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/swapops.h>
17 #include <asm/tlbflush.h>
18 #include "internal.h"
19
20 static int migrate_vma_collect_skip(unsigned long start,
21 unsigned long end,
22 struct mm_walk *walk)
23 {
24 struct migrate_vma *migrate = walk->private;
25 unsigned long addr;
26
27 for (addr = start; addr < end; addr += PAGE_SIZE) {
28 migrate->dst[migrate->npages] = 0;
29 migrate->src[migrate->npages++] = 0;
30 }
31
32 return 0;
33 }
34
35 static int migrate_vma_collect_hole(unsigned long start,
36 unsigned long end,
37 __always_unused int depth,
38 struct mm_walk *walk)
39 {
40 struct migrate_vma *migrate = walk->private;
41 unsigned long addr;
42
43 /* Only allow populating anonymous memory. */
44 if (!vma_is_anonymous(walk->vma))
45 return migrate_vma_collect_skip(start, end, walk);
46
47 for (addr = start; addr < end; addr += PAGE_SIZE) {
48 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
49 migrate->dst[migrate->npages] = 0;
50 migrate->npages++;
51 migrate->cpages++;
52 }
53
54 return 0;
55 }
56
57 static int migrate_vma_collect_pmd(pmd_t *pmdp,
58 unsigned long start,
59 unsigned long end,
60 struct mm_walk *walk)
61 {
62 struct migrate_vma *migrate = walk->private;
63 struct vm_area_struct *vma = walk->vma;
64 struct mm_struct *mm = vma->vm_mm;
65 unsigned long addr = start, unmapped = 0;
66 spinlock_t *ptl;
67 pte_t *ptep;
68
69 again:
70 if (pmd_none(*pmdp))
71 return migrate_vma_collect_hole(start, end, -1, walk);
72
73 if (pmd_trans_huge(*pmdp)) {
74 struct folio *folio;
75
76 ptl = pmd_lock(mm, pmdp);
77 if (unlikely(!pmd_trans_huge(*pmdp))) {
78 spin_unlock(ptl);
79 goto again;
80 }
81
82 folio = pmd_folio(*pmdp);
83 if (is_huge_zero_folio(folio)) {
84 spin_unlock(ptl);
85 split_huge_pmd(vma, pmdp, addr);
86 } else {
87 int ret;
88
89 folio_get(folio);
90 spin_unlock(ptl);
91 if (unlikely(!folio_trylock(folio)))
92 return migrate_vma_collect_skip(start, end,
93 walk);
94 ret = split_folio(folio);
95 folio_unlock(folio);
96 folio_put(folio);
97 if (ret)
98 return migrate_vma_collect_skip(start, end,
99 walk);
100 }
101 }
102
103 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
104 if (!ptep)
105 goto again;
106 arch_enter_lazy_mmu_mode();
107
108 for (; addr < end; addr += PAGE_SIZE, ptep++) {
109 unsigned long mpfn = 0, pfn;
110 struct folio *folio;
111 struct page *page;
112 swp_entry_t entry;
113 pte_t pte;
114
115 pte = ptep_get(ptep);
116
117 if (pte_none(pte)) {
118 if (vma_is_anonymous(vma)) {
119 mpfn = MIGRATE_PFN_MIGRATE;
120 migrate->cpages++;
121 }
122 goto next;
123 }
124
125 if (!pte_present(pte)) {
126 /*
127 * Only care about unaddressable device page special
128 * page table entry. Other special swap entries are not
129 * migratable, and we ignore regular swapped page.
130 */
131 entry = pte_to_swp_entry(pte);
132 if (!is_device_private_entry(entry))
133 goto next;
134
135 page = pfn_swap_entry_to_page(entry);
136 if (!(migrate->flags &
137 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
138 page->pgmap->owner != migrate->pgmap_owner)
139 goto next;
140
141 mpfn = migrate_pfn(page_to_pfn(page)) |
142 MIGRATE_PFN_MIGRATE;
143 if (is_writable_device_private_entry(entry))
144 mpfn |= MIGRATE_PFN_WRITE;
145 } else {
146 pfn = pte_pfn(pte);
147 if (is_zero_pfn(pfn) &&
148 (migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
149 mpfn = MIGRATE_PFN_MIGRATE;
150 migrate->cpages++;
151 goto next;
152 }
153 page = vm_normal_page(migrate->vma, addr, pte);
154 if (page && !is_zone_device_page(page) &&
155 !(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
156 goto next;
157 else if (page && is_device_coherent_page(page) &&
158 (!(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_COHERENT) ||
159 page->pgmap->owner != migrate->pgmap_owner))
160 goto next;
161 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
162 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
163 }
164
165 /* FIXME support THP */
166 if (!page || !page->mapping || PageTransCompound(page)) {
167 mpfn = 0;
168 goto next;
169 }
170
171 /*
172 * By getting a reference on the folio we pin it and that blocks
173 * any kind of migration. Side effect is that it "freezes" the
174 * pte.
175 *
176 * We drop this reference after isolating the folio from the lru
177 * for non device folio (device folio are not on the lru and thus
178 * can't be dropped from it).
179 */
180 folio = page_folio(page);
181 folio_get(folio);
182
183 /*
184 * We rely on folio_trylock() to avoid deadlock between
185 * concurrent migrations where each is waiting on the others
186 * folio lock. If we can't immediately lock the folio we fail this
187 * migration as it is only best effort anyway.
188 *
189 * If we can lock the folio it's safe to set up a migration entry
190 * now. In the common case where the folio is mapped once in a
191 * single process setting up the migration entry now is an
192 * optimisation to avoid walking the rmap later with
193 * try_to_migrate().
194 */
195 if (folio_trylock(folio)) {
196 bool anon_exclusive;
197 pte_t swp_pte;
198
199 flush_cache_page(vma, addr, pte_pfn(pte));
200 anon_exclusive = folio_test_anon(folio) &&
201 PageAnonExclusive(page);
202 if (anon_exclusive) {
203 pte = ptep_clear_flush(vma, addr, ptep);
204
205 if (folio_try_share_anon_rmap_pte(folio, page)) {
206 set_pte_at(mm, addr, ptep, pte);
207 folio_unlock(folio);
208 folio_put(folio);
209 mpfn = 0;
210 goto next;
211 }
212 } else {
213 pte = ptep_get_and_clear(mm, addr, ptep);
214 }
215
216 migrate->cpages++;
217
218 /* Set the dirty flag on the folio now the pte is gone. */
219 if (pte_dirty(pte))
220 folio_mark_dirty(folio);
221
222 /* Setup special migration page table entry */
223 if (mpfn & MIGRATE_PFN_WRITE)
224 entry = make_writable_migration_entry(
225 page_to_pfn(page));
226 else if (anon_exclusive)
227 entry = make_readable_exclusive_migration_entry(
228 page_to_pfn(page));
229 else
230 entry = make_readable_migration_entry(
231 page_to_pfn(page));
232 if (pte_present(pte)) {
233 if (pte_young(pte))
234 entry = make_migration_entry_young(entry);
235 if (pte_dirty(pte))
236 entry = make_migration_entry_dirty(entry);
237 }
238 swp_pte = swp_entry_to_pte(entry);
239 if (pte_present(pte)) {
240 if (pte_soft_dirty(pte))
241 swp_pte = pte_swp_mksoft_dirty(swp_pte);
242 if (pte_uffd_wp(pte))
243 swp_pte = pte_swp_mkuffd_wp(swp_pte);
244 } else {
245 if (pte_swp_soft_dirty(pte))
246 swp_pte = pte_swp_mksoft_dirty(swp_pte);
247 if (pte_swp_uffd_wp(pte))
248 swp_pte = pte_swp_mkuffd_wp(swp_pte);
249 }
250 set_pte_at(mm, addr, ptep, swp_pte);
251
252 /*
253 * This is like regular unmap: we remove the rmap and
254 * drop the folio refcount. The folio won't be freed, as
255 * we took a reference just above.
256 */
257 folio_remove_rmap_pte(folio, page, vma);
258 folio_put(folio);
259
260 if (pte_present(pte))
261 unmapped++;
262 } else {
263 folio_put(folio);
264 mpfn = 0;
265 }
266
267 next:
268 migrate->dst[migrate->npages] = 0;
269 migrate->src[migrate->npages++] = mpfn;
270 }
271
272 /* Only flush the TLB if we actually modified any entries */
273 if (unmapped)
274 flush_tlb_range(walk->vma, start, end);
275
276 arch_leave_lazy_mmu_mode();
277 pte_unmap_unlock(ptep - 1, ptl);
278
279 return 0;
280 }
281
282 static const struct mm_walk_ops migrate_vma_walk_ops = {
283 .pmd_entry = migrate_vma_collect_pmd,
284 .pte_hole = migrate_vma_collect_hole,
285 .walk_lock = PGWALK_RDLOCK,
286 };
287
288 /*
289 * migrate_vma_collect() - collect pages over a range of virtual addresses
290 * @migrate: migrate struct containing all migration information
291 *
292 * This will walk the CPU page table. For each virtual address backed by a
293 * valid page, it updates the src array and takes a reference on the page, in
294 * order to pin the page until we lock it and unmap it.
295 */
296 static void migrate_vma_collect(struct migrate_vma *migrate)
297 {
298 struct mmu_notifier_range range;
299
300 /*
301 * Note that the pgmap_owner is passed to the mmu notifier callback so
302 * that the registered device driver can skip invalidating device
303 * private page mappings that won't be migrated.
304 */
305 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
306 migrate->vma->vm_mm, migrate->start, migrate->end,
307 migrate->pgmap_owner);
308 mmu_notifier_invalidate_range_start(&range);
309
310 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
311 &migrate_vma_walk_ops, migrate);
312
313 mmu_notifier_invalidate_range_end(&range);
314 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
315 }
316
317 /*
318 * migrate_vma_check_page() - check if page is pinned or not
319 * @page: struct page to check
320 *
321 * Pinned pages cannot be migrated. This is the same test as in
322 * folio_migrate_mapping(), except that here we allow migration of a
323 * ZONE_DEVICE page.
324 */
325 static bool migrate_vma_check_page(struct page *page, struct page *fault_page)
326 {
327 struct folio *folio = page_folio(page);
328
329 /*
330 * One extra ref because caller holds an extra reference, either from
331 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
332 * a device page.
333 */
334 int extra = 1 + (page == fault_page);
335
336 /*
337 * FIXME support THP (transparent huge page), it is bit more complex to
338 * check them than regular pages, because they can be mapped with a pmd
339 * or with a pte (split pte mapping).
340 */
341 if (folio_test_large(folio))
342 return false;
343
344 /* Page from ZONE_DEVICE have one extra reference */
345 if (folio_is_zone_device(folio))
346 extra++;
347
348 /* For file back page */
349 if (folio_mapping(folio))
350 extra += 1 + folio_has_private(folio);
351
352 if ((folio_ref_count(folio) - extra) > folio_mapcount(folio))
353 return false;
354
355 return true;
356 }
357
358 /*
359 * Unmaps pages for migration. Returns number of source pfns marked as
360 * migrating.
361 */
362 static unsigned long migrate_device_unmap(unsigned long *src_pfns,
363 unsigned long npages,
364 struct page *fault_page)
365 {
366 unsigned long i, restore = 0;
367 bool allow_drain = true;
368 unsigned long unmapped = 0;
369
370 lru_add_drain();
371
372 for (i = 0; i < npages; i++) {
373 struct page *page = migrate_pfn_to_page(src_pfns[i]);
374 struct folio *folio;
375
376 if (!page) {
377 if (src_pfns[i] & MIGRATE_PFN_MIGRATE)
378 unmapped++;
379 continue;
380 }
381
382 /* ZONE_DEVICE pages are not on LRU */
383 if (!is_zone_device_page(page)) {
384 if (!PageLRU(page) && allow_drain) {
385 /* Drain CPU's lru cache */
386 lru_add_drain_all();
387 allow_drain = false;
388 }
389
390 if (!isolate_lru_page(page)) {
391 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
392 restore++;
393 continue;
394 }
395
396 /* Drop the reference we took in collect */
397 put_page(page);
398 }
399
400 folio = page_folio(page);
401 if (folio_mapped(folio))
402 try_to_migrate(folio, 0);
403
404 if (page_mapped(page) ||
405 !migrate_vma_check_page(page, fault_page)) {
406 if (!is_zone_device_page(page)) {
407 get_page(page);
408 putback_lru_page(page);
409 }
410
411 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
412 restore++;
413 continue;
414 }
415
416 unmapped++;
417 }
418
419 for (i = 0; i < npages && restore; i++) {
420 struct page *page = migrate_pfn_to_page(src_pfns[i]);
421 struct folio *folio;
422
423 if (!page || (src_pfns[i] & MIGRATE_PFN_MIGRATE))
424 continue;
425
426 folio = page_folio(page);
427 remove_migration_ptes(folio, folio, false);
428
429 src_pfns[i] = 0;
430 folio_unlock(folio);
431 folio_put(folio);
432 restore--;
433 }
434
435 return unmapped;
436 }
437
438 /*
439 * migrate_vma_unmap() - replace page mapping with special migration pte entry
440 * @migrate: migrate struct containing all migration information
441 *
442 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
443 * special migration pte entry and check if it has been pinned. Pinned pages are
444 * restored because we cannot migrate them.
445 *
446 * This is the last step before we call the device driver callback to allocate
447 * destination memory and copy contents of original page over to new page.
448 */
449 static void migrate_vma_unmap(struct migrate_vma *migrate)
450 {
451 migrate->cpages = migrate_device_unmap(migrate->src, migrate->npages,
452 migrate->fault_page);
453 }
454
455 /**
456 * migrate_vma_setup() - prepare to migrate a range of memory
457 * @args: contains the vma, start, and pfns arrays for the migration
458 *
459 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
460 * without an error.
461 *
462 * Prepare to migrate a range of memory virtual address range by collecting all
463 * the pages backing each virtual address in the range, saving them inside the
464 * src array. Then lock those pages and unmap them. Once the pages are locked
465 * and unmapped, check whether each page is pinned or not. Pages that aren't
466 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
467 * corresponding src array entry. Then restores any pages that are pinned, by
468 * remapping and unlocking those pages.
469 *
470 * The caller should then allocate destination memory and copy source memory to
471 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
472 * flag set). Once these are allocated and copied, the caller must update each
473 * corresponding entry in the dst array with the pfn value of the destination
474 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
475 * lock_page().
476 *
477 * Note that the caller does not have to migrate all the pages that are marked
478 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
479 * device memory to system memory. If the caller cannot migrate a device page
480 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
481 * consequences for the userspace process, so it must be avoided if at all
482 * possible.
483 *
484 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
485 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
486 * allowing the caller to allocate device memory for those unbacked virtual
487 * addresses. For this the caller simply has to allocate device memory and
488 * properly set the destination entry like for regular migration. Note that
489 * this can still fail, and thus inside the device driver you must check if the
490 * migration was successful for those entries after calling migrate_vma_pages(),
491 * just like for regular migration.
492 *
493 * After that, the callers must call migrate_vma_pages() to go over each entry
494 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
495 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
496 * then migrate_vma_pages() to migrate struct page information from the source
497 * struct page to the destination struct page. If it fails to migrate the
498 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
499 * src array.
500 *
501 * At this point all successfully migrated pages have an entry in the src
502 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
503 * array entry with MIGRATE_PFN_VALID flag set.
504 *
505 * Once migrate_vma_pages() returns the caller may inspect which pages were
506 * successfully migrated, and which were not. Successfully migrated pages will
507 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
508 *
509 * It is safe to update device page table after migrate_vma_pages() because
510 * both destination and source page are still locked, and the mmap_lock is held
511 * in read mode (hence no one can unmap the range being migrated).
512 *
513 * Once the caller is done cleaning up things and updating its page table (if it
514 * chose to do so, this is not an obligation) it finally calls
515 * migrate_vma_finalize() to update the CPU page table to point to new pages
516 * for successfully migrated pages or otherwise restore the CPU page table to
517 * point to the original source pages.
518 */
519 int migrate_vma_setup(struct migrate_vma *args)
520 {
521 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
522
523 args->start &= PAGE_MASK;
524 args->end &= PAGE_MASK;
525 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
526 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
527 return -EINVAL;
528 if (nr_pages <= 0)
529 return -EINVAL;
530 if (args->start < args->vma->vm_start ||
531 args->start >= args->vma->vm_end)
532 return -EINVAL;
533 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
534 return -EINVAL;
535 if (!args->src || !args->dst)
536 return -EINVAL;
537 if (args->fault_page && !is_device_private_page(args->fault_page))
538 return -EINVAL;
539
540 memset(args->src, 0, sizeof(*args->src) * nr_pages);
541 args->cpages = 0;
542 args->npages = 0;
543
544 migrate_vma_collect(args);
545
546 if (args->cpages)
547 migrate_vma_unmap(args);
548
549 /*
550 * At this point pages are locked and unmapped, and thus they have
551 * stable content and can safely be copied to destination memory that
552 * is allocated by the drivers.
553 */
554 return 0;
555
556 }
557 EXPORT_SYMBOL(migrate_vma_setup);
558
559 /*
560 * This code closely matches the code in:
561 * __handle_mm_fault()
562 * handle_pte_fault()
563 * do_anonymous_page()
564 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
565 * private or coherent page.
566 */
567 static void migrate_vma_insert_page(struct migrate_vma *migrate,
568 unsigned long addr,
569 struct page *page,
570 unsigned long *src)
571 {
572 struct folio *folio = page_folio(page);
573 struct vm_area_struct *vma = migrate->vma;
574 struct mm_struct *mm = vma->vm_mm;
575 bool flush = false;
576 spinlock_t *ptl;
577 pte_t entry;
578 pgd_t *pgdp;
579 p4d_t *p4dp;
580 pud_t *pudp;
581 pmd_t *pmdp;
582 pte_t *ptep;
583 pte_t orig_pte;
584
585 /* Only allow populating anonymous memory */
586 if (!vma_is_anonymous(vma))
587 goto abort;
588
589 pgdp = pgd_offset(mm, addr);
590 p4dp = p4d_alloc(mm, pgdp, addr);
591 if (!p4dp)
592 goto abort;
593 pudp = pud_alloc(mm, p4dp, addr);
594 if (!pudp)
595 goto abort;
596 pmdp = pmd_alloc(mm, pudp, addr);
597 if (!pmdp)
598 goto abort;
599 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
600 goto abort;
601 if (pte_alloc(mm, pmdp))
602 goto abort;
603 if (unlikely(anon_vma_prepare(vma)))
604 goto abort;
605 if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
606 goto abort;
607
608 /*
609 * The memory barrier inside __folio_mark_uptodate makes sure that
610 * preceding stores to the folio contents become visible before
611 * the set_pte_at() write.
612 */
613 __folio_mark_uptodate(folio);
614
615 if (folio_is_device_private(folio)) {
616 swp_entry_t swp_entry;
617
618 if (vma->vm_flags & VM_WRITE)
619 swp_entry = make_writable_device_private_entry(
620 page_to_pfn(page));
621 else
622 swp_entry = make_readable_device_private_entry(
623 page_to_pfn(page));
624 entry = swp_entry_to_pte(swp_entry);
625 } else {
626 if (folio_is_zone_device(folio) &&
627 !folio_is_device_coherent(folio)) {
628 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
629 goto abort;
630 }
631 entry = mk_pte(page, vma->vm_page_prot);
632 if (vma->vm_flags & VM_WRITE)
633 entry = pte_mkwrite(pte_mkdirty(entry), vma);
634 }
635
636 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
637 if (!ptep)
638 goto abort;
639 orig_pte = ptep_get(ptep);
640
641 if (check_stable_address_space(mm))
642 goto unlock_abort;
643
644 if (pte_present(orig_pte)) {
645 unsigned long pfn = pte_pfn(orig_pte);
646
647 if (!is_zero_pfn(pfn))
648 goto unlock_abort;
649 flush = true;
650 } else if (!pte_none(orig_pte))
651 goto unlock_abort;
652
653 /*
654 * Check for userfaultfd but do not deliver the fault. Instead,
655 * just back off.
656 */
657 if (userfaultfd_missing(vma))
658 goto unlock_abort;
659
660 inc_mm_counter(mm, MM_ANONPAGES);
661 folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
662 if (!folio_is_zone_device(folio))
663 folio_add_lru_vma(folio, vma);
664 folio_get(folio);
665
666 if (flush) {
667 flush_cache_page(vma, addr, pte_pfn(orig_pte));
668 ptep_clear_flush(vma, addr, ptep);
669 }
670 set_pte_at(mm, addr, ptep, entry);
671 update_mmu_cache(vma, addr, ptep);
672
673 pte_unmap_unlock(ptep, ptl);
674 *src = MIGRATE_PFN_MIGRATE;
675 return;
676
677 unlock_abort:
678 pte_unmap_unlock(ptep, ptl);
679 abort:
680 *src &= ~MIGRATE_PFN_MIGRATE;
681 }
682
683 static void __migrate_device_pages(unsigned long *src_pfns,
684 unsigned long *dst_pfns, unsigned long npages,
685 struct migrate_vma *migrate)
686 {
687 struct mmu_notifier_range range;
688 unsigned long i;
689 bool notified = false;
690
691 for (i = 0; i < npages; i++) {
692 struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
693 struct page *page = migrate_pfn_to_page(src_pfns[i]);
694 struct address_space *mapping;
695 struct folio *newfolio, *folio;
696 int r, extra_cnt = 0;
697
698 if (!newpage) {
699 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
700 continue;
701 }
702
703 if (!page) {
704 unsigned long addr;
705
706 if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE))
707 continue;
708
709 /*
710 * The only time there is no vma is when called from
711 * migrate_device_coherent_page(). However this isn't
712 * called if the page could not be unmapped.
713 */
714 VM_BUG_ON(!migrate);
715 addr = migrate->start + i*PAGE_SIZE;
716 if (!notified) {
717 notified = true;
718
719 mmu_notifier_range_init_owner(&range,
720 MMU_NOTIFY_MIGRATE, 0,
721 migrate->vma->vm_mm, addr, migrate->end,
722 migrate->pgmap_owner);
723 mmu_notifier_invalidate_range_start(&range);
724 }
725 migrate_vma_insert_page(migrate, addr, newpage,
726 &src_pfns[i]);
727 continue;
728 }
729
730 newfolio = page_folio(newpage);
731 folio = page_folio(page);
732 mapping = folio_mapping(folio);
733
734 if (folio_is_device_private(newfolio) ||
735 folio_is_device_coherent(newfolio)) {
736 if (mapping) {
737 /*
738 * For now only support anonymous memory migrating to
739 * device private or coherent memory.
740 *
741 * Try to get rid of swap cache if possible.
742 */
743 if (!folio_test_anon(folio) ||
744 !folio_free_swap(folio)) {
745 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
746 continue;
747 }
748 }
749 } else if (folio_is_zone_device(newfolio)) {
750 /*
751 * Other types of ZONE_DEVICE page are not supported.
752 */
753 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
754 continue;
755 }
756
757 BUG_ON(folio_test_writeback(folio));
758
759 if (migrate && migrate->fault_page == page)
760 extra_cnt = 1;
761 r = folio_migrate_mapping(mapping, newfolio, folio, extra_cnt);
762 if (r != MIGRATEPAGE_SUCCESS)
763 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
764 else
765 folio_migrate_flags(newfolio, folio);
766 }
767
768 if (notified)
769 mmu_notifier_invalidate_range_end(&range);
770 }
771
772 /**
773 * migrate_device_pages() - migrate meta-data from src page to dst page
774 * @src_pfns: src_pfns returned from migrate_device_range()
775 * @dst_pfns: array of pfns allocated by the driver to migrate memory to
776 * @npages: number of pages in the range
777 *
778 * Equivalent to migrate_vma_pages(). This is called to migrate struct page
779 * meta-data from source struct page to destination.
780 */
781 void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns,
782 unsigned long npages)
783 {
784 __migrate_device_pages(src_pfns, dst_pfns, npages, NULL);
785 }
786 EXPORT_SYMBOL(migrate_device_pages);
787
788 /**
789 * migrate_vma_pages() - migrate meta-data from src page to dst page
790 * @migrate: migrate struct containing all migration information
791 *
792 * This migrates struct page meta-data from source struct page to destination
793 * struct page. This effectively finishes the migration from source page to the
794 * destination page.
795 */
796 void migrate_vma_pages(struct migrate_vma *migrate)
797 {
798 __migrate_device_pages(migrate->src, migrate->dst, migrate->npages, migrate);
799 }
800 EXPORT_SYMBOL(migrate_vma_pages);
801
802 /*
803 * migrate_device_finalize() - complete page migration
804 * @src_pfns: src_pfns returned from migrate_device_range()
805 * @dst_pfns: array of pfns allocated by the driver to migrate memory to
806 * @npages: number of pages in the range
807 *
808 * Completes migration of the page by removing special migration entries.
809 * Drivers must ensure copying of page data is complete and visible to the CPU
810 * before calling this.
811 */
812 void migrate_device_finalize(unsigned long *src_pfns,
813 unsigned long *dst_pfns, unsigned long npages)
814 {
815 unsigned long i;
816
817 for (i = 0; i < npages; i++) {
818 struct folio *dst, *src;
819 struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
820 struct page *page = migrate_pfn_to_page(src_pfns[i]);
821
822 if (!page) {
823 if (newpage) {
824 unlock_page(newpage);
825 put_page(newpage);
826 }
827 continue;
828 }
829
830 if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
831 if (newpage) {
832 unlock_page(newpage);
833 put_page(newpage);
834 }
835 newpage = page;
836 }
837
838 src = page_folio(page);
839 dst = page_folio(newpage);
840 remove_migration_ptes(src, dst, false);
841 folio_unlock(src);
842
843 if (is_zone_device_page(page))
844 put_page(page);
845 else
846 putback_lru_page(page);
847
848 if (newpage != page) {
849 unlock_page(newpage);
850 if (is_zone_device_page(newpage))
851 put_page(newpage);
852 else
853 putback_lru_page(newpage);
854 }
855 }
856 }
857 EXPORT_SYMBOL(migrate_device_finalize);
858
859 /**
860 * migrate_vma_finalize() - restore CPU page table entry
861 * @migrate: migrate struct containing all migration information
862 *
863 * This replaces the special migration pte entry with either a mapping to the
864 * new page if migration was successful for that page, or to the original page
865 * otherwise.
866 *
867 * This also unlocks the pages and puts them back on the lru, or drops the extra
868 * refcount, for device pages.
869 */
870 void migrate_vma_finalize(struct migrate_vma *migrate)
871 {
872 migrate_device_finalize(migrate->src, migrate->dst, migrate->npages);
873 }
874 EXPORT_SYMBOL(migrate_vma_finalize);
875
876 /**
877 * migrate_device_range() - migrate device private pfns to normal memory.
878 * @src_pfns: array large enough to hold migrating source device private pfns.
879 * @start: starting pfn in the range to migrate.
880 * @npages: number of pages to migrate.
881 *
882 * migrate_vma_setup() is similar in concept to migrate_vma_setup() except that
883 * instead of looking up pages based on virtual address mappings a range of
884 * device pfns that should be migrated to system memory is used instead.
885 *
886 * This is useful when a driver needs to free device memory but doesn't know the
887 * virtual mappings of every page that may be in device memory. For example this
888 * is often the case when a driver is being unloaded or unbound from a device.
889 *
890 * Like migrate_vma_setup() this function will take a reference and lock any
891 * migrating pages that aren't free before unmapping them. Drivers may then
892 * allocate destination pages and start copying data from the device to CPU
893 * memory before calling migrate_device_pages().
894 */
895 int migrate_device_range(unsigned long *src_pfns, unsigned long start,
896 unsigned long npages)
897 {
898 unsigned long i, pfn;
899
900 for (pfn = start, i = 0; i < npages; pfn++, i++) {
901 struct page *page = pfn_to_page(pfn);
902
903 if (!get_page_unless_zero(page)) {
904 src_pfns[i] = 0;
905 continue;
906 }
907
908 if (!trylock_page(page)) {
909 src_pfns[i] = 0;
910 put_page(page);
911 continue;
912 }
913
914 src_pfns[i] = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
915 }
916
917 migrate_device_unmap(src_pfns, npages, NULL);
918
919 return 0;
920 }
921 EXPORT_SYMBOL(migrate_device_range);
922
923 /*
924 * Migrate a device coherent page back to normal memory. The caller should have
925 * a reference on page which will be copied to the new page if migration is
926 * successful or dropped on failure.
927 */
928 int migrate_device_coherent_page(struct page *page)
929 {
930 unsigned long src_pfn, dst_pfn = 0;
931 struct page *dpage;
932
933 WARN_ON_ONCE(PageCompound(page));
934
935 lock_page(page);
936 src_pfn = migrate_pfn(page_to_pfn(page)) | MIGRATE_PFN_MIGRATE;
937
938 /*
939 * We don't have a VMA and don't need to walk the page tables to find
940 * the source page. So call migrate_vma_unmap() directly to unmap the
941 * page as migrate_vma_setup() will fail if args.vma == NULL.
942 */
943 migrate_device_unmap(&src_pfn, 1, NULL);
944 if (!(src_pfn & MIGRATE_PFN_MIGRATE))
945 return -EBUSY;
946
947 dpage = alloc_page(GFP_USER | __GFP_NOWARN);
948 if (dpage) {
949 lock_page(dpage);
950 dst_pfn = migrate_pfn(page_to_pfn(dpage));
951 }
952
953 migrate_device_pages(&src_pfn, &dst_pfn, 1);
954 if (src_pfn & MIGRATE_PFN_MIGRATE)
955 copy_highpage(dpage, page);
956 migrate_device_finalize(&src_pfn, &dst_pfn, 1);
957
958 if (src_pfn & MIGRATE_PFN_MIGRATE)
959 return 0;
960 return -EBUSY;
961 }
962
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.