1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/mm_inline.h>
45 #include <linux/sched/mm.h>
46 #include <linux/sched/coredump.h>
47 #include <linux/sched/numa_balancing.h>
48 #include <linux/sched/task.h>
49 #include <linux/hugetlb.h>
50 #include <linux/mman.h>
51 #include <linux/swap.h>
52 #include <linux/highmem.h>
53 #include <linux/pagemap.h>
54 #include <linux/memremap.h>
55 #include <linux/kmsan.h>
56 #include <linux/ksm.h>
57 #include <linux/rmap.h>
58 #include <linux/export.h>
59 #include <linux/delayacct.h>
60 #include <linux/init.h>
61 #include <linux/pfn_t.h>
62 #include <linux/writeback.h>
63 #include <linux/memcontrol.h>
64 #include <linux/mmu_notifier.h>
65 #include <linux/swapops.h>
66 #include <linux/elf.h>
67 #include <linux/gfp.h>
68 #include <linux/migrate.h>
69 #include <linux/string.h>
70 #include <linux/memory-tiers.h>
71 #include <linux/debugfs.h>
72 #include <linux/userfaultfd_k.h>
73 #include <linux/dax.h>
74 #include <linux/oom.h>
75 #include <linux/numa.h>
76 #include <linux/perf_event.h>
77 #include <linux/ptrace.h>
78 #include <linux/vmalloc.h>
79 #include <linux/sched/sysctl.h>
80 #include <linux/net_mm.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 bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
117 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
120 return pte_marker_uffd_wp(vmf->orig_pte);
124 * A number of key systems in x86 including ioremap() rely on the assumption
125 * that high_memory defines the upper bound on direct map memory, then end
126 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
127 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
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 maple_tree *mt,
365 struct vm_area_struct *vma, unsigned long floor,
366 unsigned long ceiling, bool mm_wr_locked)
368 MA_STATE(mas, mt, vma->vm_end, vma->vm_end);
371 unsigned long addr = vma->vm_start;
372 struct vm_area_struct *next;
375 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
376 * be 0. This will underflow and is okay.
378 next = mas_find(&mas, ceiling - 1);
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);
401 vma_start_write(vma);
402 unlink_anon_vmas(vma);
403 unlink_file_vma(vma);
405 free_pgd_range(tlb, addr, vma->vm_end,
406 floor, next ? next->vm_start : ceiling);
412 void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
414 spinlock_t *ptl = pmd_lock(mm, pmd);
416 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
419 * Ensure all pte setup (eg. pte page lock and page clearing) are
420 * visible before the pte is made visible to other CPUs by being
421 * put into page tables.
423 * The other side of the story is the pointer chasing in the page
424 * table walking code (when walking the page table without locking;
425 * ie. most of the time). Fortunately, these data accesses consist
426 * of a chain of data-dependent loads, meaning most CPUs (alpha
427 * being the notable exception) will already guarantee loads are
428 * seen in-order. See the alpha page table accessors for the
429 * smp_rmb() barriers in page table walking code.
431 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
432 pmd_populate(mm, pmd, *pte);
438 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
440 pgtable_t new = pte_alloc_one(mm);
444 pmd_install(mm, pmd, &new);
450 int __pte_alloc_kernel(pmd_t *pmd)
452 pte_t *new = pte_alloc_one_kernel(&init_mm);
456 spin_lock(&init_mm.page_table_lock);
457 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
458 smp_wmb(); /* See comment in pmd_install() */
459 pmd_populate_kernel(&init_mm, pmd, new);
462 spin_unlock(&init_mm.page_table_lock);
464 pte_free_kernel(&init_mm, new);
468 static inline void init_rss_vec(int *rss)
470 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
473 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
477 if (current->mm == mm)
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);
699 static void restore_exclusive_pte(struct vm_area_struct *vma,
700 struct page *page, unsigned long address,
707 orig_pte = ptep_get(ptep);
708 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
709 if (pte_swp_soft_dirty(orig_pte))
710 pte = pte_mksoft_dirty(pte);
712 entry = pte_to_swp_entry(orig_pte);
713 if (pte_swp_uffd_wp(orig_pte))
714 pte = pte_mkuffd_wp(pte);
715 else if (is_writable_device_exclusive_entry(entry))
716 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
718 VM_BUG_ON(pte_write(pte) && !(PageAnon(page) && PageAnonExclusive(page)));
721 * No need to take a page reference as one was already
722 * created when the swap entry was made.
725 page_add_anon_rmap(page, vma, address, RMAP_NONE);
728 * Currently device exclusive access only supports anonymous
729 * memory so the entry shouldn't point to a filebacked page.
733 set_pte_at(vma->vm_mm, address, ptep, pte);
736 * No need to invalidate - it was non-present before. However
737 * secondary CPUs may have mappings that need invalidating.
739 update_mmu_cache(vma, address, ptep);
743 * Tries to restore an exclusive pte if the page lock can be acquired without
747 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
750 swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte));
751 struct page *page = pfn_swap_entry_to_page(entry);
753 if (trylock_page(page)) {
754 restore_exclusive_pte(vma, page, addr, src_pte);
763 * copy one vm_area from one task to the other. Assumes the page tables
764 * already present in the new task to be cleared in the whole range
765 * covered by this vma.
769 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
770 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
771 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
773 unsigned long vm_flags = dst_vma->vm_flags;
774 pte_t orig_pte = ptep_get(src_pte);
775 pte_t pte = orig_pte;
777 swp_entry_t entry = pte_to_swp_entry(orig_pte);
779 if (likely(!non_swap_entry(entry))) {
780 if (swap_duplicate(entry) < 0)
783 /* make sure dst_mm is on swapoff's mmlist. */
784 if (unlikely(list_empty(&dst_mm->mmlist))) {
785 spin_lock(&mmlist_lock);
786 if (list_empty(&dst_mm->mmlist))
787 list_add(&dst_mm->mmlist,
789 spin_unlock(&mmlist_lock);
791 /* Mark the swap entry as shared. */
792 if (pte_swp_exclusive(orig_pte)) {
793 pte = pte_swp_clear_exclusive(orig_pte);
794 set_pte_at(src_mm, addr, src_pte, pte);
797 } else if (is_migration_entry(entry)) {
798 page = pfn_swap_entry_to_page(entry);
800 rss[mm_counter(page)]++;
802 if (!is_readable_migration_entry(entry) &&
803 is_cow_mapping(vm_flags)) {
805 * COW mappings require pages in both parent and child
806 * to be set to read. A previously exclusive entry is
809 entry = make_readable_migration_entry(
811 pte = swp_entry_to_pte(entry);
812 if (pte_swp_soft_dirty(orig_pte))
813 pte = pte_swp_mksoft_dirty(pte);
814 if (pte_swp_uffd_wp(orig_pte))
815 pte = pte_swp_mkuffd_wp(pte);
816 set_pte_at(src_mm, addr, src_pte, pte);
818 } else if (is_device_private_entry(entry)) {
819 page = pfn_swap_entry_to_page(entry);
822 * Update rss count even for unaddressable pages, as
823 * they should treated just like normal pages in this
826 * We will likely want to have some new rss counters
827 * for unaddressable pages, at some point. But for now
828 * keep things as they are.
831 rss[mm_counter(page)]++;
832 /* Cannot fail as these pages cannot get pinned. */
833 BUG_ON(page_try_dup_anon_rmap(page, false, src_vma));
836 * We do not preserve soft-dirty information, because so
837 * far, checkpoint/restore is the only feature that
838 * requires that. And checkpoint/restore does not work
839 * when a device driver is involved (you cannot easily
840 * save and restore device driver state).
842 if (is_writable_device_private_entry(entry) &&
843 is_cow_mapping(vm_flags)) {
844 entry = make_readable_device_private_entry(
846 pte = swp_entry_to_pte(entry);
847 if (pte_swp_uffd_wp(orig_pte))
848 pte = pte_swp_mkuffd_wp(pte);
849 set_pte_at(src_mm, addr, src_pte, pte);
851 } else if (is_device_exclusive_entry(entry)) {
853 * Make device exclusive entries present by restoring the
854 * original entry then copying as for a present pte. Device
855 * exclusive entries currently only support private writable
856 * (ie. COW) mappings.
858 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
859 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
862 } else if (is_pte_marker_entry(entry)) {
863 pte_marker marker = copy_pte_marker(entry, dst_vma);
866 set_pte_at(dst_mm, addr, dst_pte,
867 make_pte_marker(marker));
870 if (!userfaultfd_wp(dst_vma))
871 pte = pte_swp_clear_uffd_wp(pte);
872 set_pte_at(dst_mm, addr, dst_pte, pte);
877 * Copy a present and normal page.
879 * NOTE! The usual case is that this isn't required;
880 * instead, the caller can just increase the page refcount
881 * and re-use the pte the traditional way.
883 * And if we need a pre-allocated page but don't yet have
884 * one, return a negative error to let the preallocation
885 * code know so that it can do so outside the page table
889 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
890 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
891 struct folio **prealloc, struct page *page)
893 struct folio *new_folio;
896 new_folio = *prealloc;
901 * We have a prealloc page, all good! Take it
902 * over and copy the page & arm it.
905 copy_user_highpage(&new_folio->page, page, addr, src_vma);
906 __folio_mark_uptodate(new_folio);
907 folio_add_new_anon_rmap(new_folio, dst_vma, addr);
908 folio_add_lru_vma(new_folio, dst_vma);
911 /* All done, just insert the new page copy in the child */
912 pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot);
913 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
914 if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte)))
915 /* Uffd-wp needs to be delivered to dest pte as well */
916 pte = pte_mkuffd_wp(pte);
917 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
922 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
923 * is required to copy this pte.
926 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
927 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
928 struct folio **prealloc)
930 struct mm_struct *src_mm = src_vma->vm_mm;
931 unsigned long vm_flags = src_vma->vm_flags;
932 pte_t pte = ptep_get(src_pte);
936 page = vm_normal_page(src_vma, addr, pte);
938 folio = page_folio(page);
939 if (page && folio_test_anon(folio)) {
941 * If this page may have been pinned by the parent process,
942 * copy the page immediately for the child so that we'll always
943 * guarantee the pinned page won't be randomly replaced in the
947 if (unlikely(page_try_dup_anon_rmap(page, false, src_vma))) {
948 /* Page may be pinned, we have to copy. */
950 return copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
951 addr, rss, prealloc, page);
956 page_dup_file_rmap(page, false);
957 rss[mm_counter_file(page)]++;
961 * If it's a COW mapping, write protect it both
962 * in the parent and the child
964 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
965 ptep_set_wrprotect(src_mm, addr, src_pte);
966 pte = pte_wrprotect(pte);
968 VM_BUG_ON(page && folio_test_anon(folio) && PageAnonExclusive(page));
971 * If it's a shared mapping, mark it clean in
974 if (vm_flags & VM_SHARED)
975 pte = pte_mkclean(pte);
976 pte = pte_mkold(pte);
978 if (!userfaultfd_wp(dst_vma))
979 pte = pte_clear_uffd_wp(pte);
981 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
985 static inline struct folio *page_copy_prealloc(struct mm_struct *src_mm,
986 struct vm_area_struct *vma, unsigned long addr)
988 struct folio *new_folio;
990 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, addr, false);
994 if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) {
995 folio_put(new_folio);
998 folio_throttle_swaprate(new_folio, GFP_KERNEL);
1004 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1005 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1008 struct mm_struct *dst_mm = dst_vma->vm_mm;
1009 struct mm_struct *src_mm = src_vma->vm_mm;
1010 pte_t *orig_src_pte, *orig_dst_pte;
1011 pte_t *src_pte, *dst_pte;
1013 spinlock_t *src_ptl, *dst_ptl;
1014 int progress, ret = 0;
1015 int rss[NR_MM_COUNTERS];
1016 swp_entry_t entry = (swp_entry_t){0};
1017 struct folio *prealloc = NULL;
1024 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1025 * error handling here, assume that exclusive mmap_lock on dst and src
1026 * protects anon from unexpected THP transitions; with shmem and file
1027 * protected by mmap_lock-less collapse skipping areas with anon_vma
1028 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1029 * can remove such assumptions later, but this is good enough for now.
1031 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1036 src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl);
1038 pte_unmap_unlock(dst_pte, dst_ptl);
1042 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1043 orig_src_pte = src_pte;
1044 orig_dst_pte = dst_pte;
1045 arch_enter_lazy_mmu_mode();
1049 * We are holding two locks at this point - either of them
1050 * could generate latencies in another task on another CPU.
1052 if (progress >= 32) {
1054 if (need_resched() ||
1055 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1058 ptent = ptep_get(src_pte);
1059 if (pte_none(ptent)) {
1063 if (unlikely(!pte_present(ptent))) {
1064 ret = copy_nonpresent_pte(dst_mm, src_mm,
1069 entry = pte_to_swp_entry(ptep_get(src_pte));
1071 } else if (ret == -EBUSY) {
1079 * Device exclusive entry restored, continue by copying
1080 * the now present pte.
1082 WARN_ON_ONCE(ret != -ENOENT);
1084 /* copy_present_pte() will clear `*prealloc' if consumed */
1085 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1086 addr, rss, &prealloc);
1088 * If we need a pre-allocated page for this pte, drop the
1089 * locks, allocate, and try again.
1091 if (unlikely(ret == -EAGAIN))
1093 if (unlikely(prealloc)) {
1095 * pre-alloc page cannot be reused by next time so as
1096 * to strictly follow mempolicy (e.g., alloc_page_vma()
1097 * will allocate page according to address). This
1098 * could only happen if one pinned pte changed.
1100 folio_put(prealloc);
1104 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1106 arch_leave_lazy_mmu_mode();
1107 pte_unmap_unlock(orig_src_pte, src_ptl);
1108 add_mm_rss_vec(dst_mm, rss);
1109 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1113 VM_WARN_ON_ONCE(!entry.val);
1114 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1119 } else if (ret == -EBUSY) {
1121 } else if (ret == -EAGAIN) {
1122 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1129 /* We've captured and resolved the error. Reset, try again. */
1135 if (unlikely(prealloc))
1136 folio_put(prealloc);
1141 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1142 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1145 struct mm_struct *dst_mm = dst_vma->vm_mm;
1146 struct mm_struct *src_mm = src_vma->vm_mm;
1147 pmd_t *src_pmd, *dst_pmd;
1150 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1153 src_pmd = pmd_offset(src_pud, addr);
1155 next = pmd_addr_end(addr, end);
1156 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1157 || pmd_devmap(*src_pmd)) {
1159 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1160 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1161 addr, dst_vma, src_vma);
1168 if (pmd_none_or_clear_bad(src_pmd))
1170 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1173 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1178 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1179 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1182 struct mm_struct *dst_mm = dst_vma->vm_mm;
1183 struct mm_struct *src_mm = src_vma->vm_mm;
1184 pud_t *src_pud, *dst_pud;
1187 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1190 src_pud = pud_offset(src_p4d, addr);
1192 next = pud_addr_end(addr, end);
1193 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1196 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1197 err = copy_huge_pud(dst_mm, src_mm,
1198 dst_pud, src_pud, addr, src_vma);
1205 if (pud_none_or_clear_bad(src_pud))
1207 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1210 } while (dst_pud++, src_pud++, addr = next, addr != end);
1215 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1216 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1219 struct mm_struct *dst_mm = dst_vma->vm_mm;
1220 p4d_t *src_p4d, *dst_p4d;
1223 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1226 src_p4d = p4d_offset(src_pgd, addr);
1228 next = p4d_addr_end(addr, end);
1229 if (p4d_none_or_clear_bad(src_p4d))
1231 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1234 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1239 * Return true if the vma needs to copy the pgtable during this fork(). Return
1240 * false when we can speed up fork() by allowing lazy page faults later until
1241 * when the child accesses the memory range.
1244 vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1247 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1248 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1249 * contains uffd-wp protection information, that's something we can't
1250 * retrieve from page cache, and skip copying will lose those info.
1252 if (userfaultfd_wp(dst_vma))
1255 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1258 if (src_vma->anon_vma)
1262 * Don't copy ptes where a page fault will fill them correctly. Fork
1263 * becomes much lighter when there are big shared or private readonly
1264 * mappings. The tradeoff is that copy_page_range is more efficient
1271 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1273 pgd_t *src_pgd, *dst_pgd;
1275 unsigned long addr = src_vma->vm_start;
1276 unsigned long end = src_vma->vm_end;
1277 struct mm_struct *dst_mm = dst_vma->vm_mm;
1278 struct mm_struct *src_mm = src_vma->vm_mm;
1279 struct mmu_notifier_range range;
1283 if (!vma_needs_copy(dst_vma, src_vma))
1286 if (is_vm_hugetlb_page(src_vma))
1287 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1289 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1291 * We do not free on error cases below as remove_vma
1292 * gets called on error from higher level routine
1294 ret = track_pfn_copy(src_vma);
1300 * We need to invalidate the secondary MMU mappings only when
1301 * there could be a permission downgrade on the ptes of the
1302 * parent mm. And a permission downgrade will only happen if
1303 * is_cow_mapping() returns true.
1305 is_cow = is_cow_mapping(src_vma->vm_flags);
1308 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1309 0, src_mm, addr, end);
1310 mmu_notifier_invalidate_range_start(&range);
1312 * Disabling preemption is not needed for the write side, as
1313 * the read side doesn't spin, but goes to the mmap_lock.
1315 * Use the raw variant of the seqcount_t write API to avoid
1316 * lockdep complaining about preemptibility.
1318 mmap_assert_write_locked(src_mm);
1319 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1323 dst_pgd = pgd_offset(dst_mm, addr);
1324 src_pgd = pgd_offset(src_mm, addr);
1326 next = pgd_addr_end(addr, end);
1327 if (pgd_none_or_clear_bad(src_pgd))
1329 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1331 untrack_pfn_clear(dst_vma);
1335 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1338 raw_write_seqcount_end(&src_mm->write_protect_seq);
1339 mmu_notifier_invalidate_range_end(&range);
1344 /* Whether we should zap all COWed (private) pages too */
1345 static inline bool should_zap_cows(struct zap_details *details)
1347 /* By default, zap all pages */
1351 /* Or, we zap COWed pages only if the caller wants to */
1352 return details->even_cows;
1355 /* Decides whether we should zap this page with the page pointer specified */
1356 static inline bool should_zap_page(struct zap_details *details, struct page *page)
1358 /* If we can make a decision without *page.. */
1359 if (should_zap_cows(details))
1362 /* E.g. the caller passes NULL for the case of a zero page */
1366 /* Otherwise we should only zap non-anon pages */
1367 return !PageAnon(page);
1370 static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1375 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1379 * This function makes sure that we'll replace the none pte with an uffd-wp
1380 * swap special pte marker when necessary. Must be with the pgtable lock held.
1383 zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1384 unsigned long addr, pte_t *pte,
1385 struct zap_details *details, pte_t pteval)
1387 /* Zap on anonymous always means dropping everything */
1388 if (vma_is_anonymous(vma))
1391 if (zap_drop_file_uffd_wp(details))
1394 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1397 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1398 struct vm_area_struct *vma, pmd_t *pmd,
1399 unsigned long addr, unsigned long end,
1400 struct zap_details *details)
1402 struct mm_struct *mm = tlb->mm;
1403 int force_flush = 0;
1404 int rss[NR_MM_COUNTERS];
1410 tlb_change_page_size(tlb, PAGE_SIZE);
1412 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1416 flush_tlb_batched_pending(mm);
1417 arch_enter_lazy_mmu_mode();
1419 pte_t ptent = ptep_get(pte);
1422 if (pte_none(ptent))
1428 if (pte_present(ptent)) {
1429 unsigned int delay_rmap;
1431 page = vm_normal_page(vma, addr, ptent);
1432 if (unlikely(!should_zap_page(details, page)))
1434 ptent = ptep_get_and_clear_full(mm, addr, pte,
1436 tlb_remove_tlb_entry(tlb, pte, addr);
1437 zap_install_uffd_wp_if_needed(vma, addr, pte, details,
1439 if (unlikely(!page)) {
1440 ksm_might_unmap_zero_page(mm, ptent);
1445 if (!PageAnon(page)) {
1446 if (pte_dirty(ptent)) {
1447 set_page_dirty(page);
1448 if (tlb_delay_rmap(tlb)) {
1453 if (pte_young(ptent) && likely(vma_has_recency(vma)))
1454 mark_page_accessed(page);
1456 rss[mm_counter(page)]--;
1458 page_remove_rmap(page, vma, false);
1459 if (unlikely(page_mapcount(page) < 0))
1460 print_bad_pte(vma, addr, ptent, page);
1462 if (unlikely(__tlb_remove_page(tlb, page, delay_rmap))) {
1470 entry = pte_to_swp_entry(ptent);
1471 if (is_device_private_entry(entry) ||
1472 is_device_exclusive_entry(entry)) {
1473 page = pfn_swap_entry_to_page(entry);
1474 if (unlikely(!should_zap_page(details, page)))
1477 * Both device private/exclusive mappings should only
1478 * work with anonymous page so far, so we don't need to
1479 * consider uffd-wp bit when zap. For more information,
1480 * see zap_install_uffd_wp_if_needed().
1482 WARN_ON_ONCE(!vma_is_anonymous(vma));
1483 rss[mm_counter(page)]--;
1484 if (is_device_private_entry(entry))
1485 page_remove_rmap(page, vma, false);
1487 } else if (!non_swap_entry(entry)) {
1488 /* Genuine swap entry, hence a private anon page */
1489 if (!should_zap_cows(details))
1492 if (unlikely(!free_swap_and_cache(entry)))
1493 print_bad_pte(vma, addr, ptent, NULL);
1494 } else if (is_migration_entry(entry)) {
1495 page = pfn_swap_entry_to_page(entry);
1496 if (!should_zap_page(details, page))
1498 rss[mm_counter(page)]--;
1499 } else if (pte_marker_entry_uffd_wp(entry)) {
1501 * For anon: always drop the marker; for file: only
1502 * drop the marker if explicitly requested.
1504 if (!vma_is_anonymous(vma) &&
1505 !zap_drop_file_uffd_wp(details))
1507 } else if (is_hwpoison_entry(entry) ||
1508 is_poisoned_swp_entry(entry)) {
1509 if (!should_zap_cows(details))
1512 /* We should have covered all the swap entry types */
1515 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1516 zap_install_uffd_wp_if_needed(vma, addr, pte, details, ptent);
1517 } while (pte++, addr += PAGE_SIZE, addr != end);
1519 add_mm_rss_vec(mm, rss);
1520 arch_leave_lazy_mmu_mode();
1522 /* Do the actual TLB flush before dropping ptl */
1524 tlb_flush_mmu_tlbonly(tlb);
1525 tlb_flush_rmaps(tlb, vma);
1527 pte_unmap_unlock(start_pte, ptl);
1530 * If we forced a TLB flush (either due to running out of
1531 * batch buffers or because we needed to flush dirty TLB
1532 * entries before releasing the ptl), free the batched
1533 * memory too. Come back again if we didn't do everything.
1541 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1542 struct vm_area_struct *vma, pud_t *pud,
1543 unsigned long addr, unsigned long end,
1544 struct zap_details *details)
1549 pmd = pmd_offset(pud, addr);
1551 next = pmd_addr_end(addr, end);
1552 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1553 if (next - addr != HPAGE_PMD_SIZE)
1554 __split_huge_pmd(vma, pmd, addr, false, NULL);
1555 else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1560 } else if (details && details->single_folio &&
1561 folio_test_pmd_mappable(details->single_folio) &&
1562 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1563 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1565 * Take and drop THP pmd lock so that we cannot return
1566 * prematurely, while zap_huge_pmd() has cleared *pmd,
1567 * but not yet decremented compound_mapcount().
1571 if (pmd_none(*pmd)) {
1575 addr = zap_pte_range(tlb, vma, pmd, addr, next, details);
1578 } while (pmd++, cond_resched(), addr != end);
1583 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1584 struct vm_area_struct *vma, p4d_t *p4d,
1585 unsigned long addr, unsigned long end,
1586 struct zap_details *details)
1591 pud = pud_offset(p4d, addr);
1593 next = pud_addr_end(addr, end);
1594 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1595 if (next - addr != HPAGE_PUD_SIZE) {
1596 mmap_assert_locked(tlb->mm);
1597 split_huge_pud(vma, pud, addr);
1598 } else if (zap_huge_pud(tlb, vma, pud, addr))
1602 if (pud_none_or_clear_bad(pud))
1604 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1607 } while (pud++, addr = next, addr != end);
1612 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1613 struct vm_area_struct *vma, pgd_t *pgd,
1614 unsigned long addr, unsigned long end,
1615 struct zap_details *details)
1620 p4d = p4d_offset(pgd, addr);
1622 next = p4d_addr_end(addr, end);
1623 if (p4d_none_or_clear_bad(p4d))
1625 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1626 } while (p4d++, addr = next, addr != end);
1631 void unmap_page_range(struct mmu_gather *tlb,
1632 struct vm_area_struct *vma,
1633 unsigned long addr, unsigned long end,
1634 struct zap_details *details)
1639 BUG_ON(addr >= end);
1640 tlb_start_vma(tlb, vma);
1641 pgd = pgd_offset(vma->vm_mm, addr);
1643 next = pgd_addr_end(addr, end);
1644 if (pgd_none_or_clear_bad(pgd))
1646 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1647 } while (pgd++, addr = next, addr != end);
1648 tlb_end_vma(tlb, vma);
1652 static void unmap_single_vma(struct mmu_gather *tlb,
1653 struct vm_area_struct *vma, unsigned long start_addr,
1654 unsigned long end_addr,
1655 struct zap_details *details, bool mm_wr_locked)
1657 unsigned long start = max(vma->vm_start, start_addr);
1660 if (start >= vma->vm_end)
1662 end = min(vma->vm_end, end_addr);
1663 if (end <= vma->vm_start)
1667 uprobe_munmap(vma, start, end);
1669 if (unlikely(vma->vm_flags & VM_PFNMAP))
1670 untrack_pfn(vma, 0, 0, mm_wr_locked);
1673 if (unlikely(is_vm_hugetlb_page(vma))) {
1675 * It is undesirable to test vma->vm_file as it
1676 * should be non-null for valid hugetlb area.
1677 * However, vm_file will be NULL in the error
1678 * cleanup path of mmap_region. When
1679 * hugetlbfs ->mmap method fails,
1680 * mmap_region() nullifies vma->vm_file
1681 * before calling this function to clean up.
1682 * Since no pte has actually been setup, it is
1683 * safe to do nothing in this case.
1686 zap_flags_t zap_flags = details ?
1687 details->zap_flags : 0;
1688 __unmap_hugepage_range_final(tlb, vma, start, end,
1692 unmap_page_range(tlb, vma, start, end, details);
1697 * unmap_vmas - unmap a range of memory covered by a list of vma's
1698 * @tlb: address of the caller's struct mmu_gather
1699 * @mt: the maple tree
1700 * @vma: the starting vma
1701 * @start_addr: virtual address at which to start unmapping
1702 * @end_addr: virtual address at which to end unmapping
1703 * @mm_wr_locked: lock flag
1705 * Unmap all pages in the vma list.
1707 * Only addresses between `start' and `end' will be unmapped.
1709 * The VMA list must be sorted in ascending virtual address order.
1711 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1712 * range after unmap_vmas() returns. So the only responsibility here is to
1713 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1714 * drops the lock and schedules.
1716 void unmap_vmas(struct mmu_gather *tlb, struct maple_tree *mt,
1717 struct vm_area_struct *vma, unsigned long start_addr,
1718 unsigned long end_addr, bool mm_wr_locked)
1720 struct mmu_notifier_range range;
1721 struct zap_details details = {
1722 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1723 /* Careful - we need to zap private pages too! */
1726 MA_STATE(mas, mt, vma->vm_end, vma->vm_end);
1728 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
1729 start_addr, end_addr);
1730 mmu_notifier_invalidate_range_start(&range);
1732 unmap_single_vma(tlb, vma, start_addr, end_addr, &details,
1734 } while ((vma = mas_find(&mas, end_addr - 1)) != NULL);
1735 mmu_notifier_invalidate_range_end(&range);
1739 * zap_page_range_single - remove user pages in a given range
1740 * @vma: vm_area_struct holding the applicable pages
1741 * @address: starting address of pages to zap
1742 * @size: number of bytes to zap
1743 * @details: details of shared cache invalidation
1745 * The range must fit into one VMA.
1747 void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1748 unsigned long size, struct zap_details *details)
1750 const unsigned long end = address + size;
1751 struct mmu_notifier_range range;
1752 struct mmu_gather tlb;
1755 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1757 if (is_vm_hugetlb_page(vma))
1758 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1760 tlb_gather_mmu(&tlb, vma->vm_mm);
1761 update_hiwater_rss(vma->vm_mm);
1762 mmu_notifier_invalidate_range_start(&range);
1764 * unmap 'address-end' not 'range.start-range.end' as range
1765 * could have been expanded for hugetlb pmd sharing.
1767 unmap_single_vma(&tlb, vma, address, end, details, false);
1768 mmu_notifier_invalidate_range_end(&range);
1769 tlb_finish_mmu(&tlb);
1773 * zap_vma_ptes - remove ptes mapping the vma
1774 * @vma: vm_area_struct holding ptes to be zapped
1775 * @address: starting address of pages to zap
1776 * @size: number of bytes to zap
1778 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1780 * The entire address range must be fully contained within the vma.
1783 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1786 if (!range_in_vma(vma, address, address + size) ||
1787 !(vma->vm_flags & VM_PFNMAP))
1790 zap_page_range_single(vma, address, size, NULL);
1792 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1794 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1801 pgd = pgd_offset(mm, addr);
1802 p4d = p4d_alloc(mm, pgd, addr);
1805 pud = pud_alloc(mm, p4d, addr);
1808 pmd = pmd_alloc(mm, pud, addr);
1812 VM_BUG_ON(pmd_trans_huge(*pmd));
1816 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1819 pmd_t *pmd = walk_to_pmd(mm, addr);
1823 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1826 static int validate_page_before_insert(struct page *page)
1828 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1830 flush_dcache_page(page);
1834 static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1835 unsigned long addr, struct page *page, pgprot_t prot)
1837 if (!pte_none(ptep_get(pte)))
1839 /* Ok, finally just insert the thing.. */
1841 inc_mm_counter(vma->vm_mm, mm_counter_file(page));
1842 page_add_file_rmap(page, vma, false);
1843 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
1848 * This is the old fallback for page remapping.
1850 * For historical reasons, it only allows reserved pages. Only
1851 * old drivers should use this, and they needed to mark their
1852 * pages reserved for the old functions anyway.
1854 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1855 struct page *page, pgprot_t prot)
1861 retval = validate_page_before_insert(page);
1865 pte = get_locked_pte(vma->vm_mm, addr, &ptl);
1868 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
1869 pte_unmap_unlock(pte, ptl);
1875 static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
1876 unsigned long addr, struct page *page, pgprot_t prot)
1880 if (!page_count(page))
1882 err = validate_page_before_insert(page);
1885 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
1888 /* insert_pages() amortizes the cost of spinlock operations
1889 * when inserting pages in a loop. Arch *must* define pte_index.
1891 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1892 struct page **pages, unsigned long *num, pgprot_t prot)
1895 pte_t *start_pte, *pte;
1896 spinlock_t *pte_lock;
1897 struct mm_struct *const mm = vma->vm_mm;
1898 unsigned long curr_page_idx = 0;
1899 unsigned long remaining_pages_total = *num;
1900 unsigned long pages_to_write_in_pmd;
1904 pmd = walk_to_pmd(mm, addr);
1908 pages_to_write_in_pmd = min_t(unsigned long,
1909 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1911 /* Allocate the PTE if necessary; takes PMD lock once only. */
1913 if (pte_alloc(mm, pmd))
1916 while (pages_to_write_in_pmd) {
1918 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1920 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1925 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1926 int err = insert_page_in_batch_locked(vma, pte,
1927 addr, pages[curr_page_idx], prot);
1928 if (unlikely(err)) {
1929 pte_unmap_unlock(start_pte, pte_lock);
1931 remaining_pages_total -= pte_idx;
1937 pte_unmap_unlock(start_pte, pte_lock);
1938 pages_to_write_in_pmd -= batch_size;
1939 remaining_pages_total -= batch_size;
1941 if (remaining_pages_total)
1945 *num = remaining_pages_total;
1948 #endif /* ifdef pte_index */
1951 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1952 * @vma: user vma to map to
1953 * @addr: target start user address of these pages
1954 * @pages: source kernel pages
1955 * @num: in: number of pages to map. out: number of pages that were *not*
1956 * mapped. (0 means all pages were successfully mapped).
1958 * Preferred over vm_insert_page() when inserting multiple pages.
1960 * In case of error, we may have mapped a subset of the provided
1961 * pages. It is the caller's responsibility to account for this case.
1963 * The same restrictions apply as in vm_insert_page().
1965 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1966 struct page **pages, unsigned long *num)
1969 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1971 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1973 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1974 BUG_ON(mmap_read_trylock(vma->vm_mm));
1975 BUG_ON(vma->vm_flags & VM_PFNMAP);
1976 vm_flags_set(vma, VM_MIXEDMAP);
1978 /* Defer page refcount checking till we're about to map that page. */
1979 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1981 unsigned long idx = 0, pgcount = *num;
1984 for (; idx < pgcount; ++idx) {
1985 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1989 *num = pgcount - idx;
1991 #endif /* ifdef pte_index */
1993 EXPORT_SYMBOL(vm_insert_pages);
1996 * vm_insert_page - insert single page into user vma
1997 * @vma: user vma to map to
1998 * @addr: target user address of this page
1999 * @page: source kernel page
2001 * This allows drivers to insert individual pages they've allocated
2004 * The page has to be a nice clean _individual_ kernel allocation.
2005 * If you allocate a compound page, you need to have marked it as
2006 * such (__GFP_COMP), or manually just split the page up yourself
2007 * (see split_page()).
2009 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2010 * took an arbitrary page protection parameter. This doesn't allow
2011 * that. Your vma protection will have to be set up correctly, which
2012 * means that if you want a shared writable mapping, you'd better
2013 * ask for a shared writable mapping!
2015 * The page does not need to be reserved.
2017 * Usually this function is called from f_op->mmap() handler
2018 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2019 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2020 * function from other places, for example from page-fault handler.
2022 * Return: %0 on success, negative error code otherwise.
2024 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2027 if (addr < vma->vm_start || addr >= vma->vm_end)
2029 if (!page_count(page))
2031 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2032 BUG_ON(mmap_read_trylock(vma->vm_mm));
2033 BUG_ON(vma->vm_flags & VM_PFNMAP);
2034 vm_flags_set(vma, VM_MIXEDMAP);
2036 return insert_page(vma, addr, page, vma->vm_page_prot);
2038 EXPORT_SYMBOL(vm_insert_page);
2041 * __vm_map_pages - maps range of kernel pages into user vma
2042 * @vma: user vma to map to
2043 * @pages: pointer to array of source kernel pages
2044 * @num: number of pages in page array
2045 * @offset: user's requested vm_pgoff
2047 * This allows drivers to map range of kernel pages into a user vma.
2049 * Return: 0 on success and error code otherwise.
2051 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2052 unsigned long num, unsigned long offset)
2054 unsigned long count = vma_pages(vma);
2055 unsigned long uaddr = vma->vm_start;
2058 /* Fail if the user requested offset is beyond the end of the object */
2062 /* Fail if the user requested size exceeds available object size */
2063 if (count > num - offset)
2066 for (i = 0; i < count; i++) {
2067 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2077 * vm_map_pages - maps range of kernel pages starts with non zero offset
2078 * @vma: user vma to map to
2079 * @pages: pointer to array of source kernel pages
2080 * @num: number of pages in page array
2082 * Maps an object consisting of @num pages, catering for the user's
2083 * requested vm_pgoff
2085 * If we fail to insert any page into the vma, the function will return
2086 * immediately leaving any previously inserted pages present. Callers
2087 * from the mmap handler may immediately return the error as their caller
2088 * will destroy the vma, removing any successfully inserted pages. Other
2089 * callers should make their own arrangements for calling unmap_region().
2091 * Context: Process context. Called by mmap handlers.
2092 * Return: 0 on success and error code otherwise.
2094 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2097 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2099 EXPORT_SYMBOL(vm_map_pages);
2102 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2103 * @vma: user vma to map to
2104 * @pages: pointer to array of source kernel pages
2105 * @num: number of pages in page array
2107 * Similar to vm_map_pages(), except that it explicitly sets the offset
2108 * to 0. This function is intended for the drivers that did not consider
2111 * Context: Process context. Called by mmap handlers.
2112 * Return: 0 on success and error code otherwise.
2114 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2117 return __vm_map_pages(vma, pages, num, 0);
2119 EXPORT_SYMBOL(vm_map_pages_zero);
2121 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2122 pfn_t pfn, pgprot_t prot, bool mkwrite)
2124 struct mm_struct *mm = vma->vm_mm;
2128 pte = get_locked_pte(mm, addr, &ptl);
2130 return VM_FAULT_OOM;
2131 entry = ptep_get(pte);
2132 if (!pte_none(entry)) {
2135 * For read faults on private mappings the PFN passed
2136 * in may not match the PFN we have mapped if the
2137 * mapped PFN is a writeable COW page. In the mkwrite
2138 * case we are creating a writable PTE for a shared
2139 * mapping and we expect the PFNs to match. If they
2140 * don't match, we are likely racing with block
2141 * allocation and mapping invalidation so just skip the
2144 if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) {
2145 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2148 entry = pte_mkyoung(entry);
2149 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2150 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2151 update_mmu_cache(vma, addr, pte);
2156 /* Ok, finally just insert the thing.. */
2157 if (pfn_t_devmap(pfn))
2158 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2160 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2163 entry = pte_mkyoung(entry);
2164 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2167 set_pte_at(mm, addr, pte, entry);
2168 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2171 pte_unmap_unlock(pte, ptl);
2172 return VM_FAULT_NOPAGE;
2176 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2177 * @vma: user vma to map to
2178 * @addr: target user address of this page
2179 * @pfn: source kernel pfn
2180 * @pgprot: pgprot flags for the inserted page
2182 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2183 * to override pgprot on a per-page basis.
2185 * This only makes sense for IO mappings, and it makes no sense for
2186 * COW mappings. In general, using multiple vmas is preferable;
2187 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2190 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2191 * caching- and encryption bits different than those of @vma->vm_page_prot,
2192 * because the caching- or encryption mode may not be known at mmap() time.
2194 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2195 * to set caching and encryption bits for those vmas (except for COW pages).
2196 * This is ensured by core vm only modifying these page table entries using
2197 * functions that don't touch caching- or encryption bits, using pte_modify()
2198 * if needed. (See for example mprotect()).
2200 * Also when new page-table entries are created, this is only done using the
2201 * fault() callback, and never using the value of vma->vm_page_prot,
2202 * except for page-table entries that point to anonymous pages as the result
2205 * Context: Process context. May allocate using %GFP_KERNEL.
2206 * Return: vm_fault_t value.
2208 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2209 unsigned long pfn, pgprot_t pgprot)
2212 * Technically, architectures with pte_special can avoid all these
2213 * restrictions (same for remap_pfn_range). However we would like
2214 * consistency in testing and feature parity among all, so we should
2215 * try to keep these invariants in place for everybody.
2217 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2218 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2219 (VM_PFNMAP|VM_MIXEDMAP));
2220 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2221 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2223 if (addr < vma->vm_start || addr >= vma->vm_end)
2224 return VM_FAULT_SIGBUS;
2226 if (!pfn_modify_allowed(pfn, pgprot))
2227 return VM_FAULT_SIGBUS;
2229 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2231 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2234 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2237 * vmf_insert_pfn - insert single pfn into user vma
2238 * @vma: user vma to map to
2239 * @addr: target user address of this page
2240 * @pfn: source kernel pfn
2242 * Similar to vm_insert_page, this allows drivers to insert individual pages
2243 * they've allocated into a user vma. Same comments apply.
2245 * This function should only be called from a vm_ops->fault handler, and
2246 * in that case the handler should return the result of this function.
2248 * vma cannot be a COW mapping.
2250 * As this is called only for pages that do not currently exist, we
2251 * do not need to flush old virtual caches or the TLB.
2253 * Context: Process context. May allocate using %GFP_KERNEL.
2254 * Return: vm_fault_t value.
2256 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2259 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2261 EXPORT_SYMBOL(vmf_insert_pfn);
2263 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2265 /* these checks mirror the abort conditions in vm_normal_page */
2266 if (vma->vm_flags & VM_MIXEDMAP)
2268 if (pfn_t_devmap(pfn))
2270 if (pfn_t_special(pfn))
2272 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2277 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2278 unsigned long addr, pfn_t pfn, bool mkwrite)
2280 pgprot_t pgprot = vma->vm_page_prot;
2283 BUG_ON(!vm_mixed_ok(vma, pfn));
2285 if (addr < vma->vm_start || addr >= vma->vm_end)
2286 return VM_FAULT_SIGBUS;
2288 track_pfn_insert(vma, &pgprot, pfn);
2290 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2291 return VM_FAULT_SIGBUS;
2294 * If we don't have pte special, then we have to use the pfn_valid()
2295 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2296 * refcount the page if pfn_valid is true (hence insert_page rather
2297 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2298 * without pte special, it would there be refcounted as a normal page.
2300 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2301 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2305 * At this point we are committed to insert_page()
2306 * regardless of whether the caller specified flags that
2307 * result in pfn_t_has_page() == false.
2309 page = pfn_to_page(pfn_t_to_pfn(pfn));
2310 err = insert_page(vma, addr, page, pgprot);
2312 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2316 return VM_FAULT_OOM;
2317 if (err < 0 && err != -EBUSY)
2318 return VM_FAULT_SIGBUS;
2320 return VM_FAULT_NOPAGE;
2323 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2326 return __vm_insert_mixed(vma, addr, pfn, false);
2328 EXPORT_SYMBOL(vmf_insert_mixed);
2331 * If the insertion of PTE failed because someone else already added a
2332 * different entry in the mean time, we treat that as success as we assume
2333 * the same entry was actually inserted.
2335 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2336 unsigned long addr, pfn_t pfn)
2338 return __vm_insert_mixed(vma, addr, pfn, true);
2340 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2343 * maps a range of physical memory into the requested pages. the old
2344 * mappings are removed. any references to nonexistent pages results
2345 * in null mappings (currently treated as "copy-on-access")
2347 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2348 unsigned long addr, unsigned long end,
2349 unsigned long pfn, pgprot_t prot)
2351 pte_t *pte, *mapped_pte;
2355 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2358 arch_enter_lazy_mmu_mode();
2360 BUG_ON(!pte_none(ptep_get(pte)));
2361 if (!pfn_modify_allowed(pfn, prot)) {
2365 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2367 } while (pte++, addr += PAGE_SIZE, addr != end);
2368 arch_leave_lazy_mmu_mode();
2369 pte_unmap_unlock(mapped_pte, ptl);
2373 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2374 unsigned long addr, unsigned long end,
2375 unsigned long pfn, pgprot_t prot)
2381 pfn -= addr >> PAGE_SHIFT;
2382 pmd = pmd_alloc(mm, pud, addr);
2385 VM_BUG_ON(pmd_trans_huge(*pmd));
2387 next = pmd_addr_end(addr, end);
2388 err = remap_pte_range(mm, pmd, addr, next,
2389 pfn + (addr >> PAGE_SHIFT), prot);
2392 } while (pmd++, addr = next, addr != end);
2396 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2397 unsigned long addr, unsigned long end,
2398 unsigned long pfn, pgprot_t prot)
2404 pfn -= addr >> PAGE_SHIFT;
2405 pud = pud_alloc(mm, p4d, addr);
2409 next = pud_addr_end(addr, end);
2410 err = remap_pmd_range(mm, pud, addr, next,
2411 pfn + (addr >> PAGE_SHIFT), prot);
2414 } while (pud++, addr = next, addr != end);
2418 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2419 unsigned long addr, unsigned long end,
2420 unsigned long pfn, pgprot_t prot)
2426 pfn -= addr >> PAGE_SHIFT;
2427 p4d = p4d_alloc(mm, pgd, addr);
2431 next = p4d_addr_end(addr, end);
2432 err = remap_pud_range(mm, p4d, addr, next,
2433 pfn + (addr >> PAGE_SHIFT), prot);
2436 } while (p4d++, addr = next, addr != end);
2441 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2442 * must have pre-validated the caching bits of the pgprot_t.
2444 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2445 unsigned long pfn, unsigned long size, pgprot_t prot)
2449 unsigned long end = addr + PAGE_ALIGN(size);
2450 struct mm_struct *mm = vma->vm_mm;
2453 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2457 * Physically remapped pages are special. Tell the
2458 * rest of the world about it:
2459 * VM_IO tells people not to look at these pages
2460 * (accesses can have side effects).
2461 * VM_PFNMAP tells the core MM that the base pages are just
2462 * raw PFN mappings, and do not have a "struct page" associated
2465 * Disable vma merging and expanding with mremap().
2467 * Omit vma from core dump, even when VM_IO turned off.
2469 * There's a horrible special case to handle copy-on-write
2470 * behaviour that some programs depend on. We mark the "original"
2471 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2472 * See vm_normal_page() for details.
2474 if (is_cow_mapping(vma->vm_flags)) {
2475 if (addr != vma->vm_start || end != vma->vm_end)
2477 vma->vm_pgoff = pfn;
2480 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
2482 BUG_ON(addr >= end);
2483 pfn -= addr >> PAGE_SHIFT;
2484 pgd = pgd_offset(mm, addr);
2485 flush_cache_range(vma, addr, end);
2487 next = pgd_addr_end(addr, end);
2488 err = remap_p4d_range(mm, pgd, addr, next,
2489 pfn + (addr >> PAGE_SHIFT), prot);
2492 } while (pgd++, addr = next, addr != end);
2498 * remap_pfn_range - remap kernel memory to userspace
2499 * @vma: user vma to map to
2500 * @addr: target page aligned user address to start at
2501 * @pfn: page frame number of kernel physical memory address
2502 * @size: size of mapping area
2503 * @prot: page protection flags for this mapping
2505 * Note: this is only safe if the mm semaphore is held when called.
2507 * Return: %0 on success, negative error code otherwise.
2509 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2510 unsigned long pfn, unsigned long size, pgprot_t prot)
2514 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2518 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2520 untrack_pfn(vma, pfn, PAGE_ALIGN(size), true);
2523 EXPORT_SYMBOL(remap_pfn_range);
2526 * vm_iomap_memory - remap memory to userspace
2527 * @vma: user vma to map to
2528 * @start: start of the physical memory to be mapped
2529 * @len: size of area
2531 * This is a simplified io_remap_pfn_range() for common driver use. The
2532 * driver just needs to give us the physical memory range to be mapped,
2533 * we'll figure out the rest from the vma information.
2535 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2536 * whatever write-combining details or similar.
2538 * Return: %0 on success, negative error code otherwise.
2540 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2542 unsigned long vm_len, pfn, pages;
2544 /* Check that the physical memory area passed in looks valid */
2545 if (start + len < start)
2548 * You *really* shouldn't map things that aren't page-aligned,
2549 * but we've historically allowed it because IO memory might
2550 * just have smaller alignment.
2552 len += start & ~PAGE_MASK;
2553 pfn = start >> PAGE_SHIFT;
2554 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2555 if (pfn + pages < pfn)
2558 /* We start the mapping 'vm_pgoff' pages into the area */
2559 if (vma->vm_pgoff > pages)
2561 pfn += vma->vm_pgoff;
2562 pages -= vma->vm_pgoff;
2564 /* Can we fit all of the mapping? */
2565 vm_len = vma->vm_end - vma->vm_start;
2566 if (vm_len >> PAGE_SHIFT > pages)
2569 /* Ok, let it rip */
2570 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2572 EXPORT_SYMBOL(vm_iomap_memory);
2574 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2575 unsigned long addr, unsigned long end,
2576 pte_fn_t fn, void *data, bool create,
2577 pgtbl_mod_mask *mask)
2579 pte_t *pte, *mapped_pte;
2584 mapped_pte = pte = (mm == &init_mm) ?
2585 pte_alloc_kernel_track(pmd, addr, mask) :
2586 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2590 mapped_pte = pte = (mm == &init_mm) ?
2591 pte_offset_kernel(pmd, addr) :
2592 pte_offset_map_lock(mm, pmd, addr, &ptl);
2597 arch_enter_lazy_mmu_mode();
2601 if (create || !pte_none(ptep_get(pte))) {
2602 err = fn(pte++, addr, data);
2606 } while (addr += PAGE_SIZE, addr != end);
2608 *mask |= PGTBL_PTE_MODIFIED;
2610 arch_leave_lazy_mmu_mode();
2613 pte_unmap_unlock(mapped_pte, ptl);
2617 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2618 unsigned long addr, unsigned long end,
2619 pte_fn_t fn, void *data, bool create,
2620 pgtbl_mod_mask *mask)
2626 BUG_ON(pud_huge(*pud));
2629 pmd = pmd_alloc_track(mm, pud, addr, mask);
2633 pmd = pmd_offset(pud, addr);
2636 next = pmd_addr_end(addr, end);
2637 if (pmd_none(*pmd) && !create)
2639 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2641 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2646 err = apply_to_pte_range(mm, pmd, addr, next,
2647 fn, data, create, mask);
2650 } while (pmd++, addr = next, addr != end);
2655 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2656 unsigned long addr, unsigned long end,
2657 pte_fn_t fn, void *data, bool create,
2658 pgtbl_mod_mask *mask)
2665 pud = pud_alloc_track(mm, p4d, addr, mask);
2669 pud = pud_offset(p4d, addr);
2672 next = pud_addr_end(addr, end);
2673 if (pud_none(*pud) && !create)
2675 if (WARN_ON_ONCE(pud_leaf(*pud)))
2677 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2682 err = apply_to_pmd_range(mm, pud, addr, next,
2683 fn, data, create, mask);
2686 } while (pud++, addr = next, addr != end);
2691 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2692 unsigned long addr, unsigned long end,
2693 pte_fn_t fn, void *data, bool create,
2694 pgtbl_mod_mask *mask)
2701 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2705 p4d = p4d_offset(pgd, addr);
2708 next = p4d_addr_end(addr, end);
2709 if (p4d_none(*p4d) && !create)
2711 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2713 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2718 err = apply_to_pud_range(mm, p4d, addr, next,
2719 fn, data, create, mask);
2722 } while (p4d++, addr = next, addr != end);
2727 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2728 unsigned long size, pte_fn_t fn,
2729 void *data, bool create)
2732 unsigned long start = addr, next;
2733 unsigned long end = addr + size;
2734 pgtbl_mod_mask mask = 0;
2737 if (WARN_ON(addr >= end))
2740 pgd = pgd_offset(mm, addr);
2742 next = pgd_addr_end(addr, end);
2743 if (pgd_none(*pgd) && !create)
2745 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2747 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2752 err = apply_to_p4d_range(mm, pgd, addr, next,
2753 fn, data, create, &mask);
2756 } while (pgd++, addr = next, addr != end);
2758 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2759 arch_sync_kernel_mappings(start, start + size);
2765 * Scan a region of virtual memory, filling in page tables as necessary
2766 * and calling a provided function on each leaf page table.
2768 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2769 unsigned long size, pte_fn_t fn, void *data)
2771 return __apply_to_page_range(mm, addr, size, fn, data, true);
2773 EXPORT_SYMBOL_GPL(apply_to_page_range);
2776 * Scan a region of virtual memory, calling a provided function on
2777 * each leaf page table where it exists.
2779 * Unlike apply_to_page_range, this does _not_ fill in page tables
2780 * where they are absent.
2782 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2783 unsigned long size, pte_fn_t fn, void *data)
2785 return __apply_to_page_range(mm, addr, size, fn, data, false);
2787 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2790 * handle_pte_fault chooses page fault handler according to an entry which was
2791 * read non-atomically. Before making any commitment, on those architectures
2792 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2793 * parts, do_swap_page must check under lock before unmapping the pte and
2794 * proceeding (but do_wp_page is only called after already making such a check;
2795 * and do_anonymous_page can safely check later on).
2797 static inline int pte_unmap_same(struct vm_fault *vmf)
2800 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2801 if (sizeof(pte_t) > sizeof(unsigned long)) {
2802 spin_lock(vmf->ptl);
2803 same = pte_same(ptep_get(vmf->pte), vmf->orig_pte);
2804 spin_unlock(vmf->ptl);
2807 pte_unmap(vmf->pte);
2814 * 0: copied succeeded
2815 * -EHWPOISON: copy failed due to hwpoison in source page
2816 * -EAGAIN: copied failed (some other reason)
2818 static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2819 struct vm_fault *vmf)
2824 struct vm_area_struct *vma = vmf->vma;
2825 struct mm_struct *mm = vma->vm_mm;
2826 unsigned long addr = vmf->address;
2829 if (copy_mc_user_highpage(dst, src, addr, vma)) {
2830 memory_failure_queue(page_to_pfn(src), 0);
2837 * If the source page was a PFN mapping, we don't have
2838 * a "struct page" for it. We do a best-effort copy by
2839 * just copying from the original user address. If that
2840 * fails, we just zero-fill it. Live with it.
2842 kaddr = kmap_atomic(dst);
2843 uaddr = (void __user *)(addr & PAGE_MASK);
2846 * On architectures with software "accessed" bits, we would
2847 * take a double page fault, so mark it accessed here.
2850 if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) {
2853 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2854 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2856 * Other thread has already handled the fault
2857 * and update local tlb only
2860 update_mmu_tlb(vma, addr, vmf->pte);
2865 entry = pte_mkyoung(vmf->orig_pte);
2866 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2867 update_mmu_cache(vma, addr, vmf->pte);
2871 * This really shouldn't fail, because the page is there
2872 * in the page tables. But it might just be unreadable,
2873 * in which case we just give up and fill the result with
2876 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2880 /* Re-validate under PTL if the page is still mapped */
2881 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2882 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2883 /* The PTE changed under us, update local tlb */
2885 update_mmu_tlb(vma, addr, vmf->pte);
2891 * The same page can be mapped back since last copy attempt.
2892 * Try to copy again under PTL.
2894 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2896 * Give a warn in case there can be some obscure
2909 pte_unmap_unlock(vmf->pte, vmf->ptl);
2910 kunmap_atomic(kaddr);
2911 flush_dcache_page(dst);
2916 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2918 struct file *vm_file = vma->vm_file;
2921 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2924 * Special mappings (e.g. VDSO) do not have any file so fake
2925 * a default GFP_KERNEL for them.
2931 * Notify the address space that the page is about to become writable so that
2932 * it can prohibit this or wait for the page to get into an appropriate state.
2934 * We do this without the lock held, so that it can sleep if it needs to.
2936 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
2939 unsigned int old_flags = vmf->flags;
2941 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2943 if (vmf->vma->vm_file &&
2944 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2945 return VM_FAULT_SIGBUS;
2947 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2948 /* Restore original flags so that caller is not surprised */
2949 vmf->flags = old_flags;
2950 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2952 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2954 if (!folio->mapping) {
2955 folio_unlock(folio);
2956 return 0; /* retry */
2958 ret |= VM_FAULT_LOCKED;
2960 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2965 * Handle dirtying of a page in shared file mapping on a write fault.
2967 * The function expects the page to be locked and unlocks it.
2969 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2971 struct vm_area_struct *vma = vmf->vma;
2972 struct address_space *mapping;
2973 struct folio *folio = page_folio(vmf->page);
2975 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2977 dirtied = folio_mark_dirty(folio);
2978 VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
2980 * Take a local copy of the address_space - folio.mapping may be zeroed
2981 * by truncate after folio_unlock(). The address_space itself remains
2982 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
2983 * release semantics to prevent the compiler from undoing this copying.
2985 mapping = folio_raw_mapping(folio);
2986 folio_unlock(folio);
2989 file_update_time(vma->vm_file);
2992 * Throttle page dirtying rate down to writeback speed.
2994 * mapping may be NULL here because some device drivers do not
2995 * set page.mapping but still dirty their pages
2997 * Drop the mmap_lock before waiting on IO, if we can. The file
2998 * is pinning the mapping, as per above.
3000 if ((dirtied || page_mkwrite) && mapping) {
3003 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
3004 balance_dirty_pages_ratelimited(mapping);
3007 return VM_FAULT_COMPLETED;
3015 * Handle write page faults for pages that can be reused in the current vma
3017 * This can happen either due to the mapping being with the VM_SHARED flag,
3018 * or due to us being the last reference standing to the page. In either
3019 * case, all we need to do here is to mark the page as writable and update
3020 * any related book-keeping.
3022 static inline void wp_page_reuse(struct vm_fault *vmf)
3023 __releases(vmf->ptl)
3025 struct vm_area_struct *vma = vmf->vma;
3026 struct page *page = vmf->page;
3029 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3030 VM_BUG_ON(page && PageAnon(page) && !PageAnonExclusive(page));
3033 * Clear the pages cpupid information as the existing
3034 * information potentially belongs to a now completely
3035 * unrelated process.
3038 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
3040 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3041 entry = pte_mkyoung(vmf->orig_pte);
3042 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3043 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
3044 update_mmu_cache(vma, vmf->address, vmf->pte);
3045 pte_unmap_unlock(vmf->pte, vmf->ptl);
3046 count_vm_event(PGREUSE);
3050 * Handle the case of a page which we actually need to copy to a new page,
3051 * either due to COW or unsharing.
3053 * Called with mmap_lock locked and the old page referenced, but
3054 * without the ptl held.
3056 * High level logic flow:
3058 * - Allocate a page, copy the content of the old page to the new one.
3059 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3060 * - Take the PTL. If the pte changed, bail out and release the allocated page
3061 * - If the pte is still the way we remember it, update the page table and all
3062 * relevant references. This includes dropping the reference the page-table
3063 * held to the old page, as well as updating the rmap.
3064 * - In any case, unlock the PTL and drop the reference we took to the old page.
3066 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3068 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3069 struct vm_area_struct *vma = vmf->vma;
3070 struct mm_struct *mm = vma->vm_mm;
3071 struct folio *old_folio = NULL;
3072 struct folio *new_folio = NULL;
3074 int page_copied = 0;
3075 struct mmu_notifier_range range;
3078 delayacct_wpcopy_start();
3081 old_folio = page_folio(vmf->page);
3082 if (unlikely(anon_vma_prepare(vma)))
3085 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
3086 new_folio = vma_alloc_zeroed_movable_folio(vma, vmf->address);
3090 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma,
3091 vmf->address, false);
3095 ret = __wp_page_copy_user(&new_folio->page, vmf->page, vmf);
3098 * COW failed, if the fault was solved by other,
3099 * it's fine. If not, userspace would re-fault on
3100 * the same address and we will handle the fault
3101 * from the second attempt.
3102 * The -EHWPOISON case will not be retried.
3104 folio_put(new_folio);
3106 folio_put(old_folio);
3108 delayacct_wpcopy_end();
3109 return ret == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3111 kmsan_copy_page_meta(&new_folio->page, vmf->page);
3114 if (mem_cgroup_charge(new_folio, mm, GFP_KERNEL))
3116 folio_throttle_swaprate(new_folio, GFP_KERNEL);
3118 __folio_mark_uptodate(new_folio);
3120 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
3121 vmf->address & PAGE_MASK,
3122 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3123 mmu_notifier_invalidate_range_start(&range);
3126 * Re-check the pte - we dropped the lock
3128 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3129 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3131 if (!folio_test_anon(old_folio)) {
3132 dec_mm_counter(mm, mm_counter_file(&old_folio->page));
3133 inc_mm_counter(mm, MM_ANONPAGES);
3136 ksm_might_unmap_zero_page(mm, vmf->orig_pte);
3137 inc_mm_counter(mm, MM_ANONPAGES);
3139 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3140 entry = mk_pte(&new_folio->page, vma->vm_page_prot);
3141 entry = pte_sw_mkyoung(entry);
3142 if (unlikely(unshare)) {
3143 if (pte_soft_dirty(vmf->orig_pte))
3144 entry = pte_mksoft_dirty(entry);
3145 if (pte_uffd_wp(vmf->orig_pte))
3146 entry = pte_mkuffd_wp(entry);
3148 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3152 * Clear the pte entry and flush it first, before updating the
3153 * pte with the new entry, to keep TLBs on different CPUs in
3154 * sync. This code used to set the new PTE then flush TLBs, but
3155 * that left a window where the new PTE could be loaded into
3156 * some TLBs while the old PTE remains in others.
3158 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3159 folio_add_new_anon_rmap(new_folio, vma, vmf->address);
3160 folio_add_lru_vma(new_folio, vma);
3162 * We call the notify macro here because, when using secondary
3163 * mmu page tables (such as kvm shadow page tables), we want the
3164 * new page to be mapped directly into the secondary page table.
3166 BUG_ON(unshare && pte_write(entry));
3167 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3168 update_mmu_cache(vma, vmf->address, vmf->pte);
3171 * Only after switching the pte to the new page may
3172 * we remove the mapcount here. Otherwise another
3173 * process may come and find the rmap count decremented
3174 * before the pte is switched to the new page, and
3175 * "reuse" the old page writing into it while our pte
3176 * here still points into it and can be read by other
3179 * The critical issue is to order this
3180 * page_remove_rmap with the ptp_clear_flush above.
3181 * Those stores are ordered by (if nothing else,)
3182 * the barrier present in the atomic_add_negative
3183 * in page_remove_rmap.
3185 * Then the TLB flush in ptep_clear_flush ensures that
3186 * no process can access the old page before the
3187 * decremented mapcount is visible. And the old page
3188 * cannot be reused until after the decremented
3189 * mapcount is visible. So transitively, TLBs to
3190 * old page will be flushed before it can be reused.
3192 page_remove_rmap(vmf->page, vma, false);
3195 /* Free the old page.. */
3196 new_folio = old_folio;
3198 pte_unmap_unlock(vmf->pte, vmf->ptl);
3199 } else if (vmf->pte) {
3200 update_mmu_tlb(vma, vmf->address, vmf->pte);
3201 pte_unmap_unlock(vmf->pte, vmf->ptl);
3205 * No need to double call mmu_notifier->invalidate_range() callback as
3206 * the above ptep_clear_flush_notify() did already call it.
3208 mmu_notifier_invalidate_range_only_end(&range);
3211 folio_put(new_folio);
3214 free_swap_cache(&old_folio->page);
3215 folio_put(old_folio);
3218 delayacct_wpcopy_end();
3221 folio_put(new_folio);
3224 folio_put(old_folio);
3226 delayacct_wpcopy_end();
3227 return VM_FAULT_OOM;
3231 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3232 * writeable once the page is prepared
3234 * @vmf: structure describing the fault
3236 * This function handles all that is needed to finish a write page fault in a
3237 * shared mapping due to PTE being read-only once the mapped page is prepared.
3238 * It handles locking of PTE and modifying it.
3240 * The function expects the page to be locked or other protection against
3241 * concurrent faults / writeback (such as DAX radix tree locks).
3243 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3244 * we acquired PTE lock.
3246 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3248 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3249 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3252 return VM_FAULT_NOPAGE;
3254 * We might have raced with another page fault while we released the
3255 * pte_offset_map_lock.
3257 if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) {
3258 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3259 pte_unmap_unlock(vmf->pte, vmf->ptl);
3260 return VM_FAULT_NOPAGE;
3267 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3270 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3272 struct vm_area_struct *vma = vmf->vma;
3274 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3277 pte_unmap_unlock(vmf->pte, vmf->ptl);
3278 vmf->flags |= FAULT_FLAG_MKWRITE;
3279 ret = vma->vm_ops->pfn_mkwrite(vmf);
3280 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3282 return finish_mkwrite_fault(vmf);
3288 static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3289 __releases(vmf->ptl)
3291 struct vm_area_struct *vma = vmf->vma;
3296 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3299 pte_unmap_unlock(vmf->pte, vmf->ptl);
3300 tmp = do_page_mkwrite(vmf, folio);
3301 if (unlikely(!tmp || (tmp &
3302 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3306 tmp = finish_mkwrite_fault(vmf);
3307 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3308 folio_unlock(folio);
3316 ret |= fault_dirty_shared_page(vmf);
3323 * This routine handles present pages, when
3324 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3325 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3326 * (FAULT_FLAG_UNSHARE)
3328 * It is done by copying the page to a new address and decrementing the
3329 * shared-page counter for the old page.
3331 * Note that this routine assumes that the protection checks have been
3332 * done by the caller (the low-level page fault routine in most cases).
3333 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3334 * done any necessary COW.
3336 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3337 * though the page will change only once the write actually happens. This
3338 * avoids a few races, and potentially makes it more efficient.
3340 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3341 * but allow concurrent faults), with pte both mapped and locked.
3342 * We return with mmap_lock still held, but pte unmapped and unlocked.
3344 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3345 __releases(vmf->ptl)
3347 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3348 struct vm_area_struct *vma = vmf->vma;
3349 struct folio *folio = NULL;
3351 if (likely(!unshare)) {
3352 if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) {
3353 pte_unmap_unlock(vmf->pte, vmf->ptl);
3354 return handle_userfault(vmf, VM_UFFD_WP);
3358 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3359 * is flushed in this case before copying.
3361 if (unlikely(userfaultfd_wp(vmf->vma) &&
3362 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3363 flush_tlb_page(vmf->vma, vmf->address);
3366 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3369 folio = page_folio(vmf->page);
3372 * Shared mapping: we are guaranteed to have VM_WRITE and
3373 * FAULT_FLAG_WRITE set at this point.
3375 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3377 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3380 * We should not cow pages in a shared writeable mapping.
3381 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3384 return wp_pfn_shared(vmf);
3385 return wp_page_shared(vmf, folio);
3389 * Private mapping: create an exclusive anonymous page copy if reuse
3390 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3392 if (folio && folio_test_anon(folio)) {
3394 * If the page is exclusive to this process we must reuse the
3395 * page without further checks.
3397 if (PageAnonExclusive(vmf->page))
3401 * We have to verify under folio lock: these early checks are
3402 * just an optimization to avoid locking the folio and freeing
3403 * the swapcache if there is little hope that we can reuse.
3405 * KSM doesn't necessarily raise the folio refcount.
3407 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3409 if (!folio_test_lru(folio))
3411 * We cannot easily detect+handle references from
3412 * remote LRU caches or references to LRU folios.
3415 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3417 if (!folio_trylock(folio))
3419 if (folio_test_swapcache(folio))
3420 folio_free_swap(folio);
3421 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3422 folio_unlock(folio);
3426 * Ok, we've got the only folio reference from our mapping
3427 * and the folio is locked, it's dark out, and we're wearing
3428 * sunglasses. Hit it.
3430 page_move_anon_rmap(vmf->page, vma);
3431 folio_unlock(folio);
3433 if (unlikely(unshare)) {
3434 pte_unmap_unlock(vmf->pte, vmf->ptl);
3442 * Ok, we need to copy. Oh, well..
3447 pte_unmap_unlock(vmf->pte, vmf->ptl);
3449 if (folio && folio_test_ksm(folio))
3450 count_vm_event(COW_KSM);
3452 return wp_page_copy(vmf);
3455 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3456 unsigned long start_addr, unsigned long end_addr,
3457 struct zap_details *details)
3459 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3462 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3463 pgoff_t first_index,
3465 struct zap_details *details)
3467 struct vm_area_struct *vma;
3468 pgoff_t vba, vea, zba, zea;
3470 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3471 vba = vma->vm_pgoff;
3472 vea = vba + vma_pages(vma) - 1;
3473 zba = max(first_index, vba);
3474 zea = min(last_index, vea);
3476 unmap_mapping_range_vma(vma,
3477 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3478 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3484 * unmap_mapping_folio() - Unmap single folio from processes.
3485 * @folio: The locked folio to be unmapped.
3487 * Unmap this folio from any userspace process which still has it mmaped.
3488 * Typically, for efficiency, the range of nearby pages has already been
3489 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3490 * truncation or invalidation holds the lock on a folio, it may find that
3491 * the page has been remapped again: and then uses unmap_mapping_folio()
3492 * to unmap it finally.
3494 void unmap_mapping_folio(struct folio *folio)
3496 struct address_space *mapping = folio->mapping;
3497 struct zap_details details = { };
3498 pgoff_t first_index;
3501 VM_BUG_ON(!folio_test_locked(folio));
3503 first_index = folio->index;
3504 last_index = folio_next_index(folio) - 1;
3506 details.even_cows = false;
3507 details.single_folio = folio;
3508 details.zap_flags = ZAP_FLAG_DROP_MARKER;
3510 i_mmap_lock_read(mapping);
3511 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3512 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3513 last_index, &details);
3514 i_mmap_unlock_read(mapping);
3518 * unmap_mapping_pages() - Unmap pages from processes.
3519 * @mapping: The address space containing pages to be unmapped.
3520 * @start: Index of first page to be unmapped.
3521 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3522 * @even_cows: Whether to unmap even private COWed pages.
3524 * Unmap the pages in this address space from any userspace process which
3525 * has them mmaped. Generally, you want to remove COWed pages as well when
3526 * a file is being truncated, but not when invalidating pages from the page
3529 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3530 pgoff_t nr, bool even_cows)
3532 struct zap_details details = { };
3533 pgoff_t first_index = start;
3534 pgoff_t last_index = start + nr - 1;
3536 details.even_cows = even_cows;
3537 if (last_index < first_index)
3538 last_index = ULONG_MAX;
3540 i_mmap_lock_read(mapping);
3541 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3542 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3543 last_index, &details);
3544 i_mmap_unlock_read(mapping);
3546 EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3549 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3550 * address_space corresponding to the specified byte range in the underlying
3553 * @mapping: the address space containing mmaps to be unmapped.
3554 * @holebegin: byte in first page to unmap, relative to the start of
3555 * the underlying file. This will be rounded down to a PAGE_SIZE
3556 * boundary. Note that this is different from truncate_pagecache(), which
3557 * must keep the partial page. In contrast, we must get rid of
3559 * @holelen: size of prospective hole in bytes. This will be rounded
3560 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3562 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3563 * but 0 when invalidating pagecache, don't throw away private data.
3565 void unmap_mapping_range(struct address_space *mapping,
3566 loff_t const holebegin, loff_t const holelen, int even_cows)
3568 pgoff_t hba = holebegin >> PAGE_SHIFT;
3569 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3571 /* Check for overflow. */
3572 if (sizeof(holelen) > sizeof(hlen)) {
3574 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3575 if (holeend & ~(long long)ULONG_MAX)
3576 hlen = ULONG_MAX - hba + 1;
3579 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3581 EXPORT_SYMBOL(unmap_mapping_range);
3584 * Restore a potential device exclusive pte to a working pte entry
3586 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3588 struct folio *folio = page_folio(vmf->page);
3589 struct vm_area_struct *vma = vmf->vma;
3590 struct mmu_notifier_range range;
3593 * We need a reference to lock the folio because we don't hold
3594 * the PTL so a racing thread can remove the device-exclusive
3595 * entry and unmap it. If the folio is free the entry must
3596 * have been removed already. If it happens to have already
3597 * been re-allocated after being freed all we do is lock and
3600 if (!folio_try_get(folio))
3603 if (!folio_lock_or_retry(folio, vma->vm_mm, vmf->flags)) {
3605 return VM_FAULT_RETRY;
3607 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
3608 vma->vm_mm, vmf->address & PAGE_MASK,
3609 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3610 mmu_notifier_invalidate_range_start(&range);
3612 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3614 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3615 restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte);
3618 pte_unmap_unlock(vmf->pte, vmf->ptl);
3619 folio_unlock(folio);
3622 mmu_notifier_invalidate_range_end(&range);
3626 static inline bool should_try_to_free_swap(struct folio *folio,
3627 struct vm_area_struct *vma,
3628 unsigned int fault_flags)
3630 if (!folio_test_swapcache(folio))
3632 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3633 folio_test_mlocked(folio))
3636 * If we want to map a page that's in the swapcache writable, we
3637 * have to detect via the refcount if we're really the exclusive
3638 * user. Try freeing the swapcache to get rid of the swapcache
3639 * reference only in case it's likely that we'll be the exlusive user.
3641 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3642 folio_ref_count(folio) == 2;
3645 static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3647 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
3648 vmf->address, &vmf->ptl);
3652 * Be careful so that we will only recover a special uffd-wp pte into a
3653 * none pte. Otherwise it means the pte could have changed, so retry.
3655 * This should also cover the case where e.g. the pte changed
3656 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3657 * So is_pte_marker() check is not enough to safely drop the pte.
3659 if (pte_same(vmf->orig_pte, ptep_get(vmf->pte)))
3660 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
3661 pte_unmap_unlock(vmf->pte, vmf->ptl);
3665 static vm_fault_t do_pte_missing(struct vm_fault *vmf)
3667 if (vma_is_anonymous(vmf->vma))
3668 return do_anonymous_page(vmf);
3670 return do_fault(vmf);
3674 * This is actually a page-missing access, but with uffd-wp special pte
3675 * installed. It means this pte was wr-protected before being unmapped.
3677 static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3680 * Just in case there're leftover special ptes even after the region
3681 * got unregistered - we can simply clear them.
3683 if (unlikely(!userfaultfd_wp(vmf->vma)))
3684 return pte_marker_clear(vmf);
3686 return do_pte_missing(vmf);
3689 static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3691 swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
3692 unsigned long marker = pte_marker_get(entry);
3695 * PTE markers should never be empty. If anything weird happened,
3696 * the best thing to do is to kill the process along with its mm.
3698 if (WARN_ON_ONCE(!marker))
3699 return VM_FAULT_SIGBUS;
3701 /* Higher priority than uffd-wp when data corrupted */
3702 if (marker & PTE_MARKER_POISONED)
3703 return VM_FAULT_HWPOISON;
3705 if (pte_marker_entry_uffd_wp(entry))
3706 return pte_marker_handle_uffd_wp(vmf);
3708 /* This is an unknown pte marker */
3709 return VM_FAULT_SIGBUS;
3713 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3714 * but allow concurrent faults), and pte mapped but not yet locked.
3715 * We return with pte unmapped and unlocked.
3717 * We return with the mmap_lock locked or unlocked in the same cases
3718 * as does filemap_fault().
3720 vm_fault_t do_swap_page(struct vm_fault *vmf)
3722 struct vm_area_struct *vma = vmf->vma;
3723 struct folio *swapcache, *folio = NULL;
3725 struct swap_info_struct *si = NULL;
3726 rmap_t rmap_flags = RMAP_NONE;
3727 bool exclusive = false;
3732 void *shadow = NULL;
3734 if (!pte_unmap_same(vmf))
3737 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3738 ret = VM_FAULT_RETRY;
3742 entry = pte_to_swp_entry(vmf->orig_pte);
3743 if (unlikely(non_swap_entry(entry))) {
3744 if (is_migration_entry(entry)) {
3745 migration_entry_wait(vma->vm_mm, vmf->pmd,
3747 } else if (is_device_exclusive_entry(entry)) {
3748 vmf->page = pfn_swap_entry_to_page(entry);
3749 ret = remove_device_exclusive_entry(vmf);
3750 } else if (is_device_private_entry(entry)) {
3751 vmf->page = pfn_swap_entry_to_page(entry);
3752 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3753 vmf->address, &vmf->ptl);
3754 if (unlikely(!vmf->pte ||
3755 !pte_same(ptep_get(vmf->pte),
3760 * Get a page reference while we know the page can't be
3763 get_page(vmf->page);
3764 pte_unmap_unlock(vmf->pte, vmf->ptl);
3765 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3766 put_page(vmf->page);
3767 } else if (is_hwpoison_entry(entry)) {
3768 ret = VM_FAULT_HWPOISON;
3769 } else if (is_pte_marker_entry(entry)) {
3770 ret = handle_pte_marker(vmf);
3772 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3773 ret = VM_FAULT_SIGBUS;
3778 /* Prevent swapoff from happening to us. */
3779 si = get_swap_device(entry);
3783 folio = swap_cache_get_folio(entry, vma, vmf->address);
3785 page = folio_file_page(folio, swp_offset(entry));
3789 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3790 __swap_count(entry) == 1) {
3791 /* skip swapcache */
3792 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0,
3793 vma, vmf->address, false);
3794 page = &folio->page;
3796 __folio_set_locked(folio);
3797 __folio_set_swapbacked(folio);
3799 if (mem_cgroup_swapin_charge_folio(folio,
3800 vma->vm_mm, GFP_KERNEL,
3805 mem_cgroup_swapin_uncharge_swap(entry);
3807 shadow = get_shadow_from_swap_cache(entry);
3809 workingset_refault(folio, shadow);
3811 folio_add_lru(folio);
3813 /* To provide entry to swap_readpage() */
3814 folio_set_swap_entry(folio, entry);
3815 swap_readpage(page, true, NULL);
3816 folio->private = NULL;
3819 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3822 folio = page_folio(page);
3828 * Back out if somebody else faulted in this pte
3829 * while we released the pte lock.
3831 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3832 vmf->address, &vmf->ptl);
3833 if (likely(vmf->pte &&
3834 pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3839 /* Had to read the page from swap area: Major fault */
3840 ret = VM_FAULT_MAJOR;
3841 count_vm_event(PGMAJFAULT);
3842 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3843 } else if (PageHWPoison(page)) {
3845 * hwpoisoned dirty swapcache pages are kept for killing
3846 * owner processes (which may be unknown at hwpoison time)
3848 ret = VM_FAULT_HWPOISON;
3852 locked = folio_lock_or_retry(folio, vma->vm_mm, vmf->flags);
3855 ret |= VM_FAULT_RETRY;
3861 * Make sure folio_free_swap() or swapoff did not release the
3862 * swapcache from under us. The page pin, and pte_same test
3863 * below, are not enough to exclude that. Even if it is still
3864 * swapcache, we need to check that the page's swap has not
3867 if (unlikely(!folio_test_swapcache(folio) ||
3868 page_private(page) != entry.val))
3872 * KSM sometimes has to copy on read faults, for example, if
3873 * page->index of !PageKSM() pages would be nonlinear inside the
3874 * anon VMA -- PageKSM() is lost on actual swapout.
3876 page = ksm_might_need_to_copy(page, vma, vmf->address);
3877 if (unlikely(!page)) {
3880 } else if (unlikely(PTR_ERR(page) == -EHWPOISON)) {
3881 ret = VM_FAULT_HWPOISON;
3884 folio = page_folio(page);
3887 * If we want to map a page that's in the swapcache writable, we
3888 * have to detect via the refcount if we're really the exclusive
3889 * owner. Try removing the extra reference from the local LRU
3890 * caches if required.
3892 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
3893 !folio_test_ksm(folio) && !folio_test_lru(folio))
3897 folio_throttle_swaprate(folio, GFP_KERNEL);
3900 * Back out if somebody else already faulted in this pte.
3902 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3904 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3907 if (unlikely(!folio_test_uptodate(folio))) {
3908 ret = VM_FAULT_SIGBUS;
3913 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
3914 * must never point at an anonymous page in the swapcache that is
3915 * PG_anon_exclusive. Sanity check that this holds and especially, that
3916 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
3917 * check after taking the PT lock and making sure that nobody
3918 * concurrently faulted in this page and set PG_anon_exclusive.
3920 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
3921 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
3924 * Check under PT lock (to protect against concurrent fork() sharing
3925 * the swap entry concurrently) for certainly exclusive pages.
3927 if (!folio_test_ksm(folio)) {
3928 exclusive = pte_swp_exclusive(vmf->orig_pte);
3929 if (folio != swapcache) {
3931 * We have a fresh page that is not exposed to the
3932 * swapcache -> certainly exclusive.
3935 } else if (exclusive && folio_test_writeback(folio) &&
3936 data_race(si->flags & SWP_STABLE_WRITES)) {
3938 * This is tricky: not all swap backends support
3939 * concurrent page modifications while under writeback.
3941 * So if we stumble over such a page in the swapcache
3942 * we must not set the page exclusive, otherwise we can
3943 * map it writable without further checks and modify it
3944 * while still under writeback.
3946 * For these problematic swap backends, simply drop the
3947 * exclusive marker: this is perfectly fine as we start
3948 * writeback only if we fully unmapped the page and
3949 * there are no unexpected references on the page after
3950 * unmapping succeeded. After fully unmapped, no
3951 * further GUP references (FOLL_GET and FOLL_PIN) can
3952 * appear, so dropping the exclusive marker and mapping
3953 * it only R/O is fine.
3960 * Some architectures may have to restore extra metadata to the page
3961 * when reading from swap. This metadata may be indexed by swap entry
3962 * so this must be called before swap_free().
3964 arch_swap_restore(entry, folio);
3967 * Remove the swap entry and conditionally try to free up the swapcache.
3968 * We're already holding a reference on the page but haven't mapped it
3972 if (should_try_to_free_swap(folio, vma, vmf->flags))
3973 folio_free_swap(folio);
3975 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
3976 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
3977 pte = mk_pte(page, vma->vm_page_prot);
3980 * Same logic as in do_wp_page(); however, optimize for pages that are
3981 * certainly not shared either because we just allocated them without
3982 * exposing them to the swapcache or because the swap entry indicates
3985 if (!folio_test_ksm(folio) &&
3986 (exclusive || folio_ref_count(folio) == 1)) {
3987 if (vmf->flags & FAULT_FLAG_WRITE) {
3988 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3989 vmf->flags &= ~FAULT_FLAG_WRITE;
3991 rmap_flags |= RMAP_EXCLUSIVE;
3993 flush_icache_page(vma, page);
3994 if (pte_swp_soft_dirty(vmf->orig_pte))
3995 pte = pte_mksoft_dirty(pte);
3996 if (pte_swp_uffd_wp(vmf->orig_pte))
3997 pte = pte_mkuffd_wp(pte);
3998 vmf->orig_pte = pte;
4000 /* ksm created a completely new copy */
4001 if (unlikely(folio != swapcache && swapcache)) {
4002 page_add_new_anon_rmap(page, vma, vmf->address);
4003 folio_add_lru_vma(folio, vma);
4005 page_add_anon_rmap(page, vma, vmf->address, rmap_flags);
4008 VM_BUG_ON(!folio_test_anon(folio) ||
4009 (pte_write(pte) && !PageAnonExclusive(page)));
4010 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4011 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
4013 folio_unlock(folio);
4014 if (folio != swapcache && swapcache) {
4016 * Hold the lock to avoid the swap entry to be reused
4017 * until we take the PT lock for the pte_same() check
4018 * (to avoid false positives from pte_same). For
4019 * further safety release the lock after the swap_free
4020 * so that the swap count won't change under a
4021 * parallel locked swapcache.
4023 folio_unlock(swapcache);
4024 folio_put(swapcache);
4027 if (vmf->flags & FAULT_FLAG_WRITE) {
4028 ret |= do_wp_page(vmf);
4029 if (ret & VM_FAULT_ERROR)
4030 ret &= VM_FAULT_ERROR;
4034 /* No need to invalidate - it was non-present before */
4035 update_mmu_cache(vma, vmf->address, vmf->pte);
4038 pte_unmap_unlock(vmf->pte, vmf->ptl);
4041 put_swap_device(si);
4045 pte_unmap_unlock(vmf->pte, vmf->ptl);
4047 folio_unlock(folio);
4050 if (folio != swapcache && swapcache) {
4051 folio_unlock(swapcache);
4052 folio_put(swapcache);
4055 put_swap_device(si);
4060 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4061 * but allow concurrent faults), and pte mapped but not yet locked.
4062 * We return with mmap_lock still held, but pte unmapped and unlocked.
4064 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4066 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4067 struct vm_area_struct *vma = vmf->vma;
4068 struct folio *folio;
4072 /* File mapping without ->vm_ops ? */
4073 if (vma->vm_flags & VM_SHARED)
4074 return VM_FAULT_SIGBUS;
4077 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4078 * be distinguished from a transient failure of pte_offset_map().
4080 if (pte_alloc(vma->vm_mm, vmf->pmd))
4081 return VM_FAULT_OOM;
4083 /* Use the zero-page for reads */
4084 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4085 !mm_forbids_zeropage(vma->vm_mm)) {
4086 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
4087 vma->vm_page_prot));
4088 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4089 vmf->address, &vmf->ptl);
4092 if (vmf_pte_changed(vmf)) {
4093 update_mmu_tlb(vma, vmf->address, vmf->pte);
4096 ret = check_stable_address_space(vma->vm_mm);
4099 /* Deliver the page fault to userland, check inside PT lock */
4100 if (userfaultfd_missing(vma)) {
4101 pte_unmap_unlock(vmf->pte, vmf->ptl);
4102 return handle_userfault(vmf, VM_UFFD_MISSING);
4107 /* Allocate our own private page. */
4108 if (unlikely(anon_vma_prepare(vma)))
4110 folio = vma_alloc_zeroed_movable_folio(vma, vmf->address);
4114 if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
4116 folio_throttle_swaprate(folio, GFP_KERNEL);
4119 * The memory barrier inside __folio_mark_uptodate makes sure that
4120 * preceding stores to the page contents become visible before
4121 * the set_pte_at() write.
4123 __folio_mark_uptodate(folio);
4125 entry = mk_pte(&folio->page, vma->vm_page_prot);
4126 entry = pte_sw_mkyoung(entry);
4127 if (vma->vm_flags & VM_WRITE)
4128 entry = pte_mkwrite(pte_mkdirty(entry));
4130 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
4134 if (vmf_pte_changed(vmf)) {
4135 update_mmu_tlb(vma, vmf->address, vmf->pte);
4139 ret = check_stable_address_space(vma->vm_mm);
4143 /* Deliver the page fault to userland, check inside PT lock */
4144 if (userfaultfd_missing(vma)) {
4145 pte_unmap_unlock(vmf->pte, vmf->ptl);
4147 return handle_userfault(vmf, VM_UFFD_MISSING);
4150 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4151 folio_add_new_anon_rmap(folio, vma, vmf->address);
4152 folio_add_lru_vma(folio, vma);
4155 entry = pte_mkuffd_wp(entry);
4156 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
4158 /* No need to invalidate - it was non-present before */
4159 update_mmu_cache(vma, vmf->address, vmf->pte);
4162 pte_unmap_unlock(vmf->pte, vmf->ptl);
4170 return VM_FAULT_OOM;
4174 * The mmap_lock must have been held on entry, and may have been
4175 * released depending on flags and vma->vm_ops->fault() return value.
4176 * See filemap_fault() and __lock_page_retry().
4178 static vm_fault_t __do_fault(struct vm_fault *vmf)
4180 struct vm_area_struct *vma = vmf->vma;
4184 * Preallocate pte before we take page_lock because this might lead to
4185 * deadlocks for memcg reclaim which waits for pages under writeback:
4187 * SetPageWriteback(A)
4193 * wait_on_page_writeback(A)
4194 * SetPageWriteback(B)
4196 * # flush A, B to clear the writeback
4198 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
4199 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4200 if (!vmf->prealloc_pte)
4201 return VM_FAULT_OOM;
4204 ret = vma->vm_ops->fault(vmf);
4205 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4206 VM_FAULT_DONE_COW)))
4209 if (unlikely(PageHWPoison(vmf->page))) {
4210 struct page *page = vmf->page;
4211 vm_fault_t poisonret = VM_FAULT_HWPOISON;
4212 if (ret & VM_FAULT_LOCKED) {
4213 if (page_mapped(page))
4214 unmap_mapping_pages(page_mapping(page),
4215 page->index, 1, false);
4216 /* Retry if a clean page was removed from the cache. */
4217 if (invalidate_inode_page(page))
4218 poisonret = VM_FAULT_NOPAGE;
4226 if (unlikely(!(ret & VM_FAULT_LOCKED)))
4227 lock_page(vmf->page);
4229 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
4234 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4235 static void deposit_prealloc_pte(struct vm_fault *vmf)
4237 struct vm_area_struct *vma = vmf->vma;
4239 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4241 * We are going to consume the prealloc table,
4242 * count that as nr_ptes.
4244 mm_inc_nr_ptes(vma->vm_mm);
4245 vmf->prealloc_pte = NULL;
4248 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4250 struct vm_area_struct *vma = vmf->vma;
4251 bool write = vmf->flags & FAULT_FLAG_WRITE;
4252 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4255 vm_fault_t ret = VM_FAULT_FALLBACK;
4257 if (!transhuge_vma_suitable(vma, haddr))
4260 page = compound_head(page);
4261 if (compound_order(page) != HPAGE_PMD_ORDER)
4265 * Just backoff if any subpage of a THP is corrupted otherwise
4266 * the corrupted page may mapped by PMD silently to escape the
4267 * check. This kind of THP just can be PTE mapped. Access to
4268 * the corrupted subpage should trigger SIGBUS as expected.
4270 if (unlikely(PageHasHWPoisoned(page)))
4274 * Archs like ppc64 need additional space to store information
4275 * related to pte entry. Use the preallocated table for that.
4277 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4278 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4279 if (!vmf->prealloc_pte)
4280 return VM_FAULT_OOM;
4283 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4284 if (unlikely(!pmd_none(*vmf->pmd)))
4287 for (i = 0; i < HPAGE_PMD_NR; i++)
4288 flush_icache_page(vma, page + i);
4290 entry = mk_huge_pmd(page, vma->vm_page_prot);
4292 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
4294 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
4295 page_add_file_rmap(page, vma, true);
4298 * deposit and withdraw with pmd lock held
4300 if (arch_needs_pgtable_deposit())
4301 deposit_prealloc_pte(vmf);
4303 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
4305 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
4307 /* fault is handled */
4309 count_vm_event(THP_FILE_MAPPED);
4311 spin_unlock(vmf->ptl);
4315 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4317 return VM_FAULT_FALLBACK;
4321 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
4323 struct vm_area_struct *vma = vmf->vma;
4324 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4325 bool write = vmf->flags & FAULT_FLAG_WRITE;
4326 bool prefault = vmf->address != addr;
4329 flush_icache_page(vma, page);
4330 entry = mk_pte(page, vma->vm_page_prot);
4332 if (prefault && arch_wants_old_prefaulted_pte())
4333 entry = pte_mkold(entry);
4335 entry = pte_sw_mkyoung(entry);
4338 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4339 if (unlikely(uffd_wp))
4340 entry = pte_mkuffd_wp(entry);
4341 /* copy-on-write page */
4342 if (write && !(vma->vm_flags & VM_SHARED)) {
4343 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4344 page_add_new_anon_rmap(page, vma, addr);
4345 lru_cache_add_inactive_or_unevictable(page, vma);
4347 inc_mm_counter(vma->vm_mm, mm_counter_file(page));
4348 page_add_file_rmap(page, vma, false);
4350 set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
4353 static bool vmf_pte_changed(struct vm_fault *vmf)
4355 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4356 return !pte_same(ptep_get(vmf->pte), vmf->orig_pte);
4358 return !pte_none(ptep_get(vmf->pte));
4362 * finish_fault - finish page fault once we have prepared the page to fault
4364 * @vmf: structure describing the fault
4366 * This function handles all that is needed to finish a page fault once the
4367 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4368 * given page, adds reverse page mapping, handles memcg charges and LRU
4371 * The function expects the page to be locked and on success it consumes a
4372 * reference of a page being mapped (for the PTE which maps it).
4374 * Return: %0 on success, %VM_FAULT_ code in case of error.
4376 vm_fault_t finish_fault(struct vm_fault *vmf)
4378 struct vm_area_struct *vma = vmf->vma;
4382 /* Did we COW the page? */
4383 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4384 page = vmf->cow_page;
4389 * check even for read faults because we might have lost our CoWed
4392 if (!(vma->vm_flags & VM_SHARED)) {
4393 ret = check_stable_address_space(vma->vm_mm);
4398 if (pmd_none(*vmf->pmd)) {
4399 if (PageTransCompound(page)) {
4400 ret = do_set_pmd(vmf, page);
4401 if (ret != VM_FAULT_FALLBACK)
4405 if (vmf->prealloc_pte)
4406 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4407 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4408 return VM_FAULT_OOM;
4411 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4412 vmf->address, &vmf->ptl);
4414 return VM_FAULT_NOPAGE;
4416 /* Re-check under ptl */
4417 if (likely(!vmf_pte_changed(vmf))) {
4418 do_set_pte(vmf, page, vmf->address);
4420 /* no need to invalidate: a not-present page won't be cached */
4421 update_mmu_cache(vma, vmf->address, vmf->pte);
4425 update_mmu_tlb(vma, vmf->address, vmf->pte);
4426 ret = VM_FAULT_NOPAGE;
4429 pte_unmap_unlock(vmf->pte, vmf->ptl);
4433 static unsigned long fault_around_pages __read_mostly =
4434 65536 >> PAGE_SHIFT;
4436 #ifdef CONFIG_DEBUG_FS
4437 static int fault_around_bytes_get(void *data, u64 *val)
4439 *val = fault_around_pages << PAGE_SHIFT;
4444 * fault_around_bytes must be rounded down to the nearest page order as it's
4445 * what do_fault_around() expects to see.
4447 static int fault_around_bytes_set(void *data, u64 val)
4449 if (val / PAGE_SIZE > PTRS_PER_PTE)
4453 * The minimum value is 1 page, however this results in no fault-around
4454 * at all. See should_fault_around().
4456 fault_around_pages = max(rounddown_pow_of_two(val) >> PAGE_SHIFT, 1UL);
4460 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4461 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4463 static int __init fault_around_debugfs(void)
4465 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4466 &fault_around_bytes_fops);
4469 late_initcall(fault_around_debugfs);
4473 * do_fault_around() tries to map few pages around the fault address. The hope
4474 * is that the pages will be needed soon and this will lower the number of
4477 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4478 * not ready to be mapped: not up-to-date, locked, etc.
4480 * This function doesn't cross VMA or page table boundaries, in order to call
4481 * map_pages() and acquire a PTE lock only once.
4483 * fault_around_pages defines how many pages we'll try to map.
4484 * do_fault_around() expects it to be set to a power of two less than or equal
4487 * The virtual address of the area that we map is naturally aligned to
4488 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4489 * (and therefore to page order). This way it's easier to guarantee
4490 * that we don't cross page table boundaries.
4492 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4494 pgoff_t nr_pages = READ_ONCE(fault_around_pages);
4495 pgoff_t pte_off = pte_index(vmf->address);
4496 /* The page offset of vmf->address within the VMA. */
4497 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
4498 pgoff_t from_pte, to_pte;
4501 /* The PTE offset of the start address, clamped to the VMA. */
4502 from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
4503 pte_off - min(pte_off, vma_off));
4505 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4506 to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
4507 pte_off + vma_pages(vmf->vma) - vma_off) - 1;
4509 if (pmd_none(*vmf->pmd)) {
4510 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4511 if (!vmf->prealloc_pte)
4512 return VM_FAULT_OOM;
4516 ret = vmf->vma->vm_ops->map_pages(vmf,
4517 vmf->pgoff + from_pte - pte_off,
4518 vmf->pgoff + to_pte - pte_off);
4524 /* Return true if we should do read fault-around, false otherwise */
4525 static inline bool should_fault_around(struct vm_fault *vmf)
4527 /* No ->map_pages? No way to fault around... */
4528 if (!vmf->vma->vm_ops->map_pages)
4531 if (uffd_disable_fault_around(vmf->vma))
4534 /* A single page implies no faulting 'around' at all. */
4535 return fault_around_pages > 1;
4538 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4541 struct folio *folio;
4544 * Let's call ->map_pages() first and use ->fault() as fallback
4545 * if page by the offset is not ready to be mapped (cold cache or
4548 if (should_fault_around(vmf)) {
4549 ret = do_fault_around(vmf);
4554 ret = __do_fault(vmf);
4555 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4558 ret |= finish_fault(vmf);
4559 folio = page_folio(vmf->page);
4560 folio_unlock(folio);
4561 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4566 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4568 struct vm_area_struct *vma = vmf->vma;
4571 if (unlikely(anon_vma_prepare(vma)))
4572 return VM_FAULT_OOM;
4574 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4576 return VM_FAULT_OOM;
4578 if (mem_cgroup_charge(page_folio(vmf->cow_page), vma->vm_mm,
4580 put_page(vmf->cow_page);
4581 return VM_FAULT_OOM;
4583 folio_throttle_swaprate(page_folio(vmf->cow_page), GFP_KERNEL);
4585 ret = __do_fault(vmf);
4586 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4588 if (ret & VM_FAULT_DONE_COW)
4591 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4592 __SetPageUptodate(vmf->cow_page);
4594 ret |= finish_fault(vmf);
4595 unlock_page(vmf->page);
4596 put_page(vmf->page);
4597 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4601 put_page(vmf->cow_page);
4605 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4607 struct vm_area_struct *vma = vmf->vma;
4608 vm_fault_t ret, tmp;
4609 struct folio *folio;
4611 ret = __do_fault(vmf);
4612 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4615 folio = page_folio(vmf->page);
4618 * Check if the backing address space wants to know that the page is
4619 * about to become writable
4621 if (vma->vm_ops->page_mkwrite) {
4622 folio_unlock(folio);
4623 tmp = do_page_mkwrite(vmf, folio);
4624 if (unlikely(!tmp ||
4625 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4631 ret |= finish_fault(vmf);
4632 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4634 folio_unlock(folio);
4639 ret |= fault_dirty_shared_page(vmf);
4644 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4645 * but allow concurrent faults).
4646 * The mmap_lock may have been released depending on flags and our
4647 * return value. See filemap_fault() and __folio_lock_or_retry().
4648 * If mmap_lock is released, vma may become invalid (for example
4649 * by other thread calling munmap()).
4651 static vm_fault_t do_fault(struct vm_fault *vmf)
4653 struct vm_area_struct *vma = vmf->vma;
4654 struct mm_struct *vm_mm = vma->vm_mm;
4658 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4660 if (!vma->vm_ops->fault) {
4661 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
4662 vmf->address, &vmf->ptl);
4663 if (unlikely(!vmf->pte))
4664 ret = VM_FAULT_SIGBUS;
4667 * Make sure this is not a temporary clearing of pte
4668 * by holding ptl and checking again. A R/M/W update
4669 * of pte involves: take ptl, clearing the pte so that
4670 * we don't have concurrent modification by hardware
4671 * followed by an update.
4673 if (unlikely(pte_none(ptep_get(vmf->pte))))
4674 ret = VM_FAULT_SIGBUS;
4676 ret = VM_FAULT_NOPAGE;
4678 pte_unmap_unlock(vmf->pte, vmf->ptl);
4680 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4681 ret = do_read_fault(vmf);
4682 else if (!(vma->vm_flags & VM_SHARED))
4683 ret = do_cow_fault(vmf);
4685 ret = do_shared_fault(vmf);
4687 /* preallocated pagetable is unused: free it */
4688 if (vmf->prealloc_pte) {
4689 pte_free(vm_mm, vmf->prealloc_pte);
4690 vmf->prealloc_pte = NULL;
4695 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4696 unsigned long addr, int page_nid, int *flags)
4700 /* Record the current PID acceesing VMA */
4701 vma_set_access_pid_bit(vma);
4703 count_vm_numa_event(NUMA_HINT_FAULTS);
4704 if (page_nid == numa_node_id()) {
4705 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4706 *flags |= TNF_FAULT_LOCAL;
4709 return mpol_misplaced(page, vma, addr);
4712 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4714 struct vm_area_struct *vma = vmf->vma;
4715 struct page *page = NULL;
4716 int page_nid = NUMA_NO_NODE;
4717 bool writable = false;
4724 * The "pte" at this point cannot be used safely without
4725 * validation through pte_unmap_same(). It's of NUMA type but
4726 * the pfn may be screwed if the read is non atomic.
4728 spin_lock(vmf->ptl);
4729 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
4730 pte_unmap_unlock(vmf->pte, vmf->ptl);
4734 /* Get the normal PTE */
4735 old_pte = ptep_get(vmf->pte);
4736 pte = pte_modify(old_pte, vma->vm_page_prot);
4739 * Detect now whether the PTE could be writable; this information
4740 * is only valid while holding the PT lock.
4742 writable = pte_write(pte);
4743 if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
4744 can_change_pte_writable(vma, vmf->address, pte))
4747 page = vm_normal_page(vma, vmf->address, pte);
4748 if (!page || is_zone_device_page(page))
4751 /* TODO: handle PTE-mapped THP */
4752 if (PageCompound(page))
4756 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4757 * much anyway since they can be in shared cache state. This misses
4758 * the case where a mapping is writable but the process never writes
4759 * to it but pte_write gets cleared during protection updates and
4760 * pte_dirty has unpredictable behaviour between PTE scan updates,
4761 * background writeback, dirty balancing and application behaviour.
4764 flags |= TNF_NO_GROUP;
4767 * Flag if the page is shared between multiple address spaces. This
4768 * is later used when determining whether to group tasks together
4770 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4771 flags |= TNF_SHARED;
4773 page_nid = page_to_nid(page);
4775 * For memory tiering mode, cpupid of slow memory page is used
4776 * to record page access time. So use default value.
4778 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4779 !node_is_toptier(page_nid))
4780 last_cpupid = (-1 & LAST_CPUPID_MASK);
4782 last_cpupid = page_cpupid_last(page);
4783 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4785 if (target_nid == NUMA_NO_NODE) {
4789 pte_unmap_unlock(vmf->pte, vmf->ptl);
4792 /* Migrate to the requested node */
4793 if (migrate_misplaced_page(page, vma, target_nid)) {
4794 page_nid = target_nid;
4795 flags |= TNF_MIGRATED;
4797 flags |= TNF_MIGRATE_FAIL;
4798 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4799 vmf->address, &vmf->ptl);
4800 if (unlikely(!vmf->pte))
4802 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
4803 pte_unmap_unlock(vmf->pte, vmf->ptl);
4810 if (page_nid != NUMA_NO_NODE)
4811 task_numa_fault(last_cpupid, page_nid, 1, flags);
4815 * Make it present again, depending on how arch implements
4816 * non-accessible ptes, some can allow access by kernel mode.
4818 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4819 pte = pte_modify(old_pte, vma->vm_page_prot);
4820 pte = pte_mkyoung(pte);
4822 pte = pte_mkwrite(pte);
4823 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4824 update_mmu_cache(vma, vmf->address, vmf->pte);
4825 pte_unmap_unlock(vmf->pte, vmf->ptl);
4829 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4831 if (vma_is_anonymous(vmf->vma))
4832 return do_huge_pmd_anonymous_page(vmf);
4833 if (vmf->vma->vm_ops->huge_fault)
4834 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4835 return VM_FAULT_FALLBACK;
4838 /* `inline' is required to avoid gcc 4.1.2 build error */
4839 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4841 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
4844 if (vma_is_anonymous(vmf->vma)) {
4845 if (likely(!unshare) &&
4846 userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd))
4847 return handle_userfault(vmf, VM_UFFD_WP);
4848 return do_huge_pmd_wp_page(vmf);
4851 if (vmf->vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4852 if (vmf->vma->vm_ops->huge_fault) {
4853 ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4854 if (!(ret & VM_FAULT_FALLBACK))
4859 /* COW or write-notify handled on pte level: split pmd. */
4860 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4862 return VM_FAULT_FALLBACK;
4865 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4867 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4868 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4869 /* No support for anonymous transparent PUD pages yet */
4870 if (vma_is_anonymous(vmf->vma))
4871 return VM_FAULT_FALLBACK;
4872 if (vmf->vma->vm_ops->huge_fault)
4873 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4874 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4875 return VM_FAULT_FALLBACK;
4878 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4880 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4881 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4884 /* No support for anonymous transparent PUD pages yet */
4885 if (vma_is_anonymous(vmf->vma))
4887 if (vmf->vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4888 if (vmf->vma->vm_ops->huge_fault) {
4889 ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4890 if (!(ret & VM_FAULT_FALLBACK))
4895 /* COW or write-notify not handled on PUD level: split pud.*/
4896 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4897 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
4898 return VM_FAULT_FALLBACK;
4902 * These routines also need to handle stuff like marking pages dirty
4903 * and/or accessed for architectures that don't do it in hardware (most
4904 * RISC architectures). The early dirtying is also good on the i386.
4906 * There is also a hook called "update_mmu_cache()" that architectures
4907 * with external mmu caches can use to update those (ie the Sparc or
4908 * PowerPC hashed page tables that act as extended TLBs).
4910 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4911 * concurrent faults).
4913 * The mmap_lock may have been released depending on flags and our return value.
4914 * See filemap_fault() and __folio_lock_or_retry().
4916 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4920 if (unlikely(pmd_none(*vmf->pmd))) {
4922 * Leave __pte_alloc() until later: because vm_ops->fault may
4923 * want to allocate huge page, and if we expose page table
4924 * for an instant, it will be difficult to retract from
4925 * concurrent faults and from rmap lookups.
4928 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
4931 * A regular pmd is established and it can't morph into a huge
4932 * pmd by anon khugepaged, since that takes mmap_lock in write
4933 * mode; but shmem or file collapse to THP could still morph
4934 * it into a huge pmd: just retry later if so.
4936 vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd,
4937 vmf->address, &vmf->ptl);
4938 if (unlikely(!vmf->pte))
4940 vmf->orig_pte = ptep_get_lockless(vmf->pte);
4941 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
4943 if (pte_none(vmf->orig_pte)) {
4944 pte_unmap(vmf->pte);
4950 return do_pte_missing(vmf);
4952 if (!pte_present(vmf->orig_pte))
4953 return do_swap_page(vmf);
4955 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4956 return do_numa_page(vmf);
4958 spin_lock(vmf->ptl);
4959 entry = vmf->orig_pte;
4960 if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
4961 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4964 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
4965 if (!pte_write(entry))
4966 return do_wp_page(vmf);
4967 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
4968 entry = pte_mkdirty(entry);
4970 entry = pte_mkyoung(entry);
4971 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4972 vmf->flags & FAULT_FLAG_WRITE)) {
4973 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4975 /* Skip spurious TLB flush for retried page fault */
4976 if (vmf->flags & FAULT_FLAG_TRIED)
4979 * This is needed only for protection faults but the arch code
4980 * is not yet telling us if this is a protection fault or not.
4981 * This still avoids useless tlb flushes for .text page faults
4984 if (vmf->flags & FAULT_FLAG_WRITE)
4985 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
4989 pte_unmap_unlock(vmf->pte, vmf->ptl);
4994 * By the time we get here, we already hold the mm semaphore
4996 * The mmap_lock may have been released depending on flags and our
4997 * return value. See filemap_fault() and __folio_lock_or_retry().
4999 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
5000 unsigned long address, unsigned int flags)
5002 struct vm_fault vmf = {
5004 .address = address & PAGE_MASK,
5005 .real_address = address,
5007 .pgoff = linear_page_index(vma, address),
5008 .gfp_mask = __get_fault_gfp_mask(vma),
5010 struct mm_struct *mm = vma->vm_mm;
5011 unsigned long vm_flags = vma->vm_flags;
5016 pgd = pgd_offset(mm, address);
5017 p4d = p4d_alloc(mm, pgd, address);
5019 return VM_FAULT_OOM;
5021 vmf.pud = pud_alloc(mm, p4d, address);
5023 return VM_FAULT_OOM;
5025 if (pud_none(*vmf.pud) &&
5026 hugepage_vma_check(vma, vm_flags, false, true, true)) {
5027 ret = create_huge_pud(&vmf);
5028 if (!(ret & VM_FAULT_FALLBACK))
5031 pud_t orig_pud = *vmf.pud;
5034 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
5037 * TODO once we support anonymous PUDs: NUMA case and
5038 * FAULT_FLAG_UNSHARE handling.
5040 if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) {
5041 ret = wp_huge_pud(&vmf, orig_pud);
5042 if (!(ret & VM_FAULT_FALLBACK))
5045 huge_pud_set_accessed(&vmf, orig_pud);
5051 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
5053 return VM_FAULT_OOM;
5055 /* Huge pud page fault raced with pmd_alloc? */
5056 if (pud_trans_unstable(vmf.pud))
5059 if (pmd_none(*vmf.pmd) &&
5060 hugepage_vma_check(vma, vm_flags, false, true, true)) {
5061 ret = create_huge_pmd(&vmf);
5062 if (!(ret & VM_FAULT_FALLBACK))
5065 vmf.orig_pmd = pmdp_get_lockless(vmf.pmd);
5067 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
5068 VM_BUG_ON(thp_migration_supported() &&
5069 !is_pmd_migration_entry(vmf.orig_pmd));
5070 if (is_pmd_migration_entry(vmf.orig_pmd))
5071 pmd_migration_entry_wait(mm, vmf.pmd);
5074 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
5075 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
5076 return do_huge_pmd_numa_page(&vmf);
5078 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5079 !pmd_write(vmf.orig_pmd)) {
5080 ret = wp_huge_pmd(&vmf);
5081 if (!(ret & VM_FAULT_FALLBACK))
5084 huge_pmd_set_accessed(&vmf);
5090 return handle_pte_fault(&vmf);
5094 * mm_account_fault - Do page fault accounting
5095 * @mm: mm from which memcg should be extracted. It can be NULL.
5096 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5097 * of perf event counters, but we'll still do the per-task accounting to
5098 * the task who triggered this page fault.
5099 * @address: the faulted address.
5100 * @flags: the fault flags.
5101 * @ret: the fault retcode.
5103 * This will take care of most of the page fault accounting. Meanwhile, it
5104 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5105 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5106 * still be in per-arch page fault handlers at the entry of page fault.
5108 static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
5109 unsigned long address, unsigned int flags,
5114 /* Incomplete faults will be accounted upon completion. */
5115 if (ret & VM_FAULT_RETRY)
5119 * To preserve the behavior of older kernels, PGFAULT counters record
5120 * both successful and failed faults, as opposed to perf counters,
5121 * which ignore failed cases.
5123 count_vm_event(PGFAULT);
5124 count_memcg_event_mm(mm, PGFAULT);
5127 * Do not account for unsuccessful faults (e.g. when the address wasn't
5128 * valid). That includes arch_vma_access_permitted() failing before
5129 * reaching here. So this is not a "this many hardware page faults"
5130 * counter. We should use the hw profiling for that.
5132 if (ret & VM_FAULT_ERROR)
5136 * We define the fault as a major fault when the final successful fault
5137 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5138 * handle it immediately previously).
5140 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5148 * If the fault is done for GUP, regs will be NULL. We only do the
5149 * accounting for the per thread fault counters who triggered the
5150 * fault, and we skip the perf event updates.
5156 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
5158 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
5161 #ifdef CONFIG_LRU_GEN
5162 static void lru_gen_enter_fault(struct vm_area_struct *vma)
5164 /* the LRU algorithm only applies to accesses with recency */
5165 current->in_lru_fault = vma_has_recency(vma);
5168 static void lru_gen_exit_fault(void)
5170 current->in_lru_fault = false;
5173 static void lru_gen_enter_fault(struct vm_area_struct *vma)
5177 static void lru_gen_exit_fault(void)
5180 #endif /* CONFIG_LRU_GEN */
5182 static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5183 unsigned int *flags)
5185 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5186 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5187 return VM_FAULT_SIGSEGV;
5189 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5190 * just treat it like an ordinary read-fault otherwise.
5192 if (!is_cow_mapping(vma->vm_flags))
5193 *flags &= ~FAULT_FLAG_UNSHARE;
5194 } else if (*flags & FAULT_FLAG_WRITE) {
5195 /* Write faults on read-only mappings are impossible ... */
5196 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5197 return VM_FAULT_SIGSEGV;
5198 /* ... and FOLL_FORCE only applies to COW mappings. */
5199 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5200 !is_cow_mapping(vma->vm_flags)))
5201 return VM_FAULT_SIGSEGV;
5207 * By the time we get here, we already hold the mm semaphore
5209 * The mmap_lock may have been released depending on flags and our
5210 * return value. See filemap_fault() and __folio_lock_or_retry().
5212 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5213 unsigned int flags, struct pt_regs *regs)
5215 /* If the fault handler drops the mmap_lock, vma may be freed */
5216 struct mm_struct *mm = vma->vm_mm;
5219 __set_current_state(TASK_RUNNING);
5221 ret = sanitize_fault_flags(vma, &flags);
5225 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
5226 flags & FAULT_FLAG_INSTRUCTION,
5227 flags & FAULT_FLAG_REMOTE)) {
5228 ret = VM_FAULT_SIGSEGV;
5233 * Enable the memcg OOM handling for faults triggered in user
5234 * space. Kernel faults are handled more gracefully.
5236 if (flags & FAULT_FLAG_USER)
5237 mem_cgroup_enter_user_fault();
5239 lru_gen_enter_fault(vma);
5241 if (unlikely(is_vm_hugetlb_page(vma)))
5242 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
5244 ret = __handle_mm_fault(vma, address, flags);
5246 lru_gen_exit_fault();
5248 if (flags & FAULT_FLAG_USER) {
5249 mem_cgroup_exit_user_fault();
5251 * The task may have entered a memcg OOM situation but
5252 * if the allocation error was handled gracefully (no
5253 * VM_FAULT_OOM), there is no need to kill anything.
5254 * Just clean up the OOM state peacefully.
5256 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5257 mem_cgroup_oom_synchronize(false);
5260 mm_account_fault(mm, regs, address, flags, ret);
5264 EXPORT_SYMBOL_GPL(handle_mm_fault);
5266 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5267 #include <linux/extable.h>
5269 static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5271 /* Even if this succeeds, make it clear we *might* have slept */
5272 if (likely(mmap_read_trylock(mm))) {
5277 if (regs && !user_mode(regs)) {
5278 unsigned long ip = instruction_pointer(regs);
5279 if (!search_exception_tables(ip))
5283 return !mmap_read_lock_killable(mm);
5286 static inline bool mmap_upgrade_trylock(struct mm_struct *mm)
5289 * We don't have this operation yet.
5291 * It should be easy enough to do: it's basically a
5292 * atomic_long_try_cmpxchg_acquire()
5293 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5294 * it also needs the proper lockdep magic etc.
5299 static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5301 mmap_read_unlock(mm);
5302 if (regs && !user_mode(regs)) {
5303 unsigned long ip = instruction_pointer(regs);
5304 if (!search_exception_tables(ip))
5307 return !mmap_write_lock_killable(mm);
5311 * Helper for page fault handling.
5313 * This is kind of equivalend to "mmap_read_lock()" followed
5314 * by "find_extend_vma()", except it's a lot more careful about
5315 * the locking (and will drop the lock on failure).
5317 * For example, if we have a kernel bug that causes a page
5318 * fault, we don't want to just use mmap_read_lock() to get
5319 * the mm lock, because that would deadlock if the bug were
5320 * to happen while we're holding the mm lock for writing.
5322 * So this checks the exception tables on kernel faults in
5323 * order to only do this all for instructions that are actually
5324 * expected to fault.
5326 * We can also actually take the mm lock for writing if we
5327 * need to extend the vma, which helps the VM layer a lot.
5329 struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm,
5330 unsigned long addr, struct pt_regs *regs)
5332 struct vm_area_struct *vma;
5334 if (!get_mmap_lock_carefully(mm, regs))
5337 vma = find_vma(mm, addr);
5338 if (likely(vma && (vma->vm_start <= addr)))
5342 * Well, dang. We might still be successful, but only
5343 * if we can extend a vma to do so.
5345 if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) {
5346 mmap_read_unlock(mm);
5351 * We can try to upgrade the mmap lock atomically,
5352 * in which case we can continue to use the vma
5353 * we already looked up.
5355 * Otherwise we'll have to drop the mmap lock and
5356 * re-take it, and also look up the vma again,
5359 if (!mmap_upgrade_trylock(mm)) {
5360 if (!upgrade_mmap_lock_carefully(mm, regs))
5363 vma = find_vma(mm, addr);
5366 if (vma->vm_start <= addr)
5368 if (!(vma->vm_flags & VM_GROWSDOWN))
5372 if (expand_stack_locked(vma, addr))
5376 mmap_write_downgrade(mm);
5380 mmap_write_unlock(mm);
5385 #ifdef CONFIG_PER_VMA_LOCK
5387 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5388 * stable and not isolated. If the VMA is not found or is being modified the
5389 * function returns NULL.
5391 struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
5392 unsigned long address)
5394 MA_STATE(mas, &mm->mm_mt, address, address);
5395 struct vm_area_struct *vma;
5399 vma = mas_walk(&mas);
5403 /* Only anonymous and tcp vmas are supported for now */
5404 if (!vma_is_anonymous(vma) && !vma_is_tcp(vma))
5407 if (!vma_start_read(vma))
5411 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5412 * This check must happen after vma_start_read(); otherwise, a
5413 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5414 * from its anon_vma.
5416 if (unlikely(!vma->anon_vma && !vma_is_tcp(vma)))
5417 goto inval_end_read;
5420 * Due to the possibility of userfault handler dropping mmap_lock, avoid
5421 * it for now and fall back to page fault handling under mmap_lock.
5423 if (userfaultfd_armed(vma))
5424 goto inval_end_read;
5426 /* Check since vm_start/vm_end might change before we lock the VMA */
5427 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5428 goto inval_end_read;
5430 /* Check if the VMA got isolated after we found it */
5431 if (vma->detached) {
5433 count_vm_vma_lock_event(VMA_LOCK_MISS);
5434 /* The area was replaced with another one */
5445 count_vm_vma_lock_event(VMA_LOCK_ABORT);
5448 #endif /* CONFIG_PER_VMA_LOCK */
5450 #ifndef __PAGETABLE_P4D_FOLDED
5452 * Allocate p4d page table.
5453 * We've already handled the fast-path in-line.
5455 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5457 p4d_t *new = p4d_alloc_one(mm, address);
5461 spin_lock(&mm->page_table_lock);
5462 if (pgd_present(*pgd)) { /* Another has populated it */
5465 smp_wmb(); /* See comment in pmd_install() */
5466 pgd_populate(mm, pgd, new);
5468 spin_unlock(&mm->page_table_lock);
5471 #endif /* __PAGETABLE_P4D_FOLDED */
5473 #ifndef __PAGETABLE_PUD_FOLDED
5475 * Allocate page upper directory.
5476 * We've already handled the fast-path in-line.
5478 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5480 pud_t *new = pud_alloc_one(mm, address);
5484 spin_lock(&mm->page_table_lock);
5485 if (!p4d_present(*p4d)) {
5487 smp_wmb(); /* See comment in pmd_install() */
5488 p4d_populate(mm, p4d, new);
5489 } else /* Another has populated it */
5491 spin_unlock(&mm->page_table_lock);
5494 #endif /* __PAGETABLE_PUD_FOLDED */
5496 #ifndef __PAGETABLE_PMD_FOLDED
5498 * Allocate page middle directory.
5499 * We've already handled the fast-path in-line.
5501 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5504 pmd_t *new = pmd_alloc_one(mm, address);
5508 ptl = pud_lock(mm, pud);
5509 if (!pud_present(*pud)) {
5511 smp_wmb(); /* See comment in pmd_install() */
5512 pud_populate(mm, pud, new);
5513 } else { /* Another has populated it */
5519 #endif /* __PAGETABLE_PMD_FOLDED */
5522 * follow_pte - look up PTE at a user virtual address
5523 * @mm: the mm_struct of the target address space
5524 * @address: user virtual address
5525 * @ptepp: location to store found PTE
5526 * @ptlp: location to store the lock for the PTE
5528 * On a successful return, the pointer to the PTE is stored in @ptepp;
5529 * the corresponding lock is taken and its location is stored in @ptlp.
5530 * The contents of the PTE are only stable until @ptlp is released;
5531 * any further use, if any, must be protected against invalidation
5532 * with MMU notifiers.
5534 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5535 * should be taken for read.
5537 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5538 * it is not a good general-purpose API.
5540 * Return: zero on success, -ve otherwise.
5542 int follow_pte(struct mm_struct *mm, unsigned long address,
5543 pte_t **ptepp, spinlock_t **ptlp)
5551 pgd = pgd_offset(mm, address);
5552 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
5555 p4d = p4d_offset(pgd, address);
5556 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
5559 pud = pud_offset(p4d, address);
5560 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
5563 pmd = pmd_offset(pud, address);
5564 VM_BUG_ON(pmd_trans_huge(*pmd));
5566 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
5569 if (!pte_present(ptep_get(ptep)))
5574 pte_unmap_unlock(ptep, *ptlp);
5578 EXPORT_SYMBOL_GPL(follow_pte);
5581 * follow_pfn - look up PFN at a user virtual address
5582 * @vma: memory mapping
5583 * @address: user virtual address
5584 * @pfn: location to store found PFN
5586 * Only IO mappings and raw PFN mappings are allowed.
5588 * This function does not allow the caller to read the permissions
5589 * of the PTE. Do not use it.
5591 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5593 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5600 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5603 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5606 *pfn = pte_pfn(ptep_get(ptep));
5607 pte_unmap_unlock(ptep, ptl);
5610 EXPORT_SYMBOL(follow_pfn);
5612 #ifdef CONFIG_HAVE_IOREMAP_PROT
5613 int follow_phys(struct vm_area_struct *vma,
5614 unsigned long address, unsigned int flags,
5615 unsigned long *prot, resource_size_t *phys)
5621 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5624 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5626 pte = ptep_get(ptep);
5628 if ((flags & FOLL_WRITE) && !pte_write(pte))
5631 *prot = pgprot_val(pte_pgprot(pte));
5632 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5636 pte_unmap_unlock(ptep, ptl);
5642 * generic_access_phys - generic implementation for iomem mmap access
5643 * @vma: the vma to access
5644 * @addr: userspace address, not relative offset within @vma
5645 * @buf: buffer to read/write
5646 * @len: length of transfer
5647 * @write: set to FOLL_WRITE when writing, otherwise reading
5649 * This is a generic implementation for &vm_operations_struct.access for an
5650 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5653 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5654 void *buf, int len, int write)
5656 resource_size_t phys_addr;
5657 unsigned long prot = 0;
5658 void __iomem *maddr;
5661 int offset = offset_in_page(addr);
5664 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5668 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5670 pte = ptep_get(ptep);
5671 pte_unmap_unlock(ptep, ptl);
5673 prot = pgprot_val(pte_pgprot(pte));
5674 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5676 if ((write & FOLL_WRITE) && !pte_write(pte))
5679 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5683 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5686 if (!pte_same(pte, ptep_get(ptep))) {
5687 pte_unmap_unlock(ptep, ptl);
5694 memcpy_toio(maddr + offset, buf, len);
5696 memcpy_fromio(buf, maddr + offset, len);
5698 pte_unmap_unlock(ptep, ptl);
5704 EXPORT_SYMBOL_GPL(generic_access_phys);
5708 * Access another process' address space as given in mm.
5710 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5711 int len, unsigned int gup_flags)
5713 void *old_buf = buf;
5714 int write = gup_flags & FOLL_WRITE;
5716 if (mmap_read_lock_killable(mm))
5719 /* Avoid triggering the temporary warning in __get_user_pages */
5720 if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
5723 /* ignore errors, just check how much was successfully transferred */
5727 struct vm_area_struct *vma = NULL;
5728 struct page *page = get_user_page_vma_remote(mm, addr,
5731 if (IS_ERR_OR_NULL(page)) {
5732 /* We might need to expand the stack to access it */
5733 vma = vma_lookup(mm, addr);
5735 vma = expand_stack(mm, addr);
5737 /* mmap_lock was dropped on failure */
5739 return buf - old_buf;
5741 /* Try again if stack expansion worked */
5747 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5748 * we can access using slightly different code.
5751 #ifdef CONFIG_HAVE_IOREMAP_PROT
5752 if (vma->vm_ops && vma->vm_ops->access)
5753 bytes = vma->vm_ops->access(vma, addr, buf,
5760 offset = addr & (PAGE_SIZE-1);
5761 if (bytes > PAGE_SIZE-offset)
5762 bytes = PAGE_SIZE-offset;
5766 copy_to_user_page(vma, page, addr,
5767 maddr + offset, buf, bytes);
5768 set_page_dirty_lock(page);
5770 copy_from_user_page(vma, page, addr,
5771 buf, maddr + offset, bytes);
5780 mmap_read_unlock(mm);
5782 return buf - old_buf;
5786 * access_remote_vm - access another process' address space
5787 * @mm: the mm_struct of the target address space
5788 * @addr: start address to access
5789 * @buf: source or destination buffer
5790 * @len: number of bytes to transfer
5791 * @gup_flags: flags modifying lookup behaviour
5793 * The caller must hold a reference on @mm.
5795 * Return: number of bytes copied from source to destination.
5797 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5798 void *buf, int len, unsigned int gup_flags)
5800 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5804 * Access another process' address space.
5805 * Source/target buffer must be kernel space,
5806 * Do not walk the page table directly, use get_user_pages
5808 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5809 void *buf, int len, unsigned int gup_flags)
5811 struct mm_struct *mm;
5814 mm = get_task_mm(tsk);
5818 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5824 EXPORT_SYMBOL_GPL(access_process_vm);
5827 * Print the name of a VMA.
5829 void print_vma_addr(char *prefix, unsigned long ip)
5831 struct mm_struct *mm = current->mm;
5832 struct vm_area_struct *vma;
5835 * we might be running from an atomic context so we cannot sleep
5837 if (!mmap_read_trylock(mm))
5840 vma = find_vma(mm, ip);
5841 if (vma && vma->vm_file) {
5842 struct file *f = vma->vm_file;
5843 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5847 p = file_path(f, buf, PAGE_SIZE);
5850 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5852 vma->vm_end - vma->vm_start);
5853 free_page((unsigned long)buf);
5856 mmap_read_unlock(mm);
5859 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5860 void __might_fault(const char *file, int line)
5862 if (pagefault_disabled())
5864 __might_sleep(file, line);
5865 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5867 might_lock_read(¤t->mm->mmap_lock);
5870 EXPORT_SYMBOL(__might_fault);
5873 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5875 * Process all subpages of the specified huge page with the specified
5876 * operation. The target subpage will be processed last to keep its
5879 static inline int process_huge_page(
5880 unsigned long addr_hint, unsigned int pages_per_huge_page,
5881 int (*process_subpage)(unsigned long addr, int idx, void *arg),
5884 int i, n, base, l, ret;
5885 unsigned long addr = addr_hint &
5886 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5888 /* Process target subpage last to keep its cache lines hot */
5890 n = (addr_hint - addr) / PAGE_SIZE;
5891 if (2 * n <= pages_per_huge_page) {
5892 /* If target subpage in first half of huge page */
5895 /* Process subpages at the end of huge page */
5896 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5898 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
5903 /* If target subpage in second half of huge page */
5904 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5905 l = pages_per_huge_page - n;
5906 /* Process subpages at the begin of huge page */
5907 for (i = 0; i < base; i++) {
5909 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
5915 * Process remaining subpages in left-right-left-right pattern
5916 * towards the target subpage
5918 for (i = 0; i < l; i++) {
5919 int left_idx = base + i;
5920 int right_idx = base + 2 * l - 1 - i;
5923 ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5927 ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5934 static void clear_gigantic_page(struct page *page,
5936 unsigned int pages_per_huge_page)
5942 for (i = 0; i < pages_per_huge_page; i++) {
5943 p = nth_page(page, i);
5945 clear_user_highpage(p, addr + i * PAGE_SIZE);
5949 static int clear_subpage(unsigned long addr, int idx, void *arg)
5951 struct page *page = arg;
5953 clear_user_highpage(page + idx, addr);
5957 void clear_huge_page(struct page *page,
5958 unsigned long addr_hint, unsigned int pages_per_huge_page)
5960 unsigned long addr = addr_hint &
5961 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5963 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5964 clear_gigantic_page(page, addr, pages_per_huge_page);
5968 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5971 static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
5973 struct vm_area_struct *vma,
5974 unsigned int pages_per_huge_page)
5977 struct page *dst_page;
5978 struct page *src_page;
5980 for (i = 0; i < pages_per_huge_page; i++) {
5981 dst_page = folio_page(dst, i);
5982 src_page = folio_page(src, i);
5985 if (copy_mc_user_highpage(dst_page, src_page,
5986 addr + i*PAGE_SIZE, vma)) {
5987 memory_failure_queue(page_to_pfn(src_page), 0);
5994 struct copy_subpage_arg {
5997 struct vm_area_struct *vma;
6000 static int copy_subpage(unsigned long addr, int idx, void *arg)
6002 struct copy_subpage_arg *copy_arg = arg;
6004 if (copy_mc_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
6005 addr, copy_arg->vma)) {
6006 memory_failure_queue(page_to_pfn(copy_arg->src + idx), 0);
6012 int copy_user_large_folio(struct folio *dst, struct folio *src,
6013 unsigned long addr_hint, struct vm_area_struct *vma)
6015 unsigned int pages_per_huge_page = folio_nr_pages(dst);
6016 unsigned long addr = addr_hint &
6017 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6018 struct copy_subpage_arg arg = {
6024 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES))
6025 return copy_user_gigantic_page(dst, src, addr, vma,
6026 pages_per_huge_page);
6028 return process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
6031 long copy_folio_from_user(struct folio *dst_folio,
6032 const void __user *usr_src,
6033 bool allow_pagefault)
6036 unsigned long i, rc = 0;
6037 unsigned int nr_pages = folio_nr_pages(dst_folio);
6038 unsigned long ret_val = nr_pages * PAGE_SIZE;
6039 struct page *subpage;
6041 for (i = 0; i < nr_pages; i++) {
6042 subpage = folio_page(dst_folio, i);
6043 kaddr = kmap_local_page(subpage);
6044 if (!allow_pagefault)
6045 pagefault_disable();
6046 rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE);
6047 if (!allow_pagefault)
6049 kunmap_local(kaddr);
6051 ret_val -= (PAGE_SIZE - rc);
6055 flush_dcache_page(subpage);
6061 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6063 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6065 static struct kmem_cache *page_ptl_cachep;
6067 void __init ptlock_cache_init(void)
6069 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
6073 bool ptlock_alloc(struct page *page)
6077 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
6084 void ptlock_free(struct page *page)
6086 kmem_cache_free(page_ptl_cachep, page->ptl);