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>
77 #include <trace/events/kmem.h>
80 #include <asm/mmu_context.h>
81 #include <asm/pgalloc.h>
82 #include <linux/uaccess.h>
84 #include <asm/tlbflush.h>
88 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
89 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
92 #ifndef CONFIG_NEED_MULTIPLE_NODES
93 /* use the per-pgdat data instead for discontigmem - mbligh */
94 unsigned long max_mapnr;
95 EXPORT_SYMBOL(max_mapnr);
98 EXPORT_SYMBOL(mem_map);
102 * A number of key systems in x86 including ioremap() rely on the assumption
103 * that high_memory defines the upper bound on direct map memory, then end
104 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
105 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
109 EXPORT_SYMBOL(high_memory);
112 * Randomize the address space (stacks, mmaps, brk, etc.).
114 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
115 * as ancient (libc5 based) binaries can segfault. )
117 int randomize_va_space __read_mostly =
118 #ifdef CONFIG_COMPAT_BRK
124 #ifndef arch_faults_on_old_pte
125 static inline bool arch_faults_on_old_pte(void)
128 * Those arches which don't have hw access flag feature need to
129 * implement their own helper. By default, "true" means pagefault
130 * will be hit on old pte.
136 static int __init disable_randmaps(char *s)
138 randomize_va_space = 0;
141 __setup("norandmaps", disable_randmaps);
143 unsigned long zero_pfn __read_mostly;
144 EXPORT_SYMBOL(zero_pfn);
146 unsigned long highest_memmap_pfn __read_mostly;
149 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
151 static int __init init_zero_pfn(void)
153 zero_pfn = page_to_pfn(ZERO_PAGE(0));
156 core_initcall(init_zero_pfn);
158 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
160 trace_rss_stat(mm, member, count);
163 #if defined(SPLIT_RSS_COUNTING)
165 void sync_mm_rss(struct mm_struct *mm)
169 for (i = 0; i < NR_MM_COUNTERS; i++) {
170 if (current->rss_stat.count[i]) {
171 add_mm_counter(mm, i, current->rss_stat.count[i]);
172 current->rss_stat.count[i] = 0;
175 current->rss_stat.events = 0;
178 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
180 struct task_struct *task = current;
182 if (likely(task->mm == mm))
183 task->rss_stat.count[member] += val;
185 add_mm_counter(mm, member, val);
187 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
188 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
190 /* sync counter once per 64 page faults */
191 #define TASK_RSS_EVENTS_THRESH (64)
192 static void check_sync_rss_stat(struct task_struct *task)
194 if (unlikely(task != current))
196 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
197 sync_mm_rss(task->mm);
199 #else /* SPLIT_RSS_COUNTING */
201 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
202 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
204 static void check_sync_rss_stat(struct task_struct *task)
208 #endif /* SPLIT_RSS_COUNTING */
211 * Note: this doesn't free the actual pages themselves. That
212 * has been handled earlier when unmapping all the memory regions.
214 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
217 pgtable_t token = pmd_pgtable(*pmd);
219 pte_free_tlb(tlb, token, addr);
220 mm_dec_nr_ptes(tlb->mm);
223 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
224 unsigned long addr, unsigned long end,
225 unsigned long floor, unsigned long ceiling)
232 pmd = pmd_offset(pud, addr);
234 next = pmd_addr_end(addr, end);
235 if (pmd_none_or_clear_bad(pmd))
237 free_pte_range(tlb, pmd, addr);
238 } while (pmd++, addr = next, addr != end);
248 if (end - 1 > ceiling - 1)
251 pmd = pmd_offset(pud, start);
253 pmd_free_tlb(tlb, pmd, start);
254 mm_dec_nr_pmds(tlb->mm);
257 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
258 unsigned long addr, unsigned long end,
259 unsigned long floor, unsigned long ceiling)
266 pud = pud_offset(p4d, addr);
268 next = pud_addr_end(addr, end);
269 if (pud_none_or_clear_bad(pud))
271 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
272 } while (pud++, addr = next, addr != end);
282 if (end - 1 > ceiling - 1)
285 pud = pud_offset(p4d, start);
287 pud_free_tlb(tlb, pud, start);
288 mm_dec_nr_puds(tlb->mm);
291 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
292 unsigned long addr, unsigned long end,
293 unsigned long floor, unsigned long ceiling)
300 p4d = p4d_offset(pgd, addr);
302 next = p4d_addr_end(addr, end);
303 if (p4d_none_or_clear_bad(p4d))
305 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
306 } while (p4d++, addr = next, addr != end);
312 ceiling &= PGDIR_MASK;
316 if (end - 1 > ceiling - 1)
319 p4d = p4d_offset(pgd, start);
321 p4d_free_tlb(tlb, p4d, start);
325 * This function frees user-level page tables of a process.
327 void free_pgd_range(struct mmu_gather *tlb,
328 unsigned long addr, unsigned long end,
329 unsigned long floor, unsigned long ceiling)
335 * The next few lines have given us lots of grief...
337 * Why are we testing PMD* at this top level? Because often
338 * there will be no work to do at all, and we'd prefer not to
339 * go all the way down to the bottom just to discover that.
341 * Why all these "- 1"s? Because 0 represents both the bottom
342 * of the address space and the top of it (using -1 for the
343 * top wouldn't help much: the masks would do the wrong thing).
344 * The rule is that addr 0 and floor 0 refer to the bottom of
345 * the address space, but end 0 and ceiling 0 refer to the top
346 * Comparisons need to use "end - 1" and "ceiling - 1" (though
347 * that end 0 case should be mythical).
349 * Wherever addr is brought up or ceiling brought down, we must
350 * be careful to reject "the opposite 0" before it confuses the
351 * subsequent tests. But what about where end is brought down
352 * by PMD_SIZE below? no, end can't go down to 0 there.
354 * Whereas we round start (addr) and ceiling down, by different
355 * masks at different levels, in order to test whether a table
356 * now has no other vmas using it, so can be freed, we don't
357 * bother to round floor or end up - the tests don't need that.
371 if (end - 1 > ceiling - 1)
376 * We add page table cache pages with PAGE_SIZE,
377 * (see pte_free_tlb()), flush the tlb if we need
379 tlb_change_page_size(tlb, PAGE_SIZE);
380 pgd = pgd_offset(tlb->mm, addr);
382 next = pgd_addr_end(addr, end);
383 if (pgd_none_or_clear_bad(pgd))
385 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
386 } while (pgd++, addr = next, addr != end);
389 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
390 unsigned long floor, unsigned long ceiling)
393 struct vm_area_struct *next = vma->vm_next;
394 unsigned long addr = vma->vm_start;
397 * Hide vma from rmap and truncate_pagecache before freeing
400 unlink_anon_vmas(vma);
401 unlink_file_vma(vma);
403 if (is_vm_hugetlb_page(vma)) {
404 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
405 floor, next ? next->vm_start : ceiling);
408 * Optimization: gather nearby vmas into one call down
410 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
411 && !is_vm_hugetlb_page(next)) {
414 unlink_anon_vmas(vma);
415 unlink_file_vma(vma);
417 free_pgd_range(tlb, addr, vma->vm_end,
418 floor, next ? next->vm_start : ceiling);
424 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
427 pgtable_t new = pte_alloc_one(mm);
432 * Ensure all pte setup (eg. pte page lock and page clearing) are
433 * visible before the pte is made visible to other CPUs by being
434 * put into page tables.
436 * The other side of the story is the pointer chasing in the page
437 * table walking code (when walking the page table without locking;
438 * ie. most of the time). Fortunately, these data accesses consist
439 * of a chain of data-dependent loads, meaning most CPUs (alpha
440 * being the notable exception) will already guarantee loads are
441 * seen in-order. See the alpha page table accessors for the
442 * smp_rmb() barriers in page table walking code.
444 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
446 ptl = pmd_lock(mm, pmd);
447 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
449 pmd_populate(mm, pmd, new);
458 int __pte_alloc_kernel(pmd_t *pmd)
460 pte_t *new = pte_alloc_one_kernel(&init_mm);
464 smp_wmb(); /* See comment in __pte_alloc */
466 spin_lock(&init_mm.page_table_lock);
467 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
468 pmd_populate_kernel(&init_mm, pmd, new);
471 spin_unlock(&init_mm.page_table_lock);
473 pte_free_kernel(&init_mm, new);
477 static inline void init_rss_vec(int *rss)
479 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
482 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
486 if (current->mm == mm)
488 for (i = 0; i < NR_MM_COUNTERS; i++)
490 add_mm_counter(mm, i, rss[i]);
494 * This function is called to print an error when a bad pte
495 * is found. For example, we might have a PFN-mapped pte in
496 * a region that doesn't allow it.
498 * The calling function must still handle the error.
500 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
501 pte_t pte, struct page *page)
503 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
504 p4d_t *p4d = p4d_offset(pgd, addr);
505 pud_t *pud = pud_offset(p4d, addr);
506 pmd_t *pmd = pmd_offset(pud, addr);
507 struct address_space *mapping;
509 static unsigned long resume;
510 static unsigned long nr_shown;
511 static unsigned long nr_unshown;
514 * Allow a burst of 60 reports, then keep quiet for that minute;
515 * or allow a steady drip of one report per second.
517 if (nr_shown == 60) {
518 if (time_before(jiffies, resume)) {
523 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
530 resume = jiffies + 60 * HZ;
532 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
533 index = linear_page_index(vma, addr);
535 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
537 (long long)pte_val(pte), (long long)pmd_val(*pmd));
539 dump_page(page, "bad pte");
540 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
541 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
542 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
544 vma->vm_ops ? vma->vm_ops->fault : NULL,
545 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
546 mapping ? mapping->a_ops->readpage : NULL);
548 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
552 * vm_normal_page -- This function gets the "struct page" associated with a pte.
554 * "Special" mappings do not wish to be associated with a "struct page" (either
555 * it doesn't exist, or it exists but they don't want to touch it). In this
556 * case, NULL is returned here. "Normal" mappings do have a struct page.
558 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
559 * pte bit, in which case this function is trivial. Secondly, an architecture
560 * may not have a spare pte bit, which requires a more complicated scheme,
563 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
564 * special mapping (even if there are underlying and valid "struct pages").
565 * COWed pages of a VM_PFNMAP are always normal.
567 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
568 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
569 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
570 * mapping will always honor the rule
572 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
574 * And for normal mappings this is false.
576 * This restricts such mappings to be a linear translation from virtual address
577 * to pfn. To get around this restriction, we allow arbitrary mappings so long
578 * as the vma is not a COW mapping; in that case, we know that all ptes are
579 * special (because none can have been COWed).
582 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
584 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
585 * page" backing, however the difference is that _all_ pages with a struct
586 * page (that is, those where pfn_valid is true) are refcounted and considered
587 * normal pages by the VM. The disadvantage is that pages are refcounted
588 * (which can be slower and simply not an option for some PFNMAP users). The
589 * advantage is that we don't have to follow the strict linearity rule of
590 * PFNMAP mappings in order to support COWable mappings.
593 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
596 unsigned long pfn = pte_pfn(pte);
598 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
599 if (likely(!pte_special(pte)))
601 if (vma->vm_ops && vma->vm_ops->find_special_page)
602 return vma->vm_ops->find_special_page(vma, addr);
603 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
605 if (is_zero_pfn(pfn))
610 print_bad_pte(vma, addr, pte, NULL);
614 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
616 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
617 if (vma->vm_flags & VM_MIXEDMAP) {
623 off = (addr - vma->vm_start) >> PAGE_SHIFT;
624 if (pfn == vma->vm_pgoff + off)
626 if (!is_cow_mapping(vma->vm_flags))
631 if (is_zero_pfn(pfn))
635 if (unlikely(pfn > highest_memmap_pfn)) {
636 print_bad_pte(vma, addr, pte, NULL);
641 * NOTE! We still have PageReserved() pages in the page tables.
642 * eg. VDSO mappings can cause them to exist.
645 return pfn_to_page(pfn);
648 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
649 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
652 unsigned long pfn = pmd_pfn(pmd);
655 * There is no pmd_special() but there may be special pmds, e.g.
656 * in a direct-access (dax) mapping, so let's just replicate the
657 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
659 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
660 if (vma->vm_flags & VM_MIXEDMAP) {
666 off = (addr - vma->vm_start) >> PAGE_SHIFT;
667 if (pfn == vma->vm_pgoff + off)
669 if (!is_cow_mapping(vma->vm_flags))
676 if (is_huge_zero_pmd(pmd))
678 if (unlikely(pfn > highest_memmap_pfn))
682 * NOTE! We still have PageReserved() pages in the page tables.
683 * eg. VDSO mappings can cause them to exist.
686 return pfn_to_page(pfn);
691 * copy one vm_area from one task to the other. Assumes the page tables
692 * already present in the new task to be cleared in the whole range
693 * covered by this vma.
696 static inline unsigned long
697 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
698 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
699 unsigned long addr, int *rss)
701 unsigned long vm_flags = vma->vm_flags;
702 pte_t pte = *src_pte;
705 /* pte contains position in swap or file, so copy. */
706 if (unlikely(!pte_present(pte))) {
707 swp_entry_t entry = pte_to_swp_entry(pte);
709 if (likely(!non_swap_entry(entry))) {
710 if (swap_duplicate(entry) < 0)
713 /* make sure dst_mm is on swapoff's mmlist. */
714 if (unlikely(list_empty(&dst_mm->mmlist))) {
715 spin_lock(&mmlist_lock);
716 if (list_empty(&dst_mm->mmlist))
717 list_add(&dst_mm->mmlist,
719 spin_unlock(&mmlist_lock);
722 } else if (is_migration_entry(entry)) {
723 page = migration_entry_to_page(entry);
725 rss[mm_counter(page)]++;
727 if (is_write_migration_entry(entry) &&
728 is_cow_mapping(vm_flags)) {
730 * COW mappings require pages in both
731 * parent and child to be set to read.
733 make_migration_entry_read(&entry);
734 pte = swp_entry_to_pte(entry);
735 if (pte_swp_soft_dirty(*src_pte))
736 pte = pte_swp_mksoft_dirty(pte);
737 if (pte_swp_uffd_wp(*src_pte))
738 pte = pte_swp_mkuffd_wp(pte);
739 set_pte_at(src_mm, addr, src_pte, pte);
741 } else if (is_device_private_entry(entry)) {
742 page = device_private_entry_to_page(entry);
745 * Update rss count even for unaddressable pages, as
746 * they should treated just like normal pages in this
749 * We will likely want to have some new rss counters
750 * for unaddressable pages, at some point. But for now
751 * keep things as they are.
754 rss[mm_counter(page)]++;
755 page_dup_rmap(page, false);
758 * We do not preserve soft-dirty information, because so
759 * far, checkpoint/restore is the only feature that
760 * requires that. And checkpoint/restore does not work
761 * when a device driver is involved (you cannot easily
762 * save and restore device driver state).
764 if (is_write_device_private_entry(entry) &&
765 is_cow_mapping(vm_flags)) {
766 make_device_private_entry_read(&entry);
767 pte = swp_entry_to_pte(entry);
768 if (pte_swp_uffd_wp(*src_pte))
769 pte = pte_swp_mkuffd_wp(pte);
770 set_pte_at(src_mm, addr, src_pte, pte);
777 * If it's a COW mapping, write protect it both
778 * in the parent and the child
780 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
781 ptep_set_wrprotect(src_mm, addr, src_pte);
782 pte = pte_wrprotect(pte);
786 * If it's a shared mapping, mark it clean in
789 if (vm_flags & VM_SHARED)
790 pte = pte_mkclean(pte);
791 pte = pte_mkold(pte);
794 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
795 * does not have the VM_UFFD_WP, which means that the uffd
796 * fork event is not enabled.
798 if (!(vm_flags & VM_UFFD_WP))
799 pte = pte_clear_uffd_wp(pte);
801 page = vm_normal_page(vma, addr, pte);
804 page_dup_rmap(page, false);
805 rss[mm_counter(page)]++;
809 set_pte_at(dst_mm, addr, dst_pte, pte);
813 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
814 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
815 unsigned long addr, unsigned long end)
817 pte_t *orig_src_pte, *orig_dst_pte;
818 pte_t *src_pte, *dst_pte;
819 spinlock_t *src_ptl, *dst_ptl;
821 int rss[NR_MM_COUNTERS];
822 swp_entry_t entry = (swp_entry_t){0};
827 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
830 src_pte = pte_offset_map(src_pmd, addr);
831 src_ptl = pte_lockptr(src_mm, src_pmd);
832 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
833 orig_src_pte = src_pte;
834 orig_dst_pte = dst_pte;
835 arch_enter_lazy_mmu_mode();
839 * We are holding two locks at this point - either of them
840 * could generate latencies in another task on another CPU.
842 if (progress >= 32) {
844 if (need_resched() ||
845 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
848 if (pte_none(*src_pte)) {
852 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
857 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
859 arch_leave_lazy_mmu_mode();
860 spin_unlock(src_ptl);
861 pte_unmap(orig_src_pte);
862 add_mm_rss_vec(dst_mm, rss);
863 pte_unmap_unlock(orig_dst_pte, dst_ptl);
867 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
876 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
877 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
878 unsigned long addr, unsigned long end)
880 pmd_t *src_pmd, *dst_pmd;
883 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
886 src_pmd = pmd_offset(src_pud, addr);
888 next = pmd_addr_end(addr, end);
889 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
890 || pmd_devmap(*src_pmd)) {
892 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
893 err = copy_huge_pmd(dst_mm, src_mm,
894 dst_pmd, src_pmd, addr, vma);
901 if (pmd_none_or_clear_bad(src_pmd))
903 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
906 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
910 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
911 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
912 unsigned long addr, unsigned long end)
914 pud_t *src_pud, *dst_pud;
917 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
920 src_pud = pud_offset(src_p4d, addr);
922 next = pud_addr_end(addr, end);
923 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
926 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
927 err = copy_huge_pud(dst_mm, src_mm,
928 dst_pud, src_pud, addr, vma);
935 if (pud_none_or_clear_bad(src_pud))
937 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
940 } while (dst_pud++, src_pud++, addr = next, addr != end);
944 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
945 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
946 unsigned long addr, unsigned long end)
948 p4d_t *src_p4d, *dst_p4d;
951 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
954 src_p4d = p4d_offset(src_pgd, addr);
956 next = p4d_addr_end(addr, end);
957 if (p4d_none_or_clear_bad(src_p4d))
959 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
962 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
966 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
967 struct vm_area_struct *vma)
969 pgd_t *src_pgd, *dst_pgd;
971 unsigned long addr = vma->vm_start;
972 unsigned long end = vma->vm_end;
973 struct mmu_notifier_range range;
978 * Don't copy ptes where a page fault will fill them correctly.
979 * Fork becomes much lighter when there are big shared or private
980 * readonly mappings. The tradeoff is that copy_page_range is more
981 * efficient than faulting.
983 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
987 if (is_vm_hugetlb_page(vma))
988 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
990 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
992 * We do not free on error cases below as remove_vma
993 * gets called on error from higher level routine
995 ret = track_pfn_copy(vma);
1001 * We need to invalidate the secondary MMU mappings only when
1002 * there could be a permission downgrade on the ptes of the
1003 * parent mm. And a permission downgrade will only happen if
1004 * is_cow_mapping() returns true.
1006 is_cow = is_cow_mapping(vma->vm_flags);
1009 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1010 0, vma, src_mm, addr, end);
1011 mmu_notifier_invalidate_range_start(&range);
1015 dst_pgd = pgd_offset(dst_mm, addr);
1016 src_pgd = pgd_offset(src_mm, addr);
1018 next = pgd_addr_end(addr, end);
1019 if (pgd_none_or_clear_bad(src_pgd))
1021 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1022 vma, addr, next))) {
1026 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1029 mmu_notifier_invalidate_range_end(&range);
1033 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1034 struct vm_area_struct *vma, pmd_t *pmd,
1035 unsigned long addr, unsigned long end,
1036 struct zap_details *details)
1038 struct mm_struct *mm = tlb->mm;
1039 int force_flush = 0;
1040 int rss[NR_MM_COUNTERS];
1046 tlb_change_page_size(tlb, PAGE_SIZE);
1049 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1051 flush_tlb_batched_pending(mm);
1052 arch_enter_lazy_mmu_mode();
1055 if (pte_none(ptent))
1061 if (pte_present(ptent)) {
1064 page = vm_normal_page(vma, addr, ptent);
1065 if (unlikely(details) && page) {
1067 * unmap_shared_mapping_pages() wants to
1068 * invalidate cache without truncating:
1069 * unmap shared but keep private pages.
1071 if (details->check_mapping &&
1072 details->check_mapping != page_rmapping(page))
1075 ptent = ptep_get_and_clear_full(mm, addr, pte,
1077 tlb_remove_tlb_entry(tlb, pte, addr);
1078 if (unlikely(!page))
1081 if (!PageAnon(page)) {
1082 if (pte_dirty(ptent)) {
1084 set_page_dirty(page);
1086 if (pte_young(ptent) &&
1087 likely(!(vma->vm_flags & VM_SEQ_READ)))
1088 mark_page_accessed(page);
1090 rss[mm_counter(page)]--;
1091 page_remove_rmap(page, false);
1092 if (unlikely(page_mapcount(page) < 0))
1093 print_bad_pte(vma, addr, ptent, page);
1094 if (unlikely(__tlb_remove_page(tlb, page))) {
1102 entry = pte_to_swp_entry(ptent);
1103 if (is_device_private_entry(entry)) {
1104 struct page *page = device_private_entry_to_page(entry);
1106 if (unlikely(details && details->check_mapping)) {
1108 * unmap_shared_mapping_pages() wants to
1109 * invalidate cache without truncating:
1110 * unmap shared but keep private pages.
1112 if (details->check_mapping !=
1113 page_rmapping(page))
1117 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1118 rss[mm_counter(page)]--;
1119 page_remove_rmap(page, false);
1124 /* If details->check_mapping, we leave swap entries. */
1125 if (unlikely(details))
1128 if (!non_swap_entry(entry))
1130 else if (is_migration_entry(entry)) {
1133 page = migration_entry_to_page(entry);
1134 rss[mm_counter(page)]--;
1136 if (unlikely(!free_swap_and_cache(entry)))
1137 print_bad_pte(vma, addr, ptent, NULL);
1138 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1139 } while (pte++, addr += PAGE_SIZE, addr != end);
1141 add_mm_rss_vec(mm, rss);
1142 arch_leave_lazy_mmu_mode();
1144 /* Do the actual TLB flush before dropping ptl */
1146 tlb_flush_mmu_tlbonly(tlb);
1147 pte_unmap_unlock(start_pte, ptl);
1150 * If we forced a TLB flush (either due to running out of
1151 * batch buffers or because we needed to flush dirty TLB
1152 * entries before releasing the ptl), free the batched
1153 * memory too. Restart if we didn't do everything.
1168 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1169 struct vm_area_struct *vma, pud_t *pud,
1170 unsigned long addr, unsigned long end,
1171 struct zap_details *details)
1176 pmd = pmd_offset(pud, addr);
1178 next = pmd_addr_end(addr, end);
1179 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1180 if (next - addr != HPAGE_PMD_SIZE)
1181 __split_huge_pmd(vma, pmd, addr, false, NULL);
1182 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1187 * Here there can be other concurrent MADV_DONTNEED or
1188 * trans huge page faults running, and if the pmd is
1189 * none or trans huge it can change under us. This is
1190 * because MADV_DONTNEED holds the mmap_lock in read
1193 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1195 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1198 } while (pmd++, addr = next, addr != end);
1203 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1204 struct vm_area_struct *vma, p4d_t *p4d,
1205 unsigned long addr, unsigned long end,
1206 struct zap_details *details)
1211 pud = pud_offset(p4d, addr);
1213 next = pud_addr_end(addr, end);
1214 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1215 if (next - addr != HPAGE_PUD_SIZE) {
1216 mmap_assert_locked(tlb->mm);
1217 split_huge_pud(vma, pud, addr);
1218 } else if (zap_huge_pud(tlb, vma, pud, addr))
1222 if (pud_none_or_clear_bad(pud))
1224 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1227 } while (pud++, addr = next, addr != end);
1232 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1233 struct vm_area_struct *vma, pgd_t *pgd,
1234 unsigned long addr, unsigned long end,
1235 struct zap_details *details)
1240 p4d = p4d_offset(pgd, addr);
1242 next = p4d_addr_end(addr, end);
1243 if (p4d_none_or_clear_bad(p4d))
1245 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1246 } while (p4d++, addr = next, addr != end);
1251 void unmap_page_range(struct mmu_gather *tlb,
1252 struct vm_area_struct *vma,
1253 unsigned long addr, unsigned long end,
1254 struct zap_details *details)
1259 BUG_ON(addr >= end);
1260 tlb_start_vma(tlb, vma);
1261 pgd = pgd_offset(vma->vm_mm, addr);
1263 next = pgd_addr_end(addr, end);
1264 if (pgd_none_or_clear_bad(pgd))
1266 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1267 } while (pgd++, addr = next, addr != end);
1268 tlb_end_vma(tlb, vma);
1272 static void unmap_single_vma(struct mmu_gather *tlb,
1273 struct vm_area_struct *vma, unsigned long start_addr,
1274 unsigned long end_addr,
1275 struct zap_details *details)
1277 unsigned long start = max(vma->vm_start, start_addr);
1280 if (start >= vma->vm_end)
1282 end = min(vma->vm_end, end_addr);
1283 if (end <= vma->vm_start)
1287 uprobe_munmap(vma, start, end);
1289 if (unlikely(vma->vm_flags & VM_PFNMAP))
1290 untrack_pfn(vma, 0, 0);
1293 if (unlikely(is_vm_hugetlb_page(vma))) {
1295 * It is undesirable to test vma->vm_file as it
1296 * should be non-null for valid hugetlb area.
1297 * However, vm_file will be NULL in the error
1298 * cleanup path of mmap_region. When
1299 * hugetlbfs ->mmap method fails,
1300 * mmap_region() nullifies vma->vm_file
1301 * before calling this function to clean up.
1302 * Since no pte has actually been setup, it is
1303 * safe to do nothing in this case.
1306 i_mmap_lock_write(vma->vm_file->f_mapping);
1307 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1308 i_mmap_unlock_write(vma->vm_file->f_mapping);
1311 unmap_page_range(tlb, vma, start, end, details);
1316 * unmap_vmas - unmap a range of memory covered by a list of vma's
1317 * @tlb: address of the caller's struct mmu_gather
1318 * @vma: the starting vma
1319 * @start_addr: virtual address at which to start unmapping
1320 * @end_addr: virtual address at which to end unmapping
1322 * Unmap all pages in the vma list.
1324 * Only addresses between `start' and `end' will be unmapped.
1326 * The VMA list must be sorted in ascending virtual address order.
1328 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1329 * range after unmap_vmas() returns. So the only responsibility here is to
1330 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1331 * drops the lock and schedules.
1333 void unmap_vmas(struct mmu_gather *tlb,
1334 struct vm_area_struct *vma, unsigned long start_addr,
1335 unsigned long end_addr)
1337 struct mmu_notifier_range range;
1339 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1340 start_addr, end_addr);
1341 mmu_notifier_invalidate_range_start(&range);
1342 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1343 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1344 mmu_notifier_invalidate_range_end(&range);
1348 * zap_page_range - remove user pages in a given range
1349 * @vma: vm_area_struct holding the applicable pages
1350 * @start: starting address of pages to zap
1351 * @size: number of bytes to zap
1353 * Caller must protect the VMA list
1355 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1358 struct mmu_notifier_range range;
1359 struct mmu_gather tlb;
1362 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1363 start, start + size);
1364 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1365 update_hiwater_rss(vma->vm_mm);
1366 mmu_notifier_invalidate_range_start(&range);
1367 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1368 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1369 mmu_notifier_invalidate_range_end(&range);
1370 tlb_finish_mmu(&tlb, start, range.end);
1374 * zap_page_range_single - remove user pages in a given range
1375 * @vma: vm_area_struct holding the applicable pages
1376 * @address: starting address of pages to zap
1377 * @size: number of bytes to zap
1378 * @details: details of shared cache invalidation
1380 * The range must fit into one VMA.
1382 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1383 unsigned long size, struct zap_details *details)
1385 struct mmu_notifier_range range;
1386 struct mmu_gather tlb;
1389 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1390 address, address + size);
1391 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1392 update_hiwater_rss(vma->vm_mm);
1393 mmu_notifier_invalidate_range_start(&range);
1394 unmap_single_vma(&tlb, vma, address, range.end, details);
1395 mmu_notifier_invalidate_range_end(&range);
1396 tlb_finish_mmu(&tlb, address, range.end);
1400 * zap_vma_ptes - remove ptes mapping the vma
1401 * @vma: vm_area_struct holding ptes to be zapped
1402 * @address: starting address of pages to zap
1403 * @size: number of bytes to zap
1405 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1407 * The entire address range must be fully contained within the vma.
1410 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1413 if (address < vma->vm_start || address + size > vma->vm_end ||
1414 !(vma->vm_flags & VM_PFNMAP))
1417 zap_page_range_single(vma, address, size, NULL);
1419 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1421 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1428 pgd = pgd_offset(mm, addr);
1429 p4d = p4d_alloc(mm, pgd, addr);
1432 pud = pud_alloc(mm, p4d, addr);
1435 pmd = pmd_alloc(mm, pud, addr);
1439 VM_BUG_ON(pmd_trans_huge(*pmd));
1443 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1446 pmd_t *pmd = walk_to_pmd(mm, addr);
1450 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1453 static int validate_page_before_insert(struct page *page)
1455 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1457 flush_dcache_page(page);
1461 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1462 unsigned long addr, struct page *page, pgprot_t prot)
1464 if (!pte_none(*pte))
1466 /* Ok, finally just insert the thing.. */
1468 inc_mm_counter_fast(mm, mm_counter_file(page));
1469 page_add_file_rmap(page, false);
1470 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1475 * This is the old fallback for page remapping.
1477 * For historical reasons, it only allows reserved pages. Only
1478 * old drivers should use this, and they needed to mark their
1479 * pages reserved for the old functions anyway.
1481 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1482 struct page *page, pgprot_t prot)
1484 struct mm_struct *mm = vma->vm_mm;
1489 retval = validate_page_before_insert(page);
1493 pte = get_locked_pte(mm, addr, &ptl);
1496 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1497 pte_unmap_unlock(pte, ptl);
1503 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1504 unsigned long addr, struct page *page, pgprot_t prot)
1508 if (!page_count(page))
1510 err = validate_page_before_insert(page);
1513 return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1516 /* insert_pages() amortizes the cost of spinlock operations
1517 * when inserting pages in a loop. Arch *must* define pte_index.
1519 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1520 struct page **pages, unsigned long *num, pgprot_t prot)
1523 pte_t *start_pte, *pte;
1524 spinlock_t *pte_lock;
1525 struct mm_struct *const mm = vma->vm_mm;
1526 unsigned long curr_page_idx = 0;
1527 unsigned long remaining_pages_total = *num;
1528 unsigned long pages_to_write_in_pmd;
1532 pmd = walk_to_pmd(mm, addr);
1536 pages_to_write_in_pmd = min_t(unsigned long,
1537 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1539 /* Allocate the PTE if necessary; takes PMD lock once only. */
1541 if (pte_alloc(mm, pmd))
1544 while (pages_to_write_in_pmd) {
1546 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1548 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1549 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1550 int err = insert_page_in_batch_locked(mm, pte,
1551 addr, pages[curr_page_idx], prot);
1552 if (unlikely(err)) {
1553 pte_unmap_unlock(start_pte, pte_lock);
1555 remaining_pages_total -= pte_idx;
1561 pte_unmap_unlock(start_pte, pte_lock);
1562 pages_to_write_in_pmd -= batch_size;
1563 remaining_pages_total -= batch_size;
1565 if (remaining_pages_total)
1569 *num = remaining_pages_total;
1572 #endif /* ifdef pte_index */
1575 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1576 * @vma: user vma to map to
1577 * @addr: target start user address of these pages
1578 * @pages: source kernel pages
1579 * @num: in: number of pages to map. out: number of pages that were *not*
1580 * mapped. (0 means all pages were successfully mapped).
1582 * Preferred over vm_insert_page() when inserting multiple pages.
1584 * In case of error, we may have mapped a subset of the provided
1585 * pages. It is the caller's responsibility to account for this case.
1587 * The same restrictions apply as in vm_insert_page().
1589 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1590 struct page **pages, unsigned long *num)
1593 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1595 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1597 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1598 BUG_ON(mmap_read_trylock(vma->vm_mm));
1599 BUG_ON(vma->vm_flags & VM_PFNMAP);
1600 vma->vm_flags |= VM_MIXEDMAP;
1602 /* Defer page refcount checking till we're about to map that page. */
1603 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1605 unsigned long idx = 0, pgcount = *num;
1608 for (; idx < pgcount; ++idx) {
1609 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1613 *num = pgcount - idx;
1615 #endif /* ifdef pte_index */
1617 EXPORT_SYMBOL(vm_insert_pages);
1620 * vm_insert_page - insert single page into user vma
1621 * @vma: user vma to map to
1622 * @addr: target user address of this page
1623 * @page: source kernel page
1625 * This allows drivers to insert individual pages they've allocated
1628 * The page has to be a nice clean _individual_ kernel allocation.
1629 * If you allocate a compound page, you need to have marked it as
1630 * such (__GFP_COMP), or manually just split the page up yourself
1631 * (see split_page()).
1633 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1634 * took an arbitrary page protection parameter. This doesn't allow
1635 * that. Your vma protection will have to be set up correctly, which
1636 * means that if you want a shared writable mapping, you'd better
1637 * ask for a shared writable mapping!
1639 * The page does not need to be reserved.
1641 * Usually this function is called from f_op->mmap() handler
1642 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1643 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1644 * function from other places, for example from page-fault handler.
1646 * Return: %0 on success, negative error code otherwise.
1648 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1651 if (addr < vma->vm_start || addr >= vma->vm_end)
1653 if (!page_count(page))
1655 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1656 BUG_ON(mmap_read_trylock(vma->vm_mm));
1657 BUG_ON(vma->vm_flags & VM_PFNMAP);
1658 vma->vm_flags |= VM_MIXEDMAP;
1660 return insert_page(vma, addr, page, vma->vm_page_prot);
1662 EXPORT_SYMBOL(vm_insert_page);
1665 * __vm_map_pages - maps range of kernel pages into user vma
1666 * @vma: user vma to map to
1667 * @pages: pointer to array of source kernel pages
1668 * @num: number of pages in page array
1669 * @offset: user's requested vm_pgoff
1671 * This allows drivers to map range of kernel pages into a user vma.
1673 * Return: 0 on success and error code otherwise.
1675 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1676 unsigned long num, unsigned long offset)
1678 unsigned long count = vma_pages(vma);
1679 unsigned long uaddr = vma->vm_start;
1682 /* Fail if the user requested offset is beyond the end of the object */
1686 /* Fail if the user requested size exceeds available object size */
1687 if (count > num - offset)
1690 for (i = 0; i < count; i++) {
1691 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1701 * vm_map_pages - maps range of kernel pages starts with non zero offset
1702 * @vma: user vma to map to
1703 * @pages: pointer to array of source kernel pages
1704 * @num: number of pages in page array
1706 * Maps an object consisting of @num pages, catering for the user's
1707 * requested vm_pgoff
1709 * If we fail to insert any page into the vma, the function will return
1710 * immediately leaving any previously inserted pages present. Callers
1711 * from the mmap handler may immediately return the error as their caller
1712 * will destroy the vma, removing any successfully inserted pages. Other
1713 * callers should make their own arrangements for calling unmap_region().
1715 * Context: Process context. Called by mmap handlers.
1716 * Return: 0 on success and error code otherwise.
1718 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1721 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1723 EXPORT_SYMBOL(vm_map_pages);
1726 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1727 * @vma: user vma to map to
1728 * @pages: pointer to array of source kernel pages
1729 * @num: number of pages in page array
1731 * Similar to vm_map_pages(), except that it explicitly sets the offset
1732 * to 0. This function is intended for the drivers that did not consider
1735 * Context: Process context. Called by mmap handlers.
1736 * Return: 0 on success and error code otherwise.
1738 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1741 return __vm_map_pages(vma, pages, num, 0);
1743 EXPORT_SYMBOL(vm_map_pages_zero);
1745 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1746 pfn_t pfn, pgprot_t prot, bool mkwrite)
1748 struct mm_struct *mm = vma->vm_mm;
1752 pte = get_locked_pte(mm, addr, &ptl);
1754 return VM_FAULT_OOM;
1755 if (!pte_none(*pte)) {
1758 * For read faults on private mappings the PFN passed
1759 * in may not match the PFN we have mapped if the
1760 * mapped PFN is a writeable COW page. In the mkwrite
1761 * case we are creating a writable PTE for a shared
1762 * mapping and we expect the PFNs to match. If they
1763 * don't match, we are likely racing with block
1764 * allocation and mapping invalidation so just skip the
1767 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1768 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1771 entry = pte_mkyoung(*pte);
1772 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1773 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1774 update_mmu_cache(vma, addr, pte);
1779 /* Ok, finally just insert the thing.. */
1780 if (pfn_t_devmap(pfn))
1781 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1783 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1786 entry = pte_mkyoung(entry);
1787 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1790 set_pte_at(mm, addr, pte, entry);
1791 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1794 pte_unmap_unlock(pte, ptl);
1795 return VM_FAULT_NOPAGE;
1799 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1800 * @vma: user vma to map to
1801 * @addr: target user address of this page
1802 * @pfn: source kernel pfn
1803 * @pgprot: pgprot flags for the inserted page
1805 * This is exactly like vmf_insert_pfn(), except that it allows drivers
1806 * to override pgprot on a per-page basis.
1808 * This only makes sense for IO mappings, and it makes no sense for
1809 * COW mappings. In general, using multiple vmas is preferable;
1810 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1813 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1814 * a value of @pgprot different from that of @vma->vm_page_prot.
1816 * Context: Process context. May allocate using %GFP_KERNEL.
1817 * Return: vm_fault_t value.
1819 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1820 unsigned long pfn, pgprot_t pgprot)
1823 * Technically, architectures with pte_special can avoid all these
1824 * restrictions (same for remap_pfn_range). However we would like
1825 * consistency in testing and feature parity among all, so we should
1826 * try to keep these invariants in place for everybody.
1828 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1829 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1830 (VM_PFNMAP|VM_MIXEDMAP));
1831 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1832 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1834 if (addr < vma->vm_start || addr >= vma->vm_end)
1835 return VM_FAULT_SIGBUS;
1837 if (!pfn_modify_allowed(pfn, pgprot))
1838 return VM_FAULT_SIGBUS;
1840 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1842 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1845 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1848 * vmf_insert_pfn - insert single pfn into user vma
1849 * @vma: user vma to map to
1850 * @addr: target user address of this page
1851 * @pfn: source kernel pfn
1853 * Similar to vm_insert_page, this allows drivers to insert individual pages
1854 * they've allocated into a user vma. Same comments apply.
1856 * This function should only be called from a vm_ops->fault handler, and
1857 * in that case the handler should return the result of this function.
1859 * vma cannot be a COW mapping.
1861 * As this is called only for pages that do not currently exist, we
1862 * do not need to flush old virtual caches or the TLB.
1864 * Context: Process context. May allocate using %GFP_KERNEL.
1865 * Return: vm_fault_t value.
1867 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1870 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1872 EXPORT_SYMBOL(vmf_insert_pfn);
1874 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1876 /* these checks mirror the abort conditions in vm_normal_page */
1877 if (vma->vm_flags & VM_MIXEDMAP)
1879 if (pfn_t_devmap(pfn))
1881 if (pfn_t_special(pfn))
1883 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1888 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1889 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
1894 BUG_ON(!vm_mixed_ok(vma, pfn));
1896 if (addr < vma->vm_start || addr >= vma->vm_end)
1897 return VM_FAULT_SIGBUS;
1899 track_pfn_insert(vma, &pgprot, pfn);
1901 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1902 return VM_FAULT_SIGBUS;
1905 * If we don't have pte special, then we have to use the pfn_valid()
1906 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1907 * refcount the page if pfn_valid is true (hence insert_page rather
1908 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1909 * without pte special, it would there be refcounted as a normal page.
1911 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1912 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1916 * At this point we are committed to insert_page()
1917 * regardless of whether the caller specified flags that
1918 * result in pfn_t_has_page() == false.
1920 page = pfn_to_page(pfn_t_to_pfn(pfn));
1921 err = insert_page(vma, addr, page, pgprot);
1923 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1927 return VM_FAULT_OOM;
1928 if (err < 0 && err != -EBUSY)
1929 return VM_FAULT_SIGBUS;
1931 return VM_FAULT_NOPAGE;
1935 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
1936 * @vma: user vma to map to
1937 * @addr: target user address of this page
1938 * @pfn: source kernel pfn
1939 * @pgprot: pgprot flags for the inserted page
1941 * This is exactly like vmf_insert_mixed(), except that it allows drivers
1942 * to override pgprot on a per-page basis.
1944 * Typically this function should be used by drivers to set caching- and
1945 * encryption bits different than those of @vma->vm_page_prot, because
1946 * the caching- or encryption mode may not be known at mmap() time.
1947 * This is ok as long as @vma->vm_page_prot is not used by the core vm
1948 * to set caching and encryption bits for those vmas (except for COW pages).
1949 * This is ensured by core vm only modifying these page table entries using
1950 * functions that don't touch caching- or encryption bits, using pte_modify()
1951 * if needed. (See for example mprotect()).
1952 * Also when new page-table entries are created, this is only done using the
1953 * fault() callback, and never using the value of vma->vm_page_prot,
1954 * except for page-table entries that point to anonymous pages as the result
1957 * Context: Process context. May allocate using %GFP_KERNEL.
1958 * Return: vm_fault_t value.
1960 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
1961 pfn_t pfn, pgprot_t pgprot)
1963 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
1965 EXPORT_SYMBOL(vmf_insert_mixed_prot);
1967 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1970 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
1972 EXPORT_SYMBOL(vmf_insert_mixed);
1975 * If the insertion of PTE failed because someone else already added a
1976 * different entry in the mean time, we treat that as success as we assume
1977 * the same entry was actually inserted.
1979 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1980 unsigned long addr, pfn_t pfn)
1982 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
1984 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1987 * maps a range of physical memory into the requested pages. the old
1988 * mappings are removed. any references to nonexistent pages results
1989 * in null mappings (currently treated as "copy-on-access")
1991 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1992 unsigned long addr, unsigned long end,
1993 unsigned long pfn, pgprot_t prot)
1999 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2002 arch_enter_lazy_mmu_mode();
2004 BUG_ON(!pte_none(*pte));
2005 if (!pfn_modify_allowed(pfn, prot)) {
2009 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2011 } while (pte++, addr += PAGE_SIZE, addr != end);
2012 arch_leave_lazy_mmu_mode();
2013 pte_unmap_unlock(pte - 1, ptl);
2017 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2018 unsigned long addr, unsigned long end,
2019 unsigned long pfn, pgprot_t prot)
2025 pfn -= addr >> PAGE_SHIFT;
2026 pmd = pmd_alloc(mm, pud, addr);
2029 VM_BUG_ON(pmd_trans_huge(*pmd));
2031 next = pmd_addr_end(addr, end);
2032 err = remap_pte_range(mm, pmd, addr, next,
2033 pfn + (addr >> PAGE_SHIFT), prot);
2036 } while (pmd++, addr = next, addr != end);
2040 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2041 unsigned long addr, unsigned long end,
2042 unsigned long pfn, pgprot_t prot)
2048 pfn -= addr >> PAGE_SHIFT;
2049 pud = pud_alloc(mm, p4d, addr);
2053 next = pud_addr_end(addr, end);
2054 err = remap_pmd_range(mm, pud, addr, next,
2055 pfn + (addr >> PAGE_SHIFT), prot);
2058 } while (pud++, addr = next, addr != end);
2062 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2063 unsigned long addr, unsigned long end,
2064 unsigned long pfn, pgprot_t prot)
2070 pfn -= addr >> PAGE_SHIFT;
2071 p4d = p4d_alloc(mm, pgd, addr);
2075 next = p4d_addr_end(addr, end);
2076 err = remap_pud_range(mm, p4d, addr, next,
2077 pfn + (addr >> PAGE_SHIFT), prot);
2080 } while (p4d++, addr = next, addr != end);
2085 * remap_pfn_range - remap kernel memory to userspace
2086 * @vma: user vma to map to
2087 * @addr: target page aligned user address to start at
2088 * @pfn: page frame number of kernel physical memory address
2089 * @size: size of mapping area
2090 * @prot: page protection flags for this mapping
2092 * Note: this is only safe if the mm semaphore is held when called.
2094 * Return: %0 on success, negative error code otherwise.
2096 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2097 unsigned long pfn, unsigned long size, pgprot_t prot)
2101 unsigned long end = addr + PAGE_ALIGN(size);
2102 struct mm_struct *mm = vma->vm_mm;
2103 unsigned long remap_pfn = pfn;
2106 if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2110 * Physically remapped pages are special. Tell the
2111 * rest of the world about it:
2112 * VM_IO tells people not to look at these pages
2113 * (accesses can have side effects).
2114 * VM_PFNMAP tells the core MM that the base pages are just
2115 * raw PFN mappings, and do not have a "struct page" associated
2118 * Disable vma merging and expanding with mremap().
2120 * Omit vma from core dump, even when VM_IO turned off.
2122 * There's a horrible special case to handle copy-on-write
2123 * behaviour that some programs depend on. We mark the "original"
2124 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2125 * See vm_normal_page() for details.
2127 if (is_cow_mapping(vma->vm_flags)) {
2128 if (addr != vma->vm_start || end != vma->vm_end)
2130 vma->vm_pgoff = pfn;
2133 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2137 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2139 BUG_ON(addr >= end);
2140 pfn -= addr >> PAGE_SHIFT;
2141 pgd = pgd_offset(mm, addr);
2142 flush_cache_range(vma, addr, end);
2144 next = pgd_addr_end(addr, end);
2145 err = remap_p4d_range(mm, pgd, addr, next,
2146 pfn + (addr >> PAGE_SHIFT), prot);
2149 } while (pgd++, addr = next, addr != end);
2152 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2156 EXPORT_SYMBOL(remap_pfn_range);
2159 * vm_iomap_memory - remap memory to userspace
2160 * @vma: user vma to map to
2161 * @start: start of the physical memory to be mapped
2162 * @len: size of area
2164 * This is a simplified io_remap_pfn_range() for common driver use. The
2165 * driver just needs to give us the physical memory range to be mapped,
2166 * we'll figure out the rest from the vma information.
2168 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2169 * whatever write-combining details or similar.
2171 * Return: %0 on success, negative error code otherwise.
2173 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2175 unsigned long vm_len, pfn, pages;
2177 /* Check that the physical memory area passed in looks valid */
2178 if (start + len < start)
2181 * You *really* shouldn't map things that aren't page-aligned,
2182 * but we've historically allowed it because IO memory might
2183 * just have smaller alignment.
2185 len += start & ~PAGE_MASK;
2186 pfn = start >> PAGE_SHIFT;
2187 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2188 if (pfn + pages < pfn)
2191 /* We start the mapping 'vm_pgoff' pages into the area */
2192 if (vma->vm_pgoff > pages)
2194 pfn += vma->vm_pgoff;
2195 pages -= vma->vm_pgoff;
2197 /* Can we fit all of the mapping? */
2198 vm_len = vma->vm_end - vma->vm_start;
2199 if (vm_len >> PAGE_SHIFT > pages)
2202 /* Ok, let it rip */
2203 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2205 EXPORT_SYMBOL(vm_iomap_memory);
2207 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2208 unsigned long addr, unsigned long end,
2209 pte_fn_t fn, void *data, bool create)
2216 pte = (mm == &init_mm) ?
2217 pte_alloc_kernel(pmd, addr) :
2218 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2222 pte = (mm == &init_mm) ?
2223 pte_offset_kernel(pmd, addr) :
2224 pte_offset_map_lock(mm, pmd, addr, &ptl);
2227 BUG_ON(pmd_huge(*pmd));
2229 arch_enter_lazy_mmu_mode();
2232 if (create || !pte_none(*pte)) {
2233 err = fn(pte++, addr, data);
2237 } while (addr += PAGE_SIZE, addr != end);
2239 arch_leave_lazy_mmu_mode();
2242 pte_unmap_unlock(pte-1, ptl);
2246 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2247 unsigned long addr, unsigned long end,
2248 pte_fn_t fn, void *data, bool create)
2254 BUG_ON(pud_huge(*pud));
2257 pmd = pmd_alloc(mm, pud, addr);
2261 pmd = pmd_offset(pud, addr);
2264 next = pmd_addr_end(addr, end);
2265 if (create || !pmd_none_or_clear_bad(pmd)) {
2266 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2271 } while (pmd++, addr = next, addr != end);
2275 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2276 unsigned long addr, unsigned long end,
2277 pte_fn_t fn, void *data, bool create)
2284 pud = pud_alloc(mm, p4d, addr);
2288 pud = pud_offset(p4d, addr);
2291 next = pud_addr_end(addr, end);
2292 if (create || !pud_none_or_clear_bad(pud)) {
2293 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2298 } while (pud++, addr = next, addr != end);
2302 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2303 unsigned long addr, unsigned long end,
2304 pte_fn_t fn, void *data, bool create)
2311 p4d = p4d_alloc(mm, pgd, addr);
2315 p4d = p4d_offset(pgd, addr);
2318 next = p4d_addr_end(addr, end);
2319 if (create || !p4d_none_or_clear_bad(p4d)) {
2320 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2325 } while (p4d++, addr = next, addr != end);
2329 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2330 unsigned long size, pte_fn_t fn,
2331 void *data, bool create)
2335 unsigned long end = addr + size;
2338 if (WARN_ON(addr >= end))
2341 pgd = pgd_offset(mm, addr);
2343 next = pgd_addr_end(addr, end);
2344 if (!create && pgd_none_or_clear_bad(pgd))
2346 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create);
2349 } while (pgd++, addr = next, addr != end);
2355 * Scan a region of virtual memory, filling in page tables as necessary
2356 * and calling a provided function on each leaf page table.
2358 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2359 unsigned long size, pte_fn_t fn, void *data)
2361 return __apply_to_page_range(mm, addr, size, fn, data, true);
2363 EXPORT_SYMBOL_GPL(apply_to_page_range);
2366 * Scan a region of virtual memory, calling a provided function on
2367 * each leaf page table where it exists.
2369 * Unlike apply_to_page_range, this does _not_ fill in page tables
2370 * where they are absent.
2372 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2373 unsigned long size, pte_fn_t fn, void *data)
2375 return __apply_to_page_range(mm, addr, size, fn, data, false);
2377 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2380 * handle_pte_fault chooses page fault handler according to an entry which was
2381 * read non-atomically. Before making any commitment, on those architectures
2382 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2383 * parts, do_swap_page must check under lock before unmapping the pte and
2384 * proceeding (but do_wp_page is only called after already making such a check;
2385 * and do_anonymous_page can safely check later on).
2387 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2388 pte_t *page_table, pte_t orig_pte)
2391 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2392 if (sizeof(pte_t) > sizeof(unsigned long)) {
2393 spinlock_t *ptl = pte_lockptr(mm, pmd);
2395 same = pte_same(*page_table, orig_pte);
2399 pte_unmap(page_table);
2403 static inline bool cow_user_page(struct page *dst, struct page *src,
2404 struct vm_fault *vmf)
2409 bool locked = false;
2410 struct vm_area_struct *vma = vmf->vma;
2411 struct mm_struct *mm = vma->vm_mm;
2412 unsigned long addr = vmf->address;
2414 debug_dma_assert_idle(src);
2417 copy_user_highpage(dst, src, addr, vma);
2422 * If the source page was a PFN mapping, we don't have
2423 * a "struct page" for it. We do a best-effort copy by
2424 * just copying from the original user address. If that
2425 * fails, we just zero-fill it. Live with it.
2427 kaddr = kmap_atomic(dst);
2428 uaddr = (void __user *)(addr & PAGE_MASK);
2431 * On architectures with software "accessed" bits, we would
2432 * take a double page fault, so mark it accessed here.
2434 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2437 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2439 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2441 * Other thread has already handled the fault
2442 * and update local tlb only
2444 update_mmu_tlb(vma, addr, vmf->pte);
2449 entry = pte_mkyoung(vmf->orig_pte);
2450 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2451 update_mmu_cache(vma, addr, vmf->pte);
2455 * This really shouldn't fail, because the page is there
2456 * in the page tables. But it might just be unreadable,
2457 * in which case we just give up and fill the result with
2460 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2464 /* Re-validate under PTL if the page is still mapped */
2465 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2467 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2468 /* The PTE changed under us, update local tlb */
2469 update_mmu_tlb(vma, addr, vmf->pte);
2475 * The same page can be mapped back since last copy attempt.
2476 * Try to copy again under PTL.
2478 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2480 * Give a warn in case there can be some obscure
2493 pte_unmap_unlock(vmf->pte, vmf->ptl);
2494 kunmap_atomic(kaddr);
2495 flush_dcache_page(dst);
2500 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2502 struct file *vm_file = vma->vm_file;
2505 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2508 * Special mappings (e.g. VDSO) do not have any file so fake
2509 * a default GFP_KERNEL for them.
2515 * Notify the address space that the page is about to become writable so that
2516 * it can prohibit this or wait for the page to get into an appropriate state.
2518 * We do this without the lock held, so that it can sleep if it needs to.
2520 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2523 struct page *page = vmf->page;
2524 unsigned int old_flags = vmf->flags;
2526 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2528 if (vmf->vma->vm_file &&
2529 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2530 return VM_FAULT_SIGBUS;
2532 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2533 /* Restore original flags so that caller is not surprised */
2534 vmf->flags = old_flags;
2535 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2537 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2539 if (!page->mapping) {
2541 return 0; /* retry */
2543 ret |= VM_FAULT_LOCKED;
2545 VM_BUG_ON_PAGE(!PageLocked(page), page);
2550 * Handle dirtying of a page in shared file mapping on a write fault.
2552 * The function expects the page to be locked and unlocks it.
2554 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2556 struct vm_area_struct *vma = vmf->vma;
2557 struct address_space *mapping;
2558 struct page *page = vmf->page;
2560 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2562 dirtied = set_page_dirty(page);
2563 VM_BUG_ON_PAGE(PageAnon(page), page);
2565 * Take a local copy of the address_space - page.mapping may be zeroed
2566 * by truncate after unlock_page(). The address_space itself remains
2567 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2568 * release semantics to prevent the compiler from undoing this copying.
2570 mapping = page_rmapping(page);
2574 file_update_time(vma->vm_file);
2577 * Throttle page dirtying rate down to writeback speed.
2579 * mapping may be NULL here because some device drivers do not
2580 * set page.mapping but still dirty their pages
2582 * Drop the mmap_lock before waiting on IO, if we can. The file
2583 * is pinning the mapping, as per above.
2585 if ((dirtied || page_mkwrite) && mapping) {
2588 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2589 balance_dirty_pages_ratelimited(mapping);
2592 return VM_FAULT_RETRY;
2600 * Handle write page faults for pages that can be reused in the current vma
2602 * This can happen either due to the mapping being with the VM_SHARED flag,
2603 * or due to us being the last reference standing to the page. In either
2604 * case, all we need to do here is to mark the page as writable and update
2605 * any related book-keeping.
2607 static inline void wp_page_reuse(struct vm_fault *vmf)
2608 __releases(vmf->ptl)
2610 struct vm_area_struct *vma = vmf->vma;
2611 struct page *page = vmf->page;
2614 * Clear the pages cpupid information as the existing
2615 * information potentially belongs to a now completely
2616 * unrelated process.
2619 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2621 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2622 entry = pte_mkyoung(vmf->orig_pte);
2623 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2624 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2625 update_mmu_cache(vma, vmf->address, vmf->pte);
2626 pte_unmap_unlock(vmf->pte, vmf->ptl);
2630 * Handle the case of a page which we actually need to copy to a new page.
2632 * Called with mmap_lock locked and the old page referenced, but
2633 * without the ptl held.
2635 * High level logic flow:
2637 * - Allocate a page, copy the content of the old page to the new one.
2638 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2639 * - Take the PTL. If the pte changed, bail out and release the allocated page
2640 * - If the pte is still the way we remember it, update the page table and all
2641 * relevant references. This includes dropping the reference the page-table
2642 * held to the old page, as well as updating the rmap.
2643 * - In any case, unlock the PTL and drop the reference we took to the old page.
2645 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2647 struct vm_area_struct *vma = vmf->vma;
2648 struct mm_struct *mm = vma->vm_mm;
2649 struct page *old_page = vmf->page;
2650 struct page *new_page = NULL;
2652 int page_copied = 0;
2653 struct mmu_notifier_range range;
2655 if (unlikely(anon_vma_prepare(vma)))
2658 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2659 new_page = alloc_zeroed_user_highpage_movable(vma,
2664 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2669 if (!cow_user_page(new_page, old_page, vmf)) {
2671 * COW failed, if the fault was solved by other,
2672 * it's fine. If not, userspace would re-fault on
2673 * the same address and we will handle the fault
2674 * from the second attempt.
2683 if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2685 cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2687 __SetPageUptodate(new_page);
2689 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2690 vmf->address & PAGE_MASK,
2691 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2692 mmu_notifier_invalidate_range_start(&range);
2695 * Re-check the pte - we dropped the lock
2697 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2698 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2700 if (!PageAnon(old_page)) {
2701 dec_mm_counter_fast(mm,
2702 mm_counter_file(old_page));
2703 inc_mm_counter_fast(mm, MM_ANONPAGES);
2706 inc_mm_counter_fast(mm, MM_ANONPAGES);
2708 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2709 entry = mk_pte(new_page, vma->vm_page_prot);
2710 entry = pte_sw_mkyoung(entry);
2711 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2713 * Clear the pte entry and flush it first, before updating the
2714 * pte with the new entry. This will avoid a race condition
2715 * seen in the presence of one thread doing SMC and another
2718 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2719 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2720 lru_cache_add_inactive_or_unevictable(new_page, vma);
2722 * We call the notify macro here because, when using secondary
2723 * mmu page tables (such as kvm shadow page tables), we want the
2724 * new page to be mapped directly into the secondary page table.
2726 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2727 update_mmu_cache(vma, vmf->address, vmf->pte);
2730 * Only after switching the pte to the new page may
2731 * we remove the mapcount here. Otherwise another
2732 * process may come and find the rmap count decremented
2733 * before the pte is switched to the new page, and
2734 * "reuse" the old page writing into it while our pte
2735 * here still points into it and can be read by other
2738 * The critical issue is to order this
2739 * page_remove_rmap with the ptp_clear_flush above.
2740 * Those stores are ordered by (if nothing else,)
2741 * the barrier present in the atomic_add_negative
2742 * in page_remove_rmap.
2744 * Then the TLB flush in ptep_clear_flush ensures that
2745 * no process can access the old page before the
2746 * decremented mapcount is visible. And the old page
2747 * cannot be reused until after the decremented
2748 * mapcount is visible. So transitively, TLBs to
2749 * old page will be flushed before it can be reused.
2751 page_remove_rmap(old_page, false);
2754 /* Free the old page.. */
2755 new_page = old_page;
2758 update_mmu_tlb(vma, vmf->address, vmf->pte);
2764 pte_unmap_unlock(vmf->pte, vmf->ptl);
2766 * No need to double call mmu_notifier->invalidate_range() callback as
2767 * the above ptep_clear_flush_notify() did already call it.
2769 mmu_notifier_invalidate_range_only_end(&range);
2772 * Don't let another task, with possibly unlocked vma,
2773 * keep the mlocked page.
2775 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2776 lock_page(old_page); /* LRU manipulation */
2777 if (PageMlocked(old_page))
2778 munlock_vma_page(old_page);
2779 unlock_page(old_page);
2783 return page_copied ? VM_FAULT_WRITE : 0;
2789 return VM_FAULT_OOM;
2793 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2794 * writeable once the page is prepared
2796 * @vmf: structure describing the fault
2798 * This function handles all that is needed to finish a write page fault in a
2799 * shared mapping due to PTE being read-only once the mapped page is prepared.
2800 * It handles locking of PTE and modifying it.
2802 * The function expects the page to be locked or other protection against
2803 * concurrent faults / writeback (such as DAX radix tree locks).
2805 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2806 * we acquired PTE lock.
2808 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2810 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2811 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2814 * We might have raced with another page fault while we released the
2815 * pte_offset_map_lock.
2817 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2818 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
2819 pte_unmap_unlock(vmf->pte, vmf->ptl);
2820 return VM_FAULT_NOPAGE;
2827 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2830 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2832 struct vm_area_struct *vma = vmf->vma;
2834 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2837 pte_unmap_unlock(vmf->pte, vmf->ptl);
2838 vmf->flags |= FAULT_FLAG_MKWRITE;
2839 ret = vma->vm_ops->pfn_mkwrite(vmf);
2840 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2842 return finish_mkwrite_fault(vmf);
2845 return VM_FAULT_WRITE;
2848 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2849 __releases(vmf->ptl)
2851 struct vm_area_struct *vma = vmf->vma;
2852 vm_fault_t ret = VM_FAULT_WRITE;
2854 get_page(vmf->page);
2856 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2859 pte_unmap_unlock(vmf->pte, vmf->ptl);
2860 tmp = do_page_mkwrite(vmf);
2861 if (unlikely(!tmp || (tmp &
2862 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2863 put_page(vmf->page);
2866 tmp = finish_mkwrite_fault(vmf);
2867 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2868 unlock_page(vmf->page);
2869 put_page(vmf->page);
2874 lock_page(vmf->page);
2876 ret |= fault_dirty_shared_page(vmf);
2877 put_page(vmf->page);
2883 * This routine handles present pages, when users try to write
2884 * to a shared page. It is done by copying the page to a new address
2885 * and decrementing the shared-page counter for the old page.
2887 * Note that this routine assumes that the protection checks have been
2888 * done by the caller (the low-level page fault routine in most cases).
2889 * Thus we can safely just mark it writable once we've done any necessary
2892 * We also mark the page dirty at this point even though the page will
2893 * change only once the write actually happens. This avoids a few races,
2894 * and potentially makes it more efficient.
2896 * We enter with non-exclusive mmap_lock (to exclude vma changes,
2897 * but allow concurrent faults), with pte both mapped and locked.
2898 * We return with mmap_lock still held, but pte unmapped and unlocked.
2900 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2901 __releases(vmf->ptl)
2903 struct vm_area_struct *vma = vmf->vma;
2905 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
2906 pte_unmap_unlock(vmf->pte, vmf->ptl);
2907 return handle_userfault(vmf, VM_UFFD_WP);
2910 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2913 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2916 * We should not cow pages in a shared writeable mapping.
2917 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2919 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2920 (VM_WRITE|VM_SHARED))
2921 return wp_pfn_shared(vmf);
2923 pte_unmap_unlock(vmf->pte, vmf->ptl);
2924 return wp_page_copy(vmf);
2928 * Take out anonymous pages first, anonymous shared vmas are
2929 * not dirty accountable.
2931 if (PageAnon(vmf->page)) {
2932 int total_map_swapcount;
2933 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2934 page_count(vmf->page) != 1))
2936 if (!trylock_page(vmf->page)) {
2937 get_page(vmf->page);
2938 pte_unmap_unlock(vmf->pte, vmf->ptl);
2939 lock_page(vmf->page);
2940 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2941 vmf->address, &vmf->ptl);
2942 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2943 update_mmu_tlb(vma, vmf->address, vmf->pte);
2944 unlock_page(vmf->page);
2945 pte_unmap_unlock(vmf->pte, vmf->ptl);
2946 put_page(vmf->page);
2949 put_page(vmf->page);
2951 if (PageKsm(vmf->page)) {
2952 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2954 unlock_page(vmf->page);
2958 return VM_FAULT_WRITE;
2960 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2961 if (total_map_swapcount == 1) {
2963 * The page is all ours. Move it to
2964 * our anon_vma so the rmap code will
2965 * not search our parent or siblings.
2966 * Protected against the rmap code by
2969 page_move_anon_rmap(vmf->page, vma);
2971 unlock_page(vmf->page);
2973 return VM_FAULT_WRITE;
2975 unlock_page(vmf->page);
2976 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2977 (VM_WRITE|VM_SHARED))) {
2978 return wp_page_shared(vmf);
2982 * Ok, we need to copy. Oh, well..
2984 get_page(vmf->page);
2986 pte_unmap_unlock(vmf->pte, vmf->ptl);
2987 return wp_page_copy(vmf);
2990 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2991 unsigned long start_addr, unsigned long end_addr,
2992 struct zap_details *details)
2994 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2997 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2998 struct zap_details *details)
3000 struct vm_area_struct *vma;
3001 pgoff_t vba, vea, zba, zea;
3003 vma_interval_tree_foreach(vma, root,
3004 details->first_index, details->last_index) {
3006 vba = vma->vm_pgoff;
3007 vea = vba + vma_pages(vma) - 1;
3008 zba = details->first_index;
3011 zea = details->last_index;
3015 unmap_mapping_range_vma(vma,
3016 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3017 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3023 * unmap_mapping_pages() - Unmap pages from processes.
3024 * @mapping: The address space containing pages to be unmapped.
3025 * @start: Index of first page to be unmapped.
3026 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3027 * @even_cows: Whether to unmap even private COWed pages.
3029 * Unmap the pages in this address space from any userspace process which
3030 * has them mmaped. Generally, you want to remove COWed pages as well when
3031 * a file is being truncated, but not when invalidating pages from the page
3034 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3035 pgoff_t nr, bool even_cows)
3037 struct zap_details details = { };
3039 details.check_mapping = even_cows ? NULL : mapping;
3040 details.first_index = start;
3041 details.last_index = start + nr - 1;
3042 if (details.last_index < details.first_index)
3043 details.last_index = ULONG_MAX;
3045 i_mmap_lock_write(mapping);
3046 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3047 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3048 i_mmap_unlock_write(mapping);
3052 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3053 * address_space corresponding to the specified byte range in the underlying
3056 * @mapping: the address space containing mmaps to be unmapped.
3057 * @holebegin: byte in first page to unmap, relative to the start of
3058 * the underlying file. This will be rounded down to a PAGE_SIZE
3059 * boundary. Note that this is different from truncate_pagecache(), which
3060 * must keep the partial page. In contrast, we must get rid of
3062 * @holelen: size of prospective hole in bytes. This will be rounded
3063 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3065 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3066 * but 0 when invalidating pagecache, don't throw away private data.
3068 void unmap_mapping_range(struct address_space *mapping,
3069 loff_t const holebegin, loff_t const holelen, int even_cows)
3071 pgoff_t hba = holebegin >> PAGE_SHIFT;
3072 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3074 /* Check for overflow. */
3075 if (sizeof(holelen) > sizeof(hlen)) {
3077 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3078 if (holeend & ~(long long)ULONG_MAX)
3079 hlen = ULONG_MAX - hba + 1;
3082 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3084 EXPORT_SYMBOL(unmap_mapping_range);
3087 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3088 * but allow concurrent faults), and pte mapped but not yet locked.
3089 * We return with pte unmapped and unlocked.
3091 * We return with the mmap_lock locked or unlocked in the same cases
3092 * as does filemap_fault().
3094 vm_fault_t do_swap_page(struct vm_fault *vmf)
3096 struct vm_area_struct *vma = vmf->vma;
3097 struct page *page = NULL, *swapcache;
3103 void *shadow = NULL;
3105 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3108 entry = pte_to_swp_entry(vmf->orig_pte);
3109 if (unlikely(non_swap_entry(entry))) {
3110 if (is_migration_entry(entry)) {
3111 migration_entry_wait(vma->vm_mm, vmf->pmd,
3113 } else if (is_device_private_entry(entry)) {
3114 vmf->page = device_private_entry_to_page(entry);
3115 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3116 } else if (is_hwpoison_entry(entry)) {
3117 ret = VM_FAULT_HWPOISON;
3119 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3120 ret = VM_FAULT_SIGBUS;
3126 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3127 page = lookup_swap_cache(entry, vma, vmf->address);
3131 struct swap_info_struct *si = swp_swap_info(entry);
3133 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3134 __swap_count(entry) == 1) {
3135 /* skip swapcache */
3136 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3141 __SetPageLocked(page);
3142 __SetPageSwapBacked(page);
3143 set_page_private(page, entry.val);
3145 /* Tell memcg to use swap ownership records */
3146 SetPageSwapCache(page);
3147 err = mem_cgroup_charge(page, vma->vm_mm,
3149 ClearPageSwapCache(page);
3155 shadow = get_shadow_from_swap_cache(entry);
3157 workingset_refault(page, shadow);
3159 lru_cache_add(page);
3160 swap_readpage(page, true);
3163 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3170 * Back out if somebody else faulted in this pte
3171 * while we released the pte lock.
3173 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3174 vmf->address, &vmf->ptl);
3175 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3177 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3181 /* Had to read the page from swap area: Major fault */
3182 ret = VM_FAULT_MAJOR;
3183 count_vm_event(PGMAJFAULT);
3184 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3185 } else if (PageHWPoison(page)) {
3187 * hwpoisoned dirty swapcache pages are kept for killing
3188 * owner processes (which may be unknown at hwpoison time)
3190 ret = VM_FAULT_HWPOISON;
3191 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3195 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3197 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3199 ret |= VM_FAULT_RETRY;
3204 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3205 * release the swapcache from under us. The page pin, and pte_same
3206 * test below, are not enough to exclude that. Even if it is still
3207 * swapcache, we need to check that the page's swap has not changed.
3209 if (unlikely((!PageSwapCache(page) ||
3210 page_private(page) != entry.val)) && swapcache)
3213 page = ksm_might_need_to_copy(page, vma, vmf->address);
3214 if (unlikely(!page)) {
3220 cgroup_throttle_swaprate(page, GFP_KERNEL);
3223 * Back out if somebody else already faulted in this pte.
3225 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3227 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3230 if (unlikely(!PageUptodate(page))) {
3231 ret = VM_FAULT_SIGBUS;
3236 * The page isn't present yet, go ahead with the fault.
3238 * Be careful about the sequence of operations here.
3239 * To get its accounting right, reuse_swap_page() must be called
3240 * while the page is counted on swap but not yet in mapcount i.e.
3241 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3242 * must be called after the swap_free(), or it will never succeed.
3245 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3246 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3247 pte = mk_pte(page, vma->vm_page_prot);
3248 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3249 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3250 vmf->flags &= ~FAULT_FLAG_WRITE;
3251 ret |= VM_FAULT_WRITE;
3252 exclusive = RMAP_EXCLUSIVE;
3254 flush_icache_page(vma, page);
3255 if (pte_swp_soft_dirty(vmf->orig_pte))
3256 pte = pte_mksoft_dirty(pte);
3257 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3258 pte = pte_mkuffd_wp(pte);
3259 pte = pte_wrprotect(pte);
3261 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3262 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3263 vmf->orig_pte = pte;
3265 /* ksm created a completely new copy */
3266 if (unlikely(page != swapcache && swapcache)) {
3267 page_add_new_anon_rmap(page, vma, vmf->address, false);
3268 lru_cache_add_inactive_or_unevictable(page, vma);
3270 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3274 if (mem_cgroup_swap_full(page) ||
3275 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3276 try_to_free_swap(page);
3278 if (page != swapcache && swapcache) {
3280 * Hold the lock to avoid the swap entry to be reused
3281 * until we take the PT lock for the pte_same() check
3282 * (to avoid false positives from pte_same). For
3283 * further safety release the lock after the swap_free
3284 * so that the swap count won't change under a
3285 * parallel locked swapcache.
3287 unlock_page(swapcache);
3288 put_page(swapcache);
3291 if (vmf->flags & FAULT_FLAG_WRITE) {
3292 ret |= do_wp_page(vmf);
3293 if (ret & VM_FAULT_ERROR)
3294 ret &= VM_FAULT_ERROR;
3298 /* No need to invalidate - it was non-present before */
3299 update_mmu_cache(vma, vmf->address, vmf->pte);
3301 pte_unmap_unlock(vmf->pte, vmf->ptl);
3305 pte_unmap_unlock(vmf->pte, vmf->ptl);
3310 if (page != swapcache && swapcache) {
3311 unlock_page(swapcache);
3312 put_page(swapcache);
3318 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3319 * but allow concurrent faults), and pte mapped but not yet locked.
3320 * We return with mmap_lock still held, but pte unmapped and unlocked.
3322 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3324 struct vm_area_struct *vma = vmf->vma;
3329 /* File mapping without ->vm_ops ? */
3330 if (vma->vm_flags & VM_SHARED)
3331 return VM_FAULT_SIGBUS;
3334 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3335 * pte_offset_map() on pmds where a huge pmd might be created
3336 * from a different thread.
3338 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3339 * parallel threads are excluded by other means.
3341 * Here we only have mmap_read_lock(mm).
3343 if (pte_alloc(vma->vm_mm, vmf->pmd))
3344 return VM_FAULT_OOM;
3346 /* See the comment in pte_alloc_one_map() */
3347 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3350 /* Use the zero-page for reads */
3351 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3352 !mm_forbids_zeropage(vma->vm_mm)) {
3353 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3354 vma->vm_page_prot));
3355 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3356 vmf->address, &vmf->ptl);
3357 if (!pte_none(*vmf->pte)) {
3358 update_mmu_tlb(vma, vmf->address, vmf->pte);
3361 ret = check_stable_address_space(vma->vm_mm);
3364 /* Deliver the page fault to userland, check inside PT lock */
3365 if (userfaultfd_missing(vma)) {
3366 pte_unmap_unlock(vmf->pte, vmf->ptl);
3367 return handle_userfault(vmf, VM_UFFD_MISSING);
3372 /* Allocate our own private page. */
3373 if (unlikely(anon_vma_prepare(vma)))
3375 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3379 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3381 cgroup_throttle_swaprate(page, GFP_KERNEL);
3384 * The memory barrier inside __SetPageUptodate makes sure that
3385 * preceding stores to the page contents become visible before
3386 * the set_pte_at() write.
3388 __SetPageUptodate(page);
3390 entry = mk_pte(page, vma->vm_page_prot);
3391 entry = pte_sw_mkyoung(entry);
3392 if (vma->vm_flags & VM_WRITE)
3393 entry = pte_mkwrite(pte_mkdirty(entry));
3395 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3397 if (!pte_none(*vmf->pte)) {
3398 update_mmu_cache(vma, vmf->address, vmf->pte);
3402 ret = check_stable_address_space(vma->vm_mm);
3406 /* Deliver the page fault to userland, check inside PT lock */
3407 if (userfaultfd_missing(vma)) {
3408 pte_unmap_unlock(vmf->pte, vmf->ptl);
3410 return handle_userfault(vmf, VM_UFFD_MISSING);
3413 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3414 page_add_new_anon_rmap(page, vma, vmf->address, false);
3415 lru_cache_add_inactive_or_unevictable(page, vma);
3417 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3419 /* No need to invalidate - it was non-present before */
3420 update_mmu_cache(vma, vmf->address, vmf->pte);
3422 pte_unmap_unlock(vmf->pte, vmf->ptl);
3430 return VM_FAULT_OOM;
3434 * The mmap_lock must have been held on entry, and may have been
3435 * released depending on flags and vma->vm_ops->fault() return value.
3436 * See filemap_fault() and __lock_page_retry().
3438 static vm_fault_t __do_fault(struct vm_fault *vmf)
3440 struct vm_area_struct *vma = vmf->vma;
3444 * Preallocate pte before we take page_lock because this might lead to
3445 * deadlocks for memcg reclaim which waits for pages under writeback:
3447 * SetPageWriteback(A)
3453 * wait_on_page_writeback(A)
3454 * SetPageWriteback(B)
3456 * # flush A, B to clear the writeback
3458 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3459 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3460 if (!vmf->prealloc_pte)
3461 return VM_FAULT_OOM;
3462 smp_wmb(); /* See comment in __pte_alloc() */
3465 ret = vma->vm_ops->fault(vmf);
3466 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3467 VM_FAULT_DONE_COW)))
3470 if (unlikely(PageHWPoison(vmf->page))) {
3471 if (ret & VM_FAULT_LOCKED)
3472 unlock_page(vmf->page);
3473 put_page(vmf->page);
3475 return VM_FAULT_HWPOISON;
3478 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3479 lock_page(vmf->page);
3481 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3487 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3488 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3489 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3490 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3492 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3494 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3497 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3499 struct vm_area_struct *vma = vmf->vma;
3501 if (!pmd_none(*vmf->pmd))
3503 if (vmf->prealloc_pte) {
3504 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3505 if (unlikely(!pmd_none(*vmf->pmd))) {
3506 spin_unlock(vmf->ptl);
3510 mm_inc_nr_ptes(vma->vm_mm);
3511 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3512 spin_unlock(vmf->ptl);
3513 vmf->prealloc_pte = NULL;
3514 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3515 return VM_FAULT_OOM;
3519 * If a huge pmd materialized under us just retry later. Use
3520 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3521 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3522 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3523 * running immediately after a huge pmd fault in a different thread of
3524 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3525 * All we have to ensure is that it is a regular pmd that we can walk
3526 * with pte_offset_map() and we can do that through an atomic read in
3527 * C, which is what pmd_trans_unstable() provides.
3529 if (pmd_devmap_trans_unstable(vmf->pmd))
3530 return VM_FAULT_NOPAGE;
3533 * At this point we know that our vmf->pmd points to a page of ptes
3534 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3535 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3536 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3537 * be valid and we will re-check to make sure the vmf->pte isn't
3538 * pte_none() under vmf->ptl protection when we return to
3541 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3546 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3547 static void deposit_prealloc_pte(struct vm_fault *vmf)
3549 struct vm_area_struct *vma = vmf->vma;
3551 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3553 * We are going to consume the prealloc table,
3554 * count that as nr_ptes.
3556 mm_inc_nr_ptes(vma->vm_mm);
3557 vmf->prealloc_pte = NULL;
3560 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3562 struct vm_area_struct *vma = vmf->vma;
3563 bool write = vmf->flags & FAULT_FLAG_WRITE;
3564 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3569 if (!transhuge_vma_suitable(vma, haddr))
3570 return VM_FAULT_FALLBACK;
3572 ret = VM_FAULT_FALLBACK;
3573 page = compound_head(page);
3576 * Archs like ppc64 need additonal space to store information
3577 * related to pte entry. Use the preallocated table for that.
3579 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3580 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3581 if (!vmf->prealloc_pte)
3582 return VM_FAULT_OOM;
3583 smp_wmb(); /* See comment in __pte_alloc() */
3586 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3587 if (unlikely(!pmd_none(*vmf->pmd)))
3590 for (i = 0; i < HPAGE_PMD_NR; i++)
3591 flush_icache_page(vma, page + i);
3593 entry = mk_huge_pmd(page, vma->vm_page_prot);
3595 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3597 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3598 page_add_file_rmap(page, true);
3600 * deposit and withdraw with pmd lock held
3602 if (arch_needs_pgtable_deposit())
3603 deposit_prealloc_pte(vmf);
3605 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3607 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3609 /* fault is handled */
3611 count_vm_event(THP_FILE_MAPPED);
3613 spin_unlock(vmf->ptl);
3617 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3625 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3626 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3628 * @vmf: fault environment
3629 * @page: page to map
3631 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3634 * Target users are page handler itself and implementations of
3635 * vm_ops->map_pages.
3637 * Return: %0 on success, %VM_FAULT_ code in case of error.
3639 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct page *page)
3641 struct vm_area_struct *vma = vmf->vma;
3642 bool write = vmf->flags & FAULT_FLAG_WRITE;
3646 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3647 ret = do_set_pmd(vmf, page);
3648 if (ret != VM_FAULT_FALLBACK)
3653 ret = pte_alloc_one_map(vmf);
3658 /* Re-check under ptl */
3659 if (unlikely(!pte_none(*vmf->pte))) {
3660 update_mmu_tlb(vma, vmf->address, vmf->pte);
3661 return VM_FAULT_NOPAGE;
3664 flush_icache_page(vma, page);
3665 entry = mk_pte(page, vma->vm_page_prot);
3666 entry = pte_sw_mkyoung(entry);
3668 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3669 /* copy-on-write page */
3670 if (write && !(vma->vm_flags & VM_SHARED)) {
3671 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3672 page_add_new_anon_rmap(page, vma, vmf->address, false);
3673 lru_cache_add_inactive_or_unevictable(page, vma);
3675 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3676 page_add_file_rmap(page, false);
3678 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3680 /* no need to invalidate: a not-present page won't be cached */
3681 update_mmu_cache(vma, vmf->address, vmf->pte);
3688 * finish_fault - finish page fault once we have prepared the page to fault
3690 * @vmf: structure describing the fault
3692 * This function handles all that is needed to finish a page fault once the
3693 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3694 * given page, adds reverse page mapping, handles memcg charges and LRU
3697 * The function expects the page to be locked and on success it consumes a
3698 * reference of a page being mapped (for the PTE which maps it).
3700 * Return: %0 on success, %VM_FAULT_ code in case of error.
3702 vm_fault_t finish_fault(struct vm_fault *vmf)
3707 /* Did we COW the page? */
3708 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3709 !(vmf->vma->vm_flags & VM_SHARED))
3710 page = vmf->cow_page;
3715 * check even for read faults because we might have lost our CoWed
3718 if (!(vmf->vma->vm_flags & VM_SHARED))
3719 ret = check_stable_address_space(vmf->vma->vm_mm);
3721 ret = alloc_set_pte(vmf, page);
3723 pte_unmap_unlock(vmf->pte, vmf->ptl);
3727 static unsigned long fault_around_bytes __read_mostly =
3728 rounddown_pow_of_two(65536);
3730 #ifdef CONFIG_DEBUG_FS
3731 static int fault_around_bytes_get(void *data, u64 *val)
3733 *val = fault_around_bytes;
3738 * fault_around_bytes must be rounded down to the nearest page order as it's
3739 * what do_fault_around() expects to see.
3741 static int fault_around_bytes_set(void *data, u64 val)
3743 if (val / PAGE_SIZE > PTRS_PER_PTE)
3745 if (val > PAGE_SIZE)
3746 fault_around_bytes = rounddown_pow_of_two(val);
3748 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3751 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3752 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3754 static int __init fault_around_debugfs(void)
3756 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3757 &fault_around_bytes_fops);
3760 late_initcall(fault_around_debugfs);
3764 * do_fault_around() tries to map few pages around the fault address. The hope
3765 * is that the pages will be needed soon and this will lower the number of
3768 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3769 * not ready to be mapped: not up-to-date, locked, etc.
3771 * This function is called with the page table lock taken. In the split ptlock
3772 * case the page table lock only protects only those entries which belong to
3773 * the page table corresponding to the fault address.
3775 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3778 * fault_around_bytes defines how many bytes we'll try to map.
3779 * do_fault_around() expects it to be set to a power of two less than or equal
3782 * The virtual address of the area that we map is naturally aligned to
3783 * fault_around_bytes rounded down to the machine page size
3784 * (and therefore to page order). This way it's easier to guarantee
3785 * that we don't cross page table boundaries.
3787 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3789 unsigned long address = vmf->address, nr_pages, mask;
3790 pgoff_t start_pgoff = vmf->pgoff;
3795 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3796 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3798 vmf->address = max(address & mask, vmf->vma->vm_start);
3799 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3803 * end_pgoff is either the end of the page table, the end of
3804 * the vma or nr_pages from start_pgoff, depending what is nearest.
3806 end_pgoff = start_pgoff -
3807 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3809 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3810 start_pgoff + nr_pages - 1);
3812 if (pmd_none(*vmf->pmd)) {
3813 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3814 if (!vmf->prealloc_pte)
3816 smp_wmb(); /* See comment in __pte_alloc() */
3819 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3821 /* Huge page is mapped? Page fault is solved */
3822 if (pmd_trans_huge(*vmf->pmd)) {
3823 ret = VM_FAULT_NOPAGE;
3827 /* ->map_pages() haven't done anything useful. Cold page cache? */
3831 /* check if the page fault is solved */
3832 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3833 if (!pte_none(*vmf->pte))
3834 ret = VM_FAULT_NOPAGE;
3835 pte_unmap_unlock(vmf->pte, vmf->ptl);
3837 vmf->address = address;
3842 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3844 struct vm_area_struct *vma = vmf->vma;
3848 * Let's call ->map_pages() first and use ->fault() as fallback
3849 * if page by the offset is not ready to be mapped (cold cache or
3852 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3853 ret = do_fault_around(vmf);
3858 ret = __do_fault(vmf);
3859 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3862 ret |= finish_fault(vmf);
3863 unlock_page(vmf->page);
3864 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3865 put_page(vmf->page);
3869 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3871 struct vm_area_struct *vma = vmf->vma;
3874 if (unlikely(anon_vma_prepare(vma)))
3875 return VM_FAULT_OOM;
3877 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3879 return VM_FAULT_OOM;
3881 if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
3882 put_page(vmf->cow_page);
3883 return VM_FAULT_OOM;
3885 cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
3887 ret = __do_fault(vmf);
3888 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3890 if (ret & VM_FAULT_DONE_COW)
3893 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3894 __SetPageUptodate(vmf->cow_page);
3896 ret |= finish_fault(vmf);
3897 unlock_page(vmf->page);
3898 put_page(vmf->page);
3899 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3903 put_page(vmf->cow_page);
3907 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3909 struct vm_area_struct *vma = vmf->vma;
3910 vm_fault_t ret, tmp;
3912 ret = __do_fault(vmf);
3913 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3917 * Check if the backing address space wants to know that the page is
3918 * about to become writable
3920 if (vma->vm_ops->page_mkwrite) {
3921 unlock_page(vmf->page);
3922 tmp = do_page_mkwrite(vmf);
3923 if (unlikely(!tmp ||
3924 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3925 put_page(vmf->page);
3930 ret |= finish_fault(vmf);
3931 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3933 unlock_page(vmf->page);
3934 put_page(vmf->page);
3938 ret |= fault_dirty_shared_page(vmf);
3943 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3944 * but allow concurrent faults).
3945 * The mmap_lock may have been released depending on flags and our
3946 * return value. See filemap_fault() and __lock_page_or_retry().
3947 * If mmap_lock is released, vma may become invalid (for example
3948 * by other thread calling munmap()).
3950 static vm_fault_t do_fault(struct vm_fault *vmf)
3952 struct vm_area_struct *vma = vmf->vma;
3953 struct mm_struct *vm_mm = vma->vm_mm;
3957 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3959 if (!vma->vm_ops->fault) {
3961 * If we find a migration pmd entry or a none pmd entry, which
3962 * should never happen, return SIGBUS
3964 if (unlikely(!pmd_present(*vmf->pmd)))
3965 ret = VM_FAULT_SIGBUS;
3967 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3972 * Make sure this is not a temporary clearing of pte
3973 * by holding ptl and checking again. A R/M/W update
3974 * of pte involves: take ptl, clearing the pte so that
3975 * we don't have concurrent modification by hardware
3976 * followed by an update.
3978 if (unlikely(pte_none(*vmf->pte)))
3979 ret = VM_FAULT_SIGBUS;
3981 ret = VM_FAULT_NOPAGE;
3983 pte_unmap_unlock(vmf->pte, vmf->ptl);
3985 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3986 ret = do_read_fault(vmf);
3987 else if (!(vma->vm_flags & VM_SHARED))
3988 ret = do_cow_fault(vmf);
3990 ret = do_shared_fault(vmf);
3992 /* preallocated pagetable is unused: free it */
3993 if (vmf->prealloc_pte) {
3994 pte_free(vm_mm, vmf->prealloc_pte);
3995 vmf->prealloc_pte = NULL;
4000 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4001 unsigned long addr, int page_nid,
4006 count_vm_numa_event(NUMA_HINT_FAULTS);
4007 if (page_nid == numa_node_id()) {
4008 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4009 *flags |= TNF_FAULT_LOCAL;
4012 return mpol_misplaced(page, vma, addr);
4015 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4017 struct vm_area_struct *vma = vmf->vma;
4018 struct page *page = NULL;
4019 int page_nid = NUMA_NO_NODE;
4022 bool migrated = false;
4024 bool was_writable = pte_savedwrite(vmf->orig_pte);
4028 * The "pte" at this point cannot be used safely without
4029 * validation through pte_unmap_same(). It's of NUMA type but
4030 * the pfn may be screwed if the read is non atomic.
4032 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4033 spin_lock(vmf->ptl);
4034 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4035 pte_unmap_unlock(vmf->pte, vmf->ptl);
4040 * Make it present again, Depending on how arch implementes non
4041 * accessible ptes, some can allow access by kernel mode.
4043 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4044 pte = pte_modify(old_pte, vma->vm_page_prot);
4045 pte = pte_mkyoung(pte);
4047 pte = pte_mkwrite(pte);
4048 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4049 update_mmu_cache(vma, vmf->address, vmf->pte);
4051 page = vm_normal_page(vma, vmf->address, pte);
4053 pte_unmap_unlock(vmf->pte, vmf->ptl);
4057 /* TODO: handle PTE-mapped THP */
4058 if (PageCompound(page)) {
4059 pte_unmap_unlock(vmf->pte, vmf->ptl);
4064 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4065 * much anyway since they can be in shared cache state. This misses
4066 * the case where a mapping is writable but the process never writes
4067 * to it but pte_write gets cleared during protection updates and
4068 * pte_dirty has unpredictable behaviour between PTE scan updates,
4069 * background writeback, dirty balancing and application behaviour.
4071 if (!pte_write(pte))
4072 flags |= TNF_NO_GROUP;
4075 * Flag if the page is shared between multiple address spaces. This
4076 * is later used when determining whether to group tasks together
4078 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4079 flags |= TNF_SHARED;
4081 last_cpupid = page_cpupid_last(page);
4082 page_nid = page_to_nid(page);
4083 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4085 pte_unmap_unlock(vmf->pte, vmf->ptl);
4086 if (target_nid == NUMA_NO_NODE) {
4091 /* Migrate to the requested node */
4092 migrated = migrate_misplaced_page(page, vma, target_nid);
4094 page_nid = target_nid;
4095 flags |= TNF_MIGRATED;
4097 flags |= TNF_MIGRATE_FAIL;
4100 if (page_nid != NUMA_NO_NODE)
4101 task_numa_fault(last_cpupid, page_nid, 1, flags);
4105 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4107 if (vma_is_anonymous(vmf->vma))
4108 return do_huge_pmd_anonymous_page(vmf);
4109 if (vmf->vma->vm_ops->huge_fault)
4110 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4111 return VM_FAULT_FALLBACK;
4114 /* `inline' is required to avoid gcc 4.1.2 build error */
4115 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4117 if (vma_is_anonymous(vmf->vma)) {
4118 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4119 return handle_userfault(vmf, VM_UFFD_WP);
4120 return do_huge_pmd_wp_page(vmf, orig_pmd);
4122 if (vmf->vma->vm_ops->huge_fault) {
4123 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4125 if (!(ret & VM_FAULT_FALLBACK))
4129 /* COW or write-notify handled on pte level: split pmd. */
4130 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4132 return VM_FAULT_FALLBACK;
4135 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4137 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4138 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4139 /* No support for anonymous transparent PUD pages yet */
4140 if (vma_is_anonymous(vmf->vma))
4142 if (vmf->vma->vm_ops->huge_fault) {
4143 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4145 if (!(ret & VM_FAULT_FALLBACK))
4149 /* COW or write-notify not handled on PUD level: split pud.*/
4150 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4151 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4152 return VM_FAULT_FALLBACK;
4155 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4157 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4158 /* No support for anonymous transparent PUD pages yet */
4159 if (vma_is_anonymous(vmf->vma))
4160 return VM_FAULT_FALLBACK;
4161 if (vmf->vma->vm_ops->huge_fault)
4162 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4163 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4164 return VM_FAULT_FALLBACK;
4168 * These routines also need to handle stuff like marking pages dirty
4169 * and/or accessed for architectures that don't do it in hardware (most
4170 * RISC architectures). The early dirtying is also good on the i386.
4172 * There is also a hook called "update_mmu_cache()" that architectures
4173 * with external mmu caches can use to update those (ie the Sparc or
4174 * PowerPC hashed page tables that act as extended TLBs).
4176 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4177 * concurrent faults).
4179 * The mmap_lock may have been released depending on flags and our return value.
4180 * See filemap_fault() and __lock_page_or_retry().
4182 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4186 if (unlikely(pmd_none(*vmf->pmd))) {
4188 * Leave __pte_alloc() until later: because vm_ops->fault may
4189 * want to allocate huge page, and if we expose page table
4190 * for an instant, it will be difficult to retract from
4191 * concurrent faults and from rmap lookups.
4195 /* See comment in pte_alloc_one_map() */
4196 if (pmd_devmap_trans_unstable(vmf->pmd))
4199 * A regular pmd is established and it can't morph into a huge
4200 * pmd from under us anymore at this point because we hold the
4201 * mmap_lock read mode and khugepaged takes it in write mode.
4202 * So now it's safe to run pte_offset_map().
4204 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4205 vmf->orig_pte = *vmf->pte;
4208 * some architectures can have larger ptes than wordsize,
4209 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4210 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4211 * accesses. The code below just needs a consistent view
4212 * for the ifs and we later double check anyway with the
4213 * ptl lock held. So here a barrier will do.
4216 if (pte_none(vmf->orig_pte)) {
4217 pte_unmap(vmf->pte);
4223 if (vma_is_anonymous(vmf->vma))
4224 return do_anonymous_page(vmf);
4226 return do_fault(vmf);
4229 if (!pte_present(vmf->orig_pte))
4230 return do_swap_page(vmf);
4232 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4233 return do_numa_page(vmf);
4235 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4236 spin_lock(vmf->ptl);
4237 entry = vmf->orig_pte;
4238 if (unlikely(!pte_same(*vmf->pte, entry))) {
4239 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4242 if (vmf->flags & FAULT_FLAG_WRITE) {
4243 if (!pte_write(entry))
4244 return do_wp_page(vmf);
4245 entry = pte_mkdirty(entry);
4247 entry = pte_mkyoung(entry);
4248 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4249 vmf->flags & FAULT_FLAG_WRITE)) {
4250 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4253 * This is needed only for protection faults but the arch code
4254 * is not yet telling us if this is a protection fault or not.
4255 * This still avoids useless tlb flushes for .text page faults
4258 if (vmf->flags & FAULT_FLAG_WRITE)
4259 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4262 pte_unmap_unlock(vmf->pte, vmf->ptl);
4267 * By the time we get here, we already hold the mm semaphore
4269 * The mmap_lock may have been released depending on flags and our
4270 * return value. See filemap_fault() and __lock_page_or_retry().
4272 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4273 unsigned long address, unsigned int flags)
4275 struct vm_fault vmf = {
4277 .address = address & PAGE_MASK,
4279 .pgoff = linear_page_index(vma, address),
4280 .gfp_mask = __get_fault_gfp_mask(vma),
4282 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4283 struct mm_struct *mm = vma->vm_mm;
4288 pgd = pgd_offset(mm, address);
4289 p4d = p4d_alloc(mm, pgd, address);
4291 return VM_FAULT_OOM;
4293 vmf.pud = pud_alloc(mm, p4d, address);
4295 return VM_FAULT_OOM;
4297 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4298 ret = create_huge_pud(&vmf);
4299 if (!(ret & VM_FAULT_FALLBACK))
4302 pud_t orig_pud = *vmf.pud;
4305 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4307 /* NUMA case for anonymous PUDs would go here */
4309 if (dirty && !pud_write(orig_pud)) {
4310 ret = wp_huge_pud(&vmf, orig_pud);
4311 if (!(ret & VM_FAULT_FALLBACK))
4314 huge_pud_set_accessed(&vmf, orig_pud);
4320 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4322 return VM_FAULT_OOM;
4324 /* Huge pud page fault raced with pmd_alloc? */
4325 if (pud_trans_unstable(vmf.pud))
4328 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4329 ret = create_huge_pmd(&vmf);
4330 if (!(ret & VM_FAULT_FALLBACK))
4333 pmd_t orig_pmd = *vmf.pmd;
4336 if (unlikely(is_swap_pmd(orig_pmd))) {
4337 VM_BUG_ON(thp_migration_supported() &&
4338 !is_pmd_migration_entry(orig_pmd));
4339 if (is_pmd_migration_entry(orig_pmd))
4340 pmd_migration_entry_wait(mm, vmf.pmd);
4343 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4344 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4345 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4347 if (dirty && !pmd_write(orig_pmd)) {
4348 ret = wp_huge_pmd(&vmf, orig_pmd);
4349 if (!(ret & VM_FAULT_FALLBACK))
4352 huge_pmd_set_accessed(&vmf, orig_pmd);
4358 return handle_pte_fault(&vmf);
4362 * mm_account_fault - Do page fault accountings
4364 * @regs: the pt_regs struct pointer. When set to NULL, will skip accounting
4365 * of perf event counters, but we'll still do the per-task accounting to
4366 * the task who triggered this page fault.
4367 * @address: the faulted address.
4368 * @flags: the fault flags.
4369 * @ret: the fault retcode.
4371 * This will take care of most of the page fault accountings. Meanwhile, it
4372 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4373 * updates. However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4374 * still be in per-arch page fault handlers at the entry of page fault.
4376 static inline void mm_account_fault(struct pt_regs *regs,
4377 unsigned long address, unsigned int flags,
4383 * We don't do accounting for some specific faults:
4385 * - Unsuccessful faults (e.g. when the address wasn't valid). That
4386 * includes arch_vma_access_permitted() failing before reaching here.
4387 * So this is not a "this many hardware page faults" counter. We
4388 * should use the hw profiling for that.
4390 * - Incomplete faults (VM_FAULT_RETRY). They will only be counted
4391 * once they're completed.
4393 if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4397 * We define the fault as a major fault when the final successful fault
4398 * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4399 * handle it immediately previously).
4401 major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4409 * If the fault is done for GUP, regs will be NULL. We only do the
4410 * accounting for the per thread fault counters who triggered the
4411 * fault, and we skip the perf event updates.
4417 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4419 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4423 * By the time we get here, we already hold the mm semaphore
4425 * The mmap_lock may have been released depending on flags and our
4426 * return value. See filemap_fault() and __lock_page_or_retry().
4428 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4429 unsigned int flags, struct pt_regs *regs)
4433 __set_current_state(TASK_RUNNING);
4435 count_vm_event(PGFAULT);
4436 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4438 /* do counter updates before entering really critical section. */
4439 check_sync_rss_stat(current);
4441 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4442 flags & FAULT_FLAG_INSTRUCTION,
4443 flags & FAULT_FLAG_REMOTE))
4444 return VM_FAULT_SIGSEGV;
4447 * Enable the memcg OOM handling for faults triggered in user
4448 * space. Kernel faults are handled more gracefully.
4450 if (flags & FAULT_FLAG_USER)
4451 mem_cgroup_enter_user_fault();
4453 if (unlikely(is_vm_hugetlb_page(vma)))
4454 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4456 ret = __handle_mm_fault(vma, address, flags);
4458 if (flags & FAULT_FLAG_USER) {
4459 mem_cgroup_exit_user_fault();
4461 * The task may have entered a memcg OOM situation but
4462 * if the allocation error was handled gracefully (no
4463 * VM_FAULT_OOM), there is no need to kill anything.
4464 * Just clean up the OOM state peacefully.
4466 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4467 mem_cgroup_oom_synchronize(false);
4470 mm_account_fault(regs, address, flags, ret);
4474 EXPORT_SYMBOL_GPL(handle_mm_fault);
4476 #ifndef __PAGETABLE_P4D_FOLDED
4478 * Allocate p4d page table.
4479 * We've already handled the fast-path in-line.
4481 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4483 p4d_t *new = p4d_alloc_one(mm, address);
4487 smp_wmb(); /* See comment in __pte_alloc */
4489 spin_lock(&mm->page_table_lock);
4490 if (pgd_present(*pgd)) /* Another has populated it */
4493 pgd_populate(mm, pgd, new);
4494 spin_unlock(&mm->page_table_lock);
4497 #endif /* __PAGETABLE_P4D_FOLDED */
4499 #ifndef __PAGETABLE_PUD_FOLDED
4501 * Allocate page upper directory.
4502 * We've already handled the fast-path in-line.
4504 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4506 pud_t *new = pud_alloc_one(mm, address);
4510 smp_wmb(); /* See comment in __pte_alloc */
4512 spin_lock(&mm->page_table_lock);
4513 if (!p4d_present(*p4d)) {
4515 p4d_populate(mm, p4d, new);
4516 } else /* Another has populated it */
4518 spin_unlock(&mm->page_table_lock);
4521 #endif /* __PAGETABLE_PUD_FOLDED */
4523 #ifndef __PAGETABLE_PMD_FOLDED
4525 * Allocate page middle directory.
4526 * We've already handled the fast-path in-line.
4528 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4531 pmd_t *new = pmd_alloc_one(mm, address);
4535 smp_wmb(); /* See comment in __pte_alloc */
4537 ptl = pud_lock(mm, pud);
4538 if (!pud_present(*pud)) {
4540 pud_populate(mm, pud, new);
4541 } else /* Another has populated it */
4546 #endif /* __PAGETABLE_PMD_FOLDED */
4548 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4549 struct mmu_notifier_range *range,
4550 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4558 pgd = pgd_offset(mm, address);
4559 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4562 p4d = p4d_offset(pgd, address);
4563 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4566 pud = pud_offset(p4d, address);
4567 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4570 pmd = pmd_offset(pud, address);
4571 VM_BUG_ON(pmd_trans_huge(*pmd));
4573 if (pmd_huge(*pmd)) {
4578 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4579 NULL, mm, address & PMD_MASK,
4580 (address & PMD_MASK) + PMD_SIZE);
4581 mmu_notifier_invalidate_range_start(range);
4583 *ptlp = pmd_lock(mm, pmd);
4584 if (pmd_huge(*pmd)) {
4590 mmu_notifier_invalidate_range_end(range);
4593 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4597 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4598 address & PAGE_MASK,
4599 (address & PAGE_MASK) + PAGE_SIZE);
4600 mmu_notifier_invalidate_range_start(range);
4602 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4603 if (!pte_present(*ptep))
4608 pte_unmap_unlock(ptep, *ptlp);
4610 mmu_notifier_invalidate_range_end(range);
4615 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4616 pte_t **ptepp, spinlock_t **ptlp)
4620 /* (void) is needed to make gcc happy */
4621 (void) __cond_lock(*ptlp,
4622 !(res = __follow_pte_pmd(mm, address, NULL,
4623 ptepp, NULL, ptlp)));
4627 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4628 struct mmu_notifier_range *range,
4629 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4633 /* (void) is needed to make gcc happy */
4634 (void) __cond_lock(*ptlp,
4635 !(res = __follow_pte_pmd(mm, address, range,
4636 ptepp, pmdpp, ptlp)));
4639 EXPORT_SYMBOL(follow_pte_pmd);
4642 * follow_pfn - look up PFN at a user virtual address
4643 * @vma: memory mapping
4644 * @address: user virtual address
4645 * @pfn: location to store found PFN
4647 * Only IO mappings and raw PFN mappings are allowed.
4649 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4651 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4658 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4661 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4664 *pfn = pte_pfn(*ptep);
4665 pte_unmap_unlock(ptep, ptl);
4668 EXPORT_SYMBOL(follow_pfn);
4670 #ifdef CONFIG_HAVE_IOREMAP_PROT
4671 int follow_phys(struct vm_area_struct *vma,
4672 unsigned long address, unsigned int flags,
4673 unsigned long *prot, resource_size_t *phys)
4679 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4682 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4686 if ((flags & FOLL_WRITE) && !pte_write(pte))
4689 *prot = pgprot_val(pte_pgprot(pte));
4690 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4694 pte_unmap_unlock(ptep, ptl);
4699 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4700 void *buf, int len, int write)
4702 resource_size_t phys_addr;
4703 unsigned long prot = 0;
4704 void __iomem *maddr;
4705 int offset = addr & (PAGE_SIZE-1);
4707 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4710 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4715 memcpy_toio(maddr + offset, buf, len);
4717 memcpy_fromio(buf, maddr + offset, len);
4722 EXPORT_SYMBOL_GPL(generic_access_phys);
4726 * Access another process' address space as given in mm. If non-NULL, use the
4727 * given task for page fault accounting.
4729 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4730 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4732 struct vm_area_struct *vma;
4733 void *old_buf = buf;
4734 int write = gup_flags & FOLL_WRITE;
4736 if (mmap_read_lock_killable(mm))
4739 /* ignore errors, just check how much was successfully transferred */
4741 int bytes, ret, offset;
4743 struct page *page = NULL;
4745 ret = get_user_pages_remote(mm, addr, 1,
4746 gup_flags, &page, &vma, NULL);
4748 #ifndef CONFIG_HAVE_IOREMAP_PROT
4752 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4753 * we can access using slightly different code.
4755 vma = find_vma(mm, addr);
4756 if (!vma || vma->vm_start > addr)
4758 if (vma->vm_ops && vma->vm_ops->access)
4759 ret = vma->vm_ops->access(vma, addr, buf,
4767 offset = addr & (PAGE_SIZE-1);
4768 if (bytes > PAGE_SIZE-offset)
4769 bytes = PAGE_SIZE-offset;
4773 copy_to_user_page(vma, page, addr,
4774 maddr + offset, buf, bytes);
4775 set_page_dirty_lock(page);
4777 copy_from_user_page(vma, page, addr,
4778 buf, maddr + offset, bytes);
4787 mmap_read_unlock(mm);
4789 return buf - old_buf;
4793 * access_remote_vm - access another process' address space
4794 * @mm: the mm_struct of the target address space
4795 * @addr: start address to access
4796 * @buf: source or destination buffer
4797 * @len: number of bytes to transfer
4798 * @gup_flags: flags modifying lookup behaviour
4800 * The caller must hold a reference on @mm.
4802 * Return: number of bytes copied from source to destination.
4804 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4805 void *buf, int len, unsigned int gup_flags)
4807 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4811 * Access another process' address space.
4812 * Source/target buffer must be kernel space,
4813 * Do not walk the page table directly, use get_user_pages
4815 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4816 void *buf, int len, unsigned int gup_flags)
4818 struct mm_struct *mm;
4821 mm = get_task_mm(tsk);
4825 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4831 EXPORT_SYMBOL_GPL(access_process_vm);
4834 * Print the name of a VMA.
4836 void print_vma_addr(char *prefix, unsigned long ip)
4838 struct mm_struct *mm = current->mm;
4839 struct vm_area_struct *vma;
4842 * we might be running from an atomic context so we cannot sleep
4844 if (!mmap_read_trylock(mm))
4847 vma = find_vma(mm, ip);
4848 if (vma && vma->vm_file) {
4849 struct file *f = vma->vm_file;
4850 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4854 p = file_path(f, buf, PAGE_SIZE);
4857 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4859 vma->vm_end - vma->vm_start);
4860 free_page((unsigned long)buf);
4863 mmap_read_unlock(mm);
4866 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4867 void __might_fault(const char *file, int line)
4870 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4871 * holding the mmap_lock, this is safe because kernel memory doesn't
4872 * get paged out, therefore we'll never actually fault, and the
4873 * below annotations will generate false positives.
4875 if (uaccess_kernel())
4877 if (pagefault_disabled())
4879 __might_sleep(file, line, 0);
4880 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4882 might_lock_read(¤t->mm->mmap_lock);
4885 EXPORT_SYMBOL(__might_fault);
4888 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4890 * Process all subpages of the specified huge page with the specified
4891 * operation. The target subpage will be processed last to keep its
4894 static inline void process_huge_page(
4895 unsigned long addr_hint, unsigned int pages_per_huge_page,
4896 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4900 unsigned long addr = addr_hint &
4901 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4903 /* Process target subpage last to keep its cache lines hot */
4905 n = (addr_hint - addr) / PAGE_SIZE;
4906 if (2 * n <= pages_per_huge_page) {
4907 /* If target subpage in first half of huge page */
4910 /* Process subpages at the end of huge page */
4911 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4913 process_subpage(addr + i * PAGE_SIZE, i, arg);
4916 /* If target subpage in second half of huge page */
4917 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4918 l = pages_per_huge_page - n;
4919 /* Process subpages at the begin of huge page */
4920 for (i = 0; i < base; i++) {
4922 process_subpage(addr + i * PAGE_SIZE, i, arg);
4926 * Process remaining subpages in left-right-left-right pattern
4927 * towards the target subpage
4929 for (i = 0; i < l; i++) {
4930 int left_idx = base + i;
4931 int right_idx = base + 2 * l - 1 - i;
4934 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4936 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4940 static void clear_gigantic_page(struct page *page,
4942 unsigned int pages_per_huge_page)
4945 struct page *p = page;
4948 for (i = 0; i < pages_per_huge_page;
4949 i++, p = mem_map_next(p, page, i)) {
4951 clear_user_highpage(p, addr + i * PAGE_SIZE);
4955 static void clear_subpage(unsigned long addr, int idx, void *arg)
4957 struct page *page = arg;
4959 clear_user_highpage(page + idx, addr);
4962 void clear_huge_page(struct page *page,
4963 unsigned long addr_hint, unsigned int pages_per_huge_page)
4965 unsigned long addr = addr_hint &
4966 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4968 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4969 clear_gigantic_page(page, addr, pages_per_huge_page);
4973 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4976 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4978 struct vm_area_struct *vma,
4979 unsigned int pages_per_huge_page)
4982 struct page *dst_base = dst;
4983 struct page *src_base = src;
4985 for (i = 0; i < pages_per_huge_page; ) {
4987 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4990 dst = mem_map_next(dst, dst_base, i);
4991 src = mem_map_next(src, src_base, i);
4995 struct copy_subpage_arg {
4998 struct vm_area_struct *vma;
5001 static void copy_subpage(unsigned long addr, int idx, void *arg)
5003 struct copy_subpage_arg *copy_arg = arg;
5005 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5006 addr, copy_arg->vma);
5009 void copy_user_huge_page(struct page *dst, struct page *src,
5010 unsigned long addr_hint, struct vm_area_struct *vma,
5011 unsigned int pages_per_huge_page)
5013 unsigned long addr = addr_hint &
5014 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5015 struct copy_subpage_arg arg = {
5021 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5022 copy_user_gigantic_page(dst, src, addr, vma,
5023 pages_per_huge_page);
5027 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5030 long copy_huge_page_from_user(struct page *dst_page,
5031 const void __user *usr_src,
5032 unsigned int pages_per_huge_page,
5033 bool allow_pagefault)
5035 void *src = (void *)usr_src;
5037 unsigned long i, rc = 0;
5038 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5040 for (i = 0; i < pages_per_huge_page; i++) {
5041 if (allow_pagefault)
5042 page_kaddr = kmap(dst_page + i);
5044 page_kaddr = kmap_atomic(dst_page + i);
5045 rc = copy_from_user(page_kaddr,
5046 (const void __user *)(src + i * PAGE_SIZE),
5048 if (allow_pagefault)
5049 kunmap(dst_page + i);
5051 kunmap_atomic(page_kaddr);
5053 ret_val -= (PAGE_SIZE - rc);
5061 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5063 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5065 static struct kmem_cache *page_ptl_cachep;
5067 void __init ptlock_cache_init(void)
5069 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5073 bool ptlock_alloc(struct page *page)
5077 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5084 void ptlock_free(struct page *page)
5086 kmem_cache_free(page_ptl_cachep, page->ptl);