1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
73 #include <linux/perf_event.h>
74 #include <linux/ptrace.h>
75 #include <linux/vmalloc.h>
77 #include <trace/events/kmem.h>
80 #include <asm/mmu_context.h>
81 #include <asm/pgalloc.h>
82 #include <linux/uaccess.h>
84 #include <asm/tlbflush.h>
86 #include "pgalloc-track.h"
89 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
90 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
94 unsigned long max_mapnr;
95 EXPORT_SYMBOL(max_mapnr);
98 EXPORT_SYMBOL(mem_map);
102 * A number of key systems in x86 including ioremap() rely on the assumption
103 * that high_memory defines the upper bound on direct map memory, then end
104 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
105 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
109 EXPORT_SYMBOL(high_memory);
112 * Randomize the address space (stacks, mmaps, brk, etc.).
114 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
115 * as ancient (libc5 based) binaries can segfault. )
117 int randomize_va_space __read_mostly =
118 #ifdef CONFIG_COMPAT_BRK
124 #ifndef arch_faults_on_old_pte
125 static inline bool arch_faults_on_old_pte(void)
128 * Those arches which don't have hw access flag feature need to
129 * implement their own helper. By default, "true" means pagefault
130 * will be hit on old pte.
136 #ifndef arch_wants_old_prefaulted_pte
137 static inline bool arch_wants_old_prefaulted_pte(void)
140 * Transitioning a PTE from 'old' to 'young' can be expensive on
141 * some architectures, even if it's performed in hardware. By
142 * default, "false" means prefaulted entries will be 'young'.
148 static int __init disable_randmaps(char *s)
150 randomize_va_space = 0;
153 __setup("norandmaps", disable_randmaps);
155 unsigned long zero_pfn __read_mostly;
156 EXPORT_SYMBOL(zero_pfn);
158 unsigned long highest_memmap_pfn __read_mostly;
161 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
163 static int __init init_zero_pfn(void)
165 zero_pfn = page_to_pfn(ZERO_PAGE(0));
168 early_initcall(init_zero_pfn);
170 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
172 trace_rss_stat(mm, member, count);
175 #if defined(SPLIT_RSS_COUNTING)
177 void sync_mm_rss(struct mm_struct *mm)
181 for (i = 0; i < NR_MM_COUNTERS; i++) {
182 if (current->rss_stat.count[i]) {
183 add_mm_counter(mm, i, current->rss_stat.count[i]);
184 current->rss_stat.count[i] = 0;
187 current->rss_stat.events = 0;
190 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
192 struct task_struct *task = current;
194 if (likely(task->mm == mm))
195 task->rss_stat.count[member] += val;
197 add_mm_counter(mm, member, val);
199 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
200 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
202 /* sync counter once per 64 page faults */
203 #define TASK_RSS_EVENTS_THRESH (64)
204 static void check_sync_rss_stat(struct task_struct *task)
206 if (unlikely(task != current))
208 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
209 sync_mm_rss(task->mm);
211 #else /* SPLIT_RSS_COUNTING */
213 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
214 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
216 static void check_sync_rss_stat(struct task_struct *task)
220 #endif /* SPLIT_RSS_COUNTING */
223 * Note: this doesn't free the actual pages themselves. That
224 * has been handled earlier when unmapping all the memory regions.
226 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
229 pgtable_t token = pmd_pgtable(*pmd);
231 pte_free_tlb(tlb, token, addr);
232 mm_dec_nr_ptes(tlb->mm);
235 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
236 unsigned long addr, unsigned long end,
237 unsigned long floor, unsigned long ceiling)
244 pmd = pmd_offset(pud, addr);
246 next = pmd_addr_end(addr, end);
247 if (pmd_none_or_clear_bad(pmd))
249 free_pte_range(tlb, pmd, addr);
250 } while (pmd++, addr = next, addr != end);
260 if (end - 1 > ceiling - 1)
263 pmd = pmd_offset(pud, start);
265 pmd_free_tlb(tlb, pmd, start);
266 mm_dec_nr_pmds(tlb->mm);
269 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
270 unsigned long addr, unsigned long end,
271 unsigned long floor, unsigned long ceiling)
278 pud = pud_offset(p4d, addr);
280 next = pud_addr_end(addr, end);
281 if (pud_none_or_clear_bad(pud))
283 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
284 } while (pud++, addr = next, addr != end);
294 if (end - 1 > ceiling - 1)
297 pud = pud_offset(p4d, start);
299 pud_free_tlb(tlb, pud, start);
300 mm_dec_nr_puds(tlb->mm);
303 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
304 unsigned long addr, unsigned long end,
305 unsigned long floor, unsigned long ceiling)
312 p4d = p4d_offset(pgd, addr);
314 next = p4d_addr_end(addr, end);
315 if (p4d_none_or_clear_bad(p4d))
317 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
318 } while (p4d++, addr = next, addr != end);
324 ceiling &= PGDIR_MASK;
328 if (end - 1 > ceiling - 1)
331 p4d = p4d_offset(pgd, start);
333 p4d_free_tlb(tlb, p4d, start);
337 * This function frees user-level page tables of a process.
339 void free_pgd_range(struct mmu_gather *tlb,
340 unsigned long addr, unsigned long end,
341 unsigned long floor, unsigned long ceiling)
347 * The next few lines have given us lots of grief...
349 * Why are we testing PMD* at this top level? Because often
350 * there will be no work to do at all, and we'd prefer not to
351 * go all the way down to the bottom just to discover that.
353 * Why all these "- 1"s? Because 0 represents both the bottom
354 * of the address space and the top of it (using -1 for the
355 * top wouldn't help much: the masks would do the wrong thing).
356 * The rule is that addr 0 and floor 0 refer to the bottom of
357 * the address space, but end 0 and ceiling 0 refer to the top
358 * Comparisons need to use "end - 1" and "ceiling - 1" (though
359 * that end 0 case should be mythical).
361 * Wherever addr is brought up or ceiling brought down, we must
362 * be careful to reject "the opposite 0" before it confuses the
363 * subsequent tests. But what about where end is brought down
364 * by PMD_SIZE below? no, end can't go down to 0 there.
366 * Whereas we round start (addr) and ceiling down, by different
367 * masks at different levels, in order to test whether a table
368 * now has no other vmas using it, so can be freed, we don't
369 * bother to round floor or end up - the tests don't need that.
383 if (end - 1 > ceiling - 1)
388 * We add page table cache pages with PAGE_SIZE,
389 * (see pte_free_tlb()), flush the tlb if we need
391 tlb_change_page_size(tlb, PAGE_SIZE);
392 pgd = pgd_offset(tlb->mm, addr);
394 next = pgd_addr_end(addr, end);
395 if (pgd_none_or_clear_bad(pgd))
397 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
398 } while (pgd++, addr = next, addr != end);
401 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
402 unsigned long floor, unsigned long ceiling)
405 struct vm_area_struct *next = vma->vm_next;
406 unsigned long addr = vma->vm_start;
409 * Hide vma from rmap and truncate_pagecache before freeing
412 unlink_anon_vmas(vma);
413 unlink_file_vma(vma);
415 if (is_vm_hugetlb_page(vma)) {
416 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
417 floor, next ? next->vm_start : ceiling);
420 * Optimization: gather nearby vmas into one call down
422 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
423 && !is_vm_hugetlb_page(next)) {
426 unlink_anon_vmas(vma);
427 unlink_file_vma(vma);
429 free_pgd_range(tlb, addr, vma->vm_end,
430 floor, next ? next->vm_start : ceiling);
436 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
439 pgtable_t new = pte_alloc_one(mm);
444 * Ensure all pte setup (eg. pte page lock and page clearing) are
445 * visible before the pte is made visible to other CPUs by being
446 * put into page tables.
448 * The other side of the story is the pointer chasing in the page
449 * table walking code (when walking the page table without locking;
450 * ie. most of the time). Fortunately, these data accesses consist
451 * of a chain of data-dependent loads, meaning most CPUs (alpha
452 * being the notable exception) will already guarantee loads are
453 * seen in-order. See the alpha page table accessors for the
454 * smp_rmb() barriers in page table walking code.
456 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
458 ptl = pmd_lock(mm, pmd);
459 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
461 pmd_populate(mm, pmd, new);
470 int __pte_alloc_kernel(pmd_t *pmd)
472 pte_t *new = pte_alloc_one_kernel(&init_mm);
476 smp_wmb(); /* See comment in __pte_alloc */
478 spin_lock(&init_mm.page_table_lock);
479 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
480 pmd_populate_kernel(&init_mm, pmd, new);
483 spin_unlock(&init_mm.page_table_lock);
485 pte_free_kernel(&init_mm, new);
489 static inline void init_rss_vec(int *rss)
491 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
494 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
498 if (current->mm == mm)
500 for (i = 0; i < NR_MM_COUNTERS; i++)
502 add_mm_counter(mm, i, rss[i]);
506 * This function is called to print an error when a bad pte
507 * is found. For example, we might have a PFN-mapped pte in
508 * a region that doesn't allow it.
510 * The calling function must still handle the error.
512 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
513 pte_t pte, struct page *page)
515 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
516 p4d_t *p4d = p4d_offset(pgd, addr);
517 pud_t *pud = pud_offset(p4d, addr);
518 pmd_t *pmd = pmd_offset(pud, addr);
519 struct address_space *mapping;
521 static unsigned long resume;
522 static unsigned long nr_shown;
523 static unsigned long nr_unshown;
526 * Allow a burst of 60 reports, then keep quiet for that minute;
527 * or allow a steady drip of one report per second.
529 if (nr_shown == 60) {
530 if (time_before(jiffies, resume)) {
535 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
542 resume = jiffies + 60 * HZ;
544 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
545 index = linear_page_index(vma, addr);
547 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
549 (long long)pte_val(pte), (long long)pmd_val(*pmd));
551 dump_page(page, "bad pte");
552 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
553 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
554 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
556 vma->vm_ops ? vma->vm_ops->fault : NULL,
557 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
558 mapping ? mapping->a_ops->readpage : NULL);
560 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
564 * vm_normal_page -- This function gets the "struct page" associated with a pte.
566 * "Special" mappings do not wish to be associated with a "struct page" (either
567 * it doesn't exist, or it exists but they don't want to touch it). In this
568 * case, NULL is returned here. "Normal" mappings do have a struct page.
570 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
571 * pte bit, in which case this function is trivial. Secondly, an architecture
572 * may not have a spare pte bit, which requires a more complicated scheme,
575 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
576 * special mapping (even if there are underlying and valid "struct pages").
577 * COWed pages of a VM_PFNMAP are always normal.
579 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
580 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
581 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
582 * mapping will always honor the rule
584 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
586 * And for normal mappings this is false.
588 * This restricts such mappings to be a linear translation from virtual address
589 * to pfn. To get around this restriction, we allow arbitrary mappings so long
590 * as the vma is not a COW mapping; in that case, we know that all ptes are
591 * special (because none can have been COWed).
594 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
596 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
597 * page" backing, however the difference is that _all_ pages with a struct
598 * page (that is, those where pfn_valid is true) are refcounted and considered
599 * normal pages by the VM. The disadvantage is that pages are refcounted
600 * (which can be slower and simply not an option for some PFNMAP users). The
601 * advantage is that we don't have to follow the strict linearity rule of
602 * PFNMAP mappings in order to support COWable mappings.
605 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
608 unsigned long pfn = pte_pfn(pte);
610 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
611 if (likely(!pte_special(pte)))
613 if (vma->vm_ops && vma->vm_ops->find_special_page)
614 return vma->vm_ops->find_special_page(vma, addr);
615 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
617 if (is_zero_pfn(pfn))
622 print_bad_pte(vma, addr, pte, NULL);
626 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
628 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
629 if (vma->vm_flags & VM_MIXEDMAP) {
635 off = (addr - vma->vm_start) >> PAGE_SHIFT;
636 if (pfn == vma->vm_pgoff + off)
638 if (!is_cow_mapping(vma->vm_flags))
643 if (is_zero_pfn(pfn))
647 if (unlikely(pfn > highest_memmap_pfn)) {
648 print_bad_pte(vma, addr, pte, NULL);
653 * NOTE! We still have PageReserved() pages in the page tables.
654 * eg. VDSO mappings can cause them to exist.
657 return pfn_to_page(pfn);
660 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
661 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
664 unsigned long pfn = pmd_pfn(pmd);
667 * There is no pmd_special() but there may be special pmds, e.g.
668 * in a direct-access (dax) mapping, so let's just replicate the
669 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
671 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
672 if (vma->vm_flags & VM_MIXEDMAP) {
678 off = (addr - vma->vm_start) >> PAGE_SHIFT;
679 if (pfn == vma->vm_pgoff + off)
681 if (!is_cow_mapping(vma->vm_flags))
688 if (is_huge_zero_pmd(pmd))
690 if (unlikely(pfn > highest_memmap_pfn))
694 * NOTE! We still have PageReserved() pages in the page tables.
695 * eg. VDSO mappings can cause them to exist.
698 return pfn_to_page(pfn);
702 static void restore_exclusive_pte(struct vm_area_struct *vma,
703 struct page *page, unsigned long address,
709 pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
710 if (pte_swp_soft_dirty(*ptep))
711 pte = pte_mksoft_dirty(pte);
713 entry = pte_to_swp_entry(*ptep);
714 if (pte_swp_uffd_wp(*ptep))
715 pte = pte_mkuffd_wp(pte);
716 else if (is_writable_device_exclusive_entry(entry))
717 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
719 set_pte_at(vma->vm_mm, address, ptep, pte);
722 * No need to take a page reference as one was already
723 * created when the swap entry was made.
726 page_add_anon_rmap(page, vma, address, false);
729 * Currently device exclusive access only supports anonymous
730 * memory so the entry shouldn't point to a filebacked page.
732 WARN_ON_ONCE(!PageAnon(page));
734 if (vma->vm_flags & VM_LOCKED)
735 mlock_vma_page(page);
738 * No need to invalidate - it was non-present before. However
739 * secondary CPUs may have mappings that need invalidating.
741 update_mmu_cache(vma, address, ptep);
745 * Tries to restore an exclusive pte if the page lock can be acquired without
749 try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
752 swp_entry_t entry = pte_to_swp_entry(*src_pte);
753 struct page *page = pfn_swap_entry_to_page(entry);
755 if (trylock_page(page)) {
756 restore_exclusive_pte(vma, page, addr, src_pte);
765 * copy one vm_area from one task to the other. Assumes the page tables
766 * already present in the new task to be cleared in the whole range
767 * covered by this vma.
771 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
772 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
773 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
775 unsigned long vm_flags = dst_vma->vm_flags;
776 pte_t pte = *src_pte;
778 swp_entry_t entry = pte_to_swp_entry(pte);
780 if (likely(!non_swap_entry(entry))) {
781 if (swap_duplicate(entry) < 0)
784 /* make sure dst_mm is on swapoff's mmlist. */
785 if (unlikely(list_empty(&dst_mm->mmlist))) {
786 spin_lock(&mmlist_lock);
787 if (list_empty(&dst_mm->mmlist))
788 list_add(&dst_mm->mmlist,
790 spin_unlock(&mmlist_lock);
793 } else if (is_migration_entry(entry)) {
794 page = pfn_swap_entry_to_page(entry);
796 rss[mm_counter(page)]++;
798 if (is_writable_migration_entry(entry) &&
799 is_cow_mapping(vm_flags)) {
801 * COW mappings require pages in both
802 * parent and child to be set to read.
804 entry = make_readable_migration_entry(
806 pte = swp_entry_to_pte(entry);
807 if (pte_swp_soft_dirty(*src_pte))
808 pte = pte_swp_mksoft_dirty(pte);
809 if (pte_swp_uffd_wp(*src_pte))
810 pte = pte_swp_mkuffd_wp(pte);
811 set_pte_at(src_mm, addr, src_pte, pte);
813 } else if (is_device_private_entry(entry)) {
814 page = pfn_swap_entry_to_page(entry);
817 * Update rss count even for unaddressable pages, as
818 * they should treated just like normal pages in this
821 * We will likely want to have some new rss counters
822 * for unaddressable pages, at some point. But for now
823 * keep things as they are.
826 rss[mm_counter(page)]++;
827 page_dup_rmap(page, false);
830 * We do not preserve soft-dirty information, because so
831 * far, checkpoint/restore is the only feature that
832 * requires that. And checkpoint/restore does not work
833 * when a device driver is involved (you cannot easily
834 * save and restore device driver state).
836 if (is_writable_device_private_entry(entry) &&
837 is_cow_mapping(vm_flags)) {
838 entry = make_readable_device_private_entry(
840 pte = swp_entry_to_pte(entry);
841 if (pte_swp_uffd_wp(*src_pte))
842 pte = pte_swp_mkuffd_wp(pte);
843 set_pte_at(src_mm, addr, src_pte, pte);
845 } else if (is_device_exclusive_entry(entry)) {
847 * Make device exclusive entries present by restoring the
848 * original entry then copying as for a present pte. Device
849 * exclusive entries currently only support private writable
850 * (ie. COW) mappings.
852 VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
853 if (try_restore_exclusive_pte(src_pte, src_vma, addr))
857 if (!userfaultfd_wp(dst_vma))
858 pte = pte_swp_clear_uffd_wp(pte);
859 set_pte_at(dst_mm, addr, dst_pte, pte);
864 * Copy a present and normal page if necessary.
866 * NOTE! The usual case is that this doesn't need to do
867 * anything, and can just return a positive value. That
868 * will let the caller know that it can just increase
869 * the page refcount and re-use the pte the traditional
872 * But _if_ we need to copy it because it needs to be
873 * pinned in the parent (and the child should get its own
874 * copy rather than just a reference to the same page),
875 * we'll do that here and return zero to let the caller
878 * And if we need a pre-allocated page but don't yet have
879 * one, return a negative error to let the preallocation
880 * code know so that it can do so outside the page table
884 copy_present_page(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, pte_t pte, struct page *page)
888 struct page *new_page;
891 * What we want to do is to check whether this page may
892 * have been pinned by the parent process. If so,
893 * instead of wrprotect the pte on both sides, we copy
894 * the page immediately so that we'll always guarantee
895 * the pinned page won't be randomly replaced in the
898 * The page pinning checks are just "has this mm ever
899 * seen pinning", along with the (inexact) check of
900 * the page count. That might give false positives for
901 * for pinning, but it will work correctly.
903 if (likely(!page_needs_cow_for_dma(src_vma, page)))
906 new_page = *prealloc;
911 * We have a prealloc page, all good! Take it
912 * over and copy the page & arm it.
915 copy_user_highpage(new_page, page, addr, src_vma);
916 __SetPageUptodate(new_page);
917 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
918 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
919 rss[mm_counter(new_page)]++;
921 /* All done, just insert the new page copy in the child */
922 pte = mk_pte(new_page, dst_vma->vm_page_prot);
923 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
924 if (userfaultfd_pte_wp(dst_vma, *src_pte))
925 /* Uffd-wp needs to be delivered to dest pte as well */
926 pte = pte_wrprotect(pte_mkuffd_wp(pte));
927 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
932 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
933 * is required to copy this pte.
936 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
937 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
938 struct page **prealloc)
940 struct mm_struct *src_mm = src_vma->vm_mm;
941 unsigned long vm_flags = src_vma->vm_flags;
942 pte_t pte = *src_pte;
945 page = vm_normal_page(src_vma, addr, pte);
949 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
950 addr, rss, prealloc, pte, page);
955 page_dup_rmap(page, false);
956 rss[mm_counter(page)]++;
960 * If it's a COW mapping, write protect it both
961 * in the parent and the child
963 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
964 ptep_set_wrprotect(src_mm, addr, src_pte);
965 pte = pte_wrprotect(pte);
969 * If it's a shared mapping, mark it clean in
972 if (vm_flags & VM_SHARED)
973 pte = pte_mkclean(pte);
974 pte = pte_mkold(pte);
976 if (!userfaultfd_wp(dst_vma))
977 pte = pte_clear_uffd_wp(pte);
979 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
983 static inline struct page *
984 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
987 struct page *new_page;
989 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
993 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
997 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
1003 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1004 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1007 struct mm_struct *dst_mm = dst_vma->vm_mm;
1008 struct mm_struct *src_mm = src_vma->vm_mm;
1009 pte_t *orig_src_pte, *orig_dst_pte;
1010 pte_t *src_pte, *dst_pte;
1011 spinlock_t *src_ptl, *dst_ptl;
1012 int progress, ret = 0;
1013 int rss[NR_MM_COUNTERS];
1014 swp_entry_t entry = (swp_entry_t){0};
1015 struct page *prealloc = NULL;
1021 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1026 src_pte = pte_offset_map(src_pmd, addr);
1027 src_ptl = pte_lockptr(src_mm, src_pmd);
1028 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1029 orig_src_pte = src_pte;
1030 orig_dst_pte = dst_pte;
1031 arch_enter_lazy_mmu_mode();
1035 * We are holding two locks at this point - either of them
1036 * could generate latencies in another task on another CPU.
1038 if (progress >= 32) {
1040 if (need_resched() ||
1041 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1044 if (pte_none(*src_pte)) {
1048 if (unlikely(!pte_present(*src_pte))) {
1049 ret = copy_nonpresent_pte(dst_mm, src_mm,
1054 entry = pte_to_swp_entry(*src_pte);
1056 } else if (ret == -EBUSY) {
1064 * Device exclusive entry restored, continue by copying
1065 * the now present pte.
1067 WARN_ON_ONCE(ret != -ENOENT);
1069 /* copy_present_pte() will clear `*prealloc' if consumed */
1070 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1071 addr, rss, &prealloc);
1073 * If we need a pre-allocated page for this pte, drop the
1074 * locks, allocate, and try again.
1076 if (unlikely(ret == -EAGAIN))
1078 if (unlikely(prealloc)) {
1080 * pre-alloc page cannot be reused by next time so as
1081 * to strictly follow mempolicy (e.g., alloc_page_vma()
1082 * will allocate page according to address). This
1083 * could only happen if one pinned pte changed.
1089 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1091 arch_leave_lazy_mmu_mode();
1092 spin_unlock(src_ptl);
1093 pte_unmap(orig_src_pte);
1094 add_mm_rss_vec(dst_mm, rss);
1095 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1099 VM_WARN_ON_ONCE(!entry.val);
1100 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1105 } else if (ret == -EBUSY) {
1107 } else if (ret == -EAGAIN) {
1108 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1115 /* We've captured and resolved the error. Reset, try again. */
1121 if (unlikely(prealloc))
1127 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1128 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1131 struct mm_struct *dst_mm = dst_vma->vm_mm;
1132 struct mm_struct *src_mm = src_vma->vm_mm;
1133 pmd_t *src_pmd, *dst_pmd;
1136 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1139 src_pmd = pmd_offset(src_pud, addr);
1141 next = pmd_addr_end(addr, end);
1142 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1143 || pmd_devmap(*src_pmd)) {
1145 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1146 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1147 addr, dst_vma, src_vma);
1154 if (pmd_none_or_clear_bad(src_pmd))
1156 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1159 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1164 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1165 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1168 struct mm_struct *dst_mm = dst_vma->vm_mm;
1169 struct mm_struct *src_mm = src_vma->vm_mm;
1170 pud_t *src_pud, *dst_pud;
1173 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1176 src_pud = pud_offset(src_p4d, addr);
1178 next = pud_addr_end(addr, end);
1179 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1182 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1183 err = copy_huge_pud(dst_mm, src_mm,
1184 dst_pud, src_pud, addr, src_vma);
1191 if (pud_none_or_clear_bad(src_pud))
1193 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1196 } while (dst_pud++, src_pud++, addr = next, addr != end);
1201 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1202 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1205 struct mm_struct *dst_mm = dst_vma->vm_mm;
1206 p4d_t *src_p4d, *dst_p4d;
1209 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1212 src_p4d = p4d_offset(src_pgd, addr);
1214 next = p4d_addr_end(addr, end);
1215 if (p4d_none_or_clear_bad(src_p4d))
1217 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1220 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1225 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1227 pgd_t *src_pgd, *dst_pgd;
1229 unsigned long addr = src_vma->vm_start;
1230 unsigned long end = src_vma->vm_end;
1231 struct mm_struct *dst_mm = dst_vma->vm_mm;
1232 struct mm_struct *src_mm = src_vma->vm_mm;
1233 struct mmu_notifier_range range;
1238 * Don't copy ptes where a page fault will fill them correctly.
1239 * Fork becomes much lighter when there are big shared or private
1240 * readonly mappings. The tradeoff is that copy_page_range is more
1241 * efficient than faulting.
1243 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1247 if (is_vm_hugetlb_page(src_vma))
1248 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1250 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1252 * We do not free on error cases below as remove_vma
1253 * gets called on error from higher level routine
1255 ret = track_pfn_copy(src_vma);
1261 * We need to invalidate the secondary MMU mappings only when
1262 * there could be a permission downgrade on the ptes of the
1263 * parent mm. And a permission downgrade will only happen if
1264 * is_cow_mapping() returns true.
1266 is_cow = is_cow_mapping(src_vma->vm_flags);
1269 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1270 0, src_vma, src_mm, addr, end);
1271 mmu_notifier_invalidate_range_start(&range);
1273 * Disabling preemption is not needed for the write side, as
1274 * the read side doesn't spin, but goes to the mmap_lock.
1276 * Use the raw variant of the seqcount_t write API to avoid
1277 * lockdep complaining about preemptibility.
1279 mmap_assert_write_locked(src_mm);
1280 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1284 dst_pgd = pgd_offset(dst_mm, addr);
1285 src_pgd = pgd_offset(src_mm, addr);
1287 next = pgd_addr_end(addr, end);
1288 if (pgd_none_or_clear_bad(src_pgd))
1290 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1295 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1298 raw_write_seqcount_end(&src_mm->write_protect_seq);
1299 mmu_notifier_invalidate_range_end(&range);
1304 /* Whether we should zap all COWed (private) pages too */
1305 static inline bool should_zap_cows(struct zap_details *details)
1307 /* By default, zap all pages */
1311 /* Or, we zap COWed pages only if the caller wants to */
1312 return !details->check_mapping;
1315 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1316 struct vm_area_struct *vma, pmd_t *pmd,
1317 unsigned long addr, unsigned long end,
1318 struct zap_details *details)
1320 struct mm_struct *mm = tlb->mm;
1321 int force_flush = 0;
1322 int rss[NR_MM_COUNTERS];
1328 tlb_change_page_size(tlb, PAGE_SIZE);
1331 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1333 flush_tlb_batched_pending(mm);
1334 arch_enter_lazy_mmu_mode();
1337 if (pte_none(ptent))
1343 if (pte_present(ptent)) {
1346 page = vm_normal_page(vma, addr, ptent);
1347 if (unlikely(details) && page) {
1349 * unmap_shared_mapping_pages() wants to
1350 * invalidate cache without truncating:
1351 * unmap shared but keep private pages.
1353 if (details->check_mapping &&
1354 details->check_mapping != page_rmapping(page))
1357 ptent = ptep_get_and_clear_full(mm, addr, pte,
1359 tlb_remove_tlb_entry(tlb, pte, addr);
1360 if (unlikely(!page))
1363 if (!PageAnon(page)) {
1364 if (pte_dirty(ptent)) {
1366 set_page_dirty(page);
1368 if (pte_young(ptent) &&
1369 likely(!(vma->vm_flags & VM_SEQ_READ)))
1370 mark_page_accessed(page);
1372 rss[mm_counter(page)]--;
1373 page_remove_rmap(page, false);
1374 if (unlikely(page_mapcount(page) < 0))
1375 print_bad_pte(vma, addr, ptent, page);
1376 if (unlikely(__tlb_remove_page(tlb, page))) {
1384 entry = pte_to_swp_entry(ptent);
1385 if (is_device_private_entry(entry) ||
1386 is_device_exclusive_entry(entry)) {
1387 struct page *page = pfn_swap_entry_to_page(entry);
1389 if (unlikely(details && details->check_mapping)) {
1391 * unmap_shared_mapping_pages() wants to
1392 * invalidate cache without truncating:
1393 * unmap shared but keep private pages.
1395 if (details->check_mapping !=
1396 page_rmapping(page))
1400 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1401 rss[mm_counter(page)]--;
1403 if (is_device_private_entry(entry))
1404 page_remove_rmap(page, false);
1410 if (!non_swap_entry(entry)) {
1411 /* Genuine swap entry, hence a private anon page */
1412 if (!should_zap_cows(details))
1415 } else if (is_migration_entry(entry)) {
1418 page = pfn_swap_entry_to_page(entry);
1419 if (details && details->check_mapping &&
1420 details->check_mapping != page_rmapping(page))
1422 rss[mm_counter(page)]--;
1424 if (unlikely(!free_swap_and_cache(entry)))
1425 print_bad_pte(vma, addr, ptent, NULL);
1426 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1427 } while (pte++, addr += PAGE_SIZE, addr != end);
1429 add_mm_rss_vec(mm, rss);
1430 arch_leave_lazy_mmu_mode();
1432 /* Do the actual TLB flush before dropping ptl */
1434 tlb_flush_mmu_tlbonly(tlb);
1435 pte_unmap_unlock(start_pte, ptl);
1438 * If we forced a TLB flush (either due to running out of
1439 * batch buffers or because we needed to flush dirty TLB
1440 * entries before releasing the ptl), free the batched
1441 * memory too. Restart if we didn't do everything.
1456 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1457 struct vm_area_struct *vma, pud_t *pud,
1458 unsigned long addr, unsigned long end,
1459 struct zap_details *details)
1464 pmd = pmd_offset(pud, addr);
1466 next = pmd_addr_end(addr, end);
1467 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1468 if (next - addr != HPAGE_PMD_SIZE)
1469 __split_huge_pmd(vma, pmd, addr, false, NULL);
1470 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1473 } else if (details && details->single_page &&
1474 PageTransCompound(details->single_page) &&
1475 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1476 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1478 * Take and drop THP pmd lock so that we cannot return
1479 * prematurely, while zap_huge_pmd() has cleared *pmd,
1480 * but not yet decremented compound_mapcount().
1486 * Here there can be other concurrent MADV_DONTNEED or
1487 * trans huge page faults running, and if the pmd is
1488 * none or trans huge it can change under us. This is
1489 * because MADV_DONTNEED holds the mmap_lock in read
1492 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1494 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1497 } while (pmd++, addr = next, addr != end);
1502 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1503 struct vm_area_struct *vma, p4d_t *p4d,
1504 unsigned long addr, unsigned long end,
1505 struct zap_details *details)
1510 pud = pud_offset(p4d, addr);
1512 next = pud_addr_end(addr, end);
1513 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1514 if (next - addr != HPAGE_PUD_SIZE) {
1515 mmap_assert_locked(tlb->mm);
1516 split_huge_pud(vma, pud, addr);
1517 } else if (zap_huge_pud(tlb, vma, pud, addr))
1521 if (pud_none_or_clear_bad(pud))
1523 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1526 } while (pud++, addr = next, addr != end);
1531 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1532 struct vm_area_struct *vma, pgd_t *pgd,
1533 unsigned long addr, unsigned long end,
1534 struct zap_details *details)
1539 p4d = p4d_offset(pgd, addr);
1541 next = p4d_addr_end(addr, end);
1542 if (p4d_none_or_clear_bad(p4d))
1544 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1545 } while (p4d++, addr = next, addr != end);
1550 void unmap_page_range(struct mmu_gather *tlb,
1551 struct vm_area_struct *vma,
1552 unsigned long addr, unsigned long end,
1553 struct zap_details *details)
1558 BUG_ON(addr >= end);
1559 tlb_start_vma(tlb, vma);
1560 pgd = pgd_offset(vma->vm_mm, addr);
1562 next = pgd_addr_end(addr, end);
1563 if (pgd_none_or_clear_bad(pgd))
1565 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1566 } while (pgd++, addr = next, addr != end);
1567 tlb_end_vma(tlb, vma);
1571 static void unmap_single_vma(struct mmu_gather *tlb,
1572 struct vm_area_struct *vma, unsigned long start_addr,
1573 unsigned long end_addr,
1574 struct zap_details *details)
1576 unsigned long start = max(vma->vm_start, start_addr);
1579 if (start >= vma->vm_end)
1581 end = min(vma->vm_end, end_addr);
1582 if (end <= vma->vm_start)
1586 uprobe_munmap(vma, start, end);
1588 if (unlikely(vma->vm_flags & VM_PFNMAP))
1589 untrack_pfn(vma, 0, 0);
1592 if (unlikely(is_vm_hugetlb_page(vma))) {
1594 * It is undesirable to test vma->vm_file as it
1595 * should be non-null for valid hugetlb area.
1596 * However, vm_file will be NULL in the error
1597 * cleanup path of mmap_region. When
1598 * hugetlbfs ->mmap method fails,
1599 * mmap_region() nullifies vma->vm_file
1600 * before calling this function to clean up.
1601 * Since no pte has actually been setup, it is
1602 * safe to do nothing in this case.
1605 i_mmap_lock_write(vma->vm_file->f_mapping);
1606 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1607 i_mmap_unlock_write(vma->vm_file->f_mapping);
1610 unmap_page_range(tlb, vma, start, end, details);
1615 * unmap_vmas - unmap a range of memory covered by a list of vma's
1616 * @tlb: address of the caller's struct mmu_gather
1617 * @vma: the starting vma
1618 * @start_addr: virtual address at which to start unmapping
1619 * @end_addr: virtual address at which to end unmapping
1621 * Unmap all pages in the vma list.
1623 * Only addresses between `start' and `end' will be unmapped.
1625 * The VMA list must be sorted in ascending virtual address order.
1627 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1628 * range after unmap_vmas() returns. So the only responsibility here is to
1629 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1630 * drops the lock and schedules.
1632 void unmap_vmas(struct mmu_gather *tlb,
1633 struct vm_area_struct *vma, unsigned long start_addr,
1634 unsigned long end_addr)
1636 struct mmu_notifier_range range;
1638 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1639 start_addr, end_addr);
1640 mmu_notifier_invalidate_range_start(&range);
1641 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1642 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1643 mmu_notifier_invalidate_range_end(&range);
1647 * zap_page_range - remove user pages in a given range
1648 * @vma: vm_area_struct holding the applicable pages
1649 * @start: starting address of pages to zap
1650 * @size: number of bytes to zap
1652 * Caller must protect the VMA list
1654 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1657 struct mmu_notifier_range range;
1658 struct mmu_gather tlb;
1661 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1662 start, start + size);
1663 tlb_gather_mmu(&tlb, vma->vm_mm);
1664 update_hiwater_rss(vma->vm_mm);
1665 mmu_notifier_invalidate_range_start(&range);
1666 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1667 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1668 mmu_notifier_invalidate_range_end(&range);
1669 tlb_finish_mmu(&tlb);
1673 * zap_page_range_single - remove user pages in a given range
1674 * @vma: vm_area_struct holding the applicable pages
1675 * @address: starting address of pages to zap
1676 * @size: number of bytes to zap
1677 * @details: details of shared cache invalidation
1679 * The range must fit into one VMA.
1681 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1682 unsigned long size, struct zap_details *details)
1684 struct mmu_notifier_range range;
1685 struct mmu_gather tlb;
1688 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1689 address, address + size);
1690 tlb_gather_mmu(&tlb, vma->vm_mm);
1691 update_hiwater_rss(vma->vm_mm);
1692 mmu_notifier_invalidate_range_start(&range);
1693 unmap_single_vma(&tlb, vma, address, range.end, details);
1694 mmu_notifier_invalidate_range_end(&range);
1695 tlb_finish_mmu(&tlb);
1699 * zap_vma_ptes - remove ptes mapping the vma
1700 * @vma: vm_area_struct holding ptes to be zapped
1701 * @address: starting address of pages to zap
1702 * @size: number of bytes to zap
1704 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1706 * The entire address range must be fully contained within the vma.
1709 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1712 if (address < vma->vm_start || address + size > vma->vm_end ||
1713 !(vma->vm_flags & VM_PFNMAP))
1716 zap_page_range_single(vma, address, size, NULL);
1718 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1720 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1727 pgd = pgd_offset(mm, addr);
1728 p4d = p4d_alloc(mm, pgd, addr);
1731 pud = pud_alloc(mm, p4d, addr);
1734 pmd = pmd_alloc(mm, pud, addr);
1738 VM_BUG_ON(pmd_trans_huge(*pmd));
1742 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1745 pmd_t *pmd = walk_to_pmd(mm, addr);
1749 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1752 static int validate_page_before_insert(struct page *page)
1754 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1756 flush_dcache_page(page);
1760 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1761 unsigned long addr, struct page *page, pgprot_t prot)
1763 if (!pte_none(*pte))
1765 /* Ok, finally just insert the thing.. */
1767 inc_mm_counter_fast(mm, mm_counter_file(page));
1768 page_add_file_rmap(page, false);
1769 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1774 * This is the old fallback for page remapping.
1776 * For historical reasons, it only allows reserved pages. Only
1777 * old drivers should use this, and they needed to mark their
1778 * pages reserved for the old functions anyway.
1780 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1781 struct page *page, pgprot_t prot)
1783 struct mm_struct *mm = vma->vm_mm;
1788 retval = validate_page_before_insert(page);
1792 pte = get_locked_pte(mm, addr, &ptl);
1795 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1796 pte_unmap_unlock(pte, ptl);
1802 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1803 unsigned long addr, struct page *page, pgprot_t prot)
1807 if (!page_count(page))
1809 err = validate_page_before_insert(page);
1812 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1815 /* insert_pages() amortizes the cost of spinlock operations
1816 * when inserting pages in a loop. Arch *must* define pte_index.
1818 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1819 struct page **pages, unsigned long *num, pgprot_t prot)
1822 pte_t *start_pte, *pte;
1823 spinlock_t *pte_lock;
1824 struct mm_struct *const mm = vma->vm_mm;
1825 unsigned long curr_page_idx = 0;
1826 unsigned long remaining_pages_total = *num;
1827 unsigned long pages_to_write_in_pmd;
1831 pmd = walk_to_pmd(mm, addr);
1835 pages_to_write_in_pmd = min_t(unsigned long,
1836 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1838 /* Allocate the PTE if necessary; takes PMD lock once only. */
1840 if (pte_alloc(mm, pmd))
1843 while (pages_to_write_in_pmd) {
1845 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1847 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1848 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1849 int err = insert_page_in_batch_locked(mm, pte,
1850 addr, pages[curr_page_idx], prot);
1851 if (unlikely(err)) {
1852 pte_unmap_unlock(start_pte, pte_lock);
1854 remaining_pages_total -= pte_idx;
1860 pte_unmap_unlock(start_pte, pte_lock);
1861 pages_to_write_in_pmd -= batch_size;
1862 remaining_pages_total -= batch_size;
1864 if (remaining_pages_total)
1868 *num = remaining_pages_total;
1871 #endif /* ifdef pte_index */
1874 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1875 * @vma: user vma to map to
1876 * @addr: target start user address of these pages
1877 * @pages: source kernel pages
1878 * @num: in: number of pages to map. out: number of pages that were *not*
1879 * mapped. (0 means all pages were successfully mapped).
1881 * Preferred over vm_insert_page() when inserting multiple pages.
1883 * In case of error, we may have mapped a subset of the provided
1884 * pages. It is the caller's responsibility to account for this case.
1886 * The same restrictions apply as in vm_insert_page().
1888 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1889 struct page **pages, unsigned long *num)
1892 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1894 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1896 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1897 BUG_ON(mmap_read_trylock(vma->vm_mm));
1898 BUG_ON(vma->vm_flags & VM_PFNMAP);
1899 vma->vm_flags |= VM_MIXEDMAP;
1901 /* Defer page refcount checking till we're about to map that page. */
1902 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1904 unsigned long idx = 0, pgcount = *num;
1907 for (; idx < pgcount; ++idx) {
1908 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1912 *num = pgcount - idx;
1914 #endif /* ifdef pte_index */
1916 EXPORT_SYMBOL(vm_insert_pages);
1919 * vm_insert_page - insert single page into user vma
1920 * @vma: user vma to map to
1921 * @addr: target user address of this page
1922 * @page: source kernel page
1924 * This allows drivers to insert individual pages they've allocated
1927 * The page has to be a nice clean _individual_ kernel allocation.
1928 * If you allocate a compound page, you need to have marked it as
1929 * such (__GFP_COMP), or manually just split the page up yourself
1930 * (see split_page()).
1932 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1933 * took an arbitrary page protection parameter. This doesn't allow
1934 * that. Your vma protection will have to be set up correctly, which
1935 * means that if you want a shared writable mapping, you'd better
1936 * ask for a shared writable mapping!
1938 * The page does not need to be reserved.
1940 * Usually this function is called from f_op->mmap() handler
1941 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1942 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1943 * function from other places, for example from page-fault handler.
1945 * Return: %0 on success, negative error code otherwise.
1947 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1950 if (addr < vma->vm_start || addr >= vma->vm_end)
1952 if (!page_count(page))
1954 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1955 BUG_ON(mmap_read_trylock(vma->vm_mm));
1956 BUG_ON(vma->vm_flags & VM_PFNMAP);
1957 vma->vm_flags |= VM_MIXEDMAP;
1959 return insert_page(vma, addr, page, vma->vm_page_prot);
1961 EXPORT_SYMBOL(vm_insert_page);
1964 * __vm_map_pages - maps range of kernel pages into user vma
1965 * @vma: user vma to map to
1966 * @pages: pointer to array of source kernel pages
1967 * @num: number of pages in page array
1968 * @offset: user's requested vm_pgoff
1970 * This allows drivers to map range of kernel pages into a user vma.
1972 * Return: 0 on success and error code otherwise.
1974 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1975 unsigned long num, unsigned long offset)
1977 unsigned long count = vma_pages(vma);
1978 unsigned long uaddr = vma->vm_start;
1981 /* Fail if the user requested offset is beyond the end of the object */
1985 /* Fail if the user requested size exceeds available object size */
1986 if (count > num - offset)
1989 for (i = 0; i < count; i++) {
1990 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
2000 * vm_map_pages - maps range of kernel pages starts with non zero offset
2001 * @vma: user vma to map to
2002 * @pages: pointer to array of source kernel pages
2003 * @num: number of pages in page array
2005 * Maps an object consisting of @num pages, catering for the user's
2006 * requested vm_pgoff
2008 * If we fail to insert any page into the vma, the function will return
2009 * immediately leaving any previously inserted pages present. Callers
2010 * from the mmap handler may immediately return the error as their caller
2011 * will destroy the vma, removing any successfully inserted pages. Other
2012 * callers should make their own arrangements for calling unmap_region().
2014 * Context: Process context. Called by mmap handlers.
2015 * Return: 0 on success and error code otherwise.
2017 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2020 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2022 EXPORT_SYMBOL(vm_map_pages);
2025 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2026 * @vma: user vma to map to
2027 * @pages: pointer to array of source kernel pages
2028 * @num: number of pages in page array
2030 * Similar to vm_map_pages(), except that it explicitly sets the offset
2031 * to 0. This function is intended for the drivers that did not consider
2034 * Context: Process context. Called by mmap handlers.
2035 * Return: 0 on success and error code otherwise.
2037 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2040 return __vm_map_pages(vma, pages, num, 0);
2042 EXPORT_SYMBOL(vm_map_pages_zero);
2044 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2045 pfn_t pfn, pgprot_t prot, bool mkwrite)
2047 struct mm_struct *mm = vma->vm_mm;
2051 pte = get_locked_pte(mm, addr, &ptl);
2053 return VM_FAULT_OOM;
2054 if (!pte_none(*pte)) {
2057 * For read faults on private mappings the PFN passed
2058 * in may not match the PFN we have mapped if the
2059 * mapped PFN is a writeable COW page. In the mkwrite
2060 * case we are creating a writable PTE for a shared
2061 * mapping and we expect the PFNs to match. If they
2062 * don't match, we are likely racing with block
2063 * allocation and mapping invalidation so just skip the
2066 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
2067 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
2070 entry = pte_mkyoung(*pte);
2071 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2072 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2073 update_mmu_cache(vma, addr, pte);
2078 /* Ok, finally just insert the thing.. */
2079 if (pfn_t_devmap(pfn))
2080 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2082 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2085 entry = pte_mkyoung(entry);
2086 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2089 set_pte_at(mm, addr, pte, entry);
2090 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2093 pte_unmap_unlock(pte, ptl);
2094 return VM_FAULT_NOPAGE;
2098 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2099 * @vma: user vma to map to
2100 * @addr: target user address of this page
2101 * @pfn: source kernel pfn
2102 * @pgprot: pgprot flags for the inserted page
2104 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2105 * to override pgprot on a per-page basis.
2107 * This only makes sense for IO mappings, and it makes no sense for
2108 * COW mappings. In general, using multiple vmas is preferable;
2109 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2112 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2113 * a value of @pgprot different from that of @vma->vm_page_prot.
2115 * Context: Process context. May allocate using %GFP_KERNEL.
2116 * Return: vm_fault_t value.
2118 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2119 unsigned long pfn, pgprot_t pgprot)
2122 * Technically, architectures with pte_special can avoid all these
2123 * restrictions (same for remap_pfn_range). However we would like
2124 * consistency in testing and feature parity among all, so we should
2125 * try to keep these invariants in place for everybody.
2127 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2128 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2129 (VM_PFNMAP|VM_MIXEDMAP));
2130 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2131 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2133 if (addr < vma->vm_start || addr >= vma->vm_end)
2134 return VM_FAULT_SIGBUS;
2136 if (!pfn_modify_allowed(pfn, pgprot))
2137 return VM_FAULT_SIGBUS;
2139 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2141 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2144 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2147 * vmf_insert_pfn - insert single pfn into user vma
2148 * @vma: user vma to map to
2149 * @addr: target user address of this page
2150 * @pfn: source kernel pfn
2152 * Similar to vm_insert_page, this allows drivers to insert individual pages
2153 * they've allocated into a user vma. Same comments apply.
2155 * This function should only be called from a vm_ops->fault handler, and
2156 * in that case the handler should return the result of this function.
2158 * vma cannot be a COW mapping.
2160 * As this is called only for pages that do not currently exist, we
2161 * do not need to flush old virtual caches or the TLB.
2163 * Context: Process context. May allocate using %GFP_KERNEL.
2164 * Return: vm_fault_t value.
2166 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2169 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2171 EXPORT_SYMBOL(vmf_insert_pfn);
2173 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2175 /* these checks mirror the abort conditions in vm_normal_page */
2176 if (vma->vm_flags & VM_MIXEDMAP)
2178 if (pfn_t_devmap(pfn))
2180 if (pfn_t_special(pfn))
2182 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2187 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2188 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2193 BUG_ON(!vm_mixed_ok(vma, pfn));
2195 if (addr < vma->vm_start || addr >= vma->vm_end)
2196 return VM_FAULT_SIGBUS;
2198 track_pfn_insert(vma, &pgprot, pfn);
2200 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2201 return VM_FAULT_SIGBUS;
2204 * If we don't have pte special, then we have to use the pfn_valid()
2205 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2206 * refcount the page if pfn_valid is true (hence insert_page rather
2207 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2208 * without pte special, it would there be refcounted as a normal page.
2210 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2211 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2215 * At this point we are committed to insert_page()
2216 * regardless of whether the caller specified flags that
2217 * result in pfn_t_has_page() == false.
2219 page = pfn_to_page(pfn_t_to_pfn(pfn));
2220 err = insert_page(vma, addr, page, pgprot);
2222 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2226 return VM_FAULT_OOM;
2227 if (err < 0 && err != -EBUSY)
2228 return VM_FAULT_SIGBUS;
2230 return VM_FAULT_NOPAGE;
2234 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2235 * @vma: user vma to map to
2236 * @addr: target user address of this page
2237 * @pfn: source kernel pfn
2238 * @pgprot: pgprot flags for the inserted page
2240 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2241 * to override pgprot on a per-page basis.
2243 * Typically this function should be used by drivers to set caching- and
2244 * encryption bits different than those of @vma->vm_page_prot, because
2245 * the caching- or encryption mode may not be known at mmap() time.
2246 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2247 * to set caching and encryption bits for those vmas (except for COW pages).
2248 * This is ensured by core vm only modifying these page table entries using
2249 * functions that don't touch caching- or encryption bits, using pte_modify()
2250 * if needed. (See for example mprotect()).
2251 * Also when new page-table entries are created, this is only done using the
2252 * fault() callback, and never using the value of vma->vm_page_prot,
2253 * except for page-table entries that point to anonymous pages as the result
2256 * Context: Process context. May allocate using %GFP_KERNEL.
2257 * Return: vm_fault_t value.
2259 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2260 pfn_t pfn, pgprot_t pgprot)
2262 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2264 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2266 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2269 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2271 EXPORT_SYMBOL(vmf_insert_mixed);
2274 * If the insertion of PTE failed because someone else already added a
2275 * different entry in the mean time, we treat that as success as we assume
2276 * the same entry was actually inserted.
2278 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2279 unsigned long addr, pfn_t pfn)
2281 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2283 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2286 * maps a range of physical memory into the requested pages. the old
2287 * mappings are removed. any references to nonexistent pages results
2288 * in null mappings (currently treated as "copy-on-access")
2290 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2291 unsigned long addr, unsigned long end,
2292 unsigned long pfn, pgprot_t prot)
2294 pte_t *pte, *mapped_pte;
2298 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2301 arch_enter_lazy_mmu_mode();
2303 BUG_ON(!pte_none(*pte));
2304 if (!pfn_modify_allowed(pfn, prot)) {
2308 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2310 } while (pte++, addr += PAGE_SIZE, addr != end);
2311 arch_leave_lazy_mmu_mode();
2312 pte_unmap_unlock(mapped_pte, ptl);
2316 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2317 unsigned long addr, unsigned long end,
2318 unsigned long pfn, pgprot_t prot)
2324 pfn -= addr >> PAGE_SHIFT;
2325 pmd = pmd_alloc(mm, pud, addr);
2328 VM_BUG_ON(pmd_trans_huge(*pmd));
2330 next = pmd_addr_end(addr, end);
2331 err = remap_pte_range(mm, pmd, addr, next,
2332 pfn + (addr >> PAGE_SHIFT), prot);
2335 } while (pmd++, addr = next, addr != end);
2339 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2340 unsigned long addr, unsigned long end,
2341 unsigned long pfn, pgprot_t prot)
2347 pfn -= addr >> PAGE_SHIFT;
2348 pud = pud_alloc(mm, p4d, addr);
2352 next = pud_addr_end(addr, end);
2353 err = remap_pmd_range(mm, pud, addr, next,
2354 pfn + (addr >> PAGE_SHIFT), prot);
2357 } while (pud++, addr = next, addr != end);
2361 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2362 unsigned long addr, unsigned long end,
2363 unsigned long pfn, pgprot_t prot)
2369 pfn -= addr >> PAGE_SHIFT;
2370 p4d = p4d_alloc(mm, pgd, addr);
2374 next = p4d_addr_end(addr, end);
2375 err = remap_pud_range(mm, p4d, addr, next,
2376 pfn + (addr >> PAGE_SHIFT), prot);
2379 } while (p4d++, addr = next, addr != end);
2384 * Variant of remap_pfn_range that does not call track_pfn_remap. The caller
2385 * must have pre-validated the caching bits of the pgprot_t.
2387 int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2388 unsigned long pfn, unsigned long size, pgprot_t prot)
2392 unsigned long end = addr + PAGE_ALIGN(size);
2393 struct mm_struct *mm = vma->vm_mm;
2396 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2400 * Physically remapped pages are special. Tell the
2401 * rest of the world about it:
2402 * VM_IO tells people not to look at these pages
2403 * (accesses can have side effects).
2404 * VM_PFNMAP tells the core MM that the base pages are just
2405 * raw PFN mappings, and do not have a "struct page" associated
2408 * Disable vma merging and expanding with mremap().
2410 * Omit vma from core dump, even when VM_IO turned off.
2412 * There's a horrible special case to handle copy-on-write
2413 * behaviour that some programs depend on. We mark the "original"
2414 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2415 * See vm_normal_page() for details.
2417 if (is_cow_mapping(vma->vm_flags)) {
2418 if (addr != vma->vm_start || end != vma->vm_end)
2420 vma->vm_pgoff = pfn;
2423 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2425 BUG_ON(addr >= end);
2426 pfn -= addr >> PAGE_SHIFT;
2427 pgd = pgd_offset(mm, addr);
2428 flush_cache_range(vma, addr, end);
2430 next = pgd_addr_end(addr, end);
2431 err = remap_p4d_range(mm, pgd, addr, next,
2432 pfn + (addr >> PAGE_SHIFT), prot);
2435 } while (pgd++, addr = next, addr != end);
2441 * remap_pfn_range - remap kernel memory to userspace
2442 * @vma: user vma to map to
2443 * @addr: target page aligned user address to start at
2444 * @pfn: page frame number of kernel physical memory address
2445 * @size: size of mapping area
2446 * @prot: page protection flags for this mapping
2448 * Note: this is only safe if the mm semaphore is held when called.
2450 * Return: %0 on success, negative error code otherwise.
2452 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2453 unsigned long pfn, unsigned long size, pgprot_t prot)
2457 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2461 err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2463 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2466 EXPORT_SYMBOL(remap_pfn_range);
2469 * vm_iomap_memory - remap memory to userspace
2470 * @vma: user vma to map to
2471 * @start: start of the physical memory to be mapped
2472 * @len: size of area
2474 * This is a simplified io_remap_pfn_range() for common driver use. The
2475 * driver just needs to give us the physical memory range to be mapped,
2476 * we'll figure out the rest from the vma information.
2478 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2479 * whatever write-combining details or similar.
2481 * Return: %0 on success, negative error code otherwise.
2483 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2485 unsigned long vm_len, pfn, pages;
2487 /* Check that the physical memory area passed in looks valid */
2488 if (start + len < start)
2491 * You *really* shouldn't map things that aren't page-aligned,
2492 * but we've historically allowed it because IO memory might
2493 * just have smaller alignment.
2495 len += start & ~PAGE_MASK;
2496 pfn = start >> PAGE_SHIFT;
2497 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2498 if (pfn + pages < pfn)
2501 /* We start the mapping 'vm_pgoff' pages into the area */
2502 if (vma->vm_pgoff > pages)
2504 pfn += vma->vm_pgoff;
2505 pages -= vma->vm_pgoff;
2507 /* Can we fit all of the mapping? */
2508 vm_len = vma->vm_end - vma->vm_start;
2509 if (vm_len >> PAGE_SHIFT > pages)
2512 /* Ok, let it rip */
2513 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2515 EXPORT_SYMBOL(vm_iomap_memory);
2517 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2518 unsigned long addr, unsigned long end,
2519 pte_fn_t fn, void *data, bool create,
2520 pgtbl_mod_mask *mask)
2522 pte_t *pte, *mapped_pte;
2527 mapped_pte = pte = (mm == &init_mm) ?
2528 pte_alloc_kernel_track(pmd, addr, mask) :
2529 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2533 mapped_pte = pte = (mm == &init_mm) ?
2534 pte_offset_kernel(pmd, addr) :
2535 pte_offset_map_lock(mm, pmd, addr, &ptl);
2538 BUG_ON(pmd_huge(*pmd));
2540 arch_enter_lazy_mmu_mode();
2544 if (create || !pte_none(*pte)) {
2545 err = fn(pte++, addr, data);
2549 } while (addr += PAGE_SIZE, addr != end);
2551 *mask |= PGTBL_PTE_MODIFIED;
2553 arch_leave_lazy_mmu_mode();
2556 pte_unmap_unlock(mapped_pte, ptl);
2560 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2561 unsigned long addr, unsigned long end,
2562 pte_fn_t fn, void *data, bool create,
2563 pgtbl_mod_mask *mask)
2569 BUG_ON(pud_huge(*pud));
2572 pmd = pmd_alloc_track(mm, pud, addr, mask);
2576 pmd = pmd_offset(pud, addr);
2579 next = pmd_addr_end(addr, end);
2580 if (pmd_none(*pmd) && !create)
2582 if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2584 if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2589 err = apply_to_pte_range(mm, pmd, addr, next,
2590 fn, data, create, mask);
2593 } while (pmd++, addr = next, addr != end);
2598 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2599 unsigned long addr, unsigned long end,
2600 pte_fn_t fn, void *data, bool create,
2601 pgtbl_mod_mask *mask)
2608 pud = pud_alloc_track(mm, p4d, addr, mask);
2612 pud = pud_offset(p4d, addr);
2615 next = pud_addr_end(addr, end);
2616 if (pud_none(*pud) && !create)
2618 if (WARN_ON_ONCE(pud_leaf(*pud)))
2620 if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2625 err = apply_to_pmd_range(mm, pud, addr, next,
2626 fn, data, create, mask);
2629 } while (pud++, addr = next, addr != end);
2634 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2635 unsigned long addr, unsigned long end,
2636 pte_fn_t fn, void *data, bool create,
2637 pgtbl_mod_mask *mask)
2644 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2648 p4d = p4d_offset(pgd, addr);
2651 next = p4d_addr_end(addr, end);
2652 if (p4d_none(*p4d) && !create)
2654 if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2656 if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2661 err = apply_to_pud_range(mm, p4d, addr, next,
2662 fn, data, create, mask);
2665 } while (p4d++, addr = next, addr != end);
2670 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2671 unsigned long size, pte_fn_t fn,
2672 void *data, bool create)
2675 unsigned long start = addr, next;
2676 unsigned long end = addr + size;
2677 pgtbl_mod_mask mask = 0;
2680 if (WARN_ON(addr >= end))
2683 pgd = pgd_offset(mm, addr);
2685 next = pgd_addr_end(addr, end);
2686 if (pgd_none(*pgd) && !create)
2688 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2690 if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2695 err = apply_to_p4d_range(mm, pgd, addr, next,
2696 fn, data, create, &mask);
2699 } while (pgd++, addr = next, addr != end);
2701 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2702 arch_sync_kernel_mappings(start, start + size);
2708 * Scan a region of virtual memory, filling in page tables as necessary
2709 * and calling a provided function on each leaf page table.
2711 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2712 unsigned long size, pte_fn_t fn, void *data)
2714 return __apply_to_page_range(mm, addr, size, fn, data, true);
2716 EXPORT_SYMBOL_GPL(apply_to_page_range);
2719 * Scan a region of virtual memory, calling a provided function on
2720 * each leaf page table where it exists.
2722 * Unlike apply_to_page_range, this does _not_ fill in page tables
2723 * where they are absent.
2725 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2726 unsigned long size, pte_fn_t fn, void *data)
2728 return __apply_to_page_range(mm, addr, size, fn, data, false);
2730 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2733 * handle_pte_fault chooses page fault handler according to an entry which was
2734 * read non-atomically. Before making any commitment, on those architectures
2735 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2736 * parts, do_swap_page must check under lock before unmapping the pte and
2737 * proceeding (but do_wp_page is only called after already making such a check;
2738 * and do_anonymous_page can safely check later on).
2740 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2741 pte_t *page_table, pte_t orig_pte)
2744 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2745 if (sizeof(pte_t) > sizeof(unsigned long)) {
2746 spinlock_t *ptl = pte_lockptr(mm, pmd);
2748 same = pte_same(*page_table, orig_pte);
2752 pte_unmap(page_table);
2756 static inline bool cow_user_page(struct page *dst, struct page *src,
2757 struct vm_fault *vmf)
2762 bool locked = false;
2763 struct vm_area_struct *vma = vmf->vma;
2764 struct mm_struct *mm = vma->vm_mm;
2765 unsigned long addr = vmf->address;
2768 copy_user_highpage(dst, src, addr, vma);
2773 * If the source page was a PFN mapping, we don't have
2774 * a "struct page" for it. We do a best-effort copy by
2775 * just copying from the original user address. If that
2776 * fails, we just zero-fill it. Live with it.
2778 kaddr = kmap_atomic(dst);
2779 uaddr = (void __user *)(addr & PAGE_MASK);
2782 * On architectures with software "accessed" bits, we would
2783 * take a double page fault, so mark it accessed here.
2785 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2788 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2790 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2792 * Other thread has already handled the fault
2793 * and update local tlb only
2795 update_mmu_tlb(vma, addr, vmf->pte);
2800 entry = pte_mkyoung(vmf->orig_pte);
2801 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2802 update_mmu_cache(vma, addr, vmf->pte);
2806 * This really shouldn't fail, because the page is there
2807 * in the page tables. But it might just be unreadable,
2808 * in which case we just give up and fill the result with
2811 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2815 /* Re-validate under PTL if the page is still mapped */
2816 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2818 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2819 /* The PTE changed under us, update local tlb */
2820 update_mmu_tlb(vma, addr, vmf->pte);
2826 * The same page can be mapped back since last copy attempt.
2827 * Try to copy again under PTL.
2829 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2831 * Give a warn in case there can be some obscure
2844 pte_unmap_unlock(vmf->pte, vmf->ptl);
2845 kunmap_atomic(kaddr);
2846 flush_dcache_page(dst);
2851 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2853 struct file *vm_file = vma->vm_file;
2856 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2859 * Special mappings (e.g. VDSO) do not have any file so fake
2860 * a default GFP_KERNEL for them.
2866 * Notify the address space that the page is about to become writable so that
2867 * it can prohibit this or wait for the page to get into an appropriate state.
2869 * We do this without the lock held, so that it can sleep if it needs to.
2871 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2874 struct page *page = vmf->page;
2875 unsigned int old_flags = vmf->flags;
2877 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2879 if (vmf->vma->vm_file &&
2880 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2881 return VM_FAULT_SIGBUS;
2883 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2884 /* Restore original flags so that caller is not surprised */
2885 vmf->flags = old_flags;
2886 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2888 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2890 if (!page->mapping) {
2892 return 0; /* retry */
2894 ret |= VM_FAULT_LOCKED;
2896 VM_BUG_ON_PAGE(!PageLocked(page), page);
2901 * Handle dirtying of a page in shared file mapping on a write fault.
2903 * The function expects the page to be locked and unlocks it.
2905 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2907 struct vm_area_struct *vma = vmf->vma;
2908 struct address_space *mapping;
2909 struct page *page = vmf->page;
2911 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2913 dirtied = set_page_dirty(page);
2914 VM_BUG_ON_PAGE(PageAnon(page), page);
2916 * Take a local copy of the address_space - page.mapping may be zeroed
2917 * by truncate after unlock_page(). The address_space itself remains
2918 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2919 * release semantics to prevent the compiler from undoing this copying.
2921 mapping = page_rmapping(page);
2925 file_update_time(vma->vm_file);
2928 * Throttle page dirtying rate down to writeback speed.
2930 * mapping may be NULL here because some device drivers do not
2931 * set page.mapping but still dirty their pages
2933 * Drop the mmap_lock before waiting on IO, if we can. The file
2934 * is pinning the mapping, as per above.
2936 if ((dirtied || page_mkwrite) && mapping) {
2939 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2940 balance_dirty_pages_ratelimited(mapping);
2943 return VM_FAULT_RETRY;
2951 * Handle write page faults for pages that can be reused in the current vma
2953 * This can happen either due to the mapping being with the VM_SHARED flag,
2954 * or due to us being the last reference standing to the page. In either
2955 * case, all we need to do here is to mark the page as writable and update
2956 * any related book-keeping.
2958 static inline void wp_page_reuse(struct vm_fault *vmf)
2959 __releases(vmf->ptl)
2961 struct vm_area_struct *vma = vmf->vma;
2962 struct page *page = vmf->page;
2965 * Clear the pages cpupid information as the existing
2966 * information potentially belongs to a now completely
2967 * unrelated process.
2970 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2972 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2973 entry = pte_mkyoung(vmf->orig_pte);
2974 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2975 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2976 update_mmu_cache(vma, vmf->address, vmf->pte);
2977 pte_unmap_unlock(vmf->pte, vmf->ptl);
2978 count_vm_event(PGREUSE);
2982 * Handle the case of a page which we actually need to copy to a new page.
2984 * Called with mmap_lock locked and the old page referenced, but
2985 * without the ptl held.
2987 * High level logic flow:
2989 * - Allocate a page, copy the content of the old page to the new one.
2990 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2991 * - Take the PTL. If the pte changed, bail out and release the allocated page
2992 * - If the pte is still the way we remember it, update the page table and all
2993 * relevant references. This includes dropping the reference the page-table
2994 * held to the old page, as well as updating the rmap.
2995 * - In any case, unlock the PTL and drop the reference we took to the old page.
2997 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2999 struct vm_area_struct *vma = vmf->vma;
3000 struct mm_struct *mm = vma->vm_mm;
3001 struct page *old_page = vmf->page;
3002 struct page *new_page = NULL;
3004 int page_copied = 0;
3005 struct mmu_notifier_range range;
3007 if (unlikely(anon_vma_prepare(vma)))
3010 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
3011 new_page = alloc_zeroed_user_highpage_movable(vma,
3016 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3021 if (!cow_user_page(new_page, old_page, vmf)) {
3023 * COW failed, if the fault was solved by other,
3024 * it's fine. If not, userspace would re-fault on
3025 * the same address and we will handle the fault
3026 * from the second attempt.
3035 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
3037 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
3039 __SetPageUptodate(new_page);
3041 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
3042 vmf->address & PAGE_MASK,
3043 (vmf->address & PAGE_MASK) + PAGE_SIZE);
3044 mmu_notifier_invalidate_range_start(&range);
3047 * Re-check the pte - we dropped the lock
3049 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3050 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
3052 if (!PageAnon(old_page)) {
3053 dec_mm_counter_fast(mm,
3054 mm_counter_file(old_page));
3055 inc_mm_counter_fast(mm, MM_ANONPAGES);
3058 inc_mm_counter_fast(mm, MM_ANONPAGES);
3060 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3061 entry = mk_pte(new_page, vma->vm_page_prot);
3062 entry = pte_sw_mkyoung(entry);
3063 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3066 * Clear the pte entry and flush it first, before updating the
3067 * pte with the new entry, to keep TLBs on different CPUs in
3068 * sync. This code used to set the new PTE then flush TLBs, but
3069 * that left a window where the new PTE could be loaded into
3070 * some TLBs while the old PTE remains in others.
3072 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3073 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
3074 lru_cache_add_inactive_or_unevictable(new_page, vma);
3076 * We call the notify macro here because, when using secondary
3077 * mmu page tables (such as kvm shadow page tables), we want the
3078 * new page to be mapped directly into the secondary page table.
3080 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3081 update_mmu_cache(vma, vmf->address, vmf->pte);
3084 * Only after switching the pte to the new page may
3085 * we remove the mapcount here. Otherwise another
3086 * process may come and find the rmap count decremented
3087 * before the pte is switched to the new page, and
3088 * "reuse" the old page writing into it while our pte
3089 * here still points into it and can be read by other
3092 * The critical issue is to order this
3093 * page_remove_rmap with the ptp_clear_flush above.
3094 * Those stores are ordered by (if nothing else,)
3095 * the barrier present in the atomic_add_negative
3096 * in page_remove_rmap.
3098 * Then the TLB flush in ptep_clear_flush ensures that
3099 * no process can access the old page before the
3100 * decremented mapcount is visible. And the old page
3101 * cannot be reused until after the decremented
3102 * mapcount is visible. So transitively, TLBs to
3103 * old page will be flushed before it can be reused.
3105 page_remove_rmap(old_page, false);
3108 /* Free the old page.. */
3109 new_page = old_page;
3112 update_mmu_tlb(vma, vmf->address, vmf->pte);
3118 pte_unmap_unlock(vmf->pte, vmf->ptl);
3120 * No need to double call mmu_notifier->invalidate_range() callback as
3121 * the above ptep_clear_flush_notify() did already call it.
3123 mmu_notifier_invalidate_range_only_end(&range);
3126 * Don't let another task, with possibly unlocked vma,
3127 * keep the mlocked page.
3129 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
3130 lock_page(old_page); /* LRU manipulation */
3131 if (PageMlocked(old_page))
3132 munlock_vma_page(old_page);
3133 unlock_page(old_page);
3136 free_swap_cache(old_page);
3139 return page_copied ? VM_FAULT_WRITE : 0;
3145 return VM_FAULT_OOM;
3149 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3150 * writeable once the page is prepared
3152 * @vmf: structure describing the fault
3154 * This function handles all that is needed to finish a write page fault in a
3155 * shared mapping due to PTE being read-only once the mapped page is prepared.
3156 * It handles locking of PTE and modifying it.
3158 * The function expects the page to be locked or other protection against
3159 * concurrent faults / writeback (such as DAX radix tree locks).
3161 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3162 * we acquired PTE lock.
3164 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3166 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3167 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3170 * We might have raced with another page fault while we released the
3171 * pte_offset_map_lock.
3173 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3174 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3175 pte_unmap_unlock(vmf->pte, vmf->ptl);
3176 return VM_FAULT_NOPAGE;
3183 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3186 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3188 struct vm_area_struct *vma = vmf->vma;
3190 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3193 pte_unmap_unlock(vmf->pte, vmf->ptl);
3194 vmf->flags |= FAULT_FLAG_MKWRITE;
3195 ret = vma->vm_ops->pfn_mkwrite(vmf);
3196 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3198 return finish_mkwrite_fault(vmf);
3201 return VM_FAULT_WRITE;
3204 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3205 __releases(vmf->ptl)
3207 struct vm_area_struct *vma = vmf->vma;
3208 vm_fault_t ret = VM_FAULT_WRITE;
3210 get_page(vmf->page);
3212 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3215 pte_unmap_unlock(vmf->pte, vmf->ptl);
3216 tmp = do_page_mkwrite(vmf);
3217 if (unlikely(!tmp || (tmp &
3218 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3219 put_page(vmf->page);
3222 tmp = finish_mkwrite_fault(vmf);
3223 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3224 unlock_page(vmf->page);
3225 put_page(vmf->page);
3230 lock_page(vmf->page);
3232 ret |= fault_dirty_shared_page(vmf);
3233 put_page(vmf->page);
3239 * This routine handles present pages, when users try to write
3240 * to a shared page. It is done by copying the page to a new address
3241 * and decrementing the shared-page counter for the old page.
3243 * Note that this routine assumes that the protection checks have been
3244 * done by the caller (the low-level page fault routine in most cases).
3245 * Thus we can safely just mark it writable once we've done any necessary
3248 * We also mark the page dirty at this point even though the page will
3249 * change only once the write actually happens. This avoids a few races,
3250 * and potentially makes it more efficient.
3252 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3253 * but allow concurrent faults), with pte both mapped and locked.
3254 * We return with mmap_lock still held, but pte unmapped and unlocked.
3256 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3257 __releases(vmf->ptl)
3259 struct vm_area_struct *vma = vmf->vma;
3261 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3262 pte_unmap_unlock(vmf->pte, vmf->ptl);
3263 return handle_userfault(vmf, VM_UFFD_WP);
3267 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3268 * is flushed in this case before copying.
3270 if (unlikely(userfaultfd_wp(vmf->vma) &&
3271 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3272 flush_tlb_page(vmf->vma, vmf->address);
3274 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3277 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3280 * We should not cow pages in a shared writeable mapping.
3281 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3283 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3284 (VM_WRITE|VM_SHARED))
3285 return wp_pfn_shared(vmf);
3287 pte_unmap_unlock(vmf->pte, vmf->ptl);
3288 return wp_page_copy(vmf);
3292 * Take out anonymous pages first, anonymous shared vmas are
3293 * not dirty accountable.
3295 if (PageAnon(vmf->page)) {
3296 struct page *page = vmf->page;
3298 /* PageKsm() doesn't necessarily raise the page refcount */
3299 if (PageKsm(page) || page_count(page) != 1)
3301 if (!trylock_page(page))
3303 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3308 * Ok, we've got the only map reference, and the only
3309 * page count reference, and the page is locked,
3310 * it's dark out, and we're wearing sunglasses. Hit it.
3314 return VM_FAULT_WRITE;
3315 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3316 (VM_WRITE|VM_SHARED))) {
3317 return wp_page_shared(vmf);
3321 * Ok, we need to copy. Oh, well..
3323 get_page(vmf->page);
3325 pte_unmap_unlock(vmf->pte, vmf->ptl);
3326 return wp_page_copy(vmf);
3329 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3330 unsigned long start_addr, unsigned long end_addr,
3331 struct zap_details *details)
3333 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3336 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3337 struct zap_details *details)
3339 struct vm_area_struct *vma;
3340 pgoff_t vba, vea, zba, zea;
3342 vma_interval_tree_foreach(vma, root,
3343 details->first_index, details->last_index) {
3345 vba = vma->vm_pgoff;
3346 vea = vba + vma_pages(vma) - 1;
3347 zba = details->first_index;
3350 zea = details->last_index;
3354 unmap_mapping_range_vma(vma,
3355 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3356 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3362 * unmap_mapping_page() - Unmap single page from processes.
3363 * @page: The locked page to be unmapped.
3365 * Unmap this page from any userspace process which still has it mmaped.
3366 * Typically, for efficiency, the range of nearby pages has already been
3367 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3368 * truncation or invalidation holds the lock on a page, it may find that
3369 * the page has been remapped again: and then uses unmap_mapping_page()
3370 * to unmap it finally.
3372 void unmap_mapping_page(struct page *page)
3374 struct address_space *mapping = page->mapping;
3375 struct zap_details details = { };
3377 VM_BUG_ON(!PageLocked(page));
3378 VM_BUG_ON(PageTail(page));
3380 details.check_mapping = mapping;
3381 details.first_index = page->index;
3382 details.last_index = page->index + thp_nr_pages(page) - 1;
3383 details.single_page = page;
3385 i_mmap_lock_write(mapping);
3386 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3387 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3388 i_mmap_unlock_write(mapping);
3392 * unmap_mapping_pages() - Unmap pages from processes.
3393 * @mapping: The address space containing pages to be unmapped.
3394 * @start: Index of first page to be unmapped.
3395 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3396 * @even_cows: Whether to unmap even private COWed pages.
3398 * Unmap the pages in this address space from any userspace process which
3399 * has them mmaped. Generally, you want to remove COWed pages as well when
3400 * a file is being truncated, but not when invalidating pages from the page
3403 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3404 pgoff_t nr, bool even_cows)
3406 struct zap_details details = { };
3408 details.check_mapping = even_cows ? NULL : mapping;
3409 details.first_index = start;
3410 details.last_index = start + nr - 1;
3411 if (details.last_index < details.first_index)
3412 details.last_index = ULONG_MAX;
3414 i_mmap_lock_write(mapping);
3415 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3416 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3417 i_mmap_unlock_write(mapping);
3419 EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3422 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3423 * address_space corresponding to the specified byte range in the underlying
3426 * @mapping: the address space containing mmaps to be unmapped.
3427 * @holebegin: byte in first page to unmap, relative to the start of
3428 * the underlying file. This will be rounded down to a PAGE_SIZE
3429 * boundary. Note that this is different from truncate_pagecache(), which
3430 * must keep the partial page. In contrast, we must get rid of
3432 * @holelen: size of prospective hole in bytes. This will be rounded
3433 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3435 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3436 * but 0 when invalidating pagecache, don't throw away private data.
3438 void unmap_mapping_range(struct address_space *mapping,
3439 loff_t const holebegin, loff_t const holelen, int even_cows)
3441 pgoff_t hba = holebegin >> PAGE_SHIFT;
3442 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3444 /* Check for overflow. */
3445 if (sizeof(holelen) > sizeof(hlen)) {
3447 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3448 if (holeend & ~(long long)ULONG_MAX)
3449 hlen = ULONG_MAX - hba + 1;
3452 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3454 EXPORT_SYMBOL(unmap_mapping_range);
3457 * Restore a potential device exclusive pte to a working pte entry
3459 static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3461 struct page *page = vmf->page;
3462 struct vm_area_struct *vma = vmf->vma;
3463 struct mmu_notifier_range range;
3465 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
3466 return VM_FAULT_RETRY;
3467 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
3468 vma->vm_mm, vmf->address & PAGE_MASK,
3469 (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3470 mmu_notifier_invalidate_range_start(&range);
3472 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3474 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3475 restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
3477 pte_unmap_unlock(vmf->pte, vmf->ptl);
3480 mmu_notifier_invalidate_range_end(&range);
3485 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3486 * but allow concurrent faults), and pte mapped but not yet locked.
3487 * We return with pte unmapped and unlocked.
3489 * We return with the mmap_lock locked or unlocked in the same cases
3490 * as does filemap_fault().
3492 vm_fault_t do_swap_page(struct vm_fault *vmf)
3494 struct vm_area_struct *vma = vmf->vma;
3495 struct page *page = NULL, *swapcache;
3496 struct swap_info_struct *si = NULL;
3502 void *shadow = NULL;
3504 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3507 entry = pte_to_swp_entry(vmf->orig_pte);
3508 if (unlikely(non_swap_entry(entry))) {
3509 if (is_migration_entry(entry)) {
3510 migration_entry_wait(vma->vm_mm, vmf->pmd,
3512 } else if (is_device_exclusive_entry(entry)) {
3513 vmf->page = pfn_swap_entry_to_page(entry);
3514 ret = remove_device_exclusive_entry(vmf);
3515 } else if (is_device_private_entry(entry)) {
3516 vmf->page = pfn_swap_entry_to_page(entry);
3517 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3518 } else if (is_hwpoison_entry(entry)) {
3519 ret = VM_FAULT_HWPOISON;
3521 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3522 ret = VM_FAULT_SIGBUS;
3527 /* Prevent swapoff from happening to us. */
3528 si = get_swap_device(entry);
3532 delayacct_set_flag(current, DELAYACCT_PF_SWAPIN);
3533 page = lookup_swap_cache(entry, vma, vmf->address);
3537 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3538 __swap_count(entry) == 1) {
3539 /* skip swapcache */
3540 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3543 __SetPageLocked(page);
3544 __SetPageSwapBacked(page);
3546 if (mem_cgroup_swapin_charge_page(page,
3547 vma->vm_mm, GFP_KERNEL, entry)) {
3551 mem_cgroup_swapin_uncharge_swap(entry);
3553 shadow = get_shadow_from_swap_cache(entry);
3555 workingset_refault(page, shadow);
3557 lru_cache_add(page);
3559 /* To provide entry to swap_readpage() */
3560 set_page_private(page, entry.val);
3561 swap_readpage(page, true);
3562 set_page_private(page, 0);
3565 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3572 * Back out if somebody else faulted in this pte
3573 * while we released the pte lock.
3575 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3576 vmf->address, &vmf->ptl);
3577 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3579 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3583 /* Had to read the page from swap area: Major fault */
3584 ret = VM_FAULT_MAJOR;
3585 count_vm_event(PGMAJFAULT);
3586 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3587 } else if (PageHWPoison(page)) {
3589 * hwpoisoned dirty swapcache pages are kept for killing
3590 * owner processes (which may be unknown at hwpoison time)
3592 ret = VM_FAULT_HWPOISON;
3593 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3597 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3599 delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3601 ret |= VM_FAULT_RETRY;
3606 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3607 * release the swapcache from under us. The page pin, and pte_same
3608 * test below, are not enough to exclude that. Even if it is still
3609 * swapcache, we need to check that the page's swap has not changed.
3611 if (unlikely((!PageSwapCache(page) ||
3612 page_private(page) != entry.val)) && swapcache)
3615 page = ksm_might_need_to_copy(page, vma, vmf->address);
3616 if (unlikely(!page)) {
3622 cgroup_throttle_swaprate(page, GFP_KERNEL);
3625 * Back out if somebody else already faulted in this pte.
3627 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3629 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3632 if (unlikely(!PageUptodate(page))) {
3633 ret = VM_FAULT_SIGBUS;
3638 * The page isn't present yet, go ahead with the fault.
3640 * Be careful about the sequence of operations here.
3641 * To get its accounting right, reuse_swap_page() must be called
3642 * while the page is counted on swap but not yet in mapcount i.e.
3643 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3644 * must be called after the swap_free(), or it will never succeed.
3647 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3648 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3649 pte = mk_pte(page, vma->vm_page_prot);
3650 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3651 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3652 vmf->flags &= ~FAULT_FLAG_WRITE;
3653 ret |= VM_FAULT_WRITE;
3654 exclusive = RMAP_EXCLUSIVE;
3656 flush_icache_page(vma, page);
3657 if (pte_swp_soft_dirty(vmf->orig_pte))
3658 pte = pte_mksoft_dirty(pte);
3659 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3660 pte = pte_mkuffd_wp(pte);
3661 pte = pte_wrprotect(pte);
3663 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3664 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3665 vmf->orig_pte = pte;
3667 /* ksm created a completely new copy */
3668 if (unlikely(page != swapcache && swapcache)) {
3669 page_add_new_anon_rmap(page, vma, vmf->address, false);
3670 lru_cache_add_inactive_or_unevictable(page, vma);
3672 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3676 if (mem_cgroup_swap_full(page) ||
3677 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3678 try_to_free_swap(page);
3680 if (page != swapcache && swapcache) {
3682 * Hold the lock to avoid the swap entry to be reused
3683 * until we take the PT lock for the pte_same() check
3684 * (to avoid false positives from pte_same). For
3685 * further safety release the lock after the swap_free
3686 * so that the swap count won't change under a
3687 * parallel locked swapcache.
3689 unlock_page(swapcache);
3690 put_page(swapcache);
3693 if (vmf->flags & FAULT_FLAG_WRITE) {
3694 ret |= do_wp_page(vmf);
3695 if (ret & VM_FAULT_ERROR)
3696 ret &= VM_FAULT_ERROR;
3700 /* No need to invalidate - it was non-present before */
3701 update_mmu_cache(vma, vmf->address, vmf->pte);
3703 pte_unmap_unlock(vmf->pte, vmf->ptl);
3706 put_swap_device(si);
3709 pte_unmap_unlock(vmf->pte, vmf->ptl);
3714 if (page != swapcache && swapcache) {
3715 unlock_page(swapcache);
3716 put_page(swapcache);
3719 put_swap_device(si);
3724 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3725 * but allow concurrent faults), and pte mapped but not yet locked.
3726 * We return with mmap_lock still held, but pte unmapped and unlocked.
3728 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3730 struct vm_area_struct *vma = vmf->vma;
3735 /* File mapping without ->vm_ops ? */
3736 if (vma->vm_flags & VM_SHARED)
3737 return VM_FAULT_SIGBUS;
3740 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3741 * pte_offset_map() on pmds where a huge pmd might be created
3742 * from a different thread.
3744 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3745 * parallel threads are excluded by other means.
3747 * Here we only have mmap_read_lock(mm).
3749 if (pte_alloc(vma->vm_mm, vmf->pmd))
3750 return VM_FAULT_OOM;
3752 /* See comment in handle_pte_fault() */
3753 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3756 /* Use the zero-page for reads */
3757 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3758 !mm_forbids_zeropage(vma->vm_mm)) {
3759 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3760 vma->vm_page_prot));
3761 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3762 vmf->address, &vmf->ptl);
3763 if (!pte_none(*vmf->pte)) {
3764 update_mmu_tlb(vma, vmf->address, vmf->pte);
3767 ret = check_stable_address_space(vma->vm_mm);
3770 /* Deliver the page fault to userland, check inside PT lock */
3771 if (userfaultfd_missing(vma)) {
3772 pte_unmap_unlock(vmf->pte, vmf->ptl);
3773 return handle_userfault(vmf, VM_UFFD_MISSING);
3778 /* Allocate our own private page. */
3779 if (unlikely(anon_vma_prepare(vma)))
3781 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3785 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3787 cgroup_throttle_swaprate(page, GFP_KERNEL);
3790 * The memory barrier inside __SetPageUptodate makes sure that
3791 * preceding stores to the page contents become visible before
3792 * the set_pte_at() write.
3794 __SetPageUptodate(page);
3796 entry = mk_pte(page, vma->vm_page_prot);
3797 entry = pte_sw_mkyoung(entry);
3798 if (vma->vm_flags & VM_WRITE)
3799 entry = pte_mkwrite(pte_mkdirty(entry));
3801 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3803 if (!pte_none(*vmf->pte)) {
3804 update_mmu_cache(vma, vmf->address, vmf->pte);
3808 ret = check_stable_address_space(vma->vm_mm);
3812 /* Deliver the page fault to userland, check inside PT lock */
3813 if (userfaultfd_missing(vma)) {
3814 pte_unmap_unlock(vmf->pte, vmf->ptl);
3816 return handle_userfault(vmf, VM_UFFD_MISSING);
3819 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3820 page_add_new_anon_rmap(page, vma, vmf->address, false);
3821 lru_cache_add_inactive_or_unevictable(page, vma);
3823 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3825 /* No need to invalidate - it was non-present before */
3826 update_mmu_cache(vma, vmf->address, vmf->pte);
3828 pte_unmap_unlock(vmf->pte, vmf->ptl);
3836 return VM_FAULT_OOM;
3840 * The mmap_lock must have been held on entry, and may have been
3841 * released depending on flags and vma->vm_ops->fault() return value.
3842 * See filemap_fault() and __lock_page_retry().
3844 static vm_fault_t __do_fault(struct vm_fault *vmf)
3846 struct vm_area_struct *vma = vmf->vma;
3850 * Preallocate pte before we take page_lock because this might lead to
3851 * deadlocks for memcg reclaim which waits for pages under writeback:
3853 * SetPageWriteback(A)
3859 * wait_on_page_writeback(A)
3860 * SetPageWriteback(B)
3862 * # flush A, B to clear the writeback
3864 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3865 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3866 if (!vmf->prealloc_pte)
3867 return VM_FAULT_OOM;
3868 smp_wmb(); /* See comment in __pte_alloc() */
3871 ret = vma->vm_ops->fault(vmf);
3872 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3873 VM_FAULT_DONE_COW)))
3876 if (unlikely(PageHWPoison(vmf->page))) {
3877 struct page *page = vmf->page;
3878 vm_fault_t poisonret = VM_FAULT_HWPOISON;
3879 if (ret & VM_FAULT_LOCKED) {
3880 if (page_mapped(page))
3881 unmap_mapping_pages(page_mapping(page),
3882 page->index, 1, false);
3883 /* Retry if a clean page was removed from the cache. */
3884 if (invalidate_inode_page(page))
3885 poisonret = VM_FAULT_NOPAGE;
3893 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3894 lock_page(vmf->page);
3896 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3901 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3902 static void deposit_prealloc_pte(struct vm_fault *vmf)
3904 struct vm_area_struct *vma = vmf->vma;
3906 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3908 * We are going to consume the prealloc table,
3909 * count that as nr_ptes.
3911 mm_inc_nr_ptes(vma->vm_mm);
3912 vmf->prealloc_pte = NULL;
3915 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3917 struct vm_area_struct *vma = vmf->vma;
3918 bool write = vmf->flags & FAULT_FLAG_WRITE;
3919 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3922 vm_fault_t ret = VM_FAULT_FALLBACK;
3924 if (!transhuge_vma_suitable(vma, haddr))
3927 page = compound_head(page);
3928 if (compound_order(page) != HPAGE_PMD_ORDER)
3932 * Just backoff if any subpage of a THP is corrupted otherwise
3933 * the corrupted page may mapped by PMD silently to escape the
3934 * check. This kind of THP just can be PTE mapped. Access to
3935 * the corrupted subpage should trigger SIGBUS as expected.
3937 if (unlikely(PageHasHWPoisoned(page)))
3941 * Archs like ppc64 need additional space to store information
3942 * related to pte entry. Use the preallocated table for that.
3944 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3945 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3946 if (!vmf->prealloc_pte)
3947 return VM_FAULT_OOM;
3948 smp_wmb(); /* See comment in __pte_alloc() */
3951 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3952 if (unlikely(!pmd_none(*vmf->pmd)))
3955 for (i = 0; i < HPAGE_PMD_NR; i++)
3956 flush_icache_page(vma, page + i);
3958 entry = mk_huge_pmd(page, vma->vm_page_prot);
3960 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3962 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3963 page_add_file_rmap(page, true);
3965 * deposit and withdraw with pmd lock held
3967 if (arch_needs_pgtable_deposit())
3968 deposit_prealloc_pte(vmf);
3970 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3972 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3974 /* fault is handled */
3976 count_vm_event(THP_FILE_MAPPED);
3978 spin_unlock(vmf->ptl);
3982 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3984 return VM_FAULT_FALLBACK;
3988 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3990 struct vm_area_struct *vma = vmf->vma;
3991 bool write = vmf->flags & FAULT_FLAG_WRITE;
3992 bool prefault = vmf->address != addr;
3995 flush_icache_page(vma, page);
3996 entry = mk_pte(page, vma->vm_page_prot);
3998 if (prefault && arch_wants_old_prefaulted_pte())
3999 entry = pte_mkold(entry);
4001 entry = pte_sw_mkyoung(entry);
4004 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
4005 /* copy-on-write page */
4006 if (write && !(vma->vm_flags & VM_SHARED)) {
4007 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
4008 page_add_new_anon_rmap(page, vma, addr, false);
4009 lru_cache_add_inactive_or_unevictable(page, vma);
4011 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
4012 page_add_file_rmap(page, false);
4014 set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
4018 * finish_fault - finish page fault once we have prepared the page to fault
4020 * @vmf: structure describing the fault
4022 * This function handles all that is needed to finish a page fault once the
4023 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4024 * given page, adds reverse page mapping, handles memcg charges and LRU
4027 * The function expects the page to be locked and on success it consumes a
4028 * reference of a page being mapped (for the PTE which maps it).
4030 * Return: %0 on success, %VM_FAULT_ code in case of error.
4032 vm_fault_t finish_fault(struct vm_fault *vmf)
4034 struct vm_area_struct *vma = vmf->vma;
4038 /* Did we COW the page? */
4039 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4040 page = vmf->cow_page;
4045 * check even for read faults because we might have lost our CoWed
4048 if (!(vma->vm_flags & VM_SHARED)) {
4049 ret = check_stable_address_space(vma->vm_mm);
4054 if (pmd_none(*vmf->pmd)) {
4055 if (PageTransCompound(page)) {
4056 ret = do_set_pmd(vmf, page);
4057 if (ret != VM_FAULT_FALLBACK)
4061 if (vmf->prealloc_pte) {
4062 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4063 if (likely(pmd_none(*vmf->pmd))) {
4064 mm_inc_nr_ptes(vma->vm_mm);
4065 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4066 vmf->prealloc_pte = NULL;
4068 spin_unlock(vmf->ptl);
4069 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
4070 return VM_FAULT_OOM;
4075 * See comment in handle_pte_fault() for how this scenario happens, we
4076 * need to return NOPAGE so that we drop this page.
4078 if (pmd_devmap_trans_unstable(vmf->pmd))
4079 return VM_FAULT_NOPAGE;
4081 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4082 vmf->address, &vmf->ptl);
4084 /* Re-check under ptl */
4085 if (likely(pte_none(*vmf->pte)))
4086 do_set_pte(vmf, page, vmf->address);
4088 ret = VM_FAULT_NOPAGE;
4090 update_mmu_tlb(vma, vmf->address, vmf->pte);
4091 pte_unmap_unlock(vmf->pte, vmf->ptl);
4095 static unsigned long fault_around_bytes __read_mostly =
4096 rounddown_pow_of_two(65536);
4098 #ifdef CONFIG_DEBUG_FS
4099 static int fault_around_bytes_get(void *data, u64 *val)
4101 *val = fault_around_bytes;
4106 * fault_around_bytes must be rounded down to the nearest page order as it's
4107 * what do_fault_around() expects to see.
4109 static int fault_around_bytes_set(void *data, u64 val)
4111 if (val / PAGE_SIZE > PTRS_PER_PTE)
4113 if (val > PAGE_SIZE)
4114 fault_around_bytes = rounddown_pow_of_two(val);
4116 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
4119 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4120 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4122 static int __init fault_around_debugfs(void)
4124 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4125 &fault_around_bytes_fops);
4128 late_initcall(fault_around_debugfs);
4132 * do_fault_around() tries to map few pages around the fault address. The hope
4133 * is that the pages will be needed soon and this will lower the number of
4136 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4137 * not ready to be mapped: not up-to-date, locked, etc.
4139 * This function is called with the page table lock taken. In the split ptlock
4140 * case the page table lock only protects only those entries which belong to
4141 * the page table corresponding to the fault address.
4143 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4146 * fault_around_bytes defines how many bytes we'll try to map.
4147 * do_fault_around() expects it to be set to a power of two less than or equal
4150 * The virtual address of the area that we map is naturally aligned to
4151 * fault_around_bytes rounded down to the machine page size
4152 * (and therefore to page order). This way it's easier to guarantee
4153 * that we don't cross page table boundaries.
4155 static vm_fault_t do_fault_around(struct vm_fault *vmf)
4157 unsigned long address = vmf->address, nr_pages, mask;
4158 pgoff_t start_pgoff = vmf->pgoff;
4162 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4163 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4165 address = max(address & mask, vmf->vma->vm_start);
4166 off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4170 * end_pgoff is either the end of the page table, the end of
4171 * the vma or nr_pages from start_pgoff, depending what is nearest.
4173 end_pgoff = start_pgoff -
4174 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4176 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4177 start_pgoff + nr_pages - 1);
4179 if (pmd_none(*vmf->pmd)) {
4180 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4181 if (!vmf->prealloc_pte)
4182 return VM_FAULT_OOM;
4183 smp_wmb(); /* See comment in __pte_alloc() */
4186 return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4189 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4191 struct vm_area_struct *vma = vmf->vma;
4195 * Let's call ->map_pages() first and use ->fault() as fallback
4196 * if page by the offset is not ready to be mapped (cold cache or
4199 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4200 if (likely(!userfaultfd_minor(vmf->vma))) {
4201 ret = do_fault_around(vmf);
4207 ret = __do_fault(vmf);
4208 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4211 ret |= finish_fault(vmf);
4212 unlock_page(vmf->page);
4213 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4214 put_page(vmf->page);
4218 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4220 struct vm_area_struct *vma = vmf->vma;
4223 if (unlikely(anon_vma_prepare(vma)))
4224 return VM_FAULT_OOM;
4226 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4228 return VM_FAULT_OOM;
4230 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4231 put_page(vmf->cow_page);
4232 return VM_FAULT_OOM;
4234 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4236 ret = __do_fault(vmf);
4237 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4239 if (ret & VM_FAULT_DONE_COW)
4242 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4243 __SetPageUptodate(vmf->cow_page);
4245 ret |= finish_fault(vmf);
4246 unlock_page(vmf->page);
4247 put_page(vmf->page);
4248 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4252 put_page(vmf->cow_page);
4256 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4258 struct vm_area_struct *vma = vmf->vma;
4259 vm_fault_t ret, tmp;
4261 ret = __do_fault(vmf);
4262 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4266 * Check if the backing address space wants to know that the page is
4267 * about to become writable
4269 if (vma->vm_ops->page_mkwrite) {
4270 unlock_page(vmf->page);
4271 tmp = do_page_mkwrite(vmf);
4272 if (unlikely(!tmp ||
4273 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4274 put_page(vmf->page);
4279 ret |= finish_fault(vmf);
4280 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4282 unlock_page(vmf->page);
4283 put_page(vmf->page);
4287 ret |= fault_dirty_shared_page(vmf);
4292 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4293 * but allow concurrent faults).
4294 * The mmap_lock may have been released depending on flags and our
4295 * return value. See filemap_fault() and __lock_page_or_retry().
4296 * If mmap_lock is released, vma may become invalid (for example
4297 * by other thread calling munmap()).
4299 static vm_fault_t do_fault(struct vm_fault *vmf)
4301 struct vm_area_struct *vma = vmf->vma;
4302 struct mm_struct *vm_mm = vma->vm_mm;
4306 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4308 if (!vma->vm_ops->fault) {
4310 * If we find a migration pmd entry or a none pmd entry, which
4311 * should never happen, return SIGBUS
4313 if (unlikely(!pmd_present(*vmf->pmd)))
4314 ret = VM_FAULT_SIGBUS;
4316 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4321 * Make sure this is not a temporary clearing of pte
4322 * by holding ptl and checking again. A R/M/W update
4323 * of pte involves: take ptl, clearing the pte so that
4324 * we don't have concurrent modification by hardware
4325 * followed by an update.
4327 if (unlikely(pte_none(*vmf->pte)))
4328 ret = VM_FAULT_SIGBUS;
4330 ret = VM_FAULT_NOPAGE;
4332 pte_unmap_unlock(vmf->pte, vmf->ptl);
4334 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4335 ret = do_read_fault(vmf);
4336 else if (!(vma->vm_flags & VM_SHARED))
4337 ret = do_cow_fault(vmf);
4339 ret = do_shared_fault(vmf);
4341 /* preallocated pagetable is unused: free it */
4342 if (vmf->prealloc_pte) {
4343 pte_free(vm_mm, vmf->prealloc_pte);
4344 vmf->prealloc_pte = NULL;
4349 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4350 unsigned long addr, int page_nid, int *flags)
4354 count_vm_numa_event(NUMA_HINT_FAULTS);
4355 if (page_nid == numa_node_id()) {
4356 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4357 *flags |= TNF_FAULT_LOCAL;
4360 return mpol_misplaced(page, vma, addr);
4363 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4365 struct vm_area_struct *vma = vmf->vma;
4366 struct page *page = NULL;
4367 int page_nid = NUMA_NO_NODE;
4371 bool was_writable = pte_savedwrite(vmf->orig_pte);
4375 * The "pte" at this point cannot be used safely without
4376 * validation through pte_unmap_same(). It's of NUMA type but
4377 * the pfn may be screwed if the read is non atomic.
4379 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4380 spin_lock(vmf->ptl);
4381 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4382 pte_unmap_unlock(vmf->pte, vmf->ptl);
4386 /* Get the normal PTE */
4387 old_pte = ptep_get(vmf->pte);
4388 pte = pte_modify(old_pte, vma->vm_page_prot);
4390 page = vm_normal_page(vma, vmf->address, pte);
4394 /* TODO: handle PTE-mapped THP */
4395 if (PageCompound(page))
4399 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4400 * much anyway since they can be in shared cache state. This misses
4401 * the case where a mapping is writable but the process never writes
4402 * to it but pte_write gets cleared during protection updates and
4403 * pte_dirty has unpredictable behaviour between PTE scan updates,
4404 * background writeback, dirty balancing and application behaviour.
4407 flags |= TNF_NO_GROUP;
4410 * Flag if the page is shared between multiple address spaces. This
4411 * is later used when determining whether to group tasks together
4413 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4414 flags |= TNF_SHARED;
4416 last_cpupid = page_cpupid_last(page);
4417 page_nid = page_to_nid(page);
4418 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4420 if (target_nid == NUMA_NO_NODE) {
4424 pte_unmap_unlock(vmf->pte, vmf->ptl);
4426 /* Migrate to the requested node */
4427 if (migrate_misplaced_page(page, vma, target_nid)) {
4428 page_nid = target_nid;
4429 flags |= TNF_MIGRATED;
4431 flags |= TNF_MIGRATE_FAIL;
4432 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4433 spin_lock(vmf->ptl);
4434 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4435 pte_unmap_unlock(vmf->pte, vmf->ptl);
4442 if (page_nid != NUMA_NO_NODE)
4443 task_numa_fault(last_cpupid, page_nid, 1, flags);
4447 * Make it present again, depending on how arch implements
4448 * non-accessible ptes, some can allow access by kernel mode.
4450 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4451 pte = pte_modify(old_pte, vma->vm_page_prot);
4452 pte = pte_mkyoung(pte);
4454 pte = pte_mkwrite(pte);
4455 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4456 update_mmu_cache(vma, vmf->address, vmf->pte);
4457 pte_unmap_unlock(vmf->pte, vmf->ptl);
4461 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4463 if (vma_is_anonymous(vmf->vma))
4464 return do_huge_pmd_anonymous_page(vmf);
4465 if (vmf->vma->vm_ops->huge_fault)
4466 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4467 return VM_FAULT_FALLBACK;
4470 /* `inline' is required to avoid gcc 4.1.2 build error */
4471 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4473 if (vma_is_anonymous(vmf->vma)) {
4474 if (userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd))
4475 return handle_userfault(vmf, VM_UFFD_WP);
4476 return do_huge_pmd_wp_page(vmf);
4478 if (vmf->vma->vm_ops->huge_fault) {
4479 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4481 if (!(ret & VM_FAULT_FALLBACK))
4485 /* COW or write-notify handled on pte level: split pmd. */
4486 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4488 return VM_FAULT_FALLBACK;
4491 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4493 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4494 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4495 /* No support for anonymous transparent PUD pages yet */
4496 if (vma_is_anonymous(vmf->vma))
4497 return VM_FAULT_FALLBACK;
4498 if (vmf->vma->vm_ops->huge_fault)
4499 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4500 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4501 return VM_FAULT_FALLBACK;
4504 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4506 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4507 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4508 /* No support for anonymous transparent PUD pages yet */
4509 if (vma_is_anonymous(vmf->vma))
4511 if (vmf->vma->vm_ops->huge_fault) {
4512 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4514 if (!(ret & VM_FAULT_FALLBACK))
4518 /* COW or write-notify not handled on PUD level: split pud.*/
4519 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4520 #endif /* CONFIG_TRANSPARENT_HUGEPAGE && CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
4521 return VM_FAULT_FALLBACK;
4525 * These routines also need to handle stuff like marking pages dirty
4526 * and/or accessed for architectures that don't do it in hardware (most
4527 * RISC architectures). The early dirtying is also good on the i386.
4529 * There is also a hook called "update_mmu_cache()" that architectures
4530 * with external mmu caches can use to update those (ie the Sparc or
4531 * PowerPC hashed page tables that act as extended TLBs).
4533 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4534 * concurrent faults).
4536 * The mmap_lock may have been released depending on flags and our return value.
4537 * See filemap_fault() and __lock_page_or_retry().
4539 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4543 if (unlikely(pmd_none(*vmf->pmd))) {
4545 * Leave __pte_alloc() until later: because vm_ops->fault may
4546 * want to allocate huge page, and if we expose page table
4547 * for an instant, it will be difficult to retract from
4548 * concurrent faults and from rmap lookups.
4553 * If a huge pmd materialized under us just retry later. Use
4554 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4555 * of pmd_trans_huge() to ensure the pmd didn't become
4556 * pmd_trans_huge under us and then back to pmd_none, as a
4557 * result of MADV_DONTNEED running immediately after a huge pmd
4558 * fault in a different thread of this mm, in turn leading to a
4559 * misleading pmd_trans_huge() retval. All we have to ensure is
4560 * that it is a regular pmd that we can walk with
4561 * pte_offset_map() and we can do that through an atomic read
4562 * in C, which is what pmd_trans_unstable() provides.
4564 if (pmd_devmap_trans_unstable(vmf->pmd))
4567 * A regular pmd is established and it can't morph into a huge
4568 * pmd from under us anymore at this point because we hold the
4569 * mmap_lock read mode and khugepaged takes it in write mode.
4570 * So now it's safe to run pte_offset_map().
4572 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4573 vmf->orig_pte = *vmf->pte;
4576 * some architectures can have larger ptes than wordsize,
4577 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4578 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4579 * accesses. The code below just needs a consistent view
4580 * for the ifs and we later double check anyway with the
4581 * ptl lock held. So here a barrier will do.
4584 if (pte_none(vmf->orig_pte)) {
4585 pte_unmap(vmf->pte);
4591 if (vma_is_anonymous(vmf->vma))
4592 return do_anonymous_page(vmf);
4594 return do_fault(vmf);
4597 if (!pte_present(vmf->orig_pte))
4598 return do_swap_page(vmf);
4600 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4601 return do_numa_page(vmf);
4603 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4604 spin_lock(vmf->ptl);
4605 entry = vmf->orig_pte;
4606 if (unlikely(!pte_same(*vmf->pte, entry))) {
4607 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4610 if (vmf->flags & FAULT_FLAG_WRITE) {
4611 if (!pte_write(entry))
4612 return do_wp_page(vmf);
4613 entry = pte_mkdirty(entry);
4615 entry = pte_mkyoung(entry);
4616 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4617 vmf->flags & FAULT_FLAG_WRITE)) {
4618 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4620 /* Skip spurious TLB flush for retried page fault */
4621 if (vmf->flags & FAULT_FLAG_TRIED)
4624 * This is needed only for protection faults but the arch code
4625 * is not yet telling us if this is a protection fault or not.
4626 * This still avoids useless tlb flushes for .text page faults
4629 if (vmf->flags & FAULT_FLAG_WRITE)
4630 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4633 pte_unmap_unlock(vmf->pte, vmf->ptl);
4638 * By the time we get here, we already hold the mm semaphore
4640 * The mmap_lock may have been released depending on flags and our
4641 * return value. See filemap_fault() and __lock_page_or_retry().
4643 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4644 unsigned long address, unsigned int flags)
4646 struct vm_fault vmf = {
4648 .address = address & PAGE_MASK,
4650 .pgoff = linear_page_index(vma, address),
4651 .gfp_mask = __get_fault_gfp_mask(vma),
4653 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4654 struct mm_struct *mm = vma->vm_mm;
4659 pgd = pgd_offset(mm, address);
4660 p4d = p4d_alloc(mm, pgd, address);
4662 return VM_FAULT_OOM;
4664 vmf.pud = pud_alloc(mm, p4d, address);
4666 return VM_FAULT_OOM;
4668 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4669 ret = create_huge_pud(&vmf);
4670 if (!(ret & VM_FAULT_FALLBACK))
4673 pud_t orig_pud = *vmf.pud;
4676 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4678 /* NUMA case for anonymous PUDs would go here */
4680 if (dirty && !pud_write(orig_pud)) {
4681 ret = wp_huge_pud(&vmf, orig_pud);
4682 if (!(ret & VM_FAULT_FALLBACK))
4685 huge_pud_set_accessed(&vmf, orig_pud);
4691 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4693 return VM_FAULT_OOM;
4695 /* Huge pud page fault raced with pmd_alloc? */
4696 if (pud_trans_unstable(vmf.pud))
4699 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4700 ret = create_huge_pmd(&vmf);
4701 if (!(ret & VM_FAULT_FALLBACK))
4704 vmf.orig_pmd = *vmf.pmd;
4707 if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
4708 VM_BUG_ON(thp_migration_supported() &&
4709 !is_pmd_migration_entry(vmf.orig_pmd));
4710 if (is_pmd_migration_entry(vmf.orig_pmd))
4711 pmd_migration_entry_wait(mm, vmf.pmd);
4714 if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
4715 if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
4716 return do_huge_pmd_numa_page(&vmf);
4718 if (dirty && !pmd_write(vmf.orig_pmd)) {
4719 ret = wp_huge_pmd(&vmf);
4720 if (!(ret & VM_FAULT_FALLBACK))
4723 huge_pmd_set_accessed(&vmf);
4729 return handle_pte_fault(&vmf);
4733 * mm_account_fault - Do page fault accounting
4735 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4736 * of perf event counters, but we'll still do the per-task accounting to
4737 * the task who triggered this page fault.
4738 * @address: the faulted address.
4739 * @flags: the fault flags.
4740 * @ret: the fault retcode.
4742 * This will take care of most of the page fault accounting. Meanwhile, it
4743 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4744 * updates. However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4745 * still be in per-arch page fault handlers at the entry of page fault.
4747 static inline void mm_account_fault(struct pt_regs *regs,
4748 unsigned long address, unsigned int flags,
4754 * We don't do accounting for some specific faults:
4756 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4757 * includes arch_vma_access_permitted() failing before reaching here.
4758 * So this is not a "this many hardware page faults" counter. We
4759 * should use the hw profiling for that.
4761 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4762 * once they're completed.
4764 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4768 * We define the fault as a major fault when the final successful fault
4769 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4770 * handle it immediately previously).
4772 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4780 * If the fault is done for GUP, regs will be NULL. We only do the
4781 * accounting for the per thread fault counters who triggered the
4782 * fault, and we skip the perf event updates.
4788 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4790 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4794 * By the time we get here, we already hold the mm semaphore
4796 * The mmap_lock may have been released depending on flags and our
4797 * return value. See filemap_fault() and __lock_page_or_retry().
4799 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4800 unsigned int flags, struct pt_regs *regs)
4804 __set_current_state(TASK_RUNNING);
4806 count_vm_event(PGFAULT);
4807 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4809 /* do counter updates before entering really critical section. */
4810 check_sync_rss_stat(current);
4812 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4813 flags & FAULT_FLAG_INSTRUCTION,
4814 flags & FAULT_FLAG_REMOTE))
4815 return VM_FAULT_SIGSEGV;
4818 * Enable the memcg OOM handling for faults triggered in user
4819 * space. Kernel faults are handled more gracefully.
4821 if (flags & FAULT_FLAG_USER)
4822 mem_cgroup_enter_user_fault();
4824 if (unlikely(is_vm_hugetlb_page(vma)))
4825 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4827 ret = __handle_mm_fault(vma, address, flags);
4829 if (flags & FAULT_FLAG_USER) {
4830 mem_cgroup_exit_user_fault();
4832 * The task may have entered a memcg OOM situation but
4833 * if the allocation error was handled gracefully (no
4834 * VM_FAULT_OOM), there is no need to kill anything.
4835 * Just clean up the OOM state peacefully.
4837 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4838 mem_cgroup_oom_synchronize(false);
4841 mm_account_fault(regs, address, flags, ret);
4845 EXPORT_SYMBOL_GPL(handle_mm_fault);
4847 #ifndef __PAGETABLE_P4D_FOLDED
4849 * Allocate p4d page table.
4850 * We've already handled the fast-path in-line.
4852 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4854 p4d_t *new = p4d_alloc_one(mm, address);
4858 smp_wmb(); /* See comment in __pte_alloc */
4860 spin_lock(&mm->page_table_lock);
4861 if (pgd_present(*pgd)) /* Another has populated it */
4864 pgd_populate(mm, pgd, new);
4865 spin_unlock(&mm->page_table_lock);
4868 #endif /* __PAGETABLE_P4D_FOLDED */
4870 #ifndef __PAGETABLE_PUD_FOLDED
4872 * Allocate page upper directory.
4873 * We've already handled the fast-path in-line.
4875 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4877 pud_t *new = pud_alloc_one(mm, address);
4881 smp_wmb(); /* See comment in __pte_alloc */
4883 spin_lock(&mm->page_table_lock);
4884 if (!p4d_present(*p4d)) {
4886 p4d_populate(mm, p4d, new);
4887 } else /* Another has populated it */
4889 spin_unlock(&mm->page_table_lock);
4892 #endif /* __PAGETABLE_PUD_FOLDED */
4894 #ifndef __PAGETABLE_PMD_FOLDED
4896 * Allocate page middle directory.
4897 * We've already handled the fast-path in-line.
4899 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4902 pmd_t *new = pmd_alloc_one(mm, address);
4906 smp_wmb(); /* See comment in __pte_alloc */
4908 ptl = pud_lock(mm, pud);
4909 if (!pud_present(*pud)) {
4911 pud_populate(mm, pud, new);
4912 } else /* Another has populated it */
4917 #endif /* __PAGETABLE_PMD_FOLDED */
4919 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4920 struct mmu_notifier_range *range, pte_t **ptepp,
4921 pmd_t **pmdpp, spinlock_t **ptlp)
4929 pgd = pgd_offset(mm, address);
4930 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4933 p4d = p4d_offset(pgd, address);
4934 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4937 pud = pud_offset(p4d, address);
4938 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4941 pmd = pmd_offset(pud, address);
4942 VM_BUG_ON(pmd_trans_huge(*pmd));
4944 if (pmd_huge(*pmd)) {
4949 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4950 NULL, mm, address & PMD_MASK,
4951 (address & PMD_MASK) + PMD_SIZE);
4952 mmu_notifier_invalidate_range_start(range);
4954 *ptlp = pmd_lock(mm, pmd);
4955 if (pmd_huge(*pmd)) {
4961 mmu_notifier_invalidate_range_end(range);
4964 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4968 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4969 address & PAGE_MASK,
4970 (address & PAGE_MASK) + PAGE_SIZE);
4971 mmu_notifier_invalidate_range_start(range);
4973 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4974 if (!pte_present(*ptep))
4979 pte_unmap_unlock(ptep, *ptlp);
4981 mmu_notifier_invalidate_range_end(range);
4987 * follow_pte - look up PTE at a user virtual address
4988 * @mm: the mm_struct of the target address space
4989 * @address: user virtual address
4990 * @ptepp: location to store found PTE
4991 * @ptlp: location to store the lock for the PTE
4993 * On a successful return, the pointer to the PTE is stored in @ptepp;
4994 * the corresponding lock is taken and its location is stored in @ptlp.
4995 * The contents of the PTE are only stable until @ptlp is released;
4996 * any further use, if any, must be protected against invalidation
4997 * with MMU notifiers.
4999 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
5000 * should be taken for read.
5002 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
5003 * it is not a good general-purpose API.
5005 * Return: zero on success, -ve otherwise.
5007 int follow_pte(struct mm_struct *mm, unsigned long address,
5008 pte_t **ptepp, spinlock_t **ptlp)
5010 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
5012 EXPORT_SYMBOL_GPL(follow_pte);
5015 * follow_pfn - look up PFN at a user virtual address
5016 * @vma: memory mapping
5017 * @address: user virtual address
5018 * @pfn: location to store found PFN
5020 * Only IO mappings and raw PFN mappings are allowed.
5022 * This function does not allow the caller to read the permissions
5023 * of the PTE. Do not use it.
5025 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5027 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5034 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5037 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5040 *pfn = pte_pfn(*ptep);
5041 pte_unmap_unlock(ptep, ptl);
5044 EXPORT_SYMBOL(follow_pfn);
5046 #ifdef CONFIG_HAVE_IOREMAP_PROT
5047 int follow_phys(struct vm_area_struct *vma,
5048 unsigned long address, unsigned int flags,
5049 unsigned long *prot, resource_size_t *phys)
5055 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5058 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5062 if ((flags & FOLL_WRITE) && !pte_write(pte))
5065 *prot = pgprot_val(pte_pgprot(pte));
5066 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5070 pte_unmap_unlock(ptep, ptl);
5076 * generic_access_phys - generic implementation for iomem mmap access
5077 * @vma: the vma to access
5078 * @addr: userspace address, not relative offset within @vma
5079 * @buf: buffer to read/write
5080 * @len: length of transfer
5081 * @write: set to FOLL_WRITE when writing, otherwise reading
5083 * This is a generic implementation for &vm_operations_struct.access for an
5084 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5087 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5088 void *buf, int len, int write)
5090 resource_size_t phys_addr;
5091 unsigned long prot = 0;
5092 void __iomem *maddr;
5095 int offset = offset_in_page(addr);
5098 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5102 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5105 pte_unmap_unlock(ptep, ptl);
5107 prot = pgprot_val(pte_pgprot(pte));
5108 phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5110 if ((write & FOLL_WRITE) && !pte_write(pte))
5113 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5117 if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5120 if (!pte_same(pte, *ptep)) {
5121 pte_unmap_unlock(ptep, ptl);
5128 memcpy_toio(maddr + offset, buf, len);
5130 memcpy_fromio(buf, maddr + offset, len);
5132 pte_unmap_unlock(ptep, ptl);
5138 EXPORT_SYMBOL_GPL(generic_access_phys);
5142 * Access another process' address space as given in mm.
5144 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5145 int len, unsigned int gup_flags)
5147 struct vm_area_struct *vma;
5148 void *old_buf = buf;
5149 int write = gup_flags & FOLL_WRITE;
5151 if (mmap_read_lock_killable(mm))
5154 /* ignore errors, just check how much was successfully transferred */
5156 int bytes, ret, offset;
5158 struct page *page = NULL;
5160 ret = get_user_pages_remote(mm, addr, 1,
5161 gup_flags, &page, &vma, NULL);
5163 #ifndef CONFIG_HAVE_IOREMAP_PROT
5167 * Check if this is a VM_IO | VM_PFNMAP VMA, which
5168 * we can access using slightly different code.
5170 vma = vma_lookup(mm, addr);
5173 if (vma->vm_ops && vma->vm_ops->access)
5174 ret = vma->vm_ops->access(vma, addr, buf,
5182 offset = addr & (PAGE_SIZE-1);
5183 if (bytes > PAGE_SIZE-offset)
5184 bytes = PAGE_SIZE-offset;
5188 copy_to_user_page(vma, page, addr,
5189 maddr + offset, buf, bytes);
5190 set_page_dirty_lock(page);
5192 copy_from_user_page(vma, page, addr,
5193 buf, maddr + offset, bytes);
5202 mmap_read_unlock(mm);
5204 return buf - old_buf;
5208 * access_remote_vm - access another process' address space
5209 * @mm: the mm_struct of the target address space
5210 * @addr: start address to access
5211 * @buf: source or destination buffer
5212 * @len: number of bytes to transfer
5213 * @gup_flags: flags modifying lookup behaviour
5215 * The caller must hold a reference on @mm.
5217 * Return: number of bytes copied from source to destination.
5219 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5220 void *buf, int len, unsigned int gup_flags)
5222 return __access_remote_vm(mm, addr, buf, len, gup_flags);
5226 * Access another process' address space.
5227 * Source/target buffer must be kernel space,
5228 * Do not walk the page table directly, use get_user_pages
5230 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5231 void *buf, int len, unsigned int gup_flags)
5233 struct mm_struct *mm;
5236 mm = get_task_mm(tsk);
5240 ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5246 EXPORT_SYMBOL_GPL(access_process_vm);
5249 * Print the name of a VMA.
5251 void print_vma_addr(char *prefix, unsigned long ip)
5253 struct mm_struct *mm = current->mm;
5254 struct vm_area_struct *vma;
5257 * we might be running from an atomic context so we cannot sleep
5259 if (!mmap_read_trylock(mm))
5262 vma = find_vma(mm, ip);
5263 if (vma && vma->vm_file) {
5264 struct file *f = vma->vm_file;
5265 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5269 p = file_path(f, buf, PAGE_SIZE);
5272 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5274 vma->vm_end - vma->vm_start);
5275 free_page((unsigned long)buf);
5278 mmap_read_unlock(mm);
5281 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5282 void __might_fault(const char *file, int line)
5285 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5286 * holding the mmap_lock, this is safe because kernel memory doesn't
5287 * get paged out, therefore we'll never actually fault, and the
5288 * below annotations will generate false positives.
5290 if (uaccess_kernel())
5292 if (pagefault_disabled())
5294 __might_sleep(file, line, 0);
5295 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5297 might_lock_read(¤t->mm->mmap_lock);
5300 EXPORT_SYMBOL(__might_fault);
5303 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5305 * Process all subpages of the specified huge page with the specified
5306 * operation. The target subpage will be processed last to keep its
5309 static inline void process_huge_page(
5310 unsigned long addr_hint, unsigned int pages_per_huge_page,
5311 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5315 unsigned long addr = addr_hint &
5316 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5318 /* Process target subpage last to keep its cache lines hot */
5320 n = (addr_hint - addr) / PAGE_SIZE;
5321 if (2 * n <= pages_per_huge_page) {
5322 /* If target subpage in first half of huge page */
5325 /* Process subpages at the end of huge page */
5326 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5328 process_subpage(addr + i * PAGE_SIZE, i, arg);
5331 /* If target subpage in second half of huge page */
5332 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5333 l = pages_per_huge_page - n;
5334 /* Process subpages at the begin of huge page */
5335 for (i = 0; i < base; i++) {
5337 process_subpage(addr + i * PAGE_SIZE, i, arg);
5341 * Process remaining subpages in left-right-left-right pattern
5342 * towards the target subpage
5344 for (i = 0; i < l; i++) {
5345 int left_idx = base + i;
5346 int right_idx = base + 2 * l - 1 - i;
5349 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5351 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5355 static void clear_gigantic_page(struct page *page,
5357 unsigned int pages_per_huge_page)
5360 struct page *p = page;
5363 for (i = 0; i < pages_per_huge_page;
5364 i++, p = mem_map_next(p, page, i)) {
5366 clear_user_highpage(p, addr + i * PAGE_SIZE);
5370 static void clear_subpage(unsigned long addr, int idx, void *arg)
5372 struct page *page = arg;
5374 clear_user_highpage(page + idx, addr);
5377 void clear_huge_page(struct page *page,
5378 unsigned long addr_hint, unsigned int pages_per_huge_page)
5380 unsigned long addr = addr_hint &
5381 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5383 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5384 clear_gigantic_page(page, addr, pages_per_huge_page);
5388 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5391 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5393 struct vm_area_struct *vma,
5394 unsigned int pages_per_huge_page)
5397 struct page *dst_base = dst;
5398 struct page *src_base = src;
5400 for (i = 0; i < pages_per_huge_page; ) {
5402 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5405 dst = mem_map_next(dst, dst_base, i);
5406 src = mem_map_next(src, src_base, i);
5410 struct copy_subpage_arg {
5413 struct vm_area_struct *vma;
5416 static void copy_subpage(unsigned long addr, int idx, void *arg)
5418 struct copy_subpage_arg *copy_arg = arg;
5420 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5421 addr, copy_arg->vma);
5424 void copy_user_huge_page(struct page *dst, struct page *src,
5425 unsigned long addr_hint, struct vm_area_struct *vma,
5426 unsigned int pages_per_huge_page)
5428 unsigned long addr = addr_hint &
5429 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5430 struct copy_subpage_arg arg = {
5436 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5437 copy_user_gigantic_page(dst, src, addr, vma,
5438 pages_per_huge_page);
5442 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5445 long copy_huge_page_from_user(struct page *dst_page,
5446 const void __user *usr_src,
5447 unsigned int pages_per_huge_page,
5448 bool allow_pagefault)
5450 void *src = (void *)usr_src;
5452 unsigned long i, rc = 0;
5453 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5454 struct page *subpage = dst_page;
5456 for (i = 0; i < pages_per_huge_page;
5457 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5458 if (allow_pagefault)
5459 page_kaddr = kmap(subpage);
5461 page_kaddr = kmap_atomic(subpage);
5462 rc = copy_from_user(page_kaddr,
5463 (const void __user *)(src + i * PAGE_SIZE),
5465 if (allow_pagefault)
5468 kunmap_atomic(page_kaddr);
5470 ret_val -= (PAGE_SIZE - rc);
5474 flush_dcache_page(subpage);
5480 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5482 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5484 static struct kmem_cache *page_ptl_cachep;
5486 void __init ptlock_cache_init(void)
5488 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5492 bool ptlock_alloc(struct page *page)
5496 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5503 void ptlock_free(struct page *page)
5505 kmem_cache_free(page_ptl_cachep, page->ptl);