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.
93 #ifndef CONFIG_NEED_MULTIPLE_NODES
94 /* use the per-pgdat data instead for discontigmem - mbligh */
95 unsigned long max_mapnr;
96 EXPORT_SYMBOL(max_mapnr);
99 EXPORT_SYMBOL(mem_map);
103 * A number of key systems in x86 including ioremap() rely on the assumption
104 * that high_memory defines the upper bound on direct map memory, then end
105 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
106 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
110 EXPORT_SYMBOL(high_memory);
113 * Randomize the address space (stacks, mmaps, brk, etc.).
115 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
116 * as ancient (libc5 based) binaries can segfault. )
118 int randomize_va_space __read_mostly =
119 #ifdef CONFIG_COMPAT_BRK
125 #ifndef arch_faults_on_old_pte
126 static inline bool arch_faults_on_old_pte(void)
129 * Those arches which don't have hw access flag feature need to
130 * implement their own helper. By default, "true" means pagefault
131 * will be hit on old pte.
137 static int __init disable_randmaps(char *s)
139 randomize_va_space = 0;
142 __setup("norandmaps", disable_randmaps);
144 unsigned long zero_pfn __read_mostly;
145 EXPORT_SYMBOL(zero_pfn);
147 unsigned long highest_memmap_pfn __read_mostly;
150 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
152 static int __init init_zero_pfn(void)
154 zero_pfn = page_to_pfn(ZERO_PAGE(0));
157 early_initcall(init_zero_pfn);
159 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
161 trace_rss_stat(mm, member, count);
164 #if defined(SPLIT_RSS_COUNTING)
166 void sync_mm_rss(struct mm_struct *mm)
170 for (i = 0; i < NR_MM_COUNTERS; i++) {
171 if (current->rss_stat.count[i]) {
172 add_mm_counter(mm, i, current->rss_stat.count[i]);
173 current->rss_stat.count[i] = 0;
176 current->rss_stat.events = 0;
179 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
181 struct task_struct *task = current;
183 if (likely(task->mm == mm))
184 task->rss_stat.count[member] += val;
186 add_mm_counter(mm, member, val);
188 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
189 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
191 /* sync counter once per 64 page faults */
192 #define TASK_RSS_EVENTS_THRESH (64)
193 static void check_sync_rss_stat(struct task_struct *task)
195 if (unlikely(task != current))
197 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
198 sync_mm_rss(task->mm);
200 #else /* SPLIT_RSS_COUNTING */
202 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
203 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
205 static void check_sync_rss_stat(struct task_struct *task)
209 #endif /* SPLIT_RSS_COUNTING */
212 * Note: this doesn't free the actual pages themselves. That
213 * has been handled earlier when unmapping all the memory regions.
215 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
218 pgtable_t token = pmd_pgtable(*pmd);
220 pte_free_tlb(tlb, token, addr);
221 mm_dec_nr_ptes(tlb->mm);
224 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
225 unsigned long addr, unsigned long end,
226 unsigned long floor, unsigned long ceiling)
233 pmd = pmd_offset(pud, addr);
235 next = pmd_addr_end(addr, end);
236 if (pmd_none_or_clear_bad(pmd))
238 free_pte_range(tlb, pmd, addr);
239 } while (pmd++, addr = next, addr != end);
249 if (end - 1 > ceiling - 1)
252 pmd = pmd_offset(pud, start);
254 pmd_free_tlb(tlb, pmd, start);
255 mm_dec_nr_pmds(tlb->mm);
258 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
259 unsigned long addr, unsigned long end,
260 unsigned long floor, unsigned long ceiling)
267 pud = pud_offset(p4d, addr);
269 next = pud_addr_end(addr, end);
270 if (pud_none_or_clear_bad(pud))
272 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
273 } while (pud++, addr = next, addr != end);
283 if (end - 1 > ceiling - 1)
286 pud = pud_offset(p4d, start);
288 pud_free_tlb(tlb, pud, start);
289 mm_dec_nr_puds(tlb->mm);
292 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
293 unsigned long addr, unsigned long end,
294 unsigned long floor, unsigned long ceiling)
301 p4d = p4d_offset(pgd, addr);
303 next = p4d_addr_end(addr, end);
304 if (p4d_none_or_clear_bad(p4d))
306 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
307 } while (p4d++, addr = next, addr != end);
313 ceiling &= PGDIR_MASK;
317 if (end - 1 > ceiling - 1)
320 p4d = p4d_offset(pgd, start);
322 p4d_free_tlb(tlb, p4d, start);
326 * This function frees user-level page tables of a process.
328 void free_pgd_range(struct mmu_gather *tlb,
329 unsigned long addr, unsigned long end,
330 unsigned long floor, unsigned long ceiling)
336 * The next few lines have given us lots of grief...
338 * Why are we testing PMD* at this top level? Because often
339 * there will be no work to do at all, and we'd prefer not to
340 * go all the way down to the bottom just to discover that.
342 * Why all these "- 1"s? Because 0 represents both the bottom
343 * of the address space and the top of it (using -1 for the
344 * top wouldn't help much: the masks would do the wrong thing).
345 * The rule is that addr 0 and floor 0 refer to the bottom of
346 * the address space, but end 0 and ceiling 0 refer to the top
347 * Comparisons need to use "end - 1" and "ceiling - 1" (though
348 * that end 0 case should be mythical).
350 * Wherever addr is brought up or ceiling brought down, we must
351 * be careful to reject "the opposite 0" before it confuses the
352 * subsequent tests. But what about where end is brought down
353 * by PMD_SIZE below? no, end can't go down to 0 there.
355 * Whereas we round start (addr) and ceiling down, by different
356 * masks at different levels, in order to test whether a table
357 * now has no other vmas using it, so can be freed, we don't
358 * bother to round floor or end up - the tests don't need that.
372 if (end - 1 > ceiling - 1)
377 * We add page table cache pages with PAGE_SIZE,
378 * (see pte_free_tlb()), flush the tlb if we need
380 tlb_change_page_size(tlb, PAGE_SIZE);
381 pgd = pgd_offset(tlb->mm, addr);
383 next = pgd_addr_end(addr, end);
384 if (pgd_none_or_clear_bad(pgd))
386 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
387 } while (pgd++, addr = next, addr != end);
390 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
391 unsigned long floor, unsigned long ceiling)
394 struct vm_area_struct *next = vma->vm_next;
395 unsigned long addr = vma->vm_start;
398 * Hide vma from rmap and truncate_pagecache before freeing
401 unlink_anon_vmas(vma);
402 unlink_file_vma(vma);
404 if (is_vm_hugetlb_page(vma)) {
405 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
406 floor, next ? next->vm_start : ceiling);
409 * Optimization: gather nearby vmas into one call down
411 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
412 && !is_vm_hugetlb_page(next)) {
415 unlink_anon_vmas(vma);
416 unlink_file_vma(vma);
418 free_pgd_range(tlb, addr, vma->vm_end,
419 floor, next ? next->vm_start : ceiling);
425 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
428 pgtable_t new = pte_alloc_one(mm);
433 * Ensure all pte setup (eg. pte page lock and page clearing) are
434 * visible before the pte is made visible to other CPUs by being
435 * put into page tables.
437 * The other side of the story is the pointer chasing in the page
438 * table walking code (when walking the page table without locking;
439 * ie. most of the time). Fortunately, these data accesses consist
440 * of a chain of data-dependent loads, meaning most CPUs (alpha
441 * being the notable exception) will already guarantee loads are
442 * seen in-order. See the alpha page table accessors for the
443 * smp_rmb() barriers in page table walking code.
445 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
447 ptl = pmd_lock(mm, pmd);
448 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
450 pmd_populate(mm, pmd, new);
459 int __pte_alloc_kernel(pmd_t *pmd)
461 pte_t *new = pte_alloc_one_kernel(&init_mm);
465 smp_wmb(); /* See comment in __pte_alloc */
467 spin_lock(&init_mm.page_table_lock);
468 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
469 pmd_populate_kernel(&init_mm, pmd, new);
472 spin_unlock(&init_mm.page_table_lock);
474 pte_free_kernel(&init_mm, new);
478 static inline void init_rss_vec(int *rss)
480 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
483 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
487 if (current->mm == mm)
489 for (i = 0; i < NR_MM_COUNTERS; i++)
491 add_mm_counter(mm, i, rss[i]);
495 * This function is called to print an error when a bad pte
496 * is found. For example, we might have a PFN-mapped pte in
497 * a region that doesn't allow it.
499 * The calling function must still handle the error.
501 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
502 pte_t pte, struct page *page)
504 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
505 p4d_t *p4d = p4d_offset(pgd, addr);
506 pud_t *pud = pud_offset(p4d, addr);
507 pmd_t *pmd = pmd_offset(pud, addr);
508 struct address_space *mapping;
510 static unsigned long resume;
511 static unsigned long nr_shown;
512 static unsigned long nr_unshown;
515 * Allow a burst of 60 reports, then keep quiet for that minute;
516 * or allow a steady drip of one report per second.
518 if (nr_shown == 60) {
519 if (time_before(jiffies, resume)) {
524 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
531 resume = jiffies + 60 * HZ;
533 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
534 index = linear_page_index(vma, addr);
536 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
538 (long long)pte_val(pte), (long long)pmd_val(*pmd));
540 dump_page(page, "bad pte");
541 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
542 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
543 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
545 vma->vm_ops ? vma->vm_ops->fault : NULL,
546 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
547 mapping ? mapping->a_ops->readpage : NULL);
549 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
553 * vm_normal_page -- This function gets the "struct page" associated with a pte.
555 * "Special" mappings do not wish to be associated with a "struct page" (either
556 * it doesn't exist, or it exists but they don't want to touch it). In this
557 * case, NULL is returned here. "Normal" mappings do have a struct page.
559 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
560 * pte bit, in which case this function is trivial. Secondly, an architecture
561 * may not have a spare pte bit, which requires a more complicated scheme,
564 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
565 * special mapping (even if there are underlying and valid "struct pages").
566 * COWed pages of a VM_PFNMAP are always normal.
568 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
569 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
570 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
571 * mapping will always honor the rule
573 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
575 * And for normal mappings this is false.
577 * This restricts such mappings to be a linear translation from virtual address
578 * to pfn. To get around this restriction, we allow arbitrary mappings so long
579 * as the vma is not a COW mapping; in that case, we know that all ptes are
580 * special (because none can have been COWed).
583 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
585 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
586 * page" backing, however the difference is that _all_ pages with a struct
587 * page (that is, those where pfn_valid is true) are refcounted and considered
588 * normal pages by the VM. The disadvantage is that pages are refcounted
589 * (which can be slower and simply not an option for some PFNMAP users). The
590 * advantage is that we don't have to follow the strict linearity rule of
591 * PFNMAP mappings in order to support COWable mappings.
594 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
597 unsigned long pfn = pte_pfn(pte);
599 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
600 if (likely(!pte_special(pte)))
602 if (vma->vm_ops && vma->vm_ops->find_special_page)
603 return vma->vm_ops->find_special_page(vma, addr);
604 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
606 if (is_zero_pfn(pfn))
611 print_bad_pte(vma, addr, pte, NULL);
615 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
617 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
618 if (vma->vm_flags & VM_MIXEDMAP) {
624 off = (addr - vma->vm_start) >> PAGE_SHIFT;
625 if (pfn == vma->vm_pgoff + off)
627 if (!is_cow_mapping(vma->vm_flags))
632 if (is_zero_pfn(pfn))
636 if (unlikely(pfn > highest_memmap_pfn)) {
637 print_bad_pte(vma, addr, pte, NULL);
642 * NOTE! We still have PageReserved() pages in the page tables.
643 * eg. VDSO mappings can cause them to exist.
646 return pfn_to_page(pfn);
649 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
650 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
653 unsigned long pfn = pmd_pfn(pmd);
656 * There is no pmd_special() but there may be special pmds, e.g.
657 * in a direct-access (dax) mapping, so let's just replicate the
658 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
660 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
661 if (vma->vm_flags & VM_MIXEDMAP) {
667 off = (addr - vma->vm_start) >> PAGE_SHIFT;
668 if (pfn == vma->vm_pgoff + off)
670 if (!is_cow_mapping(vma->vm_flags))
677 if (is_huge_zero_pmd(pmd))
679 if (unlikely(pfn > highest_memmap_pfn))
683 * NOTE! We still have PageReserved() pages in the page tables.
684 * eg. VDSO mappings can cause them to exist.
687 return pfn_to_page(pfn);
692 * copy one vm_area from one task to the other. Assumes the page tables
693 * already present in the new task to be cleared in the whole range
694 * covered by this vma.
698 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
699 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
700 struct vm_area_struct *src_vma, unsigned long addr, int *rss)
702 unsigned long vm_flags = dst_vma->vm_flags;
703 pte_t pte = *src_pte;
705 swp_entry_t entry = pte_to_swp_entry(pte);
707 if (likely(!non_swap_entry(entry))) {
708 if (swap_duplicate(entry) < 0)
711 /* make sure dst_mm is on swapoff's mmlist. */
712 if (unlikely(list_empty(&dst_mm->mmlist))) {
713 spin_lock(&mmlist_lock);
714 if (list_empty(&dst_mm->mmlist))
715 list_add(&dst_mm->mmlist,
717 spin_unlock(&mmlist_lock);
720 } else if (is_migration_entry(entry)) {
721 page = migration_entry_to_page(entry);
723 rss[mm_counter(page)]++;
725 if (is_write_migration_entry(entry) &&
726 is_cow_mapping(vm_flags)) {
728 * COW mappings require pages in both
729 * parent and child to be set to read.
731 make_migration_entry_read(&entry);
732 pte = swp_entry_to_pte(entry);
733 if (pte_swp_soft_dirty(*src_pte))
734 pte = pte_swp_mksoft_dirty(pte);
735 if (pte_swp_uffd_wp(*src_pte))
736 pte = pte_swp_mkuffd_wp(pte);
737 set_pte_at(src_mm, addr, src_pte, pte);
739 } else if (is_device_private_entry(entry)) {
740 page = device_private_entry_to_page(entry);
743 * Update rss count even for unaddressable pages, as
744 * they should treated just like normal pages in this
747 * We will likely want to have some new rss counters
748 * for unaddressable pages, at some point. But for now
749 * keep things as they are.
752 rss[mm_counter(page)]++;
753 page_dup_rmap(page, false);
756 * We do not preserve soft-dirty information, because so
757 * far, checkpoint/restore is the only feature that
758 * requires that. And checkpoint/restore does not work
759 * when a device driver is involved (you cannot easily
760 * save and restore device driver state).
762 if (is_write_device_private_entry(entry) &&
763 is_cow_mapping(vm_flags)) {
764 make_device_private_entry_read(&entry);
765 pte = swp_entry_to_pte(entry);
766 if (pte_swp_uffd_wp(*src_pte))
767 pte = pte_swp_mkuffd_wp(pte);
768 set_pte_at(src_mm, addr, src_pte, pte);
771 if (!userfaultfd_wp(dst_vma))
772 pte = pte_swp_clear_uffd_wp(pte);
773 set_pte_at(dst_mm, addr, dst_pte, pte);
778 * Copy a present and normal page if necessary.
780 * NOTE! The usual case is that this doesn't need to do
781 * anything, and can just return a positive value. That
782 * will let the caller know that it can just increase
783 * the page refcount and re-use the pte the traditional
786 * But _if_ we need to copy it because it needs to be
787 * pinned in the parent (and the child should get its own
788 * copy rather than just a reference to the same page),
789 * we'll do that here and return zero to let the caller
792 * And if we need a pre-allocated page but don't yet have
793 * one, return a negative error to let the preallocation
794 * code know so that it can do so outside the page table
798 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
799 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
800 struct page **prealloc, pte_t pte, struct page *page)
802 struct mm_struct *src_mm = src_vma->vm_mm;
803 struct page *new_page;
805 if (!is_cow_mapping(src_vma->vm_flags))
809 * What we want to do is to check whether this page may
810 * have been pinned by the parent process. If so,
811 * instead of wrprotect the pte on both sides, we copy
812 * the page immediately so that we'll always guarantee
813 * the pinned page won't be randomly replaced in the
816 * The page pinning checks are just "has this mm ever
817 * seen pinning", along with the (inexact) check of
818 * the page count. That might give false positives for
819 * for pinning, but it will work correctly.
821 if (likely(!atomic_read(&src_mm->has_pinned)))
823 if (likely(!page_maybe_dma_pinned(page)))
826 new_page = *prealloc;
831 * We have a prealloc page, all good! Take it
832 * over and copy the page & arm it.
835 copy_user_highpage(new_page, page, addr, src_vma);
836 __SetPageUptodate(new_page);
837 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
838 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
839 rss[mm_counter(new_page)]++;
841 /* All done, just insert the new page copy in the child */
842 pte = mk_pte(new_page, dst_vma->vm_page_prot);
843 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
844 if (userfaultfd_pte_wp(dst_vma, *src_pte))
845 /* Uffd-wp needs to be delivered to dest pte as well */
846 pte = pte_wrprotect(pte_mkuffd_wp(pte));
847 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
852 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
853 * is required to copy this pte.
856 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
857 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
858 struct page **prealloc)
860 struct mm_struct *src_mm = src_vma->vm_mm;
861 unsigned long vm_flags = src_vma->vm_flags;
862 pte_t pte = *src_pte;
865 page = vm_normal_page(src_vma, addr, pte);
869 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
870 addr, rss, prealloc, pte, page);
875 page_dup_rmap(page, false);
876 rss[mm_counter(page)]++;
880 * If it's a COW mapping, write protect it both
881 * in the parent and the child
883 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
884 ptep_set_wrprotect(src_mm, addr, src_pte);
885 pte = pte_wrprotect(pte);
889 * If it's a shared mapping, mark it clean in
892 if (vm_flags & VM_SHARED)
893 pte = pte_mkclean(pte);
894 pte = pte_mkold(pte);
896 if (!userfaultfd_wp(dst_vma))
897 pte = pte_clear_uffd_wp(pte);
899 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
903 static inline struct page *
904 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
907 struct page *new_page;
909 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
913 if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
917 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
923 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
924 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
927 struct mm_struct *dst_mm = dst_vma->vm_mm;
928 struct mm_struct *src_mm = src_vma->vm_mm;
929 pte_t *orig_src_pte, *orig_dst_pte;
930 pte_t *src_pte, *dst_pte;
931 spinlock_t *src_ptl, *dst_ptl;
932 int progress, ret = 0;
933 int rss[NR_MM_COUNTERS];
934 swp_entry_t entry = (swp_entry_t){0};
935 struct page *prealloc = NULL;
941 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
946 src_pte = pte_offset_map(src_pmd, addr);
947 src_ptl = pte_lockptr(src_mm, src_pmd);
948 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
949 orig_src_pte = src_pte;
950 orig_dst_pte = dst_pte;
951 arch_enter_lazy_mmu_mode();
955 * We are holding two locks at this point - either of them
956 * could generate latencies in another task on another CPU.
958 if (progress >= 32) {
960 if (need_resched() ||
961 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
964 if (pte_none(*src_pte)) {
968 if (unlikely(!pte_present(*src_pte))) {
969 entry.val = copy_nonpresent_pte(dst_mm, src_mm,
978 /* copy_present_pte() will clear `*prealloc' if consumed */
979 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
980 addr, rss, &prealloc);
982 * If we need a pre-allocated page for this pte, drop the
983 * locks, allocate, and try again.
985 if (unlikely(ret == -EAGAIN))
987 if (unlikely(prealloc)) {
989 * pre-alloc page cannot be reused by next time so as
990 * to strictly follow mempolicy (e.g., alloc_page_vma()
991 * will allocate page according to address). This
992 * could only happen if one pinned pte changed.
998 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1000 arch_leave_lazy_mmu_mode();
1001 spin_unlock(src_ptl);
1002 pte_unmap(orig_src_pte);
1003 add_mm_rss_vec(dst_mm, rss);
1004 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1008 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1014 WARN_ON_ONCE(ret != -EAGAIN);
1015 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1018 /* We've captured and resolved the error. Reset, try again. */
1024 if (unlikely(prealloc))
1030 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1031 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1034 struct mm_struct *dst_mm = dst_vma->vm_mm;
1035 struct mm_struct *src_mm = src_vma->vm_mm;
1036 pmd_t *src_pmd, *dst_pmd;
1039 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1042 src_pmd = pmd_offset(src_pud, addr);
1044 next = pmd_addr_end(addr, end);
1045 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1046 || pmd_devmap(*src_pmd)) {
1048 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1049 err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1050 addr, dst_vma, src_vma);
1057 if (pmd_none_or_clear_bad(src_pmd))
1059 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1062 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1067 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1068 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1071 struct mm_struct *dst_mm = dst_vma->vm_mm;
1072 struct mm_struct *src_mm = src_vma->vm_mm;
1073 pud_t *src_pud, *dst_pud;
1076 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1079 src_pud = pud_offset(src_p4d, addr);
1081 next = pud_addr_end(addr, end);
1082 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1085 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1086 err = copy_huge_pud(dst_mm, src_mm,
1087 dst_pud, src_pud, addr, src_vma);
1094 if (pud_none_or_clear_bad(src_pud))
1096 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1099 } while (dst_pud++, src_pud++, addr = next, addr != end);
1104 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1105 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1108 struct mm_struct *dst_mm = dst_vma->vm_mm;
1109 p4d_t *src_p4d, *dst_p4d;
1112 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1115 src_p4d = p4d_offset(src_pgd, addr);
1117 next = p4d_addr_end(addr, end);
1118 if (p4d_none_or_clear_bad(src_p4d))
1120 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1123 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1128 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1130 pgd_t *src_pgd, *dst_pgd;
1132 unsigned long addr = src_vma->vm_start;
1133 unsigned long end = src_vma->vm_end;
1134 struct mm_struct *dst_mm = dst_vma->vm_mm;
1135 struct mm_struct *src_mm = src_vma->vm_mm;
1136 struct mmu_notifier_range range;
1141 * Don't copy ptes where a page fault will fill them correctly.
1142 * Fork becomes much lighter when there are big shared or private
1143 * readonly mappings. The tradeoff is that copy_page_range is more
1144 * efficient than faulting.
1146 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1150 if (is_vm_hugetlb_page(src_vma))
1151 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1153 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1155 * We do not free on error cases below as remove_vma
1156 * gets called on error from higher level routine
1158 ret = track_pfn_copy(src_vma);
1164 * We need to invalidate the secondary MMU mappings only when
1165 * there could be a permission downgrade on the ptes of the
1166 * parent mm. And a permission downgrade will only happen if
1167 * is_cow_mapping() returns true.
1169 is_cow = is_cow_mapping(src_vma->vm_flags);
1172 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1173 0, src_vma, src_mm, addr, end);
1174 mmu_notifier_invalidate_range_start(&range);
1176 * Disabling preemption is not needed for the write side, as
1177 * the read side doesn't spin, but goes to the mmap_lock.
1179 * Use the raw variant of the seqcount_t write API to avoid
1180 * lockdep complaining about preemptibility.
1182 mmap_assert_write_locked(src_mm);
1183 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1187 dst_pgd = pgd_offset(dst_mm, addr);
1188 src_pgd = pgd_offset(src_mm, addr);
1190 next = pgd_addr_end(addr, end);
1191 if (pgd_none_or_clear_bad(src_pgd))
1193 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1198 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1201 raw_write_seqcount_end(&src_mm->write_protect_seq);
1202 mmu_notifier_invalidate_range_end(&range);
1207 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1208 struct vm_area_struct *vma, pmd_t *pmd,
1209 unsigned long addr, unsigned long end,
1210 struct zap_details *details)
1212 struct mm_struct *mm = tlb->mm;
1213 int force_flush = 0;
1214 int rss[NR_MM_COUNTERS];
1220 tlb_change_page_size(tlb, PAGE_SIZE);
1223 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1225 flush_tlb_batched_pending(mm);
1226 arch_enter_lazy_mmu_mode();
1229 if (pte_none(ptent))
1235 if (pte_present(ptent)) {
1238 page = vm_normal_page(vma, addr, ptent);
1239 if (unlikely(details) && page) {
1241 * unmap_shared_mapping_pages() wants to
1242 * invalidate cache without truncating:
1243 * unmap shared but keep private pages.
1245 if (details->check_mapping &&
1246 details->check_mapping != page_rmapping(page))
1249 ptent = ptep_get_and_clear_full(mm, addr, pte,
1251 tlb_remove_tlb_entry(tlb, pte, addr);
1252 if (unlikely(!page))
1255 if (!PageAnon(page)) {
1256 if (pte_dirty(ptent)) {
1258 set_page_dirty(page);
1260 if (pte_young(ptent) &&
1261 likely(!(vma->vm_flags & VM_SEQ_READ)))
1262 mark_page_accessed(page);
1264 rss[mm_counter(page)]--;
1265 page_remove_rmap(page, false);
1266 if (unlikely(page_mapcount(page) < 0))
1267 print_bad_pte(vma, addr, ptent, page);
1268 if (unlikely(__tlb_remove_page(tlb, page))) {
1276 entry = pte_to_swp_entry(ptent);
1277 if (is_device_private_entry(entry)) {
1278 struct page *page = device_private_entry_to_page(entry);
1280 if (unlikely(details && details->check_mapping)) {
1282 * unmap_shared_mapping_pages() wants to
1283 * invalidate cache without truncating:
1284 * unmap shared but keep private pages.
1286 if (details->check_mapping !=
1287 page_rmapping(page))
1291 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1292 rss[mm_counter(page)]--;
1293 page_remove_rmap(page, false);
1298 /* If details->check_mapping, we leave swap entries. */
1299 if (unlikely(details))
1302 if (!non_swap_entry(entry))
1304 else if (is_migration_entry(entry)) {
1307 page = migration_entry_to_page(entry);
1308 rss[mm_counter(page)]--;
1310 if (unlikely(!free_swap_and_cache(entry)))
1311 print_bad_pte(vma, addr, ptent, NULL);
1312 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1313 } while (pte++, addr += PAGE_SIZE, addr != end);
1315 add_mm_rss_vec(mm, rss);
1316 arch_leave_lazy_mmu_mode();
1318 /* Do the actual TLB flush before dropping ptl */
1320 tlb_flush_mmu_tlbonly(tlb);
1321 pte_unmap_unlock(start_pte, ptl);
1324 * If we forced a TLB flush (either due to running out of
1325 * batch buffers or because we needed to flush dirty TLB
1326 * entries before releasing the ptl), free the batched
1327 * memory too. Restart if we didn't do everything.
1342 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1343 struct vm_area_struct *vma, pud_t *pud,
1344 unsigned long addr, unsigned long end,
1345 struct zap_details *details)
1350 pmd = pmd_offset(pud, addr);
1352 next = pmd_addr_end(addr, end);
1353 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1354 if (next - addr != HPAGE_PMD_SIZE)
1355 __split_huge_pmd(vma, pmd, addr, false, NULL);
1356 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1359 } else if (details && details->single_page &&
1360 PageTransCompound(details->single_page) &&
1361 next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1362 spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1364 * Take and drop THP pmd lock so that we cannot return
1365 * prematurely, while zap_huge_pmd() has cleared *pmd,
1366 * but not yet decremented compound_mapcount().
1372 * Here there can be other concurrent MADV_DONTNEED or
1373 * trans huge page faults running, and if the pmd is
1374 * none or trans huge it can change under us. This is
1375 * because MADV_DONTNEED holds the mmap_lock in read
1378 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1380 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1383 } while (pmd++, addr = next, addr != end);
1388 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1389 struct vm_area_struct *vma, p4d_t *p4d,
1390 unsigned long addr, unsigned long end,
1391 struct zap_details *details)
1396 pud = pud_offset(p4d, addr);
1398 next = pud_addr_end(addr, end);
1399 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1400 if (next - addr != HPAGE_PUD_SIZE) {
1401 mmap_assert_locked(tlb->mm);
1402 split_huge_pud(vma, pud, addr);
1403 } else if (zap_huge_pud(tlb, vma, pud, addr))
1407 if (pud_none_or_clear_bad(pud))
1409 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1412 } while (pud++, addr = next, addr != end);
1417 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1418 struct vm_area_struct *vma, pgd_t *pgd,
1419 unsigned long addr, unsigned long end,
1420 struct zap_details *details)
1425 p4d = p4d_offset(pgd, addr);
1427 next = p4d_addr_end(addr, end);
1428 if (p4d_none_or_clear_bad(p4d))
1430 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1431 } while (p4d++, addr = next, addr != end);
1436 void unmap_page_range(struct mmu_gather *tlb,
1437 struct vm_area_struct *vma,
1438 unsigned long addr, unsigned long end,
1439 struct zap_details *details)
1444 BUG_ON(addr >= end);
1445 tlb_start_vma(tlb, vma);
1446 pgd = pgd_offset(vma->vm_mm, addr);
1448 next = pgd_addr_end(addr, end);
1449 if (pgd_none_or_clear_bad(pgd))
1451 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1452 } while (pgd++, addr = next, addr != end);
1453 tlb_end_vma(tlb, vma);
1457 static void unmap_single_vma(struct mmu_gather *tlb,
1458 struct vm_area_struct *vma, unsigned long start_addr,
1459 unsigned long end_addr,
1460 struct zap_details *details)
1462 unsigned long start = max(vma->vm_start, start_addr);
1465 if (start >= vma->vm_end)
1467 end = min(vma->vm_end, end_addr);
1468 if (end <= vma->vm_start)
1472 uprobe_munmap(vma, start, end);
1474 if (unlikely(vma->vm_flags & VM_PFNMAP))
1475 untrack_pfn(vma, 0, 0);
1478 if (unlikely(is_vm_hugetlb_page(vma))) {
1480 * It is undesirable to test vma->vm_file as it
1481 * should be non-null for valid hugetlb area.
1482 * However, vm_file will be NULL in the error
1483 * cleanup path of mmap_region. When
1484 * hugetlbfs ->mmap method fails,
1485 * mmap_region() nullifies vma->vm_file
1486 * before calling this function to clean up.
1487 * Since no pte has actually been setup, it is
1488 * safe to do nothing in this case.
1491 i_mmap_lock_write(vma->vm_file->f_mapping);
1492 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1493 i_mmap_unlock_write(vma->vm_file->f_mapping);
1496 unmap_page_range(tlb, vma, start, end, details);
1501 * unmap_vmas - unmap a range of memory covered by a list of vma's
1502 * @tlb: address of the caller's struct mmu_gather
1503 * @vma: the starting vma
1504 * @start_addr: virtual address at which to start unmapping
1505 * @end_addr: virtual address at which to end unmapping
1507 * Unmap all pages in the vma list.
1509 * Only addresses between `start' and `end' will be unmapped.
1511 * The VMA list must be sorted in ascending virtual address order.
1513 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1514 * range after unmap_vmas() returns. So the only responsibility here is to
1515 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1516 * drops the lock and schedules.
1518 void unmap_vmas(struct mmu_gather *tlb,
1519 struct vm_area_struct *vma, unsigned long start_addr,
1520 unsigned long end_addr)
1522 struct mmu_notifier_range range;
1524 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1525 start_addr, end_addr);
1526 mmu_notifier_invalidate_range_start(&range);
1527 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1528 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1529 mmu_notifier_invalidate_range_end(&range);
1533 * zap_page_range - remove user pages in a given range
1534 * @vma: vm_area_struct holding the applicable pages
1535 * @start: starting address of pages to zap
1536 * @size: number of bytes to zap
1538 * Caller must protect the VMA list
1540 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1543 struct mmu_notifier_range range;
1544 struct mmu_gather tlb;
1547 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1548 start, start + size);
1549 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1550 update_hiwater_rss(vma->vm_mm);
1551 mmu_notifier_invalidate_range_start(&range);
1552 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1553 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1554 mmu_notifier_invalidate_range_end(&range);
1555 tlb_finish_mmu(&tlb, start, range.end);
1559 * zap_page_range_single - remove user pages in a given range
1560 * @vma: vm_area_struct holding the applicable pages
1561 * @address: starting address of pages to zap
1562 * @size: number of bytes to zap
1563 * @details: details of shared cache invalidation
1565 * The range must fit into one VMA.
1567 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1568 unsigned long size, struct zap_details *details)
1570 struct mmu_notifier_range range;
1571 struct mmu_gather tlb;
1574 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1575 address, address + size);
1576 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1577 update_hiwater_rss(vma->vm_mm);
1578 mmu_notifier_invalidate_range_start(&range);
1579 unmap_single_vma(&tlb, vma, address, range.end, details);
1580 mmu_notifier_invalidate_range_end(&range);
1581 tlb_finish_mmu(&tlb, address, range.end);
1585 * zap_vma_ptes - remove ptes mapping the vma
1586 * @vma: vm_area_struct holding ptes to be zapped
1587 * @address: starting address of pages to zap
1588 * @size: number of bytes to zap
1590 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1592 * The entire address range must be fully contained within the vma.
1595 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1598 if (address < vma->vm_start || address + size > vma->vm_end ||
1599 !(vma->vm_flags & VM_PFNMAP))
1602 zap_page_range_single(vma, address, size, NULL);
1604 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1606 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1613 pgd = pgd_offset(mm, addr);
1614 p4d = p4d_alloc(mm, pgd, addr);
1617 pud = pud_alloc(mm, p4d, addr);
1620 pmd = pmd_alloc(mm, pud, addr);
1624 VM_BUG_ON(pmd_trans_huge(*pmd));
1628 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1631 pmd_t *pmd = walk_to_pmd(mm, addr);
1635 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1638 static int validate_page_before_insert(struct page *page)
1640 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1642 flush_dcache_page(page);
1646 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1647 unsigned long addr, struct page *page, pgprot_t prot)
1649 if (!pte_none(*pte))
1651 /* Ok, finally just insert the thing.. */
1653 inc_mm_counter_fast(mm, mm_counter_file(page));
1654 page_add_file_rmap(page, false);
1655 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1660 * This is the old fallback for page remapping.
1662 * For historical reasons, it only allows reserved pages. Only
1663 * old drivers should use this, and they needed to mark their
1664 * pages reserved for the old functions anyway.
1666 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1667 struct page *page, pgprot_t prot)
1669 struct mm_struct *mm = vma->vm_mm;
1674 retval = validate_page_before_insert(page);
1678 pte = get_locked_pte(mm, addr, &ptl);
1681 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1682 pte_unmap_unlock(pte, ptl);
1688 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1689 unsigned long addr, struct page *page, pgprot_t prot)
1693 if (!page_count(page))
1695 err = validate_page_before_insert(page);
1698 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1701 /* insert_pages() amortizes the cost of spinlock operations
1702 * when inserting pages in a loop. Arch *must* define pte_index.
1704 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1705 struct page **pages, unsigned long *num, pgprot_t prot)
1708 pte_t *start_pte, *pte;
1709 spinlock_t *pte_lock;
1710 struct mm_struct *const mm = vma->vm_mm;
1711 unsigned long curr_page_idx = 0;
1712 unsigned long remaining_pages_total = *num;
1713 unsigned long pages_to_write_in_pmd;
1717 pmd = walk_to_pmd(mm, addr);
1721 pages_to_write_in_pmd = min_t(unsigned long,
1722 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1724 /* Allocate the PTE if necessary; takes PMD lock once only. */
1726 if (pte_alloc(mm, pmd))
1729 while (pages_to_write_in_pmd) {
1731 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1733 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1734 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1735 int err = insert_page_in_batch_locked(mm, pte,
1736 addr, pages[curr_page_idx], prot);
1737 if (unlikely(err)) {
1738 pte_unmap_unlock(start_pte, pte_lock);
1740 remaining_pages_total -= pte_idx;
1746 pte_unmap_unlock(start_pte, pte_lock);
1747 pages_to_write_in_pmd -= batch_size;
1748 remaining_pages_total -= batch_size;
1750 if (remaining_pages_total)
1754 *num = remaining_pages_total;
1757 #endif /* ifdef pte_index */
1760 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1761 * @vma: user vma to map to
1762 * @addr: target start user address of these pages
1763 * @pages: source kernel pages
1764 * @num: in: number of pages to map. out: number of pages that were *not*
1765 * mapped. (0 means all pages were successfully mapped).
1767 * Preferred over vm_insert_page() when inserting multiple pages.
1769 * In case of error, we may have mapped a subset of the provided
1770 * pages. It is the caller's responsibility to account for this case.
1772 * The same restrictions apply as in vm_insert_page().
1774 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1775 struct page **pages, unsigned long *num)
1778 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1780 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1782 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1783 BUG_ON(mmap_read_trylock(vma->vm_mm));
1784 BUG_ON(vma->vm_flags & VM_PFNMAP);
1785 vma->vm_flags |= VM_MIXEDMAP;
1787 /* Defer page refcount checking till we're about to map that page. */
1788 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1790 unsigned long idx = 0, pgcount = *num;
1793 for (; idx < pgcount; ++idx) {
1794 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1798 *num = pgcount - idx;
1800 #endif /* ifdef pte_index */
1802 EXPORT_SYMBOL(vm_insert_pages);
1805 * vm_insert_page - insert single page into user vma
1806 * @vma: user vma to map to
1807 * @addr: target user address of this page
1808 * @page: source kernel page
1810 * This allows drivers to insert individual pages they've allocated
1813 * The page has to be a nice clean _individual_ kernel allocation.
1814 * If you allocate a compound page, you need to have marked it as
1815 * such (__GFP_COMP), or manually just split the page up yourself
1816 * (see split_page()).
1818 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1819 * took an arbitrary page protection parameter. This doesn't allow
1820 * that. Your vma protection will have to be set up correctly, which
1821 * means that if you want a shared writable mapping, you'd better
1822 * ask for a shared writable mapping!
1824 * The page does not need to be reserved.
1826 * Usually this function is called from f_op->mmap() handler
1827 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1828 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1829 * function from other places, for example from page-fault handler.
1831 * Return: %0 on success, negative error code otherwise.
1833 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1836 if (addr < vma->vm_start || addr >= vma->vm_end)
1838 if (!page_count(page))
1840 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1841 BUG_ON(mmap_read_trylock(vma->vm_mm));
1842 BUG_ON(vma->vm_flags & VM_PFNMAP);
1843 vma->vm_flags |= VM_MIXEDMAP;
1845 return insert_page(vma, addr, page, vma->vm_page_prot);
1847 EXPORT_SYMBOL(vm_insert_page);
1850 * __vm_map_pages - maps range of kernel pages into user vma
1851 * @vma: user vma to map to
1852 * @pages: pointer to array of source kernel pages
1853 * @num: number of pages in page array
1854 * @offset: user's requested vm_pgoff
1856 * This allows drivers to map range of kernel pages into a user vma.
1858 * Return: 0 on success and error code otherwise.
1860 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1861 unsigned long num, unsigned long offset)
1863 unsigned long count = vma_pages(vma);
1864 unsigned long uaddr = vma->vm_start;
1867 /* Fail if the user requested offset is beyond the end of the object */
1871 /* Fail if the user requested size exceeds available object size */
1872 if (count > num - offset)
1875 for (i = 0; i < count; i++) {
1876 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1886 * vm_map_pages - maps range of kernel pages starts with non zero offset
1887 * @vma: user vma to map to
1888 * @pages: pointer to array of source kernel pages
1889 * @num: number of pages in page array
1891 * Maps an object consisting of @num pages, catering for the user's
1892 * requested vm_pgoff
1894 * If we fail to insert any page into the vma, the function will return
1895 * immediately leaving any previously inserted pages present. Callers
1896 * from the mmap handler may immediately return the error as their caller
1897 * will destroy the vma, removing any successfully inserted pages. Other
1898 * callers should make their own arrangements for calling unmap_region().
1900 * Context: Process context. Called by mmap handlers.
1901 * Return: 0 on success and error code otherwise.
1903 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1906 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1908 EXPORT_SYMBOL(vm_map_pages);
1911 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1912 * @vma: user vma to map to
1913 * @pages: pointer to array of source kernel pages
1914 * @num: number of pages in page array
1916 * Similar to vm_map_pages(), except that it explicitly sets the offset
1917 * to 0. This function is intended for the drivers that did not consider
1920 * Context: Process context. Called by mmap handlers.
1921 * Return: 0 on success and error code otherwise.
1923 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1926 return __vm_map_pages(vma, pages, num, 0);
1928 EXPORT_SYMBOL(vm_map_pages_zero);
1930 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1931 pfn_t pfn, pgprot_t prot, bool mkwrite)
1933 struct mm_struct *mm = vma->vm_mm;
1937 pte = get_locked_pte(mm, addr, &ptl);
1939 return VM_FAULT_OOM;
1940 if (!pte_none(*pte)) {
1943 * For read faults on private mappings the PFN passed
1944 * in may not match the PFN we have mapped if the
1945 * mapped PFN is a writeable COW page. In the mkwrite
1946 * case we are creating a writable PTE for a shared
1947 * mapping and we expect the PFNs to match. If they
1948 * don't match, we are likely racing with block
1949 * allocation and mapping invalidation so just skip the
1952 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1953 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1956 entry = pte_mkyoung(*pte);
1957 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1958 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1959 update_mmu_cache(vma, addr, pte);
1964 /* Ok, finally just insert the thing.. */
1965 if (pfn_t_devmap(pfn))
1966 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1968 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1971 entry = pte_mkyoung(entry);
1972 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1975 set_pte_at(mm, addr, pte, entry);
1976 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1979 pte_unmap_unlock(pte, ptl);
1980 return VM_FAULT_NOPAGE;
1984 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1985 * @vma: user vma to map to
1986 * @addr: target user address of this page
1987 * @pfn: source kernel pfn
1988 * @pgprot: pgprot flags for the inserted page
1990 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1991 * to override pgprot on a per-page basis.
1993 * This only makes sense for IO mappings, and it makes no sense for
1994 * COW mappings. In general, using multiple vmas is preferable;
1995 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1998 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1999 * a value of @pgprot different from that of @vma->vm_page_prot.
2001 * Context: Process context. May allocate using %GFP_KERNEL.
2002 * Return: vm_fault_t value.
2004 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2005 unsigned long pfn, pgprot_t pgprot)
2008 * Technically, architectures with pte_special can avoid all these
2009 * restrictions (same for remap_pfn_range). However we would like
2010 * consistency in testing and feature parity among all, so we should
2011 * try to keep these invariants in place for everybody.
2013 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2014 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2015 (VM_PFNMAP|VM_MIXEDMAP));
2016 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2017 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2019 if (addr < vma->vm_start || addr >= vma->vm_end)
2020 return VM_FAULT_SIGBUS;
2022 if (!pfn_modify_allowed(pfn, pgprot))
2023 return VM_FAULT_SIGBUS;
2025 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2027 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2030 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2033 * vmf_insert_pfn - insert single pfn into user vma
2034 * @vma: user vma to map to
2035 * @addr: target user address of this page
2036 * @pfn: source kernel pfn
2038 * Similar to vm_insert_page, this allows drivers to insert individual pages
2039 * they've allocated into a user vma. Same comments apply.
2041 * This function should only be called from a vm_ops->fault handler, and
2042 * in that case the handler should return the result of this function.
2044 * vma cannot be a COW mapping.
2046 * As this is called only for pages that do not currently exist, we
2047 * do not need to flush old virtual caches or the TLB.
2049 * Context: Process context. May allocate using %GFP_KERNEL.
2050 * Return: vm_fault_t value.
2052 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2055 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2057 EXPORT_SYMBOL(vmf_insert_pfn);
2059 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2061 /* these checks mirror the abort conditions in vm_normal_page */
2062 if (vma->vm_flags & VM_MIXEDMAP)
2064 if (pfn_t_devmap(pfn))
2066 if (pfn_t_special(pfn))
2068 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2073 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2074 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2079 BUG_ON(!vm_mixed_ok(vma, pfn));
2081 if (addr < vma->vm_start || addr >= vma->vm_end)
2082 return VM_FAULT_SIGBUS;
2084 track_pfn_insert(vma, &pgprot, pfn);
2086 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2087 return VM_FAULT_SIGBUS;
2090 * If we don't have pte special, then we have to use the pfn_valid()
2091 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2092 * refcount the page if pfn_valid is true (hence insert_page rather
2093 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2094 * without pte special, it would there be refcounted as a normal page.
2096 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2097 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2101 * At this point we are committed to insert_page()
2102 * regardless of whether the caller specified flags that
2103 * result in pfn_t_has_page() == false.
2105 page = pfn_to_page(pfn_t_to_pfn(pfn));
2106 err = insert_page(vma, addr, page, pgprot);
2108 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2112 return VM_FAULT_OOM;
2113 if (err < 0 && err != -EBUSY)
2114 return VM_FAULT_SIGBUS;
2116 return VM_FAULT_NOPAGE;
2120 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2121 * @vma: user vma to map to
2122 * @addr: target user address of this page
2123 * @pfn: source kernel pfn
2124 * @pgprot: pgprot flags for the inserted page
2126 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2127 * to override pgprot on a per-page basis.
2129 * Typically this function should be used by drivers to set caching- and
2130 * encryption bits different than those of @vma->vm_page_prot, because
2131 * the caching- or encryption mode may not be known at mmap() time.
2132 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2133 * to set caching and encryption bits for those vmas (except for COW pages).
2134 * This is ensured by core vm only modifying these page table entries using
2135 * functions that don't touch caching- or encryption bits, using pte_modify()
2136 * if needed. (See for example mprotect()).
2137 * Also when new page-table entries are created, this is only done using the
2138 * fault() callback, and never using the value of vma->vm_page_prot,
2139 * except for page-table entries that point to anonymous pages as the result
2142 * Context: Process context. May allocate using %GFP_KERNEL.
2143 * Return: vm_fault_t value.
2145 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2146 pfn_t pfn, pgprot_t pgprot)
2148 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2150 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2152 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2155 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2157 EXPORT_SYMBOL(vmf_insert_mixed);
2160 * If the insertion of PTE failed because someone else already added a
2161 * different entry in the mean time, we treat that as success as we assume
2162 * the same entry was actually inserted.
2164 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2165 unsigned long addr, pfn_t pfn)
2167 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2169 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2172 * maps a range of physical memory into the requested pages. the old
2173 * mappings are removed. any references to nonexistent pages results
2174 * in null mappings (currently treated as "copy-on-access")
2176 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2177 unsigned long addr, unsigned long end,
2178 unsigned long pfn, pgprot_t prot)
2180 pte_t *pte, *mapped_pte;
2184 mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2187 arch_enter_lazy_mmu_mode();
2189 BUG_ON(!pte_none(*pte));
2190 if (!pfn_modify_allowed(pfn, prot)) {
2194 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2196 } while (pte++, addr += PAGE_SIZE, addr != end);
2197 arch_leave_lazy_mmu_mode();
2198 pte_unmap_unlock(mapped_pte, ptl);
2202 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2203 unsigned long addr, unsigned long end,
2204 unsigned long pfn, pgprot_t prot)
2210 pfn -= addr >> PAGE_SHIFT;
2211 pmd = pmd_alloc(mm, pud, addr);
2214 VM_BUG_ON(pmd_trans_huge(*pmd));
2216 next = pmd_addr_end(addr, end);
2217 err = remap_pte_range(mm, pmd, addr, next,
2218 pfn + (addr >> PAGE_SHIFT), prot);
2221 } while (pmd++, addr = next, addr != end);
2225 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2226 unsigned long addr, unsigned long end,
2227 unsigned long pfn, pgprot_t prot)
2233 pfn -= addr >> PAGE_SHIFT;
2234 pud = pud_alloc(mm, p4d, addr);
2238 next = pud_addr_end(addr, end);
2239 err = remap_pmd_range(mm, pud, addr, next,
2240 pfn + (addr >> PAGE_SHIFT), prot);
2243 } while (pud++, addr = next, addr != end);
2247 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2248 unsigned long addr, unsigned long end,
2249 unsigned long pfn, pgprot_t prot)
2255 pfn -= addr >> PAGE_SHIFT;
2256 p4d = p4d_alloc(mm, pgd, addr);
2260 next = p4d_addr_end(addr, end);
2261 err = remap_pud_range(mm, p4d, addr, next,
2262 pfn + (addr >> PAGE_SHIFT), prot);
2265 } while (p4d++, addr = next, addr != end);
2270 * remap_pfn_range - remap kernel memory to userspace
2271 * @vma: user vma to map to
2272 * @addr: target page aligned user address to start at
2273 * @pfn: page frame number of kernel physical memory address
2274 * @size: size of mapping area
2275 * @prot: page protection flags for this mapping
2277 * Note: this is only safe if the mm semaphore is held when called.
2279 * Return: %0 on success, negative error code otherwise.
2281 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2282 unsigned long pfn, unsigned long size, pgprot_t prot)
2286 unsigned long end = addr + PAGE_ALIGN(size);
2287 struct mm_struct *mm = vma->vm_mm;
2288 unsigned long remap_pfn = pfn;
2291 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2295 * Physically remapped pages are special. Tell the
2296 * rest of the world about it:
2297 * VM_IO tells people not to look at these pages
2298 * (accesses can have side effects).
2299 * VM_PFNMAP tells the core MM that the base pages are just
2300 * raw PFN mappings, and do not have a "struct page" associated
2303 * Disable vma merging and expanding with mremap().
2305 * Omit vma from core dump, even when VM_IO turned off.
2307 * There's a horrible special case to handle copy-on-write
2308 * behaviour that some programs depend on. We mark the "original"
2309 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2310 * See vm_normal_page() for details.
2312 if (is_cow_mapping(vma->vm_flags)) {
2313 if (addr != vma->vm_start || end != vma->vm_end)
2315 vma->vm_pgoff = pfn;
2318 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2322 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2324 BUG_ON(addr >= end);
2325 pfn -= addr >> PAGE_SHIFT;
2326 pgd = pgd_offset(mm, addr);
2327 flush_cache_range(vma, addr, end);
2329 next = pgd_addr_end(addr, end);
2330 err = remap_p4d_range(mm, pgd, addr, next,
2331 pfn + (addr >> PAGE_SHIFT), prot);
2334 } while (pgd++, addr = next, addr != end);
2337 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2341 EXPORT_SYMBOL(remap_pfn_range);
2344 * vm_iomap_memory - remap memory to userspace
2345 * @vma: user vma to map to
2346 * @start: start of the physical memory to be mapped
2347 * @len: size of area
2349 * This is a simplified io_remap_pfn_range() for common driver use. The
2350 * driver just needs to give us the physical memory range to be mapped,
2351 * we'll figure out the rest from the vma information.
2353 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2354 * whatever write-combining details or similar.
2356 * Return: %0 on success, negative error code otherwise.
2358 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2360 unsigned long vm_len, pfn, pages;
2362 /* Check that the physical memory area passed in looks valid */
2363 if (start + len < start)
2366 * You *really* shouldn't map things that aren't page-aligned,
2367 * but we've historically allowed it because IO memory might
2368 * just have smaller alignment.
2370 len += start & ~PAGE_MASK;
2371 pfn = start >> PAGE_SHIFT;
2372 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2373 if (pfn + pages < pfn)
2376 /* We start the mapping 'vm_pgoff' pages into the area */
2377 if (vma->vm_pgoff > pages)
2379 pfn += vma->vm_pgoff;
2380 pages -= vma->vm_pgoff;
2382 /* Can we fit all of the mapping? */
2383 vm_len = vma->vm_end - vma->vm_start;
2384 if (vm_len >> PAGE_SHIFT > pages)
2387 /* Ok, let it rip */
2388 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2390 EXPORT_SYMBOL(vm_iomap_memory);
2392 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2393 unsigned long addr, unsigned long end,
2394 pte_fn_t fn, void *data, bool create,
2395 pgtbl_mod_mask *mask)
2402 pte = (mm == &init_mm) ?
2403 pte_alloc_kernel_track(pmd, addr, mask) :
2404 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2408 pte = (mm == &init_mm) ?
2409 pte_offset_kernel(pmd, addr) :
2410 pte_offset_map_lock(mm, pmd, addr, &ptl);
2413 BUG_ON(pmd_huge(*pmd));
2415 arch_enter_lazy_mmu_mode();
2419 if (create || !pte_none(*pte)) {
2420 err = fn(pte++, addr, data);
2424 } while (addr += PAGE_SIZE, addr != end);
2426 *mask |= PGTBL_PTE_MODIFIED;
2428 arch_leave_lazy_mmu_mode();
2431 pte_unmap_unlock(pte-1, ptl);
2435 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2436 unsigned long addr, unsigned long end,
2437 pte_fn_t fn, void *data, bool create,
2438 pgtbl_mod_mask *mask)
2444 BUG_ON(pud_huge(*pud));
2447 pmd = pmd_alloc_track(mm, pud, addr, mask);
2451 pmd = pmd_offset(pud, addr);
2454 next = pmd_addr_end(addr, end);
2455 if (create || !pmd_none_or_clear_bad(pmd)) {
2456 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2461 } while (pmd++, addr = next, addr != end);
2465 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2466 unsigned long addr, unsigned long end,
2467 pte_fn_t fn, void *data, bool create,
2468 pgtbl_mod_mask *mask)
2475 pud = pud_alloc_track(mm, p4d, addr, mask);
2479 pud = pud_offset(p4d, addr);
2482 next = pud_addr_end(addr, end);
2483 if (create || !pud_none_or_clear_bad(pud)) {
2484 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2489 } while (pud++, addr = next, addr != end);
2493 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2494 unsigned long addr, unsigned long end,
2495 pte_fn_t fn, void *data, bool create,
2496 pgtbl_mod_mask *mask)
2503 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2507 p4d = p4d_offset(pgd, addr);
2510 next = p4d_addr_end(addr, end);
2511 if (create || !p4d_none_or_clear_bad(p4d)) {
2512 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2517 } while (p4d++, addr = next, addr != end);
2521 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2522 unsigned long size, pte_fn_t fn,
2523 void *data, bool create)
2526 unsigned long start = addr, next;
2527 unsigned long end = addr + size;
2528 pgtbl_mod_mask mask = 0;
2531 if (WARN_ON(addr >= end))
2534 pgd = pgd_offset(mm, addr);
2536 next = pgd_addr_end(addr, end);
2537 if (!create && pgd_none_or_clear_bad(pgd))
2539 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2542 } while (pgd++, addr = next, addr != end);
2544 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2545 arch_sync_kernel_mappings(start, start + size);
2551 * Scan a region of virtual memory, filling in page tables as necessary
2552 * and calling a provided function on each leaf page table.
2554 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2555 unsigned long size, pte_fn_t fn, void *data)
2557 return __apply_to_page_range(mm, addr, size, fn, data, true);
2559 EXPORT_SYMBOL_GPL(apply_to_page_range);
2562 * Scan a region of virtual memory, calling a provided function on
2563 * each leaf page table where it exists.
2565 * Unlike apply_to_page_range, this does _not_ fill in page tables
2566 * where they are absent.
2568 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2569 unsigned long size, pte_fn_t fn, void *data)
2571 return __apply_to_page_range(mm, addr, size, fn, data, false);
2573 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2576 * handle_pte_fault chooses page fault handler according to an entry which was
2577 * read non-atomically. Before making any commitment, on those architectures
2578 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2579 * parts, do_swap_page must check under lock before unmapping the pte and
2580 * proceeding (but do_wp_page is only called after already making such a check;
2581 * and do_anonymous_page can safely check later on).
2583 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2584 pte_t *page_table, pte_t orig_pte)
2587 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2588 if (sizeof(pte_t) > sizeof(unsigned long)) {
2589 spinlock_t *ptl = pte_lockptr(mm, pmd);
2591 same = pte_same(*page_table, orig_pte);
2595 pte_unmap(page_table);
2599 static inline bool cow_user_page(struct page *dst, struct page *src,
2600 struct vm_fault *vmf)
2605 bool locked = false;
2606 struct vm_area_struct *vma = vmf->vma;
2607 struct mm_struct *mm = vma->vm_mm;
2608 unsigned long addr = vmf->address;
2611 copy_user_highpage(dst, src, addr, vma);
2616 * If the source page was a PFN mapping, we don't have
2617 * a "struct page" for it. We do a best-effort copy by
2618 * just copying from the original user address. If that
2619 * fails, we just zero-fill it. Live with it.
2621 kaddr = kmap_atomic(dst);
2622 uaddr = (void __user *)(addr & PAGE_MASK);
2625 * On architectures with software "accessed" bits, we would
2626 * take a double page fault, so mark it accessed here.
2628 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2631 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2633 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2635 * Other thread has already handled the fault
2636 * and update local tlb only
2638 update_mmu_tlb(vma, addr, vmf->pte);
2643 entry = pte_mkyoung(vmf->orig_pte);
2644 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2645 update_mmu_cache(vma, addr, vmf->pte);
2649 * This really shouldn't fail, because the page is there
2650 * in the page tables. But it might just be unreadable,
2651 * in which case we just give up and fill the result with
2654 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2658 /* Re-validate under PTL if the page is still mapped */
2659 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2661 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2662 /* The PTE changed under us, update local tlb */
2663 update_mmu_tlb(vma, addr, vmf->pte);
2669 * The same page can be mapped back since last copy attempt.
2670 * Try to copy again under PTL.
2672 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2674 * Give a warn in case there can be some obscure
2687 pte_unmap_unlock(vmf->pte, vmf->ptl);
2688 kunmap_atomic(kaddr);
2689 flush_dcache_page(dst);
2694 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2696 struct file *vm_file = vma->vm_file;
2699 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2702 * Special mappings (e.g. VDSO) do not have any file so fake
2703 * a default GFP_KERNEL for them.
2709 * Notify the address space that the page is about to become writable so that
2710 * it can prohibit this or wait for the page to get into an appropriate state.
2712 * We do this without the lock held, so that it can sleep if it needs to.
2714 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2717 struct page *page = vmf->page;
2718 unsigned int old_flags = vmf->flags;
2720 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2722 if (vmf->vma->vm_file &&
2723 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2724 return VM_FAULT_SIGBUS;
2726 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2727 /* Restore original flags so that caller is not surprised */
2728 vmf->flags = old_flags;
2729 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2731 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2733 if (!page->mapping) {
2735 return 0; /* retry */
2737 ret |= VM_FAULT_LOCKED;
2739 VM_BUG_ON_PAGE(!PageLocked(page), page);
2744 * Handle dirtying of a page in shared file mapping on a write fault.
2746 * The function expects the page to be locked and unlocks it.
2748 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2750 struct vm_area_struct *vma = vmf->vma;
2751 struct address_space *mapping;
2752 struct page *page = vmf->page;
2754 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2756 dirtied = set_page_dirty(page);
2757 VM_BUG_ON_PAGE(PageAnon(page), page);
2759 * Take a local copy of the address_space - page.mapping may be zeroed
2760 * by truncate after unlock_page(). The address_space itself remains
2761 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2762 * release semantics to prevent the compiler from undoing this copying.
2764 mapping = page_rmapping(page);
2768 file_update_time(vma->vm_file);
2771 * Throttle page dirtying rate down to writeback speed.
2773 * mapping may be NULL here because some device drivers do not
2774 * set page.mapping but still dirty their pages
2776 * Drop the mmap_lock before waiting on IO, if we can. The file
2777 * is pinning the mapping, as per above.
2779 if ((dirtied || page_mkwrite) && mapping) {
2782 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2783 balance_dirty_pages_ratelimited(mapping);
2786 return VM_FAULT_RETRY;
2794 * Handle write page faults for pages that can be reused in the current vma
2796 * This can happen either due to the mapping being with the VM_SHARED flag,
2797 * or due to us being the last reference standing to the page. In either
2798 * case, all we need to do here is to mark the page as writable and update
2799 * any related book-keeping.
2801 static inline void wp_page_reuse(struct vm_fault *vmf)
2802 __releases(vmf->ptl)
2804 struct vm_area_struct *vma = vmf->vma;
2805 struct page *page = vmf->page;
2808 * Clear the pages cpupid information as the existing
2809 * information potentially belongs to a now completely
2810 * unrelated process.
2813 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2815 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2816 entry = pte_mkyoung(vmf->orig_pte);
2817 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2818 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2819 update_mmu_cache(vma, vmf->address, vmf->pte);
2820 pte_unmap_unlock(vmf->pte, vmf->ptl);
2821 count_vm_event(PGREUSE);
2825 * Handle the case of a page which we actually need to copy to a new page.
2827 * Called with mmap_lock locked and the old page referenced, but
2828 * without the ptl held.
2830 * High level logic flow:
2832 * - Allocate a page, copy the content of the old page to the new one.
2833 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2834 * - Take the PTL. If the pte changed, bail out and release the allocated page
2835 * - If the pte is still the way we remember it, update the page table and all
2836 * relevant references. This includes dropping the reference the page-table
2837 * held to the old page, as well as updating the rmap.
2838 * - In any case, unlock the PTL and drop the reference we took to the old page.
2840 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2842 struct vm_area_struct *vma = vmf->vma;
2843 struct mm_struct *mm = vma->vm_mm;
2844 struct page *old_page = vmf->page;
2845 struct page *new_page = NULL;
2847 int page_copied = 0;
2848 struct mmu_notifier_range range;
2850 if (unlikely(anon_vma_prepare(vma)))
2853 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2854 new_page = alloc_zeroed_user_highpage_movable(vma,
2859 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2864 if (!cow_user_page(new_page, old_page, vmf)) {
2866 * COW failed, if the fault was solved by other,
2867 * it's fine. If not, userspace would re-fault on
2868 * the same address and we will handle the fault
2869 * from the second attempt.
2878 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2880 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2882 __SetPageUptodate(new_page);
2884 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2885 vmf->address & PAGE_MASK,
2886 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2887 mmu_notifier_invalidate_range_start(&range);
2890 * Re-check the pte - we dropped the lock
2892 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2893 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2895 if (!PageAnon(old_page)) {
2896 dec_mm_counter_fast(mm,
2897 mm_counter_file(old_page));
2898 inc_mm_counter_fast(mm, MM_ANONPAGES);
2901 inc_mm_counter_fast(mm, MM_ANONPAGES);
2903 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2904 entry = mk_pte(new_page, vma->vm_page_prot);
2905 entry = pte_sw_mkyoung(entry);
2906 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2908 * Clear the pte entry and flush it first, before updating the
2909 * pte with the new entry. This will avoid a race condition
2910 * seen in the presence of one thread doing SMC and another
2913 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2914 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2915 lru_cache_add_inactive_or_unevictable(new_page, vma);
2917 * We call the notify macro here because, when using secondary
2918 * mmu page tables (such as kvm shadow page tables), we want the
2919 * new page to be mapped directly into the secondary page table.
2921 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2922 update_mmu_cache(vma, vmf->address, vmf->pte);
2925 * Only after switching the pte to the new page may
2926 * we remove the mapcount here. Otherwise another
2927 * process may come and find the rmap count decremented
2928 * before the pte is switched to the new page, and
2929 * "reuse" the old page writing into it while our pte
2930 * here still points into it and can be read by other
2933 * The critical issue is to order this
2934 * page_remove_rmap with the ptp_clear_flush above.
2935 * Those stores are ordered by (if nothing else,)
2936 * the barrier present in the atomic_add_negative
2937 * in page_remove_rmap.
2939 * Then the TLB flush in ptep_clear_flush ensures that
2940 * no process can access the old page before the
2941 * decremented mapcount is visible. And the old page
2942 * cannot be reused until after the decremented
2943 * mapcount is visible. So transitively, TLBs to
2944 * old page will be flushed before it can be reused.
2946 page_remove_rmap(old_page, false);
2949 /* Free the old page.. */
2950 new_page = old_page;
2953 update_mmu_tlb(vma, vmf->address, vmf->pte);
2959 pte_unmap_unlock(vmf->pte, vmf->ptl);
2961 * No need to double call mmu_notifier->invalidate_range() callback as
2962 * the above ptep_clear_flush_notify() did already call it.
2964 mmu_notifier_invalidate_range_only_end(&range);
2967 * Don't let another task, with possibly unlocked vma,
2968 * keep the mlocked page.
2970 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2971 lock_page(old_page); /* LRU manipulation */
2972 if (PageMlocked(old_page))
2973 munlock_vma_page(old_page);
2974 unlock_page(old_page);
2978 return page_copied ? VM_FAULT_WRITE : 0;
2984 return VM_FAULT_OOM;
2988 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2989 * writeable once the page is prepared
2991 * @vmf: structure describing the fault
2993 * This function handles all that is needed to finish a write page fault in a
2994 * shared mapping due to PTE being read-only once the mapped page is prepared.
2995 * It handles locking of PTE and modifying it.
2997 * The function expects the page to be locked or other protection against
2998 * concurrent faults / writeback (such as DAX radix tree locks).
3000 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
3001 * we acquired PTE lock.
3003 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3005 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3006 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3009 * We might have raced with another page fault while we released the
3010 * pte_offset_map_lock.
3012 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3013 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3014 pte_unmap_unlock(vmf->pte, vmf->ptl);
3015 return VM_FAULT_NOPAGE;
3022 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3025 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3027 struct vm_area_struct *vma = vmf->vma;
3029 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3032 pte_unmap_unlock(vmf->pte, vmf->ptl);
3033 vmf->flags |= FAULT_FLAG_MKWRITE;
3034 ret = vma->vm_ops->pfn_mkwrite(vmf);
3035 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3037 return finish_mkwrite_fault(vmf);
3040 return VM_FAULT_WRITE;
3043 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3044 __releases(vmf->ptl)
3046 struct vm_area_struct *vma = vmf->vma;
3047 vm_fault_t ret = VM_FAULT_WRITE;
3049 get_page(vmf->page);
3051 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3054 pte_unmap_unlock(vmf->pte, vmf->ptl);
3055 tmp = do_page_mkwrite(vmf);
3056 if (unlikely(!tmp || (tmp &
3057 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3058 put_page(vmf->page);
3061 tmp = finish_mkwrite_fault(vmf);
3062 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3063 unlock_page(vmf->page);
3064 put_page(vmf->page);
3069 lock_page(vmf->page);
3071 ret |= fault_dirty_shared_page(vmf);
3072 put_page(vmf->page);
3078 * This routine handles present pages, when users try to write
3079 * to a shared page. It is done by copying the page to a new address
3080 * and decrementing the shared-page counter for the old page.
3082 * Note that this routine assumes that the protection checks have been
3083 * done by the caller (the low-level page fault routine in most cases).
3084 * Thus we can safely just mark it writable once we've done any necessary
3087 * We also mark the page dirty at this point even though the page will
3088 * change only once the write actually happens. This avoids a few races,
3089 * and potentially makes it more efficient.
3091 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3092 * but allow concurrent faults), with pte both mapped and locked.
3093 * We return with mmap_lock still held, but pte unmapped and unlocked.
3095 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3096 __releases(vmf->ptl)
3098 struct vm_area_struct *vma = vmf->vma;
3100 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3101 pte_unmap_unlock(vmf->pte, vmf->ptl);
3102 return handle_userfault(vmf, VM_UFFD_WP);
3106 * Userfaultfd write-protect can defer flushes. Ensure the TLB
3107 * is flushed in this case before copying.
3109 if (unlikely(userfaultfd_wp(vmf->vma) &&
3110 mm_tlb_flush_pending(vmf->vma->vm_mm)))
3111 flush_tlb_page(vmf->vma, vmf->address);
3113 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3116 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3119 * We should not cow pages in a shared writeable mapping.
3120 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3122 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3123 (VM_WRITE|VM_SHARED))
3124 return wp_pfn_shared(vmf);
3126 pte_unmap_unlock(vmf->pte, vmf->ptl);
3127 return wp_page_copy(vmf);
3131 * Take out anonymous pages first, anonymous shared vmas are
3132 * not dirty accountable.
3134 if (PageAnon(vmf->page)) {
3135 struct page *page = vmf->page;
3137 /* PageKsm() doesn't necessarily raise the page refcount */
3138 if (PageKsm(page) || page_count(page) != 1)
3140 if (!trylock_page(page))
3142 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3147 * Ok, we've got the only map reference, and the only
3148 * page count reference, and the page is locked,
3149 * it's dark out, and we're wearing sunglasses. Hit it.
3153 return VM_FAULT_WRITE;
3154 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3155 (VM_WRITE|VM_SHARED))) {
3156 return wp_page_shared(vmf);
3160 * Ok, we need to copy. Oh, well..
3162 get_page(vmf->page);
3164 pte_unmap_unlock(vmf->pte, vmf->ptl);
3165 return wp_page_copy(vmf);
3168 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3169 unsigned long start_addr, unsigned long end_addr,
3170 struct zap_details *details)
3172 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3175 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3176 struct zap_details *details)
3178 struct vm_area_struct *vma;
3179 pgoff_t vba, vea, zba, zea;
3181 vma_interval_tree_foreach(vma, root,
3182 details->first_index, details->last_index) {
3184 vba = vma->vm_pgoff;
3185 vea = vba + vma_pages(vma) - 1;
3186 zba = details->first_index;
3189 zea = details->last_index;
3193 unmap_mapping_range_vma(vma,
3194 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3195 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3201 * unmap_mapping_page() - Unmap single page from processes.
3202 * @page: The locked page to be unmapped.
3204 * Unmap this page from any userspace process which still has it mmaped.
3205 * Typically, for efficiency, the range of nearby pages has already been
3206 * unmapped by unmap_mapping_pages() or unmap_mapping_range(). But once
3207 * truncation or invalidation holds the lock on a page, it may find that
3208 * the page has been remapped again: and then uses unmap_mapping_page()
3209 * to unmap it finally.
3211 void unmap_mapping_page(struct page *page)
3213 struct address_space *mapping = page->mapping;
3214 struct zap_details details = { };
3216 VM_BUG_ON(!PageLocked(page));
3217 VM_BUG_ON(PageTail(page));
3219 details.check_mapping = mapping;
3220 details.first_index = page->index;
3221 details.last_index = page->index + thp_nr_pages(page) - 1;
3222 details.single_page = page;
3224 i_mmap_lock_write(mapping);
3225 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3226 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3227 i_mmap_unlock_write(mapping);
3231 * unmap_mapping_pages() - Unmap pages from processes.
3232 * @mapping: The address space containing pages to be unmapped.
3233 * @start: Index of first page to be unmapped.
3234 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3235 * @even_cows: Whether to unmap even private COWed pages.
3237 * Unmap the pages in this address space from any userspace process which
3238 * has them mmaped. Generally, you want to remove COWed pages as well when
3239 * a file is being truncated, but not when invalidating pages from the page
3242 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3243 pgoff_t nr, bool even_cows)
3245 struct zap_details details = { };
3247 details.check_mapping = even_cows ? NULL : mapping;
3248 details.first_index = start;
3249 details.last_index = start + nr - 1;
3250 if (details.last_index < details.first_index)
3251 details.last_index = ULONG_MAX;
3253 i_mmap_lock_write(mapping);
3254 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3255 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3256 i_mmap_unlock_write(mapping);
3260 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3261 * address_space corresponding to the specified byte range in the underlying
3264 * @mapping: the address space containing mmaps to be unmapped.
3265 * @holebegin: byte in first page to unmap, relative to the start of
3266 * the underlying file. This will be rounded down to a PAGE_SIZE
3267 * boundary. Note that this is different from truncate_pagecache(), which
3268 * must keep the partial page. In contrast, we must get rid of
3270 * @holelen: size of prospective hole in bytes. This will be rounded
3271 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3273 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3274 * but 0 when invalidating pagecache, don't throw away private data.
3276 void unmap_mapping_range(struct address_space *mapping,
3277 loff_t const holebegin, loff_t const holelen, int even_cows)
3279 pgoff_t hba = holebegin >> PAGE_SHIFT;
3280 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3282 /* Check for overflow. */
3283 if (sizeof(holelen) > sizeof(hlen)) {
3285 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3286 if (holeend & ~(long long)ULONG_MAX)
3287 hlen = ULONG_MAX - hba + 1;
3290 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3292 EXPORT_SYMBOL(unmap_mapping_range);
3295 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3296 * but allow concurrent faults), and pte mapped but not yet locked.
3297 * We return with pte unmapped and unlocked.
3299 * We return with the mmap_lock locked or unlocked in the same cases
3300 * as does filemap_fault().
3302 vm_fault_t do_swap_page(struct vm_fault *vmf)
3304 struct vm_area_struct *vma = vmf->vma;
3305 struct page *page = NULL, *swapcache;
3311 void *shadow = NULL;
3313 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3316 entry = pte_to_swp_entry(vmf->orig_pte);
3317 if (unlikely(non_swap_entry(entry))) {
3318 if (is_migration_entry(entry)) {
3319 migration_entry_wait(vma->vm_mm, vmf->pmd,
3321 } else if (is_device_private_entry(entry)) {
3322 vmf->page = device_private_entry_to_page(entry);
3323 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3324 } else if (is_hwpoison_entry(entry)) {
3325 ret = VM_FAULT_HWPOISON;
3327 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3328 ret = VM_FAULT_SIGBUS;
3334 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3335 page = lookup_swap_cache(entry, vma, vmf->address);
3339 struct swap_info_struct *si = swp_swap_info(entry);
3341 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3342 __swap_count(entry) == 1) {
3343 /* skip swapcache */
3344 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3349 __SetPageLocked(page);
3350 __SetPageSwapBacked(page);
3351 set_page_private(page, entry.val);
3353 /* Tell memcg to use swap ownership records */
3354 SetPageSwapCache(page);
3355 err = mem_cgroup_charge(page, vma->vm_mm,
3357 ClearPageSwapCache(page);
3363 shadow = get_shadow_from_swap_cache(entry);
3365 workingset_refault(page, shadow);
3367 lru_cache_add(page);
3368 swap_readpage(page, true);
3371 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3378 * Back out if somebody else faulted in this pte
3379 * while we released the pte lock.
3381 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3382 vmf->address, &vmf->ptl);
3383 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3385 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3389 /* Had to read the page from swap area: Major fault */
3390 ret = VM_FAULT_MAJOR;
3391 count_vm_event(PGMAJFAULT);
3392 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3393 } else if (PageHWPoison(page)) {
3395 * hwpoisoned dirty swapcache pages are kept for killing
3396 * owner processes (which may be unknown at hwpoison time)
3398 ret = VM_FAULT_HWPOISON;
3399 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3403 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3405 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3407 ret |= VM_FAULT_RETRY;
3412 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3413 * release the swapcache from under us. The page pin, and pte_same
3414 * test below, are not enough to exclude that. Even if it is still
3415 * swapcache, we need to check that the page's swap has not changed.
3417 if (unlikely((!PageSwapCache(page) ||
3418 page_private(page) != entry.val)) && swapcache)
3421 page = ksm_might_need_to_copy(page, vma, vmf->address);
3422 if (unlikely(!page)) {
3428 cgroup_throttle_swaprate(page, GFP_KERNEL);
3431 * Back out if somebody else already faulted in this pte.
3433 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3435 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3438 if (unlikely(!PageUptodate(page))) {
3439 ret = VM_FAULT_SIGBUS;
3444 * The page isn't present yet, go ahead with the fault.
3446 * Be careful about the sequence of operations here.
3447 * To get its accounting right, reuse_swap_page() must be called
3448 * while the page is counted on swap but not yet in mapcount i.e.
3449 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3450 * must be called after the swap_free(), or it will never succeed.
3453 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3454 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3455 pte = mk_pte(page, vma->vm_page_prot);
3456 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3457 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3458 vmf->flags &= ~FAULT_FLAG_WRITE;
3459 ret |= VM_FAULT_WRITE;
3460 exclusive = RMAP_EXCLUSIVE;
3462 flush_icache_page(vma, page);
3463 if (pte_swp_soft_dirty(vmf->orig_pte))
3464 pte = pte_mksoft_dirty(pte);
3465 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3466 pte = pte_mkuffd_wp(pte);
3467 pte = pte_wrprotect(pte);
3469 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3470 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3471 vmf->orig_pte = pte;
3473 /* ksm created a completely new copy */
3474 if (unlikely(page != swapcache && swapcache)) {
3475 page_add_new_anon_rmap(page, vma, vmf->address, false);
3476 lru_cache_add_inactive_or_unevictable(page, vma);
3478 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3482 if (mem_cgroup_swap_full(page) ||
3483 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3484 try_to_free_swap(page);
3486 if (page != swapcache && swapcache) {
3488 * Hold the lock to avoid the swap entry to be reused
3489 * until we take the PT lock for the pte_same() check
3490 * (to avoid false positives from pte_same). For
3491 * further safety release the lock after the swap_free
3492 * so that the swap count won't change under a
3493 * parallel locked swapcache.
3495 unlock_page(swapcache);
3496 put_page(swapcache);
3499 if (vmf->flags & FAULT_FLAG_WRITE) {
3500 ret |= do_wp_page(vmf);
3501 if (ret & VM_FAULT_ERROR)
3502 ret &= VM_FAULT_ERROR;
3506 /* No need to invalidate - it was non-present before */
3507 update_mmu_cache(vma, vmf->address, vmf->pte);
3509 pte_unmap_unlock(vmf->pte, vmf->ptl);
3513 pte_unmap_unlock(vmf->pte, vmf->ptl);
3518 if (page != swapcache && swapcache) {
3519 unlock_page(swapcache);
3520 put_page(swapcache);
3526 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3527 * but allow concurrent faults), and pte mapped but not yet locked.
3528 * We return with mmap_lock still held, but pte unmapped and unlocked.
3530 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3532 struct vm_area_struct *vma = vmf->vma;
3537 /* File mapping without ->vm_ops ? */
3538 if (vma->vm_flags & VM_SHARED)
3539 return VM_FAULT_SIGBUS;
3542 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3543 * pte_offset_map() on pmds where a huge pmd might be created
3544 * from a different thread.
3546 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3547 * parallel threads are excluded by other means.
3549 * Here we only have mmap_read_lock(mm).
3551 if (pte_alloc(vma->vm_mm, vmf->pmd))
3552 return VM_FAULT_OOM;
3554 /* See the comment in pte_alloc_one_map() */
3555 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3558 /* Use the zero-page for reads */
3559 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3560 !mm_forbids_zeropage(vma->vm_mm)) {
3561 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3562 vma->vm_page_prot));
3563 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3564 vmf->address, &vmf->ptl);
3565 if (!pte_none(*vmf->pte)) {
3566 update_mmu_tlb(vma, vmf->address, vmf->pte);
3569 ret = check_stable_address_space(vma->vm_mm);
3572 /* Deliver the page fault to userland, check inside PT lock */
3573 if (userfaultfd_missing(vma)) {
3574 pte_unmap_unlock(vmf->pte, vmf->ptl);
3575 return handle_userfault(vmf, VM_UFFD_MISSING);
3580 /* Allocate our own private page. */
3581 if (unlikely(anon_vma_prepare(vma)))
3583 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3587 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3589 cgroup_throttle_swaprate(page, GFP_KERNEL);
3592 * The memory barrier inside __SetPageUptodate makes sure that
3593 * preceding stores to the page contents become visible before
3594 * the set_pte_at() write.
3596 __SetPageUptodate(page);
3598 entry = mk_pte(page, vma->vm_page_prot);
3599 entry = pte_sw_mkyoung(entry);
3600 if (vma->vm_flags & VM_WRITE)
3601 entry = pte_mkwrite(pte_mkdirty(entry));
3603 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3605 if (!pte_none(*vmf->pte)) {
3606 update_mmu_cache(vma, vmf->address, vmf->pte);
3610 ret = check_stable_address_space(vma->vm_mm);
3614 /* Deliver the page fault to userland, check inside PT lock */
3615 if (userfaultfd_missing(vma)) {
3616 pte_unmap_unlock(vmf->pte, vmf->ptl);
3618 return handle_userfault(vmf, VM_UFFD_MISSING);
3621 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3622 page_add_new_anon_rmap(page, vma, vmf->address, false);
3623 lru_cache_add_inactive_or_unevictable(page, vma);
3625 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3627 /* No need to invalidate - it was non-present before */
3628 update_mmu_cache(vma, vmf->address, vmf->pte);
3630 pte_unmap_unlock(vmf->pte, vmf->ptl);
3638 return VM_FAULT_OOM;
3642 * The mmap_lock must have been held on entry, and may have been
3643 * released depending on flags and vma->vm_ops->fault() return value.
3644 * See filemap_fault() and __lock_page_retry().
3646 static vm_fault_t __do_fault(struct vm_fault *vmf)
3648 struct vm_area_struct *vma = vmf->vma;
3652 * Preallocate pte before we take page_lock because this might lead to
3653 * deadlocks for memcg reclaim which waits for pages under writeback:
3655 * SetPageWriteback(A)
3661 * wait_on_page_writeback(A)
3662 * SetPageWriteback(B)
3664 * # flush A, B to clear the writeback
3666 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3667 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3668 if (!vmf->prealloc_pte)
3669 return VM_FAULT_OOM;
3670 smp_wmb(); /* See comment in __pte_alloc() */
3673 ret = vma->vm_ops->fault(vmf);
3674 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3675 VM_FAULT_DONE_COW)))
3678 if (unlikely(PageHWPoison(vmf->page))) {
3679 if (ret & VM_FAULT_LOCKED)
3680 unlock_page(vmf->page);
3681 put_page(vmf->page);
3683 return VM_FAULT_HWPOISON;
3686 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3687 lock_page(vmf->page);
3689 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3695 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3696 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3697 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3698 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3700 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3702 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3705 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3707 struct vm_area_struct *vma = vmf->vma;
3709 if (!pmd_none(*vmf->pmd))
3711 if (vmf->prealloc_pte) {
3712 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3713 if (unlikely(!pmd_none(*vmf->pmd))) {
3714 spin_unlock(vmf->ptl);
3718 mm_inc_nr_ptes(vma->vm_mm);
3719 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3720 spin_unlock(vmf->ptl);
3721 vmf->prealloc_pte = NULL;
3722 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3723 return VM_FAULT_OOM;
3727 * If a huge pmd materialized under us just retry later. Use
3728 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3729 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3730 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3731 * running immediately after a huge pmd fault in a different thread of
3732 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3733 * All we have to ensure is that it is a regular pmd that we can walk
3734 * with pte_offset_map() and we can do that through an atomic read in
3735 * C, which is what pmd_trans_unstable() provides.
3737 if (pmd_devmap_trans_unstable(vmf->pmd))
3738 return VM_FAULT_NOPAGE;
3741 * At this point we know that our vmf->pmd points to a page of ptes
3742 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3743 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3744 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3745 * be valid and we will re-check to make sure the vmf->pte isn't
3746 * pte_none() under vmf->ptl protection when we return to
3749 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3754 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3755 static void deposit_prealloc_pte(struct vm_fault *vmf)
3757 struct vm_area_struct *vma = vmf->vma;
3759 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3761 * We are going to consume the prealloc table,
3762 * count that as nr_ptes.
3764 mm_inc_nr_ptes(vma->vm_mm);
3765 vmf->prealloc_pte = NULL;
3768 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3770 struct vm_area_struct *vma = vmf->vma;
3771 bool write = vmf->flags & FAULT_FLAG_WRITE;
3772 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3775 vm_fault_t ret = VM_FAULT_FALLBACK;
3777 if (!transhuge_vma_suitable(vma, haddr))
3780 page = compound_head(page);
3781 if (compound_order(page) != HPAGE_PMD_ORDER)
3785 * Archs like ppc64 need additonal space to store information
3786 * related to pte entry. Use the preallocated table for that.
3788 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3789 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3790 if (!vmf->prealloc_pte)
3791 return VM_FAULT_OOM;
3792 smp_wmb(); /* See comment in __pte_alloc() */
3795 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3796 if (unlikely(!pmd_none(*vmf->pmd)))
3799 for (i = 0; i < HPAGE_PMD_NR; i++)
3800 flush_icache_page(vma, page + i);
3802 entry = mk_huge_pmd(page, vma->vm_page_prot);
3804 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3806 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3807 page_add_file_rmap(page, true);
3809 * deposit and withdraw with pmd lock held
3811 if (arch_needs_pgtable_deposit())
3812 deposit_prealloc_pte(vmf);
3814 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3816 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3818 /* fault is handled */
3820 count_vm_event(THP_FILE_MAPPED);
3822 spin_unlock(vmf->ptl);
3826 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3834 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3835 * mapping. If needed, the function allocates page table or use pre-allocated.
3837 * @vmf: fault environment
3838 * @page: page to map
3840 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3843 * Target users are page handler itself and implementations of
3844 * vm_ops->map_pages.
3846 * Return: %0 on success, %VM_FAULT_ code in case of error.
3848 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
3850 struct vm_area_struct *vma = vmf->vma;
3851 bool write = vmf->flags & FAULT_FLAG_WRITE;
3855 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3856 ret = do_set_pmd(vmf, page);
3857 if (ret != VM_FAULT_FALLBACK)
3862 ret = pte_alloc_one_map(vmf);
3867 /* Re-check under ptl */
3868 if (unlikely(!pte_none(*vmf->pte))) {
3869 update_mmu_tlb(vma, vmf->address, vmf->pte);
3870 return VM_FAULT_NOPAGE;
3873 flush_icache_page(vma, page);
3874 entry = mk_pte(page, vma->vm_page_prot);
3875 entry = pte_sw_mkyoung(entry);
3877 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3878 /* copy-on-write page */
3879 if (write && !(vma->vm_flags & VM_SHARED)) {
3880 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3881 page_add_new_anon_rmap(page, vma, vmf->address, false);
3882 lru_cache_add_inactive_or_unevictable(page, vma);
3884 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3885 page_add_file_rmap(page, false);
3887 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3889 /* no need to invalidate: a not-present page won't be cached */
3890 update_mmu_cache(vma, vmf->address, vmf->pte);
3897 * finish_fault - finish page fault once we have prepared the page to fault
3899 * @vmf: structure describing the fault
3901 * This function handles all that is needed to finish a page fault once the
3902 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3903 * given page, adds reverse page mapping, handles memcg charges and LRU
3906 * The function expects the page to be locked and on success it consumes a
3907 * reference of a page being mapped (for the PTE which maps it).
3909 * Return: %0 on success, %VM_FAULT_ code in case of error.
3911 vm_fault_t finish_fault(struct vm_fault *vmf)
3916 /* Did we COW the page? */
3917 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3918 !(vmf->vma->vm_flags & VM_SHARED))
3919 page = vmf->cow_page;
3924 * check even for read faults because we might have lost our CoWed
3927 if (!(vmf->vma->vm_flags & VM_SHARED))
3928 ret = check_stable_address_space(vmf->vma->vm_mm);
3930 ret = alloc_set_pte(vmf, page);
3932 pte_unmap_unlock(vmf->pte, vmf->ptl);
3936 static unsigned long fault_around_bytes __read_mostly =
3937 rounddown_pow_of_two(65536);
3939 #ifdef CONFIG_DEBUG_FS
3940 static int fault_around_bytes_get(void *data, u64 *val)
3942 *val = fault_around_bytes;
3947 * fault_around_bytes must be rounded down to the nearest page order as it's
3948 * what do_fault_around() expects to see.
3950 static int fault_around_bytes_set(void *data, u64 val)
3952 if (val / PAGE_SIZE > PTRS_PER_PTE)
3954 if (val > PAGE_SIZE)
3955 fault_around_bytes = rounddown_pow_of_two(val);
3957 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3960 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3961 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3963 static int __init fault_around_debugfs(void)
3965 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3966 &fault_around_bytes_fops);
3969 late_initcall(fault_around_debugfs);
3973 * do_fault_around() tries to map few pages around the fault address. The hope
3974 * is that the pages will be needed soon and this will lower the number of
3977 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3978 * not ready to be mapped: not up-to-date, locked, etc.
3980 * This function is called with the page table lock taken. In the split ptlock
3981 * case the page table lock only protects only those entries which belong to
3982 * the page table corresponding to the fault address.
3984 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3987 * fault_around_bytes defines how many bytes we'll try to map.
3988 * do_fault_around() expects it to be set to a power of two less than or equal
3991 * The virtual address of the area that we map is naturally aligned to
3992 * fault_around_bytes rounded down to the machine page size
3993 * (and therefore to page order). This way it's easier to guarantee
3994 * that we don't cross page table boundaries.
3996 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3998 unsigned long address = vmf->address, nr_pages, mask;
3999 pgoff_t start_pgoff = vmf->pgoff;
4004 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4005 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4007 vmf->address = max(address & mask, vmf->vma->vm_start);
4008 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4012 * end_pgoff is either the end of the page table, the end of
4013 * the vma or nr_pages from start_pgoff, depending what is nearest.
4015 end_pgoff = start_pgoff -
4016 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4018 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4019 start_pgoff + nr_pages - 1);
4021 if (pmd_none(*vmf->pmd)) {
4022 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4023 if (!vmf->prealloc_pte)
4025 smp_wmb(); /* See comment in __pte_alloc() */
4028 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4030 /* Huge page is mapped? Page fault is solved */
4031 if (pmd_trans_huge(*vmf->pmd)) {
4032 ret = VM_FAULT_NOPAGE;
4036 /* ->map_pages() haven't done anything useful. Cold page cache? */
4040 /* check if the page fault is solved */
4041 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
4042 if (!pte_none(*vmf->pte))
4043 ret = VM_FAULT_NOPAGE;
4044 pte_unmap_unlock(vmf->pte, vmf->ptl);
4046 vmf->address = address;
4051 static vm_fault_t do_read_fault(struct vm_fault *vmf)
4053 struct vm_area_struct *vma = vmf->vma;
4057 * Let's call ->map_pages() first and use ->fault() as fallback
4058 * if page by the offset is not ready to be mapped (cold cache or
4061 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4062 ret = do_fault_around(vmf);
4067 ret = __do_fault(vmf);
4068 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4071 ret |= finish_fault(vmf);
4072 unlock_page(vmf->page);
4073 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4074 put_page(vmf->page);
4078 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4080 struct vm_area_struct *vma = vmf->vma;
4083 if (unlikely(anon_vma_prepare(vma)))
4084 return VM_FAULT_OOM;
4086 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4088 return VM_FAULT_OOM;
4090 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4091 put_page(vmf->cow_page);
4092 return VM_FAULT_OOM;
4094 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4096 ret = __do_fault(vmf);
4097 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4099 if (ret & VM_FAULT_DONE_COW)
4102 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4103 __SetPageUptodate(vmf->cow_page);
4105 ret |= finish_fault(vmf);
4106 unlock_page(vmf->page);
4107 put_page(vmf->page);
4108 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4112 put_page(vmf->cow_page);
4116 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4118 struct vm_area_struct *vma = vmf->vma;
4119 vm_fault_t ret, tmp;
4121 ret = __do_fault(vmf);
4122 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4126 * Check if the backing address space wants to know that the page is
4127 * about to become writable
4129 if (vma->vm_ops->page_mkwrite) {
4130 unlock_page(vmf->page);
4131 tmp = do_page_mkwrite(vmf);
4132 if (unlikely(!tmp ||
4133 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4134 put_page(vmf->page);
4139 ret |= finish_fault(vmf);
4140 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4142 unlock_page(vmf->page);
4143 put_page(vmf->page);
4147 ret |= fault_dirty_shared_page(vmf);
4152 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4153 * but allow concurrent faults).
4154 * The mmap_lock may have been released depending on flags and our
4155 * return value. See filemap_fault() and __lock_page_or_retry().
4156 * If mmap_lock is released, vma may become invalid (for example
4157 * by other thread calling munmap()).
4159 static vm_fault_t do_fault(struct vm_fault *vmf)
4161 struct vm_area_struct *vma = vmf->vma;
4162 struct mm_struct *vm_mm = vma->vm_mm;
4166 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4168 if (!vma->vm_ops->fault) {
4170 * If we find a migration pmd entry or a none pmd entry, which
4171 * should never happen, return SIGBUS
4173 if (unlikely(!pmd_present(*vmf->pmd)))
4174 ret = VM_FAULT_SIGBUS;
4176 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4181 * Make sure this is not a temporary clearing of pte
4182 * by holding ptl and checking again. A R/M/W update
4183 * of pte involves: take ptl, clearing the pte so that
4184 * we don't have concurrent modification by hardware
4185 * followed by an update.
4187 if (unlikely(pte_none(*vmf->pte)))
4188 ret = VM_FAULT_SIGBUS;
4190 ret = VM_FAULT_NOPAGE;
4192 pte_unmap_unlock(vmf->pte, vmf->ptl);
4194 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4195 ret = do_read_fault(vmf);
4196 else if (!(vma->vm_flags & VM_SHARED))
4197 ret = do_cow_fault(vmf);
4199 ret = do_shared_fault(vmf);
4201 /* preallocated pagetable is unused: free it */
4202 if (vmf->prealloc_pte) {
4203 pte_free(vm_mm, vmf->prealloc_pte);
4204 vmf->prealloc_pte = NULL;
4209 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4210 unsigned long addr, int page_nid,
4215 count_vm_numa_event(NUMA_HINT_FAULTS);
4216 if (page_nid == numa_node_id()) {
4217 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4218 *flags |= TNF_FAULT_LOCAL;
4221 return mpol_misplaced(page, vma, addr);
4224 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4226 struct vm_area_struct *vma = vmf->vma;
4227 struct page *page = NULL;
4228 int page_nid = NUMA_NO_NODE;
4231 bool migrated = false;
4233 bool was_writable = pte_savedwrite(vmf->orig_pte);
4237 * The "pte" at this point cannot be used safely without
4238 * validation through pte_unmap_same(). It's of NUMA type but
4239 * the pfn may be screwed if the read is non atomic.
4241 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4242 spin_lock(vmf->ptl);
4243 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4244 pte_unmap_unlock(vmf->pte, vmf->ptl);
4249 * Make it present again, Depending on how arch implementes non
4250 * accessible ptes, some can allow access by kernel mode.
4252 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4253 pte = pte_modify(old_pte, vma->vm_page_prot);
4254 pte = pte_mkyoung(pte);
4256 pte = pte_mkwrite(pte);
4257 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4258 update_mmu_cache(vma, vmf->address, vmf->pte);
4260 page = vm_normal_page(vma, vmf->address, pte);
4262 pte_unmap_unlock(vmf->pte, vmf->ptl);
4266 /* TODO: handle PTE-mapped THP */
4267 if (PageCompound(page)) {
4268 pte_unmap_unlock(vmf->pte, vmf->ptl);
4273 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4274 * much anyway since they can be in shared cache state. This misses
4275 * the case where a mapping is writable but the process never writes
4276 * to it but pte_write gets cleared during protection updates and
4277 * pte_dirty has unpredictable behaviour between PTE scan updates,
4278 * background writeback, dirty balancing and application behaviour.
4280 if (!pte_write(pte))
4281 flags |= TNF_NO_GROUP;
4284 * Flag if the page is shared between multiple address spaces. This
4285 * is later used when determining whether to group tasks together
4287 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4288 flags |= TNF_SHARED;
4290 last_cpupid = page_cpupid_last(page);
4291 page_nid = page_to_nid(page);
4292 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4294 pte_unmap_unlock(vmf->pte, vmf->ptl);
4295 if (target_nid == NUMA_NO_NODE) {
4300 /* Migrate to the requested node */
4301 migrated = migrate_misplaced_page(page, vma, target_nid);
4303 page_nid = target_nid;
4304 flags |= TNF_MIGRATED;
4306 flags |= TNF_MIGRATE_FAIL;
4309 if (page_nid != NUMA_NO_NODE)
4310 task_numa_fault(last_cpupid, page_nid, 1, flags);
4314 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4316 if (vma_is_anonymous(vmf->vma))
4317 return do_huge_pmd_anonymous_page(vmf);
4318 if (vmf->vma->vm_ops->huge_fault)
4319 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4320 return VM_FAULT_FALLBACK;
4323 /* `inline' is required to avoid gcc 4.1.2 build error */
4324 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4326 if (vma_is_anonymous(vmf->vma)) {
4327 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4328 return handle_userfault(vmf, VM_UFFD_WP);
4329 return do_huge_pmd_wp_page(vmf, orig_pmd);
4331 if (vmf->vma->vm_ops->huge_fault) {
4332 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4334 if (!(ret & VM_FAULT_FALLBACK))
4338 /* COW or write-notify handled on pte level: split pmd. */
4339 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4341 return VM_FAULT_FALLBACK;
4344 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4346 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4347 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4348 /* No support for anonymous transparent PUD pages yet */
4349 if (vma_is_anonymous(vmf->vma))
4351 if (vmf->vma->vm_ops->huge_fault) {
4352 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4354 if (!(ret & VM_FAULT_FALLBACK))
4358 /* COW or write-notify not handled on PUD level: split pud.*/
4359 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4360 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4361 return VM_FAULT_FALLBACK;
4364 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4366 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4367 /* No support for anonymous transparent PUD pages yet */
4368 if (vma_is_anonymous(vmf->vma))
4369 return VM_FAULT_FALLBACK;
4370 if (vmf->vma->vm_ops->huge_fault)
4371 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4372 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4373 return VM_FAULT_FALLBACK;
4377 * These routines also need to handle stuff like marking pages dirty
4378 * and/or accessed for architectures that don't do it in hardware (most
4379 * RISC architectures). The early dirtying is also good on the i386.
4381 * There is also a hook called "update_mmu_cache()" that architectures
4382 * with external mmu caches can use to update those (ie the Sparc or
4383 * PowerPC hashed page tables that act as extended TLBs).
4385 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4386 * concurrent faults).
4388 * The mmap_lock may have been released depending on flags and our return value.
4389 * See filemap_fault() and __lock_page_or_retry().
4391 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4395 if (unlikely(pmd_none(*vmf->pmd))) {
4397 * Leave __pte_alloc() until later: because vm_ops->fault may
4398 * want to allocate huge page, and if we expose page table
4399 * for an instant, it will be difficult to retract from
4400 * concurrent faults and from rmap lookups.
4404 /* See comment in pte_alloc_one_map() */
4405 if (pmd_devmap_trans_unstable(vmf->pmd))
4408 * A regular pmd is established and it can't morph into a huge
4409 * pmd from under us anymore at this point because we hold the
4410 * mmap_lock read mode and khugepaged takes it in write mode.
4411 * So now it's safe to run pte_offset_map().
4413 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4414 vmf->orig_pte = *vmf->pte;
4417 * some architectures can have larger ptes than wordsize,
4418 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4419 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4420 * accesses. The code below just needs a consistent view
4421 * for the ifs and we later double check anyway with the
4422 * ptl lock held. So here a barrier will do.
4425 if (pte_none(vmf->orig_pte)) {
4426 pte_unmap(vmf->pte);
4432 if (vma_is_anonymous(vmf->vma))
4433 return do_anonymous_page(vmf);
4435 return do_fault(vmf);
4438 if (!pte_present(vmf->orig_pte))
4439 return do_swap_page(vmf);
4441 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4442 return do_numa_page(vmf);
4444 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4445 spin_lock(vmf->ptl);
4446 entry = vmf->orig_pte;
4447 if (unlikely(!pte_same(*vmf->pte, entry))) {
4448 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4451 if (vmf->flags & FAULT_FLAG_WRITE) {
4452 if (!pte_write(entry))
4453 return do_wp_page(vmf);
4454 entry = pte_mkdirty(entry);
4456 entry = pte_mkyoung(entry);
4457 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4458 vmf->flags & FAULT_FLAG_WRITE)) {
4459 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4461 /* Skip spurious TLB flush for retried page fault */
4462 if (vmf->flags & FAULT_FLAG_TRIED)
4465 * This is needed only for protection faults but the arch code
4466 * is not yet telling us if this is a protection fault or not.
4467 * This still avoids useless tlb flushes for .text page faults
4470 if (vmf->flags & FAULT_FLAG_WRITE)
4471 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4474 pte_unmap_unlock(vmf->pte, vmf->ptl);
4479 * By the time we get here, we already hold the mm semaphore
4481 * The mmap_lock may have been released depending on flags and our
4482 * return value. See filemap_fault() and __lock_page_or_retry().
4484 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4485 unsigned long address, unsigned int flags)
4487 struct vm_fault vmf = {
4489 .address = address & PAGE_MASK,
4491 .pgoff = linear_page_index(vma, address),
4492 .gfp_mask = __get_fault_gfp_mask(vma),
4494 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4495 struct mm_struct *mm = vma->vm_mm;
4500 pgd = pgd_offset(mm, address);
4501 p4d = p4d_alloc(mm, pgd, address);
4503 return VM_FAULT_OOM;
4505 vmf.pud = pud_alloc(mm, p4d, address);
4507 return VM_FAULT_OOM;
4509 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4510 ret = create_huge_pud(&vmf);
4511 if (!(ret & VM_FAULT_FALLBACK))
4514 pud_t orig_pud = *vmf.pud;
4517 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4519 /* NUMA case for anonymous PUDs would go here */
4521 if (dirty && !pud_write(orig_pud)) {
4522 ret = wp_huge_pud(&vmf, orig_pud);
4523 if (!(ret & VM_FAULT_FALLBACK))
4526 huge_pud_set_accessed(&vmf, orig_pud);
4532 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4534 return VM_FAULT_OOM;
4536 /* Huge pud page fault raced with pmd_alloc? */
4537 if (pud_trans_unstable(vmf.pud))
4540 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4541 ret = create_huge_pmd(&vmf);
4542 if (!(ret & VM_FAULT_FALLBACK))
4545 pmd_t orig_pmd = *vmf.pmd;
4548 if (unlikely(is_swap_pmd(orig_pmd))) {
4549 VM_BUG_ON(thp_migration_supported() &&
4550 !is_pmd_migration_entry(orig_pmd));
4551 if (is_pmd_migration_entry(orig_pmd))
4552 pmd_migration_entry_wait(mm, vmf.pmd);
4555 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4556 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4557 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4559 if (dirty && !pmd_write(orig_pmd)) {
4560 ret = wp_huge_pmd(&vmf, orig_pmd);
4561 if (!(ret & VM_FAULT_FALLBACK))
4564 huge_pmd_set_accessed(&vmf, orig_pmd);
4570 return handle_pte_fault(&vmf);
4574 * mm_account_fault - Do page fault accountings
4576 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4577 * of perf event counters, but we'll still do the per-task accounting to
4578 * the task who triggered this page fault.
4579 * @address: the faulted address.
4580 * @flags: the fault flags.
4581 * @ret: the fault retcode.
4583 * This will take care of most of the page fault accountings. Meanwhile, it
4584 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4585 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4586 * still be in per-arch page fault handlers at the entry of page fault.
4588 static inline void mm_account_fault(struct pt_regs *regs,
4589 unsigned long address, unsigned int flags,
4595 * We don't do accounting for some specific faults:
4597 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4598 * includes arch_vma_access_permitted() failing before reaching here.
4599 * So this is not a "this many hardware page faults" counter. We
4600 * should use the hw profiling for that.
4602 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4603 * once they're completed.
4605 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4609 * We define the fault as a major fault when the final successful fault
4610 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4611 * handle it immediately previously).
4613 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4621 * If the fault is done for GUP, regs will be NULL. We only do the
4622 * accounting for the per thread fault counters who triggered the
4623 * fault, and we skip the perf event updates.
4629 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4631 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4635 * By the time we get here, we already hold the mm semaphore
4637 * The mmap_lock may have been released depending on flags and our
4638 * return value. See filemap_fault() and __lock_page_or_retry().
4640 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4641 unsigned int flags, struct pt_regs *regs)
4645 __set_current_state(TASK_RUNNING);
4647 count_vm_event(PGFAULT);
4648 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4650 /* do counter updates before entering really critical section. */
4651 check_sync_rss_stat(current);
4653 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4654 flags & FAULT_FLAG_INSTRUCTION,
4655 flags & FAULT_FLAG_REMOTE))
4656 return VM_FAULT_SIGSEGV;
4659 * Enable the memcg OOM handling for faults triggered in user
4660 * space. Kernel faults are handled more gracefully.
4662 if (flags & FAULT_FLAG_USER)
4663 mem_cgroup_enter_user_fault();
4665 if (unlikely(is_vm_hugetlb_page(vma)))
4666 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4668 ret = __handle_mm_fault(vma, address, flags);
4670 if (flags & FAULT_FLAG_USER) {
4671 mem_cgroup_exit_user_fault();
4673 * The task may have entered a memcg OOM situation but
4674 * if the allocation error was handled gracefully (no
4675 * VM_FAULT_OOM), there is no need to kill anything.
4676 * Just clean up the OOM state peacefully.
4678 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4679 mem_cgroup_oom_synchronize(false);
4682 mm_account_fault(regs, address, flags, ret);
4686 EXPORT_SYMBOL_GPL(handle_mm_fault);
4688 #ifndef __PAGETABLE_P4D_FOLDED
4690 * Allocate p4d page table.
4691 * We've already handled the fast-path in-line.
4693 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4695 p4d_t *new = p4d_alloc_one(mm, address);
4699 smp_wmb(); /* See comment in __pte_alloc */
4701 spin_lock(&mm->page_table_lock);
4702 if (pgd_present(*pgd)) /* Another has populated it */
4705 pgd_populate(mm, pgd, new);
4706 spin_unlock(&mm->page_table_lock);
4709 #endif /* __PAGETABLE_P4D_FOLDED */
4711 #ifndef __PAGETABLE_PUD_FOLDED
4713 * Allocate page upper directory.
4714 * We've already handled the fast-path in-line.
4716 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4718 pud_t *new = pud_alloc_one(mm, address);
4722 smp_wmb(); /* See comment in __pte_alloc */
4724 spin_lock(&mm->page_table_lock);
4725 if (!p4d_present(*p4d)) {
4727 p4d_populate(mm, p4d, new);
4728 } else /* Another has populated it */
4730 spin_unlock(&mm->page_table_lock);
4733 #endif /* __PAGETABLE_PUD_FOLDED */
4735 #ifndef __PAGETABLE_PMD_FOLDED
4737 * Allocate page middle directory.
4738 * We've already handled the fast-path in-line.
4740 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4743 pmd_t *new = pmd_alloc_one(mm, address);
4747 smp_wmb(); /* See comment in __pte_alloc */
4749 ptl = pud_lock(mm, pud);
4750 if (!pud_present(*pud)) {
4752 pud_populate(mm, pud, new);
4753 } else /* Another has populated it */
4758 #endif /* __PAGETABLE_PMD_FOLDED */
4760 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4761 struct mmu_notifier_range *range, pte_t **ptepp,
4762 pmd_t **pmdpp, spinlock_t **ptlp)
4770 pgd = pgd_offset(mm, address);
4771 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4774 p4d = p4d_offset(pgd, address);
4775 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4778 pud = pud_offset(p4d, address);
4779 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4782 pmd = pmd_offset(pud, address);
4783 VM_BUG_ON(pmd_trans_huge(*pmd));
4785 if (pmd_huge(*pmd)) {
4790 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4791 NULL, mm, address & PMD_MASK,
4792 (address & PMD_MASK) + PMD_SIZE);
4793 mmu_notifier_invalidate_range_start(range);
4795 *ptlp = pmd_lock(mm, pmd);
4796 if (pmd_huge(*pmd)) {
4802 mmu_notifier_invalidate_range_end(range);
4805 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4809 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4810 address & PAGE_MASK,
4811 (address & PAGE_MASK) + PAGE_SIZE);
4812 mmu_notifier_invalidate_range_start(range);
4814 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4815 if (!pte_present(*ptep))
4820 pte_unmap_unlock(ptep, *ptlp);
4822 mmu_notifier_invalidate_range_end(range);
4828 * follow_pte - look up PTE at a user virtual address
4829 * @mm: the mm_struct of the target address space
4830 * @address: user virtual address
4831 * @ptepp: location to store found PTE
4832 * @ptlp: location to store the lock for the PTE
4834 * On a successful return, the pointer to the PTE is stored in @ptepp;
4835 * the corresponding lock is taken and its location is stored in @ptlp.
4836 * The contents of the PTE are only stable until @ptlp is released;
4837 * any further use, if any, must be protected against invalidation
4838 * with MMU notifiers.
4840 * Only IO mappings and raw PFN mappings are allowed. The mmap semaphore
4841 * should be taken for read.
4843 * KVM uses this function. While it is arguably less bad than ``follow_pfn``,
4844 * it is not a good general-purpose API.
4846 * Return: zero on success, -ve otherwise.
4848 int follow_pte(struct mm_struct *mm, unsigned long address,
4849 pte_t **ptepp, spinlock_t **ptlp)
4851 return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4853 EXPORT_SYMBOL_GPL(follow_pte);
4856 * follow_pfn - look up PFN at a user virtual address
4857 * @vma: memory mapping
4858 * @address: user virtual address
4859 * @pfn: location to store found PFN
4861 * Only IO mappings and raw PFN mappings are allowed.
4863 * This function does not allow the caller to read the permissions
4864 * of the PTE. Do not use it.
4866 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4868 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4875 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4878 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4881 *pfn = pte_pfn(*ptep);
4882 pte_unmap_unlock(ptep, ptl);
4885 EXPORT_SYMBOL(follow_pfn);
4887 #ifdef CONFIG_HAVE_IOREMAP_PROT
4888 int follow_phys(struct vm_area_struct *vma,
4889 unsigned long address, unsigned int flags,
4890 unsigned long *prot, resource_size_t *phys)
4896 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4899 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4903 if ((flags & FOLL_WRITE) && !pte_write(pte))
4906 *prot = pgprot_val(pte_pgprot(pte));
4907 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4911 pte_unmap_unlock(ptep, ptl);
4916 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4917 void *buf, int len, int write)
4919 resource_size_t phys_addr;
4920 unsigned long prot = 0;
4921 void __iomem *maddr;
4922 int offset = addr & (PAGE_SIZE-1);
4924 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4927 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4932 memcpy_toio(maddr + offset, buf, len);
4934 memcpy_fromio(buf, maddr + offset, len);
4939 EXPORT_SYMBOL_GPL(generic_access_phys);
4943 * Access another process' address space as given in mm. If non-NULL, use the
4944 * given task for page fault accounting.
4946 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4947 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4949 struct vm_area_struct *vma;
4950 void *old_buf = buf;
4951 int write = gup_flags & FOLL_WRITE;
4953 if (mmap_read_lock_killable(mm))
4956 /* ignore errors, just check how much was successfully transferred */
4958 int bytes, ret, offset;
4960 struct page *page = NULL;
4962 ret = get_user_pages_remote(mm, addr, 1,
4963 gup_flags, &page, &vma, NULL);
4965 #ifndef CONFIG_HAVE_IOREMAP_PROT
4969 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4970 * we can access using slightly different code.
4972 vma = find_vma(mm, addr);
4973 if (!vma || vma->vm_start > addr)
4975 if (vma->vm_ops && vma->vm_ops->access)
4976 ret = vma->vm_ops->access(vma, addr, buf,
4984 offset = addr & (PAGE_SIZE-1);
4985 if (bytes > PAGE_SIZE-offset)
4986 bytes = PAGE_SIZE-offset;
4990 copy_to_user_page(vma, page, addr,
4991 maddr + offset, buf, bytes);
4992 set_page_dirty_lock(page);
4994 copy_from_user_page(vma, page, addr,
4995 buf, maddr + offset, bytes);
5004 mmap_read_unlock(mm);
5006 return buf - old_buf;
5010 * access_remote_vm - access another process' address space
5011 * @mm: the mm_struct of the target address space
5012 * @addr: start address to access
5013 * @buf: source or destination buffer
5014 * @len: number of bytes to transfer
5015 * @gup_flags: flags modifying lookup behaviour
5017 * The caller must hold a reference on @mm.
5019 * Return: number of bytes copied from source to destination.
5021 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5022 void *buf, int len, unsigned int gup_flags)
5024 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
5028 * Access another process' address space.
5029 * Source/target buffer must be kernel space,
5030 * Do not walk the page table directly, use get_user_pages
5032 int access_process_vm(struct task_struct *tsk, unsigned long addr,
5033 void *buf, int len, unsigned int gup_flags)
5035 struct mm_struct *mm;
5038 mm = get_task_mm(tsk);
5042 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
5048 EXPORT_SYMBOL_GPL(access_process_vm);
5051 * Print the name of a VMA.
5053 void print_vma_addr(char *prefix, unsigned long ip)
5055 struct mm_struct *mm = current->mm;
5056 struct vm_area_struct *vma;
5059 * we might be running from an atomic context so we cannot sleep
5061 if (!mmap_read_trylock(mm))
5064 vma = find_vma(mm, ip);
5065 if (vma && vma->vm_file) {
5066 struct file *f = vma->vm_file;
5067 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5071 p = file_path(f, buf, PAGE_SIZE);
5074 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5076 vma->vm_end - vma->vm_start);
5077 free_page((unsigned long)buf);
5080 mmap_read_unlock(mm);
5083 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5084 void __might_fault(const char *file, int line)
5087 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5088 * holding the mmap_lock, this is safe because kernel memory doesn't
5089 * get paged out, therefore we'll never actually fault, and the
5090 * below annotations will generate false positives.
5092 if (uaccess_kernel())
5094 if (pagefault_disabled())
5096 __might_sleep(file, line, 0);
5097 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5099 might_lock_read(¤t->mm->mmap_lock);
5102 EXPORT_SYMBOL(__might_fault);
5105 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5107 * Process all subpages of the specified huge page with the specified
5108 * operation. The target subpage will be processed last to keep its
5111 static inline void process_huge_page(
5112 unsigned long addr_hint, unsigned int pages_per_huge_page,
5113 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5117 unsigned long addr = addr_hint &
5118 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5120 /* Process target subpage last to keep its cache lines hot */
5122 n = (addr_hint - addr) / PAGE_SIZE;
5123 if (2 * n <= pages_per_huge_page) {
5124 /* If target subpage in first half of huge page */
5127 /* Process subpages at the end of huge page */
5128 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5130 process_subpage(addr + i * PAGE_SIZE, i, arg);
5133 /* If target subpage in second half of huge page */
5134 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5135 l = pages_per_huge_page - n;
5136 /* Process subpages at the begin of huge page */
5137 for (i = 0; i < base; i++) {
5139 process_subpage(addr + i * PAGE_SIZE, i, arg);
5143 * Process remaining subpages in left-right-left-right pattern
5144 * towards the target subpage
5146 for (i = 0; i < l; i++) {
5147 int left_idx = base + i;
5148 int right_idx = base + 2 * l - 1 - i;
5151 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5153 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5157 static void clear_gigantic_page(struct page *page,
5159 unsigned int pages_per_huge_page)
5162 struct page *p = page;
5165 for (i = 0; i < pages_per_huge_page;
5166 i++, p = mem_map_next(p, page, i)) {
5168 clear_user_highpage(p, addr + i * PAGE_SIZE);
5172 static void clear_subpage(unsigned long addr, int idx, void *arg)
5174 struct page *page = arg;
5176 clear_user_highpage(page + idx, addr);
5179 void clear_huge_page(struct page *page,
5180 unsigned long addr_hint, unsigned int pages_per_huge_page)
5182 unsigned long addr = addr_hint &
5183 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5185 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5186 clear_gigantic_page(page, addr, pages_per_huge_page);
5190 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5193 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5195 struct vm_area_struct *vma,
5196 unsigned int pages_per_huge_page)
5199 struct page *dst_base = dst;
5200 struct page *src_base = src;
5202 for (i = 0; i < pages_per_huge_page; ) {
5204 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5207 dst = mem_map_next(dst, dst_base, i);
5208 src = mem_map_next(src, src_base, i);
5212 struct copy_subpage_arg {
5215 struct vm_area_struct *vma;
5218 static void copy_subpage(unsigned long addr, int idx, void *arg)
5220 struct copy_subpage_arg *copy_arg = arg;
5222 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5223 addr, copy_arg->vma);
5226 void copy_user_huge_page(struct page *dst, struct page *src,
5227 unsigned long addr_hint, struct vm_area_struct *vma,
5228 unsigned int pages_per_huge_page)
5230 unsigned long addr = addr_hint &
5231 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5232 struct copy_subpage_arg arg = {
5238 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5239 copy_user_gigantic_page(dst, src, addr, vma,
5240 pages_per_huge_page);
5244 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5247 long copy_huge_page_from_user(struct page *dst_page,
5248 const void __user *usr_src,
5249 unsigned int pages_per_huge_page,
5250 bool allow_pagefault)
5252 void *src = (void *)usr_src;
5254 unsigned long i, rc = 0;
5255 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5256 struct page *subpage = dst_page;
5258 for (i = 0; i < pages_per_huge_page;
5259 i++, subpage = mem_map_next(subpage, dst_page, i)) {
5260 if (allow_pagefault)
5261 page_kaddr = kmap(subpage);
5263 page_kaddr = kmap_atomic(subpage);
5264 rc = copy_from_user(page_kaddr,
5265 (const void __user *)(src + i * PAGE_SIZE),
5267 if (allow_pagefault)
5270 kunmap_atomic(page_kaddr);
5272 ret_val -= (PAGE_SIZE - rc);
5280 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5282 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5284 static struct kmem_cache *page_ptl_cachep;
5286 void __init ptlock_cache_init(void)
5288 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5292 bool ptlock_alloc(struct page *page)
5296 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5303 void ptlock_free(struct page *page)
5305 kmem_cache_free(page_ptl_cachep, page->ptl);