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>
76 #include <asm/mmu_context.h>
77 #include <asm/pgalloc.h>
78 #include <linux/uaccess.h>
80 #include <asm/tlbflush.h>
81 #include <asm/pgtable.h>
85 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
86 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
89 #ifndef CONFIG_NEED_MULTIPLE_NODES
90 /* use the per-pgdat data instead for discontigmem - mbligh */
91 unsigned long max_mapnr;
92 EXPORT_SYMBOL(max_mapnr);
95 EXPORT_SYMBOL(mem_map);
99 * A number of key systems in x86 including ioremap() rely on the assumption
100 * that high_memory defines the upper bound on direct map memory, then end
101 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
102 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
106 EXPORT_SYMBOL(high_memory);
109 * Randomize the address space (stacks, mmaps, brk, etc.).
111 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
112 * as ancient (libc5 based) binaries can segfault. )
114 int randomize_va_space __read_mostly =
115 #ifdef CONFIG_COMPAT_BRK
121 #ifndef arch_faults_on_old_pte
122 static inline bool arch_faults_on_old_pte(void)
125 * Those arches which don't have hw access flag feature need to
126 * implement their own helper. By default, "true" means pagefault
127 * will be hit on old pte.
133 static int __init disable_randmaps(char *s)
135 randomize_va_space = 0;
138 __setup("norandmaps", disable_randmaps);
140 unsigned long zero_pfn __read_mostly;
141 EXPORT_SYMBOL(zero_pfn);
143 unsigned long highest_memmap_pfn __read_mostly;
146 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
148 static int __init init_zero_pfn(void)
150 zero_pfn = page_to_pfn(ZERO_PAGE(0));
153 core_initcall(init_zero_pfn);
156 #if defined(SPLIT_RSS_COUNTING)
158 void sync_mm_rss(struct mm_struct *mm)
162 for (i = 0; i < NR_MM_COUNTERS; i++) {
163 if (current->rss_stat.count[i]) {
164 add_mm_counter(mm, i, current->rss_stat.count[i]);
165 current->rss_stat.count[i] = 0;
168 current->rss_stat.events = 0;
171 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
173 struct task_struct *task = current;
175 if (likely(task->mm == mm))
176 task->rss_stat.count[member] += val;
178 add_mm_counter(mm, member, val);
180 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
181 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
183 /* sync counter once per 64 page faults */
184 #define TASK_RSS_EVENTS_THRESH (64)
185 static void check_sync_rss_stat(struct task_struct *task)
187 if (unlikely(task != current))
189 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
190 sync_mm_rss(task->mm);
192 #else /* SPLIT_RSS_COUNTING */
194 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
195 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
197 static void check_sync_rss_stat(struct task_struct *task)
201 #endif /* SPLIT_RSS_COUNTING */
204 * Note: this doesn't free the actual pages themselves. That
205 * has been handled earlier when unmapping all the memory regions.
207 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
210 pgtable_t token = pmd_pgtable(*pmd);
212 pte_free_tlb(tlb, token, addr);
213 mm_dec_nr_ptes(tlb->mm);
216 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
217 unsigned long addr, unsigned long end,
218 unsigned long floor, unsigned long ceiling)
225 pmd = pmd_offset(pud, addr);
227 next = pmd_addr_end(addr, end);
228 if (pmd_none_or_clear_bad(pmd))
230 free_pte_range(tlb, pmd, addr);
231 } while (pmd++, addr = next, addr != end);
241 if (end - 1 > ceiling - 1)
244 pmd = pmd_offset(pud, start);
246 pmd_free_tlb(tlb, pmd, start);
247 mm_dec_nr_pmds(tlb->mm);
250 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
251 unsigned long addr, unsigned long end,
252 unsigned long floor, unsigned long ceiling)
259 pud = pud_offset(p4d, addr);
261 next = pud_addr_end(addr, end);
262 if (pud_none_or_clear_bad(pud))
264 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
265 } while (pud++, addr = next, addr != end);
275 if (end - 1 > ceiling - 1)
278 pud = pud_offset(p4d, start);
280 pud_free_tlb(tlb, pud, start);
281 mm_dec_nr_puds(tlb->mm);
284 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
285 unsigned long addr, unsigned long end,
286 unsigned long floor, unsigned long ceiling)
293 p4d = p4d_offset(pgd, addr);
295 next = p4d_addr_end(addr, end);
296 if (p4d_none_or_clear_bad(p4d))
298 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
299 } while (p4d++, addr = next, addr != end);
305 ceiling &= PGDIR_MASK;
309 if (end - 1 > ceiling - 1)
312 p4d = p4d_offset(pgd, start);
314 p4d_free_tlb(tlb, p4d, start);
318 * This function frees user-level page tables of a process.
320 void free_pgd_range(struct mmu_gather *tlb,
321 unsigned long addr, unsigned long end,
322 unsigned long floor, unsigned long ceiling)
328 * The next few lines have given us lots of grief...
330 * Why are we testing PMD* at this top level? Because often
331 * there will be no work to do at all, and we'd prefer not to
332 * go all the way down to the bottom just to discover that.
334 * Why all these "- 1"s? Because 0 represents both the bottom
335 * of the address space and the top of it (using -1 for the
336 * top wouldn't help much: the masks would do the wrong thing).
337 * The rule is that addr 0 and floor 0 refer to the bottom of
338 * the address space, but end 0 and ceiling 0 refer to the top
339 * Comparisons need to use "end - 1" and "ceiling - 1" (though
340 * that end 0 case should be mythical).
342 * Wherever addr is brought up or ceiling brought down, we must
343 * be careful to reject "the opposite 0" before it confuses the
344 * subsequent tests. But what about where end is brought down
345 * by PMD_SIZE below? no, end can't go down to 0 there.
347 * Whereas we round start (addr) and ceiling down, by different
348 * masks at different levels, in order to test whether a table
349 * now has no other vmas using it, so can be freed, we don't
350 * bother to round floor or end up - the tests don't need that.
364 if (end - 1 > ceiling - 1)
369 * We add page table cache pages with PAGE_SIZE,
370 * (see pte_free_tlb()), flush the tlb if we need
372 tlb_change_page_size(tlb, PAGE_SIZE);
373 pgd = pgd_offset(tlb->mm, addr);
375 next = pgd_addr_end(addr, end);
376 if (pgd_none_or_clear_bad(pgd))
378 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
379 } while (pgd++, addr = next, addr != end);
382 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
383 unsigned long floor, unsigned long ceiling)
386 struct vm_area_struct *next = vma->vm_next;
387 unsigned long addr = vma->vm_start;
390 * Hide vma from rmap and truncate_pagecache before freeing
393 unlink_anon_vmas(vma);
394 unlink_file_vma(vma);
396 if (is_vm_hugetlb_page(vma)) {
397 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
398 floor, next ? next->vm_start : ceiling);
401 * Optimization: gather nearby vmas into one call down
403 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
404 && !is_vm_hugetlb_page(next)) {
407 unlink_anon_vmas(vma);
408 unlink_file_vma(vma);
410 free_pgd_range(tlb, addr, vma->vm_end,
411 floor, next ? next->vm_start : ceiling);
417 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
420 pgtable_t new = pte_alloc_one(mm);
425 * Ensure all pte setup (eg. pte page lock and page clearing) are
426 * visible before the pte is made visible to other CPUs by being
427 * put into page tables.
429 * The other side of the story is the pointer chasing in the page
430 * table walking code (when walking the page table without locking;
431 * ie. most of the time). Fortunately, these data accesses consist
432 * of a chain of data-dependent loads, meaning most CPUs (alpha
433 * being the notable exception) will already guarantee loads are
434 * seen in-order. See the alpha page table accessors for the
435 * smp_read_barrier_depends() barriers in page table walking code.
437 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
439 ptl = pmd_lock(mm, pmd);
440 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
442 pmd_populate(mm, pmd, new);
451 int __pte_alloc_kernel(pmd_t *pmd)
453 pte_t *new = pte_alloc_one_kernel(&init_mm);
457 smp_wmb(); /* See comment in __pte_alloc */
459 spin_lock(&init_mm.page_table_lock);
460 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
461 pmd_populate_kernel(&init_mm, pmd, new);
464 spin_unlock(&init_mm.page_table_lock);
466 pte_free_kernel(&init_mm, new);
470 static inline void init_rss_vec(int *rss)
472 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
475 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
479 if (current->mm == mm)
481 for (i = 0; i < NR_MM_COUNTERS; i++)
483 add_mm_counter(mm, i, rss[i]);
487 * This function is called to print an error when a bad pte
488 * is found. For example, we might have a PFN-mapped pte in
489 * a region that doesn't allow it.
491 * The calling function must still handle the error.
493 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
494 pte_t pte, struct page *page)
496 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
497 p4d_t *p4d = p4d_offset(pgd, addr);
498 pud_t *pud = pud_offset(p4d, addr);
499 pmd_t *pmd = pmd_offset(pud, addr);
500 struct address_space *mapping;
502 static unsigned long resume;
503 static unsigned long nr_shown;
504 static unsigned long nr_unshown;
507 * Allow a burst of 60 reports, then keep quiet for that minute;
508 * or allow a steady drip of one report per second.
510 if (nr_shown == 60) {
511 if (time_before(jiffies, resume)) {
516 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
523 resume = jiffies + 60 * HZ;
525 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
526 index = linear_page_index(vma, addr);
528 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
530 (long long)pte_val(pte), (long long)pmd_val(*pmd));
532 dump_page(page, "bad pte");
533 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
534 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
535 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
537 vma->vm_ops ? vma->vm_ops->fault : NULL,
538 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
539 mapping ? mapping->a_ops->readpage : NULL);
541 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
545 * vm_normal_page -- This function gets the "struct page" associated with a pte.
547 * "Special" mappings do not wish to be associated with a "struct page" (either
548 * it doesn't exist, or it exists but they don't want to touch it). In this
549 * case, NULL is returned here. "Normal" mappings do have a struct page.
551 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
552 * pte bit, in which case this function is trivial. Secondly, an architecture
553 * may not have a spare pte bit, which requires a more complicated scheme,
556 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
557 * special mapping (even if there are underlying and valid "struct pages").
558 * COWed pages of a VM_PFNMAP are always normal.
560 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
561 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
562 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
563 * mapping will always honor the rule
565 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
567 * And for normal mappings this is false.
569 * This restricts such mappings to be a linear translation from virtual address
570 * to pfn. To get around this restriction, we allow arbitrary mappings so long
571 * as the vma is not a COW mapping; in that case, we know that all ptes are
572 * special (because none can have been COWed).
575 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
577 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
578 * page" backing, however the difference is that _all_ pages with a struct
579 * page (that is, those where pfn_valid is true) are refcounted and considered
580 * normal pages by the VM. The disadvantage is that pages are refcounted
581 * (which can be slower and simply not an option for some PFNMAP users). The
582 * advantage is that we don't have to follow the strict linearity rule of
583 * PFNMAP mappings in order to support COWable mappings.
586 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
589 unsigned long pfn = pte_pfn(pte);
591 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
592 if (likely(!pte_special(pte)))
594 if (vma->vm_ops && vma->vm_ops->find_special_page)
595 return vma->vm_ops->find_special_page(vma, addr);
596 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
598 if (is_zero_pfn(pfn))
603 print_bad_pte(vma, addr, pte, NULL);
607 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
609 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
610 if (vma->vm_flags & VM_MIXEDMAP) {
616 off = (addr - vma->vm_start) >> PAGE_SHIFT;
617 if (pfn == vma->vm_pgoff + off)
619 if (!is_cow_mapping(vma->vm_flags))
624 if (is_zero_pfn(pfn))
628 if (unlikely(pfn > highest_memmap_pfn)) {
629 print_bad_pte(vma, addr, pte, NULL);
634 * NOTE! We still have PageReserved() pages in the page tables.
635 * eg. VDSO mappings can cause them to exist.
638 return pfn_to_page(pfn);
641 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
642 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
645 unsigned long pfn = pmd_pfn(pmd);
648 * There is no pmd_special() but there may be special pmds, e.g.
649 * in a direct-access (dax) mapping, so let's just replicate the
650 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
652 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
653 if (vma->vm_flags & VM_MIXEDMAP) {
659 off = (addr - vma->vm_start) >> PAGE_SHIFT;
660 if (pfn == vma->vm_pgoff + off)
662 if (!is_cow_mapping(vma->vm_flags))
669 if (is_zero_pfn(pfn))
671 if (unlikely(pfn > highest_memmap_pfn))
675 * NOTE! We still have PageReserved() pages in the page tables.
676 * eg. VDSO mappings can cause them to exist.
679 return pfn_to_page(pfn);
684 * copy one vm_area from one task to the other. Assumes the page tables
685 * already present in the new task to be cleared in the whole range
686 * covered by this vma.
689 static inline unsigned long
690 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
691 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
692 unsigned long addr, int *rss)
694 unsigned long vm_flags = vma->vm_flags;
695 pte_t pte = *src_pte;
698 /* pte contains position in swap or file, so copy. */
699 if (unlikely(!pte_present(pte))) {
700 swp_entry_t entry = pte_to_swp_entry(pte);
702 if (likely(!non_swap_entry(entry))) {
703 if (swap_duplicate(entry) < 0)
706 /* make sure dst_mm is on swapoff's mmlist. */
707 if (unlikely(list_empty(&dst_mm->mmlist))) {
708 spin_lock(&mmlist_lock);
709 if (list_empty(&dst_mm->mmlist))
710 list_add(&dst_mm->mmlist,
712 spin_unlock(&mmlist_lock);
715 } else if (is_migration_entry(entry)) {
716 page = migration_entry_to_page(entry);
718 rss[mm_counter(page)]++;
720 if (is_write_migration_entry(entry) &&
721 is_cow_mapping(vm_flags)) {
723 * COW mappings require pages in both
724 * parent and child to be set to read.
726 make_migration_entry_read(&entry);
727 pte = swp_entry_to_pte(entry);
728 if (pte_swp_soft_dirty(*src_pte))
729 pte = pte_swp_mksoft_dirty(pte);
730 set_pte_at(src_mm, addr, src_pte, pte);
732 } else if (is_device_private_entry(entry)) {
733 page = device_private_entry_to_page(entry);
736 * Update rss count even for unaddressable pages, as
737 * they should treated just like normal pages in this
740 * We will likely want to have some new rss counters
741 * for unaddressable pages, at some point. But for now
742 * keep things as they are.
745 rss[mm_counter(page)]++;
746 page_dup_rmap(page, false);
749 * We do not preserve soft-dirty information, because so
750 * far, checkpoint/restore is the only feature that
751 * requires that. And checkpoint/restore does not work
752 * when a device driver is involved (you cannot easily
753 * save and restore device driver state).
755 if (is_write_device_private_entry(entry) &&
756 is_cow_mapping(vm_flags)) {
757 make_device_private_entry_read(&entry);
758 pte = swp_entry_to_pte(entry);
759 set_pte_at(src_mm, addr, src_pte, pte);
766 * If it's a COW mapping, write protect it both
767 * in the parent and the child
769 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
770 ptep_set_wrprotect(src_mm, addr, src_pte);
771 pte = pte_wrprotect(pte);
775 * If it's a shared mapping, mark it clean in
778 if (vm_flags & VM_SHARED)
779 pte = pte_mkclean(pte);
780 pte = pte_mkold(pte);
782 page = vm_normal_page(vma, addr, pte);
785 page_dup_rmap(page, false);
786 rss[mm_counter(page)]++;
787 } else if (pte_devmap(pte)) {
788 page = pte_page(pte);
792 set_pte_at(dst_mm, addr, dst_pte, pte);
796 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
797 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
798 unsigned long addr, unsigned long end)
800 pte_t *orig_src_pte, *orig_dst_pte;
801 pte_t *src_pte, *dst_pte;
802 spinlock_t *src_ptl, *dst_ptl;
804 int rss[NR_MM_COUNTERS];
805 swp_entry_t entry = (swp_entry_t){0};
810 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
813 src_pte = pte_offset_map(src_pmd, addr);
814 src_ptl = pte_lockptr(src_mm, src_pmd);
815 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
816 orig_src_pte = src_pte;
817 orig_dst_pte = dst_pte;
818 arch_enter_lazy_mmu_mode();
822 * We are holding two locks at this point - either of them
823 * could generate latencies in another task on another CPU.
825 if (progress >= 32) {
827 if (need_resched() ||
828 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
831 if (pte_none(*src_pte)) {
835 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
840 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
842 arch_leave_lazy_mmu_mode();
843 spin_unlock(src_ptl);
844 pte_unmap(orig_src_pte);
845 add_mm_rss_vec(dst_mm, rss);
846 pte_unmap_unlock(orig_dst_pte, dst_ptl);
850 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
859 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
860 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
861 unsigned long addr, unsigned long end)
863 pmd_t *src_pmd, *dst_pmd;
866 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
869 src_pmd = pmd_offset(src_pud, addr);
871 next = pmd_addr_end(addr, end);
872 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
873 || pmd_devmap(*src_pmd)) {
875 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
876 err = copy_huge_pmd(dst_mm, src_mm,
877 dst_pmd, src_pmd, addr, vma);
884 if (pmd_none_or_clear_bad(src_pmd))
886 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
889 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
893 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
894 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
895 unsigned long addr, unsigned long end)
897 pud_t *src_pud, *dst_pud;
900 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
903 src_pud = pud_offset(src_p4d, addr);
905 next = pud_addr_end(addr, end);
906 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
909 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
910 err = copy_huge_pud(dst_mm, src_mm,
911 dst_pud, src_pud, addr, vma);
918 if (pud_none_or_clear_bad(src_pud))
920 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
923 } while (dst_pud++, src_pud++, addr = next, addr != end);
927 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
928 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
929 unsigned long addr, unsigned long end)
931 p4d_t *src_p4d, *dst_p4d;
934 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
937 src_p4d = p4d_offset(src_pgd, addr);
939 next = p4d_addr_end(addr, end);
940 if (p4d_none_or_clear_bad(src_p4d))
942 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
945 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
949 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
950 struct vm_area_struct *vma)
952 pgd_t *src_pgd, *dst_pgd;
954 unsigned long addr = vma->vm_start;
955 unsigned long end = vma->vm_end;
956 struct mmu_notifier_range range;
961 * Don't copy ptes where a page fault will fill them correctly.
962 * Fork becomes much lighter when there are big shared or private
963 * readonly mappings. The tradeoff is that copy_page_range is more
964 * efficient than faulting.
966 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
970 if (is_vm_hugetlb_page(vma))
971 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
973 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
975 * We do not free on error cases below as remove_vma
976 * gets called on error from higher level routine
978 ret = track_pfn_copy(vma);
984 * We need to invalidate the secondary MMU mappings only when
985 * there could be a permission downgrade on the ptes of the
986 * parent mm. And a permission downgrade will only happen if
987 * is_cow_mapping() returns true.
989 is_cow = is_cow_mapping(vma->vm_flags);
992 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
993 0, vma, src_mm, addr, end);
994 mmu_notifier_invalidate_range_start(&range);
998 dst_pgd = pgd_offset(dst_mm, addr);
999 src_pgd = pgd_offset(src_mm, addr);
1001 next = pgd_addr_end(addr, end);
1002 if (pgd_none_or_clear_bad(src_pgd))
1004 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1005 vma, addr, next))) {
1009 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1012 mmu_notifier_invalidate_range_end(&range);
1016 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1017 struct vm_area_struct *vma, pmd_t *pmd,
1018 unsigned long addr, unsigned long end,
1019 struct zap_details *details)
1021 struct mm_struct *mm = tlb->mm;
1022 int force_flush = 0;
1023 int rss[NR_MM_COUNTERS];
1029 tlb_change_page_size(tlb, PAGE_SIZE);
1032 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1034 flush_tlb_batched_pending(mm);
1035 arch_enter_lazy_mmu_mode();
1038 if (pte_none(ptent))
1044 if (pte_present(ptent)) {
1047 page = vm_normal_page(vma, addr, ptent);
1048 if (unlikely(details) && page) {
1050 * unmap_shared_mapping_pages() wants to
1051 * invalidate cache without truncating:
1052 * unmap shared but keep private pages.
1054 if (details->check_mapping &&
1055 details->check_mapping != page_rmapping(page))
1058 ptent = ptep_get_and_clear_full(mm, addr, pte,
1060 tlb_remove_tlb_entry(tlb, pte, addr);
1061 if (unlikely(!page))
1064 if (!PageAnon(page)) {
1065 if (pte_dirty(ptent)) {
1067 set_page_dirty(page);
1069 if (pte_young(ptent) &&
1070 likely(!(vma->vm_flags & VM_SEQ_READ)))
1071 mark_page_accessed(page);
1073 rss[mm_counter(page)]--;
1074 page_remove_rmap(page, false);
1075 if (unlikely(page_mapcount(page) < 0))
1076 print_bad_pte(vma, addr, ptent, page);
1077 if (unlikely(__tlb_remove_page(tlb, page))) {
1085 entry = pte_to_swp_entry(ptent);
1086 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1087 struct page *page = device_private_entry_to_page(entry);
1089 if (unlikely(details && details->check_mapping)) {
1091 * unmap_shared_mapping_pages() wants to
1092 * invalidate cache without truncating:
1093 * unmap shared but keep private pages.
1095 if (details->check_mapping !=
1096 page_rmapping(page))
1100 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1101 rss[mm_counter(page)]--;
1102 page_remove_rmap(page, false);
1107 /* If details->check_mapping, we leave swap entries. */
1108 if (unlikely(details))
1111 if (!non_swap_entry(entry))
1113 else if (is_migration_entry(entry)) {
1116 page = migration_entry_to_page(entry);
1117 rss[mm_counter(page)]--;
1119 if (unlikely(!free_swap_and_cache(entry)))
1120 print_bad_pte(vma, addr, ptent, NULL);
1121 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1122 } while (pte++, addr += PAGE_SIZE, addr != end);
1124 add_mm_rss_vec(mm, rss);
1125 arch_leave_lazy_mmu_mode();
1127 /* Do the actual TLB flush before dropping ptl */
1129 tlb_flush_mmu_tlbonly(tlb);
1130 pte_unmap_unlock(start_pte, ptl);
1133 * If we forced a TLB flush (either due to running out of
1134 * batch buffers or because we needed to flush dirty TLB
1135 * entries before releasing the ptl), free the batched
1136 * memory too. Restart if we didn't do everything.
1151 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1152 struct vm_area_struct *vma, pud_t *pud,
1153 unsigned long addr, unsigned long end,
1154 struct zap_details *details)
1159 pmd = pmd_offset(pud, addr);
1161 next = pmd_addr_end(addr, end);
1162 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1163 if (next - addr != HPAGE_PMD_SIZE)
1164 __split_huge_pmd(vma, pmd, addr, false, NULL);
1165 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1170 * Here there can be other concurrent MADV_DONTNEED or
1171 * trans huge page faults running, and if the pmd is
1172 * none or trans huge it can change under us. This is
1173 * because MADV_DONTNEED holds the mmap_sem in read
1176 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1178 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1181 } while (pmd++, addr = next, addr != end);
1186 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1187 struct vm_area_struct *vma, p4d_t *p4d,
1188 unsigned long addr, unsigned long end,
1189 struct zap_details *details)
1194 pud = pud_offset(p4d, addr);
1196 next = pud_addr_end(addr, end);
1197 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1198 if (next - addr != HPAGE_PUD_SIZE) {
1199 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1200 split_huge_pud(vma, pud, addr);
1201 } else if (zap_huge_pud(tlb, vma, pud, addr))
1205 if (pud_none_or_clear_bad(pud))
1207 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1210 } while (pud++, addr = next, addr != end);
1215 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1216 struct vm_area_struct *vma, pgd_t *pgd,
1217 unsigned long addr, unsigned long end,
1218 struct zap_details *details)
1223 p4d = p4d_offset(pgd, addr);
1225 next = p4d_addr_end(addr, end);
1226 if (p4d_none_or_clear_bad(p4d))
1228 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1229 } while (p4d++, addr = next, addr != end);
1234 void unmap_page_range(struct mmu_gather *tlb,
1235 struct vm_area_struct *vma,
1236 unsigned long addr, unsigned long end,
1237 struct zap_details *details)
1242 BUG_ON(addr >= end);
1243 tlb_start_vma(tlb, vma);
1244 pgd = pgd_offset(vma->vm_mm, addr);
1246 next = pgd_addr_end(addr, end);
1247 if (pgd_none_or_clear_bad(pgd))
1249 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1250 } while (pgd++, addr = next, addr != end);
1251 tlb_end_vma(tlb, vma);
1255 static void unmap_single_vma(struct mmu_gather *tlb,
1256 struct vm_area_struct *vma, unsigned long start_addr,
1257 unsigned long end_addr,
1258 struct zap_details *details)
1260 unsigned long start = max(vma->vm_start, start_addr);
1263 if (start >= vma->vm_end)
1265 end = min(vma->vm_end, end_addr);
1266 if (end <= vma->vm_start)
1270 uprobe_munmap(vma, start, end);
1272 if (unlikely(vma->vm_flags & VM_PFNMAP))
1273 untrack_pfn(vma, 0, 0);
1276 if (unlikely(is_vm_hugetlb_page(vma))) {
1278 * It is undesirable to test vma->vm_file as it
1279 * should be non-null for valid hugetlb area.
1280 * However, vm_file will be NULL in the error
1281 * cleanup path of mmap_region. When
1282 * hugetlbfs ->mmap method fails,
1283 * mmap_region() nullifies vma->vm_file
1284 * before calling this function to clean up.
1285 * Since no pte has actually been setup, it is
1286 * safe to do nothing in this case.
1289 i_mmap_lock_write(vma->vm_file->f_mapping);
1290 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1291 i_mmap_unlock_write(vma->vm_file->f_mapping);
1294 unmap_page_range(tlb, vma, start, end, details);
1299 * unmap_vmas - unmap a range of memory covered by a list of vma's
1300 * @tlb: address of the caller's struct mmu_gather
1301 * @vma: the starting vma
1302 * @start_addr: virtual address at which to start unmapping
1303 * @end_addr: virtual address at which to end unmapping
1305 * Unmap all pages in the vma list.
1307 * Only addresses between `start' and `end' will be unmapped.
1309 * The VMA list must be sorted in ascending virtual address order.
1311 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1312 * range after unmap_vmas() returns. So the only responsibility here is to
1313 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1314 * drops the lock and schedules.
1316 void unmap_vmas(struct mmu_gather *tlb,
1317 struct vm_area_struct *vma, unsigned long start_addr,
1318 unsigned long end_addr)
1320 struct mmu_notifier_range range;
1322 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1323 start_addr, end_addr);
1324 mmu_notifier_invalidate_range_start(&range);
1325 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1326 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1327 mmu_notifier_invalidate_range_end(&range);
1331 * zap_page_range - remove user pages in a given range
1332 * @vma: vm_area_struct holding the applicable pages
1333 * @start: starting address of pages to zap
1334 * @size: number of bytes to zap
1336 * Caller must protect the VMA list
1338 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1341 struct mmu_notifier_range range;
1342 struct mmu_gather tlb;
1345 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1346 start, start + size);
1347 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1348 update_hiwater_rss(vma->vm_mm);
1349 mmu_notifier_invalidate_range_start(&range);
1350 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1351 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1352 mmu_notifier_invalidate_range_end(&range);
1353 tlb_finish_mmu(&tlb, start, range.end);
1357 * zap_page_range_single - remove user pages in a given range
1358 * @vma: vm_area_struct holding the applicable pages
1359 * @address: starting address of pages to zap
1360 * @size: number of bytes to zap
1361 * @details: details of shared cache invalidation
1363 * The range must fit into one VMA.
1365 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1366 unsigned long size, struct zap_details *details)
1368 struct mmu_notifier_range range;
1369 struct mmu_gather tlb;
1372 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1373 address, address + size);
1374 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1375 update_hiwater_rss(vma->vm_mm);
1376 mmu_notifier_invalidate_range_start(&range);
1377 unmap_single_vma(&tlb, vma, address, range.end, details);
1378 mmu_notifier_invalidate_range_end(&range);
1379 tlb_finish_mmu(&tlb, address, range.end);
1383 * zap_vma_ptes - remove ptes mapping the vma
1384 * @vma: vm_area_struct holding ptes to be zapped
1385 * @address: starting address of pages to zap
1386 * @size: number of bytes to zap
1388 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1390 * The entire address range must be fully contained within the vma.
1393 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1396 if (address < vma->vm_start || address + size > vma->vm_end ||
1397 !(vma->vm_flags & VM_PFNMAP))
1400 zap_page_range_single(vma, address, size, NULL);
1402 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1404 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1412 pgd = pgd_offset(mm, addr);
1413 p4d = p4d_alloc(mm, pgd, addr);
1416 pud = pud_alloc(mm, p4d, addr);
1419 pmd = pmd_alloc(mm, pud, addr);
1423 VM_BUG_ON(pmd_trans_huge(*pmd));
1424 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1428 * This is the old fallback for page remapping.
1430 * For historical reasons, it only allows reserved pages. Only
1431 * old drivers should use this, and they needed to mark their
1432 * pages reserved for the old functions anyway.
1434 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1435 struct page *page, pgprot_t prot)
1437 struct mm_struct *mm = vma->vm_mm;
1443 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1446 flush_dcache_page(page);
1447 pte = get_locked_pte(mm, addr, &ptl);
1451 if (!pte_none(*pte))
1454 /* Ok, finally just insert the thing.. */
1456 inc_mm_counter_fast(mm, mm_counter_file(page));
1457 page_add_file_rmap(page, false);
1458 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1462 pte_unmap_unlock(pte, ptl);
1468 * vm_insert_page - insert single page into user vma
1469 * @vma: user vma to map to
1470 * @addr: target user address of this page
1471 * @page: source kernel page
1473 * This allows drivers to insert individual pages they've allocated
1476 * The page has to be a nice clean _individual_ kernel allocation.
1477 * If you allocate a compound page, you need to have marked it as
1478 * such (__GFP_COMP), or manually just split the page up yourself
1479 * (see split_page()).
1481 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1482 * took an arbitrary page protection parameter. This doesn't allow
1483 * that. Your vma protection will have to be set up correctly, which
1484 * means that if you want a shared writable mapping, you'd better
1485 * ask for a shared writable mapping!
1487 * The page does not need to be reserved.
1489 * Usually this function is called from f_op->mmap() handler
1490 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1491 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1492 * function from other places, for example from page-fault handler.
1494 * Return: %0 on success, negative error code otherwise.
1496 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1499 if (addr < vma->vm_start || addr >= vma->vm_end)
1501 if (!page_count(page))
1503 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1504 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1505 BUG_ON(vma->vm_flags & VM_PFNMAP);
1506 vma->vm_flags |= VM_MIXEDMAP;
1508 return insert_page(vma, addr, page, vma->vm_page_prot);
1510 EXPORT_SYMBOL(vm_insert_page);
1513 * __vm_map_pages - maps range of kernel pages into user vma
1514 * @vma: user vma to map to
1515 * @pages: pointer to array of source kernel pages
1516 * @num: number of pages in page array
1517 * @offset: user's requested vm_pgoff
1519 * This allows drivers to map range of kernel pages into a user vma.
1521 * Return: 0 on success and error code otherwise.
1523 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1524 unsigned long num, unsigned long offset)
1526 unsigned long count = vma_pages(vma);
1527 unsigned long uaddr = vma->vm_start;
1530 /* Fail if the user requested offset is beyond the end of the object */
1534 /* Fail if the user requested size exceeds available object size */
1535 if (count > num - offset)
1538 for (i = 0; i < count; i++) {
1539 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1549 * vm_map_pages - maps range of kernel pages starts with non zero offset
1550 * @vma: user vma to map to
1551 * @pages: pointer to array of source kernel pages
1552 * @num: number of pages in page array
1554 * Maps an object consisting of @num pages, catering for the user's
1555 * requested vm_pgoff
1557 * If we fail to insert any page into the vma, the function will return
1558 * immediately leaving any previously inserted pages present. Callers
1559 * from the mmap handler may immediately return the error as their caller
1560 * will destroy the vma, removing any successfully inserted pages. Other
1561 * callers should make their own arrangements for calling unmap_region().
1563 * Context: Process context. Called by mmap handlers.
1564 * Return: 0 on success and error code otherwise.
1566 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1569 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1571 EXPORT_SYMBOL(vm_map_pages);
1574 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1575 * @vma: user vma to map to
1576 * @pages: pointer to array of source kernel pages
1577 * @num: number of pages in page array
1579 * Similar to vm_map_pages(), except that it explicitly sets the offset
1580 * to 0. This function is intended for the drivers that did not consider
1583 * Context: Process context. Called by mmap handlers.
1584 * Return: 0 on success and error code otherwise.
1586 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1589 return __vm_map_pages(vma, pages, num, 0);
1591 EXPORT_SYMBOL(vm_map_pages_zero);
1593 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1594 pfn_t pfn, pgprot_t prot, bool mkwrite)
1596 struct mm_struct *mm = vma->vm_mm;
1600 pte = get_locked_pte(mm, addr, &ptl);
1602 return VM_FAULT_OOM;
1603 if (!pte_none(*pte)) {
1606 * For read faults on private mappings the PFN passed
1607 * in may not match the PFN we have mapped if the
1608 * mapped PFN is a writeable COW page. In the mkwrite
1609 * case we are creating a writable PTE for a shared
1610 * mapping and we expect the PFNs to match. If they
1611 * don't match, we are likely racing with block
1612 * allocation and mapping invalidation so just skip the
1615 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1616 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1619 entry = pte_mkyoung(*pte);
1620 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1621 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1622 update_mmu_cache(vma, addr, pte);
1627 /* Ok, finally just insert the thing.. */
1628 if (pfn_t_devmap(pfn))
1629 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1631 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1634 entry = pte_mkyoung(entry);
1635 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1638 set_pte_at(mm, addr, pte, entry);
1639 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1642 pte_unmap_unlock(pte, ptl);
1643 return VM_FAULT_NOPAGE;
1647 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1648 * @vma: user vma to map to
1649 * @addr: target user address of this page
1650 * @pfn: source kernel pfn
1651 * @pgprot: pgprot flags for the inserted page
1653 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1654 * to override pgprot on a per-page basis.
1656 * This only makes sense for IO mappings, and it makes no sense for
1657 * COW mappings. In general, using multiple vmas is preferable;
1658 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1661 * Context: Process context. May allocate using %GFP_KERNEL.
1662 * Return: vm_fault_t value.
1664 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1665 unsigned long pfn, pgprot_t pgprot)
1668 * Technically, architectures with pte_special can avoid all these
1669 * restrictions (same for remap_pfn_range). However we would like
1670 * consistency in testing and feature parity among all, so we should
1671 * try to keep these invariants in place for everybody.
1673 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1674 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1675 (VM_PFNMAP|VM_MIXEDMAP));
1676 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1677 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1679 if (addr < vma->vm_start || addr >= vma->vm_end)
1680 return VM_FAULT_SIGBUS;
1682 if (!pfn_modify_allowed(pfn, pgprot))
1683 return VM_FAULT_SIGBUS;
1685 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1687 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1690 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1693 * vmf_insert_pfn - insert single pfn into user vma
1694 * @vma: user vma to map to
1695 * @addr: target user address of this page
1696 * @pfn: source kernel pfn
1698 * Similar to vm_insert_page, this allows drivers to insert individual pages
1699 * they've allocated into a user vma. Same comments apply.
1701 * This function should only be called from a vm_ops->fault handler, and
1702 * in that case the handler should return the result of this function.
1704 * vma cannot be a COW mapping.
1706 * As this is called only for pages that do not currently exist, we
1707 * do not need to flush old virtual caches or the TLB.
1709 * Context: Process context. May allocate using %GFP_KERNEL.
1710 * Return: vm_fault_t value.
1712 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1715 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1717 EXPORT_SYMBOL(vmf_insert_pfn);
1719 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1721 /* these checks mirror the abort conditions in vm_normal_page */
1722 if (vma->vm_flags & VM_MIXEDMAP)
1724 if (pfn_t_devmap(pfn))
1726 if (pfn_t_special(pfn))
1728 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1733 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1734 unsigned long addr, pfn_t pfn, bool mkwrite)
1736 pgprot_t pgprot = vma->vm_page_prot;
1739 BUG_ON(!vm_mixed_ok(vma, pfn));
1741 if (addr < vma->vm_start || addr >= vma->vm_end)
1742 return VM_FAULT_SIGBUS;
1744 track_pfn_insert(vma, &pgprot, pfn);
1746 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1747 return VM_FAULT_SIGBUS;
1750 * If we don't have pte special, then we have to use the pfn_valid()
1751 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1752 * refcount the page if pfn_valid is true (hence insert_page rather
1753 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1754 * without pte special, it would there be refcounted as a normal page.
1756 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1757 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1761 * At this point we are committed to insert_page()
1762 * regardless of whether the caller specified flags that
1763 * result in pfn_t_has_page() == false.
1765 page = pfn_to_page(pfn_t_to_pfn(pfn));
1766 err = insert_page(vma, addr, page, pgprot);
1768 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1772 return VM_FAULT_OOM;
1773 if (err < 0 && err != -EBUSY)
1774 return VM_FAULT_SIGBUS;
1776 return VM_FAULT_NOPAGE;
1779 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1782 return __vm_insert_mixed(vma, addr, pfn, false);
1784 EXPORT_SYMBOL(vmf_insert_mixed);
1787 * If the insertion of PTE failed because someone else already added a
1788 * different entry in the mean time, we treat that as success as we assume
1789 * the same entry was actually inserted.
1791 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1792 unsigned long addr, pfn_t pfn)
1794 return __vm_insert_mixed(vma, addr, pfn, true);
1796 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1799 * maps a range of physical memory into the requested pages. the old
1800 * mappings are removed. any references to nonexistent pages results
1801 * in null mappings (currently treated as "copy-on-access")
1803 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1804 unsigned long addr, unsigned long end,
1805 unsigned long pfn, pgprot_t prot)
1811 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1814 arch_enter_lazy_mmu_mode();
1816 BUG_ON(!pte_none(*pte));
1817 if (!pfn_modify_allowed(pfn, prot)) {
1821 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1823 } while (pte++, addr += PAGE_SIZE, addr != end);
1824 arch_leave_lazy_mmu_mode();
1825 pte_unmap_unlock(pte - 1, ptl);
1829 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1830 unsigned long addr, unsigned long end,
1831 unsigned long pfn, pgprot_t prot)
1837 pfn -= addr >> PAGE_SHIFT;
1838 pmd = pmd_alloc(mm, pud, addr);
1841 VM_BUG_ON(pmd_trans_huge(*pmd));
1843 next = pmd_addr_end(addr, end);
1844 err = remap_pte_range(mm, pmd, addr, next,
1845 pfn + (addr >> PAGE_SHIFT), prot);
1848 } while (pmd++, addr = next, addr != end);
1852 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1853 unsigned long addr, unsigned long end,
1854 unsigned long pfn, pgprot_t prot)
1860 pfn -= addr >> PAGE_SHIFT;
1861 pud = pud_alloc(mm, p4d, addr);
1865 next = pud_addr_end(addr, end);
1866 err = remap_pmd_range(mm, pud, addr, next,
1867 pfn + (addr >> PAGE_SHIFT), prot);
1870 } while (pud++, addr = next, addr != end);
1874 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1875 unsigned long addr, unsigned long end,
1876 unsigned long pfn, pgprot_t prot)
1882 pfn -= addr >> PAGE_SHIFT;
1883 p4d = p4d_alloc(mm, pgd, addr);
1887 next = p4d_addr_end(addr, end);
1888 err = remap_pud_range(mm, p4d, addr, next,
1889 pfn + (addr >> PAGE_SHIFT), prot);
1892 } while (p4d++, addr = next, addr != end);
1897 * remap_pfn_range - remap kernel memory to userspace
1898 * @vma: user vma to map to
1899 * @addr: target user address to start at
1900 * @pfn: physical address of kernel memory
1901 * @size: size of map area
1902 * @prot: page protection flags for this mapping
1904 * Note: this is only safe if the mm semaphore is held when called.
1906 * Return: %0 on success, negative error code otherwise.
1908 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1909 unsigned long pfn, unsigned long size, pgprot_t prot)
1913 unsigned long end = addr + PAGE_ALIGN(size);
1914 struct mm_struct *mm = vma->vm_mm;
1915 unsigned long remap_pfn = pfn;
1919 * Physically remapped pages are special. Tell the
1920 * rest of the world about it:
1921 * VM_IO tells people not to look at these pages
1922 * (accesses can have side effects).
1923 * VM_PFNMAP tells the core MM that the base pages are just
1924 * raw PFN mappings, and do not have a "struct page" associated
1927 * Disable vma merging and expanding with mremap().
1929 * Omit vma from core dump, even when VM_IO turned off.
1931 * There's a horrible special case to handle copy-on-write
1932 * behaviour that some programs depend on. We mark the "original"
1933 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1934 * See vm_normal_page() for details.
1936 if (is_cow_mapping(vma->vm_flags)) {
1937 if (addr != vma->vm_start || end != vma->vm_end)
1939 vma->vm_pgoff = pfn;
1942 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1946 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1948 BUG_ON(addr >= end);
1949 pfn -= addr >> PAGE_SHIFT;
1950 pgd = pgd_offset(mm, addr);
1951 flush_cache_range(vma, addr, end);
1953 next = pgd_addr_end(addr, end);
1954 err = remap_p4d_range(mm, pgd, addr, next,
1955 pfn + (addr >> PAGE_SHIFT), prot);
1958 } while (pgd++, addr = next, addr != end);
1961 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1965 EXPORT_SYMBOL(remap_pfn_range);
1968 * vm_iomap_memory - remap memory to userspace
1969 * @vma: user vma to map to
1970 * @start: start of area
1971 * @len: size of area
1973 * This is a simplified io_remap_pfn_range() for common driver use. The
1974 * driver just needs to give us the physical memory range to be mapped,
1975 * we'll figure out the rest from the vma information.
1977 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1978 * whatever write-combining details or similar.
1980 * Return: %0 on success, negative error code otherwise.
1982 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1984 unsigned long vm_len, pfn, pages;
1986 /* Check that the physical memory area passed in looks valid */
1987 if (start + len < start)
1990 * You *really* shouldn't map things that aren't page-aligned,
1991 * but we've historically allowed it because IO memory might
1992 * just have smaller alignment.
1994 len += start & ~PAGE_MASK;
1995 pfn = start >> PAGE_SHIFT;
1996 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1997 if (pfn + pages < pfn)
2000 /* We start the mapping 'vm_pgoff' pages into the area */
2001 if (vma->vm_pgoff > pages)
2003 pfn += vma->vm_pgoff;
2004 pages -= vma->vm_pgoff;
2006 /* Can we fit all of the mapping? */
2007 vm_len = vma->vm_end - vma->vm_start;
2008 if (vm_len >> PAGE_SHIFT > pages)
2011 /* Ok, let it rip */
2012 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2014 EXPORT_SYMBOL(vm_iomap_memory);
2016 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2017 unsigned long addr, unsigned long end,
2018 pte_fn_t fn, void *data)
2022 spinlock_t *uninitialized_var(ptl);
2024 pte = (mm == &init_mm) ?
2025 pte_alloc_kernel(pmd, addr) :
2026 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2030 BUG_ON(pmd_huge(*pmd));
2032 arch_enter_lazy_mmu_mode();
2035 err = fn(pte++, addr, data);
2038 } while (addr += PAGE_SIZE, addr != end);
2040 arch_leave_lazy_mmu_mode();
2043 pte_unmap_unlock(pte-1, ptl);
2047 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2048 unsigned long addr, unsigned long end,
2049 pte_fn_t fn, void *data)
2055 BUG_ON(pud_huge(*pud));
2057 pmd = pmd_alloc(mm, pud, addr);
2061 next = pmd_addr_end(addr, end);
2062 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2065 } while (pmd++, addr = next, addr != end);
2069 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2070 unsigned long addr, unsigned long end,
2071 pte_fn_t fn, void *data)
2077 pud = pud_alloc(mm, p4d, addr);
2081 next = pud_addr_end(addr, end);
2082 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2085 } while (pud++, addr = next, addr != end);
2089 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2090 unsigned long addr, unsigned long end,
2091 pte_fn_t fn, void *data)
2097 p4d = p4d_alloc(mm, pgd, addr);
2101 next = p4d_addr_end(addr, end);
2102 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2105 } while (p4d++, addr = next, addr != end);
2110 * Scan a region of virtual memory, filling in page tables as necessary
2111 * and calling a provided function on each leaf page table.
2113 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2114 unsigned long size, pte_fn_t fn, void *data)
2118 unsigned long end = addr + size;
2121 if (WARN_ON(addr >= end))
2124 pgd = pgd_offset(mm, addr);
2126 next = pgd_addr_end(addr, end);
2127 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2130 } while (pgd++, addr = next, addr != end);
2134 EXPORT_SYMBOL_GPL(apply_to_page_range);
2137 * handle_pte_fault chooses page fault handler according to an entry which was
2138 * read non-atomically. Before making any commitment, on those architectures
2139 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2140 * parts, do_swap_page must check under lock before unmapping the pte and
2141 * proceeding (but do_wp_page is only called after already making such a check;
2142 * and do_anonymous_page can safely check later on).
2144 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2145 pte_t *page_table, pte_t orig_pte)
2148 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2149 if (sizeof(pte_t) > sizeof(unsigned long)) {
2150 spinlock_t *ptl = pte_lockptr(mm, pmd);
2152 same = pte_same(*page_table, orig_pte);
2156 pte_unmap(page_table);
2160 static inline bool cow_user_page(struct page *dst, struct page *src,
2161 struct vm_fault *vmf)
2167 struct vm_area_struct *vma = vmf->vma;
2168 struct mm_struct *mm = vma->vm_mm;
2169 unsigned long addr = vmf->address;
2171 debug_dma_assert_idle(src);
2174 copy_user_highpage(dst, src, addr, vma);
2179 * If the source page was a PFN mapping, we don't have
2180 * a "struct page" for it. We do a best-effort copy by
2181 * just copying from the original user address. If that
2182 * fails, we just zero-fill it. Live with it.
2184 kaddr = kmap_atomic(dst);
2185 uaddr = (void __user *)(addr & PAGE_MASK);
2188 * On architectures with software "accessed" bits, we would
2189 * take a double page fault, so mark it accessed here.
2191 force_mkyoung = arch_faults_on_old_pte() && !pte_young(vmf->orig_pte);
2192 if (force_mkyoung) {
2195 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2196 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2198 * Other thread has already handled the fault
2199 * and we don't need to do anything. If it's
2200 * not the case, the fault will be triggered
2201 * again on the same address.
2207 entry = pte_mkyoung(vmf->orig_pte);
2208 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2209 update_mmu_cache(vma, addr, vmf->pte);
2213 * This really shouldn't fail, because the page is there
2214 * in the page tables. But it might just be unreadable,
2215 * in which case we just give up and fill the result with
2218 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2220 * Give a warn in case there can be some obscure
2231 pte_unmap_unlock(vmf->pte, vmf->ptl);
2232 kunmap_atomic(kaddr);
2233 flush_dcache_page(dst);
2238 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2240 struct file *vm_file = vma->vm_file;
2243 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2246 * Special mappings (e.g. VDSO) do not have any file so fake
2247 * a default GFP_KERNEL for them.
2253 * Notify the address space that the page is about to become writable so that
2254 * it can prohibit this or wait for the page to get into an appropriate state.
2256 * We do this without the lock held, so that it can sleep if it needs to.
2258 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2261 struct page *page = vmf->page;
2262 unsigned int old_flags = vmf->flags;
2264 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2266 if (vmf->vma->vm_file &&
2267 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2268 return VM_FAULT_SIGBUS;
2270 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2271 /* Restore original flags so that caller is not surprised */
2272 vmf->flags = old_flags;
2273 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2275 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2277 if (!page->mapping) {
2279 return 0; /* retry */
2281 ret |= VM_FAULT_LOCKED;
2283 VM_BUG_ON_PAGE(!PageLocked(page), page);
2288 * Handle dirtying of a page in shared file mapping on a write fault.
2290 * The function expects the page to be locked and unlocks it.
2292 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2295 struct address_space *mapping;
2297 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2299 dirtied = set_page_dirty(page);
2300 VM_BUG_ON_PAGE(PageAnon(page), page);
2302 * Take a local copy of the address_space - page.mapping may be zeroed
2303 * by truncate after unlock_page(). The address_space itself remains
2304 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2305 * release semantics to prevent the compiler from undoing this copying.
2307 mapping = page_rmapping(page);
2310 if ((dirtied || page_mkwrite) && mapping) {
2312 * Some device drivers do not set page.mapping
2313 * but still dirty their pages
2315 balance_dirty_pages_ratelimited(mapping);
2319 file_update_time(vma->vm_file);
2323 * Handle write page faults for pages that can be reused in the current vma
2325 * This can happen either due to the mapping being with the VM_SHARED flag,
2326 * or due to us being the last reference standing to the page. In either
2327 * case, all we need to do here is to mark the page as writable and update
2328 * any related book-keeping.
2330 static inline void wp_page_reuse(struct vm_fault *vmf)
2331 __releases(vmf->ptl)
2333 struct vm_area_struct *vma = vmf->vma;
2334 struct page *page = vmf->page;
2337 * Clear the pages cpupid information as the existing
2338 * information potentially belongs to a now completely
2339 * unrelated process.
2342 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2344 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2345 entry = pte_mkyoung(vmf->orig_pte);
2346 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2347 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2348 update_mmu_cache(vma, vmf->address, vmf->pte);
2349 pte_unmap_unlock(vmf->pte, vmf->ptl);
2353 * Handle the case of a page which we actually need to copy to a new page.
2355 * Called with mmap_sem locked and the old page referenced, but
2356 * without the ptl held.
2358 * High level logic flow:
2360 * - Allocate a page, copy the content of the old page to the new one.
2361 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2362 * - Take the PTL. If the pte changed, bail out and release the allocated page
2363 * - If the pte is still the way we remember it, update the page table and all
2364 * relevant references. This includes dropping the reference the page-table
2365 * held to the old page, as well as updating the rmap.
2366 * - In any case, unlock the PTL and drop the reference we took to the old page.
2368 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2370 struct vm_area_struct *vma = vmf->vma;
2371 struct mm_struct *mm = vma->vm_mm;
2372 struct page *old_page = vmf->page;
2373 struct page *new_page = NULL;
2375 int page_copied = 0;
2376 struct mem_cgroup *memcg;
2377 struct mmu_notifier_range range;
2379 if (unlikely(anon_vma_prepare(vma)))
2382 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2383 new_page = alloc_zeroed_user_highpage_movable(vma,
2388 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2393 if (!cow_user_page(new_page, old_page, vmf)) {
2395 * COW failed, if the fault was solved by other,
2396 * it's fine. If not, userspace would re-fault on
2397 * the same address and we will handle the fault
2398 * from the second attempt.
2407 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2410 __SetPageUptodate(new_page);
2412 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2413 vmf->address & PAGE_MASK,
2414 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2415 mmu_notifier_invalidate_range_start(&range);
2418 * Re-check the pte - we dropped the lock
2420 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2421 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2423 if (!PageAnon(old_page)) {
2424 dec_mm_counter_fast(mm,
2425 mm_counter_file(old_page));
2426 inc_mm_counter_fast(mm, MM_ANONPAGES);
2429 inc_mm_counter_fast(mm, MM_ANONPAGES);
2431 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2432 entry = mk_pte(new_page, vma->vm_page_prot);
2433 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2435 * Clear the pte entry and flush it first, before updating the
2436 * pte with the new entry. This will avoid a race condition
2437 * seen in the presence of one thread doing SMC and another
2440 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2441 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2442 mem_cgroup_commit_charge(new_page, memcg, false, false);
2443 lru_cache_add_active_or_unevictable(new_page, vma);
2445 * We call the notify macro here because, when using secondary
2446 * mmu page tables (such as kvm shadow page tables), we want the
2447 * new page to be mapped directly into the secondary page table.
2449 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2450 update_mmu_cache(vma, vmf->address, vmf->pte);
2453 * Only after switching the pte to the new page may
2454 * we remove the mapcount here. Otherwise another
2455 * process may come and find the rmap count decremented
2456 * before the pte is switched to the new page, and
2457 * "reuse" the old page writing into it while our pte
2458 * here still points into it and can be read by other
2461 * The critical issue is to order this
2462 * page_remove_rmap with the ptp_clear_flush above.
2463 * Those stores are ordered by (if nothing else,)
2464 * the barrier present in the atomic_add_negative
2465 * in page_remove_rmap.
2467 * Then the TLB flush in ptep_clear_flush ensures that
2468 * no process can access the old page before the
2469 * decremented mapcount is visible. And the old page
2470 * cannot be reused until after the decremented
2471 * mapcount is visible. So transitively, TLBs to
2472 * old page will be flushed before it can be reused.
2474 page_remove_rmap(old_page, false);
2477 /* Free the old page.. */
2478 new_page = old_page;
2481 mem_cgroup_cancel_charge(new_page, memcg, false);
2487 pte_unmap_unlock(vmf->pte, vmf->ptl);
2489 * No need to double call mmu_notifier->invalidate_range() callback as
2490 * the above ptep_clear_flush_notify() did already call it.
2492 mmu_notifier_invalidate_range_only_end(&range);
2495 * Don't let another task, with possibly unlocked vma,
2496 * keep the mlocked page.
2498 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2499 lock_page(old_page); /* LRU manipulation */
2500 if (PageMlocked(old_page))
2501 munlock_vma_page(old_page);
2502 unlock_page(old_page);
2506 return page_copied ? VM_FAULT_WRITE : 0;
2512 return VM_FAULT_OOM;
2516 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2517 * writeable once the page is prepared
2519 * @vmf: structure describing the fault
2521 * This function handles all that is needed to finish a write page fault in a
2522 * shared mapping due to PTE being read-only once the mapped page is prepared.
2523 * It handles locking of PTE and modifying it.
2525 * The function expects the page to be locked or other protection against
2526 * concurrent faults / writeback (such as DAX radix tree locks).
2528 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2529 * we acquired PTE lock.
2531 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2533 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2534 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2537 * We might have raced with another page fault while we released the
2538 * pte_offset_map_lock.
2540 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2541 pte_unmap_unlock(vmf->pte, vmf->ptl);
2542 return VM_FAULT_NOPAGE;
2549 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2552 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2554 struct vm_area_struct *vma = vmf->vma;
2556 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2559 pte_unmap_unlock(vmf->pte, vmf->ptl);
2560 vmf->flags |= FAULT_FLAG_MKWRITE;
2561 ret = vma->vm_ops->pfn_mkwrite(vmf);
2562 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2564 return finish_mkwrite_fault(vmf);
2567 return VM_FAULT_WRITE;
2570 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2571 __releases(vmf->ptl)
2573 struct vm_area_struct *vma = vmf->vma;
2575 get_page(vmf->page);
2577 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2580 pte_unmap_unlock(vmf->pte, vmf->ptl);
2581 tmp = do_page_mkwrite(vmf);
2582 if (unlikely(!tmp || (tmp &
2583 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2584 put_page(vmf->page);
2587 tmp = finish_mkwrite_fault(vmf);
2588 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2589 unlock_page(vmf->page);
2590 put_page(vmf->page);
2595 lock_page(vmf->page);
2597 fault_dirty_shared_page(vma, vmf->page);
2598 put_page(vmf->page);
2600 return VM_FAULT_WRITE;
2604 * This routine handles present pages, when users try to write
2605 * to a shared page. It is done by copying the page to a new address
2606 * and decrementing the shared-page counter for the old page.
2608 * Note that this routine assumes that the protection checks have been
2609 * done by the caller (the low-level page fault routine in most cases).
2610 * Thus we can safely just mark it writable once we've done any necessary
2613 * We also mark the page dirty at this point even though the page will
2614 * change only once the write actually happens. This avoids a few races,
2615 * and potentially makes it more efficient.
2617 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2618 * but allow concurrent faults), with pte both mapped and locked.
2619 * We return with mmap_sem still held, but pte unmapped and unlocked.
2621 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2622 __releases(vmf->ptl)
2624 struct vm_area_struct *vma = vmf->vma;
2626 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2629 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2632 * We should not cow pages in a shared writeable mapping.
2633 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2635 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2636 (VM_WRITE|VM_SHARED))
2637 return wp_pfn_shared(vmf);
2639 pte_unmap_unlock(vmf->pte, vmf->ptl);
2640 return wp_page_copy(vmf);
2644 * Take out anonymous pages first, anonymous shared vmas are
2645 * not dirty accountable.
2647 if (PageAnon(vmf->page)) {
2648 int total_map_swapcount;
2649 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2650 page_count(vmf->page) != 1))
2652 if (!trylock_page(vmf->page)) {
2653 get_page(vmf->page);
2654 pte_unmap_unlock(vmf->pte, vmf->ptl);
2655 lock_page(vmf->page);
2656 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2657 vmf->address, &vmf->ptl);
2658 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2659 unlock_page(vmf->page);
2660 pte_unmap_unlock(vmf->pte, vmf->ptl);
2661 put_page(vmf->page);
2664 put_page(vmf->page);
2666 if (PageKsm(vmf->page)) {
2667 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2669 unlock_page(vmf->page);
2673 return VM_FAULT_WRITE;
2675 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2676 if (total_map_swapcount == 1) {
2678 * The page is all ours. Move it to
2679 * our anon_vma so the rmap code will
2680 * not search our parent or siblings.
2681 * Protected against the rmap code by
2684 page_move_anon_rmap(vmf->page, vma);
2686 unlock_page(vmf->page);
2688 return VM_FAULT_WRITE;
2690 unlock_page(vmf->page);
2691 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2692 (VM_WRITE|VM_SHARED))) {
2693 return wp_page_shared(vmf);
2697 * Ok, we need to copy. Oh, well..
2699 get_page(vmf->page);
2701 pte_unmap_unlock(vmf->pte, vmf->ptl);
2702 return wp_page_copy(vmf);
2705 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2706 unsigned long start_addr, unsigned long end_addr,
2707 struct zap_details *details)
2709 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2712 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2713 struct zap_details *details)
2715 struct vm_area_struct *vma;
2716 pgoff_t vba, vea, zba, zea;
2718 vma_interval_tree_foreach(vma, root,
2719 details->first_index, details->last_index) {
2721 vba = vma->vm_pgoff;
2722 vea = vba + vma_pages(vma) - 1;
2723 zba = details->first_index;
2726 zea = details->last_index;
2730 unmap_mapping_range_vma(vma,
2731 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2732 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2738 * unmap_mapping_pages() - Unmap pages from processes.
2739 * @mapping: The address space containing pages to be unmapped.
2740 * @start: Index of first page to be unmapped.
2741 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2742 * @even_cows: Whether to unmap even private COWed pages.
2744 * Unmap the pages in this address space from any userspace process which
2745 * has them mmaped. Generally, you want to remove COWed pages as well when
2746 * a file is being truncated, but not when invalidating pages from the page
2749 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2750 pgoff_t nr, bool even_cows)
2752 struct zap_details details = { };
2754 details.check_mapping = even_cows ? NULL : mapping;
2755 details.first_index = start;
2756 details.last_index = start + nr - 1;
2757 if (details.last_index < details.first_index)
2758 details.last_index = ULONG_MAX;
2760 i_mmap_lock_write(mapping);
2761 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2762 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2763 i_mmap_unlock_write(mapping);
2767 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2768 * address_space corresponding to the specified byte range in the underlying
2771 * @mapping: the address space containing mmaps to be unmapped.
2772 * @holebegin: byte in first page to unmap, relative to the start of
2773 * the underlying file. This will be rounded down to a PAGE_SIZE
2774 * boundary. Note that this is different from truncate_pagecache(), which
2775 * must keep the partial page. In contrast, we must get rid of
2777 * @holelen: size of prospective hole in bytes. This will be rounded
2778 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2780 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2781 * but 0 when invalidating pagecache, don't throw away private data.
2783 void unmap_mapping_range(struct address_space *mapping,
2784 loff_t const holebegin, loff_t const holelen, int even_cows)
2786 pgoff_t hba = holebegin >> PAGE_SHIFT;
2787 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2789 /* Check for overflow. */
2790 if (sizeof(holelen) > sizeof(hlen)) {
2792 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2793 if (holeend & ~(long long)ULONG_MAX)
2794 hlen = ULONG_MAX - hba + 1;
2797 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2799 EXPORT_SYMBOL(unmap_mapping_range);
2802 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2803 * but allow concurrent faults), and pte mapped but not yet locked.
2804 * We return with pte unmapped and unlocked.
2806 * We return with the mmap_sem locked or unlocked in the same cases
2807 * as does filemap_fault().
2809 vm_fault_t do_swap_page(struct vm_fault *vmf)
2811 struct vm_area_struct *vma = vmf->vma;
2812 struct page *page = NULL, *swapcache;
2813 struct mem_cgroup *memcg;
2820 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2823 entry = pte_to_swp_entry(vmf->orig_pte);
2824 if (unlikely(non_swap_entry(entry))) {
2825 if (is_migration_entry(entry)) {
2826 migration_entry_wait(vma->vm_mm, vmf->pmd,
2828 } else if (is_device_private_entry(entry)) {
2829 vmf->page = device_private_entry_to_page(entry);
2830 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
2831 } else if (is_hwpoison_entry(entry)) {
2832 ret = VM_FAULT_HWPOISON;
2834 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2835 ret = VM_FAULT_SIGBUS;
2841 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2842 page = lookup_swap_cache(entry, vma, vmf->address);
2846 struct swap_info_struct *si = swp_swap_info(entry);
2848 if (si->flags & SWP_SYNCHRONOUS_IO &&
2849 __swap_count(entry) == 1) {
2850 /* skip swapcache */
2851 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2854 __SetPageLocked(page);
2855 __SetPageSwapBacked(page);
2856 set_page_private(page, entry.val);
2857 lru_cache_add_anon(page);
2858 swap_readpage(page, true);
2861 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2868 * Back out if somebody else faulted in this pte
2869 * while we released the pte lock.
2871 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2872 vmf->address, &vmf->ptl);
2873 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2875 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2879 /* Had to read the page from swap area: Major fault */
2880 ret = VM_FAULT_MAJOR;
2881 count_vm_event(PGMAJFAULT);
2882 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2883 } else if (PageHWPoison(page)) {
2885 * hwpoisoned dirty swapcache pages are kept for killing
2886 * owner processes (which may be unknown at hwpoison time)
2888 ret = VM_FAULT_HWPOISON;
2889 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2893 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2895 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2897 ret |= VM_FAULT_RETRY;
2902 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2903 * release the swapcache from under us. The page pin, and pte_same
2904 * test below, are not enough to exclude that. Even if it is still
2905 * swapcache, we need to check that the page's swap has not changed.
2907 if (unlikely((!PageSwapCache(page) ||
2908 page_private(page) != entry.val)) && swapcache)
2911 page = ksm_might_need_to_copy(page, vma, vmf->address);
2912 if (unlikely(!page)) {
2918 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2925 * Back out if somebody else already faulted in this pte.
2927 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2929 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2932 if (unlikely(!PageUptodate(page))) {
2933 ret = VM_FAULT_SIGBUS;
2938 * The page isn't present yet, go ahead with the fault.
2940 * Be careful about the sequence of operations here.
2941 * To get its accounting right, reuse_swap_page() must be called
2942 * while the page is counted on swap but not yet in mapcount i.e.
2943 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2944 * must be called after the swap_free(), or it will never succeed.
2947 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2948 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2949 pte = mk_pte(page, vma->vm_page_prot);
2950 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2951 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2952 vmf->flags &= ~FAULT_FLAG_WRITE;
2953 ret |= VM_FAULT_WRITE;
2954 exclusive = RMAP_EXCLUSIVE;
2956 flush_icache_page(vma, page);
2957 if (pte_swp_soft_dirty(vmf->orig_pte))
2958 pte = pte_mksoft_dirty(pte);
2959 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2960 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2961 vmf->orig_pte = pte;
2963 /* ksm created a completely new copy */
2964 if (unlikely(page != swapcache && swapcache)) {
2965 page_add_new_anon_rmap(page, vma, vmf->address, false);
2966 mem_cgroup_commit_charge(page, memcg, false, false);
2967 lru_cache_add_active_or_unevictable(page, vma);
2969 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2970 mem_cgroup_commit_charge(page, memcg, true, false);
2971 activate_page(page);
2975 if (mem_cgroup_swap_full(page) ||
2976 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2977 try_to_free_swap(page);
2979 if (page != swapcache && swapcache) {
2981 * Hold the lock to avoid the swap entry to be reused
2982 * until we take the PT lock for the pte_same() check
2983 * (to avoid false positives from pte_same). For
2984 * further safety release the lock after the swap_free
2985 * so that the swap count won't change under a
2986 * parallel locked swapcache.
2988 unlock_page(swapcache);
2989 put_page(swapcache);
2992 if (vmf->flags & FAULT_FLAG_WRITE) {
2993 ret |= do_wp_page(vmf);
2994 if (ret & VM_FAULT_ERROR)
2995 ret &= VM_FAULT_ERROR;
2999 /* No need to invalidate - it was non-present before */
3000 update_mmu_cache(vma, vmf->address, vmf->pte);
3002 pte_unmap_unlock(vmf->pte, vmf->ptl);
3006 mem_cgroup_cancel_charge(page, memcg, false);
3007 pte_unmap_unlock(vmf->pte, vmf->ptl);
3012 if (page != swapcache && swapcache) {
3013 unlock_page(swapcache);
3014 put_page(swapcache);
3020 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3021 * but allow concurrent faults), and pte mapped but not yet locked.
3022 * We return with mmap_sem still held, but pte unmapped and unlocked.
3024 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3026 struct vm_area_struct *vma = vmf->vma;
3027 struct mem_cgroup *memcg;
3032 /* File mapping without ->vm_ops ? */
3033 if (vma->vm_flags & VM_SHARED)
3034 return VM_FAULT_SIGBUS;
3037 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3038 * pte_offset_map() on pmds where a huge pmd might be created
3039 * from a different thread.
3041 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3042 * parallel threads are excluded by other means.
3044 * Here we only have down_read(mmap_sem).
3046 if (pte_alloc(vma->vm_mm, vmf->pmd))
3047 return VM_FAULT_OOM;
3049 /* See the comment in pte_alloc_one_map() */
3050 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3053 /* Use the zero-page for reads */
3054 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3055 !mm_forbids_zeropage(vma->vm_mm)) {
3056 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3057 vma->vm_page_prot));
3058 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3059 vmf->address, &vmf->ptl);
3060 if (!pte_none(*vmf->pte))
3062 ret = check_stable_address_space(vma->vm_mm);
3065 /* Deliver the page fault to userland, check inside PT lock */
3066 if (userfaultfd_missing(vma)) {
3067 pte_unmap_unlock(vmf->pte, vmf->ptl);
3068 return handle_userfault(vmf, VM_UFFD_MISSING);
3073 /* Allocate our own private page. */
3074 if (unlikely(anon_vma_prepare(vma)))
3076 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3080 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3085 * The memory barrier inside __SetPageUptodate makes sure that
3086 * preceeding stores to the page contents become visible before
3087 * the set_pte_at() write.
3089 __SetPageUptodate(page);
3091 entry = mk_pte(page, vma->vm_page_prot);
3092 if (vma->vm_flags & VM_WRITE)
3093 entry = pte_mkwrite(pte_mkdirty(entry));
3095 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3097 if (!pte_none(*vmf->pte))
3100 ret = check_stable_address_space(vma->vm_mm);
3104 /* Deliver the page fault to userland, check inside PT lock */
3105 if (userfaultfd_missing(vma)) {
3106 pte_unmap_unlock(vmf->pte, vmf->ptl);
3107 mem_cgroup_cancel_charge(page, memcg, false);
3109 return handle_userfault(vmf, VM_UFFD_MISSING);
3112 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3113 page_add_new_anon_rmap(page, vma, vmf->address, false);
3114 mem_cgroup_commit_charge(page, memcg, false, false);
3115 lru_cache_add_active_or_unevictable(page, vma);
3117 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3119 /* No need to invalidate - it was non-present before */
3120 update_mmu_cache(vma, vmf->address, vmf->pte);
3122 pte_unmap_unlock(vmf->pte, vmf->ptl);
3125 mem_cgroup_cancel_charge(page, memcg, false);
3131 return VM_FAULT_OOM;
3135 * The mmap_sem must have been held on entry, and may have been
3136 * released depending on flags and vma->vm_ops->fault() return value.
3137 * See filemap_fault() and __lock_page_retry().
3139 static vm_fault_t __do_fault(struct vm_fault *vmf)
3141 struct vm_area_struct *vma = vmf->vma;
3145 * Preallocate pte before we take page_lock because this might lead to
3146 * deadlocks for memcg reclaim which waits for pages under writeback:
3148 * SetPageWriteback(A)
3154 * wait_on_page_writeback(A)
3155 * SetPageWriteback(B)
3157 * # flush A, B to clear the writeback
3159 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3160 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3161 if (!vmf->prealloc_pte)
3162 return VM_FAULT_OOM;
3163 smp_wmb(); /* See comment in __pte_alloc() */
3166 ret = vma->vm_ops->fault(vmf);
3167 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3168 VM_FAULT_DONE_COW)))
3171 if (unlikely(PageHWPoison(vmf->page))) {
3172 if (ret & VM_FAULT_LOCKED)
3173 unlock_page(vmf->page);
3174 put_page(vmf->page);
3176 return VM_FAULT_HWPOISON;
3179 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3180 lock_page(vmf->page);
3182 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3188 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3189 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3190 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3191 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3193 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3195 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3198 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3200 struct vm_area_struct *vma = vmf->vma;
3202 if (!pmd_none(*vmf->pmd))
3204 if (vmf->prealloc_pte) {
3205 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3206 if (unlikely(!pmd_none(*vmf->pmd))) {
3207 spin_unlock(vmf->ptl);
3211 mm_inc_nr_ptes(vma->vm_mm);
3212 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3213 spin_unlock(vmf->ptl);
3214 vmf->prealloc_pte = NULL;
3215 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3216 return VM_FAULT_OOM;
3220 * If a huge pmd materialized under us just retry later. Use
3221 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3222 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3223 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3224 * running immediately after a huge pmd fault in a different thread of
3225 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3226 * All we have to ensure is that it is a regular pmd that we can walk
3227 * with pte_offset_map() and we can do that through an atomic read in
3228 * C, which is what pmd_trans_unstable() provides.
3230 if (pmd_devmap_trans_unstable(vmf->pmd))
3231 return VM_FAULT_NOPAGE;
3234 * At this point we know that our vmf->pmd points to a page of ptes
3235 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3236 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3237 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3238 * be valid and we will re-check to make sure the vmf->pte isn't
3239 * pte_none() under vmf->ptl protection when we return to
3242 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3247 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3248 static void deposit_prealloc_pte(struct vm_fault *vmf)
3250 struct vm_area_struct *vma = vmf->vma;
3252 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3254 * We are going to consume the prealloc table,
3255 * count that as nr_ptes.
3257 mm_inc_nr_ptes(vma->vm_mm);
3258 vmf->prealloc_pte = NULL;
3261 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3263 struct vm_area_struct *vma = vmf->vma;
3264 bool write = vmf->flags & FAULT_FLAG_WRITE;
3265 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3270 if (!transhuge_vma_suitable(vma, haddr))
3271 return VM_FAULT_FALLBACK;
3273 ret = VM_FAULT_FALLBACK;
3274 page = compound_head(page);
3277 * Archs like ppc64 need additonal space to store information
3278 * related to pte entry. Use the preallocated table for that.
3280 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3281 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3282 if (!vmf->prealloc_pte)
3283 return VM_FAULT_OOM;
3284 smp_wmb(); /* See comment in __pte_alloc() */
3287 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3288 if (unlikely(!pmd_none(*vmf->pmd)))
3291 for (i = 0; i < HPAGE_PMD_NR; i++)
3292 flush_icache_page(vma, page + i);
3294 entry = mk_huge_pmd(page, vma->vm_page_prot);
3296 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3298 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3299 page_add_file_rmap(page, true);
3301 * deposit and withdraw with pmd lock held
3303 if (arch_needs_pgtable_deposit())
3304 deposit_prealloc_pte(vmf);
3306 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3308 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3310 /* fault is handled */
3312 count_vm_event(THP_FILE_MAPPED);
3314 spin_unlock(vmf->ptl);
3318 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3326 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3327 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3329 * @vmf: fault environment
3330 * @memcg: memcg to charge page (only for private mappings)
3331 * @page: page to map
3333 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3336 * Target users are page handler itself and implementations of
3337 * vm_ops->map_pages.
3339 * Return: %0 on success, %VM_FAULT_ code in case of error.
3341 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3344 struct vm_area_struct *vma = vmf->vma;
3345 bool write = vmf->flags & FAULT_FLAG_WRITE;
3349 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3350 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3352 VM_BUG_ON_PAGE(memcg, page);
3354 ret = do_set_pmd(vmf, page);
3355 if (ret != VM_FAULT_FALLBACK)
3360 ret = pte_alloc_one_map(vmf);
3365 /* Re-check under ptl */
3366 if (unlikely(!pte_none(*vmf->pte)))
3367 return VM_FAULT_NOPAGE;
3369 flush_icache_page(vma, page);
3370 entry = mk_pte(page, vma->vm_page_prot);
3372 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3373 /* copy-on-write page */
3374 if (write && !(vma->vm_flags & VM_SHARED)) {
3375 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3376 page_add_new_anon_rmap(page, vma, vmf->address, false);
3377 mem_cgroup_commit_charge(page, memcg, false, false);
3378 lru_cache_add_active_or_unevictable(page, vma);
3380 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3381 page_add_file_rmap(page, false);
3383 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3385 /* no need to invalidate: a not-present page won't be cached */
3386 update_mmu_cache(vma, vmf->address, vmf->pte);
3393 * finish_fault - finish page fault once we have prepared the page to fault
3395 * @vmf: structure describing the fault
3397 * This function handles all that is needed to finish a page fault once the
3398 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3399 * given page, adds reverse page mapping, handles memcg charges and LRU
3402 * The function expects the page to be locked and on success it consumes a
3403 * reference of a page being mapped (for the PTE which maps it).
3405 * Return: %0 on success, %VM_FAULT_ code in case of error.
3407 vm_fault_t finish_fault(struct vm_fault *vmf)
3412 /* Did we COW the page? */
3413 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3414 !(vmf->vma->vm_flags & VM_SHARED))
3415 page = vmf->cow_page;
3420 * check even for read faults because we might have lost our CoWed
3423 if (!(vmf->vma->vm_flags & VM_SHARED))
3424 ret = check_stable_address_space(vmf->vma->vm_mm);
3426 ret = alloc_set_pte(vmf, vmf->memcg, page);
3428 pte_unmap_unlock(vmf->pte, vmf->ptl);
3432 static unsigned long fault_around_bytes __read_mostly =
3433 rounddown_pow_of_two(65536);
3435 #ifdef CONFIG_DEBUG_FS
3436 static int fault_around_bytes_get(void *data, u64 *val)
3438 *val = fault_around_bytes;
3443 * fault_around_bytes must be rounded down to the nearest page order as it's
3444 * what do_fault_around() expects to see.
3446 static int fault_around_bytes_set(void *data, u64 val)
3448 if (val / PAGE_SIZE > PTRS_PER_PTE)
3450 if (val > PAGE_SIZE)
3451 fault_around_bytes = rounddown_pow_of_two(val);
3453 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3456 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3457 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3459 static int __init fault_around_debugfs(void)
3461 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3462 &fault_around_bytes_fops);
3465 late_initcall(fault_around_debugfs);
3469 * do_fault_around() tries to map few pages around the fault address. The hope
3470 * is that the pages will be needed soon and this will lower the number of
3473 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3474 * not ready to be mapped: not up-to-date, locked, etc.
3476 * This function is called with the page table lock taken. In the split ptlock
3477 * case the page table lock only protects only those entries which belong to
3478 * the page table corresponding to the fault address.
3480 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3483 * fault_around_bytes defines how many bytes we'll try to map.
3484 * do_fault_around() expects it to be set to a power of two less than or equal
3487 * The virtual address of the area that we map is naturally aligned to
3488 * fault_around_bytes rounded down to the machine page size
3489 * (and therefore to page order). This way it's easier to guarantee
3490 * that we don't cross page table boundaries.
3492 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3494 unsigned long address = vmf->address, nr_pages, mask;
3495 pgoff_t start_pgoff = vmf->pgoff;
3500 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3501 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3503 vmf->address = max(address & mask, vmf->vma->vm_start);
3504 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3508 * end_pgoff is either the end of the page table, the end of
3509 * the vma or nr_pages from start_pgoff, depending what is nearest.
3511 end_pgoff = start_pgoff -
3512 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3514 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3515 start_pgoff + nr_pages - 1);
3517 if (pmd_none(*vmf->pmd)) {
3518 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3519 if (!vmf->prealloc_pte)
3521 smp_wmb(); /* See comment in __pte_alloc() */
3524 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3526 /* Huge page is mapped? Page fault is solved */
3527 if (pmd_trans_huge(*vmf->pmd)) {
3528 ret = VM_FAULT_NOPAGE;
3532 /* ->map_pages() haven't done anything useful. Cold page cache? */
3536 /* check if the page fault is solved */
3537 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3538 if (!pte_none(*vmf->pte))
3539 ret = VM_FAULT_NOPAGE;
3540 pte_unmap_unlock(vmf->pte, vmf->ptl);
3542 vmf->address = address;
3547 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3549 struct vm_area_struct *vma = vmf->vma;
3553 * Let's call ->map_pages() first and use ->fault() as fallback
3554 * if page by the offset is not ready to be mapped (cold cache or
3557 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3558 ret = do_fault_around(vmf);
3563 ret = __do_fault(vmf);
3564 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3567 ret |= finish_fault(vmf);
3568 unlock_page(vmf->page);
3569 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3570 put_page(vmf->page);
3574 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3576 struct vm_area_struct *vma = vmf->vma;
3579 if (unlikely(anon_vma_prepare(vma)))
3580 return VM_FAULT_OOM;
3582 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3584 return VM_FAULT_OOM;
3586 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3587 &vmf->memcg, false)) {
3588 put_page(vmf->cow_page);
3589 return VM_FAULT_OOM;
3592 ret = __do_fault(vmf);
3593 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3595 if (ret & VM_FAULT_DONE_COW)
3598 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3599 __SetPageUptodate(vmf->cow_page);
3601 ret |= finish_fault(vmf);
3602 unlock_page(vmf->page);
3603 put_page(vmf->page);
3604 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3608 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3609 put_page(vmf->cow_page);
3613 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3615 struct vm_area_struct *vma = vmf->vma;
3616 vm_fault_t ret, tmp;
3618 ret = __do_fault(vmf);
3619 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3623 * Check if the backing address space wants to know that the page is
3624 * about to become writable
3626 if (vma->vm_ops->page_mkwrite) {
3627 unlock_page(vmf->page);
3628 tmp = do_page_mkwrite(vmf);
3629 if (unlikely(!tmp ||
3630 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3631 put_page(vmf->page);
3636 ret |= finish_fault(vmf);
3637 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3639 unlock_page(vmf->page);
3640 put_page(vmf->page);
3644 fault_dirty_shared_page(vma, vmf->page);
3649 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3650 * but allow concurrent faults).
3651 * The mmap_sem may have been released depending on flags and our
3652 * return value. See filemap_fault() and __lock_page_or_retry().
3653 * If mmap_sem is released, vma may become invalid (for example
3654 * by other thread calling munmap()).
3656 static vm_fault_t do_fault(struct vm_fault *vmf)
3658 struct vm_area_struct *vma = vmf->vma;
3659 struct mm_struct *vm_mm = vma->vm_mm;
3663 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3665 if (!vma->vm_ops->fault) {
3667 * If we find a migration pmd entry or a none pmd entry, which
3668 * should never happen, return SIGBUS
3670 if (unlikely(!pmd_present(*vmf->pmd)))
3671 ret = VM_FAULT_SIGBUS;
3673 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3678 * Make sure this is not a temporary clearing of pte
3679 * by holding ptl and checking again. A R/M/W update
3680 * of pte involves: take ptl, clearing the pte so that
3681 * we don't have concurrent modification by hardware
3682 * followed by an update.
3684 if (unlikely(pte_none(*vmf->pte)))
3685 ret = VM_FAULT_SIGBUS;
3687 ret = VM_FAULT_NOPAGE;
3689 pte_unmap_unlock(vmf->pte, vmf->ptl);
3691 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3692 ret = do_read_fault(vmf);
3693 else if (!(vma->vm_flags & VM_SHARED))
3694 ret = do_cow_fault(vmf);
3696 ret = do_shared_fault(vmf);
3698 /* preallocated pagetable is unused: free it */
3699 if (vmf->prealloc_pte) {
3700 pte_free(vm_mm, vmf->prealloc_pte);
3701 vmf->prealloc_pte = NULL;
3706 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3707 unsigned long addr, int page_nid,
3712 count_vm_numa_event(NUMA_HINT_FAULTS);
3713 if (page_nid == numa_node_id()) {
3714 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3715 *flags |= TNF_FAULT_LOCAL;
3718 return mpol_misplaced(page, vma, addr);
3721 static vm_fault_t do_numa_page(struct vm_fault *vmf)
3723 struct vm_area_struct *vma = vmf->vma;
3724 struct page *page = NULL;
3725 int page_nid = NUMA_NO_NODE;
3728 bool migrated = false;
3730 bool was_writable = pte_savedwrite(vmf->orig_pte);
3734 * The "pte" at this point cannot be used safely without
3735 * validation through pte_unmap_same(). It's of NUMA type but
3736 * the pfn may be screwed if the read is non atomic.
3738 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3739 spin_lock(vmf->ptl);
3740 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3741 pte_unmap_unlock(vmf->pte, vmf->ptl);
3746 * Make it present again, Depending on how arch implementes non
3747 * accessible ptes, some can allow access by kernel mode.
3749 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
3750 pte = pte_modify(old_pte, vma->vm_page_prot);
3751 pte = pte_mkyoung(pte);
3753 pte = pte_mkwrite(pte);
3754 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
3755 update_mmu_cache(vma, vmf->address, vmf->pte);
3757 page = vm_normal_page(vma, vmf->address, pte);
3759 pte_unmap_unlock(vmf->pte, vmf->ptl);
3763 /* TODO: handle PTE-mapped THP */
3764 if (PageCompound(page)) {
3765 pte_unmap_unlock(vmf->pte, vmf->ptl);
3770 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3771 * much anyway since they can be in shared cache state. This misses
3772 * the case where a mapping is writable but the process never writes
3773 * to it but pte_write gets cleared during protection updates and
3774 * pte_dirty has unpredictable behaviour between PTE scan updates,
3775 * background writeback, dirty balancing and application behaviour.
3777 if (!pte_write(pte))
3778 flags |= TNF_NO_GROUP;
3781 * Flag if the page is shared between multiple address spaces. This
3782 * is later used when determining whether to group tasks together
3784 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3785 flags |= TNF_SHARED;
3787 last_cpupid = page_cpupid_last(page);
3788 page_nid = page_to_nid(page);
3789 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3791 pte_unmap_unlock(vmf->pte, vmf->ptl);
3792 if (target_nid == NUMA_NO_NODE) {
3797 /* Migrate to the requested node */
3798 migrated = migrate_misplaced_page(page, vma, target_nid);
3800 page_nid = target_nid;
3801 flags |= TNF_MIGRATED;
3803 flags |= TNF_MIGRATE_FAIL;
3806 if (page_nid != NUMA_NO_NODE)
3807 task_numa_fault(last_cpupid, page_nid, 1, flags);
3811 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3813 if (vma_is_anonymous(vmf->vma))
3814 return do_huge_pmd_anonymous_page(vmf);
3815 if (vmf->vma->vm_ops->huge_fault)
3816 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3817 return VM_FAULT_FALLBACK;
3820 /* `inline' is required to avoid gcc 4.1.2 build error */
3821 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3823 if (vma_is_anonymous(vmf->vma))
3824 return do_huge_pmd_wp_page(vmf, orig_pmd);
3825 if (vmf->vma->vm_ops->huge_fault)
3826 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3828 /* COW handled on pte level: split pmd */
3829 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3830 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3832 return VM_FAULT_FALLBACK;
3835 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3837 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3840 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3842 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3843 /* No support for anonymous transparent PUD pages yet */
3844 if (vma_is_anonymous(vmf->vma))
3845 return VM_FAULT_FALLBACK;
3846 if (vmf->vma->vm_ops->huge_fault)
3847 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3848 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3849 return VM_FAULT_FALLBACK;
3852 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3854 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3855 /* No support for anonymous transparent PUD pages yet */
3856 if (vma_is_anonymous(vmf->vma))
3857 return VM_FAULT_FALLBACK;
3858 if (vmf->vma->vm_ops->huge_fault)
3859 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3860 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3861 return VM_FAULT_FALLBACK;
3865 * These routines also need to handle stuff like marking pages dirty
3866 * and/or accessed for architectures that don't do it in hardware (most
3867 * RISC architectures). The early dirtying is also good on the i386.
3869 * There is also a hook called "update_mmu_cache()" that architectures
3870 * with external mmu caches can use to update those (ie the Sparc or
3871 * PowerPC hashed page tables that act as extended TLBs).
3873 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3874 * concurrent faults).
3876 * The mmap_sem may have been released depending on flags and our return value.
3877 * See filemap_fault() and __lock_page_or_retry().
3879 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
3883 if (unlikely(pmd_none(*vmf->pmd))) {
3885 * Leave __pte_alloc() until later: because vm_ops->fault may
3886 * want to allocate huge page, and if we expose page table
3887 * for an instant, it will be difficult to retract from
3888 * concurrent faults and from rmap lookups.
3892 /* See comment in pte_alloc_one_map() */
3893 if (pmd_devmap_trans_unstable(vmf->pmd))
3896 * A regular pmd is established and it can't morph into a huge
3897 * pmd from under us anymore at this point because we hold the
3898 * mmap_sem read mode and khugepaged takes it in write mode.
3899 * So now it's safe to run pte_offset_map().
3901 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3902 vmf->orig_pte = *vmf->pte;
3905 * some architectures can have larger ptes than wordsize,
3906 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3907 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3908 * accesses. The code below just needs a consistent view
3909 * for the ifs and we later double check anyway with the
3910 * ptl lock held. So here a barrier will do.
3913 if (pte_none(vmf->orig_pte)) {
3914 pte_unmap(vmf->pte);
3920 if (vma_is_anonymous(vmf->vma))
3921 return do_anonymous_page(vmf);
3923 return do_fault(vmf);
3926 if (!pte_present(vmf->orig_pte))
3927 return do_swap_page(vmf);
3929 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3930 return do_numa_page(vmf);
3932 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3933 spin_lock(vmf->ptl);
3934 entry = vmf->orig_pte;
3935 if (unlikely(!pte_same(*vmf->pte, entry)))
3937 if (vmf->flags & FAULT_FLAG_WRITE) {
3938 if (!pte_write(entry))
3939 return do_wp_page(vmf);
3940 entry = pte_mkdirty(entry);
3942 entry = pte_mkyoung(entry);
3943 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3944 vmf->flags & FAULT_FLAG_WRITE)) {
3945 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3948 * This is needed only for protection faults but the arch code
3949 * is not yet telling us if this is a protection fault or not.
3950 * This still avoids useless tlb flushes for .text page faults
3953 if (vmf->flags & FAULT_FLAG_WRITE)
3954 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3957 pte_unmap_unlock(vmf->pte, vmf->ptl);
3962 * By the time we get here, we already hold the mm semaphore
3964 * The mmap_sem may have been released depending on flags and our
3965 * return value. See filemap_fault() and __lock_page_or_retry().
3967 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
3968 unsigned long address, unsigned int flags)
3970 struct vm_fault vmf = {
3972 .address = address & PAGE_MASK,
3974 .pgoff = linear_page_index(vma, address),
3975 .gfp_mask = __get_fault_gfp_mask(vma),
3977 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3978 struct mm_struct *mm = vma->vm_mm;
3983 pgd = pgd_offset(mm, address);
3984 p4d = p4d_alloc(mm, pgd, address);
3986 return VM_FAULT_OOM;
3988 vmf.pud = pud_alloc(mm, p4d, address);
3990 return VM_FAULT_OOM;
3991 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
3992 ret = create_huge_pud(&vmf);
3993 if (!(ret & VM_FAULT_FALLBACK))
3996 pud_t orig_pud = *vmf.pud;
3999 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4001 /* NUMA case for anonymous PUDs would go here */
4003 if (dirty && !pud_write(orig_pud)) {
4004 ret = wp_huge_pud(&vmf, orig_pud);
4005 if (!(ret & VM_FAULT_FALLBACK))
4008 huge_pud_set_accessed(&vmf, orig_pud);
4014 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4016 return VM_FAULT_OOM;
4017 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4018 ret = create_huge_pmd(&vmf);
4019 if (!(ret & VM_FAULT_FALLBACK))
4022 pmd_t orig_pmd = *vmf.pmd;
4025 if (unlikely(is_swap_pmd(orig_pmd))) {
4026 VM_BUG_ON(thp_migration_supported() &&
4027 !is_pmd_migration_entry(orig_pmd));
4028 if (is_pmd_migration_entry(orig_pmd))
4029 pmd_migration_entry_wait(mm, vmf.pmd);
4032 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4033 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4034 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4036 if (dirty && !pmd_write(orig_pmd)) {
4037 ret = wp_huge_pmd(&vmf, orig_pmd);
4038 if (!(ret & VM_FAULT_FALLBACK))
4041 huge_pmd_set_accessed(&vmf, orig_pmd);
4047 return handle_pte_fault(&vmf);
4051 * By the time we get here, we already hold the mm semaphore
4053 * The mmap_sem may have been released depending on flags and our
4054 * return value. See filemap_fault() and __lock_page_or_retry().
4056 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4061 __set_current_state(TASK_RUNNING);
4063 count_vm_event(PGFAULT);
4064 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4066 /* do counter updates before entering really critical section. */
4067 check_sync_rss_stat(current);
4069 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4070 flags & FAULT_FLAG_INSTRUCTION,
4071 flags & FAULT_FLAG_REMOTE))
4072 return VM_FAULT_SIGSEGV;
4075 * Enable the memcg OOM handling for faults triggered in user
4076 * space. Kernel faults are handled more gracefully.
4078 if (flags & FAULT_FLAG_USER)
4079 mem_cgroup_enter_user_fault();
4081 if (unlikely(is_vm_hugetlb_page(vma)))
4082 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4084 ret = __handle_mm_fault(vma, address, flags);
4086 if (flags & FAULT_FLAG_USER) {
4087 mem_cgroup_exit_user_fault();
4089 * The task may have entered a memcg OOM situation but
4090 * if the allocation error was handled gracefully (no
4091 * VM_FAULT_OOM), there is no need to kill anything.
4092 * Just clean up the OOM state peacefully.
4094 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4095 mem_cgroup_oom_synchronize(false);
4100 EXPORT_SYMBOL_GPL(handle_mm_fault);
4102 #ifndef __PAGETABLE_P4D_FOLDED
4104 * Allocate p4d page table.
4105 * We've already handled the fast-path in-line.
4107 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4109 p4d_t *new = p4d_alloc_one(mm, address);
4113 smp_wmb(); /* See comment in __pte_alloc */
4115 spin_lock(&mm->page_table_lock);
4116 if (pgd_present(*pgd)) /* Another has populated it */
4119 pgd_populate(mm, pgd, new);
4120 spin_unlock(&mm->page_table_lock);
4123 #endif /* __PAGETABLE_P4D_FOLDED */
4125 #ifndef __PAGETABLE_PUD_FOLDED
4127 * Allocate page upper directory.
4128 * We've already handled the fast-path in-line.
4130 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4132 pud_t *new = pud_alloc_one(mm, address);
4136 smp_wmb(); /* See comment in __pte_alloc */
4138 spin_lock(&mm->page_table_lock);
4139 #ifndef __ARCH_HAS_5LEVEL_HACK
4140 if (!p4d_present(*p4d)) {
4142 p4d_populate(mm, p4d, new);
4143 } else /* Another has populated it */
4146 if (!pgd_present(*p4d)) {
4148 pgd_populate(mm, p4d, new);
4149 } else /* Another has populated it */
4151 #endif /* __ARCH_HAS_5LEVEL_HACK */
4152 spin_unlock(&mm->page_table_lock);
4155 #endif /* __PAGETABLE_PUD_FOLDED */
4157 #ifndef __PAGETABLE_PMD_FOLDED
4159 * Allocate page middle directory.
4160 * We've already handled the fast-path in-line.
4162 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4165 pmd_t *new = pmd_alloc_one(mm, address);
4169 smp_wmb(); /* See comment in __pte_alloc */
4171 ptl = pud_lock(mm, pud);
4172 #ifndef __ARCH_HAS_4LEVEL_HACK
4173 if (!pud_present(*pud)) {
4175 pud_populate(mm, pud, new);
4176 } else /* Another has populated it */
4179 if (!pgd_present(*pud)) {
4181 pgd_populate(mm, pud, new);
4182 } else /* Another has populated it */
4184 #endif /* __ARCH_HAS_4LEVEL_HACK */
4188 #endif /* __PAGETABLE_PMD_FOLDED */
4190 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4191 struct mmu_notifier_range *range,
4192 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4200 pgd = pgd_offset(mm, address);
4201 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4204 p4d = p4d_offset(pgd, address);
4205 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4208 pud = pud_offset(p4d, address);
4209 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4212 pmd = pmd_offset(pud, address);
4213 VM_BUG_ON(pmd_trans_huge(*pmd));
4215 if (pmd_huge(*pmd)) {
4220 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4221 NULL, mm, address & PMD_MASK,
4222 (address & PMD_MASK) + PMD_SIZE);
4223 mmu_notifier_invalidate_range_start(range);
4225 *ptlp = pmd_lock(mm, pmd);
4226 if (pmd_huge(*pmd)) {
4232 mmu_notifier_invalidate_range_end(range);
4235 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4239 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4240 address & PAGE_MASK,
4241 (address & PAGE_MASK) + PAGE_SIZE);
4242 mmu_notifier_invalidate_range_start(range);
4244 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4245 if (!pte_present(*ptep))
4250 pte_unmap_unlock(ptep, *ptlp);
4252 mmu_notifier_invalidate_range_end(range);
4257 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4258 pte_t **ptepp, spinlock_t **ptlp)
4262 /* (void) is needed to make gcc happy */
4263 (void) __cond_lock(*ptlp,
4264 !(res = __follow_pte_pmd(mm, address, NULL,
4265 ptepp, NULL, ptlp)));
4269 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4270 struct mmu_notifier_range *range,
4271 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4275 /* (void) is needed to make gcc happy */
4276 (void) __cond_lock(*ptlp,
4277 !(res = __follow_pte_pmd(mm, address, range,
4278 ptepp, pmdpp, ptlp)));
4281 EXPORT_SYMBOL(follow_pte_pmd);
4284 * follow_pfn - look up PFN at a user virtual address
4285 * @vma: memory mapping
4286 * @address: user virtual address
4287 * @pfn: location to store found PFN
4289 * Only IO mappings and raw PFN mappings are allowed.
4291 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4293 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4300 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4303 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4306 *pfn = pte_pfn(*ptep);
4307 pte_unmap_unlock(ptep, ptl);
4310 EXPORT_SYMBOL(follow_pfn);
4312 #ifdef CONFIG_HAVE_IOREMAP_PROT
4313 int follow_phys(struct vm_area_struct *vma,
4314 unsigned long address, unsigned int flags,
4315 unsigned long *prot, resource_size_t *phys)
4321 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4324 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4328 if ((flags & FOLL_WRITE) && !pte_write(pte))
4331 *prot = pgprot_val(pte_pgprot(pte));
4332 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4336 pte_unmap_unlock(ptep, ptl);
4341 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4342 void *buf, int len, int write)
4344 resource_size_t phys_addr;
4345 unsigned long prot = 0;
4346 void __iomem *maddr;
4347 int offset = addr & (PAGE_SIZE-1);
4349 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4352 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4357 memcpy_toio(maddr + offset, buf, len);
4359 memcpy_fromio(buf, maddr + offset, len);
4364 EXPORT_SYMBOL_GPL(generic_access_phys);
4368 * Access another process' address space as given in mm. If non-NULL, use the
4369 * given task for page fault accounting.
4371 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4372 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4374 struct vm_area_struct *vma;
4375 void *old_buf = buf;
4376 int write = gup_flags & FOLL_WRITE;
4378 if (down_read_killable(&mm->mmap_sem))
4381 /* ignore errors, just check how much was successfully transferred */
4383 int bytes, ret, offset;
4385 struct page *page = NULL;
4387 ret = get_user_pages_remote(tsk, mm, addr, 1,
4388 gup_flags, &page, &vma, NULL);
4390 #ifndef CONFIG_HAVE_IOREMAP_PROT
4394 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4395 * we can access using slightly different code.
4397 vma = find_vma(mm, addr);
4398 if (!vma || vma->vm_start > addr)
4400 if (vma->vm_ops && vma->vm_ops->access)
4401 ret = vma->vm_ops->access(vma, addr, buf,
4409 offset = addr & (PAGE_SIZE-1);
4410 if (bytes > PAGE_SIZE-offset)
4411 bytes = PAGE_SIZE-offset;
4415 copy_to_user_page(vma, page, addr,
4416 maddr + offset, buf, bytes);
4417 set_page_dirty_lock(page);
4419 copy_from_user_page(vma, page, addr,
4420 buf, maddr + offset, bytes);
4429 up_read(&mm->mmap_sem);
4431 return buf - old_buf;
4435 * access_remote_vm - access another process' address space
4436 * @mm: the mm_struct of the target address space
4437 * @addr: start address to access
4438 * @buf: source or destination buffer
4439 * @len: number of bytes to transfer
4440 * @gup_flags: flags modifying lookup behaviour
4442 * The caller must hold a reference on @mm.
4444 * Return: number of bytes copied from source to destination.
4446 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4447 void *buf, int len, unsigned int gup_flags)
4449 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4453 * Access another process' address space.
4454 * Source/target buffer must be kernel space,
4455 * Do not walk the page table directly, use get_user_pages
4457 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4458 void *buf, int len, unsigned int gup_flags)
4460 struct mm_struct *mm;
4463 mm = get_task_mm(tsk);
4467 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4473 EXPORT_SYMBOL_GPL(access_process_vm);
4476 * Print the name of a VMA.
4478 void print_vma_addr(char *prefix, unsigned long ip)
4480 struct mm_struct *mm = current->mm;
4481 struct vm_area_struct *vma;
4484 * we might be running from an atomic context so we cannot sleep
4486 if (!down_read_trylock(&mm->mmap_sem))
4489 vma = find_vma(mm, ip);
4490 if (vma && vma->vm_file) {
4491 struct file *f = vma->vm_file;
4492 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4496 p = file_path(f, buf, PAGE_SIZE);
4499 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4501 vma->vm_end - vma->vm_start);
4502 free_page((unsigned long)buf);
4505 up_read(&mm->mmap_sem);
4508 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4509 void __might_fault(const char *file, int line)
4512 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4513 * holding the mmap_sem, this is safe because kernel memory doesn't
4514 * get paged out, therefore we'll never actually fault, and the
4515 * below annotations will generate false positives.
4517 if (uaccess_kernel())
4519 if (pagefault_disabled())
4521 __might_sleep(file, line, 0);
4522 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4524 might_lock_read(¤t->mm->mmap_sem);
4527 EXPORT_SYMBOL(__might_fault);
4530 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4532 * Process all subpages of the specified huge page with the specified
4533 * operation. The target subpage will be processed last to keep its
4536 static inline void process_huge_page(
4537 unsigned long addr_hint, unsigned int pages_per_huge_page,
4538 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4542 unsigned long addr = addr_hint &
4543 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4545 /* Process target subpage last to keep its cache lines hot */
4547 n = (addr_hint - addr) / PAGE_SIZE;
4548 if (2 * n <= pages_per_huge_page) {
4549 /* If target subpage in first half of huge page */
4552 /* Process subpages at the end of huge page */
4553 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4555 process_subpage(addr + i * PAGE_SIZE, i, arg);
4558 /* If target subpage in second half of huge page */
4559 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4560 l = pages_per_huge_page - n;
4561 /* Process subpages at the begin of huge page */
4562 for (i = 0; i < base; i++) {
4564 process_subpage(addr + i * PAGE_SIZE, i, arg);
4568 * Process remaining subpages in left-right-left-right pattern
4569 * towards the target subpage
4571 for (i = 0; i < l; i++) {
4572 int left_idx = base + i;
4573 int right_idx = base + 2 * l - 1 - i;
4576 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4578 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4582 static void clear_gigantic_page(struct page *page,
4584 unsigned int pages_per_huge_page)
4587 struct page *p = page;
4590 for (i = 0; i < pages_per_huge_page;
4591 i++, p = mem_map_next(p, page, i)) {
4593 clear_user_highpage(p, addr + i * PAGE_SIZE);
4597 static void clear_subpage(unsigned long addr, int idx, void *arg)
4599 struct page *page = arg;
4601 clear_user_highpage(page + idx, addr);
4604 void clear_huge_page(struct page *page,
4605 unsigned long addr_hint, unsigned int pages_per_huge_page)
4607 unsigned long addr = addr_hint &
4608 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4610 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4611 clear_gigantic_page(page, addr, pages_per_huge_page);
4615 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4618 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4620 struct vm_area_struct *vma,
4621 unsigned int pages_per_huge_page)
4624 struct page *dst_base = dst;
4625 struct page *src_base = src;
4627 for (i = 0; i < pages_per_huge_page; ) {
4629 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4632 dst = mem_map_next(dst, dst_base, i);
4633 src = mem_map_next(src, src_base, i);
4637 struct copy_subpage_arg {
4640 struct vm_area_struct *vma;
4643 static void copy_subpage(unsigned long addr, int idx, void *arg)
4645 struct copy_subpage_arg *copy_arg = arg;
4647 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4648 addr, copy_arg->vma);
4651 void copy_user_huge_page(struct page *dst, struct page *src,
4652 unsigned long addr_hint, struct vm_area_struct *vma,
4653 unsigned int pages_per_huge_page)
4655 unsigned long addr = addr_hint &
4656 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4657 struct copy_subpage_arg arg = {
4663 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4664 copy_user_gigantic_page(dst, src, addr, vma,
4665 pages_per_huge_page);
4669 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4672 long copy_huge_page_from_user(struct page *dst_page,
4673 const void __user *usr_src,
4674 unsigned int pages_per_huge_page,
4675 bool allow_pagefault)
4677 void *src = (void *)usr_src;
4679 unsigned long i, rc = 0;
4680 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4682 for (i = 0; i < pages_per_huge_page; i++) {
4683 if (allow_pagefault)
4684 page_kaddr = kmap(dst_page + i);
4686 page_kaddr = kmap_atomic(dst_page + i);
4687 rc = copy_from_user(page_kaddr,
4688 (const void __user *)(src + i * PAGE_SIZE),
4690 if (allow_pagefault)
4691 kunmap(dst_page + i);
4693 kunmap_atomic(page_kaddr);
4695 ret_val -= (PAGE_SIZE - rc);
4703 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4705 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4707 static struct kmem_cache *page_ptl_cachep;
4709 void __init ptlock_cache_init(void)
4711 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4715 bool ptlock_alloc(struct page *page)
4719 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4726 void ptlock_free(struct page *page)
4728 kmem_cache_free(page_ptl_cachep, page->ptl);