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 core_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 *vma,
700 unsigned long addr, int *rss)
702 unsigned long vm_flags = 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 set_pte_at(dst_mm, addr, dst_pte, pte);
776 * Copy a present and normal page if necessary.
778 * NOTE! The usual case is that this doesn't need to do
779 * anything, and can just return a positive value. That
780 * will let the caller know that it can just increase
781 * the page refcount and re-use the pte the traditional
784 * But _if_ we need to copy it because it needs to be
785 * pinned in the parent (and the child should get its own
786 * copy rather than just a reference to the same page),
787 * we'll do that here and return zero to let the caller
790 * And if we need a pre-allocated page but don't yet have
791 * one, return a negative error to let the preallocation
792 * code know so that it can do so outside the page table
796 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
797 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
798 struct page **prealloc, pte_t pte, struct page *page)
800 struct mm_struct *src_mm = src_vma->vm_mm;
801 struct page *new_page;
803 if (!is_cow_mapping(src_vma->vm_flags))
807 * What we want to do is to check whether this page may
808 * have been pinned by the parent process. If so,
809 * instead of wrprotect the pte on both sides, we copy
810 * the page immediately so that we'll always guarantee
811 * the pinned page won't be randomly replaced in the
814 * The page pinning checks are just "has this mm ever
815 * seen pinning", along with the (inexact) check of
816 * the page count. That might give false positives for
817 * for pinning, but it will work correctly.
819 if (likely(!atomic_read(&src_mm->has_pinned)))
821 if (likely(!page_maybe_dma_pinned(page)))
824 new_page = *prealloc;
829 * We have a prealloc page, all good! Take it
830 * over and copy the page & arm it.
833 copy_user_highpage(new_page, page, addr, src_vma);
834 __SetPageUptodate(new_page);
835 page_add_new_anon_rmap(new_page, dst_vma, addr, false);
836 lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
837 rss[mm_counter(new_page)]++;
839 /* All done, just insert the new page copy in the child */
840 pte = mk_pte(new_page, dst_vma->vm_page_prot);
841 pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
842 set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
847 * Copy one pte. Returns 0 if succeeded, or -EAGAIN if one preallocated page
848 * is required to copy this pte.
851 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
852 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
853 struct page **prealloc)
855 struct mm_struct *src_mm = src_vma->vm_mm;
856 unsigned long vm_flags = src_vma->vm_flags;
857 pte_t pte = *src_pte;
860 page = vm_normal_page(src_vma, addr, pte);
864 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
865 addr, rss, prealloc, pte, page);
870 page_dup_rmap(page, false);
871 rss[mm_counter(page)]++;
875 * If it's a COW mapping, write protect it both
876 * in the parent and the child
878 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
879 ptep_set_wrprotect(src_mm, addr, src_pte);
880 pte = pte_wrprotect(pte);
884 * If it's a shared mapping, mark it clean in
887 if (vm_flags & VM_SHARED)
888 pte = pte_mkclean(pte);
889 pte = pte_mkold(pte);
892 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
893 * does not have the VM_UFFD_WP, which means that the uffd
894 * fork event is not enabled.
896 if (!(vm_flags & VM_UFFD_WP))
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,
977 /* copy_present_pte() will clear `*prealloc' if consumed */
978 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
979 addr, rss, &prealloc);
981 * If we need a pre-allocated page for this pte, drop the
982 * locks, allocate, and try again.
984 if (unlikely(ret == -EAGAIN))
986 if (unlikely(prealloc)) {
988 * pre-alloc page cannot be reused by next time so as
989 * to strictly follow mempolicy (e.g., alloc_page_vma()
990 * will allocate page according to address). This
991 * could only happen if one pinned pte changed.
997 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
999 arch_leave_lazy_mmu_mode();
1000 spin_unlock(src_ptl);
1001 pte_unmap(orig_src_pte);
1002 add_mm_rss_vec(dst_mm, rss);
1003 pte_unmap_unlock(orig_dst_pte, dst_ptl);
1007 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1013 WARN_ON_ONCE(ret != -EAGAIN);
1014 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1017 /* We've captured and resolved the error. Reset, try again. */
1023 if (unlikely(prealloc))
1029 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1030 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1033 struct mm_struct *dst_mm = dst_vma->vm_mm;
1034 struct mm_struct *src_mm = src_vma->vm_mm;
1035 pmd_t *src_pmd, *dst_pmd;
1038 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1041 src_pmd = pmd_offset(src_pud, addr);
1043 next = pmd_addr_end(addr, end);
1044 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1045 || pmd_devmap(*src_pmd)) {
1047 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1048 err = copy_huge_pmd(dst_mm, src_mm,
1049 dst_pmd, src_pmd, addr, src_vma);
1056 if (pmd_none_or_clear_bad(src_pmd))
1058 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1061 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1066 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1067 p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1070 struct mm_struct *dst_mm = dst_vma->vm_mm;
1071 struct mm_struct *src_mm = src_vma->vm_mm;
1072 pud_t *src_pud, *dst_pud;
1075 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1078 src_pud = pud_offset(src_p4d, addr);
1080 next = pud_addr_end(addr, end);
1081 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1084 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1085 err = copy_huge_pud(dst_mm, src_mm,
1086 dst_pud, src_pud, addr, src_vma);
1093 if (pud_none_or_clear_bad(src_pud))
1095 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1098 } while (dst_pud++, src_pud++, addr = next, addr != end);
1103 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1104 pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1107 struct mm_struct *dst_mm = dst_vma->vm_mm;
1108 p4d_t *src_p4d, *dst_p4d;
1111 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1114 src_p4d = p4d_offset(src_pgd, addr);
1116 next = p4d_addr_end(addr, end);
1117 if (p4d_none_or_clear_bad(src_p4d))
1119 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1122 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1127 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1129 pgd_t *src_pgd, *dst_pgd;
1131 unsigned long addr = src_vma->vm_start;
1132 unsigned long end = src_vma->vm_end;
1133 struct mm_struct *dst_mm = dst_vma->vm_mm;
1134 struct mm_struct *src_mm = src_vma->vm_mm;
1135 struct mmu_notifier_range range;
1140 * Don't copy ptes where a page fault will fill them correctly.
1141 * Fork becomes much lighter when there are big shared or private
1142 * readonly mappings. The tradeoff is that copy_page_range is more
1143 * efficient than faulting.
1145 if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1149 if (is_vm_hugetlb_page(src_vma))
1150 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1152 if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1154 * We do not free on error cases below as remove_vma
1155 * gets called on error from higher level routine
1157 ret = track_pfn_copy(src_vma);
1163 * We need to invalidate the secondary MMU mappings only when
1164 * there could be a permission downgrade on the ptes of the
1165 * parent mm. And a permission downgrade will only happen if
1166 * is_cow_mapping() returns true.
1168 is_cow = is_cow_mapping(src_vma->vm_flags);
1171 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1172 0, src_vma, src_mm, addr, end);
1173 mmu_notifier_invalidate_range_start(&range);
1177 dst_pgd = pgd_offset(dst_mm, addr);
1178 src_pgd = pgd_offset(src_mm, addr);
1180 next = pgd_addr_end(addr, end);
1181 if (pgd_none_or_clear_bad(src_pgd))
1183 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1188 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1191 mmu_notifier_invalidate_range_end(&range);
1195 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1196 struct vm_area_struct *vma, pmd_t *pmd,
1197 unsigned long addr, unsigned long end,
1198 struct zap_details *details)
1200 struct mm_struct *mm = tlb->mm;
1201 int force_flush = 0;
1202 int rss[NR_MM_COUNTERS];
1208 tlb_change_page_size(tlb, PAGE_SIZE);
1211 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1213 flush_tlb_batched_pending(mm);
1214 arch_enter_lazy_mmu_mode();
1217 if (pte_none(ptent))
1223 if (pte_present(ptent)) {
1226 page = vm_normal_page(vma, addr, ptent);
1227 if (unlikely(details) && page) {
1229 * unmap_shared_mapping_pages() wants to
1230 * invalidate cache without truncating:
1231 * unmap shared but keep private pages.
1233 if (details->check_mapping &&
1234 details->check_mapping != page_rmapping(page))
1237 ptent = ptep_get_and_clear_full(mm, addr, pte,
1239 tlb_remove_tlb_entry(tlb, pte, addr);
1240 if (unlikely(!page))
1243 if (!PageAnon(page)) {
1244 if (pte_dirty(ptent)) {
1246 set_page_dirty(page);
1248 if (pte_young(ptent) &&
1249 likely(!(vma->vm_flags & VM_SEQ_READ)))
1250 mark_page_accessed(page);
1252 rss[mm_counter(page)]--;
1253 page_remove_rmap(page, false);
1254 if (unlikely(page_mapcount(page) < 0))
1255 print_bad_pte(vma, addr, ptent, page);
1256 if (unlikely(__tlb_remove_page(tlb, page))) {
1264 entry = pte_to_swp_entry(ptent);
1265 if (is_device_private_entry(entry)) {
1266 struct page *page = device_private_entry_to_page(entry);
1268 if (unlikely(details && details->check_mapping)) {
1270 * unmap_shared_mapping_pages() wants to
1271 * invalidate cache without truncating:
1272 * unmap shared but keep private pages.
1274 if (details->check_mapping !=
1275 page_rmapping(page))
1279 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1280 rss[mm_counter(page)]--;
1281 page_remove_rmap(page, false);
1286 /* If details->check_mapping, we leave swap entries. */
1287 if (unlikely(details))
1290 if (!non_swap_entry(entry))
1292 else if (is_migration_entry(entry)) {
1295 page = migration_entry_to_page(entry);
1296 rss[mm_counter(page)]--;
1298 if (unlikely(!free_swap_and_cache(entry)))
1299 print_bad_pte(vma, addr, ptent, NULL);
1300 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1301 } while (pte++, addr += PAGE_SIZE, addr != end);
1303 add_mm_rss_vec(mm, rss);
1304 arch_leave_lazy_mmu_mode();
1306 /* Do the actual TLB flush before dropping ptl */
1308 tlb_flush_mmu_tlbonly(tlb);
1309 pte_unmap_unlock(start_pte, ptl);
1312 * If we forced a TLB flush (either due to running out of
1313 * batch buffers or because we needed to flush dirty TLB
1314 * entries before releasing the ptl), free the batched
1315 * memory too. Restart if we didn't do everything.
1330 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1331 struct vm_area_struct *vma, pud_t *pud,
1332 unsigned long addr, unsigned long end,
1333 struct zap_details *details)
1338 pmd = pmd_offset(pud, addr);
1340 next = pmd_addr_end(addr, end);
1341 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1342 if (next - addr != HPAGE_PMD_SIZE)
1343 __split_huge_pmd(vma, pmd, addr, false, NULL);
1344 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1349 * Here there can be other concurrent MADV_DONTNEED or
1350 * trans huge page faults running, and if the pmd is
1351 * none or trans huge it can change under us. This is
1352 * because MADV_DONTNEED holds the mmap_lock in read
1355 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1357 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1360 } while (pmd++, addr = next, addr != end);
1365 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1366 struct vm_area_struct *vma, p4d_t *p4d,
1367 unsigned long addr, unsigned long end,
1368 struct zap_details *details)
1373 pud = pud_offset(p4d, addr);
1375 next = pud_addr_end(addr, end);
1376 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1377 if (next - addr != HPAGE_PUD_SIZE) {
1378 mmap_assert_locked(tlb->mm);
1379 split_huge_pud(vma, pud, addr);
1380 } else if (zap_huge_pud(tlb, vma, pud, addr))
1384 if (pud_none_or_clear_bad(pud))
1386 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1389 } while (pud++, addr = next, addr != end);
1394 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1395 struct vm_area_struct *vma, pgd_t *pgd,
1396 unsigned long addr, unsigned long end,
1397 struct zap_details *details)
1402 p4d = p4d_offset(pgd, addr);
1404 next = p4d_addr_end(addr, end);
1405 if (p4d_none_or_clear_bad(p4d))
1407 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1408 } while (p4d++, addr = next, addr != end);
1413 void unmap_page_range(struct mmu_gather *tlb,
1414 struct vm_area_struct *vma,
1415 unsigned long addr, unsigned long end,
1416 struct zap_details *details)
1421 BUG_ON(addr >= end);
1422 tlb_start_vma(tlb, vma);
1423 pgd = pgd_offset(vma->vm_mm, addr);
1425 next = pgd_addr_end(addr, end);
1426 if (pgd_none_or_clear_bad(pgd))
1428 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1429 } while (pgd++, addr = next, addr != end);
1430 tlb_end_vma(tlb, vma);
1434 static void unmap_single_vma(struct mmu_gather *tlb,
1435 struct vm_area_struct *vma, unsigned long start_addr,
1436 unsigned long end_addr,
1437 struct zap_details *details)
1439 unsigned long start = max(vma->vm_start, start_addr);
1442 if (start >= vma->vm_end)
1444 end = min(vma->vm_end, end_addr);
1445 if (end <= vma->vm_start)
1449 uprobe_munmap(vma, start, end);
1451 if (unlikely(vma->vm_flags & VM_PFNMAP))
1452 untrack_pfn(vma, 0, 0);
1455 if (unlikely(is_vm_hugetlb_page(vma))) {
1457 * It is undesirable to test vma->vm_file as it
1458 * should be non-null for valid hugetlb area.
1459 * However, vm_file will be NULL in the error
1460 * cleanup path of mmap_region. When
1461 * hugetlbfs ->mmap method fails,
1462 * mmap_region() nullifies vma->vm_file
1463 * before calling this function to clean up.
1464 * Since no pte has actually been setup, it is
1465 * safe to do nothing in this case.
1468 i_mmap_lock_write(vma->vm_file->f_mapping);
1469 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1470 i_mmap_unlock_write(vma->vm_file->f_mapping);
1473 unmap_page_range(tlb, vma, start, end, details);
1478 * unmap_vmas - unmap a range of memory covered by a list of vma's
1479 * @tlb: address of the caller's struct mmu_gather
1480 * @vma: the starting vma
1481 * @start_addr: virtual address at which to start unmapping
1482 * @end_addr: virtual address at which to end unmapping
1484 * Unmap all pages in the vma list.
1486 * Only addresses between `start' and `end' will be unmapped.
1488 * The VMA list must be sorted in ascending virtual address order.
1490 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1491 * range after unmap_vmas() returns. So the only responsibility here is to
1492 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1493 * drops the lock and schedules.
1495 void unmap_vmas(struct mmu_gather *tlb,
1496 struct vm_area_struct *vma, unsigned long start_addr,
1497 unsigned long end_addr)
1499 struct mmu_notifier_range range;
1501 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1502 start_addr, end_addr);
1503 mmu_notifier_invalidate_range_start(&range);
1504 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1505 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1506 mmu_notifier_invalidate_range_end(&range);
1510 * zap_page_range - remove user pages in a given range
1511 * @vma: vm_area_struct holding the applicable pages
1512 * @start: starting address of pages to zap
1513 * @size: number of bytes to zap
1515 * Caller must protect the VMA list
1517 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1520 struct mmu_notifier_range range;
1521 struct mmu_gather tlb;
1524 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1525 start, start + size);
1526 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1527 update_hiwater_rss(vma->vm_mm);
1528 mmu_notifier_invalidate_range_start(&range);
1529 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1530 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1531 mmu_notifier_invalidate_range_end(&range);
1532 tlb_finish_mmu(&tlb, start, range.end);
1536 * zap_page_range_single - remove user pages in a given range
1537 * @vma: vm_area_struct holding the applicable pages
1538 * @address: starting address of pages to zap
1539 * @size: number of bytes to zap
1540 * @details: details of shared cache invalidation
1542 * The range must fit into one VMA.
1544 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1545 unsigned long size, struct zap_details *details)
1547 struct mmu_notifier_range range;
1548 struct mmu_gather tlb;
1551 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1552 address, address + size);
1553 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1554 update_hiwater_rss(vma->vm_mm);
1555 mmu_notifier_invalidate_range_start(&range);
1556 unmap_single_vma(&tlb, vma, address, range.end, details);
1557 mmu_notifier_invalidate_range_end(&range);
1558 tlb_finish_mmu(&tlb, address, range.end);
1562 * zap_vma_ptes - remove ptes mapping the vma
1563 * @vma: vm_area_struct holding ptes to be zapped
1564 * @address: starting address of pages to zap
1565 * @size: number of bytes to zap
1567 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1569 * The entire address range must be fully contained within the vma.
1572 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1575 if (address < vma->vm_start || address + size > vma->vm_end ||
1576 !(vma->vm_flags & VM_PFNMAP))
1579 zap_page_range_single(vma, address, size, NULL);
1581 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1583 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1590 pgd = pgd_offset(mm, addr);
1591 p4d = p4d_alloc(mm, pgd, addr);
1594 pud = pud_alloc(mm, p4d, addr);
1597 pmd = pmd_alloc(mm, pud, addr);
1601 VM_BUG_ON(pmd_trans_huge(*pmd));
1605 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1608 pmd_t *pmd = walk_to_pmd(mm, addr);
1612 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1615 static int validate_page_before_insert(struct page *page)
1617 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1619 flush_dcache_page(page);
1623 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1624 unsigned long addr, struct page *page, pgprot_t prot)
1626 if (!pte_none(*pte))
1628 /* Ok, finally just insert the thing.. */
1630 inc_mm_counter_fast(mm, mm_counter_file(page));
1631 page_add_file_rmap(page, false);
1632 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1637 * This is the old fallback for page remapping.
1639 * For historical reasons, it only allows reserved pages. Only
1640 * old drivers should use this, and they needed to mark their
1641 * pages reserved for the old functions anyway.
1643 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1644 struct page *page, pgprot_t prot)
1646 struct mm_struct *mm = vma->vm_mm;
1651 retval = validate_page_before_insert(page);
1655 pte = get_locked_pte(mm, addr, &ptl);
1658 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1659 pte_unmap_unlock(pte, ptl);
1665 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1666 unsigned long addr, struct page *page, pgprot_t prot)
1670 if (!page_count(page))
1672 err = validate_page_before_insert(page);
1675 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1678 /* insert_pages() amortizes the cost of spinlock operations
1679 * when inserting pages in a loop. Arch *must* define pte_index.
1681 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1682 struct page **pages, unsigned long *num, pgprot_t prot)
1685 pte_t *start_pte, *pte;
1686 spinlock_t *pte_lock;
1687 struct mm_struct *const mm = vma->vm_mm;
1688 unsigned long curr_page_idx = 0;
1689 unsigned long remaining_pages_total = *num;
1690 unsigned long pages_to_write_in_pmd;
1694 pmd = walk_to_pmd(mm, addr);
1698 pages_to_write_in_pmd = min_t(unsigned long,
1699 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1701 /* Allocate the PTE if necessary; takes PMD lock once only. */
1703 if (pte_alloc(mm, pmd))
1706 while (pages_to_write_in_pmd) {
1708 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1710 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1711 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1712 int err = insert_page_in_batch_locked(mm, pte,
1713 addr, pages[curr_page_idx], prot);
1714 if (unlikely(err)) {
1715 pte_unmap_unlock(start_pte, pte_lock);
1717 remaining_pages_total -= pte_idx;
1723 pte_unmap_unlock(start_pte, pte_lock);
1724 pages_to_write_in_pmd -= batch_size;
1725 remaining_pages_total -= batch_size;
1727 if (remaining_pages_total)
1731 *num = remaining_pages_total;
1734 #endif /* ifdef pte_index */
1737 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1738 * @vma: user vma to map to
1739 * @addr: target start user address of these pages
1740 * @pages: source kernel pages
1741 * @num: in: number of pages to map. out: number of pages that were *not*
1742 * mapped. (0 means all pages were successfully mapped).
1744 * Preferred over vm_insert_page() when inserting multiple pages.
1746 * In case of error, we may have mapped a subset of the provided
1747 * pages. It is the caller's responsibility to account for this case.
1749 * The same restrictions apply as in vm_insert_page().
1751 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1752 struct page **pages, unsigned long *num)
1755 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1757 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1759 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1760 BUG_ON(mmap_read_trylock(vma->vm_mm));
1761 BUG_ON(vma->vm_flags & VM_PFNMAP);
1762 vma->vm_flags |= VM_MIXEDMAP;
1764 /* Defer page refcount checking till we're about to map that page. */
1765 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1767 unsigned long idx = 0, pgcount = *num;
1770 for (; idx < pgcount; ++idx) {
1771 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1775 *num = pgcount - idx;
1777 #endif /* ifdef pte_index */
1779 EXPORT_SYMBOL(vm_insert_pages);
1782 * vm_insert_page - insert single page into user vma
1783 * @vma: user vma to map to
1784 * @addr: target user address of this page
1785 * @page: source kernel page
1787 * This allows drivers to insert individual pages they've allocated
1790 * The page has to be a nice clean _individual_ kernel allocation.
1791 * If you allocate a compound page, you need to have marked it as
1792 * such (__GFP_COMP), or manually just split the page up yourself
1793 * (see split_page()).
1795 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1796 * took an arbitrary page protection parameter. This doesn't allow
1797 * that. Your vma protection will have to be set up correctly, which
1798 * means that if you want a shared writable mapping, you'd better
1799 * ask for a shared writable mapping!
1801 * The page does not need to be reserved.
1803 * Usually this function is called from f_op->mmap() handler
1804 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1805 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1806 * function from other places, for example from page-fault handler.
1808 * Return: %0 on success, negative error code otherwise.
1810 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1813 if (addr < vma->vm_start || addr >= vma->vm_end)
1815 if (!page_count(page))
1817 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1818 BUG_ON(mmap_read_trylock(vma->vm_mm));
1819 BUG_ON(vma->vm_flags & VM_PFNMAP);
1820 vma->vm_flags |= VM_MIXEDMAP;
1822 return insert_page(vma, addr, page, vma->vm_page_prot);
1824 EXPORT_SYMBOL(vm_insert_page);
1827 * __vm_map_pages - maps range of kernel pages into user vma
1828 * @vma: user vma to map to
1829 * @pages: pointer to array of source kernel pages
1830 * @num: number of pages in page array
1831 * @offset: user's requested vm_pgoff
1833 * This allows drivers to map range of kernel pages into a user vma.
1835 * Return: 0 on success and error code otherwise.
1837 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1838 unsigned long num, unsigned long offset)
1840 unsigned long count = vma_pages(vma);
1841 unsigned long uaddr = vma->vm_start;
1844 /* Fail if the user requested offset is beyond the end of the object */
1848 /* Fail if the user requested size exceeds available object size */
1849 if (count > num - offset)
1852 for (i = 0; i < count; i++) {
1853 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1863 * vm_map_pages - maps range of kernel pages starts with non zero offset
1864 * @vma: user vma to map to
1865 * @pages: pointer to array of source kernel pages
1866 * @num: number of pages in page array
1868 * Maps an object consisting of @num pages, catering for the user's
1869 * requested vm_pgoff
1871 * If we fail to insert any page into the vma, the function will return
1872 * immediately leaving any previously inserted pages present. Callers
1873 * from the mmap handler may immediately return the error as their caller
1874 * will destroy the vma, removing any successfully inserted pages. Other
1875 * callers should make their own arrangements for calling unmap_region().
1877 * Context: Process context. Called by mmap handlers.
1878 * Return: 0 on success and error code otherwise.
1880 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1883 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1885 EXPORT_SYMBOL(vm_map_pages);
1888 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1889 * @vma: user vma to map to
1890 * @pages: pointer to array of source kernel pages
1891 * @num: number of pages in page array
1893 * Similar to vm_map_pages(), except that it explicitly sets the offset
1894 * to 0. This function is intended for the drivers that did not consider
1897 * Context: Process context. Called by mmap handlers.
1898 * Return: 0 on success and error code otherwise.
1900 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1903 return __vm_map_pages(vma, pages, num, 0);
1905 EXPORT_SYMBOL(vm_map_pages_zero);
1907 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1908 pfn_t pfn, pgprot_t prot, bool mkwrite)
1910 struct mm_struct *mm = vma->vm_mm;
1914 pte = get_locked_pte(mm, addr, &ptl);
1916 return VM_FAULT_OOM;
1917 if (!pte_none(*pte)) {
1920 * For read faults on private mappings the PFN passed
1921 * in may not match the PFN we have mapped if the
1922 * mapped PFN is a writeable COW page. In the mkwrite
1923 * case we are creating a writable PTE for a shared
1924 * mapping and we expect the PFNs to match. If they
1925 * don't match, we are likely racing with block
1926 * allocation and mapping invalidation so just skip the
1929 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1930 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1933 entry = pte_mkyoung(*pte);
1934 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1935 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1936 update_mmu_cache(vma, addr, pte);
1941 /* Ok, finally just insert the thing.. */
1942 if (pfn_t_devmap(pfn))
1943 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1945 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1948 entry = pte_mkyoung(entry);
1949 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1952 set_pte_at(mm, addr, pte, entry);
1953 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1956 pte_unmap_unlock(pte, ptl);
1957 return VM_FAULT_NOPAGE;
1961 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1962 * @vma: user vma to map to
1963 * @addr: target user address of this page
1964 * @pfn: source kernel pfn
1965 * @pgprot: pgprot flags for the inserted page
1967 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1968 * to override pgprot on a per-page basis.
1970 * This only makes sense for IO mappings, and it makes no sense for
1971 * COW mappings. In general, using multiple vmas is preferable;
1972 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1975 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1976 * a value of @pgprot different from that of @vma->vm_page_prot.
1978 * Context: Process context. May allocate using %GFP_KERNEL.
1979 * Return: vm_fault_t value.
1981 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1982 unsigned long pfn, pgprot_t pgprot)
1985 * Technically, architectures with pte_special can avoid all these
1986 * restrictions (same for remap_pfn_range). However we would like
1987 * consistency in testing and feature parity among all, so we should
1988 * try to keep these invariants in place for everybody.
1990 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1991 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1992 (VM_PFNMAP|VM_MIXEDMAP));
1993 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1994 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1996 if (addr < vma->vm_start || addr >= vma->vm_end)
1997 return VM_FAULT_SIGBUS;
1999 if (!pfn_modify_allowed(pfn, pgprot))
2000 return VM_FAULT_SIGBUS;
2002 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2004 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2007 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2010 * vmf_insert_pfn - insert single pfn into user vma
2011 * @vma: user vma to map to
2012 * @addr: target user address of this page
2013 * @pfn: source kernel pfn
2015 * Similar to vm_insert_page, this allows drivers to insert individual pages
2016 * they've allocated into a user vma. Same comments apply.
2018 * This function should only be called from a vm_ops->fault handler, and
2019 * in that case the handler should return the result of this function.
2021 * vma cannot be a COW mapping.
2023 * As this is called only for pages that do not currently exist, we
2024 * do not need to flush old virtual caches or the TLB.
2026 * Context: Process context. May allocate using %GFP_KERNEL.
2027 * Return: vm_fault_t value.
2029 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2032 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2034 EXPORT_SYMBOL(vmf_insert_pfn);
2036 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2038 /* these checks mirror the abort conditions in vm_normal_page */
2039 if (vma->vm_flags & VM_MIXEDMAP)
2041 if (pfn_t_devmap(pfn))
2043 if (pfn_t_special(pfn))
2045 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2050 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2051 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2056 BUG_ON(!vm_mixed_ok(vma, pfn));
2058 if (addr < vma->vm_start || addr >= vma->vm_end)
2059 return VM_FAULT_SIGBUS;
2061 track_pfn_insert(vma, &pgprot, pfn);
2063 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2064 return VM_FAULT_SIGBUS;
2067 * If we don't have pte special, then we have to use the pfn_valid()
2068 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2069 * refcount the page if pfn_valid is true (hence insert_page rather
2070 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2071 * without pte special, it would there be refcounted as a normal page.
2073 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2074 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2078 * At this point we are committed to insert_page()
2079 * regardless of whether the caller specified flags that
2080 * result in pfn_t_has_page() == false.
2082 page = pfn_to_page(pfn_t_to_pfn(pfn));
2083 err = insert_page(vma, addr, page, pgprot);
2085 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2089 return VM_FAULT_OOM;
2090 if (err < 0 && err != -EBUSY)
2091 return VM_FAULT_SIGBUS;
2093 return VM_FAULT_NOPAGE;
2097 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2098 * @vma: user vma to map to
2099 * @addr: target user address of this page
2100 * @pfn: source kernel pfn
2101 * @pgprot: pgprot flags for the inserted page
2103 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2104 * to override pgprot on a per-page basis.
2106 * Typically this function should be used by drivers to set caching- and
2107 * encryption bits different than those of @vma->vm_page_prot, because
2108 * the caching- or encryption mode may not be known at mmap() time.
2109 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2110 * to set caching and encryption bits for those vmas (except for COW pages).
2111 * This is ensured by core vm only modifying these page table entries using
2112 * functions that don't touch caching- or encryption bits, using pte_modify()
2113 * if needed. (See for example mprotect()).
2114 * Also when new page-table entries are created, this is only done using the
2115 * fault() callback, and never using the value of vma->vm_page_prot,
2116 * except for page-table entries that point to anonymous pages as the result
2119 * Context: Process context. May allocate using %GFP_KERNEL.
2120 * Return: vm_fault_t value.
2122 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2123 pfn_t pfn, pgprot_t pgprot)
2125 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2127 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2129 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2132 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2134 EXPORT_SYMBOL(vmf_insert_mixed);
2137 * If the insertion of PTE failed because someone else already added a
2138 * different entry in the mean time, we treat that as success as we assume
2139 * the same entry was actually inserted.
2141 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2142 unsigned long addr, pfn_t pfn)
2144 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2146 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2149 * maps a range of physical memory into the requested pages. the old
2150 * mappings are removed. any references to nonexistent pages results
2151 * in null mappings (currently treated as "copy-on-access")
2153 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2154 unsigned long addr, unsigned long end,
2155 unsigned long pfn, pgprot_t prot)
2161 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2164 arch_enter_lazy_mmu_mode();
2166 BUG_ON(!pte_none(*pte));
2167 if (!pfn_modify_allowed(pfn, prot)) {
2171 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2173 } while (pte++, addr += PAGE_SIZE, addr != end);
2174 arch_leave_lazy_mmu_mode();
2175 pte_unmap_unlock(pte - 1, ptl);
2179 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2180 unsigned long addr, unsigned long end,
2181 unsigned long pfn, pgprot_t prot)
2187 pfn -= addr >> PAGE_SHIFT;
2188 pmd = pmd_alloc(mm, pud, addr);
2191 VM_BUG_ON(pmd_trans_huge(*pmd));
2193 next = pmd_addr_end(addr, end);
2194 err = remap_pte_range(mm, pmd, addr, next,
2195 pfn + (addr >> PAGE_SHIFT), prot);
2198 } while (pmd++, addr = next, addr != end);
2202 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2203 unsigned long addr, unsigned long end,
2204 unsigned long pfn, pgprot_t prot)
2210 pfn -= addr >> PAGE_SHIFT;
2211 pud = pud_alloc(mm, p4d, addr);
2215 next = pud_addr_end(addr, end);
2216 err = remap_pmd_range(mm, pud, addr, next,
2217 pfn + (addr >> PAGE_SHIFT), prot);
2220 } while (pud++, addr = next, addr != end);
2224 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2225 unsigned long addr, unsigned long end,
2226 unsigned long pfn, pgprot_t prot)
2232 pfn -= addr >> PAGE_SHIFT;
2233 p4d = p4d_alloc(mm, pgd, addr);
2237 next = p4d_addr_end(addr, end);
2238 err = remap_pud_range(mm, p4d, addr, next,
2239 pfn + (addr >> PAGE_SHIFT), prot);
2242 } while (p4d++, addr = next, addr != end);
2247 * remap_pfn_range - remap kernel memory to userspace
2248 * @vma: user vma to map to
2249 * @addr: target page aligned user address to start at
2250 * @pfn: page frame number of kernel physical memory address
2251 * @size: size of mapping area
2252 * @prot: page protection flags for this mapping
2254 * Note: this is only safe if the mm semaphore is held when called.
2256 * Return: %0 on success, negative error code otherwise.
2258 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2259 unsigned long pfn, unsigned long size, pgprot_t prot)
2263 unsigned long end = addr + PAGE_ALIGN(size);
2264 struct mm_struct *mm = vma->vm_mm;
2265 unsigned long remap_pfn = pfn;
2268 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2272 * Physically remapped pages are special. Tell the
2273 * rest of the world about it:
2274 * VM_IO tells people not to look at these pages
2275 * (accesses can have side effects).
2276 * VM_PFNMAP tells the core MM that the base pages are just
2277 * raw PFN mappings, and do not have a "struct page" associated
2280 * Disable vma merging and expanding with mremap().
2282 * Omit vma from core dump, even when VM_IO turned off.
2284 * There's a horrible special case to handle copy-on-write
2285 * behaviour that some programs depend on. We mark the "original"
2286 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2287 * See vm_normal_page() for details.
2289 if (is_cow_mapping(vma->vm_flags)) {
2290 if (addr != vma->vm_start || end != vma->vm_end)
2292 vma->vm_pgoff = pfn;
2295 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2299 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2301 BUG_ON(addr >= end);
2302 pfn -= addr >> PAGE_SHIFT;
2303 pgd = pgd_offset(mm, addr);
2304 flush_cache_range(vma, addr, end);
2306 next = pgd_addr_end(addr, end);
2307 err = remap_p4d_range(mm, pgd, addr, next,
2308 pfn + (addr >> PAGE_SHIFT), prot);
2311 } while (pgd++, addr = next, addr != end);
2314 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2318 EXPORT_SYMBOL(remap_pfn_range);
2321 * vm_iomap_memory - remap memory to userspace
2322 * @vma: user vma to map to
2323 * @start: start of the physical memory to be mapped
2324 * @len: size of area
2326 * This is a simplified io_remap_pfn_range() for common driver use. The
2327 * driver just needs to give us the physical memory range to be mapped,
2328 * we'll figure out the rest from the vma information.
2330 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2331 * whatever write-combining details or similar.
2333 * Return: %0 on success, negative error code otherwise.
2335 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2337 unsigned long vm_len, pfn, pages;
2339 /* Check that the physical memory area passed in looks valid */
2340 if (start + len < start)
2343 * You *really* shouldn't map things that aren't page-aligned,
2344 * but we've historically allowed it because IO memory might
2345 * just have smaller alignment.
2347 len += start & ~PAGE_MASK;
2348 pfn = start >> PAGE_SHIFT;
2349 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2350 if (pfn + pages < pfn)
2353 /* We start the mapping 'vm_pgoff' pages into the area */
2354 if (vma->vm_pgoff > pages)
2356 pfn += vma->vm_pgoff;
2357 pages -= vma->vm_pgoff;
2359 /* Can we fit all of the mapping? */
2360 vm_len = vma->vm_end - vma->vm_start;
2361 if (vm_len >> PAGE_SHIFT > pages)
2364 /* Ok, let it rip */
2365 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2367 EXPORT_SYMBOL(vm_iomap_memory);
2369 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2370 unsigned long addr, unsigned long end,
2371 pte_fn_t fn, void *data, bool create,
2372 pgtbl_mod_mask *mask)
2379 pte = (mm == &init_mm) ?
2380 pte_alloc_kernel_track(pmd, addr, mask) :
2381 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2385 pte = (mm == &init_mm) ?
2386 pte_offset_kernel(pmd, addr) :
2387 pte_offset_map_lock(mm, pmd, addr, &ptl);
2390 BUG_ON(pmd_huge(*pmd));
2392 arch_enter_lazy_mmu_mode();
2395 if (create || !pte_none(*pte)) {
2396 err = fn(pte++, addr, data);
2400 } while (addr += PAGE_SIZE, addr != end);
2401 *mask |= PGTBL_PTE_MODIFIED;
2403 arch_leave_lazy_mmu_mode();
2406 pte_unmap_unlock(pte-1, ptl);
2410 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2411 unsigned long addr, unsigned long end,
2412 pte_fn_t fn, void *data, bool create,
2413 pgtbl_mod_mask *mask)
2419 BUG_ON(pud_huge(*pud));
2422 pmd = pmd_alloc_track(mm, pud, addr, mask);
2426 pmd = pmd_offset(pud, addr);
2429 next = pmd_addr_end(addr, end);
2430 if (create || !pmd_none_or_clear_bad(pmd)) {
2431 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2436 } while (pmd++, addr = next, addr != end);
2440 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2441 unsigned long addr, unsigned long end,
2442 pte_fn_t fn, void *data, bool create,
2443 pgtbl_mod_mask *mask)
2450 pud = pud_alloc_track(mm, p4d, addr, mask);
2454 pud = pud_offset(p4d, addr);
2457 next = pud_addr_end(addr, end);
2458 if (create || !pud_none_or_clear_bad(pud)) {
2459 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2464 } while (pud++, addr = next, addr != end);
2468 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2469 unsigned long addr, unsigned long end,
2470 pte_fn_t fn, void *data, bool create,
2471 pgtbl_mod_mask *mask)
2478 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2482 p4d = p4d_offset(pgd, addr);
2485 next = p4d_addr_end(addr, end);
2486 if (create || !p4d_none_or_clear_bad(p4d)) {
2487 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2492 } while (p4d++, addr = next, addr != end);
2496 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2497 unsigned long size, pte_fn_t fn,
2498 void *data, bool create)
2501 unsigned long start = addr, next;
2502 unsigned long end = addr + size;
2503 pgtbl_mod_mask mask = 0;
2506 if (WARN_ON(addr >= end))
2509 pgd = pgd_offset(mm, addr);
2511 next = pgd_addr_end(addr, end);
2512 if (!create && pgd_none_or_clear_bad(pgd))
2514 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2517 } while (pgd++, addr = next, addr != end);
2519 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2520 arch_sync_kernel_mappings(start, start + size);
2526 * Scan a region of virtual memory, filling in page tables as necessary
2527 * and calling a provided function on each leaf page table.
2529 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2530 unsigned long size, pte_fn_t fn, void *data)
2532 return __apply_to_page_range(mm, addr, size, fn, data, true);
2534 EXPORT_SYMBOL_GPL(apply_to_page_range);
2537 * Scan a region of virtual memory, calling a provided function on
2538 * each leaf page table where it exists.
2540 * Unlike apply_to_page_range, this does _not_ fill in page tables
2541 * where they are absent.
2543 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2544 unsigned long size, pte_fn_t fn, void *data)
2546 return __apply_to_page_range(mm, addr, size, fn, data, false);
2548 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2551 * handle_pte_fault chooses page fault handler according to an entry which was
2552 * read non-atomically. Before making any commitment, on those architectures
2553 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2554 * parts, do_swap_page must check under lock before unmapping the pte and
2555 * proceeding (but do_wp_page is only called after already making such a check;
2556 * and do_anonymous_page can safely check later on).
2558 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2559 pte_t *page_table, pte_t orig_pte)
2562 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2563 if (sizeof(pte_t) > sizeof(unsigned long)) {
2564 spinlock_t *ptl = pte_lockptr(mm, pmd);
2566 same = pte_same(*page_table, orig_pte);
2570 pte_unmap(page_table);
2574 static inline bool cow_user_page(struct page *dst, struct page *src,
2575 struct vm_fault *vmf)
2580 bool locked = false;
2581 struct vm_area_struct *vma = vmf->vma;
2582 struct mm_struct *mm = vma->vm_mm;
2583 unsigned long addr = vmf->address;
2586 copy_user_highpage(dst, src, addr, vma);
2591 * If the source page was a PFN mapping, we don't have
2592 * a "struct page" for it. We do a best-effort copy by
2593 * just copying from the original user address. If that
2594 * fails, we just zero-fill it. Live with it.
2596 kaddr = kmap_atomic(dst);
2597 uaddr = (void __user *)(addr & PAGE_MASK);
2600 * On architectures with software "accessed" bits, we would
2601 * take a double page fault, so mark it accessed here.
2603 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2606 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2608 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2610 * Other thread has already handled the fault
2611 * and update local tlb only
2613 update_mmu_tlb(vma, addr, vmf->pte);
2618 entry = pte_mkyoung(vmf->orig_pte);
2619 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2620 update_mmu_cache(vma, addr, vmf->pte);
2624 * This really shouldn't fail, because the page is there
2625 * in the page tables. But it might just be unreadable,
2626 * in which case we just give up and fill the result with
2629 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2633 /* Re-validate under PTL if the page is still mapped */
2634 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2636 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2637 /* The PTE changed under us, update local tlb */
2638 update_mmu_tlb(vma, addr, vmf->pte);
2644 * The same page can be mapped back since last copy attempt.
2645 * Try to copy again under PTL.
2647 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2649 * Give a warn in case there can be some obscure
2662 pte_unmap_unlock(vmf->pte, vmf->ptl);
2663 kunmap_atomic(kaddr);
2664 flush_dcache_page(dst);
2669 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2671 struct file *vm_file = vma->vm_file;
2674 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2677 * Special mappings (e.g. VDSO) do not have any file so fake
2678 * a default GFP_KERNEL for them.
2684 * Notify the address space that the page is about to become writable so that
2685 * it can prohibit this or wait for the page to get into an appropriate state.
2687 * We do this without the lock held, so that it can sleep if it needs to.
2689 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2692 struct page *page = vmf->page;
2693 unsigned int old_flags = vmf->flags;
2695 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2697 if (vmf->vma->vm_file &&
2698 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2699 return VM_FAULT_SIGBUS;
2701 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2702 /* Restore original flags so that caller is not surprised */
2703 vmf->flags = old_flags;
2704 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2706 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2708 if (!page->mapping) {
2710 return 0; /* retry */
2712 ret |= VM_FAULT_LOCKED;
2714 VM_BUG_ON_PAGE(!PageLocked(page), page);
2719 * Handle dirtying of a page in shared file mapping on a write fault.
2721 * The function expects the page to be locked and unlocks it.
2723 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2725 struct vm_area_struct *vma = vmf->vma;
2726 struct address_space *mapping;
2727 struct page *page = vmf->page;
2729 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2731 dirtied = set_page_dirty(page);
2732 VM_BUG_ON_PAGE(PageAnon(page), page);
2734 * Take a local copy of the address_space - page.mapping may be zeroed
2735 * by truncate after unlock_page(). The address_space itself remains
2736 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2737 * release semantics to prevent the compiler from undoing this copying.
2739 mapping = page_rmapping(page);
2743 file_update_time(vma->vm_file);
2746 * Throttle page dirtying rate down to writeback speed.
2748 * mapping may be NULL here because some device drivers do not
2749 * set page.mapping but still dirty their pages
2751 * Drop the mmap_lock before waiting on IO, if we can. The file
2752 * is pinning the mapping, as per above.
2754 if ((dirtied || page_mkwrite) && mapping) {
2757 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2758 balance_dirty_pages_ratelimited(mapping);
2761 return VM_FAULT_RETRY;
2769 * Handle write page faults for pages that can be reused in the current vma
2771 * This can happen either due to the mapping being with the VM_SHARED flag,
2772 * or due to us being the last reference standing to the page. In either
2773 * case, all we need to do here is to mark the page as writable and update
2774 * any related book-keeping.
2776 static inline void wp_page_reuse(struct vm_fault *vmf)
2777 __releases(vmf->ptl)
2779 struct vm_area_struct *vma = vmf->vma;
2780 struct page *page = vmf->page;
2783 * Clear the pages cpupid information as the existing
2784 * information potentially belongs to a now completely
2785 * unrelated process.
2788 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2790 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2791 entry = pte_mkyoung(vmf->orig_pte);
2792 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2793 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2794 update_mmu_cache(vma, vmf->address, vmf->pte);
2795 pte_unmap_unlock(vmf->pte, vmf->ptl);
2796 count_vm_event(PGREUSE);
2800 * Handle the case of a page which we actually need to copy to a new page.
2802 * Called with mmap_lock locked and the old page referenced, but
2803 * without the ptl held.
2805 * High level logic flow:
2807 * - Allocate a page, copy the content of the old page to the new one.
2808 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2809 * - Take the PTL. If the pte changed, bail out and release the allocated page
2810 * - If the pte is still the way we remember it, update the page table and all
2811 * relevant references. This includes dropping the reference the page-table
2812 * held to the old page, as well as updating the rmap.
2813 * - In any case, unlock the PTL and drop the reference we took to the old page.
2815 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2817 struct vm_area_struct *vma = vmf->vma;
2818 struct mm_struct *mm = vma->vm_mm;
2819 struct page *old_page = vmf->page;
2820 struct page *new_page = NULL;
2822 int page_copied = 0;
2823 struct mmu_notifier_range range;
2825 if (unlikely(anon_vma_prepare(vma)))
2828 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2829 new_page = alloc_zeroed_user_highpage_movable(vma,
2834 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2839 if (!cow_user_page(new_page, old_page, vmf)) {
2841 * COW failed, if the fault was solved by other,
2842 * it's fine. If not, userspace would re-fault on
2843 * the same address and we will handle the fault
2844 * from the second attempt.
2853 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2855 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2857 __SetPageUptodate(new_page);
2859 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2860 vmf->address & PAGE_MASK,
2861 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2862 mmu_notifier_invalidate_range_start(&range);
2865 * Re-check the pte - we dropped the lock
2867 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2868 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2870 if (!PageAnon(old_page)) {
2871 dec_mm_counter_fast(mm,
2872 mm_counter_file(old_page));
2873 inc_mm_counter_fast(mm, MM_ANONPAGES);
2876 inc_mm_counter_fast(mm, MM_ANONPAGES);
2878 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2879 entry = mk_pte(new_page, vma->vm_page_prot);
2880 entry = pte_sw_mkyoung(entry);
2881 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2883 * Clear the pte entry and flush it first, before updating the
2884 * pte with the new entry. This will avoid a race condition
2885 * seen in the presence of one thread doing SMC and another
2888 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2889 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2890 lru_cache_add_inactive_or_unevictable(new_page, vma);
2892 * We call the notify macro here because, when using secondary
2893 * mmu page tables (such as kvm shadow page tables), we want the
2894 * new page to be mapped directly into the secondary page table.
2896 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2897 update_mmu_cache(vma, vmf->address, vmf->pte);
2900 * Only after switching the pte to the new page may
2901 * we remove the mapcount here. Otherwise another
2902 * process may come and find the rmap count decremented
2903 * before the pte is switched to the new page, and
2904 * "reuse" the old page writing into it while our pte
2905 * here still points into it and can be read by other
2908 * The critical issue is to order this
2909 * page_remove_rmap with the ptp_clear_flush above.
2910 * Those stores are ordered by (if nothing else,)
2911 * the barrier present in the atomic_add_negative
2912 * in page_remove_rmap.
2914 * Then the TLB flush in ptep_clear_flush ensures that
2915 * no process can access the old page before the
2916 * decremented mapcount is visible. And the old page
2917 * cannot be reused until after the decremented
2918 * mapcount is visible. So transitively, TLBs to
2919 * old page will be flushed before it can be reused.
2921 page_remove_rmap(old_page, false);
2924 /* Free the old page.. */
2925 new_page = old_page;
2928 update_mmu_tlb(vma, vmf->address, vmf->pte);
2934 pte_unmap_unlock(vmf->pte, vmf->ptl);
2936 * No need to double call mmu_notifier->invalidate_range() callback as
2937 * the above ptep_clear_flush_notify() did already call it.
2939 mmu_notifier_invalidate_range_only_end(&range);
2942 * Don't let another task, with possibly unlocked vma,
2943 * keep the mlocked page.
2945 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2946 lock_page(old_page); /* LRU manipulation */
2947 if (PageMlocked(old_page))
2948 munlock_vma_page(old_page);
2949 unlock_page(old_page);
2953 return page_copied ? VM_FAULT_WRITE : 0;
2959 return VM_FAULT_OOM;
2963 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2964 * writeable once the page is prepared
2966 * @vmf: structure describing the fault
2968 * This function handles all that is needed to finish a write page fault in a
2969 * shared mapping due to PTE being read-only once the mapped page is prepared.
2970 * It handles locking of PTE and modifying it.
2972 * The function expects the page to be locked or other protection against
2973 * concurrent faults / writeback (such as DAX radix tree locks).
2975 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2976 * we acquired PTE lock.
2978 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2980 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2981 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2984 * We might have raced with another page fault while we released the
2985 * pte_offset_map_lock.
2987 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2988 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
2989 pte_unmap_unlock(vmf->pte, vmf->ptl);
2990 return VM_FAULT_NOPAGE;
2997 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3000 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3002 struct vm_area_struct *vma = vmf->vma;
3004 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3007 pte_unmap_unlock(vmf->pte, vmf->ptl);
3008 vmf->flags |= FAULT_FLAG_MKWRITE;
3009 ret = vma->vm_ops->pfn_mkwrite(vmf);
3010 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3012 return finish_mkwrite_fault(vmf);
3015 return VM_FAULT_WRITE;
3018 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3019 __releases(vmf->ptl)
3021 struct vm_area_struct *vma = vmf->vma;
3022 vm_fault_t ret = VM_FAULT_WRITE;
3024 get_page(vmf->page);
3026 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3029 pte_unmap_unlock(vmf->pte, vmf->ptl);
3030 tmp = do_page_mkwrite(vmf);
3031 if (unlikely(!tmp || (tmp &
3032 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3033 put_page(vmf->page);
3036 tmp = finish_mkwrite_fault(vmf);
3037 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3038 unlock_page(vmf->page);
3039 put_page(vmf->page);
3044 lock_page(vmf->page);
3046 ret |= fault_dirty_shared_page(vmf);
3047 put_page(vmf->page);
3053 * This routine handles present pages, when users try to write
3054 * to a shared page. It is done by copying the page to a new address
3055 * and decrementing the shared-page counter for the old page.
3057 * Note that this routine assumes that the protection checks have been
3058 * done by the caller (the low-level page fault routine in most cases).
3059 * Thus we can safely just mark it writable once we've done any necessary
3062 * We also mark the page dirty at this point even though the page will
3063 * change only once the write actually happens. This avoids a few races,
3064 * and potentially makes it more efficient.
3066 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3067 * but allow concurrent faults), with pte both mapped and locked.
3068 * We return with mmap_lock still held, but pte unmapped and unlocked.
3070 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3071 __releases(vmf->ptl)
3073 struct vm_area_struct *vma = vmf->vma;
3075 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3076 pte_unmap_unlock(vmf->pte, vmf->ptl);
3077 return handle_userfault(vmf, VM_UFFD_WP);
3080 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3083 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3086 * We should not cow pages in a shared writeable mapping.
3087 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3089 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3090 (VM_WRITE|VM_SHARED))
3091 return wp_pfn_shared(vmf);
3093 pte_unmap_unlock(vmf->pte, vmf->ptl);
3094 return wp_page_copy(vmf);
3098 * Take out anonymous pages first, anonymous shared vmas are
3099 * not dirty accountable.
3101 if (PageAnon(vmf->page)) {
3102 struct page *page = vmf->page;
3104 /* PageKsm() doesn't necessarily raise the page refcount */
3105 if (PageKsm(page) || page_count(page) != 1)
3107 if (!trylock_page(page))
3109 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3114 * Ok, we've got the only map reference, and the only
3115 * page count reference, and the page is locked,
3116 * it's dark out, and we're wearing sunglasses. Hit it.
3120 return VM_FAULT_WRITE;
3121 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3122 (VM_WRITE|VM_SHARED))) {
3123 return wp_page_shared(vmf);
3127 * Ok, we need to copy. Oh, well..
3129 get_page(vmf->page);
3131 pte_unmap_unlock(vmf->pte, vmf->ptl);
3132 return wp_page_copy(vmf);
3135 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3136 unsigned long start_addr, unsigned long end_addr,
3137 struct zap_details *details)
3139 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3142 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3143 struct zap_details *details)
3145 struct vm_area_struct *vma;
3146 pgoff_t vba, vea, zba, zea;
3148 vma_interval_tree_foreach(vma, root,
3149 details->first_index, details->last_index) {
3151 vba = vma->vm_pgoff;
3152 vea = vba + vma_pages(vma) - 1;
3153 zba = details->first_index;
3156 zea = details->last_index;
3160 unmap_mapping_range_vma(vma,
3161 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3162 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3168 * unmap_mapping_pages() - Unmap pages from processes.
3169 * @mapping: The address space containing pages to be unmapped.
3170 * @start: Index of first page to be unmapped.
3171 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3172 * @even_cows: Whether to unmap even private COWed pages.
3174 * Unmap the pages in this address space from any userspace process which
3175 * has them mmaped. Generally, you want to remove COWed pages as well when
3176 * a file is being truncated, but not when invalidating pages from the page
3179 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3180 pgoff_t nr, bool even_cows)
3182 struct zap_details details = { };
3184 details.check_mapping = even_cows ? NULL : mapping;
3185 details.first_index = start;
3186 details.last_index = start + nr - 1;
3187 if (details.last_index < details.first_index)
3188 details.last_index = ULONG_MAX;
3190 i_mmap_lock_write(mapping);
3191 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3192 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3193 i_mmap_unlock_write(mapping);
3197 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3198 * address_space corresponding to the specified byte range in the underlying
3201 * @mapping: the address space containing mmaps to be unmapped.
3202 * @holebegin: byte in first page to unmap, relative to the start of
3203 * the underlying file. This will be rounded down to a PAGE_SIZE
3204 * boundary. Note that this is different from truncate_pagecache(), which
3205 * must keep the partial page. In contrast, we must get rid of
3207 * @holelen: size of prospective hole in bytes. This will be rounded
3208 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3210 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3211 * but 0 when invalidating pagecache, don't throw away private data.
3213 void unmap_mapping_range(struct address_space *mapping,
3214 loff_t const holebegin, loff_t const holelen, int even_cows)
3216 pgoff_t hba = holebegin >> PAGE_SHIFT;
3217 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3219 /* Check for overflow. */
3220 if (sizeof(holelen) > sizeof(hlen)) {
3222 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3223 if (holeend & ~(long long)ULONG_MAX)
3224 hlen = ULONG_MAX - hba + 1;
3227 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3229 EXPORT_SYMBOL(unmap_mapping_range);
3232 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3233 * but allow concurrent faults), and pte mapped but not yet locked.
3234 * We return with pte unmapped and unlocked.
3236 * We return with the mmap_lock locked or unlocked in the same cases
3237 * as does filemap_fault().
3239 vm_fault_t do_swap_page(struct vm_fault *vmf)
3241 struct vm_area_struct *vma = vmf->vma;
3242 struct page *page = NULL, *swapcache;
3248 void *shadow = NULL;
3250 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3253 entry = pte_to_swp_entry(vmf->orig_pte);
3254 if (unlikely(non_swap_entry(entry))) {
3255 if (is_migration_entry(entry)) {
3256 migration_entry_wait(vma->vm_mm, vmf->pmd,
3258 } else if (is_device_private_entry(entry)) {
3259 vmf->page = device_private_entry_to_page(entry);
3260 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3261 } else if (is_hwpoison_entry(entry)) {
3262 ret = VM_FAULT_HWPOISON;
3264 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3265 ret = VM_FAULT_SIGBUS;
3271 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3272 page = lookup_swap_cache(entry, vma, vmf->address);
3276 struct swap_info_struct *si = swp_swap_info(entry);
3278 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3279 __swap_count(entry) == 1) {
3280 /* skip swapcache */
3281 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3286 __SetPageLocked(page);
3287 __SetPageSwapBacked(page);
3288 set_page_private(page, entry.val);
3290 /* Tell memcg to use swap ownership records */
3291 SetPageSwapCache(page);
3292 err = mem_cgroup_charge(page, vma->vm_mm,
3294 ClearPageSwapCache(page);
3300 shadow = get_shadow_from_swap_cache(entry);
3302 workingset_refault(page, shadow);
3304 lru_cache_add(page);
3305 swap_readpage(page, true);
3308 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3315 * Back out if somebody else faulted in this pte
3316 * while we released the pte lock.
3318 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3319 vmf->address, &vmf->ptl);
3320 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3322 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3326 /* Had to read the page from swap area: Major fault */
3327 ret = VM_FAULT_MAJOR;
3328 count_vm_event(PGMAJFAULT);
3329 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3330 } else if (PageHWPoison(page)) {
3332 * hwpoisoned dirty swapcache pages are kept for killing
3333 * owner processes (which may be unknown at hwpoison time)
3335 ret = VM_FAULT_HWPOISON;
3336 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3340 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3342 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3344 ret |= VM_FAULT_RETRY;
3349 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3350 * release the swapcache from under us. The page pin, and pte_same
3351 * test below, are not enough to exclude that. Even if it is still
3352 * swapcache, we need to check that the page's swap has not changed.
3354 if (unlikely((!PageSwapCache(page) ||
3355 page_private(page) != entry.val)) && swapcache)
3358 page = ksm_might_need_to_copy(page, vma, vmf->address);
3359 if (unlikely(!page)) {
3365 cgroup_throttle_swaprate(page, GFP_KERNEL);
3368 * Back out if somebody else already faulted in this pte.
3370 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3372 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3375 if (unlikely(!PageUptodate(page))) {
3376 ret = VM_FAULT_SIGBUS;
3381 * The page isn't present yet, go ahead with the fault.
3383 * Be careful about the sequence of operations here.
3384 * To get its accounting right, reuse_swap_page() must be called
3385 * while the page is counted on swap but not yet in mapcount i.e.
3386 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3387 * must be called after the swap_free(), or it will never succeed.
3390 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3391 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3392 pte = mk_pte(page, vma->vm_page_prot);
3393 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3394 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3395 vmf->flags &= ~FAULT_FLAG_WRITE;
3396 ret |= VM_FAULT_WRITE;
3397 exclusive = RMAP_EXCLUSIVE;
3399 flush_icache_page(vma, page);
3400 if (pte_swp_soft_dirty(vmf->orig_pte))
3401 pte = pte_mksoft_dirty(pte);
3402 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3403 pte = pte_mkuffd_wp(pte);
3404 pte = pte_wrprotect(pte);
3406 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3407 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3408 vmf->orig_pte = pte;
3410 /* ksm created a completely new copy */
3411 if (unlikely(page != swapcache && swapcache)) {
3412 page_add_new_anon_rmap(page, vma, vmf->address, false);
3413 lru_cache_add_inactive_or_unevictable(page, vma);
3415 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3419 if (mem_cgroup_swap_full(page) ||
3420 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3421 try_to_free_swap(page);
3423 if (page != swapcache && swapcache) {
3425 * Hold the lock to avoid the swap entry to be reused
3426 * until we take the PT lock for the pte_same() check
3427 * (to avoid false positives from pte_same). For
3428 * further safety release the lock after the swap_free
3429 * so that the swap count won't change under a
3430 * parallel locked swapcache.
3432 unlock_page(swapcache);
3433 put_page(swapcache);
3436 if (vmf->flags & FAULT_FLAG_WRITE) {
3437 ret |= do_wp_page(vmf);
3438 if (ret & VM_FAULT_ERROR)
3439 ret &= VM_FAULT_ERROR;
3443 /* No need to invalidate - it was non-present before */
3444 update_mmu_cache(vma, vmf->address, vmf->pte);
3446 pte_unmap_unlock(vmf->pte, vmf->ptl);
3450 pte_unmap_unlock(vmf->pte, vmf->ptl);
3455 if (page != swapcache && swapcache) {
3456 unlock_page(swapcache);
3457 put_page(swapcache);
3463 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3464 * but allow concurrent faults), and pte mapped but not yet locked.
3465 * We return with mmap_lock still held, but pte unmapped and unlocked.
3467 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3469 struct vm_area_struct *vma = vmf->vma;
3474 /* File mapping without ->vm_ops ? */
3475 if (vma->vm_flags & VM_SHARED)
3476 return VM_FAULT_SIGBUS;
3479 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3480 * pte_offset_map() on pmds where a huge pmd might be created
3481 * from a different thread.
3483 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3484 * parallel threads are excluded by other means.
3486 * Here we only have mmap_read_lock(mm).
3488 if (pte_alloc(vma->vm_mm, vmf->pmd))
3489 return VM_FAULT_OOM;
3491 /* See the comment in pte_alloc_one_map() */
3492 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3495 /* Use the zero-page for reads */
3496 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3497 !mm_forbids_zeropage(vma->vm_mm)) {
3498 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3499 vma->vm_page_prot));
3500 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3501 vmf->address, &vmf->ptl);
3502 if (!pte_none(*vmf->pte)) {
3503 update_mmu_tlb(vma, vmf->address, vmf->pte);
3506 ret = check_stable_address_space(vma->vm_mm);
3509 /* Deliver the page fault to userland, check inside PT lock */
3510 if (userfaultfd_missing(vma)) {
3511 pte_unmap_unlock(vmf->pte, vmf->ptl);
3512 return handle_userfault(vmf, VM_UFFD_MISSING);
3517 /* Allocate our own private page. */
3518 if (unlikely(anon_vma_prepare(vma)))
3520 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3524 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3526 cgroup_throttle_swaprate(page, GFP_KERNEL);
3529 * The memory barrier inside __SetPageUptodate makes sure that
3530 * preceding stores to the page contents become visible before
3531 * the set_pte_at() write.
3533 __SetPageUptodate(page);
3535 entry = mk_pte(page, vma->vm_page_prot);
3536 entry = pte_sw_mkyoung(entry);
3537 if (vma->vm_flags & VM_WRITE)
3538 entry = pte_mkwrite(pte_mkdirty(entry));
3540 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3542 if (!pte_none(*vmf->pte)) {
3543 update_mmu_cache(vma, vmf->address, vmf->pte);
3547 ret = check_stable_address_space(vma->vm_mm);
3551 /* Deliver the page fault to userland, check inside PT lock */
3552 if (userfaultfd_missing(vma)) {
3553 pte_unmap_unlock(vmf->pte, vmf->ptl);
3555 return handle_userfault(vmf, VM_UFFD_MISSING);
3558 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3559 page_add_new_anon_rmap(page, vma, vmf->address, false);
3560 lru_cache_add_inactive_or_unevictable(page, vma);
3562 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3564 /* No need to invalidate - it was non-present before */
3565 update_mmu_cache(vma, vmf->address, vmf->pte);
3567 pte_unmap_unlock(vmf->pte, vmf->ptl);
3575 return VM_FAULT_OOM;
3579 * The mmap_lock must have been held on entry, and may have been
3580 * released depending on flags and vma->vm_ops->fault() return value.
3581 * See filemap_fault() and __lock_page_retry().
3583 static vm_fault_t __do_fault(struct vm_fault *vmf)
3585 struct vm_area_struct *vma = vmf->vma;
3589 * Preallocate pte before we take page_lock because this might lead to
3590 * deadlocks for memcg reclaim which waits for pages under writeback:
3592 * SetPageWriteback(A)
3598 * wait_on_page_writeback(A)
3599 * SetPageWriteback(B)
3601 * # flush A, B to clear the writeback
3603 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3604 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3605 if (!vmf->prealloc_pte)
3606 return VM_FAULT_OOM;
3607 smp_wmb(); /* See comment in __pte_alloc() */
3610 ret = vma->vm_ops->fault(vmf);
3611 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3612 VM_FAULT_DONE_COW)))
3615 if (unlikely(PageHWPoison(vmf->page))) {
3616 if (ret & VM_FAULT_LOCKED)
3617 unlock_page(vmf->page);
3618 put_page(vmf->page);
3620 return VM_FAULT_HWPOISON;
3623 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3624 lock_page(vmf->page);
3626 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3632 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3633 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3634 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3635 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3637 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3639 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3642 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3644 struct vm_area_struct *vma = vmf->vma;
3646 if (!pmd_none(*vmf->pmd))
3648 if (vmf->prealloc_pte) {
3649 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3650 if (unlikely(!pmd_none(*vmf->pmd))) {
3651 spin_unlock(vmf->ptl);
3655 mm_inc_nr_ptes(vma->vm_mm);
3656 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3657 spin_unlock(vmf->ptl);
3658 vmf->prealloc_pte = NULL;
3659 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3660 return VM_FAULT_OOM;
3664 * If a huge pmd materialized under us just retry later. Use
3665 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3666 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3667 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3668 * running immediately after a huge pmd fault in a different thread of
3669 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3670 * All we have to ensure is that it is a regular pmd that we can walk
3671 * with pte_offset_map() and we can do that through an atomic read in
3672 * C, which is what pmd_trans_unstable() provides.
3674 if (pmd_devmap_trans_unstable(vmf->pmd))
3675 return VM_FAULT_NOPAGE;
3678 * At this point we know that our vmf->pmd points to a page of ptes
3679 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3680 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3681 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3682 * be valid and we will re-check to make sure the vmf->pte isn't
3683 * pte_none() under vmf->ptl protection when we return to
3686 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3691 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3692 static void deposit_prealloc_pte(struct vm_fault *vmf)
3694 struct vm_area_struct *vma = vmf->vma;
3696 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3698 * We are going to consume the prealloc table,
3699 * count that as nr_ptes.
3701 mm_inc_nr_ptes(vma->vm_mm);
3702 vmf->prealloc_pte = NULL;
3705 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3707 struct vm_area_struct *vma = vmf->vma;
3708 bool write = vmf->flags & FAULT_FLAG_WRITE;
3709 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3712 vm_fault_t ret = VM_FAULT_FALLBACK;
3714 if (!transhuge_vma_suitable(vma, haddr))
3717 page = compound_head(page);
3718 if (compound_order(page) != HPAGE_PMD_ORDER)
3722 * Archs like ppc64 need additonal space to store information
3723 * related to pte entry. Use the preallocated table for that.
3725 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3726 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3727 if (!vmf->prealloc_pte)
3728 return VM_FAULT_OOM;
3729 smp_wmb(); /* See comment in __pte_alloc() */
3732 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3733 if (unlikely(!pmd_none(*vmf->pmd)))
3736 for (i = 0; i < HPAGE_PMD_NR; i++)
3737 flush_icache_page(vma, page + i);
3739 entry = mk_huge_pmd(page, vma->vm_page_prot);
3741 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3743 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3744 page_add_file_rmap(page, true);
3746 * deposit and withdraw with pmd lock held
3748 if (arch_needs_pgtable_deposit())
3749 deposit_prealloc_pte(vmf);
3751 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3753 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3755 /* fault is handled */
3757 count_vm_event(THP_FILE_MAPPED);
3759 spin_unlock(vmf->ptl);
3763 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3771 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3772 * mapping. If needed, the function allocates page table or use pre-allocated.
3774 * @vmf: fault environment
3775 * @page: page to map
3777 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3780 * Target users are page handler itself and implementations of
3781 * vm_ops->map_pages.
3783 * Return: %0 on success, %VM_FAULT_ code in case of error.
3785 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
3787 struct vm_area_struct *vma = vmf->vma;
3788 bool write = vmf->flags & FAULT_FLAG_WRITE;
3792 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3793 ret = do_set_pmd(vmf, page);
3794 if (ret != VM_FAULT_FALLBACK)
3799 ret = pte_alloc_one_map(vmf);
3804 /* Re-check under ptl */
3805 if (unlikely(!pte_none(*vmf->pte))) {
3806 update_mmu_tlb(vma, vmf->address, vmf->pte);
3807 return VM_FAULT_NOPAGE;
3810 flush_icache_page(vma, page);
3811 entry = mk_pte(page, vma->vm_page_prot);
3812 entry = pte_sw_mkyoung(entry);
3814 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3815 /* copy-on-write page */
3816 if (write && !(vma->vm_flags & VM_SHARED)) {
3817 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3818 page_add_new_anon_rmap(page, vma, vmf->address, false);
3819 lru_cache_add_inactive_or_unevictable(page, vma);
3821 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3822 page_add_file_rmap(page, false);
3824 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3826 /* no need to invalidate: a not-present page won't be cached */
3827 update_mmu_cache(vma, vmf->address, vmf->pte);
3834 * finish_fault - finish page fault once we have prepared the page to fault
3836 * @vmf: structure describing the fault
3838 * This function handles all that is needed to finish a page fault once the
3839 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3840 * given page, adds reverse page mapping, handles memcg charges and LRU
3843 * The function expects the page to be locked and on success it consumes a
3844 * reference of a page being mapped (for the PTE which maps it).
3846 * Return: %0 on success, %VM_FAULT_ code in case of error.
3848 vm_fault_t finish_fault(struct vm_fault *vmf)
3853 /* Did we COW the page? */
3854 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3855 !(vmf->vma->vm_flags & VM_SHARED))
3856 page = vmf->cow_page;
3861 * check even for read faults because we might have lost our CoWed
3864 if (!(vmf->vma->vm_flags & VM_SHARED))
3865 ret = check_stable_address_space(vmf->vma->vm_mm);
3867 ret = alloc_set_pte(vmf, page);
3869 pte_unmap_unlock(vmf->pte, vmf->ptl);
3873 static unsigned long fault_around_bytes __read_mostly =
3874 rounddown_pow_of_two(65536);
3876 #ifdef CONFIG_DEBUG_FS
3877 static int fault_around_bytes_get(void *data, u64 *val)
3879 *val = fault_around_bytes;
3884 * fault_around_bytes must be rounded down to the nearest page order as it's
3885 * what do_fault_around() expects to see.
3887 static int fault_around_bytes_set(void *data, u64 val)
3889 if (val / PAGE_SIZE > PTRS_PER_PTE)
3891 if (val > PAGE_SIZE)
3892 fault_around_bytes = rounddown_pow_of_two(val);
3894 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3897 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3898 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3900 static int __init fault_around_debugfs(void)
3902 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3903 &fault_around_bytes_fops);
3906 late_initcall(fault_around_debugfs);
3910 * do_fault_around() tries to map few pages around the fault address. The hope
3911 * is that the pages will be needed soon and this will lower the number of
3914 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3915 * not ready to be mapped: not up-to-date, locked, etc.
3917 * This function is called with the page table lock taken. In the split ptlock
3918 * case the page table lock only protects only those entries which belong to
3919 * the page table corresponding to the fault address.
3921 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3924 * fault_around_bytes defines how many bytes we'll try to map.
3925 * do_fault_around() expects it to be set to a power of two less than or equal
3928 * The virtual address of the area that we map is naturally aligned to
3929 * fault_around_bytes rounded down to the machine page size
3930 * (and therefore to page order). This way it's easier to guarantee
3931 * that we don't cross page table boundaries.
3933 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3935 unsigned long address = vmf->address, nr_pages, mask;
3936 pgoff_t start_pgoff = vmf->pgoff;
3941 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3942 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3944 vmf->address = max(address & mask, vmf->vma->vm_start);
3945 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3949 * end_pgoff is either the end of the page table, the end of
3950 * the vma or nr_pages from start_pgoff, depending what is nearest.
3952 end_pgoff = start_pgoff -
3953 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3955 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3956 start_pgoff + nr_pages - 1);
3958 if (pmd_none(*vmf->pmd)) {
3959 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3960 if (!vmf->prealloc_pte)
3962 smp_wmb(); /* See comment in __pte_alloc() */
3965 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3967 /* Huge page is mapped? Page fault is solved */
3968 if (pmd_trans_huge(*vmf->pmd)) {
3969 ret = VM_FAULT_NOPAGE;
3973 /* ->map_pages() haven't done anything useful. Cold page cache? */
3977 /* check if the page fault is solved */
3978 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3979 if (!pte_none(*vmf->pte))
3980 ret = VM_FAULT_NOPAGE;
3981 pte_unmap_unlock(vmf->pte, vmf->ptl);
3983 vmf->address = address;
3988 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3990 struct vm_area_struct *vma = vmf->vma;
3994 * Let's call ->map_pages() first and use ->fault() as fallback
3995 * if page by the offset is not ready to be mapped (cold cache or
3998 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3999 ret = do_fault_around(vmf);
4004 ret = __do_fault(vmf);
4005 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4008 ret |= finish_fault(vmf);
4009 unlock_page(vmf->page);
4010 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4011 put_page(vmf->page);
4015 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4017 struct vm_area_struct *vma = vmf->vma;
4020 if (unlikely(anon_vma_prepare(vma)))
4021 return VM_FAULT_OOM;
4023 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4025 return VM_FAULT_OOM;
4027 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4028 put_page(vmf->cow_page);
4029 return VM_FAULT_OOM;
4031 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4033 ret = __do_fault(vmf);
4034 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4036 if (ret & VM_FAULT_DONE_COW)
4039 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4040 __SetPageUptodate(vmf->cow_page);
4042 ret |= finish_fault(vmf);
4043 unlock_page(vmf->page);
4044 put_page(vmf->page);
4045 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4049 put_page(vmf->cow_page);
4053 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4055 struct vm_area_struct *vma = vmf->vma;
4056 vm_fault_t ret, tmp;
4058 ret = __do_fault(vmf);
4059 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4063 * Check if the backing address space wants to know that the page is
4064 * about to become writable
4066 if (vma->vm_ops->page_mkwrite) {
4067 unlock_page(vmf->page);
4068 tmp = do_page_mkwrite(vmf);
4069 if (unlikely(!tmp ||
4070 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4071 put_page(vmf->page);
4076 ret |= finish_fault(vmf);
4077 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4079 unlock_page(vmf->page);
4080 put_page(vmf->page);
4084 ret |= fault_dirty_shared_page(vmf);
4089 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4090 * but allow concurrent faults).
4091 * The mmap_lock may have been released depending on flags and our
4092 * return value. See filemap_fault() and __lock_page_or_retry().
4093 * If mmap_lock is released, vma may become invalid (for example
4094 * by other thread calling munmap()).
4096 static vm_fault_t do_fault(struct vm_fault *vmf)
4098 struct vm_area_struct *vma = vmf->vma;
4099 struct mm_struct *vm_mm = vma->vm_mm;
4103 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4105 if (!vma->vm_ops->fault) {
4107 * If we find a migration pmd entry or a none pmd entry, which
4108 * should never happen, return SIGBUS
4110 if (unlikely(!pmd_present(*vmf->pmd)))
4111 ret = VM_FAULT_SIGBUS;
4113 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4118 * Make sure this is not a temporary clearing of pte
4119 * by holding ptl and checking again. A R/M/W update
4120 * of pte involves: take ptl, clearing the pte so that
4121 * we don't have concurrent modification by hardware
4122 * followed by an update.
4124 if (unlikely(pte_none(*vmf->pte)))
4125 ret = VM_FAULT_SIGBUS;
4127 ret = VM_FAULT_NOPAGE;
4129 pte_unmap_unlock(vmf->pte, vmf->ptl);
4131 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4132 ret = do_read_fault(vmf);
4133 else if (!(vma->vm_flags & VM_SHARED))
4134 ret = do_cow_fault(vmf);
4136 ret = do_shared_fault(vmf);
4138 /* preallocated pagetable is unused: free it */
4139 if (vmf->prealloc_pte) {
4140 pte_free(vm_mm, vmf->prealloc_pte);
4141 vmf->prealloc_pte = NULL;
4146 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4147 unsigned long addr, int page_nid,
4152 count_vm_numa_event(NUMA_HINT_FAULTS);
4153 if (page_nid == numa_node_id()) {
4154 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4155 *flags |= TNF_FAULT_LOCAL;
4158 return mpol_misplaced(page, vma, addr);
4161 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4163 struct vm_area_struct *vma = vmf->vma;
4164 struct page *page = NULL;
4165 int page_nid = NUMA_NO_NODE;
4168 bool migrated = false;
4170 bool was_writable = pte_savedwrite(vmf->orig_pte);
4174 * The "pte" at this point cannot be used safely without
4175 * validation through pte_unmap_same(). It's of NUMA type but
4176 * the pfn may be screwed if the read is non atomic.
4178 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4179 spin_lock(vmf->ptl);
4180 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4181 pte_unmap_unlock(vmf->pte, vmf->ptl);
4186 * Make it present again, Depending on how arch implementes non
4187 * accessible ptes, some can allow access by kernel mode.
4189 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4190 pte = pte_modify(old_pte, vma->vm_page_prot);
4191 pte = pte_mkyoung(pte);
4193 pte = pte_mkwrite(pte);
4194 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4195 update_mmu_cache(vma, vmf->address, vmf->pte);
4197 page = vm_normal_page(vma, vmf->address, pte);
4199 pte_unmap_unlock(vmf->pte, vmf->ptl);
4203 /* TODO: handle PTE-mapped THP */
4204 if (PageCompound(page)) {
4205 pte_unmap_unlock(vmf->pte, vmf->ptl);
4210 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4211 * much anyway since they can be in shared cache state. This misses
4212 * the case where a mapping is writable but the process never writes
4213 * to it but pte_write gets cleared during protection updates and
4214 * pte_dirty has unpredictable behaviour between PTE scan updates,
4215 * background writeback, dirty balancing and application behaviour.
4217 if (!pte_write(pte))
4218 flags |= TNF_NO_GROUP;
4221 * Flag if the page is shared between multiple address spaces. This
4222 * is later used when determining whether to group tasks together
4224 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4225 flags |= TNF_SHARED;
4227 last_cpupid = page_cpupid_last(page);
4228 page_nid = page_to_nid(page);
4229 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4231 pte_unmap_unlock(vmf->pte, vmf->ptl);
4232 if (target_nid == NUMA_NO_NODE) {
4237 /* Migrate to the requested node */
4238 migrated = migrate_misplaced_page(page, vma, target_nid);
4240 page_nid = target_nid;
4241 flags |= TNF_MIGRATED;
4243 flags |= TNF_MIGRATE_FAIL;
4246 if (page_nid != NUMA_NO_NODE)
4247 task_numa_fault(last_cpupid, page_nid, 1, flags);
4251 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4253 if (vma_is_anonymous(vmf->vma))
4254 return do_huge_pmd_anonymous_page(vmf);
4255 if (vmf->vma->vm_ops->huge_fault)
4256 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4257 return VM_FAULT_FALLBACK;
4260 /* `inline' is required to avoid gcc 4.1.2 build error */
4261 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4263 if (vma_is_anonymous(vmf->vma)) {
4264 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4265 return handle_userfault(vmf, VM_UFFD_WP);
4266 return do_huge_pmd_wp_page(vmf, orig_pmd);
4268 if (vmf->vma->vm_ops->huge_fault) {
4269 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4271 if (!(ret & VM_FAULT_FALLBACK))
4275 /* COW or write-notify handled on pte level: split pmd. */
4276 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4278 return VM_FAULT_FALLBACK;
4281 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4283 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4284 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4285 /* No support for anonymous transparent PUD pages yet */
4286 if (vma_is_anonymous(vmf->vma))
4288 if (vmf->vma->vm_ops->huge_fault) {
4289 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4291 if (!(ret & VM_FAULT_FALLBACK))
4295 /* COW or write-notify not handled on PUD level: split pud.*/
4296 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4297 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4298 return VM_FAULT_FALLBACK;
4301 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4303 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4304 /* No support for anonymous transparent PUD pages yet */
4305 if (vma_is_anonymous(vmf->vma))
4306 return VM_FAULT_FALLBACK;
4307 if (vmf->vma->vm_ops->huge_fault)
4308 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4309 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4310 return VM_FAULT_FALLBACK;
4314 * These routines also need to handle stuff like marking pages dirty
4315 * and/or accessed for architectures that don't do it in hardware (most
4316 * RISC architectures). The early dirtying is also good on the i386.
4318 * There is also a hook called "update_mmu_cache()" that architectures
4319 * with external mmu caches can use to update those (ie the Sparc or
4320 * PowerPC hashed page tables that act as extended TLBs).
4322 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4323 * concurrent faults).
4325 * The mmap_lock may have been released depending on flags and our return value.
4326 * See filemap_fault() and __lock_page_or_retry().
4328 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4332 if (unlikely(pmd_none(*vmf->pmd))) {
4334 * Leave __pte_alloc() until later: because vm_ops->fault may
4335 * want to allocate huge page, and if we expose page table
4336 * for an instant, it will be difficult to retract from
4337 * concurrent faults and from rmap lookups.
4341 /* See comment in pte_alloc_one_map() */
4342 if (pmd_devmap_trans_unstable(vmf->pmd))
4345 * A regular pmd is established and it can't morph into a huge
4346 * pmd from under us anymore at this point because we hold the
4347 * mmap_lock read mode and khugepaged takes it in write mode.
4348 * So now it's safe to run pte_offset_map().
4350 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4351 vmf->orig_pte = *vmf->pte;
4354 * some architectures can have larger ptes than wordsize,
4355 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4356 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4357 * accesses. The code below just needs a consistent view
4358 * for the ifs and we later double check anyway with the
4359 * ptl lock held. So here a barrier will do.
4362 if (pte_none(vmf->orig_pte)) {
4363 pte_unmap(vmf->pte);
4369 if (vma_is_anonymous(vmf->vma))
4370 return do_anonymous_page(vmf);
4372 return do_fault(vmf);
4375 if (!pte_present(vmf->orig_pte))
4376 return do_swap_page(vmf);
4378 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4379 return do_numa_page(vmf);
4381 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4382 spin_lock(vmf->ptl);
4383 entry = vmf->orig_pte;
4384 if (unlikely(!pte_same(*vmf->pte, entry))) {
4385 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4388 if (vmf->flags & FAULT_FLAG_WRITE) {
4389 if (!pte_write(entry))
4390 return do_wp_page(vmf);
4391 entry = pte_mkdirty(entry);
4393 entry = pte_mkyoung(entry);
4394 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4395 vmf->flags & FAULT_FLAG_WRITE)) {
4396 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4398 /* Skip spurious TLB flush for retried page fault */
4399 if (vmf->flags & FAULT_FLAG_TRIED)
4402 * This is needed only for protection faults but the arch code
4403 * is not yet telling us if this is a protection fault or not.
4404 * This still avoids useless tlb flushes for .text page faults
4407 if (vmf->flags & FAULT_FLAG_WRITE)
4408 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4411 pte_unmap_unlock(vmf->pte, vmf->ptl);
4416 * By the time we get here, we already hold the mm semaphore
4418 * The mmap_lock may have been released depending on flags and our
4419 * return value. See filemap_fault() and __lock_page_or_retry().
4421 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4422 unsigned long address, unsigned int flags)
4424 struct vm_fault vmf = {
4426 .address = address & PAGE_MASK,
4428 .pgoff = linear_page_index(vma, address),
4429 .gfp_mask = __get_fault_gfp_mask(vma),
4431 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4432 struct mm_struct *mm = vma->vm_mm;
4437 pgd = pgd_offset(mm, address);
4438 p4d = p4d_alloc(mm, pgd, address);
4440 return VM_FAULT_OOM;
4442 vmf.pud = pud_alloc(mm, p4d, address);
4444 return VM_FAULT_OOM;
4446 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4447 ret = create_huge_pud(&vmf);
4448 if (!(ret & VM_FAULT_FALLBACK))
4451 pud_t orig_pud = *vmf.pud;
4454 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4456 /* NUMA case for anonymous PUDs would go here */
4458 if (dirty && !pud_write(orig_pud)) {
4459 ret = wp_huge_pud(&vmf, orig_pud);
4460 if (!(ret & VM_FAULT_FALLBACK))
4463 huge_pud_set_accessed(&vmf, orig_pud);
4469 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4471 return VM_FAULT_OOM;
4473 /* Huge pud page fault raced with pmd_alloc? */
4474 if (pud_trans_unstable(vmf.pud))
4477 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4478 ret = create_huge_pmd(&vmf);
4479 if (!(ret & VM_FAULT_FALLBACK))
4482 pmd_t orig_pmd = *vmf.pmd;
4485 if (unlikely(is_swap_pmd(orig_pmd))) {
4486 VM_BUG_ON(thp_migration_supported() &&
4487 !is_pmd_migration_entry(orig_pmd));
4488 if (is_pmd_migration_entry(orig_pmd))
4489 pmd_migration_entry_wait(mm, vmf.pmd);
4492 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4493 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4494 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4496 if (dirty && !pmd_write(orig_pmd)) {
4497 ret = wp_huge_pmd(&vmf, orig_pmd);
4498 if (!(ret & VM_FAULT_FALLBACK))
4501 huge_pmd_set_accessed(&vmf, orig_pmd);
4507 return handle_pte_fault(&vmf);
4511 * mm_account_fault - Do page fault accountings
4513 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4514 * of perf event counters, but we'll still do the per-task accounting to
4515 * the task who triggered this page fault.
4516 * @address: the faulted address.
4517 * @flags: the fault flags.
4518 * @ret: the fault retcode.
4520 * This will take care of most of the page fault accountings. Meanwhile, it
4521 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4522 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4523 * still be in per-arch page fault handlers at the entry of page fault.
4525 static inline void mm_account_fault(struct pt_regs *regs,
4526 unsigned long address, unsigned int flags,
4532 * We don't do accounting for some specific faults:
4534 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4535 * includes arch_vma_access_permitted() failing before reaching here.
4536 * So this is not a "this many hardware page faults" counter. We
4537 * should use the hw profiling for that.
4539 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4540 * once they're completed.
4542 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4546 * We define the fault as a major fault when the final successful fault
4547 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4548 * handle it immediately previously).
4550 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4558 * If the fault is done for GUP, regs will be NULL. We only do the
4559 * accounting for the per thread fault counters who triggered the
4560 * fault, and we skip the perf event updates.
4566 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4568 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4572 * By the time we get here, we already hold the mm semaphore
4574 * The mmap_lock may have been released depending on flags and our
4575 * return value. See filemap_fault() and __lock_page_or_retry().
4577 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4578 unsigned int flags, struct pt_regs *regs)
4582 __set_current_state(TASK_RUNNING);
4584 count_vm_event(PGFAULT);
4585 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4587 /* do counter updates before entering really critical section. */
4588 check_sync_rss_stat(current);
4590 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4591 flags & FAULT_FLAG_INSTRUCTION,
4592 flags & FAULT_FLAG_REMOTE))
4593 return VM_FAULT_SIGSEGV;
4596 * Enable the memcg OOM handling for faults triggered in user
4597 * space. Kernel faults are handled more gracefully.
4599 if (flags & FAULT_FLAG_USER)
4600 mem_cgroup_enter_user_fault();
4602 if (unlikely(is_vm_hugetlb_page(vma)))
4603 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4605 ret = __handle_mm_fault(vma, address, flags);
4607 if (flags & FAULT_FLAG_USER) {
4608 mem_cgroup_exit_user_fault();
4610 * The task may have entered a memcg OOM situation but
4611 * if the allocation error was handled gracefully (no
4612 * VM_FAULT_OOM), there is no need to kill anything.
4613 * Just clean up the OOM state peacefully.
4615 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4616 mem_cgroup_oom_synchronize(false);
4619 mm_account_fault(regs, address, flags, ret);
4623 EXPORT_SYMBOL_GPL(handle_mm_fault);
4625 #ifndef __PAGETABLE_P4D_FOLDED
4627 * Allocate p4d page table.
4628 * We've already handled the fast-path in-line.
4630 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4632 p4d_t *new = p4d_alloc_one(mm, address);
4636 smp_wmb(); /* See comment in __pte_alloc */
4638 spin_lock(&mm->page_table_lock);
4639 if (pgd_present(*pgd)) /* Another has populated it */
4642 pgd_populate(mm, pgd, new);
4643 spin_unlock(&mm->page_table_lock);
4646 #endif /* __PAGETABLE_P4D_FOLDED */
4648 #ifndef __PAGETABLE_PUD_FOLDED
4650 * Allocate page upper directory.
4651 * We've already handled the fast-path in-line.
4653 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4655 pud_t *new = pud_alloc_one(mm, address);
4659 smp_wmb(); /* See comment in __pte_alloc */
4661 spin_lock(&mm->page_table_lock);
4662 if (!p4d_present(*p4d)) {
4664 p4d_populate(mm, p4d, new);
4665 } else /* Another has populated it */
4667 spin_unlock(&mm->page_table_lock);
4670 #endif /* __PAGETABLE_PUD_FOLDED */
4672 #ifndef __PAGETABLE_PMD_FOLDED
4674 * Allocate page middle directory.
4675 * We've already handled the fast-path in-line.
4677 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4680 pmd_t *new = pmd_alloc_one(mm, address);
4684 smp_wmb(); /* See comment in __pte_alloc */
4686 ptl = pud_lock(mm, pud);
4687 if (!pud_present(*pud)) {
4689 pud_populate(mm, pud, new);
4690 } else /* Another has populated it */
4695 #endif /* __PAGETABLE_PMD_FOLDED */
4697 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4698 struct mmu_notifier_range *range,
4699 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4707 pgd = pgd_offset(mm, address);
4708 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4711 p4d = p4d_offset(pgd, address);
4712 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4715 pud = pud_offset(p4d, address);
4716 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4719 pmd = pmd_offset(pud, address);
4720 VM_BUG_ON(pmd_trans_huge(*pmd));
4722 if (pmd_huge(*pmd)) {
4727 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4728 NULL, mm, address & PMD_MASK,
4729 (address & PMD_MASK) + PMD_SIZE);
4730 mmu_notifier_invalidate_range_start(range);
4732 *ptlp = pmd_lock(mm, pmd);
4733 if (pmd_huge(*pmd)) {
4739 mmu_notifier_invalidate_range_end(range);
4742 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4746 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4747 address & PAGE_MASK,
4748 (address & PAGE_MASK) + PAGE_SIZE);
4749 mmu_notifier_invalidate_range_start(range);
4751 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4752 if (!pte_present(*ptep))
4757 pte_unmap_unlock(ptep, *ptlp);
4759 mmu_notifier_invalidate_range_end(range);
4764 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4765 pte_t **ptepp, spinlock_t **ptlp)
4769 /* (void) is needed to make gcc happy */
4770 (void) __cond_lock(*ptlp,
4771 !(res = __follow_pte_pmd(mm, address, NULL,
4772 ptepp, NULL, ptlp)));
4776 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4777 struct mmu_notifier_range *range,
4778 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4782 /* (void) is needed to make gcc happy */
4783 (void) __cond_lock(*ptlp,
4784 !(res = __follow_pte_pmd(mm, address, range,
4785 ptepp, pmdpp, ptlp)));
4788 EXPORT_SYMBOL(follow_pte_pmd);
4791 * follow_pfn - look up PFN at a user virtual address
4792 * @vma: memory mapping
4793 * @address: user virtual address
4794 * @pfn: location to store found PFN
4796 * Only IO mappings and raw PFN mappings are allowed.
4798 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4800 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4807 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4810 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4813 *pfn = pte_pfn(*ptep);
4814 pte_unmap_unlock(ptep, ptl);
4817 EXPORT_SYMBOL(follow_pfn);
4819 #ifdef CONFIG_HAVE_IOREMAP_PROT
4820 int follow_phys(struct vm_area_struct *vma,
4821 unsigned long address, unsigned int flags,
4822 unsigned long *prot, resource_size_t *phys)
4828 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4831 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4835 if ((flags & FOLL_WRITE) && !pte_write(pte))
4838 *prot = pgprot_val(pte_pgprot(pte));
4839 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4843 pte_unmap_unlock(ptep, ptl);
4848 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4849 void *buf, int len, int write)
4851 resource_size_t phys_addr;
4852 unsigned long prot = 0;
4853 void __iomem *maddr;
4854 int offset = addr & (PAGE_SIZE-1);
4856 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4859 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4864 memcpy_toio(maddr + offset, buf, len);
4866 memcpy_fromio(buf, maddr + offset, len);
4871 EXPORT_SYMBOL_GPL(generic_access_phys);
4875 * Access another process' address space as given in mm. If non-NULL, use the
4876 * given task for page fault accounting.
4878 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4879 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4881 struct vm_area_struct *vma;
4882 void *old_buf = buf;
4883 int write = gup_flags & FOLL_WRITE;
4885 if (mmap_read_lock_killable(mm))
4888 /* ignore errors, just check how much was successfully transferred */
4890 int bytes, ret, offset;
4892 struct page *page = NULL;
4894 ret = get_user_pages_remote(mm, addr, 1,
4895 gup_flags, &page, &vma, NULL);
4897 #ifndef CONFIG_HAVE_IOREMAP_PROT
4901 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4902 * we can access using slightly different code.
4904 vma = find_vma(mm, addr);
4905 if (!vma || vma->vm_start > addr)
4907 if (vma->vm_ops && vma->vm_ops->access)
4908 ret = vma->vm_ops->access(vma, addr, buf,
4916 offset = addr & (PAGE_SIZE-1);
4917 if (bytes > PAGE_SIZE-offset)
4918 bytes = PAGE_SIZE-offset;
4922 copy_to_user_page(vma, page, addr,
4923 maddr + offset, buf, bytes);
4924 set_page_dirty_lock(page);
4926 copy_from_user_page(vma, page, addr,
4927 buf, maddr + offset, bytes);
4936 mmap_read_unlock(mm);
4938 return buf - old_buf;
4942 * access_remote_vm - access another process' address space
4943 * @mm: the mm_struct of the target address space
4944 * @addr: start address to access
4945 * @buf: source or destination buffer
4946 * @len: number of bytes to transfer
4947 * @gup_flags: flags modifying lookup behaviour
4949 * The caller must hold a reference on @mm.
4951 * Return: number of bytes copied from source to destination.
4953 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4954 void *buf, int len, unsigned int gup_flags)
4956 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4960 * Access another process' address space.
4961 * Source/target buffer must be kernel space,
4962 * Do not walk the page table directly, use get_user_pages
4964 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4965 void *buf, int len, unsigned int gup_flags)
4967 struct mm_struct *mm;
4970 mm = get_task_mm(tsk);
4974 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4980 EXPORT_SYMBOL_GPL(access_process_vm);
4983 * Print the name of a VMA.
4985 void print_vma_addr(char *prefix, unsigned long ip)
4987 struct mm_struct *mm = current->mm;
4988 struct vm_area_struct *vma;
4991 * we might be running from an atomic context so we cannot sleep
4993 if (!mmap_read_trylock(mm))
4996 vma = find_vma(mm, ip);
4997 if (vma && vma->vm_file) {
4998 struct file *f = vma->vm_file;
4999 char *buf = (char *)__get_free_page(GFP_NOWAIT);
5003 p = file_path(f, buf, PAGE_SIZE);
5006 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5008 vma->vm_end - vma->vm_start);
5009 free_page((unsigned long)buf);
5012 mmap_read_unlock(mm);
5015 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5016 void __might_fault(const char *file, int line)
5019 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5020 * holding the mmap_lock, this is safe because kernel memory doesn't
5021 * get paged out, therefore we'll never actually fault, and the
5022 * below annotations will generate false positives.
5024 if (uaccess_kernel())
5026 if (pagefault_disabled())
5028 __might_sleep(file, line, 0);
5029 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5031 might_lock_read(¤t->mm->mmap_lock);
5034 EXPORT_SYMBOL(__might_fault);
5037 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5039 * Process all subpages of the specified huge page with the specified
5040 * operation. The target subpage will be processed last to keep its
5043 static inline void process_huge_page(
5044 unsigned long addr_hint, unsigned int pages_per_huge_page,
5045 void (*process_subpage)(unsigned long addr, int idx, void *arg),
5049 unsigned long addr = addr_hint &
5050 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5052 /* Process target subpage last to keep its cache lines hot */
5054 n = (addr_hint - addr) / PAGE_SIZE;
5055 if (2 * n <= pages_per_huge_page) {
5056 /* If target subpage in first half of huge page */
5059 /* Process subpages at the end of huge page */
5060 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5062 process_subpage(addr + i * PAGE_SIZE, i, arg);
5065 /* If target subpage in second half of huge page */
5066 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5067 l = pages_per_huge_page - n;
5068 /* Process subpages at the begin of huge page */
5069 for (i = 0; i < base; i++) {
5071 process_subpage(addr + i * PAGE_SIZE, i, arg);
5075 * Process remaining subpages in left-right-left-right pattern
5076 * towards the target subpage
5078 for (i = 0; i < l; i++) {
5079 int left_idx = base + i;
5080 int right_idx = base + 2 * l - 1 - i;
5083 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5085 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5089 static void clear_gigantic_page(struct page *page,
5091 unsigned int pages_per_huge_page)
5094 struct page *p = page;
5097 for (i = 0; i < pages_per_huge_page;
5098 i++, p = mem_map_next(p, page, i)) {
5100 clear_user_highpage(p, addr + i * PAGE_SIZE);
5104 static void clear_subpage(unsigned long addr, int idx, void *arg)
5106 struct page *page = arg;
5108 clear_user_highpage(page + idx, addr);
5111 void clear_huge_page(struct page *page,
5112 unsigned long addr_hint, unsigned int pages_per_huge_page)
5114 unsigned long addr = addr_hint &
5115 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5117 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5118 clear_gigantic_page(page, addr, pages_per_huge_page);
5122 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5125 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5127 struct vm_area_struct *vma,
5128 unsigned int pages_per_huge_page)
5131 struct page *dst_base = dst;
5132 struct page *src_base = src;
5134 for (i = 0; i < pages_per_huge_page; ) {
5136 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5139 dst = mem_map_next(dst, dst_base, i);
5140 src = mem_map_next(src, src_base, i);
5144 struct copy_subpage_arg {
5147 struct vm_area_struct *vma;
5150 static void copy_subpage(unsigned long addr, int idx, void *arg)
5152 struct copy_subpage_arg *copy_arg = arg;
5154 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5155 addr, copy_arg->vma);
5158 void copy_user_huge_page(struct page *dst, struct page *src,
5159 unsigned long addr_hint, struct vm_area_struct *vma,
5160 unsigned int pages_per_huge_page)
5162 unsigned long addr = addr_hint &
5163 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5164 struct copy_subpage_arg arg = {
5170 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5171 copy_user_gigantic_page(dst, src, addr, vma,
5172 pages_per_huge_page);
5176 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5179 long copy_huge_page_from_user(struct page *dst_page,
5180 const void __user *usr_src,
5181 unsigned int pages_per_huge_page,
5182 bool allow_pagefault)
5184 void *src = (void *)usr_src;
5186 unsigned long i, rc = 0;
5187 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5189 for (i = 0; i < pages_per_huge_page; i++) {
5190 if (allow_pagefault)
5191 page_kaddr = kmap(dst_page + i);
5193 page_kaddr = kmap_atomic(dst_page + i);
5194 rc = copy_from_user(page_kaddr,
5195 (const void __user *)(src + i * PAGE_SIZE),
5197 if (allow_pagefault)
5198 kunmap(dst_page + i);
5200 kunmap_atomic(page_kaddr);
5202 ret_val -= (PAGE_SIZE - rc);
5210 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5212 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5214 static struct kmem_cache *page_ptl_cachep;
5216 void __init ptlock_cache_init(void)
5218 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5222 bool ptlock_alloc(struct page *page)
5226 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5233 void ptlock_free(struct page *page)
5235 kmem_cache_free(page_ptl_cachep, page->ptl);