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/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
73 #include <linux/perf_event.h>
74 #include <linux/ptrace.h>
75 #include <linux/vmalloc.h>
77 #include <trace/events/kmem.h>
80 #include <asm/mmu_context.h>
81 #include <asm/pgalloc.h>
82 #include <linux/uaccess.h>
84 #include <asm/tlbflush.h>
86 #include "pgalloc-track.h"
89 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
90 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
93 #ifndef CONFIG_NEED_MULTIPLE_NODES
94 /* use the per-pgdat data instead for discontigmem - mbligh */
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 static int __init disable_randmaps(char *s)
139 randomize_va_space = 0;
142 __setup("norandmaps", disable_randmaps);
144 unsigned long zero_pfn __read_mostly;
145 EXPORT_SYMBOL(zero_pfn);
147 unsigned long highest_memmap_pfn __read_mostly;
150 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
152 static int __init init_zero_pfn(void)
154 zero_pfn = page_to_pfn(ZERO_PAGE(0));
157 early_initcall(init_zero_pfn);
159 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
161 trace_rss_stat(mm, member, count);
164 #if defined(SPLIT_RSS_COUNTING)
166 void sync_mm_rss(struct mm_struct *mm)
170 for (i = 0; i < NR_MM_COUNTERS; i++) {
171 if (current->rss_stat.count[i]) {
172 add_mm_counter(mm, i, current->rss_stat.count[i]);
173 current->rss_stat.count[i] = 0;
176 current->rss_stat.events = 0;
179 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
181 struct task_struct *task = current;
183 if (likely(task->mm == mm))
184 task->rss_stat.count[member] += val;
186 add_mm_counter(mm, member, val);
188 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
189 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
191 /* sync counter once per 64 page faults */
192 #define TASK_RSS_EVENTS_THRESH (64)
193 static void check_sync_rss_stat(struct task_struct *task)
195 if (unlikely(task != current))
197 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
198 sync_mm_rss(task->mm);
200 #else /* SPLIT_RSS_COUNTING */
202 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
203 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
205 static void check_sync_rss_stat(struct task_struct *task)
209 #endif /* SPLIT_RSS_COUNTING */
212 * Note: this doesn't free the actual pages themselves. That
213 * has been handled earlier when unmapping all the memory regions.
215 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
218 pgtable_t token = pmd_pgtable(*pmd);
220 pte_free_tlb(tlb, token, addr);
221 mm_dec_nr_ptes(tlb->mm);
224 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
225 unsigned long addr, unsigned long end,
226 unsigned long floor, unsigned long ceiling)
233 pmd = pmd_offset(pud, addr);
235 next = pmd_addr_end(addr, end);
236 if (pmd_none_or_clear_bad(pmd))
238 free_pte_range(tlb, pmd, addr);
239 } while (pmd++, addr = next, addr != end);
249 if (end - 1 > ceiling - 1)
252 pmd = pmd_offset(pud, start);
254 pmd_free_tlb(tlb, pmd, start);
255 mm_dec_nr_pmds(tlb->mm);
258 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
259 unsigned long addr, unsigned long end,
260 unsigned long floor, unsigned long ceiling)
267 pud = pud_offset(p4d, addr);
269 next = pud_addr_end(addr, end);
270 if (pud_none_or_clear_bad(pud))
272 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
273 } while (pud++, addr = next, addr != end);
283 if (end - 1 > ceiling - 1)
286 pud = pud_offset(p4d, start);
288 pud_free_tlb(tlb, pud, start);
289 mm_dec_nr_puds(tlb->mm);
292 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
293 unsigned long addr, unsigned long end,
294 unsigned long floor, unsigned long ceiling)
301 p4d = p4d_offset(pgd, addr);
303 next = p4d_addr_end(addr, end);
304 if (p4d_none_or_clear_bad(p4d))
306 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
307 } while (p4d++, addr = next, addr != end);
313 ceiling &= PGDIR_MASK;
317 if (end - 1 > ceiling - 1)
320 p4d = p4d_offset(pgd, start);
322 p4d_free_tlb(tlb, p4d, start);
326 * This function frees user-level page tables of a process.
328 void free_pgd_range(struct mmu_gather *tlb,
329 unsigned long addr, unsigned long end,
330 unsigned long floor, unsigned long ceiling)
336 * The next few lines have given us lots of grief...
338 * Why are we testing PMD* at this top level? Because often
339 * there will be no work to do at all, and we'd prefer not to
340 * go all the way down to the bottom just to discover that.
342 * Why all these "- 1"s? Because 0 represents both the bottom
343 * of the address space and the top of it (using -1 for the
344 * top wouldn't help much: the masks would do the wrong thing).
345 * The rule is that addr 0 and floor 0 refer to the bottom of
346 * the address space, but end 0 and ceiling 0 refer to the top
347 * Comparisons need to use "end - 1" and "ceiling - 1" (though
348 * that end 0 case should be mythical).
350 * Wherever addr is brought up or ceiling brought down, we must
351 * be careful to reject "the opposite 0" before it confuses the
352 * subsequent tests. But what about where end is brought down
353 * by PMD_SIZE below? no, end can't go down to 0 there.
355 * Whereas we round start (addr) and ceiling down, by different
356 * masks at different levels, in order to test whether a table
357 * now has no other vmas using it, so can be freed, we don't
358 * bother to round floor or end up - the tests don't need that.
372 if (end - 1 > ceiling - 1)
377 * We add page table cache pages with PAGE_SIZE,
378 * (see pte_free_tlb()), flush the tlb if we need
380 tlb_change_page_size(tlb, PAGE_SIZE);
381 pgd = pgd_offset(tlb->mm, addr);
383 next = pgd_addr_end(addr, end);
384 if (pgd_none_or_clear_bad(pgd))
386 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
387 } while (pgd++, addr = next, addr != end);
390 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
391 unsigned long floor, unsigned long ceiling)
394 struct vm_area_struct *next = vma->vm_next;
395 unsigned long addr = vma->vm_start;
398 * Hide vma from rmap and truncate_pagecache before freeing
401 unlink_anon_vmas(vma);
402 unlink_file_vma(vma);
404 if (is_vm_hugetlb_page(vma)) {
405 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
406 floor, next ? next->vm_start : ceiling);
409 * Optimization: gather nearby vmas into one call down
411 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
412 && !is_vm_hugetlb_page(next)) {
415 unlink_anon_vmas(vma);
416 unlink_file_vma(vma);
418 free_pgd_range(tlb, addr, vma->vm_end,
419 floor, next ? next->vm_start : ceiling);
425 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
428 pgtable_t new = pte_alloc_one(mm);
433 * Ensure all pte setup (eg. pte page lock and page clearing) are
434 * visible before the pte is made visible to other CPUs by being
435 * put into page tables.
437 * The other side of the story is the pointer chasing in the page
438 * table walking code (when walking the page table without locking;
439 * ie. most of the time). Fortunately, these data accesses consist
440 * of a chain of data-dependent loads, meaning most CPUs (alpha
441 * being the notable exception) will already guarantee loads are
442 * seen in-order. See the alpha page table accessors for the
443 * smp_rmb() barriers in page table walking code.
445 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
447 ptl = pmd_lock(mm, pmd);
448 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
450 pmd_populate(mm, pmd, new);
459 int __pte_alloc_kernel(pmd_t *pmd)
461 pte_t *new = pte_alloc_one_kernel(&init_mm);
465 smp_wmb(); /* See comment in __pte_alloc */
467 spin_lock(&init_mm.page_table_lock);
468 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
469 pmd_populate_kernel(&init_mm, pmd, new);
472 spin_unlock(&init_mm.page_table_lock);
474 pte_free_kernel(&init_mm, new);
478 static inline void init_rss_vec(int *rss)
480 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
483 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
487 if (current->mm == mm)
489 for (i = 0; i < NR_MM_COUNTERS; i++)
491 add_mm_counter(mm, i, rss[i]);
495 * This function is called to print an error when a bad pte
496 * is found. For example, we might have a PFN-mapped pte in
497 * a region that doesn't allow it.
499 * The calling function must still handle the error.
501 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
502 pte_t pte, struct page *page)
504 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
505 p4d_t *p4d = p4d_offset(pgd, addr);
506 pud_t *pud = pud_offset(p4d, addr);
507 pmd_t *pmd = pmd_offset(pud, addr);
508 struct address_space *mapping;
510 static unsigned long resume;
511 static unsigned long nr_shown;
512 static unsigned long nr_unshown;
515 * Allow a burst of 60 reports, then keep quiet for that minute;
516 * or allow a steady drip of one report per second.
518 if (nr_shown == 60) {
519 if (time_before(jiffies, resume)) {
524 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
531 resume = jiffies + 60 * HZ;
533 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
534 index = linear_page_index(vma, addr);
536 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
538 (long long)pte_val(pte), (long long)pmd_val(*pmd));
540 dump_page(page, "bad pte");
541 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
542 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
543 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
545 vma->vm_ops ? vma->vm_ops->fault : NULL,
546 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
547 mapping ? mapping->a_ops->readpage : NULL);
549 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
553 * vm_normal_page -- This function gets the "struct page" associated with a pte.
555 * "Special" mappings do not wish to be associated with a "struct page" (either
556 * it doesn't exist, or it exists but they don't want to touch it). In this
557 * case, NULL is returned here. "Normal" mappings do have a struct page.
559 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
560 * pte bit, in which case this function is trivial. Secondly, an architecture
561 * may not have a spare pte bit, which requires a more complicated scheme,
564 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
565 * special mapping (even if there are underlying and valid "struct pages").
566 * COWed pages of a VM_PFNMAP are always normal.
568 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
569 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
570 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
571 * mapping will always honor the rule
573 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
575 * And for normal mappings this is false.
577 * This restricts such mappings to be a linear translation from virtual address
578 * to pfn. To get around this restriction, we allow arbitrary mappings so long
579 * as the vma is not a COW mapping; in that case, we know that all ptes are
580 * special (because none can have been COWed).
583 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
585 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
586 * page" backing, however the difference is that _all_ pages with a struct
587 * page (that is, those where pfn_valid is true) are refcounted and considered
588 * normal pages by the VM. The disadvantage is that pages are refcounted
589 * (which can be slower and simply not an option for some PFNMAP users). The
590 * advantage is that we don't have to follow the strict linearity rule of
591 * PFNMAP mappings in order to support COWable mappings.
594 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
597 unsigned long pfn = pte_pfn(pte);
599 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
600 if (likely(!pte_special(pte)))
602 if (vma->vm_ops && vma->vm_ops->find_special_page)
603 return vma->vm_ops->find_special_page(vma, addr);
604 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
606 if (is_zero_pfn(pfn))
611 print_bad_pte(vma, addr, pte, NULL);
615 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
617 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
618 if (vma->vm_flags & VM_MIXEDMAP) {
624 off = (addr - vma->vm_start) >> PAGE_SHIFT;
625 if (pfn == vma->vm_pgoff + off)
627 if (!is_cow_mapping(vma->vm_flags))
632 if (is_zero_pfn(pfn))
636 if (unlikely(pfn > highest_memmap_pfn)) {
637 print_bad_pte(vma, addr, pte, NULL);
642 * NOTE! We still have PageReserved() pages in the page tables.
643 * eg. VDSO mappings can cause them to exist.
646 return pfn_to_page(pfn);
649 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
650 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
653 unsigned long pfn = pmd_pfn(pmd);
656 * There is no pmd_special() but there may be special pmds, e.g.
657 * in a direct-access (dax) mapping, so let's just replicate the
658 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
660 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
661 if (vma->vm_flags & VM_MIXEDMAP) {
667 off = (addr - vma->vm_start) >> PAGE_SHIFT;
668 if (pfn == vma->vm_pgoff + off)
670 if (!is_cow_mapping(vma->vm_flags))
677 if (is_huge_zero_pmd(pmd))
679 if (unlikely(pfn > highest_memmap_pfn))
683 * NOTE! We still have PageReserved() pages in the page tables.
684 * eg. VDSO mappings can cause them to exist.
687 return pfn_to_page(pfn);
692 * copy one vm_area from one task to the other. Assumes the page tables
693 * already present in the new task to be cleared in the whole range
694 * covered by this vma.
698 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
699 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
700 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
702 unsigned long vm_flags = dst_vma->vm_flags;
703 pte_t pte = *src_pte;
705 swp_entry_t entry = pte_to_swp_entry(pte);
707 if (likely(!non_swap_entry(entry))) {
708 if (swap_duplicate(entry) < 0)
711 /* make sure dst_mm is on swapoff's mmlist. */
712 if (unlikely(list_empty(&dst_mm->mmlist))) {
713 spin_lock(&mmlist_lock);
714 if (list_empty(&dst_mm->mmlist))
715 list_add(&dst_mm->mmlist,
717 spin_unlock(&mmlist_lock);
720 } else if (is_migration_entry(entry)) {
721 page = migration_entry_to_page(entry);
723 rss[mm_counter(page)]++;
725 if (is_write_migration_entry(entry) &&
726 is_cow_mapping(vm_flags)) {
728 * COW mappings require pages in both
729 * parent and child to be set to read.
731 make_migration_entry_read(&entry);
732 pte = swp_entry_to_pte(entry);
733 if (pte_swp_soft_dirty(*src_pte))
734 pte = pte_swp_mksoft_dirty(pte);
735 if (pte_swp_uffd_wp(*src_pte))
736 pte = pte_swp_mkuffd_wp(pte);
737 set_pte_at(src_mm, addr, src_pte, pte);
739 } else if (is_device_private_entry(entry)) {
740 page = device_private_entry_to_page(entry);
743 * Update rss count even for unaddressable pages, as
744 * they should treated just like normal pages in this
747 * We will likely want to have some new rss counters
748 * for unaddressable pages, at some point. But for now
749 * keep things as they are.
752 rss[mm_counter(page)]++;
753 page_dup_rmap(page, false);
756 * We do not preserve soft-dirty information, because so
757 * far, checkpoint/restore is the only feature that
758 * requires that. And checkpoint/restore does not work
759 * when a device driver is involved (you cannot easily
760 * save and restore device driver state).
762 if (is_write_device_private_entry(entry) &&
763 is_cow_mapping(vm_flags)) {
764 make_device_private_entry_read(&entry);
765 pte = swp_entry_to_pte(entry);
766 if (pte_swp_uffd_wp(*src_pte))
767 pte = pte_swp_mkuffd_wp(pte);
768 set_pte_at(src_mm, addr, src_pte, pte);
771 if (!userfaultfd_wp(dst_vma))
772 pte = pte_swp_clear_uffd_wp(pte);
773 set_pte_at(dst_mm, addr, dst_pte, pte);
778 * Copy a present and normal page if necessary.
780 * NOTE! The usual case is that this doesn't need to do
781 * anything, and can just return a positive value. That
782 * will let the caller know that it can just increase
783 * the page refcount and re-use the pte the traditional
786 * But _if_ we need to copy it because it needs to be
787 * pinned in the parent (and the child should get its own
788 * copy rather than just a reference to the same page),
789 * we'll do that here and return zero to let the caller
792 * And if we need a pre-allocated page but don't yet have
793 * one, return a negative error to let the preallocation
794 * code know so that it can do so outside the page table
798 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
799 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
800 struct page **prealloc, pte_t pte, struct page *page)
802 struct mm_struct *src_mm = src_vma->vm_mm;
803 struct page *new_page;
805 if (!is_cow_mapping(src_vma->vm_flags))
809 * What we want to do is to check whether this page may
810 * have been pinned by the parent process. If so,
811 * instead of wrprotect the pte on both sides, we copy
812 * the page immediately so that we'll always guarantee
813 * the pinned page won't be randomly replaced in the
816 * The page pinning checks are just "has this mm ever
817 * seen pinning", along with the (inexact) check of
818 * the page count. That might give false positives for
819 * for pinning, but it will work correctly.
821 if (likely(!atomic_read(&src_mm->has_pinned)))
823 if (likely(!page_maybe_dma_pinned(page)))
827 * The vma->anon_vma of the child process may be NULL
828 * because the entire vma does not contain anonymous pages.
829 * A BUG will occur when the copy_present_page() passes
830 * a copy of a non-anonymous page of that vma to the
831 * page_add_new_anon_rmap() to set up new anonymous rmap.
832 * Return 1 if the page is not an anonymous page.
837 new_page = *prealloc;
842 * We have a prealloc page, all good! Take it
843 * over and copy the page & arm it.
846 copy_user_highpage(new_page, page, addr, src_vma);
847 __SetPageUptodate(new_page);
848 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
849 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
850 rss[mm_counter(new_page)]++;
852 /* All done, just insert the new page copy in the child */
853 pte = mk_pte(new_page, dst_vma->vm_page_prot);
854 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
855 if (userfaultfd_pte_wp(dst_vma, *src_pte))
856 /* Uffd-wp needs to be delivered to dest pte as well */
857 pte = pte_wrprotect(pte_mkuffd_wp(pte));
858 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
863 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
864 * is required to copy this pte.
867 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
868 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
869 struct page **prealloc)
871 struct mm_struct *src_mm = src_vma->vm_mm;
872 unsigned long vm_flags = src_vma->vm_flags;
873 pte_t pte = *src_pte;
876 page = vm_normal_page(src_vma, addr, pte);
880 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
881 addr, rss, prealloc, pte, page);
886 page_dup_rmap(page, false);
887 rss[mm_counter(page)]++;
891 * If it's a COW mapping, write protect it both
892 * in the parent and the child
894 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
895 ptep_set_wrprotect(src_mm, addr, src_pte);
896 pte = pte_wrprotect(pte);
900 * If it's a shared mapping, mark it clean in
903 if (vm_flags & VM_SHARED)
904 pte = pte_mkclean(pte);
905 pte = pte_mkold(pte);
907 if (!userfaultfd_wp(dst_vma))
908 pte = pte_clear_uffd_wp(pte);
910 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
914 static inline struct page *
915 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
918 struct page *new_page;
920 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
924 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
928 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
934 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
935 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
938 struct mm_struct *dst_mm = dst_vma->vm_mm;
939 struct mm_struct *src_mm = src_vma->vm_mm;
940 pte_t *orig_src_pte, *orig_dst_pte;
941 pte_t *src_pte, *dst_pte;
942 spinlock_t *src_ptl, *dst_ptl;
943 int progress, ret = 0;
944 int rss[NR_MM_COUNTERS];
945 swp_entry_t entry = (swp_entry_t){0};
946 struct page *prealloc = NULL;
952 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
957 src_pte = pte_offset_map(src_pmd, addr);
958 src_ptl = pte_lockptr(src_mm, src_pmd);
959 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
960 orig_src_pte = src_pte;
961 orig_dst_pte = dst_pte;
962 arch_enter_lazy_mmu_mode();
966 * We are holding two locks at this point - either of them
967 * could generate latencies in another task on another CPU.
969 if (progress >= 32) {
971 if (need_resched() ||
972 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
975 if (pte_none(*src_pte)) {
979 if (unlikely(!pte_present(*src_pte))) {
980 entry.val = copy_nonpresent_pte(dst_mm, src_mm,
989 /* copy_present_pte() will clear `*prealloc' if consumed */
990 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
991 addr, rss, &prealloc);
993 * If we need a pre-allocated page for this pte, drop the
994 * locks, allocate, and try again.
996 if (unlikely(ret == -EAGAIN))
998 if (unlikely(prealloc)) {
1000 * pre-alloc page cannot be reused by next time so as
1001 * to strictly follow mempolicy (e.g., alloc_page_vma()
1002 * will allocate page according to address). This
1003 * could only happen if one pinned pte changed.
1009 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1011 arch_leave_lazy_mmu_mode();
1012 spin_unlock(src_ptl);
1013 pte_unmap(orig_src_pte);
1014 add_mm_rss_vec(dst_mm, rss);
1015 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1019 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1025 WARN_ON_ONCE(ret != -EAGAIN);
1026 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1029 /* We've captured and resolved the error. Reset, try again. */
1035 if (unlikely(prealloc))
1041 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1042 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1045 struct mm_struct *dst_mm = dst_vma->vm_mm;
1046 struct mm_struct *src_mm = src_vma->vm_mm;
1047 pmd_t *src_pmd, *dst_pmd;
1050 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1053 src_pmd = pmd_offset(src_pud, addr);
1055 next = pmd_addr_end(addr, end);
1056 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1057 || pmd_devmap(*src_pmd)) {
1059 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1060 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1061 addr, dst_vma, src_vma);
1068 if (pmd_none_or_clear_bad(src_pmd))
1070 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1073 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1078 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1079 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1082 struct mm_struct *dst_mm = dst_vma->vm_mm;
1083 struct mm_struct *src_mm = src_vma->vm_mm;
1084 pud_t *src_pud, *dst_pud;
1087 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1090 src_pud = pud_offset(src_p4d, addr);
1092 next = pud_addr_end(addr, end);
1093 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1096 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1097 err = copy_huge_pud(dst_mm, src_mm,
1098 dst_pud, src_pud, addr, src_vma);
1105 if (pud_none_or_clear_bad(src_pud))
1107 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1110 } while (dst_pud++, src_pud++, addr = next, addr != end);
1115 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1116 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1119 struct mm_struct *dst_mm = dst_vma->vm_mm;
1120 p4d_t *src_p4d, *dst_p4d;
1123 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1126 src_p4d = p4d_offset(src_pgd, addr);
1128 next = p4d_addr_end(addr, end);
1129 if (p4d_none_or_clear_bad(src_p4d))
1131 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1134 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1139 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1141 pgd_t *src_pgd, *dst_pgd;
1143 unsigned long addr = src_vma->vm_start;
1144 unsigned long end = src_vma->vm_end;
1145 struct mm_struct *dst_mm = dst_vma->vm_mm;
1146 struct mm_struct *src_mm = src_vma->vm_mm;
1147 struct mmu_notifier_range range;
1152 * Don't copy ptes where a page fault will fill them correctly.
1153 * Fork becomes much lighter when there are big shared or private
1154 * readonly mappings. The tradeoff is that copy_page_range is more
1155 * efficient than faulting.
1157 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1161 if (is_vm_hugetlb_page(src_vma))
1162 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1164 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1166 * We do not free on error cases below as remove_vma
1167 * gets called on error from higher level routine
1169 ret = track_pfn_copy(src_vma);
1175 * We need to invalidate the secondary MMU mappings only when
1176 * there could be a permission downgrade on the ptes of the
1177 * parent mm. And a permission downgrade will only happen if
1178 * is_cow_mapping() returns true.
1180 is_cow = is_cow_mapping(src_vma->vm_flags);
1183 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1184 0, src_vma, src_mm, addr, end);
1185 mmu_notifier_invalidate_range_start(&range);
1187 * Disabling preemption is not needed for the write side, as
1188 * the read side doesn't spin, but goes to the mmap_lock.
1190 * Use the raw variant of the seqcount_t write API to avoid
1191 * lockdep complaining about preemptibility.
1193 mmap_assert_write_locked(src_mm);
1194 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1198 dst_pgd = pgd_offset(dst_mm, addr);
1199 src_pgd = pgd_offset(src_mm, addr);
1201 next = pgd_addr_end(addr, end);
1202 if (pgd_none_or_clear_bad(src_pgd))
1204 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1209 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1212 raw_write_seqcount_end(&src_mm->write_protect_seq);
1213 mmu_notifier_invalidate_range_end(&range);
1218 /* Whether we should zap all COWed (private) pages too */
1219 static inline bool should_zap_cows(struct zap_details *details)
1221 /* By default, zap all pages */
1225 /* Or, we zap COWed pages only if the caller wants to */
1226 return !details->check_mapping;
1229 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1230 struct vm_area_struct *vma, pmd_t *pmd,
1231 unsigned long addr, unsigned long end,
1232 struct zap_details *details)
1234 struct mm_struct *mm = tlb->mm;
1235 int force_flush = 0;
1236 int rss[NR_MM_COUNTERS];
1242 tlb_change_page_size(tlb, PAGE_SIZE);
1245 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1247 flush_tlb_batched_pending(mm);
1248 arch_enter_lazy_mmu_mode();
1251 if (pte_none(ptent))
1257 if (pte_present(ptent)) {
1260 page = vm_normal_page(vma, addr, ptent);
1261 if (unlikely(details) && page) {
1263 * unmap_shared_mapping_pages() wants to
1264 * invalidate cache without truncating:
1265 * unmap shared but keep private pages.
1267 if (details->check_mapping &&
1268 details->check_mapping != page_rmapping(page))
1271 ptent = ptep_get_and_clear_full(mm, addr, pte,
1273 tlb_remove_tlb_entry(tlb, pte, addr);
1274 if (unlikely(!page))
1277 if (!PageAnon(page)) {
1278 if (pte_dirty(ptent)) {
1280 set_page_dirty(page);
1282 if (pte_young(ptent) &&
1283 likely(!(vma->vm_flags & VM_SEQ_READ)))
1284 mark_page_accessed(page);
1286 rss[mm_counter(page)]--;
1287 page_remove_rmap(page, false);
1288 if (unlikely(page_mapcount(page) < 0))
1289 print_bad_pte(vma, addr, ptent, page);
1290 if (unlikely(__tlb_remove_page(tlb, page))) {
1298 entry = pte_to_swp_entry(ptent);
1299 if (is_device_private_entry(entry)) {
1300 struct page *page = device_private_entry_to_page(entry);
1302 if (unlikely(details && details->check_mapping)) {
1304 * unmap_shared_mapping_pages() wants to
1305 * invalidate cache without truncating:
1306 * unmap shared but keep private pages.
1308 if (details->check_mapping !=
1309 page_rmapping(page))
1313 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1314 rss[mm_counter(page)]--;
1315 page_remove_rmap(page, false);
1320 if (!non_swap_entry(entry)) {
1321 /* Genuine swap entry, hence a private anon page */
1322 if (!should_zap_cows(details))
1325 } else if (is_migration_entry(entry)) {
1328 page = migration_entry_to_page(entry);
1329 if (details && details->check_mapping &&
1330 details->check_mapping != page_rmapping(page))
1332 rss[mm_counter(page)]--;
1334 if (unlikely(!free_swap_and_cache(entry)))
1335 print_bad_pte(vma, addr, ptent, NULL);
1336 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1337 } while (pte++, addr += PAGE_SIZE, addr != end);
1339 add_mm_rss_vec(mm, rss);
1340 arch_leave_lazy_mmu_mode();
1342 /* Do the actual TLB flush before dropping ptl */
1344 tlb_flush_mmu_tlbonly(tlb);
1345 pte_unmap_unlock(start_pte, ptl);
1348 * If we forced a TLB flush (either due to running out of
1349 * batch buffers or because we needed to flush dirty TLB
1350 * entries before releasing the ptl), free the batched
1351 * memory too. Restart if we didn't do everything.
1366 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1367 struct vm_area_struct *vma, pud_t *pud,
1368 unsigned long addr, unsigned long end,
1369 struct zap_details *details)
1374 pmd = pmd_offset(pud, addr);
1376 next = pmd_addr_end(addr, end);
1377 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1378 if (next - addr != HPAGE_PMD_SIZE)
1379 __split_huge_pmd(vma, pmd, addr, false, NULL);
1380 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1383 } else if (details && details->single_page &&
1384 PageTransCompound(details->single_page) &&
1385 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1386 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1388 * Take and drop THP pmd lock so that we cannot return
1389 * prematurely, while zap_huge_pmd() has cleared *pmd,
1390 * but not yet decremented compound_mapcount().
1396 * Here there can be other concurrent MADV_DONTNEED or
1397 * trans huge page faults running, and if the pmd is
1398 * none or trans huge it can change under us. This is
1399 * because MADV_DONTNEED holds the mmap_lock in read
1402 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1404 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1407 } while (pmd++, addr = next, addr != end);
1412 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1413 struct vm_area_struct *vma, p4d_t *p4d,
1414 unsigned long addr, unsigned long end,
1415 struct zap_details *details)
1420 pud = pud_offset(p4d, addr);
1422 next = pud_addr_end(addr, end);
1423 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1424 if (next - addr != HPAGE_PUD_SIZE) {
1425 mmap_assert_locked(tlb->mm);
1426 split_huge_pud(vma, pud, addr);
1427 } else if (zap_huge_pud(tlb, vma, pud, addr))
1431 if (pud_none_or_clear_bad(pud))
1433 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1436 } while (pud++, addr = next, addr != end);
1441 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1442 struct vm_area_struct *vma, pgd_t *pgd,
1443 unsigned long addr, unsigned long end,
1444 struct zap_details *details)
1449 p4d = p4d_offset(pgd, addr);
1451 next = p4d_addr_end(addr, end);
1452 if (p4d_none_or_clear_bad(p4d))
1454 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1455 } while (p4d++, addr = next, addr != end);
1460 void unmap_page_range(struct mmu_gather *tlb,
1461 struct vm_area_struct *vma,
1462 unsigned long addr, unsigned long end,
1463 struct zap_details *details)
1468 BUG_ON(addr >= end);
1469 tlb_start_vma(tlb, vma);
1470 pgd = pgd_offset(vma->vm_mm, addr);
1472 next = pgd_addr_end(addr, end);
1473 if (pgd_none_or_clear_bad(pgd))
1475 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1476 } while (pgd++, addr = next, addr != end);
1477 tlb_end_vma(tlb, vma);
1481 static void unmap_single_vma(struct mmu_gather *tlb,
1482 struct vm_area_struct *vma, unsigned long start_addr,
1483 unsigned long end_addr,
1484 struct zap_details *details)
1486 unsigned long start = max(vma->vm_start, start_addr);
1489 if (start >= vma->vm_end)
1491 end = min(vma->vm_end, end_addr);
1492 if (end <= vma->vm_start)
1496 uprobe_munmap(vma, start, end);
1498 if (unlikely(vma->vm_flags & VM_PFNMAP))
1499 untrack_pfn(vma, 0, 0);
1502 if (unlikely(is_vm_hugetlb_page(vma))) {
1504 * It is undesirable to test vma->vm_file as it
1505 * should be non-null for valid hugetlb area.
1506 * However, vm_file will be NULL in the error
1507 * cleanup path of mmap_region. When
1508 * hugetlbfs ->mmap method fails,
1509 * mmap_region() nullifies vma->vm_file
1510 * before calling this function to clean up.
1511 * Since no pte has actually been setup, it is
1512 * safe to do nothing in this case.
1515 i_mmap_lock_write(vma->vm_file->f_mapping);
1516 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1517 i_mmap_unlock_write(vma->vm_file->f_mapping);
1520 unmap_page_range(tlb, vma, start, end, details);
1525 * unmap_vmas - unmap a range of memory covered by a list of vma's
1526 * @tlb: address of the caller's struct mmu_gather
1527 * @vma: the starting vma
1528 * @start_addr: virtual address at which to start unmapping
1529 * @end_addr: virtual address at which to end unmapping
1531 * Unmap all pages in the vma list.
1533 * Only addresses between `start' and `end' will be unmapped.
1535 * The VMA list must be sorted in ascending virtual address order.
1537 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1538 * range after unmap_vmas() returns. So the only responsibility here is to
1539 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1540 * drops the lock and schedules.
1542 void unmap_vmas(struct mmu_gather *tlb,
1543 struct vm_area_struct *vma, unsigned long start_addr,
1544 unsigned long end_addr)
1546 struct mmu_notifier_range range;
1548 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1549 start_addr, end_addr);
1550 mmu_notifier_invalidate_range_start(&range);
1551 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1552 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1553 mmu_notifier_invalidate_range_end(&range);
1557 * zap_page_range - remove user pages in a given range
1558 * @vma: vm_area_struct holding the applicable pages
1559 * @start: starting address of pages to zap
1560 * @size: number of bytes to zap
1562 * Caller must protect the VMA list
1564 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1567 struct mmu_notifier_range range;
1568 struct mmu_gather tlb;
1571 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1572 start, start + size);
1573 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1574 update_hiwater_rss(vma->vm_mm);
1575 mmu_notifier_invalidate_range_start(&range);
1576 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1577 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1578 mmu_notifier_invalidate_range_end(&range);
1579 tlb_finish_mmu(&tlb, start, range.end);
1583 * zap_page_range_single - remove user pages in a given range
1584 * @vma: vm_area_struct holding the applicable pages
1585 * @address: starting address of pages to zap
1586 * @size: number of bytes to zap
1587 * @details: details of shared cache invalidation
1589 * The range must fit into one VMA.
1591 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1592 unsigned long size, struct zap_details *details)
1594 struct mmu_notifier_range range;
1595 struct mmu_gather tlb;
1598 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1599 address, address + size);
1600 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1601 update_hiwater_rss(vma->vm_mm);
1602 mmu_notifier_invalidate_range_start(&range);
1603 unmap_single_vma(&tlb, vma, address, range.end, details);
1604 mmu_notifier_invalidate_range_end(&range);
1605 tlb_finish_mmu(&tlb, address, range.end);
1609 * zap_vma_ptes - remove ptes mapping the vma
1610 * @vma: vm_area_struct holding ptes to be zapped
1611 * @address: starting address of pages to zap
1612 * @size: number of bytes to zap
1614 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1616 * The entire address range must be fully contained within the vma.
1619 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1622 if (address < vma->vm_start || address + size > vma->vm_end ||
1623 !(vma->vm_flags & VM_PFNMAP))
1626 zap_page_range_single(vma, address, size, NULL);
1628 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1630 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1637 pgd = pgd_offset(mm, addr);
1638 p4d = p4d_alloc(mm, pgd, addr);
1641 pud = pud_alloc(mm, p4d, addr);
1644 pmd = pmd_alloc(mm, pud, addr);
1648 VM_BUG_ON(pmd_trans_huge(*pmd));
1652 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1655 pmd_t *pmd = walk_to_pmd(mm, addr);
1659 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1662 static int validate_page_before_insert(struct page *page)
1664 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1666 flush_dcache_page(page);
1670 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1671 unsigned long addr, struct page *page, pgprot_t prot)
1673 if (!pte_none(*pte))
1675 /* Ok, finally just insert the thing.. */
1677 inc_mm_counter_fast(mm, mm_counter_file(page));
1678 page_add_file_rmap(page, false);
1679 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1684 * This is the old fallback for page remapping.
1686 * For historical reasons, it only allows reserved pages. Only
1687 * old drivers should use this, and they needed to mark their
1688 * pages reserved for the old functions anyway.
1690 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1691 struct page *page, pgprot_t prot)
1693 struct mm_struct *mm = vma->vm_mm;
1698 retval = validate_page_before_insert(page);
1702 pte = get_locked_pte(mm, addr, &ptl);
1705 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1706 pte_unmap_unlock(pte, ptl);
1712 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1713 unsigned long addr, struct page *page, pgprot_t prot)
1717 if (!page_count(page))
1719 err = validate_page_before_insert(page);
1722 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1725 /* insert_pages() amortizes the cost of spinlock operations
1726 * when inserting pages in a loop. Arch *must* define pte_index.
1728 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1729 struct page **pages, unsigned long *num, pgprot_t prot)
1732 pte_t *start_pte, *pte;
1733 spinlock_t *pte_lock;
1734 struct mm_struct *const mm = vma->vm_mm;
1735 unsigned long curr_page_idx = 0;
1736 unsigned long remaining_pages_total = *num;
1737 unsigned long pages_to_write_in_pmd;
1741 pmd = walk_to_pmd(mm, addr);
1745 pages_to_write_in_pmd = min_t(unsigned long,
1746 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1748 /* Allocate the PTE if necessary; takes PMD lock once only. */
1750 if (pte_alloc(mm, pmd))
1753 while (pages_to_write_in_pmd) {
1755 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1757 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1758 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1759 int err = insert_page_in_batch_locked(mm, pte,
1760 addr, pages[curr_page_idx], prot);
1761 if (unlikely(err)) {
1762 pte_unmap_unlock(start_pte, pte_lock);
1764 remaining_pages_total -= pte_idx;
1770 pte_unmap_unlock(start_pte, pte_lock);
1771 pages_to_write_in_pmd -= batch_size;
1772 remaining_pages_total -= batch_size;
1774 if (remaining_pages_total)
1778 *num = remaining_pages_total;
1781 #endif /* ifdef pte_index */
1784 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1785 * @vma: user vma to map to
1786 * @addr: target start user address of these pages
1787 * @pages: source kernel pages
1788 * @num: in: number of pages to map. out: number of pages that were *not*
1789 * mapped. (0 means all pages were successfully mapped).
1791 * Preferred over vm_insert_page() when inserting multiple pages.
1793 * In case of error, we may have mapped a subset of the provided
1794 * pages. It is the caller's responsibility to account for this case.
1796 * The same restrictions apply as in vm_insert_page().
1798 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1799 struct page **pages, unsigned long *num)
1802 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1804 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1806 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1807 BUG_ON(mmap_read_trylock(vma->vm_mm));
1808 BUG_ON(vma->vm_flags & VM_PFNMAP);
1809 vma->vm_flags |= VM_MIXEDMAP;
1811 /* Defer page refcount checking till we're about to map that page. */
1812 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1814 unsigned long idx = 0, pgcount = *num;
1817 for (; idx < pgcount; ++idx) {
1818 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1822 *num = pgcount - idx;
1824 #endif /* ifdef pte_index */
1826 EXPORT_SYMBOL(vm_insert_pages);
1829 * vm_insert_page - insert single page into user vma
1830 * @vma: user vma to map to
1831 * @addr: target user address of this page
1832 * @page: source kernel page
1834 * This allows drivers to insert individual pages they've allocated
1837 * The page has to be a nice clean _individual_ kernel allocation.
1838 * If you allocate a compound page, you need to have marked it as
1839 * such (__GFP_COMP), or manually just split the page up yourself
1840 * (see split_page()).
1842 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1843 * took an arbitrary page protection parameter. This doesn't allow
1844 * that. Your vma protection will have to be set up correctly, which
1845 * means that if you want a shared writable mapping, you'd better
1846 * ask for a shared writable mapping!
1848 * The page does not need to be reserved.
1850 * Usually this function is called from f_op->mmap() handler
1851 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1852 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1853 * function from other places, for example from page-fault handler.
1855 * Return: %0 on success, negative error code otherwise.
1857 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1860 if (addr < vma->vm_start || addr >= vma->vm_end)
1862 if (!page_count(page))
1864 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1865 BUG_ON(mmap_read_trylock(vma->vm_mm));
1866 BUG_ON(vma->vm_flags & VM_PFNMAP);
1867 vma->vm_flags |= VM_MIXEDMAP;
1869 return insert_page(vma, addr, page, vma->vm_page_prot);
1871 EXPORT_SYMBOL(vm_insert_page);
1874 * __vm_map_pages - maps range of kernel pages into user vma
1875 * @vma: user vma to map to
1876 * @pages: pointer to array of source kernel pages
1877 * @num: number of pages in page array
1878 * @offset: user's requested vm_pgoff
1880 * This allows drivers to map range of kernel pages into a user vma.
1882 * Return: 0 on success and error code otherwise.
1884 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1885 unsigned long num, unsigned long offset)
1887 unsigned long count = vma_pages(vma);
1888 unsigned long uaddr = vma->vm_start;
1891 /* Fail if the user requested offset is beyond the end of the object */
1895 /* Fail if the user requested size exceeds available object size */
1896 if (count > num - offset)
1899 for (i = 0; i < count; i++) {
1900 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1910 * vm_map_pages - maps range of kernel pages starts with non zero offset
1911 * @vma: user vma to map to
1912 * @pages: pointer to array of source kernel pages
1913 * @num: number of pages in page array
1915 * Maps an object consisting of @num pages, catering for the user's
1916 * requested vm_pgoff
1918 * If we fail to insert any page into the vma, the function will return
1919 * immediately leaving any previously inserted pages present. Callers
1920 * from the mmap handler may immediately return the error as their caller
1921 * will destroy the vma, removing any successfully inserted pages. Other
1922 * callers should make their own arrangements for calling unmap_region().
1924 * Context: Process context. Called by mmap handlers.
1925 * Return: 0 on success and error code otherwise.
1927 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1930 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1932 EXPORT_SYMBOL(vm_map_pages);
1935 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1936 * @vma: user vma to map to
1937 * @pages: pointer to array of source kernel pages
1938 * @num: number of pages in page array
1940 * Similar to vm_map_pages(), except that it explicitly sets the offset
1941 * to 0. This function is intended for the drivers that did not consider
1944 * Context: Process context. Called by mmap handlers.
1945 * Return: 0 on success and error code otherwise.
1947 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1950 return __vm_map_pages(vma, pages, num, 0);
1952 EXPORT_SYMBOL(vm_map_pages_zero);
1954 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1955 pfn_t pfn, pgprot_t prot, bool mkwrite)
1957 struct mm_struct *mm = vma->vm_mm;
1961 pte = get_locked_pte(mm, addr, &ptl);
1963 return VM_FAULT_OOM;
1964 if (!pte_none(*pte)) {
1967 * For read faults on private mappings the PFN passed
1968 * in may not match the PFN we have mapped if the
1969 * mapped PFN is a writeable COW page. In the mkwrite
1970 * case we are creating a writable PTE for a shared
1971 * mapping and we expect the PFNs to match. If they
1972 * don't match, we are likely racing with block
1973 * allocation and mapping invalidation so just skip the
1976 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1977 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1980 entry = pte_mkyoung(*pte);
1981 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1982 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1983 update_mmu_cache(vma, addr, pte);
1988 /* Ok, finally just insert the thing.. */
1989 if (pfn_t_devmap(pfn))
1990 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1992 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1995 entry = pte_mkyoung(entry);
1996 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1999 set_pte_at(mm, addr, pte, entry);
2000 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2003 pte_unmap_unlock(pte, ptl);
2004 return VM_FAULT_NOPAGE;
2008 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2009 * @vma: user vma to map to
2010 * @addr: target user address of this page
2011 * @pfn: source kernel pfn
2012 * @pgprot: pgprot flags for the inserted page
2014 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2015 * to override pgprot on a per-page basis.
2017 * This only makes sense for IO mappings, and it makes no sense for
2018 * COW mappings. In general, using multiple vmas is preferable;
2019 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2022 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2023 * a value of @pgprot different from that of @vma->vm_page_prot.
2025 * Context: Process context. May allocate using %GFP_KERNEL.
2026 * Return: vm_fault_t value.
2028 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2029 unsigned long pfn, pgprot_t pgprot)
2032 * Technically, architectures with pte_special can avoid all these
2033 * restrictions (same for remap_pfn_range). However we would like
2034 * consistency in testing and feature parity among all, so we should
2035 * try to keep these invariants in place for everybody.
2037 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2038 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2039 (VM_PFNMAP|VM_MIXEDMAP));
2040 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2041 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2043 if (addr < vma->vm_start || addr >= vma->vm_end)
2044 return VM_FAULT_SIGBUS;
2046 if (!pfn_modify_allowed(pfn, pgprot))
2047 return VM_FAULT_SIGBUS;
2049 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2051 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2054 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2057 * vmf_insert_pfn - insert single pfn into user vma
2058 * @vma: user vma to map to
2059 * @addr: target user address of this page
2060 * @pfn: source kernel pfn
2062 * Similar to vm_insert_page, this allows drivers to insert individual pages
2063 * they've allocated into a user vma. Same comments apply.
2065 * This function should only be called from a vm_ops->fault handler, and
2066 * in that case the handler should return the result of this function.
2068 * vma cannot be a COW mapping.
2070 * As this is called only for pages that do not currently exist, we
2071 * do not need to flush old virtual caches or the TLB.
2073 * Context: Process context. May allocate using %GFP_KERNEL.
2074 * Return: vm_fault_t value.
2076 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2079 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2081 EXPORT_SYMBOL(vmf_insert_pfn);
2083 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2085 /* these checks mirror the abort conditions in vm_normal_page */
2086 if (vma->vm_flags & VM_MIXEDMAP)
2088 if (pfn_t_devmap(pfn))
2090 if (pfn_t_special(pfn))
2092 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2097 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2098 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2103 BUG_ON(!vm_mixed_ok(vma, pfn));
2105 if (addr < vma->vm_start || addr >= vma->vm_end)
2106 return VM_FAULT_SIGBUS;
2108 track_pfn_insert(vma, &pgprot, pfn);
2110 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2111 return VM_FAULT_SIGBUS;
2114 * If we don't have pte special, then we have to use the pfn_valid()
2115 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2116 * refcount the page if pfn_valid is true (hence insert_page rather
2117 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2118 * without pte special, it would there be refcounted as a normal page.
2120 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2121 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2125 * At this point we are committed to insert_page()
2126 * regardless of whether the caller specified flags that
2127 * result in pfn_t_has_page() == false.
2129 page = pfn_to_page(pfn_t_to_pfn(pfn));
2130 err = insert_page(vma, addr, page, pgprot);
2132 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2136 return VM_FAULT_OOM;
2137 if (err < 0 && err != -EBUSY)
2138 return VM_FAULT_SIGBUS;
2140 return VM_FAULT_NOPAGE;
2144 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2145 * @vma: user vma to map to
2146 * @addr: target user address of this page
2147 * @pfn: source kernel pfn
2148 * @pgprot: pgprot flags for the inserted page
2150 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2151 * to override pgprot on a per-page basis.
2153 * Typically this function should be used by drivers to set caching- and
2154 * encryption bits different than those of @vma->vm_page_prot, because
2155 * the caching- or encryption mode may not be known at mmap() time.
2156 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2157 * to set caching and encryption bits for those vmas (except for COW pages).
2158 * This is ensured by core vm only modifying these page table entries using
2159 * functions that don't touch caching- or encryption bits, using pte_modify()
2160 * if needed. (See for example mprotect()).
2161 * Also when new page-table entries are created, this is only done using the
2162 * fault() callback, and never using the value of vma->vm_page_prot,
2163 * except for page-table entries that point to anonymous pages as the result
2166 * Context: Process context. May allocate using %GFP_KERNEL.
2167 * Return: vm_fault_t value.
2169 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2170 pfn_t pfn, pgprot_t pgprot)
2172 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2174 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2176 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2179 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2181 EXPORT_SYMBOL(vmf_insert_mixed);
2184 * If the insertion of PTE failed because someone else already added a
2185 * different entry in the mean time, we treat that as success as we assume
2186 * the same entry was actually inserted.
2188 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2189 unsigned long addr, pfn_t pfn)
2191 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2193 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2196 * maps a range of physical memory into the requested pages. the old
2197 * mappings are removed. any references to nonexistent pages results
2198 * in null mappings (currently treated as "copy-on-access")
2200 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2201 unsigned long addr, unsigned long end,
2202 unsigned long pfn, pgprot_t prot)
2204 pte_t *pte, *mapped_pte;
2208 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2211 arch_enter_lazy_mmu_mode();
2213 BUG_ON(!pte_none(*pte));
2214 if (!pfn_modify_allowed(pfn, prot)) {
2218 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2220 } while (pte++, addr += PAGE_SIZE, addr != end);
2221 arch_leave_lazy_mmu_mode();
2222 pte_unmap_unlock(mapped_pte, ptl);
2226 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2227 unsigned long addr, unsigned long end,
2228 unsigned long pfn, pgprot_t prot)
2234 pfn -= addr >> PAGE_SHIFT;
2235 pmd = pmd_alloc(mm, pud, addr);
2238 VM_BUG_ON(pmd_trans_huge(*pmd));
2240 next = pmd_addr_end(addr, end);
2241 err = remap_pte_range(mm, pmd, addr, next,
2242 pfn + (addr >> PAGE_SHIFT), prot);
2245 } while (pmd++, addr = next, addr != end);
2249 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2250 unsigned long addr, unsigned long end,
2251 unsigned long pfn, pgprot_t prot)
2257 pfn -= addr >> PAGE_SHIFT;
2258 pud = pud_alloc(mm, p4d, addr);
2262 next = pud_addr_end(addr, end);
2263 err = remap_pmd_range(mm, pud, addr, next,
2264 pfn + (addr >> PAGE_SHIFT), prot);
2267 } while (pud++, addr = next, addr != end);
2271 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2272 unsigned long addr, unsigned long end,
2273 unsigned long pfn, pgprot_t prot)
2279 pfn -= addr >> PAGE_SHIFT;
2280 p4d = p4d_alloc(mm, pgd, addr);
2284 next = p4d_addr_end(addr, end);
2285 err = remap_pud_range(mm, p4d, addr, next,
2286 pfn + (addr >> PAGE_SHIFT), prot);
2289 } while (p4d++, addr = next, addr != end);
2294 * remap_pfn_range - remap kernel memory to userspace
2295 * @vma: user vma to map to
2296 * @addr: target page aligned user address to start at
2297 * @pfn: page frame number of kernel physical memory address
2298 * @size: size of mapping area
2299 * @prot: page protection flags for this mapping
2301 * Note: this is only safe if the mm semaphore is held when called.
2303 * Return: %0 on success, negative error code otherwise.
2305 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2306 unsigned long pfn, unsigned long size, pgprot_t prot)
2310 unsigned long end = addr + PAGE_ALIGN(size);
2311 struct mm_struct *mm = vma->vm_mm;
2312 unsigned long remap_pfn = pfn;
2315 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2319 * Physically remapped pages are special. Tell the
2320 * rest of the world about it:
2321 * VM_IO tells people not to look at these pages
2322 * (accesses can have side effects).
2323 * VM_PFNMAP tells the core MM that the base pages are just
2324 * raw PFN mappings, and do not have a "struct page" associated
2327 * Disable vma merging and expanding with mremap().
2329 * Omit vma from core dump, even when VM_IO turned off.
2331 * There's a horrible special case to handle copy-on-write
2332 * behaviour that some programs depend on. We mark the "original"
2333 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2334 * See vm_normal_page() for details.
2336 if (is_cow_mapping(vma->vm_flags)) {
2337 if (addr != vma->vm_start || end != vma->vm_end)
2339 vma->vm_pgoff = pfn;
2342 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2346 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2348 BUG_ON(addr >= end);
2349 pfn -= addr >> PAGE_SHIFT;
2350 pgd = pgd_offset(mm, addr);
2351 flush_cache_range(vma, addr, end);
2353 next = pgd_addr_end(addr, end);
2354 err = remap_p4d_range(mm, pgd, addr, next,
2355 pfn + (addr >> PAGE_SHIFT), prot);
2358 } while (pgd++, addr = next, addr != end);
2361 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2365 EXPORT_SYMBOL(remap_pfn_range);
2368 * vm_iomap_memory - remap memory to userspace
2369 * @vma: user vma to map to
2370 * @start: start of the physical memory to be mapped
2371 * @len: size of area
2373 * This is a simplified io_remap_pfn_range() for common driver use. The
2374 * driver just needs to give us the physical memory range to be mapped,
2375 * we'll figure out the rest from the vma information.
2377 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2378 * whatever write-combining details or similar.
2380 * Return: %0 on success, negative error code otherwise.
2382 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2384 unsigned long vm_len, pfn, pages;
2386 /* Check that the physical memory area passed in looks valid */
2387 if (start + len < start)
2390 * You *really* shouldn't map things that aren't page-aligned,
2391 * but we've historically allowed it because IO memory might
2392 * just have smaller alignment.
2394 len += start & ~PAGE_MASK;
2395 pfn = start >> PAGE_SHIFT;
2396 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2397 if (pfn + pages < pfn)
2400 /* We start the mapping 'vm_pgoff' pages into the area */
2401 if (vma->vm_pgoff > pages)
2403 pfn += vma->vm_pgoff;
2404 pages -= vma->vm_pgoff;
2406 /* Can we fit all of the mapping? */
2407 vm_len = vma->vm_end - vma->vm_start;
2408 if (vm_len >> PAGE_SHIFT > pages)
2411 /* Ok, let it rip */
2412 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2414 EXPORT_SYMBOL(vm_iomap_memory);
2416 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2417 unsigned long addr, unsigned long end,
2418 pte_fn_t fn, void *data, bool create,
2419 pgtbl_mod_mask *mask)
2426 pte = (mm == &init_mm) ?
2427 pte_alloc_kernel_track(pmd, addr, mask) :
2428 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2432 pte = (mm == &init_mm) ?
2433 pte_offset_kernel(pmd, addr) :
2434 pte_offset_map_lock(mm, pmd, addr, &ptl);
2437 BUG_ON(pmd_huge(*pmd));
2439 arch_enter_lazy_mmu_mode();
2443 if (create || !pte_none(*pte)) {
2444 err = fn(pte++, addr, data);
2448 } while (addr += PAGE_SIZE, addr != end);
2450 *mask |= PGTBL_PTE_MODIFIED;
2452 arch_leave_lazy_mmu_mode();
2455 pte_unmap_unlock(pte-1, ptl);
2459 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2460 unsigned long addr, unsigned long end,
2461 pte_fn_t fn, void *data, bool create,
2462 pgtbl_mod_mask *mask)
2468 BUG_ON(pud_huge(*pud));
2471 pmd = pmd_alloc_track(mm, pud, addr, mask);
2475 pmd = pmd_offset(pud, addr);
2478 next = pmd_addr_end(addr, end);
2479 if (create || !pmd_none_or_clear_bad(pmd)) {
2480 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2485 } while (pmd++, addr = next, addr != end);
2489 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2490 unsigned long addr, unsigned long end,
2491 pte_fn_t fn, void *data, bool create,
2492 pgtbl_mod_mask *mask)
2499 pud = pud_alloc_track(mm, p4d, addr, mask);
2503 pud = pud_offset(p4d, addr);
2506 next = pud_addr_end(addr, end);
2507 if (create || !pud_none_or_clear_bad(pud)) {
2508 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2513 } while (pud++, addr = next, addr != end);
2517 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2518 unsigned long addr, unsigned long end,
2519 pte_fn_t fn, void *data, bool create,
2520 pgtbl_mod_mask *mask)
2527 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2531 p4d = p4d_offset(pgd, addr);
2534 next = p4d_addr_end(addr, end);
2535 if (create || !p4d_none_or_clear_bad(p4d)) {
2536 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2541 } while (p4d++, addr = next, addr != end);
2545 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2546 unsigned long size, pte_fn_t fn,
2547 void *data, bool create)
2550 unsigned long start = addr, next;
2551 unsigned long end = addr + size;
2552 pgtbl_mod_mask mask = 0;
2555 if (WARN_ON(addr >= end))
2558 pgd = pgd_offset(mm, addr);
2560 next = pgd_addr_end(addr, end);
2561 if (!create && pgd_none_or_clear_bad(pgd))
2563 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2566 } while (pgd++, addr = next, addr != end);
2568 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2569 arch_sync_kernel_mappings(start, start + size);
2575 * Scan a region of virtual memory, filling in page tables as necessary
2576 * and calling a provided function on each leaf page table.
2578 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2579 unsigned long size, pte_fn_t fn, void *data)
2581 return __apply_to_page_range(mm, addr, size, fn, data, true);
2583 EXPORT_SYMBOL_GPL(apply_to_page_range);
2586 * Scan a region of virtual memory, calling a provided function on
2587 * each leaf page table where it exists.
2589 * Unlike apply_to_page_range, this does _not_ fill in page tables
2590 * where they are absent.
2592 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2593 unsigned long size, pte_fn_t fn, void *data)
2595 return __apply_to_page_range(mm, addr, size, fn, data, false);
2597 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2600 * handle_pte_fault chooses page fault handler according to an entry which was
2601 * read non-atomically. Before making any commitment, on those architectures
2602 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2603 * parts, do_swap_page must check under lock before unmapping the pte and
2604 * proceeding (but do_wp_page is only called after already making such a check;
2605 * and do_anonymous_page can safely check later on).
2607 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2608 pte_t *page_table, pte_t orig_pte)
2611 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2612 if (sizeof(pte_t) > sizeof(unsigned long)) {
2613 spinlock_t *ptl = pte_lockptr(mm, pmd);
2615 same = pte_same(*page_table, orig_pte);
2619 pte_unmap(page_table);
2623 static inline bool cow_user_page(struct page *dst, struct page *src,
2624 struct vm_fault *vmf)
2629 bool locked = false;
2630 struct vm_area_struct *vma = vmf->vma;
2631 struct mm_struct *mm = vma->vm_mm;
2632 unsigned long addr = vmf->address;
2635 copy_user_highpage(dst, src, addr, vma);
2640 * If the source page was a PFN mapping, we don't have
2641 * a "struct page" for it. We do a best-effort copy by
2642 * just copying from the original user address. If that
2643 * fails, we just zero-fill it. Live with it.
2645 kaddr = kmap_atomic(dst);
2646 uaddr = (void __user *)(addr & PAGE_MASK);
2649 * On architectures with software "accessed" bits, we would
2650 * take a double page fault, so mark it accessed here.
2652 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2655 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2657 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2659 * Other thread has already handled the fault
2660 * and update local tlb only
2662 update_mmu_tlb(vma, addr, vmf->pte);
2667 entry = pte_mkyoung(vmf->orig_pte);
2668 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2669 update_mmu_cache(vma, addr, vmf->pte);
2673 * This really shouldn't fail, because the page is there
2674 * in the page tables. But it might just be unreadable,
2675 * in which case we just give up and fill the result with
2678 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2682 /* Re-validate under PTL if the page is still mapped */
2683 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2685 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2686 /* The PTE changed under us, update local tlb */
2687 update_mmu_tlb(vma, addr, vmf->pte);
2693 * The same page can be mapped back since last copy attempt.
2694 * Try to copy again under PTL.
2696 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2698 * Give a warn in case there can be some obscure
2711 pte_unmap_unlock(vmf->pte, vmf->ptl);
2712 kunmap_atomic(kaddr);
2713 flush_dcache_page(dst);
2718 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2720 struct file *vm_file = vma->vm_file;
2723 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2726 * Special mappings (e.g. VDSO) do not have any file so fake
2727 * a default GFP_KERNEL for them.
2733 * Notify the address space that the page is about to become writable so that
2734 * it can prohibit this or wait for the page to get into an appropriate state.
2736 * We do this without the lock held, so that it can sleep if it needs to.
2738 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2741 struct page *page = vmf->page;
2742 unsigned int old_flags = vmf->flags;
2744 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2746 if (vmf->vma->vm_file &&
2747 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2748 return VM_FAULT_SIGBUS;
2750 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2751 /* Restore original flags so that caller is not surprised */
2752 vmf->flags = old_flags;
2753 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2755 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2757 if (!page->mapping) {
2759 return 0; /* retry */
2761 ret |= VM_FAULT_LOCKED;
2763 VM_BUG_ON_PAGE(!PageLocked(page), page);
2768 * Handle dirtying of a page in shared file mapping on a write fault.
2770 * The function expects the page to be locked and unlocks it.
2772 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2774 struct vm_area_struct *vma = vmf->vma;
2775 struct address_space *mapping;
2776 struct page *page = vmf->page;
2778 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2780 dirtied = set_page_dirty(page);
2781 VM_BUG_ON_PAGE(PageAnon(page), page);
2783 * Take a local copy of the address_space - page.mapping may be zeroed
2784 * by truncate after unlock_page(). The address_space itself remains
2785 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2786 * release semantics to prevent the compiler from undoing this copying.
2788 mapping = page_rmapping(page);
2792 file_update_time(vma->vm_file);
2795 * Throttle page dirtying rate down to writeback speed.
2797 * mapping may be NULL here because some device drivers do not
2798 * set page.mapping but still dirty their pages
2800 * Drop the mmap_lock before waiting on IO, if we can. The file
2801 * is pinning the mapping, as per above.
2803 if ((dirtied || page_mkwrite) && mapping) {
2806 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2807 balance_dirty_pages_ratelimited(mapping);
2810 return VM_FAULT_RETRY;
2818 * Handle write page faults for pages that can be reused in the current vma
2820 * This can happen either due to the mapping being with the VM_SHARED flag,
2821 * or due to us being the last reference standing to the page. In either
2822 * case, all we need to do here is to mark the page as writable and update
2823 * any related book-keeping.
2825 static inline void wp_page_reuse(struct vm_fault *vmf)
2826 __releases(vmf->ptl)
2828 struct vm_area_struct *vma = vmf->vma;
2829 struct page *page = vmf->page;
2832 * Clear the pages cpupid information as the existing
2833 * information potentially belongs to a now completely
2834 * unrelated process.
2837 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2839 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2840 entry = pte_mkyoung(vmf->orig_pte);
2841 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2842 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2843 update_mmu_cache(vma, vmf->address, vmf->pte);
2844 pte_unmap_unlock(vmf->pte, vmf->ptl);
2845 count_vm_event(PGREUSE);
2849 * Handle the case of a page which we actually need to copy to a new page.
2851 * Called with mmap_lock locked and the old page referenced, but
2852 * without the ptl held.
2854 * High level logic flow:
2856 * - Allocate a page, copy the content of the old page to the new one.
2857 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2858 * - Take the PTL. If the pte changed, bail out and release the allocated page
2859 * - If the pte is still the way we remember it, update the page table and all
2860 * relevant references. This includes dropping the reference the page-table
2861 * held to the old page, as well as updating the rmap.
2862 * - In any case, unlock the PTL and drop the reference we took to the old page.
2864 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2866 struct vm_area_struct *vma = vmf->vma;
2867 struct mm_struct *mm = vma->vm_mm;
2868 struct page *old_page = vmf->page;
2869 struct page *new_page = NULL;
2871 int page_copied = 0;
2872 struct mmu_notifier_range range;
2874 if (unlikely(anon_vma_prepare(vma)))
2877 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2878 new_page = alloc_zeroed_user_highpage_movable(vma,
2883 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2888 if (!cow_user_page(new_page, old_page, vmf)) {
2890 * COW failed, if the fault was solved by other,
2891 * it's fine. If not, userspace would re-fault on
2892 * the same address and we will handle the fault
2893 * from the second attempt.
2902 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2904 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2906 __SetPageUptodate(new_page);
2908 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2909 vmf->address & PAGE_MASK,
2910 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2911 mmu_notifier_invalidate_range_start(&range);
2914 * Re-check the pte - we dropped the lock
2916 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2917 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2919 if (!PageAnon(old_page)) {
2920 dec_mm_counter_fast(mm,
2921 mm_counter_file(old_page));
2922 inc_mm_counter_fast(mm, MM_ANONPAGES);
2925 inc_mm_counter_fast(mm, MM_ANONPAGES);
2927 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2928 entry = mk_pte(new_page, vma->vm_page_prot);
2929 entry = pte_sw_mkyoung(entry);
2930 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2932 * Clear the pte entry and flush it first, before updating the
2933 * pte with the new entry. This will avoid a race condition
2934 * seen in the presence of one thread doing SMC and another
2937 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2938 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2939 lru_cache_add_inactive_or_unevictable(new_page, vma);
2941 * We call the notify macro here because, when using secondary
2942 * mmu page tables (such as kvm shadow page tables), we want the
2943 * new page to be mapped directly into the secondary page table.
2945 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2946 update_mmu_cache(vma, vmf->address, vmf->pte);
2949 * Only after switching the pte to the new page may
2950 * we remove the mapcount here. Otherwise another
2951 * process may come and find the rmap count decremented
2952 * before the pte is switched to the new page, and
2953 * "reuse" the old page writing into it while our pte
2954 * here still points into it and can be read by other
2957 * The critical issue is to order this
2958 * page_remove_rmap with the ptp_clear_flush above.
2959 * Those stores are ordered by (if nothing else,)
2960 * the barrier present in the atomic_add_negative
2961 * in page_remove_rmap.
2963 * Then the TLB flush in ptep_clear_flush ensures that
2964 * no process can access the old page before the
2965 * decremented mapcount is visible. And the old page
2966 * cannot be reused until after the decremented
2967 * mapcount is visible. So transitively, TLBs to
2968 * old page will be flushed before it can be reused.
2970 page_remove_rmap(old_page, false);
2973 /* Free the old page.. */
2974 new_page = old_page;
2977 update_mmu_tlb(vma, vmf->address, vmf->pte);
2983 pte_unmap_unlock(vmf->pte, vmf->ptl);
2985 * No need to double call mmu_notifier->invalidate_range() callback as
2986 * the above ptep_clear_flush_notify() did already call it.
2988 mmu_notifier_invalidate_range_only_end(&range);
2991 * Don't let another task, with possibly unlocked vma,
2992 * keep the mlocked page.
2994 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2995 lock_page(old_page); /* LRU manipulation */
2996 if (PageMlocked(old_page))
2997 munlock_vma_page(old_page);
2998 unlock_page(old_page);
3002 return page_copied ? VM_FAULT_WRITE : 0;
3008 return VM_FAULT_OOM;
3012 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3013 * writeable once the page is prepared
3015 * @vmf: structure describing the fault
3017 * This function handles all that is needed to finish a write page fault in a
3018 * shared mapping due to PTE being read-only once the mapped page is prepared.
3019 * It handles locking of PTE and modifying it.
3021 * The function expects the page to be locked or other protection against
3022 * concurrent faults / writeback (such as DAX radix tree locks).
3024 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
3025 * we acquired PTE lock.
3027 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3029 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3030 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3033 * We might have raced with another page fault while we released the
3034 * pte_offset_map_lock.
3036 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3037 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3038 pte_unmap_unlock(vmf->pte, vmf->ptl);
3039 return VM_FAULT_NOPAGE;
3046 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3049 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3051 struct vm_area_struct *vma = vmf->vma;
3053 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3056 pte_unmap_unlock(vmf->pte, vmf->ptl);
3057 vmf->flags |= FAULT_FLAG_MKWRITE;
3058 ret = vma->vm_ops->pfn_mkwrite(vmf);
3059 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3061 return finish_mkwrite_fault(vmf);
3064 return VM_FAULT_WRITE;
3067 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3068 __releases(vmf->ptl)
3070 struct vm_area_struct *vma = vmf->vma;
3071 vm_fault_t ret = VM_FAULT_WRITE;
3073 get_page(vmf->page);
3075 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3078 pte_unmap_unlock(vmf->pte, vmf->ptl);
3079 tmp = do_page_mkwrite(vmf);
3080 if (unlikely(!tmp || (tmp &
3081 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3082 put_page(vmf->page);
3085 tmp = finish_mkwrite_fault(vmf);
3086 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3087 unlock_page(vmf->page);
3088 put_page(vmf->page);
3093 lock_page(vmf->page);
3095 ret |= fault_dirty_shared_page(vmf);
3096 put_page(vmf->page);
3102 * This routine handles present pages, when users try to write
3103 * to a shared page. It is done by copying the page to a new address
3104 * and decrementing the shared-page counter for the old page.
3106 * Note that this routine assumes that the protection checks have been
3107 * done by the caller (the low-level page fault routine in most cases).
3108 * Thus we can safely just mark it writable once we've done any necessary
3111 * We also mark the page dirty at this point even though the page will
3112 * change only once the write actually happens. This avoids a few races,
3113 * and potentially makes it more efficient.
3115 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3116 * but allow concurrent faults), with pte both mapped and locked.
3117 * We return with mmap_lock still held, but pte unmapped and unlocked.
3119 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3120 __releases(vmf->ptl)
3122 struct vm_area_struct *vma = vmf->vma;
3124 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3125 pte_unmap_unlock(vmf->pte, vmf->ptl);
3126 return handle_userfault(vmf, VM_UFFD_WP);
3130 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3131 * is flushed in this case before copying.
3133 if (unlikely(userfaultfd_wp(vmf->vma) &&
3134 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3135 flush_tlb_page(vmf->vma, vmf->address);
3137 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3140 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3143 * We should not cow pages in a shared writeable mapping.
3144 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3146 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3147 (VM_WRITE|VM_SHARED))
3148 return wp_pfn_shared(vmf);
3150 pte_unmap_unlock(vmf->pte, vmf->ptl);
3151 return wp_page_copy(vmf);
3155 * Take out anonymous pages first, anonymous shared vmas are
3156 * not dirty accountable.
3158 if (PageAnon(vmf->page)) {
3159 struct page *page = vmf->page;
3161 /* PageKsm() doesn't necessarily raise the page refcount */
3162 if (PageKsm(page) || page_count(page) != 1)
3164 if (!trylock_page(page))
3166 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3171 * Ok, we've got the only map reference, and the only
3172 * page count reference, and the page is locked,
3173 * it's dark out, and we're wearing sunglasses. Hit it.
3177 return VM_FAULT_WRITE;
3178 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3179 (VM_WRITE|VM_SHARED))) {
3180 return wp_page_shared(vmf);
3184 * Ok, we need to copy. Oh, well..
3186 get_page(vmf->page);
3188 pte_unmap_unlock(vmf->pte, vmf->ptl);
3189 return wp_page_copy(vmf);
3192 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3193 unsigned long start_addr, unsigned long end_addr,
3194 struct zap_details *details)
3196 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3199 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3200 struct zap_details *details)
3202 struct vm_area_struct *vma;
3203 pgoff_t vba, vea, zba, zea;
3205 vma_interval_tree_foreach(vma, root,
3206 details->first_index, details->last_index) {
3208 vba = vma->vm_pgoff;
3209 vea = vba + vma_pages(vma) - 1;
3210 zba = details->first_index;
3213 zea = details->last_index;
3217 unmap_mapping_range_vma(vma,
3218 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3219 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3225 * unmap_mapping_page() - Unmap single page from processes.
3226 * @page: The locked page to be unmapped.
3228 * Unmap this page from any userspace process which still has it mmaped.
3229 * Typically, for efficiency, the range of nearby pages has already been
3230 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3231 * truncation or invalidation holds the lock on a page, it may find that
3232 * the page has been remapped again: and then uses unmap_mapping_page()
3233 * to unmap it finally.
3235 void unmap_mapping_page(struct page *page)
3237 struct address_space *mapping = page->mapping;
3238 struct zap_details details = { };
3240 VM_BUG_ON(!PageLocked(page));
3241 VM_BUG_ON(PageTail(page));
3243 details.check_mapping = mapping;
3244 details.first_index = page->index;
3245 details.last_index = page->index + thp_nr_pages(page) - 1;
3246 details.single_page = page;
3248 i_mmap_lock_write(mapping);
3249 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3250 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3251 i_mmap_unlock_write(mapping);
3255 * unmap_mapping_pages() - Unmap pages from processes.
3256 * @mapping: The address space containing pages to be unmapped.
3257 * @start: Index of first page to be unmapped.
3258 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3259 * @even_cows: Whether to unmap even private COWed pages.
3261 * Unmap the pages in this address space from any userspace process which
3262 * has them mmaped. Generally, you want to remove COWed pages as well when
3263 * a file is being truncated, but not when invalidating pages from the page
3266 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3267 pgoff_t nr, bool even_cows)
3269 struct zap_details details = { };
3271 details.check_mapping = even_cows ? NULL : mapping;
3272 details.first_index = start;
3273 details.last_index = start + nr - 1;
3274 if (details.last_index < details.first_index)
3275 details.last_index = ULONG_MAX;
3277 i_mmap_lock_write(mapping);
3278 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3279 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3280 i_mmap_unlock_write(mapping);
3284 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3285 * address_space corresponding to the specified byte range in the underlying
3288 * @mapping: the address space containing mmaps to be unmapped.
3289 * @holebegin: byte in first page to unmap, relative to the start of
3290 * the underlying file. This will be rounded down to a PAGE_SIZE
3291 * boundary. Note that this is different from truncate_pagecache(), which
3292 * must keep the partial page. In contrast, we must get rid of
3294 * @holelen: size of prospective hole in bytes. This will be rounded
3295 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3297 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3298 * but 0 when invalidating pagecache, don't throw away private data.
3300 void unmap_mapping_range(struct address_space *mapping,
3301 loff_t const holebegin, loff_t const holelen, int even_cows)
3303 pgoff_t hba = (pgoff_t)(holebegin) >> PAGE_SHIFT;
3304 pgoff_t hlen = ((pgoff_t)(holelen) + PAGE_SIZE - 1) >> PAGE_SHIFT;
3306 /* Check for overflow. */
3307 if (sizeof(holelen) > sizeof(hlen)) {
3309 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3310 if (holeend & ~(long long)ULONG_MAX)
3311 hlen = ULONG_MAX - hba + 1;
3314 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3316 EXPORT_SYMBOL(unmap_mapping_range);
3319 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3320 * but allow concurrent faults), and pte mapped but not yet locked.
3321 * We return with pte unmapped and unlocked.
3323 * We return with the mmap_lock locked or unlocked in the same cases
3324 * as does filemap_fault().
3326 vm_fault_t do_swap_page(struct vm_fault *vmf)
3328 struct vm_area_struct *vma = vmf->vma;
3329 struct page *page = NULL, *swapcache;
3335 void *shadow = NULL;
3337 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3340 entry = pte_to_swp_entry(vmf->orig_pte);
3341 if (unlikely(non_swap_entry(entry))) {
3342 if (is_migration_entry(entry)) {
3343 migration_entry_wait(vma->vm_mm, vmf->pmd,
3345 } else if (is_device_private_entry(entry)) {
3346 vmf->page = device_private_entry_to_page(entry);
3347 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3348 } else if (is_hwpoison_entry(entry)) {
3349 ret = VM_FAULT_HWPOISON;
3351 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3352 ret = VM_FAULT_SIGBUS;
3358 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3359 page = lookup_swap_cache(entry, vma, vmf->address);
3363 struct swap_info_struct *si = swp_swap_info(entry);
3365 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3366 __swap_count(entry) == 1) {
3367 /* skip swapcache */
3368 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3373 __SetPageLocked(page);
3374 __SetPageSwapBacked(page);
3375 set_page_private(page, entry.val);
3377 /* Tell memcg to use swap ownership records */
3378 SetPageSwapCache(page);
3379 err = mem_cgroup_charge(page, vma->vm_mm,
3381 ClearPageSwapCache(page);
3387 shadow = get_shadow_from_swap_cache(entry);
3389 workingset_refault(page, shadow);
3391 lru_cache_add(page);
3392 swap_readpage(page, true);
3395 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3402 * Back out if somebody else faulted in this pte
3403 * while we released the pte lock.
3405 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3406 vmf->address, &vmf->ptl);
3407 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3409 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3413 /* Had to read the page from swap area: Major fault */
3414 ret = VM_FAULT_MAJOR;
3415 count_vm_event(PGMAJFAULT);
3416 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3417 } else if (PageHWPoison(page)) {
3419 * hwpoisoned dirty swapcache pages are kept for killing
3420 * owner processes (which may be unknown at hwpoison time)
3422 ret = VM_FAULT_HWPOISON;
3423 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3427 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3429 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3431 ret |= VM_FAULT_RETRY;
3436 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3437 * release the swapcache from under us. The page pin, and pte_same
3438 * test below, are not enough to exclude that. Even if it is still
3439 * swapcache, we need to check that the page's swap has not changed.
3441 if (unlikely((!PageSwapCache(page) ||
3442 page_private(page) != entry.val)) && swapcache)
3445 page = ksm_might_need_to_copy(page, vma, vmf->address);
3446 if (unlikely(!page)) {
3452 cgroup_throttle_swaprate(page, GFP_KERNEL);
3455 * Back out if somebody else already faulted in this pte.
3457 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3459 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3462 if (unlikely(!PageUptodate(page))) {
3463 ret = VM_FAULT_SIGBUS;
3468 * The page isn't present yet, go ahead with the fault.
3470 * Be careful about the sequence of operations here.
3471 * To get its accounting right, reuse_swap_page() must be called
3472 * while the page is counted on swap but not yet in mapcount i.e.
3473 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3474 * must be called after the swap_free(), or it will never succeed.
3477 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3478 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3479 pte = mk_pte(page, vma->vm_page_prot);
3480 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3481 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3482 vmf->flags &= ~FAULT_FLAG_WRITE;
3483 ret |= VM_FAULT_WRITE;
3484 exclusive = RMAP_EXCLUSIVE;
3486 flush_icache_page(vma, page);
3487 if (pte_swp_soft_dirty(vmf->orig_pte))
3488 pte = pte_mksoft_dirty(pte);
3489 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3490 pte = pte_mkuffd_wp(pte);
3491 pte = pte_wrprotect(pte);
3493 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3494 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3495 vmf->orig_pte = pte;
3497 /* ksm created a completely new copy */
3498 if (unlikely(page != swapcache && swapcache)) {
3499 page_add_new_anon_rmap(page, vma, vmf->address, false);
3500 lru_cache_add_inactive_or_unevictable(page, vma);
3502 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3506 if (mem_cgroup_swap_full(page) ||
3507 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3508 try_to_free_swap(page);
3510 if (page != swapcache && swapcache) {
3512 * Hold the lock to avoid the swap entry to be reused
3513 * until we take the PT lock for the pte_same() check
3514 * (to avoid false positives from pte_same). For
3515 * further safety release the lock after the swap_free
3516 * so that the swap count won't change under a
3517 * parallel locked swapcache.
3519 unlock_page(swapcache);
3520 put_page(swapcache);
3523 if (vmf->flags & FAULT_FLAG_WRITE) {
3524 ret |= do_wp_page(vmf);
3525 if (ret & VM_FAULT_ERROR)
3526 ret &= VM_FAULT_ERROR;
3530 /* No need to invalidate - it was non-present before */
3531 update_mmu_cache(vma, vmf->address, vmf->pte);
3533 pte_unmap_unlock(vmf->pte, vmf->ptl);
3537 pte_unmap_unlock(vmf->pte, vmf->ptl);
3542 if (page != swapcache && swapcache) {
3543 unlock_page(swapcache);
3544 put_page(swapcache);
3550 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3551 * but allow concurrent faults), and pte mapped but not yet locked.
3552 * We return with mmap_lock still held, but pte unmapped and unlocked.
3554 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3556 struct vm_area_struct *vma = vmf->vma;
3561 /* File mapping without ->vm_ops ? */
3562 if (vma->vm_flags & VM_SHARED)
3563 return VM_FAULT_SIGBUS;
3566 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3567 * pte_offset_map() on pmds where a huge pmd might be created
3568 * from a different thread.
3570 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3571 * parallel threads are excluded by other means.
3573 * Here we only have mmap_read_lock(mm).
3575 if (pte_alloc(vma->vm_mm, vmf->pmd))
3576 return VM_FAULT_OOM;
3578 /* See the comment in pte_alloc_one_map() */
3579 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3582 /* Use the zero-page for reads */
3583 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3584 !mm_forbids_zeropage(vma->vm_mm)) {
3585 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3586 vma->vm_page_prot));
3587 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3588 vmf->address, &vmf->ptl);
3589 if (!pte_none(*vmf->pte)) {
3590 update_mmu_tlb(vma, vmf->address, vmf->pte);
3593 ret = check_stable_address_space(vma->vm_mm);
3596 /* Deliver the page fault to userland, check inside PT lock */
3597 if (userfaultfd_missing(vma)) {
3598 pte_unmap_unlock(vmf->pte, vmf->ptl);
3599 return handle_userfault(vmf, VM_UFFD_MISSING);
3604 /* Allocate our own private page. */
3605 if (unlikely(anon_vma_prepare(vma)))
3607 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3611 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3613 cgroup_throttle_swaprate(page, GFP_KERNEL);
3616 * The memory barrier inside __SetPageUptodate makes sure that
3617 * preceding stores to the page contents become visible before
3618 * the set_pte_at() write.
3620 __SetPageUptodate(page);
3622 entry = mk_pte(page, vma->vm_page_prot);
3623 entry = pte_sw_mkyoung(entry);
3624 if (vma->vm_flags & VM_WRITE)
3625 entry = pte_mkwrite(pte_mkdirty(entry));
3627 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3629 if (!pte_none(*vmf->pte)) {
3630 update_mmu_cache(vma, vmf->address, vmf->pte);
3634 ret = check_stable_address_space(vma->vm_mm);
3638 /* Deliver the page fault to userland, check inside PT lock */
3639 if (userfaultfd_missing(vma)) {
3640 pte_unmap_unlock(vmf->pte, vmf->ptl);
3642 return handle_userfault(vmf, VM_UFFD_MISSING);
3645 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3646 page_add_new_anon_rmap(page, vma, vmf->address, false);
3647 lru_cache_add_inactive_or_unevictable(page, vma);
3649 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3651 /* No need to invalidate - it was non-present before */
3652 update_mmu_cache(vma, vmf->address, vmf->pte);
3654 pte_unmap_unlock(vmf->pte, vmf->ptl);
3662 return VM_FAULT_OOM;
3666 * The mmap_lock must have been held on entry, and may have been
3667 * released depending on flags and vma->vm_ops->fault() return value.
3668 * See filemap_fault() and __lock_page_retry().
3670 static vm_fault_t __do_fault(struct vm_fault *vmf)
3672 struct vm_area_struct *vma = vmf->vma;
3676 * Preallocate pte before we take page_lock because this might lead to
3677 * deadlocks for memcg reclaim which waits for pages under writeback:
3679 * SetPageWriteback(A)
3685 * wait_on_page_writeback(A)
3686 * SetPageWriteback(B)
3688 * # flush A, B to clear the writeback
3690 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3691 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3692 if (!vmf->prealloc_pte)
3693 return VM_FAULT_OOM;
3694 smp_wmb(); /* See comment in __pte_alloc() */
3697 ret = vma->vm_ops->fault(vmf);
3698 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3699 VM_FAULT_DONE_COW)))
3702 if (unlikely(PageHWPoison(vmf->page))) {
3703 struct page *page = vmf->page;
3704 vm_fault_t poisonret = VM_FAULT_HWPOISON;
3705 if (ret & VM_FAULT_LOCKED) {
3706 if (page_mapped(page))
3707 unmap_mapping_pages(page_mapping(page),
3708 page->index, 1, false);
3709 /* Retry if a clean page was removed from the cache. */
3710 if (invalidate_inode_page(page))
3711 poisonret = VM_FAULT_NOPAGE;
3719 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3720 lock_page(vmf->page);
3722 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3728 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3729 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3730 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3731 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3733 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3735 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3738 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3740 struct vm_area_struct *vma = vmf->vma;
3742 if (!pmd_none(*vmf->pmd))
3744 if (vmf->prealloc_pte) {
3745 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3746 if (unlikely(!pmd_none(*vmf->pmd))) {
3747 spin_unlock(vmf->ptl);
3751 mm_inc_nr_ptes(vma->vm_mm);
3752 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3753 spin_unlock(vmf->ptl);
3754 vmf->prealloc_pte = NULL;
3755 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3756 return VM_FAULT_OOM;
3760 * If a huge pmd materialized under us just retry later. Use
3761 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3762 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3763 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3764 * running immediately after a huge pmd fault in a different thread of
3765 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3766 * All we have to ensure is that it is a regular pmd that we can walk
3767 * with pte_offset_map() and we can do that through an atomic read in
3768 * C, which is what pmd_trans_unstable() provides.
3770 if (pmd_devmap_trans_unstable(vmf->pmd))
3771 return VM_FAULT_NOPAGE;
3774 * At this point we know that our vmf->pmd points to a page of ptes
3775 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3776 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3777 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3778 * be valid and we will re-check to make sure the vmf->pte isn't
3779 * pte_none() under vmf->ptl protection when we return to
3782 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3787 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3788 static void deposit_prealloc_pte(struct vm_fault *vmf)
3790 struct vm_area_struct *vma = vmf->vma;
3792 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3794 * We are going to consume the prealloc table,
3795 * count that as nr_ptes.
3797 mm_inc_nr_ptes(vma->vm_mm);
3798 vmf->prealloc_pte = NULL;
3801 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3803 struct vm_area_struct *vma = vmf->vma;
3804 bool write = vmf->flags & FAULT_FLAG_WRITE;
3805 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3808 vm_fault_t ret = VM_FAULT_FALLBACK;
3810 if (!transhuge_vma_suitable(vma, haddr))
3813 page = compound_head(page);
3814 if (compound_order(page) != HPAGE_PMD_ORDER)
3818 * Archs like ppc64 need additonal space to store information
3819 * related to pte entry. Use the preallocated table for that.
3821 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3822 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3823 if (!vmf->prealloc_pte)
3824 return VM_FAULT_OOM;
3825 smp_wmb(); /* See comment in __pte_alloc() */
3828 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3829 if (unlikely(!pmd_none(*vmf->pmd)))
3832 for (i = 0; i < HPAGE_PMD_NR; i++)
3833 flush_icache_page(vma, page + i);
3835 entry = mk_huge_pmd(page, vma->vm_page_prot);
3837 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3839 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3840 page_add_file_rmap(page, true);
3842 * deposit and withdraw with pmd lock held
3844 if (arch_needs_pgtable_deposit())
3845 deposit_prealloc_pte(vmf);
3847 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3849 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3851 /* fault is handled */
3853 count_vm_event(THP_FILE_MAPPED);
3855 spin_unlock(vmf->ptl);
3859 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3867 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3868 * mapping. If needed, the function allocates page table or use pre-allocated.
3870 * @vmf: fault environment
3871 * @page: page to map
3873 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3876 * Target users are page handler itself and implementations of
3877 * vm_ops->map_pages.
3879 * Return: %0 on success, %VM_FAULT_ code in case of error.
3881 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
3883 struct vm_area_struct *vma = vmf->vma;
3884 bool write = vmf->flags & FAULT_FLAG_WRITE;
3888 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3889 ret = do_set_pmd(vmf, page);
3890 if (ret != VM_FAULT_FALLBACK)
3895 ret = pte_alloc_one_map(vmf);
3900 /* Re-check under ptl */
3901 if (unlikely(!pte_none(*vmf->pte))) {
3902 update_mmu_tlb(vma, vmf->address, vmf->pte);
3903 return VM_FAULT_NOPAGE;
3906 flush_icache_page(vma, page);
3907 entry = mk_pte(page, vma->vm_page_prot);
3908 entry = pte_sw_mkyoung(entry);
3910 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3911 /* copy-on-write page */
3912 if (write && !(vma->vm_flags & VM_SHARED)) {
3913 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3914 page_add_new_anon_rmap(page, vma, vmf->address, false);
3915 lru_cache_add_inactive_or_unevictable(page, vma);
3917 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3918 page_add_file_rmap(page, false);
3920 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3922 /* no need to invalidate: a not-present page won't be cached */
3923 update_mmu_cache(vma, vmf->address, vmf->pte);
3930 * finish_fault - finish page fault once we have prepared the page to fault
3932 * @vmf: structure describing the fault
3934 * This function handles all that is needed to finish a page fault once the
3935 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3936 * given page, adds reverse page mapping, handles memcg charges and LRU
3939 * The function expects the page to be locked and on success it consumes a
3940 * reference of a page being mapped (for the PTE which maps it).
3942 * Return: %0 on success, %VM_FAULT_ code in case of error.
3944 vm_fault_t finish_fault(struct vm_fault *vmf)
3949 /* Did we COW the page? */
3950 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3951 !(vmf->vma->vm_flags & VM_SHARED))
3952 page = vmf->cow_page;
3957 * check even for read faults because we might have lost our CoWed
3960 if (!(vmf->vma->vm_flags & VM_SHARED))
3961 ret = check_stable_address_space(vmf->vma->vm_mm);
3963 ret = alloc_set_pte(vmf, page);
3965 pte_unmap_unlock(vmf->pte, vmf->ptl);
3969 static unsigned long fault_around_bytes __read_mostly =
3970 rounddown_pow_of_two(65536);
3972 #ifdef CONFIG_DEBUG_FS
3973 static int fault_around_bytes_get(void *data, u64 *val)
3975 *val = fault_around_bytes;
3980 * fault_around_bytes must be rounded down to the nearest page order as it's
3981 * what do_fault_around() expects to see.
3983 static int fault_around_bytes_set(void *data, u64 val)
3985 if (val / PAGE_SIZE > PTRS_PER_PTE)
3987 if (val > PAGE_SIZE)
3988 fault_around_bytes = rounddown_pow_of_two(val);
3990 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3993 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3994 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3996 static int __init fault_around_debugfs(void)
3998 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3999 &fault_around_bytes_fops);
4002 late_initcall(fault_around_debugfs);
4006 * do_fault_around() tries to map few pages around the fault address. The hope
4007 * is that the pages will be needed soon and this will lower the number of
4010 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4011 * not ready to be mapped: not up-to-date, locked, etc.
4013 * This function is called with the page table lock taken. In the split ptlock
4014 * case the page table lock only protects only those entries which belong to
4015 * the page table corresponding to the fault address.
4017 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4020 * fault_around_bytes defines how many bytes we'll try to map.
4021 * do_fault_around() expects it to be set to a power of two less than or equal
4024 * The virtual address of the area that we map is naturally aligned to
4025 * fault_around_bytes rounded down to the machine page size
4026 * (and therefore to page order). This way it's easier to guarantee
4027 * that we don't cross page table boundaries.
4029 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4031 unsigned long address = vmf->address, nr_pages, mask;
4032 pgoff_t start_pgoff = vmf->pgoff;
4037 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4038 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4040 vmf->address = max(address & mask, vmf->vma->vm_start);
4041 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4045 * end_pgoff is either the end of the page table, the end of
4046 * the vma or nr_pages from start_pgoff, depending what is nearest.
4048 end_pgoff = start_pgoff -
4049 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4051 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4052 start_pgoff + nr_pages - 1);
4054 if (pmd_none(*vmf->pmd)) {
4055 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4056 if (!vmf->prealloc_pte)
4058 smp_wmb(); /* See comment in __pte_alloc() */
4061 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4063 /* Huge page is mapped? Page fault is solved */
4064 if (pmd_trans_huge(*vmf->pmd)) {
4065 ret = VM_FAULT_NOPAGE;
4069 /* ->map_pages() haven't done anything useful. Cold page cache? */
4073 /* check if the page fault is solved */
4074 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
4075 if (!pte_none(*vmf->pte))
4076 ret = VM_FAULT_NOPAGE;
4077 pte_unmap_unlock(vmf->pte, vmf->ptl);
4079 vmf->address = address;
4084 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4086 struct vm_area_struct *vma = vmf->vma;
4090 * Let's call ->map_pages() first and use ->fault() as fallback
4091 * if page by the offset is not ready to be mapped (cold cache or
4094 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4095 ret = do_fault_around(vmf);
4100 ret = __do_fault(vmf);
4101 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4104 ret |= finish_fault(vmf);
4105 unlock_page(vmf->page);
4106 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4107 put_page(vmf->page);
4111 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4113 struct vm_area_struct *vma = vmf->vma;
4116 if (unlikely(anon_vma_prepare(vma)))
4117 return VM_FAULT_OOM;
4119 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4121 return VM_FAULT_OOM;
4123 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4124 put_page(vmf->cow_page);
4125 return VM_FAULT_OOM;
4127 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4129 ret = __do_fault(vmf);
4130 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4132 if (ret & VM_FAULT_DONE_COW)
4135 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4136 __SetPageUptodate(vmf->cow_page);
4138 ret |= finish_fault(vmf);
4139 unlock_page(vmf->page);
4140 put_page(vmf->page);
4141 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4145 put_page(vmf->cow_page);
4149 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4151 struct vm_area_struct *vma = vmf->vma;
4152 vm_fault_t ret, tmp;
4154 ret = __do_fault(vmf);
4155 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4159 * Check if the backing address space wants to know that the page is
4160 * about to become writable
4162 if (vma->vm_ops->page_mkwrite) {
4163 unlock_page(vmf->page);
4164 tmp = do_page_mkwrite(vmf);
4165 if (unlikely(!tmp ||
4166 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4167 put_page(vmf->page);
4172 ret |= finish_fault(vmf);
4173 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4175 unlock_page(vmf->page);
4176 put_page(vmf->page);
4180 ret |= fault_dirty_shared_page(vmf);
4185 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4186 * but allow concurrent faults).
4187 * The mmap_lock may have been released depending on flags and our
4188 * return value. See filemap_fault() and __lock_page_or_retry().
4189 * If mmap_lock is released, vma may become invalid (for example
4190 * by other thread calling munmap()).
4192 static vm_fault_t do_fault(struct vm_fault *vmf)
4194 struct vm_area_struct *vma = vmf->vma;
4195 struct mm_struct *vm_mm = vma->vm_mm;
4199 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4201 if (!vma->vm_ops->fault) {
4203 * If we find a migration pmd entry or a none pmd entry, which
4204 * should never happen, return SIGBUS
4206 if (unlikely(!pmd_present(*vmf->pmd)))
4207 ret = VM_FAULT_SIGBUS;
4209 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4214 * Make sure this is not a temporary clearing of pte
4215 * by holding ptl and checking again. A R/M/W update
4216 * of pte involves: take ptl, clearing the pte so that
4217 * we don't have concurrent modification by hardware
4218 * followed by an update.
4220 if (unlikely(pte_none(*vmf->pte)))
4221 ret = VM_FAULT_SIGBUS;
4223 ret = VM_FAULT_NOPAGE;
4225 pte_unmap_unlock(vmf->pte, vmf->ptl);
4227 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4228 ret = do_read_fault(vmf);
4229 else if (!(vma->vm_flags & VM_SHARED))
4230 ret = do_cow_fault(vmf);
4232 ret = do_shared_fault(vmf);
4234 /* preallocated pagetable is unused: free it */
4235 if (vmf->prealloc_pte) {
4236 pte_free(vm_mm, vmf->prealloc_pte);
4237 vmf->prealloc_pte = NULL;
4242 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4243 unsigned long addr, int page_nid,
4248 count_vm_numa_event(NUMA_HINT_FAULTS);
4249 if (page_nid == numa_node_id()) {
4250 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4251 *flags |= TNF_FAULT_LOCAL;
4254 return mpol_misplaced(page, vma, addr);
4257 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4259 struct vm_area_struct *vma = vmf->vma;
4260 struct page *page = NULL;
4261 int page_nid = NUMA_NO_NODE;
4264 bool migrated = false;
4266 bool was_writable = pte_savedwrite(vmf->orig_pte);
4270 * The "pte" at this point cannot be used safely without
4271 * validation through pte_unmap_same(). It's of NUMA type but
4272 * the pfn may be screwed if the read is non atomic.
4274 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4275 spin_lock(vmf->ptl);
4276 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4277 pte_unmap_unlock(vmf->pte, vmf->ptl);
4282 * Make it present again, Depending on how arch implementes non
4283 * accessible ptes, some can allow access by kernel mode.
4285 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4286 pte = pte_modify(old_pte, vma->vm_page_prot);
4287 pte = pte_mkyoung(pte);
4289 pte = pte_mkwrite(pte);
4290 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4291 update_mmu_cache(vma, vmf->address, vmf->pte);
4293 page = vm_normal_page(vma, vmf->address, pte);
4295 pte_unmap_unlock(vmf->pte, vmf->ptl);
4299 /* TODO: handle PTE-mapped THP */
4300 if (PageCompound(page)) {
4301 pte_unmap_unlock(vmf->pte, vmf->ptl);
4306 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4307 * much anyway since they can be in shared cache state. This misses
4308 * the case where a mapping is writable but the process never writes
4309 * to it but pte_write gets cleared during protection updates and
4310 * pte_dirty has unpredictable behaviour between PTE scan updates,
4311 * background writeback, dirty balancing and application behaviour.
4313 if (!pte_write(pte))
4314 flags |= TNF_NO_GROUP;
4317 * Flag if the page is shared between multiple address spaces. This
4318 * is later used when determining whether to group tasks together
4320 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4321 flags |= TNF_SHARED;
4323 last_cpupid = page_cpupid_last(page);
4324 page_nid = page_to_nid(page);
4325 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4327 pte_unmap_unlock(vmf->pte, vmf->ptl);
4328 if (target_nid == NUMA_NO_NODE) {
4333 /* Migrate to the requested node */
4334 migrated = migrate_misplaced_page(page, vma, target_nid);
4336 page_nid = target_nid;
4337 flags |= TNF_MIGRATED;
4339 flags |= TNF_MIGRATE_FAIL;
4342 if (page_nid != NUMA_NO_NODE)
4343 task_numa_fault(last_cpupid, page_nid, 1, flags);
4347 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4349 if (vma_is_anonymous(vmf->vma))
4350 return do_huge_pmd_anonymous_page(vmf);
4351 if (vmf->vma->vm_ops->huge_fault)
4352 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4353 return VM_FAULT_FALLBACK;
4356 /* `inline' is required to avoid gcc 4.1.2 build error */
4357 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4359 if (vma_is_anonymous(vmf->vma)) {
4360 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4361 return handle_userfault(vmf, VM_UFFD_WP);
4362 return do_huge_pmd_wp_page(vmf, orig_pmd);
4364 if (vmf->vma->vm_ops->huge_fault) {
4365 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4367 if (!(ret & VM_FAULT_FALLBACK))
4371 /* COW or write-notify handled on pte level: split pmd. */
4372 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4374 return VM_FAULT_FALLBACK;
4377 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4379 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4380 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4381 /* No support for anonymous transparent PUD pages yet */
4382 if (vma_is_anonymous(vmf->vma))
4383 return VM_FAULT_FALLBACK;
4384 if (vmf->vma->vm_ops->huge_fault)
4385 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4386 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4387 return VM_FAULT_FALLBACK;
4390 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4392 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4393 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4394 /* No support for anonymous transparent PUD pages yet */
4395 if (vma_is_anonymous(vmf->vma))
4397 if (vmf->vma->vm_ops->huge_fault) {
4398 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4400 if (!(ret & VM_FAULT_FALLBACK))
4404 /* COW or write-notify not handled on PUD level: split pud.*/
4405 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4406 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
4407 return VM_FAULT_FALLBACK;
4411 * These routines also need to handle stuff like marking pages dirty
4412 * and/or accessed for architectures that don't do it in hardware (most
4413 * RISC architectures). The early dirtying is also good on the i386.
4415 * There is also a hook called "update_mmu_cache()" that architectures
4416 * with external mmu caches can use to update those (ie the Sparc or
4417 * PowerPC hashed page tables that act as extended TLBs).
4419 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4420 * concurrent faults).
4422 * The mmap_lock may have been released depending on flags and our return value.
4423 * See filemap_fault() and __lock_page_or_retry().
4425 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4429 if (unlikely(pmd_none(*vmf->pmd))) {
4431 * Leave __pte_alloc() until later: because vm_ops->fault may
4432 * want to allocate huge page, and if we expose page table
4433 * for an instant, it will be difficult to retract from
4434 * concurrent faults and from rmap lookups.
4438 /* See comment in pte_alloc_one_map() */
4439 if (pmd_devmap_trans_unstable(vmf->pmd))
4442 * A regular pmd is established and it can't morph into a huge
4443 * pmd from under us anymore at this point because we hold the
4444 * mmap_lock read mode and khugepaged takes it in write mode.
4445 * So now it's safe to run pte_offset_map().
4447 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4448 vmf->orig_pte = *vmf->pte;
4451 * some architectures can have larger ptes than wordsize,
4452 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4453 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4454 * accesses. The code below just needs a consistent view
4455 * for the ifs and we later double check anyway with the
4456 * ptl lock held. So here a barrier will do.
4459 if (pte_none(vmf->orig_pte)) {
4460 pte_unmap(vmf->pte);
4466 if (vma_is_anonymous(vmf->vma))
4467 return do_anonymous_page(vmf);
4469 return do_fault(vmf);
4472 if (!pte_present(vmf->orig_pte))
4473 return do_swap_page(vmf);
4475 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4476 return do_numa_page(vmf);
4478 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4479 spin_lock(vmf->ptl);
4480 entry = vmf->orig_pte;
4481 if (unlikely(!pte_same(*vmf->pte, entry))) {
4482 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4485 if (vmf->flags & FAULT_FLAG_WRITE) {
4486 if (!pte_write(entry))
4487 return do_wp_page(vmf);
4488 entry = pte_mkdirty(entry);
4490 entry = pte_mkyoung(entry);
4491 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4492 vmf->flags & FAULT_FLAG_WRITE)) {
4493 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4495 /* Skip spurious TLB flush for retried page fault */
4496 if (vmf->flags & FAULT_FLAG_TRIED)
4499 * This is needed only for protection faults but the arch code
4500 * is not yet telling us if this is a protection fault or not.
4501 * This still avoids useless tlb flushes for .text page faults
4504 if (vmf->flags & FAULT_FLAG_WRITE)
4505 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4508 pte_unmap_unlock(vmf->pte, vmf->ptl);
4513 * By the time we get here, we already hold the mm semaphore
4515 * The mmap_lock may have been released depending on flags and our
4516 * return value. See filemap_fault() and __lock_page_or_retry().
4518 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4519 unsigned long address, unsigned int flags)
4521 struct vm_fault vmf = {
4523 .address = address & PAGE_MASK,
4525 .pgoff = linear_page_index(vma, address),
4526 .gfp_mask = __get_fault_gfp_mask(vma),
4528 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4529 struct mm_struct *mm = vma->vm_mm;
4534 pgd = pgd_offset(mm, address);
4535 p4d = p4d_alloc(mm, pgd, address);
4537 return VM_FAULT_OOM;
4539 vmf.pud = pud_alloc(mm, p4d, address);
4541 return VM_FAULT_OOM;
4543 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4544 ret = create_huge_pud(&vmf);
4545 if (!(ret & VM_FAULT_FALLBACK))
4548 pud_t orig_pud = *vmf.pud;
4551 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4553 /* NUMA case for anonymous PUDs would go here */
4555 if (dirty && !pud_write(orig_pud)) {
4556 ret = wp_huge_pud(&vmf, orig_pud);
4557 if (!(ret & VM_FAULT_FALLBACK))
4560 huge_pud_set_accessed(&vmf, orig_pud);
4566 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4568 return VM_FAULT_OOM;
4570 /* Huge pud page fault raced with pmd_alloc? */
4571 if (pud_trans_unstable(vmf.pud))
4574 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4575 ret = create_huge_pmd(&vmf);
4576 if (!(ret & VM_FAULT_FALLBACK))
4579 pmd_t orig_pmd = *vmf.pmd;
4582 if (unlikely(is_swap_pmd(orig_pmd))) {
4583 VM_BUG_ON(thp_migration_supported() &&
4584 !is_pmd_migration_entry(orig_pmd));
4585 if (is_pmd_migration_entry(orig_pmd))
4586 pmd_migration_entry_wait(mm, vmf.pmd);
4589 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4590 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4591 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4593 if (dirty && !pmd_write(orig_pmd)) {
4594 ret = wp_huge_pmd(&vmf, orig_pmd);
4595 if (!(ret & VM_FAULT_FALLBACK))
4598 huge_pmd_set_accessed(&vmf, orig_pmd);
4604 return handle_pte_fault(&vmf);
4608 * mm_account_fault - Do page fault accountings
4610 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4611 * of perf event counters, but we'll still do the per-task accounting to
4612 * the task who triggered this page fault.
4613 * @address: the faulted address.
4614 * @flags: the fault flags.
4615 * @ret: the fault retcode.
4617 * This will take care of most of the page fault accountings. Meanwhile, it
4618 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4619 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4620 * still be in per-arch page fault handlers at the entry of page fault.
4622 static inline void mm_account_fault(struct pt_regs *regs,
4623 unsigned long address, unsigned int flags,
4629 * We don't do accounting for some specific faults:
4631 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4632 * includes arch_vma_access_permitted() failing before reaching here.
4633 * So this is not a "this many hardware page faults" counter. We
4634 * should use the hw profiling for that.
4636 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4637 * once they're completed.
4639 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4643 * We define the fault as a major fault when the final successful fault
4644 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4645 * handle it immediately previously).
4647 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4655 * If the fault is done for GUP, regs will be NULL. We only do the
4656 * accounting for the per thread fault counters who triggered the
4657 * fault, and we skip the perf event updates.
4663 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4665 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4669 * By the time we get here, we already hold the mm semaphore
4671 * The mmap_lock may have been released depending on flags and our
4672 * return value. See filemap_fault() and __lock_page_or_retry().
4674 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4675 unsigned int flags, struct pt_regs *regs)
4679 __set_current_state(TASK_RUNNING);
4681 count_vm_event(PGFAULT);
4682 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4684 /* do counter updates before entering really critical section. */
4685 check_sync_rss_stat(current);
4687 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4688 flags & FAULT_FLAG_INSTRUCTION,
4689 flags & FAULT_FLAG_REMOTE))
4690 return VM_FAULT_SIGSEGV;
4693 * Enable the memcg OOM handling for faults triggered in user
4694 * space. Kernel faults are handled more gracefully.
4696 if (flags & FAULT_FLAG_USER)
4697 mem_cgroup_enter_user_fault();
4699 if (unlikely(is_vm_hugetlb_page(vma)))
4700 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4702 ret = __handle_mm_fault(vma, address, flags);
4704 if (flags & FAULT_FLAG_USER) {
4705 mem_cgroup_exit_user_fault();
4707 * The task may have entered a memcg OOM situation but
4708 * if the allocation error was handled gracefully (no
4709 * VM_FAULT_OOM), there is no need to kill anything.
4710 * Just clean up the OOM state peacefully.
4712 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4713 mem_cgroup_oom_synchronize(false);
4716 mm_account_fault(regs, address, flags, ret);
4720 EXPORT_SYMBOL_GPL(handle_mm_fault);
4722 #ifndef __PAGETABLE_P4D_FOLDED
4724 * Allocate p4d page table.
4725 * We've already handled the fast-path in-line.
4727 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4729 p4d_t *new = p4d_alloc_one(mm, address);
4733 smp_wmb(); /* See comment in __pte_alloc */
4735 spin_lock(&mm->page_table_lock);
4736 if (pgd_present(*pgd)) /* Another has populated it */
4739 pgd_populate(mm, pgd, new);
4740 spin_unlock(&mm->page_table_lock);
4743 #endif /* __PAGETABLE_P4D_FOLDED */
4745 #ifndef __PAGETABLE_PUD_FOLDED
4747 * Allocate page upper directory.
4748 * We've already handled the fast-path in-line.
4750 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4752 pud_t *new = pud_alloc_one(mm, address);
4756 smp_wmb(); /* See comment in __pte_alloc */
4758 spin_lock(&mm->page_table_lock);
4759 if (!p4d_present(*p4d)) {
4761 p4d_populate(mm, p4d, new);
4762 } else /* Another has populated it */
4764 spin_unlock(&mm->page_table_lock);
4767 #endif /* __PAGETABLE_PUD_FOLDED */
4769 #ifndef __PAGETABLE_PMD_FOLDED
4771 * Allocate page middle directory.
4772 * We've already handled the fast-path in-line.
4774 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4777 pmd_t *new = pmd_alloc_one(mm, address);
4781 smp_wmb(); /* See comment in __pte_alloc */
4783 ptl = pud_lock(mm, pud);
4784 if (!pud_present(*pud)) {
4786 pud_populate(mm, pud, new);
4787 } else /* Another has populated it */
4792 #endif /* __PAGETABLE_PMD_FOLDED */
4794 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4795 struct mmu_notifier_range *range, pte_t **ptepp,
4796 pmd_t **pmdpp, spinlock_t **ptlp)
4804 pgd = pgd_offset(mm, address);
4805 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4808 p4d = p4d_offset(pgd, address);
4809 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4812 pud = pud_offset(p4d, address);
4813 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4816 pmd = pmd_offset(pud, address);
4817 VM_BUG_ON(pmd_trans_huge(*pmd));
4819 if (pmd_huge(*pmd)) {
4824 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4825 NULL, mm, address & PMD_MASK,
4826 (address & PMD_MASK) + PMD_SIZE);
4827 mmu_notifier_invalidate_range_start(range);
4829 *ptlp = pmd_lock(mm, pmd);
4830 if (pmd_huge(*pmd)) {
4836 mmu_notifier_invalidate_range_end(range);
4839 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4843 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4844 address & PAGE_MASK,
4845 (address & PAGE_MASK) + PAGE_SIZE);
4846 mmu_notifier_invalidate_range_start(range);
4848 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4849 if (!pte_present(*ptep))
4854 pte_unmap_unlock(ptep, *ptlp);
4856 mmu_notifier_invalidate_range_end(range);
4862 * follow_pte - look up PTE at a user virtual address
4863 * @mm: the mm_struct of the target address space
4864 * @address: user virtual address
4865 * @ptepp: location to store found PTE
4866 * @ptlp: location to store the lock for the PTE
4868 * On a successful return, the pointer to the PTE is stored in @ptepp;
4869 * the corresponding lock is taken and its location is stored in @ptlp.
4870 * The contents of the PTE are only stable until @ptlp is released;
4871 * any further use, if any, must be protected against invalidation
4872 * with MMU notifiers.
4874 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4875 * should be taken for read.
4877 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4878 * it is not a good general-purpose API.
4880 * Return: zero on success, -ve otherwise.
4882 int follow_pte(struct mm_struct *mm, unsigned long address,
4883 pte_t **ptepp, spinlock_t **ptlp)
4885 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4887 EXPORT_SYMBOL_GPL(follow_pte);
4890 * follow_pfn - look up PFN at a user virtual address
4891 * @vma: memory mapping
4892 * @address: user virtual address
4893 * @pfn: location to store found PFN
4895 * Only IO mappings and raw PFN mappings are allowed.
4897 * This function does not allow the caller to read the permissions
4898 * of the PTE. Do not use it.
4900 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4902 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4909 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4912 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4915 *pfn = pte_pfn(*ptep);
4916 pte_unmap_unlock(ptep, ptl);
4919 EXPORT_SYMBOL(follow_pfn);
4921 #ifdef CONFIG_HAVE_IOREMAP_PROT
4922 int follow_phys(struct vm_area_struct *vma,
4923 unsigned long address, unsigned int flags,
4924 unsigned long *prot, resource_size_t *phys)
4930 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4933 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4937 /* Never return PFNs of anon folios in COW mappings. */
4938 if (vm_normal_page(vma, address, pte))
4941 if ((flags & FOLL_WRITE) && !pte_write(pte))
4944 *prot = pgprot_val(pte_pgprot(pte));
4945 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4949 pte_unmap_unlock(ptep, ptl);
4954 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4955 void *buf, int len, int write)
4957 resource_size_t phys_addr;
4958 unsigned long prot = 0;
4959 void __iomem *maddr;
4960 int offset = addr & (PAGE_SIZE-1);
4962 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4965 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4970 memcpy_toio(maddr + offset, buf, len);
4972 memcpy_fromio(buf, maddr + offset, len);
4977 EXPORT_SYMBOL_GPL(generic_access_phys);
4981 * Access another process' address space as given in mm. If non-NULL, use the
4982 * given task for page fault accounting.
4984 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4985 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4987 struct vm_area_struct *vma;
4988 void *old_buf = buf;
4989 int write = gup_flags & FOLL_WRITE;
4991 if (mmap_read_lock_killable(mm))
4994 /* ignore errors, just check how much was successfully transferred */
4996 int bytes, ret, offset;
4998 struct page *page = NULL;
5000 ret = get_user_pages_remote(mm, addr, 1,
5001 gup_flags, &page, &vma, NULL);
5003 #ifndef CONFIG_HAVE_IOREMAP_PROT
5007 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5008 * we can access using slightly different code.
5010 vma = find_vma(mm, addr);
5011 if (!vma || vma->vm_start > addr)
5013 if (vma->vm_ops && vma->vm_ops->access)
5014 ret = vma->vm_ops->access(vma, addr, buf,
5022 offset = addr & (PAGE_SIZE-1);
5023 if (bytes > PAGE_SIZE-offset)
5024 bytes = PAGE_SIZE-offset;
5028 copy_to_user_page(vma, page, addr,
5029 maddr + offset, buf, bytes);
5030 set_page_dirty_lock(page);
5032 copy_from_user_page(vma, page, addr,
5033 buf, maddr + offset, bytes);
5042 mmap_read_unlock(mm);
5044 return buf - old_buf;
5048 * access_remote_vm - access another process' address space
5049 * @mm: the mm_struct of the target address space
5050 * @addr: start address to access
5051 * @buf: source or destination buffer
5052 * @len: number of bytes to transfer
5053 * @gup_flags: flags modifying lookup behaviour
5055 * The caller must hold a reference on @mm.
5057 * Return: number of bytes copied from source to destination.
5059 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5060 void *buf, int len, unsigned int gup_flags)
5062 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
5066 * Access another process' address space.
5067 * Source/target buffer must be kernel space,
5068 * Do not walk the page table directly, use get_user_pages
5070 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5071 void *buf, int len, unsigned int gup_flags)
5073 struct mm_struct *mm;
5076 mm = get_task_mm(tsk);
5080 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
5086 EXPORT_SYMBOL_GPL(access_process_vm);
5089 * Print the name of a VMA.
5091 void print_vma_addr(char *prefix, unsigned long ip)
5093 struct mm_struct *mm = current->mm;
5094 struct vm_area_struct *vma;
5097 * we might be running from an atomic context so we cannot sleep
5099 if (!mmap_read_trylock(mm))
5102 vma = find_vma(mm, ip);
5103 if (vma && vma->vm_file) {
5104 struct file *f = vma->vm_file;
5105 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5109 p = file_path(f, buf, PAGE_SIZE);
5112 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5114 vma->vm_end - vma->vm_start);
5115 free_page((unsigned long)buf);
5118 mmap_read_unlock(mm);
5121 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5122 void __might_fault(const char *file, int line)
5125 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5126 * holding the mmap_lock, this is safe because kernel memory doesn't
5127 * get paged out, therefore we'll never actually fault, and the
5128 * below annotations will generate false positives.
5130 if (uaccess_kernel())
5132 if (pagefault_disabled())
5134 __might_sleep(file, line, 0);
5135 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5137 might_lock_read(¤t->mm->mmap_lock);
5140 EXPORT_SYMBOL(__might_fault);
5143 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5145 * Process all subpages of the specified huge page with the specified
5146 * operation. The target subpage will be processed last to keep its
5149 static inline void process_huge_page(
5150 unsigned long addr_hint, unsigned int pages_per_huge_page,
5151 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5155 unsigned long addr = addr_hint &
5156 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5158 /* Process target subpage last to keep its cache lines hot */
5160 n = (addr_hint - addr) / PAGE_SIZE;
5161 if (2 * n <= pages_per_huge_page) {
5162 /* If target subpage in first half of huge page */
5165 /* Process subpages at the end of huge page */
5166 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5168 process_subpage(addr + i * PAGE_SIZE, i, arg);
5171 /* If target subpage in second half of huge page */
5172 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5173 l = pages_per_huge_page - n;
5174 /* Process subpages at the begin of huge page */
5175 for (i = 0; i < base; i++) {
5177 process_subpage(addr + i * PAGE_SIZE, i, arg);
5181 * Process remaining subpages in left-right-left-right pattern
5182 * towards the target subpage
5184 for (i = 0; i < l; i++) {
5185 int left_idx = base + i;
5186 int right_idx = base + 2 * l - 1 - i;
5189 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5191 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5195 static void clear_gigantic_page(struct page *page,
5197 unsigned int pages_per_huge_page)
5200 struct page *p = page;
5203 for (i = 0; i < pages_per_huge_page;
5204 i++, p = mem_map_next(p, page, i)) {
5206 clear_user_highpage(p, addr + i * PAGE_SIZE);
5210 static void clear_subpage(unsigned long addr, int idx, void *arg)
5212 struct page *page = arg;
5214 clear_user_highpage(page + idx, addr);
5217 void clear_huge_page(struct page *page,
5218 unsigned long addr_hint, unsigned int pages_per_huge_page)
5220 unsigned long addr = addr_hint &
5221 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5223 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5224 clear_gigantic_page(page, addr, pages_per_huge_page);
5228 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5231 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5233 struct vm_area_struct *vma,
5234 unsigned int pages_per_huge_page)
5237 struct page *dst_base = dst;
5238 struct page *src_base = src;
5240 for (i = 0; i < pages_per_huge_page; ) {
5242 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5245 dst = mem_map_next(dst, dst_base, i);
5246 src = mem_map_next(src, src_base, i);
5250 struct copy_subpage_arg {
5253 struct vm_area_struct *vma;
5256 static void copy_subpage(unsigned long addr, int idx, void *arg)
5258 struct copy_subpage_arg *copy_arg = arg;
5260 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5261 addr, copy_arg->vma);
5264 void copy_user_huge_page(struct page *dst, struct page *src,
5265 unsigned long addr_hint, struct vm_area_struct *vma,
5266 unsigned int pages_per_huge_page)
5268 unsigned long addr = addr_hint &
5269 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5270 struct copy_subpage_arg arg = {
5276 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5277 copy_user_gigantic_page(dst, src, addr, vma,
5278 pages_per_huge_page);
5282 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5285 long copy_huge_page_from_user(struct page *dst_page,
5286 const void __user *usr_src,
5287 unsigned int pages_per_huge_page,
5288 bool allow_pagefault)
5290 void *src = (void *)usr_src;
5292 unsigned long i, rc = 0;
5293 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5294 struct page *subpage = dst_page;
5296 for (i = 0; i < pages_per_huge_page;
5297 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5298 if (allow_pagefault)
5299 page_kaddr = kmap(subpage);
5301 page_kaddr = kmap_atomic(subpage);
5302 rc = copy_from_user(page_kaddr,
5303 (const void __user *)(src + i * PAGE_SIZE),
5305 if (allow_pagefault)
5308 kunmap_atomic(page_kaddr);
5310 ret_val -= (PAGE_SIZE - rc);
5314 flush_dcache_page(subpage);
5320 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5322 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5324 static struct kmem_cache *page_ptl_cachep;
5326 void __init ptlock_cache_init(void)
5328 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5332 bool ptlock_alloc(struct page *page)
5336 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5343 void ptlock_free(struct page *page)
5345 kmem_cache_free(page_ptl_cachep, page->ptl);