd4d7
MH
1491 while (nr-- > 0) {
1492 if (put_user(value, status + start))
1493 return -EFAULT;
1494 start++;
1495 }
1496
1497 return 0;
1498}
1499
1500static int do_move_pages_to_node(struct mm_struct *mm,
1501 struct list_head *pagelist, int node)
1502{
1503 int err;
1504
1505 if (list_empty(pagelist))
1506 return 0;
1507
1508 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1509 MIGRATE_SYNC, MR_SYSCALL);
1510 if (err)
1511 putback_movable_pages(pagelist);
1512 return err;
742755a1
CL
1513}
1514
1515/*
a49bd4d7
MH
1516 * Resolves the given address to a struct page, isolates it from the LRU and
1517 * puts it to the given pagelist.
1518 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1519 * queued or the page doesn't need to be migrated because it is already on
1520 * the target node
742755a1 1521 */
a49bd4d7
MH
1522static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1523 int node, struct list_head *pagelist, bool migrate_all)
742755a1 1524{
a49bd4d7
MH
1525 struct vm_area_struct *vma;
1526 struct page *page;
1527 unsigned int follflags;
742755a1 1528 int err;
742755a1
CL
1529
1530 down_read(&mm->mmap_sem);
a49bd4d7
MH
1531 err = -EFAULT;
1532 vma = find_vma(mm, addr);
1533 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1534 goto out;
742755a1 1535
a49bd4d7
MH
1536 /* FOLL_DUMP to ignore special (like zero) pages */
1537 follflags = FOLL_GET | FOLL_DUMP;
a49bd4d7 1538 page = follow_page(vma, addr, follflags);
89f5b7da 1539
a49bd4d7
MH
1540 err = PTR_ERR(page);
1541 if (IS_ERR(page))
1542 goto out;
89f5b7da 1543
a49bd4d7
MH
1544 err = -ENOENT;
1545 if (!page)
1546 goto out;
742755a1 1547
a49bd4d7
MH
1548 err = 0;
1549 if (page_to_nid(page) == node)
1550 goto out_putpage;
742755a1 1551
a49bd4d7
MH
1552 err = -EACCES;
1553 if (page_mapcount(page) > 1 && !migrate_all)
1554 goto out_putpage;
742755a1 1555
a49bd4d7
MH
1556 if (PageHuge(page)) {
1557 if (PageHead(page)) {
1558 isolate_huge_page(page, pagelist);
1559 err = 0;
e632a938 1560 }
a49bd4d7
MH
1561 } else {
1562 struct page *head;
e632a938 1563
e8db67eb
NH
1564 head = compound_head(page);
1565 err = isolate_lru_page(head);
cf608ac1 1566 if (err)
a49bd4d7 1567 goto out_putpage;
742755a1 1568
a49bd4d7
MH
1569 err = 0;
1570 list_add_tail(&head->lru, pagelist);
1571 mod_node_page_state(page_pgdat(head),
1572 NR_ISOLATED_ANON + page_is_file_cache(head),
1573 hpage_nr_pages(head));
1574 }
1575out_putpage:
1576 /*
1577 * Either remove the duplicate refcount from
1578 * isolate_lru_page() or drop the page ref if it was
1579 * not isolated.
1580 */
1581 put_page(page);
1582out:
742755a1
CL
1583 up_read(&mm->mmap_sem);
1584 return err;
1585}
1586
5e9a0f02
BG
1587/*
1588 * Migrate an array of page address onto an array of nodes and fill
1589 * the corresponding array of status.
1590 */
3268c63e 1591static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1592 unsigned long nr_pages,
1593 const void __user * __user *pages,
1594 const int __user *nodes,
1595 int __user *status, int flags)
1596{
a49bd4d7
MH
1597 int current_node = NUMA_NO_NODE;
1598 LIST_HEAD(pagelist);
1599 int start, i;
1600 int err = 0, err1;
35282a2d
BG
1601
1602 migrate_prep();
1603
a49bd4d7
MH
1604 for (i = start = 0; i < nr_pages; i++) {
1605 const void __user *p;
1606 unsigned long addr;
1607 int node;
3140a227 1608
a49bd4d7
MH
1609 err = -EFAULT;
1610 if (get_user(p, pages + i))
1611 goto out_flush;
1612 if (get_user(node, nodes + i))
1613 goto out_flush;
1614 addr = (unsigned long)p;
1615
1616 err = -ENODEV;
1617 if (node < 0 || node >= MAX_NUMNODES)
1618 goto out_flush;
1619 if (!node_state(node, N_MEMORY))
1620 goto out_flush;
5e9a0f02 1621
a49bd4d7
MH
1622 err = -EACCES;
1623 if (!node_isset(node, task_nodes))
1624 goto out_flush;
1625
1626 if (current_node == NUMA_NO_NODE) {
1627 current_node = node;
1628 start = i;
1629 } else if (node != current_node) {
1630 err = do_move_pages_to_node(mm, &pagelist, current_node);
1631 if (err)
1632 goto out;
1633 err = store_status(status, start, current_node, i - start);
1634 if (err)
1635 goto out;
1636 start = i;
1637 current_node = node;
3140a227
BG
1638 }
1639
a49bd4d7
MH
1640 /*
1641 * Errors in the page lookup or isolation are not fatal and we simply
1642 * report them via status
1643 */
1644 err = add_page_for_migration(mm, addr, current_node,
1645 &pagelist, flags & MPOL_MF_MOVE_ALL);
1646 if (!err)
1647 continue;
3140a227 1648
a49bd4d7
MH
1649 err = store_status(status, i, err, 1);
1650 if (err)
1651 goto out_flush;
5e9a0f02 1652
a49bd4d7
MH
1653 err = do_move_pages_to_node(mm, &pagelist, current_node);
1654 if (err)
1655 goto out;
1656 if (i > start) {
1657 err = store_status(status, start, current_node, i - start);
1658 if (err)
1659 goto out;
1660 }
1661 current_node = NUMA_NO_NODE;
3140a227 1662 }
a49bd4d7 1663out_flush:
8f175cf5
MH
1664 if (list_empty(&pagelist))
1665 return err;
1666
a49bd4d7
MH
1667 /* Make sure we do not overwrite the existing error */
1668 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1669 if (!err1)
1670 err1 = store_status(status, start, current_node, i - start);
1671 if (!err)
1672 err = err1;
5e9a0f02
BG
1673out:
1674 return err;
1675}
1676
742755a1 1677/*
2f007e74 1678 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1679 */
80bba129
BG
1680static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1681 const void __user **pages, int *status)
742755a1 1682{
2f007e74 1683 unsigned long i;
2f007e74 1684
742755a1
CL
1685 down_read(&mm->mmap_sem);
1686
2f007e74 1687 for (i = 0; i < nr_pages; i++) {
80bba129 1688 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1689 struct vm_area_struct *vma;
1690 struct page *page;
c095adbc 1691 int err = -EFAULT;
2f007e74
BG
1692
1693 vma = find_vma(mm, addr);
70384dc6 1694 if (!vma || addr < vma->vm_start)
742755a1
CL
1695 goto set_status;
1696
d899844e
KS
1697 /* FOLL_DUMP to ignore special (like zero) pages */
1698 page = follow_page(vma, addr, FOLL_DUMP);
89f5b7da
LT
1699
1700 err = PTR_ERR(page);
1701 if (IS_ERR(page))
1702 goto set_status;
1703
d899844e 1704 err = page ? page_to_nid(page) : -ENOENT;
742755a1 1705set_status:
80bba129
BG
1706 *status = err;
1707
1708 pages++;
1709 status++;
1710 }
1711
1712 up_read(&mm->mmap_sem);
1713}
1714
1715/*
1716 * Determine the nodes of a user array of pages and store it in
1717 * a user array of status.
1718 */
1719static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1720 const void __user * __user *pages,
1721 int __user *status)
1722{
1723#define DO_PAGES_STAT_CHUNK_NR 16
1724 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1725 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1726
87b8d1ad
PA
1727 while (nr_pages) {
1728 unsigned long chunk_nr;
80bba129 1729
87b8d1ad
PA
1730 chunk_nr = nr_pages;
1731 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1732 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1733
1734 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1735 break;
80bba129
BG
1736
1737 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1738
87b8d1ad
PA
1739 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1740 break;
742755a1 1741
87b8d1ad
PA
1742 pages += chunk_nr;
1743 status += chunk_nr;
1744 nr_pages -= chunk_nr;
1745 }
1746 return nr_pages ? -EFAULT : 0;
742755a1
CL
1747}
1748
1749/*
1750 * Move a list of pages in the address space of the currently executing
1751 * process.
1752 */
7addf443
DB
1753static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1754 const void __user * __user *pages,
1755 const int __user *nodes,
1756 int __user *status, int flags)
742755a1 1757{
742755a1 1758 struct task_struct *task;
742755a1 1759 struct mm_struct *mm;
5e9a0f02 1760 int err;
3268c63e 1761 nodemask_t task_nodes;
742755a1
CL
1762
1763 /* Check flags */
1764 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1765 return -EINVAL;
1766
1767 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1768 return -EPERM;
1769
1770 /* Find the mm_struct */
a879bf58 1771 rcu_read_lock();
228ebcbe 1772 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1773 if (!task) {
a879bf58 1774 rcu_read_unlock();
742755a1
CL
1775 return -ESRCH;
1776 }
3268c63e 1777 get_task_struct(task);
742755a1
CL
1778
1779 /*
1780 * Check if this process has the right to modify the specified
197e7e52 1781 * process. Use the regular "ptrace_may_access()" checks.
742755a1 1782 */
197e7e52 1783 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
c69e8d9c 1784 rcu_read_unlock();
742755a1 1785 err = -EPERM;
5e9a0f02 1786 goto out;
742755a1 1787 }
c69e8d9c 1788 rcu_read_unlock();
742755a1 1789
86c3a764
DQ
1790 err = security_task_movememory(task);
1791 if (err)
5e9a0f02 1792 goto out;
86c3a764 1793
3268c63e
CL
1794 task_nodes = cpuset_mems_allowed(task);
1795 mm = get_task_mm(task);
1796 put_task_struct(task);
1797
6e8b09ea
SL
1798 if (!mm)
1799 return -EINVAL;
1800
1801 if (nodes)
1802 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1803 nodes, status, flags);
1804 else
1805 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1806
742755a1
CL
1807 mmput(mm);
1808 return err;
3268c63e
CL
1809
1810out:
1811 put_task_struct(task);
1812 return err;
742755a1 1813}
742755a1 1814
7addf443
DB
1815SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1816 const void __user * __user *, pages,
1817 const int __user *, nodes,
1818 int __user *, status, int, flags)
1819{
1820 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1821}
1822
1823#ifdef CONFIG_COMPAT
1824COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1825 compat_uptr_t __user *, pages32,
1826 const int __user *, nodes,
1827 int __user *, status,
1828 int, flags)
1829{
1830 const void __user * __user *pages;
1831 int i;
1832
1833 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1834 for (i = 0; i < nr_pages; i++) {
1835 compat_uptr_t p;
1836
1837 if (get_user(p, pages32 + i) ||
1838 put_user(compat_ptr(p), pages + i))
1839 return -EFAULT;
1840 }
1841 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1842}
1843#endif /* CONFIG_COMPAT */
1844
7039e1db
PZ
1845#ifdef CONFIG_NUMA_BALANCING
1846/*
1847 * Returns true if this is a safe migration target node for misplaced NUMA
1848 * pages. Currently it only checks the watermarks which crude
1849 */
1850static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1851 unsigned long nr_migrate_pages)
7039e1db
PZ
1852{
1853 int z;
599d0c95 1854
7039e1db
PZ
1855 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1856 struct zone *zone = pgdat->node_zones + z;
1857
1858 if (!populated_zone(zone))
1859 continue;
1860
7039e1db
PZ
1861 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1862 if (!zone_watermark_ok(zone, 0,
1863 high_wmark_pages(zone) +
1864 nr_migrate_pages,
1865 0, 0))
1866 continue;
1867 return true;
1868 }
1869 return false;
1870}
1871
1872static struct page *alloc_misplaced_dst_page(struct page *page,
666feb21 1873 unsigned long data)
7039e1db
PZ
1874{
1875 int nid = (int) data;
1876 struct page *newpage;
1877
96db800f 1878 newpage = __alloc_pages_node(nid,
e97ca8e5
JW
1879 (GFP_HIGHUSER_MOVABLE |
1880 __GFP_THISNODE | __GFP_NOMEMALLOC |
1881 __GFP_NORETRY | __GFP_NOWARN) &
8479eba7 1882 ~__GFP_RECLAIM, 0);
bac0382c 1883
7039e1db
PZ
1884 return newpage;
1885}
1886
1c30e017 1887static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
b32967ff 1888{
340ef390 1889 int page_lru;
a8f60772 1890
309381fe 1891 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
3abef4e6 1892
7039e1db 1893 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1894 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1895 return 0;
7039e1db 1896
340ef390
HD
1897 if (isolate_lru_page(page))
1898 return 0;
7039e1db 1899
340ef390
HD
1900 /*
1901 * migrate_misplaced_transhuge_page() skips page migration's usual
1902 * check on page_count(), so we must do it here, now that the page
1903 * has been isolated: a GUP pin, or any other pin, prevents migration.
1904 * The expected page count is 3: 1 for page's mapcount and 1 for the
1905 * caller's pin and 1 for the reference taken by isolate_lru_page().
1906 */
1907 if (PageTransHuge(page) && page_count(page) != 3) {
1908 putback_lru_page(page);
1909 return 0;
7039e1db
PZ
1910 }
1911
340ef390 1912 page_lru = page_is_file_cache(page);
599d0c95 1913 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
340ef390
HD
1914 hpage_nr_pages(page));
1915
149c33e1 1916 /*
340ef390
HD
1917 * Isolating the page has taken another reference, so the
1918 * caller's reference can be safely dropped without the page
1919 * disappearing underneath us during migration.
149c33e1
MG
1920 */
1921 put_page(page);
340ef390 1922 return 1;
b32967ff
MG
1923}
1924
de466bd6
MG
1925bool pmd_trans_migrating(pmd_t pmd)
1926{
1927 struct page *page = pmd_page(pmd);
1928 return PageLocked(page);
1929}
1930
b32967ff
MG
1931/*
1932 * Attempt to migrate a misplaced page to the specified destination
1933 * node. Caller is expected to have an elevated reference count on
1934 * the page that will be dropped by this function before returning.
1935 */
1bc115d8
MG
1936int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1937 int node)
b32967ff
MG
1938{
1939 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1940 int isolated;
b32967ff
MG
1941 int nr_remaining;
1942 LIST_HEAD(migratepages);
1943
1944 /*
1bc115d8
MG
1945 * Don't migrate file pages that are mapped in multiple processes
1946 * with execute permissions as they are probably shared libraries.
b32967ff 1947 */
1bc115d8
MG
1948 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1949 (vma->vm_flags & VM_EXEC))
b32967ff 1950 goto out;
b32967ff 1951
09a913a7
MG
1952 /*
1953 * Also do not migrate dirty pages as not all filesystems can move
1954 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1955 */
1956 if (page_is_file_cache(page) && PageDirty(page))
1957 goto out;
1958
b32967ff
MG
1959 isolated = numamigrate_isolate_page(pgdat, page);
1960 if (!isolated)
1961 goto out;
1962
1963 list_add(&page->lru, &migratepages);
9c620e2b 1964 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
68711a74
DR
1965 NULL, node, MIGRATE_ASYNC,
1966 MR_NUMA_MISPLACED);
b32967ff 1967 if (nr_remaining) {
59c82b70
JK
1968 if (!list_empty(&migratepages)) {
1969 list_del(&page->lru);
599d0c95 1970 dec_node_page_state(page, NR_ISOLATED_ANON +
59c82b70
JK
1971 page_is_file_cache(page));
1972 putback_lru_page(page);
1973 }
b32967ff
MG
1974 isolated = 0;
1975 } else
1976 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1977 BUG_ON(!list_empty(&migratepages));
7039e1db 1978 return isolated;
340ef390
HD
1979
1980out:
1981 put_page(page);
1982 return 0;
7039e1db 1983}
220018d3 1984#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1985
220018d3 1986#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1987/*
1988 * Migrates a THP to a given target node. page must be locked and is unlocked
1989 * before returning.
1990 */
b32967ff
MG
1991int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1992 struct vm_area_struct *vma,
1993 pmd_t *pmd, pmd_t entry,
1994 unsigned long address,
1995 struct page *page, int node)
1996{
c4088ebd 1997 spinlock_t *ptl;
b32967ff
MG
1998 pg_data_t *pgdat = NODE_DATA(node);
1999 int isolated = 0;
2000 struct page *new_page = NULL;
b32967ff 2001 int page_lru = page_is_file_cache(page);
7066f0f9 2002 unsigned long start = address & HPAGE_PMD_MASK;
b32967ff 2003
b32967ff 2004 new_page = alloc_pages_node(node,
25160354 2005 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
e97ca8e5 2006 HPAGE_PMD_ORDER);
340ef390
HD
2007 if (!new_page)
2008 goto out_fail;
9a982250 2009 prep_transhuge_page(new_page);
340ef390 2010
b32967ff 2011 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 2012 if (!isolated) {
b32967ff 2013 put_page(new_page);
340ef390 2014 goto out_fail;
b32967ff 2015 }
b0943d61 2016
b32967ff 2017 /* Prepare a page as a migration target */
48c935ad 2018 __SetPageLocked(new_page);
d44d363f
SL
2019 if (PageSwapBacked(page))
2020 __SetPageSwapBacked(new_page);
b32967ff
MG
2021
2022 /* anon mapping, we can simply copy page->mapping to the new page: */
2023 new_page->mapping = page->mapping;
2024 new_page->index = page->index;
7eef5f97
AA
2025 /* flush the cache before copying using the kernel virtual address */
2026 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
b32967ff
MG
2027 migrate_page_copy(new_page, page);
2028 WARN_ON(PageLRU(new_page));
2029
2030 /* Recheck the target PMD */
c4088ebd 2031 ptl = pmd_lock(mm, pmd);
f4e177d1 2032 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
c4088ebd 2033 spin_unlock(ptl);
b32967ff
MG
2034
2035 /* Reverse changes made by migrate_page_copy() */
2036 if (TestClearPageActive(new_page))
2037 SetPageActive(page);
2038 if (TestClearPageUnevictable(new_page))
2039 SetPageUnevictable(page);
b32967ff
MG
2040
2041 unlock_page(new_page);
2042 put_page(new_page); /* Free it */
2043
a54a407f
MG
2044 /* Retake the callers reference and putback on LRU */
2045 get_page(page);
b32967ff 2046 putback_lru_page(page);
599d0c95 2047 mod_node_page_state(page_pgdat(page),
a54a407f 2048 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
eb4489f6
MG
2049
2050 goto out_unlock;
b32967ff
MG
2051 }
2052
10102459 2053 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
f55e1014 2054 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
b32967ff 2055
2b4847e7 2056 /*
d7c33934
AA
2057 * Overwrite the old entry under pagetable lock and establish
2058 * the new PTE. Any parallel GUP will either observe the old
2059 * page blocking on the page lock, block on the page table
2060 * lock or observe the new page. The SetPageUptodate on the
2061 * new page and page_add_new_anon_rmap guarantee the copy is
2062 * visible before the pagetable update.
2b4847e7 2063 */
7066f0f9 2064 page_add_anon_rmap(new_page, vma, start, true);
d7c33934
AA
2065 /*
2066 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2067 * has already been flushed globally. So no TLB can be currently
2068 * caching this non present pmd mapping. There's no need to clear the
2069 * pmd before doing set_pmd_at(), nor to flush the TLB after
2070 * set_pmd_at(). Clearing the pmd here would introduce a race
2071 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2072 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2073 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2074 * pmd.
2075 */
7066f0f9 2076 set_pmd_at(mm, start, pmd, entry);
ce4a9cc5 2077 update_mmu_cache_pmd(vma, address, &entry);
2b4847e7 2078
f4e177d1 2079 page_ref_unfreeze(page, 2);
51afb12b 2080 mlock_migrate_page(new_page, page);
d281ee61 2081 page_remove_rmap(page, true);
7cd12b4a 2082 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2b4847e7 2083
c4088ebd 2084 spin_unlock(ptl);
b32967ff 2085
11de9927
MG
2086 /* Take an "isolate" reference and put new page on the LRU. */
2087 get_page(new_page);
2088 putback_lru_page(new_page);
2089
b32967ff
MG
2090 unlock_page(new_page);
2091 unlock_page(page);
2092 put_page(page); /* Drop the rmap reference */
2093 put_page(page); /* Drop the LRU isolation reference */
2094
2095 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2096 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2097
599d0c95 2098 mod_node_page_state(page_pgdat(page),
b32967ff
MG
2099 NR_ISOLATED_ANON + page_lru,
2100 -HPAGE_PMD_NR);
2101 return isolated;
2102
340ef390
HD
2103out_fail:
2104 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2b4847e7
MG
2105 ptl = pmd_lock(mm, pmd);
2106 if (pmd_same(*pmd, entry)) {
4d942466 2107 entry = pmd_modify(entry, vma->vm_page_prot);
7066f0f9 2108 set_pmd_at(mm, start, pmd, entry);
2b4847e7
MG
2109 update_mmu_cache_pmd(vma, address, &entry);
2110 }
2111 spin_unlock(ptl);
a54a407f 2112
eb4489f6 2113out_unlock:
340ef390 2114 unlock_page(page);
b32967ff 2115 put_page(page);
b32967ff
MG
2116 return 0;
2117}
7039e1db
PZ
2118#endif /* CONFIG_NUMA_BALANCING */
2119
2120#endif /* CONFIG_NUMA */
8763cb45 2121
9b2ed9cb 2122#ifdef CONFIG_DEVICE_PRIVATE
8763cb45
JG
2123static int migrate_vma_collect_hole(unsigned long start,
2124 unsigned long end,
2125 struct mm_walk *walk)
2126{
2127 struct migrate_vma *migrate = walk->private;
2128 unsigned long addr;
2129
8315ada7 2130 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
e20d103b 2131 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
8315ada7 2132 migrate->dst[migrate->npages] = 0;
e20d103b 2133 migrate->npages++;
8315ada7
JG
2134 migrate->cpages++;
2135 }
2136
2137 return 0;
2138}
2139
2140static int migrate_vma_collect_skip(unsigned long start,
2141 unsigned long end,
2142 struct mm_walk *walk)
2143{
2144 struct migrate_vma *migrate = walk->private;
2145 unsigned long addr;
2146
8763cb45
JG
2147 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2148 migrate->dst[migrate->npages] = 0;
2149 migrate->src[migrate->npages++] = 0;
2150 }
2151
2152 return 0;
2153}
2154
2155static int migrate_vma_collect_pmd(pmd_t *pmdp,
2156 unsigned long start,
2157 unsigned long end,
2158 struct mm_walk *walk)
2159{
2160 struct migrate_vma *migrate = walk->private;
2161 struct vm_area_struct *vma = walk->vma;
2162 struct mm_struct *mm = vma->vm_mm;
8c3328f1 2163 unsigned long addr = start, unmapped = 0;
8763cb45
JG
2164 spinlock_t *ptl;
2165 pte_t *ptep;
2166
2167again:
2168 if (pmd_none(*pmdp))
2169 return migrate_vma_collect_hole(start, end, walk);
2170
2171 if (pmd_trans_huge(*pmdp)) {
2172 struct page *page;
2173
2174 ptl = pmd_lock(mm, pmdp);
2175 if (unlikely(!pmd_trans_huge(*pmdp))) {
2176 spin_unlock(ptl);
2177 goto again;
2178 }
2179
2180 page = pmd_page(*pmdp);
2181 if (is_huge_zero_page(page)) {
2182 spin_unlock(ptl);
2183 split_huge_pmd(vma, pmdp, addr);
2184 if (pmd_trans_unstable(pmdp))
8315ada7 2185 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2186 walk);
2187 } else {
2188 int ret;
2189
2190 get_page(page);
2191 spin_unlock(ptl);
2192 if (unlikely(!trylock_page(page)))
8315ada7 2193 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2194 walk);
2195 ret = split_huge_page(page);
2196 unlock_page(page);
2197 put_page(page);
8315ada7
JG
2198 if (ret)
2199 return migrate_vma_collect_skip(start, end,
2200 walk);
2201 if (pmd_none(*pmdp))
8763cb45
JG
2202 return migrate_vma_collect_hole(start, end,
2203 walk);
2204 }
2205 }
2206
2207 if (unlikely(pmd_bad(*pmdp)))
8315ada7 2208 return migrate_vma_collect_skip(start, end, walk);
8763cb45
JG
2209
2210 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
8c3328f1
JG
2211 arch_enter_lazy_mmu_mode();
2212
8763cb45
JG
2213 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2214 unsigned long mpfn, pfn;
2215 struct page *page;
8c3328f1 2216 swp_entry_t entry;
8763cb45
JG
2217 pte_t pte;
2218
2219 pte = *ptep;
2220 pfn = pte_pfn(pte);
2221
a5430dda 2222 if (pte_none(pte)) {
8315ada7
JG
2223 mpfn = MIGRATE_PFN_MIGRATE;
2224 migrate->cpages++;
2225 pfn = 0;
8763cb45
JG
2226 goto next;
2227 }
2228
a5430dda
JG
2229 if (!pte_present(pte)) {
2230 mpfn = pfn = 0;
2231
2232 /*
2233 * Only care about unaddressable device page special
2234 * page table entry. Other special swap entries are not
2235 * migratable, and we ignore regular swapped page.
2236 */
2237 entry = pte_to_swp_entry(pte);
2238 if (!is_device_private_entry(entry))
2239 goto next;
2240
2241 page = device_private_entry_to_page(entry);
06d462be
CH
2242 mpfn = migrate_pfn(page_to_pfn(page)) |
2243 MIGRATE_PFN_MIGRATE;
a5430dda
JG
2244 if (is_write_device_private_entry(entry))
2245 mpfn |= MIGRATE_PFN_WRITE;
2246 } else {
8315ada7
JG
2247 if (is_zero_pfn(pfn)) {
2248 mpfn = MIGRATE_PFN_MIGRATE;
2249 migrate->cpages++;
2250 pfn = 0;
2251 goto next;
2252 }
25b2995a 2253 page = vm_normal_page(migrate->vma, addr, pte);
a5430dda
JG
2254 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2255 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2256 }
2257
8763cb45 2258 /* FIXME support THP */
8763cb45
JG
2259 if (!page || !page->mapping || PageTransCompound(page)) {
2260 mpfn = pfn = 0;
2261 goto next;
2262 }
a5430dda 2263 pfn = page_to_pfn(page);
8763cb45
JG
2264
2265 /*
2266 * By getting a reference on the page we pin it and that blocks
2267 * any kind of migration. Side effect is that it "freezes" the
2268 * pte.
2269 *
2270 * We drop this reference after isolating the page from the lru
2271 * for non device page (device page are not on the lru and thus
2272 * can't be dropped from it).
2273 */
2274 get_page(page);
2275 migrate->cpages++;
8763cb45 2276
8c3328f1
JG
2277 /*
2278 * Optimize for the common case where page is only mapped once
2279 * in one process. If we can lock the page, then we can safely
2280 * set up a special migration page table entry now.
2281 */
2282 if (trylock_page(page)) {
2283 pte_t swp_pte;
2284
2285 mpfn |= MIGRATE_PFN_LOCKED;
2286 ptep_get_and_clear(mm, addr, ptep);
2287
2288 /* Setup special migration page table entry */
07707125
RC
2289 entry = make_migration_entry(page, mpfn &
2290 MIGRATE_PFN_WRITE);
8c3328f1
JG
2291 swp_pte = swp_entry_to_pte(entry);
2292 if (pte_soft_dirty(pte))
2293 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2294 set_pte_at(mm, addr, ptep, swp_pte);
2295
2296 /*
2297 * This is like regular unmap: we remove the rmap and
2298 * drop page refcount. Page won't be freed, as we took
2299 * a reference just above.
2300 */
2301 page_remove_rmap(page, false);
2302 put_page(page);
a5430dda
JG
2303
2304 if (pte_present(pte))
2305 unmapped++;
8c3328f1
JG
2306 }
2307
8763cb45 2308next:
a5430dda 2309 migrate->dst[migrate->npages] = 0;
8763cb45
JG
2310 migrate->src[migrate->npages++] = mpfn;
2311 }
8c3328f1 2312 arch_leave_lazy_mmu_mode();
8763cb45
JG
2313 pte_unmap_unlock(ptep - 1, ptl);
2314
8c3328f1
JG
2315 /* Only flush the TLB if we actually modified any entries */
2316 if (unmapped)
2317 flush_tlb_range(walk->vma, start, end);
2318
8763cb45
JG
2319 return 0;
2320}
2321
2322/*
2323 * migrate_vma_collect() - collect pages over a range of virtual addresses
2324 * @migrate: migrate struct containing all migration information
2325 *
2326 * This will walk the CPU page table. For each virtual address backed by a
2327 * valid page, it updates the src array and takes a reference on the page, in
2328 * order to pin the page until we lock it and unmap it.
2329 */
2330static void migrate_vma_collect(struct migrate_vma *migrate)
2331{
ac46d4f3 2332 struct mmu_notifier_range range;
7b358c6f
RC
2333 struct mm_walk mm_walk = {
2334 .pmd_entry = migrate_vma_collect_pmd,
2335 .pte_hole = migrate_vma_collect_hole,
2336 .vma = migrate->vma,
2337 .mm = migrate->vma->vm_mm,
2338 .private = migrate,
2339 };
8763cb45 2340
7269f999 2341 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm_walk.mm,
6f4f13e8 2342 migrate->start,
ac46d4f3
JG
2343 migrate->end);
2344 mmu_notifier_invalidate_range_start(&range);
8763cb45 2345 walk_page_range(migrate->start, migrate->end, &mm_walk);
ac46d4f3 2346 mmu_notifier_invalidate_range_end(&range);
8763cb45
JG
2347
2348 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2349}
2350
2351/*
2352 * migrate_vma_check_page() - check if page is pinned or not
2353 * @page: struct page to check
2354 *
2355 * Pinned pages cannot be migrated. This is the same test as in
2356 * migrate_page_move_mapping(), except that here we allow migration of a
2357 * ZONE_DEVICE page.
2358 */
2359static bool migrate_vma_check_page(struct page *page)
2360{
2361 /*
2362 * One extra ref because caller holds an extra reference, either from
2363 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2364 * a device page.
2365 */
2366 int extra = 1;
2367
2368 /*
2369 * FIXME support THP (transparent huge page), it is bit more complex to
2370 * check them than regular pages, because they can be mapped with a pmd
2371 * or with a pte (split pte mapping).
2372 */
2373 if (PageCompound(page))
2374 return false;
2375
a5430dda
JG
2376 /* Page from ZONE_DEVICE have one extra reference */
2377 if (is_zone_device_page(page)) {
2378 /*
2379 * Private page can never be pin as they have no valid pte and
2380 * GUP will fail for those. Yet if there is a pending migration
2381 * a thread might try to wait on the pte migration entry and
2382 * will bump the page reference count. Sadly there is no way to
2383 * differentiate a regular pin from migration wait. Hence to
2384 * avoid 2 racing thread trying to migrate back to CPU to enter
2385 * infinite loop (one stoping migration because the other is
2386 * waiting on pte migration entry). We always return true here.
2387 *
2388 * FIXME proper solution is to rework migration_entry_wait() so
2389 * it does not need to take a reference on page.
2390 */
25b2995a 2391 return is_device_private_page(page);
a5430dda
JG
2392 }
2393
df6ad698
JG
2394 /* For file back page */
2395 if (page_mapping(page))
2396 extra += 1 + page_has_private(page);
2397
8763cb45
JG
2398 if ((page_count(page) - extra) > page_mapcount(page))
2399 return false;
2400
2401 return true;
2402}
2403
2404/*
2405 * migrate_vma_prepare() - lock pages and isolate them from the lru
2406 * @migrate: migrate struct containing all migration information
2407 *
2408 * This locks pages that have been collected by migrate_vma_collect(). Once each
2409 * page is locked it is isolated from the lru (for non-device pages). Finally,
2410 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2411 * migrated by concurrent kernel threads.
2412 */
2413static void migrate_vma_prepare(struct migrate_vma *migrate)
2414{
2415 const unsigned long npages = migrate->npages;
8c3328f1
JG
2416 const unsigned long start = migrate->start;
2417 unsigned long addr, i, restore = 0;
8763cb45 2418 bool allow_drain = true;
8763cb45
JG
2419
2420 lru_add_drain();
2421
2422 for (i = 0; (i < npages) && migrate->cpages; i++) {
2423 struct page *page = migrate_pfn_to_page(migrate->src[i]);
8c3328f1 2424 bool remap = true;
8763cb45
JG
2425
2426 if (!page)
2427 continue;
2428
8c3328f1
JG
2429 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2430 /*
2431 * Because we are migrating several pages there can be
2432 * a deadlock between 2 concurrent migration where each
2433 * are waiting on each other page lock.
2434 *
2435 * Make migrate_vma() a best effort thing and backoff
2436 * for any page we can not lock right away.
2437 */
2438 if (!trylock_page(page)) {
2439 migrate->src[i] = 0;
2440 migrate->cpages--;
2441 put_page(page);
2442 continue;
2443 }
2444 remap = false;
2445 migrate->src[i] |= MIGRATE_PFN_LOCKED;
8763cb45 2446 }
8763cb45 2447
a5430dda
JG
2448 /* ZONE_DEVICE pages are not on LRU */
2449 if (!is_zone_device_page(page)) {
2450 if (!PageLRU(page) && allow_drain) {
2451 /* Drain CPU's pagevec */
2452 lru_add_drain_all();
2453 allow_drain = false;
2454 }
8763cb45 2455
a5430dda
JG
2456 if (isolate_lru_page(page)) {
2457 if (remap) {
2458 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2459 migrate->cpages--;
2460 restore++;
2461 } else {
2462 migrate->src[i] = 0;
2463 unlock_page(page);
2464 migrate->cpages--;
2465 put_page(page);
2466 }
2467 continue;
8c3328f1 2468 }
a5430dda
JG
2469
2470 /* Drop the reference we took in collect */
2471 put_page(page);
8763cb45
JG
2472 }
2473
2474 if (!migrate_vma_check_page(page)) {
8c3328f1
JG
2475 if (remap) {
2476 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2477 migrate->cpages--;
2478 restore++;
8763cb45 2479
a5430dda
JG
2480 if (!is_zone_device_page(page)) {
2481 get_page(page);
2482 putback_lru_page(page);
2483 }
8c3328f1
JG
2484 } else {
2485 migrate->src[i] = 0;
2486 unlock_page(page);
2487 migrate->cpages--;
2488
a5430dda
JG
2489 if (!is_zone_device_page(page))
2490 putback_lru_page(page);
2491 else
2492 put_page(page);
8c3328f1 2493 }
8763cb45
JG
2494 }
2495 }
8c3328f1
JG
2496
2497 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2498 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2499
2500 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2501 continue;
2502
2503 remove_migration_pte(page, migrate->vma, addr, page);
2504
2505 migrate->src[i] = 0;
2506 unlock_page(page);
2507 put_page(page);
2508 restore--;
2509 }
8763cb45
JG
2510}
2511
2512/*
2513 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2514 * @migrate: migrate struct containing all migration information
2515 *
2516 * Replace page mapping (CPU page table pte) with a special migration pte entry
2517 * and check again if it has been pinned. Pinned pages are restored because we
2518 * cannot migrate them.
2519 *
2520 * This is the last step before we call the device driver callback to allocate
2521 * destination memory and copy contents of original page over to new page.
2522 */
2523static void migrate_vma_unmap(struct migrate_vma *migrate)
2524{
2525 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2526 const unsigned long npages = migrate->npages;
2527 const unsigned long start = migrate->start;
2528 unsigned long addr, i, restore = 0;
2529
2530 for (i = 0; i < npages; i++) {
2531 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2532
2533 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2534 continue;
2535
8c3328f1
JG
2536 if (page_mapped(page)) {
2537 try_to_unmap(page, flags);
2538 if (page_mapped(page))
2539 goto restore;
8763cb45 2540 }
8c3328f1
JG
2541
2542 if (migrate_vma_check_page(page))
2543 continue;
2544
2545restore:
2546 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2547 migrate->cpages--;
2548 restore++;
8763cb45
JG
2549 }
2550
2551 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2552 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2553
2554 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2555 continue;
2556
2557 remove_migration_ptes(page, page, false);
2558
2559 migrate->src[i] = 0;
2560 unlock_page(page);
2561 restore--;
2562
a5430dda
JG
2563 if (is_zone_device_page(page))
2564 put_page(page);
2565 else
2566 putback_lru_page(page);
8763cb45
JG
2567 }
2568}
2569
a7d1f22b
CH
2570/**
2571 * migrate_vma_setup() - prepare to migrate a range of memory
2572 * @args: contains the vma, start, and and pfns arrays for the migration
2573 *
2574 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2575 * without an error.
2576 *
2577 * Prepare to migrate a range of memory virtual address range by collecting all
2578 * the pages backing each virtual address in the range, saving them inside the
2579 * src array. Then lock those pages and unmap them. Once the pages are locked
2580 * and unmapped, check whether each page is pinned or not. Pages that aren't
2581 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2582 * corresponding src array entry. Then restores any pages that are pinned, by
2583 * remapping and unlocking those pages.
2584 *
2585 * The caller should then allocate destination memory and copy source memory to
2586 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2587 * flag set). Once these are allocated and copied, the caller must update each
2588 * corresponding entry in the dst array with the pfn value of the destination
2589 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2590 * (destination pages must have their struct pages locked, via lock_page()).
2591 *
2592 * Note that the caller does not have to migrate all the pages that are marked
2593 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2594 * device memory to system memory. If the caller cannot migrate a device page
2595 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2596 * consequences for the userspace process, so it must be avoided if at all
2597 * possible.
2598 *
2599 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2600 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2601 * allowing the caller to allocate device memory for those unback virtual
2602 * address. For this the caller simply has to allocate device memory and
2603 * properly set the destination entry like for regular migration. Note that
2604 * this can still fails and thus inside the device driver must check if the
2605 * migration was successful for those entries after calling migrate_vma_pages()
2606 * just like for regular migration.
2607 *
2608 * After that, the callers must call migrate_vma_pages() to go over each entry
2609 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2610 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2611 * then migrate_vma_pages() to migrate struct page information from the source
2612 * struct page to the destination struct page. If it fails to migrate the
2613 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2614 * src array.
2615 *
2616 * At this point all successfully migrated pages have an entry in the src
2617 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2618 * array entry with MIGRATE_PFN_VALID flag set.
2619 *
2620 * Once migrate_vma_pages() returns the caller may inspect which pages were
2621 * successfully migrated, and which were not. Successfully migrated pages will
2622 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2623 *
2624 * It is safe to update device page table after migrate_vma_pages() because
2625 * both destination and source page are still locked, and the mmap_sem is held
2626 * in read mode (hence no one can unmap the range being migrated).
2627 *
2628 * Once the caller is done cleaning up things and updating its page table (if it
2629 * chose to do so, this is not an obligation) it finally calls
2630 * migrate_vma_finalize() to update the CPU page table to point to new pages
2631 * for successfully migrated pages or otherwise restore the CPU page table to
2632 * point to the original source pages.
2633 */
2634int migrate_vma_setup(struct migrate_vma *args)
2635{
2636 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2637
2638 args->start &= PAGE_MASK;
2639 args->end &= PAGE_MASK;
2640 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2641 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2642 return -EINVAL;
2643 if (nr_pages <= 0)
2644 return -EINVAL;
2645 if (args->start < args->vma->vm_start ||
2646 args->start >= args->vma->vm_end)
2647 return -EINVAL;
2648 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2649 return -EINVAL;
2650 if (!args->src || !args->dst)
2651 return -EINVAL;
2652
2653 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2654 args->cpages = 0;
2655 args->npages = 0;
2656
2657 migrate_vma_collect(args);
2658
2659 if (args->cpages)
2660 migrate_vma_prepare(args);
2661 if (args->cpages)
2662 migrate_vma_unmap(args);
2663
2664 /*
2665 * At this point pages are locked and unmapped, and thus they have
2666 * stable content and can safely be copied to destination memory that
2667 * is allocated by the drivers.
2668 */
2669 return 0;
2670
2671}
2672EXPORT_SYMBOL(migrate_vma_setup);
2673
8315ada7
JG
2674static void migrate_vma_insert_page(struct migrate_vma *migrate,
2675 unsigned long addr,
2676 struct page *page,
2677 unsigned long *src,
2678 unsigned long *dst)
2679{
2680 struct vm_area_struct *vma = migrate->vma;
2681 struct mm_struct *mm = vma->vm_mm;
2682 struct mem_cgroup *memcg;
2683 bool flush = false;
2684 spinlock_t *ptl;
2685 pte_t entry;
2686 pgd_t *pgdp;
2687 p4d_t *p4dp;
2688 pud_t *pudp;
2689 pmd_t *pmdp;
2690 pte_t *ptep;
2691
2692 /* Only allow populating anonymous memory */
2693 if (!vma_is_anonymous(vma))
2694 goto abort;
2695
2696 pgdp = pgd_offset(mm, addr);
2697 p4dp = p4d_alloc(mm, pgdp, addr);
2698 if (!p4dp)
2699 goto abort;
2700 pudp = pud_alloc(mm, p4dp, addr);
2701 if (!pudp)
2702 goto abort;
2703 pmdp = pmd_alloc(mm, pudp, addr);
2704 if (!pmdp)
2705 goto abort;
2706
2707 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2708 goto abort;
2709
2710 /*
2711 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2712 * pte_offset_map() on pmds where a huge pmd might be created
2713 * from a different thread.
2714 *
2715 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2716 * parallel threads are excluded by other means.
2717 *
2718 * Here we only have down_read(mmap_sem).
2719 */
4cf58924 2720 if (pte_alloc(mm, pmdp))
8315ada7
JG
2721 goto abort;
2722
2723 /* See the comment in pte_alloc_one_map() */
2724 if (unlikely(pmd_trans_unstable(pmdp)))
2725 goto abort;
2726
2727 if (unlikely(anon_vma_prepare(vma)))
2728 goto abort;
2729 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2730 goto abort;
2731
2732 /*
2733 * The memory barrier inside __SetPageUptodate makes sure that
2734 * preceding stores to the page contents become visible before
2735 * the set_pte_at() write.
2736 */
2737 __SetPageUptodate(page);
2738
df6ad698
JG
2739 if (is_zone_device_page(page)) {
2740 if (is_device_private_page(page)) {
2741 swp_entry_t swp_entry;
2742
2743 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2744 entry = swp_entry_to_pte(swp_entry);
df6ad698 2745 }
8315ada7
JG
2746 } else {
2747 entry = mk_pte(page, vma->vm_page_prot);
2748 if (vma->vm_flags & VM_WRITE)
2749 entry = pte_mkwrite(pte_mkdirty(entry));
2750 }
2751
2752 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2753
2754 if (pte_present(*ptep)) {
2755 unsigned long pfn = pte_pfn(*ptep);
2756
2757 if (!is_zero_pfn(pfn)) {
2758 pte_unmap_unlock(ptep, ptl);
2759 mem_cgroup_cancel_charge(page, memcg, false);
2760 goto abort;
2761 }
2762 flush = true;
2763 } else if (!pte_none(*ptep)) {
2764 pte_unmap_unlock(ptep, ptl);
2765 mem_cgroup_cancel_charge(page, memcg, false);
2766 goto abort;
2767 }
2768
2769 /*
2770 * Check for usefaultfd but do not deliver the fault. Instead,
2771 * just back off.
2772 */
2773 if (userfaultfd_missing(vma)) {
2774 pte_unmap_unlock(ptep, ptl);
2775 mem_cgroup_cancel_charge(page, memcg, false);
2776 goto abort;
2777 }
2778
2779 inc_mm_counter(mm, MM_ANONPAGES);
2780 page_add_new_anon_rmap(page, vma, addr, false);
2781 mem_cgroup_commit_charge(page, memcg, false, false);
2782 if (!is_zone_device_page(page))
2783 lru_cache_add_active_or_unevictable(page, vma);
2784 get_page(page);
2785
2786 if (flush) {
2787 flush_cache_page(vma, addr, pte_pfn(*ptep));
2788 ptep_clear_flush_notify(vma, addr, ptep);
2789 set_pte_at_notify(mm, addr, ptep, entry);
2790 update_mmu_cache(vma, addr, ptep);
2791 } else {
2792 /* No need to invalidate - it was non-present before */
2793 set_pte_at(mm, addr, ptep, entry);
2794 update_mmu_cache(vma, addr, ptep);
2795 }
2796
2797 pte_unmap_unlock(ptep, ptl);
2798 *src = MIGRATE_PFN_MIGRATE;
2799 return;
2800
2801abort:
2802 *src &= ~MIGRATE_PFN_MIGRATE;
2803}
2804
a7d1f22b 2805/**
8763cb45
JG
2806 * migrate_vma_pages() - migrate meta-data from src page to dst page
2807 * @migrate: migrate struct containing all migration information
2808 *
2809 * This migrates struct page meta-data from source struct page to destination
2810 * struct page. This effectively finishes the migration from source page to the
2811 * destination page.
2812 */
a7d1f22b 2813void migrate_vma_pages(struct migrate_vma *migrate)
8763cb45
JG
2814{
2815 const unsigned long npages = migrate->npages;
2816 const unsigned long start = migrate->start;
ac46d4f3
JG
2817 struct mmu_notifier_range range;
2818 unsigned long addr, i;
8315ada7 2819 bool notified = false;
8763cb45
JG
2820
2821 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2822 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2823 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2824 struct address_space *mapping;
2825 int r;
2826
8315ada7
JG
2827 if (!newpage) {
2828 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
8763cb45 2829 continue;
8315ada7
JG
2830 }
2831
2832 if (!page) {
2833 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2834 continue;
2835 }
2836 if (!notified) {
8315ada7 2837 notified = true;
ac46d4f3
JG
2838
2839 mmu_notifier_range_init(&range,
7269f999 2840 MMU_NOTIFY_CLEAR, 0,
6f4f13e8 2841 NULL,
ac46d4f3
JG
2842 migrate->vma->vm_mm,
2843 addr, migrate->end);
2844 mmu_notifier_invalidate_range_start(&range);
8315ada7
JG
2845 }
2846 migrate_vma_insert_page(migrate, addr, newpage,
2847 &migrate->src[i],
2848 &migrate->dst[i]);
8763cb45 2849 continue;
8315ada7 2850 }
8763cb45
JG
2851
2852 mapping = page_mapping(page);
2853
a5430dda
JG
2854 if (is_zone_device_page(newpage)) {
2855 if (is_device_private_page(newpage)) {
2856 /*
2857 * For now only support private anonymous when
2858 * migrating to un-addressable device memory.
2859 */
2860 if (mapping) {
2861 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2862 continue;
2863 }
25b2995a 2864 } else {
a5430dda
JG
2865 /*
2866 * Other types of ZONE_DEVICE page are not
2867 * supported.
2868 */
2869 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2870 continue;
2871 }
2872 }
2873
8763cb45
JG
2874 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2875 if (r != MIGRATEPAGE_SUCCESS)
2876 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2877 }
8315ada7 2878
4645b9fe
JG
2879 /*
2880 * No need to double call mmu_notifier->invalidate_range() callback as
2881 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2882 * did already call it.
2883 */
8315ada7 2884 if (notified)
ac46d4f3 2885 mmu_notifier_invalidate_range_only_end(&range);
8763cb45 2886}
a7d1f22b 2887EXPORT_SYMBOL(migrate_vma_pages);
8763cb45 2888
a7d1f22b 2889/**
8763cb45
JG
2890 * migrate_vma_finalize() - restore CPU page table entry
2891 * @migrate: migrate struct containing all migration information
2892 *
2893 * This replaces the special migration pte entry with either a mapping to the
2894 * new page if migration was successful for that page, or to the original page
2895 * otherwise.
2896 *
2897 * This also unlocks the pages and puts them back on the lru, or drops the extra
2898 * refcount, for device pages.
2899 */
a7d1f22b 2900void migrate_vma_finalize(struct migrate_vma *migrate)
8763cb45
JG
2901{
2902 const unsigned long npages = migrate->npages;
2903 unsigned long i;
2904
2905 for (i = 0; i < npages; i++) {
2906 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2907 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2908
8315ada7
JG
2909 if (!page) {
2910 if (newpage) {
2911 unlock_page(newpage);
2912 put_page(newpage);
2913 }
8763cb45 2914 continue;
8315ada7
JG
2915 }
2916
8763cb45
JG
2917 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2918 if (newpage) {
2919 unlock_page(newpage);
2920 put_page(newpage);
2921 }
2922 newpage = page;
2923 }
2924
2925 remove_migration_ptes(page, newpage, false);
2926 unlock_page(page);
2927 migrate->cpages--;
2928
a5430dda
JG
2929 if (is_zone_device_page(page))
2930 put_page(page);
2931 else
2932 putback_lru_page(page);
8763cb45
JG
2933
2934 if (newpage != page) {
2935 unlock_page(newpage);
a5430dda
JG
2936 if (is_zone_device_page(newpage))
2937 put_page(newpage);
2938 else
2939 putback_lru_page(newpage);
8763cb45
JG
2940 }
2941 }
2942}
a7d1f22b 2943EXPORT_SYMBOL(migrate_vma_finalize);
9b2ed9cb 2944#endif /* CONFIG_DEVICE_PRIVATE */