projects/linux·blob master · refs

1530 lines · 41.2 KiB

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.h>
11#include <linux/mm_inline.h>
12#include <linux/mmu_notifier.h>
13#include <linux/oom.h>
14#include <linux/pagewalk.h>
15#include <linux/rmap.h>
16#include <linux/leafops.h>
17#include <linux/pgalloc.h>
18#include <asm/tlbflush.h>
19#include "internal.h"
20
21static int migrate_vma_collect_skip(unsigned long start,
22 unsigned long end,
23 struct mm_walk *walk)
24{
25 struct migrate_vma *migrate = walk->private;
26 unsigned long addr;
27
28 for (addr = start; addr < end; addr += PAGE_SIZE) {
29 migrate->dst[migrate->npages] = 0;
30 migrate->src[migrate->npages++] = 0;
31 }
32
33 return 0;
34}
35
36static int migrate_vma_collect_hole(unsigned long start,
37 unsigned long end,
38 __always_unused int depth,
39 struct mm_walk *walk)
40{
41 struct migrate_vma *migrate = walk->private;
42 unsigned long addr;
43
44 /* Only allow populating anonymous memory. */
45 if (!vma_is_anonymous(walk->vma))
46 return migrate_vma_collect_skip(start, end, walk);
47
48 if (thp_migration_supported() &&
49 (migrate->flags & MIGRATE_VMA_SELECT_COMPOUND) &&
50 (IS_ALIGNED(start, HPAGE_PMD_SIZE) &&
51 IS_ALIGNED(end, HPAGE_PMD_SIZE))) {
52 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE |
53 MIGRATE_PFN_COMPOUND;
54 migrate->dst[migrate->npages] = 0;
55 migrate->npages++;
56 migrate->cpages++;
57
58 /*
59 * Collect the remaining entries as holes, in case we
60 * need to split later
61 */
62 return migrate_vma_collect_skip(start + PAGE_SIZE, end, walk);
63 }
64
65 for (addr = start; addr < end; addr += PAGE_SIZE) {
66 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
67 migrate->dst[migrate->npages] = 0;
68 migrate->npages++;
69 migrate->cpages++;
70 }
71
72 return 0;
73}
74
75/**
76 * migrate_vma_split_folio() - Helper function to split a THP folio
77 * @folio: the folio to split
78 * @fault_page: struct page associated with the fault if any
79 *
80 * If @folio is not the folio containing @fault_page, the caller must hold a
81 * reference on @folio. The helper consumes that reference.
82 *
83 * Returns 0 on success
84 */
85static int migrate_vma_split_folio(struct folio *folio,
86 struct page *fault_page)
87{
88 int ret;
89 struct folio *fault_folio = fault_page ? page_folio(fault_page) : NULL;
90 struct folio *new_fault_folio = NULL;
91
92 if (folio != fault_folio)
93 folio_lock(folio);
94
95 ret = split_folio(folio);
96 if (ret) {
97 if (folio != fault_folio) {
98 folio_unlock(folio);
99 folio_put(folio);
100 }
101 return ret;
102 }
103
104 new_fault_folio = fault_page ? page_folio(fault_page) : NULL;
105
106 /*
107 * Ensure the lock is held on the correct
108 * folio after the split
109 */
110 if (!new_fault_folio) {
111 folio_unlock(folio);
112 folio_put(folio);
113 } else if (folio != new_fault_folio) {
114 if (new_fault_folio != fault_folio) {
115 folio_get(new_fault_folio);
116 folio_lock(new_fault_folio);
117 }
118 folio_unlock(folio);
119 folio_put(folio);
120 }
121
122 return 0;
123}
124
125/** migrate_vma_collect_huge_pmd - collect THP pages without splitting the
126 * folio for device private pages.
127 * @pmdp: pointer to pmd entry
128 * @start: start address of the range for migration
129 * @end: end address of the range for migration
130 * @walk: mm_walk callback structure
131 * @fault_folio: folio associated with the fault if any
132 *
133 * Collect the huge pmd entry at @pmdp for migration and set the
134 * MIGRATE_PFN_COMPOUND flag in the migrate src entry to indicate that
135 * migration will occur at HPAGE_PMD granularity
136 */
137static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start,
138 unsigned long end, struct mm_walk *walk,
139 struct folio *fault_folio)
140{
141 struct mm_struct *mm = walk->mm;
142 struct folio *folio;
143 struct migrate_vma *migrate = walk->private;
144 spinlock_t *ptl;
145 int ret;
146 unsigned long write = 0;
147
148 ptl = pmd_lock(mm, pmdp);
149 if (pmd_none(*pmdp)) {
150 spin_unlock(ptl);
151 return migrate_vma_collect_hole(start, end, -1, walk);
152 }
153
154 if (pmd_trans_huge(*pmdp)) {
155 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
156 spin_unlock(ptl);
157 return migrate_vma_collect_skip(start, end, walk);
158 }
159
160 folio = pmd_folio(*pmdp);
161 if (is_huge_zero_folio(folio)) {
162 spin_unlock(ptl);
163 return migrate_vma_collect_hole(start, end, -1, walk);
164 }
165 if (pmd_write(*pmdp))
166 write = MIGRATE_PFN_WRITE;
167 } else if (!pmd_present(*pmdp)) {
168 const softleaf_t entry = softleaf_from_pmd(*pmdp);
169
170 if (!softleaf_is_device_private(entry) ||
171 !(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_PRIVATE)) {
172 spin_unlock(ptl);
173 return migrate_vma_collect_skip(start, end, walk);
174 }
175
176 folio = softleaf_to_folio(entry);
177 if (folio->pgmap->owner != migrate->pgmap_owner) {
178 spin_unlock(ptl);
179 return migrate_vma_collect_skip(start, end, walk);
180 }
181
182 if (softleaf_is_device_private_write(entry))
183 write = MIGRATE_PFN_WRITE;
184 } else {
185 spin_unlock(ptl);
186 return -EAGAIN;
187 }
188
189 folio_get(folio);
190 if (folio != fault_folio && unlikely(!folio_trylock(folio))) {
191 spin_unlock(ptl);
192 folio_put(folio);
193 return migrate_vma_collect_skip(start, end, walk);
194 }
195
196 if (thp_migration_supported() &&
197 (migrate->flags & MIGRATE_VMA_SELECT_COMPOUND) &&
198 (IS_ALIGNED(start, HPAGE_PMD_SIZE) &&
199 IS_ALIGNED(end, HPAGE_PMD_SIZE))) {
200
201 struct page_vma_mapped_walk pvmw = {
202 .ptl = ptl,
203 .address = start,
204 .pmd = pmdp,
205 .vma = walk->vma,
206 };
207
208 unsigned long pfn = page_to_pfn(folio_page(folio, 0));
209
210 migrate->src[migrate->npages] = migrate_pfn(pfn) | write
211 | MIGRATE_PFN_MIGRATE
212 | MIGRATE_PFN_COMPOUND;
213 migrate->dst[migrate->npages++] = 0;
214 migrate->cpages++;
215 ret = set_pmd_migration_entry(&pvmw, folio_page(folio, 0));
216 if (ret) {
217 migrate->npages--;
218 migrate->cpages--;
219 migrate->src[migrate->npages] = 0;
220 migrate->dst[migrate->npages] = 0;
221 goto fallback;
222 }
223 migrate_vma_collect_skip(start + PAGE_SIZE, end, walk);
224 spin_unlock(ptl);
225 return 0;
226 }
227
228fallback:
229 spin_unlock(ptl);
230 if (!folio_test_large(folio))
231 goto done;
232 ret = split_folio(folio);
233 if (fault_folio != folio)
234 folio_unlock(folio);
235 folio_put(folio);
236 if (ret)
237 return migrate_vma_collect_skip(start, end, walk);
238 if (pmd_none(pmdp_get_lockless(pmdp)))
239 return migrate_vma_collect_hole(start, end, -1, walk);
240
241done:
242 return -ENOENT;
243}
244
245static int migrate_vma_collect_pmd(pmd_t *pmdp,
246 unsigned long start,
247 unsigned long end,
248 struct mm_walk *walk)
249{
250 struct migrate_vma *migrate = walk->private;
251 struct vm_area_struct *vma = walk->vma;
252 struct mm_struct *mm = vma->vm_mm;
253 unsigned long addr = start, unmapped = 0;
254 spinlock_t *ptl;
255 struct folio *fault_folio = migrate->fault_page ?
256 page_folio(migrate->fault_page) : NULL;
257 pte_t *ptep;
258
259again:
260 if (pmd_trans_huge(*pmdp) || !pmd_present(*pmdp)) {
261 int ret = migrate_vma_collect_huge_pmd(pmdp, start, end, walk, fault_folio);
262
263 if (ret == -EAGAIN)
264 goto again;
265 if (ret == 0)
266 return 0;
267 }
268
269 ptep = pte_offset_map_lock(mm, pmdp, start, &ptl);
270 if (!ptep)
271 goto again;
272 lazy_mmu_mode_enable();
273 ptep += (addr - start) / PAGE_SIZE;
274
275 for (; addr < end; addr += PAGE_SIZE, ptep++) {
276 struct dev_pagemap *pgmap;
277 unsigned long mpfn = 0, pfn;
278 struct folio *folio;
279 struct page *page;
280 softleaf_t entry;
281 pte_t pte;
282
283 pte = ptep_get(ptep);
284
285 if (pte_none(pte)) {
286 if (vma_is_anonymous(vma)) {
287 mpfn = MIGRATE_PFN_MIGRATE;
288 migrate->cpages++;
289 }
290 goto next;
291 }
292
293 if (!pte_present(pte)) {
294 /*
295 * Only care about unaddressable device page special
296 * page table entry. Other special swap entries are not
297 * migratable, and we ignore regular swapped page.
298 */
299 entry = softleaf_from_pte(pte);
300 if (!softleaf_is_device_private(entry))
301 goto next;
302
303 page = softleaf_to_page(entry);
304 pgmap = page_pgmap(page);
305 if (!(migrate->flags &
306 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
307 pgmap->owner != migrate->pgmap_owner)
308 goto next;
309
310 folio = page_folio(page);
311 if (folio_test_large(folio)) {
312 int ret;
313
314 /* migrate_vma_split_folio() consumes this reference */
315 if (folio != fault_folio)
316 folio_get(folio);
317 lazy_mmu_mode_disable();
318 pte_unmap_unlock(ptep, ptl);
319 ret = migrate_vma_split_folio(folio,
320 migrate->fault_page);
321
322 if (ret) {
323 if (unmapped)
324 flush_tlb_range(walk->vma, start, end);
325
326 return migrate_vma_collect_skip(addr, end, walk);
327 }
328
329 goto again;
330 }
331
332 mpfn = migrate_pfn(page_to_pfn(page)) |
333 MIGRATE_PFN_MIGRATE;
334 if (softleaf_is_device_private_write(entry))
335 mpfn |= MIGRATE_PFN_WRITE;
336 } else {
337 pfn = pte_pfn(pte);
338 if (is_zero_pfn(pfn) &&
339 (migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
340 mpfn = MIGRATE_PFN_MIGRATE;
341 migrate->cpages++;
342 goto next;
343 }
344 page = vm_normal_page(migrate->vma, addr, pte);
345 if (page && !is_zone_device_page(page) &&
346 !(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
347 goto next;
348 } else if (page && is_device_coherent_page(page)) {
349 pgmap = page_pgmap(page);
350
351 if (!(migrate->flags &
352 MIGRATE_VMA_SELECT_DEVICE_COHERENT) ||
353 pgmap->owner != migrate->pgmap_owner)
354 goto next;
355 }
356 folio = page ? page_folio(page) : NULL;
357 if (folio && folio_test_large(folio)) {
358 int ret;
359
360 /* migrate_vma_split_folio() consumes this reference */
361 if (folio != fault_folio)
362 folio_get(folio);
363 lazy_mmu_mode_disable();
364 pte_unmap_unlock(ptep, ptl);
365 ret = migrate_vma_split_folio(folio,
366 migrate->fault_page);
367
368 if (ret) {
369 if (unmapped)
370 flush_tlb_range(walk->vma, start, end);
371
372 return migrate_vma_collect_skip(addr, end, walk);
373 }
374
375 goto again;
376 }
377 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
378 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
379 }
380
381 if (!page || !page->mapping) {
382 mpfn = 0;
383 goto next;
384 }
385
386 /*
387 * By getting a reference on the folio we pin it and that blocks
388 * any kind of migration. Side effect is that it "freezes" the
389 * pte.
390 *
391 * We drop this reference after isolating the folio from the lru
392 * for non device folio (device folio are not on the lru and thus
393 * can't be dropped from it).
394 */
395 folio = page_folio(page);
396 folio_get(folio);
397
398 /*
399 * We rely on folio_trylock() to avoid deadlock between
400 * concurrent migrations where each is waiting on the others
401 * folio lock. If we can't immediately lock the folio we fail this
402 * migration as it is only best effort anyway.
403 *
404 * If we can lock the folio it's safe to set up a migration entry
405 * now. In the common case where the folio is mapped once in a
406 * single process setting up the migration entry now is an
407 * optimisation to avoid walking the rmap later with
408 * try_to_migrate().
409 */
410 if (fault_folio == folio || folio_trylock(folio)) {
411 bool anon_exclusive;
412 pte_t swp_pte;
413
414 if (pte_present(pte))
415 flush_cache_page(vma, addr, pte_pfn(pte));
416 anon_exclusive = folio_test_anon(folio) &&
417 PageAnonExclusive(page);
418 if (anon_exclusive) {
419 pte = ptep_clear_flush(vma, addr, ptep);
420
421 if (folio_try_share_anon_rmap_pte(folio, page)) {
422 set_pte_at(mm, addr, ptep, pte);
423 if (fault_folio != folio)
424 folio_unlock(folio);
425 folio_put(folio);
426 mpfn = 0;
427 goto next;
428 }
429 } else {
430 pte = ptep_get_and_clear(mm, addr, ptep);
431 }
432
433 migrate->cpages++;
434
435 /* Set the dirty flag on the folio now the pte is gone. */
436 if (pte_present(pte) && pte_dirty(pte))
437 folio_mark_dirty(folio);
438
439 /* Setup special migration page table entry */
440 if (mpfn & MIGRATE_PFN_WRITE)
441 entry = make_writable_migration_entry(
442 page_to_pfn(page));
443 else if (anon_exclusive)
444 entry = make_readable_exclusive_migration_entry(
445 page_to_pfn(page));
446 else
447 entry = make_readable_migration_entry(
448 page_to_pfn(page));
449 if (pte_present(pte)) {
450 if (pte_young(pte))
451 entry = make_migration_entry_young(entry);
452 if (pte_dirty(pte))
453 entry = make_migration_entry_dirty(entry);
454 }
455 swp_pte = swp_entry_to_pte(entry);
456 if (pte_present(pte)) {
457 if (pte_soft_dirty(pte))
458 swp_pte = pte_swp_mksoft_dirty(swp_pte);
459 if (pte_uffd(pte))
460 swp_pte = pte_swp_mkuffd(swp_pte);
461 } else {
462 if (pte_swp_soft_dirty(pte))
463 swp_pte = pte_swp_mksoft_dirty(swp_pte);
464 if (pte_swp_uffd(pte))
465 swp_pte = pte_swp_mkuffd(swp_pte);
466 }
467 set_pte_at(mm, addr, ptep, swp_pte);
468
469 /*
470 * This is like regular unmap: we remove the rmap and
471 * drop the folio refcount. The folio won't be freed, as
472 * we took a reference just above.
473 */
474 folio_remove_rmap_pte(folio, page, vma);
475 folio_put(folio);
476
477 if (pte_present(pte))
478 unmapped++;
479 } else {
480 folio_put(folio);
481 mpfn = 0;
482 }
483
484next:
485 migrate->dst[migrate->npages] = 0;
486 migrate->src[migrate->npages++] = mpfn;
487 }
488
489 /* Only flush the TLB if we actually modified any entries */
490 if (unmapped)
491 flush_tlb_range(walk->vma, start, end);
492
493 lazy_mmu_mode_disable();
494 pte_unmap_unlock(ptep - 1, ptl);
495
496 return 0;
497}
498
499static const struct mm_walk_ops migrate_vma_walk_ops = {
500 .pmd_entry = migrate_vma_collect_pmd,
501 .pte_hole = migrate_vma_collect_hole,
502 .walk_lock = PGWALK_RDLOCK,
503};
504
505/*
506 * migrate_vma_collect() - collect pages over a range of virtual addresses
507 * @migrate: migrate struct containing all migration information
508 *
509 * This will walk the CPU page table. For each virtual address backed by a
510 * valid page, it updates the src array and takes a reference on the page, in
511 * order to pin the page until we lock it and unmap it.
512 */
513static void migrate_vma_collect(struct migrate_vma *migrate)
514{
515 struct mmu_notifier_range range;
516
517 /*
518 * Note that the pgmap_owner is passed to the mmu notifier callback so
519 * that the registered device driver can skip invalidating device
520 * private page mappings that won't be migrated.
521 */
522 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
523 migrate->vma->vm_mm, migrate->start, migrate->end,
524 migrate->pgmap_owner);
525 mmu_notifier_invalidate_range_start(&range);
526
527 walk_page_range_vma(migrate->vma, migrate->start, migrate->end,
528 &migrate_vma_walk_ops, migrate);
529
530 mmu_notifier_invalidate_range_end(&range);
531 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
532}
533
534/*
535 * migrate_vma_check_page() - check if page is pinned or not
536 * @page: struct page to check
537 *
538 * Pinned pages cannot be migrated. This is the same test as in
539 * folio_migrate_mapping(), except that here we allow migration of a
540 * ZONE_DEVICE page.
541 */
542static bool migrate_vma_check_page(struct page *page, struct page *fault_page)
543{
544 struct folio *folio = page_folio(page);
545
546 /*
547 * One extra ref because caller holds an extra reference, either from
548 * folio_isolate_lru() for a regular folio, or migrate_vma_collect() for
549 * a device folio.
550 */
551 int extra = 1 + (page == fault_page);
552
553 /* Page from ZONE_DEVICE have one extra reference */
554 if (folio_is_zone_device(folio))
555 extra++;
556
557 /* For file back page */
558 if (folio_mapping(folio))
559 extra += 1 + folio_has_private(folio);
560
561 if ((folio_ref_count(folio) - extra) > folio_mapcount(folio))
562 return false;
563
564 return true;
565}
566
567/*
568 * Unmaps pages for migration. Returns number of source pfns marked as
569 * migrating.
570 */
571static unsigned long migrate_device_unmap(unsigned long *src_pfns,
572 unsigned long npages,
573 struct page *fault_page)
574{
575 struct folio *fault_folio = fault_page ?
576 page_folio(fault_page) : NULL;
577 unsigned long i, restore = 0;
578 bool allow_drain = true;
579 unsigned long unmapped = 0;
580
581 lru_add_drain();
582
583 for (i = 0; i < npages; ) {
584 struct page *page = migrate_pfn_to_page(src_pfns[i]);
585 struct folio *folio;
586 unsigned int nr = 1;
587
588 if (!page) {
589 if (src_pfns[i] & MIGRATE_PFN_MIGRATE)
590 unmapped++;
591 goto next;
592 }
593
594 folio = page_folio(page);
595 nr = folio_nr_pages(folio);
596
597 if (nr > 1)
598 src_pfns[i] |= MIGRATE_PFN_COMPOUND;
599
600
601 /* ZONE_DEVICE folios are not on LRU */
602 if (!folio_is_zone_device(folio)) {
603 if (!folio_test_lru(folio) && allow_drain) {
604 /* Drain CPU's lru cache */
605 lru_add_drain_all();
606 allow_drain = false;
607 }
608
609 if (!folio_isolate_lru(folio)) {
610 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
611 restore++;
612 goto next;
613 }
614
615 /* Drop the reference we took in collect */
616 folio_put(folio);
617 }
618
619 if (folio_mapped(folio))
620 try_to_migrate(folio, 0);
621
622 if (folio_mapped(folio) ||
623 !migrate_vma_check_page(page, fault_page)) {
624 if (!folio_is_zone_device(folio)) {
625 folio_get(folio);
626 folio_putback_lru(folio);
627 }
628
629 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
630 restore++;
631 goto next;
632 }
633
634 unmapped++;
635next:
636 i += nr;
637 }
638
639 for (i = 0; i < npages && restore; i++) {
640 struct page *page = migrate_pfn_to_page(src_pfns[i]);
641 struct folio *folio;
642
643 if (!page || (src_pfns[i] & MIGRATE_PFN_MIGRATE))
644 continue;
645
646 folio = page_folio(page);
647 remove_migration_ptes(folio, folio, 0);
648
649 src_pfns[i] = 0;
650 if (fault_folio != folio)
651 folio_unlock(folio);
652 folio_put(folio);
653 restore--;
654 }
655
656 return unmapped;
657}
658
659/*
660 * migrate_vma_unmap() - replace page mapping with special migration pte entry
661 * @migrate: migrate struct containing all migration information
662 *
663 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
664 * special migration pte entry and check if it has been pinned. Pinned pages are
665 * restored because we cannot migrate them.
666 *
667 * This is the last step before we call the device driver callback to allocate
668 * destination memory and copy contents of original page over to new page.
669 */
670static void migrate_vma_unmap(struct migrate_vma *migrate)
671{
672 migrate->cpages = migrate_device_unmap(migrate->src, migrate->npages,
673 migrate->fault_page);
674}
675
676/**
677 * migrate_vma_setup() - prepare to migrate a range of memory
678 * @args: contains the vma, start, and pfns arrays for the migration
679 *
680 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
681 * without an error.
682 *
683 * Prepare to migrate a range of memory virtual address range by collecting all
684 * the pages backing each virtual address in the range, saving them inside the
685 * src array. Then lock those pages and unmap them. Once the pages are locked
686 * and unmapped, check whether each page is pinned or not. Pages that aren't
687 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
688 * corresponding src array entry. Then restores any pages that are pinned, by
689 * remapping and unlocking those pages.
690 *
691 * The caller should then allocate destination memory and copy source memory to
692 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
693 * flag set). Once these are allocated and copied, the caller must update each
694 * corresponding entry in the dst array with the pfn value of the destination
695 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
696 * lock_page().
697 *
698 * Note that the caller does not have to migrate all the pages that are marked
699 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
700 * device memory to system memory. If the caller cannot migrate a device page
701 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
702 * consequences for the userspace process, so it must be avoided if at all
703 * possible.
704 *
705 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
706 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
707 * allowing the caller to allocate device memory for those unbacked virtual
708 * addresses. For this the caller simply has to allocate device memory and
709 * properly set the destination entry like for regular migration. Note that
710 * this can still fail, and thus inside the device driver you must check if the
711 * migration was successful for those entries after calling migrate_vma_pages(),
712 * just like for regular migration.
713 *
714 * After that, the callers must call migrate_vma_pages() to go over each entry
715 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
716 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
717 * then migrate_vma_pages() to migrate struct page information from the source
718 * struct page to the destination struct page. If it fails to migrate the
719 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
720 * src array.
721 *
722 * At this point all successfully migrated pages have an entry in the src
723 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
724 * array entry with MIGRATE_PFN_VALID flag set.
725 *
726 * Once migrate_vma_pages() returns the caller may inspect which pages were
727 * successfully migrated, and which were not. Successfully migrated pages will
728 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
729 *
730 * It is safe to update device page table after migrate_vma_pages() because
731 * both destination and source page are still locked, and the mmap_lock is held
732 * in read mode (hence no one can unmap the range being migrated).
733 *
734 * Once the caller is done cleaning up things and updating its page table (if it
735 * chose to do so, this is not an obligation) it finally calls
736 * migrate_vma_finalize() to update the CPU page table to point to new pages
737 * for successfully migrated pages or otherwise restore the CPU page table to
738 * point to the original source pages.
739 */
740int migrate_vma_setup(struct migrate_vma *args)
741{
742 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
743
744 args->start &= PAGE_MASK;
745 args->end &= PAGE_MASK;
746 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
747 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
748 return -EINVAL;
749 if (nr_pages <= 0)
750 return -EINVAL;
751 if (args->start < args->vma->vm_start ||
752 args->start >= args->vma->vm_end)
753 return -EINVAL;
754 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
755 return -EINVAL;
756 if (!args->src || !args->dst)
757 return -EINVAL;
758 if (args->fault_page && !is_device_private_page(args->fault_page))
759 return -EINVAL;
760 if (args->fault_page && !PageLocked(args->fault_page))
761 return -EINVAL;
762
763 memset(args->src, 0, sizeof(*args->src) * nr_pages);
764 args->cpages = 0;
765 args->npages = 0;
766
767 migrate_vma_collect(args);
768
769 if (args->cpages)
770 migrate_vma_unmap(args);
771
772 /*
773 * At this point pages are locked and unmapped, and thus they have
774 * stable content and can safely be copied to destination memory that
775 * is allocated by the drivers.
776 */
777 return 0;
778
779}
780EXPORT_SYMBOL(migrate_vma_setup);
781
782#ifdef CONFIG_ARCH_HAS_PMD_SOFTLEAVES
783/**
784 * migrate_vma_insert_huge_pmd_page: Insert a huge folio into @migrate->vma->vm_mm
785 * at @addr. folio is already allocated as a part of the migration process with
786 * large page.
787 *
788 * @page needs to be initialized and setup after it's allocated. The code bits
789 * here follow closely the code in __do_huge_pmd_anonymous_page(). This API does
790 * not support THP zero pages.
791 *
792 * @migrate: migrate_vma arguments
793 * @addr: address where the folio will be inserted
794 * @page: page to be inserted at @addr
795 * @src: src pfn which is being migrated
796 * @pmdp: pointer to the pmd
797 */
798static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
799 unsigned long addr,
800 struct page *page,
801 unsigned long *src,
802 pmd_t *pmdp)
803{
804 struct vm_area_struct *vma = migrate->vma;
805 gfp_t gfp = vma_thp_gfp_mask(vma);
806 struct folio *folio = page_folio(page);
807 int ret;
808 vm_fault_t csa_ret;
809 spinlock_t *ptl;
810 pgtable_t pgtable;
811 pmd_t entry;
812 bool flush = false;
813 unsigned long i;
814
815 VM_WARN_ON_ONCE(!folio);
816
817 if (!thp_vma_suitable_order(vma, addr, HPAGE_PMD_ORDER))
818 return -EINVAL;
819
820 ret = anon_vma_prepare(vma);
821 if (ret)
822 return ret;
823
824 folio_set_order(folio, HPAGE_PMD_ORDER);
825 folio_set_large_rmappable(folio);
826
827 if (mem_cgroup_charge(folio, migrate->vma->vm_mm, gfp)) {
828 count_vm_event(THP_FAULT_FALLBACK);
829 count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
830 ret = -ENOMEM;
831 goto abort;
832 }
833
834 __folio_mark_uptodate(folio);
835
836 pgtable = pte_alloc_one(vma->vm_mm);
837 if (unlikely(!pgtable))
838 goto abort;
839
840 if (folio_is_device_private(folio)) {
841 swp_entry_t swp_entry;
842
843 if (vma->vm_flags & VM_WRITE)
844 swp_entry = make_writable_device_private_entry(
845 page_to_pfn(page));
846 else
847 swp_entry = make_readable_device_private_entry(
848 page_to_pfn(page));
849 entry = softleaf_to_pmd(swp_entry);
850 } else {
851 if (folio_is_zone_device(folio) &&
852 !folio_is_device_coherent(folio)) {
853 goto free_abort;
854 }
855 entry = folio_mk_pmd(folio, vma->vm_page_prot);
856 if (vma->vm_flags & VM_WRITE)
857 entry = pmd_mkwrite(pmd_mkdirty(entry), vma);
858 }
859
860 ptl = pmd_lock(vma->vm_mm, pmdp);
861 csa_ret = check_stable_address_space(vma->vm_mm);
862 if (csa_ret)
863 goto unlock_abort;
864
865 /*
866 * Check for userfaultfd but do not deliver the fault. Instead,
867 * just back off.
868 */
869 if (userfaultfd_missing(vma))
870 goto unlock_abort;
871
872 if (is_huge_zero_pmd(*pmdp))
873 flush = true;
874 else if (!pmd_none(*pmdp))
875 goto unlock_abort;
876
877 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
878 folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
879 if (!folio_is_zone_device(folio))
880 folio_add_lru_vma(folio, vma);
881 folio_get(folio);
882
883 if (flush) {
884 pte_free(vma->vm_mm, pgtable);
885 flush_cache_range(vma, addr, addr + HPAGE_PMD_SIZE);
886 pmdp_invalidate(vma, addr, pmdp);
887 } else {
888 pgtable_trans_huge_deposit(vma->vm_mm, pmdp, pgtable);
889 mm_inc_nr_ptes(vma->vm_mm);
890 }
891 set_pmd_at(vma->vm_mm, addr, pmdp, entry);
892 update_mmu_cache_pmd(vma, addr, pmdp);
893
894 spin_unlock(ptl);
895
896 count_vm_event(THP_FAULT_ALLOC);
897 count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_ALLOC);
898 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
899
900 return 0;
901
902unlock_abort:
903 spin_unlock(ptl);
904free_abort:
905 pte_free(vma->vm_mm, pgtable);
906abort:
907 for (i = 0; i < HPAGE_PMD_NR; i++)
908 src[i] &= ~MIGRATE_PFN_MIGRATE;
909 return 0;
910}
911
912static int migrate_vma_split_unmapped_folio(struct migrate_vma *migrate,
913 unsigned long idx, unsigned long addr,
914 struct folio *folio)
915{
916 unsigned long i;
917 unsigned long pfn;
918 unsigned long flags;
919 int ret = 0;
920
921 /*
922 * take a reference, since split_huge_pmd_address() with freeze = true
923 * drops a reference at the end.
924 */
925 folio_get(folio);
926 split_huge_pmd_address(migrate->vma, addr, true);
927 ret = folio_split_unmapped(folio, 0);
928 if (ret)
929 return ret;
930 migrate->src[idx] &= ~MIGRATE_PFN_COMPOUND;
931 flags = migrate->src[idx] & ((1UL << MIGRATE_PFN_SHIFT) - 1);
932 pfn = migrate->src[idx] >> MIGRATE_PFN_SHIFT;
933 for (i = 1; i < HPAGE_PMD_NR; i++)
934 migrate->src[i+idx] = migrate_pfn(pfn + i) | flags;
935 return ret;
936}
937#else /* !CONFIG_ARCH_HAS_PMD_SOFTLEAVES */
938static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
939 unsigned long addr,
940 struct page *page,
941 unsigned long *src,
942 pmd_t *pmdp)
943{
944 return 0;
945}
946
947static int migrate_vma_split_unmapped_folio(struct migrate_vma *migrate,
948 unsigned long idx, unsigned long addr,
949 struct folio *folio)
950{
951 return 0;
952}
953#endif
954
955static unsigned long migrate_vma_nr_pages(unsigned long *src)
956{
957 unsigned long nr = 1;
958#ifdef CONFIG_ARCH_HAS_PMD_SOFTLEAVES
959 if (*src & MIGRATE_PFN_COMPOUND)
960 nr = HPAGE_PMD_NR;
961#else
962 if (*src & MIGRATE_PFN_COMPOUND)
963 VM_WARN_ON_ONCE(true);
964#endif
965 return nr;
966}
967
968/*
969 * This code closely matches the code in:
970 * __handle_mm_fault()
971 * handle_pte_fault()
972 * do_anonymous_page()
973 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
974 * private or coherent page.
975 */
976static void migrate_vma_insert_page(struct migrate_vma *migrate,
977 unsigned long addr,
978 unsigned long *dst,
979 unsigned long *src)
980{
981 struct page *page = migrate_pfn_to_page(*dst);
982 struct folio *folio = page_folio(page);
983 struct vm_area_struct *vma = migrate->vma;
984 struct mm_struct *mm = vma->vm_mm;
985 bool flush = false;
986 spinlock_t *ptl;
987 pte_t entry;
988 pgd_t *pgdp;
989 p4d_t *p4dp;
990 pud_t *pudp;
991 pmd_t *pmdp;
992 pte_t *ptep;
993 pte_t orig_pte;
994
995 /* Only allow populating anonymous memory */
996 if (!vma_is_anonymous(vma))
997 goto abort;
998
999 pgdp = pgd_offset(mm, addr);
1000 p4dp = p4d_alloc(mm, pgdp, addr);
1001 if (!p4dp)
1002 goto abort;
1003 pudp = pud_alloc(mm, p4dp, addr);
1004 if (!pudp)
1005 goto abort;
1006 pmdp = pmd_alloc(mm, pudp, addr);
1007 if (!pmdp)
1008 goto abort;
1009
1010 if (thp_migration_supported() && (*dst & MIGRATE_PFN_COMPOUND)) {
1011 int ret = migrate_vma_insert_huge_pmd_page(migrate, addr, page,
1012 src, pmdp);
1013 if (ret)
1014 goto abort;
1015 return;
1016 }
1017
1018 if (!pmd_none(*pmdp)) {
1019 if (pmd_trans_huge(*pmdp)) {
1020 if (!is_huge_zero_pmd(*pmdp))
1021 goto abort;
1022 split_huge_pmd(vma, pmdp, addr);
1023 } else if (pmd_leaf(*pmdp))
1024 goto abort;
1025 }
1026
1027 if (pte_alloc(mm, pmdp))
1028 goto abort;
1029 if (unlikely(anon_vma_prepare(vma)))
1030 goto abort;
1031 if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
1032 goto abort;
1033
1034 /*
1035 * The memory barrier inside __folio_mark_uptodate makes sure that
1036 * preceding stores to the folio contents become visible before
1037 * the set_pte_at() write.
1038 */
1039 __folio_mark_uptodate(folio);
1040
1041 if (folio_is_device_private(folio)) {
1042 swp_entry_t swp_entry;
1043
1044 if (vma->vm_flags & VM_WRITE)
1045 swp_entry = make_writable_device_private_entry(
1046 page_to_pfn(page));
1047 else
1048 swp_entry = make_readable_device_private_entry(
1049 page_to_pfn(page));
1050 entry = swp_entry_to_pte(swp_entry);
1051 } else {
1052 if (folio_is_zone_device(folio) &&
1053 !folio_is_device_coherent(folio)) {
1054 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
1055 goto abort;
1056 }
1057 entry = mk_pte(page, vma->vm_page_prot);
1058 if (vma->vm_flags & VM_WRITE)
1059 entry = pte_mkwrite(pte_mkdirty(entry), vma);
1060 }
1061
1062 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
1063 if (!ptep)
1064 goto abort;
1065 orig_pte = ptep_get(ptep);
1066
1067 if (check_stable_address_space(mm))
1068 goto unlock_abort;
1069
1070 if (pte_present(orig_pte)) {
1071 unsigned long pfn = pte_pfn(orig_pte);
1072
1073 if (!is_zero_pfn(pfn))
1074 goto unlock_abort;
1075 flush = true;
1076 } else if (!pte_none(orig_pte))
1077 goto unlock_abort;
1078
1079 /*
1080 * Check for userfaultfd but do not deliver the fault. Instead,
1081 * just back off.
1082 */
1083 if (userfaultfd_missing(vma))
1084 goto unlock_abort;
1085
1086 inc_mm_counter(mm, MM_ANONPAGES);
1087 folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
1088 if (!folio_is_zone_device(folio))
1089 folio_add_lru_vma(folio, vma);
1090 folio_get(folio);
1091
1092 if (flush) {
1093 flush_cache_page(vma, addr, pte_pfn(orig_pte));
1094 ptep_clear_flush(vma, addr, ptep);
1095 }
1096 set_pte_at(mm, addr, ptep, entry);
1097 update_mmu_cache(vma, addr, ptep);
1098
1099 pte_unmap_unlock(ptep, ptl);
1100 *src = MIGRATE_PFN_MIGRATE;
1101 return;
1102
1103unlock_abort:
1104 pte_unmap_unlock(ptep, ptl);
1105abort:
1106 *src &= ~MIGRATE_PFN_MIGRATE;
1107}
1108
1109static void __migrate_device_pages(unsigned long *src_pfns,
1110 unsigned long *dst_pfns, unsigned long npages,
1111 struct migrate_vma *migrate)
1112{
1113 struct mmu_notifier_range range;
1114 unsigned long i, j;
1115 bool notified = false;
1116 unsigned long addr;
1117
1118 for (i = 0; i < npages; ) {
1119 struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
1120 struct page *page = migrate_pfn_to_page(src_pfns[i]);
1121 struct address_space *mapping;
1122 struct folio *newfolio, *folio;
1123 int r, extra_cnt = 0;
1124 unsigned long nr = 1;
1125
1126 if (!newpage) {
1127 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
1128 goto next;
1129 }
1130
1131 if (!page) {
1132 unsigned long addr;
1133
1134 if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE))
1135 goto next;
1136
1137 /*
1138 * The only time there is no vma is when called from
1139 * migrate_device_coherent_folio(). However this isn't
1140 * called if the page could not be unmapped.
1141 */
1142 VM_BUG_ON(!migrate);
1143 addr = migrate->start + i*PAGE_SIZE;
1144 if (!notified) {
1145 notified = true;
1146
1147 mmu_notifier_range_init_owner(&range,
1148 MMU_NOTIFY_MIGRATE, 0,
1149 migrate->vma->vm_mm, addr, migrate->end,
1150 migrate->pgmap_owner);
1151 mmu_notifier_invalidate_range_start(&range);
1152 }
1153
1154 if ((src_pfns[i] & MIGRATE_PFN_COMPOUND) &&
1155 (!(dst_pfns[i] & MIGRATE_PFN_COMPOUND))) {
1156 nr = migrate_vma_nr_pages(&src_pfns[i]);
1157 src_pfns[i] &= ~MIGRATE_PFN_COMPOUND;
1158 } else {
1159 nr = 1;
1160 }
1161
1162 for (j = 0; j < nr && i + j < npages; j++) {
1163 src_pfns[i+j] |= MIGRATE_PFN_MIGRATE;
1164 migrate_vma_insert_page(migrate,
1165 addr + j * PAGE_SIZE,
1166 &dst_pfns[i+j], &src_pfns[i+j]);
1167 }
1168 goto next;
1169 }
1170
1171 newfolio = page_folio(newpage);
1172 folio = page_folio(page);
1173 mapping = folio_mapping(folio);
1174
1175 /*
1176 * If THP migration is enabled, check if both src and dst
1177 * can migrate large pages
1178 */
1179 if (thp_migration_supported()) {
1180 if ((src_pfns[i] & MIGRATE_PFN_MIGRATE) &&
1181 (src_pfns[i] & MIGRATE_PFN_COMPOUND) &&
1182 !(dst_pfns[i] & MIGRATE_PFN_COMPOUND)) {
1183
1184 if (!migrate) {
1185 src_pfns[i] &= ~(MIGRATE_PFN_MIGRATE |
1186 MIGRATE_PFN_COMPOUND);
1187 goto next;
1188 }
1189 nr = 1 << folio_order(folio);
1190 addr = migrate->start + i * PAGE_SIZE;
1191 if (migrate_vma_split_unmapped_folio(migrate, i, addr, folio)) {
1192 src_pfns[i] &= ~(MIGRATE_PFN_MIGRATE |
1193 MIGRATE_PFN_COMPOUND);
1194 goto next;
1195 }
1196
1197 /*
1198 * reset nr so that only first after-split folio
1199 * is processed below
1200 */
1201 VM_WARN_ON_ONCE(folio_test_large(folio));
1202 nr = 1;
1203 } else if ((src_pfns[i] & MIGRATE_PFN_MIGRATE) &&
1204 (dst_pfns[i] & MIGRATE_PFN_COMPOUND) &&
1205 !(src_pfns[i] & MIGRATE_PFN_COMPOUND)) {
1206 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
1207 }
1208 }
1209
1210
1211 if (folio_is_device_private(newfolio) ||
1212 folio_is_device_coherent(newfolio)) {
1213 if (mapping) {
1214 /*
1215 * For now only support anonymous memory migrating to
1216 * device private or coherent memory.
1217 *
1218 * Try to get rid of swap cache if possible.
1219 */
1220 if (!folio_test_anon(folio) ||
1221 !folio_free_swap(folio)) {
1222 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
1223 goto next;
1224 }
1225 }
1226 } else if (folio_is_zone_device(newfolio)) {
1227 /*
1228 * Other types of ZONE_DEVICE page are not supported.
1229 */
1230 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
1231 goto next;
1232 }
1233
1234 BUG_ON(folio_test_writeback(folio));
1235
1236 if (migrate && migrate->fault_page == page)
1237 extra_cnt = 1;
1238 for (j = 0; j < nr && i + j < npages; j++) {
1239 folio = page_folio(migrate_pfn_to_page(src_pfns[i+j]));
1240 newfolio = page_folio(migrate_pfn_to_page(dst_pfns[i+j]));
1241
1242 /*
1243 * folio_free_swap() removed the folio from the swap
1244 * cache. Refresh the saved mapping before migration.
1245 */
1246 mapping = folio_mapping(folio);
1247
1248 r = folio_migrate_mapping(mapping, newfolio, folio, extra_cnt);
1249 if (r)
1250 src_pfns[i+j] &= ~MIGRATE_PFN_MIGRATE;
1251 else
1252 folio_migrate_flags(newfolio, folio);
1253 }
1254next:
1255 i += nr;
1256 }
1257
1258 if (notified)
1259 mmu_notifier_invalidate_range_end(&range);
1260}
1261
1262/**
1263 * migrate_device_pages() - migrate meta-data from src page to dst page
1264 * @src_pfns: src_pfns returned from migrate_device_range()
1265 * @dst_pfns: array of pfns allocated by the driver to migrate memory to
1266 * @npages: number of pages in the range
1267 *
1268 * Equivalent to migrate_vma_pages(). This is called to migrate struct page
1269 * meta-data from source struct page to destination.
1270 */
1271void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns,
1272 unsigned long npages)
1273{
1274 __migrate_device_pages(src_pfns, dst_pfns, npages, NULL);
1275}
1276EXPORT_SYMBOL(migrate_device_pages);
1277
1278/**
1279 * migrate_vma_pages() - migrate meta-data from src page to dst page
1280 * @migrate: migrate struct containing all migration information
1281 *
1282 * This migrates struct page meta-data from source struct page to destination
1283 * struct page. This effectively finishes the migration from source page to the
1284 * destination page.
1285 */
1286void migrate_vma_pages(struct migrate_vma *migrate)
1287{
1288 __migrate_device_pages(migrate->src, migrate->dst, migrate->npages, migrate);
1289}
1290EXPORT_SYMBOL(migrate_vma_pages);
1291
1292static void __migrate_device_finalize(unsigned long *src_pfns,
1293 unsigned long *dst_pfns,
1294 unsigned long npages,
1295 struct page *fault_page)
1296{
1297 struct folio *fault_folio = fault_page ?
1298 page_folio(fault_page) : NULL;
1299 unsigned long i;
1300
1301 for (i = 0; i < npages; i++) {
1302 struct folio *dst = NULL, *src = NULL;
1303 struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
1304 struct page *page = migrate_pfn_to_page(src_pfns[i]);
1305
1306 if (newpage)
1307 dst = page_folio(newpage);
1308
1309 if (!page) {
1310 if (dst) {
1311 WARN_ON_ONCE(fault_folio == dst);
1312 folio_unlock(dst);
1313 folio_put(dst);
1314 }
1315 continue;
1316 }
1317
1318 src = page_folio(page);
1319
1320 if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE) || !dst) {
1321 if (dst) {
1322 WARN_ON_ONCE(fault_folio == dst);
1323 folio_unlock(dst);
1324 folio_put(dst);
1325 }
1326 dst = src;
1327 }
1328
1329 if (!folio_is_zone_device(dst))
1330 folio_add_lru(dst);
1331 remove_migration_ptes(src, dst, 0);
1332 if (fault_folio != src)
1333 folio_unlock(src);
1334 folio_put(src);
1335
1336 if (dst != src) {
1337 WARN_ON_ONCE(fault_folio == dst);
1338 folio_unlock(dst);
1339 folio_put(dst);
1340 }
1341 }
1342}
1343
1344/*
1345 * migrate_device_finalize() - complete page migration
1346 * @src_pfns: src_pfns returned from migrate_device_range()
1347 * @dst_pfns: array of pfns allocated by the driver to migrate memory to
1348 * @npages: number of pages in the range
1349 *
1350 * Completes migration of the page by removing special migration entries.
1351 * Drivers must ensure copying of page data is complete and visible to the CPU
1352 * before calling this.
1353 */
1354void migrate_device_finalize(unsigned long *src_pfns,
1355 unsigned long *dst_pfns, unsigned long npages)
1356{
1357 return __migrate_device_finalize(src_pfns, dst_pfns, npages, NULL);
1358}
1359EXPORT_SYMBOL(migrate_device_finalize);
1360
1361/**
1362 * migrate_vma_finalize() - restore CPU page table entry
1363 * @migrate: migrate struct containing all migration information
1364 *
1365 * This replaces the special migration pte entry with either a mapping to the
1366 * new page if migration was successful for that page, or to the original page
1367 * otherwise.
1368 *
1369 * This also unlocks the pages and puts them back on the lru, or drops the extra
1370 * refcount, for device pages.
1371 */
1372void migrate_vma_finalize(struct migrate_vma *migrate)
1373{
1374 __migrate_device_finalize(migrate->src, migrate->dst, migrate->npages,
1375 migrate->fault_page);
1376}
1377EXPORT_SYMBOL(migrate_vma_finalize);
1378
1379static unsigned long migrate_device_pfn_lock(unsigned long pfn)
1380{
1381 struct folio *folio;
1382
1383 folio = folio_get_nontail_page(pfn_to_page(pfn));
1384 if (!folio)
1385 return 0;
1386
1387 if (!folio_trylock(folio)) {
1388 folio_put(folio);
1389 return 0;
1390 }
1391
1392 return migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
1393}
1394
1395/**
1396 * migrate_device_range() - migrate device private pfns to normal memory.
1397 * @src_pfns: array large enough to hold migrating source device private pfns.
1398 * @start: starting pfn in the range to migrate.
1399 * @npages: number of pages to migrate.
1400 *
1401 * migrate_vma_setup() is similar in concept to migrate_vma_setup() except that
1402 * instead of looking up pages based on virtual address mappings a range of
1403 * device pfns that should be migrated to system memory is used instead.
1404 *
1405 * This is useful when a driver needs to free device memory but doesn't know the
1406 * virtual mappings of every page that may be in device memory. For example this
1407 * is often the case when a driver is being unloaded or unbound from a device.
1408 *
1409 * Like migrate_vma_setup() this function will take a reference and lock any
1410 * migrating pages that aren't free before unmapping them. Drivers may then
1411 * allocate destination pages and start copying data from the device to CPU
1412 * memory before calling migrate_device_pages().
1413 */
1414int migrate_device_range(unsigned long *src_pfns, unsigned long start,
1415 unsigned long npages)
1416{
1417 unsigned long i, j, pfn;
1418
1419 for (pfn = start, i = 0; i < npages; pfn++, i++) {
1420 struct page *page = pfn_to_page(pfn);
1421 struct folio *folio = page_folio(page);
1422 unsigned int nr = 1;
1423
1424 src_pfns[i] = migrate_device_pfn_lock(pfn);
1425 nr = folio_nr_pages(folio);
1426 if (nr > npages - i) {
1427 if (src_pfns[i] & MIGRATE_PFN_MIGRATE) {
1428 folio_unlock(folio);
1429 folio_put(folio);
1430 }
1431 memset(&src_pfns[i], 0,
1432 (npages - i) * sizeof(*src_pfns));
1433 break;
1434 }
1435 if (nr > 1) {
1436 src_pfns[i] |= MIGRATE_PFN_COMPOUND;
1437 for (j = 1; j < nr; j++)
1438 src_pfns[i+j] = 0;
1439 i += j - 1;
1440 pfn += j - 1;
1441 }
1442 }
1443
1444 migrate_device_unmap(src_pfns, npages, NULL);
1445
1446 return 0;
1447}
1448EXPORT_SYMBOL(migrate_device_range);
1449
1450/**
1451 * migrate_device_pfns() - migrate device private pfns to normal memory.
1452 * @src_pfns: pre-populated array of source device private pfns to migrate.
1453 * @npages: number of pages to migrate.
1454 *
1455 * Similar to migrate_device_range() but supports non-contiguous pre-populated
1456 * array of device pages to migrate.
1457 */
1458int migrate_device_pfns(unsigned long *src_pfns, unsigned long npages)
1459{
1460 unsigned long i, j;
1461
1462 for (i = 0; i < npages; i++) {
1463 struct page *page = pfn_to_page(src_pfns[i]);
1464 struct folio *folio = page_folio(page);
1465 unsigned int nr = 1;
1466
1467 src_pfns[i] = migrate_device_pfn_lock(src_pfns[i]);
1468 nr = folio_nr_pages(folio);
1469 if (nr > npages - i) {
1470 if (src_pfns[i] & MIGRATE_PFN_MIGRATE) {
1471 folio_unlock(folio);
1472 folio_put(folio);
1473 }
1474 memset(&src_pfns[i], 0,
1475 (npages - i) * sizeof(*src_pfns));
1476 break;
1477 }
1478 if (nr > 1) {
1479 src_pfns[i] |= MIGRATE_PFN_COMPOUND;
1480 for (j = 1; j < nr; j++)
1481 src_pfns[i+j] = 0;
1482 i += j - 1;
1483 }
1484 }
1485
1486 migrate_device_unmap(src_pfns, npages, NULL);
1487
1488 return 0;
1489}
1490EXPORT_SYMBOL(migrate_device_pfns);
1491
1492/*
1493 * Migrate a device coherent folio back to normal memory. The caller should have
1494 * a reference on folio which will be copied to the new folio if migration is
1495 * successful or dropped on failure.
1496 */
1497int migrate_device_coherent_folio(struct folio *folio)
1498{
1499 unsigned long src_pfn, dst_pfn = 0;
1500 struct folio *dfolio;
1501
1502 WARN_ON_ONCE(folio_test_large(folio));
1503
1504 folio_lock(folio);
1505 src_pfn = migrate_pfn(folio_pfn(folio)) | MIGRATE_PFN_MIGRATE;
1506
1507 /*
1508 * We don't have a VMA and don't need to walk the page tables to find
1509 * the source folio. So call migrate_vma_unmap() directly to unmap the
1510 * folio as migrate_vma_setup() will fail if args.vma == NULL.
1511 */
1512 migrate_device_unmap(&src_pfn, 1, NULL);
1513 if (!(src_pfn & MIGRATE_PFN_MIGRATE))
1514 return -EBUSY;
1515
1516 dfolio = folio_alloc(GFP_USER | __GFP_NOWARN, 0);
1517 if (dfolio) {
1518 folio_lock(dfolio);
1519 dst_pfn = migrate_pfn(folio_pfn(dfolio));
1520 }
1521
1522 migrate_device_pages(&src_pfn, &dst_pfn, 1);
1523 if (src_pfn & MIGRATE_PFN_MIGRATE)
1524 folio_copy(dfolio, folio);
1525 migrate_device_finalize(&src_pfn, &dst_pfn, 1);
1526
1527 if (src_pfn & MIGRATE_PFN_MIGRATE)
1528 return 0;
1529 return -EBUSY;
1530}