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/ksm.h>
56 #include <linux/rmap.h>
57 #include <linux/export.h>
58 #include <linux/delayacct.h>
59 #include <linux/init.h>
60 #include <linux/pfn_t.h>
61 #include <linux/writeback.h>
62 #include <linux/memcontrol.h>
63 #include <linux/mmu_notifier.h>
64 #include <linux/swapops.h>
65 #include <linux/elf.h>
66 #include <linux/gfp.h>
67 #include <linux/migrate.h>
68 #include <linux/string.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
73 #include <linux/numa.h>
74 #include <linux/perf_event.h>
75 #include <linux/ptrace.h>
76 #include <linux/vmalloc.h>
78 #include <trace/events/kmem.h>
81 #include <asm/mmu_context.h>
82 #include <asm/pgalloc.h>
83 #include <linux/uaccess.h>
85 #include <asm/tlbflush.h>
87 #include "pgalloc-track.h"
90 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
91 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
95 unsigned long max_mapnr;
96 EXPORT_SYMBOL(max_mapnr);
99 EXPORT_SYMBOL(mem_map);
103 * A number of key systems in x86 including ioremap() rely on the assumption
104 * that high_memory defines the upper bound on direct map memory, then end
105 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
106 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
110 EXPORT_SYMBOL(high_memory);
113 * Randomize the address space (stacks, mmaps, brk, etc.).
115 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
116 * as ancient (libc5 based) binaries can segfault. )
118 int randomize_va_space __read_mostly =
119 #ifdef CONFIG_COMPAT_BRK
125 #ifndef arch_faults_on_old_pte
126 static inline bool arch_faults_on_old_pte(void)
129 * Those arches which don't have hw access flag feature need to
130 * implement their own helper. By default, "true" means pagefault
131 * will be hit on old pte.
137 #ifndef arch_wants_old_prefaulted_pte
138 static inline bool arch_wants_old_prefaulted_pte(void)
141 * Transitioning a PTE from 'old' to 'young' can be expensive on
142 * some architectures, even if it's performed in hardware. By
143 * default, "false" means prefaulted entries will be 'young'.
149 static int __init disable_randmaps(char *s)
151 randomize_va_space = 0;
154 __setup("norandmaps", disable_randmaps);
156 unsigned long zero_pfn __read_mostly;
157 EXPORT_SYMBOL(zero_pfn);
159 unsigned long highest_memmap_pfn __read_mostly;
162 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
164 static int __init init_zero_pfn(void)
166 zero_pfn = page_to_pfn(ZERO_PAGE(0));
169 early_initcall(init_zero_pfn);
171 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
173 trace_rss_stat(mm, member, count);
176 #if defined(SPLIT_RSS_COUNTING)
178 void sync_mm_rss(struct mm_struct *mm)
182 for (i = 0; i < NR_MM_COUNTERS; i++) {
183 if (current->rss_stat.count[i]) {
184 add_mm_counter(mm, i, current->rss_stat.count[i]);
185 current->rss_stat.count[i] = 0;
188 current->rss_stat.events = 0;
191 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
193 struct task_struct *task = current;
195 if (likely(task->mm == mm))
196 task->rss_stat.count[member] += val;
198 add_mm_counter(mm, member, val);
200 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
201 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
203 /* sync counter once per 64 page faults */
204 #define TASK_RSS_EVENTS_THRESH (64)
205 static void check_sync_rss_stat(struct task_struct *task)
207 if (unlikely(task != current))
209 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
210 sync_mm_rss(task->mm);
212 #else /* SPLIT_RSS_COUNTING */
214 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
215 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
217 static void check_sync_rss_stat(struct task_struct *task)
221 #endif /* SPLIT_RSS_COUNTING */
224 * Note: this doesn't free the actual pages themselves. That
225 * has been handled earlier when unmapping all the memory regions.
227 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
230 pgtable_t token = pmd_pgtable(*pmd);
232 pte_free_tlb(tlb, token, addr);
233 mm_dec_nr_ptes(tlb->mm);
236 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
237 unsigned long addr, unsigned long end,
238 unsigned long floor, unsigned long ceiling)
245 pmd = pmd_offset(pud, addr);
247 next = pmd_addr_end(addr, end);
248 if (pmd_none_or_clear_bad(pmd))
250 free_pte_range(tlb, pmd, addr);
251 } while (pmd++, addr = next, addr != end);
261 if (end - 1 > ceiling - 1)
264 pmd = pmd_offset(pud, start);
266 pmd_free_tlb(tlb, pmd, start);
267 mm_dec_nr_pmds(tlb->mm);
270 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
271 unsigned long addr, unsigned long end,
272 unsigned long floor, unsigned long ceiling)
279 pud = pud_offset(p4d, addr);
281 next = pud_addr_end(addr, end);
282 if (pud_none_or_clear_bad(pud))
284 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
285 } while (pud++, addr = next, addr != end);
295 if (end - 1 > ceiling - 1)
298 pud = pud_offset(p4d, start);
300 pud_free_tlb(tlb, pud, start);
301 mm_dec_nr_puds(tlb->mm);
304 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
305 unsigned long addr, unsigned long end,
306 unsigned long floor, unsigned long ceiling)
313 p4d = p4d_offset(pgd, addr);
315 next = p4d_addr_end(addr, end);
316 if (p4d_none_or_clear_bad(p4d))
318 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
319 } while (p4d++, addr = next, addr != end);
325 ceiling &= PGDIR_MASK;
329 if (end - 1 > ceiling - 1)
332 p4d = p4d_offset(pgd, start);
334 p4d_free_tlb(tlb, p4d, start);
338 * This function frees user-level page tables of a process.
340 void free_pgd_range(struct mmu_gather *tlb,
341 unsigned long addr, unsigned long end,
342 unsigned long floor, unsigned long ceiling)
348 * The next few lines have given us lots of grief...
350 * Why are we testing PMD* at this top level? Because often
351 * there will be no work to do at all, and we'd prefer not to
352 * go all the way down to the bottom just to discover that.
354 * Why all these "- 1"s? Because 0 represents both the bottom
355 * of the address space and the top of it (using -1 for the
356 * top wouldn't help much: the masks would do the wrong thing).
357 * The rule is that addr 0 and floor 0 refer to the bottom of
358 * the address space, but end 0 and ceiling 0 refer to the top
359 * Comparisons need to use "end - 1" and "ceiling - 1" (though
360 * that end 0 case should be mythical).
362 * Wherever addr is brought up or ceiling brought down, we must
363 * be careful to reject "the opposite 0" before it confuses the
364 * subsequent tests. But what about where end is brought down
365 * by PMD_SIZE below? no, end can't go down to 0 there.
367 * Whereas we round start (addr) and ceiling down, by different
368 * masks at different levels, in order to test whether a table
369 * now has no other vmas using it, so can be freed, we don't
370 * bother to round floor or end up - the tests don't need that.
384 if (end - 1 > ceiling - 1)
389 * We add page table cache pages with PAGE_SIZE,
390 * (see pte_free_tlb()), flush the tlb if we need
392 tlb_change_page_size(tlb, PAGE_SIZE);
393 pgd = pgd_offset(tlb->mm, addr);
395 next = pgd_addr_end(addr, end);
396 if (pgd_none_or_clear_bad(pgd))
398 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
399 } while (pgd++, addr = next, addr != end);
402 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
403 unsigned long floor, unsigned long ceiling)
406 struct vm_area_struct *next = vma->vm_next;
407 unsigned long addr = vma->vm_start;
410 * Hide vma from rmap and truncate_pagecache before freeing
413 unlink_anon_vmas(vma);
414 unlink_file_vma(vma);
416 if (is_vm_hugetlb_page(vma)) {
417 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
418 floor, next ? next->vm_start : ceiling);
421 * Optimization: gather nearby vmas into one call down
423 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
424 && !is_vm_hugetlb_page(next)) {
427 unlink_anon_vmas(vma);
428 unlink_file_vma(vma);
430 free_pgd_range(tlb, addr, vma->vm_end,
431 floor, next ? next->vm_start : ceiling);
437 void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte)
439 spinlock_t *ptl = pmd_lock(mm, pmd);
441 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
444 * Ensure all pte setup (eg. pte page lock and page clearing) are
445 * visible before the pte is made visible to other CPUs by being
446 * put into page tables.
448 * The other side of the story is the pointer chasing in the page
449 * table walking code (when walking the page table without locking;
450 * ie. most of the time). Fortunately, these data accesses consist
451 * of a chain of data-dependent loads, meaning most CPUs (alpha
452 * being the notable exception) will already guarantee loads are
453 * seen in-order. See the alpha page table accessors for the
454 * smp_rmb() barriers in page table walking code.
456 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
457 pmd_populate(mm, pmd, *pte);
463 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
465 pgtable_t new = pte_alloc_one(mm);
469 pmd_install(mm, pmd, &new);
475 int __pte_alloc_kernel(pmd_t *pmd)
477 pte_t *new = pte_alloc_one_kernel(&init_mm);
481 spin_lock(&init_mm.page_table_lock);
482 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
483 smp_wmb(); /* See comment in pmd_install() */
484 pmd_populate_kernel(&init_mm, pmd, new);
487 spin_unlock(&init_mm.page_table_lock);
489 pte_free_kernel(&init_mm, new);
493 static inline void init_rss_vec(int *rss)
495 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
498 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
502 if (current->mm == mm)
504 for (i = 0; i < NR_MM_COUNTERS; i++)
506 add_mm_counter(mm, i, rss[i]);
510 * This function is called to print an error when a bad pte
511 * is found. For example, we might have a PFN-mapped pte in
512 * a region that doesn't allow it.
514 * The calling function must still handle the error.
516 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
517 pte_t pte, struct page *page)
519 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
520 p4d_t *p4d = p4d_offset(pgd, addr);
521 pud_t *pud = pud_offset(p4d, addr);
522 pmd_t *pmd = pmd_offset(pud, addr);
523 struct address_space *mapping;
525 static unsigned long resume;
526 static unsigned long nr_shown;
527 static unsigned long nr_unshown;
530 * Allow a burst of 60 reports, then keep quiet for that minute;
531 * or allow a steady drip of one report per second.
533 if (nr_shown == 60) {
534 if (time_before(jiffies, resume)) {
539 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
546 resume = jiffies + 60 * HZ;
548 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
549 index = linear_page_index(vma, addr);
551 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
553 (long long)pte_val(pte), (long long)pmd_val(*pmd));
555 dump_page(page, "bad pte");
556 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
557 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
558 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
560 vma->vm_ops ? vma->vm_ops->fault : NULL,
561 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
562 mapping ? mapping->a_ops->readpage : NULL);
564 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
568 * vm_normal_page -- This function gets the "struct page" associated with a pte.
570 * "Special" mappings do not wish to be associated with a "struct page" (either
571 * it doesn't exist, or it exists but they don't want to touch it). In this
572 * case, NULL is returned here. "Normal" mappings do have a struct page.
574 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
575 * pte bit, in which case this function is trivial. Secondly, an architecture
576 * may not have a spare pte bit, which requires a more complicated scheme,
579 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
580 * special mapping (even if there are underlying and valid "struct pages").
581 * COWed pages of a VM_PFNMAP are always normal.
583 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
584 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
585 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
586 * mapping will always honor the rule
588 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
590 * And for normal mappings this is false.
592 * This restricts such mappings to be a linear translation from virtual address
593 * to pfn. To get around this restriction, we allow arbitrary mappings so long
594 * as the vma is not a COW mapping; in that case, we know that all ptes are
595 * special (because none can have been COWed).
598 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
600 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
601 * page" backing, however the difference is that _all_ pages with a struct
602 * page (that is, those where pfn_valid is true) are refcounted and considered
603 * normal pages by the VM. The disadvantage is that pages are refcounted
604 * (which can be slower and simply not an option for some PFNMAP users). The
605 * advantage is that we don't have to follow the strict linearity rule of
606 * PFNMAP mappings in order to support COWable mappings.
609 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
612 unsigned long pfn = pte_pfn(pte);
614 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
615 if (likely(!pte_special(pte)))
617 if (vma->vm_ops && vma->vm_ops->find_special_page)
618 return vma->vm_ops->find_special_page(vma, addr);
619 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
621 if (is_zero_pfn(pfn))
626 print_bad_pte(vma, addr, pte, NULL);
630 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
632 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
633 if (vma->vm_flags & VM_MIXEDMAP) {
639 off = (addr - vma->vm_start) >> PAGE_SHIFT;
640 if (pfn == vma->vm_pgoff + off)
642 if (!is_cow_mapping(vma->vm_flags))
647 if (is_zero_pfn(pfn))
651 if (unlikely(pfn > highest_memmap_pfn)) {
652 print_bad_pte(vma, addr, pte, NULL);
657 * NOTE! We still have PageReserved() pages in the page tables.
658 * eg. VDSO mappings can cause them to exist.
661 return pfn_to_page(pfn);
664 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
665 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
668 unsigned long pfn = pmd_pfn(pmd);
671 * There is no pmd_special() but there may be special pmds, e.g.
672 * in a direct-access (dax) mapping, so let's just replicate the
673 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
675 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
676 if (vma->vm_flags & VM_MIXEDMAP) {
682 off = (addr - vma->vm_start) >> PAGE_SHIFT;
683 if (pfn == vma->vm_pgoff + off)
685 if (!is_cow_mapping(vma->vm_flags))
692 if (is_huge_zero_pmd(pmd))
694 if (unlikely(pfn > highest_memmap_pfn))
698 * NOTE! We still have PageReserved() pages in the page tables.
699 * eg. VDSO mappings can cause them to exist.
702 return pfn_to_page(pfn);
706 static void restore_exclusive_pte(struct vm_area_struct *vma,
707 struct page *page, unsigned long address,
713 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
714 if (pte_swp_soft_dirty(*ptep))
715 pte = pte_mksoft_dirty(pte);
717 entry = pte_to_swp_entry(*ptep);
718 if (pte_swp_uffd_wp(*ptep))
719 pte = pte_mkuffd_wp(pte);
720 else if (is_writable_device_exclusive_entry(entry))
721 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
724 * No need to take a page reference as one was already
725 * created when the swap entry was made.
728 page_add_anon_rmap(page, vma, address, false);
731 * Currently device exclusive access only supports anonymous
732 * memory so the entry shouldn't point to a filebacked page.
734 WARN_ON_ONCE(!PageAnon(page));
736 set_pte_at(vma->vm_mm, address, ptep, pte);
739 * No need to invalidate - it was non-present before. However
740 * secondary CPUs may have mappings that need invalidating.
742 update_mmu_cache(vma, address, ptep);
746 * Tries to restore an exclusive pte if the page lock can be acquired without
750 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
753 swp_entry_t entry = pte_to_swp_entry(*src_pte);
754 struct page *page = pfn_swap_entry_to_page(entry);
756 if (trylock_page(page)) {
757 restore_exclusive_pte(vma, page, addr, src_pte);
766 * copy one vm_area from one task to the other. Assumes the page tables
767 * already present in the new task to be cleared in the whole range
768 * covered by this vma.
772 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
773 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
774 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
776 unsigned long vm_flags = dst_vma->vm_flags;
777 pte_t pte = *src_pte;
779 swp_entry_t entry = pte_to_swp_entry(pte);
781 if (likely(!non_swap_entry(entry))) {
782 if (swap_duplicate(entry) < 0)
785 /* make sure dst_mm is on swapoff's mmlist. */
786 if (unlikely(list_empty(&dst_mm->mmlist))) {
787 spin_lock(&mmlist_lock);
788 if (list_empty(&dst_mm->mmlist))
789 list_add(&dst_mm->mmlist,
791 spin_unlock(&mmlist_lock);
794 } else if (is_migration_entry(entry)) {
795 page = pfn_swap_entry_to_page(entry);
797 rss[mm_counter(page)]++;
799 if (is_writable_migration_entry(entry) &&
800 is_cow_mapping(vm_flags)) {
802 * COW mappings require pages in both
803 * parent and child to be set to read.
805 entry = make_readable_migration_entry(
807 pte = swp_entry_to_pte(entry);
808 if (pte_swp_soft_dirty(*src_pte))
809 pte = pte_swp_mksoft_dirty(pte);
810 if (pte_swp_uffd_wp(*src_pte))
811 pte = pte_swp_mkuffd_wp(pte);
812 set_pte_at(src_mm, addr, src_pte, pte);
814 } else if (is_device_private_entry(entry)) {
815 page = pfn_swap_entry_to_page(entry);
818 * Update rss count even for unaddressable pages, as
819 * they should treated just like normal pages in this
822 * We will likely want to have some new rss counters
823 * for unaddressable pages, at some point. But for now
824 * keep things as they are.
827 rss[mm_counter(page)]++;
828 page_dup_rmap(page, false);
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(*src_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))
858 if (!userfaultfd_wp(dst_vma))
859 pte = pte_swp_clear_uffd_wp(pte);
860 set_pte_at(dst_mm, addr, dst_pte, pte);
865 * Copy a present and normal page if necessary.
867 * NOTE! The usual case is that this doesn't need to do
868 * anything, and can just return a positive value. That
869 * will let the caller know that it can just increase
870 * the page refcount and re-use the pte the traditional
873 * But _if_ we need to copy it because it needs to be
874 * pinned in the parent (and the child should get its own
875 * copy rather than just a reference to the same page),
876 * we'll do that here and return zero to let the caller
879 * And if we need a pre-allocated page but don't yet have
880 * one, return a negative error to let the preallocation
881 * code know so that it can do so outside the page table
885 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
886 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
887 struct page **prealloc, pte_t pte, struct page *page)
889 struct page *new_page;
892 * What we want to do is to check whether this page may
893 * have been pinned by the parent process. If so,
894 * instead of wrprotect the pte on both sides, we copy
895 * the page immediately so that we'll always guarantee
896 * the pinned page won't be randomly replaced in the
899 * The page pinning checks are just "has this mm ever
900 * seen pinning", along with the (inexact) check of
901 * the page count. That might give false positives for
902 * for pinning, but it will work correctly.
904 if (likely(!page_needs_cow_for_dma(src_vma, page)))
907 new_page = *prealloc;
912 * We have a prealloc page, all good! Take it
913 * over and copy the page & arm it.
916 copy_user_highpage(new_page, page, addr, src_vma);
917 __SetPageUptodate(new_page);
918 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
919 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
920 rss[mm_counter(new_page)]++;
922 /* All done, just insert the new page copy in the child */
923 pte = mk_pte(new_page, dst_vma->vm_page_prot);
924 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
925 if (userfaultfd_pte_wp(dst_vma, *src_pte))
926 /* Uffd-wp needs to be delivered to dest pte as well */
927 pte = pte_wrprotect(pte_mkuffd_wp(pte));
928 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
933 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
934 * is required to copy this pte.
937 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
938 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
939 struct page **prealloc)
941 struct mm_struct *src_mm = src_vma->vm_mm;
942 unsigned long vm_flags = src_vma->vm_flags;
943 pte_t pte = *src_pte;
946 page = vm_normal_page(src_vma, addr, pte);
950 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
951 addr, rss, prealloc, pte, page);
956 page_dup_rmap(page, false);
957 rss[mm_counter(page)]++;
961 * If it's a COW mapping, write protect it both
962 * in the parent and the child
964 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
965 ptep_set_wrprotect(src_mm, addr, src_pte);
966 pte = pte_wrprotect(pte);
970 * If it's a shared mapping, mark it clean in
973 if (vm_flags & VM_SHARED)
974 pte = pte_mkclean(pte);
975 pte = pte_mkold(pte);
977 if (!userfaultfd_wp(dst_vma))
978 pte = pte_clear_uffd_wp(pte);
980 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
984 static inline struct page *
985 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
988 struct page *new_page;
990 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
994 if (mem_cgroup_charge(page_folio(new_page), src_mm, GFP_KERNEL)) {
998 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
1004 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1005 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1008 struct mm_struct *dst_mm = dst_vma->vm_mm;
1009 struct mm_struct *src_mm = src_vma->vm_mm;
1010 pte_t *orig_src_pte, *orig_dst_pte;
1011 pte_t *src_pte, *dst_pte;
1012 spinlock_t *src_ptl, *dst_ptl;
1013 int progress, ret = 0;
1014 int rss[NR_MM_COUNTERS];
1015 swp_entry_t entry = (swp_entry_t){0};
1016 struct page *prealloc = NULL;
1022 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1027 src_pte = pte_offset_map(src_pmd, addr);
1028 src_ptl = pte_lockptr(src_mm, src_pmd);
1029 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1030 orig_src_pte = src_pte;
1031 orig_dst_pte = dst_pte;
1032 arch_enter_lazy_mmu_mode();
1036 * We are holding two locks at this point - either of them
1037 * could generate latencies in another task on another CPU.
1039 if (progress >= 32) {
1041 if (need_resched() ||
1042 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1045 if (pte_none(*src_pte)) {
1049 if (unlikely(!pte_present(*src_pte))) {
1050 ret = copy_nonpresent_pte(dst_mm, src_mm,
1055 entry = pte_to_swp_entry(*src_pte);
1057 } else if (ret == -EBUSY) {
1065 * Device exclusive entry restored, continue by copying
1066 * the now present pte.
1068 WARN_ON_ONCE(ret != -ENOENT);
1070 /* copy_present_pte() will clear `*prealloc' if consumed */
1071 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1072 addr, rss, &prealloc);
1074 * If we need a pre-allocated page for this pte, drop the
1075 * locks, allocate, and try again.
1077 if (unlikely(ret == -EAGAIN))
1079 if (unlikely(prealloc)) {
1081 * pre-alloc page cannot be reused by next time so as
1082 * to strictly follow mempolicy (e.g., alloc_page_vma()
1083 * will allocate page according to address). This
1084 * could only happen if one pinned pte changed.
1090 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1092 arch_leave_lazy_mmu_mode();
1093 spin_unlock(src_ptl);
1094 pte_unmap(orig_src_pte);
1095 add_mm_rss_vec(dst_mm, rss);
1096 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1100 VM_WARN_ON_ONCE(!entry.val);
1101 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1106 } else if (ret == -EBUSY) {
1108 } else if (ret == -EAGAIN) {
1109 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1116 /* We've captured and resolved the error. Reset, try again. */
1122 if (unlikely(prealloc))
1128 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1129 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1132 struct mm_struct *dst_mm = dst_vma->vm_mm;
1133 struct mm_struct *src_mm = src_vma->vm_mm;
1134 pmd_t *src_pmd, *dst_pmd;
1137 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1140 src_pmd = pmd_offset(src_pud, addr);
1142 next = pmd_addr_end(addr, end);
1143 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1144 || pmd_devmap(*src_pmd)) {
1146 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1147 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1148 addr, dst_vma, src_vma);
1155 if (pmd_none_or_clear_bad(src_pmd))
1157 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1160 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1165 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1166 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1169 struct mm_struct *dst_mm = dst_vma->vm_mm;
1170 struct mm_struct *src_mm = src_vma->vm_mm;
1171 pud_t *src_pud, *dst_pud;
1174 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1177 src_pud = pud_offset(src_p4d, addr);
1179 next = pud_addr_end(addr, end);
1180 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1183 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1184 err = copy_huge_pud(dst_mm, src_mm,
1185 dst_pud, src_pud, addr, src_vma);
1192 if (pud_none_or_clear_bad(src_pud))
1194 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1197 } while (dst_pud++, src_pud++, addr = next, addr != end);
1202 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1203 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1206 struct mm_struct *dst_mm = dst_vma->vm_mm;
1207 p4d_t *src_p4d, *dst_p4d;
1210 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1213 src_p4d = p4d_offset(src_pgd, addr);
1215 next = p4d_addr_end(addr, end);
1216 if (p4d_none_or_clear_bad(src_p4d))
1218 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1221 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1226 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1228 pgd_t *src_pgd, *dst_pgd;
1230 unsigned long addr = src_vma->vm_start;
1231 unsigned long end = src_vma->vm_end;
1232 struct mm_struct *dst_mm = dst_vma->vm_mm;
1233 struct mm_struct *src_mm = src_vma->vm_mm;
1234 struct mmu_notifier_range range;
1239 * Don't copy ptes where a page fault will fill them correctly.
1240 * Fork becomes much lighter when there are big shared or private
1241 * readonly mappings. The tradeoff is that copy_page_range is more
1242 * efficient than faulting.
1244 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1248 if (is_vm_hugetlb_page(src_vma))
1249 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1251 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1253 * We do not free on error cases below as remove_vma
1254 * gets called on error from higher level routine
1256 ret = track_pfn_copy(src_vma);
1262 * We need to invalidate the secondary MMU mappings only when
1263 * there could be a permission downgrade on the ptes of the
1264 * parent mm. And a permission downgrade will only happen if
1265 * is_cow_mapping() returns true.
1267 is_cow = is_cow_mapping(src_vma->vm_flags);
1270 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1271 0, src_vma, src_mm, addr, end);
1272 mmu_notifier_invalidate_range_start(&range);
1274 * Disabling preemption is not needed for the write side, as
1275 * the read side doesn't spin, but goes to the mmap_lock.
1277 * Use the raw variant of the seqcount_t write API to avoid
1278 * lockdep complaining about preemptibility.
1280 mmap_assert_write_locked(src_mm);
1281 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1285 dst_pgd = pgd_offset(dst_mm, addr);
1286 src_pgd = pgd_offset(src_mm, addr);
1288 next = pgd_addr_end(addr, end);
1289 if (pgd_none_or_clear_bad(src_pgd))
1291 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1296 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1299 raw_write_seqcount_end(&src_mm->write_protect_seq);
1300 mmu_notifier_invalidate_range_end(&range);
1306 * Parameter block passed down to zap_pte_range in exceptional cases.
1308 struct zap_details {
1309 struct folio *single_folio; /* Locked folio to be unmapped */
1310 bool even_cows; /* Zap COWed private pages too? */
1313 /* Whether we should zap all COWed (private) pages too */
1314 static inline bool should_zap_cows(struct zap_details *details)
1316 /* By default, zap all pages */
1320 /* Or, we zap COWed pages only if the caller wants to */
1321 return details->even_cows;
1324 /* Decides whether we should zap this page with the page pointer specified */
1325 static inline bool should_zap_page(struct zap_details *details, struct page *page)
1327 /* If we can make a decision without *page.. */
1328 if (should_zap_cows(details))
1331 /* E.g. the caller passes NULL for the case of a zero page */
1335 /* Otherwise we should only zap non-anon pages */
1336 return !PageAnon(page);
1339 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1340 struct vm_area_struct *vma, pmd_t *pmd,
1341 unsigned long addr, unsigned long end,
1342 struct zap_details *details)
1344 struct mm_struct *mm = tlb->mm;
1345 int force_flush = 0;
1346 int rss[NR_MM_COUNTERS];
1352 tlb_change_page_size(tlb, PAGE_SIZE);
1355 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1357 flush_tlb_batched_pending(mm);
1358 arch_enter_lazy_mmu_mode();
1363 if (pte_none(ptent))
1369 if (pte_present(ptent)) {
1370 page = vm_normal_page(vma, addr, ptent);
1371 if (unlikely(!should_zap_page(details, page)))
1373 ptent = ptep_get_and_clear_full(mm, addr, pte,
1375 tlb_remove_tlb_entry(tlb, pte, addr);
1376 if (unlikely(!page))
1379 if (!PageAnon(page)) {
1380 if (pte_dirty(ptent)) {
1382 set_page_dirty(page);
1384 if (pte_young(ptent) &&
1385 likely(!(vma->vm_flags & VM_SEQ_READ)))
1386 mark_page_accessed(page);
1388 rss[mm_counter(page)]--;
1389 page_remove_rmap(page, vma, false);
1390 if (unlikely(page_mapcount(page) < 0))
1391 print_bad_pte(vma, addr, ptent, page);
1392 if (unlikely(__tlb_remove_page(tlb, page))) {
1400 entry = pte_to_swp_entry(ptent);
1401 if (is_device_private_entry(entry) ||
1402 is_device_exclusive_entry(entry)) {
1403 page = pfn_swap_entry_to_page(entry);
1404 if (unlikely(!should_zap_page(details, page)))
1406 rss[mm_counter(page)]--;
1407 if (is_device_private_entry(entry))
1408 page_remove_rmap(page, vma, false);
1410 } else if (!non_swap_entry(entry)) {
1411 /* Genuine swap entry, hence a private anon page */
1412 if (!should_zap_cows(details))
1415 if (unlikely(!free_swap_and_cache(entry)))
1416 print_bad_pte(vma, addr, ptent, NULL);
1417 } else if (is_migration_entry(entry)) {
1418 page = pfn_swap_entry_to_page(entry);
1419 if (!should_zap_page(details, page))
1421 rss[mm_counter(page)]--;
1422 } else if (is_hwpoison_entry(entry)) {
1423 if (!should_zap_cows(details))
1426 /* We should have covered all the swap entry types */
1429 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1430 } while (pte++, addr += PAGE_SIZE, addr != end);
1432 add_mm_rss_vec(mm, rss);
1433 arch_leave_lazy_mmu_mode();
1435 /* Do the actual TLB flush before dropping ptl */
1437 tlb_flush_mmu_tlbonly(tlb);
1438 pte_unmap_unlock(start_pte, ptl);
1441 * If we forced a TLB flush (either due to running out of
1442 * batch buffers or because we needed to flush dirty TLB
1443 * entries before releasing the ptl), free the batched
1444 * memory too. Restart if we didn't do everything.
1459 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1460 struct vm_area_struct *vma, pud_t *pud,
1461 unsigned long addr, unsigned long end,
1462 struct zap_details *details)
1467 pmd = pmd_offset(pud, addr);
1469 next = pmd_addr_end(addr, end);
1470 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1471 if (next - addr != HPAGE_PMD_SIZE)
1472 __split_huge_pmd(vma, pmd, addr, false, NULL);
1473 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1476 } else if (details && details->single_folio &&
1477 folio_test_pmd_mappable(details->single_folio) &&
1478 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1479 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1481 * Take and drop THP pmd lock so that we cannot return
1482 * prematurely, while zap_huge_pmd() has cleared *pmd,
1483 * but not yet decremented compound_mapcount().
1489 * Here there can be other concurrent MADV_DONTNEED or
1490 * trans huge page faults running, and if the pmd is
1491 * none or trans huge it can change under us. This is
1492 * because MADV_DONTNEED holds the mmap_lock in read
1495 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1497 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1500 } while (pmd++, addr = next, addr != end);
1505 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1506 struct vm_area_struct *vma, p4d_t *p4d,
1507 unsigned long addr, unsigned long end,
1508 struct zap_details *details)
1513 pud = pud_offset(p4d, addr);
1515 next = pud_addr_end(addr, end);
1516 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1517 if (next - addr != HPAGE_PUD_SIZE) {
1518 mmap_assert_locked(tlb->mm);
1519 split_huge_pud(vma, pud, addr);
1520 } else if (zap_huge_pud(tlb, vma, pud, addr))
1524 if (pud_none_or_clear_bad(pud))
1526 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1529 } while (pud++, addr = next, addr != end);
1534 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1535 struct vm_area_struct *vma, pgd_t *pgd,
1536 unsigned long addr, unsigned long end,
1537 struct zap_details *details)
1542 p4d = p4d_offset(pgd, addr);
1544 next = p4d_addr_end(addr, end);
1545 if (p4d_none_or_clear_bad(p4d))
1547 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1548 } while (p4d++, addr = next, addr != end);
1553 void unmap_page_range(struct mmu_gather *tlb,
1554 struct vm_area_struct *vma,
1555 unsigned long addr, unsigned long end,
1556 struct zap_details *details)
1561 BUG_ON(addr >= end);
1562 tlb_start_vma(tlb, vma);
1563 pgd = pgd_offset(vma->vm_mm, addr);
1565 next = pgd_addr_end(addr, end);
1566 if (pgd_none_or_clear_bad(pgd))
1568 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1569 } while (pgd++, addr = next, addr != end);
1570 tlb_end_vma(tlb, vma);
1574 static void unmap_single_vma(struct mmu_gather *tlb,
1575 struct vm_area_struct *vma, unsigned long start_addr,
1576 unsigned long end_addr,
1577 struct zap_details *details)
1579 unsigned long start = max(vma->vm_start, start_addr);
1582 if (start >= vma->vm_end)
1584 end = min(vma->vm_end, end_addr);
1585 if (end <= vma->vm_start)
1589 uprobe_munmap(vma, start, end);
1591 if (unlikely(vma->vm_flags & VM_PFNMAP))
1592 untrack_pfn(vma, 0, 0);
1595 if (unlikely(is_vm_hugetlb_page(vma))) {
1597 * It is undesirable to test vma->vm_file as it
1598 * should be non-null for valid hugetlb area.
1599 * However, vm_file will be NULL in the error
1600 * cleanup path of mmap_region. When
1601 * hugetlbfs ->mmap method fails,
1602 * mmap_region() nullifies vma->vm_file
1603 * before calling this function to clean up.
1604 * Since no pte has actually been setup, it is
1605 * safe to do nothing in this case.
1608 i_mmap_lock_write(vma->vm_file->f_mapping);
1609 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1610 i_mmap_unlock_write(vma->vm_file->f_mapping);
1613 unmap_page_range(tlb, vma, start, end, details);
1618 * unmap_vmas - unmap a range of memory covered by a list of vma's
1619 * @tlb: address of the caller's struct mmu_gather
1620 * @vma: the starting vma
1621 * @start_addr: virtual address at which to start unmapping
1622 * @end_addr: virtual address at which to end unmapping
1624 * Unmap all pages in the vma list.
1626 * Only addresses between `start' and `end' will be unmapped.
1628 * The VMA list must be sorted in ascending virtual address order.
1630 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1631 * range after unmap_vmas() returns. So the only responsibility here is to
1632 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1633 * drops the lock and schedules.
1635 void unmap_vmas(struct mmu_gather *tlb,
1636 struct vm_area_struct *vma, unsigned long start_addr,
1637 unsigned long end_addr)
1639 struct mmu_notifier_range range;
1641 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1642 start_addr, end_addr);
1643 mmu_notifier_invalidate_range_start(&range);
1644 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1645 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1646 mmu_notifier_invalidate_range_end(&range);
1650 * zap_page_range - remove user pages in a given range
1651 * @vma: vm_area_struct holding the applicable pages
1652 * @start: starting address of pages to zap
1653 * @size: number of bytes to zap
1655 * Caller must protect the VMA list
1657 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1660 struct mmu_notifier_range range;
1661 struct mmu_gather tlb;
1664 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1665 start, start + size);
1666 tlb_gather_mmu(&tlb, vma->vm_mm);
1667 update_hiwater_rss(vma->vm_mm);
1668 mmu_notifier_invalidate_range_start(&range);
1669 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1670 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1671 mmu_notifier_invalidate_range_end(&range);
1672 tlb_finish_mmu(&tlb);
1676 * zap_page_range_single - remove user pages in a given range
1677 * @vma: vm_area_struct holding the applicable pages
1678 * @address: starting address of pages to zap
1679 * @size: number of bytes to zap
1680 * @details: details of shared cache invalidation
1682 * The range must fit into one VMA.
1684 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1685 unsigned long size, struct zap_details *details)
1687 struct mmu_notifier_range range;
1688 struct mmu_gather tlb;
1691 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1692 address, address + size);
1693 tlb_gather_mmu(&tlb, vma->vm_mm);
1694 update_hiwater_rss(vma->vm_mm);
1695 mmu_notifier_invalidate_range_start(&range);
1696 unmap_single_vma(&tlb, vma, address, range.end, details);
1697 mmu_notifier_invalidate_range_end(&range);
1698 tlb_finish_mmu(&tlb);
1702 * zap_vma_ptes - remove ptes mapping the vma
1703 * @vma: vm_area_struct holding ptes to be zapped
1704 * @address: starting address of pages to zap
1705 * @size: number of bytes to zap
1707 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1709 * The entire address range must be fully contained within the vma.
1712 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1715 if (!range_in_vma(vma, address, address + size) ||
1716 !(vma->vm_flags & VM_PFNMAP))
1719 zap_page_range_single(vma, address, size, NULL);
1721 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1723 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1730 pgd = pgd_offset(mm, addr);
1731 p4d = p4d_alloc(mm, pgd, addr);
1734 pud = pud_alloc(mm, p4d, addr);
1737 pmd = pmd_alloc(mm, pud, addr);
1741 VM_BUG_ON(pmd_trans_huge(*pmd));
1745 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1748 pmd_t *pmd = walk_to_pmd(mm, addr);
1752 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1755 static int validate_page_before_insert(struct page *page)
1757 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1759 flush_dcache_page(page);
1763 static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
1764 unsigned long addr, struct page *page, pgprot_t prot)
1766 if (!pte_none(*pte))
1768 /* Ok, finally just insert the thing.. */
1770 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
1771 page_add_file_rmap(page, vma, false);
1772 set_pte_at(vma->vm_mm, addr, pte, mk_pte(page, prot));
1777 * This is the old fallback for page remapping.
1779 * For historical reasons, it only allows reserved pages. Only
1780 * old drivers should use this, and they needed to mark their
1781 * pages reserved for the old functions anyway.
1783 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1784 struct page *page, pgprot_t prot)
1790 retval = validate_page_before_insert(page);
1794 pte = get_locked_pte(vma->vm_mm, addr, &ptl);
1797 retval = insert_page_into_pte_locked(vma, pte, addr, page, prot);
1798 pte_unmap_unlock(pte, ptl);
1804 static int insert_page_in_batch_locked(struct vm_area_struct *vma, pte_t *pte,
1805 unsigned long addr, struct page *page, pgprot_t prot)
1809 if (!page_count(page))
1811 err = validate_page_before_insert(page);
1814 return insert_page_into_pte_locked(vma, pte, addr, page, prot);
1817 /* insert_pages() amortizes the cost of spinlock operations
1818 * when inserting pages in a loop. Arch *must* define pte_index.
1820 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1821 struct page **pages, unsigned long *num, pgprot_t prot)
1824 pte_t *start_pte, *pte;
1825 spinlock_t *pte_lock;
1826 struct mm_struct *const mm = vma->vm_mm;
1827 unsigned long curr_page_idx = 0;
1828 unsigned long remaining_pages_total = *num;
1829 unsigned long pages_to_write_in_pmd;
1833 pmd = walk_to_pmd(mm, addr);
1837 pages_to_write_in_pmd = min_t(unsigned long,
1838 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1840 /* Allocate the PTE if necessary; takes PMD lock once only. */
1842 if (pte_alloc(mm, pmd))
1845 while (pages_to_write_in_pmd) {
1847 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1849 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1850 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1851 int err = insert_page_in_batch_locked(vma, pte,
1852 addr, pages[curr_page_idx], prot);
1853 if (unlikely(err)) {
1854 pte_unmap_unlock(start_pte, pte_lock);
1856 remaining_pages_total -= pte_idx;
1862 pte_unmap_unlock(start_pte, pte_lock);
1863 pages_to_write_in_pmd -= batch_size;
1864 remaining_pages_total -= batch_size;
1866 if (remaining_pages_total)
1870 *num = remaining_pages_total;
1873 #endif /* ifdef pte_index */
1876 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1877 * @vma: user vma to map to
1878 * @addr: target start user address of these pages
1879 * @pages: source kernel pages
1880 * @num: in: number of pages to map. out: number of pages that were *not*
1881 * mapped. (0 means all pages were successfully mapped).
1883 * Preferred over vm_insert_page() when inserting multiple pages.
1885 * In case of error, we may have mapped a subset of the provided
1886 * pages. It is the caller's responsibility to account for this case.
1888 * The same restrictions apply as in vm_insert_page().
1890 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1891 struct page **pages, unsigned long *num)
1894 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1896 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1898 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1899 BUG_ON(mmap_read_trylock(vma->vm_mm));
1900 BUG_ON(vma->vm_flags & VM_PFNMAP);
1901 vma->vm_flags |= VM_MIXEDMAP;
1903 /* Defer page refcount checking till we're about to map that page. */
1904 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1906 unsigned long idx = 0, pgcount = *num;
1909 for (; idx < pgcount; ++idx) {
1910 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1914 *num = pgcount - idx;
1916 #endif /* ifdef pte_index */
1918 EXPORT_SYMBOL(vm_insert_pages);
1921 * vm_insert_page - insert single page into user vma
1922 * @vma: user vma to map to
1923 * @addr: target user address of this page
1924 * @page: source kernel page
1926 * This allows drivers to insert individual pages they've allocated
1929 * The page has to be a nice clean _individual_ kernel allocation.
1930 * If you allocate a compound page, you need to have marked it as
1931 * such (__GFP_COMP), or manually just split the page up yourself
1932 * (see split_page()).
1934 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1935 * took an arbitrary page protection parameter. This doesn't allow
1936 * that. Your vma protection will have to be set up correctly, which
1937 * means that if you want a shared writable mapping, you'd better
1938 * ask for a shared writable mapping!
1940 * The page does not need to be reserved.
1942 * Usually this function is called from f_op->mmap() handler
1943 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1944 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1945 * function from other places, for example from page-fault handler.
1947 * Return: %0 on success, negative error code otherwise.
1949 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1952 if (addr < vma->vm_start || addr >= vma->vm_end)
1954 if (!page_count(page))
1956 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1957 BUG_ON(mmap_read_trylock(vma->vm_mm));
1958 BUG_ON(vma->vm_flags & VM_PFNMAP);
1959 vma->vm_flags |= VM_MIXEDMAP;
1961 return insert_page(vma, addr, page, vma->vm_page_prot);
1963 EXPORT_SYMBOL(vm_insert_page);
1966 * __vm_map_pages - maps range of kernel pages into user vma
1967 * @vma: user vma to map to
1968 * @pages: pointer to array of source kernel pages
1969 * @num: number of pages in page array
1970 * @offset: user's requested vm_pgoff
1972 * This allows drivers to map range of kernel pages into a user vma.
1974 * Return: 0 on success and error code otherwise.
1976 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1977 unsigned long num, unsigned long offset)
1979 unsigned long count = vma_pages(vma);
1980 unsigned long uaddr = vma->vm_start;
1983 /* Fail if the user requested offset is beyond the end of the object */
1987 /* Fail if the user requested size exceeds available object size */
1988 if (count > num - offset)
1991 for (i = 0; i < count; i++) {
1992 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2002 * vm_map_pages - maps range of kernel pages starts with non zero offset
2003 * @vma: user vma to map to
2004 * @pages: pointer to array of source kernel pages
2005 * @num: number of pages in page array
2007 * Maps an object consisting of @num pages, catering for the user's
2008 * requested vm_pgoff
2010 * If we fail to insert any page into the vma, the function will return
2011 * immediately leaving any previously inserted pages present. Callers
2012 * from the mmap handler may immediately return the error as their caller
2013 * will destroy the vma, removing any successfully inserted pages. Other
2014 * callers should make their own arrangements for calling unmap_region().
2016 * Context: Process context. Called by mmap handlers.
2017 * Return: 0 on success and error code otherwise.
2019 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2022 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2024 EXPORT_SYMBOL(vm_map_pages);
2027 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2028 * @vma: user vma to map to
2029 * @pages: pointer to array of source kernel pages
2030 * @num: number of pages in page array
2032 * Similar to vm_map_pages(), except that it explicitly sets the offset
2033 * to 0. This function is intended for the drivers that did not consider
2036 * Context: Process context. Called by mmap handlers.
2037 * Return: 0 on success and error code otherwise.
2039 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2042 return __vm_map_pages(vma, pages, num, 0);
2044 EXPORT_SYMBOL(vm_map_pages_zero);
2046 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2047 pfn_t pfn, pgprot_t prot, bool mkwrite)
2049 struct mm_struct *mm = vma->vm_mm;
2053 pte = get_locked_pte(mm, addr, &ptl);
2055 return VM_FAULT_OOM;
2056 if (!pte_none(*pte)) {
2059 * For read faults on private mappings the PFN passed
2060 * in may not match the PFN we have mapped if the
2061 * mapped PFN is a writeable COW page. In the mkwrite
2062 * case we are creating a writable PTE for a shared
2063 * mapping and we expect the PFNs to match. If they
2064 * don't match, we are likely racing with block
2065 * allocation and mapping invalidation so just skip the
2068 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
2069 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
2072 entry = pte_mkyoung(*pte);
2073 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2074 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2075 update_mmu_cache(vma, addr, pte);
2080 /* Ok, finally just insert the thing.. */
2081 if (pfn_t_devmap(pfn))
2082 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2084 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2087 entry = pte_mkyoung(entry);
2088 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2091 set_pte_at(mm, addr, pte, entry);
2092 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2095 pte_unmap_unlock(pte, ptl);
2096 return VM_FAULT_NOPAGE;
2100 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2101 * @vma: user vma to map to
2102 * @addr: target user address of this page
2103 * @pfn: source kernel pfn
2104 * @pgprot: pgprot flags for the inserted page
2106 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2107 * to override pgprot on a per-page basis.
2109 * This only makes sense for IO mappings, and it makes no sense for
2110 * COW mappings. In general, using multiple vmas is preferable;
2111 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2114 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2115 * a value of @pgprot different from that of @vma->vm_page_prot.
2117 * Context: Process context. May allocate using %GFP_KERNEL.
2118 * Return: vm_fault_t value.
2120 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2121 unsigned long pfn, pgprot_t pgprot)
2124 * Technically, architectures with pte_special can avoid all these
2125 * restrictions (same for remap_pfn_range). However we would like
2126 * consistency in testing and feature parity among all, so we should
2127 * try to keep these invariants in place for everybody.
2129 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2130 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2131 (VM_PFNMAP|VM_MIXEDMAP));
2132 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2133 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2135 if (addr < vma->vm_start || addr >= vma->vm_end)
2136 return VM_FAULT_SIGBUS;
2138 if (!pfn_modify_allowed(pfn, pgprot))
2139 return VM_FAULT_SIGBUS;
2141 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2143 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2146 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2149 * vmf_insert_pfn - insert single pfn into user vma
2150 * @vma: user vma to map to
2151 * @addr: target user address of this page
2152 * @pfn: source kernel pfn
2154 * Similar to vm_insert_page, this allows drivers to insert individual pages
2155 * they've allocated into a user vma. Same comments apply.
2157 * This function should only be called from a vm_ops->fault handler, and
2158 * in that case the handler should return the result of this function.
2160 * vma cannot be a COW mapping.
2162 * As this is called only for pages that do not currently exist, we
2163 * do not need to flush old virtual caches or the TLB.
2165 * Context: Process context. May allocate using %GFP_KERNEL.
2166 * Return: vm_fault_t value.
2168 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2171 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2173 EXPORT_SYMBOL(vmf_insert_pfn);
2175 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2177 /* these checks mirror the abort conditions in vm_normal_page */
2178 if (vma->vm_flags & VM_MIXEDMAP)
2180 if (pfn_t_devmap(pfn))
2182 if (pfn_t_special(pfn))
2184 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2189 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2190 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2195 BUG_ON(!vm_mixed_ok(vma, pfn));
2197 if (addr < vma->vm_start || addr >= vma->vm_end)
2198 return VM_FAULT_SIGBUS;
2200 track_pfn_insert(vma, &pgprot, pfn);
2202 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2203 return VM_FAULT_SIGBUS;
2206 * If we don't have pte special, then we have to use the pfn_valid()
2207 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2208 * refcount the page if pfn_valid is true (hence insert_page rather
2209 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2210 * without pte special, it would there be refcounted as a normal page.
2212 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2213 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2217 * At this point we are committed to insert_page()
2218 * regardless of whether the caller specified flags that
2219 * result in pfn_t_has_page() == false.
2221 page = pfn_to_page(pfn_t_to_pfn(pfn));
2222 err = insert_page(vma, addr, page, pgprot);
2224 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2228 return VM_FAULT_OOM;
2229 if (err < 0 && err != -EBUSY)
2230 return VM_FAULT_SIGBUS;
2232 return VM_FAULT_NOPAGE;
2236 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2237 * @vma: user vma to map to
2238 * @addr: target user address of this page
2239 * @pfn: source kernel pfn
2240 * @pgprot: pgprot flags for the inserted page
2242 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2243 * to override pgprot on a per-page basis.
2245 * Typically this function should be used by drivers to set caching- and
2246 * encryption bits different than those of @vma->vm_page_prot, because
2247 * the caching- or encryption mode may not be known at mmap() time.
2248 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2249 * to set caching and encryption bits for those vmas (except for COW pages).
2250 * This is ensured by core vm only modifying these page table entries using
2251 * functions that don't touch caching- or encryption bits, using pte_modify()
2252 * if needed. (See for example mprotect()).
2253 * Also when new page-table entries are created, this is only done using the
2254 * fault() callback, and never using the value of vma->vm_page_prot,
2255 * except for page-table entries that point to anonymous pages as the result
2258 * Context: Process context. May allocate using %GFP_KERNEL.
2259 * Return: vm_fault_t value.
2261 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2262 pfn_t pfn, pgprot_t pgprot)
2264 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2266 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2268 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2271 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2273 EXPORT_SYMBOL(vmf_insert_mixed);
2276 * If the insertion of PTE failed because someone else already added a
2277 * different entry in the mean time, we treat that as success as we assume
2278 * the same entry was actually inserted.
2280 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2281 unsigned long addr, pfn_t pfn)
2283 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2285 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2288 * maps a range of physical memory into the requested pages. the old
2289 * mappings are removed. any references to nonexistent pages results
2290 * in null mappings (currently treated as "copy-on-access")
2292 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2293 unsigned long addr, unsigned long end,
2294 unsigned long pfn, pgprot_t prot)
2296 pte_t *pte, *mapped_pte;
2300 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2303 arch_enter_lazy_mmu_mode();
2305 BUG_ON(!pte_none(*pte));
2306 if (!pfn_modify_allowed(pfn, prot)) {
2310 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2312 } while (pte++, addr += PAGE_SIZE, addr != end);
2313 arch_leave_lazy_mmu_mode();
2314 pte_unmap_unlock(mapped_pte, ptl);
2318 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2319 unsigned long addr, unsigned long end,
2320 unsigned long pfn, pgprot_t prot)
2326 pfn -= addr >> PAGE_SHIFT;
2327 pmd = pmd_alloc(mm, pud, addr);
2330 VM_BUG_ON(pmd_trans_huge(*pmd));
2332 next = pmd_addr_end(addr, end);
2333 err = remap_pte_range(mm, pmd, addr, next,
2334 pfn + (addr >> PAGE_SHIFT), prot);
2337 } while (pmd++, addr = next, addr != end);
2341 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2342 unsigned long addr, unsigned long end,
2343 unsigned long pfn, pgprot_t prot)
2349 pfn -= addr >> PAGE_SHIFT;
2350 pud = pud_alloc(mm, p4d, addr);
2354 next = pud_addr_end(addr, end);
2355 err = remap_pmd_range(mm, pud, addr, next,
2356 pfn + (addr >> PAGE_SHIFT), prot);
2359 } while (pud++, addr = next, addr != end);
2363 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2364 unsigned long addr, unsigned long end,
2365 unsigned long pfn, pgprot_t prot)
2371 pfn -= addr >> PAGE_SHIFT;
2372 p4d = p4d_alloc(mm, pgd, addr);
2376 next = p4d_addr_end(addr, end);
2377 err = remap_pud_range(mm, p4d, addr, next,
2378 pfn + (addr >> PAGE_SHIFT), prot);
2381 } while (p4d++, addr = next, addr != end);
2386 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2387 * must have pre-validated the caching bits of the pgprot_t.
2389 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2390 unsigned long pfn, unsigned long size, pgprot_t prot)
2394 unsigned long end = addr + PAGE_ALIGN(size);
2395 struct mm_struct *mm = vma->vm_mm;
2398 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2402 * Physically remapped pages are special. Tell the
2403 * rest of the world about it:
2404 * VM_IO tells people not to look at these pages
2405 * (accesses can have side effects).
2406 * VM_PFNMAP tells the core MM that the base pages are just
2407 * raw PFN mappings, and do not have a "struct page" associated
2410 * Disable vma merging and expanding with mremap().
2412 * Omit vma from core dump, even when VM_IO turned off.
2414 * There's a horrible special case to handle copy-on-write
2415 * behaviour that some programs depend on. We mark the "original"
2416 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2417 * See vm_normal_page() for details.
2419 if (is_cow_mapping(vma->vm_flags)) {
2420 if (addr != vma->vm_start || end != vma->vm_end)
2422 vma->vm_pgoff = pfn;
2425 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2427 BUG_ON(addr >= end);
2428 pfn -= addr >> PAGE_SHIFT;
2429 pgd = pgd_offset(mm, addr);
2430 flush_cache_range(vma, addr, end);
2432 next = pgd_addr_end(addr, end);
2433 err = remap_p4d_range(mm, pgd, addr, next,
2434 pfn + (addr >> PAGE_SHIFT), prot);
2437 } while (pgd++, addr = next, addr != end);
2443 * remap_pfn_range - remap kernel memory to userspace
2444 * @vma: user vma to map to
2445 * @addr: target page aligned user address to start at
2446 * @pfn: page frame number of kernel physical memory address
2447 * @size: size of mapping area
2448 * @prot: page protection flags for this mapping
2450 * Note: this is only safe if the mm semaphore is held when called.
2452 * Return: %0 on success, negative error code otherwise.
2454 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2455 unsigned long pfn, unsigned long size, pgprot_t prot)
2459 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2463 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2465 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2468 EXPORT_SYMBOL(remap_pfn_range);
2471 * vm_iomap_memory - remap memory to userspace
2472 * @vma: user vma to map to
2473 * @start: start of the physical memory to be mapped
2474 * @len: size of area
2476 * This is a simplified io_remap_pfn_range() for common driver use. The
2477 * driver just needs to give us the physical memory range to be mapped,
2478 * we'll figure out the rest from the vma information.
2480 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2481 * whatever write-combining details or similar.
2483 * Return: %0 on success, negative error code otherwise.
2485 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2487 unsigned long vm_len, pfn, pages;
2489 /* Check that the physical memory area passed in looks valid */
2490 if (start + len < start)
2493 * You *really* shouldn't map things that aren't page-aligned,
2494 * but we've historically allowed it because IO memory might
2495 * just have smaller alignment.
2497 len += start & ~PAGE_MASK;
2498 pfn = start >> PAGE_SHIFT;
2499 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2500 if (pfn + pages < pfn)
2503 /* We start the mapping 'vm_pgoff' pages into the area */
2504 if (vma->vm_pgoff > pages)
2506 pfn += vma->vm_pgoff;
2507 pages -= vma->vm_pgoff;
2509 /* Can we fit all of the mapping? */
2510 vm_len = vma->vm_end - vma->vm_start;
2511 if (vm_len >> PAGE_SHIFT > pages)
2514 /* Ok, let it rip */
2515 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2517 EXPORT_SYMBOL(vm_iomap_memory);
2519 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2520 unsigned long addr, unsigned long end,
2521 pte_fn_t fn, void *data, bool create,
2522 pgtbl_mod_mask *mask)
2524 pte_t *pte, *mapped_pte;
2529 mapped_pte = pte = (mm == &init_mm) ?
2530 pte_alloc_kernel_track(pmd, addr, mask) :
2531 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2535 mapped_pte = pte = (mm == &init_mm) ?
2536 pte_offset_kernel(pmd, addr) :
2537 pte_offset_map_lock(mm, pmd, addr, &ptl);
2540 BUG_ON(pmd_huge(*pmd));
2542 arch_enter_lazy_mmu_mode();
2546 if (create || !pte_none(*pte)) {
2547 err = fn(pte++, addr, data);
2551 } while (addr += PAGE_SIZE, addr != end);
2553 *mask |= PGTBL_PTE_MODIFIED;
2555 arch_leave_lazy_mmu_mode();
2558 pte_unmap_unlock(mapped_pte, ptl);
2562 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2563 unsigned long addr, unsigned long end,
2564 pte_fn_t fn, void *data, bool create,
2565 pgtbl_mod_mask *mask)
2571 BUG_ON(pud_huge(*pud));
2574 pmd = pmd_alloc_track(mm, pud, addr, mask);
2578 pmd = pmd_offset(pud, addr);
2581 next = pmd_addr_end(addr, end);
2582 if (pmd_none(*pmd) && !create)
2584 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2586 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2591 err = apply_to_pte_range(mm, pmd, addr, next,
2592 fn, data, create, mask);
2595 } while (pmd++, addr = next, addr != end);
2600 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2601 unsigned long addr, unsigned long end,
2602 pte_fn_t fn, void *data, bool create,
2603 pgtbl_mod_mask *mask)
2610 pud = pud_alloc_track(mm, p4d, addr, mask);
2614 pud = pud_offset(p4d, addr);
2617 next = pud_addr_end(addr, end);
2618 if (pud_none(*pud) && !create)
2620 if (WARN_ON_ONCE(pud_leaf(*pud)))
2622 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2627 err = apply_to_pmd_range(mm, pud, addr, next,
2628 fn, data, create, mask);
2631 } while (pud++, addr = next, addr != end);
2636 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2637 unsigned long addr, unsigned long end,
2638 pte_fn_t fn, void *data, bool create,
2639 pgtbl_mod_mask *mask)
2646 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2650 p4d = p4d_offset(pgd, addr);
2653 next = p4d_addr_end(addr, end);
2654 if (p4d_none(*p4d) && !create)
2656 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2658 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2663 err = apply_to_pud_range(mm, p4d, addr, next,
2664 fn, data, create, mask);
2667 } while (p4d++, addr = next, addr != end);
2672 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2673 unsigned long size, pte_fn_t fn,
2674 void *data, bool create)
2677 unsigned long start = addr, next;
2678 unsigned long end = addr + size;
2679 pgtbl_mod_mask mask = 0;
2682 if (WARN_ON(addr >= end))
2685 pgd = pgd_offset(mm, addr);
2687 next = pgd_addr_end(addr, end);
2688 if (pgd_none(*pgd) && !create)
2690 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2692 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2697 err = apply_to_p4d_range(mm, pgd, addr, next,
2698 fn, data, create, &mask);
2701 } while (pgd++, addr = next, addr != end);
2703 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2704 arch_sync_kernel_mappings(start, start + size);
2710 * Scan a region of virtual memory, filling in page tables as necessary
2711 * and calling a provided function on each leaf page table.
2713 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2714 unsigned long size, pte_fn_t fn, void *data)
2716 return __apply_to_page_range(mm, addr, size, fn, data, true);
2718 EXPORT_SYMBOL_GPL(apply_to_page_range);
2721 * Scan a region of virtual memory, calling a provided function on
2722 * each leaf page table where it exists.
2724 * Unlike apply_to_page_range, this does _not_ fill in page tables
2725 * where they are absent.
2727 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2728 unsigned long size, pte_fn_t fn, void *data)
2730 return __apply_to_page_range(mm, addr, size, fn, data, false);
2732 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2735 * handle_pte_fault chooses page fault handler according to an entry which was
2736 * read non-atomically. Before making any commitment, on those architectures
2737 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2738 * parts, do_swap_page must check under lock before unmapping the pte and
2739 * proceeding (but do_wp_page is only called after already making such a check;
2740 * and do_anonymous_page can safely check later on).
2742 static inline int pte_unmap_same(struct vm_fault *vmf)
2745 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2746 if (sizeof(pte_t) > sizeof(unsigned long)) {
2747 spinlock_t *ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
2749 same = pte_same(*vmf->pte, vmf->orig_pte);
2753 pte_unmap(vmf->pte);
2758 static inline bool cow_user_page(struct page *dst, struct page *src,
2759 struct vm_fault *vmf)
2764 bool locked = false;
2765 struct vm_area_struct *vma = vmf->vma;
2766 struct mm_struct *mm = vma->vm_mm;
2767 unsigned long addr = vmf->address;
2770 copy_user_highpage(dst, src, addr, vma);
2775 * If the source page was a PFN mapping, we don't have
2776 * a "struct page" for it. We do a best-effort copy by
2777 * just copying from the original user address. If that
2778 * fails, we just zero-fill it. Live with it.
2780 kaddr = kmap_atomic(dst);
2781 uaddr = (void __user *)(addr & PAGE_MASK);
2784 * On architectures with software "accessed" bits, we would
2785 * take a double page fault, so mark it accessed here.
2787 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2790 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2792 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2794 * Other thread has already handled the fault
2795 * and update local tlb only
2797 update_mmu_tlb(vma, addr, vmf->pte);
2802 entry = pte_mkyoung(vmf->orig_pte);
2803 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2804 update_mmu_cache(vma, addr, vmf->pte);
2808 * This really shouldn't fail, because the page is there
2809 * in the page tables. But it might just be unreadable,
2810 * in which case we just give up and fill the result with
2813 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2817 /* Re-validate under PTL if the page is still mapped */
2818 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2820 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2821 /* The PTE changed under us, update local tlb */
2822 update_mmu_tlb(vma, addr, vmf->pte);
2828 * The same page can be mapped back since last copy attempt.
2829 * Try to copy again under PTL.
2831 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2833 * Give a warn in case there can be some obscure
2846 pte_unmap_unlock(vmf->pte, vmf->ptl);
2847 kunmap_atomic(kaddr);
2848 flush_dcache_page(dst);
2853 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2855 struct file *vm_file = vma->vm_file;
2858 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2861 * Special mappings (e.g. VDSO) do not have any file so fake
2862 * a default GFP_KERNEL for them.
2868 * Notify the address space that the page is about to become writable so that
2869 * it can prohibit this or wait for the page to get into an appropriate state.
2871 * We do this without the lock held, so that it can sleep if it needs to.
2873 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2876 struct page *page = vmf->page;
2877 unsigned int old_flags = vmf->flags;
2879 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2881 if (vmf->vma->vm_file &&
2882 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2883 return VM_FAULT_SIGBUS;
2885 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2886 /* Restore original flags so that caller is not surprised */
2887 vmf->flags = old_flags;
2888 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2890 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2892 if (!page->mapping) {
2894 return 0; /* retry */
2896 ret |= VM_FAULT_LOCKED;
2898 VM_BUG_ON_PAGE(!PageLocked(page), page);
2903 * Handle dirtying of a page in shared file mapping on a write fault.
2905 * The function expects the page to be locked and unlocks it.
2907 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2909 struct vm_area_struct *vma = vmf->vma;
2910 struct address_space *mapping;
2911 struct page *page = vmf->page;
2913 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2915 dirtied = set_page_dirty(page);
2916 VM_BUG_ON_PAGE(PageAnon(page), page);
2918 * Take a local copy of the address_space - page.mapping may be zeroed
2919 * by truncate after unlock_page(). The address_space itself remains
2920 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2921 * release semantics to prevent the compiler from undoing this copying.
2923 mapping = page_rmapping(page);
2927 file_update_time(vma->vm_file);
2930 * Throttle page dirtying rate down to writeback speed.
2932 * mapping may be NULL here because some device drivers do not
2933 * set page.mapping but still dirty their pages
2935 * Drop the mmap_lock before waiting on IO, if we can. The file
2936 * is pinning the mapping, as per above.
2938 if ((dirtied || page_mkwrite) && mapping) {
2941 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2942 balance_dirty_pages_ratelimited(mapping);
2945 return VM_FAULT_RETRY;
2953 * Handle write page faults for pages that can be reused in the current vma
2955 * This can happen either due to the mapping being with the VM_SHARED flag,
2956 * or due to us being the last reference standing to the page. In either
2957 * case, all we need to do here is to mark the page as writable and update
2958 * any related book-keeping.
2960 static inline void wp_page_reuse(struct vm_fault *vmf)
2961 __releases(vmf->ptl)
2963 struct vm_area_struct *vma = vmf->vma;
2964 struct page *page = vmf->page;
2967 * Clear the pages cpupid information as the existing
2968 * information potentially belongs to a now completely
2969 * unrelated process.
2972 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2974 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2975 entry = pte_mkyoung(vmf->orig_pte);
2976 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2977 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2978 update_mmu_cache(vma, vmf->address, vmf->pte);
2979 pte_unmap_unlock(vmf->pte, vmf->ptl);
2980 count_vm_event(PGREUSE);
2984 * Handle the case of a page which we actually need to copy to a new page.
2986 * Called with mmap_lock locked and the old page referenced, but
2987 * without the ptl held.
2989 * High level logic flow:
2991 * - Allocate a page, copy the content of the old page to the new one.
2992 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2993 * - Take the PTL. If the pte changed, bail out and release the allocated page
2994 * - If the pte is still the way we remember it, update the page table and all
2995 * relevant references. This includes dropping the reference the page-table
2996 * held to the old page, as well as updating the rmap.
2997 * - In any case, unlock the PTL and drop the reference we took to the old page.
2999 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3001 struct vm_area_struct *vma = vmf->vma;
3002 struct mm_struct *mm = vma->vm_mm;
3003 struct page *old_page = vmf->page;
3004 struct page *new_page = NULL;
3006 int page_copied = 0;
3007 struct mmu_notifier_range range;
3009 if (unlikely(anon_vma_prepare(vma)))
3012 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
3013 new_page = alloc_zeroed_user_highpage_movable(vma,
3018 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3023 if (!cow_user_page(new_page, old_page, vmf)) {
3025 * COW failed, if the fault was solved by other,
3026 * it's fine. If not, userspace would re-fault on
3027 * the same address and we will handle the fault
3028 * from the second attempt.
3037 if (mem_cgroup_charge(page_folio(new_page), mm, GFP_KERNEL))
3039 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
3041 __SetPageUptodate(new_page);
3043 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
3044 vmf->address & PAGE_MASK,
3045 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3046 mmu_notifier_invalidate_range_start(&range);
3049 * Re-check the pte - we dropped the lock
3051 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3052 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
3054 if (!PageAnon(old_page)) {
3055 dec_mm_counter_fast(mm,
3056 mm_counter_file(old_page));
3057 inc_mm_counter_fast(mm, MM_ANONPAGES);
3060 inc_mm_counter_fast(mm, MM_ANONPAGES);
3062 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3063 entry = mk_pte(new_page, vma->vm_page_prot);
3064 entry = pte_sw_mkyoung(entry);
3065 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3068 * Clear the pte entry and flush it first, before updating the
3069 * pte with the new entry, to keep TLBs on different CPUs in
3070 * sync. This code used to set the new PTE then flush TLBs, but
3071 * that left a window where the new PTE could be loaded into
3072 * some TLBs while the old PTE remains in others.
3074 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3075 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
3076 lru_cache_add_inactive_or_unevictable(new_page, vma);
3078 * We call the notify macro here because, when using secondary
3079 * mmu page tables (such as kvm shadow page tables), we want the
3080 * new page to be mapped directly into the secondary page table.
3082 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3083 update_mmu_cache(vma, vmf->address, vmf->pte);
3086 * Only after switching the pte to the new page may
3087 * we remove the mapcount here. Otherwise another
3088 * process may come and find the rmap count decremented
3089 * before the pte is switched to the new page, and
3090 * "reuse" the old page writing into it while our pte
3091 * here still points into it and can be read by other
3094 * The critical issue is to order this
3095 * page_remove_rmap with the ptp_clear_flush above.
3096 * Those stores are ordered by (if nothing else,)
3097 * the barrier present in the atomic_add_negative
3098 * in page_remove_rmap.
3100 * Then the TLB flush in ptep_clear_flush ensures that
3101 * no process can access the old page before the
3102 * decremented mapcount is visible. And the old page
3103 * cannot be reused until after the decremented
3104 * mapcount is visible. So transitively, TLBs to
3105 * old page will be flushed before it can be reused.
3107 page_remove_rmap(old_page, vma, false);
3110 /* Free the old page.. */
3111 new_page = old_page;
3114 update_mmu_tlb(vma, vmf->address, vmf->pte);
3120 pte_unmap_unlock(vmf->pte, vmf->ptl);
3122 * No need to double call mmu_notifier->invalidate_range() callback as
3123 * the above ptep_clear_flush_notify() did already call it.
3125 mmu_notifier_invalidate_range_only_end(&range);
3128 free_swap_cache(old_page);
3131 return page_copied ? VM_FAULT_WRITE : 0;
3137 return VM_FAULT_OOM;
3141 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3142 * writeable once the page is prepared
3144 * @vmf: structure describing the fault
3146 * This function handles all that is needed to finish a write page fault in a
3147 * shared mapping due to PTE being read-only once the mapped page is prepared.
3148 * It handles locking of PTE and modifying it.
3150 * The function expects the page to be locked or other protection against
3151 * concurrent faults / writeback (such as DAX radix tree locks).
3153 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3154 * we acquired PTE lock.
3156 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3158 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3159 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3162 * We might have raced with another page fault while we released the
3163 * pte_offset_map_lock.
3165 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3166 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3167 pte_unmap_unlock(vmf->pte, vmf->ptl);
3168 return VM_FAULT_NOPAGE;
3175 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3178 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3180 struct vm_area_struct *vma = vmf->vma;
3182 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3185 pte_unmap_unlock(vmf->pte, vmf->ptl);
3186 vmf->flags |= FAULT_FLAG_MKWRITE;
3187 ret = vma->vm_ops->pfn_mkwrite(vmf);
3188 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3190 return finish_mkwrite_fault(vmf);
3193 return VM_FAULT_WRITE;
3196 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3197 __releases(vmf->ptl)
3199 struct vm_area_struct *vma = vmf->vma;
3200 vm_fault_t ret = VM_FAULT_WRITE;
3202 get_page(vmf->page);
3204 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3207 pte_unmap_unlock(vmf->pte, vmf->ptl);
3208 tmp = do_page_mkwrite(vmf);
3209 if (unlikely(!tmp || (tmp &
3210 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3211 put_page(vmf->page);
3214 tmp = finish_mkwrite_fault(vmf);
3215 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3216 unlock_page(vmf->page);
3217 put_page(vmf->page);
3222 lock_page(vmf->page);
3224 ret |= fault_dirty_shared_page(vmf);
3225 put_page(vmf->page);
3231 * This routine handles present pages, when users try to write
3232 * to a shared page. It is done by copying the page to a new address
3233 * and decrementing the shared-page counter for the old page.
3235 * Note that this routine assumes that the protection checks have been
3236 * done by the caller (the low-level page fault routine in most cases).
3237 * Thus we can safely just mark it writable once we've done any necessary
3240 * We also mark the page dirty at this point even though the page will
3241 * change only once the write actually happens. This avoids a few races,
3242 * and potentially makes it more efficient.
3244 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3245 * but allow concurrent faults), with pte both mapped and locked.
3246 * We return with mmap_lock still held, but pte unmapped and unlocked.
3248 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3249 __releases(vmf->ptl)
3251 struct vm_area_struct *vma = vmf->vma;
3253 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3254 pte_unmap_unlock(vmf->pte, vmf->ptl);
3255 return handle_userfault(vmf, VM_UFFD_WP);
3259 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3260 * is flushed in this case before copying.
3262 if (unlikely(userfaultfd_wp(vmf->vma) &&
3263 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3264 flush_tlb_page(vmf->vma, vmf->address);
3266 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3269 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3272 * We should not cow pages in a shared writeable mapping.
3273 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3275 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3276 (VM_WRITE|VM_SHARED))
3277 return wp_pfn_shared(vmf);
3279 pte_unmap_unlock(vmf->pte, vmf->ptl);
3280 return wp_page_copy(vmf);
3284 * Take out anonymous pages first, anonymous shared vmas are
3285 * not dirty accountable.
3287 if (PageAnon(vmf->page)) {
3288 struct page *page = vmf->page;
3290 /* PageKsm() doesn't necessarily raise the page refcount */
3291 if (PageKsm(page) || page_count(page) != 1)
3293 if (!trylock_page(page))
3295 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3300 * Ok, we've got the only map reference, and the only
3301 * page count reference, and the page is locked,
3302 * it's dark out, and we're wearing sunglasses. Hit it.
3306 return VM_FAULT_WRITE;
3307 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3308 (VM_WRITE|VM_SHARED))) {
3309 return wp_page_shared(vmf);
3313 * Ok, we need to copy. Oh, well..
3315 get_page(vmf->page);
3317 pte_unmap_unlock(vmf->pte, vmf->ptl);
3318 return wp_page_copy(vmf);
3321 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3322 unsigned long start_addr, unsigned long end_addr,
3323 struct zap_details *details)
3325 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3328 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3329 pgoff_t first_index,
3331 struct zap_details *details)
3333 struct vm_area_struct *vma;
3334 pgoff_t vba, vea, zba, zea;
3336 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3337 vba = vma->vm_pgoff;
3338 vea = vba + vma_pages(vma) - 1;
3339 zba = max(first_index, vba);
3340 zea = min(last_index, vea);
3342 unmap_mapping_range_vma(vma,
3343 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3344 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3350 * unmap_mapping_folio() - Unmap single folio from processes.
3351 * @folio: The locked folio to be unmapped.
3353 * Unmap this folio from any userspace process which still has it mmaped.
3354 * Typically, for efficiency, the range of nearby pages has already been
3355 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3356 * truncation or invalidation holds the lock on a folio, it may find that
3357 * the page has been remapped again: and then uses unmap_mapping_folio()
3358 * to unmap it finally.
3360 void unmap_mapping_folio(struct folio *folio)
3362 struct address_space *mapping = folio->mapping;
3363 struct zap_details details = { };
3364 pgoff_t first_index;
3367 VM_BUG_ON(!folio_test_locked(folio));
3369 first_index = folio->index;
3370 last_index = folio->index + folio_nr_pages(folio) - 1;
3372 details.even_cows = false;
3373 details.single_folio = folio;
3375 i_mmap_lock_write(mapping);
3376 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3377 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3378 last_index, &details);
3379 i_mmap_unlock_write(mapping);
3383 * unmap_mapping_pages() - Unmap pages from processes.
3384 * @mapping: The address space containing pages to be unmapped.
3385 * @start: Index of first page to be unmapped.
3386 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3387 * @even_cows: Whether to unmap even private COWed pages.
3389 * Unmap the pages in this address space from any userspace process which
3390 * has them mmaped. Generally, you want to remove COWed pages as well when
3391 * a file is being truncated, but not when invalidating pages from the page
3394 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3395 pgoff_t nr, bool even_cows)
3397 struct zap_details details = { };
3398 pgoff_t first_index = start;
3399 pgoff_t last_index = start + nr - 1;
3401 details.even_cows = even_cows;
3402 if (last_index < first_index)
3403 last_index = ULONG_MAX;
3405 i_mmap_lock_write(mapping);
3406 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3407 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3408 last_index, &details);
3409 i_mmap_unlock_write(mapping);
3411 EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3414 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3415 * address_space corresponding to the specified byte range in the underlying
3418 * @mapping: the address space containing mmaps to be unmapped.
3419 * @holebegin: byte in first page to unmap, relative to the start of
3420 * the underlying file. This will be rounded down to a PAGE_SIZE
3421 * boundary. Note that this is different from truncate_pagecache(), which
3422 * must keep the partial page. In contrast, we must get rid of
3424 * @holelen: size of prospective hole in bytes. This will be rounded
3425 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3427 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3428 * but 0 when invalidating pagecache, don't throw away private data.
3430 void unmap_mapping_range(struct address_space *mapping,
3431 loff_t const holebegin, loff_t const holelen, int even_cows)
3433 pgoff_t hba = holebegin >> PAGE_SHIFT;
3434 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3436 /* Check for overflow. */
3437 if (sizeof(holelen) > sizeof(hlen)) {
3439 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3440 if (holeend & ~(long long)ULONG_MAX)
3441 hlen = ULONG_MAX - hba + 1;
3444 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3446 EXPORT_SYMBOL(unmap_mapping_range);
3449 * Restore a potential device exclusive pte to a working pte entry
3451 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3453 struct page *page = vmf->page;
3454 struct vm_area_struct *vma = vmf->vma;
3455 struct mmu_notifier_range range;
3457 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
3458 return VM_FAULT_RETRY;
3459 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
3460 vma->vm_mm, vmf->address & PAGE_MASK,
3461 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3462 mmu_notifier_invalidate_range_start(&range);
3464 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3466 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3467 restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
3469 pte_unmap_unlock(vmf->pte, vmf->ptl);
3472 mmu_notifier_invalidate_range_end(&range);
3477 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3478 * but allow concurrent faults), and pte mapped but not yet locked.
3479 * We return with pte unmapped and unlocked.
3481 * We return with the mmap_lock locked or unlocked in the same cases
3482 * as does filemap_fault().
3484 vm_fault_t do_swap_page(struct vm_fault *vmf)
3486 struct vm_area_struct *vma = vmf->vma;
3487 struct page *page = NULL, *swapcache;
3488 struct swap_info_struct *si = NULL;
3494 void *shadow = NULL;
3496 if (!pte_unmap_same(vmf))
3499 entry = pte_to_swp_entry(vmf->orig_pte);
3500 if (unlikely(non_swap_entry(entry))) {
3501 if (is_migration_entry(entry)) {
3502 migration_entry_wait(vma->vm_mm, vmf->pmd,
3504 } else if (is_device_exclusive_entry(entry)) {
3505 vmf->page = pfn_swap_entry_to_page(entry);
3506 ret = remove_device_exclusive_entry(vmf);
3507 } else if (is_device_private_entry(entry)) {
3508 vmf->page = pfn_swap_entry_to_page(entry);
3509 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3510 } else if (is_hwpoison_entry(entry)) {
3511 ret = VM_FAULT_HWPOISON;
3513 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3514 ret = VM_FAULT_SIGBUS;
3519 /* Prevent swapoff from happening to us. */
3520 si = get_swap_device(entry);
3524 page = lookup_swap_cache(entry, vma, vmf->address);
3528 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3529 __swap_count(entry) == 1) {
3530 /* skip swapcache */
3531 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3534 __SetPageLocked(page);
3535 __SetPageSwapBacked(page);
3537 if (mem_cgroup_swapin_charge_page(page,
3538 vma->vm_mm, GFP_KERNEL, entry)) {
3542 mem_cgroup_swapin_uncharge_swap(entry);
3544 shadow = get_shadow_from_swap_cache(entry);
3546 workingset_refault(page_folio(page),
3549 lru_cache_add(page);
3551 /* To provide entry to swap_readpage() */
3552 set_page_private(page, entry.val);
3553 swap_readpage(page, true);
3554 set_page_private(page, 0);
3557 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3564 * Back out if somebody else faulted in this pte
3565 * while we released the pte lock.
3567 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3568 vmf->address, &vmf->ptl);
3569 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3574 /* Had to read the page from swap area: Major fault */
3575 ret = VM_FAULT_MAJOR;
3576 count_vm_event(PGMAJFAULT);
3577 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3578 } else if (PageHWPoison(page)) {
3580 * hwpoisoned dirty swapcache pages are kept for killing
3581 * owner processes (which may be unknown at hwpoison time)
3583 ret = VM_FAULT_HWPOISON;
3587 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3590 ret |= VM_FAULT_RETRY;
3595 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3596 * release the swapcache from under us. The page pin, and pte_same
3597 * test below, are not enough to exclude that. Even if it is still
3598 * swapcache, we need to check that the page's swap has not changed.
3600 if (unlikely((!PageSwapCache(page) ||
3601 page_private(page) != entry.val)) && swapcache)
3604 page = ksm_might_need_to_copy(page, vma, vmf->address);
3605 if (unlikely(!page)) {
3611 cgroup_throttle_swaprate(page, GFP_KERNEL);
3614 * Back out if somebody else already faulted in this pte.
3616 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3618 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3621 if (unlikely(!PageUptodate(page))) {
3622 ret = VM_FAULT_SIGBUS;
3627 * The page isn't present yet, go ahead with the fault.
3629 * Be careful about the sequence of operations here.
3630 * To get its accounting right, reuse_swap_page() must be called
3631 * while the page is counted on swap but not yet in mapcount i.e.
3632 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3633 * must be called after the swap_free(), or it will never succeed.
3636 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3637 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3638 pte = mk_pte(page, vma->vm_page_prot);
3639 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3640 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3641 vmf->flags &= ~FAULT_FLAG_WRITE;
3642 ret |= VM_FAULT_WRITE;
3643 exclusive = RMAP_EXCLUSIVE;
3645 flush_icache_page(vma, page);
3646 if (pte_swp_soft_dirty(vmf->orig_pte))
3647 pte = pte_mksoft_dirty(pte);
3648 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3649 pte = pte_mkuffd_wp(pte);
3650 pte = pte_wrprotect(pte);
3652 vmf->orig_pte = pte;
3654 /* ksm created a completely new copy */
3655 if (unlikely(page != swapcache && swapcache)) {
3656 page_add_new_anon_rmap(page, vma, vmf->address, false);
3657 lru_cache_add_inactive_or_unevictable(page, vma);
3659 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3662 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3663 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3666 if (mem_cgroup_swap_full(page) ||
3667 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3668 try_to_free_swap(page);
3670 if (page != swapcache && swapcache) {
3672 * Hold the lock to avoid the swap entry to be reused
3673 * until we take the PT lock for the pte_same() check
3674 * (to avoid false positives from pte_same). For
3675 * further safety release the lock after the swap_free
3676 * so that the swap count won't change under a
3677 * parallel locked swapcache.
3679 unlock_page(swapcache);
3680 put_page(swapcache);
3683 if (vmf->flags & FAULT_FLAG_WRITE) {
3684 ret |= do_wp_page(vmf);
3685 if (ret & VM_FAULT_ERROR)
3686 ret &= VM_FAULT_ERROR;
3690 /* No need to invalidate - it was non-present before */
3691 update_mmu_cache(vma, vmf->address, vmf->pte);
3693 pte_unmap_unlock(vmf->pte, vmf->ptl);
3696 put_swap_device(si);
3699 pte_unmap_unlock(vmf->pte, vmf->ptl);
3704 if (page != swapcache && swapcache) {
3705 unlock_page(swapcache);
3706 put_page(swapcache);
3709 put_swap_device(si);
3714 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3715 * but allow concurrent faults), and pte mapped but not yet locked.
3716 * We return with mmap_lock still held, but pte unmapped and unlocked.
3718 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3720 struct vm_area_struct *vma = vmf->vma;
3725 /* File mapping without ->vm_ops ? */
3726 if (vma->vm_flags & VM_SHARED)
3727 return VM_FAULT_SIGBUS;
3730 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3731 * pte_offset_map() on pmds where a huge pmd might be created
3732 * from a different thread.
3734 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3735 * parallel threads are excluded by other means.
3737 * Here we only have mmap_read_lock(mm).
3739 if (pte_alloc(vma->vm_mm, vmf->pmd))
3740 return VM_FAULT_OOM;
3742 /* See comment in handle_pte_fault() */
3743 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3746 /* Use the zero-page for reads */
3747 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3748 !mm_forbids_zeropage(vma->vm_mm)) {
3749 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3750 vma->vm_page_prot));
3751 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3752 vmf->address, &vmf->ptl);
3753 if (!pte_none(*vmf->pte)) {
3754 update_mmu_tlb(vma, vmf->address, vmf->pte);
3757 ret = check_stable_address_space(vma->vm_mm);
3760 /* Deliver the page fault to userland, check inside PT lock */
3761 if (userfaultfd_missing(vma)) {
3762 pte_unmap_unlock(vmf->pte, vmf->ptl);
3763 return handle_userfault(vmf, VM_UFFD_MISSING);
3768 /* Allocate our own private page. */
3769 if (unlikely(anon_vma_prepare(vma)))
3771 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3775 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
3777 cgroup_throttle_swaprate(page, GFP_KERNEL);
3780 * The memory barrier inside __SetPageUptodate makes sure that
3781 * preceding stores to the page contents become visible before
3782 * the set_pte_at() write.
3784 __SetPageUptodate(page);
3786 entry = mk_pte(page, vma->vm_page_prot);
3787 entry = pte_sw_mkyoung(entry);
3788 if (vma->vm_flags & VM_WRITE)
3789 entry = pte_mkwrite(pte_mkdirty(entry));
3791 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3793 if (!pte_none(*vmf->pte)) {
3794 update_mmu_cache(vma, vmf->address, vmf->pte);
3798 ret = check_stable_address_space(vma->vm_mm);
3802 /* Deliver the page fault to userland, check inside PT lock */
3803 if (userfaultfd_missing(vma)) {
3804 pte_unmap_unlock(vmf->pte, vmf->ptl);
3806 return handle_userfault(vmf, VM_UFFD_MISSING);
3809 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3810 page_add_new_anon_rmap(page, vma, vmf->address, false);
3811 lru_cache_add_inactive_or_unevictable(page, vma);
3813 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3815 /* No need to invalidate - it was non-present before */
3816 update_mmu_cache(vma, vmf->address, vmf->pte);
3818 pte_unmap_unlock(vmf->pte, vmf->ptl);
3826 return VM_FAULT_OOM;
3830 * The mmap_lock must have been held on entry, and may have been
3831 * released depending on flags and vma->vm_ops->fault() return value.
3832 * See filemap_fault() and __lock_page_retry().
3834 static vm_fault_t __do_fault(struct vm_fault *vmf)
3836 struct vm_area_struct *vma = vmf->vma;
3840 * Preallocate pte before we take page_lock because this might lead to
3841 * deadlocks for memcg reclaim which waits for pages under writeback:
3843 * SetPageWriteback(A)
3849 * wait_on_page_writeback(A)
3850 * SetPageWriteback(B)
3852 * # flush A, B to clear the writeback
3854 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3855 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3856 if (!vmf->prealloc_pte)
3857 return VM_FAULT_OOM;
3860 ret = vma->vm_ops->fault(vmf);
3861 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3862 VM_FAULT_DONE_COW)))
3865 if (unlikely(PageHWPoison(vmf->page))) {
3866 vm_fault_t poisonret = VM_FAULT_HWPOISON;
3867 if (ret & VM_FAULT_LOCKED) {
3868 /* Retry if a clean page was removed from the cache. */
3869 if (invalidate_inode_page(vmf->page))
3871 unlock_page(vmf->page);
3873 put_page(vmf->page);
3878 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3879 lock_page(vmf->page);
3881 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3886 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3887 static void deposit_prealloc_pte(struct vm_fault *vmf)
3889 struct vm_area_struct *vma = vmf->vma;
3891 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3893 * We are going to consume the prealloc table,
3894 * count that as nr_ptes.
3896 mm_inc_nr_ptes(vma->vm_mm);
3897 vmf->prealloc_pte = NULL;
3900 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3902 struct vm_area_struct *vma = vmf->vma;
3903 bool write = vmf->flags & FAULT_FLAG_WRITE;
3904 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3907 vm_fault_t ret = VM_FAULT_FALLBACK;
3909 if (!transhuge_vma_suitable(vma, haddr))
3912 page = compound_head(page);
3913 if (compound_order(page) != HPAGE_PMD_ORDER)
3917 * Just backoff if any subpage of a THP is corrupted otherwise
3918 * the corrupted page may mapped by PMD silently to escape the
3919 * check. This kind of THP just can be PTE mapped. Access to
3920 * the corrupted subpage should trigger SIGBUS as expected.
3922 if (unlikely(PageHasHWPoisoned(page)))
3926 * Archs like ppc64 need additional space to store information
3927 * related to pte entry. Use the preallocated table for that.
3929 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3930 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3931 if (!vmf->prealloc_pte)
3932 return VM_FAULT_OOM;
3935 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3936 if (unlikely(!pmd_none(*vmf->pmd)))
3939 for (i = 0; i < HPAGE_PMD_NR; i++)
3940 flush_icache_page(vma, page + i);
3942 entry = mk_huge_pmd(page, vma->vm_page_prot);
3944 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3946 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3947 page_add_file_rmap(page, vma, true);
3950 * deposit and withdraw with pmd lock held
3952 if (arch_needs_pgtable_deposit())
3953 deposit_prealloc_pte(vmf);
3955 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3957 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3959 /* fault is handled */
3961 count_vm_event(THP_FILE_MAPPED);
3963 spin_unlock(vmf->ptl);
3967 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3969 return VM_FAULT_FALLBACK;
3973 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3975 struct vm_area_struct *vma = vmf->vma;
3976 bool write = vmf->flags & FAULT_FLAG_WRITE;
3977 bool prefault = vmf->address != addr;
3980 flush_icache_page(vma, page);
3981 entry = mk_pte(page, vma->vm_page_prot);
3983 if (prefault && arch_wants_old_prefaulted_pte())
3984 entry = pte_mkold(entry);
3986 entry = pte_sw_mkyoung(entry);
3989 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3990 /* copy-on-write page */
3991 if (write && !(vma->vm_flags & VM_SHARED)) {
3992 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3993 page_add_new_anon_rmap(page, vma, addr, false);
3994 lru_cache_add_inactive_or_unevictable(page, vma);
3996 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3997 page_add_file_rmap(page, vma, false);
3999 set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
4003 * finish_fault - finish page fault once we have prepared the page to fault
4005 * @vmf: structure describing the fault
4007 * This function handles all that is needed to finish a page fault once the
4008 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4009 * given page, adds reverse page mapping, handles memcg charges and LRU
4012 * The function expects the page to be locked and on success it consumes a
4013 * reference of a page being mapped (for the PTE which maps it).
4015 * Return: %0 on success, %VM_FAULT_ code in case of error.
4017 vm_fault_t finish_fault(struct vm_fault *vmf)
4019 struct vm_area_struct *vma = vmf->vma;
4023 /* Did we COW the page? */
4024 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4025 page = vmf->cow_page;
4030 * check even for read faults because we might have lost our CoWed
4033 if (!(vma->vm_flags & VM_SHARED)) {
4034 ret = check_stable_address_space(vma->vm_mm);
4039 if (pmd_none(*vmf->pmd)) {
4040 if (PageTransCompound(page)) {
4041 ret = do_set_pmd(vmf, page);
4042 if (ret != VM_FAULT_FALLBACK)
4046 if (vmf->prealloc_pte)
4047 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4048 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4049 return VM_FAULT_OOM;
4052 /* See comment in handle_pte_fault() */
4053 if (pmd_devmap_trans_unstable(vmf->pmd))
4056 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4057 vmf->address, &vmf->ptl);
4059 /* Re-check under ptl */
4060 if (likely(pte_none(*vmf->pte)))
4061 do_set_pte(vmf, page, vmf->address);
4063 ret = VM_FAULT_NOPAGE;
4065 update_mmu_tlb(vma, vmf->address, vmf->pte);
4066 pte_unmap_unlock(vmf->pte, vmf->ptl);
4070 static unsigned long fault_around_bytes __read_mostly =
4071 rounddown_pow_of_two(65536);
4073 #ifdef CONFIG_DEBUG_FS
4074 static int fault_around_bytes_get(void *data, u64 *val)
4076 *val = fault_around_bytes;
4081 * fault_around_bytes must be rounded down to the nearest page order as it's
4082 * what do_fault_around() expects to see.
4084 static int fault_around_bytes_set(void *data, u64 val)
4086 if (val / PAGE_SIZE > PTRS_PER_PTE)
4088 if (val > PAGE_SIZE)
4089 fault_around_bytes = rounddown_pow_of_two(val);
4091 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
4094 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4095 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4097 static int __init fault_around_debugfs(void)
4099 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4100 &fault_around_bytes_fops);
4103 late_initcall(fault_around_debugfs);
4107 * do_fault_around() tries to map few pages around the fault address. The hope
4108 * is that the pages will be needed soon and this will lower the number of
4111 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4112 * not ready to be mapped: not up-to-date, locked, etc.
4114 * This function is called with the page table lock taken. In the split ptlock
4115 * case the page table lock only protects only those entries which belong to
4116 * the page table corresponding to the fault address.
4118 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4121 * fault_around_bytes defines how many bytes we'll try to map.
4122 * do_fault_around() expects it to be set to a power of two less than or equal
4125 * The virtual address of the area that we map is naturally aligned to
4126 * fault_around_bytes rounded down to the machine page size
4127 * (and therefore to page order). This way it's easier to guarantee
4128 * that we don't cross page table boundaries.
4130 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4132 unsigned long address = vmf->address, nr_pages, mask;
4133 pgoff_t start_pgoff = vmf->pgoff;
4137 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4138 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4140 address = max(address & mask, vmf->vma->vm_start);
4141 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4145 * end_pgoff is either the end of the page table, the end of
4146 * the vma or nr_pages from start_pgoff, depending what is nearest.
4148 end_pgoff = start_pgoff -
4149 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4151 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4152 start_pgoff + nr_pages - 1);
4154 if (pmd_none(*vmf->pmd)) {
4155 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4156 if (!vmf->prealloc_pte)
4157 return VM_FAULT_OOM;
4160 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4163 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4165 struct vm_area_struct *vma = vmf->vma;
4169 * Let's call ->map_pages() first and use ->fault() as fallback
4170 * if page by the offset is not ready to be mapped (cold cache or
4173 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4174 if (likely(!userfaultfd_minor(vmf->vma))) {
4175 ret = do_fault_around(vmf);
4181 ret = __do_fault(vmf);
4182 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4185 ret |= finish_fault(vmf);
4186 unlock_page(vmf->page);
4187 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4188 put_page(vmf->page);
4192 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4194 struct vm_area_struct *vma = vmf->vma;
4197 if (unlikely(anon_vma_prepare(vma)))
4198 return VM_FAULT_OOM;
4200 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4202 return VM_FAULT_OOM;
4204 if (mem_cgroup_charge(page_folio(vmf->cow_page), vma->vm_mm,
4206 put_page(vmf->cow_page);
4207 return VM_FAULT_OOM;
4209 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4211 ret = __do_fault(vmf);
4212 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4214 if (ret & VM_FAULT_DONE_COW)
4217 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4218 __SetPageUptodate(vmf->cow_page);
4220 ret |= finish_fault(vmf);
4221 unlock_page(vmf->page);
4222 put_page(vmf->page);
4223 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4227 put_page(vmf->cow_page);
4231 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4233 struct vm_area_struct *vma = vmf->vma;
4234 vm_fault_t ret, tmp;
4236 ret = __do_fault(vmf);
4237 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4241 * Check if the backing address space wants to know that the page is
4242 * about to become writable
4244 if (vma->vm_ops->page_mkwrite) {
4245 unlock_page(vmf->page);
4246 tmp = do_page_mkwrite(vmf);
4247 if (unlikely(!tmp ||
4248 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4249 put_page(vmf->page);
4254 ret |= finish_fault(vmf);
4255 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4257 unlock_page(vmf->page);
4258 put_page(vmf->page);
4262 ret |= fault_dirty_shared_page(vmf);
4267 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4268 * but allow concurrent faults).
4269 * The mmap_lock may have been released depending on flags and our
4270 * return value. See filemap_fault() and __folio_lock_or_retry().
4271 * If mmap_lock is released, vma may become invalid (for example
4272 * by other thread calling munmap()).
4274 static vm_fault_t do_fault(struct vm_fault *vmf)
4276 struct vm_area_struct *vma = vmf->vma;
4277 struct mm_struct *vm_mm = vma->vm_mm;
4281 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4283 if (!vma->vm_ops->fault) {
4285 * If we find a migration pmd entry or a none pmd entry, which
4286 * should never happen, return SIGBUS
4288 if (unlikely(!pmd_present(*vmf->pmd)))
4289 ret = VM_FAULT_SIGBUS;
4291 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4296 * Make sure this is not a temporary clearing of pte
4297 * by holding ptl and checking again. A R/M/W update
4298 * of pte involves: take ptl, clearing the pte so that
4299 * we don't have concurrent modification by hardware
4300 * followed by an update.
4302 if (unlikely(pte_none(*vmf->pte)))
4303 ret = VM_FAULT_SIGBUS;
4305 ret = VM_FAULT_NOPAGE;
4307 pte_unmap_unlock(vmf->pte, vmf->ptl);
4309 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4310 ret = do_read_fault(vmf);
4311 else if (!(vma->vm_flags & VM_SHARED))
4312 ret = do_cow_fault(vmf);
4314 ret = do_shared_fault(vmf);
4316 /* preallocated pagetable is unused: free it */
4317 if (vmf->prealloc_pte) {
4318 pte_free(vm_mm, vmf->prealloc_pte);
4319 vmf->prealloc_pte = NULL;
4324 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4325 unsigned long addr, int page_nid, int *flags)
4329 count_vm_numa_event(NUMA_HINT_FAULTS);
4330 if (page_nid == numa_node_id()) {
4331 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4332 *flags |= TNF_FAULT_LOCAL;
4335 return mpol_misplaced(page, vma, addr);
4338 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4340 struct vm_area_struct *vma = vmf->vma;
4341 struct page *page = NULL;
4342 int page_nid = NUMA_NO_NODE;
4346 bool was_writable = pte_savedwrite(vmf->orig_pte);
4350 * The "pte" at this point cannot be used safely without
4351 * validation through pte_unmap_same(). It's of NUMA type but
4352 * the pfn may be screwed if the read is non atomic.
4354 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4355 spin_lock(vmf->ptl);
4356 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4357 pte_unmap_unlock(vmf->pte, vmf->ptl);
4361 /* Get the normal PTE */
4362 old_pte = ptep_get(vmf->pte);
4363 pte = pte_modify(old_pte, vma->vm_page_prot);
4365 page = vm_normal_page(vma, vmf->address, pte);
4369 /* TODO: handle PTE-mapped THP */
4370 if (PageCompound(page))
4374 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4375 * much anyway since they can be in shared cache state. This misses
4376 * the case where a mapping is writable but the process never writes
4377 * to it but pte_write gets cleared during protection updates and
4378 * pte_dirty has unpredictable behaviour between PTE scan updates,
4379 * background writeback, dirty balancing and application behaviour.
4382 flags |= TNF_NO_GROUP;
4385 * Flag if the page is shared between multiple address spaces. This
4386 * is later used when determining whether to group tasks together
4388 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4389 flags |= TNF_SHARED;
4391 last_cpupid = page_cpupid_last(page);
4392 page_nid = page_to_nid(page);
4393 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4395 if (target_nid == NUMA_NO_NODE) {
4399 pte_unmap_unlock(vmf->pte, vmf->ptl);
4401 /* Migrate to the requested node */
4402 if (migrate_misplaced_page(page, vma, target_nid)) {
4403 page_nid = target_nid;
4404 flags |= TNF_MIGRATED;
4406 flags |= TNF_MIGRATE_FAIL;
4407 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4408 spin_lock(vmf->ptl);
4409 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4410 pte_unmap_unlock(vmf->pte, vmf->ptl);
4417 if (page_nid != NUMA_NO_NODE)
4418 task_numa_fault(last_cpupid, page_nid, 1, flags);
4422 * Make it present again, depending on how arch implements
4423 * non-accessible ptes, some can allow access by kernel mode.
4425 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4426 pte = pte_modify(old_pte, vma->vm_page_prot);
4427 pte = pte_mkyoung(pte);
4429 pte = pte_mkwrite(pte);
4430 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4431 update_mmu_cache(vma, vmf->address, vmf->pte);
4432 pte_unmap_unlock(vmf->pte, vmf->ptl);
4436 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4438 if (vma_is_anonymous(vmf->vma))
4439 return do_huge_pmd_anonymous_page(vmf);
4440 if (vmf->vma->vm_ops->huge_fault)
4441 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4442 return VM_FAULT_FALLBACK;
4445 /* `inline' is required to avoid gcc 4.1.2 build error */
4446 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4448 if (vma_is_anonymous(vmf->vma)) {
4449 if (userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd))
4450 return handle_userfault(vmf, VM_UFFD_WP);
4451 return do_huge_pmd_wp_page(vmf);
4453 if (vmf->vma->vm_ops->huge_fault) {
4454 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4456 if (!(ret & VM_FAULT_FALLBACK))
4460 /* COW or write-notify handled on pte level: split pmd. */
4461 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4463 return VM_FAULT_FALLBACK;
4466 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4468 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4469 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4470 /* No support for anonymous transparent PUD pages yet */
4471 if (vma_is_anonymous(vmf->vma))
4473 if (vmf->vma->vm_ops->huge_fault) {
4474 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4476 if (!(ret & VM_FAULT_FALLBACK))
4480 /* COW or write-notify not handled on PUD level: split pud.*/
4481 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4482 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4483 return VM_FAULT_FALLBACK;
4486 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4488 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4489 /* No support for anonymous transparent PUD pages yet */
4490 if (vma_is_anonymous(vmf->vma))
4491 return VM_FAULT_FALLBACK;
4492 if (vmf->vma->vm_ops->huge_fault)
4493 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4494 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4495 return VM_FAULT_FALLBACK;
4499 * These routines also need to handle stuff like marking pages dirty
4500 * and/or accessed for architectures that don't do it in hardware (most
4501 * RISC architectures). The early dirtying is also good on the i386.
4503 * There is also a hook called "update_mmu_cache()" that architectures
4504 * with external mmu caches can use to update those (ie the Sparc or
4505 * PowerPC hashed page tables that act as extended TLBs).
4507 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4508 * concurrent faults).
4510 * The mmap_lock may have been released depending on flags and our return value.
4511 * See filemap_fault() and __folio_lock_or_retry().
4513 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4517 if (unlikely(pmd_none(*vmf->pmd))) {
4519 * Leave __pte_alloc() until later: because vm_ops->fault may
4520 * want to allocate huge page, and if we expose page table
4521 * for an instant, it will be difficult to retract from
4522 * concurrent faults and from rmap lookups.
4527 * If a huge pmd materialized under us just retry later. Use
4528 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4529 * of pmd_trans_huge() to ensure the pmd didn't become
4530 * pmd_trans_huge under us and then back to pmd_none, as a
4531 * result of MADV_DONTNEED running immediately after a huge pmd
4532 * fault in a different thread of this mm, in turn leading to a
4533 * misleading pmd_trans_huge() retval. All we have to ensure is
4534 * that it is a regular pmd that we can walk with
4535 * pte_offset_map() and we can do that through an atomic read
4536 * in C, which is what pmd_trans_unstable() provides.
4538 if (pmd_devmap_trans_unstable(vmf->pmd))
4541 * A regular pmd is established and it can't morph into a huge
4542 * pmd from under us anymore at this point because we hold the
4543 * mmap_lock read mode and khugepaged takes it in write mode.
4544 * So now it's safe to run pte_offset_map().
4546 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4547 vmf->orig_pte = *vmf->pte;
4550 * some architectures can have larger ptes than wordsize,
4551 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4552 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4553 * accesses. The code below just needs a consistent view
4554 * for the ifs and we later double check anyway with the
4555 * ptl lock held. So here a barrier will do.
4558 if (pte_none(vmf->orig_pte)) {
4559 pte_unmap(vmf->pte);
4565 if (vma_is_anonymous(vmf->vma))
4566 return do_anonymous_page(vmf);
4568 return do_fault(vmf);
4571 if (!pte_present(vmf->orig_pte))
4572 return do_swap_page(vmf);
4574 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4575 return do_numa_page(vmf);
4577 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4578 spin_lock(vmf->ptl);
4579 entry = vmf->orig_pte;
4580 if (unlikely(!pte_same(*vmf->pte, entry))) {
4581 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4584 if (vmf->flags & FAULT_FLAG_WRITE) {
4585 if (!pte_write(entry))
4586 return do_wp_page(vmf);
4587 entry = pte_mkdirty(entry);
4589 entry = pte_mkyoung(entry);
4590 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4591 vmf->flags & FAULT_FLAG_WRITE)) {
4592 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4594 /* Skip spurious TLB flush for retried page fault */
4595 if (vmf->flags & FAULT_FLAG_TRIED)
4598 * This is needed only for protection faults but the arch code
4599 * is not yet telling us if this is a protection fault or not.
4600 * This still avoids useless tlb flushes for .text page faults
4603 if (vmf->flags & FAULT_FLAG_WRITE)
4604 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4607 pte_unmap_unlock(vmf->pte, vmf->ptl);
4612 * By the time we get here, we already hold the mm semaphore
4614 * The mmap_lock may have been released depending on flags and our
4615 * return value. See filemap_fault() and __folio_lock_or_retry().
4617 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4618 unsigned long address, unsigned int flags)
4620 struct vm_fault vmf = {
4622 .address = address & PAGE_MASK,
4623 .real_address = address,
4625 .pgoff = linear_page_index(vma, address),
4626 .gfp_mask = __get_fault_gfp_mask(vma),
4628 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4629 struct mm_struct *mm = vma->vm_mm;
4634 pgd = pgd_offset(mm, address);
4635 p4d = p4d_alloc(mm, pgd, address);
4637 return VM_FAULT_OOM;
4639 vmf.pud = pud_alloc(mm, p4d, address);
4641 return VM_FAULT_OOM;
4643 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4644 ret = create_huge_pud(&vmf);
4645 if (!(ret & VM_FAULT_FALLBACK))
4648 pud_t orig_pud = *vmf.pud;
4651 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4653 /* NUMA case for anonymous PUDs would go here */
4655 if (dirty && !pud_write(orig_pud)) {
4656 ret = wp_huge_pud(&vmf, orig_pud);
4657 if (!(ret & VM_FAULT_FALLBACK))
4660 huge_pud_set_accessed(&vmf, orig_pud);
4666 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4668 return VM_FAULT_OOM;
4670 /* Huge pud page fault raced with pmd_alloc? */
4671 if (pud_trans_unstable(vmf.pud))
4674 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4675 ret = create_huge_pmd(&vmf);
4676 if (!(ret & VM_FAULT_FALLBACK))
4679 vmf.orig_pmd = *vmf.pmd;
4682 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
4683 VM_BUG_ON(thp_migration_supported() &&
4684 !is_pmd_migration_entry(vmf.orig_pmd));
4685 if (is_pmd_migration_entry(vmf.orig_pmd))
4686 pmd_migration_entry_wait(mm, vmf.pmd);
4689 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
4690 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
4691 return do_huge_pmd_numa_page(&vmf);
4693 if (dirty && !pmd_write(vmf.orig_pmd)) {
4694 ret = wp_huge_pmd(&vmf);
4695 if (!(ret & VM_FAULT_FALLBACK))
4698 huge_pmd_set_accessed(&vmf);
4704 return handle_pte_fault(&vmf);
4708 * mm_account_fault - Do page fault accounting
4710 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4711 * of perf event counters, but we'll still do the per-task accounting to
4712 * the task who triggered this page fault.
4713 * @address: the faulted address.
4714 * @flags: the fault flags.
4715 * @ret: the fault retcode.
4717 * This will take care of most of the page fault accounting. Meanwhile, it
4718 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4719 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4720 * still be in per-arch page fault handlers at the entry of page fault.
4722 static inline void mm_account_fault(struct pt_regs *regs,
4723 unsigned long address, unsigned int flags,
4729 * We don't do accounting for some specific faults:
4731 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4732 * includes arch_vma_access_permitted() failing before reaching here.
4733 * So this is not a "this many hardware page faults" counter. We
4734 * should use the hw profiling for that.
4736 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4737 * once they're completed.
4739 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4743 * We define the fault as a major fault when the final successful fault
4744 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4745 * handle it immediately previously).
4747 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4755 * If the fault is done for GUP, regs will be NULL. We only do the
4756 * accounting for the per thread fault counters who triggered the
4757 * fault, and we skip the perf event updates.
4763 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4765 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4769 * By the time we get here, we already hold the mm semaphore
4771 * The mmap_lock may have been released depending on flags and our
4772 * return value. See filemap_fault() and __folio_lock_or_retry().
4774 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4775 unsigned int flags, struct pt_regs *regs)
4779 __set_current_state(TASK_RUNNING);
4781 count_vm_event(PGFAULT);
4782 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4784 /* do counter updates before entering really critical section. */
4785 check_sync_rss_stat(current);
4787 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4788 flags & FAULT_FLAG_INSTRUCTION,
4789 flags & FAULT_FLAG_REMOTE))
4790 return VM_FAULT_SIGSEGV;
4793 * Enable the memcg OOM handling for faults triggered in user
4794 * space. Kernel faults are handled more gracefully.
4796 if (flags & FAULT_FLAG_USER)
4797 mem_cgroup_enter_user_fault();
4799 if (unlikely(is_vm_hugetlb_page(vma)))
4800 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4802 ret = __handle_mm_fault(vma, address, flags);
4804 if (flags & FAULT_FLAG_USER) {
4805 mem_cgroup_exit_user_fault();
4807 * The task may have entered a memcg OOM situation but
4808 * if the allocation error was handled gracefully (no
4809 * VM_FAULT_OOM), there is no need to kill anything.
4810 * Just clean up the OOM state peacefully.
4812 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4813 mem_cgroup_oom_synchronize(false);
4816 mm_account_fault(regs, address, flags, ret);
4820 EXPORT_SYMBOL_GPL(handle_mm_fault);
4822 #ifndef __PAGETABLE_P4D_FOLDED
4824 * Allocate p4d page table.
4825 * We've already handled the fast-path in-line.
4827 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4829 p4d_t *new = p4d_alloc_one(mm, address);
4833 spin_lock(&mm->page_table_lock);
4834 if (pgd_present(*pgd)) { /* Another has populated it */
4837 smp_wmb(); /* See comment in pmd_install() */
4838 pgd_populate(mm, pgd, new);
4840 spin_unlock(&mm->page_table_lock);
4843 #endif /* __PAGETABLE_P4D_FOLDED */
4845 #ifndef __PAGETABLE_PUD_FOLDED
4847 * Allocate page upper directory.
4848 * We've already handled the fast-path in-line.
4850 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4852 pud_t *new = pud_alloc_one(mm, address);
4856 spin_lock(&mm->page_table_lock);
4857 if (!p4d_present(*p4d)) {
4859 smp_wmb(); /* See comment in pmd_install() */
4860 p4d_populate(mm, p4d, new);
4861 } else /* Another has populated it */
4863 spin_unlock(&mm->page_table_lock);
4866 #endif /* __PAGETABLE_PUD_FOLDED */
4868 #ifndef __PAGETABLE_PMD_FOLDED
4870 * Allocate page middle directory.
4871 * We've already handled the fast-path in-line.
4873 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4876 pmd_t *new = pmd_alloc_one(mm, address);
4880 ptl = pud_lock(mm, pud);
4881 if (!pud_present(*pud)) {
4883 smp_wmb(); /* See comment in pmd_install() */
4884 pud_populate(mm, pud, new);
4885 } else { /* Another has populated it */
4891 #endif /* __PAGETABLE_PMD_FOLDED */
4893 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4894 struct mmu_notifier_range *range, pte_t **ptepp,
4895 pmd_t **pmdpp, spinlock_t **ptlp)
4903 pgd = pgd_offset(mm, address);
4904 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4907 p4d = p4d_offset(pgd, address);
4908 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4911 pud = pud_offset(p4d, address);
4912 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4915 pmd = pmd_offset(pud, address);
4916 VM_BUG_ON(pmd_trans_huge(*pmd));
4918 if (pmd_huge(*pmd)) {
4923 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4924 NULL, mm, address & PMD_MASK,
4925 (address & PMD_MASK) + PMD_SIZE);
4926 mmu_notifier_invalidate_range_start(range);
4928 *ptlp = pmd_lock(mm, pmd);
4929 if (pmd_huge(*pmd)) {
4935 mmu_notifier_invalidate_range_end(range);
4938 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4942 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4943 address & PAGE_MASK,
4944 (address & PAGE_MASK) + PAGE_SIZE);
4945 mmu_notifier_invalidate_range_start(range);
4947 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4948 if (!pte_present(*ptep))
4953 pte_unmap_unlock(ptep, *ptlp);
4955 mmu_notifier_invalidate_range_end(range);
4961 * follow_pte - look up PTE at a user virtual address
4962 * @mm: the mm_struct of the target address space
4963 * @address: user virtual address
4964 * @ptepp: location to store found PTE
4965 * @ptlp: location to store the lock for the PTE
4967 * On a successful return, the pointer to the PTE is stored in @ptepp;
4968 * the corresponding lock is taken and its location is stored in @ptlp.
4969 * The contents of the PTE are only stable until @ptlp is released;
4970 * any further use, if any, must be protected against invalidation
4971 * with MMU notifiers.
4973 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4974 * should be taken for read.
4976 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4977 * it is not a good general-purpose API.
4979 * Return: zero on success, -ve otherwise.
4981 int follow_pte(struct mm_struct *mm, unsigned long address,
4982 pte_t **ptepp, spinlock_t **ptlp)
4984 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4986 EXPORT_SYMBOL_GPL(follow_pte);
4989 * follow_pfn - look up PFN at a user virtual address
4990 * @vma: memory mapping
4991 * @address: user virtual address
4992 * @pfn: location to store found PFN
4994 * Only IO mappings and raw PFN mappings are allowed.
4996 * This function does not allow the caller to read the permissions
4997 * of the PTE. Do not use it.
4999 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5001 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5008 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5011 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5014 *pfn = pte_pfn(*ptep);
5015 pte_unmap_unlock(ptep, ptl);
5018 EXPORT_SYMBOL(follow_pfn);
5020 #ifdef CONFIG_HAVE_IOREMAP_PROT
5021 int follow_phys(struct vm_area_struct *vma,
5022 unsigned long address, unsigned int flags,
5023 unsigned long *prot, resource_size_t *phys)
5029 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5032 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5036 if ((flags & FOLL_WRITE) && !pte_write(pte))
5039 *prot = pgprot_val(pte_pgprot(pte));
5040 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5044 pte_unmap_unlock(ptep, ptl);
5050 * generic_access_phys - generic implementation for iomem mmap access
5051 * @vma: the vma to access
5052 * @addr: userspace address, not relative offset within @vma
5053 * @buf: buffer to read/write
5054 * @len: length of transfer
5055 * @write: set to FOLL_WRITE when writing, otherwise reading
5057 * This is a generic implementation for &vm_operations_struct.access for an
5058 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5061 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5062 void *buf, int len, int write)
5064 resource_size_t phys_addr;
5065 unsigned long prot = 0;
5066 void __iomem *maddr;
5069 int offset = offset_in_page(addr);
5072 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5076 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5079 pte_unmap_unlock(ptep, ptl);
5081 prot = pgprot_val(pte_pgprot(pte));
5082 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5084 if ((write & FOLL_WRITE) && !pte_write(pte))
5087 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5091 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5094 if (!pte_same(pte, *ptep)) {
5095 pte_unmap_unlock(ptep, ptl);
5102 memcpy_toio(maddr + offset, buf, len);
5104 memcpy_fromio(buf, maddr + offset, len);
5106 pte_unmap_unlock(ptep, ptl);
5112 EXPORT_SYMBOL_GPL(generic_access_phys);
5116 * Access another process' address space as given in mm.
5118 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5119 int len, unsigned int gup_flags)
5121 struct vm_area_struct *vma;
5122 void *old_buf = buf;
5123 int write = gup_flags & FOLL_WRITE;
5125 if (mmap_read_lock_killable(mm))
5128 /* ignore errors, just check how much was successfully transferred */
5130 int bytes, ret, offset;
5132 struct page *page = NULL;
5134 ret = get_user_pages_remote(mm, addr, 1,
5135 gup_flags, &page, &vma, NULL);
5137 #ifndef CONFIG_HAVE_IOREMAP_PROT
5141 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5142 * we can access using slightly different code.
5144 vma = vma_lookup(mm, addr);
5147 if (vma->vm_ops && vma->vm_ops->access)
5148 ret = vma->vm_ops->access(vma, addr, buf,
5156 offset = addr & (PAGE_SIZE-1);
5157 if (bytes > PAGE_SIZE-offset)
5158 bytes = PAGE_SIZE-offset;
5162 copy_to_user_page(vma, page, addr,
5163 maddr + offset, buf, bytes);
5164 set_page_dirty_lock(page);
5166 copy_from_user_page(vma, page, addr,
5167 buf, maddr + offset, bytes);
5176 mmap_read_unlock(mm);
5178 return buf - old_buf;
5182 * access_remote_vm - access another process' address space
5183 * @mm: the mm_struct of the target address space
5184 * @addr: start address to access
5185 * @buf: source or destination buffer
5186 * @len: number of bytes to transfer
5187 * @gup_flags: flags modifying lookup behaviour
5189 * The caller must hold a reference on @mm.
5191 * Return: number of bytes copied from source to destination.
5193 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5194 void *buf, int len, unsigned int gup_flags)
5196 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5200 * Access another process' address space.
5201 * Source/target buffer must be kernel space,
5202 * Do not walk the page table directly, use get_user_pages
5204 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5205 void *buf, int len, unsigned int gup_flags)
5207 struct mm_struct *mm;
5210 mm = get_task_mm(tsk);
5214 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5220 EXPORT_SYMBOL_GPL(access_process_vm);
5223 * Print the name of a VMA.
5225 void print_vma_addr(char *prefix, unsigned long ip)
5227 struct mm_struct *mm = current->mm;
5228 struct vm_area_struct *vma;
5231 * we might be running from an atomic context so we cannot sleep
5233 if (!mmap_read_trylock(mm))
5236 vma = find_vma(mm, ip);
5237 if (vma && vma->vm_file) {
5238 struct file *f = vma->vm_file;
5239 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5243 p = file_path(f, buf, PAGE_SIZE);
5246 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5248 vma->vm_end - vma->vm_start);
5249 free_page((unsigned long)buf);
5252 mmap_read_unlock(mm);
5255 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5256 void __might_fault(const char *file, int line)
5259 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5260 * holding the mmap_lock, this is safe because kernel memory doesn't
5261 * get paged out, therefore we'll never actually fault, and the
5262 * below annotations will generate false positives.
5264 if (uaccess_kernel())
5266 if (pagefault_disabled())
5268 __might_sleep(file, line);
5269 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5271 might_lock_read(¤t->mm->mmap_lock);
5274 EXPORT_SYMBOL(__might_fault);
5277 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5279 * Process all subpages of the specified huge page with the specified
5280 * operation. The target subpage will be processed last to keep its
5283 static inline void process_huge_page(
5284 unsigned long addr_hint, unsigned int pages_per_huge_page,
5285 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5289 unsigned long addr = addr_hint &
5290 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5292 /* Process target subpage last to keep its cache lines hot */
5294 n = (addr_hint - addr) / PAGE_SIZE;
5295 if (2 * n <= pages_per_huge_page) {
5296 /* If target subpage in first half of huge page */
5299 /* Process subpages at the end of huge page */
5300 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5302 process_subpage(addr + i * PAGE_SIZE, i, arg);
5305 /* If target subpage in second half of huge page */
5306 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5307 l = pages_per_huge_page - n;
5308 /* Process subpages at the begin of huge page */
5309 for (i = 0; i < base; i++) {
5311 process_subpage(addr + i * PAGE_SIZE, i, arg);
5315 * Process remaining subpages in left-right-left-right pattern
5316 * towards the target subpage
5318 for (i = 0; i < l; i++) {
5319 int left_idx = base + i;
5320 int right_idx = base + 2 * l - 1 - i;
5323 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5325 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5329 static void clear_gigantic_page(struct page *page,
5331 unsigned int pages_per_huge_page)
5334 struct page *p = page;
5337 for (i = 0; i < pages_per_huge_page;
5338 i++, p = mem_map_next(p, page, i)) {
5340 clear_user_highpage(p, addr + i * PAGE_SIZE);
5344 static void clear_subpage(unsigned long addr, int idx, void *arg)
5346 struct page *page = arg;
5348 clear_user_highpage(page + idx, addr);
5351 void clear_huge_page(struct page *page,
5352 unsigned long addr_hint, unsigned int pages_per_huge_page)
5354 unsigned long addr = addr_hint &
5355 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5357 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5358 clear_gigantic_page(page, addr, pages_per_huge_page);
5362 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5365 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5367 struct vm_area_struct *vma,
5368 unsigned int pages_per_huge_page)
5371 struct page *dst_base = dst;
5372 struct page *src_base = src;
5374 for (i = 0; i < pages_per_huge_page; ) {
5376 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5379 dst = mem_map_next(dst, dst_base, i);
5380 src = mem_map_next(src, src_base, i);
5384 struct copy_subpage_arg {
5387 struct vm_area_struct *vma;
5390 static void copy_subpage(unsigned long addr, int idx, void *arg)
5392 struct copy_subpage_arg *copy_arg = arg;
5394 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5395 addr, copy_arg->vma);
5398 void copy_user_huge_page(struct page *dst, struct page *src,
5399 unsigned long addr_hint, struct vm_area_struct *vma,
5400 unsigned int pages_per_huge_page)
5402 unsigned long addr = addr_hint &
5403 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5404 struct copy_subpage_arg arg = {
5410 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5411 copy_user_gigantic_page(dst, src, addr, vma,
5412 pages_per_huge_page);
5416 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5419 long copy_huge_page_from_user(struct page *dst_page,
5420 const void __user *usr_src,
5421 unsigned int pages_per_huge_page,
5422 bool allow_pagefault)
5425 unsigned long i, rc = 0;
5426 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5427 struct page *subpage = dst_page;
5429 for (i = 0; i < pages_per_huge_page;
5430 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5431 if (allow_pagefault)
5432 page_kaddr = kmap(subpage);
5434 page_kaddr = kmap_atomic(subpage);
5435 rc = copy_from_user(page_kaddr,
5436 usr_src + i * PAGE_SIZE, PAGE_SIZE);
5437 if (allow_pagefault)
5440 kunmap_atomic(page_kaddr);
5442 ret_val -= (PAGE_SIZE - rc);
5446 flush_dcache_page(subpage);
5452 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5454 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5456 static struct kmem_cache *page_ptl_cachep;
5458 void __init ptlock_cache_init(void)
5460 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5464 bool ptlock_alloc(struct page *page)
5468 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5475 void ptlock_free(struct page *page)
5477 kmem_cache_free(page_ptl_cachep, page->ptl);