1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
42 #include <linux/kernel_stat.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/dma-debug.h>
69 #include <linux/debugfs.h>
70 #include <linux/userfaultfd_k.h>
71 #include <linux/dax.h>
72 #include <linux/oom.h>
73 #include <linux/numa.h>
74 #include <linux/perf_event.h>
75 #include <linux/ptrace.h>
76 #include <linux/vmalloc.h>
78 #include <trace/events/kmem.h>
81 #include <asm/mmu_context.h>
82 #include <asm/pgalloc.h>
83 #include <linux/uaccess.h>
85 #include <asm/tlbflush.h>
87 #include "pgalloc-track.h"
90 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
91 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
94 #ifndef CONFIG_NEED_MULTIPLE_NODES
95 /* use the per-pgdat data instead for discontigmem - mbligh */
96 unsigned long max_mapnr;
97 EXPORT_SYMBOL(max_mapnr);
100 EXPORT_SYMBOL(mem_map);
104 * A number of key systems in x86 including ioremap() rely on the assumption
105 * that high_memory defines the upper bound on direct map memory, then end
106 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
107 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
111 EXPORT_SYMBOL(high_memory);
114 * Randomize the address space (stacks, mmaps, brk, etc.).
116 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
117 * as ancient (libc5 based) binaries can segfault. )
119 int randomize_va_space __read_mostly =
120 #ifdef CONFIG_COMPAT_BRK
126 #ifndef arch_faults_on_old_pte
127 static inline bool arch_faults_on_old_pte(void)
130 * Those arches which don't have hw access flag feature need to
131 * implement their own helper. By default, "true" means pagefault
132 * will be hit on old pte.
138 static int __init disable_randmaps(char *s)
140 randomize_va_space = 0;
143 __setup("norandmaps", disable_randmaps);
145 unsigned long zero_pfn __read_mostly;
146 EXPORT_SYMBOL(zero_pfn);
148 unsigned long highest_memmap_pfn __read_mostly;
151 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
153 static int __init init_zero_pfn(void)
155 zero_pfn = page_to_pfn(ZERO_PAGE(0));
158 core_initcall(init_zero_pfn);
160 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
162 trace_rss_stat(mm, member, count);
165 #if defined(SPLIT_RSS_COUNTING)
167 void sync_mm_rss(struct mm_struct *mm)
171 for (i = 0; i < NR_MM_COUNTERS; i++) {
172 if (current->rss_stat.count[i]) {
173 add_mm_counter(mm, i, current->rss_stat.count[i]);
174 current->rss_stat.count[i] = 0;
177 current->rss_stat.events = 0;
180 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
182 struct task_struct *task = current;
184 if (likely(task->mm == mm))
185 task->rss_stat.count[member] += val;
187 add_mm_counter(mm, member, val);
189 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
190 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
192 /* sync counter once per 64 page faults */
193 #define TASK_RSS_EVENTS_THRESH (64)
194 static void check_sync_rss_stat(struct task_struct *task)
196 if (unlikely(task != current))
198 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
199 sync_mm_rss(task->mm);
201 #else /* SPLIT_RSS_COUNTING */
203 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
204 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
206 static void check_sync_rss_stat(struct task_struct *task)
210 #endif /* SPLIT_RSS_COUNTING */
213 * Note: this doesn't free the actual pages themselves. That
214 * has been handled earlier when unmapping all the memory regions.
216 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
219 pgtable_t token = pmd_pgtable(*pmd);
221 pte_free_tlb(tlb, token, addr);
222 mm_dec_nr_ptes(tlb->mm);
225 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
226 unsigned long addr, unsigned long end,
227 unsigned long floor, unsigned long ceiling)
234 pmd = pmd_offset(pud, addr);
236 next = pmd_addr_end(addr, end);
237 if (pmd_none_or_clear_bad(pmd))
239 free_pte_range(tlb, pmd, addr);
240 } while (pmd++, addr = next, addr != end);
250 if (end - 1 > ceiling - 1)
253 pmd = pmd_offset(pud, start);
255 pmd_free_tlb(tlb, pmd, start);
256 mm_dec_nr_pmds(tlb->mm);
259 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
260 unsigned long addr, unsigned long end,
261 unsigned long floor, unsigned long ceiling)
268 pud = pud_offset(p4d, addr);
270 next = pud_addr_end(addr, end);
271 if (pud_none_or_clear_bad(pud))
273 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
274 } while (pud++, addr = next, addr != end);
284 if (end - 1 > ceiling - 1)
287 pud = pud_offset(p4d, start);
289 pud_free_tlb(tlb, pud, start);
290 mm_dec_nr_puds(tlb->mm);
293 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
294 unsigned long addr, unsigned long end,
295 unsigned long floor, unsigned long ceiling)
302 p4d = p4d_offset(pgd, addr);
304 next = p4d_addr_end(addr, end);
305 if (p4d_none_or_clear_bad(p4d))
307 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
308 } while (p4d++, addr = next, addr != end);
314 ceiling &= PGDIR_MASK;
318 if (end - 1 > ceiling - 1)
321 p4d = p4d_offset(pgd, start);
323 p4d_free_tlb(tlb, p4d, start);
327 * This function frees user-level page tables of a process.
329 void free_pgd_range(struct mmu_gather *tlb,
330 unsigned long addr, unsigned long end,
331 unsigned long floor, unsigned long ceiling)
337 * The next few lines have given us lots of grief...
339 * Why are we testing PMD* at this top level? Because often
340 * there will be no work to do at all, and we'd prefer not to
341 * go all the way down to the bottom just to discover that.
343 * Why all these "- 1"s? Because 0 represents both the bottom
344 * of the address space and the top of it (using -1 for the
345 * top wouldn't help much: the masks would do the wrong thing).
346 * The rule is that addr 0 and floor 0 refer to the bottom of
347 * the address space, but end 0 and ceiling 0 refer to the top
348 * Comparisons need to use "end - 1" and "ceiling - 1" (though
349 * that end 0 case should be mythical).
351 * Wherever addr is brought up or ceiling brought down, we must
352 * be careful to reject "the opposite 0" before it confuses the
353 * subsequent tests. But what about where end is brought down
354 * by PMD_SIZE below? no, end can't go down to 0 there.
356 * Whereas we round start (addr) and ceiling down, by different
357 * masks at different levels, in order to test whether a table
358 * now has no other vmas using it, so can be freed, we don't
359 * bother to round floor or end up - the tests don't need that.
373 if (end - 1 > ceiling - 1)
378 * We add page table cache pages with PAGE_SIZE,
379 * (see pte_free_tlb()), flush the tlb if we need
381 tlb_change_page_size(tlb, PAGE_SIZE);
382 pgd = pgd_offset(tlb->mm, addr);
384 next = pgd_addr_end(addr, end);
385 if (pgd_none_or_clear_bad(pgd))
387 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
388 } while (pgd++, addr = next, addr != end);
391 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
392 unsigned long floor, unsigned long ceiling)
395 struct vm_area_struct *next = vma->vm_next;
396 unsigned long addr = vma->vm_start;
399 * Hide vma from rmap and truncate_pagecache before freeing
402 unlink_anon_vmas(vma);
403 unlink_file_vma(vma);
405 if (is_vm_hugetlb_page(vma)) {
406 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
407 floor, next ? next->vm_start : ceiling);
410 * Optimization: gather nearby vmas into one call down
412 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
413 && !is_vm_hugetlb_page(next)) {
416 unlink_anon_vmas(vma);
417 unlink_file_vma(vma);
419 free_pgd_range(tlb, addr, vma->vm_end,
420 floor, next ? next->vm_start : ceiling);
426 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
429 pgtable_t new = pte_alloc_one(mm);
434 * Ensure all pte setup (eg. pte page lock and page clearing) are
435 * visible before the pte is made visible to other CPUs by being
436 * put into page tables.
438 * The other side of the story is the pointer chasing in the page
439 * table walking code (when walking the page table without locking;
440 * ie. most of the time). Fortunately, these data accesses consist
441 * of a chain of data-dependent loads, meaning most CPUs (alpha
442 * being the notable exception) will already guarantee loads are
443 * seen in-order. See the alpha page table accessors for the
444 * smp_rmb() barriers in page table walking code.
446 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
448 ptl = pmd_lock(mm, pmd);
449 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
451 pmd_populate(mm, pmd, new);
460 int __pte_alloc_kernel(pmd_t *pmd)
462 pte_t *new = pte_alloc_one_kernel(&init_mm);
466 smp_wmb(); /* See comment in __pte_alloc */
468 spin_lock(&init_mm.page_table_lock);
469 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
470 pmd_populate_kernel(&init_mm, pmd, new);
473 spin_unlock(&init_mm.page_table_lock);
475 pte_free_kernel(&init_mm, new);
479 static inline void init_rss_vec(int *rss)
481 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
484 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
488 if (current->mm == mm)
490 for (i = 0; i < NR_MM_COUNTERS; i++)
492 add_mm_counter(mm, i, rss[i]);
496 * This function is called to print an error when a bad pte
497 * is found. For example, we might have a PFN-mapped pte in
498 * a region that doesn't allow it.
500 * The calling function must still handle the error.
502 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
503 pte_t pte, struct page *page)
505 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
506 p4d_t *p4d = p4d_offset(pgd, addr);
507 pud_t *pud = pud_offset(p4d, addr);
508 pmd_t *pmd = pmd_offset(pud, addr);
509 struct address_space *mapping;
511 static unsigned long resume;
512 static unsigned long nr_shown;
513 static unsigned long nr_unshown;
516 * Allow a burst of 60 reports, then keep quiet for that minute;
517 * or allow a steady drip of one report per second.
519 if (nr_shown == 60) {
520 if (time_before(jiffies, resume)) {
525 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
532 resume = jiffies + 60 * HZ;
534 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
535 index = linear_page_index(vma, addr);
537 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
539 (long long)pte_val(pte), (long long)pmd_val(*pmd));
541 dump_page(page, "bad pte");
542 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
543 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
544 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
546 vma->vm_ops ? vma->vm_ops->fault : NULL,
547 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
548 mapping ? mapping->a_ops->readpage : NULL);
550 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
554 * vm_normal_page -- This function gets the "struct page" associated with a pte.
556 * "Special" mappings do not wish to be associated with a "struct page" (either
557 * it doesn't exist, or it exists but they don't want to touch it). In this
558 * case, NULL is returned here. "Normal" mappings do have a struct page.
560 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
561 * pte bit, in which case this function is trivial. Secondly, an architecture
562 * may not have a spare pte bit, which requires a more complicated scheme,
565 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
566 * special mapping (even if there are underlying and valid "struct pages").
567 * COWed pages of a VM_PFNMAP are always normal.
569 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
570 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
571 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
572 * mapping will always honor the rule
574 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
576 * And for normal mappings this is false.
578 * This restricts such mappings to be a linear translation from virtual address
579 * to pfn. To get around this restriction, we allow arbitrary mappings so long
580 * as the vma is not a COW mapping; in that case, we know that all ptes are
581 * special (because none can have been COWed).
584 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
586 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
587 * page" backing, however the difference is that _all_ pages with a struct
588 * page (that is, those where pfn_valid is true) are refcounted and considered
589 * normal pages by the VM. The disadvantage is that pages are refcounted
590 * (which can be slower and simply not an option for some PFNMAP users). The
591 * advantage is that we don't have to follow the strict linearity rule of
592 * PFNMAP mappings in order to support COWable mappings.
595 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
598 unsigned long pfn = pte_pfn(pte);
600 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
601 if (likely(!pte_special(pte)))
603 if (vma->vm_ops && vma->vm_ops->find_special_page)
604 return vma->vm_ops->find_special_page(vma, addr);
605 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
607 if (is_zero_pfn(pfn))
612 print_bad_pte(vma, addr, pte, NULL);
616 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
618 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
619 if (vma->vm_flags & VM_MIXEDMAP) {
625 off = (addr - vma->vm_start) >> PAGE_SHIFT;
626 if (pfn == vma->vm_pgoff + off)
628 if (!is_cow_mapping(vma->vm_flags))
633 if (is_zero_pfn(pfn))
637 if (unlikely(pfn > highest_memmap_pfn)) {
638 print_bad_pte(vma, addr, pte, NULL);
643 * NOTE! We still have PageReserved() pages in the page tables.
644 * eg. VDSO mappings can cause them to exist.
647 return pfn_to_page(pfn);
650 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
651 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
654 unsigned long pfn = pmd_pfn(pmd);
657 * There is no pmd_special() but there may be special pmds, e.g.
658 * in a direct-access (dax) mapping, so let's just replicate the
659 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
661 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
662 if (vma->vm_flags & VM_MIXEDMAP) {
668 off = (addr - vma->vm_start) >> PAGE_SHIFT;
669 if (pfn == vma->vm_pgoff + off)
671 if (!is_cow_mapping(vma->vm_flags))
678 if (is_huge_zero_pmd(pmd))
680 if (unlikely(pfn > highest_memmap_pfn))
684 * NOTE! We still have PageReserved() pages in the page tables.
685 * eg. VDSO mappings can cause them to exist.
688 return pfn_to_page(pfn);
693 * copy one vm_area from one task to the other. Assumes the page tables
694 * already present in the new task to be cleared in the whole range
695 * covered by this vma.
699 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
700 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
701 unsigned long addr, int *rss)
703 unsigned long vm_flags = vma->vm_flags;
704 pte_t pte = *src_pte;
706 swp_entry_t entry = pte_to_swp_entry(pte);
708 if (likely(!non_swap_entry(entry))) {
709 if (swap_duplicate(entry) < 0)
712 /* make sure dst_mm is on swapoff's mmlist. */
713 if (unlikely(list_empty(&dst_mm->mmlist))) {
714 spin_lock(&mmlist_lock);
715 if (list_empty(&dst_mm->mmlist))
716 list_add(&dst_mm->mmlist,
718 spin_unlock(&mmlist_lock);
721 } else if (is_migration_entry(entry)) {
722 page = migration_entry_to_page(entry);
724 rss[mm_counter(page)]++;
726 if (is_write_migration_entry(entry) &&
727 is_cow_mapping(vm_flags)) {
729 * COW mappings require pages in both
730 * parent and child to be set to read.
732 make_migration_entry_read(&entry);
733 pte = swp_entry_to_pte(entry);
734 if (pte_swp_soft_dirty(*src_pte))
735 pte = pte_swp_mksoft_dirty(pte);
736 if (pte_swp_uffd_wp(*src_pte))
737 pte = pte_swp_mkuffd_wp(pte);
738 set_pte_at(src_mm, addr, src_pte, pte);
740 } else if (is_device_private_entry(entry)) {
741 page = device_private_entry_to_page(entry);
744 * Update rss count even for unaddressable pages, as
745 * they should treated just like normal pages in this
748 * We will likely want to have some new rss counters
749 * for unaddressable pages, at some point. But for now
750 * keep things as they are.
753 rss[mm_counter(page)]++;
754 page_dup_rmap(page, false);
757 * We do not preserve soft-dirty information, because so
758 * far, checkpoint/restore is the only feature that
759 * requires that. And checkpoint/restore does not work
760 * when a device driver is involved (you cannot easily
761 * save and restore device driver state).
763 if (is_write_device_private_entry(entry) &&
764 is_cow_mapping(vm_flags)) {
765 make_device_private_entry_read(&entry);
766 pte = swp_entry_to_pte(entry);
767 if (pte_swp_uffd_wp(*src_pte))
768 pte = pte_swp_mkuffd_wp(pte);
769 set_pte_at(src_mm, addr, src_pte, pte);
772 set_pte_at(dst_mm, addr, dst_pte, pte);
777 copy_present_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
778 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
779 unsigned long addr, int *rss)
781 unsigned long vm_flags = vma->vm_flags;
782 pte_t pte = *src_pte;
786 * If it's a COW mapping, write protect it both
787 * in the parent and the child
789 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
790 ptep_set_wrprotect(src_mm, addr, src_pte);
791 pte = pte_wrprotect(pte);
795 * If it's a shared mapping, mark it clean in
798 if (vm_flags & VM_SHARED)
799 pte = pte_mkclean(pte);
800 pte = pte_mkold(pte);
803 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
804 * does not have the VM_UFFD_WP, which means that the uffd
805 * fork event is not enabled.
807 if (!(vm_flags & VM_UFFD_WP))
808 pte = pte_clear_uffd_wp(pte);
810 page = vm_normal_page(vma, addr, pte);
813 page_dup_rmap(page, false);
814 rss[mm_counter(page)]++;
817 set_pte_at(dst_mm, addr, dst_pte, pte);
820 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
821 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
822 struct vm_area_struct *new,
823 unsigned long addr, unsigned long end)
825 pte_t *orig_src_pte, *orig_dst_pte;
826 pte_t *src_pte, *dst_pte;
827 spinlock_t *src_ptl, *dst_ptl;
829 int rss[NR_MM_COUNTERS];
830 swp_entry_t entry = (swp_entry_t){0};
835 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
838 src_pte = pte_offset_map(src_pmd, addr);
839 src_ptl = pte_lockptr(src_mm, src_pmd);
840 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
841 orig_src_pte = src_pte;
842 orig_dst_pte = dst_pte;
843 arch_enter_lazy_mmu_mode();
847 * We are holding two locks at this point - either of them
848 * could generate latencies in another task on another CPU.
850 if (progress >= 32) {
852 if (need_resched() ||
853 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
856 if (pte_none(*src_pte)) {
860 if (unlikely(!pte_present(*src_pte))) {
861 entry.val = copy_nonpresent_pte(dst_mm, src_mm,
869 copy_present_pte(dst_mm, src_mm, dst_pte, src_pte,
872 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
874 arch_leave_lazy_mmu_mode();
875 spin_unlock(src_ptl);
876 pte_unmap(orig_src_pte);
877 add_mm_rss_vec(dst_mm, rss);
878 pte_unmap_unlock(orig_dst_pte, dst_ptl);
882 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
891 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
892 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
893 struct vm_area_struct *new,
894 unsigned long addr, unsigned long end)
896 pmd_t *src_pmd, *dst_pmd;
899 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
902 src_pmd = pmd_offset(src_pud, addr);
904 next = pmd_addr_end(addr, end);
905 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
906 || pmd_devmap(*src_pmd)) {
908 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
909 err = copy_huge_pmd(dst_mm, src_mm,
910 dst_pmd, src_pmd, addr, vma);
917 if (pmd_none_or_clear_bad(src_pmd))
919 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
920 vma, new, addr, next))
922 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
926 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
927 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
928 struct vm_area_struct *new,
929 unsigned long addr, unsigned long end)
931 pud_t *src_pud, *dst_pud;
934 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
937 src_pud = pud_offset(src_p4d, addr);
939 next = pud_addr_end(addr, end);
940 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
943 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
944 err = copy_huge_pud(dst_mm, src_mm,
945 dst_pud, src_pud, addr, vma);
952 if (pud_none_or_clear_bad(src_pud))
954 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
955 vma, new, addr, next))
957 } while (dst_pud++, src_pud++, addr = next, addr != end);
961 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
962 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
963 struct vm_area_struct *new,
964 unsigned long addr, unsigned long end)
966 p4d_t *src_p4d, *dst_p4d;
969 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
972 src_p4d = p4d_offset(src_pgd, addr);
974 next = p4d_addr_end(addr, end);
975 if (p4d_none_or_clear_bad(src_p4d))
977 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
978 vma, new, addr, next))
980 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
984 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
985 struct vm_area_struct *vma, struct vm_area_struct *new)
987 pgd_t *src_pgd, *dst_pgd;
989 unsigned long addr = vma->vm_start;
990 unsigned long end = vma->vm_end;
991 struct mmu_notifier_range range;
996 * Don't copy ptes where a page fault will fill them correctly.
997 * Fork becomes much lighter when there are big shared or private
998 * readonly mappings. The tradeoff is that copy_page_range is more
999 * efficient than faulting.
1001 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1005 if (is_vm_hugetlb_page(vma))
1006 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1008 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1010 * We do not free on error cases below as remove_vma
1011 * gets called on error from higher level routine
1013 ret = track_pfn_copy(vma);
1019 * We need to invalidate the secondary MMU mappings only when
1020 * there could be a permission downgrade on the ptes of the
1021 * parent mm. And a permission downgrade will only happen if
1022 * is_cow_mapping() returns true.
1024 is_cow = is_cow_mapping(vma->vm_flags);
1027 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1028 0, vma, src_mm, addr, end);
1029 mmu_notifier_invalidate_range_start(&range);
1033 dst_pgd = pgd_offset(dst_mm, addr);
1034 src_pgd = pgd_offset(src_mm, addr);
1036 next = pgd_addr_end(addr, end);
1037 if (pgd_none_or_clear_bad(src_pgd))
1039 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1040 vma, new, addr, next))) {
1044 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1047 mmu_notifier_invalidate_range_end(&range);
1051 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1052 struct vm_area_struct *vma, pmd_t *pmd,
1053 unsigned long addr, unsigned long end,
1054 struct zap_details *details)
1056 struct mm_struct *mm = tlb->mm;
1057 int force_flush = 0;
1058 int rss[NR_MM_COUNTERS];
1064 tlb_change_page_size(tlb, PAGE_SIZE);
1067 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1069 flush_tlb_batched_pending(mm);
1070 arch_enter_lazy_mmu_mode();
1073 if (pte_none(ptent))
1079 if (pte_present(ptent)) {
1082 page = vm_normal_page(vma, addr, ptent);
1083 if (unlikely(details) && page) {
1085 * unmap_shared_mapping_pages() wants to
1086 * invalidate cache without truncating:
1087 * unmap shared but keep private pages.
1089 if (details->check_mapping &&
1090 details->check_mapping != page_rmapping(page))
1093 ptent = ptep_get_and_clear_full(mm, addr, pte,
1095 tlb_remove_tlb_entry(tlb, pte, addr);
1096 if (unlikely(!page))
1099 if (!PageAnon(page)) {
1100 if (pte_dirty(ptent)) {
1102 set_page_dirty(page);
1104 if (pte_young(ptent) &&
1105 likely(!(vma->vm_flags & VM_SEQ_READ)))
1106 mark_page_accessed(page);
1108 rss[mm_counter(page)]--;
1109 page_remove_rmap(page, false);
1110 if (unlikely(page_mapcount(page) < 0))
1111 print_bad_pte(vma, addr, ptent, page);
1112 if (unlikely(__tlb_remove_page(tlb, page))) {
1120 entry = pte_to_swp_entry(ptent);
1121 if (is_device_private_entry(entry)) {
1122 struct page *page = device_private_entry_to_page(entry);
1124 if (unlikely(details && details->check_mapping)) {
1126 * unmap_shared_mapping_pages() wants to
1127 * invalidate cache without truncating:
1128 * unmap shared but keep private pages.
1130 if (details->check_mapping !=
1131 page_rmapping(page))
1135 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1136 rss[mm_counter(page)]--;
1137 page_remove_rmap(page, false);
1142 /* If details->check_mapping, we leave swap entries. */
1143 if (unlikely(details))
1146 if (!non_swap_entry(entry))
1148 else if (is_migration_entry(entry)) {
1151 page = migration_entry_to_page(entry);
1152 rss[mm_counter(page)]--;
1154 if (unlikely(!free_swap_and_cache(entry)))
1155 print_bad_pte(vma, addr, ptent, NULL);
1156 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1157 } while (pte++, addr += PAGE_SIZE, addr != end);
1159 add_mm_rss_vec(mm, rss);
1160 arch_leave_lazy_mmu_mode();
1162 /* Do the actual TLB flush before dropping ptl */
1164 tlb_flush_mmu_tlbonly(tlb);
1165 pte_unmap_unlock(start_pte, ptl);
1168 * If we forced a TLB flush (either due to running out of
1169 * batch buffers or because we needed to flush dirty TLB
1170 * entries before releasing the ptl), free the batched
1171 * memory too. Restart if we didn't do everything.
1186 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1187 struct vm_area_struct *vma, pud_t *pud,
1188 unsigned long addr, unsigned long end,
1189 struct zap_details *details)
1194 pmd = pmd_offset(pud, addr);
1196 next = pmd_addr_end(addr, end);
1197 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1198 if (next - addr != HPAGE_PMD_SIZE)
1199 __split_huge_pmd(vma, pmd, addr, false, NULL);
1200 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1205 * Here there can be other concurrent MADV_DONTNEED or
1206 * trans huge page faults running, and if the pmd is
1207 * none or trans huge it can change under us. This is
1208 * because MADV_DONTNEED holds the mmap_lock in read
1211 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1213 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1216 } while (pmd++, addr = next, addr != end);
1221 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1222 struct vm_area_struct *vma, p4d_t *p4d,
1223 unsigned long addr, unsigned long end,
1224 struct zap_details *details)
1229 pud = pud_offset(p4d, addr);
1231 next = pud_addr_end(addr, end);
1232 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1233 if (next - addr != HPAGE_PUD_SIZE) {
1234 mmap_assert_locked(tlb->mm);
1235 split_huge_pud(vma, pud, addr);
1236 } else if (zap_huge_pud(tlb, vma, pud, addr))
1240 if (pud_none_or_clear_bad(pud))
1242 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1245 } while (pud++, addr = next, addr != end);
1250 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1251 struct vm_area_struct *vma, pgd_t *pgd,
1252 unsigned long addr, unsigned long end,
1253 struct zap_details *details)
1258 p4d = p4d_offset(pgd, addr);
1260 next = p4d_addr_end(addr, end);
1261 if (p4d_none_or_clear_bad(p4d))
1263 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1264 } while (p4d++, addr = next, addr != end);
1269 void unmap_page_range(struct mmu_gather *tlb,
1270 struct vm_area_struct *vma,
1271 unsigned long addr, unsigned long end,
1272 struct zap_details *details)
1277 BUG_ON(addr >= end);
1278 tlb_start_vma(tlb, vma);
1279 pgd = pgd_offset(vma->vm_mm, addr);
1281 next = pgd_addr_end(addr, end);
1282 if (pgd_none_or_clear_bad(pgd))
1284 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1285 } while (pgd++, addr = next, addr != end);
1286 tlb_end_vma(tlb, vma);
1290 static void unmap_single_vma(struct mmu_gather *tlb,
1291 struct vm_area_struct *vma, unsigned long start_addr,
1292 unsigned long end_addr,
1293 struct zap_details *details)
1295 unsigned long start = max(vma->vm_start, start_addr);
1298 if (start >= vma->vm_end)
1300 end = min(vma->vm_end, end_addr);
1301 if (end <= vma->vm_start)
1305 uprobe_munmap(vma, start, end);
1307 if (unlikely(vma->vm_flags & VM_PFNMAP))
1308 untrack_pfn(vma, 0, 0);
1311 if (unlikely(is_vm_hugetlb_page(vma))) {
1313 * It is undesirable to test vma->vm_file as it
1314 * should be non-null for valid hugetlb area.
1315 * However, vm_file will be NULL in the error
1316 * cleanup path of mmap_region. When
1317 * hugetlbfs ->mmap method fails,
1318 * mmap_region() nullifies vma->vm_file
1319 * before calling this function to clean up.
1320 * Since no pte has actually been setup, it is
1321 * safe to do nothing in this case.
1324 i_mmap_lock_write(vma->vm_file->f_mapping);
1325 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1326 i_mmap_unlock_write(vma->vm_file->f_mapping);
1329 unmap_page_range(tlb, vma, start, end, details);
1334 * unmap_vmas - unmap a range of memory covered by a list of vma's
1335 * @tlb: address of the caller's struct mmu_gather
1336 * @vma: the starting vma
1337 * @start_addr: virtual address at which to start unmapping
1338 * @end_addr: virtual address at which to end unmapping
1340 * Unmap all pages in the vma list.
1342 * Only addresses between `start' and `end' will be unmapped.
1344 * The VMA list must be sorted in ascending virtual address order.
1346 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1347 * range after unmap_vmas() returns. So the only responsibility here is to
1348 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1349 * drops the lock and schedules.
1351 void unmap_vmas(struct mmu_gather *tlb,
1352 struct vm_area_struct *vma, unsigned long start_addr,
1353 unsigned long end_addr)
1355 struct mmu_notifier_range range;
1357 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1358 start_addr, end_addr);
1359 mmu_notifier_invalidate_range_start(&range);
1360 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1361 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1362 mmu_notifier_invalidate_range_end(&range);
1366 * zap_page_range - remove user pages in a given range
1367 * @vma: vm_area_struct holding the applicable pages
1368 * @start: starting address of pages to zap
1369 * @size: number of bytes to zap
1371 * Caller must protect the VMA list
1373 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1376 struct mmu_notifier_range range;
1377 struct mmu_gather tlb;
1380 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1381 start, start + size);
1382 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1383 update_hiwater_rss(vma->vm_mm);
1384 mmu_notifier_invalidate_range_start(&range);
1385 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1386 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1387 mmu_notifier_invalidate_range_end(&range);
1388 tlb_finish_mmu(&tlb, start, range.end);
1392 * zap_page_range_single - remove user pages in a given range
1393 * @vma: vm_area_struct holding the applicable pages
1394 * @address: starting address of pages to zap
1395 * @size: number of bytes to zap
1396 * @details: details of shared cache invalidation
1398 * The range must fit into one VMA.
1400 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1401 unsigned long size, struct zap_details *details)
1403 struct mmu_notifier_range range;
1404 struct mmu_gather tlb;
1407 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1408 address, address + size);
1409 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1410 update_hiwater_rss(vma->vm_mm);
1411 mmu_notifier_invalidate_range_start(&range);
1412 unmap_single_vma(&tlb, vma, address, range.end, details);
1413 mmu_notifier_invalidate_range_end(&range);
1414 tlb_finish_mmu(&tlb, address, range.end);
1418 * zap_vma_ptes - remove ptes mapping the vma
1419 * @vma: vm_area_struct holding ptes to be zapped
1420 * @address: starting address of pages to zap
1421 * @size: number of bytes to zap
1423 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1425 * The entire address range must be fully contained within the vma.
1428 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1431 if (address < vma->vm_start || address + size > vma->vm_end ||
1432 !(vma->vm_flags & VM_PFNMAP))
1435 zap_page_range_single(vma, address, size, NULL);
1437 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1439 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1446 pgd = pgd_offset(mm, addr);
1447 p4d = p4d_alloc(mm, pgd, addr);
1450 pud = pud_alloc(mm, p4d, addr);
1453 pmd = pmd_alloc(mm, pud, addr);
1457 VM_BUG_ON(pmd_trans_huge(*pmd));
1461 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1464 pmd_t *pmd = walk_to_pmd(mm, addr);
1468 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1471 static int validate_page_before_insert(struct page *page)
1473 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1475 flush_dcache_page(page);
1479 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1480 unsigned long addr, struct page *page, pgprot_t prot)
1482 if (!pte_none(*pte))
1484 /* Ok, finally just insert the thing.. */
1486 inc_mm_counter_fast(mm, mm_counter_file(page));
1487 page_add_file_rmap(page, false);
1488 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1493 * This is the old fallback for page remapping.
1495 * For historical reasons, it only allows reserved pages. Only
1496 * old drivers should use this, and they needed to mark their
1497 * pages reserved for the old functions anyway.
1499 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1500 struct page *page, pgprot_t prot)
1502 struct mm_struct *mm = vma->vm_mm;
1507 retval = validate_page_before_insert(page);
1511 pte = get_locked_pte(mm, addr, &ptl);
1514 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1515 pte_unmap_unlock(pte, ptl);
1521 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1522 unsigned long addr, struct page *page, pgprot_t prot)
1526 if (!page_count(page))
1528 err = validate_page_before_insert(page);
1531 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1534 /* insert_pages() amortizes the cost of spinlock operations
1535 * when inserting pages in a loop. Arch *must* define pte_index.
1537 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1538 struct page **pages, unsigned long *num, pgprot_t prot)
1541 pte_t *start_pte, *pte;
1542 spinlock_t *pte_lock;
1543 struct mm_struct *const mm = vma->vm_mm;
1544 unsigned long curr_page_idx = 0;
1545 unsigned long remaining_pages_total = *num;
1546 unsigned long pages_to_write_in_pmd;
1550 pmd = walk_to_pmd(mm, addr);
1554 pages_to_write_in_pmd = min_t(unsigned long,
1555 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1557 /* Allocate the PTE if necessary; takes PMD lock once only. */
1559 if (pte_alloc(mm, pmd))
1562 while (pages_to_write_in_pmd) {
1564 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1566 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1567 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1568 int err = insert_page_in_batch_locked(mm, pte,
1569 addr, pages[curr_page_idx], prot);
1570 if (unlikely(err)) {
1571 pte_unmap_unlock(start_pte, pte_lock);
1573 remaining_pages_total -= pte_idx;
1579 pte_unmap_unlock(start_pte, pte_lock);
1580 pages_to_write_in_pmd -= batch_size;
1581 remaining_pages_total -= batch_size;
1583 if (remaining_pages_total)
1587 *num = remaining_pages_total;
1590 #endif /* ifdef pte_index */
1593 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1594 * @vma: user vma to map to
1595 * @addr: target start user address of these pages
1596 * @pages: source kernel pages
1597 * @num: in: number of pages to map. out: number of pages that were *not*
1598 * mapped. (0 means all pages were successfully mapped).
1600 * Preferred over vm_insert_page() when inserting multiple pages.
1602 * In case of error, we may have mapped a subset of the provided
1603 * pages. It is the caller's responsibility to account for this case.
1605 * The same restrictions apply as in vm_insert_page().
1607 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1608 struct page **pages, unsigned long *num)
1611 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1613 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1615 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1616 BUG_ON(mmap_read_trylock(vma->vm_mm));
1617 BUG_ON(vma->vm_flags & VM_PFNMAP);
1618 vma->vm_flags |= VM_MIXEDMAP;
1620 /* Defer page refcount checking till we're about to map that page. */
1621 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1623 unsigned long idx = 0, pgcount = *num;
1626 for (; idx < pgcount; ++idx) {
1627 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1631 *num = pgcount - idx;
1633 #endif /* ifdef pte_index */
1635 EXPORT_SYMBOL(vm_insert_pages);
1638 * vm_insert_page - insert single page into user vma
1639 * @vma: user vma to map to
1640 * @addr: target user address of this page
1641 * @page: source kernel page
1643 * This allows drivers to insert individual pages they've allocated
1646 * The page has to be a nice clean _individual_ kernel allocation.
1647 * If you allocate a compound page, you need to have marked it as
1648 * such (__GFP_COMP), or manually just split the page up yourself
1649 * (see split_page()).
1651 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1652 * took an arbitrary page protection parameter. This doesn't allow
1653 * that. Your vma protection will have to be set up correctly, which
1654 * means that if you want a shared writable mapping, you'd better
1655 * ask for a shared writable mapping!
1657 * The page does not need to be reserved.
1659 * Usually this function is called from f_op->mmap() handler
1660 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1661 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1662 * function from other places, for example from page-fault handler.
1664 * Return: %0 on success, negative error code otherwise.
1666 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1669 if (addr < vma->vm_start || addr >= vma->vm_end)
1671 if (!page_count(page))
1673 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1674 BUG_ON(mmap_read_trylock(vma->vm_mm));
1675 BUG_ON(vma->vm_flags & VM_PFNMAP);
1676 vma->vm_flags |= VM_MIXEDMAP;
1678 return insert_page(vma, addr, page, vma->vm_page_prot);
1680 EXPORT_SYMBOL(vm_insert_page);
1683 * __vm_map_pages - maps range of kernel pages into user vma
1684 * @vma: user vma to map to
1685 * @pages: pointer to array of source kernel pages
1686 * @num: number of pages in page array
1687 * @offset: user's requested vm_pgoff
1689 * This allows drivers to map range of kernel pages into a user vma.
1691 * Return: 0 on success and error code otherwise.
1693 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1694 unsigned long num, unsigned long offset)
1696 unsigned long count = vma_pages(vma);
1697 unsigned long uaddr = vma->vm_start;
1700 /* Fail if the user requested offset is beyond the end of the object */
1704 /* Fail if the user requested size exceeds available object size */
1705 if (count > num - offset)
1708 for (i = 0; i < count; i++) {
1709 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1719 * vm_map_pages - maps range of kernel pages starts with non zero offset
1720 * @vma: user vma to map to
1721 * @pages: pointer to array of source kernel pages
1722 * @num: number of pages in page array
1724 * Maps an object consisting of @num pages, catering for the user's
1725 * requested vm_pgoff
1727 * If we fail to insert any page into the vma, the function will return
1728 * immediately leaving any previously inserted pages present. Callers
1729 * from the mmap handler may immediately return the error as their caller
1730 * will destroy the vma, removing any successfully inserted pages. Other
1731 * callers should make their own arrangements for calling unmap_region().
1733 * Context: Process context. Called by mmap handlers.
1734 * Return: 0 on success and error code otherwise.
1736 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1739 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1741 EXPORT_SYMBOL(vm_map_pages);
1744 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1745 * @vma: user vma to map to
1746 * @pages: pointer to array of source kernel pages
1747 * @num: number of pages in page array
1749 * Similar to vm_map_pages(), except that it explicitly sets the offset
1750 * to 0. This function is intended for the drivers that did not consider
1753 * Context: Process context. Called by mmap handlers.
1754 * Return: 0 on success and error code otherwise.
1756 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1759 return __vm_map_pages(vma, pages, num, 0);
1761 EXPORT_SYMBOL(vm_map_pages_zero);
1763 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1764 pfn_t pfn, pgprot_t prot, bool mkwrite)
1766 struct mm_struct *mm = vma->vm_mm;
1770 pte = get_locked_pte(mm, addr, &ptl);
1772 return VM_FAULT_OOM;
1773 if (!pte_none(*pte)) {
1776 * For read faults on private mappings the PFN passed
1777 * in may not match the PFN we have mapped if the
1778 * mapped PFN is a writeable COW page. In the mkwrite
1779 * case we are creating a writable PTE for a shared
1780 * mapping and we expect the PFNs to match. If they
1781 * don't match, we are likely racing with block
1782 * allocation and mapping invalidation so just skip the
1785 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1786 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1789 entry = pte_mkyoung(*pte);
1790 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1791 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1792 update_mmu_cache(vma, addr, pte);
1797 /* Ok, finally just insert the thing.. */
1798 if (pfn_t_devmap(pfn))
1799 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1801 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1804 entry = pte_mkyoung(entry);
1805 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1808 set_pte_at(mm, addr, pte, entry);
1809 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1812 pte_unmap_unlock(pte, ptl);
1813 return VM_FAULT_NOPAGE;
1817 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1818 * @vma: user vma to map to
1819 * @addr: target user address of this page
1820 * @pfn: source kernel pfn
1821 * @pgprot: pgprot flags for the inserted page
1823 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1824 * to override pgprot on a per-page basis.
1826 * This only makes sense for IO mappings, and it makes no sense for
1827 * COW mappings. In general, using multiple vmas is preferable;
1828 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1831 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1832 * a value of @pgprot different from that of @vma->vm_page_prot.
1834 * Context: Process context. May allocate using %GFP_KERNEL.
1835 * Return: vm_fault_t value.
1837 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1838 unsigned long pfn, pgprot_t pgprot)
1841 * Technically, architectures with pte_special can avoid all these
1842 * restrictions (same for remap_pfn_range). However we would like
1843 * consistency in testing and feature parity among all, so we should
1844 * try to keep these invariants in place for everybody.
1846 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1847 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1848 (VM_PFNMAP|VM_MIXEDMAP));
1849 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1850 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1852 if (addr < vma->vm_start || addr >= vma->vm_end)
1853 return VM_FAULT_SIGBUS;
1855 if (!pfn_modify_allowed(pfn, pgprot))
1856 return VM_FAULT_SIGBUS;
1858 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1860 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1863 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1866 * vmf_insert_pfn - insert single pfn into user vma
1867 * @vma: user vma to map to
1868 * @addr: target user address of this page
1869 * @pfn: source kernel pfn
1871 * Similar to vm_insert_page, this allows drivers to insert individual pages
1872 * they've allocated into a user vma. Same comments apply.
1874 * This function should only be called from a vm_ops->fault handler, and
1875 * in that case the handler should return the result of this function.
1877 * vma cannot be a COW mapping.
1879 * As this is called only for pages that do not currently exist, we
1880 * do not need to flush old virtual caches or the TLB.
1882 * Context: Process context. May allocate using %GFP_KERNEL.
1883 * Return: vm_fault_t value.
1885 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1888 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1890 EXPORT_SYMBOL(vmf_insert_pfn);
1892 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1894 /* these checks mirror the abort conditions in vm_normal_page */
1895 if (vma->vm_flags & VM_MIXEDMAP)
1897 if (pfn_t_devmap(pfn))
1899 if (pfn_t_special(pfn))
1901 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1906 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1907 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
1912 BUG_ON(!vm_mixed_ok(vma, pfn));
1914 if (addr < vma->vm_start || addr >= vma->vm_end)
1915 return VM_FAULT_SIGBUS;
1917 track_pfn_insert(vma, &pgprot, pfn);
1919 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1920 return VM_FAULT_SIGBUS;
1923 * If we don't have pte special, then we have to use the pfn_valid()
1924 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1925 * refcount the page if pfn_valid is true (hence insert_page rather
1926 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1927 * without pte special, it would there be refcounted as a normal page.
1929 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1930 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1934 * At this point we are committed to insert_page()
1935 * regardless of whether the caller specified flags that
1936 * result in pfn_t_has_page() == false.
1938 page = pfn_to_page(pfn_t_to_pfn(pfn));
1939 err = insert_page(vma, addr, page, pgprot);
1941 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1945 return VM_FAULT_OOM;
1946 if (err < 0 && err != -EBUSY)
1947 return VM_FAULT_SIGBUS;
1949 return VM_FAULT_NOPAGE;
1953 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
1954 * @vma: user vma to map to
1955 * @addr: target user address of this page
1956 * @pfn: source kernel pfn
1957 * @pgprot: pgprot flags for the inserted page
1959 * This is exactly like vmf_insert_mixed(), except that it allows drivers
1960 * to override pgprot on a per-page basis.
1962 * Typically this function should be used by drivers to set caching- and
1963 * encryption bits different than those of @vma->vm_page_prot, because
1964 * the caching- or encryption mode may not be known at mmap() time.
1965 * This is ok as long as @vma->vm_page_prot is not used by the core vm
1966 * to set caching and encryption bits for those vmas (except for COW pages).
1967 * This is ensured by core vm only modifying these page table entries using
1968 * functions that don't touch caching- or encryption bits, using pte_modify()
1969 * if needed. (See for example mprotect()).
1970 * Also when new page-table entries are created, this is only done using the
1971 * fault() callback, and never using the value of vma->vm_page_prot,
1972 * except for page-table entries that point to anonymous pages as the result
1975 * Context: Process context. May allocate using %GFP_KERNEL.
1976 * Return: vm_fault_t value.
1978 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
1979 pfn_t pfn, pgprot_t pgprot)
1981 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
1983 EXPORT_SYMBOL(vmf_insert_mixed_prot);
1985 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1988 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
1990 EXPORT_SYMBOL(vmf_insert_mixed);
1993 * If the insertion of PTE failed because someone else already added a
1994 * different entry in the mean time, we treat that as success as we assume
1995 * the same entry was actually inserted.
1997 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1998 unsigned long addr, pfn_t pfn)
2000 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2002 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2005 * maps a range of physical memory into the requested pages. the old
2006 * mappings are removed. any references to nonexistent pages results
2007 * in null mappings (currently treated as "copy-on-access")
2009 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2010 unsigned long addr, unsigned long end,
2011 unsigned long pfn, pgprot_t prot)
2017 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2020 arch_enter_lazy_mmu_mode();
2022 BUG_ON(!pte_none(*pte));
2023 if (!pfn_modify_allowed(pfn, prot)) {
2027 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2029 } while (pte++, addr += PAGE_SIZE, addr != end);
2030 arch_leave_lazy_mmu_mode();
2031 pte_unmap_unlock(pte - 1, ptl);
2035 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2036 unsigned long addr, unsigned long end,
2037 unsigned long pfn, pgprot_t prot)
2043 pfn -= addr >> PAGE_SHIFT;
2044 pmd = pmd_alloc(mm, pud, addr);
2047 VM_BUG_ON(pmd_trans_huge(*pmd));
2049 next = pmd_addr_end(addr, end);
2050 err = remap_pte_range(mm, pmd, addr, next,
2051 pfn + (addr >> PAGE_SHIFT), prot);
2054 } while (pmd++, addr = next, addr != end);
2058 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2059 unsigned long addr, unsigned long end,
2060 unsigned long pfn, pgprot_t prot)
2066 pfn -= addr >> PAGE_SHIFT;
2067 pud = pud_alloc(mm, p4d, addr);
2071 next = pud_addr_end(addr, end);
2072 err = remap_pmd_range(mm, pud, addr, next,
2073 pfn + (addr >> PAGE_SHIFT), prot);
2076 } while (pud++, addr = next, addr != end);
2080 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2081 unsigned long addr, unsigned long end,
2082 unsigned long pfn, pgprot_t prot)
2088 pfn -= addr >> PAGE_SHIFT;
2089 p4d = p4d_alloc(mm, pgd, addr);
2093 next = p4d_addr_end(addr, end);
2094 err = remap_pud_range(mm, p4d, addr, next,
2095 pfn + (addr >> PAGE_SHIFT), prot);
2098 } while (p4d++, addr = next, addr != end);
2103 * remap_pfn_range - remap kernel memory to userspace
2104 * @vma: user vma to map to
2105 * @addr: target page aligned user address to start at
2106 * @pfn: page frame number of kernel physical memory address
2107 * @size: size of mapping area
2108 * @prot: page protection flags for this mapping
2110 * Note: this is only safe if the mm semaphore is held when called.
2112 * Return: %0 on success, negative error code otherwise.
2114 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2115 unsigned long pfn, unsigned long size, pgprot_t prot)
2119 unsigned long end = addr + PAGE_ALIGN(size);
2120 struct mm_struct *mm = vma->vm_mm;
2121 unsigned long remap_pfn = pfn;
2124 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2128 * Physically remapped pages are special. Tell the
2129 * rest of the world about it:
2130 * VM_IO tells people not to look at these pages
2131 * (accesses can have side effects).
2132 * VM_PFNMAP tells the core MM that the base pages are just
2133 * raw PFN mappings, and do not have a "struct page" associated
2136 * Disable vma merging and expanding with mremap().
2138 * Omit vma from core dump, even when VM_IO turned off.
2140 * There's a horrible special case to handle copy-on-write
2141 * behaviour that some programs depend on. We mark the "original"
2142 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2143 * See vm_normal_page() for details.
2145 if (is_cow_mapping(vma->vm_flags)) {
2146 if (addr != vma->vm_start || end != vma->vm_end)
2148 vma->vm_pgoff = pfn;
2151 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2155 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2157 BUG_ON(addr >= end);
2158 pfn -= addr >> PAGE_SHIFT;
2159 pgd = pgd_offset(mm, addr);
2160 flush_cache_range(vma, addr, end);
2162 next = pgd_addr_end(addr, end);
2163 err = remap_p4d_range(mm, pgd, addr, next,
2164 pfn + (addr >> PAGE_SHIFT), prot);
2167 } while (pgd++, addr = next, addr != end);
2170 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2174 EXPORT_SYMBOL(remap_pfn_range);
2177 * vm_iomap_memory - remap memory to userspace
2178 * @vma: user vma to map to
2179 * @start: start of the physical memory to be mapped
2180 * @len: size of area
2182 * This is a simplified io_remap_pfn_range() for common driver use. The
2183 * driver just needs to give us the physical memory range to be mapped,
2184 * we'll figure out the rest from the vma information.
2186 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2187 * whatever write-combining details or similar.
2189 * Return: %0 on success, negative error code otherwise.
2191 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2193 unsigned long vm_len, pfn, pages;
2195 /* Check that the physical memory area passed in looks valid */
2196 if (start + len < start)
2199 * You *really* shouldn't map things that aren't page-aligned,
2200 * but we've historically allowed it because IO memory might
2201 * just have smaller alignment.
2203 len += start & ~PAGE_MASK;
2204 pfn = start >> PAGE_SHIFT;
2205 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2206 if (pfn + pages < pfn)
2209 /* We start the mapping 'vm_pgoff' pages into the area */
2210 if (vma->vm_pgoff > pages)
2212 pfn += vma->vm_pgoff;
2213 pages -= vma->vm_pgoff;
2215 /* Can we fit all of the mapping? */
2216 vm_len = vma->vm_end - vma->vm_start;
2217 if (vm_len >> PAGE_SHIFT > pages)
2220 /* Ok, let it rip */
2221 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2223 EXPORT_SYMBOL(vm_iomap_memory);
2225 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2226 unsigned long addr, unsigned long end,
2227 pte_fn_t fn, void *data, bool create,
2228 pgtbl_mod_mask *mask)
2235 pte = (mm == &init_mm) ?
2236 pte_alloc_kernel_track(pmd, addr, mask) :
2237 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2241 pte = (mm == &init_mm) ?
2242 pte_offset_kernel(pmd, addr) :
2243 pte_offset_map_lock(mm, pmd, addr, &ptl);
2246 BUG_ON(pmd_huge(*pmd));
2248 arch_enter_lazy_mmu_mode();
2251 if (create || !pte_none(*pte)) {
2252 err = fn(pte++, addr, data);
2256 } while (addr += PAGE_SIZE, addr != end);
2257 *mask |= PGTBL_PTE_MODIFIED;
2259 arch_leave_lazy_mmu_mode();
2262 pte_unmap_unlock(pte-1, ptl);
2266 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2267 unsigned long addr, unsigned long end,
2268 pte_fn_t fn, void *data, bool create,
2269 pgtbl_mod_mask *mask)
2275 BUG_ON(pud_huge(*pud));
2278 pmd = pmd_alloc_track(mm, pud, addr, mask);
2282 pmd = pmd_offset(pud, addr);
2285 next = pmd_addr_end(addr, end);
2286 if (create || !pmd_none_or_clear_bad(pmd)) {
2287 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2292 } while (pmd++, addr = next, addr != end);
2296 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2297 unsigned long addr, unsigned long end,
2298 pte_fn_t fn, void *data, bool create,
2299 pgtbl_mod_mask *mask)
2306 pud = pud_alloc_track(mm, p4d, addr, mask);
2310 pud = pud_offset(p4d, addr);
2313 next = pud_addr_end(addr, end);
2314 if (create || !pud_none_or_clear_bad(pud)) {
2315 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2320 } while (pud++, addr = next, addr != end);
2324 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2325 unsigned long addr, unsigned long end,
2326 pte_fn_t fn, void *data, bool create,
2327 pgtbl_mod_mask *mask)
2334 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2338 p4d = p4d_offset(pgd, addr);
2341 next = p4d_addr_end(addr, end);
2342 if (create || !p4d_none_or_clear_bad(p4d)) {
2343 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2348 } while (p4d++, addr = next, addr != end);
2352 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2353 unsigned long size, pte_fn_t fn,
2354 void *data, bool create)
2357 unsigned long start = addr, next;
2358 unsigned long end = addr + size;
2359 pgtbl_mod_mask mask = 0;
2362 if (WARN_ON(addr >= end))
2365 pgd = pgd_offset(mm, addr);
2367 next = pgd_addr_end(addr, end);
2368 if (!create && pgd_none_or_clear_bad(pgd))
2370 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2373 } while (pgd++, addr = next, addr != end);
2375 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2376 arch_sync_kernel_mappings(start, start + size);
2382 * Scan a region of virtual memory, filling in page tables as necessary
2383 * and calling a provided function on each leaf page table.
2385 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2386 unsigned long size, pte_fn_t fn, void *data)
2388 return __apply_to_page_range(mm, addr, size, fn, data, true);
2390 EXPORT_SYMBOL_GPL(apply_to_page_range);
2393 * Scan a region of virtual memory, calling a provided function on
2394 * each leaf page table where it exists.
2396 * Unlike apply_to_page_range, this does _not_ fill in page tables
2397 * where they are absent.
2399 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2400 unsigned long size, pte_fn_t fn, void *data)
2402 return __apply_to_page_range(mm, addr, size, fn, data, false);
2404 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2407 * handle_pte_fault chooses page fault handler according to an entry which was
2408 * read non-atomically. Before making any commitment, on those architectures
2409 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2410 * parts, do_swap_page must check under lock before unmapping the pte and
2411 * proceeding (but do_wp_page is only called after already making such a check;
2412 * and do_anonymous_page can safely check later on).
2414 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2415 pte_t *page_table, pte_t orig_pte)
2418 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2419 if (sizeof(pte_t) > sizeof(unsigned long)) {
2420 spinlock_t *ptl = pte_lockptr(mm, pmd);
2422 same = pte_same(*page_table, orig_pte);
2426 pte_unmap(page_table);
2430 static inline bool cow_user_page(struct page *dst, struct page *src,
2431 struct vm_fault *vmf)
2436 bool locked = false;
2437 struct vm_area_struct *vma = vmf->vma;
2438 struct mm_struct *mm = vma->vm_mm;
2439 unsigned long addr = vmf->address;
2442 copy_user_highpage(dst, src, addr, vma);
2447 * If the source page was a PFN mapping, we don't have
2448 * a "struct page" for it. We do a best-effort copy by
2449 * just copying from the original user address. If that
2450 * fails, we just zero-fill it. Live with it.
2452 kaddr = kmap_atomic(dst);
2453 uaddr = (void __user *)(addr & PAGE_MASK);
2456 * On architectures with software "accessed" bits, we would
2457 * take a double page fault, so mark it accessed here.
2459 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2462 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2464 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2466 * Other thread has already handled the fault
2467 * and update local tlb only
2469 update_mmu_tlb(vma, addr, vmf->pte);
2474 entry = pte_mkyoung(vmf->orig_pte);
2475 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2476 update_mmu_cache(vma, addr, vmf->pte);
2480 * This really shouldn't fail, because the page is there
2481 * in the page tables. But it might just be unreadable,
2482 * in which case we just give up and fill the result with
2485 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2489 /* Re-validate under PTL if the page is still mapped */
2490 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2492 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2493 /* The PTE changed under us, update local tlb */
2494 update_mmu_tlb(vma, addr, vmf->pte);
2500 * The same page can be mapped back since last copy attempt.
2501 * Try to copy again under PTL.
2503 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2505 * Give a warn in case there can be some obscure
2518 pte_unmap_unlock(vmf->pte, vmf->ptl);
2519 kunmap_atomic(kaddr);
2520 flush_dcache_page(dst);
2525 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2527 struct file *vm_file = vma->vm_file;
2530 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2533 * Special mappings (e.g. VDSO) do not have any file so fake
2534 * a default GFP_KERNEL for them.
2540 * Notify the address space that the page is about to become writable so that
2541 * it can prohibit this or wait for the page to get into an appropriate state.
2543 * We do this without the lock held, so that it can sleep if it needs to.
2545 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2548 struct page *page = vmf->page;
2549 unsigned int old_flags = vmf->flags;
2551 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2553 if (vmf->vma->vm_file &&
2554 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2555 return VM_FAULT_SIGBUS;
2557 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2558 /* Restore original flags so that caller is not surprised */
2559 vmf->flags = old_flags;
2560 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2562 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2564 if (!page->mapping) {
2566 return 0; /* retry */
2568 ret |= VM_FAULT_LOCKED;
2570 VM_BUG_ON_PAGE(!PageLocked(page), page);
2575 * Handle dirtying of a page in shared file mapping on a write fault.
2577 * The function expects the page to be locked and unlocks it.
2579 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2581 struct vm_area_struct *vma = vmf->vma;
2582 struct address_space *mapping;
2583 struct page *page = vmf->page;
2585 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2587 dirtied = set_page_dirty(page);
2588 VM_BUG_ON_PAGE(PageAnon(page), page);
2590 * Take a local copy of the address_space - page.mapping may be zeroed
2591 * by truncate after unlock_page(). The address_space itself remains
2592 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2593 * release semantics to prevent the compiler from undoing this copying.
2595 mapping = page_rmapping(page);
2599 file_update_time(vma->vm_file);
2602 * Throttle page dirtying rate down to writeback speed.
2604 * mapping may be NULL here because some device drivers do not
2605 * set page.mapping but still dirty their pages
2607 * Drop the mmap_lock before waiting on IO, if we can. The file
2608 * is pinning the mapping, as per above.
2610 if ((dirtied || page_mkwrite) && mapping) {
2613 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2614 balance_dirty_pages_ratelimited(mapping);
2617 return VM_FAULT_RETRY;
2625 * Handle write page faults for pages that can be reused in the current vma
2627 * This can happen either due to the mapping being with the VM_SHARED flag,
2628 * or due to us being the last reference standing to the page. In either
2629 * case, all we need to do here is to mark the page as writable and update
2630 * any related book-keeping.
2632 static inline void wp_page_reuse(struct vm_fault *vmf)
2633 __releases(vmf->ptl)
2635 struct vm_area_struct *vma = vmf->vma;
2636 struct page *page = vmf->page;
2639 * Clear the pages cpupid information as the existing
2640 * information potentially belongs to a now completely
2641 * unrelated process.
2644 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2646 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2647 entry = pte_mkyoung(vmf->orig_pte);
2648 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2649 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2650 update_mmu_cache(vma, vmf->address, vmf->pte);
2651 pte_unmap_unlock(vmf->pte, vmf->ptl);
2652 count_vm_event(PGREUSE);
2656 * Handle the case of a page which we actually need to copy to a new page.
2658 * Called with mmap_lock locked and the old page referenced, but
2659 * without the ptl held.
2661 * High level logic flow:
2663 * - Allocate a page, copy the content of the old page to the new one.
2664 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2665 * - Take the PTL. If the pte changed, bail out and release the allocated page
2666 * - If the pte is still the way we remember it, update the page table and all
2667 * relevant references. This includes dropping the reference the page-table
2668 * held to the old page, as well as updating the rmap.
2669 * - In any case, unlock the PTL and drop the reference we took to the old page.
2671 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2673 struct vm_area_struct *vma = vmf->vma;
2674 struct mm_struct *mm = vma->vm_mm;
2675 struct page *old_page = vmf->page;
2676 struct page *new_page = NULL;
2678 int page_copied = 0;
2679 struct mmu_notifier_range range;
2681 if (unlikely(anon_vma_prepare(vma)))
2684 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2685 new_page = alloc_zeroed_user_highpage_movable(vma,
2690 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2695 if (!cow_user_page(new_page, old_page, vmf)) {
2697 * COW failed, if the fault was solved by other,
2698 * it's fine. If not, userspace would re-fault on
2699 * the same address and we will handle the fault
2700 * from the second attempt.
2709 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2711 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2713 __SetPageUptodate(new_page);
2715 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2716 vmf->address & PAGE_MASK,
2717 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2718 mmu_notifier_invalidate_range_start(&range);
2721 * Re-check the pte - we dropped the lock
2723 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2724 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2726 if (!PageAnon(old_page)) {
2727 dec_mm_counter_fast(mm,
2728 mm_counter_file(old_page));
2729 inc_mm_counter_fast(mm, MM_ANONPAGES);
2732 inc_mm_counter_fast(mm, MM_ANONPAGES);
2734 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2735 entry = mk_pte(new_page, vma->vm_page_prot);
2736 entry = pte_sw_mkyoung(entry);
2737 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2739 * Clear the pte entry and flush it first, before updating the
2740 * pte with the new entry. This will avoid a race condition
2741 * seen in the presence of one thread doing SMC and another
2744 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2745 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2746 lru_cache_add_inactive_or_unevictable(new_page, vma);
2748 * We call the notify macro here because, when using secondary
2749 * mmu page tables (such as kvm shadow page tables), we want the
2750 * new page to be mapped directly into the secondary page table.
2752 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2753 update_mmu_cache(vma, vmf->address, vmf->pte);
2756 * Only after switching the pte to the new page may
2757 * we remove the mapcount here. Otherwise another
2758 * process may come and find the rmap count decremented
2759 * before the pte is switched to the new page, and
2760 * "reuse" the old page writing into it while our pte
2761 * here still points into it and can be read by other
2764 * The critical issue is to order this
2765 * page_remove_rmap with the ptp_clear_flush above.
2766 * Those stores are ordered by (if nothing else,)
2767 * the barrier present in the atomic_add_negative
2768 * in page_remove_rmap.
2770 * Then the TLB flush in ptep_clear_flush ensures that
2771 * no process can access the old page before the
2772 * decremented mapcount is visible. And the old page
2773 * cannot be reused until after the decremented
2774 * mapcount is visible. So transitively, TLBs to
2775 * old page will be flushed before it can be reused.
2777 page_remove_rmap(old_page, false);
2780 /* Free the old page.. */
2781 new_page = old_page;
2784 update_mmu_tlb(vma, vmf->address, vmf->pte);
2790 pte_unmap_unlock(vmf->pte, vmf->ptl);
2792 * No need to double call mmu_notifier->invalidate_range() callback as
2793 * the above ptep_clear_flush_notify() did already call it.
2795 mmu_notifier_invalidate_range_only_end(&range);
2798 * Don't let another task, with possibly unlocked vma,
2799 * keep the mlocked page.
2801 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2802 lock_page(old_page); /* LRU manipulation */
2803 if (PageMlocked(old_page))
2804 munlock_vma_page(old_page);
2805 unlock_page(old_page);
2809 return page_copied ? VM_FAULT_WRITE : 0;
2815 return VM_FAULT_OOM;
2819 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2820 * writeable once the page is prepared
2822 * @vmf: structure describing the fault
2824 * This function handles all that is needed to finish a write page fault in a
2825 * shared mapping due to PTE being read-only once the mapped page is prepared.
2826 * It handles locking of PTE and modifying it.
2828 * The function expects the page to be locked or other protection against
2829 * concurrent faults / writeback (such as DAX radix tree locks).
2831 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2832 * we acquired PTE lock.
2834 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2836 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2837 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2840 * We might have raced with another page fault while we released the
2841 * pte_offset_map_lock.
2843 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2844 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
2845 pte_unmap_unlock(vmf->pte, vmf->ptl);
2846 return VM_FAULT_NOPAGE;
2853 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2856 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2858 struct vm_area_struct *vma = vmf->vma;
2860 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2863 pte_unmap_unlock(vmf->pte, vmf->ptl);
2864 vmf->flags |= FAULT_FLAG_MKWRITE;
2865 ret = vma->vm_ops->pfn_mkwrite(vmf);
2866 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2868 return finish_mkwrite_fault(vmf);
2871 return VM_FAULT_WRITE;
2874 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2875 __releases(vmf->ptl)
2877 struct vm_area_struct *vma = vmf->vma;
2878 vm_fault_t ret = VM_FAULT_WRITE;
2880 get_page(vmf->page);
2882 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2885 pte_unmap_unlock(vmf->pte, vmf->ptl);
2886 tmp = do_page_mkwrite(vmf);
2887 if (unlikely(!tmp || (tmp &
2888 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2889 put_page(vmf->page);
2892 tmp = finish_mkwrite_fault(vmf);
2893 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2894 unlock_page(vmf->page);
2895 put_page(vmf->page);
2900 lock_page(vmf->page);
2902 ret |= fault_dirty_shared_page(vmf);
2903 put_page(vmf->page);
2909 * This routine handles present pages, when users try to write
2910 * to a shared page. It is done by copying the page to a new address
2911 * and decrementing the shared-page counter for the old page.
2913 * Note that this routine assumes that the protection checks have been
2914 * done by the caller (the low-level page fault routine in most cases).
2915 * Thus we can safely just mark it writable once we've done any necessary
2918 * We also mark the page dirty at this point even though the page will
2919 * change only once the write actually happens. This avoids a few races,
2920 * and potentially makes it more efficient.
2922 * We enter with non-exclusive mmap_lock (to exclude vma changes,
2923 * but allow concurrent faults), with pte both mapped and locked.
2924 * We return with mmap_lock still held, but pte unmapped and unlocked.
2926 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2927 __releases(vmf->ptl)
2929 struct vm_area_struct *vma = vmf->vma;
2931 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
2932 pte_unmap_unlock(vmf->pte, vmf->ptl);
2933 return handle_userfault(vmf, VM_UFFD_WP);
2936 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2939 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2942 * We should not cow pages in a shared writeable mapping.
2943 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2945 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2946 (VM_WRITE|VM_SHARED))
2947 return wp_pfn_shared(vmf);
2949 pte_unmap_unlock(vmf->pte, vmf->ptl);
2950 return wp_page_copy(vmf);
2954 * Take out anonymous pages first, anonymous shared vmas are
2955 * not dirty accountable.
2957 if (PageAnon(vmf->page)) {
2958 struct page *page = vmf->page;
2960 /* PageKsm() doesn't necessarily raise the page refcount */
2961 if (PageKsm(page) || page_count(page) != 1)
2963 if (!trylock_page(page))
2965 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
2970 * Ok, we've got the only map reference, and the only
2971 * page count reference, and the page is locked,
2972 * it's dark out, and we're wearing sunglasses. Hit it.
2976 return VM_FAULT_WRITE;
2977 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2978 (VM_WRITE|VM_SHARED))) {
2979 return wp_page_shared(vmf);
2983 * Ok, we need to copy. Oh, well..
2985 get_page(vmf->page);
2987 pte_unmap_unlock(vmf->pte, vmf->ptl);
2988 return wp_page_copy(vmf);
2991 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2992 unsigned long start_addr, unsigned long end_addr,
2993 struct zap_details *details)
2995 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2998 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2999 struct zap_details *details)
3001 struct vm_area_struct *vma;
3002 pgoff_t vba, vea, zba, zea;
3004 vma_interval_tree_foreach(vma, root,
3005 details->first_index, details->last_index) {
3007 vba = vma->vm_pgoff;
3008 vea = vba + vma_pages(vma) - 1;
3009 zba = details->first_index;
3012 zea = details->last_index;
3016 unmap_mapping_range_vma(vma,
3017 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3018 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3024 * unmap_mapping_pages() - Unmap pages from processes.
3025 * @mapping: The address space containing pages to be unmapped.
3026 * @start: Index of first page to be unmapped.
3027 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3028 * @even_cows: Whether to unmap even private COWed pages.
3030 * Unmap the pages in this address space from any userspace process which
3031 * has them mmaped. Generally, you want to remove COWed pages as well when
3032 * a file is being truncated, but not when invalidating pages from the page
3035 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3036 pgoff_t nr, bool even_cows)
3038 struct zap_details details = { };
3040 details.check_mapping = even_cows ? NULL : mapping;
3041 details.first_index = start;
3042 details.last_index = start + nr - 1;
3043 if (details.last_index < details.first_index)
3044 details.last_index = ULONG_MAX;
3046 i_mmap_lock_write(mapping);
3047 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3048 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3049 i_mmap_unlock_write(mapping);
3053 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3054 * address_space corresponding to the specified byte range in the underlying
3057 * @mapping: the address space containing mmaps to be unmapped.
3058 * @holebegin: byte in first page to unmap, relative to the start of
3059 * the underlying file. This will be rounded down to a PAGE_SIZE
3060 * boundary. Note that this is different from truncate_pagecache(), which
3061 * must keep the partial page. In contrast, we must get rid of
3063 * @holelen: size of prospective hole in bytes. This will be rounded
3064 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3066 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3067 * but 0 when invalidating pagecache, don't throw away private data.
3069 void unmap_mapping_range(struct address_space *mapping,
3070 loff_t const holebegin, loff_t const holelen, int even_cows)
3072 pgoff_t hba = holebegin >> PAGE_SHIFT;
3073 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3075 /* Check for overflow. */
3076 if (sizeof(holelen) > sizeof(hlen)) {
3078 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3079 if (holeend & ~(long long)ULONG_MAX)
3080 hlen = ULONG_MAX - hba + 1;
3083 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3085 EXPORT_SYMBOL(unmap_mapping_range);
3088 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3089 * but allow concurrent faults), and pte mapped but not yet locked.
3090 * We return with pte unmapped and unlocked.
3092 * We return with the mmap_lock locked or unlocked in the same cases
3093 * as does filemap_fault().
3095 vm_fault_t do_swap_page(struct vm_fault *vmf)
3097 struct vm_area_struct *vma = vmf->vma;
3098 struct page *page = NULL, *swapcache;
3104 void *shadow = NULL;
3106 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3109 entry = pte_to_swp_entry(vmf->orig_pte);
3110 if (unlikely(non_swap_entry(entry))) {
3111 if (is_migration_entry(entry)) {
3112 migration_entry_wait(vma->vm_mm, vmf->pmd,
3114 } else if (is_device_private_entry(entry)) {
3115 vmf->page = device_private_entry_to_page(entry);
3116 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3117 } else if (is_hwpoison_entry(entry)) {
3118 ret = VM_FAULT_HWPOISON;
3120 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3121 ret = VM_FAULT_SIGBUS;
3127 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3128 page = lookup_swap_cache(entry, vma, vmf->address);
3132 struct swap_info_struct *si = swp_swap_info(entry);
3134 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3135 __swap_count(entry) == 1) {
3136 /* skip swapcache */
3137 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3142 __SetPageLocked(page);
3143 __SetPageSwapBacked(page);
3144 set_page_private(page, entry.val);
3146 /* Tell memcg to use swap ownership records */
3147 SetPageSwapCache(page);
3148 err = mem_cgroup_charge(page, vma->vm_mm,
3150 ClearPageSwapCache(page);
3156 shadow = get_shadow_from_swap_cache(entry);
3158 workingset_refault(page, shadow);
3160 lru_cache_add(page);
3161 swap_readpage(page, true);
3164 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3171 * Back out if somebody else faulted in this pte
3172 * while we released the pte lock.
3174 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3175 vmf->address, &vmf->ptl);
3176 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3178 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3182 /* Had to read the page from swap area: Major fault */
3183 ret = VM_FAULT_MAJOR;
3184 count_vm_event(PGMAJFAULT);
3185 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3186 } else if (PageHWPoison(page)) {
3188 * hwpoisoned dirty swapcache pages are kept for killing
3189 * owner processes (which may be unknown at hwpoison time)
3191 ret = VM_FAULT_HWPOISON;
3192 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3196 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3198 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3200 ret |= VM_FAULT_RETRY;
3205 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3206 * release the swapcache from under us. The page pin, and pte_same
3207 * test below, are not enough to exclude that. Even if it is still
3208 * swapcache, we need to check that the page's swap has not changed.
3210 if (unlikely((!PageSwapCache(page) ||
3211 page_private(page) != entry.val)) && swapcache)
3214 page = ksm_might_need_to_copy(page, vma, vmf->address);
3215 if (unlikely(!page)) {
3221 cgroup_throttle_swaprate(page, GFP_KERNEL);
3224 * Back out if somebody else already faulted in this pte.
3226 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3228 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3231 if (unlikely(!PageUptodate(page))) {
3232 ret = VM_FAULT_SIGBUS;
3237 * The page isn't present yet, go ahead with the fault.
3239 * Be careful about the sequence of operations here.
3240 * To get its accounting right, reuse_swap_page() must be called
3241 * while the page is counted on swap but not yet in mapcount i.e.
3242 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3243 * must be called after the swap_free(), or it will never succeed.
3246 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3247 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3248 pte = mk_pte(page, vma->vm_page_prot);
3249 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3250 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3251 vmf->flags &= ~FAULT_FLAG_WRITE;
3252 ret |= VM_FAULT_WRITE;
3253 exclusive = RMAP_EXCLUSIVE;
3255 flush_icache_page(vma, page);
3256 if (pte_swp_soft_dirty(vmf->orig_pte))
3257 pte = pte_mksoft_dirty(pte);
3258 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3259 pte = pte_mkuffd_wp(pte);
3260 pte = pte_wrprotect(pte);
3262 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3263 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3264 vmf->orig_pte = pte;
3266 /* ksm created a completely new copy */
3267 if (unlikely(page != swapcache && swapcache)) {
3268 page_add_new_anon_rmap(page, vma, vmf->address, false);
3269 lru_cache_add_inactive_or_unevictable(page, vma);
3271 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3275 if (mem_cgroup_swap_full(page) ||
3276 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3277 try_to_free_swap(page);
3279 if (page != swapcache && swapcache) {
3281 * Hold the lock to avoid the swap entry to be reused
3282 * until we take the PT lock for the pte_same() check
3283 * (to avoid false positives from pte_same). For
3284 * further safety release the lock after the swap_free
3285 * so that the swap count won't change under a
3286 * parallel locked swapcache.
3288 unlock_page(swapcache);
3289 put_page(swapcache);
3292 if (vmf->flags & FAULT_FLAG_WRITE) {
3293 ret |= do_wp_page(vmf);
3294 if (ret & VM_FAULT_ERROR)
3295 ret &= VM_FAULT_ERROR;
3299 /* No need to invalidate - it was non-present before */
3300 update_mmu_cache(vma, vmf->address, vmf->pte);
3302 pte_unmap_unlock(vmf->pte, vmf->ptl);
3306 pte_unmap_unlock(vmf->pte, vmf->ptl);
3311 if (page != swapcache && swapcache) {
3312 unlock_page(swapcache);
3313 put_page(swapcache);
3319 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3320 * but allow concurrent faults), and pte mapped but not yet locked.
3321 * We return with mmap_lock still held, but pte unmapped and unlocked.
3323 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3325 struct vm_area_struct *vma = vmf->vma;
3330 /* File mapping without ->vm_ops ? */
3331 if (vma->vm_flags & VM_SHARED)
3332 return VM_FAULT_SIGBUS;
3335 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3336 * pte_offset_map() on pmds where a huge pmd might be created
3337 * from a different thread.
3339 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3340 * parallel threads are excluded by other means.
3342 * Here we only have mmap_read_lock(mm).
3344 if (pte_alloc(vma->vm_mm, vmf->pmd))
3345 return VM_FAULT_OOM;
3347 /* See the comment in pte_alloc_one_map() */
3348 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3351 /* Use the zero-page for reads */
3352 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3353 !mm_forbids_zeropage(vma->vm_mm)) {
3354 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3355 vma->vm_page_prot));
3356 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3357 vmf->address, &vmf->ptl);
3358 if (!pte_none(*vmf->pte)) {
3359 update_mmu_tlb(vma, vmf->address, vmf->pte);
3362 ret = check_stable_address_space(vma->vm_mm);
3365 /* Deliver the page fault to userland, check inside PT lock */
3366 if (userfaultfd_missing(vma)) {
3367 pte_unmap_unlock(vmf->pte, vmf->ptl);
3368 return handle_userfault(vmf, VM_UFFD_MISSING);
3373 /* Allocate our own private page. */
3374 if (unlikely(anon_vma_prepare(vma)))
3376 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3380 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3382 cgroup_throttle_swaprate(page, GFP_KERNEL);
3385 * The memory barrier inside __SetPageUptodate makes sure that
3386 * preceding stores to the page contents become visible before
3387 * the set_pte_at() write.
3389 __SetPageUptodate(page);
3391 entry = mk_pte(page, vma->vm_page_prot);
3392 entry = pte_sw_mkyoung(entry);
3393 if (vma->vm_flags & VM_WRITE)
3394 entry = pte_mkwrite(pte_mkdirty(entry));
3396 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3398 if (!pte_none(*vmf->pte)) {
3399 update_mmu_cache(vma, vmf->address, vmf->pte);
3403 ret = check_stable_address_space(vma->vm_mm);
3407 /* Deliver the page fault to userland, check inside PT lock */
3408 if (userfaultfd_missing(vma)) {
3409 pte_unmap_unlock(vmf->pte, vmf->ptl);
3411 return handle_userfault(vmf, VM_UFFD_MISSING);
3414 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3415 page_add_new_anon_rmap(page, vma, vmf->address, false);
3416 lru_cache_add_inactive_or_unevictable(page, vma);
3418 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3420 /* No need to invalidate - it was non-present before */
3421 update_mmu_cache(vma, vmf->address, vmf->pte);
3423 pte_unmap_unlock(vmf->pte, vmf->ptl);
3431 return VM_FAULT_OOM;
3435 * The mmap_lock must have been held on entry, and may have been
3436 * released depending on flags and vma->vm_ops->fault() return value.
3437 * See filemap_fault() and __lock_page_retry().
3439 static vm_fault_t __do_fault(struct vm_fault *vmf)
3441 struct vm_area_struct *vma = vmf->vma;
3445 * Preallocate pte before we take page_lock because this might lead to
3446 * deadlocks for memcg reclaim which waits for pages under writeback:
3448 * SetPageWriteback(A)
3454 * wait_on_page_writeback(A)
3455 * SetPageWriteback(B)
3457 * # flush A, B to clear the writeback
3459 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3460 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3461 if (!vmf->prealloc_pte)
3462 return VM_FAULT_OOM;
3463 smp_wmb(); /* See comment in __pte_alloc() */
3466 ret = vma->vm_ops->fault(vmf);
3467 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3468 VM_FAULT_DONE_COW)))
3471 if (unlikely(PageHWPoison(vmf->page))) {
3472 if (ret & VM_FAULT_LOCKED)
3473 unlock_page(vmf->page);
3474 put_page(vmf->page);
3476 return VM_FAULT_HWPOISON;
3479 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3480 lock_page(vmf->page);
3482 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3488 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3489 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3490 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3491 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3493 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3495 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3498 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3500 struct vm_area_struct *vma = vmf->vma;
3502 if (!pmd_none(*vmf->pmd))
3504 if (vmf->prealloc_pte) {
3505 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3506 if (unlikely(!pmd_none(*vmf->pmd))) {
3507 spin_unlock(vmf->ptl);
3511 mm_inc_nr_ptes(vma->vm_mm);
3512 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3513 spin_unlock(vmf->ptl);
3514 vmf->prealloc_pte = NULL;
3515 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3516 return VM_FAULT_OOM;
3520 * If a huge pmd materialized under us just retry later. Use
3521 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3522 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3523 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3524 * running immediately after a huge pmd fault in a different thread of
3525 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3526 * All we have to ensure is that it is a regular pmd that we can walk
3527 * with pte_offset_map() and we can do that through an atomic read in
3528 * C, which is what pmd_trans_unstable() provides.
3530 if (pmd_devmap_trans_unstable(vmf->pmd))
3531 return VM_FAULT_NOPAGE;
3534 * At this point we know that our vmf->pmd points to a page of ptes
3535 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3536 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3537 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3538 * be valid and we will re-check to make sure the vmf->pte isn't
3539 * pte_none() under vmf->ptl protection when we return to
3542 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3547 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3548 static void deposit_prealloc_pte(struct vm_fault *vmf)
3550 struct vm_area_struct *vma = vmf->vma;
3552 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3554 * We are going to consume the prealloc table,
3555 * count that as nr_ptes.
3557 mm_inc_nr_ptes(vma->vm_mm);
3558 vmf->prealloc_pte = NULL;
3561 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3563 struct vm_area_struct *vma = vmf->vma;
3564 bool write = vmf->flags & FAULT_FLAG_WRITE;
3565 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3570 if (!transhuge_vma_suitable(vma, haddr))
3571 return VM_FAULT_FALLBACK;
3573 ret = VM_FAULT_FALLBACK;
3574 page = compound_head(page);
3577 * Archs like ppc64 need additonal space to store information
3578 * related to pte entry. Use the preallocated table for that.
3580 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3581 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3582 if (!vmf->prealloc_pte)
3583 return VM_FAULT_OOM;
3584 smp_wmb(); /* See comment in __pte_alloc() */
3587 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3588 if (unlikely(!pmd_none(*vmf->pmd)))
3591 for (i = 0; i < HPAGE_PMD_NR; i++)
3592 flush_icache_page(vma, page + i);
3594 entry = mk_huge_pmd(page, vma->vm_page_prot);
3596 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3598 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3599 page_add_file_rmap(page, true);
3601 * deposit and withdraw with pmd lock held
3603 if (arch_needs_pgtable_deposit())
3604 deposit_prealloc_pte(vmf);
3606 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3608 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3610 /* fault is handled */
3612 count_vm_event(THP_FILE_MAPPED);
3614 spin_unlock(vmf->ptl);
3618 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3626 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3627 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3629 * @vmf: fault environment
3630 * @page: page to map
3632 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3635 * Target users are page handler itself and implementations of
3636 * vm_ops->map_pages.
3638 * Return: %0 on success, %VM_FAULT_ code in case of error.
3640 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
3642 struct vm_area_struct *vma = vmf->vma;
3643 bool write = vmf->flags & FAULT_FLAG_WRITE;
3647 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3648 ret = do_set_pmd(vmf, page);
3649 if (ret != VM_FAULT_FALLBACK)
3654 ret = pte_alloc_one_map(vmf);
3659 /* Re-check under ptl */
3660 if (unlikely(!pte_none(*vmf->pte))) {
3661 update_mmu_tlb(vma, vmf->address, vmf->pte);
3662 return VM_FAULT_NOPAGE;
3665 flush_icache_page(vma, page);
3666 entry = mk_pte(page, vma->vm_page_prot);
3667 entry = pte_sw_mkyoung(entry);
3669 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3670 /* copy-on-write page */
3671 if (write && !(vma->vm_flags & VM_SHARED)) {
3672 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3673 page_add_new_anon_rmap(page, vma, vmf->address, false);
3674 lru_cache_add_inactive_or_unevictable(page, vma);
3676 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3677 page_add_file_rmap(page, false);
3679 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3681 /* no need to invalidate: a not-present page won't be cached */
3682 update_mmu_cache(vma, vmf->address, vmf->pte);
3689 * finish_fault - finish page fault once we have prepared the page to fault
3691 * @vmf: structure describing the fault
3693 * This function handles all that is needed to finish a page fault once the
3694 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3695 * given page, adds reverse page mapping, handles memcg charges and LRU
3698 * The function expects the page to be locked and on success it consumes a
3699 * reference of a page being mapped (for the PTE which maps it).
3701 * Return: %0 on success, %VM_FAULT_ code in case of error.
3703 vm_fault_t finish_fault(struct vm_fault *vmf)
3708 /* Did we COW the page? */
3709 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3710 !(vmf->vma->vm_flags & VM_SHARED))
3711 page = vmf->cow_page;
3716 * check even for read faults because we might have lost our CoWed
3719 if (!(vmf->vma->vm_flags & VM_SHARED))
3720 ret = check_stable_address_space(vmf->vma->vm_mm);
3722 ret = alloc_set_pte(vmf, page);
3724 pte_unmap_unlock(vmf->pte, vmf->ptl);
3728 static unsigned long fault_around_bytes __read_mostly =
3729 rounddown_pow_of_two(65536);
3731 #ifdef CONFIG_DEBUG_FS
3732 static int fault_around_bytes_get(void *data, u64 *val)
3734 *val = fault_around_bytes;
3739 * fault_around_bytes must be rounded down to the nearest page order as it's
3740 * what do_fault_around() expects to see.
3742 static int fault_around_bytes_set(void *data, u64 val)
3744 if (val / PAGE_SIZE > PTRS_PER_PTE)
3746 if (val > PAGE_SIZE)
3747 fault_around_bytes = rounddown_pow_of_two(val);
3749 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3752 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3753 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3755 static int __init fault_around_debugfs(void)
3757 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3758 &fault_around_bytes_fops);
3761 late_initcall(fault_around_debugfs);
3765 * do_fault_around() tries to map few pages around the fault address. The hope
3766 * is that the pages will be needed soon and this will lower the number of
3769 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3770 * not ready to be mapped: not up-to-date, locked, etc.
3772 * This function is called with the page table lock taken. In the split ptlock
3773 * case the page table lock only protects only those entries which belong to
3774 * the page table corresponding to the fault address.
3776 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3779 * fault_around_bytes defines how many bytes we'll try to map.
3780 * do_fault_around() expects it to be set to a power of two less than or equal
3783 * The virtual address of the area that we map is naturally aligned to
3784 * fault_around_bytes rounded down to the machine page size
3785 * (and therefore to page order). This way it's easier to guarantee
3786 * that we don't cross page table boundaries.
3788 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3790 unsigned long address = vmf->address, nr_pages, mask;
3791 pgoff_t start_pgoff = vmf->pgoff;
3796 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3797 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3799 vmf->address = max(address & mask, vmf->vma->vm_start);
3800 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3804 * end_pgoff is either the end of the page table, the end of
3805 * the vma or nr_pages from start_pgoff, depending what is nearest.
3807 end_pgoff = start_pgoff -
3808 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3810 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3811 start_pgoff + nr_pages - 1);
3813 if (pmd_none(*vmf->pmd)) {
3814 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3815 if (!vmf->prealloc_pte)
3817 smp_wmb(); /* See comment in __pte_alloc() */
3820 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3822 /* Huge page is mapped? Page fault is solved */
3823 if (pmd_trans_huge(*vmf->pmd)) {
3824 ret = VM_FAULT_NOPAGE;
3828 /* ->map_pages() haven't done anything useful. Cold page cache? */
3832 /* check if the page fault is solved */
3833 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3834 if (!pte_none(*vmf->pte))
3835 ret = VM_FAULT_NOPAGE;
3836 pte_unmap_unlock(vmf->pte, vmf->ptl);
3838 vmf->address = address;
3843 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3845 struct vm_area_struct *vma = vmf->vma;
3849 * Let's call ->map_pages() first and use ->fault() as fallback
3850 * if page by the offset is not ready to be mapped (cold cache or
3853 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3854 ret = do_fault_around(vmf);
3859 ret = __do_fault(vmf);
3860 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3863 ret |= finish_fault(vmf);
3864 unlock_page(vmf->page);
3865 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3866 put_page(vmf->page);
3870 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3872 struct vm_area_struct *vma = vmf->vma;
3875 if (unlikely(anon_vma_prepare(vma)))
3876 return VM_FAULT_OOM;
3878 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3880 return VM_FAULT_OOM;
3882 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
3883 put_page(vmf->cow_page);
3884 return VM_FAULT_OOM;
3886 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
3888 ret = __do_fault(vmf);
3889 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3891 if (ret & VM_FAULT_DONE_COW)
3894 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3895 __SetPageUptodate(vmf->cow_page);
3897 ret |= finish_fault(vmf);
3898 unlock_page(vmf->page);
3899 put_page(vmf->page);
3900 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3904 put_page(vmf->cow_page);
3908 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3910 struct vm_area_struct *vma = vmf->vma;
3911 vm_fault_t ret, tmp;
3913 ret = __do_fault(vmf);
3914 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3918 * Check if the backing address space wants to know that the page is
3919 * about to become writable
3921 if (vma->vm_ops->page_mkwrite) {
3922 unlock_page(vmf->page);
3923 tmp = do_page_mkwrite(vmf);
3924 if (unlikely(!tmp ||
3925 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3926 put_page(vmf->page);
3931 ret |= finish_fault(vmf);
3932 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3934 unlock_page(vmf->page);
3935 put_page(vmf->page);
3939 ret |= fault_dirty_shared_page(vmf);
3944 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3945 * but allow concurrent faults).
3946 * The mmap_lock may have been released depending on flags and our
3947 * return value. See filemap_fault() and __lock_page_or_retry().
3948 * If mmap_lock is released, vma may become invalid (for example
3949 * by other thread calling munmap()).
3951 static vm_fault_t do_fault(struct vm_fault *vmf)
3953 struct vm_area_struct *vma = vmf->vma;
3954 struct mm_struct *vm_mm = vma->vm_mm;
3958 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3960 if (!vma->vm_ops->fault) {
3962 * If we find a migration pmd entry or a none pmd entry, which
3963 * should never happen, return SIGBUS
3965 if (unlikely(!pmd_present(*vmf->pmd)))
3966 ret = VM_FAULT_SIGBUS;
3968 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3973 * Make sure this is not a temporary clearing of pte
3974 * by holding ptl and checking again. A R/M/W update
3975 * of pte involves: take ptl, clearing the pte so that
3976 * we don't have concurrent modification by hardware
3977 * followed by an update.
3979 if (unlikely(pte_none(*vmf->pte)))
3980 ret = VM_FAULT_SIGBUS;
3982 ret = VM_FAULT_NOPAGE;
3984 pte_unmap_unlock(vmf->pte, vmf->ptl);
3986 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3987 ret = do_read_fault(vmf);
3988 else if (!(vma->vm_flags & VM_SHARED))
3989 ret = do_cow_fault(vmf);
3991 ret = do_shared_fault(vmf);
3993 /* preallocated pagetable is unused: free it */
3994 if (vmf->prealloc_pte) {
3995 pte_free(vm_mm, vmf->prealloc_pte);
3996 vmf->prealloc_pte = NULL;
4001 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4002 unsigned long addr, int page_nid,
4007 count_vm_numa_event(NUMA_HINT_FAULTS);
4008 if (page_nid == numa_node_id()) {
4009 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4010 *flags |= TNF_FAULT_LOCAL;
4013 return mpol_misplaced(page, vma, addr);
4016 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4018 struct vm_area_struct *vma = vmf->vma;
4019 struct page *page = NULL;
4020 int page_nid = NUMA_NO_NODE;
4023 bool migrated = false;
4025 bool was_writable = pte_savedwrite(vmf->orig_pte);
4029 * The "pte" at this point cannot be used safely without
4030 * validation through pte_unmap_same(). It's of NUMA type but
4031 * the pfn may be screwed if the read is non atomic.
4033 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4034 spin_lock(vmf->ptl);
4035 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4036 pte_unmap_unlock(vmf->pte, vmf->ptl);
4041 * Make it present again, Depending on how arch implementes non
4042 * accessible ptes, some can allow access by kernel mode.
4044 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4045 pte = pte_modify(old_pte, vma->vm_page_prot);
4046 pte = pte_mkyoung(pte);
4048 pte = pte_mkwrite(pte);
4049 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4050 update_mmu_cache(vma, vmf->address, vmf->pte);
4052 page = vm_normal_page(vma, vmf->address, pte);
4054 pte_unmap_unlock(vmf->pte, vmf->ptl);
4058 /* TODO: handle PTE-mapped THP */
4059 if (PageCompound(page)) {
4060 pte_unmap_unlock(vmf->pte, vmf->ptl);
4065 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4066 * much anyway since they can be in shared cache state. This misses
4067 * the case where a mapping is writable but the process never writes
4068 * to it but pte_write gets cleared during protection updates and
4069 * pte_dirty has unpredictable behaviour between PTE scan updates,
4070 * background writeback, dirty balancing and application behaviour.
4072 if (!pte_write(pte))
4073 flags |= TNF_NO_GROUP;
4076 * Flag if the page is shared between multiple address spaces. This
4077 * is later used when determining whether to group tasks together
4079 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4080 flags |= TNF_SHARED;
4082 last_cpupid = page_cpupid_last(page);
4083 page_nid = page_to_nid(page);
4084 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4086 pte_unmap_unlock(vmf->pte, vmf->ptl);
4087 if (target_nid == NUMA_NO_NODE) {
4092 /* Migrate to the requested node */
4093 migrated = migrate_misplaced_page(page, vma, target_nid);
4095 page_nid = target_nid;
4096 flags |= TNF_MIGRATED;
4098 flags |= TNF_MIGRATE_FAIL;
4101 if (page_nid != NUMA_NO_NODE)
4102 task_numa_fault(last_cpupid, page_nid, 1, flags);
4106 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4108 if (vma_is_anonymous(vmf->vma))
4109 return do_huge_pmd_anonymous_page(vmf);
4110 if (vmf->vma->vm_ops->huge_fault)
4111 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4112 return VM_FAULT_FALLBACK;
4115 /* `inline' is required to avoid gcc 4.1.2 build error */
4116 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4118 if (vma_is_anonymous(vmf->vma)) {
4119 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4120 return handle_userfault(vmf, VM_UFFD_WP);
4121 return do_huge_pmd_wp_page(vmf, orig_pmd);
4123 if (vmf->vma->vm_ops->huge_fault) {
4124 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4126 if (!(ret & VM_FAULT_FALLBACK))
4130 /* COW or write-notify handled on pte level: split pmd. */
4131 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4133 return VM_FAULT_FALLBACK;
4136 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4138 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4139 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4140 /* No support for anonymous transparent PUD pages yet */
4141 if (vma_is_anonymous(vmf->vma))
4143 if (vmf->vma->vm_ops->huge_fault) {
4144 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4146 if (!(ret & VM_FAULT_FALLBACK))
4150 /* COW or write-notify not handled on PUD level: split pud.*/
4151 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4152 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4153 return VM_FAULT_FALLBACK;
4156 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4158 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4159 /* No support for anonymous transparent PUD pages yet */
4160 if (vma_is_anonymous(vmf->vma))
4161 return VM_FAULT_FALLBACK;
4162 if (vmf->vma->vm_ops->huge_fault)
4163 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4164 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4165 return VM_FAULT_FALLBACK;
4169 * These routines also need to handle stuff like marking pages dirty
4170 * and/or accessed for architectures that don't do it in hardware (most
4171 * RISC architectures). The early dirtying is also good on the i386.
4173 * There is also a hook called "update_mmu_cache()" that architectures
4174 * with external mmu caches can use to update those (ie the Sparc or
4175 * PowerPC hashed page tables that act as extended TLBs).
4177 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4178 * concurrent faults).
4180 * The mmap_lock may have been released depending on flags and our return value.
4181 * See filemap_fault() and __lock_page_or_retry().
4183 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4187 if (unlikely(pmd_none(*vmf->pmd))) {
4189 * Leave __pte_alloc() until later: because vm_ops->fault may
4190 * want to allocate huge page, and if we expose page table
4191 * for an instant, it will be difficult to retract from
4192 * concurrent faults and from rmap lookups.
4196 /* See comment in pte_alloc_one_map() */
4197 if (pmd_devmap_trans_unstable(vmf->pmd))
4200 * A regular pmd is established and it can't morph into a huge
4201 * pmd from under us anymore at this point because we hold the
4202 * mmap_lock read mode and khugepaged takes it in write mode.
4203 * So now it's safe to run pte_offset_map().
4205 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4206 vmf->orig_pte = *vmf->pte;
4209 * some architectures can have larger ptes than wordsize,
4210 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4211 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4212 * accesses. The code below just needs a consistent view
4213 * for the ifs and we later double check anyway with the
4214 * ptl lock held. So here a barrier will do.
4217 if (pte_none(vmf->orig_pte)) {
4218 pte_unmap(vmf->pte);
4224 if (vma_is_anonymous(vmf->vma))
4225 return do_anonymous_page(vmf);
4227 return do_fault(vmf);
4230 if (!pte_present(vmf->orig_pte))
4231 return do_swap_page(vmf);
4233 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4234 return do_numa_page(vmf);
4236 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4237 spin_lock(vmf->ptl);
4238 entry = vmf->orig_pte;
4239 if (unlikely(!pte_same(*vmf->pte, entry))) {
4240 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4243 if (vmf->flags & FAULT_FLAG_WRITE) {
4244 if (!pte_write(entry))
4245 return do_wp_page(vmf);
4246 entry = pte_mkdirty(entry);
4248 entry = pte_mkyoung(entry);
4249 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4250 vmf->flags & FAULT_FLAG_WRITE)) {
4251 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4253 /* Skip spurious TLB flush for retried page fault */
4254 if (vmf->flags & FAULT_FLAG_TRIED)
4257 * This is needed only for protection faults but the arch code
4258 * is not yet telling us if this is a protection fault or not.
4259 * This still avoids useless tlb flushes for .text page faults
4262 if (vmf->flags & FAULT_FLAG_WRITE)
4263 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4266 pte_unmap_unlock(vmf->pte, vmf->ptl);
4271 * By the time we get here, we already hold the mm semaphore
4273 * The mmap_lock may have been released depending on flags and our
4274 * return value. See filemap_fault() and __lock_page_or_retry().
4276 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4277 unsigned long address, unsigned int flags)
4279 struct vm_fault vmf = {
4281 .address = address & PAGE_MASK,
4283 .pgoff = linear_page_index(vma, address),
4284 .gfp_mask = __get_fault_gfp_mask(vma),
4286 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4287 struct mm_struct *mm = vma->vm_mm;
4292 pgd = pgd_offset(mm, address);
4293 p4d = p4d_alloc(mm, pgd, address);
4295 return VM_FAULT_OOM;
4297 vmf.pud = pud_alloc(mm, p4d, address);
4299 return VM_FAULT_OOM;
4301 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4302 ret = create_huge_pud(&vmf);
4303 if (!(ret & VM_FAULT_FALLBACK))
4306 pud_t orig_pud = *vmf.pud;
4309 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4311 /* NUMA case for anonymous PUDs would go here */
4313 if (dirty && !pud_write(orig_pud)) {
4314 ret = wp_huge_pud(&vmf, orig_pud);
4315 if (!(ret & VM_FAULT_FALLBACK))
4318 huge_pud_set_accessed(&vmf, orig_pud);
4324 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4326 return VM_FAULT_OOM;
4328 /* Huge pud page fault raced with pmd_alloc? */
4329 if (pud_trans_unstable(vmf.pud))
4332 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4333 ret = create_huge_pmd(&vmf);
4334 if (!(ret & VM_FAULT_FALLBACK))
4337 pmd_t orig_pmd = *vmf.pmd;
4340 if (unlikely(is_swap_pmd(orig_pmd))) {
4341 VM_BUG_ON(thp_migration_supported() &&
4342 !is_pmd_migration_entry(orig_pmd));
4343 if (is_pmd_migration_entry(orig_pmd))
4344 pmd_migration_entry_wait(mm, vmf.pmd);
4347 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4348 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4349 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4351 if (dirty && !pmd_write(orig_pmd)) {
4352 ret = wp_huge_pmd(&vmf, orig_pmd);
4353 if (!(ret & VM_FAULT_FALLBACK))
4356 huge_pmd_set_accessed(&vmf, orig_pmd);
4362 return handle_pte_fault(&vmf);
4366 * mm_account_fault - Do page fault accountings
4368 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4369 * of perf event counters, but we'll still do the per-task accounting to
4370 * the task who triggered this page fault.
4371 * @address: the faulted address.
4372 * @flags: the fault flags.
4373 * @ret: the fault retcode.
4375 * This will take care of most of the page fault accountings. Meanwhile, it
4376 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4377 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4378 * still be in per-arch page fault handlers at the entry of page fault.
4380 static inline void mm_account_fault(struct pt_regs *regs,
4381 unsigned long address, unsigned int flags,
4387 * We don't do accounting for some specific faults:
4389 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4390 * includes arch_vma_access_permitted() failing before reaching here.
4391 * So this is not a "this many hardware page faults" counter. We
4392 * should use the hw profiling for that.
4394 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4395 * once they're completed.
4397 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4401 * We define the fault as a major fault when the final successful fault
4402 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4403 * handle it immediately previously).
4405 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4413 * If the fault is done for GUP, regs will be NULL. We only do the
4414 * accounting for the per thread fault counters who triggered the
4415 * fault, and we skip the perf event updates.
4421 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4423 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4427 * By the time we get here, we already hold the mm semaphore
4429 * The mmap_lock may have been released depending on flags and our
4430 * return value. See filemap_fault() and __lock_page_or_retry().
4432 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4433 unsigned int flags, struct pt_regs *regs)
4437 __set_current_state(TASK_RUNNING);
4439 count_vm_event(PGFAULT);
4440 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4442 /* do counter updates before entering really critical section. */
4443 check_sync_rss_stat(current);
4445 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4446 flags & FAULT_FLAG_INSTRUCTION,
4447 flags & FAULT_FLAG_REMOTE))
4448 return VM_FAULT_SIGSEGV;
4451 * Enable the memcg OOM handling for faults triggered in user
4452 * space. Kernel faults are handled more gracefully.
4454 if (flags & FAULT_FLAG_USER)
4455 mem_cgroup_enter_user_fault();
4457 if (unlikely(is_vm_hugetlb_page(vma)))
4458 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4460 ret = __handle_mm_fault(vma, address, flags);
4462 if (flags & FAULT_FLAG_USER) {
4463 mem_cgroup_exit_user_fault();
4465 * The task may have entered a memcg OOM situation but
4466 * if the allocation error was handled gracefully (no
4467 * VM_FAULT_OOM), there is no need to kill anything.
4468 * Just clean up the OOM state peacefully.
4470 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4471 mem_cgroup_oom_synchronize(false);
4474 mm_account_fault(regs, address, flags, ret);
4478 EXPORT_SYMBOL_GPL(handle_mm_fault);
4480 #ifndef __PAGETABLE_P4D_FOLDED
4482 * Allocate p4d page table.
4483 * We've already handled the fast-path in-line.
4485 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4487 p4d_t *new = p4d_alloc_one(mm, address);
4491 smp_wmb(); /* See comment in __pte_alloc */
4493 spin_lock(&mm->page_table_lock);
4494 if (pgd_present(*pgd)) /* Another has populated it */
4497 pgd_populate(mm, pgd, new);
4498 spin_unlock(&mm->page_table_lock);
4501 #endif /* __PAGETABLE_P4D_FOLDED */
4503 #ifndef __PAGETABLE_PUD_FOLDED
4505 * Allocate page upper directory.
4506 * We've already handled the fast-path in-line.
4508 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4510 pud_t *new = pud_alloc_one(mm, address);
4514 smp_wmb(); /* See comment in __pte_alloc */
4516 spin_lock(&mm->page_table_lock);
4517 if (!p4d_present(*p4d)) {
4519 p4d_populate(mm, p4d, new);
4520 } else /* Another has populated it */
4522 spin_unlock(&mm->page_table_lock);
4525 #endif /* __PAGETABLE_PUD_FOLDED */
4527 #ifndef __PAGETABLE_PMD_FOLDED
4529 * Allocate page middle directory.
4530 * We've already handled the fast-path in-line.
4532 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4535 pmd_t *new = pmd_alloc_one(mm, address);
4539 smp_wmb(); /* See comment in __pte_alloc */
4541 ptl = pud_lock(mm, pud);
4542 if (!pud_present(*pud)) {
4544 pud_populate(mm, pud, new);
4545 } else /* Another has populated it */
4550 #endif /* __PAGETABLE_PMD_FOLDED */
4552 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4553 struct mmu_notifier_range *range,
4554 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4562 pgd = pgd_offset(mm, address);
4563 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4566 p4d = p4d_offset(pgd, address);
4567 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4570 pud = pud_offset(p4d, address);
4571 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4574 pmd = pmd_offset(pud, address);
4575 VM_BUG_ON(pmd_trans_huge(*pmd));
4577 if (pmd_huge(*pmd)) {
4582 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4583 NULL, mm, address & PMD_MASK,
4584 (address & PMD_MASK) + PMD_SIZE);
4585 mmu_notifier_invalidate_range_start(range);
4587 *ptlp = pmd_lock(mm, pmd);
4588 if (pmd_huge(*pmd)) {
4594 mmu_notifier_invalidate_range_end(range);
4597 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4601 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4602 address & PAGE_MASK,
4603 (address & PAGE_MASK) + PAGE_SIZE);
4604 mmu_notifier_invalidate_range_start(range);
4606 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4607 if (!pte_present(*ptep))
4612 pte_unmap_unlock(ptep, *ptlp);
4614 mmu_notifier_invalidate_range_end(range);
4619 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4620 pte_t **ptepp, spinlock_t **ptlp)
4624 /* (void) is needed to make gcc happy */
4625 (void) __cond_lock(*ptlp,
4626 !(res = __follow_pte_pmd(mm, address, NULL,
4627 ptepp, NULL, ptlp)));
4631 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4632 struct mmu_notifier_range *range,
4633 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4637 /* (void) is needed to make gcc happy */
4638 (void) __cond_lock(*ptlp,
4639 !(res = __follow_pte_pmd(mm, address, range,
4640 ptepp, pmdpp, ptlp)));
4643 EXPORT_SYMBOL(follow_pte_pmd);
4646 * follow_pfn - look up PFN at a user virtual address
4647 * @vma: memory mapping
4648 * @address: user virtual address
4649 * @pfn: location to store found PFN
4651 * Only IO mappings and raw PFN mappings are allowed.
4653 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4655 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4662 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4665 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4668 *pfn = pte_pfn(*ptep);
4669 pte_unmap_unlock(ptep, ptl);
4672 EXPORT_SYMBOL(follow_pfn);
4674 #ifdef CONFIG_HAVE_IOREMAP_PROT
4675 int follow_phys(struct vm_area_struct *vma,
4676 unsigned long address, unsigned int flags,
4677 unsigned long *prot, resource_size_t *phys)
4683 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4686 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4690 if ((flags & FOLL_WRITE) && !pte_write(pte))
4693 *prot = pgprot_val(pte_pgprot(pte));
4694 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4698 pte_unmap_unlock(ptep, ptl);
4703 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4704 void *buf, int len, int write)
4706 resource_size_t phys_addr;
4707 unsigned long prot = 0;
4708 void __iomem *maddr;
4709 int offset = addr & (PAGE_SIZE-1);
4711 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4714 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4719 memcpy_toio(maddr + offset, buf, len);
4721 memcpy_fromio(buf, maddr + offset, len);
4726 EXPORT_SYMBOL_GPL(generic_access_phys);
4730 * Access another process' address space as given in mm. If non-NULL, use the
4731 * given task for page fault accounting.
4733 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4734 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4736 struct vm_area_struct *vma;
4737 void *old_buf = buf;
4738 int write = gup_flags & FOLL_WRITE;
4740 if (mmap_read_lock_killable(mm))
4743 /* ignore errors, just check how much was successfully transferred */
4745 int bytes, ret, offset;
4747 struct page *page = NULL;
4749 ret = get_user_pages_remote(mm, addr, 1,
4750 gup_flags, &page, &vma, NULL);
4752 #ifndef CONFIG_HAVE_IOREMAP_PROT
4756 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4757 * we can access using slightly different code.
4759 vma = find_vma(mm, addr);
4760 if (!vma || vma->vm_start > addr)
4762 if (vma->vm_ops && vma->vm_ops->access)
4763 ret = vma->vm_ops->access(vma, addr, buf,
4771 offset = addr & (PAGE_SIZE-1);
4772 if (bytes > PAGE_SIZE-offset)
4773 bytes = PAGE_SIZE-offset;
4777 copy_to_user_page(vma, page, addr,
4778 maddr + offset, buf, bytes);
4779 set_page_dirty_lock(page);
4781 copy_from_user_page(vma, page, addr,
4782 buf, maddr + offset, bytes);
4791 mmap_read_unlock(mm);
4793 return buf - old_buf;
4797 * access_remote_vm - access another process' address space
4798 * @mm: the mm_struct of the target address space
4799 * @addr: start address to access
4800 * @buf: source or destination buffer
4801 * @len: number of bytes to transfer
4802 * @gup_flags: flags modifying lookup behaviour
4804 * The caller must hold a reference on @mm.
4806 * Return: number of bytes copied from source to destination.
4808 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4809 void *buf, int len, unsigned int gup_flags)
4811 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4815 * Access another process' address space.
4816 * Source/target buffer must be kernel space,
4817 * Do not walk the page table directly, use get_user_pages
4819 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4820 void *buf, int len, unsigned int gup_flags)
4822 struct mm_struct *mm;
4825 mm = get_task_mm(tsk);
4829 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4835 EXPORT_SYMBOL_GPL(access_process_vm);
4838 * Print the name of a VMA.
4840 void print_vma_addr(char *prefix, unsigned long ip)
4842 struct mm_struct *mm = current->mm;
4843 struct vm_area_struct *vma;
4846 * we might be running from an atomic context so we cannot sleep
4848 if (!mmap_read_trylock(mm))
4851 vma = find_vma(mm, ip);
4852 if (vma && vma->vm_file) {
4853 struct file *f = vma->vm_file;
4854 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4858 p = file_path(f, buf, PAGE_SIZE);
4861 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4863 vma->vm_end - vma->vm_start);
4864 free_page((unsigned long)buf);
4867 mmap_read_unlock(mm);
4870 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4871 void __might_fault(const char *file, int line)
4874 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4875 * holding the mmap_lock, this is safe because kernel memory doesn't
4876 * get paged out, therefore we'll never actually fault, and the
4877 * below annotations will generate false positives.
4879 if (uaccess_kernel())
4881 if (pagefault_disabled())
4883 __might_sleep(file, line, 0);
4884 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4886 might_lock_read(¤t->mm->mmap_lock);
4889 EXPORT_SYMBOL(__might_fault);
4892 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4894 * Process all subpages of the specified huge page with the specified
4895 * operation. The target subpage will be processed last to keep its
4898 static inline void process_huge_page(
4899 unsigned long addr_hint, unsigned int pages_per_huge_page,
4900 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4904 unsigned long addr = addr_hint &
4905 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4907 /* Process target subpage last to keep its cache lines hot */
4909 n = (addr_hint - addr) / PAGE_SIZE;
4910 if (2 * n <= pages_per_huge_page) {
4911 /* If target subpage in first half of huge page */
4914 /* Process subpages at the end of huge page */
4915 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4917 process_subpage(addr + i * PAGE_SIZE, i, arg);
4920 /* If target subpage in second half of huge page */
4921 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4922 l = pages_per_huge_page - n;
4923 /* Process subpages at the begin of huge page */
4924 for (i = 0; i < base; i++) {
4926 process_subpage(addr + i * PAGE_SIZE, i, arg);
4930 * Process remaining subpages in left-right-left-right pattern
4931 * towards the target subpage
4933 for (i = 0; i < l; i++) {
4934 int left_idx = base + i;
4935 int right_idx = base + 2 * l - 1 - i;
4938 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4940 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4944 static void clear_gigantic_page(struct page *page,
4946 unsigned int pages_per_huge_page)
4949 struct page *p = page;
4952 for (i = 0; i < pages_per_huge_page;
4953 i++, p = mem_map_next(p, page, i)) {
4955 clear_user_highpage(p, addr + i * PAGE_SIZE);
4959 static void clear_subpage(unsigned long addr, int idx, void *arg)
4961 struct page *page = arg;
4963 clear_user_highpage(page + idx, addr);
4966 void clear_huge_page(struct page *page,
4967 unsigned long addr_hint, unsigned int pages_per_huge_page)
4969 unsigned long addr = addr_hint &
4970 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4972 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4973 clear_gigantic_page(page, addr, pages_per_huge_page);
4977 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4980 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4982 struct vm_area_struct *vma,
4983 unsigned int pages_per_huge_page)
4986 struct page *dst_base = dst;
4987 struct page *src_base = src;
4989 for (i = 0; i < pages_per_huge_page; ) {
4991 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4994 dst = mem_map_next(dst, dst_base, i);
4995 src = mem_map_next(src, src_base, i);
4999 struct copy_subpage_arg {
5002 struct vm_area_struct *vma;
5005 static void copy_subpage(unsigned long addr, int idx, void *arg)
5007 struct copy_subpage_arg *copy_arg = arg;
5009 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5010 addr, copy_arg->vma);
5013 void copy_user_huge_page(struct page *dst, struct page *src,
5014 unsigned long addr_hint, struct vm_area_struct *vma,
5015 unsigned int pages_per_huge_page)
5017 unsigned long addr = addr_hint &
5018 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5019 struct copy_subpage_arg arg = {
5025 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5026 copy_user_gigantic_page(dst, src, addr, vma,
5027 pages_per_huge_page);
5031 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5034 long copy_huge_page_from_user(struct page *dst_page,
5035 const void __user *usr_src,
5036 unsigned int pages_per_huge_page,
5037 bool allow_pagefault)
5039 void *src = (void *)usr_src;
5041 unsigned long i, rc = 0;
5042 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5044 for (i = 0; i < pages_per_huge_page; i++) {
5045 if (allow_pagefault)
5046 page_kaddr = kmap(dst_page + i);
5048 page_kaddr = kmap_atomic(dst_page + i);
5049 rc = copy_from_user(page_kaddr,
5050 (const void __user *)(src + i * PAGE_SIZE),
5052 if (allow_pagefault)
5053 kunmap(dst_page + i);
5055 kunmap_atomic(page_kaddr);
5057 ret_val -= (PAGE_SIZE - rc);
5065 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5067 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5069 static struct kmem_cache *page_ptl_cachep;
5071 void __init ptlock_cache_init(void)
5073 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5077 bool ptlock_alloc(struct page *page)
5081 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5088 void ptlock_free(struct page *page)
5090 kmem_cache_free(page_ptl_cachep, page->ptl);