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