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 static int __init disable_randmaps(char *s)
123 randomize_va_space = 0;
126 __setup("norandmaps", disable_randmaps);
128 unsigned long zero_pfn __read_mostly;
129 EXPORT_SYMBOL(zero_pfn);
131 unsigned long highest_memmap_pfn __read_mostly;
134 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
136 static int __init init_zero_pfn(void)
138 zero_pfn = page_to_pfn(ZERO_PAGE(0));
141 core_initcall(init_zero_pfn);
144 #if defined(SPLIT_RSS_COUNTING)
146 void sync_mm_rss(struct mm_struct *mm)
150 for (i = 0; i < NR_MM_COUNTERS; i++) {
151 if (current->rss_stat.count[i]) {
152 add_mm_counter(mm, i, current->rss_stat.count[i]);
153 current->rss_stat.count[i] = 0;
156 current->rss_stat.events = 0;
159 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
161 struct task_struct *task = current;
163 if (likely(task->mm == mm))
164 task->rss_stat.count[member] += val;
166 add_mm_counter(mm, member, val);
168 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
169 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
171 /* sync counter once per 64 page faults */
172 #define TASK_RSS_EVENTS_THRESH (64)
173 static void check_sync_rss_stat(struct task_struct *task)
175 if (unlikely(task != current))
177 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
178 sync_mm_rss(task->mm);
180 #else /* SPLIT_RSS_COUNTING */
182 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
183 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
185 static void check_sync_rss_stat(struct task_struct *task)
189 #endif /* SPLIT_RSS_COUNTING */
192 * Note: this doesn't free the actual pages themselves. That
193 * has been handled earlier when unmapping all the memory regions.
195 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
198 pgtable_t token = pmd_pgtable(*pmd);
200 pte_free_tlb(tlb, token, addr);
201 mm_dec_nr_ptes(tlb->mm);
204 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
205 unsigned long addr, unsigned long end,
206 unsigned long floor, unsigned long ceiling)
213 pmd = pmd_offset(pud, addr);
215 next = pmd_addr_end(addr, end);
216 if (pmd_none_or_clear_bad(pmd))
218 free_pte_range(tlb, pmd, addr);
219 } while (pmd++, addr = next, addr != end);
229 if (end - 1 > ceiling - 1)
232 pmd = pmd_offset(pud, start);
234 pmd_free_tlb(tlb, pmd, start);
235 mm_dec_nr_pmds(tlb->mm);
238 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
239 unsigned long addr, unsigned long end,
240 unsigned long floor, unsigned long ceiling)
247 pud = pud_offset(p4d, addr);
249 next = pud_addr_end(addr, end);
250 if (pud_none_or_clear_bad(pud))
252 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
253 } while (pud++, addr = next, addr != end);
263 if (end - 1 > ceiling - 1)
266 pud = pud_offset(p4d, start);
268 pud_free_tlb(tlb, pud, start);
269 mm_dec_nr_puds(tlb->mm);
272 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
273 unsigned long addr, unsigned long end,
274 unsigned long floor, unsigned long ceiling)
281 p4d = p4d_offset(pgd, addr);
283 next = p4d_addr_end(addr, end);
284 if (p4d_none_or_clear_bad(p4d))
286 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
287 } while (p4d++, addr = next, addr != end);
293 ceiling &= PGDIR_MASK;
297 if (end - 1 > ceiling - 1)
300 p4d = p4d_offset(pgd, start);
302 p4d_free_tlb(tlb, p4d, start);
306 * This function frees user-level page tables of a process.
308 void free_pgd_range(struct mmu_gather *tlb,
309 unsigned long addr, unsigned long end,
310 unsigned long floor, unsigned long ceiling)
316 * The next few lines have given us lots of grief...
318 * Why are we testing PMD* at this top level? Because often
319 * there will be no work to do at all, and we'd prefer not to
320 * go all the way down to the bottom just to discover that.
322 * Why all these "- 1"s? Because 0 represents both the bottom
323 * of the address space and the top of it (using -1 for the
324 * top wouldn't help much: the masks would do the wrong thing).
325 * The rule is that addr 0 and floor 0 refer to the bottom of
326 * the address space, but end 0 and ceiling 0 refer to the top
327 * Comparisons need to use "end - 1" and "ceiling - 1" (though
328 * that end 0 case should be mythical).
330 * Wherever addr is brought up or ceiling brought down, we must
331 * be careful to reject "the opposite 0" before it confuses the
332 * subsequent tests. But what about where end is brought down
333 * by PMD_SIZE below? no, end can't go down to 0 there.
335 * Whereas we round start (addr) and ceiling down, by different
336 * masks at different levels, in order to test whether a table
337 * now has no other vmas using it, so can be freed, we don't
338 * bother to round floor or end up - the tests don't need that.
352 if (end - 1 > ceiling - 1)
357 * We add page table cache pages with PAGE_SIZE,
358 * (see pte_free_tlb()), flush the tlb if we need
360 tlb_change_page_size(tlb, PAGE_SIZE);
361 pgd = pgd_offset(tlb->mm, addr);
363 next = pgd_addr_end(addr, end);
364 if (pgd_none_or_clear_bad(pgd))
366 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
367 } while (pgd++, addr = next, addr != end);
370 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
371 unsigned long floor, unsigned long ceiling)
374 struct vm_area_struct *next = vma->vm_next;
375 unsigned long addr = vma->vm_start;
378 * Hide vma from rmap and truncate_pagecache before freeing
381 unlink_anon_vmas(vma);
382 unlink_file_vma(vma);
384 if (is_vm_hugetlb_page(vma)) {
385 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
386 floor, next ? next->vm_start : ceiling);
389 * Optimization: gather nearby vmas into one call down
391 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
392 && !is_vm_hugetlb_page(next)) {
395 unlink_anon_vmas(vma);
396 unlink_file_vma(vma);
398 free_pgd_range(tlb, addr, vma->vm_end,
399 floor, next ? next->vm_start : ceiling);
405 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
408 pgtable_t new = pte_alloc_one(mm);
413 * Ensure all pte setup (eg. pte page lock and page clearing) are
414 * visible before the pte is made visible to other CPUs by being
415 * put into page tables.
417 * The other side of the story is the pointer chasing in the page
418 * table walking code (when walking the page table without locking;
419 * ie. most of the time). Fortunately, these data accesses consist
420 * of a chain of data-dependent loads, meaning most CPUs (alpha
421 * being the notable exception) will already guarantee loads are
422 * seen in-order. See the alpha page table accessors for the
423 * smp_read_barrier_depends() barriers in page table walking code.
425 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
427 ptl = pmd_lock(mm, pmd);
428 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
430 pmd_populate(mm, pmd, new);
439 int __pte_alloc_kernel(pmd_t *pmd)
441 pte_t *new = pte_alloc_one_kernel(&init_mm);
445 smp_wmb(); /* See comment in __pte_alloc */
447 spin_lock(&init_mm.page_table_lock);
448 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
449 pmd_populate_kernel(&init_mm, pmd, new);
452 spin_unlock(&init_mm.page_table_lock);
454 pte_free_kernel(&init_mm, new);
458 static inline void init_rss_vec(int *rss)
460 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
463 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
467 if (current->mm == mm)
469 for (i = 0; i < NR_MM_COUNTERS; i++)
471 add_mm_counter(mm, i, rss[i]);
475 * This function is called to print an error when a bad pte
476 * is found. For example, we might have a PFN-mapped pte in
477 * a region that doesn't allow it.
479 * The calling function must still handle the error.
481 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
482 pte_t pte, struct page *page)
484 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
485 p4d_t *p4d = p4d_offset(pgd, addr);
486 pud_t *pud = pud_offset(p4d, addr);
487 pmd_t *pmd = pmd_offset(pud, addr);
488 struct address_space *mapping;
490 static unsigned long resume;
491 static unsigned long nr_shown;
492 static unsigned long nr_unshown;
495 * Allow a burst of 60 reports, then keep quiet for that minute;
496 * or allow a steady drip of one report per second.
498 if (nr_shown == 60) {
499 if (time_before(jiffies, resume)) {
504 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
511 resume = jiffies + 60 * HZ;
513 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
514 index = linear_page_index(vma, addr);
516 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
518 (long long)pte_val(pte), (long long)pmd_val(*pmd));
520 dump_page(page, "bad pte");
521 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
522 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
523 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
525 vma->vm_ops ? vma->vm_ops->fault : NULL,
526 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
527 mapping ? mapping->a_ops->readpage : NULL);
529 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
533 * vm_normal_page -- This function gets the "struct page" associated with a pte.
535 * "Special" mappings do not wish to be associated with a "struct page" (either
536 * it doesn't exist, or it exists but they don't want to touch it). In this
537 * case, NULL is returned here. "Normal" mappings do have a struct page.
539 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
540 * pte bit, in which case this function is trivial. Secondly, an architecture
541 * may not have a spare pte bit, which requires a more complicated scheme,
544 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
545 * special mapping (even if there are underlying and valid "struct pages").
546 * COWed pages of a VM_PFNMAP are always normal.
548 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
549 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
550 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
551 * mapping will always honor the rule
553 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
555 * And for normal mappings this is false.
557 * This restricts such mappings to be a linear translation from virtual address
558 * to pfn. To get around this restriction, we allow arbitrary mappings so long
559 * as the vma is not a COW mapping; in that case, we know that all ptes are
560 * special (because none can have been COWed).
563 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
565 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
566 * page" backing, however the difference is that _all_ pages with a struct
567 * page (that is, those where pfn_valid is true) are refcounted and considered
568 * normal pages by the VM. The disadvantage is that pages are refcounted
569 * (which can be slower and simply not an option for some PFNMAP users). The
570 * advantage is that we don't have to follow the strict linearity rule of
571 * PFNMAP mappings in order to support COWable mappings.
574 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
577 unsigned long pfn = pte_pfn(pte);
579 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
580 if (likely(!pte_special(pte)))
582 if (vma->vm_ops && vma->vm_ops->find_special_page)
583 return vma->vm_ops->find_special_page(vma, addr);
584 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
586 if (is_zero_pfn(pfn))
591 print_bad_pte(vma, addr, pte, NULL);
595 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
597 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
598 if (vma->vm_flags & VM_MIXEDMAP) {
604 off = (addr - vma->vm_start) >> PAGE_SHIFT;
605 if (pfn == vma->vm_pgoff + off)
607 if (!is_cow_mapping(vma->vm_flags))
612 if (is_zero_pfn(pfn))
616 if (unlikely(pfn > highest_memmap_pfn)) {
617 print_bad_pte(vma, addr, pte, NULL);
622 * NOTE! We still have PageReserved() pages in the page tables.
623 * eg. VDSO mappings can cause them to exist.
626 return pfn_to_page(pfn);
629 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
630 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
633 unsigned long pfn = pmd_pfn(pmd);
636 * There is no pmd_special() but there may be special pmds, e.g.
637 * in a direct-access (dax) mapping, so let's just replicate the
638 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
640 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
641 if (vma->vm_flags & VM_MIXEDMAP) {
647 off = (addr - vma->vm_start) >> PAGE_SHIFT;
648 if (pfn == vma->vm_pgoff + off)
650 if (!is_cow_mapping(vma->vm_flags))
657 if (is_zero_pfn(pfn))
659 if (unlikely(pfn > highest_memmap_pfn))
663 * NOTE! We still have PageReserved() pages in the page tables.
664 * eg. VDSO mappings can cause them to exist.
667 return pfn_to_page(pfn);
672 * copy one vm_area from one task to the other. Assumes the page tables
673 * already present in the new task to be cleared in the whole range
674 * covered by this vma.
677 static inline unsigned long
678 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
679 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
680 unsigned long addr, int *rss)
682 unsigned long vm_flags = vma->vm_flags;
683 pte_t pte = *src_pte;
686 /* pte contains position in swap or file, so copy. */
687 if (unlikely(!pte_present(pte))) {
688 swp_entry_t entry = pte_to_swp_entry(pte);
690 if (likely(!non_swap_entry(entry))) {
691 if (swap_duplicate(entry) < 0)
694 /* make sure dst_mm is on swapoff's mmlist. */
695 if (unlikely(list_empty(&dst_mm->mmlist))) {
696 spin_lock(&mmlist_lock);
697 if (list_empty(&dst_mm->mmlist))
698 list_add(&dst_mm->mmlist,
700 spin_unlock(&mmlist_lock);
703 } else if (is_migration_entry(entry)) {
704 page = migration_entry_to_page(entry);
706 rss[mm_counter(page)]++;
708 if (is_write_migration_entry(entry) &&
709 is_cow_mapping(vm_flags)) {
711 * COW mappings require pages in both
712 * parent and child to be set to read.
714 make_migration_entry_read(&entry);
715 pte = swp_entry_to_pte(entry);
716 if (pte_swp_soft_dirty(*src_pte))
717 pte = pte_swp_mksoft_dirty(pte);
718 set_pte_at(src_mm, addr, src_pte, pte);
720 } else if (is_device_private_entry(entry)) {
721 page = device_private_entry_to_page(entry);
724 * Update rss count even for unaddressable pages, as
725 * they should treated just like normal pages in this
728 * We will likely want to have some new rss counters
729 * for unaddressable pages, at some point. But for now
730 * keep things as they are.
733 rss[mm_counter(page)]++;
734 page_dup_rmap(page, false);
737 * We do not preserve soft-dirty information, because so
738 * far, checkpoint/restore is the only feature that
739 * requires that. And checkpoint/restore does not work
740 * when a device driver is involved (you cannot easily
741 * save and restore device driver state).
743 if (is_write_device_private_entry(entry) &&
744 is_cow_mapping(vm_flags)) {
745 make_device_private_entry_read(&entry);
746 pte = swp_entry_to_pte(entry);
747 set_pte_at(src_mm, addr, src_pte, pte);
754 * If it's a COW mapping, write protect it both
755 * in the parent and the child
757 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
758 ptep_set_wrprotect(src_mm, addr, src_pte);
759 pte = pte_wrprotect(pte);
763 * If it's a shared mapping, mark it clean in
766 if (vm_flags & VM_SHARED)
767 pte = pte_mkclean(pte);
768 pte = pte_mkold(pte);
770 page = vm_normal_page(vma, addr, pte);
773 page_dup_rmap(page, false);
774 rss[mm_counter(page)]++;
775 } else if (pte_devmap(pte)) {
776 page = pte_page(pte);
780 set_pte_at(dst_mm, addr, dst_pte, pte);
784 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
785 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
786 unsigned long addr, unsigned long end)
788 pte_t *orig_src_pte, *orig_dst_pte;
789 pte_t *src_pte, *dst_pte;
790 spinlock_t *src_ptl, *dst_ptl;
792 int rss[NR_MM_COUNTERS];
793 swp_entry_t entry = (swp_entry_t){0};
798 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
801 src_pte = pte_offset_map(src_pmd, addr);
802 src_ptl = pte_lockptr(src_mm, src_pmd);
803 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
804 orig_src_pte = src_pte;
805 orig_dst_pte = dst_pte;
806 arch_enter_lazy_mmu_mode();
810 * We are holding two locks at this point - either of them
811 * could generate latencies in another task on another CPU.
813 if (progress >= 32) {
815 if (need_resched() ||
816 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
819 if (pte_none(*src_pte)) {
823 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
828 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
830 arch_leave_lazy_mmu_mode();
831 spin_unlock(src_ptl);
832 pte_unmap(orig_src_pte);
833 add_mm_rss_vec(dst_mm, rss);
834 pte_unmap_unlock(orig_dst_pte, dst_ptl);
838 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
847 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
848 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
849 unsigned long addr, unsigned long end)
851 pmd_t *src_pmd, *dst_pmd;
854 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
857 src_pmd = pmd_offset(src_pud, addr);
859 next = pmd_addr_end(addr, end);
860 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
861 || pmd_devmap(*src_pmd)) {
863 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
864 err = copy_huge_pmd(dst_mm, src_mm,
865 dst_pmd, src_pmd, addr, vma);
872 if (pmd_none_or_clear_bad(src_pmd))
874 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
877 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
881 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
882 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
883 unsigned long addr, unsigned long end)
885 pud_t *src_pud, *dst_pud;
888 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
891 src_pud = pud_offset(src_p4d, addr);
893 next = pud_addr_end(addr, end);
894 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
897 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
898 err = copy_huge_pud(dst_mm, src_mm,
899 dst_pud, src_pud, addr, vma);
906 if (pud_none_or_clear_bad(src_pud))
908 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
911 } while (dst_pud++, src_pud++, addr = next, addr != end);
915 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
916 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
917 unsigned long addr, unsigned long end)
919 p4d_t *src_p4d, *dst_p4d;
922 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
925 src_p4d = p4d_offset(src_pgd, addr);
927 next = p4d_addr_end(addr, end);
928 if (p4d_none_or_clear_bad(src_p4d))
930 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
933 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
937 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
938 struct vm_area_struct *vma)
940 pgd_t *src_pgd, *dst_pgd;
942 unsigned long addr = vma->vm_start;
943 unsigned long end = vma->vm_end;
944 struct mmu_notifier_range range;
949 * Don't copy ptes where a page fault will fill them correctly.
950 * Fork becomes much lighter when there are big shared or private
951 * readonly mappings. The tradeoff is that copy_page_range is more
952 * efficient than faulting.
954 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
958 if (is_vm_hugetlb_page(vma))
959 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
961 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
963 * We do not free on error cases below as remove_vma
964 * gets called on error from higher level routine
966 ret = track_pfn_copy(vma);
972 * We need to invalidate the secondary MMU mappings only when
973 * there could be a permission downgrade on the ptes of the
974 * parent mm. And a permission downgrade will only happen if
975 * is_cow_mapping() returns true.
977 is_cow = is_cow_mapping(vma->vm_flags);
980 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
981 0, vma, src_mm, addr, end);
982 mmu_notifier_invalidate_range_start(&range);
986 dst_pgd = pgd_offset(dst_mm, addr);
987 src_pgd = pgd_offset(src_mm, addr);
989 next = pgd_addr_end(addr, end);
990 if (pgd_none_or_clear_bad(src_pgd))
992 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
997 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1000 mmu_notifier_invalidate_range_end(&range);
1004 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1005 struct vm_area_struct *vma, pmd_t *pmd,
1006 unsigned long addr, unsigned long end,
1007 struct zap_details *details)
1009 struct mm_struct *mm = tlb->mm;
1010 int force_flush = 0;
1011 int rss[NR_MM_COUNTERS];
1017 tlb_change_page_size(tlb, PAGE_SIZE);
1020 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1022 flush_tlb_batched_pending(mm);
1023 arch_enter_lazy_mmu_mode();
1026 if (pte_none(ptent))
1032 if (pte_present(ptent)) {
1035 page = vm_normal_page(vma, addr, ptent);
1036 if (unlikely(details) && page) {
1038 * unmap_shared_mapping_pages() wants to
1039 * invalidate cache without truncating:
1040 * unmap shared but keep private pages.
1042 if (details->check_mapping &&
1043 details->check_mapping != page_rmapping(page))
1046 ptent = ptep_get_and_clear_full(mm, addr, pte,
1048 tlb_remove_tlb_entry(tlb, pte, addr);
1049 if (unlikely(!page))
1052 if (!PageAnon(page)) {
1053 if (pte_dirty(ptent)) {
1055 set_page_dirty(page);
1057 if (pte_young(ptent) &&
1058 likely(!(vma->vm_flags & VM_SEQ_READ)))
1059 mark_page_accessed(page);
1061 rss[mm_counter(page)]--;
1062 page_remove_rmap(page, false);
1063 if (unlikely(page_mapcount(page) < 0))
1064 print_bad_pte(vma, addr, ptent, page);
1065 if (unlikely(__tlb_remove_page(tlb, page))) {
1073 entry = pte_to_swp_entry(ptent);
1074 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1075 struct page *page = device_private_entry_to_page(entry);
1077 if (unlikely(details && details->check_mapping)) {
1079 * unmap_shared_mapping_pages() wants to
1080 * invalidate cache without truncating:
1081 * unmap shared but keep private pages.
1083 if (details->check_mapping !=
1084 page_rmapping(page))
1088 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1089 rss[mm_counter(page)]--;
1090 page_remove_rmap(page, false);
1095 /* If details->check_mapping, we leave swap entries. */
1096 if (unlikely(details))
1099 if (!non_swap_entry(entry))
1101 else if (is_migration_entry(entry)) {
1104 page = migration_entry_to_page(entry);
1105 rss[mm_counter(page)]--;
1107 if (unlikely(!free_swap_and_cache(entry)))
1108 print_bad_pte(vma, addr, ptent, NULL);
1109 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1110 } while (pte++, addr += PAGE_SIZE, addr != end);
1112 add_mm_rss_vec(mm, rss);
1113 arch_leave_lazy_mmu_mode();
1115 /* Do the actual TLB flush before dropping ptl */
1117 tlb_flush_mmu_tlbonly(tlb);
1118 pte_unmap_unlock(start_pte, ptl);
1121 * If we forced a TLB flush (either due to running out of
1122 * batch buffers or because we needed to flush dirty TLB
1123 * entries before releasing the ptl), free the batched
1124 * memory too. Restart if we didn't do everything.
1139 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1140 struct vm_area_struct *vma, pud_t *pud,
1141 unsigned long addr, unsigned long end,
1142 struct zap_details *details)
1147 pmd = pmd_offset(pud, addr);
1149 next = pmd_addr_end(addr, end);
1150 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1151 if (next - addr != HPAGE_PMD_SIZE)
1152 __split_huge_pmd(vma, pmd, addr, false, NULL);
1153 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1158 * Here there can be other concurrent MADV_DONTNEED or
1159 * trans huge page faults running, and if the pmd is
1160 * none or trans huge it can change under us. This is
1161 * because MADV_DONTNEED holds the mmap_sem in read
1164 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1166 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1169 } while (pmd++, addr = next, addr != end);
1174 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1175 struct vm_area_struct *vma, p4d_t *p4d,
1176 unsigned long addr, unsigned long end,
1177 struct zap_details *details)
1182 pud = pud_offset(p4d, addr);
1184 next = pud_addr_end(addr, end);
1185 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1186 if (next - addr != HPAGE_PUD_SIZE) {
1187 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1188 split_huge_pud(vma, pud, addr);
1189 } else if (zap_huge_pud(tlb, vma, pud, addr))
1193 if (pud_none_or_clear_bad(pud))
1195 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1198 } while (pud++, addr = next, addr != end);
1203 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1204 struct vm_area_struct *vma, pgd_t *pgd,
1205 unsigned long addr, unsigned long end,
1206 struct zap_details *details)
1211 p4d = p4d_offset(pgd, addr);
1213 next = p4d_addr_end(addr, end);
1214 if (p4d_none_or_clear_bad(p4d))
1216 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1217 } while (p4d++, addr = next, addr != end);
1222 void unmap_page_range(struct mmu_gather *tlb,
1223 struct vm_area_struct *vma,
1224 unsigned long addr, unsigned long end,
1225 struct zap_details *details)
1230 BUG_ON(addr >= end);
1231 tlb_start_vma(tlb, vma);
1232 pgd = pgd_offset(vma->vm_mm, addr);
1234 next = pgd_addr_end(addr, end);
1235 if (pgd_none_or_clear_bad(pgd))
1237 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1238 } while (pgd++, addr = next, addr != end);
1239 tlb_end_vma(tlb, vma);
1243 static void unmap_single_vma(struct mmu_gather *tlb,
1244 struct vm_area_struct *vma, unsigned long start_addr,
1245 unsigned long end_addr,
1246 struct zap_details *details)
1248 unsigned long start = max(vma->vm_start, start_addr);
1251 if (start >= vma->vm_end)
1253 end = min(vma->vm_end, end_addr);
1254 if (end <= vma->vm_start)
1258 uprobe_munmap(vma, start, end);
1260 if (unlikely(vma->vm_flags & VM_PFNMAP))
1261 untrack_pfn(vma, 0, 0);
1264 if (unlikely(is_vm_hugetlb_page(vma))) {
1266 * It is undesirable to test vma->vm_file as it
1267 * should be non-null for valid hugetlb area.
1268 * However, vm_file will be NULL in the error
1269 * cleanup path of mmap_region. When
1270 * hugetlbfs ->mmap method fails,
1271 * mmap_region() nullifies vma->vm_file
1272 * before calling this function to clean up.
1273 * Since no pte has actually been setup, it is
1274 * safe to do nothing in this case.
1277 i_mmap_lock_write(vma->vm_file->f_mapping);
1278 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1279 i_mmap_unlock_write(vma->vm_file->f_mapping);
1282 unmap_page_range(tlb, vma, start, end, details);
1287 * unmap_vmas - unmap a range of memory covered by a list of vma's
1288 * @tlb: address of the caller's struct mmu_gather
1289 * @vma: the starting vma
1290 * @start_addr: virtual address at which to start unmapping
1291 * @end_addr: virtual address at which to end unmapping
1293 * Unmap all pages in the vma list.
1295 * Only addresses between `start' and `end' will be unmapped.
1297 * The VMA list must be sorted in ascending virtual address order.
1299 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1300 * range after unmap_vmas() returns. So the only responsibility here is to
1301 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1302 * drops the lock and schedules.
1304 void unmap_vmas(struct mmu_gather *tlb,
1305 struct vm_area_struct *vma, unsigned long start_addr,
1306 unsigned long end_addr)
1308 struct mmu_notifier_range range;
1310 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1311 start_addr, end_addr);
1312 mmu_notifier_invalidate_range_start(&range);
1313 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1314 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1315 mmu_notifier_invalidate_range_end(&range);
1319 * zap_page_range - remove user pages in a given range
1320 * @vma: vm_area_struct holding the applicable pages
1321 * @start: starting address of pages to zap
1322 * @size: number of bytes to zap
1324 * Caller must protect the VMA list
1326 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1329 struct mmu_notifier_range range;
1330 struct mmu_gather tlb;
1333 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1334 start, start + size);
1335 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1336 update_hiwater_rss(vma->vm_mm);
1337 mmu_notifier_invalidate_range_start(&range);
1338 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1339 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1340 mmu_notifier_invalidate_range_end(&range);
1341 tlb_finish_mmu(&tlb, start, range.end);
1345 * zap_page_range_single - remove user pages in a given range
1346 * @vma: vm_area_struct holding the applicable pages
1347 * @address: starting address of pages to zap
1348 * @size: number of bytes to zap
1349 * @details: details of shared cache invalidation
1351 * The range must fit into one VMA.
1353 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1354 unsigned long size, struct zap_details *details)
1356 struct mmu_notifier_range range;
1357 struct mmu_gather tlb;
1360 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1361 address, address + size);
1362 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1363 update_hiwater_rss(vma->vm_mm);
1364 mmu_notifier_invalidate_range_start(&range);
1365 unmap_single_vma(&tlb, vma, address, range.end, details);
1366 mmu_notifier_invalidate_range_end(&range);
1367 tlb_finish_mmu(&tlb, address, range.end);
1371 * zap_vma_ptes - remove ptes mapping the vma
1372 * @vma: vm_area_struct holding ptes to be zapped
1373 * @address: starting address of pages to zap
1374 * @size: number of bytes to zap
1376 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1378 * The entire address range must be fully contained within the vma.
1381 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1384 if (address < vma->vm_start || address + size > vma->vm_end ||
1385 !(vma->vm_flags & VM_PFNMAP))
1388 zap_page_range_single(vma, address, size, NULL);
1390 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1392 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1400 pgd = pgd_offset(mm, addr);
1401 p4d = p4d_alloc(mm, pgd, addr);
1404 pud = pud_alloc(mm, p4d, addr);
1407 pmd = pmd_alloc(mm, pud, addr);
1411 VM_BUG_ON(pmd_trans_huge(*pmd));
1412 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1416 * This is the old fallback for page remapping.
1418 * For historical reasons, it only allows reserved pages. Only
1419 * old drivers should use this, and they needed to mark their
1420 * pages reserved for the old functions anyway.
1422 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1423 struct page *page, pgprot_t prot)
1425 struct mm_struct *mm = vma->vm_mm;
1431 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1434 flush_dcache_page(page);
1435 pte = get_locked_pte(mm, addr, &ptl);
1439 if (!pte_none(*pte))
1442 /* Ok, finally just insert the thing.. */
1444 inc_mm_counter_fast(mm, mm_counter_file(page));
1445 page_add_file_rmap(page, false);
1446 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1450 pte_unmap_unlock(pte, ptl);
1456 * vm_insert_page - insert single page into user vma
1457 * @vma: user vma to map to
1458 * @addr: target user address of this page
1459 * @page: source kernel page
1461 * This allows drivers to insert individual pages they've allocated
1464 * The page has to be a nice clean _individual_ kernel allocation.
1465 * If you allocate a compound page, you need to have marked it as
1466 * such (__GFP_COMP), or manually just split the page up yourself
1467 * (see split_page()).
1469 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1470 * took an arbitrary page protection parameter. This doesn't allow
1471 * that. Your vma protection will have to be set up correctly, which
1472 * means that if you want a shared writable mapping, you'd better
1473 * ask for a shared writable mapping!
1475 * The page does not need to be reserved.
1477 * Usually this function is called from f_op->mmap() handler
1478 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1479 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1480 * function from other places, for example from page-fault handler.
1482 * Return: %0 on success, negative error code otherwise.
1484 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1487 if (addr < vma->vm_start || addr >= vma->vm_end)
1489 if (!page_count(page))
1491 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1492 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1493 BUG_ON(vma->vm_flags & VM_PFNMAP);
1494 vma->vm_flags |= VM_MIXEDMAP;
1496 return insert_page(vma, addr, page, vma->vm_page_prot);
1498 EXPORT_SYMBOL(vm_insert_page);
1501 * __vm_map_pages - maps range of kernel pages into user vma
1502 * @vma: user vma to map to
1503 * @pages: pointer to array of source kernel pages
1504 * @num: number of pages in page array
1505 * @offset: user's requested vm_pgoff
1507 * This allows drivers to map range of kernel pages into a user vma.
1509 * Return: 0 on success and error code otherwise.
1511 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1512 unsigned long num, unsigned long offset)
1514 unsigned long count = vma_pages(vma);
1515 unsigned long uaddr = vma->vm_start;
1518 /* Fail if the user requested offset is beyond the end of the object */
1522 /* Fail if the user requested size exceeds available object size */
1523 if (count > num - offset)
1526 for (i = 0; i < count; i++) {
1527 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1537 * vm_map_pages - maps range of kernel pages starts with non zero offset
1538 * @vma: user vma to map to
1539 * @pages: pointer to array of source kernel pages
1540 * @num: number of pages in page array
1542 * Maps an object consisting of @num pages, catering for the user's
1543 * requested vm_pgoff
1545 * If we fail to insert any page into the vma, the function will return
1546 * immediately leaving any previously inserted pages present. Callers
1547 * from the mmap handler may immediately return the error as their caller
1548 * will destroy the vma, removing any successfully inserted pages. Other
1549 * callers should make their own arrangements for calling unmap_region().
1551 * Context: Process context. Called by mmap handlers.
1552 * Return: 0 on success and error code otherwise.
1554 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1557 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1559 EXPORT_SYMBOL(vm_map_pages);
1562 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1563 * @vma: user vma to map to
1564 * @pages: pointer to array of source kernel pages
1565 * @num: number of pages in page array
1567 * Similar to vm_map_pages(), except that it explicitly sets the offset
1568 * to 0. This function is intended for the drivers that did not consider
1571 * Context: Process context. Called by mmap handlers.
1572 * Return: 0 on success and error code otherwise.
1574 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1577 return __vm_map_pages(vma, pages, num, 0);
1579 EXPORT_SYMBOL(vm_map_pages_zero);
1581 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1582 pfn_t pfn, pgprot_t prot, bool mkwrite)
1584 struct mm_struct *mm = vma->vm_mm;
1588 pte = get_locked_pte(mm, addr, &ptl);
1590 return VM_FAULT_OOM;
1591 if (!pte_none(*pte)) {
1594 * For read faults on private mappings the PFN passed
1595 * in may not match the PFN we have mapped if the
1596 * mapped PFN is a writeable COW page. In the mkwrite
1597 * case we are creating a writable PTE for a shared
1598 * mapping and we expect the PFNs to match. If they
1599 * don't match, we are likely racing with block
1600 * allocation and mapping invalidation so just skip the
1603 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1604 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1607 entry = pte_mkyoung(*pte);
1608 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1609 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1610 update_mmu_cache(vma, addr, pte);
1615 /* Ok, finally just insert the thing.. */
1616 if (pfn_t_devmap(pfn))
1617 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1619 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1622 entry = pte_mkyoung(entry);
1623 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1626 set_pte_at(mm, addr, pte, entry);
1627 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1630 pte_unmap_unlock(pte, ptl);
1631 return VM_FAULT_NOPAGE;
1635 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1636 * @vma: user vma to map to
1637 * @addr: target user address of this page
1638 * @pfn: source kernel pfn
1639 * @pgprot: pgprot flags for the inserted page
1641 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1642 * to override pgprot on a per-page basis.
1644 * This only makes sense for IO mappings, and it makes no sense for
1645 * COW mappings. In general, using multiple vmas is preferable;
1646 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1649 * Context: Process context. May allocate using %GFP_KERNEL.
1650 * Return: vm_fault_t value.
1652 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1653 unsigned long pfn, pgprot_t pgprot)
1656 * Technically, architectures with pte_special can avoid all these
1657 * restrictions (same for remap_pfn_range). However we would like
1658 * consistency in testing and feature parity among all, so we should
1659 * try to keep these invariants in place for everybody.
1661 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1662 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1663 (VM_PFNMAP|VM_MIXEDMAP));
1664 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1665 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1667 if (addr < vma->vm_start || addr >= vma->vm_end)
1668 return VM_FAULT_SIGBUS;
1670 if (!pfn_modify_allowed(pfn, pgprot))
1671 return VM_FAULT_SIGBUS;
1673 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1675 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1678 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1681 * vmf_insert_pfn - insert single pfn into user vma
1682 * @vma: user vma to map to
1683 * @addr: target user address of this page
1684 * @pfn: source kernel pfn
1686 * Similar to vm_insert_page, this allows drivers to insert individual pages
1687 * they've allocated into a user vma. Same comments apply.
1689 * This function should only be called from a vm_ops->fault handler, and
1690 * in that case the handler should return the result of this function.
1692 * vma cannot be a COW mapping.
1694 * As this is called only for pages that do not currently exist, we
1695 * do not need to flush old virtual caches or the TLB.
1697 * Context: Process context. May allocate using %GFP_KERNEL.
1698 * Return: vm_fault_t value.
1700 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1703 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1705 EXPORT_SYMBOL(vmf_insert_pfn);
1707 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1709 /* these checks mirror the abort conditions in vm_normal_page */
1710 if (vma->vm_flags & VM_MIXEDMAP)
1712 if (pfn_t_devmap(pfn))
1714 if (pfn_t_special(pfn))
1716 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1721 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1722 unsigned long addr, pfn_t pfn, bool mkwrite)
1724 pgprot_t pgprot = vma->vm_page_prot;
1727 BUG_ON(!vm_mixed_ok(vma, pfn));
1729 if (addr < vma->vm_start || addr >= vma->vm_end)
1730 return VM_FAULT_SIGBUS;
1732 track_pfn_insert(vma, &pgprot, pfn);
1734 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1735 return VM_FAULT_SIGBUS;
1738 * If we don't have pte special, then we have to use the pfn_valid()
1739 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1740 * refcount the page if pfn_valid is true (hence insert_page rather
1741 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1742 * without pte special, it would there be refcounted as a normal page.
1744 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1745 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1749 * At this point we are committed to insert_page()
1750 * regardless of whether the caller specified flags that
1751 * result in pfn_t_has_page() == false.
1753 page = pfn_to_page(pfn_t_to_pfn(pfn));
1754 err = insert_page(vma, addr, page, pgprot);
1756 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1760 return VM_FAULT_OOM;
1761 if (err < 0 && err != -EBUSY)
1762 return VM_FAULT_SIGBUS;
1764 return VM_FAULT_NOPAGE;
1767 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1770 return __vm_insert_mixed(vma, addr, pfn, false);
1772 EXPORT_SYMBOL(vmf_insert_mixed);
1775 * If the insertion of PTE failed because someone else already added a
1776 * different entry in the mean time, we treat that as success as we assume
1777 * the same entry was actually inserted.
1779 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1780 unsigned long addr, pfn_t pfn)
1782 return __vm_insert_mixed(vma, addr, pfn, true);
1784 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1787 * maps a range of physical memory into the requested pages. the old
1788 * mappings are removed. any references to nonexistent pages results
1789 * in null mappings (currently treated as "copy-on-access")
1791 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1792 unsigned long addr, unsigned long end,
1793 unsigned long pfn, pgprot_t prot)
1799 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1802 arch_enter_lazy_mmu_mode();
1804 BUG_ON(!pte_none(*pte));
1805 if (!pfn_modify_allowed(pfn, prot)) {
1809 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1811 } while (pte++, addr += PAGE_SIZE, addr != end);
1812 arch_leave_lazy_mmu_mode();
1813 pte_unmap_unlock(pte - 1, ptl);
1817 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1818 unsigned long addr, unsigned long end,
1819 unsigned long pfn, pgprot_t prot)
1825 pfn -= addr >> PAGE_SHIFT;
1826 pmd = pmd_alloc(mm, pud, addr);
1829 VM_BUG_ON(pmd_trans_huge(*pmd));
1831 next = pmd_addr_end(addr, end);
1832 err = remap_pte_range(mm, pmd, addr, next,
1833 pfn + (addr >> PAGE_SHIFT), prot);
1836 } while (pmd++, addr = next, addr != end);
1840 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1841 unsigned long addr, unsigned long end,
1842 unsigned long pfn, pgprot_t prot)
1848 pfn -= addr >> PAGE_SHIFT;
1849 pud = pud_alloc(mm, p4d, addr);
1853 next = pud_addr_end(addr, end);
1854 err = remap_pmd_range(mm, pud, addr, next,
1855 pfn + (addr >> PAGE_SHIFT), prot);
1858 } while (pud++, addr = next, addr != end);
1862 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1863 unsigned long addr, unsigned long end,
1864 unsigned long pfn, pgprot_t prot)
1870 pfn -= addr >> PAGE_SHIFT;
1871 p4d = p4d_alloc(mm, pgd, addr);
1875 next = p4d_addr_end(addr, end);
1876 err = remap_pud_range(mm, p4d, addr, next,
1877 pfn + (addr >> PAGE_SHIFT), prot);
1880 } while (p4d++, addr = next, addr != end);
1885 * remap_pfn_range - remap kernel memory to userspace
1886 * @vma: user vma to map to
1887 * @addr: target user address to start at
1888 * @pfn: physical address of kernel memory
1889 * @size: size of map area
1890 * @prot: page protection flags for this mapping
1892 * Note: this is only safe if the mm semaphore is held when called.
1894 * Return: %0 on success, negative error code otherwise.
1896 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1897 unsigned long pfn, unsigned long size, pgprot_t prot)
1901 unsigned long end = addr + PAGE_ALIGN(size);
1902 struct mm_struct *mm = vma->vm_mm;
1903 unsigned long remap_pfn = pfn;
1907 * Physically remapped pages are special. Tell the
1908 * rest of the world about it:
1909 * VM_IO tells people not to look at these pages
1910 * (accesses can have side effects).
1911 * VM_PFNMAP tells the core MM that the base pages are just
1912 * raw PFN mappings, and do not have a "struct page" associated
1915 * Disable vma merging and expanding with mremap().
1917 * Omit vma from core dump, even when VM_IO turned off.
1919 * There's a horrible special case to handle copy-on-write
1920 * behaviour that some programs depend on. We mark the "original"
1921 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1922 * See vm_normal_page() for details.
1924 if (is_cow_mapping(vma->vm_flags)) {
1925 if (addr != vma->vm_start || end != vma->vm_end)
1927 vma->vm_pgoff = pfn;
1930 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1934 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1936 BUG_ON(addr >= end);
1937 pfn -= addr >> PAGE_SHIFT;
1938 pgd = pgd_offset(mm, addr);
1939 flush_cache_range(vma, addr, end);
1941 next = pgd_addr_end(addr, end);
1942 err = remap_p4d_range(mm, pgd, addr, next,
1943 pfn + (addr >> PAGE_SHIFT), prot);
1946 } while (pgd++, addr = next, addr != end);
1949 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1953 EXPORT_SYMBOL(remap_pfn_range);
1956 * vm_iomap_memory - remap memory to userspace
1957 * @vma: user vma to map to
1958 * @start: start of area
1959 * @len: size of area
1961 * This is a simplified io_remap_pfn_range() for common driver use. The
1962 * driver just needs to give us the physical memory range to be mapped,
1963 * we'll figure out the rest from the vma information.
1965 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1966 * whatever write-combining details or similar.
1968 * Return: %0 on success, negative error code otherwise.
1970 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1972 unsigned long vm_len, pfn, pages;
1974 /* Check that the physical memory area passed in looks valid */
1975 if (start + len < start)
1978 * You *really* shouldn't map things that aren't page-aligned,
1979 * but we've historically allowed it because IO memory might
1980 * just have smaller alignment.
1982 len += start & ~PAGE_MASK;
1983 pfn = start >> PAGE_SHIFT;
1984 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1985 if (pfn + pages < pfn)
1988 /* We start the mapping 'vm_pgoff' pages into the area */
1989 if (vma->vm_pgoff > pages)
1991 pfn += vma->vm_pgoff;
1992 pages -= vma->vm_pgoff;
1994 /* Can we fit all of the mapping? */
1995 vm_len = vma->vm_end - vma->vm_start;
1996 if (vm_len >> PAGE_SHIFT > pages)
1999 /* Ok, let it rip */
2000 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2002 EXPORT_SYMBOL(vm_iomap_memory);
2004 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2005 unsigned long addr, unsigned long end,
2006 pte_fn_t fn, void *data)
2010 spinlock_t *uninitialized_var(ptl);
2012 pte = (mm == &init_mm) ?
2013 pte_alloc_kernel(pmd, addr) :
2014 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2018 BUG_ON(pmd_huge(*pmd));
2020 arch_enter_lazy_mmu_mode();
2023 err = fn(pte++, addr, data);
2026 } while (addr += PAGE_SIZE, addr != end);
2028 arch_leave_lazy_mmu_mode();
2031 pte_unmap_unlock(pte-1, ptl);
2035 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2036 unsigned long addr, unsigned long end,
2037 pte_fn_t fn, void *data)
2043 BUG_ON(pud_huge(*pud));
2045 pmd = pmd_alloc(mm, pud, addr);
2049 next = pmd_addr_end(addr, end);
2050 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2053 } while (pmd++, addr = next, addr != end);
2057 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2058 unsigned long addr, unsigned long end,
2059 pte_fn_t fn, void *data)
2065 pud = pud_alloc(mm, p4d, addr);
2069 next = pud_addr_end(addr, end);
2070 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2073 } while (pud++, addr = next, addr != end);
2077 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2078 unsigned long addr, unsigned long end,
2079 pte_fn_t fn, void *data)
2085 p4d = p4d_alloc(mm, pgd, addr);
2089 next = p4d_addr_end(addr, end);
2090 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2093 } while (p4d++, addr = next, addr != end);
2098 * Scan a region of virtual memory, filling in page tables as necessary
2099 * and calling a provided function on each leaf page table.
2101 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2102 unsigned long size, pte_fn_t fn, void *data)
2106 unsigned long end = addr + size;
2109 if (WARN_ON(addr >= end))
2112 pgd = pgd_offset(mm, addr);
2114 next = pgd_addr_end(addr, end);
2115 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2118 } while (pgd++, addr = next, addr != end);
2122 EXPORT_SYMBOL_GPL(apply_to_page_range);
2125 * handle_pte_fault chooses page fault handler according to an entry which was
2126 * read non-atomically. Before making any commitment, on those architectures
2127 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2128 * parts, do_swap_page must check under lock before unmapping the pte and
2129 * proceeding (but do_wp_page is only called after already making such a check;
2130 * and do_anonymous_page can safely check later on).
2132 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2133 pte_t *page_table, pte_t orig_pte)
2136 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2137 if (sizeof(pte_t) > sizeof(unsigned long)) {
2138 spinlock_t *ptl = pte_lockptr(mm, pmd);
2140 same = pte_same(*page_table, orig_pte);
2144 pte_unmap(page_table);
2148 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2150 debug_dma_assert_idle(src);
2153 * If the source page was a PFN mapping, we don't have
2154 * a "struct page" for it. We do a best-effort copy by
2155 * just copying from the original user address. If that
2156 * fails, we just zero-fill it. Live with it.
2158 if (unlikely(!src)) {
2159 void *kaddr = kmap_atomic(dst);
2160 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2163 * This really shouldn't fail, because the page is there
2164 * in the page tables. But it might just be unreadable,
2165 * in which case we just give up and fill the result with
2168 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2170 kunmap_atomic(kaddr);
2171 flush_dcache_page(dst);
2173 copy_user_highpage(dst, src, va, vma);
2176 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2178 struct file *vm_file = vma->vm_file;
2181 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2184 * Special mappings (e.g. VDSO) do not have any file so fake
2185 * a default GFP_KERNEL for them.
2191 * Notify the address space that the page is about to become writable so that
2192 * it can prohibit this or wait for the page to get into an appropriate state.
2194 * We do this without the lock held, so that it can sleep if it needs to.
2196 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2199 struct page *page = vmf->page;
2200 unsigned int old_flags = vmf->flags;
2202 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2204 if (vmf->vma->vm_file &&
2205 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2206 return VM_FAULT_SIGBUS;
2208 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2209 /* Restore original flags so that caller is not surprised */
2210 vmf->flags = old_flags;
2211 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2213 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2215 if (!page->mapping) {
2217 return 0; /* retry */
2219 ret |= VM_FAULT_LOCKED;
2221 VM_BUG_ON_PAGE(!PageLocked(page), page);
2226 * Handle dirtying of a page in shared file mapping on a write fault.
2228 * The function expects the page to be locked and unlocks it.
2230 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2233 struct address_space *mapping;
2235 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2237 dirtied = set_page_dirty(page);
2238 VM_BUG_ON_PAGE(PageAnon(page), page);
2240 * Take a local copy of the address_space - page.mapping may be zeroed
2241 * by truncate after unlock_page(). The address_space itself remains
2242 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2243 * release semantics to prevent the compiler from undoing this copying.
2245 mapping = page_rmapping(page);
2248 if ((dirtied || page_mkwrite) && mapping) {
2250 * Some device drivers do not set page.mapping
2251 * but still dirty their pages
2253 balance_dirty_pages_ratelimited(mapping);
2257 file_update_time(vma->vm_file);
2261 * Handle write page faults for pages that can be reused in the current vma
2263 * This can happen either due to the mapping being with the VM_SHARED flag,
2264 * or due to us being the last reference standing to the page. In either
2265 * case, all we need to do here is to mark the page as writable and update
2266 * any related book-keeping.
2268 static inline void wp_page_reuse(struct vm_fault *vmf)
2269 __releases(vmf->ptl)
2271 struct vm_area_struct *vma = vmf->vma;
2272 struct page *page = vmf->page;
2275 * Clear the pages cpupid information as the existing
2276 * information potentially belongs to a now completely
2277 * unrelated process.
2280 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2282 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2283 entry = pte_mkyoung(vmf->orig_pte);
2284 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2285 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2286 update_mmu_cache(vma, vmf->address, vmf->pte);
2287 pte_unmap_unlock(vmf->pte, vmf->ptl);
2291 * Handle the case of a page which we actually need to copy to a new page.
2293 * Called with mmap_sem locked and the old page referenced, but
2294 * without the ptl held.
2296 * High level logic flow:
2298 * - Allocate a page, copy the content of the old page to the new one.
2299 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2300 * - Take the PTL. If the pte changed, bail out and release the allocated page
2301 * - If the pte is still the way we remember it, update the page table and all
2302 * relevant references. This includes dropping the reference the page-table
2303 * held to the old page, as well as updating the rmap.
2304 * - In any case, unlock the PTL and drop the reference we took to the old page.
2306 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2308 struct vm_area_struct *vma = vmf->vma;
2309 struct mm_struct *mm = vma->vm_mm;
2310 struct page *old_page = vmf->page;
2311 struct page *new_page = NULL;
2313 int page_copied = 0;
2314 struct mem_cgroup *memcg;
2315 struct mmu_notifier_range range;
2317 if (unlikely(anon_vma_prepare(vma)))
2320 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2321 new_page = alloc_zeroed_user_highpage_movable(vma,
2326 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2330 cow_user_page(new_page, old_page, vmf->address, vma);
2333 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2336 __SetPageUptodate(new_page);
2338 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2339 vmf->address & PAGE_MASK,
2340 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2341 mmu_notifier_invalidate_range_start(&range);
2344 * Re-check the pte - we dropped the lock
2346 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2347 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2349 if (!PageAnon(old_page)) {
2350 dec_mm_counter_fast(mm,
2351 mm_counter_file(old_page));
2352 inc_mm_counter_fast(mm, MM_ANONPAGES);
2355 inc_mm_counter_fast(mm, MM_ANONPAGES);
2357 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2358 entry = mk_pte(new_page, vma->vm_page_prot);
2359 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2361 * Clear the pte entry and flush it first, before updating the
2362 * pte with the new entry. This will avoid a race condition
2363 * seen in the presence of one thread doing SMC and another
2366 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2367 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2368 mem_cgroup_commit_charge(new_page, memcg, false, false);
2369 lru_cache_add_active_or_unevictable(new_page, vma);
2371 * We call the notify macro here because, when using secondary
2372 * mmu page tables (such as kvm shadow page tables), we want the
2373 * new page to be mapped directly into the secondary page table.
2375 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2376 update_mmu_cache(vma, vmf->address, vmf->pte);
2379 * Only after switching the pte to the new page may
2380 * we remove the mapcount here. Otherwise another
2381 * process may come and find the rmap count decremented
2382 * before the pte is switched to the new page, and
2383 * "reuse" the old page writing into it while our pte
2384 * here still points into it and can be read by other
2387 * The critical issue is to order this
2388 * page_remove_rmap with the ptp_clear_flush above.
2389 * Those stores are ordered by (if nothing else,)
2390 * the barrier present in the atomic_add_negative
2391 * in page_remove_rmap.
2393 * Then the TLB flush in ptep_clear_flush ensures that
2394 * no process can access the old page before the
2395 * decremented mapcount is visible. And the old page
2396 * cannot be reused until after the decremented
2397 * mapcount is visible. So transitively, TLBs to
2398 * old page will be flushed before it can be reused.
2400 page_remove_rmap(old_page, false);
2403 /* Free the old page.. */
2404 new_page = old_page;
2407 mem_cgroup_cancel_charge(new_page, memcg, false);
2413 pte_unmap_unlock(vmf->pte, vmf->ptl);
2415 * No need to double call mmu_notifier->invalidate_range() callback as
2416 * the above ptep_clear_flush_notify() did already call it.
2418 mmu_notifier_invalidate_range_only_end(&range);
2421 * Don't let another task, with possibly unlocked vma,
2422 * keep the mlocked page.
2424 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2425 lock_page(old_page); /* LRU manipulation */
2426 if (PageMlocked(old_page))
2427 munlock_vma_page(old_page);
2428 unlock_page(old_page);
2432 return page_copied ? VM_FAULT_WRITE : 0;
2438 return VM_FAULT_OOM;
2442 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2443 * writeable once the page is prepared
2445 * @vmf: structure describing the fault
2447 * This function handles all that is needed to finish a write page fault in a
2448 * shared mapping due to PTE being read-only once the mapped page is prepared.
2449 * It handles locking of PTE and modifying it.
2451 * The function expects the page to be locked or other protection against
2452 * concurrent faults / writeback (such as DAX radix tree locks).
2454 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2455 * we acquired PTE lock.
2457 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2459 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2460 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2463 * We might have raced with another page fault while we released the
2464 * pte_offset_map_lock.
2466 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2467 pte_unmap_unlock(vmf->pte, vmf->ptl);
2468 return VM_FAULT_NOPAGE;
2475 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2478 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2480 struct vm_area_struct *vma = vmf->vma;
2482 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2485 pte_unmap_unlock(vmf->pte, vmf->ptl);
2486 vmf->flags |= FAULT_FLAG_MKWRITE;
2487 ret = vma->vm_ops->pfn_mkwrite(vmf);
2488 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2490 return finish_mkwrite_fault(vmf);
2493 return VM_FAULT_WRITE;
2496 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2497 __releases(vmf->ptl)
2499 struct vm_area_struct *vma = vmf->vma;
2501 get_page(vmf->page);
2503 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2506 pte_unmap_unlock(vmf->pte, vmf->ptl);
2507 tmp = do_page_mkwrite(vmf);
2508 if (unlikely(!tmp || (tmp &
2509 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2510 put_page(vmf->page);
2513 tmp = finish_mkwrite_fault(vmf);
2514 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2515 unlock_page(vmf->page);
2516 put_page(vmf->page);
2521 lock_page(vmf->page);
2523 fault_dirty_shared_page(vma, vmf->page);
2524 put_page(vmf->page);
2526 return VM_FAULT_WRITE;
2530 * This routine handles present pages, when users try to write
2531 * to a shared page. It is done by copying the page to a new address
2532 * and decrementing the shared-page counter for the old page.
2534 * Note that this routine assumes that the protection checks have been
2535 * done by the caller (the low-level page fault routine in most cases).
2536 * Thus we can safely just mark it writable once we've done any necessary
2539 * We also mark the page dirty at this point even though the page will
2540 * change only once the write actually happens. This avoids a few races,
2541 * and potentially makes it more efficient.
2543 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2544 * but allow concurrent faults), with pte both mapped and locked.
2545 * We return with mmap_sem still held, but pte unmapped and unlocked.
2547 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2548 __releases(vmf->ptl)
2550 struct vm_area_struct *vma = vmf->vma;
2552 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2555 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2558 * We should not cow pages in a shared writeable mapping.
2559 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2561 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2562 (VM_WRITE|VM_SHARED))
2563 return wp_pfn_shared(vmf);
2565 pte_unmap_unlock(vmf->pte, vmf->ptl);
2566 return wp_page_copy(vmf);
2570 * Take out anonymous pages first, anonymous shared vmas are
2571 * not dirty accountable.
2573 if (PageAnon(vmf->page)) {
2574 int total_map_swapcount;
2575 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2576 page_count(vmf->page) != 1))
2578 if (!trylock_page(vmf->page)) {
2579 get_page(vmf->page);
2580 pte_unmap_unlock(vmf->pte, vmf->ptl);
2581 lock_page(vmf->page);
2582 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2583 vmf->address, &vmf->ptl);
2584 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2585 unlock_page(vmf->page);
2586 pte_unmap_unlock(vmf->pte, vmf->ptl);
2587 put_page(vmf->page);
2590 put_page(vmf->page);
2592 if (PageKsm(vmf->page)) {
2593 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2595 unlock_page(vmf->page);
2599 return VM_FAULT_WRITE;
2601 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2602 if (total_map_swapcount == 1) {
2604 * The page is all ours. Move it to
2605 * our anon_vma so the rmap code will
2606 * not search our parent or siblings.
2607 * Protected against the rmap code by
2610 page_move_anon_rmap(vmf->page, vma);
2612 unlock_page(vmf->page);
2614 return VM_FAULT_WRITE;
2616 unlock_page(vmf->page);
2617 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2618 (VM_WRITE|VM_SHARED))) {
2619 return wp_page_shared(vmf);
2623 * Ok, we need to copy. Oh, well..
2625 get_page(vmf->page);
2627 pte_unmap_unlock(vmf->pte, vmf->ptl);
2628 return wp_page_copy(vmf);
2631 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2632 unsigned long start_addr, unsigned long end_addr,
2633 struct zap_details *details)
2635 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2638 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2639 struct zap_details *details)
2641 struct vm_area_struct *vma;
2642 pgoff_t vba, vea, zba, zea;
2644 vma_interval_tree_foreach(vma, root,
2645 details->first_index, details->last_index) {
2647 vba = vma->vm_pgoff;
2648 vea = vba + vma_pages(vma) - 1;
2649 zba = details->first_index;
2652 zea = details->last_index;
2656 unmap_mapping_range_vma(vma,
2657 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2658 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2664 * unmap_mapping_pages() - Unmap pages from processes.
2665 * @mapping: The address space containing pages to be unmapped.
2666 * @start: Index of first page to be unmapped.
2667 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2668 * @even_cows: Whether to unmap even private COWed pages.
2670 * Unmap the pages in this address space from any userspace process which
2671 * has them mmaped. Generally, you want to remove COWed pages as well when
2672 * a file is being truncated, but not when invalidating pages from the page
2675 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2676 pgoff_t nr, bool even_cows)
2678 struct zap_details details = { };
2680 details.check_mapping = even_cows ? NULL : mapping;
2681 details.first_index = start;
2682 details.last_index = start + nr - 1;
2683 if (details.last_index < details.first_index)
2684 details.last_index = ULONG_MAX;
2686 i_mmap_lock_write(mapping);
2687 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2688 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2689 i_mmap_unlock_write(mapping);
2693 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2694 * address_space corresponding to the specified byte range in the underlying
2697 * @mapping: the address space containing mmaps to be unmapped.
2698 * @holebegin: byte in first page to unmap, relative to the start of
2699 * the underlying file. This will be rounded down to a PAGE_SIZE
2700 * boundary. Note that this is different from truncate_pagecache(), which
2701 * must keep the partial page. In contrast, we must get rid of
2703 * @holelen: size of prospective hole in bytes. This will be rounded
2704 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2706 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2707 * but 0 when invalidating pagecache, don't throw away private data.
2709 void unmap_mapping_range(struct address_space *mapping,
2710 loff_t const holebegin, loff_t const holelen, int even_cows)
2712 pgoff_t hba = holebegin >> PAGE_SHIFT;
2713 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2715 /* Check for overflow. */
2716 if (sizeof(holelen) > sizeof(hlen)) {
2718 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2719 if (holeend & ~(long long)ULONG_MAX)
2720 hlen = ULONG_MAX - hba + 1;
2723 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2725 EXPORT_SYMBOL(unmap_mapping_range);
2728 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2729 * but allow concurrent faults), and pte mapped but not yet locked.
2730 * We return with pte unmapped and unlocked.
2732 * We return with the mmap_sem locked or unlocked in the same cases
2733 * as does filemap_fault().
2735 vm_fault_t do_swap_page(struct vm_fault *vmf)
2737 struct vm_area_struct *vma = vmf->vma;
2738 struct page *page = NULL, *swapcache;
2739 struct mem_cgroup *memcg;
2746 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2749 entry = pte_to_swp_entry(vmf->orig_pte);
2750 if (unlikely(non_swap_entry(entry))) {
2751 if (is_migration_entry(entry)) {
2752 migration_entry_wait(vma->vm_mm, vmf->pmd,
2754 } else if (is_device_private_entry(entry)) {
2755 vmf->page = device_private_entry_to_page(entry);
2756 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
2757 } else if (is_hwpoison_entry(entry)) {
2758 ret = VM_FAULT_HWPOISON;
2760 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2761 ret = VM_FAULT_SIGBUS;
2767 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2768 page = lookup_swap_cache(entry, vma, vmf->address);
2772 struct swap_info_struct *si = swp_swap_info(entry);
2774 if (si->flags & SWP_SYNCHRONOUS_IO &&
2775 __swap_count(entry) == 1) {
2776 /* skip swapcache */
2777 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2780 __SetPageLocked(page);
2781 __SetPageSwapBacked(page);
2782 set_page_private(page, entry.val);
2783 lru_cache_add_anon(page);
2784 swap_readpage(page, true);
2787 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2794 * Back out if somebody else faulted in this pte
2795 * while we released the pte lock.
2797 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2798 vmf->address, &vmf->ptl);
2799 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2801 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2805 /* Had to read the page from swap area: Major fault */
2806 ret = VM_FAULT_MAJOR;
2807 count_vm_event(PGMAJFAULT);
2808 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2809 } else if (PageHWPoison(page)) {
2811 * hwpoisoned dirty swapcache pages are kept for killing
2812 * owner processes (which may be unknown at hwpoison time)
2814 ret = VM_FAULT_HWPOISON;
2815 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2819 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2821 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2823 ret |= VM_FAULT_RETRY;
2828 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2829 * release the swapcache from under us. The page pin, and pte_same
2830 * test below, are not enough to exclude that. Even if it is still
2831 * swapcache, we need to check that the page's swap has not changed.
2833 if (unlikely((!PageSwapCache(page) ||
2834 page_private(page) != entry.val)) && swapcache)
2837 page = ksm_might_need_to_copy(page, vma, vmf->address);
2838 if (unlikely(!page)) {
2844 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2851 * Back out if somebody else already faulted in this pte.
2853 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2855 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2858 if (unlikely(!PageUptodate(page))) {
2859 ret = VM_FAULT_SIGBUS;
2864 * The page isn't present yet, go ahead with the fault.
2866 * Be careful about the sequence of operations here.
2867 * To get its accounting right, reuse_swap_page() must be called
2868 * while the page is counted on swap but not yet in mapcount i.e.
2869 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2870 * must be called after the swap_free(), or it will never succeed.
2873 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2874 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2875 pte = mk_pte(page, vma->vm_page_prot);
2876 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2877 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2878 vmf->flags &= ~FAULT_FLAG_WRITE;
2879 ret |= VM_FAULT_WRITE;
2880 exclusive = RMAP_EXCLUSIVE;
2882 flush_icache_page(vma, page);
2883 if (pte_swp_soft_dirty(vmf->orig_pte))
2884 pte = pte_mksoft_dirty(pte);
2885 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2886 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2887 vmf->orig_pte = pte;
2889 /* ksm created a completely new copy */
2890 if (unlikely(page != swapcache && swapcache)) {
2891 page_add_new_anon_rmap(page, vma, vmf->address, false);
2892 mem_cgroup_commit_charge(page, memcg, false, false);
2893 lru_cache_add_active_or_unevictable(page, vma);
2895 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2896 mem_cgroup_commit_charge(page, memcg, true, false);
2897 activate_page(page);
2901 if (mem_cgroup_swap_full(page) ||
2902 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2903 try_to_free_swap(page);
2905 if (page != swapcache && swapcache) {
2907 * Hold the lock to avoid the swap entry to be reused
2908 * until we take the PT lock for the pte_same() check
2909 * (to avoid false positives from pte_same). For
2910 * further safety release the lock after the swap_free
2911 * so that the swap count won't change under a
2912 * parallel locked swapcache.
2914 unlock_page(swapcache);
2915 put_page(swapcache);
2918 if (vmf->flags & FAULT_FLAG_WRITE) {
2919 ret |= do_wp_page(vmf);
2920 if (ret & VM_FAULT_ERROR)
2921 ret &= VM_FAULT_ERROR;
2925 /* No need to invalidate - it was non-present before */
2926 update_mmu_cache(vma, vmf->address, vmf->pte);
2928 pte_unmap_unlock(vmf->pte, vmf->ptl);
2932 mem_cgroup_cancel_charge(page, memcg, false);
2933 pte_unmap_unlock(vmf->pte, vmf->ptl);
2938 if (page != swapcache && swapcache) {
2939 unlock_page(swapcache);
2940 put_page(swapcache);
2946 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2947 * but allow concurrent faults), and pte mapped but not yet locked.
2948 * We return with mmap_sem still held, but pte unmapped and unlocked.
2950 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2952 struct vm_area_struct *vma = vmf->vma;
2953 struct mem_cgroup *memcg;
2958 /* File mapping without ->vm_ops ? */
2959 if (vma->vm_flags & VM_SHARED)
2960 return VM_FAULT_SIGBUS;
2963 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2964 * pte_offset_map() on pmds where a huge pmd might be created
2965 * from a different thread.
2967 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2968 * parallel threads are excluded by other means.
2970 * Here we only have down_read(mmap_sem).
2972 if (pte_alloc(vma->vm_mm, vmf->pmd))
2973 return VM_FAULT_OOM;
2975 /* See the comment in pte_alloc_one_map() */
2976 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2979 /* Use the zero-page for reads */
2980 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2981 !mm_forbids_zeropage(vma->vm_mm)) {
2982 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2983 vma->vm_page_prot));
2984 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2985 vmf->address, &vmf->ptl);
2986 if (!pte_none(*vmf->pte))
2988 ret = check_stable_address_space(vma->vm_mm);
2991 /* Deliver the page fault to userland, check inside PT lock */
2992 if (userfaultfd_missing(vma)) {
2993 pte_unmap_unlock(vmf->pte, vmf->ptl);
2994 return handle_userfault(vmf, VM_UFFD_MISSING);
2999 /* Allocate our own private page. */
3000 if (unlikely(anon_vma_prepare(vma)))
3002 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3006 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3011 * The memory barrier inside __SetPageUptodate makes sure that
3012 * preceeding stores to the page contents become visible before
3013 * the set_pte_at() write.
3015 __SetPageUptodate(page);
3017 entry = mk_pte(page, vma->vm_page_prot);
3018 if (vma->vm_flags & VM_WRITE)
3019 entry = pte_mkwrite(pte_mkdirty(entry));
3021 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3023 if (!pte_none(*vmf->pte))
3026 ret = check_stable_address_space(vma->vm_mm);
3030 /* Deliver the page fault to userland, check inside PT lock */
3031 if (userfaultfd_missing(vma)) {
3032 pte_unmap_unlock(vmf->pte, vmf->ptl);
3033 mem_cgroup_cancel_charge(page, memcg, false);
3035 return handle_userfault(vmf, VM_UFFD_MISSING);
3038 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3039 page_add_new_anon_rmap(page, vma, vmf->address, false);
3040 mem_cgroup_commit_charge(page, memcg, false, false);
3041 lru_cache_add_active_or_unevictable(page, vma);
3043 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3045 /* No need to invalidate - it was non-present before */
3046 update_mmu_cache(vma, vmf->address, vmf->pte);
3048 pte_unmap_unlock(vmf->pte, vmf->ptl);
3051 mem_cgroup_cancel_charge(page, memcg, false);
3057 return VM_FAULT_OOM;
3061 * The mmap_sem must have been held on entry, and may have been
3062 * released depending on flags and vma->vm_ops->fault() return value.
3063 * See filemap_fault() and __lock_page_retry().
3065 static vm_fault_t __do_fault(struct vm_fault *vmf)
3067 struct vm_area_struct *vma = vmf->vma;
3071 * Preallocate pte before we take page_lock because this might lead to
3072 * deadlocks for memcg reclaim which waits for pages under writeback:
3074 * SetPageWriteback(A)
3080 * wait_on_page_writeback(A)
3081 * SetPageWriteback(B)
3083 * # flush A, B to clear the writeback
3085 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3086 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3087 if (!vmf->prealloc_pte)
3088 return VM_FAULT_OOM;
3089 smp_wmb(); /* See comment in __pte_alloc() */
3092 ret = vma->vm_ops->fault(vmf);
3093 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3094 VM_FAULT_DONE_COW)))
3097 if (unlikely(PageHWPoison(vmf->page))) {
3098 if (ret & VM_FAULT_LOCKED)
3099 unlock_page(vmf->page);
3100 put_page(vmf->page);
3102 return VM_FAULT_HWPOISON;
3105 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3106 lock_page(vmf->page);
3108 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3114 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3115 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3116 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3117 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3119 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3121 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3124 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3126 struct vm_area_struct *vma = vmf->vma;
3128 if (!pmd_none(*vmf->pmd))
3130 if (vmf->prealloc_pte) {
3131 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3132 if (unlikely(!pmd_none(*vmf->pmd))) {
3133 spin_unlock(vmf->ptl);
3137 mm_inc_nr_ptes(vma->vm_mm);
3138 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3139 spin_unlock(vmf->ptl);
3140 vmf->prealloc_pte = NULL;
3141 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3142 return VM_FAULT_OOM;
3146 * If a huge pmd materialized under us just retry later. Use
3147 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3148 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3149 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3150 * running immediately after a huge pmd fault in a different thread of
3151 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3152 * All we have to ensure is that it is a regular pmd that we can walk
3153 * with pte_offset_map() and we can do that through an atomic read in
3154 * C, which is what pmd_trans_unstable() provides.
3156 if (pmd_devmap_trans_unstable(vmf->pmd))
3157 return VM_FAULT_NOPAGE;
3160 * At this point we know that our vmf->pmd points to a page of ptes
3161 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3162 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3163 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3164 * be valid and we will re-check to make sure the vmf->pte isn't
3165 * pte_none() under vmf->ptl protection when we return to
3168 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3173 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3174 static void deposit_prealloc_pte(struct vm_fault *vmf)
3176 struct vm_area_struct *vma = vmf->vma;
3178 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3180 * We are going to consume the prealloc table,
3181 * count that as nr_ptes.
3183 mm_inc_nr_ptes(vma->vm_mm);
3184 vmf->prealloc_pte = NULL;
3187 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3189 struct vm_area_struct *vma = vmf->vma;
3190 bool write = vmf->flags & FAULT_FLAG_WRITE;
3191 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3196 if (!transhuge_vma_suitable(vma, haddr))
3197 return VM_FAULT_FALLBACK;
3199 ret = VM_FAULT_FALLBACK;
3200 page = compound_head(page);
3203 * Archs like ppc64 need additonal space to store information
3204 * related to pte entry. Use the preallocated table for that.
3206 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3207 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3208 if (!vmf->prealloc_pte)
3209 return VM_FAULT_OOM;
3210 smp_wmb(); /* See comment in __pte_alloc() */
3213 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3214 if (unlikely(!pmd_none(*vmf->pmd)))
3217 for (i = 0; i < HPAGE_PMD_NR; i++)
3218 flush_icache_page(vma, page + i);
3220 entry = mk_huge_pmd(page, vma->vm_page_prot);
3222 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3224 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3225 page_add_file_rmap(page, true);
3227 * deposit and withdraw with pmd lock held
3229 if (arch_needs_pgtable_deposit())
3230 deposit_prealloc_pte(vmf);
3232 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3234 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3236 /* fault is handled */
3238 count_vm_event(THP_FILE_MAPPED);
3240 spin_unlock(vmf->ptl);
3244 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3252 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3253 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3255 * @vmf: fault environment
3256 * @memcg: memcg to charge page (only for private mappings)
3257 * @page: page to map
3259 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3262 * Target users are page handler itself and implementations of
3263 * vm_ops->map_pages.
3265 * Return: %0 on success, %VM_FAULT_ code in case of error.
3267 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3270 struct vm_area_struct *vma = vmf->vma;
3271 bool write = vmf->flags & FAULT_FLAG_WRITE;
3275 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3276 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3278 VM_BUG_ON_PAGE(memcg, page);
3280 ret = do_set_pmd(vmf, page);
3281 if (ret != VM_FAULT_FALLBACK)
3286 ret = pte_alloc_one_map(vmf);
3291 /* Re-check under ptl */
3292 if (unlikely(!pte_none(*vmf->pte)))
3293 return VM_FAULT_NOPAGE;
3295 flush_icache_page(vma, page);
3296 entry = mk_pte(page, vma->vm_page_prot);
3298 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3299 /* copy-on-write page */
3300 if (write && !(vma->vm_flags & VM_SHARED)) {
3301 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3302 page_add_new_anon_rmap(page, vma, vmf->address, false);
3303 mem_cgroup_commit_charge(page, memcg, false, false);
3304 lru_cache_add_active_or_unevictable(page, vma);
3306 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3307 page_add_file_rmap(page, false);
3309 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3311 /* no need to invalidate: a not-present page won't be cached */
3312 update_mmu_cache(vma, vmf->address, vmf->pte);
3319 * finish_fault - finish page fault once we have prepared the page to fault
3321 * @vmf: structure describing the fault
3323 * This function handles all that is needed to finish a page fault once the
3324 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3325 * given page, adds reverse page mapping, handles memcg charges and LRU
3328 * The function expects the page to be locked and on success it consumes a
3329 * reference of a page being mapped (for the PTE which maps it).
3331 * Return: %0 on success, %VM_FAULT_ code in case of error.
3333 vm_fault_t finish_fault(struct vm_fault *vmf)
3338 /* Did we COW the page? */
3339 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3340 !(vmf->vma->vm_flags & VM_SHARED))
3341 page = vmf->cow_page;
3346 * check even for read faults because we might have lost our CoWed
3349 if (!(vmf->vma->vm_flags & VM_SHARED))
3350 ret = check_stable_address_space(vmf->vma->vm_mm);
3352 ret = alloc_set_pte(vmf, vmf->memcg, page);
3354 pte_unmap_unlock(vmf->pte, vmf->ptl);
3358 static unsigned long fault_around_bytes __read_mostly =
3359 rounddown_pow_of_two(65536);
3361 #ifdef CONFIG_DEBUG_FS
3362 static int fault_around_bytes_get(void *data, u64 *val)
3364 *val = fault_around_bytes;
3369 * fault_around_bytes must be rounded down to the nearest page order as it's
3370 * what do_fault_around() expects to see.
3372 static int fault_around_bytes_set(void *data, u64 val)
3374 if (val / PAGE_SIZE > PTRS_PER_PTE)
3376 if (val > PAGE_SIZE)
3377 fault_around_bytes = rounddown_pow_of_two(val);
3379 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3382 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3383 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3385 static int __init fault_around_debugfs(void)
3387 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3388 &fault_around_bytes_fops);
3391 late_initcall(fault_around_debugfs);
3395 * do_fault_around() tries to map few pages around the fault address. The hope
3396 * is that the pages will be needed soon and this will lower the number of
3399 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3400 * not ready to be mapped: not up-to-date, locked, etc.
3402 * This function is called with the page table lock taken. In the split ptlock
3403 * case the page table lock only protects only those entries which belong to
3404 * the page table corresponding to the fault address.
3406 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3409 * fault_around_bytes defines how many bytes we'll try to map.
3410 * do_fault_around() expects it to be set to a power of two less than or equal
3413 * The virtual address of the area that we map is naturally aligned to
3414 * fault_around_bytes rounded down to the machine page size
3415 * (and therefore to page order). This way it's easier to guarantee
3416 * that we don't cross page table boundaries.
3418 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3420 unsigned long address = vmf->address, nr_pages, mask;
3421 pgoff_t start_pgoff = vmf->pgoff;
3426 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3427 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3429 vmf->address = max(address & mask, vmf->vma->vm_start);
3430 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3434 * end_pgoff is either the end of the page table, the end of
3435 * the vma or nr_pages from start_pgoff, depending what is nearest.
3437 end_pgoff = start_pgoff -
3438 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3440 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3441 start_pgoff + nr_pages - 1);
3443 if (pmd_none(*vmf->pmd)) {
3444 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3445 if (!vmf->prealloc_pte)
3447 smp_wmb(); /* See comment in __pte_alloc() */
3450 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3452 /* Huge page is mapped? Page fault is solved */
3453 if (pmd_trans_huge(*vmf->pmd)) {
3454 ret = VM_FAULT_NOPAGE;
3458 /* ->map_pages() haven't done anything useful. Cold page cache? */
3462 /* check if the page fault is solved */
3463 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3464 if (!pte_none(*vmf->pte))
3465 ret = VM_FAULT_NOPAGE;
3466 pte_unmap_unlock(vmf->pte, vmf->ptl);
3468 vmf->address = address;
3473 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3475 struct vm_area_struct *vma = vmf->vma;
3479 * Let's call ->map_pages() first and use ->fault() as fallback
3480 * if page by the offset is not ready to be mapped (cold cache or
3483 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3484 ret = do_fault_around(vmf);
3489 ret = __do_fault(vmf);
3490 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3493 ret |= finish_fault(vmf);
3494 unlock_page(vmf->page);
3495 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3496 put_page(vmf->page);
3500 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3502 struct vm_area_struct *vma = vmf->vma;
3505 if (unlikely(anon_vma_prepare(vma)))
3506 return VM_FAULT_OOM;
3508 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3510 return VM_FAULT_OOM;
3512 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3513 &vmf->memcg, false)) {
3514 put_page(vmf->cow_page);
3515 return VM_FAULT_OOM;
3518 ret = __do_fault(vmf);
3519 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3521 if (ret & VM_FAULT_DONE_COW)
3524 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3525 __SetPageUptodate(vmf->cow_page);
3527 ret |= finish_fault(vmf);
3528 unlock_page(vmf->page);
3529 put_page(vmf->page);
3530 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3534 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3535 put_page(vmf->cow_page);
3539 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3541 struct vm_area_struct *vma = vmf->vma;
3542 vm_fault_t ret, tmp;
3544 ret = __do_fault(vmf);
3545 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3549 * Check if the backing address space wants to know that the page is
3550 * about to become writable
3552 if (vma->vm_ops->page_mkwrite) {
3553 unlock_page(vmf->page);
3554 tmp = do_page_mkwrite(vmf);
3555 if (unlikely(!tmp ||
3556 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3557 put_page(vmf->page);
3562 ret |= finish_fault(vmf);
3563 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3565 unlock_page(vmf->page);
3566 put_page(vmf->page);
3570 fault_dirty_shared_page(vma, vmf->page);
3575 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3576 * but allow concurrent faults).
3577 * The mmap_sem may have been released depending on flags and our
3578 * return value. See filemap_fault() and __lock_page_or_retry().
3579 * If mmap_sem is released, vma may become invalid (for example
3580 * by other thread calling munmap()).
3582 static vm_fault_t do_fault(struct vm_fault *vmf)
3584 struct vm_area_struct *vma = vmf->vma;
3585 struct mm_struct *vm_mm = vma->vm_mm;
3589 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3591 if (!vma->vm_ops->fault) {
3593 * If we find a migration pmd entry or a none pmd entry, which
3594 * should never happen, return SIGBUS
3596 if (unlikely(!pmd_present(*vmf->pmd)))
3597 ret = VM_FAULT_SIGBUS;
3599 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3604 * Make sure this is not a temporary clearing of pte
3605 * by holding ptl and checking again. A R/M/W update
3606 * of pte involves: take ptl, clearing the pte so that
3607 * we don't have concurrent modification by hardware
3608 * followed by an update.
3610 if (unlikely(pte_none(*vmf->pte)))
3611 ret = VM_FAULT_SIGBUS;
3613 ret = VM_FAULT_NOPAGE;
3615 pte_unmap_unlock(vmf->pte, vmf->ptl);
3617 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3618 ret = do_read_fault(vmf);
3619 else if (!(vma->vm_flags & VM_SHARED))
3620 ret = do_cow_fault(vmf);
3622 ret = do_shared_fault(vmf);
3624 /* preallocated pagetable is unused: free it */
3625 if (vmf->prealloc_pte) {
3626 pte_free(vm_mm, vmf->prealloc_pte);
3627 vmf->prealloc_pte = NULL;
3632 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3633 unsigned long addr, int page_nid,
3638 count_vm_numa_event(NUMA_HINT_FAULTS);
3639 if (page_nid == numa_node_id()) {
3640 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3641 *flags |= TNF_FAULT_LOCAL;
3644 return mpol_misplaced(page, vma, addr);
3647 static vm_fault_t do_numa_page(struct vm_fault *vmf)
3649 struct vm_area_struct *vma = vmf->vma;
3650 struct page *page = NULL;
3651 int page_nid = NUMA_NO_NODE;
3654 bool migrated = false;
3656 bool was_writable = pte_savedwrite(vmf->orig_pte);
3660 * The "pte" at this point cannot be used safely without
3661 * validation through pte_unmap_same(). It's of NUMA type but
3662 * the pfn may be screwed if the read is non atomic.
3664 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3665 spin_lock(vmf->ptl);
3666 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3667 pte_unmap_unlock(vmf->pte, vmf->ptl);
3672 * Make it present again, Depending on how arch implementes non
3673 * accessible ptes, some can allow access by kernel mode.
3675 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
3676 pte = pte_modify(old_pte, vma->vm_page_prot);
3677 pte = pte_mkyoung(pte);
3679 pte = pte_mkwrite(pte);
3680 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
3681 update_mmu_cache(vma, vmf->address, vmf->pte);
3683 page = vm_normal_page(vma, vmf->address, pte);
3685 pte_unmap_unlock(vmf->pte, vmf->ptl);
3689 /* TODO: handle PTE-mapped THP */
3690 if (PageCompound(page)) {
3691 pte_unmap_unlock(vmf->pte, vmf->ptl);
3696 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3697 * much anyway since they can be in shared cache state. This misses
3698 * the case where a mapping is writable but the process never writes
3699 * to it but pte_write gets cleared during protection updates and
3700 * pte_dirty has unpredictable behaviour between PTE scan updates,
3701 * background writeback, dirty balancing and application behaviour.
3703 if (!pte_write(pte))
3704 flags |= TNF_NO_GROUP;
3707 * Flag if the page is shared between multiple address spaces. This
3708 * is later used when determining whether to group tasks together
3710 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3711 flags |= TNF_SHARED;
3713 last_cpupid = page_cpupid_last(page);
3714 page_nid = page_to_nid(page);
3715 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3717 pte_unmap_unlock(vmf->pte, vmf->ptl);
3718 if (target_nid == NUMA_NO_NODE) {
3723 /* Migrate to the requested node */
3724 migrated = migrate_misplaced_page(page, vma, target_nid);
3726 page_nid = target_nid;
3727 flags |= TNF_MIGRATED;
3729 flags |= TNF_MIGRATE_FAIL;
3732 if (page_nid != NUMA_NO_NODE)
3733 task_numa_fault(last_cpupid, page_nid, 1, flags);
3737 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3739 if (vma_is_anonymous(vmf->vma))
3740 return do_huge_pmd_anonymous_page(vmf);
3741 if (vmf->vma->vm_ops->huge_fault)
3742 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3743 return VM_FAULT_FALLBACK;
3746 /* `inline' is required to avoid gcc 4.1.2 build error */
3747 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3749 if (vma_is_anonymous(vmf->vma))
3750 return do_huge_pmd_wp_page(vmf, orig_pmd);
3751 if (vmf->vma->vm_ops->huge_fault)
3752 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3754 /* COW handled on pte level: split pmd */
3755 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3756 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3758 return VM_FAULT_FALLBACK;
3761 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3763 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3766 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3768 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3769 /* No support for anonymous transparent PUD pages yet */
3770 if (vma_is_anonymous(vmf->vma))
3771 return VM_FAULT_FALLBACK;
3772 if (vmf->vma->vm_ops->huge_fault)
3773 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3774 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3775 return VM_FAULT_FALLBACK;
3778 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3780 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3781 /* No support for anonymous transparent PUD pages yet */
3782 if (vma_is_anonymous(vmf->vma))
3783 return VM_FAULT_FALLBACK;
3784 if (vmf->vma->vm_ops->huge_fault)
3785 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3786 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3787 return VM_FAULT_FALLBACK;
3791 * These routines also need to handle stuff like marking pages dirty
3792 * and/or accessed for architectures that don't do it in hardware (most
3793 * RISC architectures). The early dirtying is also good on the i386.
3795 * There is also a hook called "update_mmu_cache()" that architectures
3796 * with external mmu caches can use to update those (ie the Sparc or
3797 * PowerPC hashed page tables that act as extended TLBs).
3799 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3800 * concurrent faults).
3802 * The mmap_sem may have been released depending on flags and our return value.
3803 * See filemap_fault() and __lock_page_or_retry().
3805 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
3809 if (unlikely(pmd_none(*vmf->pmd))) {
3811 * Leave __pte_alloc() until later: because vm_ops->fault may
3812 * want to allocate huge page, and if we expose page table
3813 * for an instant, it will be difficult to retract from
3814 * concurrent faults and from rmap lookups.
3818 /* See comment in pte_alloc_one_map() */
3819 if (pmd_devmap_trans_unstable(vmf->pmd))
3822 * A regular pmd is established and it can't morph into a huge
3823 * pmd from under us anymore at this point because we hold the
3824 * mmap_sem read mode and khugepaged takes it in write mode.
3825 * So now it's safe to run pte_offset_map().
3827 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3828 vmf->orig_pte = *vmf->pte;
3831 * some architectures can have larger ptes than wordsize,
3832 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3833 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3834 * accesses. The code below just needs a consistent view
3835 * for the ifs and we later double check anyway with the
3836 * ptl lock held. So here a barrier will do.
3839 if (pte_none(vmf->orig_pte)) {
3840 pte_unmap(vmf->pte);
3846 if (vma_is_anonymous(vmf->vma))
3847 return do_anonymous_page(vmf);
3849 return do_fault(vmf);
3852 if (!pte_present(vmf->orig_pte))
3853 return do_swap_page(vmf);
3855 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3856 return do_numa_page(vmf);
3858 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3859 spin_lock(vmf->ptl);
3860 entry = vmf->orig_pte;
3861 if (unlikely(!pte_same(*vmf->pte, entry)))
3863 if (vmf->flags & FAULT_FLAG_WRITE) {
3864 if (!pte_write(entry))
3865 return do_wp_page(vmf);
3866 entry = pte_mkdirty(entry);
3868 entry = pte_mkyoung(entry);
3869 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3870 vmf->flags & FAULT_FLAG_WRITE)) {
3871 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3874 * This is needed only for protection faults but the arch code
3875 * is not yet telling us if this is a protection fault or not.
3876 * This still avoids useless tlb flushes for .text page faults
3879 if (vmf->flags & FAULT_FLAG_WRITE)
3880 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3883 pte_unmap_unlock(vmf->pte, vmf->ptl);
3888 * By the time we get here, we already hold the mm semaphore
3890 * The mmap_sem may have been released depending on flags and our
3891 * return value. See filemap_fault() and __lock_page_or_retry().
3893 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
3894 unsigned long address, unsigned int flags)
3896 struct vm_fault vmf = {
3898 .address = address & PAGE_MASK,
3900 .pgoff = linear_page_index(vma, address),
3901 .gfp_mask = __get_fault_gfp_mask(vma),
3903 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3904 struct mm_struct *mm = vma->vm_mm;
3909 pgd = pgd_offset(mm, address);
3910 p4d = p4d_alloc(mm, pgd, address);
3912 return VM_FAULT_OOM;
3914 vmf.pud = pud_alloc(mm, p4d, address);
3916 return VM_FAULT_OOM;
3917 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
3918 ret = create_huge_pud(&vmf);
3919 if (!(ret & VM_FAULT_FALLBACK))
3922 pud_t orig_pud = *vmf.pud;
3925 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3927 /* NUMA case for anonymous PUDs would go here */
3929 if (dirty && !pud_write(orig_pud)) {
3930 ret = wp_huge_pud(&vmf, orig_pud);
3931 if (!(ret & VM_FAULT_FALLBACK))
3934 huge_pud_set_accessed(&vmf, orig_pud);
3940 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3942 return VM_FAULT_OOM;
3943 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
3944 ret = create_huge_pmd(&vmf);
3945 if (!(ret & VM_FAULT_FALLBACK))
3948 pmd_t orig_pmd = *vmf.pmd;
3951 if (unlikely(is_swap_pmd(orig_pmd))) {
3952 VM_BUG_ON(thp_migration_supported() &&
3953 !is_pmd_migration_entry(orig_pmd));
3954 if (is_pmd_migration_entry(orig_pmd))
3955 pmd_migration_entry_wait(mm, vmf.pmd);
3958 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3959 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3960 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3962 if (dirty && !pmd_write(orig_pmd)) {
3963 ret = wp_huge_pmd(&vmf, orig_pmd);
3964 if (!(ret & VM_FAULT_FALLBACK))
3967 huge_pmd_set_accessed(&vmf, orig_pmd);
3973 return handle_pte_fault(&vmf);
3977 * By the time we get here, we already hold the mm semaphore
3979 * The mmap_sem may have been released depending on flags and our
3980 * return value. See filemap_fault() and __lock_page_or_retry().
3982 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3987 __set_current_state(TASK_RUNNING);
3989 count_vm_event(PGFAULT);
3990 count_memcg_event_mm(vma->vm_mm, PGFAULT);
3992 /* do counter updates before entering really critical section. */
3993 check_sync_rss_stat(current);
3995 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3996 flags & FAULT_FLAG_INSTRUCTION,
3997 flags & FAULT_FLAG_REMOTE))
3998 return VM_FAULT_SIGSEGV;
4001 * Enable the memcg OOM handling for faults triggered in user
4002 * space. Kernel faults are handled more gracefully.
4004 if (flags & FAULT_FLAG_USER)
4005 mem_cgroup_enter_user_fault();
4007 if (unlikely(is_vm_hugetlb_page(vma)))
4008 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4010 ret = __handle_mm_fault(vma, address, flags);
4012 if (flags & FAULT_FLAG_USER) {
4013 mem_cgroup_exit_user_fault();
4015 * The task may have entered a memcg OOM situation but
4016 * if the allocation error was handled gracefully (no
4017 * VM_FAULT_OOM), there is no need to kill anything.
4018 * Just clean up the OOM state peacefully.
4020 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4021 mem_cgroup_oom_synchronize(false);
4026 EXPORT_SYMBOL_GPL(handle_mm_fault);
4028 #ifndef __PAGETABLE_P4D_FOLDED
4030 * Allocate p4d page table.
4031 * We've already handled the fast-path in-line.
4033 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4035 p4d_t *new = p4d_alloc_one(mm, address);
4039 smp_wmb(); /* See comment in __pte_alloc */
4041 spin_lock(&mm->page_table_lock);
4042 if (pgd_present(*pgd)) /* Another has populated it */
4045 pgd_populate(mm, pgd, new);
4046 spin_unlock(&mm->page_table_lock);
4049 #endif /* __PAGETABLE_P4D_FOLDED */
4051 #ifndef __PAGETABLE_PUD_FOLDED
4053 * Allocate page upper directory.
4054 * We've already handled the fast-path in-line.
4056 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4058 pud_t *new = pud_alloc_one(mm, address);
4062 smp_wmb(); /* See comment in __pte_alloc */
4064 spin_lock(&mm->page_table_lock);
4065 #ifndef __ARCH_HAS_5LEVEL_HACK
4066 if (!p4d_present(*p4d)) {
4068 p4d_populate(mm, p4d, new);
4069 } else /* Another has populated it */
4072 if (!pgd_present(*p4d)) {
4074 pgd_populate(mm, p4d, new);
4075 } else /* Another has populated it */
4077 #endif /* __ARCH_HAS_5LEVEL_HACK */
4078 spin_unlock(&mm->page_table_lock);
4081 #endif /* __PAGETABLE_PUD_FOLDED */
4083 #ifndef __PAGETABLE_PMD_FOLDED
4085 * Allocate page middle directory.
4086 * We've already handled the fast-path in-line.
4088 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4091 pmd_t *new = pmd_alloc_one(mm, address);
4095 smp_wmb(); /* See comment in __pte_alloc */
4097 ptl = pud_lock(mm, pud);
4098 #ifndef __ARCH_HAS_4LEVEL_HACK
4099 if (!pud_present(*pud)) {
4101 pud_populate(mm, pud, new);
4102 } else /* Another has populated it */
4105 if (!pgd_present(*pud)) {
4107 pgd_populate(mm, pud, new);
4108 } else /* Another has populated it */
4110 #endif /* __ARCH_HAS_4LEVEL_HACK */
4114 #endif /* __PAGETABLE_PMD_FOLDED */
4116 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4117 struct mmu_notifier_range *range,
4118 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4126 pgd = pgd_offset(mm, address);
4127 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4130 p4d = p4d_offset(pgd, address);
4131 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4134 pud = pud_offset(p4d, address);
4135 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4138 pmd = pmd_offset(pud, address);
4139 VM_BUG_ON(pmd_trans_huge(*pmd));
4141 if (pmd_huge(*pmd)) {
4146 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4147 NULL, mm, address & PMD_MASK,
4148 (address & PMD_MASK) + PMD_SIZE);
4149 mmu_notifier_invalidate_range_start(range);
4151 *ptlp = pmd_lock(mm, pmd);
4152 if (pmd_huge(*pmd)) {
4158 mmu_notifier_invalidate_range_end(range);
4161 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4165 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4166 address & PAGE_MASK,
4167 (address & PAGE_MASK) + PAGE_SIZE);
4168 mmu_notifier_invalidate_range_start(range);
4170 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4171 if (!pte_present(*ptep))
4176 pte_unmap_unlock(ptep, *ptlp);
4178 mmu_notifier_invalidate_range_end(range);
4183 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4184 pte_t **ptepp, spinlock_t **ptlp)
4188 /* (void) is needed to make gcc happy */
4189 (void) __cond_lock(*ptlp,
4190 !(res = __follow_pte_pmd(mm, address, NULL,
4191 ptepp, NULL, ptlp)));
4195 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4196 struct mmu_notifier_range *range,
4197 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4201 /* (void) is needed to make gcc happy */
4202 (void) __cond_lock(*ptlp,
4203 !(res = __follow_pte_pmd(mm, address, range,
4204 ptepp, pmdpp, ptlp)));
4207 EXPORT_SYMBOL(follow_pte_pmd);
4210 * follow_pfn - look up PFN at a user virtual address
4211 * @vma: memory mapping
4212 * @address: user virtual address
4213 * @pfn: location to store found PFN
4215 * Only IO mappings and raw PFN mappings are allowed.
4217 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4219 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4226 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4229 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4232 *pfn = pte_pfn(*ptep);
4233 pte_unmap_unlock(ptep, ptl);
4236 EXPORT_SYMBOL(follow_pfn);
4238 #ifdef CONFIG_HAVE_IOREMAP_PROT
4239 int follow_phys(struct vm_area_struct *vma,
4240 unsigned long address, unsigned int flags,
4241 unsigned long *prot, resource_size_t *phys)
4247 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4250 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4254 if ((flags & FOLL_WRITE) && !pte_write(pte))
4257 *prot = pgprot_val(pte_pgprot(pte));
4258 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4262 pte_unmap_unlock(ptep, ptl);
4267 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4268 void *buf, int len, int write)
4270 resource_size_t phys_addr;
4271 unsigned long prot = 0;
4272 void __iomem *maddr;
4273 int offset = addr & (PAGE_SIZE-1);
4275 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4278 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4283 memcpy_toio(maddr + offset, buf, len);
4285 memcpy_fromio(buf, maddr + offset, len);
4290 EXPORT_SYMBOL_GPL(generic_access_phys);
4294 * Access another process' address space as given in mm. If non-NULL, use the
4295 * given task for page fault accounting.
4297 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4298 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4300 struct vm_area_struct *vma;
4301 void *old_buf = buf;
4302 int write = gup_flags & FOLL_WRITE;
4304 if (down_read_killable(&mm->mmap_sem))
4307 /* ignore errors, just check how much was successfully transferred */
4309 int bytes, ret, offset;
4311 struct page *page = NULL;
4313 ret = get_user_pages_remote(tsk, mm, addr, 1,
4314 gup_flags, &page, &vma, NULL);
4316 #ifndef CONFIG_HAVE_IOREMAP_PROT
4320 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4321 * we can access using slightly different code.
4323 vma = find_vma(mm, addr);
4324 if (!vma || vma->vm_start > addr)
4326 if (vma->vm_ops && vma->vm_ops->access)
4327 ret = vma->vm_ops->access(vma, addr, buf,
4335 offset = addr & (PAGE_SIZE-1);
4336 if (bytes > PAGE_SIZE-offset)
4337 bytes = PAGE_SIZE-offset;
4341 copy_to_user_page(vma, page, addr,
4342 maddr + offset, buf, bytes);
4343 set_page_dirty_lock(page);
4345 copy_from_user_page(vma, page, addr,
4346 buf, maddr + offset, bytes);
4355 up_read(&mm->mmap_sem);
4357 return buf - old_buf;
4361 * access_remote_vm - access another process' address space
4362 * @mm: the mm_struct of the target address space
4363 * @addr: start address to access
4364 * @buf: source or destination buffer
4365 * @len: number of bytes to transfer
4366 * @gup_flags: flags modifying lookup behaviour
4368 * The caller must hold a reference on @mm.
4370 * Return: number of bytes copied from source to destination.
4372 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4373 void *buf, int len, unsigned int gup_flags)
4375 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4379 * Access another process' address space.
4380 * Source/target buffer must be kernel space,
4381 * Do not walk the page table directly, use get_user_pages
4383 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4384 void *buf, int len, unsigned int gup_flags)
4386 struct mm_struct *mm;
4389 mm = get_task_mm(tsk);
4393 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4399 EXPORT_SYMBOL_GPL(access_process_vm);
4402 * Print the name of a VMA.
4404 void print_vma_addr(char *prefix, unsigned long ip)
4406 struct mm_struct *mm = current->mm;
4407 struct vm_area_struct *vma;
4410 * we might be running from an atomic context so we cannot sleep
4412 if (!down_read_trylock(&mm->mmap_sem))
4415 vma = find_vma(mm, ip);
4416 if (vma && vma->vm_file) {
4417 struct file *f = vma->vm_file;
4418 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4422 p = file_path(f, buf, PAGE_SIZE);
4425 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4427 vma->vm_end - vma->vm_start);
4428 free_page((unsigned long)buf);
4431 up_read(&mm->mmap_sem);
4434 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4435 void __might_fault(const char *file, int line)
4438 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4439 * holding the mmap_sem, this is safe because kernel memory doesn't
4440 * get paged out, therefore we'll never actually fault, and the
4441 * below annotations will generate false positives.
4443 if (uaccess_kernel())
4445 if (pagefault_disabled())
4447 __might_sleep(file, line, 0);
4448 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4450 might_lock_read(¤t->mm->mmap_sem);
4453 EXPORT_SYMBOL(__might_fault);
4456 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4458 * Process all subpages of the specified huge page with the specified
4459 * operation. The target subpage will be processed last to keep its
4462 static inline void process_huge_page(
4463 unsigned long addr_hint, unsigned int pages_per_huge_page,
4464 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4468 unsigned long addr = addr_hint &
4469 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4471 /* Process target subpage last to keep its cache lines hot */
4473 n = (addr_hint - addr) / PAGE_SIZE;
4474 if (2 * n <= pages_per_huge_page) {
4475 /* If target subpage in first half of huge page */
4478 /* Process subpages at the end of huge page */
4479 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4481 process_subpage(addr + i * PAGE_SIZE, i, arg);
4484 /* If target subpage in second half of huge page */
4485 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4486 l = pages_per_huge_page - n;
4487 /* Process subpages at the begin of huge page */
4488 for (i = 0; i < base; i++) {
4490 process_subpage(addr + i * PAGE_SIZE, i, arg);
4494 * Process remaining subpages in left-right-left-right pattern
4495 * towards the target subpage
4497 for (i = 0; i < l; i++) {
4498 int left_idx = base + i;
4499 int right_idx = base + 2 * l - 1 - i;
4502 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4504 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4508 static void clear_gigantic_page(struct page *page,
4510 unsigned int pages_per_huge_page)
4513 struct page *p = page;
4516 for (i = 0; i < pages_per_huge_page;
4517 i++, p = mem_map_next(p, page, i)) {
4519 clear_user_highpage(p, addr + i * PAGE_SIZE);
4523 static void clear_subpage(unsigned long addr, int idx, void *arg)
4525 struct page *page = arg;
4527 clear_user_highpage(page + idx, addr);
4530 void clear_huge_page(struct page *page,
4531 unsigned long addr_hint, unsigned int pages_per_huge_page)
4533 unsigned long addr = addr_hint &
4534 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4536 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4537 clear_gigantic_page(page, addr, pages_per_huge_page);
4541 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4544 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4546 struct vm_area_struct *vma,
4547 unsigned int pages_per_huge_page)
4550 struct page *dst_base = dst;
4551 struct page *src_base = src;
4553 for (i = 0; i < pages_per_huge_page; ) {
4555 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4558 dst = mem_map_next(dst, dst_base, i);
4559 src = mem_map_next(src, src_base, i);
4563 struct copy_subpage_arg {
4566 struct vm_area_struct *vma;
4569 static void copy_subpage(unsigned long addr, int idx, void *arg)
4571 struct copy_subpage_arg *copy_arg = arg;
4573 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4574 addr, copy_arg->vma);
4577 void copy_user_huge_page(struct page *dst, struct page *src,
4578 unsigned long addr_hint, struct vm_area_struct *vma,
4579 unsigned int pages_per_huge_page)
4581 unsigned long addr = addr_hint &
4582 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4583 struct copy_subpage_arg arg = {
4589 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4590 copy_user_gigantic_page(dst, src, addr, vma,
4591 pages_per_huge_page);
4595 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4598 long copy_huge_page_from_user(struct page *dst_page,
4599 const void __user *usr_src,
4600 unsigned int pages_per_huge_page,
4601 bool allow_pagefault)
4603 void *src = (void *)usr_src;
4605 unsigned long i, rc = 0;
4606 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4608 for (i = 0; i < pages_per_huge_page; i++) {
4609 if (allow_pagefault)
4610 page_kaddr = kmap(dst_page + i);
4612 page_kaddr = kmap_atomic(dst_page + i);
4613 rc = copy_from_user(page_kaddr,
4614 (const void __user *)(src + i * PAGE_SIZE),
4616 if (allow_pagefault)
4617 kunmap(dst_page + i);
4619 kunmap_atomic(page_kaddr);
4621 ret_val -= (PAGE_SIZE - rc);
4629 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4631 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4633 static struct kmem_cache *page_ptl_cachep;
4635 void __init ptlock_cache_init(void)
4637 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4641 bool ptlock_alloc(struct page *page)
4645 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4652 void ptlock_free(struct page *page)
4654 kmem_cache_free(page_ptl_cachep, page->ptl);