]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/migrate.c
mm, hugetlb: integrate giga hugetlb more naturally to the allocation path
[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>
290408d4 37#include <linux/hugetlb.h>
8e6ac7fa 38#include <linux/hugetlb_cgroup.h>
5a0e3ad6 39#include <linux/gfp.h>
df6ad698 40#include <linux/pfn_t.h>
a5430dda 41#include <linux/memremap.h>
8315ada7 42#include <linux/userfaultfd_k.h>
bf6bddf1 43#include <linux/balloon_compaction.h>
f714f4f2 44#include <linux/mmu_notifier.h>
33c3fc71 45#include <linux/page_idle.h>
d435edca 46#include <linux/page_owner.h>
6e84f315 47#include <linux/sched/mm.h>
197e7e52 48#include <linux/ptrace.h>
b20a3503 49
0d1836c3
MN
50#include <asm/tlbflush.h>
51
7b2a2d4a
MG
52#define CREATE_TRACE_POINTS
53#include <trace/events/migrate.h>
54
b20a3503
CL
55#include "internal.h"
56
b20a3503 57/*
742755a1 58 * migrate_prep() needs to be called before we start compiling a list of pages
748446bb
MG
59 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
60 * undesirable, use migrate_prep_local()
b20a3503
CL
61 */
62int migrate_prep(void)
63{
b20a3503
CL
64 /*
65 * Clear the LRU lists so pages can be isolated.
66 * Note that pages may be moved off the LRU after we have
67 * drained them. Those pages will fail to migrate like other
68 * pages that may be busy.
69 */
70 lru_add_drain_all();
71
72 return 0;
73}
74
748446bb
MG
75/* Do the necessary work of migrate_prep but not if it involves other CPUs */
76int migrate_prep_local(void)
77{
78 lru_add_drain();
79
80 return 0;
81}
82
9e5bcd61 83int isolate_movable_page(struct page *page, isolate_mode_t mode)
bda807d4
MK
84{
85 struct address_space *mapping;
86
87 /*
88 * Avoid burning cycles with pages that are yet under __free_pages(),
89 * or just got freed under us.
90 *
91 * In case we 'win' a race for a movable page being freed under us and
92 * raise its refcount preventing __free_pages() from doing its job
93 * the put_page() at the end of this block will take care of
94 * release this page, thus avoiding a nasty leakage.
95 */
96 if (unlikely(!get_page_unless_zero(page)))
97 goto out;
98
99 /*
100 * Check PageMovable before holding a PG_lock because page's owner
101 * assumes anybody doesn't touch PG_lock of newly allocated page
102 * so unconditionally grapping the lock ruins page's owner side.
103 */
104 if (unlikely(!__PageMovable(page)))
105 goto out_putpage;
106 /*
107 * As movable pages are not isolated from LRU lists, concurrent
108 * compaction threads can race against page migration functions
109 * as well as race against the releasing a page.
110 *
111 * In order to avoid having an already isolated movable page
112 * being (wrongly) re-isolated while it is under migration,
113 * or to avoid attempting to isolate pages being released,
114 * lets be sure we have the page lock
115 * before proceeding with the movable page isolation steps.
116 */
117 if (unlikely(!trylock_page(page)))
118 goto out_putpage;
119
120 if (!PageMovable(page) || PageIsolated(page))
121 goto out_no_isolated;
122
123 mapping = page_mapping(page);
124 VM_BUG_ON_PAGE(!mapping, page);
125
126 if (!mapping->a_ops->isolate_page(page, mode))
127 goto out_no_isolated;
128
129 /* Driver shouldn't use PG_isolated bit of page->flags */
130 WARN_ON_ONCE(PageIsolated(page));
131 __SetPageIsolated(page);
132 unlock_page(page);
133
9e5bcd61 134 return 0;
bda807d4
MK
135
136out_no_isolated:
137 unlock_page(page);
138out_putpage:
139 put_page(page);
140out:
9e5bcd61 141 return -EBUSY;
bda807d4
MK
142}
143
144/* It should be called on page which is PG_movable */
145void putback_movable_page(struct page *page)
146{
147 struct address_space *mapping;
148
149 VM_BUG_ON_PAGE(!PageLocked(page), page);
150 VM_BUG_ON_PAGE(!PageMovable(page), page);
151 VM_BUG_ON_PAGE(!PageIsolated(page), page);
152
153 mapping = page_mapping(page);
154 mapping->a_ops->putback_page(page);
155 __ClearPageIsolated(page);
156}
157
5733c7d1
RA
158/*
159 * Put previously isolated pages back onto the appropriate lists
160 * from where they were once taken off for compaction/migration.
161 *
59c82b70
JK
162 * This function shall be used whenever the isolated pageset has been
163 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
164 * and isolate_huge_page().
5733c7d1
RA
165 */
166void putback_movable_pages(struct list_head *l)
167{
168 struct page *page;
169 struct page *page2;
170
b20a3503 171 list_for_each_entry_safe(page, page2, l, lru) {
31caf665
NH
172 if (unlikely(PageHuge(page))) {
173 putback_active_hugepage(page);
174 continue;
175 }
e24f0b8f 176 list_del(&page->lru);
bda807d4
MK
177 /*
178 * We isolated non-lru movable page so here we can use
179 * __PageMovable because LRU page's mapping cannot have
180 * PAGE_MAPPING_MOVABLE.
181 */
b1123ea6 182 if (unlikely(__PageMovable(page))) {
bda807d4
MK
183 VM_BUG_ON_PAGE(!PageIsolated(page), page);
184 lock_page(page);
185 if (PageMovable(page))
186 putback_movable_page(page);
187 else
188 __ClearPageIsolated(page);
189 unlock_page(page);
190 put_page(page);
191 } else {
e8db67eb
NH
192 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
193 page_is_file_cache(page), -hpage_nr_pages(page));
fc280fe8 194 putback_lru_page(page);
bda807d4 195 }
b20a3503 196 }
b20a3503
CL
197}
198
0697212a
CL
199/*
200 * Restore a potential migration pte to a working pte entry
201 */
e4b82222 202static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
e9995ef9 203 unsigned long addr, void *old)
0697212a 204{
3fe87967
KS
205 struct page_vma_mapped_walk pvmw = {
206 .page = old,
207 .vma = vma,
208 .address = addr,
209 .flags = PVMW_SYNC | PVMW_MIGRATION,
210 };
211 struct page *new;
212 pte_t pte;
0697212a 213 swp_entry_t entry;
0697212a 214
3fe87967
KS
215 VM_BUG_ON_PAGE(PageTail(page), page);
216 while (page_vma_mapped_walk(&pvmw)) {
4b0ece6f
NH
217 if (PageKsm(page))
218 new = page;
219 else
220 new = page - pvmw.page->index +
221 linear_page_index(vma, pvmw.address);
0697212a 222
616b8371
ZY
223#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
224 /* PMD-mapped THP migration entry */
225 if (!pvmw.pte) {
226 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
227 remove_migration_pmd(&pvmw, new);
228 continue;
229 }
230#endif
231
3fe87967
KS
232 get_page(new);
233 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
234 if (pte_swp_soft_dirty(*pvmw.pte))
235 pte = pte_mksoft_dirty(pte);
0697212a 236
3fe87967
KS
237 /*
238 * Recheck VMA as permissions can change since migration started
239 */
240 entry = pte_to_swp_entry(*pvmw.pte);
241 if (is_write_migration_entry(entry))
242 pte = maybe_mkwrite(pte, vma);
d3cb8bf6 243
df6ad698
JG
244 if (unlikely(is_zone_device_page(new))) {
245 if (is_device_private_page(new)) {
246 entry = make_device_private_entry(new, pte_write(pte));
247 pte = swp_entry_to_pte(entry);
248 } else if (is_device_public_page(new)) {
249 pte = pte_mkdevmap(pte);
250 flush_dcache_page(new);
251 }
a5430dda
JG
252 } else
253 flush_dcache_page(new);
254
3ef8fd7f 255#ifdef CONFIG_HUGETLB_PAGE
3fe87967
KS
256 if (PageHuge(new)) {
257 pte = pte_mkhuge(pte);
258 pte = arch_make_huge_pte(pte, vma, new, 0);
383321ab 259 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
3fe87967
KS
260 if (PageAnon(new))
261 hugepage_add_anon_rmap(new, vma, pvmw.address);
262 else
263 page_dup_rmap(new, true);
383321ab
AK
264 } else
265#endif
266 {
267 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
04e62a29 268
383321ab
AK
269 if (PageAnon(new))
270 page_add_anon_rmap(new, vma, pvmw.address, false);
271 else
272 page_add_file_rmap(new, false);
273 }
3fe87967
KS
274 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
275 mlock_vma_page(new);
276
277 /* No need to invalidate - it was non-present before */
278 update_mmu_cache(vma, pvmw.address, pvmw.pte);
279 }
51afb12b 280
e4b82222 281 return true;
0697212a
CL
282}
283
04e62a29
CL
284/*
285 * Get rid of all migration entries and replace them by
286 * references to the indicated page.
287 */
e388466d 288void remove_migration_ptes(struct page *old, struct page *new, bool locked)
04e62a29 289{
051ac83a
JK
290 struct rmap_walk_control rwc = {
291 .rmap_one = remove_migration_pte,
292 .arg = old,
293 };
294
e388466d
KS
295 if (locked)
296 rmap_walk_locked(new, &rwc);
297 else
298 rmap_walk(new, &rwc);
04e62a29
CL
299}
300
0697212a
CL
301/*
302 * Something used the pte of a page under migration. We need to
303 * get to the page and wait until migration is finished.
304 * When we return from this function the fault will be retried.
0697212a 305 */
e66f17ff 306void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
30dad309 307 spinlock_t *ptl)
0697212a 308{
30dad309 309 pte_t pte;
0697212a
CL
310 swp_entry_t entry;
311 struct page *page;
312
30dad309 313 spin_lock(ptl);
0697212a
CL
314 pte = *ptep;
315 if (!is_swap_pte(pte))
316 goto out;
317
318 entry = pte_to_swp_entry(pte);
319 if (!is_migration_entry(entry))
320 goto out;
321
322 page = migration_entry_to_page(entry);
323
e286781d
NP
324 /*
325 * Once radix-tree replacement of page migration started, page_count
326 * *must* be zero. And, we don't want to call wait_on_page_locked()
327 * against a page without get_page().
328 * So, we use get_page_unless_zero(), here. Even failed, page fault
329 * will occur again.
330 */
331 if (!get_page_unless_zero(page))
332 goto out;
0697212a
CL
333 pte_unmap_unlock(ptep, ptl);
334 wait_on_page_locked(page);
335 put_page(page);
336 return;
337out:
338 pte_unmap_unlock(ptep, ptl);
339}
340
30dad309
NH
341void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
342 unsigned long address)
343{
344 spinlock_t *ptl = pte_lockptr(mm, pmd);
345 pte_t *ptep = pte_offset_map(pmd, address);
346 __migration_entry_wait(mm, ptep, ptl);
347}
348
cb900f41
KS
349void migration_entry_wait_huge(struct vm_area_struct *vma,
350 struct mm_struct *mm, pte_t *pte)
30dad309 351{
cb900f41 352 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
30dad309
NH
353 __migration_entry_wait(mm, pte, ptl);
354}
355
616b8371
ZY
356#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
357void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
358{
359 spinlock_t *ptl;
360 struct page *page;
361
362 ptl = pmd_lock(mm, pmd);
363 if (!is_pmd_migration_entry(*pmd))
364 goto unlock;
365 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
366 if (!get_page_unless_zero(page))
367 goto unlock;
368 spin_unlock(ptl);
369 wait_on_page_locked(page);
370 put_page(page);
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{
42cb14b1
HD
440 struct zone *oldzone, *newzone;
441 int dirty;
8e321fef 442 int expected_count = 1 + extra_count;
7cf9c2c7 443 void **pslot;
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
19fd6231 469 spin_lock_irq(&mapping->tree_lock);
b20a3503 470
7cf9c2c7
NP
471 pslot = radix_tree_lookup_slot(&mapping->page_tree,
472 page_index(page));
b20a3503 473
8e321fef 474 expected_count += 1 + page_has_private(page);
e286781d 475 if (page_count(page) != expected_count ||
29c1f677 476 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
19fd6231 477 spin_unlock_irq(&mapping->tree_lock);
e23ca00b 478 return -EAGAIN;
b20a3503
CL
479 }
480
fe896d18 481 if (!page_ref_freeze(page, expected_count)) {
19fd6231 482 spin_unlock_irq(&mapping->tree_lock);
e286781d
NP
483 return -EAGAIN;
484 }
485
b969c4ab
MG
486 /*
487 * In the async migration case of moving a page with buffers, lock the
488 * buffers using trylock before the mapping is moved. If the mapping
489 * was moved, we later failed to lock the buffers and could not move
490 * the mapping back due to an elevated page count, we would have to
491 * block waiting on other references to be dropped.
492 */
a6bc32b8
MG
493 if (mode == MIGRATE_ASYNC && head &&
494 !buffer_migrate_lock_buffers(head, mode)) {
fe896d18 495 page_ref_unfreeze(page, expected_count);
b969c4ab
MG
496 spin_unlock_irq(&mapping->tree_lock);
497 return -EAGAIN;
498 }
499
b20a3503 500 /*
cf4b769a
HD
501 * Now we know that no one else is looking at the page:
502 * no turning back from here.
b20a3503 503 */
cf4b769a
HD
504 newpage->index = page->index;
505 newpage->mapping = page->mapping;
7cf9c2c7 506 get_page(newpage); /* add cache reference */
6326fec1
NP
507 if (PageSwapBacked(page)) {
508 __SetPageSwapBacked(newpage);
509 if (PageSwapCache(page)) {
510 SetPageSwapCache(newpage);
511 set_page_private(newpage, page_private(page));
512 }
513 } else {
514 VM_BUG_ON_PAGE(PageSwapCache(page), page);
b20a3503
CL
515 }
516
42cb14b1
HD
517 /* Move dirty while page refs frozen and newpage not yet exposed */
518 dirty = PageDirty(page);
519 if (dirty) {
520 ClearPageDirty(page);
521 SetPageDirty(newpage);
522 }
523
6d75f366 524 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
7cf9c2c7
NP
525
526 /*
937a94c9
JG
527 * Drop cache reference from old page by unfreezing
528 * to one less reference.
7cf9c2c7
NP
529 * We know this isn't the last reference.
530 */
fe896d18 531 page_ref_unfreeze(page, expected_count - 1);
7cf9c2c7 532
42cb14b1
HD
533 spin_unlock(&mapping->tree_lock);
534 /* Leave irq disabled to prevent preemption while updating stats */
535
0e8c7d0f
CL
536 /*
537 * If moved to a different zone then also account
538 * the page for that zone. Other VM counters will be
539 * taken care of when we establish references to the
540 * new page and drop references to the old page.
541 *
542 * Note that anonymous pages are accounted for
4b9d0fab 543 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
0e8c7d0f
CL
544 * are mapped to swap space.
545 */
42cb14b1 546 if (newzone != oldzone) {
11fb9989
MG
547 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
548 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
42cb14b1 549 if (PageSwapBacked(page) && !PageSwapCache(page)) {
11fb9989
MG
550 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
551 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
42cb14b1
HD
552 }
553 if (dirty && mapping_cap_account_dirty(mapping)) {
11fb9989 554 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 555 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
11fb9989 556 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
5a1c84b4 557 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
42cb14b1 558 }
4b02108a 559 }
42cb14b1 560 local_irq_enable();
b20a3503 561
78bd5209 562 return MIGRATEPAGE_SUCCESS;
b20a3503 563}
1118dce7 564EXPORT_SYMBOL(migrate_page_move_mapping);
b20a3503 565
290408d4
NH
566/*
567 * The expected number of remaining references is the same as that
568 * of migrate_page_move_mapping().
569 */
570int migrate_huge_page_move_mapping(struct address_space *mapping,
571 struct page *newpage, struct page *page)
572{
573 int expected_count;
574 void **pslot;
575
290408d4
NH
576 spin_lock_irq(&mapping->tree_lock);
577
578 pslot = radix_tree_lookup_slot(&mapping->page_tree,
579 page_index(page));
580
581 expected_count = 2 + page_has_private(page);
582 if (page_count(page) != expected_count ||
29c1f677 583 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
290408d4
NH
584 spin_unlock_irq(&mapping->tree_lock);
585 return -EAGAIN;
586 }
587
fe896d18 588 if (!page_ref_freeze(page, expected_count)) {
290408d4
NH
589 spin_unlock_irq(&mapping->tree_lock);
590 return -EAGAIN;
591 }
592
cf4b769a
HD
593 newpage->index = page->index;
594 newpage->mapping = page->mapping;
6a93ca8f 595
290408d4
NH
596 get_page(newpage);
597
6d75f366 598 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
290408d4 599
fe896d18 600 page_ref_unfreeze(page, expected_count - 1);
290408d4
NH
601
602 spin_unlock_irq(&mapping->tree_lock);
6a93ca8f 603
78bd5209 604 return MIGRATEPAGE_SUCCESS;
290408d4
NH
605}
606
30b0a105
DH
607/*
608 * Gigantic pages are so large that we do not guarantee that page++ pointer
609 * arithmetic will work across the entire page. We need something more
610 * specialized.
611 */
612static void __copy_gigantic_page(struct page *dst, struct page *src,
613 int nr_pages)
614{
615 int i;
616 struct page *dst_base = dst;
617 struct page *src_base = src;
618
619 for (i = 0; i < nr_pages; ) {
620 cond_resched();
621 copy_highpage(dst, src);
622
623 i++;
624 dst = mem_map_next(dst, dst_base, i);
625 src = mem_map_next(src, src_base, i);
626 }
627}
628
629static void copy_huge_page(struct page *dst, struct page *src)
630{
631 int i;
632 int nr_pages;
633
634 if (PageHuge(src)) {
635 /* hugetlbfs page */
636 struct hstate *h = page_hstate(src);
637 nr_pages = pages_per_huge_page(h);
638
639 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
640 __copy_gigantic_page(dst, src, nr_pages);
641 return;
642 }
643 } else {
644 /* thp page */
645 BUG_ON(!PageTransHuge(src));
646 nr_pages = hpage_nr_pages(src);
647 }
648
649 for (i = 0; i < nr_pages; i++) {
650 cond_resched();
651 copy_highpage(dst + i, src + i);
652 }
653}
654
b20a3503
CL
655/*
656 * Copy the page to its new location
657 */
2916ecc0 658void migrate_page_states(struct page *newpage, struct page *page)
b20a3503 659{
7851a45c
RR
660 int cpupid;
661
b20a3503
CL
662 if (PageError(page))
663 SetPageError(newpage);
664 if (PageReferenced(page))
665 SetPageReferenced(newpage);
666 if (PageUptodate(page))
667 SetPageUptodate(newpage);
894bc310 668 if (TestClearPageActive(page)) {
309381fe 669 VM_BUG_ON_PAGE(PageUnevictable(page), page);
b20a3503 670 SetPageActive(newpage);
418b27ef
LS
671 } else if (TestClearPageUnevictable(page))
672 SetPageUnevictable(newpage);
b20a3503
CL
673 if (PageChecked(page))
674 SetPageChecked(newpage);
675 if (PageMappedToDisk(page))
676 SetPageMappedToDisk(newpage);
677
42cb14b1
HD
678 /* Move dirty on pages not done by migrate_page_move_mapping() */
679 if (PageDirty(page))
680 SetPageDirty(newpage);
b20a3503 681
33c3fc71
VD
682 if (page_is_young(page))
683 set_page_young(newpage);
684 if (page_is_idle(page))
685 set_page_idle(newpage);
686
7851a45c
RR
687 /*
688 * Copy NUMA information to the new page, to prevent over-eager
689 * future migrations of this same page.
690 */
691 cpupid = page_cpupid_xchg_last(page, -1);
692 page_cpupid_xchg_last(newpage, cpupid);
693
e9995ef9 694 ksm_migrate_page(newpage, page);
c8d6553b
HD
695 /*
696 * Please do not reorder this without considering how mm/ksm.c's
697 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
698 */
b3b3a99c
NH
699 if (PageSwapCache(page))
700 ClearPageSwapCache(page);
b20a3503
CL
701 ClearPagePrivate(page);
702 set_page_private(page, 0);
b20a3503
CL
703
704 /*
705 * If any waiters have accumulated on the new page then
706 * wake them up.
707 */
708 if (PageWriteback(newpage))
709 end_page_writeback(newpage);
d435edca
VB
710
711 copy_page_owner(page, newpage);
74485cf2
JW
712
713 mem_cgroup_migrate(page, newpage);
b20a3503 714}
2916ecc0
JG
715EXPORT_SYMBOL(migrate_page_states);
716
717void migrate_page_copy(struct page *newpage, struct page *page)
718{
719 if (PageHuge(page) || PageTransHuge(page))
720 copy_huge_page(newpage, page);
721 else
722 copy_highpage(newpage, page);
723
724 migrate_page_states(newpage, page);
725}
1118dce7 726EXPORT_SYMBOL(migrate_page_copy);
b20a3503 727
1d8b85cc
CL
728/************************************************************
729 * Migration functions
730 ***********************************************************/
731
b20a3503 732/*
bda807d4 733 * Common logic to directly migrate a single LRU page suitable for
266cf658 734 * pages that do not use PagePrivate/PagePrivate2.
b20a3503
CL
735 *
736 * Pages are locked upon entry and exit.
737 */
2d1db3b1 738int migrate_page(struct address_space *mapping,
a6bc32b8
MG
739 struct page *newpage, struct page *page,
740 enum migrate_mode mode)
b20a3503
CL
741{
742 int rc;
743
744 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
745
8e321fef 746 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
b20a3503 747
78bd5209 748 if (rc != MIGRATEPAGE_SUCCESS)
b20a3503
CL
749 return rc;
750
2916ecc0
JG
751 if (mode != MIGRATE_SYNC_NO_COPY)
752 migrate_page_copy(newpage, page);
753 else
754 migrate_page_states(newpage, page);
78bd5209 755 return MIGRATEPAGE_SUCCESS;
b20a3503
CL
756}
757EXPORT_SYMBOL(migrate_page);
758
9361401e 759#ifdef CONFIG_BLOCK
1d8b85cc
CL
760/*
761 * Migration function for pages with buffers. This function can only be used
762 * if the underlying filesystem guarantees that no other references to "page"
763 * exist.
764 */
2d1db3b1 765int buffer_migrate_page(struct address_space *mapping,
a6bc32b8 766 struct page *newpage, struct page *page, enum migrate_mode mode)
1d8b85cc 767{
1d8b85cc
CL
768 struct buffer_head *bh, *head;
769 int rc;
770
1d8b85cc 771 if (!page_has_buffers(page))
a6bc32b8 772 return migrate_page(mapping, newpage, page, mode);
1d8b85cc
CL
773
774 head = page_buffers(page);
775
8e321fef 776 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
1d8b85cc 777
78bd5209 778 if (rc != MIGRATEPAGE_SUCCESS)
1d8b85cc
CL
779 return rc;
780
b969c4ab
MG
781 /*
782 * In the async case, migrate_page_move_mapping locked the buffers
783 * with an IRQ-safe spinlock held. In the sync case, the buffers
784 * need to be locked now
785 */
a6bc32b8
MG
786 if (mode != MIGRATE_ASYNC)
787 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
1d8b85cc
CL
788
789 ClearPagePrivate(page);
790 set_page_private(newpage, page_private(page));
791 set_page_private(page, 0);
792 put_page(page);
793 get_page(newpage);
794
795 bh = head;
796 do {
797 set_bh_page(bh, newpage, bh_offset(bh));
798 bh = bh->b_this_page;
799
800 } while (bh != head);
801
802 SetPagePrivate(newpage);
803
2916ecc0
JG
804 if (mode != MIGRATE_SYNC_NO_COPY)
805 migrate_page_copy(newpage, page);
806 else
807 migrate_page_states(newpage, page);
1d8b85cc
CL
808
809 bh = head;
810 do {
811 unlock_buffer(bh);
2916ecc0 812 put_bh(bh);
1d8b85cc
CL
813 bh = bh->b_this_page;
814
815 } while (bh != head);
816
78bd5209 817 return MIGRATEPAGE_SUCCESS;
1d8b85cc
CL
818}
819EXPORT_SYMBOL(buffer_migrate_page);
9361401e 820#endif
1d8b85cc 821
04e62a29
CL
822/*
823 * Writeback a page to clean the dirty state
824 */
825static int writeout(struct address_space *mapping, struct page *page)
8351a6e4 826{
04e62a29
CL
827 struct writeback_control wbc = {
828 .sync_mode = WB_SYNC_NONE,
829 .nr_to_write = 1,
830 .range_start = 0,
831 .range_end = LLONG_MAX,
04e62a29
CL
832 .for_reclaim = 1
833 };
834 int rc;
835
836 if (!mapping->a_ops->writepage)
837 /* No write method for the address space */
838 return -EINVAL;
839
840 if (!clear_page_dirty_for_io(page))
841 /* Someone else already triggered a write */
842 return -EAGAIN;
843
8351a6e4 844 /*
04e62a29
CL
845 * A dirty page may imply that the underlying filesystem has
846 * the page on some queue. So the page must be clean for
847 * migration. Writeout may mean we loose the lock and the
848 * page state is no longer what we checked for earlier.
849 * At this point we know that the migration attempt cannot
850 * be successful.
8351a6e4 851 */
e388466d 852 remove_migration_ptes(page, page, false);
8351a6e4 853
04e62a29 854 rc = mapping->a_ops->writepage(page, &wbc);
8351a6e4 855
04e62a29
CL
856 if (rc != AOP_WRITEPAGE_ACTIVATE)
857 /* unlocked. Relock */
858 lock_page(page);
859
bda8550d 860 return (rc < 0) ? -EIO : -EAGAIN;
04e62a29
CL
861}
862
863/*
864 * Default handling if a filesystem does not provide a migration function.
865 */
866static int fallback_migrate_page(struct address_space *mapping,
a6bc32b8 867 struct page *newpage, struct page *page, enum migrate_mode mode)
04e62a29 868{
b969c4ab 869 if (PageDirty(page)) {
a6bc32b8 870 /* Only writeback pages in full synchronous migration */
2916ecc0
JG
871 switch (mode) {
872 case MIGRATE_SYNC:
873 case MIGRATE_SYNC_NO_COPY:
874 break;
875 default:
b969c4ab 876 return -EBUSY;
2916ecc0 877 }
04e62a29 878 return writeout(mapping, page);
b969c4ab 879 }
8351a6e4
CL
880
881 /*
882 * Buffers may be managed in a filesystem specific way.
883 * We must have no buffers or drop them.
884 */
266cf658 885 if (page_has_private(page) &&
8351a6e4
CL
886 !try_to_release_page(page, GFP_KERNEL))
887 return -EAGAIN;
888
a6bc32b8 889 return migrate_page(mapping, newpage, page, mode);
8351a6e4
CL
890}
891
e24f0b8f
CL
892/*
893 * Move a page to a newly allocated page
894 * The page is locked and all ptes have been successfully removed.
895 *
896 * The new page will have replaced the old page if this function
897 * is successful.
894bc310
LS
898 *
899 * Return value:
900 * < 0 - error code
78bd5209 901 * MIGRATEPAGE_SUCCESS - success
e24f0b8f 902 */
3fe2011f 903static int move_to_new_page(struct page *newpage, struct page *page,
5c3f9a67 904 enum migrate_mode mode)
e24f0b8f
CL
905{
906 struct address_space *mapping;
bda807d4
MK
907 int rc = -EAGAIN;
908 bool is_lru = !__PageMovable(page);
e24f0b8f 909
7db7671f
HD
910 VM_BUG_ON_PAGE(!PageLocked(page), page);
911 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
e24f0b8f 912
e24f0b8f 913 mapping = page_mapping(page);
bda807d4
MK
914
915 if (likely(is_lru)) {
916 if (!mapping)
917 rc = migrate_page(mapping, newpage, page, mode);
918 else if (mapping->a_ops->migratepage)
919 /*
920 * Most pages have a mapping and most filesystems
921 * provide a migratepage callback. Anonymous pages
922 * are part of swap space which also has its own
923 * migratepage callback. This is the most common path
924 * for page migration.
925 */
926 rc = mapping->a_ops->migratepage(mapping, newpage,
927 page, mode);
928 else
929 rc = fallback_migrate_page(mapping, newpage,
930 page, mode);
931 } else {
e24f0b8f 932 /*
bda807d4
MK
933 * In case of non-lru page, it could be released after
934 * isolation step. In that case, we shouldn't try migration.
e24f0b8f 935 */
bda807d4
MK
936 VM_BUG_ON_PAGE(!PageIsolated(page), page);
937 if (!PageMovable(page)) {
938 rc = MIGRATEPAGE_SUCCESS;
939 __ClearPageIsolated(page);
940 goto out;
941 }
942
943 rc = mapping->a_ops->migratepage(mapping, newpage,
944 page, mode);
945 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
946 !PageIsolated(page));
947 }
e24f0b8f 948
5c3f9a67
HD
949 /*
950 * When successful, old pagecache page->mapping must be cleared before
951 * page is freed; but stats require that PageAnon be left as PageAnon.
952 */
953 if (rc == MIGRATEPAGE_SUCCESS) {
bda807d4
MK
954 if (__PageMovable(page)) {
955 VM_BUG_ON_PAGE(!PageIsolated(page), page);
956
957 /*
958 * We clear PG_movable under page_lock so any compactor
959 * cannot try to migrate this page.
960 */
961 __ClearPageIsolated(page);
962 }
963
964 /*
965 * Anonymous and movable page->mapping will be cleard by
966 * free_pages_prepare so don't reset it here for keeping
967 * the type to work PageAnon, for example.
968 */
969 if (!PageMappingFlags(page))
5c3f9a67 970 page->mapping = NULL;
3fe2011f 971 }
bda807d4 972out:
e24f0b8f
CL
973 return rc;
974}
975
0dabec93 976static int __unmap_and_move(struct page *page, struct page *newpage,
9c620e2b 977 int force, enum migrate_mode mode)
e24f0b8f 978{
0dabec93 979 int rc = -EAGAIN;
2ebba6b7 980 int page_was_mapped = 0;
3f6c8272 981 struct anon_vma *anon_vma = NULL;
bda807d4 982 bool is_lru = !__PageMovable(page);
95a402c3 983
529ae9aa 984 if (!trylock_page(page)) {
a6bc32b8 985 if (!force || mode == MIGRATE_ASYNC)
0dabec93 986 goto out;
3e7d3449
MG
987
988 /*
989 * It's not safe for direct compaction to call lock_page.
990 * For example, during page readahead pages are added locked
991 * to the LRU. Later, when the IO completes the pages are
992 * marked uptodate and unlocked. However, the queueing
993 * could be merging multiple pages for one bio (e.g.
994 * mpage_readpages). If an allocation happens for the
995 * second or third page, the process can end up locking
996 * the same page twice and deadlocking. Rather than
997 * trying to be clever about what pages can be locked,
998 * avoid the use of lock_page for direct compaction
999 * altogether.
1000 */
1001 if (current->flags & PF_MEMALLOC)
0dabec93 1002 goto out;
3e7d3449 1003
e24f0b8f
CL
1004 lock_page(page);
1005 }
1006
1007 if (PageWriteback(page)) {
11bc82d6 1008 /*
fed5b64a 1009 * Only in the case of a full synchronous migration is it
a6bc32b8
MG
1010 * necessary to wait for PageWriteback. In the async case,
1011 * the retry loop is too short and in the sync-light case,
1012 * the overhead of stalling is too much
11bc82d6 1013 */
2916ecc0
JG
1014 switch (mode) {
1015 case MIGRATE_SYNC:
1016 case MIGRATE_SYNC_NO_COPY:
1017 break;
1018 default:
11bc82d6 1019 rc = -EBUSY;
0a31bc97 1020 goto out_unlock;
11bc82d6
AA
1021 }
1022 if (!force)
0a31bc97 1023 goto out_unlock;
e24f0b8f
CL
1024 wait_on_page_writeback(page);
1025 }
03f15c86 1026
e24f0b8f 1027 /*
dc386d4d
KH
1028 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1029 * we cannot notice that anon_vma is freed while we migrates a page.
1ce82b69 1030 * This get_anon_vma() delays freeing anon_vma pointer until the end
dc386d4d 1031 * of migration. File cache pages are no problem because of page_lock()
989f89c5
KH
1032 * File Caches may use write_page() or lock_page() in migration, then,
1033 * just care Anon page here.
03f15c86
HD
1034 *
1035 * Only page_get_anon_vma() understands the subtleties of
1036 * getting a hold on an anon_vma from outside one of its mms.
1037 * But if we cannot get anon_vma, then we won't need it anyway,
1038 * because that implies that the anon page is no longer mapped
1039 * (and cannot be remapped so long as we hold the page lock).
dc386d4d 1040 */
03f15c86 1041 if (PageAnon(page) && !PageKsm(page))
746b18d4 1042 anon_vma = page_get_anon_vma(page);
62e1c553 1043
7db7671f
HD
1044 /*
1045 * Block others from accessing the new page when we get around to
1046 * establishing additional references. We are usually the only one
1047 * holding a reference to newpage at this point. We used to have a BUG
1048 * here if trylock_page(newpage) fails, but would like to allow for
1049 * cases where there might be a race with the previous use of newpage.
1050 * This is much like races on refcount of oldpage: just don't BUG().
1051 */
1052 if (unlikely(!trylock_page(newpage)))
1053 goto out_unlock;
1054
bda807d4
MK
1055 if (unlikely(!is_lru)) {
1056 rc = move_to_new_page(newpage, page, mode);
1057 goto out_unlock_both;
1058 }
1059
dc386d4d 1060 /*
62e1c553
SL
1061 * Corner case handling:
1062 * 1. When a new swap-cache page is read into, it is added to the LRU
1063 * and treated as swapcache but it has no rmap yet.
1064 * Calling try_to_unmap() against a page->mapping==NULL page will
1065 * trigger a BUG. So handle it here.
1066 * 2. An orphaned page (see truncate_complete_page) might have
1067 * fs-private metadata. The page can be picked up due to memory
1068 * offlining. Everywhere else except page reclaim, the page is
1069 * invisible to the vm, so the page can not be migrated. So try to
1070 * free the metadata, so the page can be freed.
e24f0b8f 1071 */
62e1c553 1072 if (!page->mapping) {
309381fe 1073 VM_BUG_ON_PAGE(PageAnon(page), page);
1ce82b69 1074 if (page_has_private(page)) {
62e1c553 1075 try_to_free_buffers(page);
7db7671f 1076 goto out_unlock_both;
62e1c553 1077 }
7db7671f
HD
1078 } else if (page_mapped(page)) {
1079 /* Establish migration ptes */
03f15c86
HD
1080 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1081 page);
2ebba6b7 1082 try_to_unmap(page,
da1b13cc 1083 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
2ebba6b7
HD
1084 page_was_mapped = 1;
1085 }
dc386d4d 1086
e6a1530d 1087 if (!page_mapped(page))
5c3f9a67 1088 rc = move_to_new_page(newpage, page, mode);
e24f0b8f 1089
5c3f9a67
HD
1090 if (page_was_mapped)
1091 remove_migration_ptes(page,
e388466d 1092 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
3f6c8272 1093
7db7671f
HD
1094out_unlock_both:
1095 unlock_page(newpage);
1096out_unlock:
3f6c8272 1097 /* Drop an anon_vma reference if we took one */
76545066 1098 if (anon_vma)
9e60109f 1099 put_anon_vma(anon_vma);
e24f0b8f 1100 unlock_page(page);
0dabec93 1101out:
c6c919eb
MK
1102 /*
1103 * If migration is successful, decrease refcount of the newpage
1104 * which will not free the page because new page owner increased
1105 * refcounter. As well, if it is LRU page, add the page to LRU
1106 * list in here.
1107 */
1108 if (rc == MIGRATEPAGE_SUCCESS) {
b1123ea6 1109 if (unlikely(__PageMovable(newpage)))
c6c919eb
MK
1110 put_page(newpage);
1111 else
1112 putback_lru_page(newpage);
1113 }
1114
0dabec93
MK
1115 return rc;
1116}
95a402c3 1117
ef2a5153
GU
1118/*
1119 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1120 * around it.
1121 */
1122#if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1123#define ICE_noinline noinline
1124#else
1125#define ICE_noinline
1126#endif
1127
0dabec93
MK
1128/*
1129 * Obtain the lock on page, remove all ptes and migrate the page
1130 * to the newly allocated page in newpage.
1131 */
ef2a5153
GU
1132static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1133 free_page_t put_new_page,
1134 unsigned long private, struct page *page,
add05cec
NH
1135 int force, enum migrate_mode mode,
1136 enum migrate_reason reason)
0dabec93 1137{
2def7424 1138 int rc = MIGRATEPAGE_SUCCESS;
0dabec93 1139 int *result = NULL;
2def7424 1140 struct page *newpage;
0dabec93 1141
2def7424 1142 newpage = get_new_page(page, private, &result);
0dabec93
MK
1143 if (!newpage)
1144 return -ENOMEM;
1145
1146 if (page_count(page) == 1) {
1147 /* page was freed from under us. So we are done. */
c6c919eb
MK
1148 ClearPageActive(page);
1149 ClearPageUnevictable(page);
bda807d4
MK
1150 if (unlikely(__PageMovable(page))) {
1151 lock_page(page);
1152 if (!PageMovable(page))
1153 __ClearPageIsolated(page);
1154 unlock_page(page);
1155 }
c6c919eb
MK
1156 if (put_new_page)
1157 put_new_page(newpage, private);
1158 else
1159 put_page(newpage);
0dabec93
MK
1160 goto out;
1161 }
1162
616b8371 1163 if (unlikely(PageTransHuge(page) && !PageTransHuge(newpage))) {
4d2fa965
KS
1164 lock_page(page);
1165 rc = split_huge_page(page);
1166 unlock_page(page);
1167 if (rc)
0dabec93 1168 goto out;
4d2fa965 1169 }
0dabec93 1170
9c620e2b 1171 rc = __unmap_and_move(page, newpage, force, mode);
c6c919eb 1172 if (rc == MIGRATEPAGE_SUCCESS)
7cd12b4a 1173 set_page_owner_migrate_reason(newpage, reason);
bf6bddf1 1174
0dabec93 1175out:
e24f0b8f 1176 if (rc != -EAGAIN) {
0dabec93
MK
1177 /*
1178 * A page that has been migrated has all references
1179 * removed and will be freed. A page that has not been
1180 * migrated will have kepts its references and be
1181 * restored.
1182 */
1183 list_del(&page->lru);
6afcf8ef
ML
1184
1185 /*
1186 * Compaction can migrate also non-LRU pages which are
1187 * not accounted to NR_ISOLATED_*. They can be recognized
1188 * as __PageMovable
1189 */
1190 if (likely(!__PageMovable(page)))
e8db67eb
NH
1191 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1192 page_is_file_cache(page), -hpage_nr_pages(page));
c6c919eb
MK
1193 }
1194
1195 /*
1196 * If migration is successful, releases reference grabbed during
1197 * isolation. Otherwise, restore the page to right list unless
1198 * we want to retry.
1199 */
1200 if (rc == MIGRATEPAGE_SUCCESS) {
1201 put_page(page);
1202 if (reason == MR_MEMORY_FAILURE) {
d7e69488 1203 /*
c6c919eb
MK
1204 * Set PG_HWPoison on just freed page
1205 * intentionally. Although it's rather weird,
1206 * it's how HWPoison flag works at the moment.
d7e69488 1207 */
da1b13cc
WL
1208 if (!test_set_page_hwpoison(page))
1209 num_poisoned_pages_inc();
c6c919eb
MK
1210 }
1211 } else {
bda807d4
MK
1212 if (rc != -EAGAIN) {
1213 if (likely(!__PageMovable(page))) {
1214 putback_lru_page(page);
1215 goto put_new;
1216 }
1217
1218 lock_page(page);
1219 if (PageMovable(page))
1220 putback_movable_page(page);
1221 else
1222 __ClearPageIsolated(page);
1223 unlock_page(page);
1224 put_page(page);
1225 }
1226put_new:
c6c919eb
MK
1227 if (put_new_page)
1228 put_new_page(newpage, private);
1229 else
1230 put_page(newpage);
e24f0b8f 1231 }
68711a74 1232
742755a1
CL
1233 if (result) {
1234 if (rc)
1235 *result = rc;
1236 else
1237 *result = page_to_nid(newpage);
1238 }
e24f0b8f
CL
1239 return rc;
1240}
1241
290408d4
NH
1242/*
1243 * Counterpart of unmap_and_move_page() for hugepage migration.
1244 *
1245 * This function doesn't wait the completion of hugepage I/O
1246 * because there is no race between I/O and migration for hugepage.
1247 * Note that currently hugepage I/O occurs only in direct I/O
1248 * where no lock is held and PG_writeback is irrelevant,
1249 * and writeback status of all subpages are counted in the reference
1250 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1251 * under direct I/O, the reference of the head page is 512 and a bit more.)
1252 * This means that when we try to migrate hugepage whose subpages are
1253 * doing direct I/O, some references remain after try_to_unmap() and
1254 * hugepage migration fails without data corruption.
1255 *
1256 * There is also no race when direct I/O is issued on the page under migration,
1257 * because then pte is replaced with migration swap entry and direct I/O code
1258 * will wait in the page fault for migration to complete.
1259 */
1260static int unmap_and_move_huge_page(new_page_t get_new_page,
68711a74
DR
1261 free_page_t put_new_page, unsigned long private,
1262 struct page *hpage, int force,
7cd12b4a 1263 enum migrate_mode mode, int reason)
290408d4 1264{
2def7424 1265 int rc = -EAGAIN;
290408d4 1266 int *result = NULL;
2ebba6b7 1267 int page_was_mapped = 0;
32665f2b 1268 struct page *new_hpage;
290408d4
NH
1269 struct anon_vma *anon_vma = NULL;
1270
83467efb
NH
1271 /*
1272 * Movability of hugepages depends on architectures and hugepage size.
1273 * This check is necessary because some callers of hugepage migration
1274 * like soft offline and memory hotremove don't walk through page
1275 * tables or check whether the hugepage is pmd-based or not before
1276 * kicking migration.
1277 */
100873d7 1278 if (!hugepage_migration_supported(page_hstate(hpage))) {
32665f2b 1279 putback_active_hugepage(hpage);
83467efb 1280 return -ENOSYS;
32665f2b 1281 }
83467efb 1282
32665f2b 1283 new_hpage = get_new_page(hpage, private, &result);
290408d4
NH
1284 if (!new_hpage)
1285 return -ENOMEM;
1286
290408d4 1287 if (!trylock_page(hpage)) {
2916ecc0 1288 if (!force)
290408d4 1289 goto out;
2916ecc0
JG
1290 switch (mode) {
1291 case MIGRATE_SYNC:
1292 case MIGRATE_SYNC_NO_COPY:
1293 break;
1294 default:
1295 goto out;
1296 }
290408d4
NH
1297 lock_page(hpage);
1298 }
1299
746b18d4
PZ
1300 if (PageAnon(hpage))
1301 anon_vma = page_get_anon_vma(hpage);
290408d4 1302
7db7671f
HD
1303 if (unlikely(!trylock_page(new_hpage)))
1304 goto put_anon;
1305
2ebba6b7
HD
1306 if (page_mapped(hpage)) {
1307 try_to_unmap(hpage,
1308 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1309 page_was_mapped = 1;
1310 }
290408d4
NH
1311
1312 if (!page_mapped(hpage))
5c3f9a67 1313 rc = move_to_new_page(new_hpage, hpage, mode);
290408d4 1314
5c3f9a67
HD
1315 if (page_was_mapped)
1316 remove_migration_ptes(hpage,
e388466d 1317 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
290408d4 1318
7db7671f
HD
1319 unlock_page(new_hpage);
1320
1321put_anon:
fd4a4663 1322 if (anon_vma)
9e60109f 1323 put_anon_vma(anon_vma);
8e6ac7fa 1324
2def7424 1325 if (rc == MIGRATEPAGE_SUCCESS) {
8e6ac7fa 1326 hugetlb_cgroup_migrate(hpage, new_hpage);
2def7424 1327 put_new_page = NULL;
7cd12b4a 1328 set_page_owner_migrate_reason(new_hpage, reason);
2def7424 1329 }
8e6ac7fa 1330
290408d4 1331 unlock_page(hpage);
09761333 1332out:
b8ec1cee
NH
1333 if (rc != -EAGAIN)
1334 putback_active_hugepage(hpage);
c3114a84
AK
1335 if (reason == MR_MEMORY_FAILURE && !test_set_page_hwpoison(hpage))
1336 num_poisoned_pages_inc();
68711a74
DR
1337
1338 /*
1339 * If migration was not successful and there's a freeing callback, use
1340 * it. Otherwise, put_page() will drop the reference grabbed during
1341 * isolation.
1342 */
2def7424 1343 if (put_new_page)
68711a74
DR
1344 put_new_page(new_hpage, private);
1345 else
3aaa76e1 1346 putback_active_hugepage(new_hpage);
68711a74 1347
290408d4
NH
1348 if (result) {
1349 if (rc)
1350 *result = rc;
1351 else
1352 *result = page_to_nid(new_hpage);
1353 }
1354 return rc;
1355}
1356
b20a3503 1357/*
c73e5c9c
SB
1358 * migrate_pages - migrate the pages specified in a list, to the free pages
1359 * supplied as the target for the page migration
b20a3503 1360 *
c73e5c9c
SB
1361 * @from: The list of pages to be migrated.
1362 * @get_new_page: The function used to allocate free pages to be used
1363 * as the target of the page migration.
68711a74
DR
1364 * @put_new_page: The function used to free target pages if migration
1365 * fails, or NULL if no special handling is necessary.
c73e5c9c
SB
1366 * @private: Private data to be passed on to get_new_page()
1367 * @mode: The migration mode that specifies the constraints for
1368 * page migration, if any.
1369 * @reason: The reason for page migration.
b20a3503 1370 *
c73e5c9c
SB
1371 * The function returns after 10 attempts or if no pages are movable any more
1372 * because the list has become empty or no retryable pages exist any more.
14e0f9bc 1373 * The caller should call putback_movable_pages() to return pages to the LRU
28bd6578 1374 * or free list only if ret != 0.
b20a3503 1375 *
c73e5c9c 1376 * Returns the number of pages that were not migrated, or an error code.
b20a3503 1377 */
9c620e2b 1378int migrate_pages(struct list_head *from, new_page_t get_new_page,
68711a74
DR
1379 free_page_t put_new_page, unsigned long private,
1380 enum migrate_mode mode, int reason)
b20a3503 1381{
e24f0b8f 1382 int retry = 1;
b20a3503 1383 int nr_failed = 0;
5647bc29 1384 int nr_succeeded = 0;
b20a3503
CL
1385 int pass = 0;
1386 struct page *page;
1387 struct page *page2;
1388 int swapwrite = current->flags & PF_SWAPWRITE;
1389 int rc;
1390
1391 if (!swapwrite)
1392 current->flags |= PF_SWAPWRITE;
1393
e24f0b8f
CL
1394 for(pass = 0; pass < 10 && retry; pass++) {
1395 retry = 0;
b20a3503 1396
e24f0b8f 1397 list_for_each_entry_safe(page, page2, from, lru) {
e24f0b8f 1398 cond_resched();
2d1db3b1 1399
31caf665
NH
1400 if (PageHuge(page))
1401 rc = unmap_and_move_huge_page(get_new_page,
68711a74 1402 put_new_page, private, page,
7cd12b4a 1403 pass > 2, mode, reason);
31caf665 1404 else
68711a74 1405 rc = unmap_and_move(get_new_page, put_new_page,
add05cec
NH
1406 private, page, pass > 2, mode,
1407 reason);
2d1db3b1 1408
e24f0b8f 1409 switch(rc) {
95a402c3 1410 case -ENOMEM:
dfef2ef4 1411 nr_failed++;
95a402c3 1412 goto out;
e24f0b8f 1413 case -EAGAIN:
2d1db3b1 1414 retry++;
e24f0b8f 1415 break;
78bd5209 1416 case MIGRATEPAGE_SUCCESS:
5647bc29 1417 nr_succeeded++;
e24f0b8f
CL
1418 break;
1419 default:
354a3363
NH
1420 /*
1421 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1422 * unlike -EAGAIN case, the failed page is
1423 * removed from migration page list and not
1424 * retried in the next outer loop.
1425 */
2d1db3b1 1426 nr_failed++;
e24f0b8f 1427 break;
2d1db3b1 1428 }
b20a3503
CL
1429 }
1430 }
f2f81fb2
VB
1431 nr_failed += retry;
1432 rc = nr_failed;
95a402c3 1433out:
5647bc29
MG
1434 if (nr_succeeded)
1435 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1436 if (nr_failed)
1437 count_vm_events(PGMIGRATE_FAIL, nr_failed);
7b2a2d4a
MG
1438 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1439
b20a3503
CL
1440 if (!swapwrite)
1441 current->flags &= ~PF_SWAPWRITE;
1442
78bd5209 1443 return rc;
b20a3503 1444}
95a402c3 1445
742755a1
CL
1446#ifdef CONFIG_NUMA
1447/*
1448 * Move a list of individual pages
1449 */
1450struct page_to_node {
1451 unsigned long addr;
1452 struct page *page;
1453 int node;
1454 int status;
1455};
1456
1457static struct page *new_page_node(struct page *p, unsigned long private,
1458 int **result)
1459{
1460 struct page_to_node *pm = (struct page_to_node *)private;
1461
1462 while (pm->node != MAX_NUMNODES && pm->page != p)
1463 pm++;
1464
1465 if (pm->node == MAX_NUMNODES)
1466 return NULL;
1467
1468 *result = &pm->status;
1469
e632a938
NH
1470 if (PageHuge(p))
1471 return alloc_huge_page_node(page_hstate(compound_head(p)),
1472 pm->node);
e8db67eb
NH
1473 else if (thp_migration_supported() && PageTransHuge(p)) {
1474 struct page *thp;
1475
1476 thp = alloc_pages_node(pm->node,
1477 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
1478 HPAGE_PMD_ORDER);
1479 if (!thp)
1480 return NULL;
1481 prep_transhuge_page(thp);
1482 return thp;
1483 } else
96db800f 1484 return __alloc_pages_node(pm->node,
e97ca8e5 1485 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
742755a1
CL
1486}
1487
1488/*
1489 * Move a set of pages as indicated in the pm array. The addr
1490 * field must be set to the virtual address of the page to be moved
1491 * and the node number must contain a valid target node.
5e9a0f02 1492 * The pm array ends with node = MAX_NUMNODES.
742755a1 1493 */
5e9a0f02
BG
1494static int do_move_page_to_node_array(struct mm_struct *mm,
1495 struct page_to_node *pm,
1496 int migrate_all)
742755a1
CL
1497{
1498 int err;
1499 struct page_to_node *pp;
1500 LIST_HEAD(pagelist);
1501
1502 down_read(&mm->mmap_sem);
1503
1504 /*
1505 * Build a list of pages to migrate
1506 */
742755a1
CL
1507 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1508 struct vm_area_struct *vma;
1509 struct page *page;
e8db67eb
NH
1510 struct page *head;
1511 unsigned int follflags;
742755a1 1512
742755a1
CL
1513 err = -EFAULT;
1514 vma = find_vma(mm, pp->addr);
70384dc6 1515 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
742755a1
CL
1516 goto set_status;
1517
d899844e 1518 /* FOLL_DUMP to ignore special (like zero) pages */
e8db67eb
NH
1519 follflags = FOLL_GET | FOLL_DUMP;
1520 if (!thp_migration_supported())
1521 follflags |= FOLL_SPLIT;
1522 page = follow_page(vma, pp->addr, follflags);
89f5b7da
LT
1523
1524 err = PTR_ERR(page);
1525 if (IS_ERR(page))
1526 goto set_status;
1527
742755a1
CL
1528 err = -ENOENT;
1529 if (!page)
1530 goto set_status;
1531
742755a1
CL
1532 err = page_to_nid(page);
1533
1534 if (err == pp->node)
1535 /*
1536 * Node already in the right place
1537 */
1538 goto put_and_set;
1539
1540 err = -EACCES;
1541 if (page_mapcount(page) > 1 &&
1542 !migrate_all)
1543 goto put_and_set;
1544
e632a938 1545 if (PageHuge(page)) {
e8db67eb 1546 if (PageHead(page)) {
e66f17ff 1547 isolate_huge_page(page, &pagelist);
e8db67eb
NH
1548 err = 0;
1549 pp->page = page;
1550 }
e632a938
NH
1551 goto put_and_set;
1552 }
1553
e8db67eb
NH
1554 pp->page = compound_head(page);
1555 head = compound_head(page);
1556 err = isolate_lru_page(head);
6d9c285a 1557 if (!err) {
e8db67eb
NH
1558 list_add_tail(&head->lru, &pagelist);
1559 mod_node_page_state(page_pgdat(head),
1560 NR_ISOLATED_ANON + page_is_file_cache(head),
1561 hpage_nr_pages(head));
6d9c285a 1562 }
742755a1
CL
1563put_and_set:
1564 /*
1565 * Either remove the duplicate refcount from
1566 * isolate_lru_page() or drop the page ref if it was
1567 * not isolated.
1568 */
1569 put_page(page);
1570set_status:
1571 pp->status = err;
1572 }
1573
e78bbfa8 1574 err = 0;
cf608ac1 1575 if (!list_empty(&pagelist)) {
68711a74 1576 err = migrate_pages(&pagelist, new_page_node, NULL,
9c620e2b 1577 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
cf608ac1 1578 if (err)
e632a938 1579 putback_movable_pages(&pagelist);
cf608ac1 1580 }
742755a1
CL
1581
1582 up_read(&mm->mmap_sem);
1583 return err;
1584}
1585
5e9a0f02
BG
1586/*
1587 * Migrate an array of page address onto an array of nodes and fill
1588 * the corresponding array of status.
1589 */
3268c63e 1590static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
5e9a0f02
BG
1591 unsigned long nr_pages,
1592 const void __user * __user *pages,
1593 const int __user *nodes,
1594 int __user *status, int flags)
1595{
3140a227 1596 struct page_to_node *pm;
3140a227
BG
1597 unsigned long chunk_nr_pages;
1598 unsigned long chunk_start;
1599 int err;
5e9a0f02 1600
3140a227
BG
1601 err = -ENOMEM;
1602 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1603 if (!pm)
5e9a0f02 1604 goto out;
35282a2d
BG
1605
1606 migrate_prep();
1607
5e9a0f02 1608 /*
3140a227
BG
1609 * Store a chunk of page_to_node array in a page,
1610 * but keep the last one as a marker
5e9a0f02 1611 */
3140a227 1612 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
5e9a0f02 1613
3140a227
BG
1614 for (chunk_start = 0;
1615 chunk_start < nr_pages;
1616 chunk_start += chunk_nr_pages) {
1617 int j;
5e9a0f02 1618
3140a227
BG
1619 if (chunk_start + chunk_nr_pages > nr_pages)
1620 chunk_nr_pages = nr_pages - chunk_start;
1621
1622 /* fill the chunk pm with addrs and nodes from user-space */
1623 for (j = 0; j < chunk_nr_pages; j++) {
1624 const void __user *p;
5e9a0f02
BG
1625 int node;
1626
3140a227
BG
1627 err = -EFAULT;
1628 if (get_user(p, pages + j + chunk_start))
1629 goto out_pm;
1630 pm[j].addr = (unsigned long) p;
1631
1632 if (get_user(node, nodes + j + chunk_start))
5e9a0f02
BG
1633 goto out_pm;
1634
1635 err = -ENODEV;
6f5a55f1
LT
1636 if (node < 0 || node >= MAX_NUMNODES)
1637 goto out_pm;
1638
389162c2 1639 if (!node_state(node, N_MEMORY))
5e9a0f02
BG
1640 goto out_pm;
1641
1642 err = -EACCES;
1643 if (!node_isset(node, task_nodes))
1644 goto out_pm;
1645
3140a227
BG
1646 pm[j].node = node;
1647 }
1648
1649 /* End marker for this chunk */
1650 pm[chunk_nr_pages].node = MAX_NUMNODES;
1651
1652 /* Migrate this chunk */
1653 err = do_move_page_to_node_array(mm, pm,
1654 flags & MPOL_MF_MOVE_ALL);
1655 if (err < 0)
1656 goto out_pm;
5e9a0f02 1657
5e9a0f02 1658 /* Return status information */
3140a227
BG
1659 for (j = 0; j < chunk_nr_pages; j++)
1660 if (put_user(pm[j].status, status + j + chunk_start)) {
5e9a0f02 1661 err = -EFAULT;
3140a227
BG
1662 goto out_pm;
1663 }
1664 }
1665 err = 0;
5e9a0f02
BG
1666
1667out_pm:
3140a227 1668 free_page((unsigned long)pm);
5e9a0f02
BG
1669out:
1670 return err;
1671}
1672
742755a1 1673/*
2f007e74 1674 * Determine the nodes of an array of pages and store it in an array of status.
742755a1 1675 */
80bba129
BG
1676static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1677 const void __user **pages, int *status)
742755a1 1678{
2f007e74 1679 unsigned long i;
2f007e74 1680
742755a1
CL
1681 down_read(&mm->mmap_sem);
1682
2f007e74 1683 for (i = 0; i < nr_pages; i++) {
80bba129 1684 unsigned long addr = (unsigned long)(*pages);
742755a1
CL
1685 struct vm_area_struct *vma;
1686 struct page *page;
c095adbc 1687 int err = -EFAULT;
2f007e74
BG
1688
1689 vma = find_vma(mm, addr);
70384dc6 1690 if (!vma || addr < vma->vm_start)
742755a1
CL
1691 goto set_status;
1692
d899844e
KS
1693 /* FOLL_DUMP to ignore special (like zero) pages */
1694 page = follow_page(vma, addr, FOLL_DUMP);
89f5b7da
LT
1695
1696 err = PTR_ERR(page);
1697 if (IS_ERR(page))
1698 goto set_status;
1699
d899844e 1700 err = page ? page_to_nid(page) : -ENOENT;
742755a1 1701set_status:
80bba129
BG
1702 *status = err;
1703
1704 pages++;
1705 status++;
1706 }
1707
1708 up_read(&mm->mmap_sem);
1709}
1710
1711/*
1712 * Determine the nodes of a user array of pages and store it in
1713 * a user array of status.
1714 */
1715static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1716 const void __user * __user *pages,
1717 int __user *status)
1718{
1719#define DO_PAGES_STAT_CHUNK_NR 16
1720 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1721 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
80bba129 1722
87b8d1ad
PA
1723 while (nr_pages) {
1724 unsigned long chunk_nr;
80bba129 1725
87b8d1ad
PA
1726 chunk_nr = nr_pages;
1727 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1728 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1729
1730 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1731 break;
80bba129
BG
1732
1733 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1734
87b8d1ad
PA
1735 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1736 break;
742755a1 1737
87b8d1ad
PA
1738 pages += chunk_nr;
1739 status += chunk_nr;
1740 nr_pages -= chunk_nr;
1741 }
1742 return nr_pages ? -EFAULT : 0;
742755a1
CL
1743}
1744
1745/*
1746 * Move a list of pages in the address space of the currently executing
1747 * process.
1748 */
938bb9f5
HC
1749SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1750 const void __user * __user *, pages,
1751 const int __user *, nodes,
1752 int __user *, status, int, flags)
742755a1 1753{
742755a1 1754 struct task_struct *task;
742755a1 1755 struct mm_struct *mm;
5e9a0f02 1756 int err;
3268c63e 1757 nodemask_t task_nodes;
742755a1
CL
1758
1759 /* Check flags */
1760 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1761 return -EINVAL;
1762
1763 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1764 return -EPERM;
1765
1766 /* Find the mm_struct */
a879bf58 1767 rcu_read_lock();
228ebcbe 1768 task = pid ? find_task_by_vpid(pid) : current;
742755a1 1769 if (!task) {
a879bf58 1770 rcu_read_unlock();
742755a1
CL
1771 return -ESRCH;
1772 }
3268c63e 1773 get_task_struct(task);
742755a1
CL
1774
1775 /*
1776 * Check if this process has the right to modify the specified
197e7e52 1777 * process. Use the regular "ptrace_may_access()" checks.
742755a1 1778 */
197e7e52 1779 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
c69e8d9c 1780 rcu_read_unlock();
742755a1 1781 err = -EPERM;
5e9a0f02 1782 goto out;
742755a1 1783 }
c69e8d9c 1784 rcu_read_unlock();
742755a1 1785
86c3a764
DQ
1786 err = security_task_movememory(task);
1787 if (err)
5e9a0f02 1788 goto out;
86c3a764 1789
3268c63e
CL
1790 task_nodes = cpuset_mems_allowed(task);
1791 mm = get_task_mm(task);
1792 put_task_struct(task);
1793
6e8b09ea
SL
1794 if (!mm)
1795 return -EINVAL;
1796
1797 if (nodes)
1798 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1799 nodes, status, flags);
1800 else
1801 err = do_pages_stat(mm, nr_pages, pages, status);
742755a1 1802
742755a1
CL
1803 mmput(mm);
1804 return err;
3268c63e
CL
1805
1806out:
1807 put_task_struct(task);
1808 return err;
742755a1 1809}
742755a1 1810
7039e1db
PZ
1811#ifdef CONFIG_NUMA_BALANCING
1812/*
1813 * Returns true if this is a safe migration target node for misplaced NUMA
1814 * pages. Currently it only checks the watermarks which crude
1815 */
1816static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
3abef4e6 1817 unsigned long nr_migrate_pages)
7039e1db
PZ
1818{
1819 int z;
599d0c95 1820
7039e1db
PZ
1821 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1822 struct zone *zone = pgdat->node_zones + z;
1823
1824 if (!populated_zone(zone))
1825 continue;
1826
7039e1db
PZ
1827 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1828 if (!zone_watermark_ok(zone, 0,
1829 high_wmark_pages(zone) +
1830 nr_migrate_pages,
1831 0, 0))
1832 continue;
1833 return true;
1834 }
1835 return false;
1836}
1837
1838static struct page *alloc_misplaced_dst_page(struct page *page,
1839 unsigned long data,
1840 int **result)
1841{
1842 int nid = (int) data;
1843 struct page *newpage;
1844
96db800f 1845 newpage = __alloc_pages_node(nid,
e97ca8e5
JW
1846 (GFP_HIGHUSER_MOVABLE |
1847 __GFP_THISNODE | __GFP_NOMEMALLOC |
1848 __GFP_NORETRY | __GFP_NOWARN) &
8479eba7 1849 ~__GFP_RECLAIM, 0);
bac0382c 1850
7039e1db
PZ
1851 return newpage;
1852}
1853
a8f60772
MG
1854/*
1855 * page migration rate limiting control.
1856 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1857 * window of time. Default here says do not migrate more than 1280M per second.
1858 */
1859static unsigned int migrate_interval_millisecs __read_mostly = 100;
1860static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1861
b32967ff 1862/* Returns true if the node is migrate rate-limited after the update */
1c30e017
MG
1863static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1864 unsigned long nr_pages)
7039e1db 1865{
a8f60772
MG
1866 /*
1867 * Rate-limit the amount of data that is being migrated to a node.
1868 * Optimal placement is no good if the memory bus is saturated and
1869 * all the time is being spent migrating!
1870 */
a8f60772 1871 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1c5e9c27 1872 spin_lock(&pgdat->numabalancing_migrate_lock);
a8f60772
MG
1873 pgdat->numabalancing_migrate_nr_pages = 0;
1874 pgdat->numabalancing_migrate_next_window = jiffies +
1875 msecs_to_jiffies(migrate_interval_millisecs);
1c5e9c27 1876 spin_unlock(&pgdat->numabalancing_migrate_lock);
a8f60772 1877 }
af1839d7
MG
1878 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1879 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1880 nr_pages);
1c5e9c27 1881 return true;
af1839d7 1882 }
1c5e9c27
MG
1883
1884 /*
1885 * This is an unlocked non-atomic update so errors are possible.
1886 * The consequences are failing to migrate when we potentiall should
1887 * have which is not severe enough to warrant locking. If it is ever
1888 * a problem, it can be converted to a per-cpu counter.
1889 */
1890 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1891 return false;
b32967ff
MG
1892}
1893
1c30e017 1894static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
b32967ff 1895{
340ef390 1896 int page_lru;
a8f60772 1897
309381fe 1898 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
3abef4e6 1899
7039e1db 1900 /* Avoid migrating to a node that is nearly full */
340ef390
HD
1901 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1902 return 0;
7039e1db 1903
340ef390
HD
1904 if (isolate_lru_page(page))
1905 return 0;
7039e1db 1906
340ef390
HD
1907 /*
1908 * migrate_misplaced_transhuge_page() skips page migration's usual
1909 * check on page_count(), so we must do it here, now that the page
1910 * has been isolated: a GUP pin, or any other pin, prevents migration.
1911 * The expected page count is 3: 1 for page's mapcount and 1 for the
1912 * caller's pin and 1 for the reference taken by isolate_lru_page().
1913 */
1914 if (PageTransHuge(page) && page_count(page) != 3) {
1915 putback_lru_page(page);
1916 return 0;
7039e1db
PZ
1917 }
1918
340ef390 1919 page_lru = page_is_file_cache(page);
599d0c95 1920 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
340ef390
HD
1921 hpage_nr_pages(page));
1922
149c33e1 1923 /*
340ef390
HD
1924 * Isolating the page has taken another reference, so the
1925 * caller's reference can be safely dropped without the page
1926 * disappearing underneath us during migration.
149c33e1
MG
1927 */
1928 put_page(page);
340ef390 1929 return 1;
b32967ff
MG
1930}
1931
de466bd6
MG
1932bool pmd_trans_migrating(pmd_t pmd)
1933{
1934 struct page *page = pmd_page(pmd);
1935 return PageLocked(page);
1936}
1937
b32967ff
MG
1938/*
1939 * Attempt to migrate a misplaced page to the specified destination
1940 * node. Caller is expected to have an elevated reference count on
1941 * the page that will be dropped by this function before returning.
1942 */
1bc115d8
MG
1943int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1944 int node)
b32967ff
MG
1945{
1946 pg_data_t *pgdat = NODE_DATA(node);
340ef390 1947 int isolated;
b32967ff
MG
1948 int nr_remaining;
1949 LIST_HEAD(migratepages);
1950
1951 /*
1bc115d8
MG
1952 * Don't migrate file pages that are mapped in multiple processes
1953 * with execute permissions as they are probably shared libraries.
b32967ff 1954 */
1bc115d8
MG
1955 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1956 (vma->vm_flags & VM_EXEC))
b32967ff 1957 goto out;
b32967ff
MG
1958
1959 /*
1960 * Rate-limit the amount of data that is being migrated to a node.
1961 * Optimal placement is no good if the memory bus is saturated and
1962 * all the time is being spent migrating!
1963 */
340ef390 1964 if (numamigrate_update_ratelimit(pgdat, 1))
b32967ff 1965 goto out;
b32967ff
MG
1966
1967 isolated = numamigrate_isolate_page(pgdat, page);
1968 if (!isolated)
1969 goto out;
1970
1971 list_add(&page->lru, &migratepages);
9c620e2b 1972 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
68711a74
DR
1973 NULL, node, MIGRATE_ASYNC,
1974 MR_NUMA_MISPLACED);
b32967ff 1975 if (nr_remaining) {
59c82b70
JK
1976 if (!list_empty(&migratepages)) {
1977 list_del(&page->lru);
599d0c95 1978 dec_node_page_state(page, NR_ISOLATED_ANON +
59c82b70
JK
1979 page_is_file_cache(page));
1980 putback_lru_page(page);
1981 }
b32967ff
MG
1982 isolated = 0;
1983 } else
1984 count_vm_numa_event(NUMA_PAGE_MIGRATE);
7039e1db 1985 BUG_ON(!list_empty(&migratepages));
7039e1db 1986 return isolated;
340ef390
HD
1987
1988out:
1989 put_page(page);
1990 return 0;
7039e1db 1991}
220018d3 1992#endif /* CONFIG_NUMA_BALANCING */
b32967ff 1993
220018d3 1994#if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
340ef390
HD
1995/*
1996 * Migrates a THP to a given target node. page must be locked and is unlocked
1997 * before returning.
1998 */
b32967ff
MG
1999int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2000 struct vm_area_struct *vma,
2001 pmd_t *pmd, pmd_t entry,
2002 unsigned long address,
2003 struct page *page, int node)
2004{
c4088ebd 2005 spinlock_t *ptl;
b32967ff
MG
2006 pg_data_t *pgdat = NODE_DATA(node);
2007 int isolated = 0;
2008 struct page *new_page = NULL;
b32967ff 2009 int page_lru = page_is_file_cache(page);
f714f4f2
MG
2010 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2011 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
b32967ff 2012
b32967ff
MG
2013 /*
2014 * Rate-limit the amount of data that is being migrated to a node.
2015 * Optimal placement is no good if the memory bus is saturated and
2016 * all the time is being spent migrating!
2017 */
d28d4335 2018 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
b32967ff
MG
2019 goto out_dropref;
2020
2021 new_page = alloc_pages_node(node,
25160354 2022 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
e97ca8e5 2023 HPAGE_PMD_ORDER);
340ef390
HD
2024 if (!new_page)
2025 goto out_fail;
9a982250 2026 prep_transhuge_page(new_page);
340ef390 2027
b32967ff 2028 isolated = numamigrate_isolate_page(pgdat, page);
340ef390 2029 if (!isolated) {
b32967ff 2030 put_page(new_page);
340ef390 2031 goto out_fail;
b32967ff 2032 }
b0943d61 2033
b32967ff 2034 /* Prepare a page as a migration target */
48c935ad 2035 __SetPageLocked(new_page);
d44d363f
SL
2036 if (PageSwapBacked(page))
2037 __SetPageSwapBacked(new_page);
b32967ff
MG
2038
2039 /* anon mapping, we can simply copy page->mapping to the new page: */
2040 new_page->mapping = page->mapping;
2041 new_page->index = page->index;
2042 migrate_page_copy(new_page, page);
2043 WARN_ON(PageLRU(new_page));
2044
2045 /* Recheck the target PMD */
f714f4f2 2046 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
c4088ebd 2047 ptl = pmd_lock(mm, pmd);
f4e177d1 2048 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
c4088ebd 2049 spin_unlock(ptl);
f714f4f2 2050 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
b32967ff
MG
2051
2052 /* Reverse changes made by migrate_page_copy() */
2053 if (TestClearPageActive(new_page))
2054 SetPageActive(page);
2055 if (TestClearPageUnevictable(new_page))
2056 SetPageUnevictable(page);
b32967ff
MG
2057
2058 unlock_page(new_page);
2059 put_page(new_page); /* Free it */
2060
a54a407f
MG
2061 /* Retake the callers reference and putback on LRU */
2062 get_page(page);
b32967ff 2063 putback_lru_page(page);
599d0c95 2064 mod_node_page_state(page_pgdat(page),
a54a407f 2065 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
eb4489f6
MG
2066
2067 goto out_unlock;
b32967ff
MG
2068 }
2069
10102459 2070 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
f55e1014 2071 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
b32967ff 2072
2b4847e7
MG
2073 /*
2074 * Clear the old entry under pagetable lock and establish the new PTE.
2075 * Any parallel GUP will either observe the old page blocking on the
2076 * page lock, block on the page table lock or observe the new page.
2077 * The SetPageUptodate on the new page and page_add_new_anon_rmap
2078 * guarantee the copy is visible before the pagetable update.
2079 */
f714f4f2 2080 flush_cache_range(vma, mmun_start, mmun_end);
d281ee61 2081 page_add_anon_rmap(new_page, vma, mmun_start, true);
8809aa2d 2082 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
f714f4f2 2083 set_pmd_at(mm, mmun_start, pmd, entry);
ce4a9cc5 2084 update_mmu_cache_pmd(vma, address, &entry);
2b4847e7 2085
f4e177d1 2086 page_ref_unfreeze(page, 2);
51afb12b 2087 mlock_migrate_page(new_page, page);
d281ee61 2088 page_remove_rmap(page, true);
7cd12b4a 2089 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2b4847e7 2090
c4088ebd 2091 spin_unlock(ptl);
4645b9fe
JG
2092 /*
2093 * No need to double call mmu_notifier->invalidate_range() callback as
2094 * the above pmdp_huge_clear_flush_notify() did already call it.
2095 */
2096 mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
b32967ff 2097
11de9927
MG
2098 /* Take an "isolate" reference and put new page on the LRU. */
2099 get_page(new_page);
2100 putback_lru_page(new_page);
2101
b32967ff
MG
2102 unlock_page(new_page);
2103 unlock_page(page);
2104 put_page(page); /* Drop the rmap reference */
2105 put_page(page); /* Drop the LRU isolation reference */
2106
2107 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2108 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2109
599d0c95 2110 mod_node_page_state(page_pgdat(page),
b32967ff
MG
2111 NR_ISOLATED_ANON + page_lru,
2112 -HPAGE_PMD_NR);
2113 return isolated;
2114
340ef390
HD
2115out_fail:
2116 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
b32967ff 2117out_dropref:
2b4847e7
MG
2118 ptl = pmd_lock(mm, pmd);
2119 if (pmd_same(*pmd, entry)) {
4d942466 2120 entry = pmd_modify(entry, vma->vm_page_prot);
f714f4f2 2121 set_pmd_at(mm, mmun_start, pmd, entry);
2b4847e7
MG
2122 update_mmu_cache_pmd(vma, address, &entry);
2123 }
2124 spin_unlock(ptl);
a54a407f 2125
eb4489f6 2126out_unlock:
340ef390 2127 unlock_page(page);
b32967ff 2128 put_page(page);
b32967ff
MG
2129 return 0;
2130}
7039e1db
PZ
2131#endif /* CONFIG_NUMA_BALANCING */
2132
2133#endif /* CONFIG_NUMA */
8763cb45 2134
6b368cd4 2135#if defined(CONFIG_MIGRATE_VMA_HELPER)
8763cb45
JG
2136struct migrate_vma {
2137 struct vm_area_struct *vma;
2138 unsigned long *dst;
2139 unsigned long *src;
2140 unsigned long cpages;
2141 unsigned long npages;
2142 unsigned long start;
2143 unsigned long end;
2144};
2145
2146static int migrate_vma_collect_hole(unsigned long start,
2147 unsigned long end,
2148 struct mm_walk *walk)
2149{
2150 struct migrate_vma *migrate = walk->private;
2151 unsigned long addr;
2152
8315ada7 2153 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
e20d103b 2154 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
8315ada7 2155 migrate->dst[migrate->npages] = 0;
e20d103b 2156 migrate->npages++;
8315ada7
JG
2157 migrate->cpages++;
2158 }
2159
2160 return 0;
2161}
2162
2163static int migrate_vma_collect_skip(unsigned long start,
2164 unsigned long end,
2165 struct mm_walk *walk)
2166{
2167 struct migrate_vma *migrate = walk->private;
2168 unsigned long addr;
2169
8763cb45
JG
2170 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2171 migrate->dst[migrate->npages] = 0;
2172 migrate->src[migrate->npages++] = 0;
2173 }
2174
2175 return 0;
2176}
2177
2178static int migrate_vma_collect_pmd(pmd_t *pmdp,
2179 unsigned long start,
2180 unsigned long end,
2181 struct mm_walk *walk)
2182{
2183 struct migrate_vma *migrate = walk->private;
2184 struct vm_area_struct *vma = walk->vma;
2185 struct mm_struct *mm = vma->vm_mm;
8c3328f1 2186 unsigned long addr = start, unmapped = 0;
8763cb45
JG
2187 spinlock_t *ptl;
2188 pte_t *ptep;
2189
2190again:
2191 if (pmd_none(*pmdp))
2192 return migrate_vma_collect_hole(start, end, walk);
2193
2194 if (pmd_trans_huge(*pmdp)) {
2195 struct page *page;
2196
2197 ptl = pmd_lock(mm, pmdp);
2198 if (unlikely(!pmd_trans_huge(*pmdp))) {
2199 spin_unlock(ptl);
2200 goto again;
2201 }
2202
2203 page = pmd_page(*pmdp);
2204 if (is_huge_zero_page(page)) {
2205 spin_unlock(ptl);
2206 split_huge_pmd(vma, pmdp, addr);
2207 if (pmd_trans_unstable(pmdp))
8315ada7 2208 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2209 walk);
2210 } else {
2211 int ret;
2212
2213 get_page(page);
2214 spin_unlock(ptl);
2215 if (unlikely(!trylock_page(page)))
8315ada7 2216 return migrate_vma_collect_skip(start, end,
8763cb45
JG
2217 walk);
2218 ret = split_huge_page(page);
2219 unlock_page(page);
2220 put_page(page);
8315ada7
JG
2221 if (ret)
2222 return migrate_vma_collect_skip(start, end,
2223 walk);
2224 if (pmd_none(*pmdp))
8763cb45
JG
2225 return migrate_vma_collect_hole(start, end,
2226 walk);
2227 }
2228 }
2229
2230 if (unlikely(pmd_bad(*pmdp)))
8315ada7 2231 return migrate_vma_collect_skip(start, end, walk);
8763cb45
JG
2232
2233 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
8c3328f1
JG
2234 arch_enter_lazy_mmu_mode();
2235
8763cb45
JG
2236 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2237 unsigned long mpfn, pfn;
2238 struct page *page;
8c3328f1 2239 swp_entry_t entry;
8763cb45
JG
2240 pte_t pte;
2241
2242 pte = *ptep;
2243 pfn = pte_pfn(pte);
2244
a5430dda 2245 if (pte_none(pte)) {
8315ada7
JG
2246 mpfn = MIGRATE_PFN_MIGRATE;
2247 migrate->cpages++;
2248 pfn = 0;
8763cb45
JG
2249 goto next;
2250 }
2251
a5430dda
JG
2252 if (!pte_present(pte)) {
2253 mpfn = pfn = 0;
2254
2255 /*
2256 * Only care about unaddressable device page special
2257 * page table entry. Other special swap entries are not
2258 * migratable, and we ignore regular swapped page.
2259 */
2260 entry = pte_to_swp_entry(pte);
2261 if (!is_device_private_entry(entry))
2262 goto next;
2263
2264 page = device_private_entry_to_page(entry);
2265 mpfn = migrate_pfn(page_to_pfn(page))|
2266 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2267 if (is_write_device_private_entry(entry))
2268 mpfn |= MIGRATE_PFN_WRITE;
2269 } else {
8315ada7
JG
2270 if (is_zero_pfn(pfn)) {
2271 mpfn = MIGRATE_PFN_MIGRATE;
2272 migrate->cpages++;
2273 pfn = 0;
2274 goto next;
2275 }
df6ad698 2276 page = _vm_normal_page(migrate->vma, addr, pte, true);
a5430dda
JG
2277 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2278 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2279 }
2280
8763cb45 2281 /* FIXME support THP */
8763cb45
JG
2282 if (!page || !page->mapping || PageTransCompound(page)) {
2283 mpfn = pfn = 0;
2284 goto next;
2285 }
a5430dda 2286 pfn = page_to_pfn(page);
8763cb45
JG
2287
2288 /*
2289 * By getting a reference on the page we pin it and that blocks
2290 * any kind of migration. Side effect is that it "freezes" the
2291 * pte.
2292 *
2293 * We drop this reference after isolating the page from the lru
2294 * for non device page (device page are not on the lru and thus
2295 * can't be dropped from it).
2296 */
2297 get_page(page);
2298 migrate->cpages++;
8763cb45 2299
8c3328f1
JG
2300 /*
2301 * Optimize for the common case where page is only mapped once
2302 * in one process. If we can lock the page, then we can safely
2303 * set up a special migration page table entry now.
2304 */
2305 if (trylock_page(page)) {
2306 pte_t swp_pte;
2307
2308 mpfn |= MIGRATE_PFN_LOCKED;
2309 ptep_get_and_clear(mm, addr, ptep);
2310
2311 /* Setup special migration page table entry */
2312 entry = make_migration_entry(page, pte_write(pte));
2313 swp_pte = swp_entry_to_pte(entry);
2314 if (pte_soft_dirty(pte))
2315 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2316 set_pte_at(mm, addr, ptep, swp_pte);
2317
2318 /*
2319 * This is like regular unmap: we remove the rmap and
2320 * drop page refcount. Page won't be freed, as we took
2321 * a reference just above.
2322 */
2323 page_remove_rmap(page, false);
2324 put_page(page);
a5430dda
JG
2325
2326 if (pte_present(pte))
2327 unmapped++;
8c3328f1
JG
2328 }
2329
8763cb45 2330next:
a5430dda 2331 migrate->dst[migrate->npages] = 0;
8763cb45
JG
2332 migrate->src[migrate->npages++] = mpfn;
2333 }
8c3328f1 2334 arch_leave_lazy_mmu_mode();
8763cb45
JG
2335 pte_unmap_unlock(ptep - 1, ptl);
2336
8c3328f1
JG
2337 /* Only flush the TLB if we actually modified any entries */
2338 if (unmapped)
2339 flush_tlb_range(walk->vma, start, end);
2340
8763cb45
JG
2341 return 0;
2342}
2343
2344/*
2345 * migrate_vma_collect() - collect pages over a range of virtual addresses
2346 * @migrate: migrate struct containing all migration information
2347 *
2348 * This will walk the CPU page table. For each virtual address backed by a
2349 * valid page, it updates the src array and takes a reference on the page, in
2350 * order to pin the page until we lock it and unmap it.
2351 */
2352static void migrate_vma_collect(struct migrate_vma *migrate)
2353{
2354 struct mm_walk mm_walk;
2355
2356 mm_walk.pmd_entry = migrate_vma_collect_pmd;
2357 mm_walk.pte_entry = NULL;
2358 mm_walk.pte_hole = migrate_vma_collect_hole;
2359 mm_walk.hugetlb_entry = NULL;
2360 mm_walk.test_walk = NULL;
2361 mm_walk.vma = migrate->vma;
2362 mm_walk.mm = migrate->vma->vm_mm;
2363 mm_walk.private = migrate;
2364
8c3328f1
JG
2365 mmu_notifier_invalidate_range_start(mm_walk.mm,
2366 migrate->start,
2367 migrate->end);
8763cb45 2368 walk_page_range(migrate->start, migrate->end, &mm_walk);
8c3328f1
JG
2369 mmu_notifier_invalidate_range_end(mm_walk.mm,
2370 migrate->start,
2371 migrate->end);
8763cb45
JG
2372
2373 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2374}
2375
2376/*
2377 * migrate_vma_check_page() - check if page is pinned or not
2378 * @page: struct page to check
2379 *
2380 * Pinned pages cannot be migrated. This is the same test as in
2381 * migrate_page_move_mapping(), except that here we allow migration of a
2382 * ZONE_DEVICE page.
2383 */
2384static bool migrate_vma_check_page(struct page *page)
2385{
2386 /*
2387 * One extra ref because caller holds an extra reference, either from
2388 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2389 * a device page.
2390 */
2391 int extra = 1;
2392
2393 /*
2394 * FIXME support THP (transparent huge page), it is bit more complex to
2395 * check them than regular pages, because they can be mapped with a pmd
2396 * or with a pte (split pte mapping).
2397 */
2398 if (PageCompound(page))
2399 return false;
2400
a5430dda
JG
2401 /* Page from ZONE_DEVICE have one extra reference */
2402 if (is_zone_device_page(page)) {
2403 /*
2404 * Private page can never be pin as they have no valid pte and
2405 * GUP will fail for those. Yet if there is a pending migration
2406 * a thread might try to wait on the pte migration entry and
2407 * will bump the page reference count. Sadly there is no way to
2408 * differentiate a regular pin from migration wait. Hence to
2409 * avoid 2 racing thread trying to migrate back to CPU to enter
2410 * infinite loop (one stoping migration because the other is
2411 * waiting on pte migration entry). We always return true here.
2412 *
2413 * FIXME proper solution is to rework migration_entry_wait() so
2414 * it does not need to take a reference on page.
2415 */
2416 if (is_device_private_page(page))
2417 return true;
2418
df6ad698
JG
2419 /*
2420 * Only allow device public page to be migrated and account for
2421 * the extra reference count imply by ZONE_DEVICE pages.
2422 */
2423 if (!is_device_public_page(page))
2424 return false;
2425 extra++;
a5430dda
JG
2426 }
2427
df6ad698
JG
2428 /* For file back page */
2429 if (page_mapping(page))
2430 extra += 1 + page_has_private(page);
2431
8763cb45
JG
2432 if ((page_count(page) - extra) > page_mapcount(page))
2433 return false;
2434
2435 return true;
2436}
2437
2438/*
2439 * migrate_vma_prepare() - lock pages and isolate them from the lru
2440 * @migrate: migrate struct containing all migration information
2441 *
2442 * This locks pages that have been collected by migrate_vma_collect(). Once each
2443 * page is locked it is isolated from the lru (for non-device pages). Finally,
2444 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2445 * migrated by concurrent kernel threads.
2446 */
2447static void migrate_vma_prepare(struct migrate_vma *migrate)
2448{
2449 const unsigned long npages = migrate->npages;
8c3328f1
JG
2450 const unsigned long start = migrate->start;
2451 unsigned long addr, i, restore = 0;
8763cb45 2452 bool allow_drain = true;
8763cb45
JG
2453
2454 lru_add_drain();
2455
2456 for (i = 0; (i < npages) && migrate->cpages; i++) {
2457 struct page *page = migrate_pfn_to_page(migrate->src[i]);
8c3328f1 2458 bool remap = true;
8763cb45
JG
2459
2460 if (!page)
2461 continue;
2462
8c3328f1
JG
2463 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2464 /*
2465 * Because we are migrating several pages there can be
2466 * a deadlock between 2 concurrent migration where each
2467 * are waiting on each other page lock.
2468 *
2469 * Make migrate_vma() a best effort thing and backoff
2470 * for any page we can not lock right away.
2471 */
2472 if (!trylock_page(page)) {
2473 migrate->src[i] = 0;
2474 migrate->cpages--;
2475 put_page(page);
2476 continue;
2477 }
2478 remap = false;
2479 migrate->src[i] |= MIGRATE_PFN_LOCKED;
8763cb45 2480 }
8763cb45 2481
a5430dda
JG
2482 /* ZONE_DEVICE pages are not on LRU */
2483 if (!is_zone_device_page(page)) {
2484 if (!PageLRU(page) && allow_drain) {
2485 /* Drain CPU's pagevec */
2486 lru_add_drain_all();
2487 allow_drain = false;
2488 }
8763cb45 2489
a5430dda
JG
2490 if (isolate_lru_page(page)) {
2491 if (remap) {
2492 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2493 migrate->cpages--;
2494 restore++;
2495 } else {
2496 migrate->src[i] = 0;
2497 unlock_page(page);
2498 migrate->cpages--;
2499 put_page(page);
2500 }
2501 continue;
8c3328f1 2502 }
a5430dda
JG
2503
2504 /* Drop the reference we took in collect */
2505 put_page(page);
8763cb45
JG
2506 }
2507
2508 if (!migrate_vma_check_page(page)) {
8c3328f1
JG
2509 if (remap) {
2510 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2511 migrate->cpages--;
2512 restore++;
8763cb45 2513
a5430dda
JG
2514 if (!is_zone_device_page(page)) {
2515 get_page(page);
2516 putback_lru_page(page);
2517 }
8c3328f1
JG
2518 } else {
2519 migrate->src[i] = 0;
2520 unlock_page(page);
2521 migrate->cpages--;
2522
a5430dda
JG
2523 if (!is_zone_device_page(page))
2524 putback_lru_page(page);
2525 else
2526 put_page(page);
8c3328f1 2527 }
8763cb45
JG
2528 }
2529 }
8c3328f1
JG
2530
2531 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2532 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2533
2534 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2535 continue;
2536
2537 remove_migration_pte(page, migrate->vma, addr, page);
2538
2539 migrate->src[i] = 0;
2540 unlock_page(page);
2541 put_page(page);
2542 restore--;
2543 }
8763cb45
JG
2544}
2545
2546/*
2547 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2548 * @migrate: migrate struct containing all migration information
2549 *
2550 * Replace page mapping (CPU page table pte) with a special migration pte entry
2551 * and check again if it has been pinned. Pinned pages are restored because we
2552 * cannot migrate them.
2553 *
2554 * This is the last step before we call the device driver callback to allocate
2555 * destination memory and copy contents of original page over to new page.
2556 */
2557static void migrate_vma_unmap(struct migrate_vma *migrate)
2558{
2559 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2560 const unsigned long npages = migrate->npages;
2561 const unsigned long start = migrate->start;
2562 unsigned long addr, i, restore = 0;
2563
2564 for (i = 0; i < npages; i++) {
2565 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2566
2567 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2568 continue;
2569
8c3328f1
JG
2570 if (page_mapped(page)) {
2571 try_to_unmap(page, flags);
2572 if (page_mapped(page))
2573 goto restore;
8763cb45 2574 }
8c3328f1
JG
2575
2576 if (migrate_vma_check_page(page))
2577 continue;
2578
2579restore:
2580 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2581 migrate->cpages--;
2582 restore++;
8763cb45
JG
2583 }
2584
2585 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2586 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2587
2588 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2589 continue;
2590
2591 remove_migration_ptes(page, page, false);
2592
2593 migrate->src[i] = 0;
2594 unlock_page(page);
2595 restore--;
2596
a5430dda
JG
2597 if (is_zone_device_page(page))
2598 put_page(page);
2599 else
2600 putback_lru_page(page);
8763cb45
JG
2601 }
2602}
2603
8315ada7
JG
2604static void migrate_vma_insert_page(struct migrate_vma *migrate,
2605 unsigned long addr,
2606 struct page *page,
2607 unsigned long *src,
2608 unsigned long *dst)
2609{
2610 struct vm_area_struct *vma = migrate->vma;
2611 struct mm_struct *mm = vma->vm_mm;
2612 struct mem_cgroup *memcg;
2613 bool flush = false;
2614 spinlock_t *ptl;
2615 pte_t entry;
2616 pgd_t *pgdp;
2617 p4d_t *p4dp;
2618 pud_t *pudp;
2619 pmd_t *pmdp;
2620 pte_t *ptep;
2621
2622 /* Only allow populating anonymous memory */
2623 if (!vma_is_anonymous(vma))
2624 goto abort;
2625
2626 pgdp = pgd_offset(mm, addr);
2627 p4dp = p4d_alloc(mm, pgdp, addr);
2628 if (!p4dp)
2629 goto abort;
2630 pudp = pud_alloc(mm, p4dp, addr);
2631 if (!pudp)
2632 goto abort;
2633 pmdp = pmd_alloc(mm, pudp, addr);
2634 if (!pmdp)
2635 goto abort;
2636
2637 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2638 goto abort;
2639
2640 /*
2641 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2642 * pte_offset_map() on pmds where a huge pmd might be created
2643 * from a different thread.
2644 *
2645 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2646 * parallel threads are excluded by other means.
2647 *
2648 * Here we only have down_read(mmap_sem).
2649 */
2650 if (pte_alloc(mm, pmdp, addr))
2651 goto abort;
2652
2653 /* See the comment in pte_alloc_one_map() */
2654 if (unlikely(pmd_trans_unstable(pmdp)))
2655 goto abort;
2656
2657 if (unlikely(anon_vma_prepare(vma)))
2658 goto abort;
2659 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2660 goto abort;
2661
2662 /*
2663 * The memory barrier inside __SetPageUptodate makes sure that
2664 * preceding stores to the page contents become visible before
2665 * the set_pte_at() write.
2666 */
2667 __SetPageUptodate(page);
2668
df6ad698
JG
2669 if (is_zone_device_page(page)) {
2670 if (is_device_private_page(page)) {
2671 swp_entry_t swp_entry;
2672
2673 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2674 entry = swp_entry_to_pte(swp_entry);
2675 } else if (is_device_public_page(page)) {
2676 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2677 if (vma->vm_flags & VM_WRITE)
2678 entry = pte_mkwrite(pte_mkdirty(entry));
2679 entry = pte_mkdevmap(entry);
2680 }
8315ada7
JG
2681 } else {
2682 entry = mk_pte(page, vma->vm_page_prot);
2683 if (vma->vm_flags & VM_WRITE)
2684 entry = pte_mkwrite(pte_mkdirty(entry));
2685 }
2686
2687 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2688
2689 if (pte_present(*ptep)) {
2690 unsigned long pfn = pte_pfn(*ptep);
2691
2692 if (!is_zero_pfn(pfn)) {
2693 pte_unmap_unlock(ptep, ptl);
2694 mem_cgroup_cancel_charge(page, memcg, false);
2695 goto abort;
2696 }
2697 flush = true;
2698 } else if (!pte_none(*ptep)) {
2699 pte_unmap_unlock(ptep, ptl);
2700 mem_cgroup_cancel_charge(page, memcg, false);
2701 goto abort;
2702 }
2703
2704 /*
2705 * Check for usefaultfd but do not deliver the fault. Instead,
2706 * just back off.
2707 */
2708 if (userfaultfd_missing(vma)) {
2709 pte_unmap_unlock(ptep, ptl);
2710 mem_cgroup_cancel_charge(page, memcg, false);
2711 goto abort;
2712 }
2713
2714 inc_mm_counter(mm, MM_ANONPAGES);
2715 page_add_new_anon_rmap(page, vma, addr, false);
2716 mem_cgroup_commit_charge(page, memcg, false, false);
2717 if (!is_zone_device_page(page))
2718 lru_cache_add_active_or_unevictable(page, vma);
2719 get_page(page);
2720
2721 if (flush) {
2722 flush_cache_page(vma, addr, pte_pfn(*ptep));
2723 ptep_clear_flush_notify(vma, addr, ptep);
2724 set_pte_at_notify(mm, addr, ptep, entry);
2725 update_mmu_cache(vma, addr, ptep);
2726 } else {
2727 /* No need to invalidate - it was non-present before */
2728 set_pte_at(mm, addr, ptep, entry);
2729 update_mmu_cache(vma, addr, ptep);
2730 }
2731
2732 pte_unmap_unlock(ptep, ptl);
2733 *src = MIGRATE_PFN_MIGRATE;
2734 return;
2735
2736abort:
2737 *src &= ~MIGRATE_PFN_MIGRATE;
2738}
2739
8763cb45
JG
2740/*
2741 * migrate_vma_pages() - migrate meta-data from src page to dst page
2742 * @migrate: migrate struct containing all migration information
2743 *
2744 * This migrates struct page meta-data from source struct page to destination
2745 * struct page. This effectively finishes the migration from source page to the
2746 * destination page.
2747 */
2748static void migrate_vma_pages(struct migrate_vma *migrate)
2749{
2750 const unsigned long npages = migrate->npages;
2751 const unsigned long start = migrate->start;
8315ada7
JG
2752 struct vm_area_struct *vma = migrate->vma;
2753 struct mm_struct *mm = vma->vm_mm;
2754 unsigned long addr, i, mmu_start;
2755 bool notified = false;
8763cb45
JG
2756
2757 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2758 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2759 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2760 struct address_space *mapping;
2761 int r;
2762
8315ada7
JG
2763 if (!newpage) {
2764 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
8763cb45 2765 continue;
8315ada7
JG
2766 }
2767
2768 if (!page) {
2769 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2770 continue;
2771 }
2772 if (!notified) {
2773 mmu_start = addr;
2774 notified = true;
2775 mmu_notifier_invalidate_range_start(mm,
2776 mmu_start,
2777 migrate->end);
2778 }
2779 migrate_vma_insert_page(migrate, addr, newpage,
2780 &migrate->src[i],
2781 &migrate->dst[i]);
8763cb45 2782 continue;
8315ada7 2783 }
8763cb45
JG
2784
2785 mapping = page_mapping(page);
2786
a5430dda
JG
2787 if (is_zone_device_page(newpage)) {
2788 if (is_device_private_page(newpage)) {
2789 /*
2790 * For now only support private anonymous when
2791 * migrating to un-addressable device memory.
2792 */
2793 if (mapping) {
2794 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2795 continue;
2796 }
df6ad698 2797 } else if (!is_device_public_page(newpage)) {
a5430dda
JG
2798 /*
2799 * Other types of ZONE_DEVICE page are not
2800 * supported.
2801 */
2802 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2803 continue;
2804 }
2805 }
2806
8763cb45
JG
2807 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2808 if (r != MIGRATEPAGE_SUCCESS)
2809 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2810 }
8315ada7 2811
4645b9fe
JG
2812 /*
2813 * No need to double call mmu_notifier->invalidate_range() callback as
2814 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2815 * did already call it.
2816 */
8315ada7 2817 if (notified)
4645b9fe
JG
2818 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2819 migrate->end);
8763cb45
JG
2820}
2821
2822/*
2823 * migrate_vma_finalize() - restore CPU page table entry
2824 * @migrate: migrate struct containing all migration information
2825 *
2826 * This replaces the special migration pte entry with either a mapping to the
2827 * new page if migration was successful for that page, or to the original page
2828 * otherwise.
2829 *
2830 * This also unlocks the pages and puts them back on the lru, or drops the extra
2831 * refcount, for device pages.
2832 */
2833static void migrate_vma_finalize(struct migrate_vma *migrate)
2834{
2835 const unsigned long npages = migrate->npages;
2836 unsigned long i;
2837
2838 for (i = 0; i < npages; i++) {
2839 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2840 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2841
8315ada7
JG
2842 if (!page) {
2843 if (newpage) {
2844 unlock_page(newpage);
2845 put_page(newpage);
2846 }
8763cb45 2847 continue;
8315ada7
JG
2848 }
2849
8763cb45
JG
2850 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2851 if (newpage) {
2852 unlock_page(newpage);
2853 put_page(newpage);
2854 }
2855 newpage = page;
2856 }
2857
2858 remove_migration_ptes(page, newpage, false);
2859 unlock_page(page);
2860 migrate->cpages--;
2861
a5430dda
JG
2862 if (is_zone_device_page(page))
2863 put_page(page);
2864 else
2865 putback_lru_page(page);
8763cb45
JG
2866
2867 if (newpage != page) {
2868 unlock_page(newpage);
a5430dda
JG
2869 if (is_zone_device_page(newpage))
2870 put_page(newpage);
2871 else
2872 putback_lru_page(newpage);
8763cb45
JG
2873 }
2874 }
2875}
2876
2877/*
2878 * migrate_vma() - migrate a range of memory inside vma
2879 *
2880 * @ops: migration callback for allocating destination memory and copying
2881 * @vma: virtual memory area containing the range to be migrated
2882 * @start: start address of the range to migrate (inclusive)
2883 * @end: end address of the range to migrate (exclusive)
2884 * @src: array of hmm_pfn_t containing source pfns
2885 * @dst: array of hmm_pfn_t containing destination pfns
2886 * @private: pointer passed back to each of the callback
2887 * Returns: 0 on success, error code otherwise
2888 *
2889 * This function tries to migrate a range of memory virtual address range, using
2890 * callbacks to allocate and copy memory from source to destination. First it
2891 * collects all the pages backing each virtual address in the range, saving this
2892 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2893 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2894 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2895 * in the corresponding src array entry. It then restores any pages that are
2896 * pinned, by remapping and unlocking those pages.
2897 *
2898 * At this point it calls the alloc_and_copy() callback. For documentation on
2899 * what is expected from that callback, see struct migrate_vma_ops comments in
2900 * include/linux/migrate.h
2901 *
2902 * After the alloc_and_copy() callback, this function goes over each entry in
2903 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2904 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2905 * then the function tries to migrate struct page information from the source
2906 * struct page to the destination struct page. If it fails to migrate the struct
2907 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2908 * array.
2909 *
2910 * At this point all successfully migrated pages have an entry in the src
2911 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2912 * array entry with MIGRATE_PFN_VALID flag set.
2913 *
2914 * It then calls the finalize_and_map() callback. See comments for "struct
2915 * migrate_vma_ops", in include/linux/migrate.h for details about
2916 * finalize_and_map() behavior.
2917 *
2918 * After the finalize_and_map() callback, for successfully migrated pages, this
2919 * function updates the CPU page table to point to new pages, otherwise it
2920 * restores the CPU page table to point to the original source pages.
2921 *
2922 * Function returns 0 after the above steps, even if no pages were migrated
2923 * (The function only returns an error if any of the arguments are invalid.)
2924 *
2925 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2926 * unsigned long entries.
2927 */
2928int migrate_vma(const struct migrate_vma_ops *ops,
2929 struct vm_area_struct *vma,
2930 unsigned long start,
2931 unsigned long end,
2932 unsigned long *src,
2933 unsigned long *dst,
2934 void *private)
2935{
2936 struct migrate_vma migrate;
2937
2938 /* Sanity check the arguments */
2939 start &= PAGE_MASK;
2940 end &= PAGE_MASK;
2941 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL))
2942 return -EINVAL;
2943 if (start < vma->vm_start || start >= vma->vm_end)
2944 return -EINVAL;
2945 if (end <= vma->vm_start || end > vma->vm_end)
2946 return -EINVAL;
2947 if (!ops || !src || !dst || start >= end)
2948 return -EINVAL;
2949
2950 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2951 migrate.src = src;
2952 migrate.dst = dst;
2953 migrate.start = start;
2954 migrate.npages = 0;
2955 migrate.cpages = 0;
2956 migrate.end = end;
2957 migrate.vma = vma;
2958
2959 /* Collect, and try to unmap source pages */
2960 migrate_vma_collect(&migrate);
2961 if (!migrate.cpages)
2962 return 0;
2963
2964 /* Lock and isolate page */
2965 migrate_vma_prepare(&migrate);
2966 if (!migrate.cpages)
2967 return 0;
2968
2969 /* Unmap pages */
2970 migrate_vma_unmap(&migrate);
2971 if (!migrate.cpages)
2972 return 0;
2973
2974 /*
2975 * At this point pages are locked and unmapped, and thus they have
2976 * stable content and can safely be copied to destination memory that
2977 * is allocated by the callback.
2978 *
2979 * Note that migration can fail in migrate_vma_struct_page() for each
2980 * individual page.
2981 */
2982 ops->alloc_and_copy(vma, src, dst, start, end, private);
2983
2984 /* This does the real migration of struct page */
2985 migrate_vma_pages(&migrate);
2986
2987 ops->finalize_and_map(vma, src, dst, start, end, private);
2988
2989 /* Unlock and remap pages */
2990 migrate_vma_finalize(&migrate);
2991
2992 return 0;
2993}
2994EXPORT_SYMBOL(migrate_vma);
6b368cd4 2995#endif /* defined(MIGRATE_VMA_HELPER) */