2 // SPDX-License-Identifier: GPL-2.0-only
6 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
10 * demand-loading started 01.12.91 - seems it is high on the list of
11 * things wanted, and it should be easy to implement. - Linus
15 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
16 * pages started 02.12.91, seems to work. - Linus.
18 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
19 * would have taken more than the 6M I have free, but it worked well as
22 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
26 * Real VM (paging to/from disk) started 18.12.91. Much more work and
27 * thought has to go into this. Oh, well..
28 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
29 * Found it. Everything seems to work now.
30 * 20.12.91 - Ok, making the swap-device changeable like the root.
34 * 05.04.94 - Multi-page memory management added for v1.1.
37 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
40 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
43 #include <linux/kernel_stat.h>
45 #include <linux/mm_inline.h>
46 #include <linux/sched/mm.h>
47 #include <linux/sched/coredump.h>
48 #include <linux/sched/numa_balancing.h>
49 #include <linux/sched/task.h>
50 #include <linux/hugetlb.h>
51 #include <linux/mman.h>
52 #include <linux/swap.h>
53 #include <linux/highmem.h>
54 #include <linux/pagemap.h>
55 #include <linux/memremap.h>
56 #include <linux/kmsan.h>
57 #include <linux/ksm.h>
58 #include <linux/rmap.h>
59 #include <linux/export.h>
60 #include <linux/delayacct.h>
61 #include <linux/init.h>
62 #include <linux/pfn_t.h>
63 #include <linux/writeback.h>
64 #include <linux/memcontrol.h>
65 #include <linux/mmu_notifier.h>
66 #include <linux/swapops.h>
67 #include <linux/elf.h>
68 #include <linux/gfp.h>
69 #include <linux/migrate.h>
70 #include <linux/string.h>
71 #include <linux/memory-tiers.h>
72 #include <linux/debugfs.h>
73 #include <linux/userfaultfd_k.h>
74 #include <linux/dax.h>
75 #include <linux/oom.h>
76 #include <linux/numa.h>
77 #include <linux/perf_event.h>
78 #include <linux/ptrace.h>
79 #include <linux/vmalloc.h>
80 #include <linux/sched/sysctl.h>
82 #include <trace/events/kmem.h>
85 #include <asm/mmu_context.h>
86 #include <asm/pgalloc.h>
87 #include <linux/uaccess.h>
89 #include <asm/tlbflush.h>
91 #include "pgalloc-track.h"
95 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
96 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
100 unsigned long max_mapnr;
101 EXPORT_SYMBOL(max_mapnr);
103 struct page *mem_map;
104 EXPORT_SYMBOL(mem_map);
107 static vm_fault_t do_fault(struct vm_fault *vmf);
108 static vm_fault_t do_anonymous_page(struct vm_fault *vmf);
109 static bool vmf_pte_changed(struct vm_fault *vmf);
112 * Return true if the original pte was a uffd-wp pte marker (so the pte was
115 static __always_inline bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
117 if (!userfaultfd_wp(vmf->vma))
119 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
122 return pte_marker_uffd_wp(vmf->orig_pte);
126 * A number of key systems in x86 including ioremap() rely on the assumption
127 * that high_memory defines the upper bound on direct map memory, then end
131 EXPORT_SYMBOL(high_memory);
134 * Randomize the address space (stacks, mmaps, brk, etc.).
136 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
137 * as ancient (libc5 based) binaries can segfault. )
139 int randomize_va_space __read_mostly =
140 #ifdef CONFIG_COMPAT_BRK
146 #ifndef arch_wants_old_prefaulted_pte
147 static inline bool arch_wants_old_prefaulted_pte(void)
150 * Transitioning a PTE from 'old' to 'young' can be expensive on
151 * some architectures, even if it's performed in hardware. By
152 * default, "false" means prefaulted entries will be 'young'.
158 static int __init disable_randmaps(char *s)
160 randomize_va_space = 0;
163 __setup("norandmaps", disable_randmaps);
165 unsigned long zero_pfn __read_mostly;
166 EXPORT_SYMBOL(zero_pfn);
168 unsigned long highest_memmap_pfn __read_mostly;
171 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
173 static int __init init_zero_pfn(void)
175 zero_pfn = page_to_pfn(ZERO_PAGE(0));
178 early_initcall(init_zero_pfn);
180 void mm_trace_rss_stat(struct mm_struct *mm, int member)
182 trace_rss_stat(mm, member);
186 * Note: this doesn't free the actual pages themselves. That
187 * has been handled earlier when unmapping all the memory regions.
189 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
192 pgtable_t token = pmd_pgtable(*pmd);
194 pte_free_tlb(tlb, token, addr);
195 mm_dec_nr_ptes(tlb->mm);
198 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
199 unsigned long addr, unsigned long end,
200 unsigned long floor, unsigned long ceiling)
207 pmd = pmd_offset(pud, addr);
209 next = pmd_addr_end(addr, end);
210 if (pmd_none_or_clear_bad(pmd))
212 free_pte_range(tlb, pmd, addr);
213 } while (pmd++, addr = next, addr != end);
223 if (end - 1 > ceiling - 1)
226 pmd = pmd_offset(pud, start);
228 pmd_free_tlb(tlb, pmd, start);
229 mm_dec_nr_pmds(tlb->mm);
232 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
233 unsigned long addr, unsigned long end,
234 unsigned long floor, unsigned long ceiling)
241 pud = pud_offset(p4d, addr);
243 next = pud_addr_end(addr, end);
244 if (pud_none_or_clear_bad(pud))
246 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
247 } while (pud++, addr = next, addr != end);
257 if (end - 1 > ceiling - 1)
260 pud = pud_offset(p4d, start);
262 pud_free_tlb(tlb, pud, start);
263 mm_dec_nr_puds(tlb->mm);
266 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
267 unsigned long addr, unsigned long end,
268 unsigned long floor, unsigned long ceiling)
275 p4d = p4d_offset(pgd, addr);
277 next = p4d_addr_end(addr, end);
278 if (p4d_none_or_clear_bad(p4d))
280 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
281 } while (p4d++, addr = next, addr != end);
287 ceiling &= PGDIR_MASK;
291 if (end - 1 > ceiling - 1)
294 p4d = p4d_offset(pgd, start);
296 p4d_free_tlb(tlb, p4d, start);
300 * This function frees user-level page tables of a process.
302 void free_pgd_range(struct mmu_gather *tlb,
303 unsigned long addr, unsigned long end,
304 unsigned long floor, unsigned long ceiling)
310 * The next few lines have given us lots of grief...
312 * Why are we testing PMD* at this top level? Because often
313 * there will be no work to do at all, and we'd prefer not to
314 * go all the way down to the bottom just to discover that.
316 * Why all these "- 1"s? Because 0 represents both the bottom
317 * of the address space and the top of it (using -1 for the
318 * top wouldn't help much: the masks would do the wrong thing).
319 * The rule is that addr 0 and floor 0 refer to the bottom of
320 * the address space, but end 0 and ceiling 0 refer to the top
321 * Comparisons need to use "end - 1" and "ceiling - 1" (though
322 * that end 0 case should be mythical).
324 * Wherever addr is brought up or ceiling brought down, we must
325 * be careful to reject "the opposite 0" before it confuses the
326 * subsequent tests. But what about where end is brought down
327 * by PMD_SIZE below? no, end can't go down to 0 there.
329 * Whereas we round start (addr) and ceiling down, by different
330 * masks at different levels, in order to test whether a table
331 * now has no other vmas using it, so can be freed, we don't
332 * bother to round floor or end up - the tests don't need that.
346 if (end - 1 > ceiling - 1)
351 * We add page table cache pages with PAGE_SIZE,
352 * (see pte_free_tlb()), flush the tlb if we need
354 tlb_change_page_size(tlb, PAGE_SIZE);
355 pgd = pgd_offset(tlb->mm, addr);
357 next = pgd_addr_end(addr, end);
358 if (pgd_none_or_clear_bad(pgd))
360 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
361 } while (pgd++, addr = next, addr != end);
364 void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
365 struct vm_area_struct *vma, unsigned long floor,
366 unsigned long ceiling, bool mm_wr_locked)
369 unsigned long addr = vma->vm_start;
370 struct vm_area_struct *next;
373 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
374 * be 0. This will underflow and is okay.
376 next = mas_find(mas, ceiling - 1);
377 if (unlikely(xa_is_zero(next)))
381 * Hide vma from rmap and truncate_pagecache before freeing
385 vma_start_write(vma);
386 unlink_anon_vmas(vma);
387 unlink_file_vma(vma);
389 if (is_vm_hugetlb_page(vma)) {
390 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
391 floor, next ? next->vm_start : ceiling);
394 * Optimization: gather nearby vmas into one call down
396 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
397 && !is_vm_hugetlb_page(next)) {
399 next = mas_find(mas, ceiling - 1);
400 if (unlikely(xa_is_zero(next)))
403 vma_start_write(vma);
404 unlink_anon_vmas(vma);
405 unlink_file_vma(vma);
407 free_pgd_range(tlb, addr, vma->vm_end,
408 floor, next ? next->vm_start : ceiling);
414 void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
416 spinlock_t *ptl = pmd_lock(mm, pmd);
418 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
421 * Ensure all pte setup (eg. pte page lock and page clearing) are
422 * visible before the pte is made visible to other CPUs by being
423 * put into page tables.
425 * The other side of the story is the pointer chasing in the page
426 * table walking code (when walking the page table without locking;
427 * ie. most of the time). Fortunately, these data accesses consist
428 * of a chain of data-dependent loads, meaning most CPUs (alpha
429 * being the notable exception) will already guarantee loads are
430 * seen in-order. See the alpha page table accessors for the
431 * smp_rmb() barriers in page table walking code.
433 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
434 pmd_populate(mm, pmd, *pte);
440 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
442 pgtable_t new = pte_alloc_one(mm);
446 pmd_install(mm, pmd, &new);
452 int __pte_alloc_kernel(pmd_t *pmd)
454 pte_t *new = pte_alloc_one_kernel(&init_mm);
458 spin_lock(&init_mm.page_table_lock);
459 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
460 smp_wmb(); /* See comment in pmd_install() */
461 pmd_populate_kernel(&init_mm, pmd, new);
464 spin_unlock(&init_mm.page_table_lock);
466 pte_free_kernel(&init_mm, new);
470 static inline void init_rss_vec(int *rss)
472 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
475 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
479 for (i = 0; i < NR_MM_COUNTERS; i++)
481 add_mm_counter(mm, i, rss[i]);
485 * This function is called to print an error when a bad pte
486 * is found. For example, we might have a PFN-mapped pte in
487 * a region that doesn't allow it.
489 * The calling function must still handle the error.
491 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
492 pte_t pte, struct page *page)
494 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
495 p4d_t *p4d = p4d_offset(pgd, addr);
496 pud_t *pud = pud_offset(p4d, addr);
497 pmd_t *pmd = pmd_offset(pud, addr);
498 struct address_space *mapping;
500 static unsigned long resume;
501 static unsigned long nr_shown;
502 static unsigned long nr_unshown;
505 * Allow a burst of 60 reports, then keep quiet for that minute;
506 * or allow a steady drip of one report per second.
508 if (nr_shown == 60) {
509 if (time_before(jiffies, resume)) {
514 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
521 resume = jiffies + 60 * HZ;
523 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
524 index = linear_page_index(vma, addr);
526 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
528 (long long)pte_val(pte), (long long)pmd_val(*pmd));
530 dump_page(page, "bad pte");
531 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
532 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
533 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
535 vma->vm_ops ? vma->vm_ops->fault : NULL,
536 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
537 mapping ? mapping->a_ops->read_folio : NULL);
539 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
543 * vm_normal_page -- This function gets the "struct page" associated with a pte.
545 * "Special" mappings do not wish to be associated with a "struct page" (either
546 * it doesn't exist, or it exists but they don't want to touch it). In this
547 * case, NULL is returned here. "Normal" mappings do have a struct page.
549 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
550 * pte bit, in which case this function is trivial. Secondly, an architecture
551 * may not have a spare pte bit, which requires a more complicated scheme,
554 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
555 * special mapping (even if there are underlying and valid "struct pages").
556 * COWed pages of a VM_PFNMAP are always normal.
558 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
559 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
560 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
561 * mapping will always honor the rule
563 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
565 * And for normal mappings this is false.
567 * This restricts such mappings to be a linear translation from virtual address
568 * to pfn. To get around this restriction, we allow arbitrary mappings so long
569 * as the vma is not a COW mapping; in that case, we know that all ptes are
570 * special (because none can have been COWed).
573 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
575 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
576 * page" backing, however the difference is that _all_ pages with a struct
577 * page (that is, those where pfn_valid is true) are refcounted and considered
578 * normal pages by the VM. The disadvantage is that pages are refcounted
579 * (which can be slower and simply not an option for some PFNMAP users). The
580 * advantage is that we don't have to follow the strict linearity rule of
581 * PFNMAP mappings in order to support COWable mappings.
584 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
587 unsigned long pfn = pte_pfn(pte);
589 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
590 if (likely(!pte_special(pte)))
592 if (vma->vm_ops && vma->vm_ops->find_special_page)
593 return vma->vm_ops->find_special_page(vma, addr);
594 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
596 if (is_zero_pfn(pfn))
600 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
601 * and will have refcounts incremented on their struct pages
602 * when they are inserted into PTEs, thus they are safe to
603 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
604 * do not have refcounts. Example of legacy ZONE_DEVICE is
605 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
609 print_bad_pte(vma, addr, pte, NULL);
613 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
615 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
616 if (vma->vm_flags & VM_MIXEDMAP) {
622 off = (addr - vma->vm_start) >> PAGE_SHIFT;
623 if (pfn == vma->vm_pgoff + off)
625 if (!is_cow_mapping(vma->vm_flags))
630 if (is_zero_pfn(pfn))
634 if (unlikely(pfn > highest_memmap_pfn)) {
635 print_bad_pte(vma, addr, pte, NULL);
640 * NOTE! We still have PageReserved() pages in the page tables.
641 * eg. VDSO mappings can cause them to exist.
644 return pfn_to_page(pfn);
647 struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr,
650 struct page *page = vm_normal_page(vma, addr, pte);
653 return page_folio(page);
657 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
658 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
661 unsigned long pfn = pmd_pfn(pmd);
664 * There is no pmd_special() but there may be special pmds, e.g.
665 * in a direct-access (dax) mapping, so let's just replicate the
666 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
668 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
669 if (vma->vm_flags & VM_MIXEDMAP) {
675 off = (addr - vma->vm_start) >> PAGE_SHIFT;
676 if (pfn == vma->vm_pgoff + off)
678 if (!is_cow_mapping(vma->vm_flags))
685 if (is_huge_zero_pmd(pmd))
687 if (unlikely(pfn > highest_memmap_pfn))
691 * NOTE! We still have PageReserved() pages in the page tables.
692 * eg. VDSO mappings can cause them to exist.
695 return pfn_to_page(pfn);
698 struct folio *vm_normal_folio_pmd(struct vm_area_struct *vma,
699 unsigned long addr, pmd_t pmd)
701 struct page *page = vm_normal_page_pmd(vma, addr, pmd);
704 return page_folio(page);
709 static void restore_exclusive_pte(struct vm_area_struct *vma,
710 struct page *page, unsigned long address,
713 struct folio *folio = page_folio(page);
718 orig_pte = ptep_get(ptep);
719 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
720 if (pte_swp_soft_dirty(orig_pte))
721 pte = pte_mksoft_dirty(pte);
723 entry = pte_to_swp_entry(orig_pte);
724 if (pte_swp_uffd_wp(orig_pte))
725 pte = pte_mkuffd_wp(pte);
726 else if (is_writable_device_exclusive_entry(entry))
727 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
729 VM_BUG_ON_FOLIO(pte_write(pte) && (!folio_test_anon(folio) &&
730 PageAnonExclusive(page)), folio);
733 * No need to take a page reference as one was already
734 * created when the swap entry was made.
736 if (folio_test_anon(folio))
737 folio_add_anon_rmap_pte(folio, page, vma, address, RMAP_NONE);
740 * Currently device exclusive access only supports anonymous
741 * memory so the entry shouldn't point to a filebacked page.
745 set_pte_at(vma->vm_mm, address, ptep, pte);
748 * No need to invalidate - it was non-present before. However
749 * secondary CPUs may have mappings that need invalidating.
751 update_mmu_cache(vma, address, ptep);
755 * Tries to restore an exclusive pte if the page lock can be acquired without
759 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
762 swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte));
763 struct page *page = pfn_swap_entry_to_page(entry);
765 if (trylock_page(page)) {
766 restore_exclusive_pte(vma, page, addr, src_pte);
775 * copy one vm_area from one task to the other. Assumes the page tables
776 * already present in the new task to be cleared in the whole range
777 * covered by this vma.
781 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
782 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
783 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
785 unsigned long vm_flags = dst_vma->vm_flags;
786 pte_t orig_pte = ptep_get(src_pte);
787 pte_t pte = orig_pte;
790 swp_entry_t entry = pte_to_swp_entry(orig_pte);
792 if (likely(!non_swap_entry(entry))) {
793 if (swap_duplicate(entry) < 0)
796 /* make sure dst_mm is on swapoff's mmlist. */
797 if (unlikely(list_empty(&dst_mm->mmlist))) {
798 spin_lock(&mmlist_lock);
799 if (list_empty(&dst_mm->mmlist))
800 list_add(&dst_mm->mmlist,
802 spin_unlock(&mmlist_lock);
804 /* Mark the swap entry as shared. */
805 if (pte_swp_exclusive(orig_pte)) {
806 pte = pte_swp_clear_exclusive(orig_pte);
807 set_pte_at(src_mm, addr, src_pte, pte);
810 } else if (is_migration_entry(entry)) {
811 folio = pfn_swap_entry_folio(entry);
813 rss[mm_counter(folio)]++;
815 if (!is_readable_migration_entry(entry) &&
816 is_cow_mapping(vm_flags)) {
818 * COW mappings require pages in both parent and child
819 * to be set to read. A previously exclusive entry is
822 entry = make_readable_migration_entry(
824 pte = swp_entry_to_pte(entry);
825 if (pte_swp_soft_dirty(orig_pte))
826 pte = pte_swp_mksoft_dirty(pte);
827 if (pte_swp_uffd_wp(orig_pte))
828 pte = pte_swp_mkuffd_wp(pte);
829 set_pte_at(src_mm, addr, src_pte, pte);
831 } else if (is_device_private_entry(entry)) {
832 page = pfn_swap_entry_to_page(entry);
833 folio = page_folio(page);
836 * Update rss count even for unaddressable pages, as
837 * they should treated just like normal pages in this
840 * We will likely want to have some new rss counters
841 * for unaddressable pages, at some point. But for now
842 * keep things as they are.
845 rss[mm_counter(folio)]++;
846 /* Cannot fail as these pages cannot get pinned. */
847 folio_try_dup_anon_rmap_pte(folio, page, src_vma);
850 * We do not preserve soft-dirty information, because so
851 * far, checkpoint/restore is the only feature that
852 * requires that. And checkpoint/restore does not work
853 * when a device driver is involved (you cannot easily
854 * save and restore device driver state).
856 if (is_writable_device_private_entry(entry) &&
857 is_cow_mapping(vm_flags)) {
858 entry = make_readable_device_private_entry(
860 pte = swp_entry_to_pte(entry);
861 if (pte_swp_uffd_wp(orig_pte))
862 pte = pte_swp_mkuffd_wp(pte);
863 set_pte_at(src_mm, addr, src_pte, pte);
865 } else if (is_device_exclusive_entry(entry)) {
867 * Make device exclusive entries present by restoring the
868 * original entry then copying as for a present pte. Device
869 * exclusive entries currently only support private writable
870 * (ie. COW) mappings.
872 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
873 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
876 } else if (is_pte_marker_entry(entry)) {
877 pte_marker marker = copy_pte_marker(entry, dst_vma);
880 set_pte_at(dst_mm, addr, dst_pte,
881 make_pte_marker(marker));
884 if (!userfaultfd_wp(dst_vma))
885 pte = pte_swp_clear_uffd_wp(pte);
886 set_pte_at(dst_mm, addr, dst_pte, pte);
891 * Copy a present and normal page.
893 * NOTE! The usual case is that this isn't required;
894 * instead, the caller can just increase the page refcount
895 * and re-use the pte the traditional way.
897 * And if we need a pre-allocated page but don't yet have
898 * one, return a negative error to let the preallocation
899 * code know so that it can do so outside the page table
903 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
904 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
905 struct folio **prealloc, struct page *page)
907 struct folio *new_folio;
910 new_folio = *prealloc;
915 * We have a prealloc page, all good! Take it
916 * over and copy the page & arm it.
919 copy_user_highpage(&new_folio->page, page, addr, src_vma);
920 __folio_mark_uptodate(new_folio);
921 folio_add_new_anon_rmap(new_folio, dst_vma, addr);
922 folio_add_lru_vma(new_folio, dst_vma);
925 /* All done, just insert the new page copy in the child */
926 pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot);
927 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
928 if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte)))
929 /* Uffd-wp needs to be delivered to dest pte as well */
930 pte = pte_mkuffd_wp(pte);
931 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
935 static __always_inline void __copy_present_ptes(struct vm_area_struct *dst_vma,
936 struct vm_area_struct *src_vma, pte_t *dst_pte, pte_t *src_pte,
937 pte_t pte, unsigned long addr, int nr)
939 struct mm_struct *src_mm = src_vma->vm_mm;
941 /* If it's a COW mapping, write protect it both processes. */
942 if (is_cow_mapping(src_vma->vm_flags) && pte_write(pte)) {
943 wrprotect_ptes(src_mm, addr, src_pte, nr);
944 pte = pte_wrprotect(pte);
947 /* If it's a shared mapping, mark it clean in the child. */
948 if (src_vma->vm_flags & VM_SHARED)
949 pte = pte_mkclean(pte);
950 pte = pte_mkold(pte);
952 if (!userfaultfd_wp(dst_vma))
953 pte = pte_clear_uffd_wp(pte);
955 set_ptes(dst_vma->vm_mm, addr, dst_pte, pte, nr);
959 * Copy one present PTE, trying to batch-process subsequent PTEs that map
960 * consecutive pages of the same folio by copying them as well.
962 * Returns -EAGAIN if one preallocated page is required to copy the next PTE.
963 * Otherwise, returns the number of copied PTEs (at least 1).
966 copy_present_ptes(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
967 pte_t *dst_pte, pte_t *src_pte, pte_t pte, unsigned long addr,
968 int max_nr, int *rss, struct folio **prealloc)
976 page = vm_normal_page(src_vma, addr, pte);
980 folio = page_folio(page);
983 * If we likely have to copy, just don't bother with batching. Make
984 * sure that the common "small folio" case is as fast as possible
985 * by keeping the batching logic separate.
987 if (unlikely(!*prealloc && folio_test_large(folio) && max_nr != 1)) {
988 if (src_vma->vm_flags & VM_SHARED)
989 flags |= FPB_IGNORE_DIRTY;
990 if (!vma_soft_dirty_enabled(src_vma))
991 flags |= FPB_IGNORE_SOFT_DIRTY;
993 nr = folio_pte_batch(folio, addr, src_pte, pte, max_nr, flags,
994 &any_writable, NULL, NULL);
995 folio_ref_add(folio, nr);
996 if (folio_test_anon(folio)) {
997 if (unlikely(folio_try_dup_anon_rmap_ptes(folio, page,
999 folio_ref_sub(folio, nr);
1002 rss[MM_ANONPAGES] += nr;
1003 VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio);
1005 folio_dup_file_rmap_ptes(folio, page, nr);
1006 rss[mm_counter_file(folio)] += nr;
1009 pte = pte_mkwrite(pte, src_vma);
1010 __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte,
1016 if (folio_test_anon(folio)) {
1018 * If this page may have been pinned by the parent process,
1019 * copy the page immediately for the child so that we'll always
1020 * guarantee the pinned page won't be randomly replaced in the
1023 if (unlikely(folio_try_dup_anon_rmap_pte(folio, page, src_vma))) {
1024 /* Page may be pinned, we have to copy. */
1026 err = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
1027 addr, rss, prealloc, page);
1028 return err ? err : 1;
1030 rss[MM_ANONPAGES]++;
1031 VM_WARN_ON_FOLIO(PageAnonExclusive(page), folio);
1033 folio_dup_file_rmap_pte(folio, page);
1034 rss[mm_counter_file(folio)]++;
1038 __copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte, pte, addr, 1);
1042 static inline struct folio *folio_prealloc(struct mm_struct *src_mm,
1043 struct vm_area_struct *vma, unsigned long addr, bool need_zero)
1045 struct folio *new_folio;
1048 new_folio = vma_alloc_zeroed_movable_folio(vma, addr);
1050 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma,
1056 if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) {
1057 folio_put(new_folio);
1060 folio_throttle_swaprate(new_folio, GFP_KERNEL);
1066 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1067 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1070 struct mm_struct *dst_mm = dst_vma->vm_mm;
1071 struct mm_struct *src_mm = src_vma->vm_mm;
1072 pte_t *orig_src_pte, *orig_dst_pte;
1073 pte_t *src_pte, *dst_pte;
1075 spinlock_t *src_ptl, *dst_ptl;
1076 int progress, max_nr, ret = 0;
1077 int rss[NR_MM_COUNTERS];
1078 swp_entry_t entry = (swp_entry_t){0};
1079 struct folio *prealloc = NULL;
1087 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1088 * error handling here, assume that exclusive mmap_lock on dst and src
1089 * protects anon from unexpected THP transitions; with shmem and file
1090 * protected by mmap_lock-less collapse skipping areas with anon_vma
1091 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1092 * can remove such assumptions later, but this is good enough for now.
1094 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1099 src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl);
1101 pte_unmap_unlock(dst_pte, dst_ptl);
1105 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1106 orig_src_pte = src_pte;
1107 orig_dst_pte = dst_pte;
1108 arch_enter_lazy_mmu_mode();
1114 * We are holding two locks at this point - either of them
1115 * could generate latencies in another task on another CPU.
1117 if (progress >= 32) {
1119 if (need_resched() ||
1120 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1123 ptent = ptep_get(src_pte);
1124 if (pte_none(ptent)) {
1128 if (unlikely(!pte_present(ptent))) {
1129 ret = copy_nonpresent_pte(dst_mm, src_mm,
1134 entry = pte_to_swp_entry(ptep_get(src_pte));
1136 } else if (ret == -EBUSY) {
1142 ptent = ptep_get(src_pte);
1143 VM_WARN_ON_ONCE(!pte_present(ptent));
1146 * Device exclusive entry restored, continue by copying
1147 * the now present pte.
1149 WARN_ON_ONCE(ret != -ENOENT);
1151 /* copy_present_ptes() will clear `*prealloc' if consumed */
1152 max_nr = (end - addr) / PAGE_SIZE;
1153 ret = copy_present_ptes(dst_vma, src_vma, dst_pte, src_pte,
1154 ptent, addr, max_nr, rss, &prealloc);
1156 * If we need a pre-allocated page for this pte, drop the
1157 * locks, allocate, and try again.
1159 if (unlikely(ret == -EAGAIN))
1161 if (unlikely(prealloc)) {
1163 * pre-alloc page cannot be reused by next time so as
1164 * to strictly follow mempolicy (e.g., alloc_page_vma()
1165 * will allocate page according to address). This
1166 * could only happen if one pinned pte changed.
1168 folio_put(prealloc);
1173 } while (dst_pte += nr, src_pte += nr, addr += PAGE_SIZE * nr,
1176 arch_leave_lazy_mmu_mode();
1177 pte_unmap_unlock(orig_src_pte, src_ptl);
1178 add_mm_rss_vec(dst_mm, rss);
1179 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1183 VM_WARN_ON_ONCE(!entry.val);
1184 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1189 } else if (ret == -EBUSY) {
1191 } else if (ret == -EAGAIN) {
1192 prealloc = folio_prealloc(src_mm, src_vma, addr, false);
1195 } else if (ret < 0) {
1199 /* We've captured and resolved the error. Reset, try again. */
1205 if (unlikely(prealloc))
1206 folio_put(prealloc);
1211 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1212 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1215 struct mm_struct *dst_mm = dst_vma->vm_mm;
1216 struct mm_struct *src_mm = src_vma->vm_mm;
1217 pmd_t *src_pmd, *dst_pmd;
1220 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1223 src_pmd = pmd_offset(src_pud, addr);
1225 next = pmd_addr_end(addr, end);
1226 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1227 || pmd_devmap(*src_pmd)) {
1229 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1230 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1231 addr, dst_vma, src_vma);
1238 if (pmd_none_or_clear_bad(src_pmd))
1240 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1243 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1248 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1249 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1252 struct mm_struct *dst_mm = dst_vma->vm_mm;
1253 struct mm_struct *src_mm = src_vma->vm_mm;
1254 pud_t *src_pud, *dst_pud;
1257 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1260 src_pud = pud_offset(src_p4d, addr);
1262 next = pud_addr_end(addr, end);
1263 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1266 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1267 err = copy_huge_pud(dst_mm, src_mm,
1268 dst_pud, src_pud, addr, src_vma);
1275 if (pud_none_or_clear_bad(src_pud))
1277 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1280 } while (dst_pud++, src_pud++, addr = next, addr != end);
1285 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1286 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1289 struct mm_struct *dst_mm = dst_vma->vm_mm;
1290 p4d_t *src_p4d, *dst_p4d;
1293 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1296 src_p4d = p4d_offset(src_pgd, addr);
1298 next = p4d_addr_end(addr, end);
1299 if (p4d_none_or_clear_bad(src_p4d))
1301 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1304 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1309 * Return true if the vma needs to copy the pgtable during this fork(). Return
1310 * false when we can speed up fork() by allowing lazy page faults later until
1311 * when the child accesses the memory range.
1314 vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1317 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1318 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1319 * contains uffd-wp protection information, that's something we can't
1320 * retrieve from page cache, and skip copying will lose those info.
1322 if (userfaultfd_wp(dst_vma))
1325 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1328 if (src_vma->anon_vma)
1332 * Don't copy ptes where a page fault will fill them correctly. Fork
1333 * becomes much lighter when there are big shared or private readonly
1334 * mappings. The tradeoff is that copy_page_range is more efficient
1341 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1343 pgd_t *src_pgd, *dst_pgd;
1345 unsigned long addr = src_vma->vm_start;
1346 unsigned long end = src_vma->vm_end;
1347 struct mm_struct *dst_mm = dst_vma->vm_mm;
1348 struct mm_struct *src_mm = src_vma->vm_mm;
1349 struct mmu_notifier_range range;
1353 if (!vma_needs_copy(dst_vma, src_vma))
1356 if (is_vm_hugetlb_page(src_vma))
1357 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1359 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1361 * We do not free on error cases below as remove_vma
1362 * gets called on error from higher level routine
1364 ret = track_pfn_copy(src_vma);
1370 * We need to invalidate the secondary MMU mappings only when
1371 * there could be a permission downgrade on the ptes of the
1372 * parent mm. And a permission downgrade will only happen if
1373 * is_cow_mapping() returns true.
1375 is_cow = is_cow_mapping(src_vma->vm_flags);
1378 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1379 0, src_mm, addr, end);
1380 mmu_notifier_invalidate_range_start(&range);
1382 * Disabling preemption is not needed for the write side, as
1383 * the read side doesn't spin, but goes to the mmap_lock.
1385 * Use the raw variant of the seqcount_t write API to avoid
1386 * lockdep complaining about preemptibility.
1388 vma_assert_write_locked(src_vma);
1389 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1393 dst_pgd = pgd_offset(dst_mm, addr);
1394 src_pgd = pgd_offset(src_mm, addr);
1396 next = pgd_addr_end(addr, end);
1397 if (pgd_none_or_clear_bad(src_pgd))
1399 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1401 untrack_pfn_clear(dst_vma);
1405 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1408 raw_write_seqcount_end(&src_mm->write_protect_seq);
1409 mmu_notifier_invalidate_range_end(&range);
1414 /* Whether we should zap all COWed (private) pages too */
1415 static inline bool should_zap_cows(struct zap_details *details)
1417 /* By default, zap all pages */
1421 /* Or, we zap COWed pages only if the caller wants to */
1422 return details->even_cows;
1425 /* Decides whether we should zap this folio with the folio pointer specified */
1426 static inline bool should_zap_folio(struct zap_details *details,
1427 struct folio *folio)
1429 /* If we can make a decision without *folio.. */
1430 if (should_zap_cows(details))
1433 /* Otherwise we should only zap non-anon folios */
1434 return !folio_test_anon(folio);
1437 static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1442 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1446 * This function makes sure that we'll replace the none pte with an uffd-wp
1447 * swap special pte marker when necessary. Must be with the pgtable lock held.
1450 zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1451 unsigned long addr, pte_t *pte, int nr,
1452 struct zap_details *details, pte_t pteval)
1454 /* Zap on anonymous always means dropping everything */
1455 if (vma_is_anonymous(vma))
1458 if (zap_drop_file_uffd_wp(details))
1462 /* the PFN in the PTE is irrelevant. */
1463 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1471 static __always_inline void zap_present_folio_ptes(struct mmu_gather *tlb,
1472 struct vm_area_struct *vma, struct folio *folio,
1473 struct page *page, pte_t *pte, pte_t ptent, unsigned int nr,
1474 unsigned long addr, struct zap_details *details, int *rss,
1475 bool *force_flush, bool *force_break)
1477 struct mm_struct *mm = tlb->mm;
1478 bool delay_rmap = false;
1480 if (!folio_test_anon(folio)) {
1481 ptent = get_and_clear_full_ptes(mm, addr, pte, nr, tlb->fullmm);
1482 if (pte_dirty(ptent)) {
1483 folio_mark_dirty(folio);
1484 if (tlb_delay_rmap(tlb)) {
1486 *force_flush = true;
1489 if (pte_young(ptent) && likely(vma_has_recency(vma)))
1490 folio_mark_accessed(folio);
1491 rss[mm_counter(folio)] -= nr;
1493 /* We don't need up-to-date accessed/dirty bits. */
1494 clear_full_ptes(mm, addr, pte, nr, tlb->fullmm);
1495 rss[MM_ANONPAGES] -= nr;
1497 /* Checking a single PTE in a batch is sufficient. */
1498 arch_check_zapped_pte(vma, ptent);
1499 tlb_remove_tlb_entries(tlb, pte, nr, addr);
1500 if (unlikely(userfaultfd_pte_wp(vma, ptent)))
1501 zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details,
1505 folio_remove_rmap_ptes(folio, page, nr, vma);
1507 if (unlikely(folio_mapcount(folio) < 0))
1508 print_bad_pte(vma, addr, ptent, page);
1511 if (want_init_mlocked_on_free() && folio_test_mlocked(folio) &&
1512 !delay_rmap && folio_test_anon(folio)) {
1513 kernel_init_pages(page, folio_nr_pages(folio));
1516 if (unlikely(__tlb_remove_folio_pages(tlb, page, nr, delay_rmap))) {
1517 *force_flush = true;
1518 *force_break = true;
1523 * Zap or skip at least one present PTE, trying to batch-process subsequent
1524 * PTEs that map consecutive pages of the same folio.
1526 * Returns the number of processed (skipped or zapped) PTEs (at least 1).
1528 static inline int zap_present_ptes(struct mmu_gather *tlb,
1529 struct vm_area_struct *vma, pte_t *pte, pte_t ptent,
1530 unsigned int max_nr, unsigned long addr,
1531 struct zap_details *details, int *rss, bool *force_flush,
1534 const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
1535 struct mm_struct *mm = tlb->mm;
1536 struct folio *folio;
1540 page = vm_normal_page(vma, addr, ptent);
1542 /* We don't need up-to-date accessed/dirty bits. */
1543 ptep_get_and_clear_full(mm, addr, pte, tlb->fullmm);
1544 arch_check_zapped_pte(vma, ptent);
1545 tlb_remove_tlb_entry(tlb, pte, addr);
1546 if (userfaultfd_pte_wp(vma, ptent))
1547 zap_install_uffd_wp_if_needed(vma, addr, pte, 1,
1549 ksm_might_unmap_zero_page(mm, ptent);
1553 folio = page_folio(page);
1554 if (unlikely(!should_zap_folio(details, folio)))
1558 * Make sure that the common "small folio" case is as fast as possible
1559 * by keeping the batching logic separate.
1561 if (unlikely(folio_test_large(folio) && max_nr != 1)) {
1562 nr = folio_pte_batch(folio, addr, pte, ptent, max_nr, fpb_flags,
1565 zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, nr,
1566 addr, details, rss, force_flush,
1570 zap_present_folio_ptes(tlb, vma, folio, page, pte, ptent, 1, addr,
1571 details, rss, force_flush, force_break);
1575 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1576 struct vm_area_struct *vma, pmd_t *pmd,
1577 unsigned long addr, unsigned long end,
1578 struct zap_details *details)
1580 bool force_flush = false, force_break = false;
1581 struct mm_struct *mm = tlb->mm;
1582 int rss[NR_MM_COUNTERS];
1589 tlb_change_page_size(tlb, PAGE_SIZE);
1591 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1595 flush_tlb_batched_pending(mm);
1596 arch_enter_lazy_mmu_mode();
1598 pte_t ptent = ptep_get(pte);
1599 struct folio *folio;
1604 if (pte_none(ptent))
1610 if (pte_present(ptent)) {
1611 max_nr = (end - addr) / PAGE_SIZE;
1612 nr = zap_present_ptes(tlb, vma, pte, ptent, max_nr,
1613 addr, details, rss, &force_flush,
1615 if (unlikely(force_break)) {
1616 addr += nr * PAGE_SIZE;
1622 entry = pte_to_swp_entry(ptent);
1623 if (is_device_private_entry(entry) ||
1624 is_device_exclusive_entry(entry)) {
1625 page = pfn_swap_entry_to_page(entry);
1626 folio = page_folio(page);
1627 if (unlikely(!should_zap_folio(details, folio)))
1630 * Both device private/exclusive mappings should only
1631 * work with anonymous page so far, so we don't need to
1632 * consider uffd-wp bit when zap. For more information,
1633 * see zap_install_uffd_wp_if_needed().
1635 WARN_ON_ONCE(!vma_is_anonymous(vma));
1636 rss[mm_counter(folio)]--;
1637 if (is_device_private_entry(entry))
1638 folio_remove_rmap_pte(folio, page, vma);
1640 } else if (!non_swap_entry(entry)) {
1641 max_nr = (end - addr) / PAGE_SIZE;
1642 nr = swap_pte_batch(pte, max_nr, ptent);
1643 /* Genuine swap entries, hence a private anon pages */
1644 if (!should_zap_cows(details))
1646 rss[MM_SWAPENTS] -= nr;
1647 free_swap_and_cache_nr(entry, nr);
1648 } else if (is_migration_entry(entry)) {
1649 folio = pfn_swap_entry_folio(entry);
1650 if (!should_zap_folio(details, folio))
1652 rss[mm_counter(folio)]--;
1653 } else if (pte_marker_entry_uffd_wp(entry)) {
1655 * For anon: always drop the marker; for file: only
1656 * drop the marker if explicitly requested.
1658 if (!vma_is_anonymous(vma) &&
1659 !zap_drop_file_uffd_wp(details))
1661 } else if (is_hwpoison_entry(entry) ||
1662 is_poisoned_swp_entry(entry)) {
1663 if (!should_zap_cows(details))
1666 /* We should have covered all the swap entry types */
1667 pr_alert("unrecognized swap entry 0x%lx\n", entry.val);
1670 clear_not_present_full_ptes(mm, addr, pte, nr, tlb->fullmm);
1671 zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details, ptent);
1672 } while (pte += nr, addr += PAGE_SIZE * nr, addr != end);
1674 add_mm_rss_vec(mm, rss);
1675 arch_leave_lazy_mmu_mode();
1677 /* Do the actual TLB flush before dropping ptl */
1679 tlb_flush_mmu_tlbonly(tlb);
1680 tlb_flush_rmaps(tlb, vma);
1682 pte_unmap_unlock(start_pte, ptl);
1685 * If we forced a TLB flush (either due to running out of
1686 * batch buffers or because we needed to flush dirty TLB
1687 * entries before releasing the ptl), free the batched
1688 * memory too. Come back again if we didn't do everything.
1696 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1697 struct vm_area_struct *vma, pud_t *pud,
1698 unsigned long addr, unsigned long end,
1699 struct zap_details *details)
1704 pmd = pmd_offset(pud, addr);
1706 next = pmd_addr_end(addr, end);
1707 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1708 if (next - addr != HPAGE_PMD_SIZE)
1709 __split_huge_pmd(vma, pmd, addr, false, NULL);
1710 else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1715 } else if (details && details->single_folio &&
1716 folio_test_pmd_mappable(details->single_folio) &&
1717 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1718 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1720 * Take and drop THP pmd lock so that we cannot return
1721 * prematurely, while zap_huge_pmd() has cleared *pmd,
1722 * but not yet decremented compound_mapcount().
1726 if (pmd_none(*pmd)) {
1730 addr = zap_pte_range(tlb, vma, pmd, addr, next, details);
1733 } while (pmd++, cond_resched(), addr != end);
1738 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1739 struct vm_area_struct *vma, p4d_t *p4d,
1740 unsigned long addr, unsigned long end,
1741 struct zap_details *details)
1746 pud = pud_offset(p4d, addr);
1748 next = pud_addr_end(addr, end);
1749 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1750 if (next - addr != HPAGE_PUD_SIZE) {
1751 mmap_assert_locked(tlb->mm);
1752 split_huge_pud(vma, pud, addr);
1753 } else if (zap_huge_pud(tlb, vma, pud, addr))
1757 if (pud_none_or_clear_bad(pud))
1759 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1762 } while (pud++, addr = next, addr != end);
1767 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1768 struct vm_area_struct *vma, pgd_t *pgd,
1769 unsigned long addr, unsigned long end,
1770 struct zap_details *details)
1775 p4d = p4d_offset(pgd, addr);
1777 next = p4d_addr_end(addr, end);
1778 if (p4d_none_or_clear_bad(p4d))
1780 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1781 } while (p4d++, addr = next, addr != end);
1786 void unmap_page_range(struct mmu_gather *tlb,
1787 struct vm_area_struct *vma,
1788 unsigned long addr, unsigned long end,
1789 struct zap_details *details)
1794 BUG_ON(addr >= end);
1795 tlb_start_vma(tlb, vma);
1796 pgd = pgd_offset(vma->vm_mm, addr);
1798 next = pgd_addr_end(addr, end);
1799 if (pgd_none_or_clear_bad(pgd))
1801 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1802 } while (pgd++, addr = next, addr != end);
1803 tlb_end_vma(tlb, vma);
1807 static void unmap_single_vma(struct mmu_gather *tlb,
1808 struct vm_area_struct *vma, unsigned long start_addr,
1809 unsigned long end_addr,
1810 struct zap_details *details, bool mm_wr_locked)
1812 unsigned long start = max(vma->vm_start, start_addr);
1815 if (start >= vma->vm_end)
1817 end = min(vma->vm_end, end_addr);
1818 if (end <= vma->vm_start)
1822 uprobe_munmap(vma, start, end);
1824 if (unlikely(vma->vm_flags & VM_PFNMAP))
1825 untrack_pfn(vma, 0, 0, mm_wr_locked);
1828 if (unlikely(is_vm_hugetlb_page(vma))) {
1830 * It is undesirable to test vma->vm_file as it
1831 * should be non-null for valid hugetlb area.
1832 * However, vm_file will be NULL in the error
1833 * cleanup path of mmap_region. When
1834 * hugetlbfs ->mmap method fails,
1835 * mmap_region() nullifies vma->vm_file
1836 * before calling this function to clean up.
1837 * Since no pte has actually been setup, it is
1838 * safe to do nothing in this case.
1841 zap_flags_t zap_flags = details ?
1842 details->zap_flags : 0;
1843 __unmap_hugepage_range(tlb, vma, start, end,
1847 unmap_page_range(tlb, vma, start, end, details);
1852 * unmap_vmas - unmap a range of memory covered by a list of vma's
1853 * @tlb: address of the caller's struct mmu_gather
1854 * @mas: the maple state
1855 * @vma: the starting vma
1856 * @start_addr: virtual address at which to start unmapping
1857 * @end_addr: virtual address at which to end unmapping
1858 * @tree_end: The maximum index to check
1859 * @mm_wr_locked: lock flag
1861 * Unmap all pages in the vma list.
1863 * Only addresses between `start' and `end' will be unmapped.
1865 * The VMA list must be sorted in ascending virtual address order.
1867 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1868 * range after unmap_vmas() returns. So the only responsibility here is to
1869 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1870 * drops the lock and schedules.
1872 void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas,
1873 struct vm_area_struct *vma, unsigned long start_addr,
1874 unsigned long end_addr, unsigned long tree_end,
1877 struct mmu_notifier_range range;
1878 struct zap_details details = {
1879 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1880 /* Careful - we need to zap private pages too! */
1884 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
1885 start_addr, end_addr);
1886 mmu_notifier_invalidate_range_start(&range);
1888 unsigned long start = start_addr;
1889 unsigned long end = end_addr;
1890 hugetlb_zap_begin(vma, &start, &end);
1891 unmap_single_vma(tlb, vma, start, end, &details,
1893 hugetlb_zap_end(vma, &details);
1894 vma = mas_find(mas, tree_end - 1);
1895 } while (vma && likely(!xa_is_zero(vma)));
1896 mmu_notifier_invalidate_range_end(&range);
1900 * zap_page_range_single - remove user pages in a given range
1901 * @vma: vm_area_struct holding the applicable pages
1902 * @address: starting address of pages to zap
1903 * @size: number of bytes to zap
1904 * @details: details of shared cache invalidation
1906 * The range must fit into one VMA.
1908 void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1909 unsigned long size, struct zap_details *details)
1911 const unsigned long end = address + size;
1912 struct mmu_notifier_range range;
1913 struct mmu_gather tlb;
1916 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1918 hugetlb_zap_begin(vma, &range.start, &range.end);
1919 tlb_gather_mmu(&tlb, vma->vm_mm);
1920 update_hiwater_rss(vma->vm_mm);
1921 mmu_notifier_invalidate_range_start(&range);
1923 * unmap 'address-end' not 'range.start-range.end' as range
1924 * could have been expanded for hugetlb pmd sharing.
1926 unmap_single_vma(&tlb, vma, address, end, details, false);
1927 mmu_notifier_invalidate_range_end(&range);
1928 tlb_finish_mmu(&tlb);
1929 hugetlb_zap_end(vma, details);
1933 * zap_vma_ptes - remove ptes mapping the vma
1934 * @vma: vm_area_struct holding ptes to be zapped
1935 * @address: starting address of pages to zap
1936 * @size: number of bytes to zap
1938 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1940 * The entire address range must be fully contained within the vma.
1943 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1946 if (!range_in_vma(vma, address, address + size) ||
1947 !(vma->vm_flags & VM_PFNMAP))
1950 zap_page_range_single(vma, address, size, NULL);
1952 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1954 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1961 pgd = pgd_offset(mm, addr);
1962 p4d = p4d_alloc(mm, pgd, addr);
1965 pud = pud_alloc(mm, p4d, addr);
1968 pmd = pmd_alloc(mm, pud, addr);
1972 VM_BUG_ON(pmd_trans_huge(*pmd));
1976 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1979 pmd_t *pmd = walk_to_pmd(mm, addr);
1983 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1986 static int validate_page_before_insert(struct page *page)
1988 struct folio *folio = page_folio(page);
1990 if (folio_test_anon(folio) || folio_test_slab(folio) ||
1991 page_has_type(page))
1993 flush_dcache_folio(folio);
1997 static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1998 unsigned long addr, struct page *page, pgprot_t prot)
2000 struct folio *folio = page_folio(page);
2002 if (!pte_none(ptep_get(pte)))
2004 /* Ok, finally just insert the thing.. */
2006 inc_mm_counter(vma->vm_mm, mm_counter_file(folio));
2007 folio_add_file_rmap_pte(folio, page, vma);
2008 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
2013 * This is the old fallback for page remapping.
2015 * For historical reasons, it only allows reserved pages. Only
2016 * old drivers should use this, and they needed to mark their
2017 * pages reserved for the old functions anyway.
2019 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2020 struct page *page, pgprot_t prot)
2026 retval = validate_page_before_insert(page);
2030 pte = get_locked_pte(vma->vm_mm, addr, &ptl);
2033 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
2034 pte_unmap_unlock(pte, ptl);
2039 static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
2040 unsigned long addr, struct page *page, pgprot_t prot)
2044 if (!page_count(page))
2046 err = validate_page_before_insert(page);
2049 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
2052 /* insert_pages() amortizes the cost of spinlock operations
2053 * when inserting pages in a loop.
2055 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
2056 struct page **pages, unsigned long *num, pgprot_t prot)
2059 pte_t *start_pte, *pte;
2060 spinlock_t *pte_lock;
2061 struct mm_struct *const mm = vma->vm_mm;
2062 unsigned long curr_page_idx = 0;
2063 unsigned long remaining_pages_total = *num;
2064 unsigned long pages_to_write_in_pmd;
2068 pmd = walk_to_pmd(mm, addr);
2072 pages_to_write_in_pmd = min_t(unsigned long,
2073 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
2075 /* Allocate the PTE if necessary; takes PMD lock once only. */
2077 if (pte_alloc(mm, pmd))
2080 while (pages_to_write_in_pmd) {
2082 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
2084 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
2089 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
2090 int err = insert_page_in_batch_locked(vma, pte,
2091 addr, pages[curr_page_idx], prot);
2092 if (unlikely(err)) {
2093 pte_unmap_unlock(start_pte, pte_lock);
2095 remaining_pages_total -= pte_idx;
2101 pte_unmap_unlock(start_pte, pte_lock);
2102 pages_to_write_in_pmd -= batch_size;
2103 remaining_pages_total -= batch_size;
2105 if (remaining_pages_total)
2109 *num = remaining_pages_total;
2114 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
2115 * @vma: user vma to map to
2116 * @addr: target start user address of these pages
2117 * @pages: source kernel pages
2118 * @num: in: number of pages to map. out: number of pages that were *not*
2119 * mapped. (0 means all pages were successfully mapped).
2121 * Preferred over vm_insert_page() when inserting multiple pages.
2123 * In case of error, we may have mapped a subset of the provided
2124 * pages. It is the caller's responsibility to account for this case.
2126 * The same restrictions apply as in vm_insert_page().
2128 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
2129 struct page **pages, unsigned long *num)
2131 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
2133 if (addr < vma->vm_start || end_addr >= vma->vm_end)
2135 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2136 BUG_ON(mmap_read_trylock(vma->vm_mm));
2137 BUG_ON(vma->vm_flags & VM_PFNMAP);
2138 vm_flags_set(vma, VM_MIXEDMAP);
2140 /* Defer page refcount checking till we're about to map that page. */
2141 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
2143 EXPORT_SYMBOL(vm_insert_pages);
2146 * vm_insert_page - insert single page into user vma
2147 * @vma: user vma to map to
2148 * @addr: target user address of this page
2149 * @page: source kernel page
2151 * This allows drivers to insert individual pages they've allocated
2154 * The page has to be a nice clean _individual_ kernel allocation.
2155 * If you allocate a compound page, you need to have marked it as
2156 * such (__GFP_COMP), or manually just split the page up yourself
2157 * (see split_page()).
2159 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2160 * took an arbitrary page protection parameter. This doesn't allow
2161 * that. Your vma protection will have to be set up correctly, which
2162 * means that if you want a shared writable mapping, you'd better
2163 * ask for a shared writable mapping!
2165 * The page does not need to be reserved.
2167 * Usually this function is called from f_op->mmap() handler
2168 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2169 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2170 * function from other places, for example from page-fault handler.
2172 * Return: %0 on success, negative error code otherwise.
2174 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2177 if (addr < vma->vm_start || addr >= vma->vm_end)
2179 if (!page_count(page))
2181 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2182 BUG_ON(mmap_read_trylock(vma->vm_mm));
2183 BUG_ON(vma->vm_flags & VM_PFNMAP);
2184 vm_flags_set(vma, VM_MIXEDMAP);
2186 return insert_page(vma, addr, page, vma->vm_page_prot);
2188 EXPORT_SYMBOL(vm_insert_page);
2191 * __vm_map_pages - maps range of kernel pages into user vma
2192 * @vma: user vma to map to
2193 * @pages: pointer to array of source kernel pages
2194 * @num: number of pages in page array
2195 * @offset: user's requested vm_pgoff
2197 * This allows drivers to map range of kernel pages into a user vma.
2199 * Return: 0 on success and error code otherwise.
2201 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2202 unsigned long num, unsigned long offset)
2204 unsigned long count = vma_pages(vma);
2205 unsigned long uaddr = vma->vm_start;
2208 /* Fail if the user requested offset is beyond the end of the object */
2212 /* Fail if the user requested size exceeds available object size */
2213 if (count > num - offset)
2216 for (i = 0; i < count; i++) {
2217 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2227 * vm_map_pages - maps range of kernel pages starts with non zero offset
2228 * @vma: user vma to map to
2229 * @pages: pointer to array of source kernel pages
2230 * @num: number of pages in page array
2232 * Maps an object consisting of @num pages, catering for the user's
2233 * requested vm_pgoff
2235 * If we fail to insert any page into the vma, the function will return
2236 * immediately leaving any previously inserted pages present. Callers
2237 * from the mmap handler may immediately return the error as their caller
2238 * will destroy the vma, removing any successfully inserted pages. Other
2239 * callers should make their own arrangements for calling unmap_region().
2241 * Context: Process context. Called by mmap handlers.
2242 * Return: 0 on success and error code otherwise.
2244 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2247 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2249 EXPORT_SYMBOL(vm_map_pages);
2252 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2253 * @vma: user vma to map to
2254 * @pages: pointer to array of source kernel pages
2255 * @num: number of pages in page array
2257 * Similar to vm_map_pages(), except that it explicitly sets the offset
2258 * to 0. This function is intended for the drivers that did not consider
2261 * Context: Process context. Called by mmap handlers.
2262 * Return: 0 on success and error code otherwise.
2264 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2267 return __vm_map_pages(vma, pages, num, 0);
2269 EXPORT_SYMBOL(vm_map_pages_zero);
2271 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2272 pfn_t pfn, pgprot_t prot, bool mkwrite)
2274 struct mm_struct *mm = vma->vm_mm;
2278 pte = get_locked_pte(mm, addr, &ptl);
2280 return VM_FAULT_OOM;
2281 entry = ptep_get(pte);
2282 if (!pte_none(entry)) {
2285 * For read faults on private mappings the PFN passed
2286 * in may not match the PFN we have mapped if the
2287 * mapped PFN is a writeable COW page. In the mkwrite
2288 * case we are creating a writable PTE for a shared
2289 * mapping and we expect the PFNs to match. If they
2290 * don't match, we are likely racing with block
2291 * allocation and mapping invalidation so just skip the
2294 if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) {
2295 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2298 entry = pte_mkyoung(entry);
2299 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2300 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2301 update_mmu_cache(vma, addr, pte);
2306 /* Ok, finally just insert the thing.. */
2307 if (pfn_t_devmap(pfn))
2308 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2310 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2313 entry = pte_mkyoung(entry);
2314 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2317 set_pte_at(mm, addr, pte, entry);
2318 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2321 pte_unmap_unlock(pte, ptl);
2322 return VM_FAULT_NOPAGE;
2326 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2327 * @vma: user vma to map to
2328 * @addr: target user address of this page
2329 * @pfn: source kernel pfn
2330 * @pgprot: pgprot flags for the inserted page
2332 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2333 * to override pgprot on a per-page basis.
2335 * This only makes sense for IO mappings, and it makes no sense for
2336 * COW mappings. In general, using multiple vmas is preferable;
2337 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2340 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2341 * caching- and encryption bits different than those of @vma->vm_page_prot,
2342 * because the caching- or encryption mode may not be known at mmap() time.
2344 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2345 * to set caching and encryption bits for those vmas (except for COW pages).
2346 * This is ensured by core vm only modifying these page table entries using
2347 * functions that don't touch caching- or encryption bits, using pte_modify()
2348 * if needed. (See for example mprotect()).
2350 * Also when new page-table entries are created, this is only done using the
2351 * fault() callback, and never using the value of vma->vm_page_prot,
2352 * except for page-table entries that point to anonymous pages as the result
2355 * Context: Process context. May allocate using %GFP_KERNEL.
2356 * Return: vm_fault_t value.
2358 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2359 unsigned long pfn, pgprot_t pgprot)
2362 * Technically, architectures with pte_special can avoid all these
2363 * restrictions (same for remap_pfn_range). However we would like
2364 * consistency in testing and feature parity among all, so we should
2365 * try to keep these invariants in place for everybody.
2367 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2368 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2369 (VM_PFNMAP|VM_MIXEDMAP));
2370 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2371 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2373 if (addr < vma->vm_start || addr >= vma->vm_end)
2374 return VM_FAULT_SIGBUS;
2376 if (!pfn_modify_allowed(pfn, pgprot))
2377 return VM_FAULT_SIGBUS;
2379 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2381 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2384 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2387 * vmf_insert_pfn - insert single pfn into user vma
2388 * @vma: user vma to map to
2389 * @addr: target user address of this page
2390 * @pfn: source kernel pfn
2392 * Similar to vm_insert_page, this allows drivers to insert individual pages
2393 * they've allocated into a user vma. Same comments apply.
2395 * This function should only be called from a vm_ops->fault handler, and
2396 * in that case the handler should return the result of this function.
2398 * vma cannot be a COW mapping.
2400 * As this is called only for pages that do not currently exist, we
2401 * do not need to flush old virtual caches or the TLB.
2403 * Context: Process context. May allocate using %GFP_KERNEL.
2404 * Return: vm_fault_t value.
2406 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2409 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2411 EXPORT_SYMBOL(vmf_insert_pfn);
2413 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2415 /* these checks mirror the abort conditions in vm_normal_page */
2416 if (vma->vm_flags & VM_MIXEDMAP)
2418 if (pfn_t_devmap(pfn))
2420 if (pfn_t_special(pfn))
2422 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2427 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2428 unsigned long addr, pfn_t pfn, bool mkwrite)
2430 pgprot_t pgprot = vma->vm_page_prot;
2433 BUG_ON(!vm_mixed_ok(vma, pfn));
2435 if (addr < vma->vm_start || addr >= vma->vm_end)
2436 return VM_FAULT_SIGBUS;
2438 track_pfn_insert(vma, &pgprot, pfn);
2440 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2441 return VM_FAULT_SIGBUS;
2444 * If we don't have pte special, then we have to use the pfn_valid()
2445 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2446 * refcount the page if pfn_valid is true (hence insert_page rather
2447 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2448 * without pte special, it would there be refcounted as a normal page.
2450 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2451 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2455 * At this point we are committed to insert_page()
2456 * regardless of whether the caller specified flags that
2457 * result in pfn_t_has_page() == false.
2459 page = pfn_to_page(pfn_t_to_pfn(pfn));
2460 err = insert_page(vma, addr, page, pgprot);
2462 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2466 return VM_FAULT_OOM;
2467 if (err < 0 && err != -EBUSY)
2468 return VM_FAULT_SIGBUS;
2470 return VM_FAULT_NOPAGE;
2473 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2476 return __vm_insert_mixed(vma, addr, pfn, false);
2478 EXPORT_SYMBOL(vmf_insert_mixed);
2481 * If the insertion of PTE failed because someone else already added a
2482 * different entry in the mean time, we treat that as success as we assume
2483 * the same entry was actually inserted.
2485 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2486 unsigned long addr, pfn_t pfn)
2488 return __vm_insert_mixed(vma, addr, pfn, true);
2490 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2493 * maps a range of physical memory into the requested pages. the old
2494 * mappings are removed. any references to nonexistent pages results
2495 * in null mappings (currently treated as "copy-on-access")
2497 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2498 unsigned long addr, unsigned long end,
2499 unsigned long pfn, pgprot_t prot)
2501 pte_t *pte, *mapped_pte;
2505 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2508 arch_enter_lazy_mmu_mode();
2510 BUG_ON(!pte_none(ptep_get(pte)));
2511 if (!pfn_modify_allowed(pfn, prot)) {
2515 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2517 } while (pte++, addr += PAGE_SIZE, addr != end);
2518 arch_leave_lazy_mmu_mode();
2519 pte_unmap_unlock(mapped_pte, ptl);
2523 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2524 unsigned long addr, unsigned long end,
2525 unsigned long pfn, pgprot_t prot)
2531 pfn -= addr >> PAGE_SHIFT;
2532 pmd = pmd_alloc(mm, pud, addr);
2535 VM_BUG_ON(pmd_trans_huge(*pmd));
2537 next = pmd_addr_end(addr, end);
2538 err = remap_pte_range(mm, pmd, addr, next,
2539 pfn + (addr >> PAGE_SHIFT), prot);
2542 } while (pmd++, addr = next, addr != end);
2546 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2547 unsigned long addr, unsigned long end,
2548 unsigned long pfn, pgprot_t prot)
2554 pfn -= addr >> PAGE_SHIFT;
2555 pud = pud_alloc(mm, p4d, addr);
2559 next = pud_addr_end(addr, end);
2560 err = remap_pmd_range(mm, pud, addr, next,
2561 pfn + (addr >> PAGE_SHIFT), prot);
2564 } while (pud++, addr = next, addr != end);
2568 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2569 unsigned long addr, unsigned long end,
2570 unsigned long pfn, pgprot_t prot)
2576 pfn -= addr >> PAGE_SHIFT;
2577 p4d = p4d_alloc(mm, pgd, addr);
2581 next = p4d_addr_end(addr, end);
2582 err = remap_pud_range(mm, p4d, addr, next,
2583 pfn + (addr >> PAGE_SHIFT), prot);
2586 } while (p4d++, addr = next, addr != end);
2591 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2592 * must have pre-validated the caching bits of the pgprot_t.
2594 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2595 unsigned long pfn, unsigned long size, pgprot_t prot)
2599 unsigned long end = addr + PAGE_ALIGN(size);
2600 struct mm_struct *mm = vma->vm_mm;
2603 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2607 * Physically remapped pages are special. Tell the
2608 * rest of the world about it:
2609 * VM_IO tells people not to look at these pages
2610 * (accesses can have side effects).
2611 * VM_PFNMAP tells the core MM that the base pages are just
2612 * raw PFN mappings, and do not have a "struct page" associated
2615 * Disable vma merging and expanding with mremap().
2617 * Omit vma from core dump, even when VM_IO turned off.
2619 * There's a horrible special case to handle copy-on-write
2620 * behaviour that some programs depend on. We mark the "original"
2621 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2622 * See vm_normal_page() for details.
2624 if (is_cow_mapping(vma->vm_flags)) {
2625 if (addr != vma->vm_start || end != vma->vm_end)
2627 vma->vm_pgoff = pfn;
2630 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
2632 BUG_ON(addr >= end);
2633 pfn -= addr >> PAGE_SHIFT;
2634 pgd = pgd_offset(mm, addr);
2635 flush_cache_range(vma, addr, end);
2637 next = pgd_addr_end(addr, end);
2638 err = remap_p4d_range(mm, pgd, addr, next,
2639 pfn + (addr >> PAGE_SHIFT), prot);
2642 } while (pgd++, addr = next, addr != end);
2648 * remap_pfn_range - remap kernel memory to userspace
2649 * @vma: user vma to map to
2650 * @addr: target page aligned user address to start at
2651 * @pfn: page frame number of kernel physical memory address
2652 * @size: size of mapping area
2653 * @prot: page protection flags for this mapping
2655 * Note: this is only safe if the mm semaphore is held when called.
2657 * Return: %0 on success, negative error code otherwise.
2659 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2660 unsigned long pfn, unsigned long size, pgprot_t prot)
2664 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2668 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2670 untrack_pfn(vma, pfn, PAGE_ALIGN(size), true);
2673 EXPORT_SYMBOL(remap_pfn_range);
2676 * vm_iomap_memory - remap memory to userspace
2677 * @vma: user vma to map to
2678 * @start: start of the physical memory to be mapped
2679 * @len: size of area
2681 * This is a simplified io_remap_pfn_range() for common driver use. The
2682 * driver just needs to give us the physical memory range to be mapped,
2683 * we'll figure out the rest from the vma information.
2685 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2686 * whatever write-combining details or similar.
2688 * Return: %0 on success, negative error code otherwise.
2690 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2692 unsigned long vm_len, pfn, pages;
2694 /* Check that the physical memory area passed in looks valid */
2695 if (start + len < start)
2698 * You *really* shouldn't map things that aren't page-aligned,
2699 * but we've historically allowed it because IO memory might
2700 * just have smaller alignment.
2702 len += start & ~PAGE_MASK;
2703 pfn = start >> PAGE_SHIFT;
2704 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2705 if (pfn + pages < pfn)
2708 /* We start the mapping 'vm_pgoff' pages into the area */
2709 if (vma->vm_pgoff > pages)
2711 pfn += vma->vm_pgoff;
2712 pages -= vma->vm_pgoff;
2714 /* Can we fit all of the mapping? */
2715 vm_len = vma->vm_end - vma->vm_start;
2716 if (vm_len >> PAGE_SHIFT > pages)
2719 /* Ok, let it rip */
2720 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2722 EXPORT_SYMBOL(vm_iomap_memory);
2724 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2725 unsigned long addr, unsigned long end,
2726 pte_fn_t fn, void *data, bool create,
2727 pgtbl_mod_mask *mask)
2729 pte_t *pte, *mapped_pte;
2734 mapped_pte = pte = (mm == &init_mm) ?
2735 pte_alloc_kernel_track(pmd, addr, mask) :
2736 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2740 mapped_pte = pte = (mm == &init_mm) ?
2741 pte_offset_kernel(pmd, addr) :
2742 pte_offset_map_lock(mm, pmd, addr, &ptl);
2747 arch_enter_lazy_mmu_mode();
2751 if (create || !pte_none(ptep_get(pte))) {
2752 err = fn(pte++, addr, data);
2756 } while (addr += PAGE_SIZE, addr != end);
2758 *mask |= PGTBL_PTE_MODIFIED;
2760 arch_leave_lazy_mmu_mode();
2763 pte_unmap_unlock(mapped_pte, ptl);
2767 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2768 unsigned long addr, unsigned long end,
2769 pte_fn_t fn, void *data, bool create,
2770 pgtbl_mod_mask *mask)
2776 BUG_ON(pud_leaf(*pud));
2779 pmd = pmd_alloc_track(mm, pud, addr, mask);
2783 pmd = pmd_offset(pud, addr);
2786 next = pmd_addr_end(addr, end);
2787 if (pmd_none(*pmd) && !create)
2789 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2791 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2796 err = apply_to_pte_range(mm, pmd, addr, next,
2797 fn, data, create, mask);
2800 } while (pmd++, addr = next, addr != end);
2805 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2806 unsigned long addr, unsigned long end,
2807 pte_fn_t fn, void *data, bool create,
2808 pgtbl_mod_mask *mask)
2815 pud = pud_alloc_track(mm, p4d, addr, mask);
2819 pud = pud_offset(p4d, addr);
2822 next = pud_addr_end(addr, end);
2823 if (pud_none(*pud) && !create)
2825 if (WARN_ON_ONCE(pud_leaf(*pud)))
2827 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2832 err = apply_to_pmd_range(mm, pud, addr, next,
2833 fn, data, create, mask);
2836 } while (pud++, addr = next, addr != end);
2841 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2842 unsigned long addr, unsigned long end,
2843 pte_fn_t fn, void *data, bool create,
2844 pgtbl_mod_mask *mask)
2851 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2855 p4d = p4d_offset(pgd, addr);
2858 next = p4d_addr_end(addr, end);
2859 if (p4d_none(*p4d) && !create)
2861 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2863 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2868 err = apply_to_pud_range(mm, p4d, addr, next,
2869 fn, data, create, mask);
2872 } while (p4d++, addr = next, addr != end);
2877 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2878 unsigned long size, pte_fn_t fn,
2879 void *data, bool create)
2882 unsigned long start = addr, next;
2883 unsigned long end = addr + size;
2884 pgtbl_mod_mask mask = 0;
2887 if (WARN_ON(addr >= end))
2890 pgd = pgd_offset(mm, addr);
2892 next = pgd_addr_end(addr, end);
2893 if (pgd_none(*pgd) && !create)
2895 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2897 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2902 err = apply_to_p4d_range(mm, pgd, addr, next,
2903 fn, data, create, &mask);
2906 } while (pgd++, addr = next, addr != end);
2908 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2909 arch_sync_kernel_mappings(start, start + size);
2915 * Scan a region of virtual memory, filling in page tables as necessary
2916 * and calling a provided function on each leaf page table.
2918 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2919 unsigned long size, pte_fn_t fn, void *data)
2921 return __apply_to_page_range(mm, addr, size, fn, data, true);
2923 EXPORT_SYMBOL_GPL(apply_to_page_range);
2926 * Scan a region of virtual memory, calling a provided function on
2927 * each leaf page table where it exists.
2929 * Unlike apply_to_page_range, this does _not_ fill in page tables
2930 * where they are absent.
2932 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2933 unsigned long size, pte_fn_t fn, void *data)
2935 return __apply_to_page_range(mm, addr, size, fn, data, false);
2937 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2940 * handle_pte_fault chooses page fault handler according to an entry which was
2941 * read non-atomically. Before making any commitment, on those architectures
2942 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2943 * parts, do_swap_page must check under lock before unmapping the pte and
2944 * proceeding (but do_wp_page is only called after already making such a check;
2945 * and do_anonymous_page can safely check later on).
2947 static inline int pte_unmap_same(struct vm_fault *vmf)
2950 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2951 if (sizeof(pte_t) > sizeof(unsigned long)) {
2952 spin_lock(vmf->ptl);
2953 same = pte_same(ptep_get(vmf->pte), vmf->orig_pte);
2954 spin_unlock(vmf->ptl);
2957 pte_unmap(vmf->pte);
2964 * 0: copied succeeded
2965 * -EHWPOISON: copy failed due to hwpoison in source page
2966 * -EAGAIN: copied failed (some other reason)
2968 static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2969 struct vm_fault *vmf)
2974 struct vm_area_struct *vma = vmf->vma;
2975 struct mm_struct *mm = vma->vm_mm;
2976 unsigned long addr = vmf->address;
2979 if (copy_mc_user_highpage(dst, src, addr, vma)) {
2980 memory_failure_queue(page_to_pfn(src), 0);
2987 * If the source page was a PFN mapping, we don't have
2988 * a "struct page" for it. We do a best-effort copy by
2989 * just copying from the original user address. If that
2990 * fails, we just zero-fill it. Live with it.
2992 kaddr = kmap_local_page(dst);
2993 pagefault_disable();
2994 uaddr = (void __user *)(addr & PAGE_MASK);
2997 * On architectures with software "accessed" bits, we would
2998 * take a double page fault, so mark it accessed here.
3001 if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) {
3004 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
3005 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3007 * Other thread has already handled the fault
3008 * and update local tlb only
3011 update_mmu_tlb(vma, addr, vmf->pte);
3016 entry = pte_mkyoung(vmf->orig_pte);
3017 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
3018 update_mmu_cache_range(vmf, vma, addr, vmf->pte, 1);
3022 * This really shouldn't fail, because the page is there
3023 * in the page tables. But it might just be unreadable,
3024 * in which case we just give up and fill the result with
3027 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
3031 /* Re-validate under PTL if the page is still mapped */
3032 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
3033 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3034 /* The PTE changed under us, update local tlb */
3036 update_mmu_tlb(vma, addr, vmf->pte);
3042 * The same page can be mapped back since last copy attempt.
3043 * Try to copy again under PTL.
3045 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
3047 * Give a warn in case there can be some obscure
3060 pte_unmap_unlock(vmf->pte, vmf->ptl);
3062 kunmap_local(kaddr);
3063 flush_dcache_page(dst);
3068 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
3070 struct file *vm_file = vma->vm_file;
3073 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
3076 * Special mappings (e.g. VDSO) do not have any file so fake
3077 * a default GFP_KERNEL for them.
3083 * Notify the address space that the page is about to become writable so that
3084 * it can prohibit this or wait for the page to get into an appropriate state.
3086 * We do this without the lock held, so that it can sleep if it needs to.
3088 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
3091 unsigned int old_flags = vmf->flags;
3093 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3095 if (vmf->vma->vm_file &&
3096 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
3097 return VM_FAULT_SIGBUS;
3099 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
3100 /* Restore original flags so that caller is not surprised */
3101 vmf->flags = old_flags;
3102 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
3104 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
3106 if (!folio->mapping) {
3107 folio_unlock(folio);
3108 return 0; /* retry */
3110 ret |= VM_FAULT_LOCKED;
3112 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3117 * Handle dirtying of a page in shared file mapping on a write fault.
3119 * The function expects the page to be locked and unlocks it.
3121 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
3123 struct vm_area_struct *vma = vmf->vma;
3124 struct address_space *mapping;
3125 struct folio *folio = page_folio(vmf->page);
3127 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
3129 dirtied = folio_mark_dirty(folio);
3130 VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
3132 * Take a local copy of the address_space - folio.mapping may be zeroed
3133 * by truncate after folio_unlock(). The address_space itself remains
3134 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
3135 * release semantics to prevent the compiler from undoing this copying.
3137 mapping = folio_raw_mapping(folio);
3138 folio_unlock(folio);
3141 file_update_time(vma->vm_file);
3144 * Throttle page dirtying rate down to writeback speed.
3146 * mapping may be NULL here because some device drivers do not
3147 * set page.mapping but still dirty their pages
3149 * Drop the mmap_lock before waiting on IO, if we can. The file
3150 * is pinning the mapping, as per above.
3152 if ((dirtied || page_mkwrite) && mapping) {
3155 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
3156 balance_dirty_pages_ratelimited(mapping);
3159 return VM_FAULT_COMPLETED;
3167 * Handle write page faults for pages that can be reused in the current vma
3169 * This can happen either due to the mapping being with the VM_SHARED flag,
3170 * or due to us being the last reference standing to the page. In either
3171 * case, all we need to do here is to mark the page as writable and update
3172 * any related book-keeping.
3174 static inline void wp_page_reuse(struct vm_fault *vmf, struct folio *folio)
3175 __releases(vmf->ptl)
3177 struct vm_area_struct *vma = vmf->vma;
3180 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3183 VM_BUG_ON(folio_test_anon(folio) &&
3184 !PageAnonExclusive(vmf->page));
3186 * Clear the folio's cpupid information as the existing
3187 * information potentially belongs to a now completely
3188 * unrelated process.
3190 folio_xchg_last_cpupid(folio, (1 << LAST_CPUPID_SHIFT) - 1);
3193 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3194 entry = pte_mkyoung(vmf->orig_pte);
3195 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3196 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
3197 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3198 pte_unmap_unlock(vmf->pte, vmf->ptl);
3199 count_vm_event(PGREUSE);
3203 * We could add a bitflag somewhere, but for now, we know that all
3204 * vm_ops that have a ->map_pages have been audited and don't need
3205 * the mmap_lock to be held.
3207 static inline vm_fault_t vmf_can_call_fault(const struct vm_fault *vmf)
3209 struct vm_area_struct *vma = vmf->vma;
3211 if (vma->vm_ops->map_pages || !(vmf->flags & FAULT_FLAG_VMA_LOCK))
3214 return VM_FAULT_RETRY;
3218 * vmf_anon_prepare - Prepare to handle an anonymous fault.
3219 * @vmf: The vm_fault descriptor passed from the fault handler.
3221 * When preparing to insert an anonymous page into a VMA from a
3222 * fault handler, call this function rather than anon_vma_prepare().
3223 * If this vma does not already have an associated anon_vma and we are
3224 * only protected by the per-VMA lock, the caller must retry with the
3225 * mmap_lock held. __anon_vma_prepare() will look at adjacent VMAs to
3226 * determine if this VMA can share its anon_vma, and that's not safe to
3227 * do with only the per-VMA lock held for this VMA.
3229 * Return: 0 if fault handling can proceed. Any other value should be
3230 * returned to the caller.
3232 vm_fault_t vmf_anon_prepare(struct vm_fault *vmf)
3234 struct vm_area_struct *vma = vmf->vma;
3237 if (likely(vma->anon_vma))
3239 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3240 if (!mmap_read_trylock(vma->vm_mm)) {
3242 return VM_FAULT_RETRY;
3245 if (__anon_vma_prepare(vma))
3247 if (vmf->flags & FAULT_FLAG_VMA_LOCK)
3248 mmap_read_unlock(vma->vm_mm);
3253 * Handle the case of a page which we actually need to copy to a new page,
3254 * either due to COW or unsharing.
3256 * Called with mmap_lock locked and the old page referenced, but
3257 * without the ptl held.
3259 * High level logic flow:
3261 * - Allocate a page, copy the content of the old page to the new one.
3262 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3263 * - Take the PTL. If the pte changed, bail out and release the allocated page
3264 * - If the pte is still the way we remember it, update the page table and all
3265 * relevant references. This includes dropping the reference the page-table
3266 * held to the old page, as well as updating the rmap.
3267 * - In any case, unlock the PTL and drop the reference we took to the old page.
3269 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3271 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3272 struct vm_area_struct *vma = vmf->vma;
3273 struct mm_struct *mm = vma->vm_mm;
3274 struct folio *old_folio = NULL;
3275 struct folio *new_folio = NULL;
3277 int page_copied = 0;
3278 struct mmu_notifier_range range;
3282 delayacct_wpcopy_start();
3285 old_folio = page_folio(vmf->page);
3286 ret = vmf_anon_prepare(vmf);
3290 pfn_is_zero = is_zero_pfn(pte_pfn(vmf->orig_pte));
3291 new_folio = folio_prealloc(mm, vma, vmf->address, pfn_is_zero);
3298 err = __wp_page_copy_user(&new_folio->page, vmf->page, vmf);
3301 * COW failed, if the fault was solved by other,
3302 * it's fine. If not, userspace would re-fault on
3303 * the same address and we will handle the fault
3304 * from the second attempt.
3305 * The -EHWPOISON case will not be retried.
3307 folio_put(new_folio);
3309 folio_put(old_folio);
3311 delayacct_wpcopy_end();
3312 return err == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3314 kmsan_copy_page_meta(&new_folio->page, vmf->page);
3317 __folio_mark_uptodate(new_folio);
3319 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
3320 vmf->address & PAGE_MASK,
3321 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3322 mmu_notifier_invalidate_range_start(&range);
3325 * Re-check the pte - we dropped the lock
3327 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3328 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3330 if (!folio_test_anon(old_folio)) {
3331 dec_mm_counter(mm, mm_counter_file(old_folio));
3332 inc_mm_counter(mm, MM_ANONPAGES);
3335 ksm_might_unmap_zero_page(mm, vmf->orig_pte);
3336 inc_mm_counter(mm, MM_ANONPAGES);
3338 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3339 entry = mk_pte(&new_folio->page, vma->vm_page_prot);
3340 entry = pte_sw_mkyoung(entry);
3341 if (unlikely(unshare)) {
3342 if (pte_soft_dirty(vmf->orig_pte))
3343 entry = pte_mksoft_dirty(entry);
3344 if (pte_uffd_wp(vmf->orig_pte))
3345 entry = pte_mkuffd_wp(entry);
3347 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3351 * Clear the pte entry and flush it first, before updating the
3352 * pte with the new entry, to keep TLBs on different CPUs in
3353 * sync. This code used to set the new PTE then flush TLBs, but
3354 * that left a window where the new PTE could be loaded into
3355 * some TLBs while the old PTE remains in others.
3357 ptep_clear_flush(vma, vmf->address, vmf->pte);
3358 folio_add_new_anon_rmap(new_folio, vma, vmf->address);
3359 folio_add_lru_vma(new_folio, vma);
3360 BUG_ON(unshare && pte_write(entry));
3361 set_pte_at(mm, vmf->address, vmf->pte, entry);
3362 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3365 * Only after switching the pte to the new page may
3366 * we remove the mapcount here. Otherwise another
3367 * process may come and find the rmap count decremented
3368 * before the pte is switched to the new page, and
3369 * "reuse" the old page writing into it while our pte
3370 * here still points into it and can be read by other
3373 * The critical issue is to order this
3374 * folio_remove_rmap_pte() with the ptp_clear_flush
3375 * above. Those stores are ordered by (if nothing else,)
3376 * the barrier present in the atomic_add_negative
3377 * in folio_remove_rmap_pte();
3379 * Then the TLB flush in ptep_clear_flush ensures that
3380 * no process can access the old page before the
3381 * decremented mapcount is visible. And the old page
3382 * cannot be reused until after the decremented
3383 * mapcount is visible. So transitively, TLBs to
3384 * old page will be flushed before it can be reused.
3386 folio_remove_rmap_pte(old_folio, vmf->page, vma);
3389 /* Free the old page.. */
3390 new_folio = old_folio;
3392 pte_unmap_unlock(vmf->pte, vmf->ptl);
3393 } else if (vmf->pte) {
3394 update_mmu_tlb(vma, vmf->address, vmf->pte);
3395 pte_unmap_unlock(vmf->pte, vmf->ptl);
3398 mmu_notifier_invalidate_range_end(&range);
3401 folio_put(new_folio);
3404 free_swap_cache(old_folio);
3405 folio_put(old_folio);
3408 delayacct_wpcopy_end();
3414 folio_put(old_folio);
3416 delayacct_wpcopy_end();
3421 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3422 * writeable once the page is prepared
3424 * @vmf: structure describing the fault
3425 * @folio: the folio of vmf->page
3427 * This function handles all that is needed to finish a write page fault in a
3428 * shared mapping due to PTE being read-only once the mapped page is prepared.
3429 * It handles locking of PTE and modifying it.
3431 * The function expects the page to be locked or other protection against
3432 * concurrent faults / writeback (such as DAX radix tree locks).
3434 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3435 * we acquired PTE lock.
3437 static vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf, struct folio *folio)
3439 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3440 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3443 return VM_FAULT_NOPAGE;
3445 * We might have raced with another page fault while we released the
3446 * pte_offset_map_lock.
3448 if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) {
3449 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3450 pte_unmap_unlock(vmf->pte, vmf->ptl);
3451 return VM_FAULT_NOPAGE;
3453 wp_page_reuse(vmf, folio);
3458 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3461 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3463 struct vm_area_struct *vma = vmf->vma;
3465 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3468 pte_unmap_unlock(vmf->pte, vmf->ptl);
3469 ret = vmf_can_call_fault(vmf);
3473 vmf->flags |= FAULT_FLAG_MKWRITE;
3474 ret = vma->vm_ops->pfn_mkwrite(vmf);
3475 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3477 return finish_mkwrite_fault(vmf, NULL);
3479 wp_page_reuse(vmf, NULL);
3483 static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3484 __releases(vmf->ptl)
3486 struct vm_area_struct *vma = vmf->vma;
3491 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3494 pte_unmap_unlock(vmf->pte, vmf->ptl);
3495 tmp = vmf_can_call_fault(vmf);
3501 tmp = do_page_mkwrite(vmf, folio);
3502 if (unlikely(!tmp || (tmp &
3503 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3507 tmp = finish_mkwrite_fault(vmf, folio);
3508 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3509 folio_unlock(folio);
3514 wp_page_reuse(vmf, folio);
3517 ret |= fault_dirty_shared_page(vmf);
3523 static bool wp_can_reuse_anon_folio(struct folio *folio,
3524 struct vm_area_struct *vma)
3527 * We could currently only reuse a subpage of a large folio if no
3528 * other subpages of the large folios are still mapped. However,
3529 * let's just consistently not reuse subpages even if we could
3530 * reuse in that scenario, and give back a large folio a bit
3533 if (folio_test_large(folio))
3537 * We have to verify under folio lock: these early checks are
3538 * just an optimization to avoid locking the folio and freeing
3539 * the swapcache if there is little hope that we can reuse.
3541 * KSM doesn't necessarily raise the folio refcount.
3543 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3545 if (!folio_test_lru(folio))
3547 * We cannot easily detect+handle references from
3548 * remote LRU caches or references to LRU folios.
3551 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3553 if (!folio_trylock(folio))
3555 if (folio_test_swapcache(folio))
3556 folio_free_swap(folio);
3557 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3558 folio_unlock(folio);
3562 * Ok, we've got the only folio reference from our mapping
3563 * and the folio is locked, it's dark out, and we're wearing
3564 * sunglasses. Hit it.
3566 folio_move_anon_rmap(folio, vma);
3567 folio_unlock(folio);
3572 * This routine handles present pages, when
3573 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3574 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3575 * (FAULT_FLAG_UNSHARE)
3577 * It is done by copying the page to a new address and decrementing the
3578 * shared-page counter for the old page.
3580 * Note that this routine assumes that the protection checks have been
3581 * done by the caller (the low-level page fault routine in most cases).
3582 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3583 * done any necessary COW.
3585 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3586 * though the page will change only once the write actually happens. This
3587 * avoids a few races, and potentially makes it more efficient.
3589 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3590 * but allow concurrent faults), with pte both mapped and locked.
3591 * We return with mmap_lock still held, but pte unmapped and unlocked.
3593 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3594 __releases(vmf->ptl)
3596 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3597 struct vm_area_struct *vma = vmf->vma;
3598 struct folio *folio = NULL;
3601 if (likely(!unshare)) {
3602 if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) {
3603 if (!userfaultfd_wp_async(vma)) {
3604 pte_unmap_unlock(vmf->pte, vmf->ptl);
3605 return handle_userfault(vmf, VM_UFFD_WP);
3609 * Nothing needed (cache flush, TLB invalidations,
3610 * etc.) because we're only removing the uffd-wp bit,
3611 * which is completely invisible to the user.
3613 pte = pte_clear_uffd_wp(ptep_get(vmf->pte));
3615 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3617 * Update this to be prepared for following up CoW
3620 vmf->orig_pte = pte;
3624 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3625 * is flushed in this case before copying.
3627 if (unlikely(userfaultfd_wp(vmf->vma) &&
3628 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3629 flush_tlb_page(vmf->vma, vmf->address);
3632 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3635 folio = page_folio(vmf->page);
3638 * Shared mapping: we are guaranteed to have VM_WRITE and
3639 * FAULT_FLAG_WRITE set at this point.
3641 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3643 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3646 * We should not cow pages in a shared writeable mapping.
3647 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3650 return wp_pfn_shared(vmf);
3651 return wp_page_shared(vmf, folio);
3655 * Private mapping: create an exclusive anonymous page copy if reuse
3656 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3658 * If we encounter a page that is marked exclusive, we must reuse
3659 * the page without further checks.
3661 if (folio && folio_test_anon(folio) &&
3662 (PageAnonExclusive(vmf->page) || wp_can_reuse_anon_folio(folio, vma))) {
3663 if (!PageAnonExclusive(vmf->page))
3664 SetPageAnonExclusive(vmf->page);
3665 if (unlikely(unshare)) {
3666 pte_unmap_unlock(vmf->pte, vmf->ptl);
3669 wp_page_reuse(vmf, folio);
3673 * Ok, we need to copy. Oh, well..
3678 pte_unmap_unlock(vmf->pte, vmf->ptl);
3680 if (folio && folio_test_ksm(folio))
3681 count_vm_event(COW_KSM);
3683 return wp_page_copy(vmf);
3686 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3687 unsigned long start_addr, unsigned long end_addr,
3688 struct zap_details *details)
3690 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3693 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3694 pgoff_t first_index,
3696 struct zap_details *details)
3698 struct vm_area_struct *vma;
3699 pgoff_t vba, vea, zba, zea;
3701 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3702 vba = vma->vm_pgoff;
3703 vea = vba + vma_pages(vma) - 1;
3704 zba = max(first_index, vba);
3705 zea = min(last_index, vea);
3707 unmap_mapping_range_vma(vma,
3708 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3709 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3715 * unmap_mapping_folio() - Unmap single folio from processes.
3716 * @folio: The locked folio to be unmapped.
3718 * Unmap this folio from any userspace process which still has it mmaped.
3719 * Typically, for efficiency, the range of nearby pages has already been
3720 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3721 * truncation or invalidation holds the lock on a folio, it may find that
3722 * the page has been remapped again: and then uses unmap_mapping_folio()
3723 * to unmap it finally.
3725 void unmap_mapping_folio(struct folio *folio)
3727 struct address_space *mapping = folio->mapping;
3728 struct zap_details details = { };
3729 pgoff_t first_index;
3732 VM_BUG_ON(!folio_test_locked(folio));
3734 first_index = folio->index;
3735 last_index = folio_next_index(folio) - 1;
3737 details.even_cows = false;
3738 details.single_folio = folio;
3739 details.zap_flags = ZAP_FLAG_DROP_MARKER;
3741 i_mmap_lock_read(mapping);
3742 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3743 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3744 last_index, &details);
3745 i_mmap_unlock_read(mapping);
3749 * unmap_mapping_pages() - Unmap pages from processes.
3750 * @mapping: The address space containing pages to be unmapped.
3751 * @start: Index of first page to be unmapped.
3752 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3753 * @even_cows: Whether to unmap even private COWed pages.
3755 * Unmap the pages in this address space from any userspace process which
3756 * has them mmaped. Generally, you want to remove COWed pages as well when
3757 * a file is being truncated, but not when invalidating pages from the page
3760 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3761 pgoff_t nr, bool even_cows)
3763 struct zap_details details = { };
3764 pgoff_t first_index = start;
3765 pgoff_t last_index = start + nr - 1;
3767 details.even_cows = even_cows;
3768 if (last_index < first_index)
3769 last_index = ULONG_MAX;
3771 i_mmap_lock_read(mapping);
3772 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3773 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3774 last_index, &details);
3775 i_mmap_unlock_read(mapping);
3777 EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3780 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3781 * address_space corresponding to the specified byte range in the underlying
3784 * @mapping: the address space containing mmaps to be unmapped.
3785 * @holebegin: byte in first page to unmap, relative to the start of
3786 * the underlying file. This will be rounded down to a PAGE_SIZE
3787 * boundary. Note that this is different from truncate_pagecache(), which
3788 * must keep the partial page. In contrast, we must get rid of
3790 * @holelen: size of prospective hole in bytes. This will be rounded
3791 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3793 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3794 * but 0 when invalidating pagecache, don't throw away private data.
3796 void unmap_mapping_range(struct address_space *mapping,
3797 loff_t const holebegin, loff_t const holelen, int even_cows)
3799 pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT;
3800 pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT;
3802 /* Check for overflow. */
3803 if (sizeof(holelen) > sizeof(hlen)) {
3805 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3806 if (holeend & ~(long long)ULONG_MAX)
3807 hlen = ULONG_MAX - hba + 1;
3810 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3812 EXPORT_SYMBOL(unmap_mapping_range);
3815 * Restore a potential device exclusive pte to a working pte entry
3817 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3819 struct folio *folio = page_folio(vmf->page);
3820 struct vm_area_struct *vma = vmf->vma;
3821 struct mmu_notifier_range range;
3825 * We need a reference to lock the folio because we don't hold
3826 * the PTL so a racing thread can remove the device-exclusive
3827 * entry and unmap it. If the folio is free the entry must
3828 * have been removed already. If it happens to have already
3829 * been re-allocated after being freed all we do is lock and
3832 if (!folio_try_get(folio))
3835 ret = folio_lock_or_retry(folio, vmf);
3840 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
3841 vma->vm_mm, vmf->address & PAGE_MASK,
3842 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3843 mmu_notifier_invalidate_range_start(&range);
3845 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3847 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3848 restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte);
3851 pte_unmap_unlock(vmf->pte, vmf->ptl);
3852 folio_unlock(folio);
3855 mmu_notifier_invalidate_range_end(&range);
3859 static inline bool should_try_to_free_swap(struct folio *folio,
3860 struct vm_area_struct *vma,
3861 unsigned int fault_flags)
3863 if (!folio_test_swapcache(folio))
3865 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3866 folio_test_mlocked(folio))
3869 * If we want to map a page that's in the swapcache writable, we
3870 * have to detect via the refcount if we're really the exclusive
3871 * user. Try freeing the swapcache to get rid of the swapcache
3872 * reference only in case it's likely that we'll be the exlusive user.
3874 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3875 folio_ref_count(folio) == 2;
3878 static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3880 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
3881 vmf->address, &vmf->ptl);
3885 * Be careful so that we will only recover a special uffd-wp pte into a
3886 * none pte. Otherwise it means the pte could have changed, so retry.
3888 * This should also cover the case where e.g. the pte changed
3889 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3890 * So is_pte_marker() check is not enough to safely drop the pte.
3892 if (pte_same(vmf->orig_pte, ptep_get(vmf->pte)))
3893 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
3894 pte_unmap_unlock(vmf->pte, vmf->ptl);
3898 static vm_fault_t do_pte_missing(struct vm_fault *vmf)
3900 if (vma_is_anonymous(vmf->vma))
3901 return do_anonymous_page(vmf);
3903 return do_fault(vmf);
3907 * This is actually a page-missing access, but with uffd-wp special pte
3908 * installed. It means this pte was wr-protected before being unmapped.
3910 static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3913 * Just in case there're leftover special ptes even after the region
3914 * got unregistered - we can simply clear them.
3916 if (unlikely(!userfaultfd_wp(vmf->vma)))
3917 return pte_marker_clear(vmf);
3919 return do_pte_missing(vmf);
3922 static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3924 swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
3925 unsigned long marker = pte_marker_get(entry);
3928 * PTE markers should never be empty. If anything weird happened,
3929 * the best thing to do is to kill the process along with its mm.
3931 if (WARN_ON_ONCE(!marker))
3932 return VM_FAULT_SIGBUS;
3934 /* Higher priority than uffd-wp when data corrupted */
3935 if (marker & PTE_MARKER_POISONED)
3936 return VM_FAULT_HWPOISON;
3938 if (pte_marker_entry_uffd_wp(entry))
3939 return pte_marker_handle_uffd_wp(vmf);
3941 /* This is an unknown pte marker */
3942 return VM_FAULT_SIGBUS;
3946 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3947 * but allow concurrent faults), and pte mapped but not yet locked.
3948 * We return with pte unmapped and unlocked.
3950 * We return with the mmap_lock locked or unlocked in the same cases
3951 * as does filemap_fault().
3953 vm_fault_t do_swap_page(struct vm_fault *vmf)
3955 struct vm_area_struct *vma = vmf->vma;
3956 struct folio *swapcache, *folio = NULL;
3958 struct swap_info_struct *si = NULL;
3959 rmap_t rmap_flags = RMAP_NONE;
3960 bool need_clear_cache = false;
3961 bool exclusive = false;
3965 void *shadow = NULL;
3967 if (!pte_unmap_same(vmf))
3970 entry = pte_to_swp_entry(vmf->orig_pte);
3971 if (unlikely(non_swap_entry(entry))) {
3972 if (is_migration_entry(entry)) {
3973 migration_entry_wait(vma->vm_mm, vmf->pmd,
3975 } else if (is_device_exclusive_entry(entry)) {
3976 vmf->page = pfn_swap_entry_to_page(entry);
3977 ret = remove_device_exclusive_entry(vmf);
3978 } else if (is_device_private_entry(entry)) {
3979 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3981 * migrate_to_ram is not yet ready to operate
3985 ret = VM_FAULT_RETRY;
3989 vmf->page = pfn_swap_entry_to_page(entry);
3990 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3991 vmf->address, &vmf->ptl);
3992 if (unlikely(!vmf->pte ||
3993 !pte_same(ptep_get(vmf->pte),
3998 * Get a page reference while we know the page can't be
4001 get_page(vmf->page);
4002 pte_unmap_unlock(vmf->pte, vmf->ptl);
4003 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
4004 put_page(vmf->page);
4005 } else if (is_hwpoison_entry(entry)) {
4006 ret = VM_FAULT_HWPOISON;
4007 } else if (is_pte_marker_entry(entry)) {
4008 ret = handle_pte_marker(vmf);
4010 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
4011 ret = VM_FAULT_SIGBUS;
4016 /* Prevent swapoff from happening to us. */
4017 si = get_swap_device(entry);
4021 folio = swap_cache_get_folio(entry, vma, vmf->address);
4023 page = folio_file_page(folio, swp_offset(entry));
4027 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
4028 __swap_count(entry) == 1) {
4030 * Prevent parallel swapin from proceeding with
4031 * the cache flag. Otherwise, another thread may
4032 * finish swapin first, free the entry, and swapout
4033 * reusing the same entry. It's undetectable as
4034 * pte_same() returns true due to entry reuse.
4036 if (swapcache_prepare(entry)) {
4037 /* Relax a bit to prevent rapid repeated page faults */
4038 schedule_timeout_uninterruptible(1);
4041 need_clear_cache = true;
4043 /* skip swapcache */
4044 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0,
4045 vma, vmf->address, false);
4046 page = &folio->page;
4048 __folio_set_locked(folio);
4049 __folio_set_swapbacked(folio);
4051 if (mem_cgroup_swapin_charge_folio(folio,
4052 vma->vm_mm, GFP_KERNEL,
4057 mem_cgroup_swapin_uncharge_swap(entry);
4059 shadow = get_shadow_from_swap_cache(entry);
4061 workingset_refault(folio, shadow);
4063 folio_add_lru(folio);
4065 /* To provide entry to swap_read_folio() */
4066 folio->swap = entry;
4067 swap_read_folio(folio, true, NULL);
4068 folio->private = NULL;
4071 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
4074 folio = page_folio(page);
4080 * Back out if somebody else faulted in this pte
4081 * while we released the pte lock.
4083 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4084 vmf->address, &vmf->ptl);
4085 if (likely(vmf->pte &&
4086 pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4091 /* Had to read the page from swap area: Major fault */
4092 ret = VM_FAULT_MAJOR;
4093 count_vm_event(PGMAJFAULT);
4094 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
4095 } else if (PageHWPoison(page)) {
4097 * hwpoisoned dirty swapcache pages are kept for killing
4098 * owner processes (which may be unknown at hwpoison time)
4100 ret = VM_FAULT_HWPOISON;
4104 ret |= folio_lock_or_retry(folio, vmf);
4105 if (ret & VM_FAULT_RETRY)
4110 * Make sure folio_free_swap() or swapoff did not release the
4111 * swapcache from under us. The page pin, and pte_same test
4112 * below, are not enough to exclude that. Even if it is still
4113 * swapcache, we need to check that the page's swap has not
4116 if (unlikely(!folio_test_swapcache(folio) ||
4117 page_swap_entry(page).val != entry.val))
4121 * KSM sometimes has to copy on read faults, for example, if
4122 * page->index of !PageKSM() pages would be nonlinear inside the
4123 * anon VMA -- PageKSM() is lost on actual swapout.
4125 folio = ksm_might_need_to_copy(folio, vma, vmf->address);
4126 if (unlikely(!folio)) {
4130 } else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
4131 ret = VM_FAULT_HWPOISON;
4135 if (folio != swapcache)
4136 page = folio_page(folio, 0);
4139 * If we want to map a page that's in the swapcache writable, we
4140 * have to detect via the refcount if we're really the exclusive
4141 * owner. Try removing the extra reference from the local LRU
4142 * caches if required.
4144 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
4145 !folio_test_ksm(folio) && !folio_test_lru(folio))
4149 folio_throttle_swaprate(folio, GFP_KERNEL);
4152 * Back out if somebody else already faulted in this pte.
4154 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
4156 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
4159 if (unlikely(!folio_test_uptodate(folio))) {
4160 ret = VM_FAULT_SIGBUS;
4165 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
4166 * must never point at an anonymous page in the swapcache that is
4167 * PG_anon_exclusive. Sanity check that this holds and especially, that
4168 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
4169 * check after taking the PT lock and making sure that nobody
4170 * concurrently faulted in this page and set PG_anon_exclusive.
4172 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
4173 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
4176 * Check under PT lock (to protect against concurrent fork() sharing
4177 * the swap entry concurrently) for certainly exclusive pages.
4179 if (!folio_test_ksm(folio)) {
4180 exclusive = pte_swp_exclusive(vmf->orig_pte);
4181 if (folio != swapcache) {
4183 * We have a fresh page that is not exposed to the
4184 * swapcache -> certainly exclusive.
4187 } else if (exclusive && folio_test_writeback(folio) &&
4188 data_race(si->flags & SWP_STABLE_WRITES)) {
4190 * This is tricky: not all swap backends support
4191 * concurrent page modifications while under writeback.
4193 * So if we stumble over such a page in the swapcache
4194 * we must not set the page exclusive, otherwise we can
4195 * map it writable without further checks and modify it
4196 * while still under writeback.
4198 * For these problematic swap backends, simply drop the
4199 * exclusive marker: this is perfectly fine as we start
4200 * writeback only if we fully unmapped the page and
4201 * there are no unexpected references on the page after
4202 * unmapping succeeded. After fully unmapped, no
4203 * further GUP references (FOLL_GET and FOLL_PIN) can
4204 * appear, so dropping the exclusive marker and mapping
4205 * it only R/O is fine.
4212 * Some architectures may have to restore extra metadata to the page
4213 * when reading from swap. This metadata may be indexed by swap entry
4214 * so this must be called before swap_free().
4216 arch_swap_restore(folio_swap(entry, folio), folio);
4219 * Remove the swap entry and conditionally try to free up the swapcache.
4220 * We're already holding a reference on the page but haven't mapped it
4224 if (should_try_to_free_swap(folio, vma, vmf->flags))
4225 folio_free_swap(folio);
4227 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4228 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
4229 pte = mk_pte(page, vma->vm_page_prot);
4232 * Same logic as in do_wp_page(); however, optimize for pages that are
4233 * certainly not shared either because we just allocated them without
4234 * exposing them to the swapcache or because the swap entry indicates
4237 if (!folio_test_ksm(folio) &&
4238 (exclusive || folio_ref_count(folio) == 1)) {
4239 if (vmf->flags & FAULT_FLAG_WRITE) {
4240 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
4241 vmf->flags &= ~FAULT_FLAG_WRITE;
4243 rmap_flags |= RMAP_EXCLUSIVE;
4245 flush_icache_page(vma, page);
4246 if (pte_swp_soft_dirty(vmf->orig_pte))
4247 pte = pte_mksoft_dirty(pte);
4248 if (pte_swp_uffd_wp(vmf->orig_pte))
4249 pte = pte_mkuffd_wp(pte);
4250 vmf->orig_pte = pte;
4252 /* ksm created a completely new copy */
4253 if (unlikely(folio != swapcache && swapcache)) {
4254 folio_add_new_anon_rmap(folio, vma, vmf->address);
4255 folio_add_lru_vma(folio, vma);
4257 folio_add_anon_rmap_pte(folio, page, vma, vmf->address,
4261 VM_BUG_ON(!folio_test_anon(folio) ||
4262 (pte_write(pte) && !PageAnonExclusive(page)));
4263 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4264 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
4266 folio_unlock(folio);
4267 if (folio != swapcache && swapcache) {
4269 * Hold the lock to avoid the swap entry to be reused
4270 * until we take the PT lock for the pte_same() check
4271 * (to avoid false positives from pte_same). For
4272 * further safety release the lock after the swap_free
4273 * so that the swap count won't change under a
4274 * parallel locked swapcache.
4276 folio_unlock(swapcache);
4277 folio_put(swapcache);
4280 if (vmf->flags & FAULT_FLAG_WRITE) {
4281 ret |= do_wp_page(vmf);
4282 if (ret & VM_FAULT_ERROR)
4283 ret &= VM_FAULT_ERROR;
4287 /* No need to invalidate - it was non-present before */
4288 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
4291 pte_unmap_unlock(vmf->pte, vmf->ptl);
4293 /* Clear the swap cache pin for direct swapin after PTL unlock */
4294 if (need_clear_cache)
4295 swapcache_clear(si, entry);
4297 put_swap_device(si);
4301 pte_unmap_unlock(vmf->pte, vmf->ptl);
4303 folio_unlock(folio);
4306 if (folio != swapcache && swapcache) {
4307 folio_unlock(swapcache);
4308 folio_put(swapcache);
4310 if (need_clear_cache)
4311 swapcache_clear(si, entry);
4313 put_swap_device(si);
4317 static bool pte_range_none(pte_t *pte, int nr_pages)
4321 for (i = 0; i < nr_pages; i++) {
4322 if (!pte_none(ptep_get_lockless(pte + i)))
4329 static struct folio *alloc_anon_folio(struct vm_fault *vmf)
4331 struct vm_area_struct *vma = vmf->vma;
4332 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4333 unsigned long orders;
4334 struct folio *folio;
4341 * If uffd is active for the vma we need per-page fault fidelity to
4342 * maintain the uffd semantics.
4344 if (unlikely(userfaultfd_armed(vma)))
4348 * Get a list of all the (large) orders below PMD_ORDER that are enabled
4349 * for this vma. Then filter out the orders that can't be allocated over
4350 * the faulting address and still be fully contained in the vma.
4352 orders = thp_vma_allowable_orders(vma, vma->vm_flags,
4353 TVA_IN_PF | TVA_ENFORCE_SYSFS, BIT(PMD_ORDER) - 1);
4354 orders = thp_vma_suitable_orders(vma, vmf->address, orders);
4359 pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK);
4361 return ERR_PTR(-EAGAIN);
4364 * Find the highest order where the aligned range is completely
4365 * pte_none(). Note that all remaining orders will be completely
4368 order = highest_order(orders);
4370 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4371 if (pte_range_none(pte + pte_index(addr), 1 << order))
4373 order = next_order(&orders, order);
4381 /* Try allocating the highest of the remaining orders. */
4382 gfp = vma_thp_gfp_mask(vma);
4384 addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
4385 folio = vma_alloc_folio(gfp, order, vma, addr, true);
4387 if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
4388 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
4392 folio_throttle_swaprate(folio, gfp);
4393 clear_huge_page(&folio->page, vmf->address, 1 << order);
4397 count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
4398 order = next_order(&orders, order);
4403 return folio_prealloc(vma->vm_mm, vma, vmf->address, true);
4407 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4408 * but allow concurrent faults), and pte mapped but not yet locked.
4409 * We return with mmap_lock still held, but pte unmapped and unlocked.
4411 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4413 struct vm_area_struct *vma = vmf->vma;
4414 unsigned long addr = vmf->address;
4415 struct folio *folio;
4421 /* File mapping without ->vm_ops ? */
4422 if (vma->vm_flags & VM_SHARED)
4423 return VM_FAULT_SIGBUS;
4426 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4427 * be distinguished from a transient failure of pte_offset_map().
4429 if (pte_alloc(vma->vm_mm, vmf->pmd))
4430 return VM_FAULT_OOM;
4432 /* Use the zero-page for reads */
4433 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4434 !mm_forbids_zeropage(vma->vm_mm)) {
4435 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
4436 vma->vm_page_prot));
4437 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4438 vmf->address, &vmf->ptl);
4441 if (vmf_pte_changed(vmf)) {
4442 update_mmu_tlb(vma, vmf->address, vmf->pte);
4445 ret = check_stable_address_space(vma->vm_mm);
4448 /* Deliver the page fault to userland, check inside PT lock */
4449 if (userfaultfd_missing(vma)) {
4450 pte_unmap_unlock(vmf->pte, vmf->ptl);
4451 return handle_userfault(vmf, VM_UFFD_MISSING);
4456 /* Allocate our own private page. */
4457 ret = vmf_anon_prepare(vmf);
4460 /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
4461 folio = alloc_anon_folio(vmf);
4467 nr_pages = folio_nr_pages(folio);
4468 addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
4471 * The memory barrier inside __folio_mark_uptodate makes sure that
4472 * preceding stores to the page contents become visible before
4473 * the set_pte_at() write.
4475 __folio_mark_uptodate(folio);
4477 entry = mk_pte(&folio->page, vma->vm_page_prot);
4478 entry = pte_sw_mkyoung(entry);
4479 if (vma->vm_flags & VM_WRITE)
4480 entry = pte_mkwrite(pte_mkdirty(entry), vma);
4482 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
4485 if (nr_pages == 1 && vmf_pte_changed(vmf)) {
4486 update_mmu_tlb(vma, addr, vmf->pte);
4488 } else if (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages)) {
4489 for (i = 0; i < nr_pages; i++)
4490 update_mmu_tlb(vma, addr + PAGE_SIZE * i, vmf->pte + i);
4494 ret = check_stable_address_space(vma->vm_mm);
4498 /* Deliver the page fault to userland, check inside PT lock */
4499 if (userfaultfd_missing(vma)) {
4500 pte_unmap_unlock(vmf->pte, vmf->ptl);
4502 return handle_userfault(vmf, VM_UFFD_MISSING);
4505 folio_ref_add(folio, nr_pages - 1);
4506 add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages);
4507 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4508 count_mthp_stat(folio_order(folio), MTHP_STAT_ANON_FAULT_ALLOC);
4510 folio_add_new_anon_rmap(folio, vma, addr);
4511 folio_add_lru_vma(folio, vma);
4513 if (vmf_orig_pte_uffd_wp(vmf))
4514 entry = pte_mkuffd_wp(entry);
4515 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages);
4517 /* No need to invalidate - it was non-present before */
4518 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages);
4521 pte_unmap_unlock(vmf->pte, vmf->ptl);
4527 return VM_FAULT_OOM;
4531 * The mmap_lock must have been held on entry, and may have been
4532 * released depending on flags and vma->vm_ops->fault() return value.
4533 * See filemap_fault() and __lock_page_retry().
4535 static vm_fault_t __do_fault(struct vm_fault *vmf)
4537 struct vm_area_struct *vma = vmf->vma;
4538 struct folio *folio;
4542 * Preallocate pte before we take page_lock because this might lead to
4543 * deadlocks for memcg reclaim which waits for pages under writeback:
4545 * SetPageWriteback(A)
4551 * wait_on_page_writeback(A)
4552 * SetPageWriteback(B)
4554 * # flush A, B to clear the writeback
4556 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
4557 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4558 if (!vmf->prealloc_pte)
4559 return VM_FAULT_OOM;
4562 ret = vma->vm_ops->fault(vmf);
4563 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4564 VM_FAULT_DONE_COW)))
4567 folio = page_folio(vmf->page);
4568 if (unlikely(PageHWPoison(vmf->page))) {
4569 vm_fault_t poisonret = VM_FAULT_HWPOISON;
4570 if (ret & VM_FAULT_LOCKED) {
4571 if (page_mapped(vmf->page))
4572 unmap_mapping_folio(folio);
4573 /* Retry if a clean folio was removed from the cache. */
4574 if (mapping_evict_folio(folio->mapping, folio))
4575 poisonret = VM_FAULT_NOPAGE;
4576 folio_unlock(folio);
4583 if (unlikely(!(ret & VM_FAULT_LOCKED)))
4586 VM_BUG_ON_PAGE(!folio_test_locked(folio), vmf->page);
4591 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4592 static void deposit_prealloc_pte(struct vm_fault *vmf)
4594 struct vm_area_struct *vma = vmf->vma;
4596 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4598 * We are going to consume the prealloc table,
4599 * count that as nr_ptes.
4601 mm_inc_nr_ptes(vma->vm_mm);
4602 vmf->prealloc_pte = NULL;
4605 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4607 struct folio *folio = page_folio(page);
4608 struct vm_area_struct *vma = vmf->vma;
4609 bool write = vmf->flags & FAULT_FLAG_WRITE;
4610 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4612 vm_fault_t ret = VM_FAULT_FALLBACK;
4614 if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
4617 if (page != &folio->page || folio_order(folio) != HPAGE_PMD_ORDER)
4621 * Just backoff if any subpage of a THP is corrupted otherwise
4622 * the corrupted page may mapped by PMD silently to escape the
4623 * check. This kind of THP just can be PTE mapped. Access to
4624 * the corrupted subpage should trigger SIGBUS as expected.
4626 if (unlikely(folio_test_has_hwpoisoned(folio)))
4630 * Archs like ppc64 need additional space to store information
4631 * related to pte entry. Use the preallocated table for that.
4633 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4634 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4635 if (!vmf->prealloc_pte)
4636 return VM_FAULT_OOM;
4639 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4640 if (unlikely(!pmd_none(*vmf->pmd)))
4643 flush_icache_pages(vma, page, HPAGE_PMD_NR);
4645 entry = mk_huge_pmd(page, vma->vm_page_prot);
4647 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
4649 add_mm_counter(vma->vm_mm, mm_counter_file(folio), HPAGE_PMD_NR);
4650 folio_add_file_rmap_pmd(folio, page, vma);
4653 * deposit and withdraw with pmd lock held
4655 if (arch_needs_pgtable_deposit())
4656 deposit_prealloc_pte(vmf);
4658 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
4660 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
4662 /* fault is handled */
4664 count_vm_event(THP_FILE_MAPPED);
4666 spin_unlock(vmf->ptl);
4670 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4672 return VM_FAULT_FALLBACK;
4677 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4678 * @vmf: Fault decription.
4679 * @folio: The folio that contains @page.
4680 * @page: The first page to create a PTE for.
4681 * @nr: The number of PTEs to create.
4682 * @addr: The first address to create a PTE for.
4684 void set_pte_range(struct vm_fault *vmf, struct folio *folio,
4685 struct page *page, unsigned int nr, unsigned long addr)
4687 struct vm_area_struct *vma = vmf->vma;
4688 bool write = vmf->flags & FAULT_FLAG_WRITE;
4689 bool prefault = in_range(vmf->address, addr, nr * PAGE_SIZE);
4692 flush_icache_pages(vma, page, nr);
4693 entry = mk_pte(page, vma->vm_page_prot);
4695 if (prefault && arch_wants_old_prefaulted_pte())
4696 entry = pte_mkold(entry);
4698 entry = pte_sw_mkyoung(entry);
4701 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4702 if (unlikely(vmf_orig_pte_uffd_wp(vmf)))
4703 entry = pte_mkuffd_wp(entry);
4704 /* copy-on-write page */
4705 if (write && !(vma->vm_flags & VM_SHARED)) {
4706 VM_BUG_ON_FOLIO(nr != 1, folio);
4707 folio_add_new_anon_rmap(folio, vma, addr);
4708 folio_add_lru_vma(folio, vma);
4710 folio_add_file_rmap_ptes(folio, page, nr, vma);
4712 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr);
4714 /* no need to invalidate: a not-present page won't be cached */
4715 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr);
4718 static bool vmf_pte_changed(struct vm_fault *vmf)
4720 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4721 return !pte_same(ptep_get(vmf->pte), vmf->orig_pte);
4723 return !pte_none(ptep_get(vmf->pte));
4727 * finish_fault - finish page fault once we have prepared the page to fault
4729 * @vmf: structure describing the fault
4731 * This function handles all that is needed to finish a page fault once the
4732 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4733 * given page, adds reverse page mapping, handles memcg charges and LRU
4736 * The function expects the page to be locked and on success it consumes a
4737 * reference of a page being mapped (for the PTE which maps it).
4739 * Return: %0 on success, %VM_FAULT_ code in case of error.
4741 vm_fault_t finish_fault(struct vm_fault *vmf)
4743 struct vm_area_struct *vma = vmf->vma;
4746 bool is_cow = (vmf->flags & FAULT_FLAG_WRITE) &&
4747 !(vma->vm_flags & VM_SHARED);
4749 /* Did we COW the page? */
4751 page = vmf->cow_page;
4756 * check even for read faults because we might have lost our CoWed
4759 if (!(vma->vm_flags & VM_SHARED)) {
4760 ret = check_stable_address_space(vma->vm_mm);
4765 if (pmd_none(*vmf->pmd)) {
4766 if (PageTransCompound(page)) {
4767 ret = do_set_pmd(vmf, page);
4768 if (ret != VM_FAULT_FALLBACK)
4772 if (vmf->prealloc_pte)
4773 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4774 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4775 return VM_FAULT_OOM;
4778 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4779 vmf->address, &vmf->ptl);
4781 return VM_FAULT_NOPAGE;
4783 /* Re-check under ptl */
4784 if (likely(!vmf_pte_changed(vmf))) {
4785 struct folio *folio = page_folio(page);
4786 int type = is_cow ? MM_ANONPAGES : mm_counter_file(folio);
4788 set_pte_range(vmf, folio, page, 1, vmf->address);
4789 add_mm_counter(vma->vm_mm, type, 1);
4792 update_mmu_tlb(vma, vmf->address, vmf->pte);
4793 ret = VM_FAULT_NOPAGE;
4796 pte_unmap_unlock(vmf->pte, vmf->ptl);
4800 static unsigned long fault_around_pages __read_mostly =
4801 65536 >> PAGE_SHIFT;
4803 #ifdef CONFIG_DEBUG_FS
4804 static int fault_around_bytes_get(void *data, u64 *val)
4806 *val = fault_around_pages << PAGE_SHIFT;
4811 * fault_around_bytes must be rounded down to the nearest page order as it's
4812 * what do_fault_around() expects to see.
4814 static int fault_around_bytes_set(void *data, u64 val)
4816 if (val / PAGE_SIZE > PTRS_PER_PTE)
4820 * The minimum value is 1 page, however this results in no fault-around
4821 * at all. See should_fault_around().
4823 val = max(val, PAGE_SIZE);
4824 fault_around_pages = rounddown_pow_of_two(val) >> PAGE_SHIFT;
4828 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4829 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4831 static int __init fault_around_debugfs(void)
4833 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4834 &fault_around_bytes_fops);
4837 late_initcall(fault_around_debugfs);
4841 * do_fault_around() tries to map few pages around the fault address. The hope
4842 * is that the pages will be needed soon and this will lower the number of
4845 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4846 * not ready to be mapped: not up-to-date, locked, etc.
4848 * This function doesn't cross VMA or page table boundaries, in order to call
4849 * map_pages() and acquire a PTE lock only once.
4851 * fault_around_pages defines how many pages we'll try to map.
4852 * do_fault_around() expects it to be set to a power of two less than or equal
4855 * The virtual address of the area that we map is naturally aligned to
4856 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4857 * (and therefore to page order). This way it's easier to guarantee
4858 * that we don't cross page table boundaries.
4860 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4862 pgoff_t nr_pages = READ_ONCE(fault_around_pages);
4863 pgoff_t pte_off = pte_index(vmf->address);
4864 /* The page offset of vmf->address within the VMA. */
4865 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
4866 pgoff_t from_pte, to_pte;
4869 /* The PTE offset of the start address, clamped to the VMA. */
4870 from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
4871 pte_off - min(pte_off, vma_off));
4873 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4874 to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
4875 pte_off + vma_pages(vmf->vma) - vma_off) - 1;
4877 if (pmd_none(*vmf->pmd)) {
4878 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4879 if (!vmf->prealloc_pte)
4880 return VM_FAULT_OOM;
4884 ret = vmf->vma->vm_ops->map_pages(vmf,
4885 vmf->pgoff + from_pte - pte_off,
4886 vmf->pgoff + to_pte - pte_off);
4892 /* Return true if we should do read fault-around, false otherwise */
4893 static inline bool should_fault_around(struct vm_fault *vmf)
4895 /* No ->map_pages? No way to fault around... */
4896 if (!vmf->vma->vm_ops->map_pages)
4899 if (uffd_disable_fault_around(vmf->vma))
4902 /* A single page implies no faulting 'around' at all. */
4903 return fault_around_pages > 1;
4906 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4909 struct folio *folio;
4912 * Let's call ->map_pages() first and use ->fault() as fallback
4913 * if page by the offset is not ready to be mapped (cold cache or
4916 if (should_fault_around(vmf)) {
4917 ret = do_fault_around(vmf);
4922 ret = vmf_can_call_fault(vmf);
4926 ret = __do_fault(vmf);
4927 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4930 ret |= finish_fault(vmf);
4931 folio = page_folio(vmf->page);
4932 folio_unlock(folio);
4933 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4938 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4940 struct vm_area_struct *vma = vmf->vma;
4941 struct folio *folio;
4944 ret = vmf_can_call_fault(vmf);
4946 ret = vmf_anon_prepare(vmf);
4950 folio = folio_prealloc(vma->vm_mm, vma, vmf->address, false);
4952 return VM_FAULT_OOM;
4954 vmf->cow_page = &folio->page;
4956 ret = __do_fault(vmf);
4957 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4959 if (ret & VM_FAULT_DONE_COW)
4962 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4963 __folio_mark_uptodate(folio);
4965 ret |= finish_fault(vmf);
4966 unlock_page(vmf->page);
4967 put_page(vmf->page);
4968 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4976 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4978 struct vm_area_struct *vma = vmf->vma;
4979 vm_fault_t ret, tmp;
4980 struct folio *folio;
4982 ret = vmf_can_call_fault(vmf);
4986 ret = __do_fault(vmf);
4987 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4990 folio = page_folio(vmf->page);
4993 * Check if the backing address space wants to know that the page is
4994 * about to become writable
4996 if (vma->vm_ops->page_mkwrite) {
4997 folio_unlock(folio);
4998 tmp = do_page_mkwrite(vmf, folio);
4999 if (unlikely(!tmp ||
5000 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
5006 ret |= finish_fault(vmf);
5007 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
5009 folio_unlock(folio);
5014 ret |= fault_dirty_shared_page(vmf);
5019 * We enter with non-exclusive mmap_lock (to exclude vma changes,
5020 * but allow concurrent faults).
5021 * The mmap_lock may have been released depending on flags and our
5022 * return value. See filemap_fault() and __folio_lock_or_retry().
5023 * If mmap_lock is released, vma may become invalid (for example
5024 * by other thread calling munmap()).
5026 static vm_fault_t do_fault(struct vm_fault *vmf)
5028 struct vm_area_struct *vma = vmf->vma;
5029 struct mm_struct *vm_mm = vma->vm_mm;
5033 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
5035 if (!vma->vm_ops->fault) {
5036 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
5037 vmf->address, &vmf->ptl);
5038 if (unlikely(!vmf->pte))
5039 ret = VM_FAULT_SIGBUS;
5042 * Make sure this is not a temporary clearing of pte
5043 * by holding ptl and checking again. A R/M/W update
5044 * of pte involves: take ptl, clearing the pte so that
5045 * we don't have concurrent modification by hardware
5046 * followed by an update.
5048 if (unlikely(pte_none(ptep_get(vmf->pte))))
5049 ret = VM_FAULT_SIGBUS;
5051 ret = VM_FAULT_NOPAGE;
5053 pte_unmap_unlock(vmf->pte, vmf->ptl);
5055 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
5056 ret = do_read_fault(vmf);
5057 else if (!(vma->vm_flags & VM_SHARED))
5058 ret = do_cow_fault(vmf);
5060 ret = do_shared_fault(vmf);
5062 /* preallocated pagetable is unused: free it */
5063 if (vmf->prealloc_pte) {
5064 pte_free(vm_mm, vmf->prealloc_pte);
5065 vmf->prealloc_pte = NULL;
5070 int numa_migrate_prep(struct folio *folio, struct vm_fault *vmf,
5071 unsigned long addr, int page_nid, int *flags)
5073 struct vm_area_struct *vma = vmf->vma;
5077 /* Record the current PID acceesing VMA */
5078 vma_set_access_pid_bit(vma);
5080 count_vm_numa_event(NUMA_HINT_FAULTS);
5081 if (page_nid == numa_node_id()) {
5082 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
5083 *flags |= TNF_FAULT_LOCAL;
5086 return mpol_misplaced(folio, vmf, addr);
5089 static void numa_rebuild_single_mapping(struct vm_fault *vmf, struct vm_area_struct *vma,
5090 unsigned long fault_addr, pte_t *fault_pte,
5095 old_pte = ptep_modify_prot_start(vma, fault_addr, fault_pte);
5096 pte = pte_modify(old_pte, vma->vm_page_prot);
5097 pte = pte_mkyoung(pte);
5099 pte = pte_mkwrite(pte, vma);
5100 ptep_modify_prot_commit(vma, fault_addr, fault_pte, old_pte, pte);
5101 update_mmu_cache_range(vmf, vma, fault_addr, fault_pte, 1);
5104 static void numa_rebuild_large_mapping(struct vm_fault *vmf, struct vm_area_struct *vma,
5105 struct folio *folio, pte_t fault_pte,
5106 bool ignore_writable, bool pte_write_upgrade)
5108 int nr = pte_pfn(fault_pte) - folio_pfn(folio);
5109 unsigned long start = max(vmf->address - nr * PAGE_SIZE, vma->vm_start);
5110 unsigned long end = min(vmf->address + (folio_nr_pages(folio) - nr) * PAGE_SIZE, vma->vm_end);
5111 pte_t *start_ptep = vmf->pte - (vmf->address - start) / PAGE_SIZE;
5114 /* Restore all PTEs' mapping of the large folio */
5115 for (addr = start; addr != end; start_ptep++, addr += PAGE_SIZE) {
5116 pte_t ptent = ptep_get(start_ptep);
5117 bool writable = false;
5119 if (!pte_present(ptent) || !pte_protnone(ptent))
5122 if (pfn_folio(pte_pfn(ptent)) != folio)
5125 if (!ignore_writable) {
5126 ptent = pte_modify(ptent, vma->vm_page_prot);
5127 writable = pte_write(ptent);
5128 if (!writable && pte_write_upgrade &&
5129 can_change_pte_writable(vma, addr, ptent))
5133 numa_rebuild_single_mapping(vmf, vma, addr, start_ptep, writable);
5137 static vm_fault_t do_numa_page(struct vm_fault *vmf)
5139 struct vm_area_struct *vma = vmf->vma;
5140 struct folio *folio = NULL;
5141 int nid = NUMA_NO_NODE;
5142 bool writable = false, ignore_writable = false;
5143 bool pte_write_upgrade = vma_wants_manual_pte_write_upgrade(vma);
5147 int flags = 0, nr_pages;
5150 * The pte cannot be used safely until we verify, while holding the page
5151 * table lock, that its contents have not changed during fault handling.
5153 spin_lock(vmf->ptl);
5154 /* Read the live PTE from the page tables: */
5155 old_pte = ptep_get(vmf->pte);
5157 if (unlikely(!pte_same(old_pte, vmf->orig_pte))) {
5158 pte_unmap_unlock(vmf->pte, vmf->ptl);
5162 pte = pte_modify(old_pte, vma->vm_page_prot);
5165 * Detect now whether the PTE could be writable; this information
5166 * is only valid while holding the PT lock.
5168 writable = pte_write(pte);
5169 if (!writable && pte_write_upgrade &&
5170 can_change_pte_writable(vma, vmf->address, pte))
5173 folio = vm_normal_folio(vma, vmf->address, pte);
5174 if (!folio || folio_is_zone_device(folio))
5178 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
5179 * much anyway since they can be in shared cache state. This misses
5180 * the case where a mapping is writable but the process never writes
5181 * to it but pte_write gets cleared during protection updates and
5182 * pte_dirty has unpredictable behaviour between PTE scan updates,
5183 * background writeback, dirty balancing and application behaviour.
5186 flags |= TNF_NO_GROUP;
5189 * Flag if the folio is shared between multiple address spaces. This
5190 * is later used when determining whether to group tasks together
5192 if (folio_likely_mapped_shared(folio) && (vma->vm_flags & VM_SHARED))
5193 flags |= TNF_SHARED;
5195 nid = folio_nid(folio);
5196 nr_pages = folio_nr_pages(folio);
5198 * For memory tiering mode, cpupid of slow memory page is used
5199 * to record page access time. So use default value.
5201 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
5202 !node_is_toptier(nid))
5203 last_cpupid = (-1 & LAST_CPUPID_MASK);
5205 last_cpupid = folio_last_cpupid(folio);
5206 target_nid = numa_migrate_prep(folio, vmf, vmf->address, nid, &flags);
5207 if (target_nid == NUMA_NO_NODE) {
5211 pte_unmap_unlock(vmf->pte, vmf->ptl);
5213 ignore_writable = true;
5215 /* Migrate to the requested node */
5216 if (migrate_misplaced_folio(folio, vma, target_nid)) {
5218 flags |= TNF_MIGRATED;
5220 flags |= TNF_MIGRATE_FAIL;
5221 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
5222 vmf->address, &vmf->ptl);
5223 if (unlikely(!vmf->pte))
5225 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
5226 pte_unmap_unlock(vmf->pte, vmf->ptl);
5233 if (nid != NUMA_NO_NODE)
5234 task_numa_fault(last_cpupid, nid, nr_pages, flags);
5238 * Make it present again, depending on how arch implements
5239 * non-accessible ptes, some can allow access by kernel mode.
5241 if (folio && folio_test_large(folio))
5242 numa_rebuild_large_mapping(vmf, vma, folio, pte, ignore_writable,
5245 numa_rebuild_single_mapping(vmf, vma, vmf->address, vmf->pte,
5247 pte_unmap_unlock(vmf->pte, vmf->ptl);
5251 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
5253 struct vm_area_struct *vma = vmf->vma;
5254 if (vma_is_anonymous(vma))
5255 return do_huge_pmd_anonymous_page(vmf);
5256 if (vma->vm_ops->huge_fault)
5257 return vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5258 return VM_FAULT_FALLBACK;
5261 /* `inline' is required to avoid gcc 4.1.2 build error */
5262 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
5264 struct vm_area_struct *vma = vmf->vma;
5265 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
5268 if (vma_is_anonymous(vma)) {
5269 if (likely(!unshare) &&
5270 userfaultfd_huge_pmd_wp(vma, vmf->orig_pmd)) {
5271 if (userfaultfd_wp_async(vmf->vma))
5273 return handle_userfault(vmf, VM_UFFD_WP);
5275 return do_huge_pmd_wp_page(vmf);
5278 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5279 if (vma->vm_ops->huge_fault) {
5280 ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER);
5281 if (!(ret & VM_FAULT_FALLBACK))
5287 /* COW or write-notify handled on pte level: split pmd. */
5288 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
5290 return VM_FAULT_FALLBACK;
5293 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
5295 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5296 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5297 struct vm_area_struct *vma = vmf->vma;
5298 /* No support for anonymous transparent PUD pages yet */
5299 if (vma_is_anonymous(vma))
5300 return VM_FAULT_FALLBACK;
5301 if (vma->vm_ops->huge_fault)
5302 return vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5303 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5304 return VM_FAULT_FALLBACK;
5307 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
5309 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
5310 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
5311 struct vm_area_struct *vma = vmf->vma;
5314 /* No support for anonymous transparent PUD pages yet */
5315 if (vma_is_anonymous(vma))
5317 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
5318 if (vma->vm_ops->huge_fault) {
5319 ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER);
5320 if (!(ret & VM_FAULT_FALLBACK))
5325 /* COW or write-notify not handled on PUD level: split pud.*/
5326 __split_huge_pud(vma, vmf->pud, vmf->address);
5327 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
5328 return VM_FAULT_FALLBACK;
5332 * These routines also need to handle stuff like marking pages dirty
5333 * and/or accessed for architectures that don't do it in hardware (most
5334 * RISC architectures). The early dirtying is also good on the i386.
5336 * There is also a hook called "update_mmu_cache()" that architectures
5337 * with external mmu caches can use to update those (ie the Sparc or
5338 * PowerPC hashed page tables that act as extended TLBs).
5340 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
5341 * concurrent faults).
5343 * The mmap_lock may have been released depending on flags and our return value.
5344 * See filemap_fault() and __folio_lock_or_retry().
5346 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
5350 if (unlikely(pmd_none(*vmf->pmd))) {
5352 * Leave __pte_alloc() until later: because vm_ops->fault may
5353 * want to allocate huge page, and if we expose page table
5354 * for an instant, it will be difficult to retract from
5355 * concurrent faults and from rmap lookups.
5358 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
5361 * A regular pmd is established and it can't morph into a huge
5362 * pmd by anon khugepaged, since that takes mmap_lock in write
5363 * mode; but shmem or file collapse to THP could still morph
5364 * it into a huge pmd: just retry later if so.
5366 vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd,
5367 vmf->address, &vmf->ptl);
5368 if (unlikely(!vmf->pte))
5370 vmf->orig_pte = ptep_get_lockless(vmf->pte);
5371 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
5373 if (pte_none(vmf->orig_pte)) {
5374 pte_unmap(vmf->pte);
5380 return do_pte_missing(vmf);
5382 if (!pte_present(vmf->orig_pte))
5383 return do_swap_page(vmf);
5385 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
5386 return do_numa_page(vmf);
5388 spin_lock(vmf->ptl);
5389 entry = vmf->orig_pte;
5390 if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
5391 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
5394 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
5395 if (!pte_write(entry))
5396 return do_wp_page(vmf);
5397 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
5398 entry = pte_mkdirty(entry);
5400 entry = pte_mkyoung(entry);
5401 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
5402 vmf->flags & FAULT_FLAG_WRITE)) {
5403 update_mmu_cache_range(vmf, vmf->vma, vmf->address,
5406 /* Skip spurious TLB flush for retried page fault */
5407 if (vmf->flags & FAULT_FLAG_TRIED)
5410 * This is needed only for protection faults but the arch code
5411 * is not yet telling us if this is a protection fault or not.
5412 * This still avoids useless tlb flushes for .text page faults
5415 if (vmf->flags & FAULT_FLAG_WRITE)
5416 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
5420 pte_unmap_unlock(vmf->pte, vmf->ptl);
5425 * On entry, we hold either the VMA lock or the mmap_lock
5426 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5427 * the result, the mmap_lock is not held on exit. See filemap_fault()
5428 * and __folio_lock_or_retry().
5430 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
5431 unsigned long address, unsigned int flags)
5433 struct vm_fault vmf = {
5435 .address = address & PAGE_MASK,
5436 .real_address = address,
5438 .pgoff = linear_page_index(vma, address),
5439 .gfp_mask = __get_fault_gfp_mask(vma),
5441 struct mm_struct *mm = vma->vm_mm;
5442 unsigned long vm_flags = vma->vm_flags;
5447 pgd = pgd_offset(mm, address);
5448 p4d = p4d_alloc(mm, pgd, address);
5450 return VM_FAULT_OOM;
5452 vmf.pud = pud_alloc(mm, p4d, address);
5454 return VM_FAULT_OOM;
5456 if (pud_none(*vmf.pud) &&
5457 thp_vma_allowable_order(vma, vm_flags,
5458 TVA_IN_PF | TVA_ENFORCE_SYSFS, PUD_ORDER)) {
5459 ret = create_huge_pud(&vmf);
5460 if (!(ret & VM_FAULT_FALLBACK))
5463 pud_t orig_pud = *vmf.pud;
5466 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
5469 * TODO once we support anonymous PUDs: NUMA case and
5470 * FAULT_FLAG_UNSHARE handling.
5472 if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) {
5473 ret = wp_huge_pud(&vmf, orig_pud);
5474 if (!(ret & VM_FAULT_FALLBACK))
5477 huge_pud_set_accessed(&vmf, orig_pud);
5483 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
5485 return VM_FAULT_OOM;
5487 /* Huge pud page fault raced with pmd_alloc? */
5488 if (pud_trans_unstable(vmf.pud))
5491 if (pmd_none(*vmf.pmd) &&
5492 thp_vma_allowable_order(vma, vm_flags,
5493 TVA_IN_PF | TVA_ENFORCE_SYSFS, PMD_ORDER)) {
5494 ret = create_huge_pmd(&vmf);
5495 if (!(ret & VM_FAULT_FALLBACK))
5498 vmf.orig_pmd = pmdp_get_lockless(vmf.pmd);
5500 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
5501 VM_BUG_ON(thp_migration_supported() &&
5502 !is_pmd_migration_entry(vmf.orig_pmd));
5503 if (is_pmd_migration_entry(vmf.orig_pmd))
5504 pmd_migration_entry_wait(mm, vmf.pmd);
5507 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
5508 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
5509 return do_huge_pmd_numa_page(&vmf);
5511 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5512 !pmd_write(vmf.orig_pmd)) {
5513 ret = wp_huge_pmd(&vmf);
5514 if (!(ret & VM_FAULT_FALLBACK))
5517 huge_pmd_set_accessed(&vmf);
5523 return handle_pte_fault(&vmf);
5527 * mm_account_fault - Do page fault accounting
5528 * @mm: mm from which memcg should be extracted. It can be NULL.
5529 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5530 * of perf event counters, but we'll still do the per-task accounting to
5531 * the task who triggered this page fault.
5532 * @address: the faulted address.
5533 * @flags: the fault flags.
5534 * @ret: the fault retcode.
5536 * This will take care of most of the page fault accounting. Meanwhile, it
5537 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5538 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5539 * still be in per-arch page fault handlers at the entry of page fault.
5541 static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
5542 unsigned long address, unsigned int flags,
5547 /* Incomplete faults will be accounted upon completion. */
5548 if (ret & VM_FAULT_RETRY)
5552 * To preserve the behavior of older kernels, PGFAULT counters record
5553 * both successful and failed faults, as opposed to perf counters,
5554 * which ignore failed cases.
5556 count_vm_event(PGFAULT);
5557 count_memcg_event_mm(mm, PGFAULT);
5560 * Do not account for unsuccessful faults (e.g. when the address wasn't
5561 * valid). That includes arch_vma_access_permitted() failing before
5562 * reaching here. So this is not a "this many hardware page faults"
5563 * counter. We should use the hw profiling for that.
5565 if (ret & VM_FAULT_ERROR)
5569 * We define the fault as a major fault when the final successful fault
5570 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5571 * handle it immediately previously).
5573 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5581 * If the fault is done for GUP, regs will be NULL. We only do the
5582 * accounting for the per thread fault counters who triggered the
5583 * fault, and we skip the perf event updates.
5589 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
5591 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
5594 #ifdef CONFIG_LRU_GEN
5595 static void lru_gen_enter_fault(struct vm_area_struct *vma)
5597 /* the LRU algorithm only applies to accesses with recency */
5598 current->in_lru_fault = vma_has_recency(vma);
5601 static void lru_gen_exit_fault(void)
5603 current->in_lru_fault = false;
5606 static void lru_gen_enter_fault(struct vm_area_struct *vma)
5610 static void lru_gen_exit_fault(void)
5613 #endif /* CONFIG_LRU_GEN */
5615 static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5616 unsigned int *flags)
5618 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5619 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5620 return VM_FAULT_SIGSEGV;
5622 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5623 * just treat it like an ordinary read-fault otherwise.
5625 if (!is_cow_mapping(vma->vm_flags))
5626 *flags &= ~FAULT_FLAG_UNSHARE;
5627 } else if (*flags & FAULT_FLAG_WRITE) {
5628 /* Write faults on read-only mappings are impossible ... */
5629 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5630 return VM_FAULT_SIGSEGV;
5631 /* ... and FOLL_FORCE only applies to COW mappings. */
5632 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5633 !is_cow_mapping(vma->vm_flags)))
5634 return VM_FAULT_SIGSEGV;
5636 #ifdef CONFIG_PER_VMA_LOCK
5638 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5639 * the assumption that lock is dropped on VM_FAULT_RETRY.
5641 if (WARN_ON_ONCE((*flags &
5642 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) ==
5643 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)))
5644 return VM_FAULT_SIGSEGV;
5651 * By the time we get here, we already hold the mm semaphore
5653 * The mmap_lock may have been released depending on flags and our
5654 * return value. See filemap_fault() and __folio_lock_or_retry().
5656 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5657 unsigned int flags, struct pt_regs *regs)
5659 /* If the fault handler drops the mmap_lock, vma may be freed */
5660 struct mm_struct *mm = vma->vm_mm;
5663 __set_current_state(TASK_RUNNING);
5665 ret = sanitize_fault_flags(vma, &flags);
5669 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
5670 flags & FAULT_FLAG_INSTRUCTION,
5671 flags & FAULT_FLAG_REMOTE)) {
5672 ret = VM_FAULT_SIGSEGV;
5677 * Enable the memcg OOM handling for faults triggered in user
5678 * space. Kernel faults are handled more gracefully.
5680 if (flags & FAULT_FLAG_USER)
5681 mem_cgroup_enter_user_fault();
5683 lru_gen_enter_fault(vma);
5685 if (unlikely(is_vm_hugetlb_page(vma)))
5686 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
5688 ret = __handle_mm_fault(vma, address, flags);
5690 lru_gen_exit_fault();
5692 if (flags & FAULT_FLAG_USER) {
5693 mem_cgroup_exit_user_fault();
5695 * The task may have entered a memcg OOM situation but
5696 * if the allocation error was handled gracefully (no
5697 * VM_FAULT_OOM), there is no need to kill anything.
5698 * Just clean up the OOM state peacefully.
5700 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5701 mem_cgroup_oom_synchronize(false);
5704 mm_account_fault(mm, regs, address, flags, ret);
5708 EXPORT_SYMBOL_GPL(handle_mm_fault);
5710 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5711 #include <linux/extable.h>
5713 static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5715 if (likely(mmap_read_trylock(mm)))
5718 if (regs && !user_mode(regs)) {
5719 unsigned long ip = exception_ip(regs);
5720 if (!search_exception_tables(ip))
5724 return !mmap_read_lock_killable(mm);
5727 static inline bool mmap_upgrade_trylock(struct mm_struct *mm)
5730 * We don't have this operation yet.
5732 * It should be easy enough to do: it's basically a
5733 * atomic_long_try_cmpxchg_acquire()
5734 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5735 * it also needs the proper lockdep magic etc.
5740 static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5742 mmap_read_unlock(mm);
5743 if (regs && !user_mode(regs)) {
5744 unsigned long ip = exception_ip(regs);
5745 if (!search_exception_tables(ip))
5748 return !mmap_write_lock_killable(mm);
5752 * Helper for page fault handling.
5754 * This is kind of equivalend to "mmap_read_lock()" followed
5755 * by "find_extend_vma()", except it's a lot more careful about
5756 * the locking (and will drop the lock on failure).
5758 * For example, if we have a kernel bug that causes a page
5759 * fault, we don't want to just use mmap_read_lock() to get
5760 * the mm lock, because that would deadlock if the bug were
5761 * to happen while we're holding the mm lock for writing.
5763 * So this checks the exception tables on kernel faults in
5764 * order to only do this all for instructions that are actually
5765 * expected to fault.
5767 * We can also actually take the mm lock for writing if we
5768 * need to extend the vma, which helps the VM layer a lot.
5770 struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm,
5771 unsigned long addr, struct pt_regs *regs)
5773 struct vm_area_struct *vma;
5775 if (!get_mmap_lock_carefully(mm, regs))
5778 vma = find_vma(mm, addr);
5779 if (likely(vma && (vma->vm_start <= addr)))
5783 * Well, dang. We might still be successful, but only
5784 * if we can extend a vma to do so.
5786 if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) {
5787 mmap_read_unlock(mm);
5792 * We can try to upgrade the mmap lock atomically,
5793 * in which case we can continue to use the vma
5794 * we already looked up.
5796 * Otherwise we'll have to drop the mmap lock and
5797 * re-take it, and also look up the vma again,
5800 if (!mmap_upgrade_trylock(mm)) {
5801 if (!upgrade_mmap_lock_carefully(mm, regs))
5804 vma = find_vma(mm, addr);
5807 if (vma->vm_start <= addr)
5809 if (!(vma->vm_flags & VM_GROWSDOWN))
5813 if (expand_stack_locked(vma, addr))
5817 mmap_write_downgrade(mm);
5821 mmap_write_unlock(mm);
5826 #ifdef CONFIG_PER_VMA_LOCK
5828 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5829 * stable and not isolated. If the VMA is not found or is being modified the
5830 * function returns NULL.
5832 struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
5833 unsigned long address)
5835 MA_STATE(mas, &mm->mm_mt, address, address);
5836 struct vm_area_struct *vma;
5840 vma = mas_walk(&mas);
5844 if (!vma_start_read(vma))
5847 /* Check since vm_start/vm_end might change before we lock the VMA */
5848 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5849 goto inval_end_read;
5851 /* Check if the VMA got isolated after we found it */
5852 if (vma->detached) {
5854 count_vm_vma_lock_event(VMA_LOCK_MISS);
5855 /* The area was replaced with another one */
5866 count_vm_vma_lock_event(VMA_LOCK_ABORT);
5869 #endif /* CONFIG_PER_VMA_LOCK */
5871 #ifndef __PAGETABLE_P4D_FOLDED
5873 * Allocate p4d page table.
5874 * We've already handled the fast-path in-line.
5876 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5878 p4d_t *new = p4d_alloc_one(mm, address);
5882 spin_lock(&mm->page_table_lock);
5883 if (pgd_present(*pgd)) { /* Another has populated it */
5886 smp_wmb(); /* See comment in pmd_install() */
5887 pgd_populate(mm, pgd, new);
5889 spin_unlock(&mm->page_table_lock);
5892 #endif /* __PAGETABLE_P4D_FOLDED */
5894 #ifndef __PAGETABLE_PUD_FOLDED
5896 * Allocate page upper directory.
5897 * We've already handled the fast-path in-line.
5899 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5901 pud_t *new = pud_alloc_one(mm, address);
5905 spin_lock(&mm->page_table_lock);
5906 if (!p4d_present(*p4d)) {
5908 smp_wmb(); /* See comment in pmd_install() */
5909 p4d_populate(mm, p4d, new);
5910 } else /* Another has populated it */
5912 spin_unlock(&mm->page_table_lock);
5915 #endif /* __PAGETABLE_PUD_FOLDED */
5917 #ifndef __PAGETABLE_PMD_FOLDED
5919 * Allocate page middle directory.
5920 * We've already handled the fast-path in-line.
5922 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5925 pmd_t *new = pmd_alloc_one(mm, address);
5929 ptl = pud_lock(mm, pud);
5930 if (!pud_present(*pud)) {
5932 smp_wmb(); /* See comment in pmd_install() */
5933 pud_populate(mm, pud, new);
5934 } else { /* Another has populated it */
5940 #endif /* __PAGETABLE_PMD_FOLDED */
5943 * follow_pte - look up PTE at a user virtual address
5944 * @vma: the memory mapping
5945 * @address: user virtual address
5946 * @ptepp: location to store found PTE
5947 * @ptlp: location to store the lock for the PTE
5949 * On a successful return, the pointer to the PTE is stored in @ptepp;
5950 * the corresponding lock is taken and its location is stored in @ptlp.
5952 * The contents of the PTE are only stable until @ptlp is released using
5953 * pte_unmap_unlock(). This function will fail if the PTE is non-present.
5954 * Present PTEs may include PTEs that map refcounted pages, such as
5955 * anonymous folios in COW mappings.
5957 * Callers must be careful when relying on PTE content after
5958 * pte_unmap_unlock(). Especially if the PTE maps a refcounted page,
5959 * callers must protect against invalidation with MMU notifiers; otherwise
5960 * access to the PFN at a later point in time can trigger use-after-free.
5962 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5963 * should be taken for read.
5965 * This function must not be used to modify PTE content.
5967 * Return: zero on success, -ve otherwise.
5969 int follow_pte(struct vm_area_struct *vma, unsigned long address,
5970 pte_t **ptepp, spinlock_t **ptlp)
5972 struct mm_struct *mm = vma->vm_mm;
5979 mmap_assert_locked(mm);
5980 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5983 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5986 pgd = pgd_offset(mm, address);
5987 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
5990 p4d = p4d_offset(pgd, address);
5991 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
5994 pud = pud_offset(p4d, address);
5995 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
5998 pmd = pmd_offset(pud, address);
5999 VM_BUG_ON(pmd_trans_huge(*pmd));
6001 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
6004 if (!pte_present(ptep_get(ptep)))
6009 pte_unmap_unlock(ptep, *ptlp);
6013 EXPORT_SYMBOL_GPL(follow_pte);
6015 #ifdef CONFIG_HAVE_IOREMAP_PROT
6017 * generic_access_phys - generic implementation for iomem mmap access
6018 * @vma: the vma to access
6019 * @addr: userspace address, not relative offset within @vma
6020 * @buf: buffer to read/write
6021 * @len: length of transfer
6022 * @write: set to FOLL_WRITE when writing, otherwise reading
6024 * This is a generic implementation for &vm_operations_struct.access for an
6025 * iomem mapping. This callback is used by access_process_vm() when the @vma is
6028 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
6029 void *buf, int len, int write)
6031 resource_size_t phys_addr;
6032 unsigned long prot = 0;
6033 void __iomem *maddr;
6036 int offset = offset_in_page(addr);
6040 if (follow_pte(vma, addr, &ptep, &ptl))
6042 pte = ptep_get(ptep);
6043 pte_unmap_unlock(ptep, ptl);
6045 prot = pgprot_val(pte_pgprot(pte));
6046 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
6048 if ((write & FOLL_WRITE) && !pte_write(pte))
6051 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
6055 if (follow_pte(vma, addr, &ptep, &ptl))
6058 if (!pte_same(pte, ptep_get(ptep))) {
6059 pte_unmap_unlock(ptep, ptl);
6066 memcpy_toio(maddr + offset, buf, len);
6068 memcpy_fromio(buf, maddr + offset, len);
6070 pte_unmap_unlock(ptep, ptl);
6076 EXPORT_SYMBOL_GPL(generic_access_phys);
6080 * Access another process' address space as given in mm.
6082 static int __access_remote_vm(struct mm_struct *mm, unsigned long addr,
6083 void *buf, int len, unsigned int gup_flags)
6085 void *old_buf = buf;
6086 int write = gup_flags & FOLL_WRITE;
6088 if (mmap_read_lock_killable(mm))
6091 /* Untag the address before looking up the VMA */
6092 addr = untagged_addr_remote(mm, addr);
6094 /* Avoid triggering the temporary warning in __get_user_pages */
6095 if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
6098 /* ignore errors, just check how much was successfully transferred */
6102 struct vm_area_struct *vma = NULL;
6103 struct page *page = get_user_page_vma_remote(mm, addr,
6107 /* We might need to expand the stack to access it */
6108 vma = vma_lookup(mm, addr);
6110 vma = expand_stack(mm, addr);
6112 /* mmap_lock was dropped on failure */
6114 return buf - old_buf;
6116 /* Try again if stack expansion worked */
6121 * Check if this is a VM_IO | VM_PFNMAP VMA, which
6122 * we can access using slightly different code.
6125 #ifdef CONFIG_HAVE_IOREMAP_PROT
6126 if (vma->vm_ops && vma->vm_ops->access)
6127 bytes = vma->vm_ops->access(vma, addr, buf,
6134 offset = addr & (PAGE_SIZE-1);
6135 if (bytes > PAGE_SIZE-offset)
6136 bytes = PAGE_SIZE-offset;
6138 maddr = kmap_local_page(page);
6140 copy_to_user_page(vma, page, addr,
6141 maddr + offset, buf, bytes);
6142 set_page_dirty_lock(page);
6144 copy_from_user_page(vma, page, addr,
6145 buf, maddr + offset, bytes);
6147 unmap_and_put_page(page, maddr);
6153 mmap_read_unlock(mm);
6155 return buf - old_buf;
6159 * access_remote_vm - access another process' address space
6160 * @mm: the mm_struct of the target address space
6161 * @addr: start address to access
6162 * @buf: source or destination buffer
6163 * @len: number of bytes to transfer
6164 * @gup_flags: flags modifying lookup behaviour
6166 * The caller must hold a reference on @mm.
6168 * Return: number of bytes copied from source to destination.
6170 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
6171 void *buf, int len, unsigned int gup_flags)
6173 return __access_remote_vm(mm, addr, buf, len, gup_flags);
6177 * Access another process' address space.
6178 * Source/target buffer must be kernel space,
6179 * Do not walk the page table directly, use get_user_pages
6181 int access_process_vm(struct task_struct *tsk, unsigned long addr,
6182 void *buf, int len, unsigned int gup_flags)
6184 struct mm_struct *mm;
6187 mm = get_task_mm(tsk);
6191 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
6197 EXPORT_SYMBOL_GPL(access_process_vm);
6200 * Print the name of a VMA.
6202 void print_vma_addr(char *prefix, unsigned long ip)
6204 struct mm_struct *mm = current->mm;
6205 struct vm_area_struct *vma;
6208 * we might be running from an atomic context so we cannot sleep
6210 if (!mmap_read_trylock(mm))
6213 vma = find_vma(mm, ip);
6214 if (vma && vma->vm_file) {
6215 struct file *f = vma->vm_file;
6216 char *buf = (char *)__get_free_page(GFP_NOWAIT);
6220 p = file_path(f, buf, PAGE_SIZE);
6223 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
6225 vma->vm_end - vma->vm_start);
6226 free_page((unsigned long)buf);
6229 mmap_read_unlock(mm);
6232 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6233 void __might_fault(const char *file, int line)
6235 if (pagefault_disabled())
6237 __might_sleep(file, line);
6238 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
6240 might_lock_read(¤t->mm->mmap_lock);
6243 EXPORT_SYMBOL(__might_fault);
6246 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
6248 * Process all subpages of the specified huge page with the specified
6249 * operation. The target subpage will be processed last to keep its
6252 static inline int process_huge_page(
6253 unsigned long addr_hint, unsigned int pages_per_huge_page,
6254 int (*process_subpage)(unsigned long addr, int idx, void *arg),
6257 int i, n, base, l, ret;
6258 unsigned long addr = addr_hint &
6259 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6261 /* Process target subpage last to keep its cache lines hot */
6263 n = (addr_hint - addr) / PAGE_SIZE;
6264 if (2 * n <= pages_per_huge_page) {
6265 /* If target subpage in first half of huge page */
6268 /* Process subpages at the end of huge page */
6269 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
6271 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6276 /* If target subpage in second half of huge page */
6277 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
6278 l = pages_per_huge_page - n;
6279 /* Process subpages at the begin of huge page */
6280 for (i = 0; i < base; i++) {
6282 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
6288 * Process remaining subpages in left-right-left-right pattern
6289 * towards the target subpage
6291 for (i = 0; i < l; i++) {
6292 int left_idx = base + i;
6293 int right_idx = base + 2 * l - 1 - i;
6296 ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
6300 ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
6307 static void clear_gigantic_page(struct page *page,
6309 unsigned int pages_per_huge_page)
6315 for (i = 0; i < pages_per_huge_page; i++) {
6316 p = nth_page(page, i);
6318 clear_user_highpage(p, addr + i * PAGE_SIZE);
6322 static int clear_subpage(unsigned long addr, int idx, void *arg)
6324 struct page *page = arg;
6326 clear_user_highpage(nth_page(page, idx), addr);
6330 void clear_huge_page(struct page *page,
6331 unsigned long addr_hint, unsigned int pages_per_huge_page)
6333 unsigned long addr = addr_hint &
6334 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6336 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
6337 clear_gigantic_page(page, addr, pages_per_huge_page);
6341 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
6344 static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
6346 struct vm_area_struct *vma,
6347 unsigned int pages_per_huge_page)
6350 struct page *dst_page;
6351 struct page *src_page;
6353 for (i = 0; i < pages_per_huge_page; i++) {
6354 dst_page = folio_page(dst, i);
6355 src_page = folio_page(src, i);
6358 if (copy_mc_user_highpage(dst_page, src_page,
6359 addr + i*PAGE_SIZE, vma)) {
6360 memory_failure_queue(page_to_pfn(src_page), 0);
6367 struct copy_subpage_arg {
6370 struct vm_area_struct *vma;
6373 static int copy_subpage(unsigned long addr, int idx, void *arg)
6375 struct copy_subpage_arg *copy_arg = arg;
6376 struct page *dst = nth_page(copy_arg->dst, idx);
6377 struct page *src = nth_page(copy_arg->src, idx);
6379 if (copy_mc_user_highpage(dst, src, addr, copy_arg->vma)) {
6380 memory_failure_queue(page_to_pfn(src), 0);
6386 int copy_user_large_folio(struct folio *dst, struct folio *src,
6387 unsigned long addr_hint, struct vm_area_struct *vma)
6389 unsigned int pages_per_huge_page = folio_nr_pages(dst);
6390 unsigned long addr = addr_hint &
6391 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6392 struct copy_subpage_arg arg = {
6398 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES))
6399 return copy_user_gigantic_page(dst, src, addr, vma,
6400 pages_per_huge_page);
6402 return process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
6405 long copy_folio_from_user(struct folio *dst_folio,
6406 const void __user *usr_src,
6407 bool allow_pagefault)
6410 unsigned long i, rc = 0;
6411 unsigned int nr_pages = folio_nr_pages(dst_folio);
6412 unsigned long ret_val = nr_pages * PAGE_SIZE;
6413 struct page *subpage;
6415 for (i = 0; i < nr_pages; i++) {
6416 subpage = folio_page(dst_folio, i);
6417 kaddr = kmap_local_page(subpage);
6418 if (!allow_pagefault)
6419 pagefault_disable();
6420 rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE);
6421 if (!allow_pagefault)
6423 kunmap_local(kaddr);
6425 ret_val -= (PAGE_SIZE - rc);
6429 flush_dcache_page(subpage);
6435 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6437 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6439 static struct kmem_cache *page_ptl_cachep;
6441 void __init ptlock_cache_init(void)
6443 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
6447 bool ptlock_alloc(struct ptdesc *ptdesc)
6451 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
6458 void ptlock_free(struct ptdesc *ptdesc)
6460 kmem_cache_free(page_ptl_cachep, ptdesc->ptl);
6464 void vma_pgtable_walk_begin(struct vm_area_struct *vma)
6466 if (is_vm_hugetlb_page(vma))
6467 hugetlb_vma_lock_read(vma);
6470 void vma_pgtable_walk_end(struct vm_area_struct *vma)
6472 if (is_vm_hugetlb_page(vma))
6473 hugetlb_vma_unlock_read(vma);