]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
mm/mmu_notifier: use structure for invalidate_range_start/end callback
[mirror_ubuntu-jammy-kernel.git] / mm / migrate.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
b20a3503 2/*
14e0f9bc 3 * Memory Migration functionality - linux/mm/migrate.c
b20a3503
CL
4 *
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 *
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
9 *
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
cde53535 13 * Christoph Lameter
b20a3503
CL
14 */
15
16#include <linux/migrate.h>
b95f1b31 17#include <linux/export.h>
b20a3503 18#include <linux/swap.h>
0697212a 19#include <linux/swapops.h>
b20a3503 20#include <linux/pagemap.h>
e23ca00b 21#include <linux/buffer_head.h>
b20a3503 22#include <linux/mm_inline.h>
b488893a 23#include <linux/nsproxy.h>
b20a3503 24#include <linux/pagevec.h>
e9995ef9 25#include <linux/ksm.h>
b20a3503
CL
26#include <linux/rmap.h>
27#include <linux/topology.h>
28#include <linux/cpu.h>
29#include <linux/cpuset.h>
04e62a29 30#include <linux/writeback.h>
742755a1
CL
31#include <linux/mempolicy.h>
32#include <linux/vmalloc.h>
86c3a764 33#include <linux/security.h>
42cb14b1 34#include <linux/backing-dev.h>
bda807d4 35#include <linux/compaction.h>
4f5ca265 36#include <linux/syscalls.h>
7addf443 37#include <linux/compat.h>
290408d4 38#include <linux/hugetlb.h>
8e6ac7fa 39#include <linux/hugetlb_cgroup.h>
5a0e3ad6 40#include <linux/gfp.h>
df6ad698 41#include <linux/pfn_t.h>
a5430dda 42#include <linux/memremap.h>
8315ada7 43#include <linux/userfaultfd_k.h>
bf6bddf1 44#include <linux/balloon_compaction.h>
f714f4f2 45#include <linux/mmu_notifier.h>
33c3fc71 46#include <linux/page_idle.h>
d435edca 47#include <linux/page_owner.h>
6e84f315 48#include <linux/sched/mm.h>
197e7e52 49#include <linux/ptrace.h>
b20a3503 50
0d1836c3
MN
51#include <asm/tlbflush.h>
52
7b2a2d4a
MG
53#define CREATE_TRACE_POINTS
54#include <trace/events/migrate.h>
55
b20a3503
CL
56#include "internal.h"
57
b20a3503 58/*
742755a1 59 * migrate_prep() needs to be called before we start compiling a list of pages
748446bb
MG
60 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
61 * undesirable, use migrate_prep_local()
b20a3503
CL
62 */
63int migrate_prep(void)
64{
b20a3503
CL
65 /*
66 * Clear the LRU lists so pages can be isolated.
67 * Note that pages may be moved off the LRU after we have
68 * drained them. Those pages will fail to migrate like other
69 * pages that may be busy.
70 */
71 lru_add_drain_all();
72
73 return 0;
74}
75
748446bb
MG
76/* Do the necessary work of migrate_prep but not if it involves other CPUs */
77int migrate_prep_local(void)
78{
79 lru_add_drain();
80
81 return 0;
82}
83
9e5bcd61 84int isolate_movable_page(struct page *page, isolate_mode_t mode)
bda807d4
MK
85{
86 struct address_space *mapping;
87
88 /*
89 * Avoid burning cycles with pages that are yet under __free_pages(),
90 * or just got freed under us.
91 *
92 * In case we 'win' a race for a movable page being freed under us and
93 * raise its refcount preventing __free_pages() from doing its job
94 * the put_page() at the end of this block will take care of
95 * release this page, thus avoiding a nasty leakage.
96 */
97 if (unlikely(!get_page_unless_zero(page)))
98 goto out;
99
100 /*
101 * Check PageMovable before holding a PG_lock because page's owner
102 * assumes anybody doesn't touch PG_lock of newly allocated page
103 * so unconditionally grapping the lock ruins page's owner side.
104 */
105 if (unlikely(!__PageMovable(page)))
106 goto out_putpage;
107 /*
108 * As movable pages are not isolated from LRU lists, concurrent
109 * compaction threads can race against page migration functions
110 * as well as race against the releasing a page.
111 *
112 * In order to avoid having an already isolated movable page
113 * being (wrongly) re-isolated while it is under migration,
114 * or to avoid attempting to isolate pages being released,
115 * lets be sure we have the page lock
116 * before proceeding with the movable page isolation steps.
117 */
118 if (unlikely(!trylock_page(page)))
119 goto out_putpage;
120
121 if (!PageMovable(page) || PageIsolated(page))
122 goto out_no_isolated;
123
124 mapping = page_mapping(page);
125 VM_BUG_ON_PAGE(!mapping, page);
126
127 if (!mapping->a_ops->isolate_page(page, mode))
128 goto out_no_isolated;
129
130 /* Driver shouldn't use PG_isolated bit of page->flags */
131 WARN_ON_ONCE(PageIsolated(page));
132 __SetPageIsolated(page);
133 unlock_page(page);
134
9e5bcd61 135 return 0;
bda807d4
MK
136
137out_no_isolated:
138 unlock_page(page);
139out_putpage:
140 put_page(page);
141out:
9e5bcd61 142 return -EBUSY;
bda807d4
MK
143}
144
145/* It should be called on page which is PG_movable */
146void putback_movable_page(struct page *page)
147{
148 struct address_space *mapping;
149
150 VM_BUG_ON_PAGE(!PageLocked(page), page);
151 VM_BUG_ON_PAGE(!PageMovable(page), page);
152 VM_BUG_ON_PAGE(!PageIsolated(page), page);
153
154 mapping = page_mapping(page);
155 mapping->a_ops->putback_page(page);
156 __ClearPageIsolated(page);
157}
158
5733c7d1
RA
159/*
160 * Put previously isolated pages back onto the appropriate lists
161 * from where they were once taken off for compaction/migration.
162 *
59c82b70
JK
163 * This function shall be used whenever the isolated pageset has been
164 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
165 * and isolate_huge_page().
5733c7d1
RA
166 */
167void putback_movable_pages(struct list_head *l)
168{
169 struct page *page;
170 struct page *page2;
171
b20a3503 172 list_for_each_entry_safe(page, page2, l, lru) {
31caf665
NH
173 if (unlikely(PageHuge(page))) {
174 putback_active_hugepage(page);
175 continue;
176 }
e24f0b8f 177 list_del(&page->lru);
bda807d4
MK
178 /*
179 * We isolated non-lru movable page so here we can use
180 * __PageMovable because LRU page's mapping cannot have
181 * PAGE_MAPPING_MOVABLE.
182 */
b1123ea6 183 if (unlikely(__PageMovable(page))) {
bda807d4
MK
184 VM_BUG_ON_PAGE(!PageIsolated(page), page);
185 lock_page(page);
186 if (PageMovable(page))
187 putback_movable_page(page);
188 else
189 __ClearPageIsolated(page);
190 unlock_page(page);
191 put_page(page);
192 } else {
e8db67eb
NH
193 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
194 page_is_file_cache(page), -hpage_nr_pages(page));
fc280fe8 195 putback_lru_page(page);
bda807d4 196 }
b20a3503 197 }
b20a3503
CL
198}
199
0697212a
CL
200/*
201 * Restore a potential migration pte to a working pte entry
202 */
e4b82222 203static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
e9995ef9 204 unsigned long addr, void *old)
0697212a 205{
3fe87967
KS
206 struct page_vma_mapped_walk pvmw = {
207 .page = old,
208 .vma = vma,
209 .address = addr,
210 .flags = PVMW_SYNC | PVMW_MIGRATION,
211 };
212 struct page *new;
213 pte_t pte;
0697212a 214 swp_entry_t entry;
0697212a 215
3fe87967
KS
216 VM_BUG_ON_PAGE(PageTail(page), page);
217 while (page_vma_mapped_walk(&pvmw)) {
4b0ece6f
NH
218 if (PageKsm(page))
219 new = page;
220 else
221 new = page - pvmw.page->index +
222 linear_page_index(vma, pvmw.address);
0697212a 223
616b8371
ZY
224#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
225 /* PMD-mapped THP migration entry */
226 if (!pvmw.pte) {
227 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
228 remove_migration_pmd(&pvmw, new);
229 continue;
230 }
231#endif
232
3fe87967
KS
233 get_page(new);
234 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
235 if (pte_swp_soft_dirty(*pvmw.pte))
236 pte = pte_mksoft_dirty(pte);
0697212a 237
3fe87967
KS
238 /*
239 * Recheck VMA as permissions can change since migration started
240 */
241 entry = pte_to_swp_entry(*pvmw.pte);
242 if (is_write_migration_entry(entry))
243 pte = maybe_mkwrite(pte, vma);
d3cb8bf6 244
df6ad698
JG
245 if (unlikely(is_zone_device_page(new))) {
246 if (is_device_private_page(new)) {
247 entry = make_device_private_entry(new, pte_write(pte));
248 pte = swp_entry_to_pte(entry);
249 } else if (is_device_public_page(new)) {
250 pte = pte_mkdevmap(pte);
251 flush_dcache_page(new);
252 }
a5430dda
JG
253 } else
254 flush_dcache_page(new);
255
3ef8fd7f 256#ifdef CONFIG_HUGETLB_PAGE
3fe87967
KS
257 if (PageHuge(new)) {
258 pte = pte_mkhuge(pte);
259 pte = arch_make_huge_pte(pte, vma, new, 0);
383321ab 260 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
3fe87967
KS
261 if (PageAnon(new))
262 hugepage_add_anon_rmap(new, vma, pvmw.address);
263 else
264 page_dup_rmap(new, true);
383321ab
AK
265 } else
266#endif
267 {
268 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
04e62a29 269
383321ab
AK
270 if (PageAnon(new))
271 page_add_anon_rmap(new, vma, pvmw.address, false);
272 else
273 page_add_file_rmap(new, false);
274 }
3fe87967
KS
275 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
276 mlock_vma_page(new);
277
e125fe40
KS
278 if (PageTransHuge(page) && PageMlocked(page))
279 clear_page_mlock(page);
280
3fe87967
KS
281 /* No need to invalidate - it was non-present before */
282 update_mmu_cache(vma, pvmw.address, pvmw.pte);
283 }
51afb12b 284
e4b82222 285 return true;
0697212a
CL
286}
287
04e62a29
CL
288/*
289 * Get rid of all migration entries and replace them by
290 * references to the indicated page.
291 */
e388466d 292void remove_migration_ptes(struct page *old, struct page *new, bool locked)
04e62a29 293{
051ac83a
JK
294 struct rmap_walk_control rwc = {
295 .rmap_one = remove_migration_pte,
296 .arg = old,
297 };
298
e388466d
KS
299 if (locked)
300 rmap_walk_locked(new, &rwc);
301 else
302 rmap_walk(new, &rwc);
04e62a29
CL
303}
304
0697212a
CL
305/*
306 * Something used the pte of a page under migration. We need to
307 * get to the page and wait until migration is finished.
308 * When we return from this function the fault will be retried.
0697212a 309 */
e66f17ff 310void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
30dad309 311 spinlock_t *ptl)
0697212a 312{
30dad309 313 pte_t pte;
0697212a
CL
314 swp_entry_t entry;
315 struct page *page;
316
30dad309 317 spin_lock(ptl);
0697212a
CL
318 pte = *ptep;
319 if (!is_swap_pte(pte))
320 goto out;
321
322 entry = pte_to_swp_entry(pte);
323 if (!is_migration_entry(entry))
324 goto out;
325
326 page = migration_entry_to_page(entry);
327
e286781d 328 /*
89eb946a 329 * Once page cache replacement of page migration started, page_count
9a1ea439
HD
330 * is zero; but we must not call put_and_wait_on_page_locked() without
331 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
e286781d
NP
332 */
333 if (!get_page_unless_zero(page))
334 goto out;
0697212a 335 pte_unmap_unlock(ptep, ptl);
9a1ea439 336 put_and_wait_on_page_locked(page);
0697212a
CL
337 return;
338out:
339 pte_unmap_unlock(ptep, ptl);
340}
341
30dad309
NH
342void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
343 unsigned long address)
344{
345 spinlock_t *ptl = pte_lockptr(mm, pmd);
346 pte_t *ptep = pte_offset_map(pmd, address);
347 __migration_entry_wait(mm, ptep, ptl);
348}
349
cb900f41
KS
350void migration_entry_wait_huge(struct vm_area_struct *vma,
351 struct mm_struct *mm, pte_t *pte)
30dad309 352{
cb900f41 353 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
30dad309
NH
354 __migration_entry_wait(mm, pte, ptl);
355}
356
616b8371
ZY
357#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
358void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
359{
360 spinlock_t *ptl;
361 struct page *page;
362
363 ptl = pmd_lock(mm, pmd);
364 if (!is_pmd_migration_entry(*pmd))
365 goto unlock;
366 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
367 if (!get_page_unless_zero(page))
368 goto unlock;
369 spin_unlock(ptl);
9a1ea439 370 put_and_wait_on_page_locked(page);
616b8371
ZY
371 return;
372unlock:
373 spin_unlock(ptl);
374}
375#endif
376
b969c4ab
MG
377#ifdef CONFIG_BLOCK
378/* Returns true if all buffers are successfully locked */
a6bc32b8
MG
379static bool buffer_migrate_lock_buffers(struct buffer_head *head,
380 enum migrate_mode mode)
b969c4ab
MG
381{
382 struct buffer_head *bh = head;
383
384 /* Simple case, sync compaction */
a6bc32b8 385 if (mode != MIGRATE_ASYNC) {
b969c4ab
MG
386 do {
387 get_bh(bh);
388 lock_buffer(bh);
389 bh = bh->b_this_page;
390
391 } while (bh != head);
392
393 return true;
394 }
395
396 /* async case, we cannot block on lock_buffer so use trylock_buffer */
397 do {
398 get_bh(bh);
399 if (!trylock_buffer(bh)) {
400 /*
401 * We failed to lock the buffer and cannot stall in
402 * async migration. Release the taken locks
403 */
404 struct buffer_head *failed_bh = bh;
405 put_bh(failed_bh);
406 bh = head;
407 while (bh != failed_bh) {
408 unlock_buffer(bh);
409 put_bh(bh);
410 bh = bh->b_this_page;
411 }
412 return false;
413 }
414
415 bh = bh->b_this_page;
416 } while (bh != head);
417 return true;
418}
419#else
420static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
a6bc32b8 421 enum migrate_mode mode)
b969c4ab
MG
422{
423 return true;
424}
425#endif /* CONFIG_BLOCK */
426
b20a3503 427/*
c3fcf8a5 428 * Replace the page in the mapping.
5b5c7120
CL
429 *
430 * The number of remaining references must be:
431 * 1 for anonymous pages without a mapping
432 * 2 for pages with a mapping
266cf658 433 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
b20a3503 434 */
36bc08cc 435int migrate_page_move_mapping(struct address_space *mapping,
b969c4ab 436 struct page *newpage, struct page *page,
8e321fef
BL
437 struct buffer_head *head, enum migrate_mode mode,
438 int extra_count)
b20a3503 439{
89eb946a 440 XA_STATE(xas, &mapping->i_pages, page_index(page));
42cb14b1
HD
441 struct zone *oldzone, *newzone;
442 int dirty;
8e321fef 443 int expected_count = 1 + extra_count;
b20a3503 444
8763cb45 445 /*
df6ad698
JG
446 * Device public or private pages have an extra refcount as they are
447 * ZONE_DEVICE pages.
8763cb45 448 */
df6ad698
JG
449 expected_count += is_device_private_page(page);
450 expected_count += is_device_public_page(page);
8763cb45 451
6c5240ae 452 if (!mapping) {
0e8c7d0f 453 /* Anonymous page without mapping */
8e321fef 454 if (page_count(page) != expected_count)
6c5240ae 455 return -EAGAIN;
cf4b769a
HD
456
457 /* No turning back from here */
cf4b769a
HD
458 newpage->index = page->index;
459 newpage->mapping = page->mapping;
460 if (PageSwapBacked(page))
fa9949da 461 __SetPageSwapBacked(newpage);
cf4b769a 462
78bd5209 463 return MIGRATEPAGE_SUCCESS;
6c5240ae
CL
464 }
465
42cb14b1
HD
466 oldzone = page_zone(page);
467 newzone = page_zone(newpage);
468
89eb946a 469 xas_lock_irq(&xas);
b20a3503 470
e71769ae 471 expected_count += hpage_nr_pages(page) + page_has_private(page);
89eb946a
MW
472 if (page_count(page) != expected_count || xas_load(&xas) != page) {
473 xas_unlock_irq(&xas);
e23ca00b 474 return -EAGAIN;
b20a3503
CL
475 }
476
fe896d18 477 if (!page_ref_freeze(page, expected_count)) {
89eb946a 478 xas_unlock_irq(&xas);
e286781d
NP
479 return -EAGAIN;
480 }
481
b969c4ab
MG
482 /*
483 * In the async migration case of moving a page with buffers, lock the
484 * buffers using trylock before the mapping is moved. If the mapping
485 * was moved, we later failed to lock the buffers and could not move
486 * the mapping back due to an elevated page count, we would have to
487 * block waiting on other references to be dropped.
488 */
a6bc32b8
MG
489 if (mode == MIGRATE_ASYNC && head &&
490 !buffer_migrate_lock_buffers(head, mode)) {
fe896d18 491 page_ref_unfreeze(page, expected_count);
89eb946a 492 xas_unlock_irq(&xas);
b969c4ab
MG
493 return -EAGAIN;
494 }
495
b20a3503 496 /*
cf4b769a
HD
497 * Now we know that no one else is looking at the page:
498 * no turning back from here.
b20a3503 499 */
cf4b769a
HD
500 newpage->index = page->index;
501 newpage->mapping = page->mapping;
e71769ae 502 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
6326fec1
NP
503 if (PageSwapBacked(page)) {
504 __SetPageSwapBacked(newpage);
505 if (PageSwapCache(page)) {
506 SetPageSwapCache(newpage);
507 set_page_private(newpage, page_private(page));
508 }
509 } else {
510 VM_BUG_ON_PAGE(PageSwapCache(page), page);
b20a3503
CL
511 }
512
42cb14b1
HD
513 /* Move dirty while page refs frozen and newpage not yet exposed */
514 dirty = PageDirty(page);
515 if (dirty) {
516 ClearPageDirty(page);
517 SetPageDirty(newpage);
518 }
519
89eb946a 520 xas_store(&xas, newpage);
e71769ae
NH
521 if (PageTransHuge(page)) {
522 int i;
e71769ae 523
013567be 524 for (i = 1; i < HPAGE_PMD_NR; i++) {
89eb946a
MW
525 xas_next(&xas);
526 xas_store(&xas, newpage + i);
e71769ae 527 }
e71769ae 528 }
7cf9c2c7
NP
529
530 /*
937a94c9
JG
531 * Drop cache reference from old page by unfreezing
532 * to one less reference.
7cf9c2c7
NP
533 * We know this isn't the last reference.
534 */
e71769ae 535 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
7cf9c2c7 536
89eb946a 537 xas_unlock(&xas);
42cb14b1
HD
538 /* Leave irq disabled to prevent preemption while updating stats */
539
0e8c7d0f
CL
540 /*
541 * If moved to a different zone then also account
542 * the page for that zone. Other VM counters will be
543 * taken care of when we establish references to the
544 * new page and drop references to the old page.
545 *
546 * Note that anonymous pages are accounted for
4b9d0fab 547 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
0e8c7d0f
CL
548 * are mapped to swap space.
549 */
42cb14b1 550 if (newzone != oldzone) {
11fb9989
MG
551 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
552 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
42cb14b1 553 if (PageSwapBacked(page) && !PageSwapCache(page)) {
11fb9989
MG
554 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
555 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
42cb14b1
HD
556 }
557 if (dirty && mapping_cap_account_dirty(mapping)) {
11fb9989 558 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 559 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
11fb9989 560 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 561 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
42cb14b1 562 }
4b02108a 563 }
42cb14b1 564 local_irq_enable();
b20a3503 565
78bd5209 566 return MIGRATEPAGE_SUCCESS;
b20a3503 567}
1118dce7 568EXPORT_SYMBOL(migrate_page_move_mapping);
b20a3503 569
290408d4
NH
570/*
571 * The expected number of remaining references is the same as that
572 * of migrate_page_move_mapping().
573 */
574int migrate_huge_page_move_mapping(struct address_space *mapping,
575 struct page *newpage, struct page *page)
576{
89eb946a 577 XA_STATE(xas, &mapping->i_pages, page_index(page));
290408d4 578 int expected_count;
290408d4 579
89eb946a 580 xas_lock_irq(&xas);
290408d4 581 expected_count = 2 + page_has_private(page);
89eb946a
MW
582 if (page_count(page) != expected_count || xas_load(&xas) != page) {
583 xas_unlock_irq(&xas);
290408d4
NH
584 return -EAGAIN;
585 }
586
fe896d18 587 if (!page_ref_freeze(page, expected_count)) {
89eb946a 588 xas_unlock_irq(&xas);
290408d4
NH
589 return -EAGAIN;
590 }
591
cf4b769a
HD
592 newpage->index = page->index;
593 newpage->mapping = page->mapping;
6a93ca8f 594
290408d4
NH
595 get_page(newpage);
596
89eb946a 597 xas_store(&xas, newpage);
290408d4 598
fe896d18 599 page_ref_unfreeze(page, expected_count - 1);
290408d4 600
89eb946a 601 xas_unlock_irq(&xas);
6a93ca8f 602
78bd5209 603 return MIGRATEPAGE_SUCCESS;
290408d4
NH
604}
605
30b0a105
DH
606/*
607 * Gigantic pages are so large that we do not guarantee that page++ pointer
608 * arithmetic will work across the entire page. We need something more
609 * specialized.
610 */
611static void __copy_gigantic_page(struct page *dst, struct page *src,
612 int nr_pages)
613{
614 int i;
615 struct page *dst_base = dst;
616 struct page *src_base = src;
617
618 for (i = 0; i < nr_pages; ) {
619 cond_resched();
620 copy_highpage(dst, src);
621
622 i++;
623 dst = mem_map_next(dst, dst_base, i);
624 src = mem_map_next(src, src_base, i);
625 }
626}
627
628static void copy_huge_page(struct page *dst, struct page *src)
629{
630 int i;
631 int nr_pages;
632
633 if (PageHuge(src)) {
634 /* hugetlbfs page */
635 struct hstate *h = page_hstate(src);
636 nr_pages = pages_per_huge_page(h);
637
638 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
639 __copy_gigantic_page(dst, src, nr_pages);
640 return;
641 }
642 } else {
643 /* thp page */
644 BUG_ON(!PageTransHuge(src));
645 nr_pages = hpage_nr_pages(src);
646 }
647
648 for (i = 0; i < nr_pages; i++) {
649 cond_resched();
650 copy_highpage(dst + i, src + i);
651 }
652}
653
b20a3503
CL
654/*
655 * Copy the page to its new location
656 */
2916ecc0 657void migrate_page_states(struct page *newpage, struct page *page)
b20a3503 658{
7851a45c
RR
659 int cpupid;
660
b20a3503
CL
661 if (PageError(page))
662 SetPageError(newpage);
663 if (PageReferenced(page))
664 SetPageReferenced(newpage);
665 if (PageUptodate(page))
666 SetPageUptodate(newpage);
894bc310 667 if (TestClearPageActive(page)) {
309381fe 668 VM_BUG_ON_PAGE(PageUnevictable(page), page);
b20a3503 669 SetPageActive(newpage);
418b27ef
LS
670 } else if (TestClearPageUnevictable(page))
671 SetPageUnevictable(newpage);
1899ad18
JW
672 if (PageWorkingset(page))
673 SetPageWorkingset(newpage);
b20a3503
CL
674 if (PageChecked(page))
675 SetPageChecked(newpage);
676 if (PageMappedToDisk(page))
677 SetPageMappedToDisk(newpage);
678
42cb14b1
HD
679 /* Move dirty on pages not done by migrate_page_move_mapping() */
680 if (PageDirty(page))
681 SetPageDirty(newpage);
b20a3503 682
33c3fc71
VD
683 if (page_is_young(page))
684 set_page_young(newpage);
685 if (page_is_idle(page))
686 set_page_idle(newpage);
687
7851a45c
RR
688 /*
689 * Copy NUMA information to the new page, to prevent over-eager
690 * future migrations of this same page.
691 */
692 cpupid = page_cpupid_xchg_last(page, -1);
693 page_cpupid_xchg_last(newpage, cpupid);
694
e9995ef9 695 ksm_migrate_page(newpage, page);
c8d6553b
HD
696 /*
697 * Please do not reorder this without considering how mm/ksm.c's
698 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
699 */
b3b3a99c
NH
700 if (PageSwapCache(page))
701 ClearPageSwapCache(page);
b20a3503
CL
702 ClearPagePrivate(page);
703 set_page_private(page, 0);
b20a3503
CL
704
705 /*
706 * If any waiters have accumulated on the new page then
707 * wake them up.
708 */
709 if (PageWriteback(newpage))
710 end_page_writeback(newpage);
d435edca
VB
711
712 copy_page_owner(page, newpage);
74485cf2
JW
713
714 mem_cgroup_migrate(page, newpage);
b20a3503 715}
2916ecc0
JG
716EXPORT_SYMBOL(migrate_page_states);
717
718void migrate_page_copy(struct page *newpage, struct page *page)
719{
720 if (PageHuge(page) || PageTransHuge(page))
721 copy_huge_page(newpage, page);
722 else
723 copy_highpage(newpage, page);
724
725 migrate_page_states(newpage, page);
726}
1118dce7 727EXPORT_SYMBOL(migrate_page_copy);
b20a3503 728
1d8b85cc
CL
729/************************************************************
730 * Migration functions
731 ***********************************************************/
732
b20a3503 733/*
bda807d4 734 * Common logic to directly migrate a single LRU page suitable for
266cf658 735 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
736 *
737 * Pages are locked upon entry and exit.
738 */
2d1db3b1 739int migrate_page(struct address_space *mapping,
a6bc32b8
MG
740 struct page *newpage, struct page *page,
741 enum migrate_mode mode)
b20a3503
CL
742{
743 int rc;
744
745 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
746
8e321fef 747 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
b20a3503 748
78bd5209 749 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
750 return rc;
751
2916ecc0
JG
752 if (mode != MIGRATE_SYNC_NO_COPY)
753 migrate_page_copy(newpage, page);
754 else
755 migrate_page_states(newpage, page);
78bd5209 756 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
757}
758EXPORT_SYMBOL(migrate_page);
759
9361401e 760#ifdef CONFIG_BLOCK
1d8b85cc
CL
761/*
762 * Migration function for pages with buffers. This function can only be used
763 * if the underlying filesystem guarantees that no other references to "page"
764 * exist.
765 */
2d1db3b1 766int buffer_migrate_page(struct address_space *mapping,
a6bc32b8 767 struct page *newpage, struct page *page, enum migrate_mode mode)
1d8b85cc 768{
1d8b85cc
CL
769 struct buffer_head *bh, *head;
770 int rc;
771
1d8b85cc 772 if (!page_has_buffers(page))
a6bc32b8 773 return migrate_page(mapping, newpage, page, mode);
1d8b85cc
CL
774
775 head = page_buffers(page);
776
8e321fef 777 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
1d8b85cc 778
78bd5209 779 if (rc != MIGRATEPAGE_SUCCESS)
1d8b85cc
CL
780 return rc;
781
b969c4ab
MG
782 /*
783 * In the async case, migrate_page_move_mapping locked the buffers
784 * with an IRQ-safe spinlock held. In the sync case, the buffers
785 * need to be locked now
786 */
a6bc32b8
MG
787 if (mode != MIGRATE_ASYNC)
788 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
1d8b85cc
CL
789
790 ClearPagePrivate(page);
791 set_page_private(newpage, page_private(page));
792 set_page_private(page, 0);
793 put_page(page);
794 get_page(newpage);
795
796 bh = head;
797 do {
798 set_bh_page(bh, newpage, bh_offset(bh));
799 bh = bh->b_this_page;
800
801 } while (bh != head);
802
803 SetPagePrivate(newpage);
804
2916ecc0
JG
805 if (mode != MIGRATE_SYNC_NO_COPY)
806 migrate_page_copy(newpage, page);
807 else
808 migrate_page_states(newpage, page);
1d8b85cc
CL
809
810 bh = head;
811 do {
812 unlock_buffer(bh);
2916ecc0 813 put_bh(bh);
1d8b85cc
CL
814 bh = bh->b_this_page;
815
816 } while (bh != head);
817
78bd5209 818 return MIGRATEPAGE_SUCCESS;
1d8b85cc
CL
819}
820EXPORT_SYMBOL(buffer_migrate_page);
9361401e 821#endif
1d8b85cc 822
04e62a29
CL
823/*
824 * Writeback a page to clean the dirty state
825 */
826static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 827{
04e62a29
CL
828 struct writeback_control wbc = {
829 .sync_mode = WB_SYNC_NONE,
830 .nr_to_write = 1,
831 .range_start = 0,
832 .range_end = LLONG_MAX,
04e62a29
CL
833 .for_reclaim = 1
834 };
835 int rc;
836
837 if (!mapping->a_ops->writepage)
838 /* No write method for the address space */
839 return -EINVAL;
840
841 if (!clear_page_dirty_for_io(page))
842 /* Someone else already triggered a write */
843 return -EAGAIN;
844
8351a6e4 845 /*
04e62a29
CL
846 * A dirty page may imply that the underlying filesystem has
847 * the page on some queue. So the page must be clean for
848 * migration. Writeout may mean we loose the lock and the
849 * page state is no longer what we checked for earlier.
850 * At this point we know that the migration attempt cannot
851 * be successful.
8351a6e4 852 */
e388466d 853 remove_migration_ptes(page, page, false);
8351a6e4 854
04e62a29 855 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 856
04e62a29
CL
857 if (rc != AOP_WRITEPAGE_ACTIVATE)
858 /* unlocked. Relock */
859 lock_page(page);
860
bda8550d 861 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
862}
863
864/*
865 * Default handling if a filesystem does not provide a migration function.
866 */
867static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 868 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 869{
b969c4ab 870 if (PageDirty(page)) {
a6bc32b8 871 /* Only writeback pages in full synchronous migration */
2916ecc0
JG
872 switch (mode) {
873 case MIGRATE_SYNC:
874 case MIGRATE_SYNC_NO_COPY:
875 break;
876 default:
b969c4ab 877 return -EBUSY;
2916ecc0 878 }
04e62a29 879 return writeout(mapping, page);
b969c4ab 880 }
8351a6e4
CL
881
882 /*
883 * Buffers may be managed in a filesystem specific way.
884 * We must have no buffers or drop them.
885 */
266cf658 886 if (page_has_private(page) &&
8351a6e4
CL
887 !try_to_release_page(page, GFP_KERNEL))
888 return -EAGAIN;
889
a6bc32b8 890 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
891}
892
e24f0b8f
CL
893/*
894 * Move a page to a newly allocated page
895 * The page is locked and all ptes have been successfully removed.
896 *
897 * The new page will have replaced the old page if this function
898 * is successful.
894bc310
LS
899 *
900 * Return value:
901 * < 0 - error code
78bd5209 902 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 903 */
3fe2011f 904static int move_to_new_page(struct page *newpage, struct page *page,
5c3f9a67 905 enum migrate_mode mode)
e24f0b8f
CL
906{
907 struct address_space *mapping;
bda807d4
MK
908 int rc = -EAGAIN;
909 bool is_lru = !__PageMovable(page);
e24f0b8f 910
7db7671f
HD
911 VM_BUG_ON_PAGE(!PageLocked(page), page);
912 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
e24f0b8f 913
e24f0b8f 914 mapping = page_mapping(page);
bda807d4
MK
915
916 if (likely(is_lru)) {
917 if (!mapping)
918 rc = migrate_page(mapping, newpage, page, mode);
919 else if (mapping->a_ops->migratepage)
920 /*
921 * Most pages have a mapping and most filesystems
922 * provide a migratepage callback. Anonymous pages
923 * are part of swap space which also has its own
924 * migratepage callback. This is the most common path
925 * for page migration.
926 */
927 rc = mapping->a_ops->migratepage(mapping, newpage,
928 page, mode);
929 else
930 rc = fallback_migrate_page(mapping, newpage,
931 page, mode);
932 } else {
e24f0b8f 933 /*
bda807d4
MK
934 * In case of non-lru page, it could be released after
935 * isolation step. In that case, we shouldn't try migration.
e24f0b8f 936 */
bda807d4
MK
937 VM_BUG_ON_PAGE(!PageIsolated(page), page);
938 if (!PageMovable(page)) {
939 rc = MIGRATEPAGE_SUCCESS;
940 __ClearPageIsolated(page);
941 goto out;
942 }
943
944 rc = mapping->a_ops->migratepage(mapping, newpage,
945 page, mode);
946 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
947 !PageIsolated(page));
948 }
e24f0b8f 949
5c3f9a67
HD
950 /*
951 * When successful, old pagecache page->mapping must be cleared before
952 * page is freed; but stats require that PageAnon be left as PageAnon.
953 */
954 if (rc == MIGRATEPAGE_SUCCESS) {
bda807d4
MK
955 if (__PageMovable(page)) {
956 VM_BUG_ON_PAGE(!PageIsolated(page), page);
957
958 /*
959 * We clear PG_movable under page_lock so any compactor
960 * cannot try to migrate this page.
961 */
962 __ClearPageIsolated(page);
963 }
964
965 /*
966 * Anonymous and movable page->mapping will be cleard by
967 * free_pages_prepare so don't reset it here for keeping
968 * the type to work PageAnon, for example.
969 */
970 if (!PageMappingFlags(page))
5c3f9a67 971 page->mapping = NULL;
3fe2011f 972 }
bda807d4 973out:
e24f0b8f
CL
974 return rc;
975}
976
0dabec93 977static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 978 int force, enum migrate_mode mode)
e24f0b8f 979{
0dabec93 980 int rc = -EAGAIN;
2ebba6b7 981 int page_was_mapped = 0;
3f6c8272 982 struct anon_vma *anon_vma = NULL;
bda807d4 983 bool is_lru = !__PageMovable(page);
95a402c3 984
529ae9aa 985 if (!trylock_page(page)) {
a6bc32b8 986 if (!force || mode == MIGRATE_ASYNC)
0dabec93 987 goto out;
3e7d3449
MG
988
989 /*
990 * It's not safe for direct compaction to call lock_page.
991 * For example, during page readahead pages are added locked
992 * to the LRU. Later, when the IO completes the pages are
993 * marked uptodate and unlocked. However, the queueing
994 * could be merging multiple pages for one bio (e.g.
995 * mpage_readpages). If an allocation happens for the
996 * second or third page, the process can end up locking
997 * the same page twice and deadlocking. Rather than
998 * trying to be clever about what pages can be locked,
999 * avoid the use of lock_page for direct compaction
1000 * altogether.
1001 */
1002 if (current->flags & PF_MEMALLOC)
0dabec93 1003 goto out;
3e7d3449 1004
e24f0b8f
CL
1005 lock_page(page);
1006 }
1007
1008 if (PageWriteback(page)) {
11bc82d6 1009 /*
fed5b64a 1010 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
1011 * necessary to wait for PageWriteback. In the async case,
1012 * the retry loop is too short and in the sync-light case,
1013 * the overhead of stalling is too much
11bc82d6 1014 */
2916ecc0
JG
1015 switch (mode) {
1016 case MIGRATE_SYNC:
1017 case MIGRATE_SYNC_NO_COPY:
1018 break;
1019 default:
11bc82d6 1020 rc = -EBUSY;
0a31bc97 1021 goto out_unlock;
11bc82d6
AA
1022 }
1023 if (!force)
0a31bc97 1024 goto out_unlock;
e24f0b8f
CL
1025 wait_on_page_writeback(page);
1026 }
03f15c86 1027
e24f0b8f 1028 /*
dc386d4d
KH
1029 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1030 * we cannot notice that anon_vma is freed while we migrates a page.
1ce82b69 1031 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 1032 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
1033 * File Caches may use write_page() or lock_page() in migration, then,
1034 * just care Anon page here.
03f15c86
HD
1035 *
1036 * Only page_get_anon_vma() understands the subtleties of
1037 * getting a hold on an anon_vma from outside one of its mms.
1038 * But if we cannot get anon_vma, then we won't need it anyway,
1039 * because that implies that the anon page is no longer mapped
1040 * (and cannot be remapped so long as we hold the page lock).
dc386d4d 1041 */
03f15c86 1042 if (PageAnon(page) && !PageKsm(page))
746b18d4 1043 anon_vma = page_get_anon_vma(page);
62e1c553 1044
7db7671f
HD
1045 /*
1046 * Block others from accessing the new page when we get around to
1047 * establishing additional references. We are usually the only one
1048 * holding a reference to newpage at this point. We used to have a BUG
1049 * here if trylock_page(newpage) fails, but would like to allow for
1050 * cases where there might be a race with the previous use of newpage.
1051 * This is much like races on refcount of oldpage: just don't BUG().
1052 */
1053 if (unlikely(!trylock_page(newpage)))
1054 goto out_unlock;
1055
bda807d4
MK
1056 if (unlikely(!is_lru)) {
1057 rc = move_to_new_page(newpage, page, mode);
1058 goto out_unlock_both;
1059 }
1060
dc386d4d 1061 /*
62e1c553
SL
1062 * Corner case handling:
1063 * 1. When a new swap-cache page is read into, it is added to the LRU
1064 * and treated as swapcache but it has no rmap yet.
1065 * Calling try_to_unmap() against a page->mapping==NULL page will
1066 * trigger a BUG. So handle it here.
1067 * 2. An orphaned page (see truncate_complete_page) might have
1068 * fs-private metadata. The page can be picked up due to memory
1069 * offlining. Everywhere else except page reclaim, the page is
1070 * invisible to the vm, so the page can not be migrated. So try to
1071 * free the metadata, so the page can be freed.
e24f0b8f 1072 */
62e1c553 1073 if (!page->mapping) {
309381fe 1074 VM_BUG_ON_PAGE(PageAnon(page), page);
1ce82b69 1075 if (page_has_private(page)) {
62e1c553 1076 try_to_free_buffers(page);
7db7671f 1077 goto out_unlock_both;
62e1c553 1078 }
7db7671f
HD
1079 } else if (page_mapped(page)) {
1080 /* Establish migration ptes */
03f15c86
HD
1081 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1082 page);
2ebba6b7 1083 try_to_unmap(page,
da1b13cc 1084 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
2ebba6b7
HD
1085 page_was_mapped = 1;
1086 }
dc386d4d 1087
e6a1530d 1088 if (!page_mapped(page))
5c3f9a67 1089 rc = move_to_new_page(newpage, page, mode);
e24f0b8f 1090
5c3f9a67
HD
1091 if (page_was_mapped)
1092 remove_migration_ptes(page,
e388466d 1093 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
3f6c8272 1094
7db7671f
HD
1095out_unlock_both:
1096 unlock_page(newpage);
1097out_unlock:
3f6c8272 1098 /* Drop an anon_vma reference if we took one */
76545066 1099 if (anon_vma)
9e60109f 1100 put_anon_vma(anon_vma);
e24f0b8f 1101 unlock_page(page);
0dabec93 1102out:
c6c919eb
MK
1103 /*
1104 * If migration is successful, decrease refcount of the newpage
1105 * which will not free the page because new page owner increased
1106 * refcounter. As well, if it is LRU page, add the page to LRU
1107 * list in here.
1108 */
1109 if (rc == MIGRATEPAGE_SUCCESS) {
b1123ea6 1110 if (unlikely(__PageMovable(newpage)))
c6c919eb
MK
1111 put_page(newpage);
1112 else
1113 putback_lru_page(newpage);
1114 }
1115
0dabec93
MK
1116 return rc;
1117}
95a402c3 1118
ef2a5153
GU
1119/*
1120 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1121 * around it.
1122 */
815f0ddb
ND
1123#if defined(CONFIG_ARM) && \
1124 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
ef2a5153
GU
1125#define ICE_noinline noinline
1126#else
1127#define ICE_noinline
1128#endif
1129
0dabec93
MK
1130/*
1131 * Obtain the lock on page, remove all ptes and migrate the page
1132 * to the newly allocated page in newpage.
1133 */
ef2a5153
GU
1134static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1135 free_page_t put_new_page,
1136 unsigned long private, struct page *page,
add05cec
NH
1137 int force, enum migrate_mode mode,
1138 enum migrate_reason reason)
0dabec93 1139{
2def7424 1140 int rc = MIGRATEPAGE_SUCCESS;
2def7424 1141 struct page *newpage;
0dabec93 1142
94723aaf
MH
1143 if (!thp_migration_supported() && PageTransHuge(page))
1144 return -ENOMEM;
1145
666feb21 1146 newpage = get_new_page(page, private);
0dabec93
MK
1147 if (!newpage)
1148 return -ENOMEM;
1149
1150 if (page_count(page) == 1) {
1151 /* page was freed from under us. So we are done. */
c6c919eb
MK
1152 ClearPageActive(page);
1153 ClearPageUnevictable(page);
bda807d4
MK
1154 if (unlikely(__PageMovable(page))) {
1155 lock_page(page);
1156 if (!PageMovable(page))
1157 __ClearPageIsolated(page);
1158 unlock_page(page);
1159 }
c6c919eb
MK
1160 if (put_new_page)
1161 put_new_page(newpage, private);
1162 else
1163 put_page(newpage);
0dabec93
MK
1164 goto out;
1165 }
1166
9c620e2b 1167 rc = __unmap_and_move(page, newpage, force, mode);
c6c919eb 1168 if (rc == MIGRATEPAGE_SUCCESS)
7cd12b4a 1169 set_page_owner_migrate_reason(newpage, reason);
bf6bddf1 1170
0dabec93 1171out:
e24f0b8f 1172 if (rc != -EAGAIN) {
0dabec93
MK
1173 /*
1174 * A page that has been migrated has all references
1175 * removed and will be freed. A page that has not been
1176 * migrated will have kepts its references and be
1177 * restored.
1178 */
1179 list_del(&page->lru);
6afcf8ef
ML
1180
1181 /*
1182 * Compaction can migrate also non-LRU pages which are
1183 * not accounted to NR_ISOLATED_*. They can be recognized
1184 * as __PageMovable
1185 */
1186 if (likely(!__PageMovable(page)))
e8db67eb
NH
1187 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1188 page_is_file_cache(page), -hpage_nr_pages(page));
c6c919eb
MK
1189 }
1190
1191 /*
1192 * If migration is successful, releases reference grabbed during
1193 * isolation. Otherwise, restore the page to right list unless
1194 * we want to retry.
1195 */
1196 if (rc == MIGRATEPAGE_SUCCESS) {
1197 put_page(page);
1198 if (reason == MR_MEMORY_FAILURE) {
d7e69488 1199 /*
c6c919eb
MK
1200 * Set PG_HWPoison on just freed page
1201 * intentionally. Although it's rather weird,
1202 * it's how HWPoison flag works at the moment.
d7e69488 1203 */
d4ae9916 1204 if (set_hwpoison_free_buddy_page(page))
da1b13cc 1205 num_poisoned_pages_inc();
c6c919eb
MK
1206 }
1207 } else {
bda807d4
MK
1208 if (rc != -EAGAIN) {
1209 if (likely(!__PageMovable(page))) {
1210 putback_lru_page(page);
1211 goto put_new;
1212 }
1213
1214 lock_page(page);
1215 if (PageMovable(page))
1216 putback_movable_page(page);
1217 else
1218 __ClearPageIsolated(page);
1219 unlock_page(page);
1220 put_page(page);
1221 }
1222put_new:
c6c919eb
MK
1223 if (put_new_page)
1224 put_new_page(newpage, private);
1225 else
1226 put_page(newpage);
e24f0b8f 1227 }
68711a74 1228
e24f0b8f
CL
1229 return rc;
1230}
1231
290408d4
NH
1232/*
1233 * Counterpart of unmap_and_move_page() for hugepage migration.
1234 *
1235 * This function doesn't wait the completion of hugepage I/O
1236 * because there is no race between I/O and migration for hugepage.
1237 * Note that currently hugepage I/O occurs only in direct I/O
1238 * where no lock is held and PG_writeback is irrelevant,
1239 * and writeback status of all subpages are counted in the reference
1240 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1241 * under direct I/O, the reference of the head page is 512 and a bit more.)
1242 * This means that when we try to migrate hugepage whose subpages are
1243 * doing direct I/O, some references remain after try_to_unmap() and
1244 * hugepage migration fails without data corruption.
1245 *
1246 * There is also no race when direct I/O is issued on the page under migration,
1247 * because then pte is replaced with migration swap entry and direct I/O code
1248 * will wait in the page fault for migration to complete.
1249 */
1250static int unmap_and_move_huge_page(new_page_t get_new_page,
68711a74
DR
1251 free_page_t put_new_page, unsigned long private,
1252 struct page *hpage, int force,
7cd12b4a 1253 enum migrate_mode mode, int reason)
290408d4 1254{
2def7424 1255 int rc = -EAGAIN;
2ebba6b7 1256 int page_was_mapped = 0;
32665f2b 1257 struct page *new_hpage;
290408d4
NH
1258 struct anon_vma *anon_vma = NULL;
1259
83467efb
NH
1260 /*
1261 * Movability of hugepages depends on architectures and hugepage size.
1262 * This check is necessary because some callers of hugepage migration
1263 * like soft offline and memory hotremove don't walk through page
1264 * tables or check whether the hugepage is pmd-based or not before
1265 * kicking migration.
1266 */
100873d7 1267 if (!hugepage_migration_supported(page_hstate(hpage))) {
32665f2b 1268 putback_active_hugepage(hpage);
83467efb 1269 return -ENOSYS;
32665f2b 1270 }
83467efb 1271
666feb21 1272 new_hpage = get_new_page(hpage, private);
290408d4
NH
1273 if (!new_hpage)
1274 return -ENOMEM;
1275
290408d4 1276 if (!trylock_page(hpage)) {
2916ecc0 1277 if (!force)
290408d4 1278 goto out;
2916ecc0
JG
1279 switch (mode) {
1280 case MIGRATE_SYNC:
1281 case MIGRATE_SYNC_NO_COPY:
1282 break;
1283 default:
1284 goto out;
1285 }
290408d4
NH
1286 lock_page(hpage);
1287 }
1288
746b18d4
PZ
1289 if (PageAnon(hpage))
1290 anon_vma = page_get_anon_vma(hpage);
290408d4 1291
7db7671f
HD
1292 if (unlikely(!trylock_page(new_hpage)))
1293 goto put_anon;
1294
2ebba6b7
HD
1295 if (page_mapped(hpage)) {
1296 try_to_unmap(hpage,
1297 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1298 page_was_mapped = 1;
1299 }
290408d4
NH
1300
1301 if (!page_mapped(hpage))
5c3f9a67 1302 rc = move_to_new_page(new_hpage, hpage, mode);
290408d4 1303
5c3f9a67
HD
1304 if (page_was_mapped)
1305 remove_migration_ptes(hpage,
e388466d 1306 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
290408d4 1307
7db7671f
HD
1308 unlock_page(new_hpage);
1309
1310put_anon:
fd4a4663 1311 if (anon_vma)
9e60109f 1312 put_anon_vma(anon_vma);
8e6ac7fa 1313
2def7424 1314 if (rc == MIGRATEPAGE_SUCCESS) {
ab5ac90a 1315 move_hugetlb_state(hpage, new_hpage, reason);
2def7424
HD
1316 put_new_page = NULL;
1317 }
8e6ac7fa 1318
290408d4 1319 unlock_page(hpage);
09761333 1320out:
b8ec1cee
NH
1321 if (rc != -EAGAIN)
1322 putback_active_hugepage(hpage);
68711a74
DR
1323
1324 /*
1325 * If migration was not successful and there's a freeing callback, use
1326 * it. Otherwise, put_page() will drop the reference grabbed during
1327 * isolation.
1328 */
2def7424 1329 if (put_new_page)
68711a74
DR
1330 put_new_page(new_hpage, private);
1331 else
3aaa76e1 1332 putback_active_hugepage(new_hpage);
68711a74 1333
290408d4
NH
1334 return rc;
1335}
1336
b20a3503 1337/*
c73e5c9c
SB
1338 * migrate_pages - migrate the pages specified in a list, to the free pages
1339 * supplied as the target for the page migration
b20a3503 1340 *
c73e5c9c
SB
1341 * @from: The list of pages to be migrated.
1342 * @get_new_page: The function used to allocate free pages to be used
1343 * as the target of the page migration.
68711a74
DR
1344 * @put_new_page: The function used to free target pages if migration
1345 * fails, or NULL if no special handling is necessary.
c73e5c9c
SB
1346 * @private: Private data to be passed on to get_new_page()
1347 * @mode: The migration mode that specifies the constraints for
1348 * page migration, if any.
1349 * @reason: The reason for page migration.
b20a3503 1350 *
c73e5c9c
SB
1351 * The function returns after 10 attempts or if no pages are movable any more
1352 * because the list has become empty or no retryable pages exist any more.
14e0f9bc 1353 * The caller should call putback_movable_pages() to return pages to the LRU
28bd6578 1354 * or free list only if ret != 0.
b20a3503 1355 *
c73e5c9c 1356 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1357 */
9c620e2b 1358int migrate_pages(struct list_head *from, new_page_t get_new_page,
68711a74
DR
1359 free_page_t put_new_page, unsigned long private,
1360 enum migrate_mode mode, int reason)
b20a3503 1361{
e24f0b8f 1362 int retry = 1;
b20a3503 1363 int nr_failed = 0;
5647bc29 1364 int nr_succeeded = 0;
b20a3503
CL
1365 int pass = 0;
1366 struct page *page;
1367 struct page *page2;
1368 int swapwrite = current->flags & PF_SWAPWRITE;
1369 int rc;
1370
1371 if (!swapwrite)
1372 current->flags |= PF_SWAPWRITE;
1373
e24f0b8f
CL
1374 for(pass = 0; pass < 10 && retry; pass++) {
1375 retry = 0;
b20a3503 1376
e24f0b8f 1377 list_for_each_entry_safe(page, page2, from, lru) {
94723aaf 1378retry:
e24f0b8f 1379 cond_resched();
2d1db3b1 1380
31caf665
NH
1381 if (PageHuge(page))
1382 rc = unmap_and_move_huge_page(get_new_page,
68711a74 1383 put_new_page, private, page,
7cd12b4a 1384 pass > 2, mode, reason);
31caf665 1385 else
68711a74 1386 rc = unmap_and_move(get_new_page, put_new_page,
add05cec
NH
1387 private, page, pass > 2, mode,
1388 reason);
2d1db3b1 1389
e24f0b8f 1390 switch(rc) {
95a402c3 1391 case -ENOMEM:
94723aaf
MH
1392 /*
1393 * THP migration might be unsupported or the
1394 * allocation could've failed so we should
1395 * retry on the same page with the THP split
1396 * to base pages.
1397 *
1398 * Head page is retried immediately and tail
1399 * pages are added to the tail of the list so
1400 * we encounter them after the rest of the list
1401 * is processed.
1402 */
e6112fc3 1403 if (PageTransHuge(page) && !PageHuge(page)) {
94723aaf
MH
1404 lock_page(page);
1405 rc = split_huge_page_to_list(page, from);
1406 unlock_page(page);
1407 if (!rc) {
1408 list_safe_reset_next(page, page2, lru);
1409 goto retry;
1410 }
1411 }
dfef2ef4 1412 nr_failed++;
95a402c3 1413 goto out;
e24f0b8f 1414 case -EAGAIN:
2d1db3b1 1415 retry++;
e24f0b8f 1416 break;
78bd5209 1417 case MIGRATEPAGE_SUCCESS:
5647bc29 1418 nr_succeeded++;
e24f0b8f
CL
1419 break;
1420 default:
354a3363
NH
1421 /*
1422 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1423 * unlike -EAGAIN case, the failed page is
1424 * removed from migration page list and not
1425 * retried in the next outer loop.
1426 */
2d1db3b1 1427 nr_failed++;
e24f0b8f 1428 break;
2d1db3b1 1429 }
b20a3503
CL
1430 }
1431 }
f2f81fb2
VB
1432 nr_failed += retry;
1433 rc = nr_failed;
95a402c3 1434out:
5647bc29
MG
1435 if (nr_succeeded)
1436 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1437 if (nr_failed)
1438 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1439 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1440
b20a3503
CL
1441 if (!swapwrite)
1442 current->flags &= ~PF_SWAPWRITE;
1443
78bd5209 1444 return rc;
b20a3503 1445}
95a402c3 1446
742755a1 1447#ifdef CONFIG_NUMA
742755a1 1448
a49bd4d7 1449static int store_status(int __user *status, int start, int value, int nr)
742755a1 1450{
a49bd4d7
MH
1451 while (nr-- > 0) {
1452 if (put_user(value, status + start))
1453 return -EFAULT;
1454 start++;
1455 }
1456
1457 return 0;
1458}
1459
1460static int do_move_pages_to_node(struct mm_struct *mm,
1461 struct list_head *pagelist, int node)
1462{
1463 int err;
1464
1465 if (list_empty(pagelist))
1466 return 0;
1467
1468 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1469 MIGRATE_SYNC, MR_SYSCALL);
1470 if (err)
1471 putback_movable_pages(pagelist);
1472 return err;
742755a1
CL
1473}
1474
1475/*
a49bd4d7
MH
1476 * Resolves the given address to a struct page, isolates it from the LRU and
1477 * puts it to the given pagelist.
1478 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1479 * queued or the page doesn't need to be migrated because it is already on
1480 * the target node
742755a1 1481 */
a49bd4d7
MH
1482static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1483 int node, struct list_head *pagelist, bool migrate_all)
742755a1 1484{
a49bd4d7
MH
1485 struct vm_area_struct *vma;
1486 struct page *page;
1487 unsigned int follflags;
742755a1 1488 int err;
742755a1
CL
1489
1490 down_read(&mm->mmap_sem);
a49bd4d7
MH
1491 err = -EFAULT;
1492 vma = find_vma(mm, addr);
1493 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1494 goto out;
742755a1 1495
a49bd4d7
MH
1496 /* FOLL_DUMP to ignore special (like zero) pages */
1497 follflags = FOLL_GET | FOLL_DUMP;
a49bd4d7 1498 page = follow_page(vma, addr, follflags);
89f5b7da 1499
a49bd4d7
MH
1500 err = PTR_ERR(page);
1501 if (IS_ERR(page))
1502 goto out;
89f5b7da 1503
a49bd4d7
MH
1504 err = -ENOENT;
1505 if (!page)
1506 goto out;
742755a1 1507
a49bd4d7
MH
1508 err = 0;
1509 if (page_to_nid(page) == node)
1510 goto out_putpage;
742755a1 1511
a49bd4d7
MH
1512 err = -EACCES;
1513 if (page_mapcount(page) > 1 && !migrate_all)
1514 goto out_putpage;
742755a1 1515
a49bd4d7
MH
1516 if (PageHuge(page)) {
1517 if (PageHead(page)) {
1518 isolate_huge_page(page, pagelist);
1519 err = 0;
e632a938 1520 }
a49bd4d7
MH
1521 } else {
1522 struct page *head;
e632a938 1523
e8db67eb
NH
1524 head = compound_head(page);
1525 err = isolate_lru_page(head);
cf608ac1 1526 if (err)
a49bd4d7 1527 goto out_putpage;
742755a1 1528
a49bd4d7
MH
1529 err = 0;
1530 list_add_tail(&head->lru, pagelist);
1531 mod_node_page_state(page_pgdat(head),
1532 NR_ISOLATED_ANON + page_is_file_cache(head),
1533 hpage_nr_pages(head));
1534 }
1535out_putpage:
1536 /*
1537 * Either remove the duplicate refcount from
1538 * isolate_lru_page() or drop the page ref if it was
1539 * not isolated.
1540 */
1541 put_page(page);
1542out:
742755a1
CL
1543 up_read(&mm->mmap_sem);
1544 return err;
1545}
1546
5e9a0f02
BG
1547/*
1548 * Migrate an array of page address onto an array of nodes and fill
1549 * the corresponding array of status.
1550 */
3268c63e 1551static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1552 unsigned long nr_pages,
1553 const void __user * __user *pages,
1554 const int __user *nodes,
1555 int __user *status, int flags)
1556{
a49bd4d7
MH
1557 int current_node = NUMA_NO_NODE;
1558 LIST_HEAD(pagelist);
1559 int start, i;
1560 int err = 0, err1;
35282a2d
BG
1561
1562 migrate_prep();
1563
a49bd4d7
MH
1564 for (i = start = 0; i < nr_pages; i++) {
1565 const void __user *p;
1566 unsigned long addr;
1567 int node;
3140a227 1568
a49bd4d7
MH
1569 err = -EFAULT;
1570 if (get_user(p, pages + i))
1571 goto out_flush;
1572 if (get_user(node, nodes + i))
1573 goto out_flush;
1574 addr = (unsigned long)p;
1575
1576 err = -ENODEV;
1577 if (node < 0 || node >= MAX_NUMNODES)
1578 goto out_flush;
1579 if (!node_state(node, N_MEMORY))
1580 goto out_flush;
5e9a0f02 1581
a49bd4d7
MH
1582 err = -EACCES;
1583 if (!node_isset(node, task_nodes))
1584 goto out_flush;
1585
1586 if (current_node == NUMA_NO_NODE) {
1587 current_node = node;
1588 start = i;
1589 } else if (node != current_node) {
1590 err = do_move_pages_to_node(mm, &pagelist, current_node);
1591 if (err)
1592 goto out;
1593 err = store_status(status, start, current_node, i - start);
1594 if (err)
1595 goto out;
1596 start = i;
1597 current_node = node;
3140a227
BG
1598 }
1599
a49bd4d7
MH
1600 /*
1601 * Errors in the page lookup or isolation are not fatal and we simply
1602 * report them via status
1603 */
1604 err = add_page_for_migration(mm, addr, current_node,
1605 &pagelist, flags & MPOL_MF_MOVE_ALL);
1606 if (!err)
1607 continue;
3140a227 1608
a49bd4d7
MH
1609 err = store_status(status, i, err, 1);
1610 if (err)
1611 goto out_flush;
5e9a0f02 1612
a49bd4d7
MH
1613 err = do_move_pages_to_node(mm, &pagelist, current_node);
1614 if (err)
1615 goto out;
1616 if (i > start) {
1617 err = store_status(status, start, current_node, i - start);
1618 if (err)
1619 goto out;
1620 }
1621 current_node = NUMA_NO_NODE;
3140a227 1622 }
a49bd4d7 1623out_flush:
8f175cf5
MH
1624 if (list_empty(&pagelist))
1625 return err;
1626
a49bd4d7
MH
1627 /* Make sure we do not overwrite the existing error */
1628 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1629 if (!err1)
1630 err1 = store_status(status, start, current_node, i - start);
1631 if (!err)
1632 err = err1;
5e9a0f02
BG
1633out:
1634 return err;
1635}
1636
742755a1 1637/*
2f007e74 1638 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1639 */
80bba129
BG
1640static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1641 const void __user **pages, int *status)
742755a1 1642{
2f007e74 1643 unsigned long i;
2f007e74 1644
742755a1
CL
1645 down_read(&mm->mmap_sem);
1646
2f007e74 1647 for (i = 0; i < nr_pages; i++) {
80bba129 1648 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1649 struct vm_area_struct *vma;
1650 struct page *page;
c095adbc 1651 int err = -EFAULT;
2f007e74
BG
1652
1653 vma = find_vma(mm, addr);
70384dc6 1654 if (!vma || addr < vma->vm_start)
742755a1
CL
1655 goto set_status;
1656
d899844e
KS
1657 /* FOLL_DUMP to ignore special (like zero) pages */
1658 page = follow_page(vma, addr, FOLL_DUMP);
89f5b7da
LT
1659
1660 err = PTR_ERR(page);
1661 if (IS_ERR(page))
1662 goto set_status;
1663
d899844e 1664 err = page ? page_to_nid(page) : -ENOENT;
742755a1 1665set_status:
80bba129
BG
1666 *status = err;
1667
1668 pages++;
1669 status++;
1670 }
1671
1672 up_read(&mm->mmap_sem);
1673}
1674
1675/*
1676 * Determine the nodes of a user array of pages and store it in
1677 * a user array of status.
1678 */
1679static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1680 const void __user * __user *pages,
1681 int __user *status)
1682{
1683#define DO_PAGES_STAT_CHUNK_NR 16
1684 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1685 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1686
87b8d1ad
PA
1687 while (nr_pages) {
1688 unsigned long chunk_nr;
80bba129 1689
87b8d1ad
PA
1690 chunk_nr = nr_pages;
1691 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1692 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1693
1694 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1695 break;
80bba129
BG
1696
1697 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1698
87b8d1ad
PA
1699 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1700 break;
742755a1 1701
87b8d1ad
PA
1702 pages += chunk_nr;
1703 status += chunk_nr;
1704 nr_pages -= chunk_nr;
1705 }
1706 return nr_pages ? -EFAULT : 0;
742755a1
CL
1707}
1708
1709/*
1710 * Move a list of pages in the address space of the currently executing
1711 * process.
1712 */
7addf443
DB
1713static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1714 const void __user * __user *pages,
1715 const int __user *nodes,
1716 int __user *status, int flags)
742755a1 1717{
742755a1 1718 struct task_struct *task;
742755a1 1719 struct mm_struct *mm;
5e9a0f02 1720 int err;
3268c63e 1721 nodemask_t task_nodes;
742755a1
CL
1722
1723 /* Check flags */
1724 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1725 return -EINVAL;
1726
1727 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1728 return -EPERM;
1729
1730 /* Find the mm_struct */
a879bf58 1731 rcu_read_lock();
228ebcbe 1732 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1733 if (!task) {
a879bf58 1734 rcu_read_unlock();
742755a1
CL
1735 return -ESRCH;
1736 }
3268c63e 1737 get_task_struct(task);
742755a1
CL
1738
1739 /*
1740 * Check if this process has the right to modify the specified
197e7e52 1741 * process. Use the regular "ptrace_may_access()" checks.
742755a1 1742 */
197e7e52 1743 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
c69e8d9c 1744 rcu_read_unlock();
742755a1 1745 err = -EPERM;
5e9a0f02 1746 goto out;
742755a1 1747 }
c69e8d9c 1748 rcu_read_unlock();
742755a1 1749
86c3a764
DQ
1750 err = security_task_movememory(task);
1751 if (err)
5e9a0f02 1752 goto out;
86c3a764 1753
3268c63e
CL
1754 task_nodes = cpuset_mems_allowed(task);
1755 mm = get_task_mm(task);
1756 put_task_struct(task);
1757
6e8b09ea
SL
1758 if (!mm)
1759 return -EINVAL;
1760
1761 if (nodes)
1762 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1763 nodes, status, flags);
1764 else
1765 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1766
742755a1
CL
1767 mmput(mm);
1768 return err;
3268c63e
CL
1769
1770out:
1771 put_task_struct(task);
1772 return err;
742755a1 1773}
742755a1 1774
7addf443
DB
1775SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1776 const void __user * __user *, pages,
1777 const int __user *, nodes,
1778 int __user *, status, int, flags)
1779{
1780 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1781}
1782
1783#ifdef CONFIG_COMPAT
1784COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1785 compat_uptr_t __user *, pages32,
1786 const int __user *, nodes,
1787 int __user *, status,
1788 int, flags)
1789{
1790 const void __user * __user *pages;
1791 int i;
1792
1793 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1794 for (i = 0; i < nr_pages; i++) {
1795 compat_uptr_t p;
1796
1797 if (get_user(p, pages32 + i) ||
1798 put_user(compat_ptr(p), pages + i))
1799 return -EFAULT;
1800 }
1801 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1802}
1803#endif /* CONFIG_COMPAT */
1804
7039e1db
PZ
1805#ifdef CONFIG_NUMA_BALANCING
1806/*
1807 * Returns true if this is a safe migration target node for misplaced NUMA
1808 * pages. Currently it only checks the watermarks which crude
1809 */
1810static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1811 unsigned long nr_migrate_pages)
7039e1db
PZ
1812{
1813 int z;
599d0c95 1814
7039e1db
PZ
1815 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1816 struct zone *zone = pgdat->node_zones + z;
1817
1818 if (!populated_zone(zone))
1819 continue;
1820
7039e1db
PZ
1821 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1822 if (!zone_watermark_ok(zone, 0,
1823 high_wmark_pages(zone) +
1824 nr_migrate_pages,
1825 0, 0))
1826 continue;
1827 return true;
1828 }
1829 return false;
1830}
1831
1832static struct page *alloc_misplaced_dst_page(struct page *page,
666feb21 1833 unsigned long data)
7039e1db
PZ
1834{
1835 int nid = (int) data;
1836 struct page *newpage;
1837
96db800f 1838 newpage = __alloc_pages_node(nid,
e97ca8e5
JW
1839 (GFP_HIGHUSER_MOVABLE |
1840 __GFP_THISNODE | __GFP_NOMEMALLOC |
1841 __GFP_NORETRY | __GFP_NOWARN) &
8479eba7 1842 ~__GFP_RECLAIM, 0);
bac0382c 1843
7039e1db
PZ
1844 return newpage;
1845}
1846
1c30e017 1847static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
b32967ff 1848{
340ef390 1849 int page_lru;
a8f60772 1850
309381fe 1851 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
3abef4e6 1852
7039e1db 1853 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1854 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1855 return 0;
7039e1db 1856
340ef390
HD
1857 if (isolate_lru_page(page))
1858 return 0;
7039e1db 1859
340ef390
HD
1860 /*
1861 * migrate_misplaced_transhuge_page() skips page migration's usual
1862 * check on page_count(), so we must do it here, now that the page
1863 * has been isolated: a GUP pin, or any other pin, prevents migration.
1864 * The expected page count is 3: 1 for page's mapcount and 1 for the
1865 * caller's pin and 1 for the reference taken by isolate_lru_page().
1866 */
1867 if (PageTransHuge(page) && page_count(page) != 3) {
1868 putback_lru_page(page);
1869 return 0;
7039e1db
PZ
1870 }
1871
340ef390 1872 page_lru = page_is_file_cache(page);
599d0c95 1873 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
340ef390
HD
1874 hpage_nr_pages(page));
1875
149c33e1 1876 /*
340ef390
HD
1877 * Isolating the page has taken another reference, so the
1878 * caller's reference can be safely dropped without the page
1879 * disappearing underneath us during migration.
149c33e1
MG
1880 */
1881 put_page(page);
340ef390 1882 return 1;
b32967ff
MG
1883}
1884
de466bd6
MG
1885bool pmd_trans_migrating(pmd_t pmd)
1886{
1887 struct page *page = pmd_page(pmd);
1888 return PageLocked(page);
1889}
1890
b32967ff
MG
1891/*
1892 * Attempt to migrate a misplaced page to the specified destination
1893 * node. Caller is expected to have an elevated reference count on
1894 * the page that will be dropped by this function before returning.
1895 */
1bc115d8
MG
1896int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1897 int node)
b32967ff
MG
1898{
1899 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1900 int isolated;
b32967ff
MG
1901 int nr_remaining;
1902 LIST_HEAD(migratepages);
1903
1904 /*
1bc115d8
MG
1905 * Don't migrate file pages that are mapped in multiple processes
1906 * with execute permissions as they are probably shared libraries.
b32967ff 1907 */
1bc115d8
MG
1908 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1909 (vma->vm_flags & VM_EXEC))
b32967ff 1910 goto out;
b32967ff 1911
09a913a7
MG
1912 /*
1913 * Also do not migrate dirty pages as not all filesystems can move
1914 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1915 */
1916 if (page_is_file_cache(page) && PageDirty(page))
1917 goto out;
1918
b32967ff
MG
1919 isolated = numamigrate_isolate_page(pgdat, page);
1920 if (!isolated)
1921 goto out;
1922
1923 list_add(&page->lru, &migratepages);
9c620e2b 1924 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
68711a74
DR
1925 NULL, node, MIGRATE_ASYNC,
1926 MR_NUMA_MISPLACED);
b32967ff 1927 if (nr_remaining) {
59c82b70
JK
1928 if (!list_empty(&migratepages)) {
1929 list_del(&page->lru);
599d0c95 1930 dec_node_page_state(page, NR_ISOLATED_ANON +
59c82b70
JK
1931 page_is_file_cache(page));
1932 putback_lru_page(page);
1933 }
b32967ff
MG
1934 isolated = 0;
1935 } else
1936 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1937 BUG_ON(!list_empty(&migratepages));
7039e1db 1938 return isolated;
340ef390
HD
1939
1940out:
1941 put_page(page);
1942 return 0;
7039e1db 1943}
220018d3 1944#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1945
220018d3 1946#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1947/*
1948 * Migrates a THP to a given target node. page must be locked and is unlocked
1949 * before returning.
1950 */
b32967ff
MG
1951int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1952 struct vm_area_struct *vma,
1953 pmd_t *pmd, pmd_t entry,
1954 unsigned long address,
1955 struct page *page, int node)
1956{
c4088ebd 1957 spinlock_t *ptl;
b32967ff
MG
1958 pg_data_t *pgdat = NODE_DATA(node);
1959 int isolated = 0;
1960 struct page *new_page = NULL;
b32967ff 1961 int page_lru = page_is_file_cache(page);
7066f0f9 1962 unsigned long start = address & HPAGE_PMD_MASK;
b32967ff 1963
b32967ff 1964 new_page = alloc_pages_node(node,
25160354 1965 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
e97ca8e5 1966 HPAGE_PMD_ORDER);
340ef390
HD
1967 if (!new_page)
1968 goto out_fail;
9a982250 1969 prep_transhuge_page(new_page);
340ef390 1970
b32967ff 1971 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 1972 if (!isolated) {
b32967ff 1973 put_page(new_page);
340ef390 1974 goto out_fail;
b32967ff 1975 }
b0943d61 1976
b32967ff 1977 /* Prepare a page as a migration target */
48c935ad 1978 __SetPageLocked(new_page);
d44d363f
SL
1979 if (PageSwapBacked(page))
1980 __SetPageSwapBacked(new_page);
b32967ff
MG
1981
1982 /* anon mapping, we can simply copy page->mapping to the new page: */
1983 new_page->mapping = page->mapping;
1984 new_page->index = page->index;
7eef5f97
AA
1985 /* flush the cache before copying using the kernel virtual address */
1986 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
b32967ff
MG
1987 migrate_page_copy(new_page, page);
1988 WARN_ON(PageLRU(new_page));
1989
1990 /* Recheck the target PMD */
c4088ebd 1991 ptl = pmd_lock(mm, pmd);
f4e177d1 1992 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
c4088ebd 1993 spin_unlock(ptl);
b32967ff
MG
1994
1995 /* Reverse changes made by migrate_page_copy() */
1996 if (TestClearPageActive(new_page))
1997 SetPageActive(page);
1998 if (TestClearPageUnevictable(new_page))
1999 SetPageUnevictable(page);
b32967ff
MG
2000
2001 unlock_page(new_page);
2002 put_page(new_page); /* Free it */
2003
a54a407f
MG
2004 /* Retake the callers reference and putback on LRU */
2005 get_page(page);
b32967ff 2006 putback_lru_page(page);
599d0c95 2007 mod_node_page_state(page_pgdat(page),
a54a407f 2008 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
eb4489f6
MG
2009
2010 goto out_unlock;
b32967ff
MG
2011 }
2012
10102459 2013 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
f55e1014 2014 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
b32967ff 2015
2b4847e7 2016 /*
d7c33934
AA
2017 * Overwrite the old entry under pagetable lock and establish
2018 * the new PTE. Any parallel GUP will either observe the old
2019 * page blocking on the page lock, block on the page table
2020 * lock or observe the new page. The SetPageUptodate on the
2021 * new page and page_add_new_anon_rmap guarantee the copy is
2022 * visible before the pagetable update.
2b4847e7 2023 */
7066f0f9 2024 page_add_anon_rmap(new_page, vma, start, true);
d7c33934
AA
2025 /*
2026 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2027 * has already been flushed globally. So no TLB can be currently
2028 * caching this non present pmd mapping. There's no need to clear the
2029 * pmd before doing set_pmd_at(), nor to flush the TLB after
2030 * set_pmd_at(). Clearing the pmd here would introduce a race
2031 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2032 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2033 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2034 * pmd.
2035 */
7066f0f9 2036 set_pmd_at(mm, start, pmd, entry);
ce4a9cc5 2037 update_mmu_cache_pmd(vma, address, &entry);
2b4847e7 2038
f4e177d1 2039 page_ref_unfreeze(page, 2);
51afb12b 2040 mlock_migrate_page(new_page, page);
d281ee61 2041 page_remove_rmap(page, true);
7cd12b4a 2042 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2b4847e7 2043
c4088ebd 2044 spin_unlock(ptl);
b32967ff 2045
11de9927
MG
2046 /* Take an "isolate" reference and put new page on the LRU. */
2047 get_page(new_page);
2048 putback_lru_page(new_page);
2049
b32967ff
MG
2050 unlock_page(new_page);
2051 unlock_page(page);
2052 put_page(page); /* Drop the rmap reference */
2053 put_page(page); /* Drop the LRU isolation reference */
2054
2055 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2056 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2057
599d0c95 2058 mod_node_page_state(page_pgdat(page),
b32967ff
MG
2059 NR_ISOLATED_ANON + page_lru,
2060 -HPAGE_PMD_NR);
2061 return isolated;
2062
340ef390
HD
2063out_fail:
2064 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2b4847e7
MG
2065 ptl = pmd_lock(mm, pmd);
2066 if (pmd_same(*pmd, entry)) {
4d942466 2067 entry = pmd_modify(entry, vma->vm_page_prot);
7066f0f9 2068 set_pmd_at(mm, start, pmd, entry);
2b4847e7
MG
2069 update_mmu_cache_pmd(vma, address, &entry);
2070 }
2071 spin_unlock(ptl);
a54a407f 2072
eb4489f6 2073out_unlock:
340ef390 2074 unlock_page(page);
b32967ff 2075 put_page(page);
b32967ff
MG
2076 return 0;
2077}
7039e1db
PZ
2078#endif /* CONFIG_NUMA_BALANCING */
2079
2080#endif /* CONFIG_NUMA */
8763cb45 2081
6b368cd4 2082#if defined(CONFIG_MIGRATE_VMA_HELPER)
8763cb45
JG
2083struct migrate_vma {
2084 struct vm_area_struct *vma;
2085 unsigned long *dst;
2086 unsigned long *src;
2087 unsigned long cpages;
2088 unsigned long npages;
2089 unsigned long start;
2090 unsigned long end;
2091};
2092
2093static int migrate_vma_collect_hole(unsigned long start,
2094 unsigned long end,
2095 struct mm_walk *walk)
2096{
2097 struct migrate_vma *migrate = walk->private;
2098 unsigned long addr;
2099
8315ada7 2100 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
e20d103b 2101 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
8315ada7 2102 migrate->dst[migrate->npages] = 0;
e20d103b 2103 migrate->npages++;
8315ada7
JG
2104 migrate->cpages++;
2105 }
2106
2107 return 0;
2108}
2109
2110static int migrate_vma_collect_skip(unsigned long start,
2111 unsigned long end,
2112 struct mm_walk *walk)
2113{
2114 struct migrate_vma *migrate = walk->private;
2115 unsigned long addr;
2116
8763cb45
JG
2117 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2118 migrate->dst[migrate->npages] = 0;
2119 migrate->src[migrate->npages++] = 0;
2120 }
2121
2122 return 0;
2123}
2124
2125static int migrate_vma_collect_pmd(pmd_t *pmdp,
2126 unsigned long start,
2127 unsigned long end,
2128 struct mm_walk *walk)
2129{
2130 struct migrate_vma *migrate = walk->private;
2131 struct vm_area_struct *vma = walk->vma;
2132 struct mm_struct *mm = vma->vm_mm;
8c3328f1 2133 unsigned long addr = start, unmapped = 0;
8763cb45
JG
2134 spinlock_t *ptl;
2135 pte_t *ptep;
2136
2137again:
2138 if (pmd_none(*pmdp))
2139 return migrate_vma_collect_hole(start, end, walk);
2140
2141 if (pmd_trans_huge(*pmdp)) {
2142 struct page *page;
2143
2144 ptl = pmd_lock(mm, pmdp);
2145 if (unlikely(!pmd_trans_huge(*pmdp))) {
2146 spin_unlock(ptl);
2147 goto again;
2148 }
2149
2150 page = pmd_page(*pmdp);
2151 if (is_huge_zero_page(page)) {
2152 spin_unlock(ptl);
2153 split_huge_pmd(vma, pmdp, addr);
2154 if (pmd_trans_unstable(pmdp))
8315ada7 2155 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2156 walk);
2157 } else {
2158 int ret;
2159
2160 get_page(page);
2161 spin_unlock(ptl);
2162 if (unlikely(!trylock_page(page)))
8315ada7 2163 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2164 walk);
2165 ret = split_huge_page(page);
2166 unlock_page(page);
2167 put_page(page);
8315ada7
JG
2168 if (ret)
2169 return migrate_vma_collect_skip(start, end,
2170 walk);
2171 if (pmd_none(*pmdp))
8763cb45
JG
2172 return migrate_vma_collect_hole(start, end,
2173 walk);
2174 }
2175 }
2176
2177 if (unlikely(pmd_bad(*pmdp)))
8315ada7 2178 return migrate_vma_collect_skip(start, end, walk);
8763cb45
JG
2179
2180 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
8c3328f1
JG
2181 arch_enter_lazy_mmu_mode();
2182
8763cb45
JG
2183 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2184 unsigned long mpfn, pfn;
2185 struct page *page;
8c3328f1 2186 swp_entry_t entry;
8763cb45
JG
2187 pte_t pte;
2188
2189 pte = *ptep;
2190 pfn = pte_pfn(pte);
2191
a5430dda 2192 if (pte_none(pte)) {
8315ada7
JG
2193 mpfn = MIGRATE_PFN_MIGRATE;
2194 migrate->cpages++;
2195 pfn = 0;
8763cb45
JG
2196 goto next;
2197 }
2198
a5430dda
JG
2199 if (!pte_present(pte)) {
2200 mpfn = pfn = 0;
2201
2202 /*
2203 * Only care about unaddressable device page special
2204 * page table entry. Other special swap entries are not
2205 * migratable, and we ignore regular swapped page.
2206 */
2207 entry = pte_to_swp_entry(pte);
2208 if (!is_device_private_entry(entry))
2209 goto next;
2210
2211 page = device_private_entry_to_page(entry);
2212 mpfn = migrate_pfn(page_to_pfn(page))|
2213 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2214 if (is_write_device_private_entry(entry))
2215 mpfn |= MIGRATE_PFN_WRITE;
2216 } else {
8315ada7
JG
2217 if (is_zero_pfn(pfn)) {
2218 mpfn = MIGRATE_PFN_MIGRATE;
2219 migrate->cpages++;
2220 pfn = 0;
2221 goto next;
2222 }
df6ad698 2223 page = _vm_normal_page(migrate->vma, addr, pte, true);
a5430dda
JG
2224 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2225 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2226 }
2227
8763cb45 2228 /* FIXME support THP */
8763cb45
JG
2229 if (!page || !page->mapping || PageTransCompound(page)) {
2230 mpfn = pfn = 0;
2231 goto next;
2232 }
a5430dda 2233 pfn = page_to_pfn(page);
8763cb45
JG
2234
2235 /*
2236 * By getting a reference on the page we pin it and that blocks
2237 * any kind of migration. Side effect is that it "freezes" the
2238 * pte.
2239 *
2240 * We drop this reference after isolating the page from the lru
2241 * for non device page (device page are not on the lru and thus
2242 * can't be dropped from it).
2243 */
2244 get_page(page);
2245 migrate->cpages++;
8763cb45 2246
8c3328f1
JG
2247 /*
2248 * Optimize for the common case where page is only mapped once
2249 * in one process. If we can lock the page, then we can safely
2250 * set up a special migration page table entry now.
2251 */
2252 if (trylock_page(page)) {
2253 pte_t swp_pte;
2254
2255 mpfn |= MIGRATE_PFN_LOCKED;
2256 ptep_get_and_clear(mm, addr, ptep);
2257
2258 /* Setup special migration page table entry */
07707125
RC
2259 entry = make_migration_entry(page, mpfn &
2260 MIGRATE_PFN_WRITE);
8c3328f1
JG
2261 swp_pte = swp_entry_to_pte(entry);
2262 if (pte_soft_dirty(pte))
2263 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2264 set_pte_at(mm, addr, ptep, swp_pte);
2265
2266 /*
2267 * This is like regular unmap: we remove the rmap and
2268 * drop page refcount. Page won't be freed, as we took
2269 * a reference just above.
2270 */
2271 page_remove_rmap(page, false);
2272 put_page(page);
a5430dda
JG
2273
2274 if (pte_present(pte))
2275 unmapped++;
8c3328f1
JG
2276 }
2277
8763cb45 2278next:
a5430dda 2279 migrate->dst[migrate->npages] = 0;
8763cb45
JG
2280 migrate->src[migrate->npages++] = mpfn;
2281 }
8c3328f1 2282 arch_leave_lazy_mmu_mode();
8763cb45
JG
2283 pte_unmap_unlock(ptep - 1, ptl);
2284
8c3328f1
JG
2285 /* Only flush the TLB if we actually modified any entries */
2286 if (unmapped)
2287 flush_tlb_range(walk->vma, start, end);
2288
8763cb45
JG
2289 return 0;
2290}
2291
2292/*
2293 * migrate_vma_collect() - collect pages over a range of virtual addresses
2294 * @migrate: migrate struct containing all migration information
2295 *
2296 * This will walk the CPU page table. For each virtual address backed by a
2297 * valid page, it updates the src array and takes a reference on the page, in
2298 * order to pin the page until we lock it and unmap it.
2299 */
2300static void migrate_vma_collect(struct migrate_vma *migrate)
2301{
2302 struct mm_walk mm_walk;
2303
2304 mm_walk.pmd_entry = migrate_vma_collect_pmd;
2305 mm_walk.pte_entry = NULL;
2306 mm_walk.pte_hole = migrate_vma_collect_hole;
2307 mm_walk.hugetlb_entry = NULL;
2308 mm_walk.test_walk = NULL;
2309 mm_walk.vma = migrate->vma;
2310 mm_walk.mm = migrate->vma->vm_mm;
2311 mm_walk.private = migrate;
2312
8c3328f1
JG
2313 mmu_notifier_invalidate_range_start(mm_walk.mm,
2314 migrate->start,
2315 migrate->end);
8763cb45 2316 walk_page_range(migrate->start, migrate->end, &mm_walk);
8c3328f1
JG
2317 mmu_notifier_invalidate_range_end(mm_walk.mm,
2318 migrate->start,
2319 migrate->end);
8763cb45
JG
2320
2321 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2322}
2323
2324/*
2325 * migrate_vma_check_page() - check if page is pinned or not
2326 * @page: struct page to check
2327 *
2328 * Pinned pages cannot be migrated. This is the same test as in
2329 * migrate_page_move_mapping(), except that here we allow migration of a
2330 * ZONE_DEVICE page.
2331 */
2332static bool migrate_vma_check_page(struct page *page)
2333{
2334 /*
2335 * One extra ref because caller holds an extra reference, either from
2336 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2337 * a device page.
2338 */
2339 int extra = 1;
2340
2341 /*
2342 * FIXME support THP (transparent huge page), it is bit more complex to
2343 * check them than regular pages, because they can be mapped with a pmd
2344 * or with a pte (split pte mapping).
2345 */
2346 if (PageCompound(page))
2347 return false;
2348
a5430dda
JG
2349 /* Page from ZONE_DEVICE have one extra reference */
2350 if (is_zone_device_page(page)) {
2351 /*
2352 * Private page can never be pin as they have no valid pte and
2353 * GUP will fail for those. Yet if there is a pending migration
2354 * a thread might try to wait on the pte migration entry and
2355 * will bump the page reference count. Sadly there is no way to
2356 * differentiate a regular pin from migration wait. Hence to
2357 * avoid 2 racing thread trying to migrate back to CPU to enter
2358 * infinite loop (one stoping migration because the other is
2359 * waiting on pte migration entry). We always return true here.
2360 *
2361 * FIXME proper solution is to rework migration_entry_wait() so
2362 * it does not need to take a reference on page.
2363 */
2364 if (is_device_private_page(page))
2365 return true;
2366
df6ad698
JG
2367 /*
2368 * Only allow device public page to be migrated and account for
2369 * the extra reference count imply by ZONE_DEVICE pages.
2370 */
2371 if (!is_device_public_page(page))
2372 return false;
2373 extra++;
a5430dda
JG
2374 }
2375
df6ad698
JG
2376 /* For file back page */
2377 if (page_mapping(page))
2378 extra += 1 + page_has_private(page);
2379
8763cb45
JG
2380 if ((page_count(page) - extra) > page_mapcount(page))
2381 return false;
2382
2383 return true;
2384}
2385
2386/*
2387 * migrate_vma_prepare() - lock pages and isolate them from the lru
2388 * @migrate: migrate struct containing all migration information
2389 *
2390 * This locks pages that have been collected by migrate_vma_collect(). Once each
2391 * page is locked it is isolated from the lru (for non-device pages). Finally,
2392 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2393 * migrated by concurrent kernel threads.
2394 */
2395static void migrate_vma_prepare(struct migrate_vma *migrate)
2396{
2397 const unsigned long npages = migrate->npages;
8c3328f1
JG
2398 const unsigned long start = migrate->start;
2399 unsigned long addr, i, restore = 0;
8763cb45 2400 bool allow_drain = true;
8763cb45
JG
2401
2402 lru_add_drain();
2403
2404 for (i = 0; (i < npages) && migrate->cpages; i++) {
2405 struct page *page = migrate_pfn_to_page(migrate->src[i]);
8c3328f1 2406 bool remap = true;
8763cb45
JG
2407
2408 if (!page)
2409 continue;
2410
8c3328f1
JG
2411 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2412 /*
2413 * Because we are migrating several pages there can be
2414 * a deadlock between 2 concurrent migration where each
2415 * are waiting on each other page lock.
2416 *
2417 * Make migrate_vma() a best effort thing and backoff
2418 * for any page we can not lock right away.
2419 */
2420 if (!trylock_page(page)) {
2421 migrate->src[i] = 0;
2422 migrate->cpages--;
2423 put_page(page);
2424 continue;
2425 }
2426 remap = false;
2427 migrate->src[i] |= MIGRATE_PFN_LOCKED;
8763cb45 2428 }
8763cb45 2429
a5430dda
JG
2430 /* ZONE_DEVICE pages are not on LRU */
2431 if (!is_zone_device_page(page)) {
2432 if (!PageLRU(page) && allow_drain) {
2433 /* Drain CPU's pagevec */
2434 lru_add_drain_all();
2435 allow_drain = false;
2436 }
8763cb45 2437
a5430dda
JG
2438 if (isolate_lru_page(page)) {
2439 if (remap) {
2440 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2441 migrate->cpages--;
2442 restore++;
2443 } else {
2444 migrate->src[i] = 0;
2445 unlock_page(page);
2446 migrate->cpages--;
2447 put_page(page);
2448 }
2449 continue;
8c3328f1 2450 }
a5430dda
JG
2451
2452 /* Drop the reference we took in collect */
2453 put_page(page);
8763cb45
JG
2454 }
2455
2456 if (!migrate_vma_check_page(page)) {
8c3328f1
JG
2457 if (remap) {
2458 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2459 migrate->cpages--;
2460 restore++;
8763cb45 2461
a5430dda
JG
2462 if (!is_zone_device_page(page)) {
2463 get_page(page);
2464 putback_lru_page(page);
2465 }
8c3328f1
JG
2466 } else {
2467 migrate->src[i] = 0;
2468 unlock_page(page);
2469 migrate->cpages--;
2470
a5430dda
JG
2471 if (!is_zone_device_page(page))
2472 putback_lru_page(page);
2473 else
2474 put_page(page);
8c3328f1 2475 }
8763cb45
JG
2476 }
2477 }
8c3328f1
JG
2478
2479 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2480 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2481
2482 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2483 continue;
2484
2485 remove_migration_pte(page, migrate->vma, addr, page);
2486
2487 migrate->src[i] = 0;
2488 unlock_page(page);
2489 put_page(page);
2490 restore--;
2491 }
8763cb45
JG
2492}
2493
2494/*
2495 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2496 * @migrate: migrate struct containing all migration information
2497 *
2498 * Replace page mapping (CPU page table pte) with a special migration pte entry
2499 * and check again if it has been pinned. Pinned pages are restored because we
2500 * cannot migrate them.
2501 *
2502 * This is the last step before we call the device driver callback to allocate
2503 * destination memory and copy contents of original page over to new page.
2504 */
2505static void migrate_vma_unmap(struct migrate_vma *migrate)
2506{
2507 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2508 const unsigned long npages = migrate->npages;
2509 const unsigned long start = migrate->start;
2510 unsigned long addr, i, restore = 0;
2511
2512 for (i = 0; i < npages; i++) {
2513 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2514
2515 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2516 continue;
2517
8c3328f1
JG
2518 if (page_mapped(page)) {
2519 try_to_unmap(page, flags);
2520 if (page_mapped(page))
2521 goto restore;
8763cb45 2522 }
8c3328f1
JG
2523
2524 if (migrate_vma_check_page(page))
2525 continue;
2526
2527restore:
2528 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2529 migrate->cpages--;
2530 restore++;
8763cb45
JG
2531 }
2532
2533 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2534 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2535
2536 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2537 continue;
2538
2539 remove_migration_ptes(page, page, false);
2540
2541 migrate->src[i] = 0;
2542 unlock_page(page);
2543 restore--;
2544
a5430dda
JG
2545 if (is_zone_device_page(page))
2546 put_page(page);
2547 else
2548 putback_lru_page(page);
8763cb45
JG
2549 }
2550}
2551
8315ada7
JG
2552static void migrate_vma_insert_page(struct migrate_vma *migrate,
2553 unsigned long addr,
2554 struct page *page,
2555 unsigned long *src,
2556 unsigned long *dst)
2557{
2558 struct vm_area_struct *vma = migrate->vma;
2559 struct mm_struct *mm = vma->vm_mm;
2560 struct mem_cgroup *memcg;
2561 bool flush = false;
2562 spinlock_t *ptl;
2563 pte_t entry;
2564 pgd_t *pgdp;
2565 p4d_t *p4dp;
2566 pud_t *pudp;
2567 pmd_t *pmdp;
2568 pte_t *ptep;
2569
2570 /* Only allow populating anonymous memory */
2571 if (!vma_is_anonymous(vma))
2572 goto abort;
2573
2574 pgdp = pgd_offset(mm, addr);
2575 p4dp = p4d_alloc(mm, pgdp, addr);
2576 if (!p4dp)
2577 goto abort;
2578 pudp = pud_alloc(mm, p4dp, addr);
2579 if (!pudp)
2580 goto abort;
2581 pmdp = pmd_alloc(mm, pudp, addr);
2582 if (!pmdp)
2583 goto abort;
2584
2585 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2586 goto abort;
2587
2588 /*
2589 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2590 * pte_offset_map() on pmds where a huge pmd might be created
2591 * from a different thread.
2592 *
2593 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2594 * parallel threads are excluded by other means.
2595 *
2596 * Here we only have down_read(mmap_sem).
2597 */
2598 if (pte_alloc(mm, pmdp, addr))
2599 goto abort;
2600
2601 /* See the comment in pte_alloc_one_map() */
2602 if (unlikely(pmd_trans_unstable(pmdp)))
2603 goto abort;
2604
2605 if (unlikely(anon_vma_prepare(vma)))
2606 goto abort;
2607 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2608 goto abort;
2609
2610 /*
2611 * The memory barrier inside __SetPageUptodate makes sure that
2612 * preceding stores to the page contents become visible before
2613 * the set_pte_at() write.
2614 */
2615 __SetPageUptodate(page);
2616
df6ad698
JG
2617 if (is_zone_device_page(page)) {
2618 if (is_device_private_page(page)) {
2619 swp_entry_t swp_entry;
2620
2621 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2622 entry = swp_entry_to_pte(swp_entry);
2623 } else if (is_device_public_page(page)) {
2624 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2625 if (vma->vm_flags & VM_WRITE)
2626 entry = pte_mkwrite(pte_mkdirty(entry));
2627 entry = pte_mkdevmap(entry);
2628 }
8315ada7
JG
2629 } else {
2630 entry = mk_pte(page, vma->vm_page_prot);
2631 if (vma->vm_flags & VM_WRITE)
2632 entry = pte_mkwrite(pte_mkdirty(entry));
2633 }
2634
2635 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2636
2637 if (pte_present(*ptep)) {
2638 unsigned long pfn = pte_pfn(*ptep);
2639
2640 if (!is_zero_pfn(pfn)) {
2641 pte_unmap_unlock(ptep, ptl);
2642 mem_cgroup_cancel_charge(page, memcg, false);
2643 goto abort;
2644 }
2645 flush = true;
2646 } else if (!pte_none(*ptep)) {
2647 pte_unmap_unlock(ptep, ptl);
2648 mem_cgroup_cancel_charge(page, memcg, false);
2649 goto abort;
2650 }
2651
2652 /*
2653 * Check for usefaultfd but do not deliver the fault. Instead,
2654 * just back off.
2655 */
2656 if (userfaultfd_missing(vma)) {
2657 pte_unmap_unlock(ptep, ptl);
2658 mem_cgroup_cancel_charge(page, memcg, false);
2659 goto abort;
2660 }
2661
2662 inc_mm_counter(mm, MM_ANONPAGES);
2663 page_add_new_anon_rmap(page, vma, addr, false);
2664 mem_cgroup_commit_charge(page, memcg, false, false);
2665 if (!is_zone_device_page(page))
2666 lru_cache_add_active_or_unevictable(page, vma);
2667 get_page(page);
2668
2669 if (flush) {
2670 flush_cache_page(vma, addr, pte_pfn(*ptep));
2671 ptep_clear_flush_notify(vma, addr, ptep);
2672 set_pte_at_notify(mm, addr, ptep, entry);
2673 update_mmu_cache(vma, addr, ptep);
2674 } else {
2675 /* No need to invalidate - it was non-present before */
2676 set_pte_at(mm, addr, ptep, entry);
2677 update_mmu_cache(vma, addr, ptep);
2678 }
2679
2680 pte_unmap_unlock(ptep, ptl);
2681 *src = MIGRATE_PFN_MIGRATE;
2682 return;
2683
2684abort:
2685 *src &= ~MIGRATE_PFN_MIGRATE;
2686}
2687
8763cb45
JG
2688/*
2689 * migrate_vma_pages() - migrate meta-data from src page to dst page
2690 * @migrate: migrate struct containing all migration information
2691 *
2692 * This migrates struct page meta-data from source struct page to destination
2693 * struct page. This effectively finishes the migration from source page to the
2694 * destination page.
2695 */
2696static void migrate_vma_pages(struct migrate_vma *migrate)
2697{
2698 const unsigned long npages = migrate->npages;
2699 const unsigned long start = migrate->start;
8315ada7
JG
2700 struct vm_area_struct *vma = migrate->vma;
2701 struct mm_struct *mm = vma->vm_mm;
2702 unsigned long addr, i, mmu_start;
2703 bool notified = false;
8763cb45
JG
2704
2705 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2706 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2707 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2708 struct address_space *mapping;
2709 int r;
2710
8315ada7
JG
2711 if (!newpage) {
2712 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
8763cb45 2713 continue;
8315ada7
JG
2714 }
2715
2716 if (!page) {
2717 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2718 continue;
2719 }
2720 if (!notified) {
2721 mmu_start = addr;
2722 notified = true;
2723 mmu_notifier_invalidate_range_start(mm,
2724 mmu_start,
2725 migrate->end);
2726 }
2727 migrate_vma_insert_page(migrate, addr, newpage,
2728 &migrate->src[i],
2729 &migrate->dst[i]);
8763cb45 2730 continue;
8315ada7 2731 }
8763cb45
JG
2732
2733 mapping = page_mapping(page);
2734
a5430dda
JG
2735 if (is_zone_device_page(newpage)) {
2736 if (is_device_private_page(newpage)) {
2737 /*
2738 * For now only support private anonymous when
2739 * migrating to un-addressable device memory.
2740 */
2741 if (mapping) {
2742 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2743 continue;
2744 }
df6ad698 2745 } else if (!is_device_public_page(newpage)) {
a5430dda
JG
2746 /*
2747 * Other types of ZONE_DEVICE page are not
2748 * supported.
2749 */
2750 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2751 continue;
2752 }
2753 }
2754
8763cb45
JG
2755 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2756 if (r != MIGRATEPAGE_SUCCESS)
2757 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2758 }
8315ada7 2759
4645b9fe
JG
2760 /*
2761 * No need to double call mmu_notifier->invalidate_range() callback as
2762 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2763 * did already call it.
2764 */
8315ada7 2765 if (notified)
4645b9fe
JG
2766 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2767 migrate->end);
8763cb45
JG
2768}
2769
2770/*
2771 * migrate_vma_finalize() - restore CPU page table entry
2772 * @migrate: migrate struct containing all migration information
2773 *
2774 * This replaces the special migration pte entry with either a mapping to the
2775 * new page if migration was successful for that page, or to the original page
2776 * otherwise.
2777 *
2778 * This also unlocks the pages and puts them back on the lru, or drops the extra
2779 * refcount, for device pages.
2780 */
2781static void migrate_vma_finalize(struct migrate_vma *migrate)
2782{
2783 const unsigned long npages = migrate->npages;
2784 unsigned long i;
2785
2786 for (i = 0; i < npages; i++) {
2787 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2788 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2789
8315ada7
JG
2790 if (!page) {
2791 if (newpage) {
2792 unlock_page(newpage);
2793 put_page(newpage);
2794 }
8763cb45 2795 continue;
8315ada7
JG
2796 }
2797
8763cb45
JG
2798 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2799 if (newpage) {
2800 unlock_page(newpage);
2801 put_page(newpage);
2802 }
2803 newpage = page;
2804 }
2805
2806 remove_migration_ptes(page, newpage, false);
2807 unlock_page(page);
2808 migrate->cpages--;
2809
a5430dda
JG
2810 if (is_zone_device_page(page))
2811 put_page(page);
2812 else
2813 putback_lru_page(page);
8763cb45
JG
2814
2815 if (newpage != page) {
2816 unlock_page(newpage);
a5430dda
JG
2817 if (is_zone_device_page(newpage))
2818 put_page(newpage);
2819 else
2820 putback_lru_page(newpage);
8763cb45
JG
2821 }
2822 }
2823}
2824
2825/*
2826 * migrate_vma() - migrate a range of memory inside vma
2827 *
2828 * @ops: migration callback for allocating destination memory and copying
2829 * @vma: virtual memory area containing the range to be migrated
2830 * @start: start address of the range to migrate (inclusive)
2831 * @end: end address of the range to migrate (exclusive)
2832 * @src: array of hmm_pfn_t containing source pfns
2833 * @dst: array of hmm_pfn_t containing destination pfns
2834 * @private: pointer passed back to each of the callback
2835 * Returns: 0 on success, error code otherwise
2836 *
2837 * This function tries to migrate a range of memory virtual address range, using
2838 * callbacks to allocate and copy memory from source to destination. First it
2839 * collects all the pages backing each virtual address in the range, saving this
2840 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2841 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2842 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2843 * in the corresponding src array entry. It then restores any pages that are
2844 * pinned, by remapping and unlocking those pages.
2845 *
2846 * At this point it calls the alloc_and_copy() callback. For documentation on
2847 * what is expected from that callback, see struct migrate_vma_ops comments in
2848 * include/linux/migrate.h
2849 *
2850 * After the alloc_and_copy() callback, this function goes over each entry in
2851 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2852 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2853 * then the function tries to migrate struct page information from the source
2854 * struct page to the destination struct page. If it fails to migrate the struct
2855 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2856 * array.
2857 *
2858 * At this point all successfully migrated pages have an entry in the src
2859 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2860 * array entry with MIGRATE_PFN_VALID flag set.
2861 *
2862 * It then calls the finalize_and_map() callback. See comments for "struct
2863 * migrate_vma_ops", in include/linux/migrate.h for details about
2864 * finalize_and_map() behavior.
2865 *
2866 * After the finalize_and_map() callback, for successfully migrated pages, this
2867 * function updates the CPU page table to point to new pages, otherwise it
2868 * restores the CPU page table to point to the original source pages.
2869 *
2870 * Function returns 0 after the above steps, even if no pages were migrated
2871 * (The function only returns an error if any of the arguments are invalid.)
2872 *
2873 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2874 * unsigned long entries.
2875 */
2876int migrate_vma(const struct migrate_vma_ops *ops,
2877 struct vm_area_struct *vma,
2878 unsigned long start,
2879 unsigned long end,
2880 unsigned long *src,
2881 unsigned long *dst,
2882 void *private)
2883{
2884 struct migrate_vma migrate;
2885
2886 /* Sanity check the arguments */
2887 start &= PAGE_MASK;
2888 end &= PAGE_MASK;
e1fb4a08
DJ
2889 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2890 vma_is_dax(vma))
8763cb45
JG
2891 return -EINVAL;
2892 if (start < vma->vm_start || start >= vma->vm_end)
2893 return -EINVAL;
2894 if (end <= vma->vm_start || end > vma->vm_end)
2895 return -EINVAL;
2896 if (!ops || !src || !dst || start >= end)
2897 return -EINVAL;
2898
2899 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2900 migrate.src = src;
2901 migrate.dst = dst;
2902 migrate.start = start;
2903 migrate.npages = 0;
2904 migrate.cpages = 0;
2905 migrate.end = end;
2906 migrate.vma = vma;
2907
2908 /* Collect, and try to unmap source pages */
2909 migrate_vma_collect(&migrate);
2910 if (!migrate.cpages)
2911 return 0;
2912
2913 /* Lock and isolate page */
2914 migrate_vma_prepare(&migrate);
2915 if (!migrate.cpages)
2916 return 0;
2917
2918 /* Unmap pages */
2919 migrate_vma_unmap(&migrate);
2920 if (!migrate.cpages)
2921 return 0;
2922
2923 /*
2924 * At this point pages are locked and unmapped, and thus they have
2925 * stable content and can safely be copied to destination memory that
2926 * is allocated by the callback.
2927 *
2928 * Note that migration can fail in migrate_vma_struct_page() for each
2929 * individual page.
2930 */
2931 ops->alloc_and_copy(vma, src, dst, start, end, private);
2932
2933 /* This does the real migration of struct page */
2934 migrate_vma_pages(&migrate);
2935
2936 ops->finalize_and_map(vma, src, dst, start, end, private);
2937
2938 /* Unlock and remap pages */
2939 migrate_vma_finalize(&migrate);
2940
2941 return 0;
2942}
2943EXPORT_SYMBOL(migrate_vma);
6b368cd4 2944#endif /* defined(MIGRATE_VMA_HELPER) */