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);
738 if (vma->vm_flags & VM_LOCKED)
739 mlock_vma_page(page);
742 * No need to invalidate - it was non-present before. However
743 * secondary CPUs may have mappings that need invalidating.
745 update_mmu_cache(vma, address, ptep);
749 * Tries to restore an exclusive pte if the page lock can be acquired without
753 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
756 swp_entry_t entry = pte_to_swp_entry(*src_pte);
757 struct page *page = pfn_swap_entry_to_page(entry);
759 if (trylock_page(page)) {
760 restore_exclusive_pte(vma, page, addr, src_pte);
769 * copy one vm_area from one task to the other. Assumes the page tables
770 * already present in the new task to be cleared in the whole range
771 * covered by this vma.
775 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
776 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
777 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
779 unsigned long vm_flags = dst_vma->vm_flags;
780 pte_t pte = *src_pte;
782 swp_entry_t entry = pte_to_swp_entry(pte);
784 if (likely(!non_swap_entry(entry))) {
785 if (swap_duplicate(entry) < 0)
788 /* make sure dst_mm is on swapoff's mmlist. */
789 if (unlikely(list_empty(&dst_mm->mmlist))) {
790 spin_lock(&mmlist_lock);
791 if (list_empty(&dst_mm->mmlist))
792 list_add(&dst_mm->mmlist,
794 spin_unlock(&mmlist_lock);
797 } else if (is_migration_entry(entry)) {
798 page = pfn_swap_entry_to_page(entry);
800 rss[mm_counter(page)]++;
802 if (is_writable_migration_entry(entry) &&
803 is_cow_mapping(vm_flags)) {
805 * COW mappings require pages in both
806 * parent and child to be set to read.
808 entry = make_readable_migration_entry(
810 pte = swp_entry_to_pte(entry);
811 if (pte_swp_soft_dirty(*src_pte))
812 pte = pte_swp_mksoft_dirty(pte);
813 if (pte_swp_uffd_wp(*src_pte))
814 pte = pte_swp_mkuffd_wp(pte);
815 set_pte_at(src_mm, addr, src_pte, pte);
817 } else if (is_device_private_entry(entry)) {
818 page = pfn_swap_entry_to_page(entry);
821 * Update rss count even for unaddressable pages, as
822 * they should treated just like normal pages in this
825 * We will likely want to have some new rss counters
826 * for unaddressable pages, at some point. But for now
827 * keep things as they are.
830 rss[mm_counter(page)]++;
831 page_dup_rmap(page, false);
834 * We do not preserve soft-dirty information, because so
835 * far, checkpoint/restore is the only feature that
836 * requires that. And checkpoint/restore does not work
837 * when a device driver is involved (you cannot easily
838 * save and restore device driver state).
840 if (is_writable_device_private_entry(entry) &&
841 is_cow_mapping(vm_flags)) {
842 entry = make_readable_device_private_entry(
844 pte = swp_entry_to_pte(entry);
845 if (pte_swp_uffd_wp(*src_pte))
846 pte = pte_swp_mkuffd_wp(pte);
847 set_pte_at(src_mm, addr, src_pte, pte);
849 } else if (is_device_exclusive_entry(entry)) {
851 * Make device exclusive entries present by restoring the
852 * original entry then copying as for a present pte. Device
853 * exclusive entries currently only support private writable
854 * (ie. COW) mappings.
856 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
857 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
861 if (!userfaultfd_wp(dst_vma))
862 pte = pte_swp_clear_uffd_wp(pte);
863 set_pte_at(dst_mm, addr, dst_pte, pte);
868 * Copy a present and normal page if necessary.
870 * NOTE! The usual case is that this doesn't need to do
871 * anything, and can just return a positive value. That
872 * will let the caller know that it can just increase
873 * the page refcount and re-use the pte the traditional
876 * But _if_ we need to copy it because it needs to be
877 * pinned in the parent (and the child should get its own
878 * copy rather than just a reference to the same page),
879 * we'll do that here and return zero to let the caller
882 * And if we need a pre-allocated page but don't yet have
883 * one, return a negative error to let the preallocation
884 * code know so that it can do so outside the page table
888 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
889 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
890 struct page **prealloc, pte_t pte, struct page *page)
892 struct page *new_page;
895 * What we want to do is to check whether this page may
896 * have been pinned by the parent process. If so,
897 * instead of wrprotect the pte on both sides, we copy
898 * the page immediately so that we'll always guarantee
899 * the pinned page won't be randomly replaced in the
902 * The page pinning checks are just "has this mm ever
903 * seen pinning", along with the (inexact) check of
904 * the page count. That might give false positives for
905 * for pinning, but it will work correctly.
907 if (likely(!page_needs_cow_for_dma(src_vma, page)))
910 new_page = *prealloc;
915 * We have a prealloc page, all good! Take it
916 * over and copy the page & arm it.
919 copy_user_highpage(new_page, page, addr, src_vma);
920 __SetPageUptodate(new_page);
921 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
922 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
923 rss[mm_counter(new_page)]++;
925 /* All done, just insert the new page copy in the child */
926 pte = mk_pte(new_page, dst_vma->vm_page_prot);
927 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
928 if (userfaultfd_pte_wp(dst_vma, *src_pte))
929 /* Uffd-wp needs to be delivered to dest pte as well */
930 pte = pte_wrprotect(pte_mkuffd_wp(pte));
931 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
936 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
937 * is required to copy this pte.
940 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
941 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
942 struct page **prealloc)
944 struct mm_struct *src_mm = src_vma->vm_mm;
945 unsigned long vm_flags = src_vma->vm_flags;
946 pte_t pte = *src_pte;
949 page = vm_normal_page(src_vma, addr, pte);
953 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
954 addr, rss, prealloc, pte, page);
959 page_dup_rmap(page, false);
960 rss[mm_counter(page)]++;
964 * If it's a COW mapping, write protect it both
965 * in the parent and the child
967 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
968 ptep_set_wrprotect(src_mm, addr, src_pte);
969 pte = pte_wrprotect(pte);
973 * If it's a shared mapping, mark it clean in
976 if (vm_flags & VM_SHARED)
977 pte = pte_mkclean(pte);
978 pte = pte_mkold(pte);
980 if (!userfaultfd_wp(dst_vma))
981 pte = pte_clear_uffd_wp(pte);
983 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
987 static inline struct page *
988 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
991 struct page *new_page;
993 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
997 if (mem_cgroup_charge(page_folio(new_page), src_mm, GFP_KERNEL)) {
1001 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
1007 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1008 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1011 struct mm_struct *dst_mm = dst_vma->vm_mm;
1012 struct mm_struct *src_mm = src_vma->vm_mm;
1013 pte_t *orig_src_pte, *orig_dst_pte;
1014 pte_t *src_pte, *dst_pte;
1015 spinlock_t *src_ptl, *dst_ptl;
1016 int progress, ret = 0;
1017 int rss[NR_MM_COUNTERS];
1018 swp_entry_t entry = (swp_entry_t){0};
1019 struct page *prealloc = NULL;
1025 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1030 src_pte = pte_offset_map(src_pmd, addr);
1031 src_ptl = pte_lockptr(src_mm, src_pmd);
1032 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1033 orig_src_pte = src_pte;
1034 orig_dst_pte = dst_pte;
1035 arch_enter_lazy_mmu_mode();
1039 * We are holding two locks at this point - either of them
1040 * could generate latencies in another task on another CPU.
1042 if (progress >= 32) {
1044 if (need_resched() ||
1045 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1048 if (pte_none(*src_pte)) {
1052 if (unlikely(!pte_present(*src_pte))) {
1053 ret = copy_nonpresent_pte(dst_mm, src_mm,
1058 entry = pte_to_swp_entry(*src_pte);
1060 } else if (ret == -EBUSY) {
1068 * Device exclusive entry restored, continue by copying
1069 * the now present pte.
1071 WARN_ON_ONCE(ret != -ENOENT);
1073 /* copy_present_pte() will clear `*prealloc' if consumed */
1074 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1075 addr, rss, &prealloc);
1077 * If we need a pre-allocated page for this pte, drop the
1078 * locks, allocate, and try again.
1080 if (unlikely(ret == -EAGAIN))
1082 if (unlikely(prealloc)) {
1084 * pre-alloc page cannot be reused by next time so as
1085 * to strictly follow mempolicy (e.g., alloc_page_vma()
1086 * will allocate page according to address). This
1087 * could only happen if one pinned pte changed.
1093 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1095 arch_leave_lazy_mmu_mode();
1096 spin_unlock(src_ptl);
1097 pte_unmap(orig_src_pte);
1098 add_mm_rss_vec(dst_mm, rss);
1099 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1103 VM_WARN_ON_ONCE(!entry.val);
1104 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1109 } else if (ret == -EBUSY) {
1111 } else if (ret == -EAGAIN) {
1112 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1119 /* We've captured and resolved the error. Reset, try again. */
1125 if (unlikely(prealloc))
1131 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1132 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1135 struct mm_struct *dst_mm = dst_vma->vm_mm;
1136 struct mm_struct *src_mm = src_vma->vm_mm;
1137 pmd_t *src_pmd, *dst_pmd;
1140 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1143 src_pmd = pmd_offset(src_pud, addr);
1145 next = pmd_addr_end(addr, end);
1146 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1147 || pmd_devmap(*src_pmd)) {
1149 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1150 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1151 addr, dst_vma, src_vma);
1158 if (pmd_none_or_clear_bad(src_pmd))
1160 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1163 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1168 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1169 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1172 struct mm_struct *dst_mm = dst_vma->vm_mm;
1173 struct mm_struct *src_mm = src_vma->vm_mm;
1174 pud_t *src_pud, *dst_pud;
1177 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1180 src_pud = pud_offset(src_p4d, addr);
1182 next = pud_addr_end(addr, end);
1183 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1186 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1187 err = copy_huge_pud(dst_mm, src_mm,
1188 dst_pud, src_pud, addr, src_vma);
1195 if (pud_none_or_clear_bad(src_pud))
1197 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1200 } while (dst_pud++, src_pud++, addr = next, addr != end);
1205 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1206 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1209 struct mm_struct *dst_mm = dst_vma->vm_mm;
1210 p4d_t *src_p4d, *dst_p4d;
1213 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1216 src_p4d = p4d_offset(src_pgd, addr);
1218 next = p4d_addr_end(addr, end);
1219 if (p4d_none_or_clear_bad(src_p4d))
1221 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1224 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1229 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1231 pgd_t *src_pgd, *dst_pgd;
1233 unsigned long addr = src_vma->vm_start;
1234 unsigned long end = src_vma->vm_end;
1235 struct mm_struct *dst_mm = dst_vma->vm_mm;
1236 struct mm_struct *src_mm = src_vma->vm_mm;
1237 struct mmu_notifier_range range;
1242 * Don't copy ptes where a page fault will fill them correctly.
1243 * Fork becomes much lighter when there are big shared or private
1244 * readonly mappings. The tradeoff is that copy_page_range is more
1245 * efficient than faulting.
1247 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1251 if (is_vm_hugetlb_page(src_vma))
1252 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1254 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1256 * We do not free on error cases below as remove_vma
1257 * gets called on error from higher level routine
1259 ret = track_pfn_copy(src_vma);
1265 * We need to invalidate the secondary MMU mappings only when
1266 * there could be a permission downgrade on the ptes of the
1267 * parent mm. And a permission downgrade will only happen if
1268 * is_cow_mapping() returns true.
1270 is_cow = is_cow_mapping(src_vma->vm_flags);
1273 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1274 0, src_vma, src_mm, addr, end);
1275 mmu_notifier_invalidate_range_start(&range);
1277 * Disabling preemption is not needed for the write side, as
1278 * the read side doesn't spin, but goes to the mmap_lock.
1280 * Use the raw variant of the seqcount_t write API to avoid
1281 * lockdep complaining about preemptibility.
1283 mmap_assert_write_locked(src_mm);
1284 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1288 dst_pgd = pgd_offset(dst_mm, addr);
1289 src_pgd = pgd_offset(src_mm, addr);
1291 next = pgd_addr_end(addr, end);
1292 if (pgd_none_or_clear_bad(src_pgd))
1294 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1299 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1302 raw_write_seqcount_end(&src_mm->write_protect_seq);
1303 mmu_notifier_invalidate_range_end(&range);
1309 * Parameter block passed down to zap_pte_range in exceptional cases.
1311 struct zap_details {
1312 struct folio *single_folio; /* Locked folio to be unmapped */
1313 bool even_cows; /* Zap COWed private pages too? */
1316 /* Whether we should zap all COWed (private) pages too */
1317 static inline bool should_zap_cows(struct zap_details *details)
1319 /* By default, zap all pages */
1323 /* Or, we zap COWed pages only if the caller wants to */
1324 return details->even_cows;
1327 /* Decides whether we should zap this page with the page pointer specified */
1328 static inline bool should_zap_page(struct zap_details *details, struct page *page)
1330 /* If we can make a decision without *page.. */
1331 if (should_zap_cows(details))
1334 /* E.g. the caller passes NULL for the case of a zero page */
1338 /* Otherwise we should only zap non-anon pages */
1339 return !PageAnon(page);
1342 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1343 struct vm_area_struct *vma, pmd_t *pmd,
1344 unsigned long addr, unsigned long end,
1345 struct zap_details *details)
1347 struct mm_struct *mm = tlb->mm;
1348 int force_flush = 0;
1349 int rss[NR_MM_COUNTERS];
1355 tlb_change_page_size(tlb, PAGE_SIZE);
1358 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1360 flush_tlb_batched_pending(mm);
1361 arch_enter_lazy_mmu_mode();
1366 if (pte_none(ptent))
1372 if (pte_present(ptent)) {
1373 page = vm_normal_page(vma, addr, ptent);
1374 if (unlikely(!should_zap_page(details, page)))
1376 ptent = ptep_get_and_clear_full(mm, addr, pte,
1378 tlb_remove_tlb_entry(tlb, pte, addr);
1379 if (unlikely(!page))
1382 if (!PageAnon(page)) {
1383 if (pte_dirty(ptent)) {
1385 set_page_dirty(page);
1387 if (pte_young(ptent) &&
1388 likely(!(vma->vm_flags & VM_SEQ_READ)))
1389 mark_page_accessed(page);
1391 rss[mm_counter(page)]--;
1392 page_remove_rmap(page, false);
1393 if (unlikely(page_mapcount(page) < 0))
1394 print_bad_pte(vma, addr, ptent, page);
1395 if (unlikely(__tlb_remove_page(tlb, page))) {
1403 entry = pte_to_swp_entry(ptent);
1404 if (is_device_private_entry(entry) ||
1405 is_device_exclusive_entry(entry)) {
1406 page = pfn_swap_entry_to_page(entry);
1407 if (unlikely(!should_zap_page(details, page)))
1409 rss[mm_counter(page)]--;
1410 if (is_device_private_entry(entry))
1411 page_remove_rmap(page, false);
1413 } else if (!non_swap_entry(entry)) {
1414 /* Genuine swap entry, hence a private anon page */
1415 if (!should_zap_cows(details))
1418 if (unlikely(!free_swap_and_cache(entry)))
1419 print_bad_pte(vma, addr, ptent, NULL);
1420 } else if (is_migration_entry(entry)) {
1421 page = pfn_swap_entry_to_page(entry);
1422 if (!should_zap_page(details, page))
1424 rss[mm_counter(page)]--;
1425 } else if (is_hwpoison_entry(entry)) {
1426 if (!should_zap_cows(details))
1429 /* We should have covered all the swap entry types */
1432 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1433 } while (pte++, addr += PAGE_SIZE, addr != end);
1435 add_mm_rss_vec(mm, rss);
1436 arch_leave_lazy_mmu_mode();
1438 /* Do the actual TLB flush before dropping ptl */
1440 tlb_flush_mmu_tlbonly(tlb);
1441 pte_unmap_unlock(start_pte, ptl);
1444 * If we forced a TLB flush (either due to running out of
1445 * batch buffers or because we needed to flush dirty TLB
1446 * entries before releasing the ptl), free the batched
1447 * memory too. Restart if we didn't do everything.
1462 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1463 struct vm_area_struct *vma, pud_t *pud,
1464 unsigned long addr, unsigned long end,
1465 struct zap_details *details)
1470 pmd = pmd_offset(pud, addr);
1472 next = pmd_addr_end(addr, end);
1473 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1474 if (next - addr != HPAGE_PMD_SIZE)
1475 __split_huge_pmd(vma, pmd, addr, false, NULL);
1476 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1479 } else if (details && details->single_folio &&
1480 folio_test_pmd_mappable(details->single_folio) &&
1481 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1482 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1484 * Take and drop THP pmd lock so that we cannot return
1485 * prematurely, while zap_huge_pmd() has cleared *pmd,
1486 * but not yet decremented compound_mapcount().
1492 * Here there can be other concurrent MADV_DONTNEED or
1493 * trans huge page faults running, and if the pmd is
1494 * none or trans huge it can change under us. This is
1495 * because MADV_DONTNEED holds the mmap_lock in read
1498 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1500 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1503 } while (pmd++, addr = next, addr != end);
1508 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1509 struct vm_area_struct *vma, p4d_t *p4d,
1510 unsigned long addr, unsigned long end,
1511 struct zap_details *details)
1516 pud = pud_offset(p4d, addr);
1518 next = pud_addr_end(addr, end);
1519 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1520 if (next - addr != HPAGE_PUD_SIZE) {
1521 mmap_assert_locked(tlb->mm);
1522 split_huge_pud(vma, pud, addr);
1523 } else if (zap_huge_pud(tlb, vma, pud, addr))
1527 if (pud_none_or_clear_bad(pud))
1529 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1532 } while (pud++, addr = next, addr != end);
1537 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1538 struct vm_area_struct *vma, pgd_t *pgd,
1539 unsigned long addr, unsigned long end,
1540 struct zap_details *details)
1545 p4d = p4d_offset(pgd, addr);
1547 next = p4d_addr_end(addr, end);
1548 if (p4d_none_or_clear_bad(p4d))
1550 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1551 } while (p4d++, addr = next, addr != end);
1556 void unmap_page_range(struct mmu_gather *tlb,
1557 struct vm_area_struct *vma,
1558 unsigned long addr, unsigned long end,
1559 struct zap_details *details)
1564 BUG_ON(addr >= end);
1565 tlb_start_vma(tlb, vma);
1566 pgd = pgd_offset(vma->vm_mm, addr);
1568 next = pgd_addr_end(addr, end);
1569 if (pgd_none_or_clear_bad(pgd))
1571 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1572 } while (pgd++, addr = next, addr != end);
1573 tlb_end_vma(tlb, vma);
1577 static void unmap_single_vma(struct mmu_gather *tlb,
1578 struct vm_area_struct *vma, unsigned long start_addr,
1579 unsigned long end_addr,
1580 struct zap_details *details)
1582 unsigned long start = max(vma->vm_start, start_addr);
1585 if (start >= vma->vm_end)
1587 end = min(vma->vm_end, end_addr);
1588 if (end <= vma->vm_start)
1592 uprobe_munmap(vma, start, end);
1594 if (unlikely(vma->vm_flags & VM_PFNMAP))
1595 untrack_pfn(vma, 0, 0);
1598 if (unlikely(is_vm_hugetlb_page(vma))) {
1600 * It is undesirable to test vma->vm_file as it
1601 * should be non-null for valid hugetlb area.
1602 * However, vm_file will be NULL in the error
1603 * cleanup path of mmap_region. When
1604 * hugetlbfs ->mmap method fails,
1605 * mmap_region() nullifies vma->vm_file
1606 * before calling this function to clean up.
1607 * Since no pte has actually been setup, it is
1608 * safe to do nothing in this case.
1611 i_mmap_lock_write(vma->vm_file->f_mapping);
1612 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1613 i_mmap_unlock_write(vma->vm_file->f_mapping);
1616 unmap_page_range(tlb, vma, start, end, details);
1621 * unmap_vmas - unmap a range of memory covered by a list of vma's
1622 * @tlb: address of the caller's struct mmu_gather
1623 * @vma: the starting vma
1624 * @start_addr: virtual address at which to start unmapping
1625 * @end_addr: virtual address at which to end unmapping
1627 * Unmap all pages in the vma list.
1629 * Only addresses between `start' and `end' will be unmapped.
1631 * The VMA list must be sorted in ascending virtual address order.
1633 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1634 * range after unmap_vmas() returns. So the only responsibility here is to
1635 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1636 * drops the lock and schedules.
1638 void unmap_vmas(struct mmu_gather *tlb,
1639 struct vm_area_struct *vma, unsigned long start_addr,
1640 unsigned long end_addr)
1642 struct mmu_notifier_range range;
1644 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1645 start_addr, end_addr);
1646 mmu_notifier_invalidate_range_start(&range);
1647 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1648 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1649 mmu_notifier_invalidate_range_end(&range);
1653 * zap_page_range - remove user pages in a given range
1654 * @vma: vm_area_struct holding the applicable pages
1655 * @start: starting address of pages to zap
1656 * @size: number of bytes to zap
1658 * Caller must protect the VMA list
1660 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1663 struct mmu_notifier_range range;
1664 struct mmu_gather tlb;
1667 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1668 start, start + size);
1669 tlb_gather_mmu(&tlb, vma->vm_mm);
1670 update_hiwater_rss(vma->vm_mm);
1671 mmu_notifier_invalidate_range_start(&range);
1672 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1673 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1674 mmu_notifier_invalidate_range_end(&range);
1675 tlb_finish_mmu(&tlb);
1679 * zap_page_range_single - remove user pages in a given range
1680 * @vma: vm_area_struct holding the applicable pages
1681 * @address: starting address of pages to zap
1682 * @size: number of bytes to zap
1683 * @details: details of shared cache invalidation
1685 * The range must fit into one VMA.
1687 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1688 unsigned long size, struct zap_details *details)
1690 struct mmu_notifier_range range;
1691 struct mmu_gather tlb;
1694 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1695 address, address + size);
1696 tlb_gather_mmu(&tlb, vma->vm_mm);
1697 update_hiwater_rss(vma->vm_mm);
1698 mmu_notifier_invalidate_range_start(&range);
1699 unmap_single_vma(&tlb, vma, address, range.end, details);
1700 mmu_notifier_invalidate_range_end(&range);
1701 tlb_finish_mmu(&tlb);
1705 * zap_vma_ptes - remove ptes mapping the vma
1706 * @vma: vm_area_struct holding ptes to be zapped
1707 * @address: starting address of pages to zap
1708 * @size: number of bytes to zap
1710 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1712 * The entire address range must be fully contained within the vma.
1715 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1718 if (!range_in_vma(vma, address, address + size) ||
1719 !(vma->vm_flags & VM_PFNMAP))
1722 zap_page_range_single(vma, address, size, NULL);
1724 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1726 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1733 pgd = pgd_offset(mm, addr);
1734 p4d = p4d_alloc(mm, pgd, addr);
1737 pud = pud_alloc(mm, p4d, addr);
1740 pmd = pmd_alloc(mm, pud, addr);
1744 VM_BUG_ON(pmd_trans_huge(*pmd));
1748 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1751 pmd_t *pmd = walk_to_pmd(mm, addr);
1755 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1758 static int validate_page_before_insert(struct page *page)
1760 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1762 flush_dcache_page(page);
1766 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1767 unsigned long addr, struct page *page, pgprot_t prot)
1769 if (!pte_none(*pte))
1771 /* Ok, finally just insert the thing.. */
1773 inc_mm_counter_fast(mm, mm_counter_file(page));
1774 page_add_file_rmap(page, false);
1775 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1780 * This is the old fallback for page remapping.
1782 * For historical reasons, it only allows reserved pages. Only
1783 * old drivers should use this, and they needed to mark their
1784 * pages reserved for the old functions anyway.
1786 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1787 struct page *page, pgprot_t prot)
1789 struct mm_struct *mm = vma->vm_mm;
1794 retval = validate_page_before_insert(page);
1798 pte = get_locked_pte(mm, addr, &ptl);
1801 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1802 pte_unmap_unlock(pte, ptl);
1808 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1809 unsigned long addr, struct page *page, pgprot_t prot)
1813 if (!page_count(page))
1815 err = validate_page_before_insert(page);
1818 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1821 /* insert_pages() amortizes the cost of spinlock operations
1822 * when inserting pages in a loop. Arch *must* define pte_index.
1824 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1825 struct page **pages, unsigned long *num, pgprot_t prot)
1828 pte_t *start_pte, *pte;
1829 spinlock_t *pte_lock;
1830 struct mm_struct *const mm = vma->vm_mm;
1831 unsigned long curr_page_idx = 0;
1832 unsigned long remaining_pages_total = *num;
1833 unsigned long pages_to_write_in_pmd;
1837 pmd = walk_to_pmd(mm, addr);
1841 pages_to_write_in_pmd = min_t(unsigned long,
1842 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1844 /* Allocate the PTE if necessary; takes PMD lock once only. */
1846 if (pte_alloc(mm, pmd))
1849 while (pages_to_write_in_pmd) {
1851 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1853 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1854 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1855 int err = insert_page_in_batch_locked(mm, pte,
1856 addr, pages[curr_page_idx], prot);
1857 if (unlikely(err)) {
1858 pte_unmap_unlock(start_pte, pte_lock);
1860 remaining_pages_total -= pte_idx;
1866 pte_unmap_unlock(start_pte, pte_lock);
1867 pages_to_write_in_pmd -= batch_size;
1868 remaining_pages_total -= batch_size;
1870 if (remaining_pages_total)
1874 *num = remaining_pages_total;
1877 #endif /* ifdef pte_index */
1880 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1881 * @vma: user vma to map to
1882 * @addr: target start user address of these pages
1883 * @pages: source kernel pages
1884 * @num: in: number of pages to map. out: number of pages that were *not*
1885 * mapped. (0 means all pages were successfully mapped).
1887 * Preferred over vm_insert_page() when inserting multiple pages.
1889 * In case of error, we may have mapped a subset of the provided
1890 * pages. It is the caller's responsibility to account for this case.
1892 * The same restrictions apply as in vm_insert_page().
1894 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1895 struct page **pages, unsigned long *num)
1898 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1900 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1902 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1903 BUG_ON(mmap_read_trylock(vma->vm_mm));
1904 BUG_ON(vma->vm_flags & VM_PFNMAP);
1905 vma->vm_flags |= VM_MIXEDMAP;
1907 /* Defer page refcount checking till we're about to map that page. */
1908 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1910 unsigned long idx = 0, pgcount = *num;
1913 for (; idx < pgcount; ++idx) {
1914 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1918 *num = pgcount - idx;
1920 #endif /* ifdef pte_index */
1922 EXPORT_SYMBOL(vm_insert_pages);
1925 * vm_insert_page - insert single page into user vma
1926 * @vma: user vma to map to
1927 * @addr: target user address of this page
1928 * @page: source kernel page
1930 * This allows drivers to insert individual pages they've allocated
1933 * The page has to be a nice clean _individual_ kernel allocation.
1934 * If you allocate a compound page, you need to have marked it as
1935 * such (__GFP_COMP), or manually just split the page up yourself
1936 * (see split_page()).
1938 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1939 * took an arbitrary page protection parameter. This doesn't allow
1940 * that. Your vma protection will have to be set up correctly, which
1941 * means that if you want a shared writable mapping, you'd better
1942 * ask for a shared writable mapping!
1944 * The page does not need to be reserved.
1946 * Usually this function is called from f_op->mmap() handler
1947 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1948 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1949 * function from other places, for example from page-fault handler.
1951 * Return: %0 on success, negative error code otherwise.
1953 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1956 if (addr < vma->vm_start || addr >= vma->vm_end)
1958 if (!page_count(page))
1960 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1961 BUG_ON(mmap_read_trylock(vma->vm_mm));
1962 BUG_ON(vma->vm_flags & VM_PFNMAP);
1963 vma->vm_flags |= VM_MIXEDMAP;
1965 return insert_page(vma, addr, page, vma->vm_page_prot);
1967 EXPORT_SYMBOL(vm_insert_page);
1970 * __vm_map_pages - maps range of kernel pages into user vma
1971 * @vma: user vma to map to
1972 * @pages: pointer to array of source kernel pages
1973 * @num: number of pages in page array
1974 * @offset: user's requested vm_pgoff
1976 * This allows drivers to map range of kernel pages into a user vma.
1978 * Return: 0 on success and error code otherwise.
1980 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1981 unsigned long num, unsigned long offset)
1983 unsigned long count = vma_pages(vma);
1984 unsigned long uaddr = vma->vm_start;
1987 /* Fail if the user requested offset is beyond the end of the object */
1991 /* Fail if the user requested size exceeds available object size */
1992 if (count > num - offset)
1995 for (i = 0; i < count; i++) {
1996 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2006 * vm_map_pages - maps range of kernel pages starts with non zero offset
2007 * @vma: user vma to map to
2008 * @pages: pointer to array of source kernel pages
2009 * @num: number of pages in page array
2011 * Maps an object consisting of @num pages, catering for the user's
2012 * requested vm_pgoff
2014 * If we fail to insert any page into the vma, the function will return
2015 * immediately leaving any previously inserted pages present. Callers
2016 * from the mmap handler may immediately return the error as their caller
2017 * will destroy the vma, removing any successfully inserted pages. Other
2018 * callers should make their own arrangements for calling unmap_region().
2020 * Context: Process context. Called by mmap handlers.
2021 * Return: 0 on success and error code otherwise.
2023 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2026 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2028 EXPORT_SYMBOL(vm_map_pages);
2031 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2032 * @vma: user vma to map to
2033 * @pages: pointer to array of source kernel pages
2034 * @num: number of pages in page array
2036 * Similar to vm_map_pages(), except that it explicitly sets the offset
2037 * to 0. This function is intended for the drivers that did not consider
2040 * Context: Process context. Called by mmap handlers.
2041 * Return: 0 on success and error code otherwise.
2043 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2046 return __vm_map_pages(vma, pages, num, 0);
2048 EXPORT_SYMBOL(vm_map_pages_zero);
2050 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2051 pfn_t pfn, pgprot_t prot, bool mkwrite)
2053 struct mm_struct *mm = vma->vm_mm;
2057 pte = get_locked_pte(mm, addr, &ptl);
2059 return VM_FAULT_OOM;
2060 if (!pte_none(*pte)) {
2063 * For read faults on private mappings the PFN passed
2064 * in may not match the PFN we have mapped if the
2065 * mapped PFN is a writeable COW page. In the mkwrite
2066 * case we are creating a writable PTE for a shared
2067 * mapping and we expect the PFNs to match. If they
2068 * don't match, we are likely racing with block
2069 * allocation and mapping invalidation so just skip the
2072 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
2073 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
2076 entry = pte_mkyoung(*pte);
2077 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2078 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2079 update_mmu_cache(vma, addr, pte);
2084 /* Ok, finally just insert the thing.. */
2085 if (pfn_t_devmap(pfn))
2086 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2088 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2091 entry = pte_mkyoung(entry);
2092 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2095 set_pte_at(mm, addr, pte, entry);
2096 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2099 pte_unmap_unlock(pte, ptl);
2100 return VM_FAULT_NOPAGE;
2104 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2105 * @vma: user vma to map to
2106 * @addr: target user address of this page
2107 * @pfn: source kernel pfn
2108 * @pgprot: pgprot flags for the inserted page
2110 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2111 * to override pgprot on a per-page basis.
2113 * This only makes sense for IO mappings, and it makes no sense for
2114 * COW mappings. In general, using multiple vmas is preferable;
2115 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2118 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2119 * a value of @pgprot different from that of @vma->vm_page_prot.
2121 * Context: Process context. May allocate using %GFP_KERNEL.
2122 * Return: vm_fault_t value.
2124 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2125 unsigned long pfn, pgprot_t pgprot)
2128 * Technically, architectures with pte_special can avoid all these
2129 * restrictions (same for remap_pfn_range). However we would like
2130 * consistency in testing and feature parity among all, so we should
2131 * try to keep these invariants in place for everybody.
2133 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2134 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2135 (VM_PFNMAP|VM_MIXEDMAP));
2136 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2137 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2139 if (addr < vma->vm_start || addr >= vma->vm_end)
2140 return VM_FAULT_SIGBUS;
2142 if (!pfn_modify_allowed(pfn, pgprot))
2143 return VM_FAULT_SIGBUS;
2145 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2147 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2150 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2153 * vmf_insert_pfn - insert single pfn into user vma
2154 * @vma: user vma to map to
2155 * @addr: target user address of this page
2156 * @pfn: source kernel pfn
2158 * Similar to vm_insert_page, this allows drivers to insert individual pages
2159 * they've allocated into a user vma. Same comments apply.
2161 * This function should only be called from a vm_ops->fault handler, and
2162 * in that case the handler should return the result of this function.
2164 * vma cannot be a COW mapping.
2166 * As this is called only for pages that do not currently exist, we
2167 * do not need to flush old virtual caches or the TLB.
2169 * Context: Process context. May allocate using %GFP_KERNEL.
2170 * Return: vm_fault_t value.
2172 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2175 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2177 EXPORT_SYMBOL(vmf_insert_pfn);
2179 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2181 /* these checks mirror the abort conditions in vm_normal_page */
2182 if (vma->vm_flags & VM_MIXEDMAP)
2184 if (pfn_t_devmap(pfn))
2186 if (pfn_t_special(pfn))
2188 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2193 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2194 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2199 BUG_ON(!vm_mixed_ok(vma, pfn));
2201 if (addr < vma->vm_start || addr >= vma->vm_end)
2202 return VM_FAULT_SIGBUS;
2204 track_pfn_insert(vma, &pgprot, pfn);
2206 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2207 return VM_FAULT_SIGBUS;
2210 * If we don't have pte special, then we have to use the pfn_valid()
2211 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2212 * refcount the page if pfn_valid is true (hence insert_page rather
2213 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2214 * without pte special, it would there be refcounted as a normal page.
2216 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2217 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2221 * At this point we are committed to insert_page()
2222 * regardless of whether the caller specified flags that
2223 * result in pfn_t_has_page() == false.
2225 page = pfn_to_page(pfn_t_to_pfn(pfn));
2226 err = insert_page(vma, addr, page, pgprot);
2228 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2232 return VM_FAULT_OOM;
2233 if (err < 0 && err != -EBUSY)
2234 return VM_FAULT_SIGBUS;
2236 return VM_FAULT_NOPAGE;
2240 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2241 * @vma: user vma to map to
2242 * @addr: target user address of this page
2243 * @pfn: source kernel pfn
2244 * @pgprot: pgprot flags for the inserted page
2246 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2247 * to override pgprot on a per-page basis.
2249 * Typically this function should be used by drivers to set caching- and
2250 * encryption bits different than those of @vma->vm_page_prot, because
2251 * the caching- or encryption mode may not be known at mmap() time.
2252 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2253 * to set caching and encryption bits for those vmas (except for COW pages).
2254 * This is ensured by core vm only modifying these page table entries using
2255 * functions that don't touch caching- or encryption bits, using pte_modify()
2256 * if needed. (See for example mprotect()).
2257 * Also when new page-table entries are created, this is only done using the
2258 * fault() callback, and never using the value of vma->vm_page_prot,
2259 * except for page-table entries that point to anonymous pages as the result
2262 * Context: Process context. May allocate using %GFP_KERNEL.
2263 * Return: vm_fault_t value.
2265 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2266 pfn_t pfn, pgprot_t pgprot)
2268 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2270 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2272 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2275 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2277 EXPORT_SYMBOL(vmf_insert_mixed);
2280 * If the insertion of PTE failed because someone else already added a
2281 * different entry in the mean time, we treat that as success as we assume
2282 * the same entry was actually inserted.
2284 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2285 unsigned long addr, pfn_t pfn)
2287 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2289 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2292 * maps a range of physical memory into the requested pages. the old
2293 * mappings are removed. any references to nonexistent pages results
2294 * in null mappings (currently treated as "copy-on-access")
2296 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2297 unsigned long addr, unsigned long end,
2298 unsigned long pfn, pgprot_t prot)
2300 pte_t *pte, *mapped_pte;
2304 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2307 arch_enter_lazy_mmu_mode();
2309 BUG_ON(!pte_none(*pte));
2310 if (!pfn_modify_allowed(pfn, prot)) {
2314 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2316 } while (pte++, addr += PAGE_SIZE, addr != end);
2317 arch_leave_lazy_mmu_mode();
2318 pte_unmap_unlock(mapped_pte, ptl);
2322 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2323 unsigned long addr, unsigned long end,
2324 unsigned long pfn, pgprot_t prot)
2330 pfn -= addr >> PAGE_SHIFT;
2331 pmd = pmd_alloc(mm, pud, addr);
2334 VM_BUG_ON(pmd_trans_huge(*pmd));
2336 next = pmd_addr_end(addr, end);
2337 err = remap_pte_range(mm, pmd, addr, next,
2338 pfn + (addr >> PAGE_SHIFT), prot);
2341 } while (pmd++, addr = next, addr != end);
2345 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2346 unsigned long addr, unsigned long end,
2347 unsigned long pfn, pgprot_t prot)
2353 pfn -= addr >> PAGE_SHIFT;
2354 pud = pud_alloc(mm, p4d, addr);
2358 next = pud_addr_end(addr, end);
2359 err = remap_pmd_range(mm, pud, addr, next,
2360 pfn + (addr >> PAGE_SHIFT), prot);
2363 } while (pud++, addr = next, addr != end);
2367 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2368 unsigned long addr, unsigned long end,
2369 unsigned long pfn, pgprot_t prot)
2375 pfn -= addr >> PAGE_SHIFT;
2376 p4d = p4d_alloc(mm, pgd, addr);
2380 next = p4d_addr_end(addr, end);
2381 err = remap_pud_range(mm, p4d, addr, next,
2382 pfn + (addr >> PAGE_SHIFT), prot);
2385 } while (p4d++, addr = next, addr != end);
2390 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2391 * must have pre-validated the caching bits of the pgprot_t.
2393 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2394 unsigned long pfn, unsigned long size, pgprot_t prot)
2398 unsigned long end = addr + PAGE_ALIGN(size);
2399 struct mm_struct *mm = vma->vm_mm;
2402 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2406 * Physically remapped pages are special. Tell the
2407 * rest of the world about it:
2408 * VM_IO tells people not to look at these pages
2409 * (accesses can have side effects).
2410 * VM_PFNMAP tells the core MM that the base pages are just
2411 * raw PFN mappings, and do not have a "struct page" associated
2414 * Disable vma merging and expanding with mremap().
2416 * Omit vma from core dump, even when VM_IO turned off.
2418 * There's a horrible special case to handle copy-on-write
2419 * behaviour that some programs depend on. We mark the "original"
2420 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2421 * See vm_normal_page() for details.
2423 if (is_cow_mapping(vma->vm_flags)) {
2424 if (addr != vma->vm_start || end != vma->vm_end)
2426 vma->vm_pgoff = pfn;
2429 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2431 BUG_ON(addr >= end);
2432 pfn -= addr >> PAGE_SHIFT;
2433 pgd = pgd_offset(mm, addr);
2434 flush_cache_range(vma, addr, end);
2436 next = pgd_addr_end(addr, end);
2437 err = remap_p4d_range(mm, pgd, addr, next,
2438 pfn + (addr >> PAGE_SHIFT), prot);
2441 } while (pgd++, addr = next, addr != end);
2447 * remap_pfn_range - remap kernel memory to userspace
2448 * @vma: user vma to map to
2449 * @addr: target page aligned user address to start at
2450 * @pfn: page frame number of kernel physical memory address
2451 * @size: size of mapping area
2452 * @prot: page protection flags for this mapping
2454 * Note: this is only safe if the mm semaphore is held when called.
2456 * Return: %0 on success, negative error code otherwise.
2458 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2459 unsigned long pfn, unsigned long size, pgprot_t prot)
2463 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2467 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2469 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2472 EXPORT_SYMBOL(remap_pfn_range);
2475 * vm_iomap_memory - remap memory to userspace
2476 * @vma: user vma to map to
2477 * @start: start of the physical memory to be mapped
2478 * @len: size of area
2480 * This is a simplified io_remap_pfn_range() for common driver use. The
2481 * driver just needs to give us the physical memory range to be mapped,
2482 * we'll figure out the rest from the vma information.
2484 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2485 * whatever write-combining details or similar.
2487 * Return: %0 on success, negative error code otherwise.
2489 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2491 unsigned long vm_len, pfn, pages;
2493 /* Check that the physical memory area passed in looks valid */
2494 if (start + len < start)
2497 * You *really* shouldn't map things that aren't page-aligned,
2498 * but we've historically allowed it because IO memory might
2499 * just have smaller alignment.
2501 len += start & ~PAGE_MASK;
2502 pfn = start >> PAGE_SHIFT;
2503 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2504 if (pfn + pages < pfn)
2507 /* We start the mapping 'vm_pgoff' pages into the area */
2508 if (vma->vm_pgoff > pages)
2510 pfn += vma->vm_pgoff;
2511 pages -= vma->vm_pgoff;
2513 /* Can we fit all of the mapping? */
2514 vm_len = vma->vm_end - vma->vm_start;
2515 if (vm_len >> PAGE_SHIFT > pages)
2518 /* Ok, let it rip */
2519 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2521 EXPORT_SYMBOL(vm_iomap_memory);
2523 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2524 unsigned long addr, unsigned long end,
2525 pte_fn_t fn, void *data, bool create,
2526 pgtbl_mod_mask *mask)
2528 pte_t *pte, *mapped_pte;
2533 mapped_pte = pte = (mm == &init_mm) ?
2534 pte_alloc_kernel_track(pmd, addr, mask) :
2535 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2539 mapped_pte = pte = (mm == &init_mm) ?
2540 pte_offset_kernel(pmd, addr) :
2541 pte_offset_map_lock(mm, pmd, addr, &ptl);
2544 BUG_ON(pmd_huge(*pmd));
2546 arch_enter_lazy_mmu_mode();
2550 if (create || !pte_none(*pte)) {
2551 err = fn(pte++, addr, data);
2555 } while (addr += PAGE_SIZE, addr != end);
2557 *mask |= PGTBL_PTE_MODIFIED;
2559 arch_leave_lazy_mmu_mode();
2562 pte_unmap_unlock(mapped_pte, ptl);
2566 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2567 unsigned long addr, unsigned long end,
2568 pte_fn_t fn, void *data, bool create,
2569 pgtbl_mod_mask *mask)
2575 BUG_ON(pud_huge(*pud));
2578 pmd = pmd_alloc_track(mm, pud, addr, mask);
2582 pmd = pmd_offset(pud, addr);
2585 next = pmd_addr_end(addr, end);
2586 if (pmd_none(*pmd) && !create)
2588 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2590 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2595 err = apply_to_pte_range(mm, pmd, addr, next,
2596 fn, data, create, mask);
2599 } while (pmd++, addr = next, addr != end);
2604 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2605 unsigned long addr, unsigned long end,
2606 pte_fn_t fn, void *data, bool create,
2607 pgtbl_mod_mask *mask)
2614 pud = pud_alloc_track(mm, p4d, addr, mask);
2618 pud = pud_offset(p4d, addr);
2621 next = pud_addr_end(addr, end);
2622 if (pud_none(*pud) && !create)
2624 if (WARN_ON_ONCE(pud_leaf(*pud)))
2626 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2631 err = apply_to_pmd_range(mm, pud, addr, next,
2632 fn, data, create, mask);
2635 } while (pud++, addr = next, addr != end);
2640 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2641 unsigned long addr, unsigned long end,
2642 pte_fn_t fn, void *data, bool create,
2643 pgtbl_mod_mask *mask)
2650 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2654 p4d = p4d_offset(pgd, addr);
2657 next = p4d_addr_end(addr, end);
2658 if (p4d_none(*p4d) && !create)
2660 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2662 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2667 err = apply_to_pud_range(mm, p4d, addr, next,
2668 fn, data, create, mask);
2671 } while (p4d++, addr = next, addr != end);
2676 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2677 unsigned long size, pte_fn_t fn,
2678 void *data, bool create)
2681 unsigned long start = addr, next;
2682 unsigned long end = addr + size;
2683 pgtbl_mod_mask mask = 0;
2686 if (WARN_ON(addr >= end))
2689 pgd = pgd_offset(mm, addr);
2691 next = pgd_addr_end(addr, end);
2692 if (pgd_none(*pgd) && !create)
2694 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2696 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2701 err = apply_to_p4d_range(mm, pgd, addr, next,
2702 fn, data, create, &mask);
2705 } while (pgd++, addr = next, addr != end);
2707 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2708 arch_sync_kernel_mappings(start, start + size);
2714 * Scan a region of virtual memory, filling in page tables as necessary
2715 * and calling a provided function on each leaf page table.
2717 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2718 unsigned long size, pte_fn_t fn, void *data)
2720 return __apply_to_page_range(mm, addr, size, fn, data, true);
2722 EXPORT_SYMBOL_GPL(apply_to_page_range);
2725 * Scan a region of virtual memory, calling a provided function on
2726 * each leaf page table where it exists.
2728 * Unlike apply_to_page_range, this does _not_ fill in page tables
2729 * where they are absent.
2731 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2732 unsigned long size, pte_fn_t fn, void *data)
2734 return __apply_to_page_range(mm, addr, size, fn, data, false);
2736 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2739 * handle_pte_fault chooses page fault handler according to an entry which was
2740 * read non-atomically. Before making any commitment, on those architectures
2741 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2742 * parts, do_swap_page must check under lock before unmapping the pte and
2743 * proceeding (but do_wp_page is only called after already making such a check;
2744 * and do_anonymous_page can safely check later on).
2746 static inline int pte_unmap_same(struct vm_fault *vmf)
2749 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2750 if (sizeof(pte_t) > sizeof(unsigned long)) {
2751 spinlock_t *ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
2753 same = pte_same(*vmf->pte, vmf->orig_pte);
2757 pte_unmap(vmf->pte);
2762 static inline bool cow_user_page(struct page *dst, struct page *src,
2763 struct vm_fault *vmf)
2768 bool locked = false;
2769 struct vm_area_struct *vma = vmf->vma;
2770 struct mm_struct *mm = vma->vm_mm;
2771 unsigned long addr = vmf->address;
2774 copy_user_highpage(dst, src, addr, vma);
2779 * If the source page was a PFN mapping, we don't have
2780 * a "struct page" for it. We do a best-effort copy by
2781 * just copying from the original user address. If that
2782 * fails, we just zero-fill it. Live with it.
2784 kaddr = kmap_atomic(dst);
2785 uaddr = (void __user *)(addr & PAGE_MASK);
2788 * On architectures with software "accessed" bits, we would
2789 * take a double page fault, so mark it accessed here.
2791 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2794 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2796 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2798 * Other thread has already handled the fault
2799 * and update local tlb only
2801 update_mmu_tlb(vma, addr, vmf->pte);
2806 entry = pte_mkyoung(vmf->orig_pte);
2807 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2808 update_mmu_cache(vma, addr, vmf->pte);
2812 * This really shouldn't fail, because the page is there
2813 * in the page tables. But it might just be unreadable,
2814 * in which case we just give up and fill the result with
2817 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2821 /* Re-validate under PTL if the page is still mapped */
2822 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2824 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2825 /* The PTE changed under us, update local tlb */
2826 update_mmu_tlb(vma, addr, vmf->pte);
2832 * The same page can be mapped back since last copy attempt.
2833 * Try to copy again under PTL.
2835 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2837 * Give a warn in case there can be some obscure
2850 pte_unmap_unlock(vmf->pte, vmf->ptl);
2851 kunmap_atomic(kaddr);
2852 flush_dcache_page(dst);
2857 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2859 struct file *vm_file = vma->vm_file;
2862 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2865 * Special mappings (e.g. VDSO) do not have any file so fake
2866 * a default GFP_KERNEL for them.
2872 * Notify the address space that the page is about to become writable so that
2873 * it can prohibit this or wait for the page to get into an appropriate state.
2875 * We do this without the lock held, so that it can sleep if it needs to.
2877 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2880 struct page *page = vmf->page;
2881 unsigned int old_flags = vmf->flags;
2883 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2885 if (vmf->vma->vm_file &&
2886 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2887 return VM_FAULT_SIGBUS;
2889 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2890 /* Restore original flags so that caller is not surprised */
2891 vmf->flags = old_flags;
2892 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2894 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2896 if (!page->mapping) {
2898 return 0; /* retry */
2900 ret |= VM_FAULT_LOCKED;
2902 VM_BUG_ON_PAGE(!PageLocked(page), page);
2907 * Handle dirtying of a page in shared file mapping on a write fault.
2909 * The function expects the page to be locked and unlocks it.
2911 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2913 struct vm_area_struct *vma = vmf->vma;
2914 struct address_space *mapping;
2915 struct page *page = vmf->page;
2917 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2919 dirtied = set_page_dirty(page);
2920 VM_BUG_ON_PAGE(PageAnon(page), page);
2922 * Take a local copy of the address_space - page.mapping may be zeroed
2923 * by truncate after unlock_page(). The address_space itself remains
2924 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2925 * release semantics to prevent the compiler from undoing this copying.
2927 mapping = page_rmapping(page);
2931 file_update_time(vma->vm_file);
2934 * Throttle page dirtying rate down to writeback speed.
2936 * mapping may be NULL here because some device drivers do not
2937 * set page.mapping but still dirty their pages
2939 * Drop the mmap_lock before waiting on IO, if we can. The file
2940 * is pinning the mapping, as per above.
2942 if ((dirtied || page_mkwrite) && mapping) {
2945 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2946 balance_dirty_pages_ratelimited(mapping);
2949 return VM_FAULT_RETRY;
2957 * Handle write page faults for pages that can be reused in the current vma
2959 * This can happen either due to the mapping being with the VM_SHARED flag,
2960 * or due to us being the last reference standing to the page. In either
2961 * case, all we need to do here is to mark the page as writable and update
2962 * any related book-keeping.
2964 static inline void wp_page_reuse(struct vm_fault *vmf)
2965 __releases(vmf->ptl)
2967 struct vm_area_struct *vma = vmf->vma;
2968 struct page *page = vmf->page;
2971 * Clear the pages cpupid information as the existing
2972 * information potentially belongs to a now completely
2973 * unrelated process.
2976 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2978 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2979 entry = pte_mkyoung(vmf->orig_pte);
2980 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2981 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2982 update_mmu_cache(vma, vmf->address, vmf->pte);
2983 pte_unmap_unlock(vmf->pte, vmf->ptl);
2984 count_vm_event(PGREUSE);
2988 * Handle the case of a page which we actually need to copy to a new page.
2990 * Called with mmap_lock locked and the old page referenced, but
2991 * without the ptl held.
2993 * High level logic flow:
2995 * - Allocate a page, copy the content of the old page to the new one.
2996 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2997 * - Take the PTL. If the pte changed, bail out and release the allocated page
2998 * - If the pte is still the way we remember it, update the page table and all
2999 * relevant references. This includes dropping the reference the page-table
3000 * held to the old page, as well as updating the rmap.
3001 * - In any case, unlock the PTL and drop the reference we took to the old page.
3003 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
3005 struct vm_area_struct *vma = vmf->vma;
3006 struct mm_struct *mm = vma->vm_mm;
3007 struct page *old_page = vmf->page;
3008 struct page *new_page = NULL;
3010 int page_copied = 0;
3011 struct mmu_notifier_range range;
3013 if (unlikely(anon_vma_prepare(vma)))
3016 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
3017 new_page = alloc_zeroed_user_highpage_movable(vma,
3022 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3027 if (!cow_user_page(new_page, old_page, vmf)) {
3029 * COW failed, if the fault was solved by other,
3030 * it's fine. If not, userspace would re-fault on
3031 * the same address and we will handle the fault
3032 * from the second attempt.
3041 if (mem_cgroup_charge(page_folio(new_page), mm, GFP_KERNEL))
3043 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
3045 __SetPageUptodate(new_page);
3047 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
3048 vmf->address & PAGE_MASK,
3049 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3050 mmu_notifier_invalidate_range_start(&range);
3053 * Re-check the pte - we dropped the lock
3055 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3056 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
3058 if (!PageAnon(old_page)) {
3059 dec_mm_counter_fast(mm,
3060 mm_counter_file(old_page));
3061 inc_mm_counter_fast(mm, MM_ANONPAGES);
3064 inc_mm_counter_fast(mm, MM_ANONPAGES);
3066 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3067 entry = mk_pte(new_page, vma->vm_page_prot);
3068 entry = pte_sw_mkyoung(entry);
3069 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3072 * Clear the pte entry and flush it first, before updating the
3073 * pte with the new entry, to keep TLBs on different CPUs in
3074 * sync. This code used to set the new PTE then flush TLBs, but
3075 * that left a window where the new PTE could be loaded into
3076 * some TLBs while the old PTE remains in others.
3078 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3079 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
3080 lru_cache_add_inactive_or_unevictable(new_page, vma);
3082 * We call the notify macro here because, when using secondary
3083 * mmu page tables (such as kvm shadow page tables), we want the
3084 * new page to be mapped directly into the secondary page table.
3086 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3087 update_mmu_cache(vma, vmf->address, vmf->pte);
3090 * Only after switching the pte to the new page may
3091 * we remove the mapcount here. Otherwise another
3092 * process may come and find the rmap count decremented
3093 * before the pte is switched to the new page, and
3094 * "reuse" the old page writing into it while our pte
3095 * here still points into it and can be read by other
3098 * The critical issue is to order this
3099 * page_remove_rmap with the ptp_clear_flush above.
3100 * Those stores are ordered by (if nothing else,)
3101 * the barrier present in the atomic_add_negative
3102 * in page_remove_rmap.
3104 * Then the TLB flush in ptep_clear_flush ensures that
3105 * no process can access the old page before the
3106 * decremented mapcount is visible. And the old page
3107 * cannot be reused until after the decremented
3108 * mapcount is visible. So transitively, TLBs to
3109 * old page will be flushed before it can be reused.
3111 page_remove_rmap(old_page, false);
3114 /* Free the old page.. */
3115 new_page = old_page;
3118 update_mmu_tlb(vma, vmf->address, vmf->pte);
3124 pte_unmap_unlock(vmf->pte, vmf->ptl);
3126 * No need to double call mmu_notifier->invalidate_range() callback as
3127 * the above ptep_clear_flush_notify() did already call it.
3129 mmu_notifier_invalidate_range_only_end(&range);
3132 * Don't let another task, with possibly unlocked vma,
3133 * keep the mlocked page.
3135 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
3136 lock_page(old_page); /* LRU manipulation */
3137 if (PageMlocked(old_page))
3138 munlock_vma_page(old_page);
3139 unlock_page(old_page);
3142 free_swap_cache(old_page);
3145 return page_copied ? VM_FAULT_WRITE : 0;
3151 return VM_FAULT_OOM;
3155 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3156 * writeable once the page is prepared
3158 * @vmf: structure describing the fault
3160 * This function handles all that is needed to finish a write page fault in a
3161 * shared mapping due to PTE being read-only once the mapped page is prepared.
3162 * It handles locking of PTE and modifying it.
3164 * The function expects the page to be locked or other protection against
3165 * concurrent faults / writeback (such as DAX radix tree locks).
3167 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3168 * we acquired PTE lock.
3170 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3172 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3173 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3176 * We might have raced with another page fault while we released the
3177 * pte_offset_map_lock.
3179 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3180 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3181 pte_unmap_unlock(vmf->pte, vmf->ptl);
3182 return VM_FAULT_NOPAGE;
3189 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3192 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3194 struct vm_area_struct *vma = vmf->vma;
3196 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3199 pte_unmap_unlock(vmf->pte, vmf->ptl);
3200 vmf->flags |= FAULT_FLAG_MKWRITE;
3201 ret = vma->vm_ops->pfn_mkwrite(vmf);
3202 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3204 return finish_mkwrite_fault(vmf);
3207 return VM_FAULT_WRITE;
3210 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3211 __releases(vmf->ptl)
3213 struct vm_area_struct *vma = vmf->vma;
3214 vm_fault_t ret = VM_FAULT_WRITE;
3216 get_page(vmf->page);
3218 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3221 pte_unmap_unlock(vmf->pte, vmf->ptl);
3222 tmp = do_page_mkwrite(vmf);
3223 if (unlikely(!tmp || (tmp &
3224 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3225 put_page(vmf->page);
3228 tmp = finish_mkwrite_fault(vmf);
3229 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3230 unlock_page(vmf->page);
3231 put_page(vmf->page);
3236 lock_page(vmf->page);
3238 ret |= fault_dirty_shared_page(vmf);
3239 put_page(vmf->page);
3245 * This routine handles present pages, when users try to write
3246 * to a shared page. It is done by copying the page to a new address
3247 * and decrementing the shared-page counter for the old page.
3249 * Note that this routine assumes that the protection checks have been
3250 * done by the caller (the low-level page fault routine in most cases).
3251 * Thus we can safely just mark it writable once we've done any necessary
3254 * We also mark the page dirty at this point even though the page will
3255 * change only once the write actually happens. This avoids a few races,
3256 * and potentially makes it more efficient.
3258 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3259 * but allow concurrent faults), with pte both mapped and locked.
3260 * We return with mmap_lock still held, but pte unmapped and unlocked.
3262 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3263 __releases(vmf->ptl)
3265 struct vm_area_struct *vma = vmf->vma;
3267 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3268 pte_unmap_unlock(vmf->pte, vmf->ptl);
3269 return handle_userfault(vmf, VM_UFFD_WP);
3273 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3274 * is flushed in this case before copying.
3276 if (unlikely(userfaultfd_wp(vmf->vma) &&
3277 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3278 flush_tlb_page(vmf->vma, vmf->address);
3280 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3283 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3286 * We should not cow pages in a shared writeable mapping.
3287 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3289 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3290 (VM_WRITE|VM_SHARED))
3291 return wp_pfn_shared(vmf);
3293 pte_unmap_unlock(vmf->pte, vmf->ptl);
3294 return wp_page_copy(vmf);
3298 * Take out anonymous pages first, anonymous shared vmas are
3299 * not dirty accountable.
3301 if (PageAnon(vmf->page)) {
3302 struct page *page = vmf->page;
3304 /* PageKsm() doesn't necessarily raise the page refcount */
3305 if (PageKsm(page) || page_count(page) != 1)
3307 if (!trylock_page(page))
3309 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3314 * Ok, we've got the only map reference, and the only
3315 * page count reference, and the page is locked,
3316 * it's dark out, and we're wearing sunglasses. Hit it.
3320 return VM_FAULT_WRITE;
3321 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3322 (VM_WRITE|VM_SHARED))) {
3323 return wp_page_shared(vmf);
3327 * Ok, we need to copy. Oh, well..
3329 get_page(vmf->page);
3331 pte_unmap_unlock(vmf->pte, vmf->ptl);
3332 return wp_page_copy(vmf);
3335 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3336 unsigned long start_addr, unsigned long end_addr,
3337 struct zap_details *details)
3339 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3342 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3343 pgoff_t first_index,
3345 struct zap_details *details)
3347 struct vm_area_struct *vma;
3348 pgoff_t vba, vea, zba, zea;
3350 vma_interval_tree_foreach(vma, root, first_index, last_index) {
3351 vba = vma->vm_pgoff;
3352 vea = vba + vma_pages(vma) - 1;
3353 zba = max(first_index, vba);
3354 zea = min(last_index, vea);
3356 unmap_mapping_range_vma(vma,
3357 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3358 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3364 * unmap_mapping_folio() - Unmap single folio from processes.
3365 * @folio: The locked folio to be unmapped.
3367 * Unmap this folio from any userspace process which still has it mmaped.
3368 * Typically, for efficiency, the range of nearby pages has already been
3369 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3370 * truncation or invalidation holds the lock on a folio, it may find that
3371 * the page has been remapped again: and then uses unmap_mapping_folio()
3372 * to unmap it finally.
3374 void unmap_mapping_folio(struct folio *folio)
3376 struct address_space *mapping = folio->mapping;
3377 struct zap_details details = { };
3378 pgoff_t first_index;
3381 VM_BUG_ON(!folio_test_locked(folio));
3383 first_index = folio->index;
3384 last_index = folio->index + folio_nr_pages(folio) - 1;
3386 details.even_cows = false;
3387 details.single_folio = folio;
3389 i_mmap_lock_write(mapping);
3390 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3391 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3392 last_index, &details);
3393 i_mmap_unlock_write(mapping);
3397 * unmap_mapping_pages() - Unmap pages from processes.
3398 * @mapping: The address space containing pages to be unmapped.
3399 * @start: Index of first page to be unmapped.
3400 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3401 * @even_cows: Whether to unmap even private COWed pages.
3403 * Unmap the pages in this address space from any userspace process which
3404 * has them mmaped. Generally, you want to remove COWed pages as well when
3405 * a file is being truncated, but not when invalidating pages from the page
3408 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3409 pgoff_t nr, bool even_cows)
3411 struct zap_details details = { };
3412 pgoff_t first_index = start;
3413 pgoff_t last_index = start + nr - 1;
3415 details.even_cows = even_cows;
3416 if (last_index < first_index)
3417 last_index = ULONG_MAX;
3419 i_mmap_lock_write(mapping);
3420 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3421 unmap_mapping_range_tree(&mapping->i_mmap, first_index,
3422 last_index, &details);
3423 i_mmap_unlock_write(mapping);
3425 EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3428 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3429 * address_space corresponding to the specified byte range in the underlying
3432 * @mapping: the address space containing mmaps to be unmapped.
3433 * @holebegin: byte in first page to unmap, relative to the start of
3434 * the underlying file. This will be rounded down to a PAGE_SIZE
3435 * boundary. Note that this is different from truncate_pagecache(), which
3436 * must keep the partial page. In contrast, we must get rid of
3438 * @holelen: size of prospective hole in bytes. This will be rounded
3439 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3441 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3442 * but 0 when invalidating pagecache, don't throw away private data.
3444 void unmap_mapping_range(struct address_space *mapping,
3445 loff_t const holebegin, loff_t const holelen, int even_cows)
3447 pgoff_t hba = holebegin >> PAGE_SHIFT;
3448 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3450 /* Check for overflow. */
3451 if (sizeof(holelen) > sizeof(hlen)) {
3453 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3454 if (holeend & ~(long long)ULONG_MAX)
3455 hlen = ULONG_MAX - hba + 1;
3458 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3460 EXPORT_SYMBOL(unmap_mapping_range);
3463 * Restore a potential device exclusive pte to a working pte entry
3465 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3467 struct page *page = vmf->page;
3468 struct vm_area_struct *vma = vmf->vma;
3469 struct mmu_notifier_range range;
3471 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
3472 return VM_FAULT_RETRY;
3473 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
3474 vma->vm_mm, vmf->address & PAGE_MASK,
3475 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3476 mmu_notifier_invalidate_range_start(&range);
3478 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3480 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3481 restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
3483 pte_unmap_unlock(vmf->pte, vmf->ptl);
3486 mmu_notifier_invalidate_range_end(&range);
3491 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3492 * but allow concurrent faults), and pte mapped but not yet locked.
3493 * We return with pte unmapped and unlocked.
3495 * We return with the mmap_lock locked or unlocked in the same cases
3496 * as does filemap_fault().
3498 vm_fault_t do_swap_page(struct vm_fault *vmf)
3500 struct vm_area_struct *vma = vmf->vma;
3501 struct page *page = NULL, *swapcache;
3502 struct swap_info_struct *si = NULL;
3508 void *shadow = NULL;
3510 if (!pte_unmap_same(vmf))
3513 entry = pte_to_swp_entry(vmf->orig_pte);
3514 if (unlikely(non_swap_entry(entry))) {
3515 if (is_migration_entry(entry)) {
3516 migration_entry_wait(vma->vm_mm, vmf->pmd,
3518 } else if (is_device_exclusive_entry(entry)) {
3519 vmf->page = pfn_swap_entry_to_page(entry);
3520 ret = remove_device_exclusive_entry(vmf);
3521 } else if (is_device_private_entry(entry)) {
3522 vmf->page = pfn_swap_entry_to_page(entry);
3523 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3524 } else if (is_hwpoison_entry(entry)) {
3525 ret = VM_FAULT_HWPOISON;
3527 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3528 ret = VM_FAULT_SIGBUS;
3533 /* Prevent swapoff from happening to us. */
3534 si = get_swap_device(entry);
3538 page = lookup_swap_cache(entry, vma, vmf->address);
3542 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3543 __swap_count(entry) == 1) {
3544 /* skip swapcache */
3545 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3548 __SetPageLocked(page);
3549 __SetPageSwapBacked(page);
3551 if (mem_cgroup_swapin_charge_page(page,
3552 vma->vm_mm, GFP_KERNEL, entry)) {
3556 mem_cgroup_swapin_uncharge_swap(entry);
3558 shadow = get_shadow_from_swap_cache(entry);
3560 workingset_refault(page_folio(page),
3563 lru_cache_add(page);
3565 /* To provide entry to swap_readpage() */
3566 set_page_private(page, entry.val);
3567 swap_readpage(page, true);
3568 set_page_private(page, 0);
3571 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3578 * Back out if somebody else faulted in this pte
3579 * while we released the pte lock.
3581 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3582 vmf->address, &vmf->ptl);
3583 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3588 /* Had to read the page from swap area: Major fault */
3589 ret = VM_FAULT_MAJOR;
3590 count_vm_event(PGMAJFAULT);
3591 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3592 } else if (PageHWPoison(page)) {
3594 * hwpoisoned dirty swapcache pages are kept for killing
3595 * owner processes (which may be unknown at hwpoison time)
3597 ret = VM_FAULT_HWPOISON;
3601 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3604 ret |= VM_FAULT_RETRY;
3609 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3610 * release the swapcache from under us. The page pin, and pte_same
3611 * test below, are not enough to exclude that. Even if it is still
3612 * swapcache, we need to check that the page's swap has not changed.
3614 if (unlikely((!PageSwapCache(page) ||
3615 page_private(page) != entry.val)) && swapcache)
3618 page = ksm_might_need_to_copy(page, vma, vmf->address);
3619 if (unlikely(!page)) {
3625 cgroup_throttle_swaprate(page, GFP_KERNEL);
3628 * Back out if somebody else already faulted in this pte.
3630 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3632 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3635 if (unlikely(!PageUptodate(page))) {
3636 ret = VM_FAULT_SIGBUS;
3641 * The page isn't present yet, go ahead with the fault.
3643 * Be careful about the sequence of operations here.
3644 * To get its accounting right, reuse_swap_page() must be called
3645 * while the page is counted on swap but not yet in mapcount i.e.
3646 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3647 * must be called after the swap_free(), or it will never succeed.
3650 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3651 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3652 pte = mk_pte(page, vma->vm_page_prot);
3653 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3654 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3655 vmf->flags &= ~FAULT_FLAG_WRITE;
3656 ret |= VM_FAULT_WRITE;
3657 exclusive = RMAP_EXCLUSIVE;
3659 flush_icache_page(vma, page);
3660 if (pte_swp_soft_dirty(vmf->orig_pte))
3661 pte = pte_mksoft_dirty(pte);
3662 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3663 pte = pte_mkuffd_wp(pte);
3664 pte = pte_wrprotect(pte);
3666 vmf->orig_pte = pte;
3668 /* ksm created a completely new copy */
3669 if (unlikely(page != swapcache && swapcache)) {
3670 page_add_new_anon_rmap(page, vma, vmf->address, false);
3671 lru_cache_add_inactive_or_unevictable(page, vma);
3673 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3676 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3677 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3680 if (mem_cgroup_swap_full(page) ||
3681 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3682 try_to_free_swap(page);
3684 if (page != swapcache && swapcache) {
3686 * Hold the lock to avoid the swap entry to be reused
3687 * until we take the PT lock for the pte_same() check
3688 * (to avoid false positives from pte_same). For
3689 * further safety release the lock after the swap_free
3690 * so that the swap count won't change under a
3691 * parallel locked swapcache.
3693 unlock_page(swapcache);
3694 put_page(swapcache);
3697 if (vmf->flags & FAULT_FLAG_WRITE) {
3698 ret |= do_wp_page(vmf);
3699 if (ret & VM_FAULT_ERROR)
3700 ret &= VM_FAULT_ERROR;
3704 /* No need to invalidate - it was non-present before */
3705 update_mmu_cache(vma, vmf->address, vmf->pte);
3707 pte_unmap_unlock(vmf->pte, vmf->ptl);
3710 put_swap_device(si);
3713 pte_unmap_unlock(vmf->pte, vmf->ptl);
3718 if (page != swapcache && swapcache) {
3719 unlock_page(swapcache);
3720 put_page(swapcache);
3723 put_swap_device(si);
3728 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3729 * but allow concurrent faults), and pte mapped but not yet locked.
3730 * We return with mmap_lock still held, but pte unmapped and unlocked.
3732 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3734 struct vm_area_struct *vma = vmf->vma;
3739 /* File mapping without ->vm_ops ? */
3740 if (vma->vm_flags & VM_SHARED)
3741 return VM_FAULT_SIGBUS;
3744 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3745 * pte_offset_map() on pmds where a huge pmd might be created
3746 * from a different thread.
3748 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3749 * parallel threads are excluded by other means.
3751 * Here we only have mmap_read_lock(mm).
3753 if (pte_alloc(vma->vm_mm, vmf->pmd))
3754 return VM_FAULT_OOM;
3756 /* See comment in handle_pte_fault() */
3757 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3760 /* Use the zero-page for reads */
3761 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3762 !mm_forbids_zeropage(vma->vm_mm)) {
3763 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3764 vma->vm_page_prot));
3765 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3766 vmf->address, &vmf->ptl);
3767 if (!pte_none(*vmf->pte)) {
3768 update_mmu_tlb(vma, vmf->address, vmf->pte);
3771 ret = check_stable_address_space(vma->vm_mm);
3774 /* Deliver the page fault to userland, check inside PT lock */
3775 if (userfaultfd_missing(vma)) {
3776 pte_unmap_unlock(vmf->pte, vmf->ptl);
3777 return handle_userfault(vmf, VM_UFFD_MISSING);
3782 /* Allocate our own private page. */
3783 if (unlikely(anon_vma_prepare(vma)))
3785 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3789 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
3791 cgroup_throttle_swaprate(page, GFP_KERNEL);
3794 * The memory barrier inside __SetPageUptodate makes sure that
3795 * preceding stores to the page contents become visible before
3796 * the set_pte_at() write.
3798 __SetPageUptodate(page);
3800 entry = mk_pte(page, vma->vm_page_prot);
3801 entry = pte_sw_mkyoung(entry);
3802 if (vma->vm_flags & VM_WRITE)
3803 entry = pte_mkwrite(pte_mkdirty(entry));
3805 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3807 if (!pte_none(*vmf->pte)) {
3808 update_mmu_cache(vma, vmf->address, vmf->pte);
3812 ret = check_stable_address_space(vma->vm_mm);
3816 /* Deliver the page fault to userland, check inside PT lock */
3817 if (userfaultfd_missing(vma)) {
3818 pte_unmap_unlock(vmf->pte, vmf->ptl);
3820 return handle_userfault(vmf, VM_UFFD_MISSING);
3823 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3824 page_add_new_anon_rmap(page, vma, vmf->address, false);
3825 lru_cache_add_inactive_or_unevictable(page, vma);
3827 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3829 /* No need to invalidate - it was non-present before */
3830 update_mmu_cache(vma, vmf->address, vmf->pte);
3832 pte_unmap_unlock(vmf->pte, vmf->ptl);
3840 return VM_FAULT_OOM;
3844 * The mmap_lock must have been held on entry, and may have been
3845 * released depending on flags and vma->vm_ops->fault() return value.
3846 * See filemap_fault() and __lock_page_retry().
3848 static vm_fault_t __do_fault(struct vm_fault *vmf)
3850 struct vm_area_struct *vma = vmf->vma;
3854 * Preallocate pte before we take page_lock because this might lead to
3855 * deadlocks for memcg reclaim which waits for pages under writeback:
3857 * SetPageWriteback(A)
3863 * wait_on_page_writeback(A)
3864 * SetPageWriteback(B)
3866 * # flush A, B to clear the writeback
3868 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3869 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3870 if (!vmf->prealloc_pte)
3871 return VM_FAULT_OOM;
3874 ret = vma->vm_ops->fault(vmf);
3875 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3876 VM_FAULT_DONE_COW)))
3879 if (unlikely(PageHWPoison(vmf->page))) {
3880 if (ret & VM_FAULT_LOCKED)
3881 unlock_page(vmf->page);
3882 put_page(vmf->page);
3884 return VM_FAULT_HWPOISON;
3887 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3888 lock_page(vmf->page);
3890 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3895 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3896 static void deposit_prealloc_pte(struct vm_fault *vmf)
3898 struct vm_area_struct *vma = vmf->vma;
3900 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3902 * We are going to consume the prealloc table,
3903 * count that as nr_ptes.
3905 mm_inc_nr_ptes(vma->vm_mm);
3906 vmf->prealloc_pte = NULL;
3909 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3911 struct vm_area_struct *vma = vmf->vma;
3912 bool write = vmf->flags & FAULT_FLAG_WRITE;
3913 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3916 vm_fault_t ret = VM_FAULT_FALLBACK;
3918 if (!transhuge_vma_suitable(vma, haddr))
3921 page = compound_head(page);
3922 if (compound_order(page) != HPAGE_PMD_ORDER)
3926 * Just backoff if any subpage of a THP is corrupted otherwise
3927 * the corrupted page may mapped by PMD silently to escape the
3928 * check. This kind of THP just can be PTE mapped. Access to
3929 * the corrupted subpage should trigger SIGBUS as expected.
3931 if (unlikely(PageHasHWPoisoned(page)))
3935 * Archs like ppc64 need additional space to store information
3936 * related to pte entry. Use the preallocated table for that.
3938 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3939 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3940 if (!vmf->prealloc_pte)
3941 return VM_FAULT_OOM;
3944 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3945 if (unlikely(!pmd_none(*vmf->pmd)))
3948 for (i = 0; i < HPAGE_PMD_NR; i++)
3949 flush_icache_page(vma, page + i);
3951 entry = mk_huge_pmd(page, vma->vm_page_prot);
3953 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3955 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3956 page_add_file_rmap(page, true);
3958 * deposit and withdraw with pmd lock held
3960 if (arch_needs_pgtable_deposit())
3961 deposit_prealloc_pte(vmf);
3963 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3965 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3967 /* fault is handled */
3969 count_vm_event(THP_FILE_MAPPED);
3971 spin_unlock(vmf->ptl);
3975 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3977 return VM_FAULT_FALLBACK;
3981 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3983 struct vm_area_struct *vma = vmf->vma;
3984 bool write = vmf->flags & FAULT_FLAG_WRITE;
3985 bool prefault = vmf->address != addr;
3988 flush_icache_page(vma, page);
3989 entry = mk_pte(page, vma->vm_page_prot);
3991 if (prefault && arch_wants_old_prefaulted_pte())
3992 entry = pte_mkold(entry);
3994 entry = pte_sw_mkyoung(entry);
3997 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3998 /* copy-on-write page */
3999 if (write && !(vma->vm_flags & VM_SHARED)) {
4000 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
4001 page_add_new_anon_rmap(page, vma, addr, false);
4002 lru_cache_add_inactive_or_unevictable(page, vma);
4004 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
4005 page_add_file_rmap(page, false);
4007 set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
4011 * finish_fault - finish page fault once we have prepared the page to fault
4013 * @vmf: structure describing the fault
4015 * This function handles all that is needed to finish a page fault once the
4016 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4017 * given page, adds reverse page mapping, handles memcg charges and LRU
4020 * The function expects the page to be locked and on success it consumes a
4021 * reference of a page being mapped (for the PTE which maps it).
4023 * Return: %0 on success, %VM_FAULT_ code in case of error.
4025 vm_fault_t finish_fault(struct vm_fault *vmf)
4027 struct vm_area_struct *vma = vmf->vma;
4031 /* Did we COW the page? */
4032 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4033 page = vmf->cow_page;
4038 * check even for read faults because we might have lost our CoWed
4041 if (!(vma->vm_flags & VM_SHARED)) {
4042 ret = check_stable_address_space(vma->vm_mm);
4047 if (pmd_none(*vmf->pmd)) {
4048 if (PageTransCompound(page)) {
4049 ret = do_set_pmd(vmf, page);
4050 if (ret != VM_FAULT_FALLBACK)
4054 if (vmf->prealloc_pte)
4055 pmd_install(vma->vm_mm, vmf->pmd, &vmf->prealloc_pte);
4056 else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
4057 return VM_FAULT_OOM;
4060 /* See comment in handle_pte_fault() */
4061 if (pmd_devmap_trans_unstable(vmf->pmd))
4064 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4065 vmf->address, &vmf->ptl);
4067 /* Re-check under ptl */
4068 if (likely(pte_none(*vmf->pte)))
4069 do_set_pte(vmf, page, vmf->address);
4071 ret = VM_FAULT_NOPAGE;
4073 update_mmu_tlb(vma, vmf->address, vmf->pte);
4074 pte_unmap_unlock(vmf->pte, vmf->ptl);
4078 static unsigned long fault_around_bytes __read_mostly =
4079 rounddown_pow_of_two(65536);
4081 #ifdef CONFIG_DEBUG_FS
4082 static int fault_around_bytes_get(void *data, u64 *val)
4084 *val = fault_around_bytes;
4089 * fault_around_bytes must be rounded down to the nearest page order as it's
4090 * what do_fault_around() expects to see.
4092 static int fault_around_bytes_set(void *data, u64 val)
4094 if (val / PAGE_SIZE > PTRS_PER_PTE)
4096 if (val > PAGE_SIZE)
4097 fault_around_bytes = rounddown_pow_of_two(val);
4099 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
4102 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4103 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4105 static int __init fault_around_debugfs(void)
4107 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4108 &fault_around_bytes_fops);
4111 late_initcall(fault_around_debugfs);
4115 * do_fault_around() tries to map few pages around the fault address. The hope
4116 * is that the pages will be needed soon and this will lower the number of
4119 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4120 * not ready to be mapped: not up-to-date, locked, etc.
4122 * This function is called with the page table lock taken. In the split ptlock
4123 * case the page table lock only protects only those entries which belong to
4124 * the page table corresponding to the fault address.
4126 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4129 * fault_around_bytes defines how many bytes we'll try to map.
4130 * do_fault_around() expects it to be set to a power of two less than or equal
4133 * The virtual address of the area that we map is naturally aligned to
4134 * fault_around_bytes rounded down to the machine page size
4135 * (and therefore to page order). This way it's easier to guarantee
4136 * that we don't cross page table boundaries.
4138 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4140 unsigned long address = vmf->address, nr_pages, mask;
4141 pgoff_t start_pgoff = vmf->pgoff;
4145 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4146 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4148 address = max(address & mask, vmf->vma->vm_start);
4149 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4153 * end_pgoff is either the end of the page table, the end of
4154 * the vma or nr_pages from start_pgoff, depending what is nearest.
4156 end_pgoff = start_pgoff -
4157 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4159 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4160 start_pgoff + nr_pages - 1);
4162 if (pmd_none(*vmf->pmd)) {
4163 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4164 if (!vmf->prealloc_pte)
4165 return VM_FAULT_OOM;
4168 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4171 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4173 struct vm_area_struct *vma = vmf->vma;
4177 * Let's call ->map_pages() first and use ->fault() as fallback
4178 * if page by the offset is not ready to be mapped (cold cache or
4181 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4182 if (likely(!userfaultfd_minor(vmf->vma))) {
4183 ret = do_fault_around(vmf);
4189 ret = __do_fault(vmf);
4190 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4193 ret |= finish_fault(vmf);
4194 unlock_page(vmf->page);
4195 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4196 put_page(vmf->page);
4200 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4202 struct vm_area_struct *vma = vmf->vma;
4205 if (unlikely(anon_vma_prepare(vma)))
4206 return VM_FAULT_OOM;
4208 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4210 return VM_FAULT_OOM;
4212 if (mem_cgroup_charge(page_folio(vmf->cow_page), vma->vm_mm,
4214 put_page(vmf->cow_page);
4215 return VM_FAULT_OOM;
4217 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4219 ret = __do_fault(vmf);
4220 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4222 if (ret & VM_FAULT_DONE_COW)
4225 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4226 __SetPageUptodate(vmf->cow_page);
4228 ret |= finish_fault(vmf);
4229 unlock_page(vmf->page);
4230 put_page(vmf->page);
4231 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4235 put_page(vmf->cow_page);
4239 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4241 struct vm_area_struct *vma = vmf->vma;
4242 vm_fault_t ret, tmp;
4244 ret = __do_fault(vmf);
4245 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4249 * Check if the backing address space wants to know that the page is
4250 * about to become writable
4252 if (vma->vm_ops->page_mkwrite) {
4253 unlock_page(vmf->page);
4254 tmp = do_page_mkwrite(vmf);
4255 if (unlikely(!tmp ||
4256 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4257 put_page(vmf->page);
4262 ret |= finish_fault(vmf);
4263 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4265 unlock_page(vmf->page);
4266 put_page(vmf->page);
4270 ret |= fault_dirty_shared_page(vmf);
4275 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4276 * but allow concurrent faults).
4277 * The mmap_lock may have been released depending on flags and our
4278 * return value. See filemap_fault() and __folio_lock_or_retry().
4279 * If mmap_lock is released, vma may become invalid (for example
4280 * by other thread calling munmap()).
4282 static vm_fault_t do_fault(struct vm_fault *vmf)
4284 struct vm_area_struct *vma = vmf->vma;
4285 struct mm_struct *vm_mm = vma->vm_mm;
4289 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4291 if (!vma->vm_ops->fault) {
4293 * If we find a migration pmd entry or a none pmd entry, which
4294 * should never happen, return SIGBUS
4296 if (unlikely(!pmd_present(*vmf->pmd)))
4297 ret = VM_FAULT_SIGBUS;
4299 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4304 * Make sure this is not a temporary clearing of pte
4305 * by holding ptl and checking again. A R/M/W update
4306 * of pte involves: take ptl, clearing the pte so that
4307 * we don't have concurrent modification by hardware
4308 * followed by an update.
4310 if (unlikely(pte_none(*vmf->pte)))
4311 ret = VM_FAULT_SIGBUS;
4313 ret = VM_FAULT_NOPAGE;
4315 pte_unmap_unlock(vmf->pte, vmf->ptl);
4317 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4318 ret = do_read_fault(vmf);
4319 else if (!(vma->vm_flags & VM_SHARED))
4320 ret = do_cow_fault(vmf);
4322 ret = do_shared_fault(vmf);
4324 /* preallocated pagetable is unused: free it */
4325 if (vmf->prealloc_pte) {
4326 pte_free(vm_mm, vmf->prealloc_pte);
4327 vmf->prealloc_pte = NULL;
4332 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4333 unsigned long addr, int page_nid, int *flags)
4337 count_vm_numa_event(NUMA_HINT_FAULTS);
4338 if (page_nid == numa_node_id()) {
4339 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4340 *flags |= TNF_FAULT_LOCAL;
4343 return mpol_misplaced(page, vma, addr);
4346 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4348 struct vm_area_struct *vma = vmf->vma;
4349 struct page *page = NULL;
4350 int page_nid = NUMA_NO_NODE;
4354 bool was_writable = pte_savedwrite(vmf->orig_pte);
4358 * The "pte" at this point cannot be used safely without
4359 * validation through pte_unmap_same(). It's of NUMA type but
4360 * the pfn may be screwed if the read is non atomic.
4362 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4363 spin_lock(vmf->ptl);
4364 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4365 pte_unmap_unlock(vmf->pte, vmf->ptl);
4369 /* Get the normal PTE */
4370 old_pte = ptep_get(vmf->pte);
4371 pte = pte_modify(old_pte, vma->vm_page_prot);
4373 page = vm_normal_page(vma, vmf->address, pte);
4377 /* TODO: handle PTE-mapped THP */
4378 if (PageCompound(page))
4382 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4383 * much anyway since they can be in shared cache state. This misses
4384 * the case where a mapping is writable but the process never writes
4385 * to it but pte_write gets cleared during protection updates and
4386 * pte_dirty has unpredictable behaviour between PTE scan updates,
4387 * background writeback, dirty balancing and application behaviour.
4390 flags |= TNF_NO_GROUP;
4393 * Flag if the page is shared between multiple address spaces. This
4394 * is later used when determining whether to group tasks together
4396 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4397 flags |= TNF_SHARED;
4399 last_cpupid = page_cpupid_last(page);
4400 page_nid = page_to_nid(page);
4401 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4403 if (target_nid == NUMA_NO_NODE) {
4407 pte_unmap_unlock(vmf->pte, vmf->ptl);
4409 /* Migrate to the requested node */
4410 if (migrate_misplaced_page(page, vma, target_nid)) {
4411 page_nid = target_nid;
4412 flags |= TNF_MIGRATED;
4414 flags |= TNF_MIGRATE_FAIL;
4415 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4416 spin_lock(vmf->ptl);
4417 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4418 pte_unmap_unlock(vmf->pte, vmf->ptl);
4425 if (page_nid != NUMA_NO_NODE)
4426 task_numa_fault(last_cpupid, page_nid, 1, flags);
4430 * Make it present again, depending on how arch implements
4431 * non-accessible ptes, some can allow access by kernel mode.
4433 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4434 pte = pte_modify(old_pte, vma->vm_page_prot);
4435 pte = pte_mkyoung(pte);
4437 pte = pte_mkwrite(pte);
4438 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4439 update_mmu_cache(vma, vmf->address, vmf->pte);
4440 pte_unmap_unlock(vmf->pte, vmf->ptl);
4444 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4446 if (vma_is_anonymous(vmf->vma))
4447 return do_huge_pmd_anonymous_page(vmf);
4448 if (vmf->vma->vm_ops->huge_fault)
4449 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4450 return VM_FAULT_FALLBACK;
4453 /* `inline' is required to avoid gcc 4.1.2 build error */
4454 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4456 if (vma_is_anonymous(vmf->vma)) {
4457 if (userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd))
4458 return handle_userfault(vmf, VM_UFFD_WP);
4459 return do_huge_pmd_wp_page(vmf);
4461 if (vmf->vma->vm_ops->huge_fault) {
4462 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4464 if (!(ret & VM_FAULT_FALLBACK))
4468 /* COW or write-notify handled on pte level: split pmd. */
4469 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4471 return VM_FAULT_FALLBACK;
4474 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4476 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4477 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4478 /* No support for anonymous transparent PUD pages yet */
4479 if (vma_is_anonymous(vmf->vma))
4481 if (vmf->vma->vm_ops->huge_fault) {
4482 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4484 if (!(ret & VM_FAULT_FALLBACK))
4488 /* COW or write-notify not handled on PUD level: split pud.*/
4489 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4490 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4491 return VM_FAULT_FALLBACK;
4494 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4496 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4497 /* No support for anonymous transparent PUD pages yet */
4498 if (vma_is_anonymous(vmf->vma))
4499 return VM_FAULT_FALLBACK;
4500 if (vmf->vma->vm_ops->huge_fault)
4501 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4502 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4503 return VM_FAULT_FALLBACK;
4507 * These routines also need to handle stuff like marking pages dirty
4508 * and/or accessed for architectures that don't do it in hardware (most
4509 * RISC architectures). The early dirtying is also good on the i386.
4511 * There is also a hook called "update_mmu_cache()" that architectures
4512 * with external mmu caches can use to update those (ie the Sparc or
4513 * PowerPC hashed page tables that act as extended TLBs).
4515 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4516 * concurrent faults).
4518 * The mmap_lock may have been released depending on flags and our return value.
4519 * See filemap_fault() and __folio_lock_or_retry().
4521 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4525 if (unlikely(pmd_none(*vmf->pmd))) {
4527 * Leave __pte_alloc() until later: because vm_ops->fault may
4528 * want to allocate huge page, and if we expose page table
4529 * for an instant, it will be difficult to retract from
4530 * concurrent faults and from rmap lookups.
4535 * If a huge pmd materialized under us just retry later. Use
4536 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4537 * of pmd_trans_huge() to ensure the pmd didn't become
4538 * pmd_trans_huge under us and then back to pmd_none, as a
4539 * result of MADV_DONTNEED running immediately after a huge pmd
4540 * fault in a different thread of this mm, in turn leading to a
4541 * misleading pmd_trans_huge() retval. All we have to ensure is
4542 * that it is a regular pmd that we can walk with
4543 * pte_offset_map() and we can do that through an atomic read
4544 * in C, which is what pmd_trans_unstable() provides.
4546 if (pmd_devmap_trans_unstable(vmf->pmd))
4549 * A regular pmd is established and it can't morph into a huge
4550 * pmd from under us anymore at this point because we hold the
4551 * mmap_lock read mode and khugepaged takes it in write mode.
4552 * So now it's safe to run pte_offset_map().
4554 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4555 vmf->orig_pte = *vmf->pte;
4558 * some architectures can have larger ptes than wordsize,
4559 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4560 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4561 * accesses. The code below just needs a consistent view
4562 * for the ifs and we later double check anyway with the
4563 * ptl lock held. So here a barrier will do.
4566 if (pte_none(vmf->orig_pte)) {
4567 pte_unmap(vmf->pte);
4573 if (vma_is_anonymous(vmf->vma))
4574 return do_anonymous_page(vmf);
4576 return do_fault(vmf);
4579 if (!pte_present(vmf->orig_pte))
4580 return do_swap_page(vmf);
4582 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4583 return do_numa_page(vmf);
4585 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4586 spin_lock(vmf->ptl);
4587 entry = vmf->orig_pte;
4588 if (unlikely(!pte_same(*vmf->pte, entry))) {
4589 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4592 if (vmf->flags & FAULT_FLAG_WRITE) {
4593 if (!pte_write(entry))
4594 return do_wp_page(vmf);
4595 entry = pte_mkdirty(entry);
4597 entry = pte_mkyoung(entry);
4598 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4599 vmf->flags & FAULT_FLAG_WRITE)) {
4600 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4602 /* Skip spurious TLB flush for retried page fault */
4603 if (vmf->flags & FAULT_FLAG_TRIED)
4606 * This is needed only for protection faults but the arch code
4607 * is not yet telling us if this is a protection fault or not.
4608 * This still avoids useless tlb flushes for .text page faults
4611 if (vmf->flags & FAULT_FLAG_WRITE)
4612 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4615 pte_unmap_unlock(vmf->pte, vmf->ptl);
4620 * By the time we get here, we already hold the mm semaphore
4622 * The mmap_lock may have been released depending on flags and our
4623 * return value. See filemap_fault() and __folio_lock_or_retry().
4625 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4626 unsigned long address, unsigned int flags)
4628 struct vm_fault vmf = {
4630 .address = address & PAGE_MASK,
4632 .pgoff = linear_page_index(vma, address),
4633 .gfp_mask = __get_fault_gfp_mask(vma),
4635 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4636 struct mm_struct *mm = vma->vm_mm;
4641 pgd = pgd_offset(mm, address);
4642 p4d = p4d_alloc(mm, pgd, address);
4644 return VM_FAULT_OOM;
4646 vmf.pud = pud_alloc(mm, p4d, address);
4648 return VM_FAULT_OOM;
4650 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4651 ret = create_huge_pud(&vmf);
4652 if (!(ret & VM_FAULT_FALLBACK))
4655 pud_t orig_pud = *vmf.pud;
4658 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4660 /* NUMA case for anonymous PUDs would go here */
4662 if (dirty && !pud_write(orig_pud)) {
4663 ret = wp_huge_pud(&vmf, orig_pud);
4664 if (!(ret & VM_FAULT_FALLBACK))
4667 huge_pud_set_accessed(&vmf, orig_pud);
4673 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4675 return VM_FAULT_OOM;
4677 /* Huge pud page fault raced with pmd_alloc? */
4678 if (pud_trans_unstable(vmf.pud))
4681 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4682 ret = create_huge_pmd(&vmf);
4683 if (!(ret & VM_FAULT_FALLBACK))
4686 vmf.orig_pmd = *vmf.pmd;
4689 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
4690 VM_BUG_ON(thp_migration_supported() &&
4691 !is_pmd_migration_entry(vmf.orig_pmd));
4692 if (is_pmd_migration_entry(vmf.orig_pmd))
4693 pmd_migration_entry_wait(mm, vmf.pmd);
4696 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
4697 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
4698 return do_huge_pmd_numa_page(&vmf);
4700 if (dirty && !pmd_write(vmf.orig_pmd)) {
4701 ret = wp_huge_pmd(&vmf);
4702 if (!(ret & VM_FAULT_FALLBACK))
4705 huge_pmd_set_accessed(&vmf);
4711 return handle_pte_fault(&vmf);
4715 * mm_account_fault - Do page fault accounting
4717 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4718 * of perf event counters, but we'll still do the per-task accounting to
4719 * the task who triggered this page fault.
4720 * @address: the faulted address.
4721 * @flags: the fault flags.
4722 * @ret: the fault retcode.
4724 * This will take care of most of the page fault accounting. Meanwhile, it
4725 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4726 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4727 * still be in per-arch page fault handlers at the entry of page fault.
4729 static inline void mm_account_fault(struct pt_regs *regs,
4730 unsigned long address, unsigned int flags,
4736 * We don't do accounting for some specific faults:
4738 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4739 * includes arch_vma_access_permitted() failing before reaching here.
4740 * So this is not a "this many hardware page faults" counter. We
4741 * should use the hw profiling for that.
4743 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4744 * once they're completed.
4746 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4750 * We define the fault as a major fault when the final successful fault
4751 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4752 * handle it immediately previously).
4754 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4762 * If the fault is done for GUP, regs will be NULL. We only do the
4763 * accounting for the per thread fault counters who triggered the
4764 * fault, and we skip the perf event updates.
4770 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4772 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4776 * By the time we get here, we already hold the mm semaphore
4778 * The mmap_lock may have been released depending on flags and our
4779 * return value. See filemap_fault() and __folio_lock_or_retry().
4781 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4782 unsigned int flags, struct pt_regs *regs)
4786 __set_current_state(TASK_RUNNING);
4788 count_vm_event(PGFAULT);
4789 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4791 /* do counter updates before entering really critical section. */
4792 check_sync_rss_stat(current);
4794 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4795 flags & FAULT_FLAG_INSTRUCTION,
4796 flags & FAULT_FLAG_REMOTE))
4797 return VM_FAULT_SIGSEGV;
4800 * Enable the memcg OOM handling for faults triggered in user
4801 * space. Kernel faults are handled more gracefully.
4803 if (flags & FAULT_FLAG_USER)
4804 mem_cgroup_enter_user_fault();
4806 if (unlikely(is_vm_hugetlb_page(vma)))
4807 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4809 ret = __handle_mm_fault(vma, address, flags);
4811 if (flags & FAULT_FLAG_USER) {
4812 mem_cgroup_exit_user_fault();
4814 * The task may have entered a memcg OOM situation but
4815 * if the allocation error was handled gracefully (no
4816 * VM_FAULT_OOM), there is no need to kill anything.
4817 * Just clean up the OOM state peacefully.
4819 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4820 mem_cgroup_oom_synchronize(false);
4823 mm_account_fault(regs, address, flags, ret);
4827 EXPORT_SYMBOL_GPL(handle_mm_fault);
4829 #ifndef __PAGETABLE_P4D_FOLDED
4831 * Allocate p4d page table.
4832 * We've already handled the fast-path in-line.
4834 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4836 p4d_t *new = p4d_alloc_one(mm, address);
4840 spin_lock(&mm->page_table_lock);
4841 if (pgd_present(*pgd)) { /* Another has populated it */
4844 smp_wmb(); /* See comment in pmd_install() */
4845 pgd_populate(mm, pgd, new);
4847 spin_unlock(&mm->page_table_lock);
4850 #endif /* __PAGETABLE_P4D_FOLDED */
4852 #ifndef __PAGETABLE_PUD_FOLDED
4854 * Allocate page upper directory.
4855 * We've already handled the fast-path in-line.
4857 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4859 pud_t *new = pud_alloc_one(mm, address);
4863 spin_lock(&mm->page_table_lock);
4864 if (!p4d_present(*p4d)) {
4866 smp_wmb(); /* See comment in pmd_install() */
4867 p4d_populate(mm, p4d, new);
4868 } else /* Another has populated it */
4870 spin_unlock(&mm->page_table_lock);
4873 #endif /* __PAGETABLE_PUD_FOLDED */
4875 #ifndef __PAGETABLE_PMD_FOLDED
4877 * Allocate page middle directory.
4878 * We've already handled the fast-path in-line.
4880 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4883 pmd_t *new = pmd_alloc_one(mm, address);
4887 ptl = pud_lock(mm, pud);
4888 if (!pud_present(*pud)) {
4890 smp_wmb(); /* See comment in pmd_install() */
4891 pud_populate(mm, pud, new);
4892 } else { /* Another has populated it */
4898 #endif /* __PAGETABLE_PMD_FOLDED */
4900 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4901 struct mmu_notifier_range *range, pte_t **ptepp,
4902 pmd_t **pmdpp, spinlock_t **ptlp)
4910 pgd = pgd_offset(mm, address);
4911 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4914 p4d = p4d_offset(pgd, address);
4915 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4918 pud = pud_offset(p4d, address);
4919 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4922 pmd = pmd_offset(pud, address);
4923 VM_BUG_ON(pmd_trans_huge(*pmd));
4925 if (pmd_huge(*pmd)) {
4930 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4931 NULL, mm, address & PMD_MASK,
4932 (address & PMD_MASK) + PMD_SIZE);
4933 mmu_notifier_invalidate_range_start(range);
4935 *ptlp = pmd_lock(mm, pmd);
4936 if (pmd_huge(*pmd)) {
4942 mmu_notifier_invalidate_range_end(range);
4945 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4949 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4950 address & PAGE_MASK,
4951 (address & PAGE_MASK) + PAGE_SIZE);
4952 mmu_notifier_invalidate_range_start(range);
4954 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4955 if (!pte_present(*ptep))
4960 pte_unmap_unlock(ptep, *ptlp);
4962 mmu_notifier_invalidate_range_end(range);
4968 * follow_pte - look up PTE at a user virtual address
4969 * @mm: the mm_struct of the target address space
4970 * @address: user virtual address
4971 * @ptepp: location to store found PTE
4972 * @ptlp: location to store the lock for the PTE
4974 * On a successful return, the pointer to the PTE is stored in @ptepp;
4975 * the corresponding lock is taken and its location is stored in @ptlp.
4976 * The contents of the PTE are only stable until @ptlp is released;
4977 * any further use, if any, must be protected against invalidation
4978 * with MMU notifiers.
4980 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4981 * should be taken for read.
4983 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4984 * it is not a good general-purpose API.
4986 * Return: zero on success, -ve otherwise.
4988 int follow_pte(struct mm_struct *mm, unsigned long address,
4989 pte_t **ptepp, spinlock_t **ptlp)
4991 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4993 EXPORT_SYMBOL_GPL(follow_pte);
4996 * follow_pfn - look up PFN at a user virtual address
4997 * @vma: memory mapping
4998 * @address: user virtual address
4999 * @pfn: location to store found PFN
5001 * Only IO mappings and raw PFN mappings are allowed.
5003 * This function does not allow the caller to read the permissions
5004 * of the PTE. Do not use it.
5006 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5008 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5015 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5018 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5021 *pfn = pte_pfn(*ptep);
5022 pte_unmap_unlock(ptep, ptl);
5025 EXPORT_SYMBOL(follow_pfn);
5027 #ifdef CONFIG_HAVE_IOREMAP_PROT
5028 int follow_phys(struct vm_area_struct *vma,
5029 unsigned long address, unsigned int flags,
5030 unsigned long *prot, resource_size_t *phys)
5036 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5039 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5043 if ((flags & FOLL_WRITE) && !pte_write(pte))
5046 *prot = pgprot_val(pte_pgprot(pte));
5047 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5051 pte_unmap_unlock(ptep, ptl);
5057 * generic_access_phys - generic implementation for iomem mmap access
5058 * @vma: the vma to access
5059 * @addr: userspace address, not relative offset within @vma
5060 * @buf: buffer to read/write
5061 * @len: length of transfer
5062 * @write: set to FOLL_WRITE when writing, otherwise reading
5064 * This is a generic implementation for &vm_operations_struct.access for an
5065 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5068 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5069 void *buf, int len, int write)
5071 resource_size_t phys_addr;
5072 unsigned long prot = 0;
5073 void __iomem *maddr;
5076 int offset = offset_in_page(addr);
5079 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5083 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5086 pte_unmap_unlock(ptep, ptl);
5088 prot = pgprot_val(pte_pgprot(pte));
5089 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5091 if ((write & FOLL_WRITE) && !pte_write(pte))
5094 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5098 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5101 if (!pte_same(pte, *ptep)) {
5102 pte_unmap_unlock(ptep, ptl);
5109 memcpy_toio(maddr + offset, buf, len);
5111 memcpy_fromio(buf, maddr + offset, len);
5113 pte_unmap_unlock(ptep, ptl);
5119 EXPORT_SYMBOL_GPL(generic_access_phys);
5123 * Access another process' address space as given in mm.
5125 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5126 int len, unsigned int gup_flags)
5128 struct vm_area_struct *vma;
5129 void *old_buf = buf;
5130 int write = gup_flags & FOLL_WRITE;
5132 if (mmap_read_lock_killable(mm))
5135 /* ignore errors, just check how much was successfully transferred */
5137 int bytes, ret, offset;
5139 struct page *page = NULL;
5141 ret = get_user_pages_remote(mm, addr, 1,
5142 gup_flags, &page, &vma, NULL);
5144 #ifndef CONFIG_HAVE_IOREMAP_PROT
5148 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5149 * we can access using slightly different code.
5151 vma = vma_lookup(mm, addr);
5154 if (vma->vm_ops && vma->vm_ops->access)
5155 ret = vma->vm_ops->access(vma, addr, buf,
5163 offset = addr & (PAGE_SIZE-1);
5164 if (bytes > PAGE_SIZE-offset)
5165 bytes = PAGE_SIZE-offset;
5169 copy_to_user_page(vma, page, addr,
5170 maddr + offset, buf, bytes);
5171 set_page_dirty_lock(page);
5173 copy_from_user_page(vma, page, addr,
5174 buf, maddr + offset, bytes);
5183 mmap_read_unlock(mm);
5185 return buf - old_buf;
5189 * access_remote_vm - access another process' address space
5190 * @mm: the mm_struct of the target address space
5191 * @addr: start address to access
5192 * @buf: source or destination buffer
5193 * @len: number of bytes to transfer
5194 * @gup_flags: flags modifying lookup behaviour
5196 * The caller must hold a reference on @mm.
5198 * Return: number of bytes copied from source to destination.
5200 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5201 void *buf, int len, unsigned int gup_flags)
5203 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5207 * Access another process' address space.
5208 * Source/target buffer must be kernel space,
5209 * Do not walk the page table directly, use get_user_pages
5211 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5212 void *buf, int len, unsigned int gup_flags)
5214 struct mm_struct *mm;
5217 mm = get_task_mm(tsk);
5221 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5227 EXPORT_SYMBOL_GPL(access_process_vm);
5230 * Print the name of a VMA.
5232 void print_vma_addr(char *prefix, unsigned long ip)
5234 struct mm_struct *mm = current->mm;
5235 struct vm_area_struct *vma;
5238 * we might be running from an atomic context so we cannot sleep
5240 if (!mmap_read_trylock(mm))
5243 vma = find_vma(mm, ip);
5244 if (vma && vma->vm_file) {
5245 struct file *f = vma->vm_file;
5246 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5250 p = file_path(f, buf, PAGE_SIZE);
5253 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5255 vma->vm_end - vma->vm_start);
5256 free_page((unsigned long)buf);
5259 mmap_read_unlock(mm);
5262 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5263 void __might_fault(const char *file, int line)
5266 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5267 * holding the mmap_lock, this is safe because kernel memory doesn't
5268 * get paged out, therefore we'll never actually fault, and the
5269 * below annotations will generate false positives.
5271 if (uaccess_kernel())
5273 if (pagefault_disabled())
5275 __might_sleep(file, line);
5276 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5278 might_lock_read(¤t->mm->mmap_lock);
5281 EXPORT_SYMBOL(__might_fault);
5284 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5286 * Process all subpages of the specified huge page with the specified
5287 * operation. The target subpage will be processed last to keep its
5290 static inline void process_huge_page(
5291 unsigned long addr_hint, unsigned int pages_per_huge_page,
5292 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5296 unsigned long addr = addr_hint &
5297 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5299 /* Process target subpage last to keep its cache lines hot */
5301 n = (addr_hint - addr) / PAGE_SIZE;
5302 if (2 * n <= pages_per_huge_page) {
5303 /* If target subpage in first half of huge page */
5306 /* Process subpages at the end of huge page */
5307 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5309 process_subpage(addr + i * PAGE_SIZE, i, arg);
5312 /* If target subpage in second half of huge page */
5313 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5314 l = pages_per_huge_page - n;
5315 /* Process subpages at the begin of huge page */
5316 for (i = 0; i < base; i++) {
5318 process_subpage(addr + i * PAGE_SIZE, i, arg);
5322 * Process remaining subpages in left-right-left-right pattern
5323 * towards the target subpage
5325 for (i = 0; i < l; i++) {
5326 int left_idx = base + i;
5327 int right_idx = base + 2 * l - 1 - i;
5330 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5332 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5336 static void clear_gigantic_page(struct page *page,
5338 unsigned int pages_per_huge_page)
5341 struct page *p = page;
5344 for (i = 0; i < pages_per_huge_page;
5345 i++, p = mem_map_next(p, page, i)) {
5347 clear_user_highpage(p, addr + i * PAGE_SIZE);
5351 static void clear_subpage(unsigned long addr, int idx, void *arg)
5353 struct page *page = arg;
5355 clear_user_highpage(page + idx, addr);
5358 void clear_huge_page(struct page *page,
5359 unsigned long addr_hint, unsigned int pages_per_huge_page)
5361 unsigned long addr = addr_hint &
5362 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5364 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5365 clear_gigantic_page(page, addr, pages_per_huge_page);
5369 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5372 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5374 struct vm_area_struct *vma,
5375 unsigned int pages_per_huge_page)
5378 struct page *dst_base = dst;
5379 struct page *src_base = src;
5381 for (i = 0; i < pages_per_huge_page; ) {
5383 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5386 dst = mem_map_next(dst, dst_base, i);
5387 src = mem_map_next(src, src_base, i);
5391 struct copy_subpage_arg {
5394 struct vm_area_struct *vma;
5397 static void copy_subpage(unsigned long addr, int idx, void *arg)
5399 struct copy_subpage_arg *copy_arg = arg;
5401 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5402 addr, copy_arg->vma);
5405 void copy_user_huge_page(struct page *dst, struct page *src,
5406 unsigned long addr_hint, struct vm_area_struct *vma,
5407 unsigned int pages_per_huge_page)
5409 unsigned long addr = addr_hint &
5410 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5411 struct copy_subpage_arg arg = {
5417 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5418 copy_user_gigantic_page(dst, src, addr, vma,
5419 pages_per_huge_page);
5423 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5426 long copy_huge_page_from_user(struct page *dst_page,
5427 const void __user *usr_src,
5428 unsigned int pages_per_huge_page,
5429 bool allow_pagefault)
5432 unsigned long i, rc = 0;
5433 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5434 struct page *subpage = dst_page;
5436 for (i = 0; i < pages_per_huge_page;
5437 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5438 if (allow_pagefault)
5439 page_kaddr = kmap(subpage);
5441 page_kaddr = kmap_atomic(subpage);
5442 rc = copy_from_user(page_kaddr,
5443 usr_src + i * PAGE_SIZE, PAGE_SIZE);
5444 if (allow_pagefault)
5447 kunmap_atomic(page_kaddr);
5449 ret_val -= (PAGE_SIZE - rc);
5453 flush_dcache_page(subpage);
5459 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5461 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5463 static struct kmem_cache *page_ptl_cachep;
5465 void __init ptlock_cache_init(void)
5467 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5471 bool ptlock_alloc(struct page *page)
5475 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5482 void ptlock_free(struct page *page)
5484 kmem_cache_free(page_ptl_cachep, page->ptl);