xref: /linux/mm/migrate_device.c (revision 1504b6f97bad166b484d6f27dc99746fdca5f467)
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 page *page;
75 
76 		ptl = pmd_lock(mm, pmdp);
77 		if (unlikely(!pmd_trans_huge(*pmdp))) {
78 			spin_unlock(ptl);
79 			goto again;
80 		}
81 
82 		page = pmd_page(*pmdp);
83 		if (is_huge_zero_page(page)) {
84 			spin_unlock(ptl);
85 			split_huge_pmd(vma, pmdp, addr);
86 			if (pmd_trans_unstable(pmdp))
87 				return migrate_vma_collect_skip(start, end,
88 								walk);
89 		} else {
90 			int ret;
91 
92 			get_page(page);
93 			spin_unlock(ptl);
94 			if (unlikely(!trylock_page(page)))
95 				return migrate_vma_collect_skip(start, end,
96 								walk);
97 			ret = split_huge_page(page);
98 			unlock_page(page);
99 			put_page(page);
100 			if (ret)
101 				return migrate_vma_collect_skip(start, end,
102 								walk);
103 			if (pmd_none(*pmdp))
104 				return migrate_vma_collect_hole(start, end, -1,
105 								walk);
106 		}
107 	}
108 
109 	if (unlikely(pmd_bad(*pmdp)))
110 		return migrate_vma_collect_skip(start, end, walk);
111 
112 	ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
113 	arch_enter_lazy_mmu_mode();
114 
115 	for (; addr < end; addr += PAGE_SIZE, ptep++) {
116 		unsigned long mpfn = 0, pfn;
117 		struct page *page;
118 		swp_entry_t entry;
119 		pte_t pte;
120 
121 		pte = *ptep;
122 
123 		if (pte_none(pte)) {
124 			if (vma_is_anonymous(vma)) {
125 				mpfn = MIGRATE_PFN_MIGRATE;
126 				migrate->cpages++;
127 			}
128 			goto next;
129 		}
130 
131 		if (!pte_present(pte)) {
132 			/*
133 			 * Only care about unaddressable device page special
134 			 * page table entry. Other special swap entries are not
135 			 * migratable, and we ignore regular swapped page.
136 			 */
137 			entry = pte_to_swp_entry(pte);
138 			if (!is_device_private_entry(entry))
139 				goto next;
140 
141 			page = pfn_swap_entry_to_page(entry);
142 			if (!(migrate->flags &
143 				MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
144 			    page->pgmap->owner != migrate->pgmap_owner)
145 				goto next;
146 
147 			mpfn = migrate_pfn(page_to_pfn(page)) |
148 					MIGRATE_PFN_MIGRATE;
149 			if (is_writable_device_private_entry(entry))
150 				mpfn |= MIGRATE_PFN_WRITE;
151 		} else {
152 			pfn = pte_pfn(pte);
153 			if (is_zero_pfn(pfn) &&
154 			    (migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
155 				mpfn = MIGRATE_PFN_MIGRATE;
156 				migrate->cpages++;
157 				goto next;
158 			}
159 			page = vm_normal_page(migrate->vma, addr, pte);
160 			if (page && !is_zone_device_page(page) &&
161 			    !(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
162 				goto next;
163 			else if (page && is_device_coherent_page(page) &&
164 			    (!(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_COHERENT) ||
165 			     page->pgmap->owner != migrate->pgmap_owner))
166 				goto next;
167 			mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
168 			mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
169 		}
170 
171 		/* FIXME support THP */
172 		if (!page || !page->mapping || PageTransCompound(page)) {
173 			mpfn = 0;
174 			goto next;
175 		}
176 
177 		/*
178 		 * By getting a reference on the page we pin it and that blocks
179 		 * any kind of migration. Side effect is that it "freezes" the
180 		 * pte.
181 		 *
182 		 * We drop this reference after isolating the page from the lru
183 		 * for non device page (device page are not on the lru and thus
184 		 * can't be dropped from it).
185 		 */
186 		get_page(page);
187 
188 		/*
189 		 * We rely on trylock_page() to avoid deadlock between
190 		 * concurrent migrations where each is waiting on the others
191 		 * page lock. If we can't immediately lock the page we fail this
192 		 * migration as it is only best effort anyway.
193 		 *
194 		 * If we can lock the page it's safe to set up a migration entry
195 		 * now. In the common case where the page is mapped once in a
196 		 * single process setting up the migration entry now is an
197 		 * optimisation to avoid walking the rmap later with
198 		 * try_to_migrate().
199 		 */
200 		if (trylock_page(page)) {
201 			bool anon_exclusive;
202 			pte_t swp_pte;
203 
204 			flush_cache_page(vma, addr, pte_pfn(*ptep));
205 			anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
206 			if (anon_exclusive) {
207 				pte = ptep_clear_flush(vma, addr, ptep);
208 
209 				if (page_try_share_anon_rmap(page)) {
210 					set_pte_at(mm, addr, ptep, pte);
211 					unlock_page(page);
212 					put_page(page);
213 					mpfn = 0;
214 					goto next;
215 				}
216 			} else {
217 				pte = ptep_get_and_clear(mm, addr, ptep);
218 			}
219 
220 			migrate->cpages++;
221 
222 			/* Set the dirty flag on the folio now the pte is gone. */
223 			if (pte_dirty(pte))
224 				folio_mark_dirty(page_folio(page));
225 
226 			/* Setup special migration page table entry */
227 			if (mpfn & MIGRATE_PFN_WRITE)
228 				entry = make_writable_migration_entry(
229 							page_to_pfn(page));
230 			else if (anon_exclusive)
231 				entry = make_readable_exclusive_migration_entry(
232 							page_to_pfn(page));
233 			else
234 				entry = make_readable_migration_entry(
235 							page_to_pfn(page));
236 			if (pte_present(pte)) {
237 				if (pte_young(pte))
238 					entry = make_migration_entry_young(entry);
239 				if (pte_dirty(pte))
240 					entry = make_migration_entry_dirty(entry);
241 			}
242 			swp_pte = swp_entry_to_pte(entry);
243 			if (pte_present(pte)) {
244 				if (pte_soft_dirty(pte))
245 					swp_pte = pte_swp_mksoft_dirty(swp_pte);
246 				if (pte_uffd_wp(pte))
247 					swp_pte = pte_swp_mkuffd_wp(swp_pte);
248 			} else {
249 				if (pte_swp_soft_dirty(pte))
250 					swp_pte = pte_swp_mksoft_dirty(swp_pte);
251 				if (pte_swp_uffd_wp(pte))
252 					swp_pte = pte_swp_mkuffd_wp(swp_pte);
253 			}
254 			set_pte_at(mm, addr, ptep, swp_pte);
255 
256 			/*
257 			 * This is like regular unmap: we remove the rmap and
258 			 * drop page refcount. Page won't be freed, as we took
259 			 * a reference just above.
260 			 */
261 			page_remove_rmap(page, vma, false);
262 			put_page(page);
263 
264 			if (pte_present(pte))
265 				unmapped++;
266 		} else {
267 			put_page(page);
268 			mpfn = 0;
269 		}
270 
271 next:
272 		migrate->dst[migrate->npages] = 0;
273 		migrate->src[migrate->npages++] = mpfn;
274 	}
275 
276 	/* Only flush the TLB if we actually modified any entries */
277 	if (unmapped)
278 		flush_tlb_range(walk->vma, start, end);
279 
280 	arch_leave_lazy_mmu_mode();
281 	pte_unmap_unlock(ptep - 1, ptl);
282 
283 	return 0;
284 }
285 
286 static const struct mm_walk_ops migrate_vma_walk_ops = {
287 	.pmd_entry		= migrate_vma_collect_pmd,
288 	.pte_hole		= migrate_vma_collect_hole,
289 };
290 
291 /*
292  * migrate_vma_collect() - collect pages over a range of virtual addresses
293  * @migrate: migrate struct containing all migration information
294  *
295  * This will walk the CPU page table. For each virtual address backed by a
296  * valid page, it updates the src array and takes a reference on the page, in
297  * order to pin the page until we lock it and unmap it.
298  */
299 static void migrate_vma_collect(struct migrate_vma *migrate)
300 {
301 	struct mmu_notifier_range range;
302 
303 	/*
304 	 * Note that the pgmap_owner is passed to the mmu notifier callback so
305 	 * that the registered device driver can skip invalidating device
306 	 * private page mappings that won't be migrated.
307 	 */
308 	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
309 		migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end,
310 		migrate->pgmap_owner);
311 	mmu_notifier_invalidate_range_start(&range);
312 
313 	walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
314 			&migrate_vma_walk_ops, migrate);
315 
316 	mmu_notifier_invalidate_range_end(&range);
317 	migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
318 }
319 
320 /*
321  * migrate_vma_check_page() - check if page is pinned or not
322  * @page: struct page to check
323  *
324  * Pinned pages cannot be migrated. This is the same test as in
325  * folio_migrate_mapping(), except that here we allow migration of a
326  * ZONE_DEVICE page.
327  */
328 static bool migrate_vma_check_page(struct page *page, struct page *fault_page)
329 {
330 	/*
331 	 * One extra ref because caller holds an extra reference, either from
332 	 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
333 	 * a device page.
334 	 */
335 	int extra = 1 + (page == fault_page);
336 
337 	/*
338 	 * FIXME support THP (transparent huge page), it is bit more complex to
339 	 * check them than regular pages, because they can be mapped with a pmd
340 	 * or with a pte (split pte mapping).
341 	 */
342 	if (PageCompound(page))
343 		return false;
344 
345 	/* Page from ZONE_DEVICE have one extra reference */
346 	if (is_zone_device_page(page))
347 		extra++;
348 
349 	/* For file back page */
350 	if (page_mapping(page))
351 		extra += 1 + page_has_private(page);
352 
353 	if ((page_count(page) - extra) > page_mapcount(page))
354 		return false;
355 
356 	return true;
357 }
358 
359 /*
360  * Unmaps pages for migration. Returns number of unmapped pages.
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 			continue;
378 
379 		/* ZONE_DEVICE pages are not on LRU */
380 		if (!is_zone_device_page(page)) {
381 			if (!PageLRU(page) && allow_drain) {
382 				/* Drain CPU's pagevec */
383 				lru_add_drain_all();
384 				allow_drain = false;
385 			}
386 
387 			if (isolate_lru_page(page)) {
388 				src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
389 				restore++;
390 				continue;
391 			}
392 
393 			/* Drop the reference we took in collect */
394 			put_page(page);
395 		}
396 
397 		folio = page_folio(page);
398 		if (folio_mapped(folio))
399 			try_to_migrate(folio, 0);
400 
401 		if (page_mapped(page) ||
402 		    !migrate_vma_check_page(page, fault_page)) {
403 			if (!is_zone_device_page(page)) {
404 				get_page(page);
405 				putback_lru_page(page);
406 			}
407 
408 			src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
409 			restore++;
410 			continue;
411 		}
412 
413 		unmapped++;
414 	}
415 
416 	for (i = 0; i < npages && restore; i++) {
417 		struct page *page = migrate_pfn_to_page(src_pfns[i]);
418 		struct folio *folio;
419 
420 		if (!page || (src_pfns[i] & MIGRATE_PFN_MIGRATE))
421 			continue;
422 
423 		folio = page_folio(page);
424 		remove_migration_ptes(folio, folio, false);
425 
426 		src_pfns[i] = 0;
427 		folio_unlock(folio);
428 		folio_put(folio);
429 		restore--;
430 	}
431 
432 	return unmapped;
433 }
434 
435 /*
436  * migrate_vma_unmap() - replace page mapping with special migration pte entry
437  * @migrate: migrate struct containing all migration information
438  *
439  * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
440  * special migration pte entry and check if it has been pinned. Pinned pages are
441  * restored because we cannot migrate them.
442  *
443  * This is the last step before we call the device driver callback to allocate
444  * destination memory and copy contents of original page over to new page.
445  */
446 static void migrate_vma_unmap(struct migrate_vma *migrate)
447 {
448 	migrate->cpages = migrate_device_unmap(migrate->src, migrate->npages,
449 					migrate->fault_page);
450 }
451 
452 /**
453  * migrate_vma_setup() - prepare to migrate a range of memory
454  * @args: contains the vma, start, and pfns arrays for the migration
455  *
456  * Returns: negative errno on failures, 0 when 0 or more pages were migrated
457  * without an error.
458  *
459  * Prepare to migrate a range of memory virtual address range by collecting all
460  * the pages backing each virtual address in the range, saving them inside the
461  * src array.  Then lock those pages and unmap them. Once the pages are locked
462  * and unmapped, check whether each page is pinned or not.  Pages that aren't
463  * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
464  * corresponding src array entry.  Then restores any pages that are pinned, by
465  * remapping and unlocking those pages.
466  *
467  * The caller should then allocate destination memory and copy source memory to
468  * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
469  * flag set).  Once these are allocated and copied, the caller must update each
470  * corresponding entry in the dst array with the pfn value of the destination
471  * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
472  * lock_page().
473  *
474  * Note that the caller does not have to migrate all the pages that are marked
475  * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
476  * device memory to system memory.  If the caller cannot migrate a device page
477  * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
478  * consequences for the userspace process, so it must be avoided if at all
479  * possible.
480  *
481  * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
482  * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
483  * allowing the caller to allocate device memory for those unbacked virtual
484  * addresses.  For this the caller simply has to allocate device memory and
485  * properly set the destination entry like for regular migration.  Note that
486  * this can still fail, and thus inside the device driver you must check if the
487  * migration was successful for those entries after calling migrate_vma_pages(),
488  * just like for regular migration.
489  *
490  * After that, the callers must call migrate_vma_pages() to go over each entry
491  * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
492  * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
493  * then migrate_vma_pages() to migrate struct page information from the source
494  * struct page to the destination struct page.  If it fails to migrate the
495  * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
496  * src array.
497  *
498  * At this point all successfully migrated pages have an entry in the src
499  * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
500  * array entry with MIGRATE_PFN_VALID flag set.
501  *
502  * Once migrate_vma_pages() returns the caller may inspect which pages were
503  * successfully migrated, and which were not.  Successfully migrated pages will
504  * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
505  *
506  * It is safe to update device page table after migrate_vma_pages() because
507  * both destination and source page are still locked, and the mmap_lock is held
508  * in read mode (hence no one can unmap the range being migrated).
509  *
510  * Once the caller is done cleaning up things and updating its page table (if it
511  * chose to do so, this is not an obligation) it finally calls
512  * migrate_vma_finalize() to update the CPU page table to point to new pages
513  * for successfully migrated pages or otherwise restore the CPU page table to
514  * point to the original source pages.
515  */
516 int migrate_vma_setup(struct migrate_vma *args)
517 {
518 	long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
519 
520 	args->start &= PAGE_MASK;
521 	args->end &= PAGE_MASK;
522 	if (!args->vma || is_vm_hugetlb_page(args->vma) ||
523 	    (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
524 		return -EINVAL;
525 	if (nr_pages <= 0)
526 		return -EINVAL;
527 	if (args->start < args->vma->vm_start ||
528 	    args->start >= args->vma->vm_end)
529 		return -EINVAL;
530 	if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
531 		return -EINVAL;
532 	if (!args->src || !args->dst)
533 		return -EINVAL;
534 	if (args->fault_page && !is_device_private_page(args->fault_page))
535 		return -EINVAL;
536 
537 	memset(args->src, 0, sizeof(*args->src) * nr_pages);
538 	args->cpages = 0;
539 	args->npages = 0;
540 
541 	migrate_vma_collect(args);
542 
543 	if (args->cpages)
544 		migrate_vma_unmap(args);
545 
546 	/*
547 	 * At this point pages are locked and unmapped, and thus they have
548 	 * stable content and can safely be copied to destination memory that
549 	 * is allocated by the drivers.
550 	 */
551 	return 0;
552 
553 }
554 EXPORT_SYMBOL(migrate_vma_setup);
555 
556 /*
557  * This code closely matches the code in:
558  *   __handle_mm_fault()
559  *     handle_pte_fault()
560  *       do_anonymous_page()
561  * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
562  * private or coherent page.
563  */
564 static void migrate_vma_insert_page(struct migrate_vma *migrate,
565 				    unsigned long addr,
566 				    struct page *page,
567 				    unsigned long *src)
568 {
569 	struct vm_area_struct *vma = migrate->vma;
570 	struct mm_struct *mm = vma->vm_mm;
571 	bool flush = false;
572 	spinlock_t *ptl;
573 	pte_t entry;
574 	pgd_t *pgdp;
575 	p4d_t *p4dp;
576 	pud_t *pudp;
577 	pmd_t *pmdp;
578 	pte_t *ptep;
579 
580 	/* Only allow populating anonymous memory */
581 	if (!vma_is_anonymous(vma))
582 		goto abort;
583 
584 	pgdp = pgd_offset(mm, addr);
585 	p4dp = p4d_alloc(mm, pgdp, addr);
586 	if (!p4dp)
587 		goto abort;
588 	pudp = pud_alloc(mm, p4dp, addr);
589 	if (!pudp)
590 		goto abort;
591 	pmdp = pmd_alloc(mm, pudp, addr);
592 	if (!pmdp)
593 		goto abort;
594 
595 	if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
596 		goto abort;
597 
598 	/*
599 	 * Use pte_alloc() instead of pte_alloc_map().  We can't run
600 	 * pte_offset_map() on pmds where a huge pmd might be created
601 	 * from a different thread.
602 	 *
603 	 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
604 	 * parallel threads are excluded by other means.
605 	 *
606 	 * Here we only have mmap_read_lock(mm).
607 	 */
608 	if (pte_alloc(mm, pmdp))
609 		goto abort;
610 
611 	/* See the comment in pte_alloc_one_map() */
612 	if (unlikely(pmd_trans_unstable(pmdp)))
613 		goto abort;
614 
615 	if (unlikely(anon_vma_prepare(vma)))
616 		goto abort;
617 	if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
618 		goto abort;
619 
620 	/*
621 	 * The memory barrier inside __SetPageUptodate makes sure that
622 	 * preceding stores to the page contents become visible before
623 	 * the set_pte_at() write.
624 	 */
625 	__SetPageUptodate(page);
626 
627 	if (is_device_private_page(page)) {
628 		swp_entry_t swp_entry;
629 
630 		if (vma->vm_flags & VM_WRITE)
631 			swp_entry = make_writable_device_private_entry(
632 						page_to_pfn(page));
633 		else
634 			swp_entry = make_readable_device_private_entry(
635 						page_to_pfn(page));
636 		entry = swp_entry_to_pte(swp_entry);
637 	} else {
638 		if (is_zone_device_page(page) &&
639 		    !is_device_coherent_page(page)) {
640 			pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
641 			goto abort;
642 		}
643 		entry = mk_pte(page, vma->vm_page_prot);
644 		if (vma->vm_flags & VM_WRITE)
645 			entry = pte_mkwrite(pte_mkdirty(entry));
646 	}
647 
648 	ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
649 
650 	if (check_stable_address_space(mm))
651 		goto unlock_abort;
652 
653 	if (pte_present(*ptep)) {
654 		unsigned long pfn = pte_pfn(*ptep);
655 
656 		if (!is_zero_pfn(pfn))
657 			goto unlock_abort;
658 		flush = true;
659 	} else if (!pte_none(*ptep))
660 		goto unlock_abort;
661 
662 	/*
663 	 * Check for userfaultfd but do not deliver the fault. Instead,
664 	 * just back off.
665 	 */
666 	if (userfaultfd_missing(vma))
667 		goto unlock_abort;
668 
669 	inc_mm_counter(mm, MM_ANONPAGES);
670 	page_add_new_anon_rmap(page, vma, addr);
671 	if (!is_zone_device_page(page))
672 		lru_cache_add_inactive_or_unevictable(page, vma);
673 	get_page(page);
674 
675 	if (flush) {
676 		flush_cache_page(vma, addr, pte_pfn(*ptep));
677 		ptep_clear_flush_notify(vma, addr, ptep);
678 		set_pte_at_notify(mm, addr, ptep, entry);
679 		update_mmu_cache(vma, addr, ptep);
680 	} else {
681 		/* No need to invalidate - it was non-present before */
682 		set_pte_at(mm, addr, ptep, entry);
683 		update_mmu_cache(vma, addr, ptep);
684 	}
685 
686 	pte_unmap_unlock(ptep, ptl);
687 	*src = MIGRATE_PFN_MIGRATE;
688 	return;
689 
690 unlock_abort:
691 	pte_unmap_unlock(ptep, ptl);
692 abort:
693 	*src &= ~MIGRATE_PFN_MIGRATE;
694 }
695 
696 static void __migrate_device_pages(unsigned long *src_pfns,
697 				unsigned long *dst_pfns, unsigned long npages,
698 				struct migrate_vma *migrate)
699 {
700 	struct mmu_notifier_range range;
701 	unsigned long i;
702 	bool notified = false;
703 
704 	for (i = 0; i < npages; i++) {
705 		struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
706 		struct page *page = migrate_pfn_to_page(src_pfns[i]);
707 		struct address_space *mapping;
708 		int r;
709 
710 		if (!newpage) {
711 			src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
712 			continue;
713 		}
714 
715 		if (!page) {
716 			unsigned long addr;
717 
718 			if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE))
719 				continue;
720 
721 			/*
722 			 * The only time there is no vma is when called from
723 			 * migrate_device_coherent_page(). However this isn't
724 			 * called if the page could not be unmapped.
725 			 */
726 			VM_BUG_ON(!migrate);
727 			addr = migrate->start + i*PAGE_SIZE;
728 			if (!notified) {
729 				notified = true;
730 
731 				mmu_notifier_range_init_owner(&range,
732 					MMU_NOTIFY_MIGRATE, 0, migrate->vma,
733 					migrate->vma->vm_mm, addr, migrate->end,
734 					migrate->pgmap_owner);
735 				mmu_notifier_invalidate_range_start(&range);
736 			}
737 			migrate_vma_insert_page(migrate, addr, newpage,
738 						&src_pfns[i]);
739 			continue;
740 		}
741 
742 		mapping = page_mapping(page);
743 
744 		if (is_device_private_page(newpage) ||
745 		    is_device_coherent_page(newpage)) {
746 			/*
747 			 * For now only support anonymous memory migrating to
748 			 * device private or coherent memory.
749 			 */
750 			if (mapping) {
751 				src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
752 				continue;
753 			}
754 		} else if (is_zone_device_page(newpage)) {
755 			/*
756 			 * Other types of ZONE_DEVICE page are not supported.
757 			 */
758 			src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
759 			continue;
760 		}
761 
762 		if (migrate && migrate->fault_page == page)
763 			r = migrate_folio_extra(mapping, page_folio(newpage),
764 						page_folio(page),
765 						MIGRATE_SYNC_NO_COPY, 1);
766 		else
767 			r = migrate_folio(mapping, page_folio(newpage),
768 					page_folio(page), MIGRATE_SYNC_NO_COPY);
769 		if (r != MIGRATEPAGE_SUCCESS)
770 			src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
771 	}
772 
773 	/*
774 	 * No need to double call mmu_notifier->invalidate_range() callback as
775 	 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
776 	 * did already call it.
777 	 */
778 	if (notified)
779 		mmu_notifier_invalidate_range_only_end(&range);
780 }
781 
782 /**
783  * migrate_device_pages() - migrate meta-data from src page to dst page
784  * @src_pfns: src_pfns returned from migrate_device_range()
785  * @dst_pfns: array of pfns allocated by the driver to migrate memory to
786  * @npages: number of pages in the range
787  *
788  * Equivalent to migrate_vma_pages(). This is called to migrate struct page
789  * meta-data from source struct page to destination.
790  */
791 void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns,
792 			unsigned long npages)
793 {
794 	__migrate_device_pages(src_pfns, dst_pfns, npages, NULL);
795 }
796 EXPORT_SYMBOL(migrate_device_pages);
797 
798 /**
799  * migrate_vma_pages() - migrate meta-data from src page to dst page
800  * @migrate: migrate struct containing all migration information
801  *
802  * This migrates struct page meta-data from source struct page to destination
803  * struct page. This effectively finishes the migration from source page to the
804  * destination page.
805  */
806 void migrate_vma_pages(struct migrate_vma *migrate)
807 {
808 	__migrate_device_pages(migrate->src, migrate->dst, migrate->npages, migrate);
809 }
810 EXPORT_SYMBOL(migrate_vma_pages);
811 
812 /*
813  * migrate_device_finalize() - complete page migration
814  * @src_pfns: src_pfns returned from migrate_device_range()
815  * @dst_pfns: array of pfns allocated by the driver to migrate memory to
816  * @npages: number of pages in the range
817  *
818  * Completes migration of the page by removing special migration entries.
819  * Drivers must ensure copying of page data is complete and visible to the CPU
820  * before calling this.
821  */
822 void migrate_device_finalize(unsigned long *src_pfns,
823 			unsigned long *dst_pfns, unsigned long npages)
824 {
825 	unsigned long i;
826 
827 	for (i = 0; i < npages; i++) {
828 		struct folio *dst, *src;
829 		struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
830 		struct page *page = migrate_pfn_to_page(src_pfns[i]);
831 
832 		if (!page) {
833 			if (newpage) {
834 				unlock_page(newpage);
835 				put_page(newpage);
836 			}
837 			continue;
838 		}
839 
840 		if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
841 			if (newpage) {
842 				unlock_page(newpage);
843 				put_page(newpage);
844 			}
845 			newpage = page;
846 		}
847 
848 		src = page_folio(page);
849 		dst = page_folio(newpage);
850 		remove_migration_ptes(src, dst, false);
851 		folio_unlock(src);
852 
853 		if (is_zone_device_page(page))
854 			put_page(page);
855 		else
856 			putback_lru_page(page);
857 
858 		if (newpage != page) {
859 			unlock_page(newpage);
860 			if (is_zone_device_page(newpage))
861 				put_page(newpage);
862 			else
863 				putback_lru_page(newpage);
864 		}
865 	}
866 }
867 EXPORT_SYMBOL(migrate_device_finalize);
868 
869 /**
870  * migrate_vma_finalize() - restore CPU page table entry
871  * @migrate: migrate struct containing all migration information
872  *
873  * This replaces the special migration pte entry with either a mapping to the
874  * new page if migration was successful for that page, or to the original page
875  * otherwise.
876  *
877  * This also unlocks the pages and puts them back on the lru, or drops the extra
878  * refcount, for device pages.
879  */
880 void migrate_vma_finalize(struct migrate_vma *migrate)
881 {
882 	migrate_device_finalize(migrate->src, migrate->dst, migrate->npages);
883 }
884 EXPORT_SYMBOL(migrate_vma_finalize);
885 
886 /**
887  * migrate_device_range() - migrate device private pfns to normal memory.
888  * @src_pfns: array large enough to hold migrating source device private pfns.
889  * @start: starting pfn in the range to migrate.
890  * @npages: number of pages to migrate.
891  *
892  * migrate_vma_setup() is similar in concept to migrate_vma_setup() except that
893  * instead of looking up pages based on virtual address mappings a range of
894  * device pfns that should be migrated to system memory is used instead.
895  *
896  * This is useful when a driver needs to free device memory but doesn't know the
897  * virtual mappings of every page that may be in device memory. For example this
898  * is often the case when a driver is being unloaded or unbound from a device.
899  *
900  * Like migrate_vma_setup() this function will take a reference and lock any
901  * migrating pages that aren't free before unmapping them. Drivers may then
902  * allocate destination pages and start copying data from the device to CPU
903  * memory before calling migrate_device_pages().
904  */
905 int migrate_device_range(unsigned long *src_pfns, unsigned long start,
906 			unsigned long npages)
907 {
908 	unsigned long i, pfn;
909 
910 	for (pfn = start, i = 0; i < npages; pfn++, i++) {
911 		struct page *page = pfn_to_page(pfn);
912 
913 		if (!get_page_unless_zero(page)) {
914 			src_pfns[i] = 0;
915 			continue;
916 		}
917 
918 		if (!trylock_page(page)) {
919 			src_pfns[i] = 0;
920 			put_page(page);
921 			continue;
922 		}
923 
924 		src_pfns[i] = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
925 	}
926 
927 	migrate_device_unmap(src_pfns, npages, NULL);
928 
929 	return 0;
930 }
931 EXPORT_SYMBOL(migrate_device_range);
932 
933 /*
934  * Migrate a device coherent page back to normal memory. The caller should have
935  * a reference on page which will be copied to the new page if migration is
936  * successful or dropped on failure.
937  */
938 int migrate_device_coherent_page(struct page *page)
939 {
940 	unsigned long src_pfn, dst_pfn = 0;
941 	struct page *dpage;
942 
943 	WARN_ON_ONCE(PageCompound(page));
944 
945 	lock_page(page);
946 	src_pfn = migrate_pfn(page_to_pfn(page)) | MIGRATE_PFN_MIGRATE;
947 
948 	/*
949 	 * We don't have a VMA and don't need to walk the page tables to find
950 	 * the source page. So call migrate_vma_unmap() directly to unmap the
951 	 * page as migrate_vma_setup() will fail if args.vma == NULL.
952 	 */
953 	migrate_device_unmap(&src_pfn, 1, NULL);
954 	if (!(src_pfn & MIGRATE_PFN_MIGRATE))
955 		return -EBUSY;
956 
957 	dpage = alloc_page(GFP_USER | __GFP_NOWARN);
958 	if (dpage) {
959 		lock_page(dpage);
960 		dst_pfn = migrate_pfn(page_to_pfn(dpage));
961 	}
962 
963 	migrate_device_pages(&src_pfn, &dst_pfn, 1);
964 	if (src_pfn & MIGRATE_PFN_MIGRATE)
965 		copy_highpage(dpage, page);
966 	migrate_device_finalize(&src_pfn, &dst_pfn, 1);
967 
968 	if (src_pfn & MIGRATE_PFN_MIGRATE)
969 		return 0;
970 	return -EBUSY;
971 }
972