1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
73 #include <linux/perf_event.h>
74 #include <linux/ptrace.h>
75 #include <linux/vmalloc.h>
77 #include <trace/events/kmem.h>
80 #include <asm/mmu_context.h>
81 #include <asm/pgalloc.h>
82 #include <linux/uaccess.h>
84 #include <asm/tlbflush.h>
85 #ifdef CONFIG_FINEGRAINED_THP
86 #include <asm/huge_mm.h>
87 #include <asm/finegrained_thp.h>
89 #include <asm-generic/huge_mm.h>
90 #include <asm-generic/finegrained_thp.h>
93 #include "pgalloc-track.h"
96 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
97 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
100 #ifndef CONFIG_NEED_MULTIPLE_NODES
101 /* use the per-pgdat data instead for discontigmem - mbligh */
102 unsigned long max_mapnr;
103 EXPORT_SYMBOL(max_mapnr);
105 struct page *mem_map;
106 EXPORT_SYMBOL(mem_map);
110 * A number of key systems in x86 including ioremap() rely on the assumption
111 * that high_memory defines the upper bound on direct map memory, then end
112 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
113 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
117 EXPORT_SYMBOL(high_memory);
120 * Randomize the address space (stacks, mmaps, brk, etc.).
122 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
123 * as ancient (libc5 based) binaries can segfault. )
125 int randomize_va_space __read_mostly =
126 #ifdef CONFIG_COMPAT_BRK
132 #ifndef arch_faults_on_old_pte
133 static inline bool arch_faults_on_old_pte(void)
136 * Those arches which don't have hw access flag feature need to
137 * implement their own helper. By default, "true" means pagefault
138 * will be hit on old pte.
144 static int __init disable_randmaps(char *s)
146 randomize_va_space = 0;
149 __setup("norandmaps", disable_randmaps);
151 unsigned long zero_pfn __read_mostly;
152 EXPORT_SYMBOL(zero_pfn);
154 unsigned long highest_memmap_pfn __read_mostly;
156 atomic_long_t nr_phys_cont_pte_pages;
157 atomic_long_t nr_phys_huge_pmd_pages;
159 unsigned long phys_cont_pte_pages(void)
161 return atomic_long_read(&nr_phys_cont_pte_pages);
164 unsigned long phys_huge_pmd_pages(void)
166 return atomic_long_read(&nr_phys_huge_pmd_pages);
170 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
172 static int __init init_zero_pfn(void)
174 zero_pfn = page_to_pfn(ZERO_PAGE(0));
177 early_initcall(init_zero_pfn);
179 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
181 trace_rss_stat(mm, member, count);
184 #if defined(SPLIT_RSS_COUNTING)
186 void sync_mm_rss(struct mm_struct *mm)
190 for (i = 0; i < NR_MM_COUNTERS; i++) {
191 if (current->rss_stat.count[i]) {
192 add_mm_counter(mm, i, current->rss_stat.count[i]);
193 current->rss_stat.count[i] = 0;
196 current->rss_stat.events = 0;
199 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
201 struct task_struct *task = current;
203 if (likely(task->mm == mm))
204 task->rss_stat.count[member] += val;
206 add_mm_counter(mm, member, val);
208 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
209 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
211 /* sync counter once per 64 page faults */
212 #define TASK_RSS_EVENTS_THRESH (64)
213 static void check_sync_rss_stat(struct task_struct *task)
215 if (unlikely(task != current))
217 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
218 sync_mm_rss(task->mm);
220 #else /* SPLIT_RSS_COUNTING */
222 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
223 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
225 static void check_sync_rss_stat(struct task_struct *task)
229 #endif /* SPLIT_RSS_COUNTING */
231 #ifdef CONFIG_FINEGRAINED_THP
232 void thp_print_cont_pte_table(struct mm_struct *mm,
233 unsigned long addr, pte_t *ptep, unsigned long line);
234 #endif /* CONFIG_FINEGRAINED_THP */
237 * Note: this doesn't free the actual pages themselves. That
238 * has been handled earlier when unmapping all the memory regions.
240 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
243 pgtable_t token = pmd_pgtable(*pmd);
245 pte_free_tlb(tlb, token, addr);
246 mm_dec_nr_ptes(tlb->mm);
249 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
250 unsigned long addr, unsigned long end,
251 unsigned long floor, unsigned long ceiling)
258 pmd = pmd_offset(pud, addr);
260 next = pmd_addr_end(addr, end);
261 if (pmd_none_or_clear_bad(pmd))
263 free_pte_range(tlb, pmd, addr);
264 } while (pmd++, addr = next, addr != end);
274 if (end - 1 > ceiling - 1)
277 pmd = pmd_offset(pud, start);
279 pmd_free_tlb(tlb, pmd, start);
280 mm_dec_nr_pmds(tlb->mm);
283 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
284 unsigned long addr, unsigned long end,
285 unsigned long floor, unsigned long ceiling)
292 pud = pud_offset(p4d, addr);
294 next = pud_addr_end(addr, end);
295 if (pud_none_or_clear_bad(pud))
297 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
298 } while (pud++, addr = next, addr != end);
308 if (end - 1 > ceiling - 1)
311 pud = pud_offset(p4d, start);
313 pud_free_tlb(tlb, pud, start);
314 mm_dec_nr_puds(tlb->mm);
317 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
318 unsigned long addr, unsigned long end,
319 unsigned long floor, unsigned long ceiling)
326 p4d = p4d_offset(pgd, addr);
328 next = p4d_addr_end(addr, end);
329 if (p4d_none_or_clear_bad(p4d))
331 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
332 } while (p4d++, addr = next, addr != end);
338 ceiling &= PGDIR_MASK;
342 if (end - 1 > ceiling - 1)
345 p4d = p4d_offset(pgd, start);
347 p4d_free_tlb(tlb, p4d, start);
351 * This function frees user-level page tables of a process.
353 void free_pgd_range(struct mmu_gather *tlb,
354 unsigned long addr, unsigned long end,
355 unsigned long floor, unsigned long ceiling)
361 * The next few lines have given us lots of grief...
363 * Why are we testing PMD* at this top level? Because often
364 * there will be no work to do at all, and we'd prefer not to
365 * go all the way down to the bottom just to discover that.
367 * Why all these "- 1"s? Because 0 represents both the bottom
368 * of the address space and the top of it (using -1 for the
369 * top wouldn't help much: the masks would do the wrong thing).
370 * The rule is that addr 0 and floor 0 refer to the bottom of
371 * the address space, but end 0 and ceiling 0 refer to the top
372 * Comparisons need to use "end - 1" and "ceiling - 1" (though
373 * that end 0 case should be mythical).
375 * Wherever addr is brought up or ceiling brought down, we must
376 * be careful to reject "the opposite 0" before it confuses the
377 * subsequent tests. But what about where end is brought down
378 * by PMD_SIZE below? no, end can't go down to 0 there.
380 * Whereas we round start (addr) and ceiling down, by different
381 * masks at different levels, in order to test whether a table
382 * now has no other vmas using it, so can be freed, we don't
383 * bother to round floor or end up - the tests don't need that.
397 if (end - 1 > ceiling - 1)
402 * We add page table cache pages with PAGE_SIZE,
403 * (see pte_free_tlb()), flush the tlb if we need
405 tlb_change_page_size(tlb, PAGE_SIZE);
406 pgd = pgd_offset(tlb->mm, addr);
408 next = pgd_addr_end(addr, end);
409 if (pgd_none_or_clear_bad(pgd))
411 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
412 } while (pgd++, addr = next, addr != end);
415 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
416 unsigned long floor, unsigned long ceiling)
419 struct vm_area_struct *next = vma->vm_next;
420 unsigned long addr = vma->vm_start;
423 * Hide vma from rmap and truncate_pagecache before freeing
426 unlink_anon_vmas(vma);
427 unlink_file_vma(vma);
429 if (is_vm_hugetlb_page(vma)) {
430 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
431 floor, next ? next->vm_start : ceiling);
434 * Optimization: gather nearby vmas into one call down
436 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
437 && !is_vm_hugetlb_page(next)) {
440 unlink_anon_vmas(vma);
441 unlink_file_vma(vma);
443 free_pgd_range(tlb, addr, vma->vm_end,
444 floor, next ? next->vm_start : ceiling);
450 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
453 pgtable_t new = pte_alloc_one(mm);
458 * Ensure all pte setup (eg. pte page lock and page clearing) are
459 * visible before the pte is made visible to other CPUs by being
460 * put into page tables.
462 * The other side of the story is the pointer chasing in the page
463 * table walking code (when walking the page table without locking;
464 * ie. most of the time). Fortunately, these data accesses consist
465 * of a chain of data-dependent loads, meaning most CPUs (alpha
466 * being the notable exception) will already guarantee loads are
467 * seen in-order. See the alpha page table accessors for the
468 * smp_rmb() barriers in page table walking code.
470 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
472 ptl = pmd_lock(mm, pmd);
473 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
475 pmd_populate(mm, pmd, new);
484 int __pte_alloc_kernel(pmd_t *pmd)
486 pte_t *new = pte_alloc_one_kernel(&init_mm);
490 smp_wmb(); /* See comment in __pte_alloc */
492 spin_lock(&init_mm.page_table_lock);
493 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
494 pmd_populate_kernel(&init_mm, pmd, new);
497 spin_unlock(&init_mm.page_table_lock);
499 pte_free_kernel(&init_mm, new);
503 static inline void init_rss_vec(int *rss)
505 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
508 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
512 if (current->mm == mm)
514 for (i = 0; i < NR_MM_COUNTERS; i++)
516 add_mm_counter(mm, i, rss[i]);
520 * This function is called to print an error when a bad pte
521 * is found. For example, we might have a PFN-mapped pte in
522 * a region that doesn't allow it.
524 * The calling function must still handle the error.
526 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
527 pte_t pte, struct page *page)
529 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
530 p4d_t *p4d = p4d_offset(pgd, addr);
531 pud_t *pud = pud_offset(p4d, addr);
532 pmd_t *pmd = pmd_offset(pud, addr);
533 struct address_space *mapping;
535 static unsigned long resume;
536 static unsigned long nr_shown;
537 static unsigned long nr_unshown;
540 * Allow a burst of 60 reports, then keep quiet for that minute;
541 * or allow a steady drip of one report per second.
543 if (nr_shown == 60) {
544 if (time_before(jiffies, resume)) {
549 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
556 resume = jiffies + 60 * HZ;
558 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
559 index = linear_page_index(vma, addr);
561 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
563 (long long)pte_val(pte), (long long)pmd_val(*pmd));
565 dump_page(page, "bad pte");
566 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
567 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
568 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
570 vma->vm_ops ? vma->vm_ops->fault : NULL,
571 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
572 mapping ? mapping->a_ops->readpage : NULL);
574 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
578 * vm_normal_page -- This function gets the "struct page" associated with a pte.
580 * "Special" mappings do not wish to be associated with a "struct page" (either
581 * it doesn't exist, or it exists but they don't want to touch it). In this
582 * case, NULL is returned here. "Normal" mappings do have a struct page.
584 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
585 * pte bit, in which case this function is trivial. Secondly, an architecture
586 * may not have a spare pte bit, which requires a more complicated scheme,
589 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
590 * special mapping (even if there are underlying and valid "struct pages").
591 * COWed pages of a VM_PFNMAP are always normal.
593 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
594 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
595 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
596 * mapping will always honor the rule
598 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
600 * And for normal mappings this is false.
602 * This restricts such mappings to be a linear translation from virtual address
603 * to pfn. To get around this restriction, we allow arbitrary mappings so long
604 * as the vma is not a COW mapping; in that case, we know that all ptes are
605 * special (because none can have been COWed).
608 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
610 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
611 * page" backing, however the difference is that _all_ pages with a struct
612 * page (that is, those where pfn_valid is true) are refcounted and considered
613 * normal pages by the VM. The disadvantage is that pages are refcounted
614 * (which can be slower and simply not an option for some PFNMAP users). The
615 * advantage is that we don't have to follow the strict linearity rule of
616 * PFNMAP mappings in order to support COWable mappings.
619 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
622 unsigned long pfn = pte_pfn(pte);
624 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
625 if (likely(!pte_special(pte)))
627 if (vma->vm_ops && vma->vm_ops->find_special_page)
628 return vma->vm_ops->find_special_page(vma, addr);
629 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
631 if (is_zero_pfn(pfn))
636 print_bad_pte(vma, addr, pte, NULL);
640 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
642 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
643 if (vma->vm_flags & VM_MIXEDMAP) {
649 off = (addr - vma->vm_start) >> PAGE_SHIFT;
650 if (pfn == vma->vm_pgoff + off)
652 if (!is_cow_mapping(vma->vm_flags))
657 if (is_zero_pfn(pfn))
661 if (unlikely(pfn > highest_memmap_pfn)) {
662 print_bad_pte(vma, addr, pte, NULL);
667 * NOTE! We still have PageReserved() pages in the page tables.
668 * eg. VDSO mappings can cause them to exist.
671 return pfn_to_page(pfn);
674 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
675 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
678 unsigned long pfn = pmd_pfn(pmd);
681 * There is no pmd_special() but there may be special pmds, e.g.
682 * in a direct-access (dax) mapping, so let's just replicate the
683 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
685 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
686 if (vma->vm_flags & VM_MIXEDMAP) {
692 off = (addr - vma->vm_start) >> PAGE_SHIFT;
693 if (pfn == vma->vm_pgoff + off)
695 if (!is_cow_mapping(vma->vm_flags))
702 if (is_huge_zero_pmd(pmd))
704 if (unlikely(pfn > highest_memmap_pfn))
708 * NOTE! We still have PageReserved() pages in the page tables.
709 * eg. VDSO mappings can cause them to exist.
712 return pfn_to_page(pfn);
717 * copy one vm_area from one task to the other. Assumes the page tables
718 * already present in the new task to be cleared in the whole range
719 * covered by this vma.
723 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
724 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
725 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
727 unsigned long vm_flags = dst_vma->vm_flags;
728 pte_t pte = *src_pte;
730 swp_entry_t entry = pte_to_swp_entry(pte);
732 if (likely(!non_swap_entry(entry))) {
733 if (swap_duplicate(entry) < 0)
736 /* make sure dst_mm is on swapoff's mmlist. */
737 if (unlikely(list_empty(&dst_mm->mmlist))) {
738 spin_lock(&mmlist_lock);
739 if (list_empty(&dst_mm->mmlist))
740 list_add(&dst_mm->mmlist,
742 spin_unlock(&mmlist_lock);
745 } else if (is_migration_entry(entry)) {
746 page = migration_entry_to_page(entry);
748 rss[mm_counter(page)]++;
750 if (is_write_migration_entry(entry) &&
751 is_cow_mapping(vm_flags)) {
753 * COW mappings require pages in both
754 * parent and child to be set to read.
756 make_migration_entry_read(&entry);
757 pte = swp_entry_to_pte(entry);
758 pte = arch_pte_clearhuge(pte);
759 if (pte_swp_soft_dirty(*src_pte))
760 pte = pte_swp_mksoft_dirty(pte);
761 if (pte_swp_uffd_wp(*src_pte))
762 pte = pte_swp_mkuffd_wp(pte);
763 set_pte_at(src_mm, addr, src_pte, pte);
765 } else if (is_device_private_entry(entry)) {
766 page = device_private_entry_to_page(entry);
769 * Update rss count even for unaddressable pages, as
770 * they should treated just like normal pages in this
773 * We will likely want to have some new rss counters
774 * for unaddressable pages, at some point. But for now
775 * keep things as they are.
778 rss[mm_counter(page)]++;
779 page_dup_rmap(page, false);
782 * We do not preserve soft-dirty information, because so
783 * far, checkpoint/restore is the only feature that
784 * requires that. And checkpoint/restore does not work
785 * when a device driver is involved (you cannot easily
786 * save and restore device driver state).
788 if (is_write_device_private_entry(entry) &&
789 is_cow_mapping(vm_flags)) {
790 make_device_private_entry_read(&entry);
791 pte = swp_entry_to_pte(entry);
792 pte = arch_pte_clearhuge(pte);
793 if (pte_swp_uffd_wp(*src_pte))
794 pte = pte_swp_mkuffd_wp(pte);
795 set_pte_at(src_mm, addr, src_pte, pte);
798 if (!userfaultfd_wp(dst_vma))
799 pte = pte_swp_clear_uffd_wp(pte);
800 pte = arch_pte_clearhuge(pte);
801 set_pte_at(dst_mm, addr, dst_pte, pte);
806 * Copy a present and normal page if necessary.
808 * NOTE! The usual case is that this doesn't need to do
809 * anything, and can just return a positive value. That
810 * will let the caller know that it can just increase
811 * the page refcount and re-use the pte the traditional
814 * But _if_ we need to copy it because it needs to be
815 * pinned in the parent (and the child should get its own
816 * copy rather than just a reference to the same page),
817 * we'll do that here and return zero to let the caller
820 * And if we need a pre-allocated page but don't yet have
821 * one, return a negative error to let the preallocation
822 * code know so that it can do so outside the page table
826 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
827 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
828 struct page **prealloc, pte_t pte, struct page *page)
830 struct mm_struct *src_mm = src_vma->vm_mm;
831 struct page *new_page;
833 if (!is_cow_mapping(src_vma->vm_flags))
837 * What we want to do is to check whether this page may
838 * have been pinned by the parent process. If so,
839 * instead of wrprotect the pte on both sides, we copy
840 * the page immediately so that we'll always guarantee
841 * the pinned page won't be randomly replaced in the
844 * The page pinning checks are just "has this mm ever
845 * seen pinning", along with the (inexact) check of
846 * the page count. That might give false positives for
847 * for pinning, but it will work correctly.
849 if (likely(!atomic_read(&src_mm->has_pinned)))
851 if (likely(!page_maybe_dma_pinned(page)))
854 new_page = *prealloc;
859 * We have a prealloc page, all good! Take it
860 * over and copy the page & arm it.
863 copy_user_highpage(new_page, page, addr, src_vma);
864 __SetPageUptodate(new_page);
865 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
866 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
867 rss[mm_counter(new_page)]++;
869 /* All done, just insert the new page copy in the child */
870 pte = mk_pte(new_page, dst_vma->vm_page_prot);
871 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
872 if (userfaultfd_pte_wp(dst_vma, *src_pte))
873 /* Uffd-wp needs to be delivered to dest pte as well */
874 pte = pte_wrprotect(pte_mkuffd_wp(pte));
875 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
880 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
881 * is required to copy this pte.
884 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
885 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
886 struct page **prealloc)
888 struct mm_struct *src_mm = src_vma->vm_mm;
889 unsigned long vm_flags = src_vma->vm_flags;
890 pte_t pte = *src_pte;
893 page = vm_normal_page(src_vma, addr, pte);
897 * when 64KB hugepage map is copied,
898 * clear contiguous bit
900 pte = arch_pte_clearhuge(pte);
902 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
903 addr, rss, prealloc, pte, page);
908 page_dup_rmap(page, false);
909 rss[mm_counter(page)]++;
913 * If it's a COW mapping, write protect it both
914 * in the parent and the child
916 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
917 ptep_set_wrprotect(src_mm, addr, src_pte);
918 pte = pte_wrprotect(pte);
922 * If it's a shared mapping, mark it clean in
925 if (vm_flags & VM_SHARED)
926 pte = pte_mkclean(pte);
927 pte = pte_mkold(pte);
929 if (!userfaultfd_wp(dst_vma))
930 pte = pte_clear_uffd_wp(pte);
932 pte = arch_pte_clearhuge(pte);
934 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
938 static inline struct page *
939 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
942 struct page *new_page;
944 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
948 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
952 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
958 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
959 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
962 struct mm_struct *dst_mm = dst_vma->vm_mm;
963 struct mm_struct *src_mm = src_vma->vm_mm;
964 pte_t *orig_src_pte, *orig_dst_pte;
965 pte_t *src_pte, *dst_pte;
966 spinlock_t *src_ptl, *dst_ptl;
967 int progress, ret = 0;
968 int rss[NR_MM_COUNTERS];
969 swp_entry_t entry = (swp_entry_t){0};
970 struct page *prealloc = NULL;
976 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
981 src_pte = pte_offset_map(src_pmd, addr);
982 src_ptl = pte_lockptr(src_mm, src_pmd);
983 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
984 orig_src_pte = src_pte;
985 orig_dst_pte = dst_pte;
986 arch_enter_lazy_mmu_mode();
990 * We are holding two locks at this point - either of them
991 * could generate latencies in another task on another CPU.
993 if (progress >= 32) {
995 if (need_resched() ||
996 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
999 if (pte_none(*src_pte)) {
1004 if (unlikely(!pte_present(*src_pte))) {
1005 entry.val = copy_nonpresent_pte(dst_mm, src_mm,
1015 /* copy_present_pte() will clear `*prealloc' if consumed */
1016 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1017 addr, rss, &prealloc);
1019 * If we need a pre-allocated page for this pte, drop the
1020 * locks, allocate, and try again.
1022 if (unlikely(ret == -EAGAIN))
1024 if (unlikely(prealloc)) {
1026 * pre-alloc page cannot be reused by next time so as
1027 * to strictly follow mempolicy (e.g., alloc_page_vma()
1028 * will allocate page according to address). This
1029 * could only happen if one pinned pte changed.
1035 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1037 arch_leave_lazy_mmu_mode();
1038 spin_unlock(src_ptl);
1039 pte_unmap(orig_src_pte);
1040 add_mm_rss_vec(dst_mm, rss);
1041 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1045 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1051 WARN_ON_ONCE(ret != -EAGAIN);
1052 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1055 /* We've captured and resolved the error. Reset, try again. */
1061 if (unlikely(prealloc))
1067 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1068 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1071 struct mm_struct *dst_mm = dst_vma->vm_mm;
1072 struct mm_struct *src_mm = src_vma->vm_mm;
1073 pmd_t *src_pmd, *dst_pmd;
1076 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1079 src_pmd = pmd_offset(src_pud, addr);
1081 next = pmd_addr_end(addr, end);
1082 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1083 || pmd_devmap(*src_pmd)) {
1085 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1086 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1087 addr, dst_vma, src_vma);
1094 if (pmd_none_or_clear_bad(src_pmd))
1096 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1099 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1104 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1105 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1108 struct mm_struct *dst_mm = dst_vma->vm_mm;
1109 struct mm_struct *src_mm = src_vma->vm_mm;
1110 pud_t *src_pud, *dst_pud;
1113 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1116 src_pud = pud_offset(src_p4d, addr);
1118 next = pud_addr_end(addr, end);
1119 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1122 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1123 err = copy_huge_pud(dst_mm, src_mm,
1124 dst_pud, src_pud, addr, src_vma);
1131 if (pud_none_or_clear_bad(src_pud))
1133 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1136 } while (dst_pud++, src_pud++, addr = next, addr != end);
1141 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1142 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1145 struct mm_struct *dst_mm = dst_vma->vm_mm;
1146 p4d_t *src_p4d, *dst_p4d;
1149 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1152 src_p4d = p4d_offset(src_pgd, addr);
1154 next = p4d_addr_end(addr, end);
1155 if (p4d_none_or_clear_bad(src_p4d))
1157 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1160 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1164 #ifdef CONFIG_FINEGRAINED_THP
1165 bool zap_cont_pte_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
1166 pmd_t *pmd, pte_t **ptep, unsigned long *addr,
1167 unsigned long end, struct page *page,
1168 int *rss, spinlock_t *ptl);
1169 #else /* CONFIG_FINEGRAINED_THP */
1170 bool zap_cont_pte_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
1171 pmd_t *pmd, pte_t **ptep, unsigned long *addr,
1172 unsigned long end, struct page *page,
1173 int *rss, spinlock_t *ptl)
1177 #endif /* CONFIG_FINEGRAINED_THP */
1180 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1182 pgd_t *src_pgd, *dst_pgd;
1184 unsigned long addr = src_vma->vm_start;
1185 unsigned long end = src_vma->vm_end;
1186 struct mm_struct *dst_mm = dst_vma->vm_mm;
1187 struct mm_struct *src_mm = src_vma->vm_mm;
1188 struct mmu_notifier_range range;
1193 * Don't copy ptes where a page fault will fill them correctly.
1194 * Fork becomes much lighter when there are big shared or private
1195 * readonly mappings. The tradeoff is that copy_page_range is more
1196 * efficient than faulting.
1198 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1202 if (is_vm_hugetlb_page(src_vma))
1203 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1205 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1207 * We do not free on error cases below as remove_vma
1208 * gets called on error from higher level routine
1210 ret = track_pfn_copy(src_vma);
1216 * We need to invalidate the secondary MMU mappings only when
1217 * there could be a permission downgrade on the ptes of the
1218 * parent mm. And a permission downgrade will only happen if
1219 * is_cow_mapping() returns true.
1221 is_cow = is_cow_mapping(src_vma->vm_flags);
1224 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1225 0, src_vma, src_mm, addr, end);
1226 mmu_notifier_invalidate_range_start(&range);
1228 * Disabling preemption is not needed for the write side, as
1229 * the read side doesn't spin, but goes to the mmap_lock.
1231 * Use the raw variant of the seqcount_t write API to avoid
1232 * lockdep complaining about preemptibility.
1234 mmap_assert_write_locked(src_mm);
1235 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1239 dst_pgd = pgd_offset(dst_mm, addr);
1240 src_pgd = pgd_offset(src_mm, addr);
1242 next = pgd_addr_end(addr, end);
1243 if (pgd_none_or_clear_bad(src_pgd))
1245 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1250 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1253 raw_write_seqcount_end(&src_mm->write_protect_seq);
1254 mmu_notifier_invalidate_range_end(&range);
1259 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1260 struct vm_area_struct *vma, pmd_t *pmd,
1261 unsigned long addr, unsigned long end,
1262 struct zap_details *details)
1264 struct mm_struct *mm = tlb->mm;
1265 int force_flush = 0;
1266 int rss[NR_MM_COUNTERS];
1272 tlb_change_page_size(tlb, PAGE_SIZE);
1275 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1277 flush_tlb_batched_pending(mm);
1278 arch_enter_lazy_mmu_mode();
1281 if (pte_none(ptent))
1287 if (pte_present(ptent)) {
1290 page = vm_normal_page(vma, addr, ptent);
1291 if (unlikely(details) && page) {
1293 * unmap_shared_mapping_pages() wants to
1294 * invalidate cache without truncating:
1295 * unmap shared but keep private pages.
1297 if (details->check_mapping &&
1298 details->check_mapping != page_rmapping(page))
1301 #ifdef CONFIG_FINEGRAINED_THP
1302 if (page && pte_cont(ptent) && PageTransHuge(compound_head(page))) {
1303 if (zap_cont_pte_range(tlb, vma, pmd, &pte,
1304 &addr, end, page, rss, ptl)) {
1308 } else if (pte_cont(ptent))
1309 atomic_long_dec(&nr_phys_cont_pte_pages);
1310 #endif /* CONFIG_FINEGRAINED_THP */
1311 ptent = ptep_get_and_clear_full(mm, addr, pte,
1313 tlb_remove_tlb_entry(tlb, pte, addr);
1314 if (unlikely(!page))
1317 if (!PageAnon(page)) {
1318 if (pte_dirty(ptent)) {
1320 set_page_dirty(page);
1322 if (pte_young(ptent) &&
1323 likely(!(vma->vm_flags & VM_SEQ_READ)))
1324 mark_page_accessed(page);
1326 rss[mm_counter(page)]--;
1327 page_remove_rmap(page, false);
1328 if (unlikely(page_mapcount(page) < 0))
1329 print_bad_pte(vma, addr, ptent, page);
1330 if (unlikely(__tlb_remove_page(tlb, page))) {
1338 entry = pte_to_swp_entry(ptent);
1339 if (is_device_private_entry(entry)) {
1340 struct page *page = device_private_entry_to_page(entry);
1342 if (unlikely(details && details->check_mapping)) {
1344 * unmap_shared_mapping_pages() wants to
1345 * invalidate cache without truncating:
1346 * unmap shared but keep private pages.
1348 if (details->check_mapping !=
1349 page_rmapping(page))
1353 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1354 rss[mm_counter(page)]--;
1355 page_remove_rmap(page, false);
1360 /* If details->check_mapping, we leave swap entries. */
1361 if (unlikely(details))
1364 if (!non_swap_entry(entry))
1366 else if (is_migration_entry(entry)) {
1369 page = migration_entry_to_page(entry);
1370 rss[mm_counter(page)]--;
1372 if (unlikely(!free_swap_and_cache(entry)))
1373 print_bad_pte(vma, addr, ptent, NULL);
1374 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1375 } while (pte++, addr += PAGE_SIZE, addr != end);
1377 add_mm_rss_vec(mm, rss);
1378 arch_leave_lazy_mmu_mode();
1380 /* Do the actual TLB flush before dropping ptl */
1382 tlb_flush_mmu_tlbonly(tlb);
1383 pte_unmap_unlock(start_pte, ptl);
1386 * If we forced a TLB flush (either due to running out of
1387 * batch buffers or because we needed to flush dirty TLB
1388 * entries before releasing the ptl), free the batched
1389 * memory too. Restart if we didn't do everything.
1404 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1405 struct vm_area_struct *vma, pud_t *pud,
1406 unsigned long addr, unsigned long end,
1407 struct zap_details *details)
1412 pmd = pmd_offset(pud, addr);
1414 next = pmd_addr_end(addr, end);
1415 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1416 if (next - addr != HPAGE_PMD_SIZE)
1417 __split_huge_pmd(vma, pmd, addr, false, NULL);
1418 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1421 } else if (details && details->single_page &&
1422 PageTransCompound(details->single_page) &&
1423 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1424 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1426 * Take and drop THP pmd lock so that we cannot return
1427 * prematurely, while zap_huge_pmd() has cleared *pmd,
1428 * but not yet decremented compound_mapcount().
1434 * Here there can be other concurrent MADV_DONTNEED or
1435 * trans huge page faults running, and if the pmd is
1436 * none or trans huge it can change under us. This is
1437 * because MADV_DONTNEED holds the mmap_lock in read
1440 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1442 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1445 } while (pmd++, addr = next, addr != end);
1450 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1451 struct vm_area_struct *vma, p4d_t *p4d,
1452 unsigned long addr, unsigned long end,
1453 struct zap_details *details)
1458 pud = pud_offset(p4d, addr);
1460 next = pud_addr_end(addr, end);
1461 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1462 if (next - addr != HPAGE_PUD_SIZE) {
1463 mmap_assert_locked(tlb->mm);
1464 split_huge_pud(vma, pud, addr);
1465 } else if (zap_huge_pud(tlb, vma, pud, addr))
1469 if (pud_none_or_clear_bad(pud))
1471 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1474 } while (pud++, addr = next, addr != end);
1479 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1480 struct vm_area_struct *vma, pgd_t *pgd,
1481 unsigned long addr, unsigned long end,
1482 struct zap_details *details)
1487 p4d = p4d_offset(pgd, addr);
1489 next = p4d_addr_end(addr, end);
1490 if (p4d_none_or_clear_bad(p4d))
1492 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1493 } while (p4d++, addr = next, addr != end);
1498 void unmap_page_range(struct mmu_gather *tlb,
1499 struct vm_area_struct *vma,
1500 unsigned long addr, unsigned long end,
1501 struct zap_details *details)
1506 BUG_ON(addr >= end);
1507 tlb_start_vma(tlb, vma);
1508 pgd = pgd_offset(vma->vm_mm, addr);
1510 next = pgd_addr_end(addr, end);
1511 if (pgd_none_or_clear_bad(pgd))
1513 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1514 } while (pgd++, addr = next, addr != end);
1515 tlb_end_vma(tlb, vma);
1519 static void unmap_single_vma(struct mmu_gather *tlb,
1520 struct vm_area_struct *vma, unsigned long start_addr,
1521 unsigned long end_addr,
1522 struct zap_details *details)
1524 unsigned long start = max(vma->vm_start, start_addr);
1527 if (start >= vma->vm_end)
1529 end = min(vma->vm_end, end_addr);
1530 if (end <= vma->vm_start)
1534 uprobe_munmap(vma, start, end);
1536 if (unlikely(vma->vm_flags & VM_PFNMAP))
1537 untrack_pfn(vma, 0, 0);
1540 if (unlikely(is_vm_hugetlb_page(vma))) {
1542 * It is undesirable to test vma->vm_file as it
1543 * should be non-null for valid hugetlb area.
1544 * However, vm_file will be NULL in the error
1545 * cleanup path of mmap_region. When
1546 * hugetlbfs ->mmap method fails,
1547 * mmap_region() nullifies vma->vm_file
1548 * before calling this function to clean up.
1549 * Since no pte has actually been setup, it is
1550 * safe to do nothing in this case.
1553 i_mmap_lock_write(vma->vm_file->f_mapping);
1554 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1555 i_mmap_unlock_write(vma->vm_file->f_mapping);
1558 unmap_page_range(tlb, vma, start, end, details);
1563 * unmap_vmas - unmap a range of memory covered by a list of vma's
1564 * @tlb: address of the caller's struct mmu_gather
1565 * @vma: the starting vma
1566 * @start_addr: virtual address at which to start unmapping
1567 * @end_addr: virtual address at which to end unmapping
1569 * Unmap all pages in the vma list.
1571 * Only addresses between `start' and `end' will be unmapped.
1573 * The VMA list must be sorted in ascending virtual address order.
1575 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1576 * range after unmap_vmas() returns. So the only responsibility here is to
1577 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1578 * drops the lock and schedules.
1580 void unmap_vmas(struct mmu_gather *tlb,
1581 struct vm_area_struct *vma, unsigned long start_addr,
1582 unsigned long end_addr)
1584 struct mmu_notifier_range range;
1586 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1587 start_addr, end_addr);
1588 mmu_notifier_invalidate_range_start(&range);
1589 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1590 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1591 mmu_notifier_invalidate_range_end(&range);
1595 * zap_page_range - remove user pages in a given range
1596 * @vma: vm_area_struct holding the applicable pages
1597 * @start: starting address of pages to zap
1598 * @size: number of bytes to zap
1600 * Caller must protect the VMA list
1602 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1605 struct mmu_notifier_range range;
1606 struct mmu_gather tlb;
1609 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1610 start, start + size);
1611 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1612 update_hiwater_rss(vma->vm_mm);
1613 mmu_notifier_invalidate_range_start(&range);
1614 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1615 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1616 mmu_notifier_invalidate_range_end(&range);
1617 tlb_finish_mmu(&tlb, start, range.end);
1621 * zap_page_range_single - remove user pages in a given range
1622 * @vma: vm_area_struct holding the applicable pages
1623 * @address: starting address of pages to zap
1624 * @size: number of bytes to zap
1625 * @details: details of shared cache invalidation
1627 * The range must fit into one VMA.
1629 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1630 unsigned long size, struct zap_details *details)
1632 struct mmu_notifier_range range;
1633 struct mmu_gather tlb;
1636 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1637 address, address + size);
1638 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1639 update_hiwater_rss(vma->vm_mm);
1640 mmu_notifier_invalidate_range_start(&range);
1641 unmap_single_vma(&tlb, vma, address, range.end, details);
1642 mmu_notifier_invalidate_range_end(&range);
1643 tlb_finish_mmu(&tlb, address, range.end);
1647 * zap_vma_ptes - remove ptes mapping the vma
1648 * @vma: vm_area_struct holding ptes to be zapped
1649 * @address: starting address of pages to zap
1650 * @size: number of bytes to zap
1652 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1654 * The entire address range must be fully contained within the vma.
1657 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1660 if (address < vma->vm_start || address + size > vma->vm_end ||
1661 !(vma->vm_flags & VM_PFNMAP))
1664 zap_page_range_single(vma, address, size, NULL);
1666 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1668 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1675 pgd = pgd_offset(mm, addr);
1676 p4d = p4d_alloc(mm, pgd, addr);
1679 pud = pud_alloc(mm, p4d, addr);
1682 pmd = pmd_alloc(mm, pud, addr);
1686 VM_BUG_ON(pmd_trans_huge(*pmd));
1690 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1693 pmd_t *pmd = walk_to_pmd(mm, addr);
1697 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1700 static int validate_page_before_insert(struct page *page)
1702 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1704 flush_dcache_page(page);
1708 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1709 unsigned long addr, struct page *page, pgprot_t prot)
1711 if (!pte_none(*pte))
1713 /* Ok, finally just insert the thing.. */
1715 inc_mm_counter_fast(mm, mm_counter_file(page));
1716 page_add_file_rmap(page, false);
1717 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1722 * This is the old fallback for page remapping.
1724 * For historical reasons, it only allows reserved pages. Only
1725 * old drivers should use this, and they needed to mark their
1726 * pages reserved for the old functions anyway.
1728 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1729 struct page *page, pgprot_t prot)
1731 struct mm_struct *mm = vma->vm_mm;
1736 retval = validate_page_before_insert(page);
1740 pte = get_locked_pte(mm, addr, &ptl);
1743 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1744 pte_unmap_unlock(pte, ptl);
1750 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1751 unsigned long addr, struct page *page, pgprot_t prot)
1755 if (!page_count(page))
1757 err = validate_page_before_insert(page);
1760 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1763 /* insert_pages() amortizes the cost of spinlock operations
1764 * when inserting pages in a loop. Arch *must* define pte_index.
1766 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1767 struct page **pages, unsigned long *num, pgprot_t prot)
1770 pte_t *start_pte, *pte;
1771 spinlock_t *pte_lock;
1772 struct mm_struct *const mm = vma->vm_mm;
1773 unsigned long curr_page_idx = 0;
1774 unsigned long remaining_pages_total = *num;
1775 unsigned long pages_to_write_in_pmd;
1779 pmd = walk_to_pmd(mm, addr);
1783 pages_to_write_in_pmd = min_t(unsigned long,
1784 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1786 /* Allocate the PTE if necessary; takes PMD lock once only. */
1788 if (pte_alloc(mm, pmd))
1791 while (pages_to_write_in_pmd) {
1793 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1795 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1796 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1797 int err = insert_page_in_batch_locked(mm, pte,
1798 addr, pages[curr_page_idx], prot);
1799 if (unlikely(err)) {
1800 pte_unmap_unlock(start_pte, pte_lock);
1802 remaining_pages_total -= pte_idx;
1808 pte_unmap_unlock(start_pte, pte_lock);
1809 pages_to_write_in_pmd -= batch_size;
1810 remaining_pages_total -= batch_size;
1812 if (remaining_pages_total)
1816 *num = remaining_pages_total;
1819 #endif /* ifdef pte_index */
1822 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1823 * @vma: user vma to map to
1824 * @addr: target start user address of these pages
1825 * @pages: source kernel pages
1826 * @num: in: number of pages to map. out: number of pages that were *not*
1827 * mapped. (0 means all pages were successfully mapped).
1829 * Preferred over vm_insert_page() when inserting multiple pages.
1831 * In case of error, we may have mapped a subset of the provided
1832 * pages. It is the caller's responsibility to account for this case.
1834 * The same restrictions apply as in vm_insert_page().
1836 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1837 struct page **pages, unsigned long *num)
1840 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1842 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1844 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1845 BUG_ON(mmap_read_trylock(vma->vm_mm));
1846 BUG_ON(vma->vm_flags & VM_PFNMAP);
1847 vma->vm_flags |= VM_MIXEDMAP;
1849 /* Defer page refcount checking till we're about to map that page. */
1850 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1852 unsigned long idx = 0, pgcount = *num;
1855 for (; idx < pgcount; ++idx) {
1856 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1860 *num = pgcount - idx;
1862 #endif /* ifdef pte_index */
1864 EXPORT_SYMBOL(vm_insert_pages);
1867 * vm_insert_page - insert single page into user vma
1868 * @vma: user vma to map to
1869 * @addr: target user address of this page
1870 * @page: source kernel page
1872 * This allows drivers to insert individual pages they've allocated
1875 * The page has to be a nice clean _individual_ kernel allocation.
1876 * If you allocate a compound page, you need to have marked it as
1877 * such (__GFP_COMP), or manually just split the page up yourself
1878 * (see split_page()).
1880 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1881 * took an arbitrary page protection parameter. This doesn't allow
1882 * that. Your vma protection will have to be set up correctly, which
1883 * means that if you want a shared writable mapping, you'd better
1884 * ask for a shared writable mapping!
1886 * The page does not need to be reserved.
1888 * Usually this function is called from f_op->mmap() handler
1889 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1890 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1891 * function from other places, for example from page-fault handler.
1893 * Return: %0 on success, negative error code otherwise.
1895 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1898 if (addr < vma->vm_start || addr >= vma->vm_end)
1900 if (!page_count(page))
1902 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1903 BUG_ON(mmap_read_trylock(vma->vm_mm));
1904 BUG_ON(vma->vm_flags & VM_PFNMAP);
1905 vma->vm_flags |= VM_MIXEDMAP;
1907 return insert_page(vma, addr, page, vma->vm_page_prot);
1909 EXPORT_SYMBOL(vm_insert_page);
1912 * __vm_map_pages - maps range of kernel pages into user vma
1913 * @vma: user vma to map to
1914 * @pages: pointer to array of source kernel pages
1915 * @num: number of pages in page array
1916 * @offset: user's requested vm_pgoff
1918 * This allows drivers to map range of kernel pages into a user vma.
1920 * Return: 0 on success and error code otherwise.
1922 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1923 unsigned long num, unsigned long offset)
1925 unsigned long count = vma_pages(vma);
1926 unsigned long uaddr = vma->vm_start;
1929 /* Fail if the user requested offset is beyond the end of the object */
1933 /* Fail if the user requested size exceeds available object size */
1934 if (count > num - offset)
1937 for (i = 0; i < count; i++) {
1938 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1948 * vm_map_pages - maps range of kernel pages starts with non zero offset
1949 * @vma: user vma to map to
1950 * @pages: pointer to array of source kernel pages
1951 * @num: number of pages in page array
1953 * Maps an object consisting of @num pages, catering for the user's
1954 * requested vm_pgoff
1956 * If we fail to insert any page into the vma, the function will return
1957 * immediately leaving any previously inserted pages present. Callers
1958 * from the mmap handler may immediately return the error as their caller
1959 * will destroy the vma, removing any successfully inserted pages. Other
1960 * callers should make their own arrangements for calling unmap_region().
1962 * Context: Process context. Called by mmap handlers.
1963 * Return: 0 on success and error code otherwise.
1965 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1968 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1970 EXPORT_SYMBOL(vm_map_pages);
1973 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1974 * @vma: user vma to map to
1975 * @pages: pointer to array of source kernel pages
1976 * @num: number of pages in page array
1978 * Similar to vm_map_pages(), except that it explicitly sets the offset
1979 * to 0. This function is intended for the drivers that did not consider
1982 * Context: Process context. Called by mmap handlers.
1983 * Return: 0 on success and error code otherwise.
1985 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1988 return __vm_map_pages(vma, pages, num, 0);
1990 EXPORT_SYMBOL(vm_map_pages_zero);
1992 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1993 pfn_t pfn, pgprot_t prot, bool mkwrite)
1995 struct mm_struct *mm = vma->vm_mm;
1999 pte = get_locked_pte(mm, addr, &ptl);
2001 return VM_FAULT_OOM;
2002 if (!pte_none(*pte)) {
2005 * For read faults on private mappings the PFN passed
2006 * in may not match the PFN we have mapped if the
2007 * mapped PFN is a writeable COW page. In the mkwrite
2008 * case we are creating a writable PTE for a shared
2009 * mapping and we expect the PFNs to match. If they
2010 * don't match, we are likely racing with block
2011 * allocation and mapping invalidation so just skip the
2014 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
2015 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
2018 entry = pte_mkyoung(*pte);
2019 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2020 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2021 update_mmu_cache(vma, addr, pte);
2026 /* Ok, finally just insert the thing.. */
2027 if (pfn_t_devmap(pfn))
2028 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2030 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2033 entry = pte_mkyoung(entry);
2034 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2037 set_pte_at(mm, addr, pte, entry);
2038 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2041 pte_unmap_unlock(pte, ptl);
2042 return VM_FAULT_NOPAGE;
2046 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2047 * @vma: user vma to map to
2048 * @addr: target user address of this page
2049 * @pfn: source kernel pfn
2050 * @pgprot: pgprot flags for the inserted page
2052 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2053 * to override pgprot on a per-page basis.
2055 * This only makes sense for IO mappings, and it makes no sense for
2056 * COW mappings. In general, using multiple vmas is preferable;
2057 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2060 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2061 * a value of @pgprot different from that of @vma->vm_page_prot.
2063 * Context: Process context. May allocate using %GFP_KERNEL.
2064 * Return: vm_fault_t value.
2066 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2067 unsigned long pfn, pgprot_t pgprot)
2070 * Technically, architectures with pte_special can avoid all these
2071 * restrictions (same for remap_pfn_range). However we would like
2072 * consistency in testing and feature parity among all, so we should
2073 * try to keep these invariants in place for everybody.
2075 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2076 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2077 (VM_PFNMAP|VM_MIXEDMAP));
2078 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2079 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2081 if (addr < vma->vm_start || addr >= vma->vm_end)
2082 return VM_FAULT_SIGBUS;
2084 if (!pfn_modify_allowed(pfn, pgprot))
2085 return VM_FAULT_SIGBUS;
2087 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2089 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2092 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2095 * vmf_insert_pfn - insert single pfn into user vma
2096 * @vma: user vma to map to
2097 * @addr: target user address of this page
2098 * @pfn: source kernel pfn
2100 * Similar to vm_insert_page, this allows drivers to insert individual pages
2101 * they've allocated into a user vma. Same comments apply.
2103 * This function should only be called from a vm_ops->fault handler, and
2104 * in that case the handler should return the result of this function.
2106 * vma cannot be a COW mapping.
2108 * As this is called only for pages that do not currently exist, we
2109 * do not need to flush old virtual caches or the TLB.
2111 * Context: Process context. May allocate using %GFP_KERNEL.
2112 * Return: vm_fault_t value.
2114 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2117 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2119 EXPORT_SYMBOL(vmf_insert_pfn);
2121 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2123 /* these checks mirror the abort conditions in vm_normal_page */
2124 if (vma->vm_flags & VM_MIXEDMAP)
2126 if (pfn_t_devmap(pfn))
2128 if (pfn_t_special(pfn))
2130 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2135 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2136 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2141 BUG_ON(!vm_mixed_ok(vma, pfn));
2143 if (addr < vma->vm_start || addr >= vma->vm_end)
2144 return VM_FAULT_SIGBUS;
2146 track_pfn_insert(vma, &pgprot, pfn);
2148 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2149 return VM_FAULT_SIGBUS;
2152 * If we don't have pte special, then we have to use the pfn_valid()
2153 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2154 * refcount the page if pfn_valid is true (hence insert_page rather
2155 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2156 * without pte special, it would there be refcounted as a normal page.
2158 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2159 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2163 * At this point we are committed to insert_page()
2164 * regardless of whether the caller specified flags that
2165 * result in pfn_t_has_page() == false.
2167 page = pfn_to_page(pfn_t_to_pfn(pfn));
2168 err = insert_page(vma, addr, page, pgprot);
2170 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2174 return VM_FAULT_OOM;
2175 if (err < 0 && err != -EBUSY)
2176 return VM_FAULT_SIGBUS;
2178 return VM_FAULT_NOPAGE;
2182 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2183 * @vma: user vma to map to
2184 * @addr: target user address of this page
2185 * @pfn: source kernel pfn
2186 * @pgprot: pgprot flags for the inserted page
2188 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2189 * to override pgprot on a per-page basis.
2191 * Typically this function should be used by drivers to set caching- and
2192 * encryption bits different than those of @vma->vm_page_prot, because
2193 * the caching- or encryption mode may not be known at mmap() time.
2194 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2195 * to set caching and encryption bits for those vmas (except for COW pages).
2196 * This is ensured by core vm only modifying these page table entries using
2197 * functions that don't touch caching- or encryption bits, using pte_modify()
2198 * if needed. (See for example mprotect()).
2199 * Also when new page-table entries are created, this is only done using the
2200 * fault() callback, and never using the value of vma->vm_page_prot,
2201 * except for page-table entries that point to anonymous pages as the result
2204 * Context: Process context. May allocate using %GFP_KERNEL.
2205 * Return: vm_fault_t value.
2207 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2208 pfn_t pfn, pgprot_t pgprot)
2210 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2212 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2214 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2217 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2219 EXPORT_SYMBOL(vmf_insert_mixed);
2222 * If the insertion of PTE failed because someone else already added a
2223 * different entry in the mean time, we treat that as success as we assume
2224 * the same entry was actually inserted.
2226 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2227 unsigned long addr, pfn_t pfn)
2229 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2231 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2235 * maps a range of physical memory into the requested pages. the old
2236 * mappings are removed. any references to nonexistent pages results
2237 * in null mappings (currently treated as "copy-on-access")
2239 #ifdef CONFIG_FINEGRAINED_THP
2240 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2241 unsigned long addr, unsigned long end,
2242 unsigned long pfn, pgprot_t prot)
2244 return arch_remap_pte_range(mm, pmd, addr, end, pfn, prot);
2247 static int remap_try_huge_pmd(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
2248 unsigned long end, unsigned long pfn,
2251 phys_addr_t phys_addr = __pfn_to_phys(pfn);
2255 if ((end - addr) != PMD_SIZE)
2258 if (!IS_ALIGNED(addr, PMD_SIZE))
2261 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
2264 /* fixme - is this correct? */
2265 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr)) {
2266 pr_info("%s %d - freed pmd page??\n", __func__, __LINE__);
2270 ptl = pmd_lock(mm, pmd);
2271 ret = pmd_set_huge(pmd, phys_addr, prot);
2274 atomic_long_add(HPAGE_PMD_NR, &nr_phys_huge_pmd_pages);
2278 #else /* CONFIG_FINEGRAINED_THP */
2279 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2280 unsigned long addr, unsigned long end,
2281 unsigned long pfn, pgprot_t prot)
2283 pte_t *pte, *mapped_pte;
2287 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2290 arch_enter_lazy_mmu_mode();
2292 BUG_ON(!pte_none(*pte));
2293 if (!pfn_modify_allowed(pfn, prot)) {
2298 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2302 } while (addr != end);
2303 arch_leave_lazy_mmu_mode();
2304 pte_unmap_unlock(mapped_pte, ptl);
2307 #endif /* CONFIG_FINEGRAINED_THP */
2309 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2310 unsigned long addr, unsigned long end,
2311 unsigned long pfn, pgprot_t prot)
2317 pfn -= addr >> PAGE_SHIFT;
2318 pmd = pmd_alloc(mm, pud, addr);
2321 VM_BUG_ON(pmd_trans_huge(*pmd));
2323 next = pmd_addr_end(addr, end);
2324 #ifdef CONFIG_FINEGRAINED_THP
2325 if (remap_try_huge_pmd(mm, pmd, addr, next,
2326 pfn + (addr >> PAGE_SHIFT), prot))
2328 #endif /* CONFIG_FINEGRAINED_THP */
2329 err = remap_pte_range(mm, pmd, addr, next,
2330 pfn + (addr >> PAGE_SHIFT), prot);
2333 } while (pmd++, addr = next, addr != end);
2337 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2338 unsigned long addr, unsigned long end,
2339 unsigned long pfn, pgprot_t prot)
2345 pfn -= addr >> PAGE_SHIFT;
2346 pud = pud_alloc(mm, p4d, addr);
2350 next = pud_addr_end(addr, end);
2351 err = remap_pmd_range(mm, pud, addr, next,
2352 pfn + (addr >> PAGE_SHIFT), prot);
2355 } while (pud++, addr = next, addr != end);
2359 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2360 unsigned long addr, unsigned long end,
2361 unsigned long pfn, pgprot_t prot)
2367 pfn -= addr >> PAGE_SHIFT;
2368 p4d = p4d_alloc(mm, pgd, addr);
2372 next = p4d_addr_end(addr, end);
2373 err = remap_pud_range(mm, p4d, addr, next,
2374 pfn + (addr >> PAGE_SHIFT), prot);
2377 } while (p4d++, addr = next, addr != end);
2382 * remap_pfn_range - remap kernel memory to userspace
2383 * @vma: user vma to map to
2384 * @addr: target page aligned user address to start at
2385 * @pfn: page frame number of kernel physical memory address
2386 * @size: size of mapping area
2387 * @prot: page protection flags for this mapping
2389 * Note: this is only safe if the mm semaphore is held when called.
2391 * Return: %0 on success, negative error code otherwise.
2393 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2394 unsigned long pfn, unsigned long size, pgprot_t prot)
2398 unsigned long end = addr + PAGE_ALIGN(size);
2399 struct mm_struct *mm = vma->vm_mm;
2400 unsigned long remap_pfn = pfn;
2403 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2407 * Physically remapped pages are special. Tell the
2408 * rest of the world about it:
2409 * VM_IO tells people not to look at these pages
2410 * (accesses can have side effects).
2411 * VM_PFNMAP tells the core MM that the base pages are just
2412 * raw PFN mappings, and do not have a "struct page" associated
2415 * Disable vma merging and expanding with mremap().
2417 * Omit vma from core dump, even when VM_IO turned off.
2419 * There's a horrible special case to handle copy-on-write
2420 * behaviour that some programs depend on. We mark the "original"
2421 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2422 * See vm_normal_page() for details.
2424 if (is_cow_mapping(vma->vm_flags)) {
2425 if (addr != vma->vm_start || end != vma->vm_end)
2427 vma->vm_pgoff = pfn;
2430 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2434 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2436 BUG_ON(addr >= end);
2437 pfn -= addr >> PAGE_SHIFT;
2438 pgd = pgd_offset(mm, addr);
2439 flush_cache_range(vma, addr, end);
2441 next = pgd_addr_end(addr, end);
2442 err = remap_p4d_range(mm, pgd, addr, next,
2443 pfn + (addr >> PAGE_SHIFT), prot);
2446 } while (pgd++, addr = next, addr != end);
2449 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2453 EXPORT_SYMBOL(remap_pfn_range);
2456 * vm_iomap_memory - remap memory to userspace
2457 * @vma: user vma to map to
2458 * @start: start of the physical memory to be mapped
2459 * @len: size of area
2461 * This is a simplified io_remap_pfn_range() for common driver use. The
2462 * driver just needs to give us the physical memory range to be mapped,
2463 * we'll figure out the rest from the vma information.
2465 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2466 * whatever write-combining details or similar.
2468 * Return: %0 on success, negative error code otherwise.
2470 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2472 unsigned long vm_len, pfn, pages;
2474 /* Check that the physical memory area passed in looks valid */
2475 if (start + len < start)
2478 * You *really* shouldn't map things that aren't page-aligned,
2479 * but we've historically allowed it because IO memory might
2480 * just have smaller alignment.
2482 len += start & ~PAGE_MASK;
2483 pfn = start >> PAGE_SHIFT;
2484 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2485 if (pfn + pages < pfn)
2488 /* We start the mapping 'vm_pgoff' pages into the area */
2489 if (vma->vm_pgoff > pages)
2491 pfn += vma->vm_pgoff;
2492 pages -= vma->vm_pgoff;
2494 /* Can we fit all of the mapping? */
2495 vm_len = vma->vm_end - vma->vm_start;
2496 if (vm_len >> PAGE_SHIFT > pages)
2499 /* Ok, let it rip */
2500 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2502 EXPORT_SYMBOL(vm_iomap_memory);
2504 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2505 unsigned long addr, unsigned long end,
2506 pte_fn_t fn, void *data, bool create,
2507 pgtbl_mod_mask *mask)
2514 pte = (mm == &init_mm) ?
2515 pte_alloc_kernel_track(pmd, addr, mask) :
2516 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2520 pte = (mm == &init_mm) ?
2521 pte_offset_kernel(pmd, addr) :
2522 pte_offset_map_lock(mm, pmd, addr, &ptl);
2525 BUG_ON(pmd_huge(*pmd));
2527 arch_enter_lazy_mmu_mode();
2531 if (create || !pte_none(*pte)) {
2532 err = fn(pte++, addr, data);
2536 } while (addr += PAGE_SIZE, addr != end);
2538 *mask |= PGTBL_PTE_MODIFIED;
2540 arch_leave_lazy_mmu_mode();
2543 pte_unmap_unlock(pte-1, ptl);
2547 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2548 unsigned long addr, unsigned long end,
2549 pte_fn_t fn, void *data, bool create,
2550 pgtbl_mod_mask *mask)
2556 BUG_ON(pud_huge(*pud));
2559 pmd = pmd_alloc_track(mm, pud, addr, mask);
2563 pmd = pmd_offset(pud, addr);
2566 next = pmd_addr_end(addr, end);
2567 if (create || !pmd_none_or_clear_bad(pmd)) {
2568 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2573 } while (pmd++, addr = next, addr != end);
2577 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2578 unsigned long addr, unsigned long end,
2579 pte_fn_t fn, void *data, bool create,
2580 pgtbl_mod_mask *mask)
2587 pud = pud_alloc_track(mm, p4d, addr, mask);
2591 pud = pud_offset(p4d, addr);
2594 next = pud_addr_end(addr, end);
2595 if (create || !pud_none_or_clear_bad(pud)) {
2596 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2601 } while (pud++, addr = next, addr != end);
2605 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2606 unsigned long addr, unsigned long end,
2607 pte_fn_t fn, void *data, bool create,
2608 pgtbl_mod_mask *mask)
2615 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2619 p4d = p4d_offset(pgd, addr);
2622 next = p4d_addr_end(addr, end);
2623 if (create || !p4d_none_or_clear_bad(p4d)) {
2624 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2629 } while (p4d++, addr = next, addr != end);
2633 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2634 unsigned long size, pte_fn_t fn,
2635 void *data, bool create)
2638 unsigned long start = addr, next;
2639 unsigned long end = addr + size;
2640 pgtbl_mod_mask mask = 0;
2643 if (WARN_ON(addr >= end))
2646 pgd = pgd_offset(mm, addr);
2648 next = pgd_addr_end(addr, end);
2649 if (!create && pgd_none_or_clear_bad(pgd))
2651 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2654 } while (pgd++, addr = next, addr != end);
2656 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2657 arch_sync_kernel_mappings(start, start + size);
2663 * Scan a region of virtual memory, filling in page tables as necessary
2664 * and calling a provided function on each leaf page table.
2666 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2667 unsigned long size, pte_fn_t fn, void *data)
2669 return __apply_to_page_range(mm, addr, size, fn, data, true);
2671 EXPORT_SYMBOL_GPL(apply_to_page_range);
2674 * Scan a region of virtual memory, calling a provided function on
2675 * each leaf page table where it exists.
2677 * Unlike apply_to_page_range, this does _not_ fill in page tables
2678 * where they are absent.
2680 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2681 unsigned long size, pte_fn_t fn, void *data)
2683 return __apply_to_page_range(mm, addr, size, fn, data, false);
2685 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2688 * handle_pte_fault chooses page fault handler according to an entry which was
2689 * read non-atomically. Before making any commitment, on those architectures
2690 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2691 * parts, do_swap_page must check under lock before unmapping the pte and
2692 * proceeding (but do_wp_page is only called after already making such a check;
2693 * and do_anonymous_page can safely check later on).
2695 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2696 pte_t *page_table, pte_t orig_pte)
2699 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2700 if (sizeof(pte_t) > sizeof(unsigned long)) {
2701 spinlock_t *ptl = pte_lockptr(mm, pmd);
2703 same = pte_same(*page_table, orig_pte);
2707 pte_unmap(page_table);
2711 static inline bool cow_user_page(struct page *dst, struct page *src,
2712 struct vm_fault *vmf)
2717 bool locked = false;
2718 struct vm_area_struct *vma = vmf->vma;
2719 struct mm_struct *mm = vma->vm_mm;
2720 unsigned long addr = vmf->address;
2723 copy_user_highpage(dst, src, addr, vma);
2728 * If the source page was a PFN mapping, we don't have
2729 * a "struct page" for it. We do a best-effort copy by
2730 * just copying from the original user address. If that
2731 * fails, we just zero-fill it. Live with it.
2733 kaddr = kmap_atomic(dst);
2734 uaddr = (void __user *)(addr & PAGE_MASK);
2737 * On architectures with software "accessed" bits, we would
2738 * take a double page fault, so mark it accessed here.
2740 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2743 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2745 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2747 * Other thread has already handled the fault
2748 * and update local tlb only
2750 update_mmu_tlb(vma, addr, vmf->pte);
2755 entry = pte_mkyoung(vmf->orig_pte);
2756 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2757 update_mmu_cache(vma, addr, vmf->pte);
2761 * This really shouldn't fail, because the page is there
2762 * in the page tables. But it might just be unreadable,
2763 * in which case we just give up and fill the result with
2766 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2770 /* Re-validate under PTL if the page is still mapped */
2771 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2773 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2774 /* The PTE changed under us, update local tlb */
2775 update_mmu_tlb(vma, addr, vmf->pte);
2781 * The same page can be mapped back since last copy attempt.
2782 * Try to copy again under PTL.
2784 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2786 * Give a warn in case there can be some obscure
2799 pte_unmap_unlock(vmf->pte, vmf->ptl);
2800 kunmap_atomic(kaddr);
2801 flush_dcache_page(dst);
2806 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2808 struct file *vm_file = vma->vm_file;
2811 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2814 * Special mappings (e.g. VDSO) do not have any file so fake
2815 * a default GFP_KERNEL for them.
2821 * Notify the address space that the page is about to become writable so that
2822 * it can prohibit this or wait for the page to get into an appropriate state.
2824 * We do this without the lock held, so that it can sleep if it needs to.
2826 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2829 struct page *page = vmf->page;
2830 unsigned int old_flags = vmf->flags;
2832 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2834 if (vmf->vma->vm_file &&
2835 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2836 return VM_FAULT_SIGBUS;
2838 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2839 /* Restore original flags so that caller is not surprised */
2840 vmf->flags = old_flags;
2841 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2843 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2845 if (!page->mapping) {
2847 return 0; /* retry */
2849 ret |= VM_FAULT_LOCKED;
2851 VM_BUG_ON_PAGE(!PageLocked(page), page);
2856 * Handle dirtying of a page in shared file mapping on a write fault.
2858 * The function expects the page to be locked and unlocks it.
2860 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2862 struct vm_area_struct *vma = vmf->vma;
2863 struct address_space *mapping;
2864 struct page *page = vmf->page;
2866 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2868 dirtied = set_page_dirty(page);
2869 VM_BUG_ON_PAGE(PageAnon(page), page);
2871 * Take a local copy of the address_space - page.mapping may be zeroed
2872 * by truncate after unlock_page(). The address_space itself remains
2873 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2874 * release semantics to prevent the compiler from undoing this copying.
2876 mapping = page_rmapping(page);
2880 file_update_time(vma->vm_file);
2883 * Throttle page dirtying rate down to writeback speed.
2885 * mapping may be NULL here because some device drivers do not
2886 * set page.mapping but still dirty their pages
2888 * Drop the mmap_lock before waiting on IO, if we can. The file
2889 * is pinning the mapping, as per above.
2891 if ((dirtied || page_mkwrite) && mapping) {
2894 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2895 balance_dirty_pages_ratelimited(mapping);
2898 return VM_FAULT_RETRY;
2906 * Handle write page faults for pages that can be reused in the current vma
2908 * This can happen either due to the mapping being with the VM_SHARED flag,
2909 * or due to us being the last reference standing to the page. In either
2910 * case, all we need to do here is to mark the page as writable and update
2911 * any related book-keeping.
2913 static inline void wp_page_reuse(struct vm_fault *vmf)
2914 __releases(vmf->ptl)
2916 struct vm_area_struct *vma = vmf->vma;
2917 struct page *page = vmf->page;
2920 * Clear the pages cpupid information as the existing
2921 * information potentially belongs to a now completely
2922 * unrelated process.
2925 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2927 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2928 entry = pte_mkyoung(vmf->orig_pte);
2929 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2930 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2931 update_mmu_cache(vma, vmf->address, vmf->pte);
2932 pte_unmap_unlock(vmf->pte, vmf->ptl);
2933 count_vm_event(PGREUSE);
2937 * Handle the case of a page which we actually need to copy to a new page.
2939 * Called with mmap_lock locked and the old page referenced, but
2940 * without the ptl held.
2942 * High level logic flow:
2944 * - Allocate a page, copy the content of the old page to the new one.
2945 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2946 * - Take the PTL. If the pte changed, bail out and release the allocated page
2947 * - If the pte is still the way we remember it, update the page table and all
2948 * relevant references. This includes dropping the reference the page-table
2949 * held to the old page, as well as updating the rmap.
2950 * - In any case, unlock the PTL and drop the reference we took to the old page.
2952 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2954 struct vm_area_struct *vma = vmf->vma;
2955 struct mm_struct *mm = vma->vm_mm;
2956 struct page *old_page = vmf->page;
2957 struct page *new_page = NULL;
2959 int page_copied = 0;
2960 struct mmu_notifier_range range;
2962 if (unlikely(anon_vma_prepare(vma)))
2965 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2966 new_page = alloc_zeroed_user_highpage_movable(vma,
2971 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2976 if (!cow_user_page(new_page, old_page, vmf)) {
2978 * COW failed, if the fault was solved by other,
2979 * it's fine. If not, userspace would re-fault on
2980 * the same address and we will handle the fault
2981 * from the second attempt.
2990 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2992 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2994 __SetPageUptodate(new_page);
2996 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2997 vmf->address & PAGE_MASK,
2998 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2999 mmu_notifier_invalidate_range_start(&range);
3002 * Re-check the pte - we dropped the lock
3004 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3005 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
3007 if (!PageAnon(old_page)) {
3008 dec_mm_counter_fast(mm,
3009 mm_counter_file(old_page));
3010 inc_mm_counter_fast(mm, MM_ANONPAGES);
3013 inc_mm_counter_fast(mm, MM_ANONPAGES);
3015 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3016 entry = mk_pte(new_page, vma->vm_page_prot);
3017 entry = pte_sw_mkyoung(entry);
3018 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3020 * Clear the pte entry and flush it first, before updating the
3021 * pte with the new entry. This will avoid a race condition
3022 * seen in the presence of one thread doing SMC and another
3025 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3026 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
3027 lru_cache_add_inactive_or_unevictable(new_page, vma);
3029 * We call the notify macro here because, when using secondary
3030 * mmu page tables (such as kvm shadow page tables), we want the
3031 * new page to be mapped directly into the secondary page table.
3033 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3034 update_mmu_cache(vma, vmf->address, vmf->pte);
3037 * Only after switching the pte to the new page may
3038 * we remove the mapcount here. Otherwise another
3039 * process may come and find the rmap count decremented
3040 * before the pte is switched to the new page, and
3041 * "reuse" the old page writing into it while our pte
3042 * here still points into it and can be read by other
3045 * The critical issue is to order this
3046 * page_remove_rmap with the ptp_clear_flush above.
3047 * Those stores are ordered by (if nothing else,)
3048 * the barrier present in the atomic_add_negative
3049 * in page_remove_rmap.
3051 * Then the TLB flush in ptep_clear_flush ensures that
3052 * no process can access the old page before the
3053 * decremented mapcount is visible. And the old page
3054 * cannot be reused until after the decremented
3055 * mapcount is visible. So transitively, TLBs to
3056 * old page will be flushed before it can be reused.
3058 page_remove_rmap(old_page, false);
3061 /* Free the old page.. */
3062 new_page = old_page;
3065 update_mmu_tlb(vma, vmf->address, vmf->pte);
3071 pte_unmap_unlock(vmf->pte, vmf->ptl);
3073 * No need to double call mmu_notifier->invalidate_range() callback as
3074 * the above ptep_clear_flush_notify() did already call it.
3076 mmu_notifier_invalidate_range_only_end(&range);
3079 * Don't let another task, with possibly unlocked vma,
3080 * keep the mlocked page.
3082 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
3083 lock_page(old_page); /* LRU manipulation */
3084 if (PageMlocked(old_page))
3085 munlock_vma_page(old_page);
3086 unlock_page(old_page);
3090 return page_copied ? VM_FAULT_WRITE : 0;
3096 return VM_FAULT_OOM;
3100 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3101 * writeable once the page is prepared
3103 * @vmf: structure describing the fault
3105 * This function handles all that is needed to finish a write page fault in a
3106 * shared mapping due to PTE being read-only once the mapped page is prepared.
3107 * It handles locking of PTE and modifying it.
3109 * The function expects the page to be locked or other protection against
3110 * concurrent faults / writeback (such as DAX radix tree locks).
3112 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
3113 * we acquired PTE lock.
3115 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3117 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3118 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3121 * We might have raced with another page fault while we released the
3122 * pte_offset_map_lock.
3124 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3125 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3126 pte_unmap_unlock(vmf->pte, vmf->ptl);
3127 return VM_FAULT_NOPAGE;
3134 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3137 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3139 struct vm_area_struct *vma = vmf->vma;
3141 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3144 pte_unmap_unlock(vmf->pte, vmf->ptl);
3145 vmf->flags |= FAULT_FLAG_MKWRITE;
3146 ret = vma->vm_ops->pfn_mkwrite(vmf);
3147 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3149 return finish_mkwrite_fault(vmf);
3152 return VM_FAULT_WRITE;
3155 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3156 __releases(vmf->ptl)
3158 struct vm_area_struct *vma = vmf->vma;
3159 vm_fault_t ret = VM_FAULT_WRITE;
3161 get_page(vmf->page);
3163 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3166 pte_unmap_unlock(vmf->pte, vmf->ptl);
3167 tmp = do_page_mkwrite(vmf);
3168 if (unlikely(!tmp || (tmp &
3169 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3170 put_page(vmf->page);
3173 tmp = finish_mkwrite_fault(vmf);
3174 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3175 unlock_page(vmf->page);
3176 put_page(vmf->page);
3181 lock_page(vmf->page);
3183 ret |= fault_dirty_shared_page(vmf);
3184 put_page(vmf->page);
3190 * This routine handles present pages, when users try to write
3191 * to a shared page. It is done by copying the page to a new address
3192 * and decrementing the shared-page counter for the old page.
3194 * Note that this routine assumes that the protection checks have been
3195 * done by the caller (the low-level page fault routine in most cases).
3196 * Thus we can safely just mark it writable once we've done any necessary
3199 * We also mark the page dirty at this point even though the page will
3200 * change only once the write actually happens. This avoids a few races,
3201 * and potentially makes it more efficient.
3203 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3204 * but allow concurrent faults), with pte both mapped and locked.
3205 * We return with mmap_lock still held, but pte unmapped and unlocked.
3207 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3208 __releases(vmf->ptl)
3210 struct vm_area_struct *vma = vmf->vma;
3212 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3213 pte_unmap_unlock(vmf->pte, vmf->ptl);
3214 return handle_userfault(vmf, VM_UFFD_WP);
3218 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3219 * is flushed in this case before copying.
3221 if (unlikely(userfaultfd_wp(vmf->vma) &&
3222 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3223 flush_tlb_page(vmf->vma, vmf->address);
3225 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3228 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3231 * We should not cow pages in a shared writeable mapping.
3232 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3234 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3235 (VM_WRITE|VM_SHARED))
3236 return wp_pfn_shared(vmf);
3238 pte_unmap_unlock(vmf->pte, vmf->ptl);
3239 return wp_page_copy(vmf);
3243 * Take out anonymous pages first, anonymous shared vmas are
3244 * not dirty accountable.
3246 if (PageAnon(vmf->page)) {
3247 struct page *page = vmf->page;
3249 /* PageKsm() doesn't necessarily raise the page refcount */
3250 if (PageKsm(page) || page_count(page) != 1)
3252 if (!trylock_page(page))
3254 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3259 * Ok, we've got the only map reference, and the only
3260 * page count reference, and the page is locked,
3261 * it's dark out, and we're wearing sunglasses. Hit it.
3265 return VM_FAULT_WRITE;
3266 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3267 (VM_WRITE|VM_SHARED))) {
3268 return wp_page_shared(vmf);
3272 * Ok, we need to copy. Oh, well..
3274 get_page(vmf->page);
3276 pte_unmap_unlock(vmf->pte, vmf->ptl);
3277 return wp_page_copy(vmf);
3280 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3281 unsigned long start_addr, unsigned long end_addr,
3282 struct zap_details *details)
3284 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3287 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3288 struct zap_details *details)
3290 struct vm_area_struct *vma;
3291 pgoff_t vba, vea, zba, zea;
3293 vma_interval_tree_foreach(vma, root,
3294 details->first_index, details->last_index) {
3296 vba = vma->vm_pgoff;
3297 vea = vba + vma_pages(vma) - 1;
3298 zba = details->first_index;
3301 zea = details->last_index;
3305 unmap_mapping_range_vma(vma,
3306 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3307 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3313 * unmap_mapping_page() - Unmap single page from processes.
3314 * @page: The locked page to be unmapped.
3316 * Unmap this page from any userspace process which still has it mmaped.
3317 * Typically, for efficiency, the range of nearby pages has already been
3318 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3319 * truncation or invalidation holds the lock on a page, it may find that
3320 * the page has been remapped again: and then uses unmap_mapping_page()
3321 * to unmap it finally.
3323 void unmap_mapping_page(struct page *page)
3325 struct address_space *mapping = page->mapping;
3326 struct zap_details details = { };
3328 VM_BUG_ON(!PageLocked(page));
3329 VM_BUG_ON(PageTail(page));
3331 details.check_mapping = mapping;
3332 details.first_index = page->index;
3333 details.last_index = page->index + thp_nr_pages(page) - 1;
3334 details.single_page = page;
3336 i_mmap_lock_write(mapping);
3337 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3338 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3339 i_mmap_unlock_write(mapping);
3343 * unmap_mapping_pages() - Unmap pages from processes.
3344 * @mapping: The address space containing pages to be unmapped.
3345 * @start: Index of first page to be unmapped.
3346 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3347 * @even_cows: Whether to unmap even private COWed pages.
3349 * Unmap the pages in this address space from any userspace process which
3350 * has them mmaped. Generally, you want to remove COWed pages as well when
3351 * a file is being truncated, but not when invalidating pages from the page
3354 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3355 pgoff_t nr, bool even_cows)
3357 struct zap_details details = { };
3359 details.check_mapping = even_cows ? NULL : mapping;
3360 details.first_index = start;
3361 details.last_index = start + nr - 1;
3362 if (details.last_index < details.first_index)
3363 details.last_index = ULONG_MAX;
3365 i_mmap_lock_write(mapping);
3366 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3367 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3368 i_mmap_unlock_write(mapping);
3372 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3373 * address_space corresponding to the specified byte range in the underlying
3376 * @mapping: the address space containing mmaps to be unmapped.
3377 * @holebegin: byte in first page to unmap, relative to the start of
3378 * the underlying file. This will be rounded down to a PAGE_SIZE
3379 * boundary. Note that this is different from truncate_pagecache(), which
3380 * must keep the partial page. In contrast, we must get rid of
3382 * @holelen: size of prospective hole in bytes. This will be rounded
3383 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3385 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3386 * but 0 when invalidating pagecache, don't throw away private data.
3388 void unmap_mapping_range(struct address_space *mapping,
3389 loff_t const holebegin, loff_t const holelen, int even_cows)
3391 pgoff_t hba = holebegin >> PAGE_SHIFT;
3392 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3394 /* Check for overflow. */
3395 if (sizeof(holelen) > sizeof(hlen)) {
3397 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3398 if (holeend & ~(long long)ULONG_MAX)
3399 hlen = ULONG_MAX - hba + 1;
3402 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3404 EXPORT_SYMBOL(unmap_mapping_range);
3407 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3408 * but allow concurrent faults), and pte mapped but not yet locked.
3409 * We return with pte unmapped and unlocked.
3411 * We return with the mmap_lock locked or unlocked in the same cases
3412 * as does filemap_fault().
3414 vm_fault_t do_swap_page(struct vm_fault *vmf)
3416 struct vm_area_struct *vma = vmf->vma;
3417 struct page *page = NULL, *swapcache;
3423 void *shadow = NULL;
3425 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3428 entry = pte_to_swp_entry(vmf->orig_pte);
3429 if (unlikely(non_swap_entry(entry))) {
3430 if (is_migration_entry(entry)) {
3431 migration_entry_wait(vma->vm_mm, vmf->pmd,
3433 } else if (is_device_private_entry(entry)) {
3434 vmf->page = device_private_entry_to_page(entry);
3435 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3436 } else if (is_hwpoison_entry(entry)) {
3437 ret = VM_FAULT_HWPOISON;
3439 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3440 ret = VM_FAULT_SIGBUS;
3446 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3447 page = lookup_swap_cache(entry, vma, vmf->address);
3451 struct swap_info_struct *si = swp_swap_info(entry);
3453 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3454 __swap_count(entry) == 1) {
3455 /* skip swapcache */
3456 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3461 __SetPageLocked(page);
3462 __SetPageSwapBacked(page);
3463 set_page_private(page, entry.val);
3465 /* Tell memcg to use swap ownership records */
3466 SetPageSwapCache(page);
3467 err = mem_cgroup_charge(page, vma->vm_mm,
3469 ClearPageSwapCache(page);
3475 shadow = get_shadow_from_swap_cache(entry);
3477 workingset_refault(page, shadow);
3479 lru_cache_add(page);
3480 swap_readpage(page, true);
3483 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3490 * Back out if somebody else faulted in this pte
3491 * while we released the pte lock.
3493 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3494 vmf->address, &vmf->ptl);
3495 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3497 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3501 /* Had to read the page from swap area: Major fault */
3502 ret = VM_FAULT_MAJOR;
3503 count_vm_event(PGMAJFAULT);
3504 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3505 } else if (PageHWPoison(page)) {
3507 * hwpoisoned dirty swapcache pages are kept for killing
3508 * owner processes (which may be unknown at hwpoison time)
3510 ret = VM_FAULT_HWPOISON;
3511 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3515 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3517 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3519 ret |= VM_FAULT_RETRY;
3524 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3525 * release the swapcache from under us. The page pin, and pte_same
3526 * test below, are not enough to exclude that. Even if it is still
3527 * swapcache, we need to check that the page's swap has not changed.
3529 if (unlikely((!PageSwapCache(page) ||
3530 page_private(page) != entry.val)) && swapcache)
3533 page = ksm_might_need_to_copy(page, vma, vmf->address);
3534 if (unlikely(!page)) {
3540 cgroup_throttle_swaprate(page, GFP_KERNEL);
3543 * Back out if somebody else already faulted in this pte.
3545 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3547 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3550 if (unlikely(!PageUptodate(page))) {
3551 ret = VM_FAULT_SIGBUS;
3556 * The page isn't present yet, go ahead with the fault.
3558 * Be careful about the sequence of operations here.
3559 * To get its accounting right, reuse_swap_page() must be called
3560 * while the page is counted on swap but not yet in mapcount i.e.
3561 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3562 * must be called after the swap_free(), or it will never succeed.
3565 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3566 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3567 pte = mk_pte(page, vma->vm_page_prot);
3568 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3569 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3570 vmf->flags &= ~FAULT_FLAG_WRITE;
3571 ret |= VM_FAULT_WRITE;
3572 exclusive = RMAP_EXCLUSIVE;
3574 flush_icache_page(vma, page);
3575 if (pte_swp_soft_dirty(vmf->orig_pte))
3576 pte = pte_mksoft_dirty(pte);
3577 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3578 pte = pte_mkuffd_wp(pte);
3579 pte = pte_wrprotect(pte);
3581 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3582 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3583 vmf->orig_pte = pte;
3585 /* ksm created a completely new copy */
3586 if (unlikely(page != swapcache && swapcache)) {
3587 page_add_new_anon_rmap(page, vma, vmf->address, false);
3588 lru_cache_add_inactive_or_unevictable(page, vma);
3590 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3594 if (mem_cgroup_swap_full(page) ||
3595 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3596 try_to_free_swap(page);
3598 if (page != swapcache && swapcache) {
3600 * Hold the lock to avoid the swap entry to be reused
3601 * until we take the PT lock for the pte_same() check
3602 * (to avoid false positives from pte_same). For
3603 * further safety release the lock after the swap_free
3604 * so that the swap count won't change under a
3605 * parallel locked swapcache.
3607 unlock_page(swapcache);
3608 put_page(swapcache);
3611 if (vmf->flags & FAULT_FLAG_WRITE) {
3612 ret |= do_wp_page(vmf);
3613 if (ret & VM_FAULT_ERROR)
3614 ret &= VM_FAULT_ERROR;
3618 /* No need to invalidate - it was non-present before */
3619 update_mmu_cache(vma, vmf->address, vmf->pte);
3621 pte_unmap_unlock(vmf->pte, vmf->ptl);
3625 pte_unmap_unlock(vmf->pte, vmf->ptl);
3630 if (page != swapcache && swapcache) {
3631 unlock_page(swapcache);
3632 put_page(swapcache);
3637 extern bool eager_allocation;
3640 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3641 * but allow concurrent faults), and pte mapped but not yet locked.
3642 * We return with mmap_lock still held, but pte unmapped and unlocked.
3644 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3646 struct vm_area_struct *vma = vmf->vma;
3651 /* File mapping without ->vm_ops ? */
3652 if (vma->vm_flags & VM_SHARED)
3653 return VM_FAULT_SIGBUS;
3656 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3657 * pte_offset_map() on pmds where a huge pmd might be created
3658 * from a different thread.
3660 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3661 * parallel threads are excluded by other means.
3663 * Here we only have mmap_read_lock(mm).
3665 if (pte_alloc(vma->vm_mm, vmf->pmd))
3666 return VM_FAULT_OOM;
3668 /* See the comment in pte_alloc_one_map() */
3669 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3672 /* Use the zero-page for reads */
3673 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3674 !mm_forbids_zeropage(vma->vm_mm)) {
3675 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3676 vma->vm_page_prot));
3677 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3678 vmf->address, &vmf->ptl);
3679 if (!pte_none(*vmf->pte)) {
3680 update_mmu_tlb(vma, vmf->address, vmf->pte);
3683 ret = check_stable_address_space(vma->vm_mm);
3686 /* Deliver the page fault to userland, check inside PT lock */
3687 if (userfaultfd_missing(vma)) {
3688 pte_unmap_unlock(vmf->pte, vmf->ptl);
3689 return handle_userfault(vmf, VM_UFFD_MISSING);
3694 /* Allocate our own private page. */
3695 if (unlikely(anon_vma_prepare(vma)))
3697 #ifdef CONFIG_FINEGRAINED_THP
3698 #ifndef CONFIG_THP_CONSERVATIVE
3700 * 64KB hugepage creation on page fault is only allowed
3701 * in an aggressive policy or a near-conservative policy
3703 if (__transparent_hugepage_enabled(vma)) {
3704 ret = arch_do_huge_pte_anonymous_page(vmf);
3705 if (!(ret & VM_FAULT_FALLBACK)) {
3710 #endif /* CONFIG_THP_CONSERVATIVE */
3711 #endif /* CONFIG_FINEGRAINED_THP */
3713 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3717 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3719 cgroup_throttle_swaprate(page, GFP_KERNEL);
3722 * The memory barrier inside __SetPageUptodate makes sure that
3723 * preceding stores to the page contents become visible before
3724 * the set_pte_at() write.
3726 __SetPageUptodate(page);
3728 entry = mk_pte(page, vma->vm_page_prot);
3729 entry = pte_sw_mkyoung(entry);
3730 if (vma->vm_flags & VM_WRITE)
3731 entry = pte_mkwrite(pte_mkdirty(entry));
3733 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3735 if (!pte_none(*vmf->pte)) {
3736 update_mmu_cache(vma, vmf->address, vmf->pte);
3740 ret = check_stable_address_space(vma->vm_mm);
3744 /* Deliver the page fault to userland, check inside PT lock */
3745 if (userfaultfd_missing(vma)) {
3746 pte_unmap_unlock(vmf->pte, vmf->ptl);
3748 return handle_userfault(vmf, VM_UFFD_MISSING);
3751 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3752 page_add_new_anon_rmap(page, vma, vmf->address, false);
3753 lru_cache_add_inactive_or_unevictable(page, vma);
3755 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3757 /* No need to invalidate - it was non-present before */
3758 update_mmu_cache(vma, vmf->address, vmf->pte);
3760 pte_unmap_unlock(vmf->pte, vmf->ptl);
3768 return VM_FAULT_OOM;
3772 * The mmap_lock must have been held on entry, and may have been
3773 * released depending on flags and vma->vm_ops->fault() return value.
3774 * See filemap_fault() and __lock_page_retry().
3776 static vm_fault_t __do_fault(struct vm_fault *vmf)
3778 struct vm_area_struct *vma = vmf->vma;
3782 * Preallocate pte before we take page_lock because this might lead to
3783 * deadlocks for memcg reclaim which waits for pages under writeback:
3785 * SetPageWriteback(A)
3791 * wait_on_page_writeback(A)
3792 * SetPageWriteback(B)
3794 * # flush A, B to clear the writeback
3796 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3797 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3798 if (!vmf->prealloc_pte)
3799 return VM_FAULT_OOM;
3800 smp_wmb(); /* See comment in __pte_alloc() */
3803 ret = vma->vm_ops->fault(vmf);
3804 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3805 VM_FAULT_DONE_COW)))
3808 if (unlikely(PageHWPoison(vmf->page))) {
3809 if (ret & VM_FAULT_LOCKED)
3810 unlock_page(vmf->page);
3811 put_page(vmf->page);
3813 return VM_FAULT_HWPOISON;
3816 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3817 lock_page(vmf->page);
3819 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3825 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3826 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3827 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3828 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3830 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3832 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3835 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3837 struct vm_area_struct *vma = vmf->vma;
3839 if (!pmd_none(*vmf->pmd))
3841 if (vmf->prealloc_pte) {
3842 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3843 if (unlikely(!pmd_none(*vmf->pmd))) {
3844 spin_unlock(vmf->ptl);
3848 mm_inc_nr_ptes(vma->vm_mm);
3849 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3850 spin_unlock(vmf->ptl);
3851 vmf->prealloc_pte = NULL;
3852 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3853 return VM_FAULT_OOM;
3857 * If a huge pmd materialized under us just retry later. Use
3858 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3859 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3860 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3861 * running immediately after a huge pmd fault in a different thread of
3862 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3863 * All we have to ensure is that it is a regular pmd that we can walk
3864 * with pte_offset_map() and we can do that through an atomic read in
3865 * C, which is what pmd_trans_unstable() provides.
3867 if (pmd_devmap_trans_unstable(vmf->pmd))
3868 return VM_FAULT_NOPAGE;
3871 * At this point we know that our vmf->pmd points to a page of ptes
3872 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3873 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3874 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3875 * be valid and we will re-check to make sure the vmf->pte isn't
3876 * pte_none() under vmf->ptl protection when we return to
3879 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3884 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3885 static void deposit_prealloc_pte(struct vm_fault *vmf)
3887 struct vm_area_struct *vma = vmf->vma;
3889 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3891 * We are going to consume the prealloc table,
3892 * count that as nr_ptes.
3894 mm_inc_nr_ptes(vma->vm_mm);
3895 vmf->prealloc_pte = NULL;
3898 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3900 struct vm_area_struct *vma = vmf->vma;
3901 bool write = vmf->flags & FAULT_FLAG_WRITE;
3902 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3905 vm_fault_t ret = VM_FAULT_FALLBACK;
3907 if (!transhuge_vma_suitable(vma, haddr))
3910 page = compound_head(page);
3911 if (compound_order(page) != HPAGE_PMD_ORDER)
3915 * Archs like ppc64 need additonal space to store information
3916 * related to pte entry. Use the preallocated table for that.
3918 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3919 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3920 if (!vmf->prealloc_pte)
3921 return VM_FAULT_OOM;
3922 smp_wmb(); /* See comment in __pte_alloc() */
3925 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3926 if (unlikely(!pmd_none(*vmf->pmd)))
3929 for (i = 0; i < HPAGE_PMD_NR; i++)
3930 flush_icache_page(vma, page + i);
3932 entry = mk_huge_pmd(page, vma->vm_page_prot);
3934 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3936 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3937 page_add_file_rmap(page, true);
3939 * deposit and withdraw with pmd lock held
3941 if (arch_needs_pgtable_deposit())
3942 deposit_prealloc_pte(vmf);
3944 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3946 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3948 /* fault is handled */
3950 count_vm_event(THP_FILE_MAPPED);
3952 spin_unlock(vmf->ptl);
3956 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3962 #ifdef CONFIG_FINEGRAINED_THP
3963 static vm_fault_t arch_do_set_huge_pte(struct vm_fault *vmf, struct page *page)
3972 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3973 * mapping. If needed, the function allocates page table or use pre-allocated.
3975 * @vmf: fault environment
3976 * @page: page to map
3978 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3981 * Target users are page handler itself and implementations of
3982 * vm_ops->map_pages.
3984 * Return: %0 on success, %VM_FAULT_ code in case of error.
3986 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
3988 struct vm_area_struct *vma = vmf->vma;
3989 bool write = vmf->flags & FAULT_FLAG_WRITE;
3993 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3994 compound_nr(compound_head(page)) == HPAGE_PMD_NR) {
3995 ret = do_set_pmd(vmf, page);
3996 if (ret != VM_FAULT_FALLBACK)
4000 #ifdef CONFIG_FINEGRAINED_THP
4001 /* PageTransHuge cannot find hugepage if the page is not a head */
4002 if (PageTransCompound(page) &&
4003 compound_nr(compound_head(page)) == HPAGE_CONT_PTE_NR) {
4004 ret = arch_do_set_huge_pte(vmf, page);
4005 if (ret != VM_FAULT_FALLBACK)
4008 #endif /* CONFIG_FINEGRAINED_THP */
4011 ret = pte_alloc_one_map(vmf);
4016 /* Re-check under ptl */
4017 if (unlikely(!pte_none(*vmf->pte))) {
4018 update_mmu_tlb(vma, vmf->address, vmf->pte);
4019 return VM_FAULT_NOPAGE;
4022 if (!strcmp(current->comm, "org.tizen.nlp.s") || !strcmp(current->comm, "memps"))
4023 pr_info("THP-wp: huge fault for addr (%lx) (%s) %s\n",
4024 vmf->address, current->comm, __func__);
4026 flush_icache_page(vma, page);
4027 entry = mk_pte(page, vma->vm_page_prot);
4028 entry = pte_sw_mkyoung(entry);
4030 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4031 /* copy-on-write page */
4032 if (write && !(vma->vm_flags & VM_SHARED)) {
4033 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
4034 page_add_new_anon_rmap(page, vma, vmf->address, false);
4035 lru_cache_add_inactive_or_unevictable(page, vma);
4037 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
4038 page_add_file_rmap(page, false);
4040 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
4042 /* no need to invalidate: a not-present page won't be cached */
4043 update_mmu_cache(vma, vmf->address, vmf->pte);
4050 * finish_fault - finish page fault once we have prepared the page to fault
4052 * @vmf: structure describing the fault
4054 * This function handles all that is needed to finish a page fault once the
4055 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4056 * given page, adds reverse page mapping, handles memcg charges and LRU
4059 * The function expects the page to be locked and on success it consumes a
4060 * reference of a page being mapped (for the PTE which maps it).
4062 * Return: %0 on success, %VM_FAULT_ code in case of error.
4064 vm_fault_t finish_fault(struct vm_fault *vmf)
4069 /* Did we COW the page? */
4070 if ((vmf->flags & FAULT_FLAG_WRITE) &&
4071 !(vmf->vma->vm_flags & VM_SHARED))
4072 page = vmf->cow_page;
4077 * check even for read faults because we might have lost our CoWed
4080 if (!(vmf->vma->vm_flags & VM_SHARED))
4081 ret = check_stable_address_space(vmf->vma->vm_mm);
4083 ret = alloc_set_pte(vmf, page);
4085 pte_unmap_unlock(vmf->pte, vmf->ptl);
4089 static unsigned long fault_around_bytes __read_mostly =
4090 rounddown_pow_of_two(4096);
4092 #ifdef CONFIG_DEBUG_FS
4093 static int fault_around_bytes_get(void *data, u64 *val)
4095 *val = fault_around_bytes;
4100 * fault_around_bytes must be rounded down to the nearest page order as it's
4101 * what do_fault_around() expects to see.
4103 static int fault_around_bytes_set(void *data, u64 val)
4105 if (val / PAGE_SIZE > PTRS_PER_PTE)
4107 if (val > PAGE_SIZE)
4108 fault_around_bytes = rounddown_pow_of_two(val);
4110 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
4113 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4114 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4116 static int __init fault_around_debugfs(void)
4118 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4119 &fault_around_bytes_fops);
4122 late_initcall(fault_around_debugfs);
4126 * do_fault_around() tries to map few pages around the fault address. The hope
4127 * is that the pages will be needed soon and this will lower the number of
4130 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4131 * not ready to be mapped: not up-to-date, locked, etc.
4133 * This function is called with the page table lock taken. In the split ptlock
4134 * case the page table lock only protects only those entries which belong to
4135 * the page table corresponding to the fault address.
4137 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4140 * fault_around_bytes defines how many bytes we'll try to map.
4141 * do_fault_around() expects it to be set to a power of two less than or equal
4144 * The virtual address of the area that we map is naturally aligned to
4145 * fault_around_bytes rounded down to the machine page size
4146 * (and therefore to page order). This way it's easier to guarantee
4147 * that we don't cross page table boundaries.
4149 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4151 unsigned long address = vmf->address, nr_pages, mask;
4152 pgoff_t start_pgoff = vmf->pgoff;
4157 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4158 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4160 vmf->address = max(address & mask, vmf->vma->vm_start);
4161 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4165 * end_pgoff is either the end of the page table, the end of
4166 * the vma or nr_pages from start_pgoff, depending what is nearest.
4168 end_pgoff = start_pgoff -
4169 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4171 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4172 start_pgoff + nr_pages - 1);
4174 if (pmd_none(*vmf->pmd)) {
4175 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4176 if (!vmf->prealloc_pte)
4178 smp_wmb(); /* See comment in __pte_alloc() */
4181 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4183 /* Huge page is mapped? Page fault is solved */
4184 if (pmd_trans_huge(*vmf->pmd)) {
4185 ret = VM_FAULT_NOPAGE;
4189 /* ->map_pages() haven't done anything useful. Cold page cache? */
4193 /* check if the page fault is solved */
4194 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
4195 if (!pte_none(*vmf->pte))
4196 ret = VM_FAULT_NOPAGE;
4197 pte_unmap_unlock(vmf->pte, vmf->ptl);
4199 vmf->address = address;
4204 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4206 struct vm_area_struct *vma = vmf->vma;
4210 * Let's call ->map_pages() first and use ->fault() as fallback
4211 * if page by the offset is not ready to be mapped (cold cache or
4214 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4215 ret = do_fault_around(vmf);
4220 ret = __do_fault(vmf);
4221 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4224 ret |= finish_fault(vmf);
4225 unlock_page(vmf->page);
4226 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4227 put_page(vmf->page);
4231 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4233 struct vm_area_struct *vma = vmf->vma;
4236 if (unlikely(anon_vma_prepare(vma)))
4237 return VM_FAULT_OOM;
4239 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4241 return VM_FAULT_OOM;
4243 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4244 put_page(vmf->cow_page);
4245 return VM_FAULT_OOM;
4247 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4249 ret = __do_fault(vmf);
4250 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4252 if (ret & VM_FAULT_DONE_COW)
4254 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4255 __SetPageUptodate(vmf->cow_page);
4257 ret |= finish_fault(vmf);
4258 unlock_page(vmf->page);
4259 put_page(vmf->page);
4260 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4264 put_page(vmf->cow_page);
4268 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4270 struct vm_area_struct *vma = vmf->vma;
4271 vm_fault_t ret, tmp;
4273 ret = __do_fault(vmf);
4274 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4278 * Check if the backing address space wants to know that the page is
4279 * about to become writable
4281 if (vma->vm_ops->page_mkwrite) {
4282 unlock_page(vmf->page);
4283 tmp = do_page_mkwrite(vmf);
4284 if (unlikely(!tmp ||
4285 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4286 put_page(vmf->page);
4291 ret |= finish_fault(vmf);
4292 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4294 unlock_page(vmf->page);
4295 put_page(vmf->page);
4299 ret |= fault_dirty_shared_page(vmf);
4304 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4305 * but allow concurrent faults).
4306 * The mmap_lock may have been released depending on flags and our
4307 * return value. See filemap_fault() and __lock_page_or_retry().
4308 * If mmap_lock is released, vma may become invalid (for example
4309 * by other thread calling munmap()).
4311 static vm_fault_t do_fault(struct vm_fault *vmf)
4313 struct vm_area_struct *vma = vmf->vma;
4314 struct mm_struct *vm_mm = vma->vm_mm;
4318 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4320 if (!vma->vm_ops->fault) {
4322 * If we find a migration pmd entry or a none pmd entry, which
4323 * should never happen, return SIGBUS
4325 if (unlikely(!pmd_present(*vmf->pmd)))
4326 ret = VM_FAULT_SIGBUS;
4328 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4333 * Make sure this is not a temporary clearing of pte
4334 * by holding ptl and checking again. A R/M/W update
4335 * of pte involves: take ptl, clearing the pte so that
4336 * we don't have concurrent modification by hardware
4337 * followed by an update.
4339 if (unlikely(pte_none(*vmf->pte)))
4340 ret = VM_FAULT_SIGBUS;
4342 ret = VM_FAULT_NOPAGE;
4344 pte_unmap_unlock(vmf->pte, vmf->ptl);
4346 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4347 ret = do_read_fault(vmf);
4348 else if (!(vma->vm_flags & VM_SHARED))
4349 ret = do_cow_fault(vmf);
4351 ret = do_shared_fault(vmf);
4353 /* preallocated pagetable is unused: free it */
4354 if (vmf->prealloc_pte) {
4355 pte_free(vm_mm, vmf->prealloc_pte);
4356 vmf->prealloc_pte = NULL;
4361 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4362 unsigned long addr, int page_nid,
4367 count_vm_numa_event(NUMA_HINT_FAULTS);
4368 if (page_nid == numa_node_id()) {
4369 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4370 *flags |= TNF_FAULT_LOCAL;
4373 return mpol_misplaced(page, vma, addr);
4376 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4378 struct vm_area_struct *vma = vmf->vma;
4379 struct page *page = NULL;
4380 int page_nid = NUMA_NO_NODE;
4383 bool migrated = false;
4385 bool was_writable = pte_savedwrite(vmf->orig_pte);
4389 * The "pte" at this point cannot be used safely without
4390 * validation through pte_unmap_same(). It's of NUMA type but
4391 * the pfn may be screwed if the read is non atomic.
4393 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4394 spin_lock(vmf->ptl);
4395 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4396 pte_unmap_unlock(vmf->pte, vmf->ptl);
4401 * Make it present again, Depending on how arch implementes non
4402 * accessible ptes, some can allow access by kernel mode.
4404 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4405 pte = pte_modify(old_pte, vma->vm_page_prot);
4406 pte = pte_mkyoung(pte);
4408 pte = pte_mkwrite(pte);
4409 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4410 update_mmu_cache(vma, vmf->address, vmf->pte);
4412 page = vm_normal_page(vma, vmf->address, pte);
4414 pte_unmap_unlock(vmf->pte, vmf->ptl);
4418 /* TODO: handle PTE-mapped THP */
4419 if (PageCompound(page)) {
4420 pte_unmap_unlock(vmf->pte, vmf->ptl);
4425 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4426 * much anyway since they can be in shared cache state. This misses
4427 * the case where a mapping is writable but the process never writes
4428 * to it but pte_write gets cleared during protection updates and
4429 * pte_dirty has unpredictable behaviour between PTE scan updates,
4430 * background writeback, dirty balancing and application behaviour.
4432 if (!pte_write(pte))
4433 flags |= TNF_NO_GROUP;
4436 * Flag if the page is shared between multiple address spaces. This
4437 * is later used when determining whether to group tasks together
4439 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4440 flags |= TNF_SHARED;
4442 last_cpupid = page_cpupid_last(page);
4443 page_nid = page_to_nid(page);
4444 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4446 pte_unmap_unlock(vmf->pte, vmf->ptl);
4447 if (target_nid == NUMA_NO_NODE) {
4452 /* Migrate to the requested node */
4453 migrated = migrate_misplaced_page(page, vma, target_nid);
4455 page_nid = target_nid;
4456 flags |= TNF_MIGRATED;
4458 flags |= TNF_MIGRATE_FAIL;
4461 if (page_nid != NUMA_NO_NODE)
4462 task_numa_fault(last_cpupid, page_nid, 1, flags);
4466 #ifdef CONFIG_FINEGRAINED_THP
4467 static inline vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf)
4469 //struct timespec64 ts, te, diff;
4472 #ifdef CONFIG_FINEGRAINED_THP
4473 return VM_FAULT_FALLBACK;
4476 //ktime_get_ts64(&ts);
4477 ret = do_huge_pmd_anonymous_page(vmf);
4479 ktime_get_ts64(&te);
4480 diff = timespec64_sub(te, ts);
4481 if (!(ret & VM_FAULT_FALLBACK))
4482 pr_info("THP-fault: 2MB hugepage takes %lu nsecs\n",
4483 timespec64_to_ns(&diff));
4487 #endif /* CONFIG_FINEGRAINED_THP */
4489 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4491 if (vma_is_anonymous(vmf->vma))
4492 #ifdef CONFIG_FINEGRAINED_THP
4493 return __do_huge_pmd_anonymous_page(vmf);
4495 return do_huge_pmd_anonymous_page(vmf);
4497 if (vmf->vma->vm_ops->huge_fault)
4498 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4499 return VM_FAULT_FALLBACK;
4502 /* `inline' is required to avoid gcc 4.1.2 build error */
4503 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4505 if (vma_is_anonymous(vmf->vma)) {
4506 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4507 return handle_userfault(vmf, VM_UFFD_WP);
4508 return do_huge_pmd_wp_page(vmf, orig_pmd);
4510 if (vmf->vma->vm_ops->huge_fault) {
4511 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4513 if (!(ret & VM_FAULT_FALLBACK))
4517 /* COW or write-notify handled on pte level: split pmd. */
4518 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4520 return VM_FAULT_FALLBACK;
4523 #ifdef CONFIG_FINEGRAINED_THP
4524 vm_fault_t wp_huge_pte(struct vm_fault *vmf, pte_t orig_pte);
4525 #endif /* CONFIG_FINEGRAINED_THP */
4527 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4529 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4530 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4531 /* No support for anonymous transparent PUD pages yet */
4532 if (vma_is_anonymous(vmf->vma))
4534 if (vmf->vma->vm_ops->huge_fault) {
4535 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4537 if (!(ret & VM_FAULT_FALLBACK))
4541 /* COW or write-notify not handled on PUD level: split pud.*/
4542 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4543 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4544 return VM_FAULT_FALLBACK;
4547 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4549 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4550 /* No support for anonymous transparent PUD pages yet */
4551 if (vma_is_anonymous(vmf->vma))
4552 return VM_FAULT_FALLBACK;
4553 if (vmf->vma->vm_ops->huge_fault)
4554 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4555 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4556 return VM_FAULT_FALLBACK;
4560 * These routines also need to handle stuff like marking pages dirty
4561 * and/or accessed for architectures that don't do it in hardware (most
4562 * RISC architectures). The early dirtying is also good on the i386.
4564 * There is also a hook called "update_mmu_cache()" that architectures
4565 * with external mmu caches can use to update those (ie the Sparc or
4566 * PowerPC hashed page tables that act as extended TLBs).
4568 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4569 * concurrent faults).
4571 * The mmap_lock may have been released depending on flags and our return value.
4572 * See filemap_fault() and __lock_page_or_retry().
4574 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4578 if (unlikely(pmd_none(*vmf->pmd))) {
4580 * Leave __pte_alloc() until later: because vm_ops->fault may
4581 * want to allocate huge page, and if we expose page table
4582 * for an instant, it will be difficult to retract from
4583 * concurrent faults and from rmap lookups.
4587 /* See comment in pte_alloc_one_map() */
4588 if (pmd_devmap_trans_unstable(vmf->pmd))
4591 * A regular pmd is established and it can't morph into a huge
4592 * pmd from under us anymore at this point because we hold the
4593 * mmap_lock read mode and khugepaged takes it in write mode.
4594 * So now it's safe to run pte_offset_map().
4596 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4597 vmf->orig_pte = *vmf->pte;
4600 * some architectures can have larger ptes than wordsize,
4601 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4602 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4603 * accesses. The code below just needs a consistent view
4604 * for the ifs and we later double check anyway with the
4605 * ptl lock held. So here a barrier will do.
4608 if (pte_none(vmf->orig_pte)) {
4609 pte_unmap(vmf->pte);
4615 if (vma_is_anonymous(vmf->vma))
4616 return do_anonymous_page(vmf);
4618 return do_fault(vmf);
4621 if (!pte_present(vmf->orig_pte))
4622 return do_swap_page(vmf);
4624 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4625 return do_numa_page(vmf);
4627 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4628 spin_lock(vmf->ptl);
4629 entry = vmf->orig_pte;
4630 if (unlikely(!pte_same(*vmf->pte, entry))) {
4631 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4634 if (vmf->flags & FAULT_FLAG_WRITE) {
4635 if (!pte_write(entry)) {
4636 int ret = arch_do_wp_page(vmf, entry);
4638 if (!(ret & VM_FAULT_FALLBACK)) {
4640 * arch_do_wp_page returns
4641 * VM_FAULT value with spin lock acquisition.
4643 spin_unlock(vmf->ptl);
4646 return do_wp_page(vmf);
4648 if (arch_huge_pte_set_accessed(vmf, entry))
4650 entry = pte_mkdirty(entry);
4652 entry = pte_mkyoung(entry);
4653 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4654 vmf->flags & FAULT_FLAG_WRITE)) {
4655 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4657 /* Skip spurious TLB flush for retried page fault */
4658 if (vmf->flags & FAULT_FLAG_TRIED)
4661 * This is needed only for protection faults but the arch code
4662 * is not yet telling us if this is a protection fault or not.
4663 * This still avoids useless tlb flushes for .text page faults
4666 if (vmf->flags & FAULT_FLAG_WRITE)
4667 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4670 pte_unmap_unlock(vmf->pte, vmf->ptl);
4675 * By the time we get here, we already hold the mm semaphore
4677 * The mmap_lock may have been released depending on flags and our
4678 * return value. See filemap_fault() and __lock_page_or_retry().
4680 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4681 unsigned long address, unsigned int flags)
4683 struct vm_fault vmf = {
4685 .address = address & PAGE_MASK,
4687 .pgoff = linear_page_index(vma, address),
4688 .gfp_mask = __get_fault_gfp_mask(vma),
4690 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4691 struct mm_struct *mm = vma->vm_mm;
4696 pgd = pgd_offset(mm, address);
4697 p4d = p4d_alloc(mm, pgd, address);
4699 return VM_FAULT_OOM;
4701 vmf.pud = pud_alloc(mm, p4d, address);
4703 return VM_FAULT_OOM;
4705 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4706 ret = create_huge_pud(&vmf);
4707 if (!(ret & VM_FAULT_FALLBACK))
4710 pud_t orig_pud = *vmf.pud;
4713 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4715 /* NUMA case for anonymous PUDs would go here */
4717 if (dirty && !pud_write(orig_pud)) {
4718 ret = wp_huge_pud(&vmf, orig_pud);
4719 if (!(ret & VM_FAULT_FALLBACK))
4722 huge_pud_set_accessed(&vmf, orig_pud);
4728 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4730 return VM_FAULT_OOM;
4732 /* Huge pud page fault raced with pmd_alloc? */
4733 if (pud_trans_unstable(vmf.pud))
4736 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4737 ret = create_huge_pmd(&vmf);
4738 if (!(ret & VM_FAULT_FALLBACK))
4741 pmd_t orig_pmd = *vmf.pmd;
4744 if (unlikely(is_swap_pmd(orig_pmd))) {
4745 VM_BUG_ON(thp_migration_supported() &&
4746 !is_pmd_migration_entry(orig_pmd));
4747 if (is_pmd_migration_entry(orig_pmd))
4748 pmd_migration_entry_wait(mm, vmf.pmd);
4751 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4752 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4753 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4755 if (dirty && !pmd_write(orig_pmd)) {
4756 ret = wp_huge_pmd(&vmf, orig_pmd);
4757 if (!(ret & VM_FAULT_FALLBACK))
4760 huge_pmd_set_accessed(&vmf, orig_pmd);
4766 return handle_pte_fault(&vmf);
4770 * mm_account_fault - Do page fault accountings
4772 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4773 * of perf event counters, but we'll still do the per-task accounting to
4774 * the task who triggered this page fault.
4775 * @address: the faulted address.
4776 * @flags: the fault flags.
4777 * @ret: the fault retcode.
4779 * This will take care of most of the page fault accountings. Meanwhile, it
4780 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4781 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4782 * still be in per-arch page fault handlers at the entry of page fault.
4784 static inline void mm_account_fault(struct pt_regs *regs,
4785 unsigned long address, unsigned int flags,
4791 * We don't do accounting for some specific faults:
4793 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4794 * includes arch_vma_access_permitted() failing before reaching here.
4795 * So this is not a "this many hardware page faults" counter. We
4796 * should use the hw profiling for that.
4798 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4799 * once they're completed.
4801 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4805 * We define the fault as a major fault when the final successful fault
4806 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4807 * handle it immediately previously).
4809 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4817 * If the fault is done for GUP, regs will be NULL. We only do the
4818 * accounting for the per thread fault counters who triggered the
4819 * fault, and we skip the perf event updates.
4825 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4827 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4831 * By the time we get here, we already hold the mm semaphore
4833 * The mmap_lock may have been released depending on flags and our
4834 * return value. See filemap_fault() and __lock_page_or_retry().
4836 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4837 unsigned int flags, struct pt_regs *regs)
4841 __set_current_state(TASK_RUNNING);
4843 count_vm_event(PGFAULT);
4844 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4846 /* do counter updates before entering really critical section. */
4847 check_sync_rss_stat(current);
4849 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4850 flags & FAULT_FLAG_INSTRUCTION,
4851 flags & FAULT_FLAG_REMOTE))
4852 return VM_FAULT_SIGSEGV;
4855 * Enable the memcg OOM handling for faults triggered in user
4856 * space. Kernel faults are handled more gracefully.
4858 if (flags & FAULT_FLAG_USER)
4859 mem_cgroup_enter_user_fault();
4861 if (unlikely(is_vm_hugetlb_page(vma)))
4862 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4864 ret = __handle_mm_fault(vma, address, flags);
4866 if (flags & FAULT_FLAG_USER) {
4867 mem_cgroup_exit_user_fault();
4869 * The task may have entered a memcg OOM situation but
4870 * if the allocation error was handled gracefully (no
4871 * VM_FAULT_OOM), there is no need to kill anything.
4872 * Just clean up the OOM state peacefully.
4874 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4875 mem_cgroup_oom_synchronize(false);
4878 mm_account_fault(regs, address, flags, ret);
4882 EXPORT_SYMBOL_GPL(handle_mm_fault);
4884 #ifndef __PAGETABLE_P4D_FOLDED
4886 * Allocate p4d page table.
4887 * We've already handled the fast-path in-line.
4889 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4891 p4d_t *new = p4d_alloc_one(mm, address);
4895 smp_wmb(); /* See comment in __pte_alloc */
4897 spin_lock(&mm->page_table_lock);
4898 if (pgd_present(*pgd)) /* Another has populated it */
4901 pgd_populate(mm, pgd, new);
4902 spin_unlock(&mm->page_table_lock);
4905 #endif /* __PAGETABLE_P4D_FOLDED */
4907 #ifndef __PAGETABLE_PUD_FOLDED
4909 * Allocate page upper directory.
4910 * We've already handled the fast-path in-line.
4912 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4914 pud_t *new = pud_alloc_one(mm, address);
4918 smp_wmb(); /* See comment in __pte_alloc */
4920 spin_lock(&mm->page_table_lock);
4921 if (!p4d_present(*p4d)) {
4923 p4d_populate(mm, p4d, new);
4924 } else /* Another has populated it */
4926 spin_unlock(&mm->page_table_lock);
4929 #endif /* __PAGETABLE_PUD_FOLDED */
4931 #ifndef __PAGETABLE_PMD_FOLDED
4933 * Allocate page middle directory.
4934 * We've already handled the fast-path in-line.
4936 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4939 pmd_t *new = pmd_alloc_one(mm, address);
4943 smp_wmb(); /* See comment in __pte_alloc */
4945 ptl = pud_lock(mm, pud);
4946 if (!pud_present(*pud)) {
4948 pud_populate(mm, pud, new);
4949 } else /* Another has populated it */
4954 #endif /* __PAGETABLE_PMD_FOLDED */
4956 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4957 struct mmu_notifier_range *range, pte_t **ptepp,
4958 pmd_t **pmdpp, spinlock_t **ptlp)
4966 pgd = pgd_offset(mm, address);
4967 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4970 p4d = p4d_offset(pgd, address);
4971 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4974 pud = pud_offset(p4d, address);
4975 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4978 pmd = pmd_offset(pud, address);
4979 VM_BUG_ON(pmd_trans_huge(*pmd));
4981 if (pmd_huge(*pmd)) {
4986 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4987 NULL, mm, address & PMD_MASK,
4988 (address & PMD_MASK) + PMD_SIZE);
4989 mmu_notifier_invalidate_range_start(range);
4991 *ptlp = pmd_lock(mm, pmd);
4992 if (pmd_huge(*pmd)) {
4998 mmu_notifier_invalidate_range_end(range);
5001 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
5005 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
5006 address & PAGE_MASK,
5007 (address & PAGE_MASK) + PAGE_SIZE);
5008 mmu_notifier_invalidate_range_start(range);
5010 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
5011 if (!pte_present(*ptep))
5016 pte_unmap_unlock(ptep, *ptlp);
5018 mmu_notifier_invalidate_range_end(range);
5024 * follow_pte - look up PTE at a user virtual address
5025 * @mm: the mm_struct of the target address space
5026 * @address: user virtual address
5027 * @ptepp: location to store found PTE
5028 * @ptlp: location to store the lock for the PTE
5030 * On a successful return, the pointer to the PTE is stored in @ptepp;
5031 * the corresponding lock is taken and its location is stored in @ptlp.
5032 * The contents of the PTE are only stable until @ptlp is released;
5033 * any further use, if any, must be protected against invalidation
5034 * with MMU notifiers.
5036 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5037 * should be taken for read.
5039 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5040 * it is not a good general-purpose API.
5042 * Return: zero on success, -ve otherwise.
5044 int follow_pte(struct mm_struct *mm, unsigned long address,
5045 pte_t **ptepp, spinlock_t **ptlp)
5047 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
5049 EXPORT_SYMBOL_GPL(follow_pte);
5052 * follow_pfn - look up PFN at a user virtual address
5053 * @vma: memory mapping
5054 * @address: user virtual address
5055 * @pfn: location to store found PFN
5057 * Only IO mappings and raw PFN mappings are allowed.
5059 * This function does not allow the caller to read the permissions
5060 * of the PTE. Do not use it.
5062 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5064 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5071 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5074 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5077 *pfn = pte_pfn(*ptep);
5078 pte_unmap_unlock(ptep, ptl);
5081 EXPORT_SYMBOL(follow_pfn);
5083 #ifdef CONFIG_HAVE_IOREMAP_PROT
5084 int follow_phys(struct vm_area_struct *vma,
5085 unsigned long address, unsigned int flags,
5086 unsigned long *prot, resource_size_t *phys)
5092 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5095 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5099 if ((flags & FOLL_WRITE) && !pte_write(pte))
5102 *prot = pgprot_val(pte_pgprot(pte));
5103 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5107 pte_unmap_unlock(ptep, ptl);
5112 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5113 void *buf, int len, int write)
5115 resource_size_t phys_addr;
5116 unsigned long prot = 0;
5117 void __iomem *maddr;
5118 int offset = addr & (PAGE_SIZE-1);
5120 if (follow_phys(vma, addr, write, &prot, &phys_addr))
5123 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5128 memcpy_toio(maddr + offset, buf, len);
5130 memcpy_fromio(buf, maddr + offset, len);
5135 EXPORT_SYMBOL_GPL(generic_access_phys);
5139 * Access another process' address space as given in mm. If non-NULL, use the
5140 * given task for page fault accounting.
5142 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
5143 unsigned long addr, void *buf, int len, unsigned int gup_flags)
5145 struct vm_area_struct *vma;
5146 void *old_buf = buf;
5147 int write = gup_flags & FOLL_WRITE;
5149 if (mmap_read_lock_killable(mm))
5152 /* ignore errors, just check how much was successfully transferred */
5154 int bytes, ret, offset;
5156 struct page *page = NULL;
5158 ret = get_user_pages_remote(mm, addr, 1,
5159 gup_flags, &page, &vma, NULL);
5161 #ifndef CONFIG_HAVE_IOREMAP_PROT
5165 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5166 * we can access using slightly different code.
5168 vma = find_vma(mm, addr);
5169 if (!vma || vma->vm_start > addr)
5171 if (vma->vm_ops && vma->vm_ops->access)
5172 ret = vma->vm_ops->access(vma, addr, buf,
5180 offset = addr & (PAGE_SIZE-1);
5181 if (bytes > PAGE_SIZE-offset)
5182 bytes = PAGE_SIZE-offset;
5186 copy_to_user_page(vma, page, addr,
5187 maddr + offset, buf, bytes);
5188 set_page_dirty_lock(page);
5190 copy_from_user_page(vma, page, addr,
5191 buf, maddr + offset, bytes);
5200 mmap_read_unlock(mm);
5202 return buf - old_buf;
5206 * access_remote_vm - access another process' address space
5207 * @mm: the mm_struct of the target address space
5208 * @addr: start address to access
5209 * @buf: source or destination buffer
5210 * @len: number of bytes to transfer
5211 * @gup_flags: flags modifying lookup behaviour
5213 * The caller must hold a reference on @mm.
5215 * Return: number of bytes copied from source to destination.
5217 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5218 void *buf, int len, unsigned int gup_flags)
5220 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
5224 * Access another process' address space.
5225 * Source/target buffer must be kernel space,
5226 * Do not walk the page table directly, use get_user_pages
5228 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5229 void *buf, int len, unsigned int gup_flags)
5231 struct mm_struct *mm;
5234 mm = get_task_mm(tsk);
5238 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
5244 EXPORT_SYMBOL_GPL(access_process_vm);
5247 * Print the name of a VMA.
5249 void print_vma_addr(char *prefix, unsigned long ip)
5251 struct mm_struct *mm = current->mm;
5252 struct vm_area_struct *vma;
5255 * we might be running from an atomic context so we cannot sleep
5257 if (!mmap_read_trylock(mm))
5260 vma = find_vma(mm, ip);
5261 if (vma && vma->vm_file) {
5262 struct file *f = vma->vm_file;
5263 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5267 p = file_path(f, buf, PAGE_SIZE);
5270 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5272 vma->vm_end - vma->vm_start);
5273 free_page((unsigned long)buf);
5276 mmap_read_unlock(mm);
5279 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5280 void __might_fault(const char *file, int line)
5283 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5284 * holding the mmap_lock, this is safe because kernel memory doesn't
5285 * get paged out, therefore we'll never actually fault, and the
5286 * below annotations will generate false positives.
5288 if (uaccess_kernel())
5290 if (pagefault_disabled())
5292 __might_sleep(file, line, 0);
5293 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5295 might_lock_read(¤t->mm->mmap_lock);
5298 EXPORT_SYMBOL(__might_fault);
5301 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5303 * Process all subpages of the specified huge page with the specified
5304 * operation. The target subpage will be processed last to keep its
5307 static inline void process_huge_page(
5308 unsigned long addr_hint, unsigned int pages_per_huge_page,
5309 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5313 unsigned long addr = addr_hint &
5314 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5316 /* Process target subpage last to keep its cache lines hot */
5318 n = (addr_hint - addr) / PAGE_SIZE;
5319 if (2 * n <= pages_per_huge_page) {
5320 /* If target subpage in first half of huge page */
5323 /* Process subpages at the end of huge page */
5324 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5326 process_subpage(addr + i * PAGE_SIZE, i, arg);
5329 /* If target subpage in second half of huge page */
5330 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5331 l = pages_per_huge_page - n;
5332 /* Process subpages at the begin of huge page */
5333 for (i = 0; i < base; i++) {
5335 process_subpage(addr + i * PAGE_SIZE, i, arg);
5339 * Process remaining subpages in left-right-left-right pattern
5340 * towards the target subpage
5342 for (i = 0; i < l; i++) {
5343 int left_idx = base + i;
5344 int right_idx = base + 2 * l - 1 - i;
5347 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5349 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5353 static void clear_gigantic_page(struct page *page,
5355 unsigned int pages_per_huge_page)
5358 struct page *p = page;
5361 for (i = 0; i < pages_per_huge_page;
5362 i++, p = mem_map_next(p, page, i)) {
5364 clear_user_highpage(p, addr + i * PAGE_SIZE);
5368 static void clear_subpage(unsigned long addr, int idx, void *arg)
5370 struct page *page = arg;
5372 clear_user_highpage(page + idx, addr);
5375 void clear_huge_page(struct page *page,
5376 unsigned long addr_hint, unsigned int pages_per_huge_page)
5378 unsigned long addr = addr_hint &
5379 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5381 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5382 clear_gigantic_page(page, addr, pages_per_huge_page);
5386 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5389 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5391 struct vm_area_struct *vma,
5392 unsigned int pages_per_huge_page)
5395 struct page *dst_base = dst;
5396 struct page *src_base = src;
5398 for (i = 0; i < pages_per_huge_page; ) {
5400 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5403 dst = mem_map_next(dst, dst_base, i);
5404 src = mem_map_next(src, src_base, i);
5408 struct copy_subpage_arg {
5411 struct vm_area_struct *vma;
5414 static void copy_subpage(unsigned long addr, int idx, void *arg)
5416 struct copy_subpage_arg *copy_arg = arg;
5418 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5419 addr, copy_arg->vma);
5422 void copy_user_huge_page(struct page *dst, struct page *src,
5423 unsigned long addr_hint, struct vm_area_struct *vma,
5424 unsigned int pages_per_huge_page)
5426 unsigned long addr = addr_hint &
5427 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5428 struct copy_subpage_arg arg = {
5434 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5435 copy_user_gigantic_page(dst, src, addr, vma,
5436 pages_per_huge_page);
5440 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5443 long copy_huge_page_from_user(struct page *dst_page,
5444 const void __user *usr_src,
5445 unsigned int pages_per_huge_page,
5446 bool allow_pagefault)
5448 void *src = (void *)usr_src;
5450 unsigned long i, rc = 0;
5451 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5452 struct page *subpage = dst_page;
5454 for (i = 0; i < pages_per_huge_page;
5455 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5456 if (allow_pagefault)
5457 page_kaddr = kmap(subpage);
5459 page_kaddr = kmap_atomic(subpage);
5460 rc = copy_from_user(page_kaddr,
5461 (const void __user *)(src + i * PAGE_SIZE),
5463 if (allow_pagefault)
5466 kunmap_atomic(page_kaddr);
5468 ret_val -= (PAGE_SIZE - rc);
5476 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5478 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5480 static struct kmem_cache *page_ptl_cachep;
5482 void __init ptlock_cache_init(void)
5484 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5488 bool ptlock_alloc(struct page *page)
5492 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5499 void ptlock_free(struct page *page)
5501 kmem_cache_free(page_ptl_cachep, page->ptl);