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>
81 #include <trace/events/kmem.h>
84 #include <asm/mmu_context.h>
85 #include <asm/pgalloc.h>
86 #include <linux/uaccess.h>
88 #include <asm/tlbflush.h>
90 #include "pgalloc-track.h"
94 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
95 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
99 unsigned long max_mapnr;
100 EXPORT_SYMBOL(max_mapnr);
102 struct page *mem_map;
103 EXPORT_SYMBOL(mem_map);
106 static vm_fault_t do_fault(struct vm_fault *vmf);
107 static vm_fault_t do_anonymous_page(struct vm_fault *vmf);
108 static bool vmf_pte_changed(struct vm_fault *vmf);
111 * Return true if the original pte was a uffd-wp pte marker (so the pte was
114 static bool vmf_orig_pte_uffd_wp(struct vm_fault *vmf)
116 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID))
119 return pte_marker_uffd_wp(vmf->orig_pte);
123 * A number of key systems in x86 including ioremap() rely on the assumption
124 * that high_memory defines the upper bound on direct map memory, then end
125 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
126 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
130 EXPORT_SYMBOL(high_memory);
133 * Randomize the address space (stacks, mmaps, brk, etc.).
135 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
136 * as ancient (libc5 based) binaries can segfault. )
138 int randomize_va_space __read_mostly =
139 #ifdef CONFIG_COMPAT_BRK
145 #ifndef arch_wants_old_prefaulted_pte
146 static inline bool arch_wants_old_prefaulted_pte(void)
149 * Transitioning a PTE from 'old' to 'young' can be expensive on
150 * some architectures, even if it's performed in hardware. By
151 * default, "false" means prefaulted entries will be 'young'.
157 static int __init disable_randmaps(char *s)
159 randomize_va_space = 0;
162 __setup("norandmaps", disable_randmaps);
164 unsigned long zero_pfn __read_mostly;
165 EXPORT_SYMBOL(zero_pfn);
167 unsigned long highest_memmap_pfn __read_mostly;
170 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
172 static int __init init_zero_pfn(void)
174 zero_pfn = page_to_pfn(ZERO_PAGE(0));
177 early_initcall(init_zero_pfn);
179 void mm_trace_rss_stat(struct mm_struct *mm, int member)
181 trace_rss_stat(mm, member);
185 * Note: this doesn't free the actual pages themselves. That
186 * has been handled earlier when unmapping all the memory regions.
188 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
191 pgtable_t token = pmd_pgtable(*pmd);
193 pte_free_tlb(tlb, token, addr);
194 mm_dec_nr_ptes(tlb->mm);
197 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
198 unsigned long addr, unsigned long end,
199 unsigned long floor, unsigned long ceiling)
206 pmd = pmd_offset(pud, addr);
208 next = pmd_addr_end(addr, end);
209 if (pmd_none_or_clear_bad(pmd))
211 free_pte_range(tlb, pmd, addr);
212 } while (pmd++, addr = next, addr != end);
222 if (end - 1 > ceiling - 1)
225 pmd = pmd_offset(pud, start);
227 pmd_free_tlb(tlb, pmd, start);
228 mm_dec_nr_pmds(tlb->mm);
231 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
232 unsigned long addr, unsigned long end,
233 unsigned long floor, unsigned long ceiling)
240 pud = pud_offset(p4d, addr);
242 next = pud_addr_end(addr, end);
243 if (pud_none_or_clear_bad(pud))
245 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
246 } while (pud++, addr = next, addr != end);
256 if (end - 1 > ceiling - 1)
259 pud = pud_offset(p4d, start);
261 pud_free_tlb(tlb, pud, start);
262 mm_dec_nr_puds(tlb->mm);
265 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
266 unsigned long addr, unsigned long end,
267 unsigned long floor, unsigned long ceiling)
274 p4d = p4d_offset(pgd, addr);
276 next = p4d_addr_end(addr, end);
277 if (p4d_none_or_clear_bad(p4d))
279 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
280 } while (p4d++, addr = next, addr != end);
286 ceiling &= PGDIR_MASK;
290 if (end - 1 > ceiling - 1)
293 p4d = p4d_offset(pgd, start);
295 p4d_free_tlb(tlb, p4d, start);
299 * This function frees user-level page tables of a process.
301 void free_pgd_range(struct mmu_gather *tlb,
302 unsigned long addr, unsigned long end,
303 unsigned long floor, unsigned long ceiling)
309 * The next few lines have given us lots of grief...
311 * Why are we testing PMD* at this top level? Because often
312 * there will be no work to do at all, and we'd prefer not to
313 * go all the way down to the bottom just to discover that.
315 * Why all these "- 1"s? Because 0 represents both the bottom
316 * of the address space and the top of it (using -1 for the
317 * top wouldn't help much: the masks would do the wrong thing).
318 * The rule is that addr 0 and floor 0 refer to the bottom of
319 * the address space, but end 0 and ceiling 0 refer to the top
320 * Comparisons need to use "end - 1" and "ceiling - 1" (though
321 * that end 0 case should be mythical).
323 * Wherever addr is brought up or ceiling brought down, we must
324 * be careful to reject "the opposite 0" before it confuses the
325 * subsequent tests. But what about where end is brought down
326 * by PMD_SIZE below? no, end can't go down to 0 there.
328 * Whereas we round start (addr) and ceiling down, by different
329 * masks at different levels, in order to test whether a table
330 * now has no other vmas using it, so can be freed, we don't
331 * bother to round floor or end up - the tests don't need that.
345 if (end - 1 > ceiling - 1)
350 * We add page table cache pages with PAGE_SIZE,
351 * (see pte_free_tlb()), flush the tlb if we need
353 tlb_change_page_size(tlb, PAGE_SIZE);
354 pgd = pgd_offset(tlb->mm, addr);
356 next = pgd_addr_end(addr, end);
357 if (pgd_none_or_clear_bad(pgd))
359 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
360 } while (pgd++, addr = next, addr != end);
363 void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
364 struct vm_area_struct *vma, unsigned long floor,
365 unsigned long ceiling, bool mm_wr_locked)
368 unsigned long addr = vma->vm_start;
369 struct vm_area_struct *next;
372 * Note: USER_PGTABLES_CEILING may be passed as ceiling and may
373 * be 0. This will underflow and is okay.
375 next = mas_find(mas, ceiling - 1);
378 * Hide vma from rmap and truncate_pagecache before freeing
382 vma_start_write(vma);
383 unlink_anon_vmas(vma);
384 unlink_file_vma(vma);
386 if (is_vm_hugetlb_page(vma)) {
387 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
388 floor, next ? next->vm_start : ceiling);
391 * Optimization: gather nearby vmas into one call down
393 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
394 && !is_vm_hugetlb_page(next)) {
396 next = mas_find(mas, ceiling - 1);
398 vma_start_write(vma);
399 unlink_anon_vmas(vma);
400 unlink_file_vma(vma);
402 free_pgd_range(tlb, addr, vma->vm_end,
403 floor, next ? next->vm_start : ceiling);
409 void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
411 spinlock_t *ptl = pmd_lock(mm, pmd);
413 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
416 * Ensure all pte setup (eg. pte page lock and page clearing) are
417 * visible before the pte is made visible to other CPUs by being
418 * put into page tables.
420 * The other side of the story is the pointer chasing in the page
421 * table walking code (when walking the page table without locking;
422 * ie. most of the time). Fortunately, these data accesses consist
423 * of a chain of data-dependent loads, meaning most CPUs (alpha
424 * being the notable exception) will already guarantee loads are
425 * seen in-order. See the alpha page table accessors for the
426 * smp_rmb() barriers in page table walking code.
428 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
429 pmd_populate(mm, pmd, *pte);
435 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
437 pgtable_t new = pte_alloc_one(mm);
441 pmd_install(mm, pmd, &new);
447 int __pte_alloc_kernel(pmd_t *pmd)
449 pte_t *new = pte_alloc_one_kernel(&init_mm);
453 spin_lock(&init_mm.page_table_lock);
454 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
455 smp_wmb(); /* See comment in pmd_install() */
456 pmd_populate_kernel(&init_mm, pmd, new);
459 spin_unlock(&init_mm.page_table_lock);
461 pte_free_kernel(&init_mm, new);
465 static inline void init_rss_vec(int *rss)
467 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
470 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
474 for (i = 0; i < NR_MM_COUNTERS; i++)
476 add_mm_counter(mm, i, rss[i]);
480 * This function is called to print an error when a bad pte
481 * is found. For example, we might have a PFN-mapped pte in
482 * a region that doesn't allow it.
484 * The calling function must still handle the error.
486 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
487 pte_t pte, struct page *page)
489 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
490 p4d_t *p4d = p4d_offset(pgd, addr);
491 pud_t *pud = pud_offset(p4d, addr);
492 pmd_t *pmd = pmd_offset(pud, addr);
493 struct address_space *mapping;
495 static unsigned long resume;
496 static unsigned long nr_shown;
497 static unsigned long nr_unshown;
500 * Allow a burst of 60 reports, then keep quiet for that minute;
501 * or allow a steady drip of one report per second.
503 if (nr_shown == 60) {
504 if (time_before(jiffies, resume)) {
509 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
516 resume = jiffies + 60 * HZ;
518 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
519 index = linear_page_index(vma, addr);
521 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
523 (long long)pte_val(pte), (long long)pmd_val(*pmd));
525 dump_page(page, "bad pte");
526 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
527 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
528 pr_alert("file:%pD fault:%ps mmap:%ps read_folio:%ps\n",
530 vma->vm_ops ? vma->vm_ops->fault : NULL,
531 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
532 mapping ? mapping->a_ops->read_folio : NULL);
534 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
538 * vm_normal_page -- This function gets the "struct page" associated with a pte.
540 * "Special" mappings do not wish to be associated with a "struct page" (either
541 * it doesn't exist, or it exists but they don't want to touch it). In this
542 * case, NULL is returned here. "Normal" mappings do have a struct page.
544 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
545 * pte bit, in which case this function is trivial. Secondly, an architecture
546 * may not have a spare pte bit, which requires a more complicated scheme,
549 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
550 * special mapping (even if there are underlying and valid "struct pages").
551 * COWed pages of a VM_PFNMAP are always normal.
553 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
554 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
555 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
556 * mapping will always honor the rule
558 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
560 * And for normal mappings this is false.
562 * This restricts such mappings to be a linear translation from virtual address
563 * to pfn. To get around this restriction, we allow arbitrary mappings so long
564 * as the vma is not a COW mapping; in that case, we know that all ptes are
565 * special (because none can have been COWed).
568 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
570 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
571 * page" backing, however the difference is that _all_ pages with a struct
572 * page (that is, those where pfn_valid is true) are refcounted and considered
573 * normal pages by the VM. The disadvantage is that pages are refcounted
574 * (which can be slower and simply not an option for some PFNMAP users). The
575 * advantage is that we don't have to follow the strict linearity rule of
576 * PFNMAP mappings in order to support COWable mappings.
579 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
582 unsigned long pfn = pte_pfn(pte);
584 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
585 if (likely(!pte_special(pte)))
587 if (vma->vm_ops && vma->vm_ops->find_special_page)
588 return vma->vm_ops->find_special_page(vma, addr);
589 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
591 if (is_zero_pfn(pfn))
595 * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
596 * and will have refcounts incremented on their struct pages
597 * when they are inserted into PTEs, thus they are safe to
598 * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
599 * do not have refcounts. Example of legacy ZONE_DEVICE is
600 * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
604 print_bad_pte(vma, addr, pte, NULL);
608 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
610 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
611 if (vma->vm_flags & VM_MIXEDMAP) {
617 off = (addr - vma->vm_start) >> PAGE_SHIFT;
618 if (pfn == vma->vm_pgoff + off)
620 if (!is_cow_mapping(vma->vm_flags))
625 if (is_zero_pfn(pfn))
629 if (unlikely(pfn > highest_memmap_pfn)) {
630 print_bad_pte(vma, addr, pte, NULL);
635 * NOTE! We still have PageReserved() pages in the page tables.
636 * eg. VDSO mappings can cause them to exist.
639 return pfn_to_page(pfn);
642 struct folio *vm_normal_folio(struct vm_area_struct *vma, unsigned long addr,
645 struct page *page = vm_normal_page(vma, addr, pte);
648 return page_folio(page);
652 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
653 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
656 unsigned long pfn = pmd_pfn(pmd);
659 * There is no pmd_special() but there may be special pmds, e.g.
660 * in a direct-access (dax) mapping, so let's just replicate the
661 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
663 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
664 if (vma->vm_flags & VM_MIXEDMAP) {
670 off = (addr - vma->vm_start) >> PAGE_SHIFT;
671 if (pfn == vma->vm_pgoff + off)
673 if (!is_cow_mapping(vma->vm_flags))
680 if (is_huge_zero_pmd(pmd))
682 if (unlikely(pfn > highest_memmap_pfn))
686 * NOTE! We still have PageReserved() pages in the page tables.
687 * eg. VDSO mappings can cause them to exist.
690 return pfn_to_page(pfn);
694 static void restore_exclusive_pte(struct vm_area_struct *vma,
695 struct page *page, unsigned long address,
702 orig_pte = ptep_get(ptep);
703 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
704 if (pte_swp_soft_dirty(orig_pte))
705 pte = pte_mksoft_dirty(pte);
707 entry = pte_to_swp_entry(orig_pte);
708 if (pte_swp_uffd_wp(orig_pte))
709 pte = pte_mkuffd_wp(pte);
710 else if (is_writable_device_exclusive_entry(entry))
711 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
713 VM_BUG_ON(pte_write(pte) && !(PageAnon(page) && PageAnonExclusive(page)));
716 * No need to take a page reference as one was already
717 * created when the swap entry was made.
720 page_add_anon_rmap(page, vma, address, RMAP_NONE);
723 * Currently device exclusive access only supports anonymous
724 * memory so the entry shouldn't point to a filebacked page.
728 set_pte_at(vma->vm_mm, address, ptep, pte);
731 * No need to invalidate - it was non-present before. However
732 * secondary CPUs may have mappings that need invalidating.
734 update_mmu_cache(vma, address, ptep);
738 * Tries to restore an exclusive pte if the page lock can be acquired without
742 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
745 swp_entry_t entry = pte_to_swp_entry(ptep_get(src_pte));
746 struct page *page = pfn_swap_entry_to_page(entry);
748 if (trylock_page(page)) {
749 restore_exclusive_pte(vma, page, addr, src_pte);
758 * copy one vm_area from one task to the other. Assumes the page tables
759 * already present in the new task to be cleared in the whole range
760 * covered by this vma.
764 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
765 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
766 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
768 unsigned long vm_flags = dst_vma->vm_flags;
769 pte_t orig_pte = ptep_get(src_pte);
770 pte_t pte = orig_pte;
772 swp_entry_t entry = pte_to_swp_entry(orig_pte);
774 if (likely(!non_swap_entry(entry))) {
775 if (swap_duplicate(entry) < 0)
778 /* make sure dst_mm is on swapoff's mmlist. */
779 if (unlikely(list_empty(&dst_mm->mmlist))) {
780 spin_lock(&mmlist_lock);
781 if (list_empty(&dst_mm->mmlist))
782 list_add(&dst_mm->mmlist,
784 spin_unlock(&mmlist_lock);
786 /* Mark the swap entry as shared. */
787 if (pte_swp_exclusive(orig_pte)) {
788 pte = pte_swp_clear_exclusive(orig_pte);
789 set_pte_at(src_mm, addr, src_pte, pte);
792 } else if (is_migration_entry(entry)) {
793 page = pfn_swap_entry_to_page(entry);
795 rss[mm_counter(page)]++;
797 if (!is_readable_migration_entry(entry) &&
798 is_cow_mapping(vm_flags)) {
800 * COW mappings require pages in both parent and child
801 * to be set to read. A previously exclusive entry is
804 entry = make_readable_migration_entry(
806 pte = swp_entry_to_pte(entry);
807 if (pte_swp_soft_dirty(orig_pte))
808 pte = pte_swp_mksoft_dirty(pte);
809 if (pte_swp_uffd_wp(orig_pte))
810 pte = pte_swp_mkuffd_wp(pte);
811 set_pte_at(src_mm, addr, src_pte, pte);
813 } else if (is_device_private_entry(entry)) {
814 page = pfn_swap_entry_to_page(entry);
817 * Update rss count even for unaddressable pages, as
818 * they should treated just like normal pages in this
821 * We will likely want to have some new rss counters
822 * for unaddressable pages, at some point. But for now
823 * keep things as they are.
826 rss[mm_counter(page)]++;
827 /* Cannot fail as these pages cannot get pinned. */
828 BUG_ON(page_try_dup_anon_rmap(page, false, src_vma));
831 * We do not preserve soft-dirty information, because so
832 * far, checkpoint/restore is the only feature that
833 * requires that. And checkpoint/restore does not work
834 * when a device driver is involved (you cannot easily
835 * save and restore device driver state).
837 if (is_writable_device_private_entry(entry) &&
838 is_cow_mapping(vm_flags)) {
839 entry = make_readable_device_private_entry(
841 pte = swp_entry_to_pte(entry);
842 if (pte_swp_uffd_wp(orig_pte))
843 pte = pte_swp_mkuffd_wp(pte);
844 set_pte_at(src_mm, addr, src_pte, pte);
846 } else if (is_device_exclusive_entry(entry)) {
848 * Make device exclusive entries present by restoring the
849 * original entry then copying as for a present pte. Device
850 * exclusive entries currently only support private writable
851 * (ie. COW) mappings.
853 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
854 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
857 } else if (is_pte_marker_entry(entry)) {
858 pte_marker marker = copy_pte_marker(entry, dst_vma);
861 set_pte_at(dst_mm, addr, dst_pte,
862 make_pte_marker(marker));
865 if (!userfaultfd_wp(dst_vma))
866 pte = pte_swp_clear_uffd_wp(pte);
867 set_pte_at(dst_mm, addr, dst_pte, pte);
872 * Copy a present and normal page.
874 * NOTE! The usual case is that this isn't required;
875 * instead, the caller can just increase the page refcount
876 * and re-use the pte the traditional way.
878 * And if we need a pre-allocated page but don't yet have
879 * one, return a negative error to let the preallocation
880 * code know so that it can do so outside the page table
884 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
885 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
886 struct folio **prealloc, struct page *page)
888 struct folio *new_folio;
891 new_folio = *prealloc;
896 * We have a prealloc page, all good! Take it
897 * over and copy the page & arm it.
900 copy_user_highpage(&new_folio->page, page, addr, src_vma);
901 __folio_mark_uptodate(new_folio);
902 folio_add_new_anon_rmap(new_folio, dst_vma, addr);
903 folio_add_lru_vma(new_folio, dst_vma);
906 /* All done, just insert the new page copy in the child */
907 pte = mk_pte(&new_folio->page, dst_vma->vm_page_prot);
908 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
909 if (userfaultfd_pte_wp(dst_vma, ptep_get(src_pte)))
910 /* Uffd-wp needs to be delivered to dest pte as well */
911 pte = pte_mkuffd_wp(pte);
912 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
917 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
918 * is required to copy this pte.
921 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
922 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
923 struct folio **prealloc)
925 struct mm_struct *src_mm = src_vma->vm_mm;
926 unsigned long vm_flags = src_vma->vm_flags;
927 pte_t pte = ptep_get(src_pte);
931 page = vm_normal_page(src_vma, addr, pte);
933 folio = page_folio(page);
934 if (page && folio_test_anon(folio)) {
936 * If this page may have been pinned by the parent process,
937 * copy the page immediately for the child so that we'll always
938 * guarantee the pinned page won't be randomly replaced in the
942 if (unlikely(page_try_dup_anon_rmap(page, false, src_vma))) {
943 /* Page may be pinned, we have to copy. */
945 return copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
946 addr, rss, prealloc, page);
951 page_dup_file_rmap(page, false);
952 rss[mm_counter_file(page)]++;
956 * If it's a COW mapping, write protect it both
957 * in the parent and the child
959 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
960 ptep_set_wrprotect(src_mm, addr, src_pte);
961 pte = pte_wrprotect(pte);
963 VM_BUG_ON(page && folio_test_anon(folio) && PageAnonExclusive(page));
966 * If it's a shared mapping, mark it clean in
969 if (vm_flags & VM_SHARED)
970 pte = pte_mkclean(pte);
971 pte = pte_mkold(pte);
973 if (!userfaultfd_wp(dst_vma))
974 pte = pte_clear_uffd_wp(pte);
976 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
980 static inline struct folio *page_copy_prealloc(struct mm_struct *src_mm,
981 struct vm_area_struct *vma, unsigned long addr)
983 struct folio *new_folio;
985 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma, addr, false);
989 if (mem_cgroup_charge(new_folio, src_mm, GFP_KERNEL)) {
990 folio_put(new_folio);
993 folio_throttle_swaprate(new_folio, GFP_KERNEL);
999 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1000 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1003 struct mm_struct *dst_mm = dst_vma->vm_mm;
1004 struct mm_struct *src_mm = src_vma->vm_mm;
1005 pte_t *orig_src_pte, *orig_dst_pte;
1006 pte_t *src_pte, *dst_pte;
1008 spinlock_t *src_ptl, *dst_ptl;
1009 int progress, ret = 0;
1010 int rss[NR_MM_COUNTERS];
1011 swp_entry_t entry = (swp_entry_t){0};
1012 struct folio *prealloc = NULL;
1019 * copy_pmd_range()'s prior pmd_none_or_clear_bad(src_pmd), and the
1020 * error handling here, assume that exclusive mmap_lock on dst and src
1021 * protects anon from unexpected THP transitions; with shmem and file
1022 * protected by mmap_lock-less collapse skipping areas with anon_vma
1023 * (whereas vma_needs_copy() skips areas without anon_vma). A rework
1024 * can remove such assumptions later, but this is good enough for now.
1026 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1031 src_pte = pte_offset_map_nolock(src_mm, src_pmd, addr, &src_ptl);
1033 pte_unmap_unlock(dst_pte, dst_ptl);
1037 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1038 orig_src_pte = src_pte;
1039 orig_dst_pte = dst_pte;
1040 arch_enter_lazy_mmu_mode();
1044 * We are holding two locks at this point - either of them
1045 * could generate latencies in another task on another CPU.
1047 if (progress >= 32) {
1049 if (need_resched() ||
1050 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1053 ptent = ptep_get(src_pte);
1054 if (pte_none(ptent)) {
1058 if (unlikely(!pte_present(ptent))) {
1059 ret = copy_nonpresent_pte(dst_mm, src_mm,
1064 entry = pte_to_swp_entry(ptep_get(src_pte));
1066 } else if (ret == -EBUSY) {
1074 * Device exclusive entry restored, continue by copying
1075 * the now present pte.
1077 WARN_ON_ONCE(ret != -ENOENT);
1079 /* copy_present_pte() will clear `*prealloc' if consumed */
1080 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1081 addr, rss, &prealloc);
1083 * If we need a pre-allocated page for this pte, drop the
1084 * locks, allocate, and try again.
1086 if (unlikely(ret == -EAGAIN))
1088 if (unlikely(prealloc)) {
1090 * pre-alloc page cannot be reused by next time so as
1091 * to strictly follow mempolicy (e.g., alloc_page_vma()
1092 * will allocate page according to address). This
1093 * could only happen if one pinned pte changed.
1095 folio_put(prealloc);
1099 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1101 arch_leave_lazy_mmu_mode();
1102 pte_unmap_unlock(orig_src_pte, src_ptl);
1103 add_mm_rss_vec(dst_mm, rss);
1104 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1108 VM_WARN_ON_ONCE(!entry.val);
1109 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1114 } else if (ret == -EBUSY) {
1116 } else if (ret == -EAGAIN) {
1117 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1124 /* We've captured and resolved the error. Reset, try again. */
1130 if (unlikely(prealloc))
1131 folio_put(prealloc);
1136 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1137 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1140 struct mm_struct *dst_mm = dst_vma->vm_mm;
1141 struct mm_struct *src_mm = src_vma->vm_mm;
1142 pmd_t *src_pmd, *dst_pmd;
1145 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1148 src_pmd = pmd_offset(src_pud, addr);
1150 next = pmd_addr_end(addr, end);
1151 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1152 || pmd_devmap(*src_pmd)) {
1154 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1155 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1156 addr, dst_vma, src_vma);
1163 if (pmd_none_or_clear_bad(src_pmd))
1165 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1168 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1173 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1174 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1177 struct mm_struct *dst_mm = dst_vma->vm_mm;
1178 struct mm_struct *src_mm = src_vma->vm_mm;
1179 pud_t *src_pud, *dst_pud;
1182 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1185 src_pud = pud_offset(src_p4d, addr);
1187 next = pud_addr_end(addr, end);
1188 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1191 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1192 err = copy_huge_pud(dst_mm, src_mm,
1193 dst_pud, src_pud, addr, src_vma);
1200 if (pud_none_or_clear_bad(src_pud))
1202 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1205 } while (dst_pud++, src_pud++, addr = next, addr != end);
1210 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1211 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1214 struct mm_struct *dst_mm = dst_vma->vm_mm;
1215 p4d_t *src_p4d, *dst_p4d;
1218 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1221 src_p4d = p4d_offset(src_pgd, addr);
1223 next = p4d_addr_end(addr, end);
1224 if (p4d_none_or_clear_bad(src_p4d))
1226 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1229 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1234 * Return true if the vma needs to copy the pgtable during this fork(). Return
1235 * false when we can speed up fork() by allowing lazy page faults later until
1236 * when the child accesses the memory range.
1239 vma_needs_copy(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1242 * Always copy pgtables when dst_vma has uffd-wp enabled even if it's
1243 * file-backed (e.g. shmem). Because when uffd-wp is enabled, pgtable
1244 * contains uffd-wp protection information, that's something we can't
1245 * retrieve from page cache, and skip copying will lose those info.
1247 if (userfaultfd_wp(dst_vma))
1250 if (src_vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
1253 if (src_vma->anon_vma)
1257 * Don't copy ptes where a page fault will fill them correctly. Fork
1258 * becomes much lighter when there are big shared or private readonly
1259 * mappings. The tradeoff is that copy_page_range is more efficient
1266 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1268 pgd_t *src_pgd, *dst_pgd;
1270 unsigned long addr = src_vma->vm_start;
1271 unsigned long end = src_vma->vm_end;
1272 struct mm_struct *dst_mm = dst_vma->vm_mm;
1273 struct mm_struct *src_mm = src_vma->vm_mm;
1274 struct mmu_notifier_range range;
1278 if (!vma_needs_copy(dst_vma, src_vma))
1281 if (is_vm_hugetlb_page(src_vma))
1282 return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
1284 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1286 * We do not free on error cases below as remove_vma
1287 * gets called on error from higher level routine
1289 ret = track_pfn_copy(src_vma);
1295 * We need to invalidate the secondary MMU mappings only when
1296 * there could be a permission downgrade on the ptes of the
1297 * parent mm. And a permission downgrade will only happen if
1298 * is_cow_mapping() returns true.
1300 is_cow = is_cow_mapping(src_vma->vm_flags);
1303 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1304 0, src_mm, addr, end);
1305 mmu_notifier_invalidate_range_start(&range);
1307 * Disabling preemption is not needed for the write side, as
1308 * the read side doesn't spin, but goes to the mmap_lock.
1310 * Use the raw variant of the seqcount_t write API to avoid
1311 * lockdep complaining about preemptibility.
1313 vma_assert_write_locked(src_vma);
1314 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1318 dst_pgd = pgd_offset(dst_mm, addr);
1319 src_pgd = pgd_offset(src_mm, addr);
1321 next = pgd_addr_end(addr, end);
1322 if (pgd_none_or_clear_bad(src_pgd))
1324 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1326 untrack_pfn_clear(dst_vma);
1330 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1333 raw_write_seqcount_end(&src_mm->write_protect_seq);
1334 mmu_notifier_invalidate_range_end(&range);
1339 /* Whether we should zap all COWed (private) pages too */
1340 static inline bool should_zap_cows(struct zap_details *details)
1342 /* By default, zap all pages */
1346 /* Or, we zap COWed pages only if the caller wants to */
1347 return details->even_cows;
1350 /* Decides whether we should zap this page with the page pointer specified */
1351 static inline bool should_zap_page(struct zap_details *details, struct page *page)
1353 /* If we can make a decision without *page.. */
1354 if (should_zap_cows(details))
1357 /* E.g. the caller passes NULL for the case of a zero page */
1361 /* Otherwise we should only zap non-anon pages */
1362 return !PageAnon(page);
1365 static inline bool zap_drop_file_uffd_wp(struct zap_details *details)
1370 return details->zap_flags & ZAP_FLAG_DROP_MARKER;
1374 * This function makes sure that we'll replace the none pte with an uffd-wp
1375 * swap special pte marker when necessary. Must be with the pgtable lock held.
1378 zap_install_uffd_wp_if_needed(struct vm_area_struct *vma,
1379 unsigned long addr, pte_t *pte,
1380 struct zap_details *details, pte_t pteval)
1382 /* Zap on anonymous always means dropping everything */
1383 if (vma_is_anonymous(vma))
1386 if (zap_drop_file_uffd_wp(details))
1389 pte_install_uffd_wp_if_needed(vma, addr, pte, pteval);
1392 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1393 struct vm_area_struct *vma, pmd_t *pmd,
1394 unsigned long addr, unsigned long end,
1395 struct zap_details *details)
1397 struct mm_struct *mm = tlb->mm;
1398 int force_flush = 0;
1399 int rss[NR_MM_COUNTERS];
1405 tlb_change_page_size(tlb, PAGE_SIZE);
1407 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1411 flush_tlb_batched_pending(mm);
1412 arch_enter_lazy_mmu_mode();
1414 pte_t ptent = ptep_get(pte);
1417 if (pte_none(ptent))
1423 if (pte_present(ptent)) {
1424 unsigned int delay_rmap;
1426 page = vm_normal_page(vma, addr, ptent);
1427 if (unlikely(!should_zap_page(details, page)))
1429 ptent = ptep_get_and_clear_full(mm, addr, pte,
1431 arch_check_zapped_pte(vma, ptent);
1432 tlb_remove_tlb_entry(tlb, pte, addr);
1433 zap_install_uffd_wp_if_needed(vma, addr, pte, details,
1435 if (unlikely(!page)) {
1436 ksm_might_unmap_zero_page(mm, ptent);
1441 if (!PageAnon(page)) {
1442 if (pte_dirty(ptent)) {
1443 set_page_dirty(page);
1444 if (tlb_delay_rmap(tlb)) {
1449 if (pte_young(ptent) && likely(vma_has_recency(vma)))
1450 mark_page_accessed(page);
1452 rss[mm_counter(page)]--;
1454 page_remove_rmap(page, vma, false);
1455 if (unlikely(page_mapcount(page) < 0))
1456 print_bad_pte(vma, addr, ptent, page);
1458 if (unlikely(__tlb_remove_page(tlb, page, delay_rmap))) {
1466 entry = pte_to_swp_entry(ptent);
1467 if (is_device_private_entry(entry) ||
1468 is_device_exclusive_entry(entry)) {
1469 page = pfn_swap_entry_to_page(entry);
1470 if (unlikely(!should_zap_page(details, page)))
1473 * Both device private/exclusive mappings should only
1474 * work with anonymous page so far, so we don't need to
1475 * consider uffd-wp bit when zap. For more information,
1476 * see zap_install_uffd_wp_if_needed().
1478 WARN_ON_ONCE(!vma_is_anonymous(vma));
1479 rss[mm_counter(page)]--;
1480 if (is_device_private_entry(entry))
1481 page_remove_rmap(page, vma, false);
1483 } else if (!non_swap_entry(entry)) {
1484 /* Genuine swap entry, hence a private anon page */
1485 if (!should_zap_cows(details))
1488 if (unlikely(!free_swap_and_cache(entry)))
1489 print_bad_pte(vma, addr, ptent, NULL);
1490 } else if (is_migration_entry(entry)) {
1491 page = pfn_swap_entry_to_page(entry);
1492 if (!should_zap_page(details, page))
1494 rss[mm_counter(page)]--;
1495 } else if (pte_marker_entry_uffd_wp(entry)) {
1497 * For anon: always drop the marker; for file: only
1498 * drop the marker if explicitly requested.
1500 if (!vma_is_anonymous(vma) &&
1501 !zap_drop_file_uffd_wp(details))
1503 } else if (is_hwpoison_entry(entry) ||
1504 is_poisoned_swp_entry(entry)) {
1505 if (!should_zap_cows(details))
1508 /* We should have covered all the swap entry types */
1511 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1512 zap_install_uffd_wp_if_needed(vma, addr, pte, details, ptent);
1513 } while (pte++, addr += PAGE_SIZE, addr != end);
1515 add_mm_rss_vec(mm, rss);
1516 arch_leave_lazy_mmu_mode();
1518 /* Do the actual TLB flush before dropping ptl */
1520 tlb_flush_mmu_tlbonly(tlb);
1521 tlb_flush_rmaps(tlb, vma);
1523 pte_unmap_unlock(start_pte, ptl);
1526 * If we forced a TLB flush (either due to running out of
1527 * batch buffers or because we needed to flush dirty TLB
1528 * entries before releasing the ptl), free the batched
1529 * memory too. Come back again if we didn't do everything.
1537 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1538 struct vm_area_struct *vma, pud_t *pud,
1539 unsigned long addr, unsigned long end,
1540 struct zap_details *details)
1545 pmd = pmd_offset(pud, addr);
1547 next = pmd_addr_end(addr, end);
1548 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1549 if (next - addr != HPAGE_PMD_SIZE)
1550 __split_huge_pmd(vma, pmd, addr, false, NULL);
1551 else if (zap_huge_pmd(tlb, vma, pmd, addr)) {
1556 } else if (details && details->single_folio &&
1557 folio_test_pmd_mappable(details->single_folio) &&
1558 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1559 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1561 * Take and drop THP pmd lock so that we cannot return
1562 * prematurely, while zap_huge_pmd() has cleared *pmd,
1563 * but not yet decremented compound_mapcount().
1567 if (pmd_none(*pmd)) {
1571 addr = zap_pte_range(tlb, vma, pmd, addr, next, details);
1574 } while (pmd++, cond_resched(), addr != end);
1579 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1580 struct vm_area_struct *vma, p4d_t *p4d,
1581 unsigned long addr, unsigned long end,
1582 struct zap_details *details)
1587 pud = pud_offset(p4d, addr);
1589 next = pud_addr_end(addr, end);
1590 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1591 if (next - addr != HPAGE_PUD_SIZE) {
1592 mmap_assert_locked(tlb->mm);
1593 split_huge_pud(vma, pud, addr);
1594 } else if (zap_huge_pud(tlb, vma, pud, addr))
1598 if (pud_none_or_clear_bad(pud))
1600 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1603 } while (pud++, addr = next, addr != end);
1608 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1609 struct vm_area_struct *vma, pgd_t *pgd,
1610 unsigned long addr, unsigned long end,
1611 struct zap_details *details)
1616 p4d = p4d_offset(pgd, addr);
1618 next = p4d_addr_end(addr, end);
1619 if (p4d_none_or_clear_bad(p4d))
1621 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1622 } while (p4d++, addr = next, addr != end);
1627 void unmap_page_range(struct mmu_gather *tlb,
1628 struct vm_area_struct *vma,
1629 unsigned long addr, unsigned long end,
1630 struct zap_details *details)
1635 BUG_ON(addr >= end);
1636 tlb_start_vma(tlb, vma);
1637 pgd = pgd_offset(vma->vm_mm, addr);
1639 next = pgd_addr_end(addr, end);
1640 if (pgd_none_or_clear_bad(pgd))
1642 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1643 } while (pgd++, addr = next, addr != end);
1644 tlb_end_vma(tlb, vma);
1648 static void unmap_single_vma(struct mmu_gather *tlb,
1649 struct vm_area_struct *vma, unsigned long start_addr,
1650 unsigned long end_addr,
1651 struct zap_details *details, bool mm_wr_locked)
1653 unsigned long start = max(vma->vm_start, start_addr);
1656 if (start >= vma->vm_end)
1658 end = min(vma->vm_end, end_addr);
1659 if (end <= vma->vm_start)
1663 uprobe_munmap(vma, start, end);
1665 if (unlikely(vma->vm_flags & VM_PFNMAP))
1666 untrack_pfn(vma, 0, 0, mm_wr_locked);
1669 if (unlikely(is_vm_hugetlb_page(vma))) {
1671 * It is undesirable to test vma->vm_file as it
1672 * should be non-null for valid hugetlb area.
1673 * However, vm_file will be NULL in the error
1674 * cleanup path of mmap_region. When
1675 * hugetlbfs ->mmap method fails,
1676 * mmap_region() nullifies vma->vm_file
1677 * before calling this function to clean up.
1678 * Since no pte has actually been setup, it is
1679 * safe to do nothing in this case.
1682 zap_flags_t zap_flags = details ?
1683 details->zap_flags : 0;
1684 __unmap_hugepage_range_final(tlb, vma, start, end,
1688 unmap_page_range(tlb, vma, start, end, details);
1693 * unmap_vmas - unmap a range of memory covered by a list of vma's
1694 * @tlb: address of the caller's struct mmu_gather
1695 * @mas: the maple state
1696 * @vma: the starting vma
1697 * @start_addr: virtual address at which to start unmapping
1698 * @end_addr: virtual address at which to end unmapping
1699 * @tree_end: The maximum index to check
1700 * @mm_wr_locked: lock flag
1702 * Unmap all pages in the vma list.
1704 * Only addresses between `start' and `end' will be unmapped.
1706 * The VMA list must be sorted in ascending virtual address order.
1708 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1709 * range after unmap_vmas() returns. So the only responsibility here is to
1710 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1711 * drops the lock and schedules.
1713 void unmap_vmas(struct mmu_gather *tlb, struct ma_state *mas,
1714 struct vm_area_struct *vma, unsigned long start_addr,
1715 unsigned long end_addr, unsigned long tree_end,
1718 struct mmu_notifier_range range;
1719 struct zap_details details = {
1720 .zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
1721 /* Careful - we need to zap private pages too! */
1725 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma->vm_mm,
1726 start_addr, end_addr);
1727 mmu_notifier_invalidate_range_start(&range);
1729 unmap_single_vma(tlb, vma, start_addr, end_addr, &details,
1731 } while ((vma = mas_find(mas, tree_end - 1)) != NULL);
1732 mmu_notifier_invalidate_range_end(&range);
1736 * zap_page_range_single - remove user pages in a given range
1737 * @vma: vm_area_struct holding the applicable pages
1738 * @address: starting address of pages to zap
1739 * @size: number of bytes to zap
1740 * @details: details of shared cache invalidation
1742 * The range must fit into one VMA.
1744 void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1745 unsigned long size, struct zap_details *details)
1747 const unsigned long end = address + size;
1748 struct mmu_notifier_range range;
1749 struct mmu_gather tlb;
1752 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1754 if (is_vm_hugetlb_page(vma))
1755 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1757 tlb_gather_mmu(&tlb, vma->vm_mm);
1758 update_hiwater_rss(vma->vm_mm);
1759 mmu_notifier_invalidate_range_start(&range);
1761 * unmap 'address-end' not 'range.start-range.end' as range
1762 * could have been expanded for hugetlb pmd sharing.
1764 unmap_single_vma(&tlb, vma, address, end, details, false);
1765 mmu_notifier_invalidate_range_end(&range);
1766 tlb_finish_mmu(&tlb);
1770 * zap_vma_ptes - remove ptes mapping the vma
1771 * @vma: vm_area_struct holding ptes to be zapped
1772 * @address: starting address of pages to zap
1773 * @size: number of bytes to zap
1775 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1777 * The entire address range must be fully contained within the vma.
1780 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1783 if (!range_in_vma(vma, address, address + size) ||
1784 !(vma->vm_flags & VM_PFNMAP))
1787 zap_page_range_single(vma, address, size, NULL);
1789 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1791 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1798 pgd = pgd_offset(mm, addr);
1799 p4d = p4d_alloc(mm, pgd, addr);
1802 pud = pud_alloc(mm, p4d, addr);
1805 pmd = pmd_alloc(mm, pud, addr);
1809 VM_BUG_ON(pmd_trans_huge(*pmd));
1813 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1816 pmd_t *pmd = walk_to_pmd(mm, addr);
1820 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1823 static int validate_page_before_insert(struct page *page)
1825 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1827 flush_dcache_page(page);
1831 static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1832 unsigned long addr, struct page *page, pgprot_t prot)
1834 if (!pte_none(ptep_get(pte)))
1836 /* Ok, finally just insert the thing.. */
1838 inc_mm_counter(vma->vm_mm, mm_counter_file(page));
1839 page_add_file_rmap(page, vma, false);
1840 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
1845 * This is the old fallback for page remapping.
1847 * For historical reasons, it only allows reserved pages. Only
1848 * old drivers should use this, and they needed to mark their
1849 * pages reserved for the old functions anyway.
1851 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1852 struct page *page, pgprot_t prot)
1858 retval = validate_page_before_insert(page);
1862 pte = get_locked_pte(vma->vm_mm, addr, &ptl);
1865 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
1866 pte_unmap_unlock(pte, ptl);
1871 static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
1872 unsigned long addr, struct page *page, pgprot_t prot)
1876 if (!page_count(page))
1878 err = validate_page_before_insert(page);
1881 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
1884 /* insert_pages() amortizes the cost of spinlock operations
1885 * when inserting pages in a loop.
1887 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1888 struct page **pages, unsigned long *num, pgprot_t prot)
1891 pte_t *start_pte, *pte;
1892 spinlock_t *pte_lock;
1893 struct mm_struct *const mm = vma->vm_mm;
1894 unsigned long curr_page_idx = 0;
1895 unsigned long remaining_pages_total = *num;
1896 unsigned long pages_to_write_in_pmd;
1900 pmd = walk_to_pmd(mm, addr);
1904 pages_to_write_in_pmd = min_t(unsigned long,
1905 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1907 /* Allocate the PTE if necessary; takes PMD lock once only. */
1909 if (pte_alloc(mm, pmd))
1912 while (pages_to_write_in_pmd) {
1914 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1916 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1921 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1922 int err = insert_page_in_batch_locked(vma, pte,
1923 addr, pages[curr_page_idx], prot);
1924 if (unlikely(err)) {
1925 pte_unmap_unlock(start_pte, pte_lock);
1927 remaining_pages_total -= pte_idx;
1933 pte_unmap_unlock(start_pte, pte_lock);
1934 pages_to_write_in_pmd -= batch_size;
1935 remaining_pages_total -= batch_size;
1937 if (remaining_pages_total)
1941 *num = remaining_pages_total;
1946 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1947 * @vma: user vma to map to
1948 * @addr: target start user address of these pages
1949 * @pages: source kernel pages
1950 * @num: in: number of pages to map. out: number of pages that were *not*
1951 * mapped. (0 means all pages were successfully mapped).
1953 * Preferred over vm_insert_page() when inserting multiple pages.
1955 * In case of error, we may have mapped a subset of the provided
1956 * pages. It is the caller's responsibility to account for this case.
1958 * The same restrictions apply as in vm_insert_page().
1960 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1961 struct page **pages, unsigned long *num)
1963 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1965 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1967 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1968 BUG_ON(mmap_read_trylock(vma->vm_mm));
1969 BUG_ON(vma->vm_flags & VM_PFNMAP);
1970 vm_flags_set(vma, VM_MIXEDMAP);
1972 /* Defer page refcount checking till we're about to map that page. */
1973 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1975 EXPORT_SYMBOL(vm_insert_pages);
1978 * vm_insert_page - insert single page into user vma
1979 * @vma: user vma to map to
1980 * @addr: target user address of this page
1981 * @page: source kernel page
1983 * This allows drivers to insert individual pages they've allocated
1986 * The page has to be a nice clean _individual_ kernel allocation.
1987 * If you allocate a compound page, you need to have marked it as
1988 * such (__GFP_COMP), or manually just split the page up yourself
1989 * (see split_page()).
1991 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1992 * took an arbitrary page protection parameter. This doesn't allow
1993 * that. Your vma protection will have to be set up correctly, which
1994 * means that if you want a shared writable mapping, you'd better
1995 * ask for a shared writable mapping!
1997 * The page does not need to be reserved.
1999 * Usually this function is called from f_op->mmap() handler
2000 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
2001 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2002 * function from other places, for example from page-fault handler.
2004 * Return: %0 on success, negative error code otherwise.
2006 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2009 if (addr < vma->vm_start || addr >= vma->vm_end)
2011 if (!page_count(page))
2013 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2014 BUG_ON(mmap_read_trylock(vma->vm_mm));
2015 BUG_ON(vma->vm_flags & VM_PFNMAP);
2016 vm_flags_set(vma, VM_MIXEDMAP);
2018 return insert_page(vma, addr, page, vma->vm_page_prot);
2020 EXPORT_SYMBOL(vm_insert_page);
2023 * __vm_map_pages - maps range of kernel pages into user vma
2024 * @vma: user vma to map to
2025 * @pages: pointer to array of source kernel pages
2026 * @num: number of pages in page array
2027 * @offset: user's requested vm_pgoff
2029 * This allows drivers to map range of kernel pages into a user vma.
2031 * Return: 0 on success and error code otherwise.
2033 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2034 unsigned long num, unsigned long offset)
2036 unsigned long count = vma_pages(vma);
2037 unsigned long uaddr = vma->vm_start;
2040 /* Fail if the user requested offset is beyond the end of the object */
2044 /* Fail if the user requested size exceeds available object size */
2045 if (count > num - offset)
2048 for (i = 0; i < count; i++) {
2049 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2059 * vm_map_pages - maps range of kernel pages starts with non zero offset
2060 * @vma: user vma to map to
2061 * @pages: pointer to array of source kernel pages
2062 * @num: number of pages in page array
2064 * Maps an object consisting of @num pages, catering for the user's
2065 * requested vm_pgoff
2067 * If we fail to insert any page into the vma, the function will return
2068 * immediately leaving any previously inserted pages present. Callers
2069 * from the mmap handler may immediately return the error as their caller
2070 * will destroy the vma, removing any successfully inserted pages. Other
2071 * callers should make their own arrangements for calling unmap_region().
2073 * Context: Process context. Called by mmap handlers.
2074 * Return: 0 on success and error code otherwise.
2076 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2079 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2081 EXPORT_SYMBOL(vm_map_pages);
2084 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2085 * @vma: user vma to map to
2086 * @pages: pointer to array of source kernel pages
2087 * @num: number of pages in page array
2089 * Similar to vm_map_pages(), except that it explicitly sets the offset
2090 * to 0. This function is intended for the drivers that did not consider
2093 * Context: Process context. Called by mmap handlers.
2094 * Return: 0 on success and error code otherwise.
2096 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2099 return __vm_map_pages(vma, pages, num, 0);
2101 EXPORT_SYMBOL(vm_map_pages_zero);
2103 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2104 pfn_t pfn, pgprot_t prot, bool mkwrite)
2106 struct mm_struct *mm = vma->vm_mm;
2110 pte = get_locked_pte(mm, addr, &ptl);
2112 return VM_FAULT_OOM;
2113 entry = ptep_get(pte);
2114 if (!pte_none(entry)) {
2117 * For read faults on private mappings the PFN passed
2118 * in may not match the PFN we have mapped if the
2119 * mapped PFN is a writeable COW page. In the mkwrite
2120 * case we are creating a writable PTE for a shared
2121 * mapping and we expect the PFNs to match. If they
2122 * don't match, we are likely racing with block
2123 * allocation and mapping invalidation so just skip the
2126 if (pte_pfn(entry) != pfn_t_to_pfn(pfn)) {
2127 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
2130 entry = pte_mkyoung(entry);
2131 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2132 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2133 update_mmu_cache(vma, addr, pte);
2138 /* Ok, finally just insert the thing.. */
2139 if (pfn_t_devmap(pfn))
2140 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2142 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2145 entry = pte_mkyoung(entry);
2146 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2149 set_pte_at(mm, addr, pte, entry);
2150 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2153 pte_unmap_unlock(pte, ptl);
2154 return VM_FAULT_NOPAGE;
2158 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2159 * @vma: user vma to map to
2160 * @addr: target user address of this page
2161 * @pfn: source kernel pfn
2162 * @pgprot: pgprot flags for the inserted page
2164 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2165 * to override pgprot on a per-page basis.
2167 * This only makes sense for IO mappings, and it makes no sense for
2168 * COW mappings. In general, using multiple vmas is preferable;
2169 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2172 * pgprot typically only differs from @vma->vm_page_prot when drivers set
2173 * caching- and encryption bits different than those of @vma->vm_page_prot,
2174 * because the caching- or encryption mode may not be known at mmap() time.
2176 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2177 * to set caching and encryption bits for those vmas (except for COW pages).
2178 * This is ensured by core vm only modifying these page table entries using
2179 * functions that don't touch caching- or encryption bits, using pte_modify()
2180 * if needed. (See for example mprotect()).
2182 * Also when new page-table entries are created, this is only done using the
2183 * fault() callback, and never using the value of vma->vm_page_prot,
2184 * except for page-table entries that point to anonymous pages as the result
2187 * Context: Process context. May allocate using %GFP_KERNEL.
2188 * Return: vm_fault_t value.
2190 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2191 unsigned long pfn, pgprot_t pgprot)
2194 * Technically, architectures with pte_special can avoid all these
2195 * restrictions (same for remap_pfn_range). However we would like
2196 * consistency in testing and feature parity among all, so we should
2197 * try to keep these invariants in place for everybody.
2199 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2200 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2201 (VM_PFNMAP|VM_MIXEDMAP));
2202 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2203 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2205 if (addr < vma->vm_start || addr >= vma->vm_end)
2206 return VM_FAULT_SIGBUS;
2208 if (!pfn_modify_allowed(pfn, pgprot))
2209 return VM_FAULT_SIGBUS;
2211 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2213 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2216 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2219 * vmf_insert_pfn - insert single pfn into user vma
2220 * @vma: user vma to map to
2221 * @addr: target user address of this page
2222 * @pfn: source kernel pfn
2224 * Similar to vm_insert_page, this allows drivers to insert individual pages
2225 * they've allocated into a user vma. Same comments apply.
2227 * This function should only be called from a vm_ops->fault handler, and
2228 * in that case the handler should return the result of this function.
2230 * vma cannot be a COW mapping.
2232 * As this is called only for pages that do not currently exist, we
2233 * do not need to flush old virtual caches or the TLB.
2235 * Context: Process context. May allocate using %GFP_KERNEL.
2236 * Return: vm_fault_t value.
2238 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2241 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2243 EXPORT_SYMBOL(vmf_insert_pfn);
2245 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2247 /* these checks mirror the abort conditions in vm_normal_page */
2248 if (vma->vm_flags & VM_MIXEDMAP)
2250 if (pfn_t_devmap(pfn))
2252 if (pfn_t_special(pfn))
2254 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2259 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2260 unsigned long addr, pfn_t pfn, bool mkwrite)
2262 pgprot_t pgprot = vma->vm_page_prot;
2265 BUG_ON(!vm_mixed_ok(vma, pfn));
2267 if (addr < vma->vm_start || addr >= vma->vm_end)
2268 return VM_FAULT_SIGBUS;
2270 track_pfn_insert(vma, &pgprot, pfn);
2272 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2273 return VM_FAULT_SIGBUS;
2276 * If we don't have pte special, then we have to use the pfn_valid()
2277 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2278 * refcount the page if pfn_valid is true (hence insert_page rather
2279 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2280 * without pte special, it would there be refcounted as a normal page.
2282 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2283 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2287 * At this point we are committed to insert_page()
2288 * regardless of whether the caller specified flags that
2289 * result in pfn_t_has_page() == false.
2291 page = pfn_to_page(pfn_t_to_pfn(pfn));
2292 err = insert_page(vma, addr, page, pgprot);
2294 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2298 return VM_FAULT_OOM;
2299 if (err < 0 && err != -EBUSY)
2300 return VM_FAULT_SIGBUS;
2302 return VM_FAULT_NOPAGE;
2305 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2308 return __vm_insert_mixed(vma, addr, pfn, false);
2310 EXPORT_SYMBOL(vmf_insert_mixed);
2313 * If the insertion of PTE failed because someone else already added a
2314 * different entry in the mean time, we treat that as success as we assume
2315 * the same entry was actually inserted.
2317 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2318 unsigned long addr, pfn_t pfn)
2320 return __vm_insert_mixed(vma, addr, pfn, true);
2322 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2325 * maps a range of physical memory into the requested pages. the old
2326 * mappings are removed. any references to nonexistent pages results
2327 * in null mappings (currently treated as "copy-on-access")
2329 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2330 unsigned long addr, unsigned long end,
2331 unsigned long pfn, pgprot_t prot)
2333 pte_t *pte, *mapped_pte;
2337 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2340 arch_enter_lazy_mmu_mode();
2342 BUG_ON(!pte_none(ptep_get(pte)));
2343 if (!pfn_modify_allowed(pfn, prot)) {
2347 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2349 } while (pte++, addr += PAGE_SIZE, addr != end);
2350 arch_leave_lazy_mmu_mode();
2351 pte_unmap_unlock(mapped_pte, ptl);
2355 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2356 unsigned long addr, unsigned long end,
2357 unsigned long pfn, pgprot_t prot)
2363 pfn -= addr >> PAGE_SHIFT;
2364 pmd = pmd_alloc(mm, pud, addr);
2367 VM_BUG_ON(pmd_trans_huge(*pmd));
2369 next = pmd_addr_end(addr, end);
2370 err = remap_pte_range(mm, pmd, addr, next,
2371 pfn + (addr >> PAGE_SHIFT), prot);
2374 } while (pmd++, addr = next, addr != end);
2378 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2379 unsigned long addr, unsigned long end,
2380 unsigned long pfn, pgprot_t prot)
2386 pfn -= addr >> PAGE_SHIFT;
2387 pud = pud_alloc(mm, p4d, addr);
2391 next = pud_addr_end(addr, end);
2392 err = remap_pmd_range(mm, pud, addr, next,
2393 pfn + (addr >> PAGE_SHIFT), prot);
2396 } while (pud++, addr = next, addr != end);
2400 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2401 unsigned long addr, unsigned long end,
2402 unsigned long pfn, pgprot_t prot)
2408 pfn -= addr >> PAGE_SHIFT;
2409 p4d = p4d_alloc(mm, pgd, addr);
2413 next = p4d_addr_end(addr, end);
2414 err = remap_pud_range(mm, p4d, addr, next,
2415 pfn + (addr >> PAGE_SHIFT), prot);
2418 } while (p4d++, addr = next, addr != end);
2423 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2424 * must have pre-validated the caching bits of the pgprot_t.
2426 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2427 unsigned long pfn, unsigned long size, pgprot_t prot)
2431 unsigned long end = addr + PAGE_ALIGN(size);
2432 struct mm_struct *mm = vma->vm_mm;
2435 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2439 * Physically remapped pages are special. Tell the
2440 * rest of the world about it:
2441 * VM_IO tells people not to look at these pages
2442 * (accesses can have side effects).
2443 * VM_PFNMAP tells the core MM that the base pages are just
2444 * raw PFN mappings, and do not have a "struct page" associated
2447 * Disable vma merging and expanding with mremap().
2449 * Omit vma from core dump, even when VM_IO turned off.
2451 * There's a horrible special case to handle copy-on-write
2452 * behaviour that some programs depend on. We mark the "original"
2453 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2454 * See vm_normal_page() for details.
2456 if (is_cow_mapping(vma->vm_flags)) {
2457 if (addr != vma->vm_start || end != vma->vm_end)
2459 vma->vm_pgoff = pfn;
2462 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
2464 BUG_ON(addr >= end);
2465 pfn -= addr >> PAGE_SHIFT;
2466 pgd = pgd_offset(mm, addr);
2467 flush_cache_range(vma, addr, end);
2469 next = pgd_addr_end(addr, end);
2470 err = remap_p4d_range(mm, pgd, addr, next,
2471 pfn + (addr >> PAGE_SHIFT), prot);
2474 } while (pgd++, addr = next, addr != end);
2480 * remap_pfn_range - remap kernel memory to userspace
2481 * @vma: user vma to map to
2482 * @addr: target page aligned user address to start at
2483 * @pfn: page frame number of kernel physical memory address
2484 * @size: size of mapping area
2485 * @prot: page protection flags for this mapping
2487 * Note: this is only safe if the mm semaphore is held when called.
2489 * Return: %0 on success, negative error code otherwise.
2491 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2492 unsigned long pfn, unsigned long size, pgprot_t prot)
2496 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2500 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2502 untrack_pfn(vma, pfn, PAGE_ALIGN(size), true);
2505 EXPORT_SYMBOL(remap_pfn_range);
2508 * vm_iomap_memory - remap memory to userspace
2509 * @vma: user vma to map to
2510 * @start: start of the physical memory to be mapped
2511 * @len: size of area
2513 * This is a simplified io_remap_pfn_range() for common driver use. The
2514 * driver just needs to give us the physical memory range to be mapped,
2515 * we'll figure out the rest from the vma information.
2517 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2518 * whatever write-combining details or similar.
2520 * Return: %0 on success, negative error code otherwise.
2522 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2524 unsigned long vm_len, pfn, pages;
2526 /* Check that the physical memory area passed in looks valid */
2527 if (start + len < start)
2530 * You *really* shouldn't map things that aren't page-aligned,
2531 * but we've historically allowed it because IO memory might
2532 * just have smaller alignment.
2534 len += start & ~PAGE_MASK;
2535 pfn = start >> PAGE_SHIFT;
2536 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2537 if (pfn + pages < pfn)
2540 /* We start the mapping 'vm_pgoff' pages into the area */
2541 if (vma->vm_pgoff > pages)
2543 pfn += vma->vm_pgoff;
2544 pages -= vma->vm_pgoff;
2546 /* Can we fit all of the mapping? */
2547 vm_len = vma->vm_end - vma->vm_start;
2548 if (vm_len >> PAGE_SHIFT > pages)
2551 /* Ok, let it rip */
2552 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2554 EXPORT_SYMBOL(vm_iomap_memory);
2556 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2557 unsigned long addr, unsigned long end,
2558 pte_fn_t fn, void *data, bool create,
2559 pgtbl_mod_mask *mask)
2561 pte_t *pte, *mapped_pte;
2566 mapped_pte = pte = (mm == &init_mm) ?
2567 pte_alloc_kernel_track(pmd, addr, mask) :
2568 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2572 mapped_pte = pte = (mm == &init_mm) ?
2573 pte_offset_kernel(pmd, addr) :
2574 pte_offset_map_lock(mm, pmd, addr, &ptl);
2579 arch_enter_lazy_mmu_mode();
2583 if (create || !pte_none(ptep_get(pte))) {
2584 err = fn(pte++, addr, data);
2588 } while (addr += PAGE_SIZE, addr != end);
2590 *mask |= PGTBL_PTE_MODIFIED;
2592 arch_leave_lazy_mmu_mode();
2595 pte_unmap_unlock(mapped_pte, ptl);
2599 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2600 unsigned long addr, unsigned long end,
2601 pte_fn_t fn, void *data, bool create,
2602 pgtbl_mod_mask *mask)
2608 BUG_ON(pud_huge(*pud));
2611 pmd = pmd_alloc_track(mm, pud, addr, mask);
2615 pmd = pmd_offset(pud, addr);
2618 next = pmd_addr_end(addr, end);
2619 if (pmd_none(*pmd) && !create)
2621 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2623 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2628 err = apply_to_pte_range(mm, pmd, addr, next,
2629 fn, data, create, mask);
2632 } while (pmd++, addr = next, addr != end);
2637 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2638 unsigned long addr, unsigned long end,
2639 pte_fn_t fn, void *data, bool create,
2640 pgtbl_mod_mask *mask)
2647 pud = pud_alloc_track(mm, p4d, addr, mask);
2651 pud = pud_offset(p4d, addr);
2654 next = pud_addr_end(addr, end);
2655 if (pud_none(*pud) && !create)
2657 if (WARN_ON_ONCE(pud_leaf(*pud)))
2659 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2664 err = apply_to_pmd_range(mm, pud, addr, next,
2665 fn, data, create, mask);
2668 } while (pud++, addr = next, addr != end);
2673 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2674 unsigned long addr, unsigned long end,
2675 pte_fn_t fn, void *data, bool create,
2676 pgtbl_mod_mask *mask)
2683 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2687 p4d = p4d_offset(pgd, addr);
2690 next = p4d_addr_end(addr, end);
2691 if (p4d_none(*p4d) && !create)
2693 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2695 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2700 err = apply_to_pud_range(mm, p4d, addr, next,
2701 fn, data, create, mask);
2704 } while (p4d++, addr = next, addr != end);
2709 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2710 unsigned long size, pte_fn_t fn,
2711 void *data, bool create)
2714 unsigned long start = addr, next;
2715 unsigned long end = addr + size;
2716 pgtbl_mod_mask mask = 0;
2719 if (WARN_ON(addr >= end))
2722 pgd = pgd_offset(mm, addr);
2724 next = pgd_addr_end(addr, end);
2725 if (pgd_none(*pgd) && !create)
2727 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2729 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2734 err = apply_to_p4d_range(mm, pgd, addr, next,
2735 fn, data, create, &mask);
2738 } while (pgd++, addr = next, addr != end);
2740 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2741 arch_sync_kernel_mappings(start, start + size);
2747 * Scan a region of virtual memory, filling in page tables as necessary
2748 * and calling a provided function on each leaf page table.
2750 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2751 unsigned long size, pte_fn_t fn, void *data)
2753 return __apply_to_page_range(mm, addr, size, fn, data, true);
2755 EXPORT_SYMBOL_GPL(apply_to_page_range);
2758 * Scan a region of virtual memory, calling a provided function on
2759 * each leaf page table where it exists.
2761 * Unlike apply_to_page_range, this does _not_ fill in page tables
2762 * where they are absent.
2764 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2765 unsigned long size, pte_fn_t fn, void *data)
2767 return __apply_to_page_range(mm, addr, size, fn, data, false);
2769 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2772 * handle_pte_fault chooses page fault handler according to an entry which was
2773 * read non-atomically. Before making any commitment, on those architectures
2774 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2775 * parts, do_swap_page must check under lock before unmapping the pte and
2776 * proceeding (but do_wp_page is only called after already making such a check;
2777 * and do_anonymous_page can safely check later on).
2779 static inline int pte_unmap_same(struct vm_fault *vmf)
2782 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2783 if (sizeof(pte_t) > sizeof(unsigned long)) {
2784 spin_lock(vmf->ptl);
2785 same = pte_same(ptep_get(vmf->pte), vmf->orig_pte);
2786 spin_unlock(vmf->ptl);
2789 pte_unmap(vmf->pte);
2796 * 0: copied succeeded
2797 * -EHWPOISON: copy failed due to hwpoison in source page
2798 * -EAGAIN: copied failed (some other reason)
2800 static inline int __wp_page_copy_user(struct page *dst, struct page *src,
2801 struct vm_fault *vmf)
2806 struct vm_area_struct *vma = vmf->vma;
2807 struct mm_struct *mm = vma->vm_mm;
2808 unsigned long addr = vmf->address;
2811 if (copy_mc_user_highpage(dst, src, addr, vma)) {
2812 memory_failure_queue(page_to_pfn(src), 0);
2819 * If the source page was a PFN mapping, we don't have
2820 * a "struct page" for it. We do a best-effort copy by
2821 * just copying from the original user address. If that
2822 * fails, we just zero-fill it. Live with it.
2824 kaddr = kmap_atomic(dst);
2825 uaddr = (void __user *)(addr & PAGE_MASK);
2828 * On architectures with software "accessed" bits, we would
2829 * take a double page fault, so mark it accessed here.
2832 if (!arch_has_hw_pte_young() && !pte_young(vmf->orig_pte)) {
2835 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2836 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2838 * Other thread has already handled the fault
2839 * and update local tlb only
2842 update_mmu_tlb(vma, addr, vmf->pte);
2847 entry = pte_mkyoung(vmf->orig_pte);
2848 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2849 update_mmu_cache_range(vmf, vma, addr, vmf->pte, 1);
2853 * This really shouldn't fail, because the page is there
2854 * in the page tables. But it might just be unreadable,
2855 * in which case we just give up and fill the result with
2858 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2862 /* Re-validate under PTL if the page is still mapped */
2863 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2864 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
2865 /* The PTE changed under us, update local tlb */
2867 update_mmu_tlb(vma, addr, vmf->pte);
2873 * The same page can be mapped back since last copy attempt.
2874 * Try to copy again under PTL.
2876 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2878 * Give a warn in case there can be some obscure
2891 pte_unmap_unlock(vmf->pte, vmf->ptl);
2892 kunmap_atomic(kaddr);
2893 flush_dcache_page(dst);
2898 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2900 struct file *vm_file = vma->vm_file;
2903 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2906 * Special mappings (e.g. VDSO) do not have any file so fake
2907 * a default GFP_KERNEL for them.
2913 * Notify the address space that the page is about to become writable so that
2914 * it can prohibit this or wait for the page to get into an appropriate state.
2916 * We do this without the lock held, so that it can sleep if it needs to.
2918 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf, struct folio *folio)
2921 unsigned int old_flags = vmf->flags;
2923 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2925 if (vmf->vma->vm_file &&
2926 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2927 return VM_FAULT_SIGBUS;
2929 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2930 /* Restore original flags so that caller is not surprised */
2931 vmf->flags = old_flags;
2932 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2934 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2936 if (!folio->mapping) {
2937 folio_unlock(folio);
2938 return 0; /* retry */
2940 ret |= VM_FAULT_LOCKED;
2942 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2947 * Handle dirtying of a page in shared file mapping on a write fault.
2949 * The function expects the page to be locked and unlocks it.
2951 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2953 struct vm_area_struct *vma = vmf->vma;
2954 struct address_space *mapping;
2955 struct folio *folio = page_folio(vmf->page);
2957 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2959 dirtied = folio_mark_dirty(folio);
2960 VM_BUG_ON_FOLIO(folio_test_anon(folio), folio);
2962 * Take a local copy of the address_space - folio.mapping may be zeroed
2963 * by truncate after folio_unlock(). The address_space itself remains
2964 * pinned by vma->vm_file's reference. We rely on folio_unlock()'s
2965 * release semantics to prevent the compiler from undoing this copying.
2967 mapping = folio_raw_mapping(folio);
2968 folio_unlock(folio);
2971 file_update_time(vma->vm_file);
2974 * Throttle page dirtying rate down to writeback speed.
2976 * mapping may be NULL here because some device drivers do not
2977 * set page.mapping but still dirty their pages
2979 * Drop the mmap_lock before waiting on IO, if we can. The file
2980 * is pinning the mapping, as per above.
2982 if ((dirtied || page_mkwrite) && mapping) {
2985 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2986 balance_dirty_pages_ratelimited(mapping);
2989 return VM_FAULT_COMPLETED;
2997 * Handle write page faults for pages that can be reused in the current vma
2999 * This can happen either due to the mapping being with the VM_SHARED flag,
3000 * or due to us being the last reference standing to the page. In either
3001 * case, all we need to do here is to mark the page as writable and update
3002 * any related book-keeping.
3004 static inline void wp_page_reuse(struct vm_fault *vmf)
3005 __releases(vmf->ptl)
3007 struct vm_area_struct *vma = vmf->vma;
3008 struct page *page = vmf->page;
3011 VM_BUG_ON(!(vmf->flags & FAULT_FLAG_WRITE));
3012 VM_BUG_ON(page && PageAnon(page) && !PageAnonExclusive(page));
3015 * Clear the pages cpupid information as the existing
3016 * information potentially belongs to a now completely
3017 * unrelated process.
3020 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
3022 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3023 entry = pte_mkyoung(vmf->orig_pte);
3024 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3025 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
3026 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3027 pte_unmap_unlock(vmf->pte, vmf->ptl);
3028 count_vm_event(PGREUSE);
3032 * Handle the case of a page which we actually need to copy to a new page,
3033 * either due to COW or unsharing.
3035 * Called with mmap_lock locked and the old page referenced, but
3036 * without the ptl held.
3038 * High level logic flow:
3040 * - Allocate a page, copy the content of the old page to the new one.
3041 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
3042 * - Take the PTL. If the pte changed, bail out and release the allocated page
3043 * - If the pte is still the way we remember it, update the page table and all
3044 * relevant references. This includes dropping the reference the page-table
3045 * held to the old page, as well as updating the rmap.
3046 * - In any case, unlock the PTL and drop the reference we took to the old page.
3048 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3050 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3051 struct vm_area_struct *vma = vmf->vma;
3052 struct mm_struct *mm = vma->vm_mm;
3053 struct folio *old_folio = NULL;
3054 struct folio *new_folio = NULL;
3056 int page_copied = 0;
3057 struct mmu_notifier_range range;
3060 delayacct_wpcopy_start();
3063 old_folio = page_folio(vmf->page);
3064 if (unlikely(anon_vma_prepare(vma)))
3067 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
3068 new_folio = vma_alloc_zeroed_movable_folio(vma, vmf->address);
3072 new_folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0, vma,
3073 vmf->address, false);
3077 ret = __wp_page_copy_user(&new_folio->page, vmf->page, vmf);
3080 * COW failed, if the fault was solved by other,
3081 * it's fine. If not, userspace would re-fault on
3082 * the same address and we will handle the fault
3083 * from the second attempt.
3084 * The -EHWPOISON case will not be retried.
3086 folio_put(new_folio);
3088 folio_put(old_folio);
3090 delayacct_wpcopy_end();
3091 return ret == -EHWPOISON ? VM_FAULT_HWPOISON : 0;
3093 kmsan_copy_page_meta(&new_folio->page, vmf->page);
3096 if (mem_cgroup_charge(new_folio, mm, GFP_KERNEL))
3098 folio_throttle_swaprate(new_folio, GFP_KERNEL);
3100 __folio_mark_uptodate(new_folio);
3102 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
3103 vmf->address & PAGE_MASK,
3104 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3105 mmu_notifier_invalidate_range_start(&range);
3108 * Re-check the pte - we dropped the lock
3110 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3111 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
3113 if (!folio_test_anon(old_folio)) {
3114 dec_mm_counter(mm, mm_counter_file(&old_folio->page));
3115 inc_mm_counter(mm, MM_ANONPAGES);
3118 ksm_might_unmap_zero_page(mm, vmf->orig_pte);
3119 inc_mm_counter(mm, MM_ANONPAGES);
3121 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3122 entry = mk_pte(&new_folio->page, vma->vm_page_prot);
3123 entry = pte_sw_mkyoung(entry);
3124 if (unlikely(unshare)) {
3125 if (pte_soft_dirty(vmf->orig_pte))
3126 entry = pte_mksoft_dirty(entry);
3127 if (pte_uffd_wp(vmf->orig_pte))
3128 entry = pte_mkuffd_wp(entry);
3130 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3134 * Clear the pte entry and flush it first, before updating the
3135 * pte with the new entry, to keep TLBs on different CPUs in
3136 * sync. This code used to set the new PTE then flush TLBs, but
3137 * that left a window where the new PTE could be loaded into
3138 * some TLBs while the old PTE remains in others.
3140 ptep_clear_flush(vma, vmf->address, vmf->pte);
3141 folio_add_new_anon_rmap(new_folio, vma, vmf->address);
3142 folio_add_lru_vma(new_folio, vma);
3144 * We call the notify macro here because, when using secondary
3145 * mmu page tables (such as kvm shadow page tables), we want the
3146 * new page to be mapped directly into the secondary page table.
3148 BUG_ON(unshare && pte_write(entry));
3149 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3150 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
3153 * Only after switching the pte to the new page may
3154 * we remove the mapcount here. Otherwise another
3155 * process may come and find the rmap count decremented
3156 * before the pte is switched to the new page, and
3157 * "reuse" the old page writing into it while our pte
3158 * here still points into it and can be read by other
3161 * The critical issue is to order this
3162 * page_remove_rmap with the ptp_clear_flush above.
3163 * Those stores are ordered by (if nothing else,)
3164 * the barrier present in the atomic_add_negative
3165 * in page_remove_rmap.
3167 * Then the TLB flush in ptep_clear_flush ensures that
3168 * no process can access the old page before the
3169 * decremented mapcount is visible. And the old page
3170 * cannot be reused until after the decremented
3171 * mapcount is visible. So transitively, TLBs to
3172 * old page will be flushed before it can be reused.
3174 page_remove_rmap(vmf->page, vma, false);
3177 /* Free the old page.. */
3178 new_folio = old_folio;
3180 pte_unmap_unlock(vmf->pte, vmf->ptl);
3181 } else if (vmf->pte) {
3182 update_mmu_tlb(vma, vmf->address, vmf->pte);
3183 pte_unmap_unlock(vmf->pte, vmf->ptl);
3186 mmu_notifier_invalidate_range_end(&range);
3189 folio_put(new_folio);
3192 free_swap_cache(&old_folio->page);
3193 folio_put(old_folio);
3196 delayacct_wpcopy_end();
3199 folio_put(new_folio);
3202 folio_put(old_folio);
3204 delayacct_wpcopy_end();
3205 return VM_FAULT_OOM;
3209 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3210 * writeable once the page is prepared
3212 * @vmf: structure describing the fault
3214 * This function handles all that is needed to finish a write page fault in a
3215 * shared mapping due to PTE being read-only once the mapped page is prepared.
3216 * It handles locking of PTE and modifying it.
3218 * The function expects the page to be locked or other protection against
3219 * concurrent faults / writeback (such as DAX radix tree locks).
3221 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3222 * we acquired PTE lock.
3224 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3226 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3227 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3230 return VM_FAULT_NOPAGE;
3232 * We might have raced with another page fault while we released the
3233 * pte_offset_map_lock.
3235 if (!pte_same(ptep_get(vmf->pte), vmf->orig_pte)) {
3236 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3237 pte_unmap_unlock(vmf->pte, vmf->ptl);
3238 return VM_FAULT_NOPAGE;
3245 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3248 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3250 struct vm_area_struct *vma = vmf->vma;
3252 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3255 pte_unmap_unlock(vmf->pte, vmf->ptl);
3256 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3257 vma_end_read(vmf->vma);
3258 return VM_FAULT_RETRY;
3261 vmf->flags |= FAULT_FLAG_MKWRITE;
3262 ret = vma->vm_ops->pfn_mkwrite(vmf);
3263 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3265 return finish_mkwrite_fault(vmf);
3271 static vm_fault_t wp_page_shared(struct vm_fault *vmf, struct folio *folio)
3272 __releases(vmf->ptl)
3274 struct vm_area_struct *vma = vmf->vma;
3279 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3282 pte_unmap_unlock(vmf->pte, vmf->ptl);
3283 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3285 vma_end_read(vmf->vma);
3286 return VM_FAULT_RETRY;
3289 tmp = do_page_mkwrite(vmf, folio);
3290 if (unlikely(!tmp || (tmp &
3291 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3295 tmp = finish_mkwrite_fault(vmf);
3296 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3297 folio_unlock(folio);
3305 ret |= fault_dirty_shared_page(vmf);
3312 * This routine handles present pages, when
3313 * * users try to write to a shared page (FAULT_FLAG_WRITE)
3314 * * GUP wants to take a R/O pin on a possibly shared anonymous page
3315 * (FAULT_FLAG_UNSHARE)
3317 * It is done by copying the page to a new address and decrementing the
3318 * shared-page counter for the old page.
3320 * Note that this routine assumes that the protection checks have been
3321 * done by the caller (the low-level page fault routine in most cases).
3322 * Thus, with FAULT_FLAG_WRITE, we can safely just mark it writable once we've
3323 * done any necessary COW.
3325 * In case of FAULT_FLAG_WRITE, we also mark the page dirty at this point even
3326 * though the page will change only once the write actually happens. This
3327 * avoids a few races, and potentially makes it more efficient.
3329 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3330 * but allow concurrent faults), with pte both mapped and locked.
3331 * We return with mmap_lock still held, but pte unmapped and unlocked.
3333 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3334 __releases(vmf->ptl)
3336 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
3337 struct vm_area_struct *vma = vmf->vma;
3338 struct folio *folio = NULL;
3340 if (likely(!unshare)) {
3341 if (userfaultfd_pte_wp(vma, ptep_get(vmf->pte))) {
3342 pte_unmap_unlock(vmf->pte, vmf->ptl);
3343 return handle_userfault(vmf, VM_UFFD_WP);
3347 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3348 * is flushed in this case before copying.
3350 if (unlikely(userfaultfd_wp(vmf->vma) &&
3351 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3352 flush_tlb_page(vmf->vma, vmf->address);
3355 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3358 folio = page_folio(vmf->page);
3361 * Shared mapping: we are guaranteed to have VM_WRITE and
3362 * FAULT_FLAG_WRITE set at this point.
3364 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
3366 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3369 * We should not cow pages in a shared writeable mapping.
3370 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3373 return wp_pfn_shared(vmf);
3374 return wp_page_shared(vmf, folio);
3378 * Private mapping: create an exclusive anonymous page copy if reuse
3379 * is impossible. We might miss VM_WRITE for FOLL_FORCE handling.
3381 if (folio && folio_test_anon(folio)) {
3383 * If the page is exclusive to this process we must reuse the
3384 * page without further checks.
3386 if (PageAnonExclusive(vmf->page))
3390 * We have to verify under folio lock: these early checks are
3391 * just an optimization to avoid locking the folio and freeing
3392 * the swapcache if there is little hope that we can reuse.
3394 * KSM doesn't necessarily raise the folio refcount.
3396 if (folio_test_ksm(folio) || folio_ref_count(folio) > 3)
3398 if (!folio_test_lru(folio))
3400 * We cannot easily detect+handle references from
3401 * remote LRU caches or references to LRU folios.
3404 if (folio_ref_count(folio) > 1 + folio_test_swapcache(folio))
3406 if (!folio_trylock(folio))
3408 if (folio_test_swapcache(folio))
3409 folio_free_swap(folio);
3410 if (folio_test_ksm(folio) || folio_ref_count(folio) != 1) {
3411 folio_unlock(folio);
3415 * Ok, we've got the only folio reference from our mapping
3416 * and the folio is locked, it's dark out, and we're wearing
3417 * sunglasses. Hit it.
3419 page_move_anon_rmap(vmf->page, vma);
3420 folio_unlock(folio);
3422 if (unlikely(unshare)) {
3423 pte_unmap_unlock(vmf->pte, vmf->ptl);
3430 if ((vmf->flags & FAULT_FLAG_VMA_LOCK) && !vma->anon_vma) {
3431 pte_unmap_unlock(vmf->pte, vmf->ptl);
3432 vma_end_read(vmf->vma);
3433 return VM_FAULT_RETRY;
3437 * Ok, we need to copy. Oh, well..
3442 pte_unmap_unlock(vmf->pte, vmf->ptl);
3444 if (folio && folio_test_ksm(folio))
3445 count_vm_event(COW_KSM);
3447 return wp_page_copy(vmf);
3450 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3451 unsigned long start_addr, unsigned long end_addr,
3452 struct zap_details *details)
3454 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3457 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3458 pgoff_t first_index,
3460 struct zap_details *details)
3462 struct vm_area_struct *vma;
3463 pgoff_t vba, vea, zba, zea;
3465 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3466 vba = vma->vm_pgoff;
3467 vea = vba + vma_pages(vma) - 1;
3468 zba = max(first_index, vba);
3469 zea = min(last_index, vea);
3471 unmap_mapping_range_vma(vma,
3472 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3473 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3479 * unmap_mapping_folio() - Unmap single folio from processes.
3480 * @folio: The locked folio to be unmapped.
3482 * Unmap this folio from any userspace process which still has it mmaped.
3483 * Typically, for efficiency, the range of nearby pages has already been
3484 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3485 * truncation or invalidation holds the lock on a folio, it may find that
3486 * the page has been remapped again: and then uses unmap_mapping_folio()
3487 * to unmap it finally.
3489 void unmap_mapping_folio(struct folio *folio)
3491 struct address_space *mapping = folio->mapping;
3492 struct zap_details details = { };
3493 pgoff_t first_index;
3496 VM_BUG_ON(!folio_test_locked(folio));
3498 first_index = folio->index;
3499 last_index = folio_next_index(folio) - 1;
3501 details.even_cows = false;
3502 details.single_folio = folio;
3503 details.zap_flags = ZAP_FLAG_DROP_MARKER;
3505 i_mmap_lock_read(mapping);
3506 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3507 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3508 last_index, &details);
3509 i_mmap_unlock_read(mapping);
3513 * unmap_mapping_pages() - Unmap pages from processes.
3514 * @mapping: The address space containing pages to be unmapped.
3515 * @start: Index of first page to be unmapped.
3516 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3517 * @even_cows: Whether to unmap even private COWed pages.
3519 * Unmap the pages in this address space from any userspace process which
3520 * has them mmaped. Generally, you want to remove COWed pages as well when
3521 * a file is being truncated, but not when invalidating pages from the page
3524 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3525 pgoff_t nr, bool even_cows)
3527 struct zap_details details = { };
3528 pgoff_t first_index = start;
3529 pgoff_t last_index = start + nr - 1;
3531 details.even_cows = even_cows;
3532 if (last_index < first_index)
3533 last_index = ULONG_MAX;
3535 i_mmap_lock_read(mapping);
3536 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3537 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3538 last_index, &details);
3539 i_mmap_unlock_read(mapping);
3541 EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3544 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3545 * address_space corresponding to the specified byte range in the underlying
3548 * @mapping: the address space containing mmaps to be unmapped.
3549 * @holebegin: byte in first page to unmap, relative to the start of
3550 * the underlying file. This will be rounded down to a PAGE_SIZE
3551 * boundary. Note that this is different from truncate_pagecache(), which
3552 * must keep the partial page. In contrast, we must get rid of
3554 * @holelen: size of prospective hole in bytes. This will be rounded
3555 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3557 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3558 * but 0 when invalidating pagecache, don't throw away private data.
3560 void unmap_mapping_range(struct address_space *mapping,
3561 loff_t const holebegin, loff_t const holelen, int even_cows)
3563 pgoff_t hba = holebegin >> PAGE_SHIFT;
3564 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3566 /* Check for overflow. */
3567 if (sizeof(holelen) > sizeof(hlen)) {
3569 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3570 if (holeend & ~(long long)ULONG_MAX)
3571 hlen = ULONG_MAX - hba + 1;
3574 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3576 EXPORT_SYMBOL(unmap_mapping_range);
3579 * Restore a potential device exclusive pte to a working pte entry
3581 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3583 struct folio *folio = page_folio(vmf->page);
3584 struct vm_area_struct *vma = vmf->vma;
3585 struct mmu_notifier_range range;
3589 * We need a reference to lock the folio because we don't hold
3590 * the PTL so a racing thread can remove the device-exclusive
3591 * entry and unmap it. If the folio is free the entry must
3592 * have been removed already. If it happens to have already
3593 * been re-allocated after being freed all we do is lock and
3596 if (!folio_try_get(folio))
3599 ret = folio_lock_or_retry(folio, vmf);
3604 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
3605 vma->vm_mm, vmf->address & PAGE_MASK,
3606 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3607 mmu_notifier_invalidate_range_start(&range);
3609 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3611 if (likely(vmf->pte && pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3612 restore_exclusive_pte(vma, vmf->page, vmf->address, vmf->pte);
3615 pte_unmap_unlock(vmf->pte, vmf->ptl);
3616 folio_unlock(folio);
3619 mmu_notifier_invalidate_range_end(&range);
3623 static inline bool should_try_to_free_swap(struct folio *folio,
3624 struct vm_area_struct *vma,
3625 unsigned int fault_flags)
3627 if (!folio_test_swapcache(folio))
3629 if (mem_cgroup_swap_full(folio) || (vma->vm_flags & VM_LOCKED) ||
3630 folio_test_mlocked(folio))
3633 * If we want to map a page that's in the swapcache writable, we
3634 * have to detect via the refcount if we're really the exclusive
3635 * user. Try freeing the swapcache to get rid of the swapcache
3636 * reference only in case it's likely that we'll be the exlusive user.
3638 return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
3639 folio_ref_count(folio) == 2;
3642 static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
3644 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
3645 vmf->address, &vmf->ptl);
3649 * Be careful so that we will only recover a special uffd-wp pte into a
3650 * none pte. Otherwise it means the pte could have changed, so retry.
3652 * This should also cover the case where e.g. the pte changed
3653 * quickly from a PTE_MARKER_UFFD_WP into PTE_MARKER_POISONED.
3654 * So is_pte_marker() check is not enough to safely drop the pte.
3656 if (pte_same(vmf->orig_pte, ptep_get(vmf->pte)))
3657 pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
3658 pte_unmap_unlock(vmf->pte, vmf->ptl);
3662 static vm_fault_t do_pte_missing(struct vm_fault *vmf)
3664 if (vma_is_anonymous(vmf->vma))
3665 return do_anonymous_page(vmf);
3667 return do_fault(vmf);
3671 * This is actually a page-missing access, but with uffd-wp special pte
3672 * installed. It means this pte was wr-protected before being unmapped.
3674 static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
3677 * Just in case there're leftover special ptes even after the region
3678 * got unregistered - we can simply clear them.
3680 if (unlikely(!userfaultfd_wp(vmf->vma)))
3681 return pte_marker_clear(vmf);
3683 return do_pte_missing(vmf);
3686 static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
3688 swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
3689 unsigned long marker = pte_marker_get(entry);
3692 * PTE markers should never be empty. If anything weird happened,
3693 * the best thing to do is to kill the process along with its mm.
3695 if (WARN_ON_ONCE(!marker))
3696 return VM_FAULT_SIGBUS;
3698 /* Higher priority than uffd-wp when data corrupted */
3699 if (marker & PTE_MARKER_POISONED)
3700 return VM_FAULT_HWPOISON;
3702 if (pte_marker_entry_uffd_wp(entry))
3703 return pte_marker_handle_uffd_wp(vmf);
3705 /* This is an unknown pte marker */
3706 return VM_FAULT_SIGBUS;
3710 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3711 * but allow concurrent faults), and pte mapped but not yet locked.
3712 * We return with pte unmapped and unlocked.
3714 * We return with the mmap_lock locked or unlocked in the same cases
3715 * as does filemap_fault().
3717 vm_fault_t do_swap_page(struct vm_fault *vmf)
3719 struct vm_area_struct *vma = vmf->vma;
3720 struct folio *swapcache, *folio = NULL;
3722 struct swap_info_struct *si = NULL;
3723 rmap_t rmap_flags = RMAP_NONE;
3724 bool exclusive = false;
3728 void *shadow = NULL;
3730 if (!pte_unmap_same(vmf))
3733 entry = pte_to_swp_entry(vmf->orig_pte);
3734 if (unlikely(non_swap_entry(entry))) {
3735 if (is_migration_entry(entry)) {
3736 migration_entry_wait(vma->vm_mm, vmf->pmd,
3738 } else if (is_device_exclusive_entry(entry)) {
3739 vmf->page = pfn_swap_entry_to_page(entry);
3740 ret = remove_device_exclusive_entry(vmf);
3741 } else if (is_device_private_entry(entry)) {
3742 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
3744 * migrate_to_ram is not yet ready to operate
3748 ret = VM_FAULT_RETRY;
3752 vmf->page = pfn_swap_entry_to_page(entry);
3753 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3754 vmf->address, &vmf->ptl);
3755 if (unlikely(!vmf->pte ||
3756 !pte_same(ptep_get(vmf->pte),
3761 * Get a page reference while we know the page can't be
3764 get_page(vmf->page);
3765 pte_unmap_unlock(vmf->pte, vmf->ptl);
3766 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3767 put_page(vmf->page);
3768 } else if (is_hwpoison_entry(entry)) {
3769 ret = VM_FAULT_HWPOISON;
3770 } else if (is_pte_marker_entry(entry)) {
3771 ret = handle_pte_marker(vmf);
3773 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3774 ret = VM_FAULT_SIGBUS;
3779 /* Prevent swapoff from happening to us. */
3780 si = get_swap_device(entry);
3784 folio = swap_cache_get_folio(entry, vma, vmf->address);
3786 page = folio_file_page(folio, swp_offset(entry));
3790 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3791 __swap_count(entry) == 1) {
3792 /* skip swapcache */
3793 folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE, 0,
3794 vma, vmf->address, false);
3795 page = &folio->page;
3797 __folio_set_locked(folio);
3798 __folio_set_swapbacked(folio);
3800 if (mem_cgroup_swapin_charge_folio(folio,
3801 vma->vm_mm, GFP_KERNEL,
3806 mem_cgroup_swapin_uncharge_swap(entry);
3808 shadow = get_shadow_from_swap_cache(entry);
3810 workingset_refault(folio, shadow);
3812 folio_add_lru(folio);
3814 /* To provide entry to swap_readpage() */
3815 folio->swap = entry;
3816 swap_readpage(page, true, NULL);
3817 folio->private = NULL;
3820 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3823 folio = page_folio(page);
3829 * Back out if somebody else faulted in this pte
3830 * while we released the pte lock.
3832 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3833 vmf->address, &vmf->ptl);
3834 if (likely(vmf->pte &&
3835 pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3840 /* Had to read the page from swap area: Major fault */
3841 ret = VM_FAULT_MAJOR;
3842 count_vm_event(PGMAJFAULT);
3843 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3844 } else if (PageHWPoison(page)) {
3846 * hwpoisoned dirty swapcache pages are kept for killing
3847 * owner processes (which may be unknown at hwpoison time)
3849 ret = VM_FAULT_HWPOISON;
3853 ret |= folio_lock_or_retry(folio, vmf);
3854 if (ret & VM_FAULT_RETRY)
3859 * Make sure folio_free_swap() or swapoff did not release the
3860 * swapcache from under us. The page pin, and pte_same test
3861 * below, are not enough to exclude that. Even if it is still
3862 * swapcache, we need to check that the page's swap has not
3865 if (unlikely(!folio_test_swapcache(folio) ||
3866 page_swap_entry(page).val != entry.val))
3870 * KSM sometimes has to copy on read faults, for example, if
3871 * page->index of !PageKSM() pages would be nonlinear inside the
3872 * anon VMA -- PageKSM() is lost on actual swapout.
3874 page = ksm_might_need_to_copy(page, vma, vmf->address);
3875 if (unlikely(!page)) {
3878 } else if (unlikely(PTR_ERR(page) == -EHWPOISON)) {
3879 ret = VM_FAULT_HWPOISON;
3882 folio = page_folio(page);
3885 * If we want to map a page that's in the swapcache writable, we
3886 * have to detect via the refcount if we're really the exclusive
3887 * owner. Try removing the extra reference from the local LRU
3888 * caches if required.
3890 if ((vmf->flags & FAULT_FLAG_WRITE) && folio == swapcache &&
3891 !folio_test_ksm(folio) && !folio_test_lru(folio))
3895 folio_throttle_swaprate(folio, GFP_KERNEL);
3898 * Back out if somebody else already faulted in this pte.
3900 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3902 if (unlikely(!vmf->pte || !pte_same(ptep_get(vmf->pte), vmf->orig_pte)))
3905 if (unlikely(!folio_test_uptodate(folio))) {
3906 ret = VM_FAULT_SIGBUS;
3911 * PG_anon_exclusive reuses PG_mappedtodisk for anon pages. A swap pte
3912 * must never point at an anonymous page in the swapcache that is
3913 * PG_anon_exclusive. Sanity check that this holds and especially, that
3914 * no filesystem set PG_mappedtodisk on a page in the swapcache. Sanity
3915 * check after taking the PT lock and making sure that nobody
3916 * concurrently faulted in this page and set PG_anon_exclusive.
3918 BUG_ON(!folio_test_anon(folio) && folio_test_mappedtodisk(folio));
3919 BUG_ON(folio_test_anon(folio) && PageAnonExclusive(page));
3922 * Check under PT lock (to protect against concurrent fork() sharing
3923 * the swap entry concurrently) for certainly exclusive pages.
3925 if (!folio_test_ksm(folio)) {
3926 exclusive = pte_swp_exclusive(vmf->orig_pte);
3927 if (folio != swapcache) {
3929 * We have a fresh page that is not exposed to the
3930 * swapcache -> certainly exclusive.
3933 } else if (exclusive && folio_test_writeback(folio) &&
3934 data_race(si->flags & SWP_STABLE_WRITES)) {
3936 * This is tricky: not all swap backends support
3937 * concurrent page modifications while under writeback.
3939 * So if we stumble over such a page in the swapcache
3940 * we must not set the page exclusive, otherwise we can
3941 * map it writable without further checks and modify it
3942 * while still under writeback.
3944 * For these problematic swap backends, simply drop the
3945 * exclusive marker: this is perfectly fine as we start
3946 * writeback only if we fully unmapped the page and
3947 * there are no unexpected references on the page after
3948 * unmapping succeeded. After fully unmapped, no
3949 * further GUP references (FOLL_GET and FOLL_PIN) can
3950 * appear, so dropping the exclusive marker and mapping
3951 * it only R/O is fine.
3958 * Some architectures may have to restore extra metadata to the page
3959 * when reading from swap. This metadata may be indexed by swap entry
3960 * so this must be called before swap_free().
3962 arch_swap_restore(entry, folio);
3965 * Remove the swap entry and conditionally try to free up the swapcache.
3966 * We're already holding a reference on the page but haven't mapped it
3970 if (should_try_to_free_swap(folio, vma, vmf->flags))
3971 folio_free_swap(folio);
3973 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
3974 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
3975 pte = mk_pte(page, vma->vm_page_prot);
3978 * Same logic as in do_wp_page(); however, optimize for pages that are
3979 * certainly not shared either because we just allocated them without
3980 * exposing them to the swapcache or because the swap entry indicates
3983 if (!folio_test_ksm(folio) &&
3984 (exclusive || folio_ref_count(folio) == 1)) {
3985 if (vmf->flags & FAULT_FLAG_WRITE) {
3986 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3987 vmf->flags &= ~FAULT_FLAG_WRITE;
3989 rmap_flags |= RMAP_EXCLUSIVE;
3991 flush_icache_page(vma, page);
3992 if (pte_swp_soft_dirty(vmf->orig_pte))
3993 pte = pte_mksoft_dirty(pte);
3994 if (pte_swp_uffd_wp(vmf->orig_pte))
3995 pte = pte_mkuffd_wp(pte);
3996 vmf->orig_pte = pte;
3998 /* ksm created a completely new copy */
3999 if (unlikely(folio != swapcache && swapcache)) {
4000 page_add_new_anon_rmap(page, vma, vmf->address);
4001 folio_add_lru_vma(folio, vma);
4003 page_add_anon_rmap(page, vma, vmf->address, rmap_flags);
4006 VM_BUG_ON(!folio_test_anon(folio) ||
4007 (pte_write(pte) && !PageAnonExclusive(page)));
4008 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
4009 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
4011 folio_unlock(folio);
4012 if (folio != swapcache && swapcache) {
4014 * Hold the lock to avoid the swap entry to be reused
4015 * until we take the PT lock for the pte_same() check
4016 * (to avoid false positives from pte_same). For
4017 * further safety release the lock after the swap_free
4018 * so that the swap count won't change under a
4019 * parallel locked swapcache.
4021 folio_unlock(swapcache);
4022 folio_put(swapcache);
4025 if (vmf->flags & FAULT_FLAG_WRITE) {
4026 ret |= do_wp_page(vmf);
4027 if (ret & VM_FAULT_ERROR)
4028 ret &= VM_FAULT_ERROR;
4032 /* No need to invalidate - it was non-present before */
4033 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
4036 pte_unmap_unlock(vmf->pte, vmf->ptl);
4039 put_swap_device(si);
4043 pte_unmap_unlock(vmf->pte, vmf->ptl);
4045 folio_unlock(folio);
4048 if (folio != swapcache && swapcache) {
4049 folio_unlock(swapcache);
4050 folio_put(swapcache);
4053 put_swap_device(si);
4058 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4059 * but allow concurrent faults), and pte mapped but not yet locked.
4060 * We return with mmap_lock still held, but pte unmapped and unlocked.
4062 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
4064 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4065 struct vm_area_struct *vma = vmf->vma;
4066 struct folio *folio;
4070 /* File mapping without ->vm_ops ? */
4071 if (vma->vm_flags & VM_SHARED)
4072 return VM_FAULT_SIGBUS;
4075 * Use pte_alloc() instead of pte_alloc_map(), so that OOM can
4076 * be distinguished from a transient failure of pte_offset_map().
4078 if (pte_alloc(vma->vm_mm, vmf->pmd))
4079 return VM_FAULT_OOM;
4081 /* Use the zero-page for reads */
4082 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
4083 !mm_forbids_zeropage(vma->vm_mm)) {
4084 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
4085 vma->vm_page_prot));
4086 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4087 vmf->address, &vmf->ptl);
4090 if (vmf_pte_changed(vmf)) {
4091 update_mmu_tlb(vma, vmf->address, vmf->pte);
4094 ret = check_stable_address_space(vma->vm_mm);
4097 /* Deliver the page fault to userland, check inside PT lock */
4098 if (userfaultfd_missing(vma)) {
4099 pte_unmap_unlock(vmf->pte, vmf->ptl);
4100 return handle_userfault(vmf, VM_UFFD_MISSING);
4105 /* Allocate our own private page. */
4106 if (unlikely(anon_vma_prepare(vma)))
4108 folio = vma_alloc_zeroed_movable_folio(vma, vmf->address);
4112 if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
4114 folio_throttle_swaprate(folio, GFP_KERNEL);
4117 * The memory barrier inside __folio_mark_uptodate makes sure that
4118 * preceding stores to the page contents become visible before
4119 * the set_pte_at() write.
4121 __folio_mark_uptodate(folio);
4123 entry = mk_pte(&folio->page, vma->vm_page_prot);
4124 entry = pte_sw_mkyoung(entry);
4125 if (vma->vm_flags & VM_WRITE)
4126 entry = pte_mkwrite(pte_mkdirty(entry), vma);
4128 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
4132 if (vmf_pte_changed(vmf)) {
4133 update_mmu_tlb(vma, vmf->address, vmf->pte);
4137 ret = check_stable_address_space(vma->vm_mm);
4141 /* Deliver the page fault to userland, check inside PT lock */
4142 if (userfaultfd_missing(vma)) {
4143 pte_unmap_unlock(vmf->pte, vmf->ptl);
4145 return handle_userfault(vmf, VM_UFFD_MISSING);
4148 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
4149 folio_add_new_anon_rmap(folio, vma, vmf->address);
4150 folio_add_lru_vma(folio, vma);
4153 entry = pte_mkuffd_wp(entry);
4154 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
4156 /* No need to invalidate - it was non-present before */
4157 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
4160 pte_unmap_unlock(vmf->pte, vmf->ptl);
4168 return VM_FAULT_OOM;
4172 * The mmap_lock must have been held on entry, and may have been
4173 * released depending on flags and vma->vm_ops->fault() return value.
4174 * See filemap_fault() and __lock_page_retry().
4176 static vm_fault_t __do_fault(struct vm_fault *vmf)
4178 struct vm_area_struct *vma = vmf->vma;
4182 * Preallocate pte before we take page_lock because this might lead to
4183 * deadlocks for memcg reclaim which waits for pages under writeback:
4185 * SetPageWriteback(A)
4191 * wait_on_page_writeback(A)
4192 * SetPageWriteback(B)
4194 * # flush A, B to clear the writeback
4196 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
4197 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4198 if (!vmf->prealloc_pte)
4199 return VM_FAULT_OOM;
4202 ret = vma->vm_ops->fault(vmf);
4203 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
4204 VM_FAULT_DONE_COW)))
4207 if (unlikely(PageHWPoison(vmf->page))) {
4208 struct page *page = vmf->page;
4209 vm_fault_t poisonret = VM_FAULT_HWPOISON;
4210 if (ret & VM_FAULT_LOCKED) {
4211 if (page_mapped(page))
4212 unmap_mapping_pages(page_mapping(page),
4213 page->index, 1, false);
4214 /* Retry if a clean page was removed from the cache. */
4215 if (invalidate_inode_page(page))
4216 poisonret = VM_FAULT_NOPAGE;
4224 if (unlikely(!(ret & VM_FAULT_LOCKED)))
4225 lock_page(vmf->page);
4227 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
4232 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4233 static void deposit_prealloc_pte(struct vm_fault *vmf)
4235 struct vm_area_struct *vma = vmf->vma;
4237 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4239 * We are going to consume the prealloc table,
4240 * count that as nr_ptes.
4242 mm_inc_nr_ptes(vma->vm_mm);
4243 vmf->prealloc_pte = NULL;
4246 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4248 struct vm_area_struct *vma = vmf->vma;
4249 bool write = vmf->flags & FAULT_FLAG_WRITE;
4250 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
4252 vm_fault_t ret = VM_FAULT_FALLBACK;
4254 if (!transhuge_vma_suitable(vma, haddr))
4257 page = compound_head(page);
4258 if (compound_order(page) != HPAGE_PMD_ORDER)
4262 * Just backoff if any subpage of a THP is corrupted otherwise
4263 * the corrupted page may mapped by PMD silently to escape the
4264 * check. This kind of THP just can be PTE mapped. Access to
4265 * the corrupted subpage should trigger SIGBUS as expected.
4267 if (unlikely(PageHasHWPoisoned(page)))
4271 * Archs like ppc64 need additional space to store information
4272 * related to pte entry. Use the preallocated table for that.
4274 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
4275 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
4276 if (!vmf->prealloc_pte)
4277 return VM_FAULT_OOM;
4280 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4281 if (unlikely(!pmd_none(*vmf->pmd)))
4284 flush_icache_pages(vma, page, HPAGE_PMD_NR);
4286 entry = mk_huge_pmd(page, vma->vm_page_prot);
4288 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
4290 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
4291 page_add_file_rmap(page, vma, true);
4294 * deposit and withdraw with pmd lock held
4296 if (arch_needs_pgtable_deposit())
4297 deposit_prealloc_pte(vmf);
4299 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
4301 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
4303 /* fault is handled */
4305 count_vm_event(THP_FILE_MAPPED);
4307 spin_unlock(vmf->ptl);
4311 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
4313 return VM_FAULT_FALLBACK;
4318 * set_pte_range - Set a range of PTEs to point to pages in a folio.
4319 * @vmf: Fault decription.
4320 * @folio: The folio that contains @page.
4321 * @page: The first page to create a PTE for.
4322 * @nr: The number of PTEs to create.
4323 * @addr: The first address to create a PTE for.
4325 void set_pte_range(struct vm_fault *vmf, struct folio *folio,
4326 struct page *page, unsigned int nr, unsigned long addr)
4328 struct vm_area_struct *vma = vmf->vma;
4329 bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
4330 bool write = vmf->flags & FAULT_FLAG_WRITE;
4331 bool prefault = in_range(vmf->address, addr, nr * PAGE_SIZE);
4334 flush_icache_pages(vma, page, nr);
4335 entry = mk_pte(page, vma->vm_page_prot);
4337 if (prefault && arch_wants_old_prefaulted_pte())
4338 entry = pte_mkold(entry);
4340 entry = pte_sw_mkyoung(entry);
4343 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4344 if (unlikely(uffd_wp))
4345 entry = pte_mkuffd_wp(entry);
4346 /* copy-on-write page */
4347 if (write && !(vma->vm_flags & VM_SHARED)) {
4348 add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr);
4349 VM_BUG_ON_FOLIO(nr != 1, folio);
4350 folio_add_new_anon_rmap(folio, vma, addr);
4351 folio_add_lru_vma(folio, vma);
4353 add_mm_counter(vma->vm_mm, mm_counter_file(page), nr);
4354 folio_add_file_rmap_range(folio, page, nr, vma, false);
4356 set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr);
4358 /* no need to invalidate: a not-present page won't be cached */
4359 update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr);
4362 static bool vmf_pte_changed(struct vm_fault *vmf)
4364 if (vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)
4365 return !pte_same(ptep_get(vmf->pte), vmf->orig_pte);
4367 return !pte_none(ptep_get(vmf->pte));
4371 * finish_fault - finish page fault once we have prepared the page to fault
4373 * @vmf: structure describing the fault
4375 * This function handles all that is needed to finish a page fault once the
4376 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4377 * given page, adds reverse page mapping, handles memcg charges and LRU
4380 * The function expects the page to be locked and on success it consumes a
4381 * reference of a page being mapped (for the PTE which maps it).
4383 * Return: %0 on success, %VM_FAULT_ code in case of error.
4385 vm_fault_t finish_fault(struct vm_fault *vmf)
4387 struct vm_area_struct *vma = vmf->vma;
4391 /* Did we COW the page? */
4392 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4393 page = vmf->cow_page;
4398 * check even for read faults because we might have lost our CoWed
4401 if (!(vma->vm_flags & VM_SHARED)) {
4402 ret = check_stable_address_space(vma->vm_mm);
4407 if (pmd_none(*vmf->pmd)) {
4408 if (PageTransCompound(page)) {
4409 ret = do_set_pmd(vmf, page);
4410 if (ret != VM_FAULT_FALLBACK)
4414 if (vmf->prealloc_pte)
4415 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4416 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4417 return VM_FAULT_OOM;
4420 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4421 vmf->address, &vmf->ptl);
4423 return VM_FAULT_NOPAGE;
4425 /* Re-check under ptl */
4426 if (likely(!vmf_pte_changed(vmf))) {
4427 struct folio *folio = page_folio(page);
4429 set_pte_range(vmf, folio, page, 1, vmf->address);
4432 update_mmu_tlb(vma, vmf->address, vmf->pte);
4433 ret = VM_FAULT_NOPAGE;
4436 pte_unmap_unlock(vmf->pte, vmf->ptl);
4440 static unsigned long fault_around_pages __read_mostly =
4441 65536 >> PAGE_SHIFT;
4443 #ifdef CONFIG_DEBUG_FS
4444 static int fault_around_bytes_get(void *data, u64 *val)
4446 *val = fault_around_pages << PAGE_SHIFT;
4451 * fault_around_bytes must be rounded down to the nearest page order as it's
4452 * what do_fault_around() expects to see.
4454 static int fault_around_bytes_set(void *data, u64 val)
4456 if (val / PAGE_SIZE > PTRS_PER_PTE)
4460 * The minimum value is 1 page, however this results in no fault-around
4461 * at all. See should_fault_around().
4463 fault_around_pages = max(rounddown_pow_of_two(val) >> PAGE_SHIFT, 1UL);
4467 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4468 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4470 static int __init fault_around_debugfs(void)
4472 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4473 &fault_around_bytes_fops);
4476 late_initcall(fault_around_debugfs);
4480 * do_fault_around() tries to map few pages around the fault address. The hope
4481 * is that the pages will be needed soon and this will lower the number of
4484 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4485 * not ready to be mapped: not up-to-date, locked, etc.
4487 * This function doesn't cross VMA or page table boundaries, in order to call
4488 * map_pages() and acquire a PTE lock only once.
4490 * fault_around_pages defines how many pages we'll try to map.
4491 * do_fault_around() expects it to be set to a power of two less than or equal
4494 * The virtual address of the area that we map is naturally aligned to
4495 * fault_around_pages * PAGE_SIZE rounded down to the machine page size
4496 * (and therefore to page order). This way it's easier to guarantee
4497 * that we don't cross page table boundaries.
4499 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4501 pgoff_t nr_pages = READ_ONCE(fault_around_pages);
4502 pgoff_t pte_off = pte_index(vmf->address);
4503 /* The page offset of vmf->address within the VMA. */
4504 pgoff_t vma_off = vmf->pgoff - vmf->vma->vm_pgoff;
4505 pgoff_t from_pte, to_pte;
4508 /* The PTE offset of the start address, clamped to the VMA. */
4509 from_pte = max(ALIGN_DOWN(pte_off, nr_pages),
4510 pte_off - min(pte_off, vma_off));
4512 /* The PTE offset of the end address, clamped to the VMA and PTE. */
4513 to_pte = min3(from_pte + nr_pages, (pgoff_t)PTRS_PER_PTE,
4514 pte_off + vma_pages(vmf->vma) - vma_off) - 1;
4516 if (pmd_none(*vmf->pmd)) {
4517 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4518 if (!vmf->prealloc_pte)
4519 return VM_FAULT_OOM;
4523 ret = vmf->vma->vm_ops->map_pages(vmf,
4524 vmf->pgoff + from_pte - pte_off,
4525 vmf->pgoff + to_pte - pte_off);
4531 /* Return true if we should do read fault-around, false otherwise */
4532 static inline bool should_fault_around(struct vm_fault *vmf)
4534 /* No ->map_pages? No way to fault around... */
4535 if (!vmf->vma->vm_ops->map_pages)
4538 if (uffd_disable_fault_around(vmf->vma))
4541 /* A single page implies no faulting 'around' at all. */
4542 return fault_around_pages > 1;
4545 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4548 struct folio *folio;
4551 * Let's call ->map_pages() first and use ->fault() as fallback
4552 * if page by the offset is not ready to be mapped (cold cache or
4555 if (should_fault_around(vmf)) {
4556 ret = do_fault_around(vmf);
4561 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
4562 vma_end_read(vmf->vma);
4563 return VM_FAULT_RETRY;
4566 ret = __do_fault(vmf);
4567 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4570 ret |= finish_fault(vmf);
4571 folio = page_folio(vmf->page);
4572 folio_unlock(folio);
4573 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4578 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4580 struct vm_area_struct *vma = vmf->vma;
4583 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
4585 return VM_FAULT_RETRY;
4588 if (unlikely(anon_vma_prepare(vma)))
4589 return VM_FAULT_OOM;
4591 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4593 return VM_FAULT_OOM;
4595 if (mem_cgroup_charge(page_folio(vmf->cow_page), vma->vm_mm,
4597 put_page(vmf->cow_page);
4598 return VM_FAULT_OOM;
4600 folio_throttle_swaprate(page_folio(vmf->cow_page), GFP_KERNEL);
4602 ret = __do_fault(vmf);
4603 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4605 if (ret & VM_FAULT_DONE_COW)
4608 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4609 __SetPageUptodate(vmf->cow_page);
4611 ret |= finish_fault(vmf);
4612 unlock_page(vmf->page);
4613 put_page(vmf->page);
4614 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4618 put_page(vmf->cow_page);
4622 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4624 struct vm_area_struct *vma = vmf->vma;
4625 vm_fault_t ret, tmp;
4626 struct folio *folio;
4628 if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
4630 return VM_FAULT_RETRY;
4633 ret = __do_fault(vmf);
4634 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4637 folio = page_folio(vmf->page);
4640 * Check if the backing address space wants to know that the page is
4641 * about to become writable
4643 if (vma->vm_ops->page_mkwrite) {
4644 folio_unlock(folio);
4645 tmp = do_page_mkwrite(vmf, folio);
4646 if (unlikely(!tmp ||
4647 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4653 ret |= finish_fault(vmf);
4654 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4656 folio_unlock(folio);
4661 ret |= fault_dirty_shared_page(vmf);
4666 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4667 * but allow concurrent faults).
4668 * The mmap_lock may have been released depending on flags and our
4669 * return value. See filemap_fault() and __folio_lock_or_retry().
4670 * If mmap_lock is released, vma may become invalid (for example
4671 * by other thread calling munmap()).
4673 static vm_fault_t do_fault(struct vm_fault *vmf)
4675 struct vm_area_struct *vma = vmf->vma;
4676 struct mm_struct *vm_mm = vma->vm_mm;
4680 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4682 if (!vma->vm_ops->fault) {
4683 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
4684 vmf->address, &vmf->ptl);
4685 if (unlikely(!vmf->pte))
4686 ret = VM_FAULT_SIGBUS;
4689 * Make sure this is not a temporary clearing of pte
4690 * by holding ptl and checking again. A R/M/W update
4691 * of pte involves: take ptl, clearing the pte so that
4692 * we don't have concurrent modification by hardware
4693 * followed by an update.
4695 if (unlikely(pte_none(ptep_get(vmf->pte))))
4696 ret = VM_FAULT_SIGBUS;
4698 ret = VM_FAULT_NOPAGE;
4700 pte_unmap_unlock(vmf->pte, vmf->ptl);
4702 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4703 ret = do_read_fault(vmf);
4704 else if (!(vma->vm_flags & VM_SHARED))
4705 ret = do_cow_fault(vmf);
4707 ret = do_shared_fault(vmf);
4709 /* preallocated pagetable is unused: free it */
4710 if (vmf->prealloc_pte) {
4711 pte_free(vm_mm, vmf->prealloc_pte);
4712 vmf->prealloc_pte = NULL;
4717 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4718 unsigned long addr, int page_nid, int *flags)
4722 /* Record the current PID acceesing VMA */
4723 vma_set_access_pid_bit(vma);
4725 count_vm_numa_event(NUMA_HINT_FAULTS);
4726 if (page_nid == numa_node_id()) {
4727 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4728 *flags |= TNF_FAULT_LOCAL;
4731 return mpol_misplaced(page, vma, addr);
4734 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4736 struct vm_area_struct *vma = vmf->vma;
4737 struct page *page = NULL;
4738 int page_nid = NUMA_NO_NODE;
4739 bool writable = false;
4746 * The "pte" at this point cannot be used safely without
4747 * validation through pte_unmap_same(). It's of NUMA type but
4748 * the pfn may be screwed if the read is non atomic.
4750 spin_lock(vmf->ptl);
4751 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
4752 pte_unmap_unlock(vmf->pte, vmf->ptl);
4756 /* Get the normal PTE */
4757 old_pte = ptep_get(vmf->pte);
4758 pte = pte_modify(old_pte, vma->vm_page_prot);
4761 * Detect now whether the PTE could be writable; this information
4762 * is only valid while holding the PT lock.
4764 writable = pte_write(pte);
4765 if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
4766 can_change_pte_writable(vma, vmf->address, pte))
4769 page = vm_normal_page(vma, vmf->address, pte);
4770 if (!page || is_zone_device_page(page))
4773 /* TODO: handle PTE-mapped THP */
4774 if (PageCompound(page))
4778 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4779 * much anyway since they can be in shared cache state. This misses
4780 * the case where a mapping is writable but the process never writes
4781 * to it but pte_write gets cleared during protection updates and
4782 * pte_dirty has unpredictable behaviour between PTE scan updates,
4783 * background writeback, dirty balancing and application behaviour.
4786 flags |= TNF_NO_GROUP;
4789 * Flag if the page is shared between multiple address spaces. This
4790 * is later used when determining whether to group tasks together
4792 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4793 flags |= TNF_SHARED;
4795 page_nid = page_to_nid(page);
4797 * For memory tiering mode, cpupid of slow memory page is used
4798 * to record page access time. So use default value.
4800 if ((sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4801 !node_is_toptier(page_nid))
4802 last_cpupid = (-1 & LAST_CPUPID_MASK);
4804 last_cpupid = page_cpupid_last(page);
4805 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4807 if (target_nid == NUMA_NO_NODE) {
4811 pte_unmap_unlock(vmf->pte, vmf->ptl);
4814 /* Migrate to the requested node */
4815 if (migrate_misplaced_page(page, vma, target_nid)) {
4816 page_nid = target_nid;
4817 flags |= TNF_MIGRATED;
4819 flags |= TNF_MIGRATE_FAIL;
4820 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4821 vmf->address, &vmf->ptl);
4822 if (unlikely(!vmf->pte))
4824 if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
4825 pte_unmap_unlock(vmf->pte, vmf->ptl);
4832 if (page_nid != NUMA_NO_NODE)
4833 task_numa_fault(last_cpupid, page_nid, 1, flags);
4837 * Make it present again, depending on how arch implements
4838 * non-accessible ptes, some can allow access by kernel mode.
4840 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4841 pte = pte_modify(old_pte, vma->vm_page_prot);
4842 pte = pte_mkyoung(pte);
4844 pte = pte_mkwrite(pte, vma);
4845 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4846 update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
4847 pte_unmap_unlock(vmf->pte, vmf->ptl);
4851 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4853 struct vm_area_struct *vma = vmf->vma;
4854 if (vma_is_anonymous(vma))
4855 return do_huge_pmd_anonymous_page(vmf);
4856 if (vma->vm_ops->huge_fault)
4857 return vma->vm_ops->huge_fault(vmf, PMD_ORDER);
4858 return VM_FAULT_FALLBACK;
4861 /* `inline' is required to avoid gcc 4.1.2 build error */
4862 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4864 struct vm_area_struct *vma = vmf->vma;
4865 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
4868 if (vma_is_anonymous(vma)) {
4869 if (likely(!unshare) &&
4870 userfaultfd_huge_pmd_wp(vma, vmf->orig_pmd))
4871 return handle_userfault(vmf, VM_UFFD_WP);
4872 return do_huge_pmd_wp_page(vmf);
4875 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4876 if (vma->vm_ops->huge_fault) {
4877 ret = vma->vm_ops->huge_fault(vmf, PMD_ORDER);
4878 if (!(ret & VM_FAULT_FALLBACK))
4883 /* COW or write-notify handled on pte level: split pmd. */
4884 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
4886 return VM_FAULT_FALLBACK;
4889 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4891 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4892 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4893 struct vm_area_struct *vma = vmf->vma;
4894 /* No support for anonymous transparent PUD pages yet */
4895 if (vma_is_anonymous(vma))
4896 return VM_FAULT_FALLBACK;
4897 if (vma->vm_ops->huge_fault)
4898 return vma->vm_ops->huge_fault(vmf, PUD_ORDER);
4899 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4900 return VM_FAULT_FALLBACK;
4903 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4905 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4906 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4907 struct vm_area_struct *vma = vmf->vma;
4910 /* No support for anonymous transparent PUD pages yet */
4911 if (vma_is_anonymous(vma))
4913 if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) {
4914 if (vma->vm_ops->huge_fault) {
4915 ret = vma->vm_ops->huge_fault(vmf, PUD_ORDER);
4916 if (!(ret & VM_FAULT_FALLBACK))
4921 /* COW or write-notify not handled on PUD level: split pud.*/
4922 __split_huge_pud(vma, vmf->pud, vmf->address);
4923 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
4924 return VM_FAULT_FALLBACK;
4928 * These routines also need to handle stuff like marking pages dirty
4929 * and/or accessed for architectures that don't do it in hardware (most
4930 * RISC architectures). The early dirtying is also good on the i386.
4932 * There is also a hook called "update_mmu_cache()" that architectures
4933 * with external mmu caches can use to update those (ie the Sparc or
4934 * PowerPC hashed page tables that act as extended TLBs).
4936 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4937 * concurrent faults).
4939 * The mmap_lock may have been released depending on flags and our return value.
4940 * See filemap_fault() and __folio_lock_or_retry().
4942 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4946 if (unlikely(pmd_none(*vmf->pmd))) {
4948 * Leave __pte_alloc() until later: because vm_ops->fault may
4949 * want to allocate huge page, and if we expose page table
4950 * for an instant, it will be difficult to retract from
4951 * concurrent faults and from rmap lookups.
4954 vmf->flags &= ~FAULT_FLAG_ORIG_PTE_VALID;
4957 * A regular pmd is established and it can't morph into a huge
4958 * pmd by anon khugepaged, since that takes mmap_lock in write
4959 * mode; but shmem or file collapse to THP could still morph
4960 * it into a huge pmd: just retry later if so.
4962 vmf->pte = pte_offset_map_nolock(vmf->vma->vm_mm, vmf->pmd,
4963 vmf->address, &vmf->ptl);
4964 if (unlikely(!vmf->pte))
4966 vmf->orig_pte = ptep_get_lockless(vmf->pte);
4967 vmf->flags |= FAULT_FLAG_ORIG_PTE_VALID;
4969 if (pte_none(vmf->orig_pte)) {
4970 pte_unmap(vmf->pte);
4976 return do_pte_missing(vmf);
4978 if (!pte_present(vmf->orig_pte))
4979 return do_swap_page(vmf);
4981 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4982 return do_numa_page(vmf);
4984 spin_lock(vmf->ptl);
4985 entry = vmf->orig_pte;
4986 if (unlikely(!pte_same(ptep_get(vmf->pte), entry))) {
4987 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4990 if (vmf->flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
4991 if (!pte_write(entry))
4992 return do_wp_page(vmf);
4993 else if (likely(vmf->flags & FAULT_FLAG_WRITE))
4994 entry = pte_mkdirty(entry);
4996 entry = pte_mkyoung(entry);
4997 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4998 vmf->flags & FAULT_FLAG_WRITE)) {
4999 update_mmu_cache_range(vmf, vmf->vma, vmf->address,
5002 /* Skip spurious TLB flush for retried page fault */
5003 if (vmf->flags & FAULT_FLAG_TRIED)
5006 * This is needed only for protection faults but the arch code
5007 * is not yet telling us if this is a protection fault or not.
5008 * This still avoids useless tlb flushes for .text page faults
5011 if (vmf->flags & FAULT_FLAG_WRITE)
5012 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address,
5016 pte_unmap_unlock(vmf->pte, vmf->ptl);
5021 * On entry, we hold either the VMA lock or the mmap_lock
5022 * (FAULT_FLAG_VMA_LOCK tells you which). If VM_FAULT_RETRY is set in
5023 * the result, the mmap_lock is not held on exit. See filemap_fault()
5024 * and __folio_lock_or_retry().
5026 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
5027 unsigned long address, unsigned int flags)
5029 struct vm_fault vmf = {
5031 .address = address & PAGE_MASK,
5032 .real_address = address,
5034 .pgoff = linear_page_index(vma, address),
5035 .gfp_mask = __get_fault_gfp_mask(vma),
5037 struct mm_struct *mm = vma->vm_mm;
5038 unsigned long vm_flags = vma->vm_flags;
5043 pgd = pgd_offset(mm, address);
5044 p4d = p4d_alloc(mm, pgd, address);
5046 return VM_FAULT_OOM;
5048 vmf.pud = pud_alloc(mm, p4d, address);
5050 return VM_FAULT_OOM;
5052 if (pud_none(*vmf.pud) &&
5053 hugepage_vma_check(vma, vm_flags, false, true, true)) {
5054 ret = create_huge_pud(&vmf);
5055 if (!(ret & VM_FAULT_FALLBACK))
5058 pud_t orig_pud = *vmf.pud;
5061 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
5064 * TODO once we support anonymous PUDs: NUMA case and
5065 * FAULT_FLAG_UNSHARE handling.
5067 if ((flags & FAULT_FLAG_WRITE) && !pud_write(orig_pud)) {
5068 ret = wp_huge_pud(&vmf, orig_pud);
5069 if (!(ret & VM_FAULT_FALLBACK))
5072 huge_pud_set_accessed(&vmf, orig_pud);
5078 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
5080 return VM_FAULT_OOM;
5082 /* Huge pud page fault raced with pmd_alloc? */
5083 if (pud_trans_unstable(vmf.pud))
5086 if (pmd_none(*vmf.pmd) &&
5087 hugepage_vma_check(vma, vm_flags, false, true, true)) {
5088 ret = create_huge_pmd(&vmf);
5089 if (!(ret & VM_FAULT_FALLBACK))
5092 vmf.orig_pmd = pmdp_get_lockless(vmf.pmd);
5094 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
5095 VM_BUG_ON(thp_migration_supported() &&
5096 !is_pmd_migration_entry(vmf.orig_pmd));
5097 if (is_pmd_migration_entry(vmf.orig_pmd))
5098 pmd_migration_entry_wait(mm, vmf.pmd);
5101 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
5102 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
5103 return do_huge_pmd_numa_page(&vmf);
5105 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5106 !pmd_write(vmf.orig_pmd)) {
5107 ret = wp_huge_pmd(&vmf);
5108 if (!(ret & VM_FAULT_FALLBACK))
5111 huge_pmd_set_accessed(&vmf);
5117 return handle_pte_fault(&vmf);
5121 * mm_account_fault - Do page fault accounting
5122 * @mm: mm from which memcg should be extracted. It can be NULL.
5123 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
5124 * of perf event counters, but we'll still do the per-task accounting to
5125 * the task who triggered this page fault.
5126 * @address: the faulted address.
5127 * @flags: the fault flags.
5128 * @ret: the fault retcode.
5130 * This will take care of most of the page fault accounting. Meanwhile, it
5131 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
5132 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
5133 * still be in per-arch page fault handlers at the entry of page fault.
5135 static inline void mm_account_fault(struct mm_struct *mm, struct pt_regs *regs,
5136 unsigned long address, unsigned int flags,
5141 /* Incomplete faults will be accounted upon completion. */
5142 if (ret & VM_FAULT_RETRY)
5146 * To preserve the behavior of older kernels, PGFAULT counters record
5147 * both successful and failed faults, as opposed to perf counters,
5148 * which ignore failed cases.
5150 count_vm_event(PGFAULT);
5151 count_memcg_event_mm(mm, PGFAULT);
5154 * Do not account for unsuccessful faults (e.g. when the address wasn't
5155 * valid). That includes arch_vma_access_permitted() failing before
5156 * reaching here. So this is not a "this many hardware page faults"
5157 * counter. We should use the hw profiling for that.
5159 if (ret & VM_FAULT_ERROR)
5163 * We define the fault as a major fault when the final successful fault
5164 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
5165 * handle it immediately previously).
5167 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
5175 * If the fault is done for GUP, regs will be NULL. We only do the
5176 * accounting for the per thread fault counters who triggered the
5177 * fault, and we skip the perf event updates.
5183 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
5185 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
5188 #ifdef CONFIG_LRU_GEN
5189 static void lru_gen_enter_fault(struct vm_area_struct *vma)
5191 /* the LRU algorithm only applies to accesses with recency */
5192 current->in_lru_fault = vma_has_recency(vma);
5195 static void lru_gen_exit_fault(void)
5197 current->in_lru_fault = false;
5200 static void lru_gen_enter_fault(struct vm_area_struct *vma)
5204 static void lru_gen_exit_fault(void)
5207 #endif /* CONFIG_LRU_GEN */
5209 static vm_fault_t sanitize_fault_flags(struct vm_area_struct *vma,
5210 unsigned int *flags)
5212 if (unlikely(*flags & FAULT_FLAG_UNSHARE)) {
5213 if (WARN_ON_ONCE(*flags & FAULT_FLAG_WRITE))
5214 return VM_FAULT_SIGSEGV;
5216 * FAULT_FLAG_UNSHARE only applies to COW mappings. Let's
5217 * just treat it like an ordinary read-fault otherwise.
5219 if (!is_cow_mapping(vma->vm_flags))
5220 *flags &= ~FAULT_FLAG_UNSHARE;
5221 } else if (*flags & FAULT_FLAG_WRITE) {
5222 /* Write faults on read-only mappings are impossible ... */
5223 if (WARN_ON_ONCE(!(vma->vm_flags & VM_MAYWRITE)))
5224 return VM_FAULT_SIGSEGV;
5225 /* ... and FOLL_FORCE only applies to COW mappings. */
5226 if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE) &&
5227 !is_cow_mapping(vma->vm_flags)))
5228 return VM_FAULT_SIGSEGV;
5230 #ifdef CONFIG_PER_VMA_LOCK
5232 * Per-VMA locks can't be used with FAULT_FLAG_RETRY_NOWAIT because of
5233 * the assumption that lock is dropped on VM_FAULT_RETRY.
5235 if (WARN_ON_ONCE((*flags &
5236 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)) ==
5237 (FAULT_FLAG_VMA_LOCK | FAULT_FLAG_RETRY_NOWAIT)))
5238 return VM_FAULT_SIGSEGV;
5245 * By the time we get here, we already hold the mm semaphore
5247 * The mmap_lock may have been released depending on flags and our
5248 * return value. See filemap_fault() and __folio_lock_or_retry().
5250 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
5251 unsigned int flags, struct pt_regs *regs)
5253 /* If the fault handler drops the mmap_lock, vma may be freed */
5254 struct mm_struct *mm = vma->vm_mm;
5257 __set_current_state(TASK_RUNNING);
5259 ret = sanitize_fault_flags(vma, &flags);
5263 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
5264 flags & FAULT_FLAG_INSTRUCTION,
5265 flags & FAULT_FLAG_REMOTE)) {
5266 ret = VM_FAULT_SIGSEGV;
5271 * Enable the memcg OOM handling for faults triggered in user
5272 * space. Kernel faults are handled more gracefully.
5274 if (flags & FAULT_FLAG_USER)
5275 mem_cgroup_enter_user_fault();
5277 lru_gen_enter_fault(vma);
5279 if (unlikely(is_vm_hugetlb_page(vma)))
5280 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
5282 ret = __handle_mm_fault(vma, address, flags);
5284 lru_gen_exit_fault();
5286 if (flags & FAULT_FLAG_USER) {
5287 mem_cgroup_exit_user_fault();
5289 * The task may have entered a memcg OOM situation but
5290 * if the allocation error was handled gracefully (no
5291 * VM_FAULT_OOM), there is no need to kill anything.
5292 * Just clean up the OOM state peacefully.
5294 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
5295 mem_cgroup_oom_synchronize(false);
5298 mm_account_fault(mm, regs, address, flags, ret);
5302 EXPORT_SYMBOL_GPL(handle_mm_fault);
5304 #ifdef CONFIG_LOCK_MM_AND_FIND_VMA
5305 #include <linux/extable.h>
5307 static inline bool get_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5309 if (likely(mmap_read_trylock(mm)))
5312 if (regs && !user_mode(regs)) {
5313 unsigned long ip = instruction_pointer(regs);
5314 if (!search_exception_tables(ip))
5318 return !mmap_read_lock_killable(mm);
5321 static inline bool mmap_upgrade_trylock(struct mm_struct *mm)
5324 * We don't have this operation yet.
5326 * It should be easy enough to do: it's basically a
5327 * atomic_long_try_cmpxchg_acquire()
5328 * from RWSEM_READER_BIAS -> RWSEM_WRITER_LOCKED, but
5329 * it also needs the proper lockdep magic etc.
5334 static inline bool upgrade_mmap_lock_carefully(struct mm_struct *mm, struct pt_regs *regs)
5336 mmap_read_unlock(mm);
5337 if (regs && !user_mode(regs)) {
5338 unsigned long ip = instruction_pointer(regs);
5339 if (!search_exception_tables(ip))
5342 return !mmap_write_lock_killable(mm);
5346 * Helper for page fault handling.
5348 * This is kind of equivalend to "mmap_read_lock()" followed
5349 * by "find_extend_vma()", except it's a lot more careful about
5350 * the locking (and will drop the lock on failure).
5352 * For example, if we have a kernel bug that causes a page
5353 * fault, we don't want to just use mmap_read_lock() to get
5354 * the mm lock, because that would deadlock if the bug were
5355 * to happen while we're holding the mm lock for writing.
5357 * So this checks the exception tables on kernel faults in
5358 * order to only do this all for instructions that are actually
5359 * expected to fault.
5361 * We can also actually take the mm lock for writing if we
5362 * need to extend the vma, which helps the VM layer a lot.
5364 struct vm_area_struct *lock_mm_and_find_vma(struct mm_struct *mm,
5365 unsigned long addr, struct pt_regs *regs)
5367 struct vm_area_struct *vma;
5369 if (!get_mmap_lock_carefully(mm, regs))
5372 vma = find_vma(mm, addr);
5373 if (likely(vma && (vma->vm_start <= addr)))
5377 * Well, dang. We might still be successful, but only
5378 * if we can extend a vma to do so.
5380 if (!vma || !(vma->vm_flags & VM_GROWSDOWN)) {
5381 mmap_read_unlock(mm);
5386 * We can try to upgrade the mmap lock atomically,
5387 * in which case we can continue to use the vma
5388 * we already looked up.
5390 * Otherwise we'll have to drop the mmap lock and
5391 * re-take it, and also look up the vma again,
5394 if (!mmap_upgrade_trylock(mm)) {
5395 if (!upgrade_mmap_lock_carefully(mm, regs))
5398 vma = find_vma(mm, addr);
5401 if (vma->vm_start <= addr)
5403 if (!(vma->vm_flags & VM_GROWSDOWN))
5407 if (expand_stack_locked(vma, addr))
5411 mmap_write_downgrade(mm);
5415 mmap_write_unlock(mm);
5420 #ifdef CONFIG_PER_VMA_LOCK
5422 * Lookup and lock a VMA under RCU protection. Returned VMA is guaranteed to be
5423 * stable and not isolated. If the VMA is not found or is being modified the
5424 * function returns NULL.
5426 struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
5427 unsigned long address)
5429 MA_STATE(mas, &mm->mm_mt, address, address);
5430 struct vm_area_struct *vma;
5434 vma = mas_walk(&mas);
5438 if (!vma_start_read(vma))
5442 * find_mergeable_anon_vma uses adjacent vmas which are not locked.
5443 * This check must happen after vma_start_read(); otherwise, a
5444 * concurrent mremap() with MREMAP_DONTUNMAP could dissociate the VMA
5445 * from its anon_vma.
5447 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma))
5448 goto inval_end_read;
5450 /* Check since vm_start/vm_end might change before we lock the VMA */
5451 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
5452 goto inval_end_read;
5454 /* Check if the VMA got isolated after we found it */
5455 if (vma->detached) {
5457 count_vm_vma_lock_event(VMA_LOCK_MISS);
5458 /* The area was replaced with another one */
5469 count_vm_vma_lock_event(VMA_LOCK_ABORT);
5472 #endif /* CONFIG_PER_VMA_LOCK */
5474 #ifndef __PAGETABLE_P4D_FOLDED
5476 * Allocate p4d page table.
5477 * We've already handled the fast-path in-line.
5479 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
5481 p4d_t *new = p4d_alloc_one(mm, address);
5485 spin_lock(&mm->page_table_lock);
5486 if (pgd_present(*pgd)) { /* Another has populated it */
5489 smp_wmb(); /* See comment in pmd_install() */
5490 pgd_populate(mm, pgd, new);
5492 spin_unlock(&mm->page_table_lock);
5495 #endif /* __PAGETABLE_P4D_FOLDED */
5497 #ifndef __PAGETABLE_PUD_FOLDED
5499 * Allocate page upper directory.
5500 * We've already handled the fast-path in-line.
5502 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
5504 pud_t *new = pud_alloc_one(mm, address);
5508 spin_lock(&mm->page_table_lock);
5509 if (!p4d_present(*p4d)) {
5511 smp_wmb(); /* See comment in pmd_install() */
5512 p4d_populate(mm, p4d, new);
5513 } else /* Another has populated it */
5515 spin_unlock(&mm->page_table_lock);
5518 #endif /* __PAGETABLE_PUD_FOLDED */
5520 #ifndef __PAGETABLE_PMD_FOLDED
5522 * Allocate page middle directory.
5523 * We've already handled the fast-path in-line.
5525 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
5528 pmd_t *new = pmd_alloc_one(mm, address);
5532 ptl = pud_lock(mm, pud);
5533 if (!pud_present(*pud)) {
5535 smp_wmb(); /* See comment in pmd_install() */
5536 pud_populate(mm, pud, new);
5537 } else { /* Another has populated it */
5543 #endif /* __PAGETABLE_PMD_FOLDED */
5546 * follow_pte - look up PTE at a user virtual address
5547 * @mm: the mm_struct of the target address space
5548 * @address: user virtual address
5549 * @ptepp: location to store found PTE
5550 * @ptlp: location to store the lock for the PTE
5552 * On a successful return, the pointer to the PTE is stored in @ptepp;
5553 * the corresponding lock is taken and its location is stored in @ptlp.
5554 * The contents of the PTE are only stable until @ptlp is released;
5555 * any further use, if any, must be protected against invalidation
5556 * with MMU notifiers.
5558 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5559 * should be taken for read.
5561 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5562 * it is not a good general-purpose API.
5564 * Return: zero on success, -ve otherwise.
5566 int follow_pte(struct mm_struct *mm, unsigned long address,
5567 pte_t **ptepp, spinlock_t **ptlp)
5575 pgd = pgd_offset(mm, address);
5576 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
5579 p4d = p4d_offset(pgd, address);
5580 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
5583 pud = pud_offset(p4d, address);
5584 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
5587 pmd = pmd_offset(pud, address);
5588 VM_BUG_ON(pmd_trans_huge(*pmd));
5590 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
5593 if (!pte_present(ptep_get(ptep)))
5598 pte_unmap_unlock(ptep, *ptlp);
5602 EXPORT_SYMBOL_GPL(follow_pte);
5605 * follow_pfn - look up PFN at a user virtual address
5606 * @vma: memory mapping
5607 * @address: user virtual address
5608 * @pfn: location to store found PFN
5610 * Only IO mappings and raw PFN mappings are allowed.
5612 * This function does not allow the caller to read the permissions
5613 * of the PTE. Do not use it.
5615 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5617 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5624 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5627 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5630 *pfn = pte_pfn(ptep_get(ptep));
5631 pte_unmap_unlock(ptep, ptl);
5634 EXPORT_SYMBOL(follow_pfn);
5636 #ifdef CONFIG_HAVE_IOREMAP_PROT
5637 int follow_phys(struct vm_area_struct *vma,
5638 unsigned long address, unsigned int flags,
5639 unsigned long *prot, resource_size_t *phys)
5645 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5648 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5650 pte = ptep_get(ptep);
5652 if ((flags & FOLL_WRITE) && !pte_write(pte))
5655 *prot = pgprot_val(pte_pgprot(pte));
5656 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5660 pte_unmap_unlock(ptep, ptl);
5666 * generic_access_phys - generic implementation for iomem mmap access
5667 * @vma: the vma to access
5668 * @addr: userspace address, not relative offset within @vma
5669 * @buf: buffer to read/write
5670 * @len: length of transfer
5671 * @write: set to FOLL_WRITE when writing, otherwise reading
5673 * This is a generic implementation for &vm_operations_struct.access for an
5674 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5677 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5678 void *buf, int len, int write)
5680 resource_size_t phys_addr;
5681 unsigned long prot = 0;
5682 void __iomem *maddr;
5685 int offset = offset_in_page(addr);
5688 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5692 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5694 pte = ptep_get(ptep);
5695 pte_unmap_unlock(ptep, ptl);
5697 prot = pgprot_val(pte_pgprot(pte));
5698 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5700 if ((write & FOLL_WRITE) && !pte_write(pte))
5703 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5707 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5710 if (!pte_same(pte, ptep_get(ptep))) {
5711 pte_unmap_unlock(ptep, ptl);
5718 memcpy_toio(maddr + offset, buf, len);
5720 memcpy_fromio(buf, maddr + offset, len);
5722 pte_unmap_unlock(ptep, ptl);
5728 EXPORT_SYMBOL_GPL(generic_access_phys);
5732 * Access another process' address space as given in mm.
5734 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5735 int len, unsigned int gup_flags)
5737 void *old_buf = buf;
5738 int write = gup_flags & FOLL_WRITE;
5740 if (mmap_read_lock_killable(mm))
5743 /* Untag the address before looking up the VMA */
5744 addr = untagged_addr_remote(mm, addr);
5746 /* Avoid triggering the temporary warning in __get_user_pages */
5747 if (!vma_lookup(mm, addr) && !expand_stack(mm, addr))
5750 /* ignore errors, just check how much was successfully transferred */
5754 struct vm_area_struct *vma = NULL;
5755 struct page *page = get_user_page_vma_remote(mm, addr,
5758 if (IS_ERR_OR_NULL(page)) {
5759 /* We might need to expand the stack to access it */
5760 vma = vma_lookup(mm, addr);
5762 vma = expand_stack(mm, addr);
5764 /* mmap_lock was dropped on failure */
5766 return buf - old_buf;
5768 /* Try again if stack expansion worked */
5774 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5775 * we can access using slightly different code.
5778 #ifdef CONFIG_HAVE_IOREMAP_PROT
5779 if (vma->vm_ops && vma->vm_ops->access)
5780 bytes = vma->vm_ops->access(vma, addr, buf,
5787 offset = addr & (PAGE_SIZE-1);
5788 if (bytes > PAGE_SIZE-offset)
5789 bytes = PAGE_SIZE-offset;
5793 copy_to_user_page(vma, page, addr,
5794 maddr + offset, buf, bytes);
5795 set_page_dirty_lock(page);
5797 copy_from_user_page(vma, page, addr,
5798 buf, maddr + offset, bytes);
5807 mmap_read_unlock(mm);
5809 return buf - old_buf;
5813 * access_remote_vm - access another process' address space
5814 * @mm: the mm_struct of the target address space
5815 * @addr: start address to access
5816 * @buf: source or destination buffer
5817 * @len: number of bytes to transfer
5818 * @gup_flags: flags modifying lookup behaviour
5820 * The caller must hold a reference on @mm.
5822 * Return: number of bytes copied from source to destination.
5824 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5825 void *buf, int len, unsigned int gup_flags)
5827 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5831 * Access another process' address space.
5832 * Source/target buffer must be kernel space,
5833 * Do not walk the page table directly, use get_user_pages
5835 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5836 void *buf, int len, unsigned int gup_flags)
5838 struct mm_struct *mm;
5841 mm = get_task_mm(tsk);
5845 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5851 EXPORT_SYMBOL_GPL(access_process_vm);
5854 * Print the name of a VMA.
5856 void print_vma_addr(char *prefix, unsigned long ip)
5858 struct mm_struct *mm = current->mm;
5859 struct vm_area_struct *vma;
5862 * we might be running from an atomic context so we cannot sleep
5864 if (!mmap_read_trylock(mm))
5867 vma = find_vma(mm, ip);
5868 if (vma && vma->vm_file) {
5869 struct file *f = vma->vm_file;
5870 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5874 p = file_path(f, buf, PAGE_SIZE);
5877 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5879 vma->vm_end - vma->vm_start);
5880 free_page((unsigned long)buf);
5883 mmap_read_unlock(mm);
5886 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5887 void __might_fault(const char *file, int line)
5889 if (pagefault_disabled())
5891 __might_sleep(file, line);
5892 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5894 might_lock_read(¤t->mm->mmap_lock);
5897 EXPORT_SYMBOL(__might_fault);
5900 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5902 * Process all subpages of the specified huge page with the specified
5903 * operation. The target subpage will be processed last to keep its
5906 static inline int process_huge_page(
5907 unsigned long addr_hint, unsigned int pages_per_huge_page,
5908 int (*process_subpage)(unsigned long addr, int idx, void *arg),
5911 int i, n, base, l, ret;
5912 unsigned long addr = addr_hint &
5913 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5915 /* Process target subpage last to keep its cache lines hot */
5917 n = (addr_hint - addr) / PAGE_SIZE;
5918 if (2 * n <= pages_per_huge_page) {
5919 /* If target subpage in first half of huge page */
5922 /* Process subpages at the end of huge page */
5923 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5925 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
5930 /* If target subpage in second half of huge page */
5931 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5932 l = pages_per_huge_page - n;
5933 /* Process subpages at the begin of huge page */
5934 for (i = 0; i < base; i++) {
5936 ret = process_subpage(addr + i * PAGE_SIZE, i, arg);
5942 * Process remaining subpages in left-right-left-right pattern
5943 * towards the target subpage
5945 for (i = 0; i < l; i++) {
5946 int left_idx = base + i;
5947 int right_idx = base + 2 * l - 1 - i;
5950 ret = process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5954 ret = process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5961 static void clear_gigantic_page(struct page *page,
5963 unsigned int pages_per_huge_page)
5969 for (i = 0; i < pages_per_huge_page; i++) {
5970 p = nth_page(page, i);
5972 clear_user_highpage(p, addr + i * PAGE_SIZE);
5976 static int clear_subpage(unsigned long addr, int idx, void *arg)
5978 struct page *page = arg;
5980 clear_user_highpage(page + idx, addr);
5984 void clear_huge_page(struct page *page,
5985 unsigned long addr_hint, unsigned int pages_per_huge_page)
5987 unsigned long addr = addr_hint &
5988 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5990 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5991 clear_gigantic_page(page, addr, pages_per_huge_page);
5995 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5998 static int copy_user_gigantic_page(struct folio *dst, struct folio *src,
6000 struct vm_area_struct *vma,
6001 unsigned int pages_per_huge_page)
6004 struct page *dst_page;
6005 struct page *src_page;
6007 for (i = 0; i < pages_per_huge_page; i++) {
6008 dst_page = folio_page(dst, i);
6009 src_page = folio_page(src, i);
6012 if (copy_mc_user_highpage(dst_page, src_page,
6013 addr + i*PAGE_SIZE, vma)) {
6014 memory_failure_queue(page_to_pfn(src_page), 0);
6021 struct copy_subpage_arg {
6024 struct vm_area_struct *vma;
6027 static int copy_subpage(unsigned long addr, int idx, void *arg)
6029 struct copy_subpage_arg *copy_arg = arg;
6031 if (copy_mc_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
6032 addr, copy_arg->vma)) {
6033 memory_failure_queue(page_to_pfn(copy_arg->src + idx), 0);
6039 int copy_user_large_folio(struct folio *dst, struct folio *src,
6040 unsigned long addr_hint, struct vm_area_struct *vma)
6042 unsigned int pages_per_huge_page = folio_nr_pages(dst);
6043 unsigned long addr = addr_hint &
6044 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
6045 struct copy_subpage_arg arg = {
6051 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES))
6052 return copy_user_gigantic_page(dst, src, addr, vma,
6053 pages_per_huge_page);
6055 return process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
6058 long copy_folio_from_user(struct folio *dst_folio,
6059 const void __user *usr_src,
6060 bool allow_pagefault)
6063 unsigned long i, rc = 0;
6064 unsigned int nr_pages = folio_nr_pages(dst_folio);
6065 unsigned long ret_val = nr_pages * PAGE_SIZE;
6066 struct page *subpage;
6068 for (i = 0; i < nr_pages; i++) {
6069 subpage = folio_page(dst_folio, i);
6070 kaddr = kmap_local_page(subpage);
6071 if (!allow_pagefault)
6072 pagefault_disable();
6073 rc = copy_from_user(kaddr, usr_src + i * PAGE_SIZE, PAGE_SIZE);
6074 if (!allow_pagefault)
6076 kunmap_local(kaddr);
6078 ret_val -= (PAGE_SIZE - rc);
6082 flush_dcache_page(subpage);
6088 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
6090 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
6092 static struct kmem_cache *page_ptl_cachep;
6094 void __init ptlock_cache_init(void)
6096 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
6100 bool ptlock_alloc(struct ptdesc *ptdesc)
6104 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
6111 void ptlock_free(struct ptdesc *ptdesc)
6113 kmem_cache_free(page_ptl_cachep, ptdesc->ptl);