]> git.ipfire.org Git - thirdparty/linux.git/blame - mm/migrate_device.c
mm/rmap: Convert try_to_unmap() to take a folio
[thirdparty/linux.git] / mm / migrate_device.c
CommitLineData
76cbbead
CH
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Device Memory Migration functionality.
4 *
5 * Originally written by Jérôme Glisse.
6 */
7#include <linux/export.h>
8#include <linux/memremap.h>
9#include <linux/migrate.h>
10#include <linux/mm_inline.h>
11#include <linux/mmu_notifier.h>
12#include <linux/oom.h>
13#include <linux/pagewalk.h>
14#include <linux/rmap.h>
15#include <linux/swapops.h>
16#include <asm/tlbflush.h>
17#include "internal.h"
18
19static int migrate_vma_collect_skip(unsigned long start,
20 unsigned long end,
21 struct mm_walk *walk)
22{
23 struct migrate_vma *migrate = walk->private;
24 unsigned long addr;
25
26 for (addr = start; addr < end; addr += PAGE_SIZE) {
27 migrate->dst[migrate->npages] = 0;
28 migrate->src[migrate->npages++] = 0;
29 }
30
31 return 0;
32}
33
34static int migrate_vma_collect_hole(unsigned long start,
35 unsigned long end,
36 __always_unused int depth,
37 struct mm_walk *walk)
38{
39 struct migrate_vma *migrate = walk->private;
40 unsigned long addr;
41
42 /* Only allow populating anonymous memory. */
43 if (!vma_is_anonymous(walk->vma))
44 return migrate_vma_collect_skip(start, end, walk);
45
46 for (addr = start; addr < end; addr += PAGE_SIZE) {
47 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
48 migrate->dst[migrate->npages] = 0;
49 migrate->npages++;
50 migrate->cpages++;
51 }
52
53 return 0;
54}
55
56static int migrate_vma_collect_pmd(pmd_t *pmdp,
57 unsigned long start,
58 unsigned long end,
59 struct mm_walk *walk)
60{
61 struct migrate_vma *migrate = walk->private;
62 struct vm_area_struct *vma = walk->vma;
63 struct mm_struct *mm = vma->vm_mm;
64 unsigned long addr = start, unmapped = 0;
65 spinlock_t *ptl;
66 pte_t *ptep;
67
68again:
69 if (pmd_none(*pmdp))
70 return migrate_vma_collect_hole(start, end, -1, walk);
71
72 if (pmd_trans_huge(*pmdp)) {
73 struct page *page;
74
75 ptl = pmd_lock(mm, pmdp);
76 if (unlikely(!pmd_trans_huge(*pmdp))) {
77 spin_unlock(ptl);
78 goto again;
79 }
80
81 page = pmd_page(*pmdp);
82 if (is_huge_zero_page(page)) {
83 spin_unlock(ptl);
84 split_huge_pmd(vma, pmdp, addr);
85 if (pmd_trans_unstable(pmdp))
86 return migrate_vma_collect_skip(start, end,
87 walk);
88 } else {
89 int ret;
90
91 get_page(page);
92 spin_unlock(ptl);
93 if (unlikely(!trylock_page(page)))
94 return migrate_vma_collect_skip(start, end,
95 walk);
96 ret = split_huge_page(page);
97 unlock_page(page);
98 put_page(page);
99 if (ret)
100 return migrate_vma_collect_skip(start, end,
101 walk);
102 if (pmd_none(*pmdp))
103 return migrate_vma_collect_hole(start, end, -1,
104 walk);
105 }
106 }
107
108 if (unlikely(pmd_bad(*pmdp)))
109 return migrate_vma_collect_skip(start, end, walk);
110
111 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
112 arch_enter_lazy_mmu_mode();
113
114 for (; addr < end; addr += PAGE_SIZE, ptep++) {
115 unsigned long mpfn = 0, pfn;
116 struct page *page;
117 swp_entry_t entry;
118 pte_t pte;
119
120 pte = *ptep;
121
122 if (pte_none(pte)) {
123 if (vma_is_anonymous(vma)) {
124 mpfn = MIGRATE_PFN_MIGRATE;
125 migrate->cpages++;
126 }
127 goto next;
128 }
129
130 if (!pte_present(pte)) {
131 /*
132 * Only care about unaddressable device page special
133 * page table entry. Other special swap entries are not
134 * migratable, and we ignore regular swapped page.
135 */
136 entry = pte_to_swp_entry(pte);
137 if (!is_device_private_entry(entry))
138 goto next;
139
140 page = pfn_swap_entry_to_page(entry);
141 if (!(migrate->flags &
142 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
143 page->pgmap->owner != migrate->pgmap_owner)
144 goto next;
145
146 mpfn = migrate_pfn(page_to_pfn(page)) |
147 MIGRATE_PFN_MIGRATE;
148 if (is_writable_device_private_entry(entry))
149 mpfn |= MIGRATE_PFN_WRITE;
150 } else {
151 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
152 goto next;
153 pfn = pte_pfn(pte);
154 if (is_zero_pfn(pfn)) {
155 mpfn = MIGRATE_PFN_MIGRATE;
156 migrate->cpages++;
157 goto next;
158 }
159 page = vm_normal_page(migrate->vma, addr, pte);
160 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
161 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
162 }
163
164 /* FIXME support THP */
165 if (!page || !page->mapping || PageTransCompound(page)) {
166 mpfn = 0;
167 goto next;
168 }
169
170 /*
171 * By getting a reference on the page we pin it and that blocks
172 * any kind of migration. Side effect is that it "freezes" the
173 * pte.
174 *
175 * We drop this reference after isolating the page from the lru
176 * for non device page (device page are not on the lru and thus
177 * can't be dropped from it).
178 */
179 get_page(page);
180
181 /*
182 * Optimize for the common case where page is only mapped once
183 * in one process. If we can lock the page, then we can safely
184 * set up a special migration page table entry now.
185 */
186 if (trylock_page(page)) {
187 pte_t swp_pte;
188
189 migrate->cpages++;
190 ptep_get_and_clear(mm, addr, ptep);
191
192 /* Setup special migration page table entry */
193 if (mpfn & MIGRATE_PFN_WRITE)
194 entry = make_writable_migration_entry(
195 page_to_pfn(page));
196 else
197 entry = make_readable_migration_entry(
198 page_to_pfn(page));
199 swp_pte = swp_entry_to_pte(entry);
200 if (pte_present(pte)) {
201 if (pte_soft_dirty(pte))
202 swp_pte = pte_swp_mksoft_dirty(swp_pte);
203 if (pte_uffd_wp(pte))
204 swp_pte = pte_swp_mkuffd_wp(swp_pte);
205 } else {
206 if (pte_swp_soft_dirty(pte))
207 swp_pte = pte_swp_mksoft_dirty(swp_pte);
208 if (pte_swp_uffd_wp(pte))
209 swp_pte = pte_swp_mkuffd_wp(swp_pte);
210 }
211 set_pte_at(mm, addr, ptep, swp_pte);
212
213 /*
214 * This is like regular unmap: we remove the rmap and
215 * drop page refcount. Page won't be freed, as we took
216 * a reference just above.
217 */
218 page_remove_rmap(page, vma, false);
219 put_page(page);
220
221 if (pte_present(pte))
222 unmapped++;
223 } else {
224 put_page(page);
225 mpfn = 0;
226 }
227
228next:
229 migrate->dst[migrate->npages] = 0;
230 migrate->src[migrate->npages++] = mpfn;
231 }
232 arch_leave_lazy_mmu_mode();
233 pte_unmap_unlock(ptep - 1, ptl);
234
235 /* Only flush the TLB if we actually modified any entries */
236 if (unmapped)
237 flush_tlb_range(walk->vma, start, end);
238
239 return 0;
240}
241
242static const struct mm_walk_ops migrate_vma_walk_ops = {
243 .pmd_entry = migrate_vma_collect_pmd,
244 .pte_hole = migrate_vma_collect_hole,
245};
246
247/*
248 * migrate_vma_collect() - collect pages over a range of virtual addresses
249 * @migrate: migrate struct containing all migration information
250 *
251 * This will walk the CPU page table. For each virtual address backed by a
252 * valid page, it updates the src array and takes a reference on the page, in
253 * order to pin the page until we lock it and unmap it.
254 */
255static void migrate_vma_collect(struct migrate_vma *migrate)
256{
257 struct mmu_notifier_range range;
258
259 /*
260 * Note that the pgmap_owner is passed to the mmu notifier callback so
261 * that the registered device driver can skip invalidating device
262 * private page mappings that won't be migrated.
263 */
264 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
265 migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end,
266 migrate->pgmap_owner);
267 mmu_notifier_invalidate_range_start(&range);
268
269 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
270 &migrate_vma_walk_ops, migrate);
271
272 mmu_notifier_invalidate_range_end(&range);
273 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
274}
275
276/*
277 * migrate_vma_check_page() - check if page is pinned or not
278 * @page: struct page to check
279 *
280 * Pinned pages cannot be migrated. This is the same test as in
281 * folio_migrate_mapping(), except that here we allow migration of a
282 * ZONE_DEVICE page.
283 */
284static bool migrate_vma_check_page(struct page *page)
285{
286 /*
287 * One extra ref because caller holds an extra reference, either from
288 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
289 * a device page.
290 */
291 int extra = 1;
292
293 /*
294 * FIXME support THP (transparent huge page), it is bit more complex to
295 * check them than regular pages, because they can be mapped with a pmd
296 * or with a pte (split pte mapping).
297 */
298 if (PageCompound(page))
299 return false;
300
301 /* Page from ZONE_DEVICE have one extra reference */
302 if (is_zone_device_page(page))
303 extra++;
304
305 /* For file back page */
306 if (page_mapping(page))
307 extra += 1 + page_has_private(page);
308
309 if ((page_count(page) - extra) > page_mapcount(page))
310 return false;
311
312 return true;
313}
314
315/*
316 * migrate_vma_unmap() - replace page mapping with special migration pte entry
317 * @migrate: migrate struct containing all migration information
318 *
319 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
320 * special migration pte entry and check if it has been pinned. Pinned pages are
321 * restored because we cannot migrate them.
322 *
323 * This is the last step before we call the device driver callback to allocate
324 * destination memory and copy contents of original page over to new page.
325 */
326static void migrate_vma_unmap(struct migrate_vma *migrate)
327{
328 const unsigned long npages = migrate->npages;
329 unsigned long i, restore = 0;
330 bool allow_drain = true;
331
332 lru_add_drain();
333
334 for (i = 0; i < npages; i++) {
335 struct page *page = migrate_pfn_to_page(migrate->src[i]);
336
337 if (!page)
338 continue;
339
340 /* ZONE_DEVICE pages are not on LRU */
341 if (!is_zone_device_page(page)) {
342 if (!PageLRU(page) && allow_drain) {
343 /* Drain CPU's pagevec */
344 lru_add_drain_all();
345 allow_drain = false;
346 }
347
348 if (isolate_lru_page(page)) {
349 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
350 migrate->cpages--;
351 restore++;
352 continue;
353 }
354
355 /* Drop the reference we took in collect */
356 put_page(page);
357 }
358
359 if (page_mapped(page))
360 try_to_migrate(page, 0);
361
362 if (page_mapped(page) || !migrate_vma_check_page(page)) {
363 if (!is_zone_device_page(page)) {
364 get_page(page);
365 putback_lru_page(page);
366 }
367
368 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
369 migrate->cpages--;
370 restore++;
371 continue;
372 }
373 }
374
375 for (i = 0; i < npages && restore; i++) {
376 struct page *page = migrate_pfn_to_page(migrate->src[i]);
377
378 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
379 continue;
380
381 remove_migration_ptes(page, page, false);
382
383 migrate->src[i] = 0;
384 unlock_page(page);
385 put_page(page);
386 restore--;
387 }
388}
389
390/**
391 * migrate_vma_setup() - prepare to migrate a range of memory
392 * @args: contains the vma, start, and pfns arrays for the migration
393 *
394 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
395 * without an error.
396 *
397 * Prepare to migrate a range of memory virtual address range by collecting all
398 * the pages backing each virtual address in the range, saving them inside the
399 * src array. Then lock those pages and unmap them. Once the pages are locked
400 * and unmapped, check whether each page is pinned or not. Pages that aren't
401 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
402 * corresponding src array entry. Then restores any pages that are pinned, by
403 * remapping and unlocking those pages.
404 *
405 * The caller should then allocate destination memory and copy source memory to
406 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
407 * flag set). Once these are allocated and copied, the caller must update each
408 * corresponding entry in the dst array with the pfn value of the destination
409 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
410 * lock_page().
411 *
412 * Note that the caller does not have to migrate all the pages that are marked
413 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
414 * device memory to system memory. If the caller cannot migrate a device page
415 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
416 * consequences for the userspace process, so it must be avoided if at all
417 * possible.
418 *
419 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
420 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
421 * allowing the caller to allocate device memory for those unbacked virtual
422 * addresses. For this the caller simply has to allocate device memory and
423 * properly set the destination entry like for regular migration. Note that
424 * this can still fail, and thus inside the device driver you must check if the
425 * migration was successful for those entries after calling migrate_vma_pages(),
426 * just like for regular migration.
427 *
428 * After that, the callers must call migrate_vma_pages() to go over each entry
429 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
430 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
431 * then migrate_vma_pages() to migrate struct page information from the source
432 * struct page to the destination struct page. If it fails to migrate the
433 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
434 * src array.
435 *
436 * At this point all successfully migrated pages have an entry in the src
437 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
438 * array entry with MIGRATE_PFN_VALID flag set.
439 *
440 * Once migrate_vma_pages() returns the caller may inspect which pages were
441 * successfully migrated, and which were not. Successfully migrated pages will
442 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
443 *
444 * It is safe to update device page table after migrate_vma_pages() because
445 * both destination and source page are still locked, and the mmap_lock is held
446 * in read mode (hence no one can unmap the range being migrated).
447 *
448 * Once the caller is done cleaning up things and updating its page table (if it
449 * chose to do so, this is not an obligation) it finally calls
450 * migrate_vma_finalize() to update the CPU page table to point to new pages
451 * for successfully migrated pages or otherwise restore the CPU page table to
452 * point to the original source pages.
453 */
454int migrate_vma_setup(struct migrate_vma *args)
455{
456 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
457
458 args->start &= PAGE_MASK;
459 args->end &= PAGE_MASK;
460 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
461 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
462 return -EINVAL;
463 if (nr_pages <= 0)
464 return -EINVAL;
465 if (args->start < args->vma->vm_start ||
466 args->start >= args->vma->vm_end)
467 return -EINVAL;
468 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
469 return -EINVAL;
470 if (!args->src || !args->dst)
471 return -EINVAL;
472
473 memset(args->src, 0, sizeof(*args->src) * nr_pages);
474 args->cpages = 0;
475 args->npages = 0;
476
477 migrate_vma_collect(args);
478
479 if (args->cpages)
480 migrate_vma_unmap(args);
481
482 /*
483 * At this point pages are locked and unmapped, and thus they have
484 * stable content and can safely be copied to destination memory that
485 * is allocated by the drivers.
486 */
487 return 0;
488
489}
490EXPORT_SYMBOL(migrate_vma_setup);
491
492/*
493 * This code closely matches the code in:
494 * __handle_mm_fault()
495 * handle_pte_fault()
496 * do_anonymous_page()
497 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
498 * private page.
499 */
500static void migrate_vma_insert_page(struct migrate_vma *migrate,
501 unsigned long addr,
502 struct page *page,
503 unsigned long *src)
504{
505 struct vm_area_struct *vma = migrate->vma;
506 struct mm_struct *mm = vma->vm_mm;
507 bool flush = false;
508 spinlock_t *ptl;
509 pte_t entry;
510 pgd_t *pgdp;
511 p4d_t *p4dp;
512 pud_t *pudp;
513 pmd_t *pmdp;
514 pte_t *ptep;
515
516 /* Only allow populating anonymous memory */
517 if (!vma_is_anonymous(vma))
518 goto abort;
519
520 pgdp = pgd_offset(mm, addr);
521 p4dp = p4d_alloc(mm, pgdp, addr);
522 if (!p4dp)
523 goto abort;
524 pudp = pud_alloc(mm, p4dp, addr);
525 if (!pudp)
526 goto abort;
527 pmdp = pmd_alloc(mm, pudp, addr);
528 if (!pmdp)
529 goto abort;
530
531 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
532 goto abort;
533
534 /*
535 * Use pte_alloc() instead of pte_alloc_map(). We can't run
536 * pte_offset_map() on pmds where a huge pmd might be created
537 * from a different thread.
538 *
539 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
540 * parallel threads are excluded by other means.
541 *
542 * Here we only have mmap_read_lock(mm).
543 */
544 if (pte_alloc(mm, pmdp))
545 goto abort;
546
547 /* See the comment in pte_alloc_one_map() */
548 if (unlikely(pmd_trans_unstable(pmdp)))
549 goto abort;
550
551 if (unlikely(anon_vma_prepare(vma)))
552 goto abort;
553 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
554 goto abort;
555
556 /*
557 * The memory barrier inside __SetPageUptodate makes sure that
558 * preceding stores to the page contents become visible before
559 * the set_pte_at() write.
560 */
561 __SetPageUptodate(page);
562
563 if (is_device_private_page(page)) {
564 swp_entry_t swp_entry;
565
566 if (vma->vm_flags & VM_WRITE)
567 swp_entry = make_writable_device_private_entry(
568 page_to_pfn(page));
569 else
570 swp_entry = make_readable_device_private_entry(
571 page_to_pfn(page));
572 entry = swp_entry_to_pte(swp_entry);
573 } else {
574 /*
575 * For now we only support migrating to un-addressable device
576 * memory.
577 */
578 if (is_zone_device_page(page)) {
579 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
580 goto abort;
581 }
582 entry = mk_pte(page, vma->vm_page_prot);
583 if (vma->vm_flags & VM_WRITE)
584 entry = pte_mkwrite(pte_mkdirty(entry));
585 }
586
587 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
588
589 if (check_stable_address_space(mm))
590 goto unlock_abort;
591
592 if (pte_present(*ptep)) {
593 unsigned long pfn = pte_pfn(*ptep);
594
595 if (!is_zero_pfn(pfn))
596 goto unlock_abort;
597 flush = true;
598 } else if (!pte_none(*ptep))
599 goto unlock_abort;
600
601 /*
602 * Check for userfaultfd but do not deliver the fault. Instead,
603 * just back off.
604 */
605 if (userfaultfd_missing(vma))
606 goto unlock_abort;
607
608 inc_mm_counter(mm, MM_ANONPAGES);
609 page_add_new_anon_rmap(page, vma, addr, false);
610 if (!is_zone_device_page(page))
611 lru_cache_add_inactive_or_unevictable(page, vma);
612 get_page(page);
613
614 if (flush) {
615 flush_cache_page(vma, addr, pte_pfn(*ptep));
616 ptep_clear_flush_notify(vma, addr, ptep);
617 set_pte_at_notify(mm, addr, ptep, entry);
618 update_mmu_cache(vma, addr, ptep);
619 } else {
620 /* No need to invalidate - it was non-present before */
621 set_pte_at(mm, addr, ptep, entry);
622 update_mmu_cache(vma, addr, ptep);
623 }
624
625 pte_unmap_unlock(ptep, ptl);
626 *src = MIGRATE_PFN_MIGRATE;
627 return;
628
629unlock_abort:
630 pte_unmap_unlock(ptep, ptl);
631abort:
632 *src &= ~MIGRATE_PFN_MIGRATE;
633}
634
635/**
636 * migrate_vma_pages() - migrate meta-data from src page to dst page
637 * @migrate: migrate struct containing all migration information
638 *
639 * This migrates struct page meta-data from source struct page to destination
640 * struct page. This effectively finishes the migration from source page to the
641 * destination page.
642 */
643void migrate_vma_pages(struct migrate_vma *migrate)
644{
645 const unsigned long npages = migrate->npages;
646 const unsigned long start = migrate->start;
647 struct mmu_notifier_range range;
648 unsigned long addr, i;
649 bool notified = false;
650
651 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
652 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
653 struct page *page = migrate_pfn_to_page(migrate->src[i]);
654 struct address_space *mapping;
655 int r;
656
657 if (!newpage) {
658 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
659 continue;
660 }
661
662 if (!page) {
663 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
664 continue;
665 if (!notified) {
666 notified = true;
667
668 mmu_notifier_range_init_owner(&range,
669 MMU_NOTIFY_MIGRATE, 0, migrate->vma,
670 migrate->vma->vm_mm, addr, migrate->end,
671 migrate->pgmap_owner);
672 mmu_notifier_invalidate_range_start(&range);
673 }
674 migrate_vma_insert_page(migrate, addr, newpage,
675 &migrate->src[i]);
676 continue;
677 }
678
679 mapping = page_mapping(page);
680
681 if (is_device_private_page(newpage)) {
682 /*
683 * For now only support private anonymous when migrating
684 * to un-addressable device memory.
685 */
686 if (mapping) {
687 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
688 continue;
689 }
690 } else if (is_zone_device_page(newpage)) {
691 /*
692 * Other types of ZONE_DEVICE page are not supported.
693 */
694 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
695 continue;
696 }
697
698 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
699 if (r != MIGRATEPAGE_SUCCESS)
700 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
701 }
702
703 /*
704 * No need to double call mmu_notifier->invalidate_range() callback as
705 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
706 * did already call it.
707 */
708 if (notified)
709 mmu_notifier_invalidate_range_only_end(&range);
710}
711EXPORT_SYMBOL(migrate_vma_pages);
712
713/**
714 * migrate_vma_finalize() - restore CPU page table entry
715 * @migrate: migrate struct containing all migration information
716 *
717 * This replaces the special migration pte entry with either a mapping to the
718 * new page if migration was successful for that page, or to the original page
719 * otherwise.
720 *
721 * This also unlocks the pages and puts them back on the lru, or drops the extra
722 * refcount, for device pages.
723 */
724void migrate_vma_finalize(struct migrate_vma *migrate)
725{
726 const unsigned long npages = migrate->npages;
727 unsigned long i;
728
729 for (i = 0; i < npages; i++) {
730 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
731 struct page *page = migrate_pfn_to_page(migrate->src[i]);
732
733 if (!page) {
734 if (newpage) {
735 unlock_page(newpage);
736 put_page(newpage);
737 }
738 continue;
739 }
740
741 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
742 if (newpage) {
743 unlock_page(newpage);
744 put_page(newpage);
745 }
746 newpage = page;
747 }
748
749 remove_migration_ptes(page, newpage, false);
750 unlock_page(page);
751
752 if (is_zone_device_page(page))
753 put_page(page);
754 else
755 putback_lru_page(page);
756
757 if (newpage != page) {
758 unlock_page(newpage);
759 if (is_zone_device_page(newpage))
760 put_page(newpage);
761 else
762 putback_lru_page(newpage);
763 }
764 }
765}
766EXPORT_SYMBOL(migrate_vma_finalize);