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/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
73 #include <linux/numa.h>
74 #include <linux/perf_event.h>
75 #include <linux/ptrace.h>
76 #include <linux/vmalloc.h>
78 #include <trace/events/kmem.h>
81 #include <asm/mmu_context.h>
82 #include <asm/pgalloc.h>
83 #include <linux/uaccess.h>
85 #include <asm/tlbflush.h>
87 #include "pgalloc-track.h"
90 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
91 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
94 #ifndef CONFIG_NEED_MULTIPLE_NODES
95 /* use the per-pgdat data instead for discontigmem - mbligh */
96 unsigned long max_mapnr;
97 EXPORT_SYMBOL(max_mapnr);
100 EXPORT_SYMBOL(mem_map);
104 * A number of key systems in x86 including ioremap() rely on the assumption
105 * that high_memory defines the upper bound on direct map memory, then end
106 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
107 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
111 EXPORT_SYMBOL(high_memory);
114 * Randomize the address space (stacks, mmaps, brk, etc.).
116 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
117 * as ancient (libc5 based) binaries can segfault. )
119 int randomize_va_space __read_mostly =
120 #ifdef CONFIG_COMPAT_BRK
126 #ifndef arch_faults_on_old_pte
127 static inline bool arch_faults_on_old_pte(void)
130 * Those arches which don't have hw access flag feature need to
131 * implement their own helper. By default, "true" means pagefault
132 * will be hit on old pte.
138 static int __init disable_randmaps(char *s)
140 randomize_va_space = 0;
143 __setup("norandmaps", disable_randmaps);
145 unsigned long zero_pfn __read_mostly;
146 EXPORT_SYMBOL(zero_pfn);
148 unsigned long highest_memmap_pfn __read_mostly;
151 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
153 static int __init init_zero_pfn(void)
155 zero_pfn = page_to_pfn(ZERO_PAGE(0));
158 core_initcall(init_zero_pfn);
160 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
162 trace_rss_stat(mm, member, count);
165 #if defined(SPLIT_RSS_COUNTING)
167 void sync_mm_rss(struct mm_struct *mm)
171 for (i = 0; i < NR_MM_COUNTERS; i++) {
172 if (current->rss_stat.count[i]) {
173 add_mm_counter(mm, i, current->rss_stat.count[i]);
174 current->rss_stat.count[i] = 0;
177 current->rss_stat.events = 0;
180 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
182 struct task_struct *task = current;
184 if (likely(task->mm == mm))
185 task->rss_stat.count[member] += val;
187 add_mm_counter(mm, member, val);
189 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
190 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
192 /* sync counter once per 64 page faults */
193 #define TASK_RSS_EVENTS_THRESH (64)
194 static void check_sync_rss_stat(struct task_struct *task)
196 if (unlikely(task != current))
198 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
199 sync_mm_rss(task->mm);
201 #else /* SPLIT_RSS_COUNTING */
203 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
204 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
206 static void check_sync_rss_stat(struct task_struct *task)
210 #endif /* SPLIT_RSS_COUNTING */
213 * Note: this doesn't free the actual pages themselves. That
214 * has been handled earlier when unmapping all the memory regions.
216 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
219 pgtable_t token = pmd_pgtable(*pmd);
221 pte_free_tlb(tlb, token, addr);
222 mm_dec_nr_ptes(tlb->mm);
225 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
226 unsigned long addr, unsigned long end,
227 unsigned long floor, unsigned long ceiling)
234 pmd = pmd_offset(pud, addr);
236 next = pmd_addr_end(addr, end);
237 if (pmd_none_or_clear_bad(pmd))
239 free_pte_range(tlb, pmd, addr);
240 } while (pmd++, addr = next, addr != end);
250 if (end - 1 > ceiling - 1)
253 pmd = pmd_offset(pud, start);
255 pmd_free_tlb(tlb, pmd, start);
256 mm_dec_nr_pmds(tlb->mm);
259 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
260 unsigned long addr, unsigned long end,
261 unsigned long floor, unsigned long ceiling)
268 pud = pud_offset(p4d, addr);
270 next = pud_addr_end(addr, end);
271 if (pud_none_or_clear_bad(pud))
273 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
274 } while (pud++, addr = next, addr != end);
284 if (end - 1 > ceiling - 1)
287 pud = pud_offset(p4d, start);
289 pud_free_tlb(tlb, pud, start);
290 mm_dec_nr_puds(tlb->mm);
293 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
294 unsigned long addr, unsigned long end,
295 unsigned long floor, unsigned long ceiling)
302 p4d = p4d_offset(pgd, addr);
304 next = p4d_addr_end(addr, end);
305 if (p4d_none_or_clear_bad(p4d))
307 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
308 } while (p4d++, addr = next, addr != end);
314 ceiling &= PGDIR_MASK;
318 if (end - 1 > ceiling - 1)
321 p4d = p4d_offset(pgd, start);
323 p4d_free_tlb(tlb, p4d, start);
327 * This function frees user-level page tables of a process.
329 void free_pgd_range(struct mmu_gather *tlb,
330 unsigned long addr, unsigned long end,
331 unsigned long floor, unsigned long ceiling)
337 * The next few lines have given us lots of grief...
339 * Why are we testing PMD* at this top level? Because often
340 * there will be no work to do at all, and we'd prefer not to
341 * go all the way down to the bottom just to discover that.
343 * Why all these "- 1"s? Because 0 represents both the bottom
344 * of the address space and the top of it (using -1 for the
345 * top wouldn't help much: the masks would do the wrong thing).
346 * The rule is that addr 0 and floor 0 refer to the bottom of
347 * the address space, but end 0 and ceiling 0 refer to the top
348 * Comparisons need to use "end - 1" and "ceiling - 1" (though
349 * that end 0 case should be mythical).
351 * Wherever addr is brought up or ceiling brought down, we must
352 * be careful to reject "the opposite 0" before it confuses the
353 * subsequent tests. But what about where end is brought down
354 * by PMD_SIZE below? no, end can't go down to 0 there.
356 * Whereas we round start (addr) and ceiling down, by different
357 * masks at different levels, in order to test whether a table
358 * now has no other vmas using it, so can be freed, we don't
359 * bother to round floor or end up - the tests don't need that.
373 if (end - 1 > ceiling - 1)
378 * We add page table cache pages with PAGE_SIZE,
379 * (see pte_free_tlb()), flush the tlb if we need
381 tlb_change_page_size(tlb, PAGE_SIZE);
382 pgd = pgd_offset(tlb->mm, addr);
384 next = pgd_addr_end(addr, end);
385 if (pgd_none_or_clear_bad(pgd))
387 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
388 } while (pgd++, addr = next, addr != end);
391 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
392 unsigned long floor, unsigned long ceiling)
395 struct vm_area_struct *next = vma->vm_next;
396 unsigned long addr = vma->vm_start;
399 * Hide vma from rmap and truncate_pagecache before freeing
402 unlink_anon_vmas(vma);
403 unlink_file_vma(vma);
405 if (is_vm_hugetlb_page(vma)) {
406 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
407 floor, next ? next->vm_start : ceiling);
410 * Optimization: gather nearby vmas into one call down
412 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
413 && !is_vm_hugetlb_page(next)) {
416 unlink_anon_vmas(vma);
417 unlink_file_vma(vma);
419 free_pgd_range(tlb, addr, vma->vm_end,
420 floor, next ? next->vm_start : ceiling);
426 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
429 pgtable_t new = pte_alloc_one(mm);
434 * Ensure all pte setup (eg. pte page lock and page clearing) are
435 * visible before the pte is made visible to other CPUs by being
436 * put into page tables.
438 * The other side of the story is the pointer chasing in the page
439 * table walking code (when walking the page table without locking;
440 * ie. most of the time). Fortunately, these data accesses consist
441 * of a chain of data-dependent loads, meaning most CPUs (alpha
442 * being the notable exception) will already guarantee loads are
443 * seen in-order. See the alpha page table accessors for the
444 * smp_rmb() barriers in page table walking code.
446 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
448 ptl = pmd_lock(mm, pmd);
449 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
451 pmd_populate(mm, pmd, new);
460 int __pte_alloc_kernel(pmd_t *pmd)
462 pte_t *new = pte_alloc_one_kernel(&init_mm);
466 smp_wmb(); /* See comment in __pte_alloc */
468 spin_lock(&init_mm.page_table_lock);
469 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
470 pmd_populate_kernel(&init_mm, pmd, new);
473 spin_unlock(&init_mm.page_table_lock);
475 pte_free_kernel(&init_mm, new);
479 static inline void init_rss_vec(int *rss)
481 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
484 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
488 if (current->mm == mm)
490 for (i = 0; i < NR_MM_COUNTERS; i++)
492 add_mm_counter(mm, i, rss[i]);
496 * This function is called to print an error when a bad pte
497 * is found. For example, we might have a PFN-mapped pte in
498 * a region that doesn't allow it.
500 * The calling function must still handle the error.
502 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
503 pte_t pte, struct page *page)
505 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
506 p4d_t *p4d = p4d_offset(pgd, addr);
507 pud_t *pud = pud_offset(p4d, addr);
508 pmd_t *pmd = pmd_offset(pud, addr);
509 struct address_space *mapping;
511 static unsigned long resume;
512 static unsigned long nr_shown;
513 static unsigned long nr_unshown;
516 * Allow a burst of 60 reports, then keep quiet for that minute;
517 * or allow a steady drip of one report per second.
519 if (nr_shown == 60) {
520 if (time_before(jiffies, resume)) {
525 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
532 resume = jiffies + 60 * HZ;
534 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
535 index = linear_page_index(vma, addr);
537 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
539 (long long)pte_val(pte), (long long)pmd_val(*pmd));
541 dump_page(page, "bad pte");
542 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
543 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
544 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
546 vma->vm_ops ? vma->vm_ops->fault : NULL,
547 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
548 mapping ? mapping->a_ops->readpage : NULL);
550 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
554 * vm_normal_page -- This function gets the "struct page" associated with a pte.
556 * "Special" mappings do not wish to be associated with a "struct page" (either
557 * it doesn't exist, or it exists but they don't want to touch it). In this
558 * case, NULL is returned here. "Normal" mappings do have a struct page.
560 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
561 * pte bit, in which case this function is trivial. Secondly, an architecture
562 * may not have a spare pte bit, which requires a more complicated scheme,
565 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
566 * special mapping (even if there are underlying and valid "struct pages").
567 * COWed pages of a VM_PFNMAP are always normal.
569 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
570 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
571 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
572 * mapping will always honor the rule
574 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
576 * And for normal mappings this is false.
578 * This restricts such mappings to be a linear translation from virtual address
579 * to pfn. To get around this restriction, we allow arbitrary mappings so long
580 * as the vma is not a COW mapping; in that case, we know that all ptes are
581 * special (because none can have been COWed).
584 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
586 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
587 * page" backing, however the difference is that _all_ pages with a struct
588 * page (that is, those where pfn_valid is true) are refcounted and considered
589 * normal pages by the VM. The disadvantage is that pages are refcounted
590 * (which can be slower and simply not an option for some PFNMAP users). The
591 * advantage is that we don't have to follow the strict linearity rule of
592 * PFNMAP mappings in order to support COWable mappings.
595 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
598 unsigned long pfn = pte_pfn(pte);
600 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
601 if (likely(!pte_special(pte)))
603 if (vma->vm_ops && vma->vm_ops->find_special_page)
604 return vma->vm_ops->find_special_page(vma, addr);
605 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
607 if (is_zero_pfn(pfn))
612 print_bad_pte(vma, addr, pte, NULL);
616 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
618 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
619 if (vma->vm_flags & VM_MIXEDMAP) {
625 off = (addr - vma->vm_start) >> PAGE_SHIFT;
626 if (pfn == vma->vm_pgoff + off)
628 if (!is_cow_mapping(vma->vm_flags))
633 if (is_zero_pfn(pfn))
637 if (unlikely(pfn > highest_memmap_pfn)) {
638 print_bad_pte(vma, addr, pte, NULL);
643 * NOTE! We still have PageReserved() pages in the page tables.
644 * eg. VDSO mappings can cause them to exist.
647 return pfn_to_page(pfn);
650 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
651 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
654 unsigned long pfn = pmd_pfn(pmd);
657 * There is no pmd_special() but there may be special pmds, e.g.
658 * in a direct-access (dax) mapping, so let's just replicate the
659 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
661 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
662 if (vma->vm_flags & VM_MIXEDMAP) {
668 off = (addr - vma->vm_start) >> PAGE_SHIFT;
669 if (pfn == vma->vm_pgoff + off)
671 if (!is_cow_mapping(vma->vm_flags))
678 if (is_huge_zero_pmd(pmd))
680 if (unlikely(pfn > highest_memmap_pfn))
684 * NOTE! We still have PageReserved() pages in the page tables.
685 * eg. VDSO mappings can cause them to exist.
688 return pfn_to_page(pfn);
693 * copy one vm_area from one task to the other. Assumes the page tables
694 * already present in the new task to be cleared in the whole range
695 * covered by this vma.
699 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
700 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
701 unsigned long addr, int *rss)
703 unsigned long vm_flags = vma->vm_flags;
704 pte_t pte = *src_pte;
706 swp_entry_t entry = pte_to_swp_entry(pte);
708 if (likely(!non_swap_entry(entry))) {
709 if (swap_duplicate(entry) < 0)
712 /* make sure dst_mm is on swapoff's mmlist. */
713 if (unlikely(list_empty(&dst_mm->mmlist))) {
714 spin_lock(&mmlist_lock);
715 if (list_empty(&dst_mm->mmlist))
716 list_add(&dst_mm->mmlist,
718 spin_unlock(&mmlist_lock);
721 } else if (is_migration_entry(entry)) {
722 page = migration_entry_to_page(entry);
724 rss[mm_counter(page)]++;
726 if (is_write_migration_entry(entry) &&
727 is_cow_mapping(vm_flags)) {
729 * COW mappings require pages in both
730 * parent and child to be set to read.
732 make_migration_entry_read(&entry);
733 pte = swp_entry_to_pte(entry);
734 if (pte_swp_soft_dirty(*src_pte))
735 pte = pte_swp_mksoft_dirty(pte);
736 if (pte_swp_uffd_wp(*src_pte))
737 pte = pte_swp_mkuffd_wp(pte);
738 set_pte_at(src_mm, addr, src_pte, pte);
740 } else if (is_device_private_entry(entry)) {
741 page = device_private_entry_to_page(entry);
744 * Update rss count even for unaddressable pages, as
745 * they should treated just like normal pages in this
748 * We will likely want to have some new rss counters
749 * for unaddressable pages, at some point. But for now
750 * keep things as they are.
753 rss[mm_counter(page)]++;
754 page_dup_rmap(page, false);
757 * We do not preserve soft-dirty information, because so
758 * far, checkpoint/restore is the only feature that
759 * requires that. And checkpoint/restore does not work
760 * when a device driver is involved (you cannot easily
761 * save and restore device driver state).
763 if (is_write_device_private_entry(entry) &&
764 is_cow_mapping(vm_flags)) {
765 make_device_private_entry_read(&entry);
766 pte = swp_entry_to_pte(entry);
767 if (pte_swp_uffd_wp(*src_pte))
768 pte = pte_swp_mkuffd_wp(pte);
769 set_pte_at(src_mm, addr, src_pte, pte);
772 set_pte_at(dst_mm, addr, dst_pte, pte);
777 * Copy a present and normal page if necessary.
779 * NOTE! The usual case is that this doesn't need to do
780 * anything, and can just return a positive value. That
781 * will let the caller know that it can just increase
782 * the page refcount and re-use the pte the traditional
785 * But _if_ we need to copy it because it needs to be
786 * pinned in the parent (and the child should get its own
787 * copy rather than just a reference to the same page),
788 * we'll do that here and return zero to let the caller
791 * And if we need a pre-allocated page but don't yet have
792 * one, return a negative error to let the preallocation
793 * code know so that it can do so outside the page table
797 copy_present_page(struct mm_struct *dst_mm, struct mm_struct *src_mm,
798 pte_t *dst_pte, pte_t *src_pte,
799 struct vm_area_struct *vma, struct vm_area_struct *new,
800 unsigned long addr, int *rss, struct page **prealloc,
801 pte_t pte, struct page *page)
803 struct page *new_page;
805 if (!is_cow_mapping(vma->vm_flags))
811 * What we want to do is to check whether this page may
812 * have been pinned by the parent process. If so,
813 * instead of wrprotect the pte on both sides, we copy
814 * the page immediately so that we'll always guarantee
815 * the pinned page won't be randomly replaced in the
818 * To achieve this, we do the following:
820 * 1. Write-protect the pte if it's writable. This is
821 * to protect concurrent write fast-gup with
822 * FOLL_PIN, so that we'll fail the fast-gup with
823 * the write bit removed.
825 * 2. Check page_maybe_dma_pinned() to see whether this
826 * page may have been pinned.
828 * The order of these steps is important to serialize
829 * against the fast-gup code (gup_pte_range()) on the
830 * pte check and try_grab_compound_head(), so that
831 * we'll make sure either we'll capture that fast-gup
832 * so we'll copy the pinned page here, or we'll fail
835 * NOTE! Even if we don't end up copying the page,
836 * we won't undo this wrprotect(), because the normal
837 * reference copy will need it anyway.
840 ptep_set_wrprotect(src_mm, addr, src_pte);
843 * These are the "normally we can just copy by reference"
846 if (likely(!atomic_read(&src_mm->has_pinned)))
848 if (likely(!page_maybe_dma_pinned(page)))
852 * Uhhuh. It looks like the page might be a pinned page,
853 * and we actually need to copy it. Now we can set the
854 * source pte back to being writable.
857 set_pte_at(src_mm, addr, src_pte, pte);
859 new_page = *prealloc;
864 * We have a prealloc page, all good! Take it
865 * over and copy the page & arm it.
868 copy_user_highpage(new_page, page, addr, vma);
869 __SetPageUptodate(new_page);
870 page_add_new_anon_rmap(new_page, new, addr, false);
871 lru_cache_add_inactive_or_unevictable(new_page, new);
872 rss[mm_counter(new_page)]++;
874 /* All done, just insert the new page copy in the child */
875 pte = mk_pte(new_page, new->vm_page_prot);
876 pte = maybe_mkwrite(pte_mkdirty(pte), new);
877 set_pte_at(dst_mm, addr, dst_pte, pte);
882 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
883 * is required to copy this pte.
886 copy_present_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
887 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
888 struct vm_area_struct *new,
889 unsigned long addr, int *rss, struct page **prealloc)
891 unsigned long vm_flags = vma->vm_flags;
892 pte_t pte = *src_pte;
895 page = vm_normal_page(vma, addr, pte);
899 retval = copy_present_page(dst_mm, src_mm,
908 page_dup_rmap(page, false);
909 rss[mm_counter(page)]++;
913 * If it's a COW mapping, write protect it both
914 * in the parent and the child
916 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
917 ptep_set_wrprotect(src_mm, addr, src_pte);
918 pte = pte_wrprotect(pte);
922 * If it's a shared mapping, mark it clean in
925 if (vm_flags & VM_SHARED)
926 pte = pte_mkclean(pte);
927 pte = pte_mkold(pte);
930 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
931 * does not have the VM_UFFD_WP, which means that the uffd
932 * fork event is not enabled.
934 if (!(vm_flags & VM_UFFD_WP))
935 pte = pte_clear_uffd_wp(pte);
937 set_pte_at(dst_mm, addr, dst_pte, pte);
941 static inline struct page *
942 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
945 struct page *new_page;
947 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
951 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
955 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
960 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
961 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
962 struct vm_area_struct *new,
963 unsigned long addr, unsigned long end)
965 pte_t *orig_src_pte, *orig_dst_pte;
966 pte_t *src_pte, *dst_pte;
967 spinlock_t *src_ptl, *dst_ptl;
968 int progress, ret = 0;
969 int rss[NR_MM_COUNTERS];
970 swp_entry_t entry = (swp_entry_t){0};
971 struct page *prealloc = NULL;
977 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
982 src_pte = pte_offset_map(src_pmd, addr);
983 src_ptl = pte_lockptr(src_mm, src_pmd);
984 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
985 orig_src_pte = src_pte;
986 orig_dst_pte = dst_pte;
987 arch_enter_lazy_mmu_mode();
991 * We are holding two locks at this point - either of them
992 * could generate latencies in another task on another CPU.
994 if (progress >= 32) {
996 if (need_resched() ||
997 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1000 if (pte_none(*src_pte)) {
1004 if (unlikely(!pte_present(*src_pte))) {
1005 entry.val = copy_nonpresent_pte(dst_mm, src_mm,
1013 /* copy_present_pte() will clear `*prealloc' if consumed */
1014 ret = copy_present_pte(dst_mm, src_mm, dst_pte, src_pte,
1015 vma, new, addr, rss, &prealloc);
1017 * If we need a pre-allocated page for this pte, drop the
1018 * locks, allocate, and try again.
1020 if (unlikely(ret == -EAGAIN))
1022 if (unlikely(prealloc)) {
1024 * pre-alloc page cannot be reused by next time so as
1025 * to strictly follow mempolicy (e.g., alloc_page_vma()
1026 * will allocate page according to address). This
1027 * could only happen if one pinned pte changed.
1033 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1035 arch_leave_lazy_mmu_mode();
1036 spin_unlock(src_ptl);
1037 pte_unmap(orig_src_pte);
1038 add_mm_rss_vec(dst_mm, rss);
1039 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1043 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1049 WARN_ON_ONCE(ret != -EAGAIN);
1050 prealloc = page_copy_prealloc(src_mm, vma, addr);
1053 /* We've captured and resolved the error. Reset, try again. */
1059 if (unlikely(prealloc))
1064 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1065 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1066 struct vm_area_struct *new,
1067 unsigned long addr, unsigned long end)
1069 pmd_t *src_pmd, *dst_pmd;
1072 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1075 src_pmd = pmd_offset(src_pud, addr);
1077 next = pmd_addr_end(addr, end);
1078 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1079 || pmd_devmap(*src_pmd)) {
1081 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1082 err = copy_huge_pmd(dst_mm, src_mm,
1083 dst_pmd, src_pmd, addr, vma);
1090 if (pmd_none_or_clear_bad(src_pmd))
1092 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1093 vma, new, addr, next))
1095 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1099 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1100 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
1101 struct vm_area_struct *new,
1102 unsigned long addr, unsigned long end)
1104 pud_t *src_pud, *dst_pud;
1107 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1110 src_pud = pud_offset(src_p4d, addr);
1112 next = pud_addr_end(addr, end);
1113 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1116 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1117 err = copy_huge_pud(dst_mm, src_mm,
1118 dst_pud, src_pud, addr, vma);
1125 if (pud_none_or_clear_bad(src_pud))
1127 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1128 vma, new, addr, next))
1130 } while (dst_pud++, src_pud++, addr = next, addr != end);
1134 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1135 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1136 struct vm_area_struct *new,
1137 unsigned long addr, unsigned long end)
1139 p4d_t *src_p4d, *dst_p4d;
1142 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1145 src_p4d = p4d_offset(src_pgd, addr);
1147 next = p4d_addr_end(addr, end);
1148 if (p4d_none_or_clear_bad(src_p4d))
1150 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
1151 vma, new, addr, next))
1153 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1157 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1158 struct vm_area_struct *vma, struct vm_area_struct *new)
1160 pgd_t *src_pgd, *dst_pgd;
1162 unsigned long addr = vma->vm_start;
1163 unsigned long end = vma->vm_end;
1164 struct mmu_notifier_range range;
1169 * Don't copy ptes where a page fault will fill them correctly.
1170 * Fork becomes much lighter when there are big shared or private
1171 * readonly mappings. The tradeoff is that copy_page_range is more
1172 * efficient than faulting.
1174 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1178 if (is_vm_hugetlb_page(vma))
1179 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1181 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1183 * We do not free on error cases below as remove_vma
1184 * gets called on error from higher level routine
1186 ret = track_pfn_copy(vma);
1192 * We need to invalidate the secondary MMU mappings only when
1193 * there could be a permission downgrade on the ptes of the
1194 * parent mm. And a permission downgrade will only happen if
1195 * is_cow_mapping() returns true.
1197 is_cow = is_cow_mapping(vma->vm_flags);
1200 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1201 0, vma, src_mm, addr, end);
1202 mmu_notifier_invalidate_range_start(&range);
1206 dst_pgd = pgd_offset(dst_mm, addr);
1207 src_pgd = pgd_offset(src_mm, addr);
1209 next = pgd_addr_end(addr, end);
1210 if (pgd_none_or_clear_bad(src_pgd))
1212 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1213 vma, new, addr, next))) {
1217 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1220 mmu_notifier_invalidate_range_end(&range);
1224 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pmd_t *pmd,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1229 struct mm_struct *mm = tlb->mm;
1230 int force_flush = 0;
1231 int rss[NR_MM_COUNTERS];
1237 tlb_change_page_size(tlb, PAGE_SIZE);
1240 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1242 flush_tlb_batched_pending(mm);
1243 arch_enter_lazy_mmu_mode();
1246 if (pte_none(ptent))
1252 if (pte_present(ptent)) {
1255 page = vm_normal_page(vma, addr, ptent);
1256 if (unlikely(details) && page) {
1258 * unmap_shared_mapping_pages() wants to
1259 * invalidate cache without truncating:
1260 * unmap shared but keep private pages.
1262 if (details->check_mapping &&
1263 details->check_mapping != page_rmapping(page))
1266 ptent = ptep_get_and_clear_full(mm, addr, pte,
1268 tlb_remove_tlb_entry(tlb, pte, addr);
1269 if (unlikely(!page))
1272 if (!PageAnon(page)) {
1273 if (pte_dirty(ptent)) {
1275 set_page_dirty(page);
1277 if (pte_young(ptent) &&
1278 likely(!(vma->vm_flags & VM_SEQ_READ)))
1279 mark_page_accessed(page);
1281 rss[mm_counter(page)]--;
1282 page_remove_rmap(page, false);
1283 if (unlikely(page_mapcount(page) < 0))
1284 print_bad_pte(vma, addr, ptent, page);
1285 if (unlikely(__tlb_remove_page(tlb, page))) {
1293 entry = pte_to_swp_entry(ptent);
1294 if (is_device_private_entry(entry)) {
1295 struct page *page = device_private_entry_to_page(entry);
1297 if (unlikely(details && details->check_mapping)) {
1299 * unmap_shared_mapping_pages() wants to
1300 * invalidate cache without truncating:
1301 * unmap shared but keep private pages.
1303 if (details->check_mapping !=
1304 page_rmapping(page))
1308 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1309 rss[mm_counter(page)]--;
1310 page_remove_rmap(page, false);
1315 /* If details->check_mapping, we leave swap entries. */
1316 if (unlikely(details))
1319 if (!non_swap_entry(entry))
1321 else if (is_migration_entry(entry)) {
1324 page = migration_entry_to_page(entry);
1325 rss[mm_counter(page)]--;
1327 if (unlikely(!free_swap_and_cache(entry)))
1328 print_bad_pte(vma, addr, ptent, NULL);
1329 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1330 } while (pte++, addr += PAGE_SIZE, addr != end);
1332 add_mm_rss_vec(mm, rss);
1333 arch_leave_lazy_mmu_mode();
1335 /* Do the actual TLB flush before dropping ptl */
1337 tlb_flush_mmu_tlbonly(tlb);
1338 pte_unmap_unlock(start_pte, ptl);
1341 * If we forced a TLB flush (either due to running out of
1342 * batch buffers or because we needed to flush dirty TLB
1343 * entries before releasing the ptl), free the batched
1344 * memory too. Restart if we didn't do everything.
1359 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1360 struct vm_area_struct *vma, pud_t *pud,
1361 unsigned long addr, unsigned long end,
1362 struct zap_details *details)
1367 pmd = pmd_offset(pud, addr);
1369 next = pmd_addr_end(addr, end);
1370 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1371 if (next - addr != HPAGE_PMD_SIZE)
1372 __split_huge_pmd(vma, pmd, addr, false, NULL);
1373 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1378 * Here there can be other concurrent MADV_DONTNEED or
1379 * trans huge page faults running, and if the pmd is
1380 * none or trans huge it can change under us. This is
1381 * because MADV_DONTNEED holds the mmap_lock in read
1384 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1386 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1389 } while (pmd++, addr = next, addr != end);
1394 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1395 struct vm_area_struct *vma, p4d_t *p4d,
1396 unsigned long addr, unsigned long end,
1397 struct zap_details *details)
1402 pud = pud_offset(p4d, addr);
1404 next = pud_addr_end(addr, end);
1405 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1406 if (next - addr != HPAGE_PUD_SIZE) {
1407 mmap_assert_locked(tlb->mm);
1408 split_huge_pud(vma, pud, addr);
1409 } else if (zap_huge_pud(tlb, vma, pud, addr))
1413 if (pud_none_or_clear_bad(pud))
1415 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1418 } while (pud++, addr = next, addr != end);
1423 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1424 struct vm_area_struct *vma, pgd_t *pgd,
1425 unsigned long addr, unsigned long end,
1426 struct zap_details *details)
1431 p4d = p4d_offset(pgd, addr);
1433 next = p4d_addr_end(addr, end);
1434 if (p4d_none_or_clear_bad(p4d))
1436 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1437 } while (p4d++, addr = next, addr != end);
1442 void unmap_page_range(struct mmu_gather *tlb,
1443 struct vm_area_struct *vma,
1444 unsigned long addr, unsigned long end,
1445 struct zap_details *details)
1450 BUG_ON(addr >= end);
1451 tlb_start_vma(tlb, vma);
1452 pgd = pgd_offset(vma->vm_mm, addr);
1454 next = pgd_addr_end(addr, end);
1455 if (pgd_none_or_clear_bad(pgd))
1457 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1458 } while (pgd++, addr = next, addr != end);
1459 tlb_end_vma(tlb, vma);
1463 static void unmap_single_vma(struct mmu_gather *tlb,
1464 struct vm_area_struct *vma, unsigned long start_addr,
1465 unsigned long end_addr,
1466 struct zap_details *details)
1468 unsigned long start = max(vma->vm_start, start_addr);
1471 if (start >= vma->vm_end)
1473 end = min(vma->vm_end, end_addr);
1474 if (end <= vma->vm_start)
1478 uprobe_munmap(vma, start, end);
1480 if (unlikely(vma->vm_flags & VM_PFNMAP))
1481 untrack_pfn(vma, 0, 0);
1484 if (unlikely(is_vm_hugetlb_page(vma))) {
1486 * It is undesirable to test vma->vm_file as it
1487 * should be non-null for valid hugetlb area.
1488 * However, vm_file will be NULL in the error
1489 * cleanup path of mmap_region. When
1490 * hugetlbfs ->mmap method fails,
1491 * mmap_region() nullifies vma->vm_file
1492 * before calling this function to clean up.
1493 * Since no pte has actually been setup, it is
1494 * safe to do nothing in this case.
1497 i_mmap_lock_write(vma->vm_file->f_mapping);
1498 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1499 i_mmap_unlock_write(vma->vm_file->f_mapping);
1502 unmap_page_range(tlb, vma, start, end, details);
1507 * unmap_vmas - unmap a range of memory covered by a list of vma's
1508 * @tlb: address of the caller's struct mmu_gather
1509 * @vma: the starting vma
1510 * @start_addr: virtual address at which to start unmapping
1511 * @end_addr: virtual address at which to end unmapping
1513 * Unmap all pages in the vma list.
1515 * Only addresses between `start' and `end' will be unmapped.
1517 * The VMA list must be sorted in ascending virtual address order.
1519 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1520 * range after unmap_vmas() returns. So the only responsibility here is to
1521 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1522 * drops the lock and schedules.
1524 void unmap_vmas(struct mmu_gather *tlb,
1525 struct vm_area_struct *vma, unsigned long start_addr,
1526 unsigned long end_addr)
1528 struct mmu_notifier_range range;
1530 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1531 start_addr, end_addr);
1532 mmu_notifier_invalidate_range_start(&range);
1533 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1534 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1535 mmu_notifier_invalidate_range_end(&range);
1539 * zap_page_range - remove user pages in a given range
1540 * @vma: vm_area_struct holding the applicable pages
1541 * @start: starting address of pages to zap
1542 * @size: number of bytes to zap
1544 * Caller must protect the VMA list
1546 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1549 struct mmu_notifier_range range;
1550 struct mmu_gather tlb;
1553 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1554 start, start + size);
1555 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1556 update_hiwater_rss(vma->vm_mm);
1557 mmu_notifier_invalidate_range_start(&range);
1558 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1559 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1560 mmu_notifier_invalidate_range_end(&range);
1561 tlb_finish_mmu(&tlb, start, range.end);
1565 * zap_page_range_single - remove user pages in a given range
1566 * @vma: vm_area_struct holding the applicable pages
1567 * @address: starting address of pages to zap
1568 * @size: number of bytes to zap
1569 * @details: details of shared cache invalidation
1571 * The range must fit into one VMA.
1573 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1574 unsigned long size, struct zap_details *details)
1576 struct mmu_notifier_range range;
1577 struct mmu_gather tlb;
1580 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1581 address, address + size);
1582 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1583 update_hiwater_rss(vma->vm_mm);
1584 mmu_notifier_invalidate_range_start(&range);
1585 unmap_single_vma(&tlb, vma, address, range.end, details);
1586 mmu_notifier_invalidate_range_end(&range);
1587 tlb_finish_mmu(&tlb, address, range.end);
1591 * zap_vma_ptes - remove ptes mapping the vma
1592 * @vma: vm_area_struct holding ptes to be zapped
1593 * @address: starting address of pages to zap
1594 * @size: number of bytes to zap
1596 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1598 * The entire address range must be fully contained within the vma.
1601 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1604 if (address < vma->vm_start || address + size > vma->vm_end ||
1605 !(vma->vm_flags & VM_PFNMAP))
1608 zap_page_range_single(vma, address, size, NULL);
1610 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1612 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1619 pgd = pgd_offset(mm, addr);
1620 p4d = p4d_alloc(mm, pgd, addr);
1623 pud = pud_alloc(mm, p4d, addr);
1626 pmd = pmd_alloc(mm, pud, addr);
1630 VM_BUG_ON(pmd_trans_huge(*pmd));
1634 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1637 pmd_t *pmd = walk_to_pmd(mm, addr);
1641 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1644 static int validate_page_before_insert(struct page *page)
1646 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1648 flush_dcache_page(page);
1652 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1653 unsigned long addr, struct page *page, pgprot_t prot)
1655 if (!pte_none(*pte))
1657 /* Ok, finally just insert the thing.. */
1659 inc_mm_counter_fast(mm, mm_counter_file(page));
1660 page_add_file_rmap(page, false);
1661 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1666 * This is the old fallback for page remapping.
1668 * For historical reasons, it only allows reserved pages. Only
1669 * old drivers should use this, and they needed to mark their
1670 * pages reserved for the old functions anyway.
1672 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1673 struct page *page, pgprot_t prot)
1675 struct mm_struct *mm = vma->vm_mm;
1680 retval = validate_page_before_insert(page);
1684 pte = get_locked_pte(mm, addr, &ptl);
1687 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1688 pte_unmap_unlock(pte, ptl);
1694 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1695 unsigned long addr, struct page *page, pgprot_t prot)
1699 if (!page_count(page))
1701 err = validate_page_before_insert(page);
1704 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1707 /* insert_pages() amortizes the cost of spinlock operations
1708 * when inserting pages in a loop. Arch *must* define pte_index.
1710 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1711 struct page **pages, unsigned long *num, pgprot_t prot)
1714 pte_t *start_pte, *pte;
1715 spinlock_t *pte_lock;
1716 struct mm_struct *const mm = vma->vm_mm;
1717 unsigned long curr_page_idx = 0;
1718 unsigned long remaining_pages_total = *num;
1719 unsigned long pages_to_write_in_pmd;
1723 pmd = walk_to_pmd(mm, addr);
1727 pages_to_write_in_pmd = min_t(unsigned long,
1728 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1730 /* Allocate the PTE if necessary; takes PMD lock once only. */
1732 if (pte_alloc(mm, pmd))
1735 while (pages_to_write_in_pmd) {
1737 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1739 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1740 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1741 int err = insert_page_in_batch_locked(mm, pte,
1742 addr, pages[curr_page_idx], prot);
1743 if (unlikely(err)) {
1744 pte_unmap_unlock(start_pte, pte_lock);
1746 remaining_pages_total -= pte_idx;
1752 pte_unmap_unlock(start_pte, pte_lock);
1753 pages_to_write_in_pmd -= batch_size;
1754 remaining_pages_total -= batch_size;
1756 if (remaining_pages_total)
1760 *num = remaining_pages_total;
1763 #endif /* ifdef pte_index */
1766 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1767 * @vma: user vma to map to
1768 * @addr: target start user address of these pages
1769 * @pages: source kernel pages
1770 * @num: in: number of pages to map. out: number of pages that were *not*
1771 * mapped. (0 means all pages were successfully mapped).
1773 * Preferred over vm_insert_page() when inserting multiple pages.
1775 * In case of error, we may have mapped a subset of the provided
1776 * pages. It is the caller's responsibility to account for this case.
1778 * The same restrictions apply as in vm_insert_page().
1780 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1781 struct page **pages, unsigned long *num)
1784 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1786 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1788 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1789 BUG_ON(mmap_read_trylock(vma->vm_mm));
1790 BUG_ON(vma->vm_flags & VM_PFNMAP);
1791 vma->vm_flags |= VM_MIXEDMAP;
1793 /* Defer page refcount checking till we're about to map that page. */
1794 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1796 unsigned long idx = 0, pgcount = *num;
1799 for (; idx < pgcount; ++idx) {
1800 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1804 *num = pgcount - idx;
1806 #endif /* ifdef pte_index */
1808 EXPORT_SYMBOL(vm_insert_pages);
1811 * vm_insert_page - insert single page into user vma
1812 * @vma: user vma to map to
1813 * @addr: target user address of this page
1814 * @page: source kernel page
1816 * This allows drivers to insert individual pages they've allocated
1819 * The page has to be a nice clean _individual_ kernel allocation.
1820 * If you allocate a compound page, you need to have marked it as
1821 * such (__GFP_COMP), or manually just split the page up yourself
1822 * (see split_page()).
1824 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1825 * took an arbitrary page protection parameter. This doesn't allow
1826 * that. Your vma protection will have to be set up correctly, which
1827 * means that if you want a shared writable mapping, you'd better
1828 * ask for a shared writable mapping!
1830 * The page does not need to be reserved.
1832 * Usually this function is called from f_op->mmap() handler
1833 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1834 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1835 * function from other places, for example from page-fault handler.
1837 * Return: %0 on success, negative error code otherwise.
1839 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1842 if (addr < vma->vm_start || addr >= vma->vm_end)
1844 if (!page_count(page))
1846 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1847 BUG_ON(mmap_read_trylock(vma->vm_mm));
1848 BUG_ON(vma->vm_flags & VM_PFNMAP);
1849 vma->vm_flags |= VM_MIXEDMAP;
1851 return insert_page(vma, addr, page, vma->vm_page_prot);
1853 EXPORT_SYMBOL(vm_insert_page);
1856 * __vm_map_pages - maps range of kernel pages into user vma
1857 * @vma: user vma to map to
1858 * @pages: pointer to array of source kernel pages
1859 * @num: number of pages in page array
1860 * @offset: user's requested vm_pgoff
1862 * This allows drivers to map range of kernel pages into a user vma.
1864 * Return: 0 on success and error code otherwise.
1866 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1867 unsigned long num, unsigned long offset)
1869 unsigned long count = vma_pages(vma);
1870 unsigned long uaddr = vma->vm_start;
1873 /* Fail if the user requested offset is beyond the end of the object */
1877 /* Fail if the user requested size exceeds available object size */
1878 if (count > num - offset)
1881 for (i = 0; i < count; i++) {
1882 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1892 * vm_map_pages - maps range of kernel pages starts with non zero offset
1893 * @vma: user vma to map to
1894 * @pages: pointer to array of source kernel pages
1895 * @num: number of pages in page array
1897 * Maps an object consisting of @num pages, catering for the user's
1898 * requested vm_pgoff
1900 * If we fail to insert any page into the vma, the function will return
1901 * immediately leaving any previously inserted pages present. Callers
1902 * from the mmap handler may immediately return the error as their caller
1903 * will destroy the vma, removing any successfully inserted pages. Other
1904 * callers should make their own arrangements for calling unmap_region().
1906 * Context: Process context. Called by mmap handlers.
1907 * Return: 0 on success and error code otherwise.
1909 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1912 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1914 EXPORT_SYMBOL(vm_map_pages);
1917 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1918 * @vma: user vma to map to
1919 * @pages: pointer to array of source kernel pages
1920 * @num: number of pages in page array
1922 * Similar to vm_map_pages(), except that it explicitly sets the offset
1923 * to 0. This function is intended for the drivers that did not consider
1926 * Context: Process context. Called by mmap handlers.
1927 * Return: 0 on success and error code otherwise.
1929 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1932 return __vm_map_pages(vma, pages, num, 0);
1934 EXPORT_SYMBOL(vm_map_pages_zero);
1936 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1937 pfn_t pfn, pgprot_t prot, bool mkwrite)
1939 struct mm_struct *mm = vma->vm_mm;
1943 pte = get_locked_pte(mm, addr, &ptl);
1945 return VM_FAULT_OOM;
1946 if (!pte_none(*pte)) {
1949 * For read faults on private mappings the PFN passed
1950 * in may not match the PFN we have mapped if the
1951 * mapped PFN is a writeable COW page. In the mkwrite
1952 * case we are creating a writable PTE for a shared
1953 * mapping and we expect the PFNs to match. If they
1954 * don't match, we are likely racing with block
1955 * allocation and mapping invalidation so just skip the
1958 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1959 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1962 entry = pte_mkyoung(*pte);
1963 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1964 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1965 update_mmu_cache(vma, addr, pte);
1970 /* Ok, finally just insert the thing.. */
1971 if (pfn_t_devmap(pfn))
1972 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1974 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1977 entry = pte_mkyoung(entry);
1978 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1981 set_pte_at(mm, addr, pte, entry);
1982 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1985 pte_unmap_unlock(pte, ptl);
1986 return VM_FAULT_NOPAGE;
1990 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1991 * @vma: user vma to map to
1992 * @addr: target user address of this page
1993 * @pfn: source kernel pfn
1994 * @pgprot: pgprot flags for the inserted page
1996 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1997 * to override pgprot on a per-page basis.
1999 * This only makes sense for IO mappings, and it makes no sense for
2000 * COW mappings. In general, using multiple vmas is preferable;
2001 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2004 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2005 * a value of @pgprot different from that of @vma->vm_page_prot.
2007 * Context: Process context. May allocate using %GFP_KERNEL.
2008 * Return: vm_fault_t value.
2010 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2011 unsigned long pfn, pgprot_t pgprot)
2014 * Technically, architectures with pte_special can avoid all these
2015 * restrictions (same for remap_pfn_range). However we would like
2016 * consistency in testing and feature parity among all, so we should
2017 * try to keep these invariants in place for everybody.
2019 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2020 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2021 (VM_PFNMAP|VM_MIXEDMAP));
2022 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2023 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2025 if (addr < vma->vm_start || addr >= vma->vm_end)
2026 return VM_FAULT_SIGBUS;
2028 if (!pfn_modify_allowed(pfn, pgprot))
2029 return VM_FAULT_SIGBUS;
2031 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2033 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2036 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2039 * vmf_insert_pfn - insert single pfn into user vma
2040 * @vma: user vma to map to
2041 * @addr: target user address of this page
2042 * @pfn: source kernel pfn
2044 * Similar to vm_insert_page, this allows drivers to insert individual pages
2045 * they've allocated into a user vma. Same comments apply.
2047 * This function should only be called from a vm_ops->fault handler, and
2048 * in that case the handler should return the result of this function.
2050 * vma cannot be a COW mapping.
2052 * As this is called only for pages that do not currently exist, we
2053 * do not need to flush old virtual caches or the TLB.
2055 * Context: Process context. May allocate using %GFP_KERNEL.
2056 * Return: vm_fault_t value.
2058 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2061 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2063 EXPORT_SYMBOL(vmf_insert_pfn);
2065 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2067 /* these checks mirror the abort conditions in vm_normal_page */
2068 if (vma->vm_flags & VM_MIXEDMAP)
2070 if (pfn_t_devmap(pfn))
2072 if (pfn_t_special(pfn))
2074 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2079 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2080 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2085 BUG_ON(!vm_mixed_ok(vma, pfn));
2087 if (addr < vma->vm_start || addr >= vma->vm_end)
2088 return VM_FAULT_SIGBUS;
2090 track_pfn_insert(vma, &pgprot, pfn);
2092 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2093 return VM_FAULT_SIGBUS;
2096 * If we don't have pte special, then we have to use the pfn_valid()
2097 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2098 * refcount the page if pfn_valid is true (hence insert_page rather
2099 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2100 * without pte special, it would there be refcounted as a normal page.
2102 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2103 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2107 * At this point we are committed to insert_page()
2108 * regardless of whether the caller specified flags that
2109 * result in pfn_t_has_page() == false.
2111 page = pfn_to_page(pfn_t_to_pfn(pfn));
2112 err = insert_page(vma, addr, page, pgprot);
2114 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2118 return VM_FAULT_OOM;
2119 if (err < 0 && err != -EBUSY)
2120 return VM_FAULT_SIGBUS;
2122 return VM_FAULT_NOPAGE;
2126 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2127 * @vma: user vma to map to
2128 * @addr: target user address of this page
2129 * @pfn: source kernel pfn
2130 * @pgprot: pgprot flags for the inserted page
2132 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2133 * to override pgprot on a per-page basis.
2135 * Typically this function should be used by drivers to set caching- and
2136 * encryption bits different than those of @vma->vm_page_prot, because
2137 * the caching- or encryption mode may not be known at mmap() time.
2138 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2139 * to set caching and encryption bits for those vmas (except for COW pages).
2140 * This is ensured by core vm only modifying these page table entries using
2141 * functions that don't touch caching- or encryption bits, using pte_modify()
2142 * if needed. (See for example mprotect()).
2143 * Also when new page-table entries are created, this is only done using the
2144 * fault() callback, and never using the value of vma->vm_page_prot,
2145 * except for page-table entries that point to anonymous pages as the result
2148 * Context: Process context. May allocate using %GFP_KERNEL.
2149 * Return: vm_fault_t value.
2151 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2152 pfn_t pfn, pgprot_t pgprot)
2154 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2156 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2158 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2161 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2163 EXPORT_SYMBOL(vmf_insert_mixed);
2166 * If the insertion of PTE failed because someone else already added a
2167 * different entry in the mean time, we treat that as success as we assume
2168 * the same entry was actually inserted.
2170 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2171 unsigned long addr, pfn_t pfn)
2173 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2175 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2178 * maps a range of physical memory into the requested pages. the old
2179 * mappings are removed. any references to nonexistent pages results
2180 * in null mappings (currently treated as "copy-on-access")
2182 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2183 unsigned long addr, unsigned long end,
2184 unsigned long pfn, pgprot_t prot)
2190 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2193 arch_enter_lazy_mmu_mode();
2195 BUG_ON(!pte_none(*pte));
2196 if (!pfn_modify_allowed(pfn, prot)) {
2200 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2202 } while (pte++, addr += PAGE_SIZE, addr != end);
2203 arch_leave_lazy_mmu_mode();
2204 pte_unmap_unlock(pte - 1, ptl);
2208 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2209 unsigned long addr, unsigned long end,
2210 unsigned long pfn, pgprot_t prot)
2216 pfn -= addr >> PAGE_SHIFT;
2217 pmd = pmd_alloc(mm, pud, addr);
2220 VM_BUG_ON(pmd_trans_huge(*pmd));
2222 next = pmd_addr_end(addr, end);
2223 err = remap_pte_range(mm, pmd, addr, next,
2224 pfn + (addr >> PAGE_SHIFT), prot);
2227 } while (pmd++, addr = next, addr != end);
2231 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2232 unsigned long addr, unsigned long end,
2233 unsigned long pfn, pgprot_t prot)
2239 pfn -= addr >> PAGE_SHIFT;
2240 pud = pud_alloc(mm, p4d, addr);
2244 next = pud_addr_end(addr, end);
2245 err = remap_pmd_range(mm, pud, addr, next,
2246 pfn + (addr >> PAGE_SHIFT), prot);
2249 } while (pud++, addr = next, addr != end);
2253 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2254 unsigned long addr, unsigned long end,
2255 unsigned long pfn, pgprot_t prot)
2261 pfn -= addr >> PAGE_SHIFT;
2262 p4d = p4d_alloc(mm, pgd, addr);
2266 next = p4d_addr_end(addr, end);
2267 err = remap_pud_range(mm, p4d, addr, next,
2268 pfn + (addr >> PAGE_SHIFT), prot);
2271 } while (p4d++, addr = next, addr != end);
2276 * remap_pfn_range - remap kernel memory to userspace
2277 * @vma: user vma to map to
2278 * @addr: target page aligned user address to start at
2279 * @pfn: page frame number of kernel physical memory address
2280 * @size: size of mapping area
2281 * @prot: page protection flags for this mapping
2283 * Note: this is only safe if the mm semaphore is held when called.
2285 * Return: %0 on success, negative error code otherwise.
2287 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2288 unsigned long pfn, unsigned long size, pgprot_t prot)
2292 unsigned long end = addr + PAGE_ALIGN(size);
2293 struct mm_struct *mm = vma->vm_mm;
2294 unsigned long remap_pfn = pfn;
2297 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2301 * Physically remapped pages are special. Tell the
2302 * rest of the world about it:
2303 * VM_IO tells people not to look at these pages
2304 * (accesses can have side effects).
2305 * VM_PFNMAP tells the core MM that the base pages are just
2306 * raw PFN mappings, and do not have a "struct page" associated
2309 * Disable vma merging and expanding with mremap().
2311 * Omit vma from core dump, even when VM_IO turned off.
2313 * There's a horrible special case to handle copy-on-write
2314 * behaviour that some programs depend on. We mark the "original"
2315 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2316 * See vm_normal_page() for details.
2318 if (is_cow_mapping(vma->vm_flags)) {
2319 if (addr != vma->vm_start || end != vma->vm_end)
2321 vma->vm_pgoff = pfn;
2324 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2328 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2330 BUG_ON(addr >= end);
2331 pfn -= addr >> PAGE_SHIFT;
2332 pgd = pgd_offset(mm, addr);
2333 flush_cache_range(vma, addr, end);
2335 next = pgd_addr_end(addr, end);
2336 err = remap_p4d_range(mm, pgd, addr, next,
2337 pfn + (addr >> PAGE_SHIFT), prot);
2340 } while (pgd++, addr = next, addr != end);
2343 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2347 EXPORT_SYMBOL(remap_pfn_range);
2350 * vm_iomap_memory - remap memory to userspace
2351 * @vma: user vma to map to
2352 * @start: start of the physical memory to be mapped
2353 * @len: size of area
2355 * This is a simplified io_remap_pfn_range() for common driver use. The
2356 * driver just needs to give us the physical memory range to be mapped,
2357 * we'll figure out the rest from the vma information.
2359 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2360 * whatever write-combining details or similar.
2362 * Return: %0 on success, negative error code otherwise.
2364 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2366 unsigned long vm_len, pfn, pages;
2368 /* Check that the physical memory area passed in looks valid */
2369 if (start + len < start)
2372 * You *really* shouldn't map things that aren't page-aligned,
2373 * but we've historically allowed it because IO memory might
2374 * just have smaller alignment.
2376 len += start & ~PAGE_MASK;
2377 pfn = start >> PAGE_SHIFT;
2378 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2379 if (pfn + pages < pfn)
2382 /* We start the mapping 'vm_pgoff' pages into the area */
2383 if (vma->vm_pgoff > pages)
2385 pfn += vma->vm_pgoff;
2386 pages -= vma->vm_pgoff;
2388 /* Can we fit all of the mapping? */
2389 vm_len = vma->vm_end - vma->vm_start;
2390 if (vm_len >> PAGE_SHIFT > pages)
2393 /* Ok, let it rip */
2394 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2396 EXPORT_SYMBOL(vm_iomap_memory);
2398 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2399 unsigned long addr, unsigned long end,
2400 pte_fn_t fn, void *data, bool create,
2401 pgtbl_mod_mask *mask)
2408 pte = (mm == &init_mm) ?
2409 pte_alloc_kernel_track(pmd, addr, mask) :
2410 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2414 pte = (mm == &init_mm) ?
2415 pte_offset_kernel(pmd, addr) :
2416 pte_offset_map_lock(mm, pmd, addr, &ptl);
2419 BUG_ON(pmd_huge(*pmd));
2421 arch_enter_lazy_mmu_mode();
2424 if (create || !pte_none(*pte)) {
2425 err = fn(pte++, addr, data);
2429 } while (addr += PAGE_SIZE, addr != end);
2430 *mask |= PGTBL_PTE_MODIFIED;
2432 arch_leave_lazy_mmu_mode();
2435 pte_unmap_unlock(pte-1, ptl);
2439 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2440 unsigned long addr, unsigned long end,
2441 pte_fn_t fn, void *data, bool create,
2442 pgtbl_mod_mask *mask)
2448 BUG_ON(pud_huge(*pud));
2451 pmd = pmd_alloc_track(mm, pud, addr, mask);
2455 pmd = pmd_offset(pud, addr);
2458 next = pmd_addr_end(addr, end);
2459 if (create || !pmd_none_or_clear_bad(pmd)) {
2460 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2465 } while (pmd++, addr = next, addr != end);
2469 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2470 unsigned long addr, unsigned long end,
2471 pte_fn_t fn, void *data, bool create,
2472 pgtbl_mod_mask *mask)
2479 pud = pud_alloc_track(mm, p4d, addr, mask);
2483 pud = pud_offset(p4d, addr);
2486 next = pud_addr_end(addr, end);
2487 if (create || !pud_none_or_clear_bad(pud)) {
2488 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2493 } while (pud++, addr = next, addr != end);
2497 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2498 unsigned long addr, unsigned long end,
2499 pte_fn_t fn, void *data, bool create,
2500 pgtbl_mod_mask *mask)
2507 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2511 p4d = p4d_offset(pgd, addr);
2514 next = p4d_addr_end(addr, end);
2515 if (create || !p4d_none_or_clear_bad(p4d)) {
2516 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2521 } while (p4d++, addr = next, addr != end);
2525 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2526 unsigned long size, pte_fn_t fn,
2527 void *data, bool create)
2530 unsigned long start = addr, next;
2531 unsigned long end = addr + size;
2532 pgtbl_mod_mask mask = 0;
2535 if (WARN_ON(addr >= end))
2538 pgd = pgd_offset(mm, addr);
2540 next = pgd_addr_end(addr, end);
2541 if (!create && pgd_none_or_clear_bad(pgd))
2543 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2546 } while (pgd++, addr = next, addr != end);
2548 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2549 arch_sync_kernel_mappings(start, start + size);
2555 * Scan a region of virtual memory, filling in page tables as necessary
2556 * and calling a provided function on each leaf page table.
2558 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2559 unsigned long size, pte_fn_t fn, void *data)
2561 return __apply_to_page_range(mm, addr, size, fn, data, true);
2563 EXPORT_SYMBOL_GPL(apply_to_page_range);
2566 * Scan a region of virtual memory, calling a provided function on
2567 * each leaf page table where it exists.
2569 * Unlike apply_to_page_range, this does _not_ fill in page tables
2570 * where they are absent.
2572 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2573 unsigned long size, pte_fn_t fn, void *data)
2575 return __apply_to_page_range(mm, addr, size, fn, data, false);
2577 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2580 * handle_pte_fault chooses page fault handler according to an entry which was
2581 * read non-atomically. Before making any commitment, on those architectures
2582 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2583 * parts, do_swap_page must check under lock before unmapping the pte and
2584 * proceeding (but do_wp_page is only called after already making such a check;
2585 * and do_anonymous_page can safely check later on).
2587 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2588 pte_t *page_table, pte_t orig_pte)
2591 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2592 if (sizeof(pte_t) > sizeof(unsigned long)) {
2593 spinlock_t *ptl = pte_lockptr(mm, pmd);
2595 same = pte_same(*page_table, orig_pte);
2599 pte_unmap(page_table);
2603 static inline bool cow_user_page(struct page *dst, struct page *src,
2604 struct vm_fault *vmf)
2609 bool locked = false;
2610 struct vm_area_struct *vma = vmf->vma;
2611 struct mm_struct *mm = vma->vm_mm;
2612 unsigned long addr = vmf->address;
2615 copy_user_highpage(dst, src, addr, vma);
2620 * If the source page was a PFN mapping, we don't have
2621 * a "struct page" for it. We do a best-effort copy by
2622 * just copying from the original user address. If that
2623 * fails, we just zero-fill it. Live with it.
2625 kaddr = kmap_atomic(dst);
2626 uaddr = (void __user *)(addr & PAGE_MASK);
2629 * On architectures with software "accessed" bits, we would
2630 * take a double page fault, so mark it accessed here.
2632 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2635 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2637 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2639 * Other thread has already handled the fault
2640 * and update local tlb only
2642 update_mmu_tlb(vma, addr, vmf->pte);
2647 entry = pte_mkyoung(vmf->orig_pte);
2648 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2649 update_mmu_cache(vma, addr, vmf->pte);
2653 * This really shouldn't fail, because the page is there
2654 * in the page tables. But it might just be unreadable,
2655 * in which case we just give up and fill the result with
2658 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2662 /* Re-validate under PTL if the page is still mapped */
2663 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2665 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2666 /* The PTE changed under us, update local tlb */
2667 update_mmu_tlb(vma, addr, vmf->pte);
2673 * The same page can be mapped back since last copy attempt.
2674 * Try to copy again under PTL.
2676 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2678 * Give a warn in case there can be some obscure
2691 pte_unmap_unlock(vmf->pte, vmf->ptl);
2692 kunmap_atomic(kaddr);
2693 flush_dcache_page(dst);
2698 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2700 struct file *vm_file = vma->vm_file;
2703 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2706 * Special mappings (e.g. VDSO) do not have any file so fake
2707 * a default GFP_KERNEL for them.
2713 * Notify the address space that the page is about to become writable so that
2714 * it can prohibit this or wait for the page to get into an appropriate state.
2716 * We do this without the lock held, so that it can sleep if it needs to.
2718 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2721 struct page *page = vmf->page;
2722 unsigned int old_flags = vmf->flags;
2724 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2726 if (vmf->vma->vm_file &&
2727 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2728 return VM_FAULT_SIGBUS;
2730 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2731 /* Restore original flags so that caller is not surprised */
2732 vmf->flags = old_flags;
2733 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2735 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2737 if (!page->mapping) {
2739 return 0; /* retry */
2741 ret |= VM_FAULT_LOCKED;
2743 VM_BUG_ON_PAGE(!PageLocked(page), page);
2748 * Handle dirtying of a page in shared file mapping on a write fault.
2750 * The function expects the page to be locked and unlocks it.
2752 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2754 struct vm_area_struct *vma = vmf->vma;
2755 struct address_space *mapping;
2756 struct page *page = vmf->page;
2758 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2760 dirtied = set_page_dirty(page);
2761 VM_BUG_ON_PAGE(PageAnon(page), page);
2763 * Take a local copy of the address_space - page.mapping may be zeroed
2764 * by truncate after unlock_page(). The address_space itself remains
2765 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2766 * release semantics to prevent the compiler from undoing this copying.
2768 mapping = page_rmapping(page);
2772 file_update_time(vma->vm_file);
2775 * Throttle page dirtying rate down to writeback speed.
2777 * mapping may be NULL here because some device drivers do not
2778 * set page.mapping but still dirty their pages
2780 * Drop the mmap_lock before waiting on IO, if we can. The file
2781 * is pinning the mapping, as per above.
2783 if ((dirtied || page_mkwrite) && mapping) {
2786 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2787 balance_dirty_pages_ratelimited(mapping);
2790 return VM_FAULT_RETRY;
2798 * Handle write page faults for pages that can be reused in the current vma
2800 * This can happen either due to the mapping being with the VM_SHARED flag,
2801 * or due to us being the last reference standing to the page. In either
2802 * case, all we need to do here is to mark the page as writable and update
2803 * any related book-keeping.
2805 static inline void wp_page_reuse(struct vm_fault *vmf)
2806 __releases(vmf->ptl)
2808 struct vm_area_struct *vma = vmf->vma;
2809 struct page *page = vmf->page;
2812 * Clear the pages cpupid information as the existing
2813 * information potentially belongs to a now completely
2814 * unrelated process.
2817 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2819 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2820 entry = pte_mkyoung(vmf->orig_pte);
2821 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2822 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2823 update_mmu_cache(vma, vmf->address, vmf->pte);
2824 pte_unmap_unlock(vmf->pte, vmf->ptl);
2825 count_vm_event(PGREUSE);
2829 * Handle the case of a page which we actually need to copy to a new page.
2831 * Called with mmap_lock locked and the old page referenced, but
2832 * without the ptl held.
2834 * High level logic flow:
2836 * - Allocate a page, copy the content of the old page to the new one.
2837 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2838 * - Take the PTL. If the pte changed, bail out and release the allocated page
2839 * - If the pte is still the way we remember it, update the page table and all
2840 * relevant references. This includes dropping the reference the page-table
2841 * held to the old page, as well as updating the rmap.
2842 * - In any case, unlock the PTL and drop the reference we took to the old page.
2844 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2846 struct vm_area_struct *vma = vmf->vma;
2847 struct mm_struct *mm = vma->vm_mm;
2848 struct page *old_page = vmf->page;
2849 struct page *new_page = NULL;
2851 int page_copied = 0;
2852 struct mmu_notifier_range range;
2854 if (unlikely(anon_vma_prepare(vma)))
2857 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2858 new_page = alloc_zeroed_user_highpage_movable(vma,
2863 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2868 if (!cow_user_page(new_page, old_page, vmf)) {
2870 * COW failed, if the fault was solved by other,
2871 * it's fine. If not, userspace would re-fault on
2872 * the same address and we will handle the fault
2873 * from the second attempt.
2882 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2884 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2886 __SetPageUptodate(new_page);
2888 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2889 vmf->address & PAGE_MASK,
2890 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2891 mmu_notifier_invalidate_range_start(&range);
2894 * Re-check the pte - we dropped the lock
2896 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2897 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2899 if (!PageAnon(old_page)) {
2900 dec_mm_counter_fast(mm,
2901 mm_counter_file(old_page));
2902 inc_mm_counter_fast(mm, MM_ANONPAGES);
2905 inc_mm_counter_fast(mm, MM_ANONPAGES);
2907 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2908 entry = mk_pte(new_page, vma->vm_page_prot);
2909 entry = pte_sw_mkyoung(entry);
2910 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2912 * Clear the pte entry and flush it first, before updating the
2913 * pte with the new entry. This will avoid a race condition
2914 * seen in the presence of one thread doing SMC and another
2917 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2918 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2919 lru_cache_add_inactive_or_unevictable(new_page, vma);
2921 * We call the notify macro here because, when using secondary
2922 * mmu page tables (such as kvm shadow page tables), we want the
2923 * new page to be mapped directly into the secondary page table.
2925 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2926 update_mmu_cache(vma, vmf->address, vmf->pte);
2929 * Only after switching the pte to the new page may
2930 * we remove the mapcount here. Otherwise another
2931 * process may come and find the rmap count decremented
2932 * before the pte is switched to the new page, and
2933 * "reuse" the old page writing into it while our pte
2934 * here still points into it and can be read by other
2937 * The critical issue is to order this
2938 * page_remove_rmap with the ptp_clear_flush above.
2939 * Those stores are ordered by (if nothing else,)
2940 * the barrier present in the atomic_add_negative
2941 * in page_remove_rmap.
2943 * Then the TLB flush in ptep_clear_flush ensures that
2944 * no process can access the old page before the
2945 * decremented mapcount is visible. And the old page
2946 * cannot be reused until after the decremented
2947 * mapcount is visible. So transitively, TLBs to
2948 * old page will be flushed before it can be reused.
2950 page_remove_rmap(old_page, false);
2953 /* Free the old page.. */
2954 new_page = old_page;
2957 update_mmu_tlb(vma, vmf->address, vmf->pte);
2963 pte_unmap_unlock(vmf->pte, vmf->ptl);
2965 * No need to double call mmu_notifier->invalidate_range() callback as
2966 * the above ptep_clear_flush_notify() did already call it.
2968 mmu_notifier_invalidate_range_only_end(&range);
2971 * Don't let another task, with possibly unlocked vma,
2972 * keep the mlocked page.
2974 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2975 lock_page(old_page); /* LRU manipulation */
2976 if (PageMlocked(old_page))
2977 munlock_vma_page(old_page);
2978 unlock_page(old_page);
2982 return page_copied ? VM_FAULT_WRITE : 0;
2988 return VM_FAULT_OOM;
2992 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2993 * writeable once the page is prepared
2995 * @vmf: structure describing the fault
2997 * This function handles all that is needed to finish a write page fault in a
2998 * shared mapping due to PTE being read-only once the mapped page is prepared.
2999 * It handles locking of PTE and modifying it.
3001 * The function expects the page to be locked or other protection against
3002 * concurrent faults / writeback (such as DAX radix tree locks).
3004 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
3005 * we acquired PTE lock.
3007 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3009 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3010 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3013 * We might have raced with another page fault while we released the
3014 * pte_offset_map_lock.
3016 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3017 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3018 pte_unmap_unlock(vmf->pte, vmf->ptl);
3019 return VM_FAULT_NOPAGE;
3026 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3029 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3031 struct vm_area_struct *vma = vmf->vma;
3033 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3036 pte_unmap_unlock(vmf->pte, vmf->ptl);
3037 vmf->flags |= FAULT_FLAG_MKWRITE;
3038 ret = vma->vm_ops->pfn_mkwrite(vmf);
3039 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3041 return finish_mkwrite_fault(vmf);
3044 return VM_FAULT_WRITE;
3047 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3048 __releases(vmf->ptl)
3050 struct vm_area_struct *vma = vmf->vma;
3051 vm_fault_t ret = VM_FAULT_WRITE;
3053 get_page(vmf->page);
3055 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3058 pte_unmap_unlock(vmf->pte, vmf->ptl);
3059 tmp = do_page_mkwrite(vmf);
3060 if (unlikely(!tmp || (tmp &
3061 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3062 put_page(vmf->page);
3065 tmp = finish_mkwrite_fault(vmf);
3066 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3067 unlock_page(vmf->page);
3068 put_page(vmf->page);
3073 lock_page(vmf->page);
3075 ret |= fault_dirty_shared_page(vmf);
3076 put_page(vmf->page);
3082 * This routine handles present pages, when users try to write
3083 * to a shared page. It is done by copying the page to a new address
3084 * and decrementing the shared-page counter for the old page.
3086 * Note that this routine assumes that the protection checks have been
3087 * done by the caller (the low-level page fault routine in most cases).
3088 * Thus we can safely just mark it writable once we've done any necessary
3091 * We also mark the page dirty at this point even though the page will
3092 * change only once the write actually happens. This avoids a few races,
3093 * and potentially makes it more efficient.
3095 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3096 * but allow concurrent faults), with pte both mapped and locked.
3097 * We return with mmap_lock still held, but pte unmapped and unlocked.
3099 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3100 __releases(vmf->ptl)
3102 struct vm_area_struct *vma = vmf->vma;
3104 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3105 pte_unmap_unlock(vmf->pte, vmf->ptl);
3106 return handle_userfault(vmf, VM_UFFD_WP);
3109 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3112 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3115 * We should not cow pages in a shared writeable mapping.
3116 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3118 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3119 (VM_WRITE|VM_SHARED))
3120 return wp_pfn_shared(vmf);
3122 pte_unmap_unlock(vmf->pte, vmf->ptl);
3123 return wp_page_copy(vmf);
3127 * Take out anonymous pages first, anonymous shared vmas are
3128 * not dirty accountable.
3130 if (PageAnon(vmf->page)) {
3131 struct page *page = vmf->page;
3133 /* PageKsm() doesn't necessarily raise the page refcount */
3134 if (PageKsm(page) || page_count(page) != 1)
3136 if (!trylock_page(page))
3138 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3143 * Ok, we've got the only map reference, and the only
3144 * page count reference, and the page is locked,
3145 * it's dark out, and we're wearing sunglasses. Hit it.
3149 return VM_FAULT_WRITE;
3150 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3151 (VM_WRITE|VM_SHARED))) {
3152 return wp_page_shared(vmf);
3156 * Ok, we need to copy. Oh, well..
3158 get_page(vmf->page);
3160 pte_unmap_unlock(vmf->pte, vmf->ptl);
3161 return wp_page_copy(vmf);
3164 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3165 unsigned long start_addr, unsigned long end_addr,
3166 struct zap_details *details)
3168 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3171 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3172 struct zap_details *details)
3174 struct vm_area_struct *vma;
3175 pgoff_t vba, vea, zba, zea;
3177 vma_interval_tree_foreach(vma, root,
3178 details->first_index, details->last_index) {
3180 vba = vma->vm_pgoff;
3181 vea = vba + vma_pages(vma) - 1;
3182 zba = details->first_index;
3185 zea = details->last_index;
3189 unmap_mapping_range_vma(vma,
3190 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3191 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3197 * unmap_mapping_pages() - Unmap pages from processes.
3198 * @mapping: The address space containing pages to be unmapped.
3199 * @start: Index of first page to be unmapped.
3200 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3201 * @even_cows: Whether to unmap even private COWed pages.
3203 * Unmap the pages in this address space from any userspace process which
3204 * has them mmaped. Generally, you want to remove COWed pages as well when
3205 * a file is being truncated, but not when invalidating pages from the page
3208 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3209 pgoff_t nr, bool even_cows)
3211 struct zap_details details = { };
3213 details.check_mapping = even_cows ? NULL : mapping;
3214 details.first_index = start;
3215 details.last_index = start + nr - 1;
3216 if (details.last_index < details.first_index)
3217 details.last_index = ULONG_MAX;
3219 i_mmap_lock_write(mapping);
3220 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3221 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3222 i_mmap_unlock_write(mapping);
3226 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3227 * address_space corresponding to the specified byte range in the underlying
3230 * @mapping: the address space containing mmaps to be unmapped.
3231 * @holebegin: byte in first page to unmap, relative to the start of
3232 * the underlying file. This will be rounded down to a PAGE_SIZE
3233 * boundary. Note that this is different from truncate_pagecache(), which
3234 * must keep the partial page. In contrast, we must get rid of
3236 * @holelen: size of prospective hole in bytes. This will be rounded
3237 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3239 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3240 * but 0 when invalidating pagecache, don't throw away private data.
3242 void unmap_mapping_range(struct address_space *mapping,
3243 loff_t const holebegin, loff_t const holelen, int even_cows)
3245 pgoff_t hba = holebegin >> PAGE_SHIFT;
3246 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3248 /* Check for overflow. */
3249 if (sizeof(holelen) > sizeof(hlen)) {
3251 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3252 if (holeend & ~(long long)ULONG_MAX)
3253 hlen = ULONG_MAX - hba + 1;
3256 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3258 EXPORT_SYMBOL(unmap_mapping_range);
3261 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3262 * but allow concurrent faults), and pte mapped but not yet locked.
3263 * We return with pte unmapped and unlocked.
3265 * We return with the mmap_lock locked or unlocked in the same cases
3266 * as does filemap_fault().
3268 vm_fault_t do_swap_page(struct vm_fault *vmf)
3270 struct vm_area_struct *vma = vmf->vma;
3271 struct page *page = NULL, *swapcache;
3277 void *shadow = NULL;
3279 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3282 entry = pte_to_swp_entry(vmf->orig_pte);
3283 if (unlikely(non_swap_entry(entry))) {
3284 if (is_migration_entry(entry)) {
3285 migration_entry_wait(vma->vm_mm, vmf->pmd,
3287 } else if (is_device_private_entry(entry)) {
3288 vmf->page = device_private_entry_to_page(entry);
3289 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3290 } else if (is_hwpoison_entry(entry)) {
3291 ret = VM_FAULT_HWPOISON;
3293 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3294 ret = VM_FAULT_SIGBUS;
3300 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3301 page = lookup_swap_cache(entry, vma, vmf->address);
3305 struct swap_info_struct *si = swp_swap_info(entry);
3307 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3308 __swap_count(entry) == 1) {
3309 /* skip swapcache */
3310 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3315 __SetPageLocked(page);
3316 __SetPageSwapBacked(page);
3317 set_page_private(page, entry.val);
3319 /* Tell memcg to use swap ownership records */
3320 SetPageSwapCache(page);
3321 err = mem_cgroup_charge(page, vma->vm_mm,
3323 ClearPageSwapCache(page);
3329 shadow = get_shadow_from_swap_cache(entry);
3331 workingset_refault(page, shadow);
3333 lru_cache_add(page);
3334 swap_readpage(page, true);
3337 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3344 * Back out if somebody else faulted in this pte
3345 * while we released the pte lock.
3347 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3348 vmf->address, &vmf->ptl);
3349 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3351 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3355 /* Had to read the page from swap area: Major fault */
3356 ret = VM_FAULT_MAJOR;
3357 count_vm_event(PGMAJFAULT);
3358 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3359 } else if (PageHWPoison(page)) {
3361 * hwpoisoned dirty swapcache pages are kept for killing
3362 * owner processes (which may be unknown at hwpoison time)
3364 ret = VM_FAULT_HWPOISON;
3365 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3369 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3371 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3373 ret |= VM_FAULT_RETRY;
3378 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3379 * release the swapcache from under us. The page pin, and pte_same
3380 * test below, are not enough to exclude that. Even if it is still
3381 * swapcache, we need to check that the page's swap has not changed.
3383 if (unlikely((!PageSwapCache(page) ||
3384 page_private(page) != entry.val)) && swapcache)
3387 page = ksm_might_need_to_copy(page, vma, vmf->address);
3388 if (unlikely(!page)) {
3394 cgroup_throttle_swaprate(page, GFP_KERNEL);
3397 * Back out if somebody else already faulted in this pte.
3399 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3401 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3404 if (unlikely(!PageUptodate(page))) {
3405 ret = VM_FAULT_SIGBUS;
3410 * The page isn't present yet, go ahead with the fault.
3412 * Be careful about the sequence of operations here.
3413 * To get its accounting right, reuse_swap_page() must be called
3414 * while the page is counted on swap but not yet in mapcount i.e.
3415 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3416 * must be called after the swap_free(), or it will never succeed.
3419 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3420 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3421 pte = mk_pte(page, vma->vm_page_prot);
3422 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3423 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3424 vmf->flags &= ~FAULT_FLAG_WRITE;
3425 ret |= VM_FAULT_WRITE;
3426 exclusive = RMAP_EXCLUSIVE;
3428 flush_icache_page(vma, page);
3429 if (pte_swp_soft_dirty(vmf->orig_pte))
3430 pte = pte_mksoft_dirty(pte);
3431 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3432 pte = pte_mkuffd_wp(pte);
3433 pte = pte_wrprotect(pte);
3435 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3436 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3437 vmf->orig_pte = pte;
3439 /* ksm created a completely new copy */
3440 if (unlikely(page != swapcache && swapcache)) {
3441 page_add_new_anon_rmap(page, vma, vmf->address, false);
3442 lru_cache_add_inactive_or_unevictable(page, vma);
3444 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3448 if (mem_cgroup_swap_full(page) ||
3449 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3450 try_to_free_swap(page);
3452 if (page != swapcache && swapcache) {
3454 * Hold the lock to avoid the swap entry to be reused
3455 * until we take the PT lock for the pte_same() check
3456 * (to avoid false positives from pte_same). For
3457 * further safety release the lock after the swap_free
3458 * so that the swap count won't change under a
3459 * parallel locked swapcache.
3461 unlock_page(swapcache);
3462 put_page(swapcache);
3465 if (vmf->flags & FAULT_FLAG_WRITE) {
3466 ret |= do_wp_page(vmf);
3467 if (ret & VM_FAULT_ERROR)
3468 ret &= VM_FAULT_ERROR;
3472 /* No need to invalidate - it was non-present before */
3473 update_mmu_cache(vma, vmf->address, vmf->pte);
3475 pte_unmap_unlock(vmf->pte, vmf->ptl);
3479 pte_unmap_unlock(vmf->pte, vmf->ptl);
3484 if (page != swapcache && swapcache) {
3485 unlock_page(swapcache);
3486 put_page(swapcache);
3492 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3493 * but allow concurrent faults), and pte mapped but not yet locked.
3494 * We return with mmap_lock still held, but pte unmapped and unlocked.
3496 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3498 struct vm_area_struct *vma = vmf->vma;
3503 /* File mapping without ->vm_ops ? */
3504 if (vma->vm_flags & VM_SHARED)
3505 return VM_FAULT_SIGBUS;
3508 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3509 * pte_offset_map() on pmds where a huge pmd might be created
3510 * from a different thread.
3512 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3513 * parallel threads are excluded by other means.
3515 * Here we only have mmap_read_lock(mm).
3517 if (pte_alloc(vma->vm_mm, vmf->pmd))
3518 return VM_FAULT_OOM;
3520 /* See the comment in pte_alloc_one_map() */
3521 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3524 /* Use the zero-page for reads */
3525 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3526 !mm_forbids_zeropage(vma->vm_mm)) {
3527 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3528 vma->vm_page_prot));
3529 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3530 vmf->address, &vmf->ptl);
3531 if (!pte_none(*vmf->pte)) {
3532 update_mmu_tlb(vma, vmf->address, vmf->pte);
3535 ret = check_stable_address_space(vma->vm_mm);
3538 /* Deliver the page fault to userland, check inside PT lock */
3539 if (userfaultfd_missing(vma)) {
3540 pte_unmap_unlock(vmf->pte, vmf->ptl);
3541 return handle_userfault(vmf, VM_UFFD_MISSING);
3546 /* Allocate our own private page. */
3547 if (unlikely(anon_vma_prepare(vma)))
3549 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3553 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3555 cgroup_throttle_swaprate(page, GFP_KERNEL);
3558 * The memory barrier inside __SetPageUptodate makes sure that
3559 * preceding stores to the page contents become visible before
3560 * the set_pte_at() write.
3562 __SetPageUptodate(page);
3564 entry = mk_pte(page, vma->vm_page_prot);
3565 entry = pte_sw_mkyoung(entry);
3566 if (vma->vm_flags & VM_WRITE)
3567 entry = pte_mkwrite(pte_mkdirty(entry));
3569 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3571 if (!pte_none(*vmf->pte)) {
3572 update_mmu_cache(vma, vmf->address, vmf->pte);
3576 ret = check_stable_address_space(vma->vm_mm);
3580 /* Deliver the page fault to userland, check inside PT lock */
3581 if (userfaultfd_missing(vma)) {
3582 pte_unmap_unlock(vmf->pte, vmf->ptl);
3584 return handle_userfault(vmf, VM_UFFD_MISSING);
3587 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3588 page_add_new_anon_rmap(page, vma, vmf->address, false);
3589 lru_cache_add_inactive_or_unevictable(page, vma);
3591 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3593 /* No need to invalidate - it was non-present before */
3594 update_mmu_cache(vma, vmf->address, vmf->pte);
3596 pte_unmap_unlock(vmf->pte, vmf->ptl);
3604 return VM_FAULT_OOM;
3608 * The mmap_lock must have been held on entry, and may have been
3609 * released depending on flags and vma->vm_ops->fault() return value.
3610 * See filemap_fault() and __lock_page_retry().
3612 static vm_fault_t __do_fault(struct vm_fault *vmf)
3614 struct vm_area_struct *vma = vmf->vma;
3618 * Preallocate pte before we take page_lock because this might lead to
3619 * deadlocks for memcg reclaim which waits for pages under writeback:
3621 * SetPageWriteback(A)
3627 * wait_on_page_writeback(A)
3628 * SetPageWriteback(B)
3630 * # flush A, B to clear the writeback
3632 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3633 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3634 if (!vmf->prealloc_pte)
3635 return VM_FAULT_OOM;
3636 smp_wmb(); /* See comment in __pte_alloc() */
3639 ret = vma->vm_ops->fault(vmf);
3640 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3641 VM_FAULT_DONE_COW)))
3644 if (unlikely(PageHWPoison(vmf->page))) {
3645 if (ret & VM_FAULT_LOCKED)
3646 unlock_page(vmf->page);
3647 put_page(vmf->page);
3649 return VM_FAULT_HWPOISON;
3652 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3653 lock_page(vmf->page);
3655 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3661 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3662 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3663 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3664 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3666 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3668 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3671 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3673 struct vm_area_struct *vma = vmf->vma;
3675 if (!pmd_none(*vmf->pmd))
3677 if (vmf->prealloc_pte) {
3678 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3679 if (unlikely(!pmd_none(*vmf->pmd))) {
3680 spin_unlock(vmf->ptl);
3684 mm_inc_nr_ptes(vma->vm_mm);
3685 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3686 spin_unlock(vmf->ptl);
3687 vmf->prealloc_pte = NULL;
3688 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3689 return VM_FAULT_OOM;
3693 * If a huge pmd materialized under us just retry later. Use
3694 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3695 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3696 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3697 * running immediately after a huge pmd fault in a different thread of
3698 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3699 * All we have to ensure is that it is a regular pmd that we can walk
3700 * with pte_offset_map() and we can do that through an atomic read in
3701 * C, which is what pmd_trans_unstable() provides.
3703 if (pmd_devmap_trans_unstable(vmf->pmd))
3704 return VM_FAULT_NOPAGE;
3707 * At this point we know that our vmf->pmd points to a page of ptes
3708 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3709 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3710 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3711 * be valid and we will re-check to make sure the vmf->pte isn't
3712 * pte_none() under vmf->ptl protection when we return to
3715 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3720 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3721 static void deposit_prealloc_pte(struct vm_fault *vmf)
3723 struct vm_area_struct *vma = vmf->vma;
3725 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3727 * We are going to consume the prealloc table,
3728 * count that as nr_ptes.
3730 mm_inc_nr_ptes(vma->vm_mm);
3731 vmf->prealloc_pte = NULL;
3734 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3736 struct vm_area_struct *vma = vmf->vma;
3737 bool write = vmf->flags & FAULT_FLAG_WRITE;
3738 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3743 if (!transhuge_vma_suitable(vma, haddr))
3744 return VM_FAULT_FALLBACK;
3746 ret = VM_FAULT_FALLBACK;
3747 page = compound_head(page);
3750 * Archs like ppc64 need additonal space to store information
3751 * related to pte entry. Use the preallocated table for that.
3753 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3754 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3755 if (!vmf->prealloc_pte)
3756 return VM_FAULT_OOM;
3757 smp_wmb(); /* See comment in __pte_alloc() */
3760 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3761 if (unlikely(!pmd_none(*vmf->pmd)))
3764 for (i = 0; i < HPAGE_PMD_NR; i++)
3765 flush_icache_page(vma, page + i);
3767 entry = mk_huge_pmd(page, vma->vm_page_prot);
3769 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3771 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3772 page_add_file_rmap(page, true);
3774 * deposit and withdraw with pmd lock held
3776 if (arch_needs_pgtable_deposit())
3777 deposit_prealloc_pte(vmf);
3779 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3781 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3783 /* fault is handled */
3785 count_vm_event(THP_FILE_MAPPED);
3787 spin_unlock(vmf->ptl);
3791 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3799 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3800 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3802 * @vmf: fault environment
3803 * @page: page to map
3805 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3808 * Target users are page handler itself and implementations of
3809 * vm_ops->map_pages.
3811 * Return: %0 on success, %VM_FAULT_ code in case of error.
3813 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
3815 struct vm_area_struct *vma = vmf->vma;
3816 bool write = vmf->flags & FAULT_FLAG_WRITE;
3820 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3821 ret = do_set_pmd(vmf, page);
3822 if (ret != VM_FAULT_FALLBACK)
3827 ret = pte_alloc_one_map(vmf);
3832 /* Re-check under ptl */
3833 if (unlikely(!pte_none(*vmf->pte))) {
3834 update_mmu_tlb(vma, vmf->address, vmf->pte);
3835 return VM_FAULT_NOPAGE;
3838 flush_icache_page(vma, page);
3839 entry = mk_pte(page, vma->vm_page_prot);
3840 entry = pte_sw_mkyoung(entry);
3842 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3843 /* copy-on-write page */
3844 if (write && !(vma->vm_flags & VM_SHARED)) {
3845 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3846 page_add_new_anon_rmap(page, vma, vmf->address, false);
3847 lru_cache_add_inactive_or_unevictable(page, vma);
3849 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3850 page_add_file_rmap(page, false);
3852 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3854 /* no need to invalidate: a not-present page won't be cached */
3855 update_mmu_cache(vma, vmf->address, vmf->pte);
3862 * finish_fault - finish page fault once we have prepared the page to fault
3864 * @vmf: structure describing the fault
3866 * This function handles all that is needed to finish a page fault once the
3867 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3868 * given page, adds reverse page mapping, handles memcg charges and LRU
3871 * The function expects the page to be locked and on success it consumes a
3872 * reference of a page being mapped (for the PTE which maps it).
3874 * Return: %0 on success, %VM_FAULT_ code in case of error.
3876 vm_fault_t finish_fault(struct vm_fault *vmf)
3881 /* Did we COW the page? */
3882 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3883 !(vmf->vma->vm_flags & VM_SHARED))
3884 page = vmf->cow_page;
3889 * check even for read faults because we might have lost our CoWed
3892 if (!(vmf->vma->vm_flags & VM_SHARED))
3893 ret = check_stable_address_space(vmf->vma->vm_mm);
3895 ret = alloc_set_pte(vmf, page);
3897 pte_unmap_unlock(vmf->pte, vmf->ptl);
3901 static unsigned long fault_around_bytes __read_mostly =
3902 rounddown_pow_of_two(65536);
3904 #ifdef CONFIG_DEBUG_FS
3905 static int fault_around_bytes_get(void *data, u64 *val)
3907 *val = fault_around_bytes;
3912 * fault_around_bytes must be rounded down to the nearest page order as it's
3913 * what do_fault_around() expects to see.
3915 static int fault_around_bytes_set(void *data, u64 val)
3917 if (val / PAGE_SIZE > PTRS_PER_PTE)
3919 if (val > PAGE_SIZE)
3920 fault_around_bytes = rounddown_pow_of_two(val);
3922 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3925 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3926 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3928 static int __init fault_around_debugfs(void)
3930 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3931 &fault_around_bytes_fops);
3934 late_initcall(fault_around_debugfs);
3938 * do_fault_around() tries to map few pages around the fault address. The hope
3939 * is that the pages will be needed soon and this will lower the number of
3942 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3943 * not ready to be mapped: not up-to-date, locked, etc.
3945 * This function is called with the page table lock taken. In the split ptlock
3946 * case the page table lock only protects only those entries which belong to
3947 * the page table corresponding to the fault address.
3949 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3952 * fault_around_bytes defines how many bytes we'll try to map.
3953 * do_fault_around() expects it to be set to a power of two less than or equal
3956 * The virtual address of the area that we map is naturally aligned to
3957 * fault_around_bytes rounded down to the machine page size
3958 * (and therefore to page order). This way it's easier to guarantee
3959 * that we don't cross page table boundaries.
3961 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3963 unsigned long address = vmf->address, nr_pages, mask;
3964 pgoff_t start_pgoff = vmf->pgoff;
3969 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3970 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3972 vmf->address = max(address & mask, vmf->vma->vm_start);
3973 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3977 * end_pgoff is either the end of the page table, the end of
3978 * the vma or nr_pages from start_pgoff, depending what is nearest.
3980 end_pgoff = start_pgoff -
3981 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3983 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3984 start_pgoff + nr_pages - 1);
3986 if (pmd_none(*vmf->pmd)) {
3987 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3988 if (!vmf->prealloc_pte)
3990 smp_wmb(); /* See comment in __pte_alloc() */
3993 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3995 /* Huge page is mapped? Page fault is solved */
3996 if (pmd_trans_huge(*vmf->pmd)) {
3997 ret = VM_FAULT_NOPAGE;
4001 /* ->map_pages() haven't done anything useful. Cold page cache? */
4005 /* check if the page fault is solved */
4006 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
4007 if (!pte_none(*vmf->pte))
4008 ret = VM_FAULT_NOPAGE;
4009 pte_unmap_unlock(vmf->pte, vmf->ptl);
4011 vmf->address = address;
4016 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4018 struct vm_area_struct *vma = vmf->vma;
4022 * Let's call ->map_pages() first and use ->fault() as fallback
4023 * if page by the offset is not ready to be mapped (cold cache or
4026 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4027 ret = do_fault_around(vmf);
4032 ret = __do_fault(vmf);
4033 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4036 ret |= finish_fault(vmf);
4037 unlock_page(vmf->page);
4038 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4039 put_page(vmf->page);
4043 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4045 struct vm_area_struct *vma = vmf->vma;
4048 if (unlikely(anon_vma_prepare(vma)))
4049 return VM_FAULT_OOM;
4051 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4053 return VM_FAULT_OOM;
4055 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4056 put_page(vmf->cow_page);
4057 return VM_FAULT_OOM;
4059 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4061 ret = __do_fault(vmf);
4062 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4064 if (ret & VM_FAULT_DONE_COW)
4067 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4068 __SetPageUptodate(vmf->cow_page);
4070 ret |= finish_fault(vmf);
4071 unlock_page(vmf->page);
4072 put_page(vmf->page);
4073 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4077 put_page(vmf->cow_page);
4081 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4083 struct vm_area_struct *vma = vmf->vma;
4084 vm_fault_t ret, tmp;
4086 ret = __do_fault(vmf);
4087 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4091 * Check if the backing address space wants to know that the page is
4092 * about to become writable
4094 if (vma->vm_ops->page_mkwrite) {
4095 unlock_page(vmf->page);
4096 tmp = do_page_mkwrite(vmf);
4097 if (unlikely(!tmp ||
4098 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4099 put_page(vmf->page);
4104 ret |= finish_fault(vmf);
4105 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4107 unlock_page(vmf->page);
4108 put_page(vmf->page);
4112 ret |= fault_dirty_shared_page(vmf);
4117 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4118 * but allow concurrent faults).
4119 * The mmap_lock may have been released depending on flags and our
4120 * return value. See filemap_fault() and __lock_page_or_retry().
4121 * If mmap_lock is released, vma may become invalid (for example
4122 * by other thread calling munmap()).
4124 static vm_fault_t do_fault(struct vm_fault *vmf)
4126 struct vm_area_struct *vma = vmf->vma;
4127 struct mm_struct *vm_mm = vma->vm_mm;
4131 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4133 if (!vma->vm_ops->fault) {
4135 * If we find a migration pmd entry or a none pmd entry, which
4136 * should never happen, return SIGBUS
4138 if (unlikely(!pmd_present(*vmf->pmd)))
4139 ret = VM_FAULT_SIGBUS;
4141 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4146 * Make sure this is not a temporary clearing of pte
4147 * by holding ptl and checking again. A R/M/W update
4148 * of pte involves: take ptl, clearing the pte so that
4149 * we don't have concurrent modification by hardware
4150 * followed by an update.
4152 if (unlikely(pte_none(*vmf->pte)))
4153 ret = VM_FAULT_SIGBUS;
4155 ret = VM_FAULT_NOPAGE;
4157 pte_unmap_unlock(vmf->pte, vmf->ptl);
4159 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4160 ret = do_read_fault(vmf);
4161 else if (!(vma->vm_flags & VM_SHARED))
4162 ret = do_cow_fault(vmf);
4164 ret = do_shared_fault(vmf);
4166 /* preallocated pagetable is unused: free it */
4167 if (vmf->prealloc_pte) {
4168 pte_free(vm_mm, vmf->prealloc_pte);
4169 vmf->prealloc_pte = NULL;
4174 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4175 unsigned long addr, int page_nid,
4180 count_vm_numa_event(NUMA_HINT_FAULTS);
4181 if (page_nid == numa_node_id()) {
4182 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4183 *flags |= TNF_FAULT_LOCAL;
4186 return mpol_misplaced(page, vma, addr);
4189 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4191 struct vm_area_struct *vma = vmf->vma;
4192 struct page *page = NULL;
4193 int page_nid = NUMA_NO_NODE;
4196 bool migrated = false;
4198 bool was_writable = pte_savedwrite(vmf->orig_pte);
4202 * The "pte" at this point cannot be used safely without
4203 * validation through pte_unmap_same(). It's of NUMA type but
4204 * the pfn may be screwed if the read is non atomic.
4206 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4207 spin_lock(vmf->ptl);
4208 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4209 pte_unmap_unlock(vmf->pte, vmf->ptl);
4214 * Make it present again, Depending on how arch implementes non
4215 * accessible ptes, some can allow access by kernel mode.
4217 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4218 pte = pte_modify(old_pte, vma->vm_page_prot);
4219 pte = pte_mkyoung(pte);
4221 pte = pte_mkwrite(pte);
4222 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4223 update_mmu_cache(vma, vmf->address, vmf->pte);
4225 page = vm_normal_page(vma, vmf->address, pte);
4227 pte_unmap_unlock(vmf->pte, vmf->ptl);
4231 /* TODO: handle PTE-mapped THP */
4232 if (PageCompound(page)) {
4233 pte_unmap_unlock(vmf->pte, vmf->ptl);
4238 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4239 * much anyway since they can be in shared cache state. This misses
4240 * the case where a mapping is writable but the process never writes
4241 * to it but pte_write gets cleared during protection updates and
4242 * pte_dirty has unpredictable behaviour between PTE scan updates,
4243 * background writeback, dirty balancing and application behaviour.
4245 if (!pte_write(pte))
4246 flags |= TNF_NO_GROUP;
4249 * Flag if the page is shared between multiple address spaces. This
4250 * is later used when determining whether to group tasks together
4252 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4253 flags |= TNF_SHARED;
4255 last_cpupid = page_cpupid_last(page);
4256 page_nid = page_to_nid(page);
4257 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4259 pte_unmap_unlock(vmf->pte, vmf->ptl);
4260 if (target_nid == NUMA_NO_NODE) {
4265 /* Migrate to the requested node */
4266 migrated = migrate_misplaced_page(page, vma, target_nid);
4268 page_nid = target_nid;
4269 flags |= TNF_MIGRATED;
4271 flags |= TNF_MIGRATE_FAIL;
4274 if (page_nid != NUMA_NO_NODE)
4275 task_numa_fault(last_cpupid, page_nid, 1, flags);
4279 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4281 if (vma_is_anonymous(vmf->vma))
4282 return do_huge_pmd_anonymous_page(vmf);
4283 if (vmf->vma->vm_ops->huge_fault)
4284 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4285 return VM_FAULT_FALLBACK;
4288 /* `inline' is required to avoid gcc 4.1.2 build error */
4289 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4291 if (vma_is_anonymous(vmf->vma)) {
4292 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4293 return handle_userfault(vmf, VM_UFFD_WP);
4294 return do_huge_pmd_wp_page(vmf, orig_pmd);
4296 if (vmf->vma->vm_ops->huge_fault) {
4297 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4299 if (!(ret & VM_FAULT_FALLBACK))
4303 /* COW or write-notify handled on pte level: split pmd. */
4304 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4306 return VM_FAULT_FALLBACK;
4309 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4311 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4312 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4313 /* No support for anonymous transparent PUD pages yet */
4314 if (vma_is_anonymous(vmf->vma))
4316 if (vmf->vma->vm_ops->huge_fault) {
4317 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4319 if (!(ret & VM_FAULT_FALLBACK))
4323 /* COW or write-notify not handled on PUD level: split pud.*/
4324 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4325 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4326 return VM_FAULT_FALLBACK;
4329 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4331 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4332 /* No support for anonymous transparent PUD pages yet */
4333 if (vma_is_anonymous(vmf->vma))
4334 return VM_FAULT_FALLBACK;
4335 if (vmf->vma->vm_ops->huge_fault)
4336 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4337 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4338 return VM_FAULT_FALLBACK;
4342 * These routines also need to handle stuff like marking pages dirty
4343 * and/or accessed for architectures that don't do it in hardware (most
4344 * RISC architectures). The early dirtying is also good on the i386.
4346 * There is also a hook called "update_mmu_cache()" that architectures
4347 * with external mmu caches can use to update those (ie the Sparc or
4348 * PowerPC hashed page tables that act as extended TLBs).
4350 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4351 * concurrent faults).
4353 * The mmap_lock may have been released depending on flags and our return value.
4354 * See filemap_fault() and __lock_page_or_retry().
4356 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4360 if (unlikely(pmd_none(*vmf->pmd))) {
4362 * Leave __pte_alloc() until later: because vm_ops->fault may
4363 * want to allocate huge page, and if we expose page table
4364 * for an instant, it will be difficult to retract from
4365 * concurrent faults and from rmap lookups.
4369 /* See comment in pte_alloc_one_map() */
4370 if (pmd_devmap_trans_unstable(vmf->pmd))
4373 * A regular pmd is established and it can't morph into a huge
4374 * pmd from under us anymore at this point because we hold the
4375 * mmap_lock read mode and khugepaged takes it in write mode.
4376 * So now it's safe to run pte_offset_map().
4378 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4379 vmf->orig_pte = *vmf->pte;
4382 * some architectures can have larger ptes than wordsize,
4383 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4384 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4385 * accesses. The code below just needs a consistent view
4386 * for the ifs and we later double check anyway with the
4387 * ptl lock held. So here a barrier will do.
4390 if (pte_none(vmf->orig_pte)) {
4391 pte_unmap(vmf->pte);
4397 if (vma_is_anonymous(vmf->vma))
4398 return do_anonymous_page(vmf);
4400 return do_fault(vmf);
4403 if (!pte_present(vmf->orig_pte))
4404 return do_swap_page(vmf);
4406 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4407 return do_numa_page(vmf);
4409 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4410 spin_lock(vmf->ptl);
4411 entry = vmf->orig_pte;
4412 if (unlikely(!pte_same(*vmf->pte, entry))) {
4413 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4416 if (vmf->flags & FAULT_FLAG_WRITE) {
4417 if (!pte_write(entry))
4418 return do_wp_page(vmf);
4419 entry = pte_mkdirty(entry);
4421 entry = pte_mkyoung(entry);
4422 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4423 vmf->flags & FAULT_FLAG_WRITE)) {
4424 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4426 /* Skip spurious TLB flush for retried page fault */
4427 if (vmf->flags & FAULT_FLAG_TRIED)
4430 * This is needed only for protection faults but the arch code
4431 * is not yet telling us if this is a protection fault or not.
4432 * This still avoids useless tlb flushes for .text page faults
4435 if (vmf->flags & FAULT_FLAG_WRITE)
4436 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4439 pte_unmap_unlock(vmf->pte, vmf->ptl);
4444 * By the time we get here, we already hold the mm semaphore
4446 * The mmap_lock may have been released depending on flags and our
4447 * return value. See filemap_fault() and __lock_page_or_retry().
4449 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4450 unsigned long address, unsigned int flags)
4452 struct vm_fault vmf = {
4454 .address = address & PAGE_MASK,
4456 .pgoff = linear_page_index(vma, address),
4457 .gfp_mask = __get_fault_gfp_mask(vma),
4459 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4460 struct mm_struct *mm = vma->vm_mm;
4465 pgd = pgd_offset(mm, address);
4466 p4d = p4d_alloc(mm, pgd, address);
4468 return VM_FAULT_OOM;
4470 vmf.pud = pud_alloc(mm, p4d, address);
4472 return VM_FAULT_OOM;
4474 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4475 ret = create_huge_pud(&vmf);
4476 if (!(ret & VM_FAULT_FALLBACK))
4479 pud_t orig_pud = *vmf.pud;
4482 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4484 /* NUMA case for anonymous PUDs would go here */
4486 if (dirty && !pud_write(orig_pud)) {
4487 ret = wp_huge_pud(&vmf, orig_pud);
4488 if (!(ret & VM_FAULT_FALLBACK))
4491 huge_pud_set_accessed(&vmf, orig_pud);
4497 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4499 return VM_FAULT_OOM;
4501 /* Huge pud page fault raced with pmd_alloc? */
4502 if (pud_trans_unstable(vmf.pud))
4505 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4506 ret = create_huge_pmd(&vmf);
4507 if (!(ret & VM_FAULT_FALLBACK))
4510 pmd_t orig_pmd = *vmf.pmd;
4513 if (unlikely(is_swap_pmd(orig_pmd))) {
4514 VM_BUG_ON(thp_migration_supported() &&
4515 !is_pmd_migration_entry(orig_pmd));
4516 if (is_pmd_migration_entry(orig_pmd))
4517 pmd_migration_entry_wait(mm, vmf.pmd);
4520 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4521 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4522 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4524 if (dirty && !pmd_write(orig_pmd)) {
4525 ret = wp_huge_pmd(&vmf, orig_pmd);
4526 if (!(ret & VM_FAULT_FALLBACK))
4529 huge_pmd_set_accessed(&vmf, orig_pmd);
4535 return handle_pte_fault(&vmf);
4539 * mm_account_fault - Do page fault accountings
4541 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4542 * of perf event counters, but we'll still do the per-task accounting to
4543 * the task who triggered this page fault.
4544 * @address: the faulted address.
4545 * @flags: the fault flags.
4546 * @ret: the fault retcode.
4548 * This will take care of most of the page fault accountings. Meanwhile, it
4549 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4550 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4551 * still be in per-arch page fault handlers at the entry of page fault.
4553 static inline void mm_account_fault(struct pt_regs *regs,
4554 unsigned long address, unsigned int flags,
4560 * We don't do accounting for some specific faults:
4562 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4563 * includes arch_vma_access_permitted() failing before reaching here.
4564 * So this is not a "this many hardware page faults" counter. We
4565 * should use the hw profiling for that.
4567 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4568 * once they're completed.
4570 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4574 * We define the fault as a major fault when the final successful fault
4575 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4576 * handle it immediately previously).
4578 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4586 * If the fault is done for GUP, regs will be NULL. We only do the
4587 * accounting for the per thread fault counters who triggered the
4588 * fault, and we skip the perf event updates.
4594 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4596 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4600 * By the time we get here, we already hold the mm semaphore
4602 * The mmap_lock may have been released depending on flags and our
4603 * return value. See filemap_fault() and __lock_page_or_retry().
4605 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4606 unsigned int flags, struct pt_regs *regs)
4610 __set_current_state(TASK_RUNNING);
4612 count_vm_event(PGFAULT);
4613 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4615 /* do counter updates before entering really critical section. */
4616 check_sync_rss_stat(current);
4618 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4619 flags & FAULT_FLAG_INSTRUCTION,
4620 flags & FAULT_FLAG_REMOTE))
4621 return VM_FAULT_SIGSEGV;
4624 * Enable the memcg OOM handling for faults triggered in user
4625 * space. Kernel faults are handled more gracefully.
4627 if (flags & FAULT_FLAG_USER)
4628 mem_cgroup_enter_user_fault();
4630 if (unlikely(is_vm_hugetlb_page(vma)))
4631 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4633 ret = __handle_mm_fault(vma, address, flags);
4635 if (flags & FAULT_FLAG_USER) {
4636 mem_cgroup_exit_user_fault();
4638 * The task may have entered a memcg OOM situation but
4639 * if the allocation error was handled gracefully (no
4640 * VM_FAULT_OOM), there is no need to kill anything.
4641 * Just clean up the OOM state peacefully.
4643 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4644 mem_cgroup_oom_synchronize(false);
4647 mm_account_fault(regs, address, flags, ret);
4651 EXPORT_SYMBOL_GPL(handle_mm_fault);
4653 #ifndef __PAGETABLE_P4D_FOLDED
4655 * Allocate p4d page table.
4656 * We've already handled the fast-path in-line.
4658 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4660 p4d_t *new = p4d_alloc_one(mm, address);
4664 smp_wmb(); /* See comment in __pte_alloc */
4666 spin_lock(&mm->page_table_lock);
4667 if (pgd_present(*pgd)) /* Another has populated it */
4670 pgd_populate(mm, pgd, new);
4671 spin_unlock(&mm->page_table_lock);
4674 #endif /* __PAGETABLE_P4D_FOLDED */
4676 #ifndef __PAGETABLE_PUD_FOLDED
4678 * Allocate page upper directory.
4679 * We've already handled the fast-path in-line.
4681 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4683 pud_t *new = pud_alloc_one(mm, address);
4687 smp_wmb(); /* See comment in __pte_alloc */
4689 spin_lock(&mm->page_table_lock);
4690 if (!p4d_present(*p4d)) {
4692 p4d_populate(mm, p4d, new);
4693 } else /* Another has populated it */
4695 spin_unlock(&mm->page_table_lock);
4698 #endif /* __PAGETABLE_PUD_FOLDED */
4700 #ifndef __PAGETABLE_PMD_FOLDED
4702 * Allocate page middle directory.
4703 * We've already handled the fast-path in-line.
4705 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4708 pmd_t *new = pmd_alloc_one(mm, address);
4712 smp_wmb(); /* See comment in __pte_alloc */
4714 ptl = pud_lock(mm, pud);
4715 if (!pud_present(*pud)) {
4717 pud_populate(mm, pud, new);
4718 } else /* Another has populated it */
4723 #endif /* __PAGETABLE_PMD_FOLDED */
4725 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4726 struct mmu_notifier_range *range,
4727 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4735 pgd = pgd_offset(mm, address);
4736 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4739 p4d = p4d_offset(pgd, address);
4740 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4743 pud = pud_offset(p4d, address);
4744 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4747 pmd = pmd_offset(pud, address);
4748 VM_BUG_ON(pmd_trans_huge(*pmd));
4750 if (pmd_huge(*pmd)) {
4755 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4756 NULL, mm, address & PMD_MASK,
4757 (address & PMD_MASK) + PMD_SIZE);
4758 mmu_notifier_invalidate_range_start(range);
4760 *ptlp = pmd_lock(mm, pmd);
4761 if (pmd_huge(*pmd)) {
4767 mmu_notifier_invalidate_range_end(range);
4770 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4774 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4775 address & PAGE_MASK,
4776 (address & PAGE_MASK) + PAGE_SIZE);
4777 mmu_notifier_invalidate_range_start(range);
4779 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4780 if (!pte_present(*ptep))
4785 pte_unmap_unlock(ptep, *ptlp);
4787 mmu_notifier_invalidate_range_end(range);
4792 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4793 pte_t **ptepp, spinlock_t **ptlp)
4797 /* (void) is needed to make gcc happy */
4798 (void) __cond_lock(*ptlp,
4799 !(res = __follow_pte_pmd(mm, address, NULL,
4800 ptepp, NULL, ptlp)));
4804 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4805 struct mmu_notifier_range *range,
4806 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4810 /* (void) is needed to make gcc happy */
4811 (void) __cond_lock(*ptlp,
4812 !(res = __follow_pte_pmd(mm, address, range,
4813 ptepp, pmdpp, ptlp)));
4816 EXPORT_SYMBOL(follow_pte_pmd);
4819 * follow_pfn - look up PFN at a user virtual address
4820 * @vma: memory mapping
4821 * @address: user virtual address
4822 * @pfn: location to store found PFN
4824 * Only IO mappings and raw PFN mappings are allowed.
4826 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4828 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4835 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4838 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4841 *pfn = pte_pfn(*ptep);
4842 pte_unmap_unlock(ptep, ptl);
4845 EXPORT_SYMBOL(follow_pfn);
4847 #ifdef CONFIG_HAVE_IOREMAP_PROT
4848 int follow_phys(struct vm_area_struct *vma,
4849 unsigned long address, unsigned int flags,
4850 unsigned long *prot, resource_size_t *phys)
4856 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4859 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4863 if ((flags & FOLL_WRITE) && !pte_write(pte))
4866 *prot = pgprot_val(pte_pgprot(pte));
4867 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4871 pte_unmap_unlock(ptep, ptl);
4876 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4877 void *buf, int len, int write)
4879 resource_size_t phys_addr;
4880 unsigned long prot = 0;
4881 void __iomem *maddr;
4882 int offset = addr & (PAGE_SIZE-1);
4884 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4887 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4892 memcpy_toio(maddr + offset, buf, len);
4894 memcpy_fromio(buf, maddr + offset, len);
4899 EXPORT_SYMBOL_GPL(generic_access_phys);
4903 * Access another process' address space as given in mm. If non-NULL, use the
4904 * given task for page fault accounting.
4906 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4907 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4909 struct vm_area_struct *vma;
4910 void *old_buf = buf;
4911 int write = gup_flags & FOLL_WRITE;
4913 if (mmap_read_lock_killable(mm))
4916 /* ignore errors, just check how much was successfully transferred */
4918 int bytes, ret, offset;
4920 struct page *page = NULL;
4922 ret = get_user_pages_remote(mm, addr, 1,
4923 gup_flags, &page, &vma, NULL);
4925 #ifndef CONFIG_HAVE_IOREMAP_PROT
4929 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4930 * we can access using slightly different code.
4932 vma = find_vma(mm, addr);
4933 if (!vma || vma->vm_start > addr)
4935 if (vma->vm_ops && vma->vm_ops->access)
4936 ret = vma->vm_ops->access(vma, addr, buf,
4944 offset = addr & (PAGE_SIZE-1);
4945 if (bytes > PAGE_SIZE-offset)
4946 bytes = PAGE_SIZE-offset;
4950 copy_to_user_page(vma, page, addr,
4951 maddr + offset, buf, bytes);
4952 set_page_dirty_lock(page);
4954 copy_from_user_page(vma, page, addr,
4955 buf, maddr + offset, bytes);
4964 mmap_read_unlock(mm);
4966 return buf - old_buf;
4970 * access_remote_vm - access another process' address space
4971 * @mm: the mm_struct of the target address space
4972 * @addr: start address to access
4973 * @buf: source or destination buffer
4974 * @len: number of bytes to transfer
4975 * @gup_flags: flags modifying lookup behaviour
4977 * The caller must hold a reference on @mm.
4979 * Return: number of bytes copied from source to destination.
4981 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4982 void *buf, int len, unsigned int gup_flags)
4984 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4988 * Access another process' address space.
4989 * Source/target buffer must be kernel space,
4990 * Do not walk the page table directly, use get_user_pages
4992 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4993 void *buf, int len, unsigned int gup_flags)
4995 struct mm_struct *mm;
4998 mm = get_task_mm(tsk);
5002 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
5008 EXPORT_SYMBOL_GPL(access_process_vm);
5011 * Print the name of a VMA.
5013 void print_vma_addr(char *prefix, unsigned long ip)
5015 struct mm_struct *mm = current->mm;
5016 struct vm_area_struct *vma;
5019 * we might be running from an atomic context so we cannot sleep
5021 if (!mmap_read_trylock(mm))
5024 vma = find_vma(mm, ip);
5025 if (vma && vma->vm_file) {
5026 struct file *f = vma->vm_file;
5027 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5031 p = file_path(f, buf, PAGE_SIZE);
5034 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5036 vma->vm_end - vma->vm_start);
5037 free_page((unsigned long)buf);
5040 mmap_read_unlock(mm);
5043 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5044 void __might_fault(const char *file, int line)
5047 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5048 * holding the mmap_lock, this is safe because kernel memory doesn't
5049 * get paged out, therefore we'll never actually fault, and the
5050 * below annotations will generate false positives.
5052 if (uaccess_kernel())
5054 if (pagefault_disabled())
5056 __might_sleep(file, line, 0);
5057 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5059 might_lock_read(¤t->mm->mmap_lock);
5062 EXPORT_SYMBOL(__might_fault);
5065 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5067 * Process all subpages of the specified huge page with the specified
5068 * operation. The target subpage will be processed last to keep its
5071 static inline void process_huge_page(
5072 unsigned long addr_hint, unsigned int pages_per_huge_page,
5073 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5077 unsigned long addr = addr_hint &
5078 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5080 /* Process target subpage last to keep its cache lines hot */
5082 n = (addr_hint - addr) / PAGE_SIZE;
5083 if (2 * n <= pages_per_huge_page) {
5084 /* If target subpage in first half of huge page */
5087 /* Process subpages at the end of huge page */
5088 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5090 process_subpage(addr + i * PAGE_SIZE, i, arg);
5093 /* If target subpage in second half of huge page */
5094 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5095 l = pages_per_huge_page - n;
5096 /* Process subpages at the begin of huge page */
5097 for (i = 0; i < base; i++) {
5099 process_subpage(addr + i * PAGE_SIZE, i, arg);
5103 * Process remaining subpages in left-right-left-right pattern
5104 * towards the target subpage
5106 for (i = 0; i < l; i++) {
5107 int left_idx = base + i;
5108 int right_idx = base + 2 * l - 1 - i;
5111 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5113 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5117 static void clear_gigantic_page(struct page *page,
5119 unsigned int pages_per_huge_page)
5122 struct page *p = page;
5125 for (i = 0; i < pages_per_huge_page;
5126 i++, p = mem_map_next(p, page, i)) {
5128 clear_user_highpage(p, addr + i * PAGE_SIZE);
5132 static void clear_subpage(unsigned long addr, int idx, void *arg)
5134 struct page *page = arg;
5136 clear_user_highpage(page + idx, addr);
5139 void clear_huge_page(struct page *page,
5140 unsigned long addr_hint, unsigned int pages_per_huge_page)
5142 unsigned long addr = addr_hint &
5143 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5145 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5146 clear_gigantic_page(page, addr, pages_per_huge_page);
5150 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5153 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5155 struct vm_area_struct *vma,
5156 unsigned int pages_per_huge_page)
5159 struct page *dst_base = dst;
5160 struct page *src_base = src;
5162 for (i = 0; i < pages_per_huge_page; ) {
5164 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5167 dst = mem_map_next(dst, dst_base, i);
5168 src = mem_map_next(src, src_base, i);
5172 struct copy_subpage_arg {
5175 struct vm_area_struct *vma;
5178 static void copy_subpage(unsigned long addr, int idx, void *arg)
5180 struct copy_subpage_arg *copy_arg = arg;
5182 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5183 addr, copy_arg->vma);
5186 void copy_user_huge_page(struct page *dst, struct page *src,
5187 unsigned long addr_hint, struct vm_area_struct *vma,
5188 unsigned int pages_per_huge_page)
5190 unsigned long addr = addr_hint &
5191 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5192 struct copy_subpage_arg arg = {
5198 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5199 copy_user_gigantic_page(dst, src, addr, vma,
5200 pages_per_huge_page);
5204 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5207 long copy_huge_page_from_user(struct page *dst_page,
5208 const void __user *usr_src,
5209 unsigned int pages_per_huge_page,
5210 bool allow_pagefault)
5212 void *src = (void *)usr_src;
5214 unsigned long i, rc = 0;
5215 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5217 for (i = 0; i < pages_per_huge_page; i++) {
5218 if (allow_pagefault)
5219 page_kaddr = kmap(dst_page + i);
5221 page_kaddr = kmap_atomic(dst_page + i);
5222 rc = copy_from_user(page_kaddr,
5223 (const void __user *)(src + i * PAGE_SIZE),
5225 if (allow_pagefault)
5226 kunmap(dst_page + i);
5228 kunmap_atomic(page_kaddr);
5230 ret_val -= (PAGE_SIZE - rc);
5238 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5240 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5242 static struct kmem_cache *page_ptl_cachep;
5244 void __init ptlock_cache_init(void)
5246 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5250 bool ptlock_alloc(struct page *page)
5254 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5261 void ptlock_free(struct page *page)
5263 kmem_cache_free(page_ptl_cachep, page->ptl);