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>
75 #include <trace/events/kmem.h>
78 #include <asm/mmu_context.h>
79 #include <asm/pgalloc.h>
80 #include <linux/uaccess.h>
82 #include <asm/tlbflush.h>
83 #include <asm/pgtable.h>
87 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
88 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
91 #ifndef CONFIG_NEED_MULTIPLE_NODES
92 /* use the per-pgdat data instead for discontigmem - mbligh */
93 unsigned long max_mapnr;
94 EXPORT_SYMBOL(max_mapnr);
97 EXPORT_SYMBOL(mem_map);
101 * A number of key systems in x86 including ioremap() rely on the assumption
102 * that high_memory defines the upper bound on direct map memory, then end
103 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
104 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
108 EXPORT_SYMBOL(high_memory);
111 * Randomize the address space (stacks, mmaps, brk, etc.).
113 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
114 * as ancient (libc5 based) binaries can segfault. )
116 int randomize_va_space __read_mostly =
117 #ifdef CONFIG_COMPAT_BRK
123 #ifndef arch_faults_on_old_pte
124 static inline bool arch_faults_on_old_pte(void)
127 * Those arches which don't have hw access flag feature need to
128 * implement their own helper. By default, "true" means pagefault
129 * will be hit on old pte.
135 static int __init disable_randmaps(char *s)
137 randomize_va_space = 0;
140 __setup("norandmaps", disable_randmaps);
142 unsigned long zero_pfn __read_mostly;
143 EXPORT_SYMBOL(zero_pfn);
145 unsigned long highest_memmap_pfn __read_mostly;
148 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
150 static int __init init_zero_pfn(void)
152 zero_pfn = page_to_pfn(ZERO_PAGE(0));
155 core_initcall(init_zero_pfn);
157 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
159 trace_rss_stat(mm, member, count);
162 #if defined(SPLIT_RSS_COUNTING)
164 void sync_mm_rss(struct mm_struct *mm)
168 for (i = 0; i < NR_MM_COUNTERS; i++) {
169 if (current->rss_stat.count[i]) {
170 add_mm_counter(mm, i, current->rss_stat.count[i]);
171 current->rss_stat.count[i] = 0;
174 current->rss_stat.events = 0;
177 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
179 struct task_struct *task = current;
181 if (likely(task->mm == mm))
182 task->rss_stat.count[member] += val;
184 add_mm_counter(mm, member, val);
186 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
187 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
189 /* sync counter once per 64 page faults */
190 #define TASK_RSS_EVENTS_THRESH (64)
191 static void check_sync_rss_stat(struct task_struct *task)
193 if (unlikely(task != current))
195 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
196 sync_mm_rss(task->mm);
198 #else /* SPLIT_RSS_COUNTING */
200 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
201 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
203 static void check_sync_rss_stat(struct task_struct *task)
207 #endif /* SPLIT_RSS_COUNTING */
210 * Note: this doesn't free the actual pages themselves. That
211 * has been handled earlier when unmapping all the memory regions.
213 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
216 pgtable_t token = pmd_pgtable(*pmd);
218 pte_free_tlb(tlb, token, addr);
219 mm_dec_nr_ptes(tlb->mm);
222 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
223 unsigned long addr, unsigned long end,
224 unsigned long floor, unsigned long ceiling)
231 pmd = pmd_offset(pud, addr);
233 next = pmd_addr_end(addr, end);
234 if (pmd_none_or_clear_bad(pmd))
236 free_pte_range(tlb, pmd, addr);
237 } while (pmd++, addr = next, addr != end);
247 if (end - 1 > ceiling - 1)
250 pmd = pmd_offset(pud, start);
252 pmd_free_tlb(tlb, pmd, start);
253 mm_dec_nr_pmds(tlb->mm);
256 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
257 unsigned long addr, unsigned long end,
258 unsigned long floor, unsigned long ceiling)
265 pud = pud_offset(p4d, addr);
267 next = pud_addr_end(addr, end);
268 if (pud_none_or_clear_bad(pud))
270 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
271 } while (pud++, addr = next, addr != end);
281 if (end - 1 > ceiling - 1)
284 pud = pud_offset(p4d, start);
286 pud_free_tlb(tlb, pud, start);
287 mm_dec_nr_puds(tlb->mm);
290 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
291 unsigned long addr, unsigned long end,
292 unsigned long floor, unsigned long ceiling)
299 p4d = p4d_offset(pgd, addr);
301 next = p4d_addr_end(addr, end);
302 if (p4d_none_or_clear_bad(p4d))
304 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
305 } while (p4d++, addr = next, addr != end);
311 ceiling &= PGDIR_MASK;
315 if (end - 1 > ceiling - 1)
318 p4d = p4d_offset(pgd, start);
320 p4d_free_tlb(tlb, p4d, start);
324 * This function frees user-level page tables of a process.
326 void free_pgd_range(struct mmu_gather *tlb,
327 unsigned long addr, unsigned long end,
328 unsigned long floor, unsigned long ceiling)
334 * The next few lines have given us lots of grief...
336 * Why are we testing PMD* at this top level? Because often
337 * there will be no work to do at all, and we'd prefer not to
338 * go all the way down to the bottom just to discover that.
340 * Why all these "- 1"s? Because 0 represents both the bottom
341 * of the address space and the top of it (using -1 for the
342 * top wouldn't help much: the masks would do the wrong thing).
343 * The rule is that addr 0 and floor 0 refer to the bottom of
344 * the address space, but end 0 and ceiling 0 refer to the top
345 * Comparisons need to use "end - 1" and "ceiling - 1" (though
346 * that end 0 case should be mythical).
348 * Wherever addr is brought up or ceiling brought down, we must
349 * be careful to reject "the opposite 0" before it confuses the
350 * subsequent tests. But what about where end is brought down
351 * by PMD_SIZE below? no, end can't go down to 0 there.
353 * Whereas we round start (addr) and ceiling down, by different
354 * masks at different levels, in order to test whether a table
355 * now has no other vmas using it, so can be freed, we don't
356 * bother to round floor or end up - the tests don't need that.
370 if (end - 1 > ceiling - 1)
375 * We add page table cache pages with PAGE_SIZE,
376 * (see pte_free_tlb()), flush the tlb if we need
378 tlb_change_page_size(tlb, PAGE_SIZE);
379 pgd = pgd_offset(tlb->mm, addr);
381 next = pgd_addr_end(addr, end);
382 if (pgd_none_or_clear_bad(pgd))
384 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
385 } while (pgd++, addr = next, addr != end);
388 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
389 unsigned long floor, unsigned long ceiling)
392 struct vm_area_struct *next = vma->vm_next;
393 unsigned long addr = vma->vm_start;
396 * Hide vma from rmap and truncate_pagecache before freeing
399 unlink_anon_vmas(vma);
400 unlink_file_vma(vma);
402 if (is_vm_hugetlb_page(vma)) {
403 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
404 floor, next ? next->vm_start : ceiling);
407 * Optimization: gather nearby vmas into one call down
409 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
410 && !is_vm_hugetlb_page(next)) {
413 unlink_anon_vmas(vma);
414 unlink_file_vma(vma);
416 free_pgd_range(tlb, addr, vma->vm_end,
417 floor, next ? next->vm_start : ceiling);
423 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
426 pgtable_t new = pte_alloc_one(mm);
431 * Ensure all pte setup (eg. pte page lock and page clearing) are
432 * visible before the pte is made visible to other CPUs by being
433 * put into page tables.
435 * The other side of the story is the pointer chasing in the page
436 * table walking code (when walking the page table without locking;
437 * ie. most of the time). Fortunately, these data accesses consist
438 * of a chain of data-dependent loads, meaning most CPUs (alpha
439 * being the notable exception) will already guarantee loads are
440 * seen in-order. See the alpha page table accessors for the
441 * smp_read_barrier_depends() barriers in page table walking code.
443 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
445 ptl = pmd_lock(mm, pmd);
446 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
448 pmd_populate(mm, pmd, new);
457 int __pte_alloc_kernel(pmd_t *pmd)
459 pte_t *new = pte_alloc_one_kernel(&init_mm);
463 smp_wmb(); /* See comment in __pte_alloc */
465 spin_lock(&init_mm.page_table_lock);
466 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
467 pmd_populate_kernel(&init_mm, pmd, new);
470 spin_unlock(&init_mm.page_table_lock);
472 pte_free_kernel(&init_mm, new);
476 static inline void init_rss_vec(int *rss)
478 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
481 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
485 if (current->mm == mm)
487 for (i = 0; i < NR_MM_COUNTERS; i++)
489 add_mm_counter(mm, i, rss[i]);
493 * This function is called to print an error when a bad pte
494 * is found. For example, we might have a PFN-mapped pte in
495 * a region that doesn't allow it.
497 * The calling function must still handle the error.
499 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
500 pte_t pte, struct page *page)
502 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
503 p4d_t *p4d = p4d_offset(pgd, addr);
504 pud_t *pud = pud_offset(p4d, addr);
505 pmd_t *pmd = pmd_offset(pud, addr);
506 struct address_space *mapping;
508 static unsigned long resume;
509 static unsigned long nr_shown;
510 static unsigned long nr_unshown;
513 * Allow a burst of 60 reports, then keep quiet for that minute;
514 * or allow a steady drip of one report per second.
516 if (nr_shown == 60) {
517 if (time_before(jiffies, resume)) {
522 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
529 resume = jiffies + 60 * HZ;
531 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
532 index = linear_page_index(vma, addr);
534 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
536 (long long)pte_val(pte), (long long)pmd_val(*pmd));
538 dump_page(page, "bad pte");
539 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
540 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
541 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
543 vma->vm_ops ? vma->vm_ops->fault : NULL,
544 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
545 mapping ? mapping->a_ops->readpage : NULL);
547 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
551 * vm_normal_page -- This function gets the "struct page" associated with a pte.
553 * "Special" mappings do not wish to be associated with a "struct page" (either
554 * it doesn't exist, or it exists but they don't want to touch it). In this
555 * case, NULL is returned here. "Normal" mappings do have a struct page.
557 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
558 * pte bit, in which case this function is trivial. Secondly, an architecture
559 * may not have a spare pte bit, which requires a more complicated scheme,
562 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
563 * special mapping (even if there are underlying and valid "struct pages").
564 * COWed pages of a VM_PFNMAP are always normal.
566 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
567 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
568 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
569 * mapping will always honor the rule
571 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
573 * And for normal mappings this is false.
575 * This restricts such mappings to be a linear translation from virtual address
576 * to pfn. To get around this restriction, we allow arbitrary mappings so long
577 * as the vma is not a COW mapping; in that case, we know that all ptes are
578 * special (because none can have been COWed).
581 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
583 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
584 * page" backing, however the difference is that _all_ pages with a struct
585 * page (that is, those where pfn_valid is true) are refcounted and considered
586 * normal pages by the VM. The disadvantage is that pages are refcounted
587 * (which can be slower and simply not an option for some PFNMAP users). The
588 * advantage is that we don't have to follow the strict linearity rule of
589 * PFNMAP mappings in order to support COWable mappings.
592 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
595 unsigned long pfn = pte_pfn(pte);
597 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
598 if (likely(!pte_special(pte)))
600 if (vma->vm_ops && vma->vm_ops->find_special_page)
601 return vma->vm_ops->find_special_page(vma, addr);
602 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
604 if (is_zero_pfn(pfn))
609 print_bad_pte(vma, addr, pte, NULL);
613 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
615 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
616 if (vma->vm_flags & VM_MIXEDMAP) {
622 off = (addr - vma->vm_start) >> PAGE_SHIFT;
623 if (pfn == vma->vm_pgoff + off)
625 if (!is_cow_mapping(vma->vm_flags))
630 if (is_zero_pfn(pfn))
634 if (unlikely(pfn > highest_memmap_pfn)) {
635 print_bad_pte(vma, addr, pte, NULL);
640 * NOTE! We still have PageReserved() pages in the page tables.
641 * eg. VDSO mappings can cause them to exist.
644 return pfn_to_page(pfn);
647 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
648 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
651 unsigned long pfn = pmd_pfn(pmd);
654 * There is no pmd_special() but there may be special pmds, e.g.
655 * in a direct-access (dax) mapping, so let's just replicate the
656 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
658 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
659 if (vma->vm_flags & VM_MIXEDMAP) {
665 off = (addr - vma->vm_start) >> PAGE_SHIFT;
666 if (pfn == vma->vm_pgoff + off)
668 if (!is_cow_mapping(vma->vm_flags))
675 if (is_huge_zero_pmd(pmd))
677 if (unlikely(pfn > highest_memmap_pfn))
681 * NOTE! We still have PageReserved() pages in the page tables.
682 * eg. VDSO mappings can cause them to exist.
685 return pfn_to_page(pfn);
690 * copy one vm_area from one task to the other. Assumes the page tables
691 * already present in the new task to be cleared in the whole range
692 * covered by this vma.
695 static inline unsigned long
696 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
697 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
698 unsigned long addr, int *rss)
700 unsigned long vm_flags = vma->vm_flags;
701 pte_t pte = *src_pte;
704 /* pte contains position in swap or file, so copy. */
705 if (unlikely(!pte_present(pte))) {
706 swp_entry_t entry = pte_to_swp_entry(pte);
708 if (likely(!non_swap_entry(entry))) {
709 if (swap_duplicate(entry) < 0)
712 /* make sure dst_mm is on swapoff's mmlist. */
713 if (unlikely(list_empty(&dst_mm->mmlist))) {
714 spin_lock(&mmlist_lock);
715 if (list_empty(&dst_mm->mmlist))
716 list_add(&dst_mm->mmlist,
718 spin_unlock(&mmlist_lock);
721 } else if (is_migration_entry(entry)) {
722 page = migration_entry_to_page(entry);
724 rss[mm_counter(page)]++;
726 if (is_write_migration_entry(entry) &&
727 is_cow_mapping(vm_flags)) {
729 * COW mappings require pages in both
730 * parent and child to be set to read.
732 make_migration_entry_read(&entry);
733 pte = swp_entry_to_pte(entry);
734 if (pte_swp_soft_dirty(*src_pte))
735 pte = pte_swp_mksoft_dirty(pte);
736 if (pte_swp_uffd_wp(*src_pte))
737 pte = pte_swp_mkuffd_wp(pte);
738 set_pte_at(src_mm, addr, src_pte, pte);
740 } else if (is_device_private_entry(entry)) {
741 page = device_private_entry_to_page(entry);
744 * Update rss count even for unaddressable pages, as
745 * they should treated just like normal pages in this
748 * We will likely want to have some new rss counters
749 * for unaddressable pages, at some point. But for now
750 * keep things as they are.
753 rss[mm_counter(page)]++;
754 page_dup_rmap(page, false);
757 * We do not preserve soft-dirty information, because so
758 * far, checkpoint/restore is the only feature that
759 * requires that. And checkpoint/restore does not work
760 * when a device driver is involved (you cannot easily
761 * save and restore device driver state).
763 if (is_write_device_private_entry(entry) &&
764 is_cow_mapping(vm_flags)) {
765 make_device_private_entry_read(&entry);
766 pte = swp_entry_to_pte(entry);
767 if (pte_swp_uffd_wp(*src_pte))
768 pte = pte_swp_mkuffd_wp(pte);
769 set_pte_at(src_mm, addr, src_pte, pte);
776 * If it's a COW mapping, write protect it both
777 * in the parent and the child
779 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
780 ptep_set_wrprotect(src_mm, addr, src_pte);
781 pte = pte_wrprotect(pte);
785 * If it's a shared mapping, mark it clean in
788 if (vm_flags & VM_SHARED)
789 pte = pte_mkclean(pte);
790 pte = pte_mkold(pte);
793 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
794 * does not have the VM_UFFD_WP, which means that the uffd
795 * fork event is not enabled.
797 if (!(vm_flags & VM_UFFD_WP))
798 pte = pte_clear_uffd_wp(pte);
800 page = vm_normal_page(vma, addr, pte);
803 page_dup_rmap(page, false);
804 rss[mm_counter(page)]++;
805 } else if (pte_devmap(pte)) {
806 page = pte_page(pte);
810 set_pte_at(dst_mm, addr, dst_pte, pte);
814 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
815 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
816 unsigned long addr, unsigned long end)
818 pte_t *orig_src_pte, *orig_dst_pte;
819 pte_t *src_pte, *dst_pte;
820 spinlock_t *src_ptl, *dst_ptl;
822 int rss[NR_MM_COUNTERS];
823 swp_entry_t entry = (swp_entry_t){0};
828 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
831 src_pte = pte_offset_map(src_pmd, addr);
832 src_ptl = pte_lockptr(src_mm, src_pmd);
833 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
834 orig_src_pte = src_pte;
835 orig_dst_pte = dst_pte;
836 arch_enter_lazy_mmu_mode();
840 * We are holding two locks at this point - either of them
841 * could generate latencies in another task on another CPU.
843 if (progress >= 32) {
845 if (need_resched() ||
846 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
849 if (pte_none(*src_pte)) {
853 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
858 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
860 arch_leave_lazy_mmu_mode();
861 spin_unlock(src_ptl);
862 pte_unmap(orig_src_pte);
863 add_mm_rss_vec(dst_mm, rss);
864 pte_unmap_unlock(orig_dst_pte, dst_ptl);
868 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
877 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
878 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
879 unsigned long addr, unsigned long end)
881 pmd_t *src_pmd, *dst_pmd;
884 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
887 src_pmd = pmd_offset(src_pud, addr);
889 next = pmd_addr_end(addr, end);
890 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
891 || pmd_devmap(*src_pmd)) {
893 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
894 err = copy_huge_pmd(dst_mm, src_mm,
895 dst_pmd, src_pmd, addr, vma);
902 if (pmd_none_or_clear_bad(src_pmd))
904 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
907 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
911 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
912 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
913 unsigned long addr, unsigned long end)
915 pud_t *src_pud, *dst_pud;
918 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
921 src_pud = pud_offset(src_p4d, addr);
923 next = pud_addr_end(addr, end);
924 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
927 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
928 err = copy_huge_pud(dst_mm, src_mm,
929 dst_pud, src_pud, addr, vma);
936 if (pud_none_or_clear_bad(src_pud))
938 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
941 } while (dst_pud++, src_pud++, addr = next, addr != end);
945 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
946 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
947 unsigned long addr, unsigned long end)
949 p4d_t *src_p4d, *dst_p4d;
952 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
955 src_p4d = p4d_offset(src_pgd, addr);
957 next = p4d_addr_end(addr, end);
958 if (p4d_none_or_clear_bad(src_p4d))
960 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
963 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
967 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
968 struct vm_area_struct *vma)
970 pgd_t *src_pgd, *dst_pgd;
972 unsigned long addr = vma->vm_start;
973 unsigned long end = vma->vm_end;
974 struct mmu_notifier_range range;
979 * Don't copy ptes where a page fault will fill them correctly.
980 * Fork becomes much lighter when there are big shared or private
981 * readonly mappings. The tradeoff is that copy_page_range is more
982 * efficient than faulting.
984 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
988 if (is_vm_hugetlb_page(vma))
989 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
991 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
993 * We do not free on error cases below as remove_vma
994 * gets called on error from higher level routine
996 ret = track_pfn_copy(vma);
1002 * We need to invalidate the secondary MMU mappings only when
1003 * there could be a permission downgrade on the ptes of the
1004 * parent mm. And a permission downgrade will only happen if
1005 * is_cow_mapping() returns true.
1007 is_cow = is_cow_mapping(vma->vm_flags);
1010 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1011 0, vma, src_mm, addr, end);
1012 mmu_notifier_invalidate_range_start(&range);
1016 dst_pgd = pgd_offset(dst_mm, addr);
1017 src_pgd = pgd_offset(src_mm, addr);
1019 next = pgd_addr_end(addr, end);
1020 if (pgd_none_or_clear_bad(src_pgd))
1022 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1023 vma, addr, next))) {
1027 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1030 mmu_notifier_invalidate_range_end(&range);
1034 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1035 struct vm_area_struct *vma, pmd_t *pmd,
1036 unsigned long addr, unsigned long end,
1037 struct zap_details *details)
1039 struct mm_struct *mm = tlb->mm;
1040 int force_flush = 0;
1041 int rss[NR_MM_COUNTERS];
1047 tlb_change_page_size(tlb, PAGE_SIZE);
1050 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1052 flush_tlb_batched_pending(mm);
1053 arch_enter_lazy_mmu_mode();
1056 if (pte_none(ptent))
1062 if (pte_present(ptent)) {
1065 page = vm_normal_page(vma, addr, ptent);
1066 if (unlikely(details) && page) {
1068 * unmap_shared_mapping_pages() wants to
1069 * invalidate cache without truncating:
1070 * unmap shared but keep private pages.
1072 if (details->check_mapping &&
1073 details->check_mapping != page_rmapping(page))
1076 ptent = ptep_get_and_clear_full(mm, addr, pte,
1078 tlb_remove_tlb_entry(tlb, pte, addr);
1079 if (unlikely(!page))
1082 if (!PageAnon(page)) {
1083 if (pte_dirty(ptent)) {
1085 set_page_dirty(page);
1087 if (pte_young(ptent) &&
1088 likely(!(vma->vm_flags & VM_SEQ_READ)))
1089 mark_page_accessed(page);
1091 rss[mm_counter(page)]--;
1092 page_remove_rmap(page, false);
1093 if (unlikely(page_mapcount(page) < 0))
1094 print_bad_pte(vma, addr, ptent, page);
1095 if (unlikely(__tlb_remove_page(tlb, page))) {
1103 entry = pte_to_swp_entry(ptent);
1104 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1105 struct page *page = device_private_entry_to_page(entry);
1107 if (unlikely(details && details->check_mapping)) {
1109 * unmap_shared_mapping_pages() wants to
1110 * invalidate cache without truncating:
1111 * unmap shared but keep private pages.
1113 if (details->check_mapping !=
1114 page_rmapping(page))
1118 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1119 rss[mm_counter(page)]--;
1120 page_remove_rmap(page, false);
1125 /* If details->check_mapping, we leave swap entries. */
1126 if (unlikely(details))
1129 if (!non_swap_entry(entry))
1131 else if (is_migration_entry(entry)) {
1134 page = migration_entry_to_page(entry);
1135 rss[mm_counter(page)]--;
1137 if (unlikely(!free_swap_and_cache(entry)))
1138 print_bad_pte(vma, addr, ptent, NULL);
1139 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1140 } while (pte++, addr += PAGE_SIZE, addr != end);
1142 add_mm_rss_vec(mm, rss);
1143 arch_leave_lazy_mmu_mode();
1145 /* Do the actual TLB flush before dropping ptl */
1147 tlb_flush_mmu_tlbonly(tlb);
1148 pte_unmap_unlock(start_pte, ptl);
1151 * If we forced a TLB flush (either due to running out of
1152 * batch buffers or because we needed to flush dirty TLB
1153 * entries before releasing the ptl), free the batched
1154 * memory too. Restart if we didn't do everything.
1169 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1170 struct vm_area_struct *vma, pud_t *pud,
1171 unsigned long addr, unsigned long end,
1172 struct zap_details *details)
1177 pmd = pmd_offset(pud, addr);
1179 next = pmd_addr_end(addr, end);
1180 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1181 if (next - addr != HPAGE_PMD_SIZE)
1182 __split_huge_pmd(vma, pmd, addr, false, NULL);
1183 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1188 * Here there can be other concurrent MADV_DONTNEED or
1189 * trans huge page faults running, and if the pmd is
1190 * none or trans huge it can change under us. This is
1191 * because MADV_DONTNEED holds the mmap_sem in read
1194 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1196 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1199 } while (pmd++, addr = next, addr != end);
1204 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1205 struct vm_area_struct *vma, p4d_t *p4d,
1206 unsigned long addr, unsigned long end,
1207 struct zap_details *details)
1212 pud = pud_offset(p4d, addr);
1214 next = pud_addr_end(addr, end);
1215 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1216 if (next - addr != HPAGE_PUD_SIZE) {
1217 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1218 split_huge_pud(vma, pud, addr);
1219 } else if (zap_huge_pud(tlb, vma, pud, addr))
1223 if (pud_none_or_clear_bad(pud))
1225 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1228 } while (pud++, addr = next, addr != end);
1233 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1234 struct vm_area_struct *vma, pgd_t *pgd,
1235 unsigned long addr, unsigned long end,
1236 struct zap_details *details)
1241 p4d = p4d_offset(pgd, addr);
1243 next = p4d_addr_end(addr, end);
1244 if (p4d_none_or_clear_bad(p4d))
1246 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1247 } while (p4d++, addr = next, addr != end);
1252 void unmap_page_range(struct mmu_gather *tlb,
1253 struct vm_area_struct *vma,
1254 unsigned long addr, unsigned long end,
1255 struct zap_details *details)
1260 BUG_ON(addr >= end);
1261 tlb_start_vma(tlb, vma);
1262 pgd = pgd_offset(vma->vm_mm, addr);
1264 next = pgd_addr_end(addr, end);
1265 if (pgd_none_or_clear_bad(pgd))
1267 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1268 } while (pgd++, addr = next, addr != end);
1269 tlb_end_vma(tlb, vma);
1273 static void unmap_single_vma(struct mmu_gather *tlb,
1274 struct vm_area_struct *vma, unsigned long start_addr,
1275 unsigned long end_addr,
1276 struct zap_details *details)
1278 unsigned long start = max(vma->vm_start, start_addr);
1281 if (start >= vma->vm_end)
1283 end = min(vma->vm_end, end_addr);
1284 if (end <= vma->vm_start)
1288 uprobe_munmap(vma, start, end);
1290 if (unlikely(vma->vm_flags & VM_PFNMAP))
1291 untrack_pfn(vma, 0, 0);
1294 if (unlikely(is_vm_hugetlb_page(vma))) {
1296 * It is undesirable to test vma->vm_file as it
1297 * should be non-null for valid hugetlb area.
1298 * However, vm_file will be NULL in the error
1299 * cleanup path of mmap_region. When
1300 * hugetlbfs ->mmap method fails,
1301 * mmap_region() nullifies vma->vm_file
1302 * before calling this function to clean up.
1303 * Since no pte has actually been setup, it is
1304 * safe to do nothing in this case.
1307 i_mmap_lock_write(vma->vm_file->f_mapping);
1308 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1309 i_mmap_unlock_write(vma->vm_file->f_mapping);
1312 unmap_page_range(tlb, vma, start, end, details);
1317 * unmap_vmas - unmap a range of memory covered by a list of vma's
1318 * @tlb: address of the caller's struct mmu_gather
1319 * @vma: the starting vma
1320 * @start_addr: virtual address at which to start unmapping
1321 * @end_addr: virtual address at which to end unmapping
1323 * Unmap all pages in the vma list.
1325 * Only addresses between `start' and `end' will be unmapped.
1327 * The VMA list must be sorted in ascending virtual address order.
1329 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1330 * range after unmap_vmas() returns. So the only responsibility here is to
1331 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1332 * drops the lock and schedules.
1334 void unmap_vmas(struct mmu_gather *tlb,
1335 struct vm_area_struct *vma, unsigned long start_addr,
1336 unsigned long end_addr)
1338 struct mmu_notifier_range range;
1340 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1341 start_addr, end_addr);
1342 mmu_notifier_invalidate_range_start(&range);
1343 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1344 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1345 mmu_notifier_invalidate_range_end(&range);
1349 * zap_page_range - remove user pages in a given range
1350 * @vma: vm_area_struct holding the applicable pages
1351 * @start: starting address of pages to zap
1352 * @size: number of bytes to zap
1354 * Caller must protect the VMA list
1356 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1359 struct mmu_notifier_range range;
1360 struct mmu_gather tlb;
1363 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1364 start, start + size);
1365 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1366 update_hiwater_rss(vma->vm_mm);
1367 mmu_notifier_invalidate_range_start(&range);
1368 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1369 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1370 mmu_notifier_invalidate_range_end(&range);
1371 tlb_finish_mmu(&tlb, start, range.end);
1375 * zap_page_range_single - remove user pages in a given range
1376 * @vma: vm_area_struct holding the applicable pages
1377 * @address: starting address of pages to zap
1378 * @size: number of bytes to zap
1379 * @details: details of shared cache invalidation
1381 * The range must fit into one VMA.
1383 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1384 unsigned long size, struct zap_details *details)
1386 struct mmu_notifier_range range;
1387 struct mmu_gather tlb;
1390 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1391 address, address + size);
1392 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1393 update_hiwater_rss(vma->vm_mm);
1394 mmu_notifier_invalidate_range_start(&range);
1395 unmap_single_vma(&tlb, vma, address, range.end, details);
1396 mmu_notifier_invalidate_range_end(&range);
1397 tlb_finish_mmu(&tlb, address, range.end);
1401 * zap_vma_ptes - remove ptes mapping the vma
1402 * @vma: vm_area_struct holding ptes to be zapped
1403 * @address: starting address of pages to zap
1404 * @size: number of bytes to zap
1406 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1408 * The entire address range must be fully contained within the vma.
1411 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1414 if (address < vma->vm_start || address + size > vma->vm_end ||
1415 !(vma->vm_flags & VM_PFNMAP))
1418 zap_page_range_single(vma, address, size, NULL);
1420 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1422 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1430 pgd = pgd_offset(mm, addr);
1431 p4d = p4d_alloc(mm, pgd, addr);
1434 pud = pud_alloc(mm, p4d, addr);
1437 pmd = pmd_alloc(mm, pud, addr);
1441 VM_BUG_ON(pmd_trans_huge(*pmd));
1442 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1445 static int validate_page_before_insert(struct page *page)
1447 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1449 flush_dcache_page(page);
1453 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1454 unsigned long addr, struct page *page, pgprot_t prot)
1456 if (!pte_none(*pte))
1458 /* Ok, finally just insert the thing.. */
1460 inc_mm_counter_fast(mm, mm_counter_file(page));
1461 page_add_file_rmap(page, false);
1462 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1467 * This is the old fallback for page remapping.
1469 * For historical reasons, it only allows reserved pages. Only
1470 * old drivers should use this, and they needed to mark their
1471 * pages reserved for the old functions anyway.
1473 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1474 struct page *page, pgprot_t prot)
1476 struct mm_struct *mm = vma->vm_mm;
1481 retval = validate_page_before_insert(page);
1485 pte = get_locked_pte(mm, addr, &ptl);
1488 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1489 pte_unmap_unlock(pte, ptl);
1495 * vm_insert_page - insert single page into user vma
1496 * @vma: user vma to map to
1497 * @addr: target user address of this page
1498 * @page: source kernel page
1500 * This allows drivers to insert individual pages they've allocated
1503 * The page has to be a nice clean _individual_ kernel allocation.
1504 * If you allocate a compound page, you need to have marked it as
1505 * such (__GFP_COMP), or manually just split the page up yourself
1506 * (see split_page()).
1508 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1509 * took an arbitrary page protection parameter. This doesn't allow
1510 * that. Your vma protection will have to be set up correctly, which
1511 * means that if you want a shared writable mapping, you'd better
1512 * ask for a shared writable mapping!
1514 * The page does not need to be reserved.
1516 * Usually this function is called from f_op->mmap() handler
1517 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1518 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1519 * function from other places, for example from page-fault handler.
1521 * Return: %0 on success, negative error code otherwise.
1523 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1526 if (addr < vma->vm_start || addr >= vma->vm_end)
1528 if (!page_count(page))
1530 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1531 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1532 BUG_ON(vma->vm_flags & VM_PFNMAP);
1533 vma->vm_flags |= VM_MIXEDMAP;
1535 return insert_page(vma, addr, page, vma->vm_page_prot);
1537 EXPORT_SYMBOL(vm_insert_page);
1540 * __vm_map_pages - maps range of kernel pages into user vma
1541 * @vma: user vma to map to
1542 * @pages: pointer to array of source kernel pages
1543 * @num: number of pages in page array
1544 * @offset: user's requested vm_pgoff
1546 * This allows drivers to map range of kernel pages into a user vma.
1548 * Return: 0 on success and error code otherwise.
1550 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1551 unsigned long num, unsigned long offset)
1553 unsigned long count = vma_pages(vma);
1554 unsigned long uaddr = vma->vm_start;
1557 /* Fail if the user requested offset is beyond the end of the object */
1561 /* Fail if the user requested size exceeds available object size */
1562 if (count > num - offset)
1565 for (i = 0; i < count; i++) {
1566 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1576 * vm_map_pages - maps range of kernel pages starts with non zero offset
1577 * @vma: user vma to map to
1578 * @pages: pointer to array of source kernel pages
1579 * @num: number of pages in page array
1581 * Maps an object consisting of @num pages, catering for the user's
1582 * requested vm_pgoff
1584 * If we fail to insert any page into the vma, the function will return
1585 * immediately leaving any previously inserted pages present. Callers
1586 * from the mmap handler may immediately return the error as their caller
1587 * will destroy the vma, removing any successfully inserted pages. Other
1588 * callers should make their own arrangements for calling unmap_region().
1590 * Context: Process context. Called by mmap handlers.
1591 * Return: 0 on success and error code otherwise.
1593 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1596 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1598 EXPORT_SYMBOL(vm_map_pages);
1601 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1602 * @vma: user vma to map to
1603 * @pages: pointer to array of source kernel pages
1604 * @num: number of pages in page array
1606 * Similar to vm_map_pages(), except that it explicitly sets the offset
1607 * to 0. This function is intended for the drivers that did not consider
1610 * Context: Process context. Called by mmap handlers.
1611 * Return: 0 on success and error code otherwise.
1613 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1616 return __vm_map_pages(vma, pages, num, 0);
1618 EXPORT_SYMBOL(vm_map_pages_zero);
1620 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1621 pfn_t pfn, pgprot_t prot, bool mkwrite)
1623 struct mm_struct *mm = vma->vm_mm;
1627 pte = get_locked_pte(mm, addr, &ptl);
1629 return VM_FAULT_OOM;
1630 if (!pte_none(*pte)) {
1633 * For read faults on private mappings the PFN passed
1634 * in may not match the PFN we have mapped if the
1635 * mapped PFN is a writeable COW page. In the mkwrite
1636 * case we are creating a writable PTE for a shared
1637 * mapping and we expect the PFNs to match. If they
1638 * don't match, we are likely racing with block
1639 * allocation and mapping invalidation so just skip the
1642 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1643 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1646 entry = pte_mkyoung(*pte);
1647 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1648 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1649 update_mmu_cache(vma, addr, pte);
1654 /* Ok, finally just insert the thing.. */
1655 if (pfn_t_devmap(pfn))
1656 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1658 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1661 entry = pte_mkyoung(entry);
1662 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1665 set_pte_at(mm, addr, pte, entry);
1666 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1669 pte_unmap_unlock(pte, ptl);
1670 return VM_FAULT_NOPAGE;
1674 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1675 * @vma: user vma to map to
1676 * @addr: target user address of this page
1677 * @pfn: source kernel pfn
1678 * @pgprot: pgprot flags for the inserted page
1680 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1681 * to override pgprot on a per-page basis.
1683 * This only makes sense for IO mappings, and it makes no sense for
1684 * COW mappings. In general, using multiple vmas is preferable;
1685 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1688 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1689 * a value of @pgprot different from that of @vma->vm_page_prot.
1691 * Context: Process context. May allocate using %GFP_KERNEL.
1692 * Return: vm_fault_t value.
1694 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1695 unsigned long pfn, pgprot_t pgprot)
1698 * Technically, architectures with pte_special can avoid all these
1699 * restrictions (same for remap_pfn_range). However we would like
1700 * consistency in testing and feature parity among all, so we should
1701 * try to keep these invariants in place for everybody.
1703 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1704 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1705 (VM_PFNMAP|VM_MIXEDMAP));
1706 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1707 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1709 if (addr < vma->vm_start || addr >= vma->vm_end)
1710 return VM_FAULT_SIGBUS;
1712 if (!pfn_modify_allowed(pfn, pgprot))
1713 return VM_FAULT_SIGBUS;
1715 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1717 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1720 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1723 * vmf_insert_pfn - insert single pfn into user vma
1724 * @vma: user vma to map to
1725 * @addr: target user address of this page
1726 * @pfn: source kernel pfn
1728 * Similar to vm_insert_page, this allows drivers to insert individual pages
1729 * they've allocated into a user vma. Same comments apply.
1731 * This function should only be called from a vm_ops->fault handler, and
1732 * in that case the handler should return the result of this function.
1734 * vma cannot be a COW mapping.
1736 * As this is called only for pages that do not currently exist, we
1737 * do not need to flush old virtual caches or the TLB.
1739 * Context: Process context. May allocate using %GFP_KERNEL.
1740 * Return: vm_fault_t value.
1742 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1745 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1747 EXPORT_SYMBOL(vmf_insert_pfn);
1749 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1751 /* these checks mirror the abort conditions in vm_normal_page */
1752 if (vma->vm_flags & VM_MIXEDMAP)
1754 if (pfn_t_devmap(pfn))
1756 if (pfn_t_special(pfn))
1758 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1763 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1764 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
1769 BUG_ON(!vm_mixed_ok(vma, pfn));
1771 if (addr < vma->vm_start || addr >= vma->vm_end)
1772 return VM_FAULT_SIGBUS;
1774 track_pfn_insert(vma, &pgprot, pfn);
1776 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1777 return VM_FAULT_SIGBUS;
1780 * If we don't have pte special, then we have to use the pfn_valid()
1781 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1782 * refcount the page if pfn_valid is true (hence insert_page rather
1783 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1784 * without pte special, it would there be refcounted as a normal page.
1786 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1787 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1791 * At this point we are committed to insert_page()
1792 * regardless of whether the caller specified flags that
1793 * result in pfn_t_has_page() == false.
1795 page = pfn_to_page(pfn_t_to_pfn(pfn));
1796 err = insert_page(vma, addr, page, pgprot);
1798 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1802 return VM_FAULT_OOM;
1803 if (err < 0 && err != -EBUSY)
1804 return VM_FAULT_SIGBUS;
1806 return VM_FAULT_NOPAGE;
1810 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
1811 * @vma: user vma to map to
1812 * @addr: target user address of this page
1813 * @pfn: source kernel pfn
1814 * @pgprot: pgprot flags for the inserted page
1816 * This is exactly like vmf_insert_mixed(), except that it allows drivers to
1817 * to override pgprot on a per-page basis.
1819 * Typically this function should be used by drivers to set caching- and
1820 * encryption bits different than those of @vma->vm_page_prot, because
1821 * the caching- or encryption mode may not be known at mmap() time.
1822 * This is ok as long as @vma->vm_page_prot is not used by the core vm
1823 * to set caching and encryption bits for those vmas (except for COW pages).
1824 * This is ensured by core vm only modifying these page table entries using
1825 * functions that don't touch caching- or encryption bits, using pte_modify()
1826 * if needed. (See for example mprotect()).
1827 * Also when new page-table entries are created, this is only done using the
1828 * fault() callback, and never using the value of vma->vm_page_prot,
1829 * except for page-table entries that point to anonymous pages as the result
1832 * Context: Process context. May allocate using %GFP_KERNEL.
1833 * Return: vm_fault_t value.
1835 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
1836 pfn_t pfn, pgprot_t pgprot)
1838 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
1840 EXPORT_SYMBOL(vmf_insert_mixed_prot);
1842 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1845 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
1847 EXPORT_SYMBOL(vmf_insert_mixed);
1850 * If the insertion of PTE failed because someone else already added a
1851 * different entry in the mean time, we treat that as success as we assume
1852 * the same entry was actually inserted.
1854 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1855 unsigned long addr, pfn_t pfn)
1857 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
1859 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1862 * maps a range of physical memory into the requested pages. the old
1863 * mappings are removed. any references to nonexistent pages results
1864 * in null mappings (currently treated as "copy-on-access")
1866 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1867 unsigned long addr, unsigned long end,
1868 unsigned long pfn, pgprot_t prot)
1874 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1877 arch_enter_lazy_mmu_mode();
1879 BUG_ON(!pte_none(*pte));
1880 if (!pfn_modify_allowed(pfn, prot)) {
1884 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1886 } while (pte++, addr += PAGE_SIZE, addr != end);
1887 arch_leave_lazy_mmu_mode();
1888 pte_unmap_unlock(pte - 1, ptl);
1892 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1893 unsigned long addr, unsigned long end,
1894 unsigned long pfn, pgprot_t prot)
1900 pfn -= addr >> PAGE_SHIFT;
1901 pmd = pmd_alloc(mm, pud, addr);
1904 VM_BUG_ON(pmd_trans_huge(*pmd));
1906 next = pmd_addr_end(addr, end);
1907 err = remap_pte_range(mm, pmd, addr, next,
1908 pfn + (addr >> PAGE_SHIFT), prot);
1911 } while (pmd++, addr = next, addr != end);
1915 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1916 unsigned long addr, unsigned long end,
1917 unsigned long pfn, pgprot_t prot)
1923 pfn -= addr >> PAGE_SHIFT;
1924 pud = pud_alloc(mm, p4d, addr);
1928 next = pud_addr_end(addr, end);
1929 err = remap_pmd_range(mm, pud, addr, next,
1930 pfn + (addr >> PAGE_SHIFT), prot);
1933 } while (pud++, addr = next, addr != end);
1937 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1938 unsigned long addr, unsigned long end,
1939 unsigned long pfn, pgprot_t prot)
1945 pfn -= addr >> PAGE_SHIFT;
1946 p4d = p4d_alloc(mm, pgd, addr);
1950 next = p4d_addr_end(addr, end);
1951 err = remap_pud_range(mm, p4d, addr, next,
1952 pfn + (addr >> PAGE_SHIFT), prot);
1955 } while (p4d++, addr = next, addr != end);
1960 * remap_pfn_range - remap kernel memory to userspace
1961 * @vma: user vma to map to
1962 * @addr: target user address to start at
1963 * @pfn: page frame number of kernel physical memory address
1964 * @size: size of mapping area
1965 * @prot: page protection flags for this mapping
1967 * Note: this is only safe if the mm semaphore is held when called.
1969 * Return: %0 on success, negative error code otherwise.
1971 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1972 unsigned long pfn, unsigned long size, pgprot_t prot)
1976 unsigned long end = addr + PAGE_ALIGN(size);
1977 struct mm_struct *mm = vma->vm_mm;
1978 unsigned long remap_pfn = pfn;
1982 * Physically remapped pages are special. Tell the
1983 * rest of the world about it:
1984 * VM_IO tells people not to look at these pages
1985 * (accesses can have side effects).
1986 * VM_PFNMAP tells the core MM that the base pages are just
1987 * raw PFN mappings, and do not have a "struct page" associated
1990 * Disable vma merging and expanding with mremap().
1992 * Omit vma from core dump, even when VM_IO turned off.
1994 * There's a horrible special case to handle copy-on-write
1995 * behaviour that some programs depend on. We mark the "original"
1996 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1997 * See vm_normal_page() for details.
1999 if (is_cow_mapping(vma->vm_flags)) {
2000 if (addr != vma->vm_start || end != vma->vm_end)
2002 vma->vm_pgoff = pfn;
2005 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2009 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2011 BUG_ON(addr >= end);
2012 pfn -= addr >> PAGE_SHIFT;
2013 pgd = pgd_offset(mm, addr);
2014 flush_cache_range(vma, addr, end);
2016 next = pgd_addr_end(addr, end);
2017 err = remap_p4d_range(mm, pgd, addr, next,
2018 pfn + (addr >> PAGE_SHIFT), prot);
2021 } while (pgd++, addr = next, addr != end);
2024 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2028 EXPORT_SYMBOL(remap_pfn_range);
2031 * vm_iomap_memory - remap memory to userspace
2032 * @vma: user vma to map to
2033 * @start: start of the physical memory to be mapped
2034 * @len: size of area
2036 * This is a simplified io_remap_pfn_range() for common driver use. The
2037 * driver just needs to give us the physical memory range to be mapped,
2038 * we'll figure out the rest from the vma information.
2040 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2041 * whatever write-combining details or similar.
2043 * Return: %0 on success, negative error code otherwise.
2045 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2047 unsigned long vm_len, pfn, pages;
2049 /* Check that the physical memory area passed in looks valid */
2050 if (start + len < start)
2053 * You *really* shouldn't map things that aren't page-aligned,
2054 * but we've historically allowed it because IO memory might
2055 * just have smaller alignment.
2057 len += start & ~PAGE_MASK;
2058 pfn = start >> PAGE_SHIFT;
2059 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2060 if (pfn + pages < pfn)
2063 /* We start the mapping 'vm_pgoff' pages into the area */
2064 if (vma->vm_pgoff > pages)
2066 pfn += vma->vm_pgoff;
2067 pages -= vma->vm_pgoff;
2069 /* Can we fit all of the mapping? */
2070 vm_len = vma->vm_end - vma->vm_start;
2071 if (vm_len >> PAGE_SHIFT > pages)
2074 /* Ok, let it rip */
2075 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2077 EXPORT_SYMBOL(vm_iomap_memory);
2079 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2080 unsigned long addr, unsigned long end,
2081 pte_fn_t fn, void *data, bool create)
2085 spinlock_t *uninitialized_var(ptl);
2088 pte = (mm == &init_mm) ?
2089 pte_alloc_kernel(pmd, addr) :
2090 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2094 pte = (mm == &init_mm) ?
2095 pte_offset_kernel(pmd, addr) :
2096 pte_offset_map_lock(mm, pmd, addr, &ptl);
2099 BUG_ON(pmd_huge(*pmd));
2101 arch_enter_lazy_mmu_mode();
2104 if (create || !pte_none(*pte)) {
2105 err = fn(pte++, addr, data);
2109 } while (addr += PAGE_SIZE, addr != end);
2111 arch_leave_lazy_mmu_mode();
2114 pte_unmap_unlock(pte-1, ptl);
2118 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2119 unsigned long addr, unsigned long end,
2120 pte_fn_t fn, void *data, bool create)
2126 BUG_ON(pud_huge(*pud));
2129 pmd = pmd_alloc(mm, pud, addr);
2133 pmd = pmd_offset(pud, addr);
2136 next = pmd_addr_end(addr, end);
2137 if (create || !pmd_none_or_clear_bad(pmd)) {
2138 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2143 } while (pmd++, addr = next, addr != end);
2147 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2148 unsigned long addr, unsigned long end,
2149 pte_fn_t fn, void *data, bool create)
2156 pud = pud_alloc(mm, p4d, addr);
2160 pud = pud_offset(p4d, addr);
2163 next = pud_addr_end(addr, end);
2164 if (create || !pud_none_or_clear_bad(pud)) {
2165 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2170 } while (pud++, addr = next, addr != end);
2174 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2175 unsigned long addr, unsigned long end,
2176 pte_fn_t fn, void *data, bool create)
2183 p4d = p4d_alloc(mm, pgd, addr);
2187 p4d = p4d_offset(pgd, addr);
2190 next = p4d_addr_end(addr, end);
2191 if (create || !p4d_none_or_clear_bad(p4d)) {
2192 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2197 } while (p4d++, addr = next, addr != end);
2201 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2202 unsigned long size, pte_fn_t fn,
2203 void *data, bool create)
2207 unsigned long end = addr + size;
2210 if (WARN_ON(addr >= end))
2213 pgd = pgd_offset(mm, addr);
2215 next = pgd_addr_end(addr, end);
2216 if (!create && pgd_none_or_clear_bad(pgd))
2218 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create);
2221 } while (pgd++, addr = next, addr != end);
2227 * Scan a region of virtual memory, filling in page tables as necessary
2228 * and calling a provided function on each leaf page table.
2230 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2231 unsigned long size, pte_fn_t fn, void *data)
2233 return __apply_to_page_range(mm, addr, size, fn, data, true);
2235 EXPORT_SYMBOL_GPL(apply_to_page_range);
2238 * Scan a region of virtual memory, calling a provided function on
2239 * each leaf page table where it exists.
2241 * Unlike apply_to_page_range, this does _not_ fill in page tables
2242 * where they are absent.
2244 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2245 unsigned long size, pte_fn_t fn, void *data)
2247 return __apply_to_page_range(mm, addr, size, fn, data, false);
2249 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2252 * handle_pte_fault chooses page fault handler according to an entry which was
2253 * read non-atomically. Before making any commitment, on those architectures
2254 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2255 * parts, do_swap_page must check under lock before unmapping the pte and
2256 * proceeding (but do_wp_page is only called after already making such a check;
2257 * and do_anonymous_page can safely check later on).
2259 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2260 pte_t *page_table, pte_t orig_pte)
2263 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2264 if (sizeof(pte_t) > sizeof(unsigned long)) {
2265 spinlock_t *ptl = pte_lockptr(mm, pmd);
2267 same = pte_same(*page_table, orig_pte);
2271 pte_unmap(page_table);
2275 static inline bool cow_user_page(struct page *dst, struct page *src,
2276 struct vm_fault *vmf)
2281 bool locked = false;
2282 struct vm_area_struct *vma = vmf->vma;
2283 struct mm_struct *mm = vma->vm_mm;
2284 unsigned long addr = vmf->address;
2286 debug_dma_assert_idle(src);
2289 copy_user_highpage(dst, src, addr, vma);
2294 * If the source page was a PFN mapping, we don't have
2295 * a "struct page" for it. We do a best-effort copy by
2296 * just copying from the original user address. If that
2297 * fails, we just zero-fill it. Live with it.
2299 kaddr = kmap_atomic(dst);
2300 uaddr = (void __user *)(addr & PAGE_MASK);
2303 * On architectures with software "accessed" bits, we would
2304 * take a double page fault, so mark it accessed here.
2306 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2309 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2311 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2313 * Other thread has already handled the fault
2314 * and we don't need to do anything. If it's
2315 * not the case, the fault will be triggered
2316 * again on the same address.
2322 entry = pte_mkyoung(vmf->orig_pte);
2323 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2324 update_mmu_cache(vma, addr, vmf->pte);
2328 * This really shouldn't fail, because the page is there
2329 * in the page tables. But it might just be unreadable,
2330 * in which case we just give up and fill the result with
2333 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2337 /* Re-validate under PTL if the page is still mapped */
2338 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2340 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2341 /* The PTE changed under us. Retry page fault. */
2347 * The same page can be mapped back since last copy attampt.
2348 * Try to copy again under PTL.
2350 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2352 * Give a warn in case there can be some obscure
2365 pte_unmap_unlock(vmf->pte, vmf->ptl);
2366 kunmap_atomic(kaddr);
2367 flush_dcache_page(dst);
2372 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2374 struct file *vm_file = vma->vm_file;
2377 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2380 * Special mappings (e.g. VDSO) do not have any file so fake
2381 * a default GFP_KERNEL for them.
2387 * Notify the address space that the page is about to become writable so that
2388 * it can prohibit this or wait for the page to get into an appropriate state.
2390 * We do this without the lock held, so that it can sleep if it needs to.
2392 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2395 struct page *page = vmf->page;
2396 unsigned int old_flags = vmf->flags;
2398 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2400 if (vmf->vma->vm_file &&
2401 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2402 return VM_FAULT_SIGBUS;
2404 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2405 /* Restore original flags so that caller is not surprised */
2406 vmf->flags = old_flags;
2407 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2409 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2411 if (!page->mapping) {
2413 return 0; /* retry */
2415 ret |= VM_FAULT_LOCKED;
2417 VM_BUG_ON_PAGE(!PageLocked(page), page);
2422 * Handle dirtying of a page in shared file mapping on a write fault.
2424 * The function expects the page to be locked and unlocks it.
2426 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2428 struct vm_area_struct *vma = vmf->vma;
2429 struct address_space *mapping;
2430 struct page *page = vmf->page;
2432 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2434 dirtied = set_page_dirty(page);
2435 VM_BUG_ON_PAGE(PageAnon(page), page);
2437 * Take a local copy of the address_space - page.mapping may be zeroed
2438 * by truncate after unlock_page(). The address_space itself remains
2439 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2440 * release semantics to prevent the compiler from undoing this copying.
2442 mapping = page_rmapping(page);
2446 file_update_time(vma->vm_file);
2449 * Throttle page dirtying rate down to writeback speed.
2451 * mapping may be NULL here because some device drivers do not
2452 * set page.mapping but still dirty their pages
2454 * Drop the mmap_sem before waiting on IO, if we can. The file
2455 * is pinning the mapping, as per above.
2457 if ((dirtied || page_mkwrite) && mapping) {
2460 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2461 balance_dirty_pages_ratelimited(mapping);
2464 return VM_FAULT_RETRY;
2472 * Handle write page faults for pages that can be reused in the current vma
2474 * This can happen either due to the mapping being with the VM_SHARED flag,
2475 * or due to us being the last reference standing to the page. In either
2476 * case, all we need to do here is to mark the page as writable and update
2477 * any related book-keeping.
2479 static inline void wp_page_reuse(struct vm_fault *vmf)
2480 __releases(vmf->ptl)
2482 struct vm_area_struct *vma = vmf->vma;
2483 struct page *page = vmf->page;
2486 * Clear the pages cpupid information as the existing
2487 * information potentially belongs to a now completely
2488 * unrelated process.
2491 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2493 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2494 entry = pte_mkyoung(vmf->orig_pte);
2495 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2496 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2497 update_mmu_cache(vma, vmf->address, vmf->pte);
2498 pte_unmap_unlock(vmf->pte, vmf->ptl);
2502 * Handle the case of a page which we actually need to copy to a new page.
2504 * Called with mmap_sem locked and the old page referenced, but
2505 * without the ptl held.
2507 * High level logic flow:
2509 * - Allocate a page, copy the content of the old page to the new one.
2510 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2511 * - Take the PTL. If the pte changed, bail out and release the allocated page
2512 * - If the pte is still the way we remember it, update the page table and all
2513 * relevant references. This includes dropping the reference the page-table
2514 * held to the old page, as well as updating the rmap.
2515 * - In any case, unlock the PTL and drop the reference we took to the old page.
2517 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2519 struct vm_area_struct *vma = vmf->vma;
2520 struct mm_struct *mm = vma->vm_mm;
2521 struct page *old_page = vmf->page;
2522 struct page *new_page = NULL;
2524 int page_copied = 0;
2525 struct mem_cgroup *memcg;
2526 struct mmu_notifier_range range;
2528 if (unlikely(anon_vma_prepare(vma)))
2531 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2532 new_page = alloc_zeroed_user_highpage_movable(vma,
2537 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2542 if (!cow_user_page(new_page, old_page, vmf)) {
2544 * COW failed, if the fault was solved by other,
2545 * it's fine. If not, userspace would re-fault on
2546 * the same address and we will handle the fault
2547 * from the second attempt.
2556 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2559 __SetPageUptodate(new_page);
2561 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2562 vmf->address & PAGE_MASK,
2563 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2564 mmu_notifier_invalidate_range_start(&range);
2567 * Re-check the pte - we dropped the lock
2569 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2570 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2572 if (!PageAnon(old_page)) {
2573 dec_mm_counter_fast(mm,
2574 mm_counter_file(old_page));
2575 inc_mm_counter_fast(mm, MM_ANONPAGES);
2578 inc_mm_counter_fast(mm, MM_ANONPAGES);
2580 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2581 entry = mk_pte(new_page, vma->vm_page_prot);
2582 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2584 * Clear the pte entry and flush it first, before updating the
2585 * pte with the new entry. This will avoid a race condition
2586 * seen in the presence of one thread doing SMC and another
2589 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2590 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2591 mem_cgroup_commit_charge(new_page, memcg, false, false);
2592 lru_cache_add_active_or_unevictable(new_page, vma);
2594 * We call the notify macro here because, when using secondary
2595 * mmu page tables (such as kvm shadow page tables), we want the
2596 * new page to be mapped directly into the secondary page table.
2598 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2599 update_mmu_cache(vma, vmf->address, vmf->pte);
2602 * Only after switching the pte to the new page may
2603 * we remove the mapcount here. Otherwise another
2604 * process may come and find the rmap count decremented
2605 * before the pte is switched to the new page, and
2606 * "reuse" the old page writing into it while our pte
2607 * here still points into it and can be read by other
2610 * The critical issue is to order this
2611 * page_remove_rmap with the ptp_clear_flush above.
2612 * Those stores are ordered by (if nothing else,)
2613 * the barrier present in the atomic_add_negative
2614 * in page_remove_rmap.
2616 * Then the TLB flush in ptep_clear_flush ensures that
2617 * no process can access the old page before the
2618 * decremented mapcount is visible. And the old page
2619 * cannot be reused until after the decremented
2620 * mapcount is visible. So transitively, TLBs to
2621 * old page will be flushed before it can be reused.
2623 page_remove_rmap(old_page, false);
2626 /* Free the old page.. */
2627 new_page = old_page;
2630 mem_cgroup_cancel_charge(new_page, memcg, false);
2636 pte_unmap_unlock(vmf->pte, vmf->ptl);
2638 * No need to double call mmu_notifier->invalidate_range() callback as
2639 * the above ptep_clear_flush_notify() did already call it.
2641 mmu_notifier_invalidate_range_only_end(&range);
2644 * Don't let another task, with possibly unlocked vma,
2645 * keep the mlocked page.
2647 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2648 lock_page(old_page); /* LRU manipulation */
2649 if (PageMlocked(old_page))
2650 munlock_vma_page(old_page);
2651 unlock_page(old_page);
2655 return page_copied ? VM_FAULT_WRITE : 0;
2661 return VM_FAULT_OOM;
2665 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2666 * writeable once the page is prepared
2668 * @vmf: structure describing the fault
2670 * This function handles all that is needed to finish a write page fault in a
2671 * shared mapping due to PTE being read-only once the mapped page is prepared.
2672 * It handles locking of PTE and modifying it.
2674 * The function expects the page to be locked or other protection against
2675 * concurrent faults / writeback (such as DAX radix tree locks).
2677 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2678 * we acquired PTE lock.
2680 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2682 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2683 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2686 * We might have raced with another page fault while we released the
2687 * pte_offset_map_lock.
2689 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2690 pte_unmap_unlock(vmf->pte, vmf->ptl);
2691 return VM_FAULT_NOPAGE;
2698 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2701 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2703 struct vm_area_struct *vma = vmf->vma;
2705 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2708 pte_unmap_unlock(vmf->pte, vmf->ptl);
2709 vmf->flags |= FAULT_FLAG_MKWRITE;
2710 ret = vma->vm_ops->pfn_mkwrite(vmf);
2711 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2713 return finish_mkwrite_fault(vmf);
2716 return VM_FAULT_WRITE;
2719 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2720 __releases(vmf->ptl)
2722 struct vm_area_struct *vma = vmf->vma;
2723 vm_fault_t ret = VM_FAULT_WRITE;
2725 get_page(vmf->page);
2727 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2730 pte_unmap_unlock(vmf->pte, vmf->ptl);
2731 tmp = do_page_mkwrite(vmf);
2732 if (unlikely(!tmp || (tmp &
2733 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2734 put_page(vmf->page);
2737 tmp = finish_mkwrite_fault(vmf);
2738 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2739 unlock_page(vmf->page);
2740 put_page(vmf->page);
2745 lock_page(vmf->page);
2747 ret |= fault_dirty_shared_page(vmf);
2748 put_page(vmf->page);
2754 * This routine handles present pages, when users try to write
2755 * to a shared page. It is done by copying the page to a new address
2756 * and decrementing the shared-page counter for the old page.
2758 * Note that this routine assumes that the protection checks have been
2759 * done by the caller (the low-level page fault routine in most cases).
2760 * Thus we can safely just mark it writable once we've done any necessary
2763 * We also mark the page dirty at this point even though the page will
2764 * change only once the write actually happens. This avoids a few races,
2765 * and potentially makes it more efficient.
2767 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2768 * but allow concurrent faults), with pte both mapped and locked.
2769 * We return with mmap_sem still held, but pte unmapped and unlocked.
2771 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2772 __releases(vmf->ptl)
2774 struct vm_area_struct *vma = vmf->vma;
2776 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
2777 pte_unmap_unlock(vmf->pte, vmf->ptl);
2778 return handle_userfault(vmf, VM_UFFD_WP);
2781 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2784 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2787 * We should not cow pages in a shared writeable mapping.
2788 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2790 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2791 (VM_WRITE|VM_SHARED))
2792 return wp_pfn_shared(vmf);
2794 pte_unmap_unlock(vmf->pte, vmf->ptl);
2795 return wp_page_copy(vmf);
2799 * Take out anonymous pages first, anonymous shared vmas are
2800 * not dirty accountable.
2802 if (PageAnon(vmf->page)) {
2803 int total_map_swapcount;
2804 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2805 page_count(vmf->page) != 1))
2807 if (!trylock_page(vmf->page)) {
2808 get_page(vmf->page);
2809 pte_unmap_unlock(vmf->pte, vmf->ptl);
2810 lock_page(vmf->page);
2811 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2812 vmf->address, &vmf->ptl);
2813 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2814 unlock_page(vmf->page);
2815 pte_unmap_unlock(vmf->pte, vmf->ptl);
2816 put_page(vmf->page);
2819 put_page(vmf->page);
2821 if (PageKsm(vmf->page)) {
2822 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2824 unlock_page(vmf->page);
2828 return VM_FAULT_WRITE;
2830 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2831 if (total_map_swapcount == 1) {
2833 * The page is all ours. Move it to
2834 * our anon_vma so the rmap code will
2835 * not search our parent or siblings.
2836 * Protected against the rmap code by
2839 page_move_anon_rmap(vmf->page, vma);
2841 unlock_page(vmf->page);
2843 return VM_FAULT_WRITE;
2845 unlock_page(vmf->page);
2846 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2847 (VM_WRITE|VM_SHARED))) {
2848 return wp_page_shared(vmf);
2852 * Ok, we need to copy. Oh, well..
2854 get_page(vmf->page);
2856 pte_unmap_unlock(vmf->pte, vmf->ptl);
2857 return wp_page_copy(vmf);
2860 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2861 unsigned long start_addr, unsigned long end_addr,
2862 struct zap_details *details)
2864 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2867 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2868 struct zap_details *details)
2870 struct vm_area_struct *vma;
2871 pgoff_t vba, vea, zba, zea;
2873 vma_interval_tree_foreach(vma, root,
2874 details->first_index, details->last_index) {
2876 vba = vma->vm_pgoff;
2877 vea = vba + vma_pages(vma) - 1;
2878 zba = details->first_index;
2881 zea = details->last_index;
2885 unmap_mapping_range_vma(vma,
2886 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2887 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2893 * unmap_mapping_pages() - Unmap pages from processes.
2894 * @mapping: The address space containing pages to be unmapped.
2895 * @start: Index of first page to be unmapped.
2896 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2897 * @even_cows: Whether to unmap even private COWed pages.
2899 * Unmap the pages in this address space from any userspace process which
2900 * has them mmaped. Generally, you want to remove COWed pages as well when
2901 * a file is being truncated, but not when invalidating pages from the page
2904 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2905 pgoff_t nr, bool even_cows)
2907 struct zap_details details = { };
2909 details.check_mapping = even_cows ? NULL : mapping;
2910 details.first_index = start;
2911 details.last_index = start + nr - 1;
2912 if (details.last_index < details.first_index)
2913 details.last_index = ULONG_MAX;
2915 i_mmap_lock_write(mapping);
2916 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2917 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2918 i_mmap_unlock_write(mapping);
2922 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2923 * address_space corresponding to the specified byte range in the underlying
2926 * @mapping: the address space containing mmaps to be unmapped.
2927 * @holebegin: byte in first page to unmap, relative to the start of
2928 * the underlying file. This will be rounded down to a PAGE_SIZE
2929 * boundary. Note that this is different from truncate_pagecache(), which
2930 * must keep the partial page. In contrast, we must get rid of
2932 * @holelen: size of prospective hole in bytes. This will be rounded
2933 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2935 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2936 * but 0 when invalidating pagecache, don't throw away private data.
2938 void unmap_mapping_range(struct address_space *mapping,
2939 loff_t const holebegin, loff_t const holelen, int even_cows)
2941 pgoff_t hba = holebegin >> PAGE_SHIFT;
2942 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2944 /* Check for overflow. */
2945 if (sizeof(holelen) > sizeof(hlen)) {
2947 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2948 if (holeend & ~(long long)ULONG_MAX)
2949 hlen = ULONG_MAX - hba + 1;
2952 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2954 EXPORT_SYMBOL(unmap_mapping_range);
2957 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2958 * but allow concurrent faults), and pte mapped but not yet locked.
2959 * We return with pte unmapped and unlocked.
2961 * We return with the mmap_sem locked or unlocked in the same cases
2962 * as does filemap_fault().
2964 vm_fault_t do_swap_page(struct vm_fault *vmf)
2966 struct vm_area_struct *vma = vmf->vma;
2967 struct page *page = NULL, *swapcache;
2968 struct mem_cgroup *memcg;
2975 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2978 entry = pte_to_swp_entry(vmf->orig_pte);
2979 if (unlikely(non_swap_entry(entry))) {
2980 if (is_migration_entry(entry)) {
2981 migration_entry_wait(vma->vm_mm, vmf->pmd,
2983 } else if (is_device_private_entry(entry)) {
2984 vmf->page = device_private_entry_to_page(entry);
2985 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
2986 } else if (is_hwpoison_entry(entry)) {
2987 ret = VM_FAULT_HWPOISON;
2989 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2990 ret = VM_FAULT_SIGBUS;
2996 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2997 page = lookup_swap_cache(entry, vma, vmf->address);
3001 struct swap_info_struct *si = swp_swap_info(entry);
3003 if (si->flags & SWP_SYNCHRONOUS_IO &&
3004 __swap_count(entry) == 1) {
3005 /* skip swapcache */
3006 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3009 __SetPageLocked(page);
3010 __SetPageSwapBacked(page);
3011 set_page_private(page, entry.val);
3012 lru_cache_add_anon(page);
3013 swap_readpage(page, true);
3016 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3023 * Back out if somebody else faulted in this pte
3024 * while we released the pte lock.
3026 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3027 vmf->address, &vmf->ptl);
3028 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3030 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3034 /* Had to read the page from swap area: Major fault */
3035 ret = VM_FAULT_MAJOR;
3036 count_vm_event(PGMAJFAULT);
3037 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3038 } else if (PageHWPoison(page)) {
3040 * hwpoisoned dirty swapcache pages are kept for killing
3041 * owner processes (which may be unknown at hwpoison time)
3043 ret = VM_FAULT_HWPOISON;
3044 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3048 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3050 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3052 ret |= VM_FAULT_RETRY;
3057 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3058 * release the swapcache from under us. The page pin, and pte_same
3059 * test below, are not enough to exclude that. Even if it is still
3060 * swapcache, we need to check that the page's swap has not changed.
3062 if (unlikely((!PageSwapCache(page) ||
3063 page_private(page) != entry.val)) && swapcache)
3066 page = ksm_might_need_to_copy(page, vma, vmf->address);
3067 if (unlikely(!page)) {
3073 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
3080 * Back out if somebody else already faulted in this pte.
3082 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3084 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3087 if (unlikely(!PageUptodate(page))) {
3088 ret = VM_FAULT_SIGBUS;
3093 * The page isn't present yet, go ahead with the fault.
3095 * Be careful about the sequence of operations here.
3096 * To get its accounting right, reuse_swap_page() must be called
3097 * while the page is counted on swap but not yet in mapcount i.e.
3098 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3099 * must be called after the swap_free(), or it will never succeed.
3102 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3103 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3104 pte = mk_pte(page, vma->vm_page_prot);
3105 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3106 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3107 vmf->flags &= ~FAULT_FLAG_WRITE;
3108 ret |= VM_FAULT_WRITE;
3109 exclusive = RMAP_EXCLUSIVE;
3111 flush_icache_page(vma, page);
3112 if (pte_swp_soft_dirty(vmf->orig_pte))
3113 pte = pte_mksoft_dirty(pte);
3114 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3115 pte = pte_mkuffd_wp(pte);
3116 pte = pte_wrprotect(pte);
3118 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3119 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3120 vmf->orig_pte = pte;
3122 /* ksm created a completely new copy */
3123 if (unlikely(page != swapcache && swapcache)) {
3124 page_add_new_anon_rmap(page, vma, vmf->address, false);
3125 mem_cgroup_commit_charge(page, memcg, false, false);
3126 lru_cache_add_active_or_unevictable(page, vma);
3128 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3129 mem_cgroup_commit_charge(page, memcg, true, false);
3130 activate_page(page);
3134 if (mem_cgroup_swap_full(page) ||
3135 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3136 try_to_free_swap(page);
3138 if (page != swapcache && swapcache) {
3140 * Hold the lock to avoid the swap entry to be reused
3141 * until we take the PT lock for the pte_same() check
3142 * (to avoid false positives from pte_same). For
3143 * further safety release the lock after the swap_free
3144 * so that the swap count won't change under a
3145 * parallel locked swapcache.
3147 unlock_page(swapcache);
3148 put_page(swapcache);
3151 if (vmf->flags & FAULT_FLAG_WRITE) {
3152 ret |= do_wp_page(vmf);
3153 if (ret & VM_FAULT_ERROR)
3154 ret &= VM_FAULT_ERROR;
3158 /* No need to invalidate - it was non-present before */
3159 update_mmu_cache(vma, vmf->address, vmf->pte);
3161 pte_unmap_unlock(vmf->pte, vmf->ptl);
3165 mem_cgroup_cancel_charge(page, memcg, false);
3166 pte_unmap_unlock(vmf->pte, vmf->ptl);
3171 if (page != swapcache && swapcache) {
3172 unlock_page(swapcache);
3173 put_page(swapcache);
3179 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3180 * but allow concurrent faults), and pte mapped but not yet locked.
3181 * We return with mmap_sem still held, but pte unmapped and unlocked.
3183 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3185 struct vm_area_struct *vma = vmf->vma;
3186 struct mem_cgroup *memcg;
3191 /* File mapping without ->vm_ops ? */
3192 if (vma->vm_flags & VM_SHARED)
3193 return VM_FAULT_SIGBUS;
3196 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3197 * pte_offset_map() on pmds where a huge pmd might be created
3198 * from a different thread.
3200 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3201 * parallel threads are excluded by other means.
3203 * Here we only have down_read(mmap_sem).
3205 if (pte_alloc(vma->vm_mm, vmf->pmd))
3206 return VM_FAULT_OOM;
3208 /* See the comment in pte_alloc_one_map() */
3209 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3212 /* Use the zero-page for reads */
3213 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3214 !mm_forbids_zeropage(vma->vm_mm)) {
3215 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3216 vma->vm_page_prot));
3217 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3218 vmf->address, &vmf->ptl);
3219 if (!pte_none(*vmf->pte))
3221 ret = check_stable_address_space(vma->vm_mm);
3224 /* Deliver the page fault to userland, check inside PT lock */
3225 if (userfaultfd_missing(vma)) {
3226 pte_unmap_unlock(vmf->pte, vmf->ptl);
3227 return handle_userfault(vmf, VM_UFFD_MISSING);
3232 /* Allocate our own private page. */
3233 if (unlikely(anon_vma_prepare(vma)))
3235 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3239 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3244 * The memory barrier inside __SetPageUptodate makes sure that
3245 * preceding stores to the page contents become visible before
3246 * the set_pte_at() write.
3248 __SetPageUptodate(page);
3250 entry = mk_pte(page, vma->vm_page_prot);
3251 if (vma->vm_flags & VM_WRITE)
3252 entry = pte_mkwrite(pte_mkdirty(entry));
3254 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3256 if (!pte_none(*vmf->pte))
3259 ret = check_stable_address_space(vma->vm_mm);
3263 /* Deliver the page fault to userland, check inside PT lock */
3264 if (userfaultfd_missing(vma)) {
3265 pte_unmap_unlock(vmf->pte, vmf->ptl);
3266 mem_cgroup_cancel_charge(page, memcg, false);
3268 return handle_userfault(vmf, VM_UFFD_MISSING);
3271 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3272 page_add_new_anon_rmap(page, vma, vmf->address, false);
3273 mem_cgroup_commit_charge(page, memcg, false, false);
3274 lru_cache_add_active_or_unevictable(page, vma);
3276 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3278 /* No need to invalidate - it was non-present before */
3279 update_mmu_cache(vma, vmf->address, vmf->pte);
3281 pte_unmap_unlock(vmf->pte, vmf->ptl);
3284 mem_cgroup_cancel_charge(page, memcg, false);
3290 return VM_FAULT_OOM;
3294 * The mmap_sem must have been held on entry, and may have been
3295 * released depending on flags and vma->vm_ops->fault() return value.
3296 * See filemap_fault() and __lock_page_retry().
3298 static vm_fault_t __do_fault(struct vm_fault *vmf)
3300 struct vm_area_struct *vma = vmf->vma;
3304 * Preallocate pte before we take page_lock because this might lead to
3305 * deadlocks for memcg reclaim which waits for pages under writeback:
3307 * SetPageWriteback(A)
3313 * wait_on_page_writeback(A)
3314 * SetPageWriteback(B)
3316 * # flush A, B to clear the writeback
3318 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3319 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3320 if (!vmf->prealloc_pte)
3321 return VM_FAULT_OOM;
3322 smp_wmb(); /* See comment in __pte_alloc() */
3325 ret = vma->vm_ops->fault(vmf);
3326 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3327 VM_FAULT_DONE_COW)))
3330 if (unlikely(PageHWPoison(vmf->page))) {
3331 if (ret & VM_FAULT_LOCKED)
3332 unlock_page(vmf->page);
3333 put_page(vmf->page);
3335 return VM_FAULT_HWPOISON;
3338 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3339 lock_page(vmf->page);
3341 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3347 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3348 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3349 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3350 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3352 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3354 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3357 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3359 struct vm_area_struct *vma = vmf->vma;
3361 if (!pmd_none(*vmf->pmd))
3363 if (vmf->prealloc_pte) {
3364 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3365 if (unlikely(!pmd_none(*vmf->pmd))) {
3366 spin_unlock(vmf->ptl);
3370 mm_inc_nr_ptes(vma->vm_mm);
3371 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3372 spin_unlock(vmf->ptl);
3373 vmf->prealloc_pte = NULL;
3374 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3375 return VM_FAULT_OOM;
3379 * If a huge pmd materialized under us just retry later. Use
3380 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3381 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3382 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3383 * running immediately after a huge pmd fault in a different thread of
3384 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3385 * All we have to ensure is that it is a regular pmd that we can walk
3386 * with pte_offset_map() and we can do that through an atomic read in
3387 * C, which is what pmd_trans_unstable() provides.
3389 if (pmd_devmap_trans_unstable(vmf->pmd))
3390 return VM_FAULT_NOPAGE;
3393 * At this point we know that our vmf->pmd points to a page of ptes
3394 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3395 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3396 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3397 * be valid and we will re-check to make sure the vmf->pte isn't
3398 * pte_none() under vmf->ptl protection when we return to
3401 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3406 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3407 static void deposit_prealloc_pte(struct vm_fault *vmf)
3409 struct vm_area_struct *vma = vmf->vma;
3411 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3413 * We are going to consume the prealloc table,
3414 * count that as nr_ptes.
3416 mm_inc_nr_ptes(vma->vm_mm);
3417 vmf->prealloc_pte = NULL;
3420 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3422 struct vm_area_struct *vma = vmf->vma;
3423 bool write = vmf->flags & FAULT_FLAG_WRITE;
3424 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3429 if (!transhuge_vma_suitable(vma, haddr))
3430 return VM_FAULT_FALLBACK;
3432 ret = VM_FAULT_FALLBACK;
3433 page = compound_head(page);
3436 * Archs like ppc64 need additonal space to store information
3437 * related to pte entry. Use the preallocated table for that.
3439 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3440 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3441 if (!vmf->prealloc_pte)
3442 return VM_FAULT_OOM;
3443 smp_wmb(); /* See comment in __pte_alloc() */
3446 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3447 if (unlikely(!pmd_none(*vmf->pmd)))
3450 for (i = 0; i < HPAGE_PMD_NR; i++)
3451 flush_icache_page(vma, page + i);
3453 entry = mk_huge_pmd(page, vma->vm_page_prot);
3455 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3457 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3458 page_add_file_rmap(page, true);
3460 * deposit and withdraw with pmd lock held
3462 if (arch_needs_pgtable_deposit())
3463 deposit_prealloc_pte(vmf);
3465 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3467 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3469 /* fault is handled */
3471 count_vm_event(THP_FILE_MAPPED);
3473 spin_unlock(vmf->ptl);
3477 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3485 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3486 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3488 * @vmf: fault environment
3489 * @memcg: memcg to charge page (only for private mappings)
3490 * @page: page to map
3492 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3495 * Target users are page handler itself and implementations of
3496 * vm_ops->map_pages.
3498 * Return: %0 on success, %VM_FAULT_ code in case of error.
3500 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3503 struct vm_area_struct *vma = vmf->vma;
3504 bool write = vmf->flags & FAULT_FLAG_WRITE;
3508 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3510 VM_BUG_ON_PAGE(memcg, page);
3512 ret = do_set_pmd(vmf, page);
3513 if (ret != VM_FAULT_FALLBACK)
3518 ret = pte_alloc_one_map(vmf);
3523 /* Re-check under ptl */
3524 if (unlikely(!pte_none(*vmf->pte)))
3525 return VM_FAULT_NOPAGE;
3527 flush_icache_page(vma, page);
3528 entry = mk_pte(page, vma->vm_page_prot);
3530 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3531 /* copy-on-write page */
3532 if (write && !(vma->vm_flags & VM_SHARED)) {
3533 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3534 page_add_new_anon_rmap(page, vma, vmf->address, false);
3535 mem_cgroup_commit_charge(page, memcg, false, false);
3536 lru_cache_add_active_or_unevictable(page, vma);
3538 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3539 page_add_file_rmap(page, false);
3541 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3543 /* no need to invalidate: a not-present page won't be cached */
3544 update_mmu_cache(vma, vmf->address, vmf->pte);
3551 * finish_fault - finish page fault once we have prepared the page to fault
3553 * @vmf: structure describing the fault
3555 * This function handles all that is needed to finish a page fault once the
3556 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3557 * given page, adds reverse page mapping, handles memcg charges and LRU
3560 * The function expects the page to be locked and on success it consumes a
3561 * reference of a page being mapped (for the PTE which maps it).
3563 * Return: %0 on success, %VM_FAULT_ code in case of error.
3565 vm_fault_t finish_fault(struct vm_fault *vmf)
3570 /* Did we COW the page? */
3571 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3572 !(vmf->vma->vm_flags & VM_SHARED))
3573 page = vmf->cow_page;
3578 * check even for read faults because we might have lost our CoWed
3581 if (!(vmf->vma->vm_flags & VM_SHARED))
3582 ret = check_stable_address_space(vmf->vma->vm_mm);
3584 ret = alloc_set_pte(vmf, vmf->memcg, page);
3586 pte_unmap_unlock(vmf->pte, vmf->ptl);
3590 static unsigned long fault_around_bytes __read_mostly =
3591 rounddown_pow_of_two(65536);
3593 #ifdef CONFIG_DEBUG_FS
3594 static int fault_around_bytes_get(void *data, u64 *val)
3596 *val = fault_around_bytes;
3601 * fault_around_bytes must be rounded down to the nearest page order as it's
3602 * what do_fault_around() expects to see.
3604 static int fault_around_bytes_set(void *data, u64 val)
3606 if (val / PAGE_SIZE > PTRS_PER_PTE)
3608 if (val > PAGE_SIZE)
3609 fault_around_bytes = rounddown_pow_of_two(val);
3611 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3614 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3615 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3617 static int __init fault_around_debugfs(void)
3619 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3620 &fault_around_bytes_fops);
3623 late_initcall(fault_around_debugfs);
3627 * do_fault_around() tries to map few pages around the fault address. The hope
3628 * is that the pages will be needed soon and this will lower the number of
3631 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3632 * not ready to be mapped: not up-to-date, locked, etc.
3634 * This function is called with the page table lock taken. In the split ptlock
3635 * case the page table lock only protects only those entries which belong to
3636 * the page table corresponding to the fault address.
3638 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3641 * fault_around_bytes defines how many bytes we'll try to map.
3642 * do_fault_around() expects it to be set to a power of two less than or equal
3645 * The virtual address of the area that we map is naturally aligned to
3646 * fault_around_bytes rounded down to the machine page size
3647 * (and therefore to page order). This way it's easier to guarantee
3648 * that we don't cross page table boundaries.
3650 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3652 unsigned long address = vmf->address, nr_pages, mask;
3653 pgoff_t start_pgoff = vmf->pgoff;
3658 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3659 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3661 vmf->address = max(address & mask, vmf->vma->vm_start);
3662 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3666 * end_pgoff is either the end of the page table, the end of
3667 * the vma or nr_pages from start_pgoff, depending what is nearest.
3669 end_pgoff = start_pgoff -
3670 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3672 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3673 start_pgoff + nr_pages - 1);
3675 if (pmd_none(*vmf->pmd)) {
3676 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3677 if (!vmf->prealloc_pte)
3679 smp_wmb(); /* See comment in __pte_alloc() */
3682 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3684 /* Huge page is mapped? Page fault is solved */
3685 if (pmd_trans_huge(*vmf->pmd)) {
3686 ret = VM_FAULT_NOPAGE;
3690 /* ->map_pages() haven't done anything useful. Cold page cache? */
3694 /* check if the page fault is solved */
3695 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3696 if (!pte_none(*vmf->pte))
3697 ret = VM_FAULT_NOPAGE;
3698 pte_unmap_unlock(vmf->pte, vmf->ptl);
3700 vmf->address = address;
3705 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3707 struct vm_area_struct *vma = vmf->vma;
3711 * Let's call ->map_pages() first and use ->fault() as fallback
3712 * if page by the offset is not ready to be mapped (cold cache or
3715 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3716 ret = do_fault_around(vmf);
3721 ret = __do_fault(vmf);
3722 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3725 ret |= finish_fault(vmf);
3726 unlock_page(vmf->page);
3727 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3728 put_page(vmf->page);
3732 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3734 struct vm_area_struct *vma = vmf->vma;
3737 if (unlikely(anon_vma_prepare(vma)))
3738 return VM_FAULT_OOM;
3740 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3742 return VM_FAULT_OOM;
3744 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3745 &vmf->memcg, false)) {
3746 put_page(vmf->cow_page);
3747 return VM_FAULT_OOM;
3750 ret = __do_fault(vmf);
3751 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3753 if (ret & VM_FAULT_DONE_COW)
3756 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3757 __SetPageUptodate(vmf->cow_page);
3759 ret |= finish_fault(vmf);
3760 unlock_page(vmf->page);
3761 put_page(vmf->page);
3762 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3766 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3767 put_page(vmf->cow_page);
3771 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3773 struct vm_area_struct *vma = vmf->vma;
3774 vm_fault_t ret, tmp;
3776 ret = __do_fault(vmf);
3777 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3781 * Check if the backing address space wants to know that the page is
3782 * about to become writable
3784 if (vma->vm_ops->page_mkwrite) {
3785 unlock_page(vmf->page);
3786 tmp = do_page_mkwrite(vmf);
3787 if (unlikely(!tmp ||
3788 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3789 put_page(vmf->page);
3794 ret |= finish_fault(vmf);
3795 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3797 unlock_page(vmf->page);
3798 put_page(vmf->page);
3802 ret |= fault_dirty_shared_page(vmf);
3807 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3808 * but allow concurrent faults).
3809 * The mmap_sem may have been released depending on flags and our
3810 * return value. See filemap_fault() and __lock_page_or_retry().
3811 * If mmap_sem is released, vma may become invalid (for example
3812 * by other thread calling munmap()).
3814 static vm_fault_t do_fault(struct vm_fault *vmf)
3816 struct vm_area_struct *vma = vmf->vma;
3817 struct mm_struct *vm_mm = vma->vm_mm;
3821 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3823 if (!vma->vm_ops->fault) {
3825 * If we find a migration pmd entry or a none pmd entry, which
3826 * should never happen, return SIGBUS
3828 if (unlikely(!pmd_present(*vmf->pmd)))
3829 ret = VM_FAULT_SIGBUS;
3831 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3836 * Make sure this is not a temporary clearing of pte
3837 * by holding ptl and checking again. A R/M/W update
3838 * of pte involves: take ptl, clearing the pte so that
3839 * we don't have concurrent modification by hardware
3840 * followed by an update.
3842 if (unlikely(pte_none(*vmf->pte)))
3843 ret = VM_FAULT_SIGBUS;
3845 ret = VM_FAULT_NOPAGE;
3847 pte_unmap_unlock(vmf->pte, vmf->ptl);
3849 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3850 ret = do_read_fault(vmf);
3851 else if (!(vma->vm_flags & VM_SHARED))
3852 ret = do_cow_fault(vmf);
3854 ret = do_shared_fault(vmf);
3856 /* preallocated pagetable is unused: free it */
3857 if (vmf->prealloc_pte) {
3858 pte_free(vm_mm, vmf->prealloc_pte);
3859 vmf->prealloc_pte = NULL;
3864 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3865 unsigned long addr, int page_nid,
3870 count_vm_numa_event(NUMA_HINT_FAULTS);
3871 if (page_nid == numa_node_id()) {
3872 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3873 *flags |= TNF_FAULT_LOCAL;
3876 return mpol_misplaced(page, vma, addr);
3879 static vm_fault_t do_numa_page(struct vm_fault *vmf)
3881 struct vm_area_struct *vma = vmf->vma;
3882 struct page *page = NULL;
3883 int page_nid = NUMA_NO_NODE;
3886 bool migrated = false;
3888 bool was_writable = pte_savedwrite(vmf->orig_pte);
3892 * The "pte" at this point cannot be used safely without
3893 * validation through pte_unmap_same(). It's of NUMA type but
3894 * the pfn may be screwed if the read is non atomic.
3896 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3897 spin_lock(vmf->ptl);
3898 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3899 pte_unmap_unlock(vmf->pte, vmf->ptl);
3904 * Make it present again, Depending on how arch implementes non
3905 * accessible ptes, some can allow access by kernel mode.
3907 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
3908 pte = pte_modify(old_pte, vma->vm_page_prot);
3909 pte = pte_mkyoung(pte);
3911 pte = pte_mkwrite(pte);
3912 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
3913 update_mmu_cache(vma, vmf->address, vmf->pte);
3915 page = vm_normal_page(vma, vmf->address, pte);
3917 pte_unmap_unlock(vmf->pte, vmf->ptl);
3921 /* TODO: handle PTE-mapped THP */
3922 if (PageCompound(page)) {
3923 pte_unmap_unlock(vmf->pte, vmf->ptl);
3928 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3929 * much anyway since they can be in shared cache state. This misses
3930 * the case where a mapping is writable but the process never writes
3931 * to it but pte_write gets cleared during protection updates and
3932 * pte_dirty has unpredictable behaviour between PTE scan updates,
3933 * background writeback, dirty balancing and application behaviour.
3935 if (!pte_write(pte))
3936 flags |= TNF_NO_GROUP;
3939 * Flag if the page is shared between multiple address spaces. This
3940 * is later used when determining whether to group tasks together
3942 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3943 flags |= TNF_SHARED;
3945 last_cpupid = page_cpupid_last(page);
3946 page_nid = page_to_nid(page);
3947 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3949 pte_unmap_unlock(vmf->pte, vmf->ptl);
3950 if (target_nid == NUMA_NO_NODE) {
3955 /* Migrate to the requested node */
3956 migrated = migrate_misplaced_page(page, vma, target_nid);
3958 page_nid = target_nid;
3959 flags |= TNF_MIGRATED;
3961 flags |= TNF_MIGRATE_FAIL;
3964 if (page_nid != NUMA_NO_NODE)
3965 task_numa_fault(last_cpupid, page_nid, 1, flags);
3969 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3971 if (vma_is_anonymous(vmf->vma))
3972 return do_huge_pmd_anonymous_page(vmf);
3973 if (vmf->vma->vm_ops->huge_fault)
3974 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3975 return VM_FAULT_FALLBACK;
3978 /* `inline' is required to avoid gcc 4.1.2 build error */
3979 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3981 if (vma_is_anonymous(vmf->vma)) {
3982 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
3983 return handle_userfault(vmf, VM_UFFD_WP);
3984 return do_huge_pmd_wp_page(vmf, orig_pmd);
3986 if (vmf->vma->vm_ops->huge_fault) {
3987 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3989 if (!(ret & VM_FAULT_FALLBACK))
3993 /* COW or write-notify handled on pte level: split pmd. */
3994 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3996 return VM_FAULT_FALLBACK;
3999 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4001 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4002 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4003 /* No support for anonymous transparent PUD pages yet */
4004 if (vma_is_anonymous(vmf->vma))
4006 if (vmf->vma->vm_ops->huge_fault) {
4007 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4009 if (!(ret & VM_FAULT_FALLBACK))
4013 /* COW or write-notify not handled on PUD level: split pud.*/
4014 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4015 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4016 return VM_FAULT_FALLBACK;
4019 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4021 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4022 /* No support for anonymous transparent PUD pages yet */
4023 if (vma_is_anonymous(vmf->vma))
4024 return VM_FAULT_FALLBACK;
4025 if (vmf->vma->vm_ops->huge_fault)
4026 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4027 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4028 return VM_FAULT_FALLBACK;
4032 * These routines also need to handle stuff like marking pages dirty
4033 * and/or accessed for architectures that don't do it in hardware (most
4034 * RISC architectures). The early dirtying is also good on the i386.
4036 * There is also a hook called "update_mmu_cache()" that architectures
4037 * with external mmu caches can use to update those (ie the Sparc or
4038 * PowerPC hashed page tables that act as extended TLBs).
4040 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
4041 * concurrent faults).
4043 * The mmap_sem may have been released depending on flags and our return value.
4044 * See filemap_fault() and __lock_page_or_retry().
4046 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4050 if (unlikely(pmd_none(*vmf->pmd))) {
4052 * Leave __pte_alloc() until later: because vm_ops->fault may
4053 * want to allocate huge page, and if we expose page table
4054 * for an instant, it will be difficult to retract from
4055 * concurrent faults and from rmap lookups.
4059 /* See comment in pte_alloc_one_map() */
4060 if (pmd_devmap_trans_unstable(vmf->pmd))
4063 * A regular pmd is established and it can't morph into a huge
4064 * pmd from under us anymore at this point because we hold the
4065 * mmap_sem read mode and khugepaged takes it in write mode.
4066 * So now it's safe to run pte_offset_map().
4068 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4069 vmf->orig_pte = *vmf->pte;
4072 * some architectures can have larger ptes than wordsize,
4073 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4074 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4075 * accesses. The code below just needs a consistent view
4076 * for the ifs and we later double check anyway with the
4077 * ptl lock held. So here a barrier will do.
4080 if (pte_none(vmf->orig_pte)) {
4081 pte_unmap(vmf->pte);
4087 if (vma_is_anonymous(vmf->vma))
4088 return do_anonymous_page(vmf);
4090 return do_fault(vmf);
4093 if (!pte_present(vmf->orig_pte))
4094 return do_swap_page(vmf);
4096 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4097 return do_numa_page(vmf);
4099 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4100 spin_lock(vmf->ptl);
4101 entry = vmf->orig_pte;
4102 if (unlikely(!pte_same(*vmf->pte, entry)))
4104 if (vmf->flags & FAULT_FLAG_WRITE) {
4105 if (!pte_write(entry))
4106 return do_wp_page(vmf);
4107 entry = pte_mkdirty(entry);
4109 entry = pte_mkyoung(entry);
4110 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4111 vmf->flags & FAULT_FLAG_WRITE)) {
4112 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4115 * This is needed only for protection faults but the arch code
4116 * is not yet telling us if this is a protection fault or not.
4117 * This still avoids useless tlb flushes for .text page faults
4120 if (vmf->flags & FAULT_FLAG_WRITE)
4121 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4124 pte_unmap_unlock(vmf->pte, vmf->ptl);
4129 * By the time we get here, we already hold the mm semaphore
4131 * The mmap_sem may have been released depending on flags and our
4132 * return value. See filemap_fault() and __lock_page_or_retry().
4134 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4135 unsigned long address, unsigned int flags)
4137 struct vm_fault vmf = {
4139 .address = address & PAGE_MASK,
4141 .pgoff = linear_page_index(vma, address),
4142 .gfp_mask = __get_fault_gfp_mask(vma),
4144 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4145 struct mm_struct *mm = vma->vm_mm;
4150 pgd = pgd_offset(mm, address);
4151 p4d = p4d_alloc(mm, pgd, address);
4153 return VM_FAULT_OOM;
4155 vmf.pud = pud_alloc(mm, p4d, address);
4157 return VM_FAULT_OOM;
4159 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4160 ret = create_huge_pud(&vmf);
4161 if (!(ret & VM_FAULT_FALLBACK))
4164 pud_t orig_pud = *vmf.pud;
4167 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4169 /* NUMA case for anonymous PUDs would go here */
4171 if (dirty && !pud_write(orig_pud)) {
4172 ret = wp_huge_pud(&vmf, orig_pud);
4173 if (!(ret & VM_FAULT_FALLBACK))
4176 huge_pud_set_accessed(&vmf, orig_pud);
4182 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4184 return VM_FAULT_OOM;
4186 /* Huge pud page fault raced with pmd_alloc? */
4187 if (pud_trans_unstable(vmf.pud))
4190 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4191 ret = create_huge_pmd(&vmf);
4192 if (!(ret & VM_FAULT_FALLBACK))
4195 pmd_t orig_pmd = *vmf.pmd;
4198 if (unlikely(is_swap_pmd(orig_pmd))) {
4199 VM_BUG_ON(thp_migration_supported() &&
4200 !is_pmd_migration_entry(orig_pmd));
4201 if (is_pmd_migration_entry(orig_pmd))
4202 pmd_migration_entry_wait(mm, vmf.pmd);
4205 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4206 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4207 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4209 if (dirty && !pmd_write(orig_pmd)) {
4210 ret = wp_huge_pmd(&vmf, orig_pmd);
4211 if (!(ret & VM_FAULT_FALLBACK))
4214 huge_pmd_set_accessed(&vmf, orig_pmd);
4220 return handle_pte_fault(&vmf);
4224 * By the time we get here, we already hold the mm semaphore
4226 * The mmap_sem may have been released depending on flags and our
4227 * return value. See filemap_fault() and __lock_page_or_retry().
4229 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4234 __set_current_state(TASK_RUNNING);
4236 count_vm_event(PGFAULT);
4237 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4239 /* do counter updates before entering really critical section. */
4240 check_sync_rss_stat(current);
4242 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4243 flags & FAULT_FLAG_INSTRUCTION,
4244 flags & FAULT_FLAG_REMOTE))
4245 return VM_FAULT_SIGSEGV;
4248 * Enable the memcg OOM handling for faults triggered in user
4249 * space. Kernel faults are handled more gracefully.
4251 if (flags & FAULT_FLAG_USER)
4252 mem_cgroup_enter_user_fault();
4254 if (unlikely(is_vm_hugetlb_page(vma)))
4255 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4257 ret = __handle_mm_fault(vma, address, flags);
4259 if (flags & FAULT_FLAG_USER) {
4260 mem_cgroup_exit_user_fault();
4262 * The task may have entered a memcg OOM situation but
4263 * if the allocation error was handled gracefully (no
4264 * VM_FAULT_OOM), there is no need to kill anything.
4265 * Just clean up the OOM state peacefully.
4267 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4268 mem_cgroup_oom_synchronize(false);
4273 EXPORT_SYMBOL_GPL(handle_mm_fault);
4275 #ifndef __PAGETABLE_P4D_FOLDED
4277 * Allocate p4d page table.
4278 * We've already handled the fast-path in-line.
4280 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4282 p4d_t *new = p4d_alloc_one(mm, address);
4286 smp_wmb(); /* See comment in __pte_alloc */
4288 spin_lock(&mm->page_table_lock);
4289 if (pgd_present(*pgd)) /* Another has populated it */
4292 pgd_populate(mm, pgd, new);
4293 spin_unlock(&mm->page_table_lock);
4296 #endif /* __PAGETABLE_P4D_FOLDED */
4298 #ifndef __PAGETABLE_PUD_FOLDED
4300 * Allocate page upper directory.
4301 * We've already handled the fast-path in-line.
4303 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4305 pud_t *new = pud_alloc_one(mm, address);
4309 smp_wmb(); /* See comment in __pte_alloc */
4311 spin_lock(&mm->page_table_lock);
4312 #ifndef __ARCH_HAS_5LEVEL_HACK
4313 if (!p4d_present(*p4d)) {
4315 p4d_populate(mm, p4d, new);
4316 } else /* Another has populated it */
4319 if (!pgd_present(*p4d)) {
4321 pgd_populate(mm, p4d, new);
4322 } else /* Another has populated it */
4324 #endif /* __ARCH_HAS_5LEVEL_HACK */
4325 spin_unlock(&mm->page_table_lock);
4328 #endif /* __PAGETABLE_PUD_FOLDED */
4330 #ifndef __PAGETABLE_PMD_FOLDED
4332 * Allocate page middle directory.
4333 * We've already handled the fast-path in-line.
4335 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4338 pmd_t *new = pmd_alloc_one(mm, address);
4342 smp_wmb(); /* See comment in __pte_alloc */
4344 ptl = pud_lock(mm, pud);
4345 if (!pud_present(*pud)) {
4347 pud_populate(mm, pud, new);
4348 } else /* Another has populated it */
4353 #endif /* __PAGETABLE_PMD_FOLDED */
4355 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4356 struct mmu_notifier_range *range,
4357 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4365 pgd = pgd_offset(mm, address);
4366 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4369 p4d = p4d_offset(pgd, address);
4370 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4373 pud = pud_offset(p4d, address);
4374 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4377 pmd = pmd_offset(pud, address);
4378 VM_BUG_ON(pmd_trans_huge(*pmd));
4380 if (pmd_huge(*pmd)) {
4385 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4386 NULL, mm, address & PMD_MASK,
4387 (address & PMD_MASK) + PMD_SIZE);
4388 mmu_notifier_invalidate_range_start(range);
4390 *ptlp = pmd_lock(mm, pmd);
4391 if (pmd_huge(*pmd)) {
4397 mmu_notifier_invalidate_range_end(range);
4400 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4404 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4405 address & PAGE_MASK,
4406 (address & PAGE_MASK) + PAGE_SIZE);
4407 mmu_notifier_invalidate_range_start(range);
4409 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4410 if (!pte_present(*ptep))
4415 pte_unmap_unlock(ptep, *ptlp);
4417 mmu_notifier_invalidate_range_end(range);
4422 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4423 pte_t **ptepp, spinlock_t **ptlp)
4427 /* (void) is needed to make gcc happy */
4428 (void) __cond_lock(*ptlp,
4429 !(res = __follow_pte_pmd(mm, address, NULL,
4430 ptepp, NULL, ptlp)));
4434 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4435 struct mmu_notifier_range *range,
4436 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4440 /* (void) is needed to make gcc happy */
4441 (void) __cond_lock(*ptlp,
4442 !(res = __follow_pte_pmd(mm, address, range,
4443 ptepp, pmdpp, ptlp)));
4446 EXPORT_SYMBOL(follow_pte_pmd);
4449 * follow_pfn - look up PFN at a user virtual address
4450 * @vma: memory mapping
4451 * @address: user virtual address
4452 * @pfn: location to store found PFN
4454 * Only IO mappings and raw PFN mappings are allowed.
4456 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4458 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4465 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4468 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4471 *pfn = pte_pfn(*ptep);
4472 pte_unmap_unlock(ptep, ptl);
4475 EXPORT_SYMBOL(follow_pfn);
4477 #ifdef CONFIG_HAVE_IOREMAP_PROT
4478 int follow_phys(struct vm_area_struct *vma,
4479 unsigned long address, unsigned int flags,
4480 unsigned long *prot, resource_size_t *phys)
4486 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4489 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4493 if ((flags & FOLL_WRITE) && !pte_write(pte))
4496 *prot = pgprot_val(pte_pgprot(pte));
4497 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4501 pte_unmap_unlock(ptep, ptl);
4506 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4507 void *buf, int len, int write)
4509 resource_size_t phys_addr;
4510 unsigned long prot = 0;
4511 void __iomem *maddr;
4512 int offset = addr & (PAGE_SIZE-1);
4514 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4517 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4522 memcpy_toio(maddr + offset, buf, len);
4524 memcpy_fromio(buf, maddr + offset, len);
4529 EXPORT_SYMBOL_GPL(generic_access_phys);
4533 * Access another process' address space as given in mm. If non-NULL, use the
4534 * given task for page fault accounting.
4536 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4537 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4539 struct vm_area_struct *vma;
4540 void *old_buf = buf;
4541 int write = gup_flags & FOLL_WRITE;
4543 if (down_read_killable(&mm->mmap_sem))
4546 /* ignore errors, just check how much was successfully transferred */
4548 int bytes, ret, offset;
4550 struct page *page = NULL;
4552 ret = get_user_pages_remote(tsk, mm, addr, 1,
4553 gup_flags, &page, &vma, NULL);
4555 #ifndef CONFIG_HAVE_IOREMAP_PROT
4559 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4560 * we can access using slightly different code.
4562 vma = find_vma(mm, addr);
4563 if (!vma || vma->vm_start > addr)
4565 if (vma->vm_ops && vma->vm_ops->access)
4566 ret = vma->vm_ops->access(vma, addr, buf,
4574 offset = addr & (PAGE_SIZE-1);
4575 if (bytes > PAGE_SIZE-offset)
4576 bytes = PAGE_SIZE-offset;
4580 copy_to_user_page(vma, page, addr,
4581 maddr + offset, buf, bytes);
4582 set_page_dirty_lock(page);
4584 copy_from_user_page(vma, page, addr,
4585 buf, maddr + offset, bytes);
4594 up_read(&mm->mmap_sem);
4596 return buf - old_buf;
4600 * access_remote_vm - access another process' address space
4601 * @mm: the mm_struct of the target address space
4602 * @addr: start address to access
4603 * @buf: source or destination buffer
4604 * @len: number of bytes to transfer
4605 * @gup_flags: flags modifying lookup behaviour
4607 * The caller must hold a reference on @mm.
4609 * Return: number of bytes copied from source to destination.
4611 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4612 void *buf, int len, unsigned int gup_flags)
4614 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4618 * Access another process' address space.
4619 * Source/target buffer must be kernel space,
4620 * Do not walk the page table directly, use get_user_pages
4622 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4623 void *buf, int len, unsigned int gup_flags)
4625 struct mm_struct *mm;
4628 mm = get_task_mm(tsk);
4632 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4638 EXPORT_SYMBOL_GPL(access_process_vm);
4641 * Print the name of a VMA.
4643 void print_vma_addr(char *prefix, unsigned long ip)
4645 struct mm_struct *mm = current->mm;
4646 struct vm_area_struct *vma;
4649 * we might be running from an atomic context so we cannot sleep
4651 if (!down_read_trylock(&mm->mmap_sem))
4654 vma = find_vma(mm, ip);
4655 if (vma && vma->vm_file) {
4656 struct file *f = vma->vm_file;
4657 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4661 p = file_path(f, buf, PAGE_SIZE);
4664 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4666 vma->vm_end - vma->vm_start);
4667 free_page((unsigned long)buf);
4670 up_read(&mm->mmap_sem);
4673 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4674 void __might_fault(const char *file, int line)
4677 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4678 * holding the mmap_sem, this is safe because kernel memory doesn't
4679 * get paged out, therefore we'll never actually fault, and the
4680 * below annotations will generate false positives.
4682 if (uaccess_kernel())
4684 if (pagefault_disabled())
4686 __might_sleep(file, line, 0);
4687 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4689 might_lock_read(¤t->mm->mmap_sem);
4692 EXPORT_SYMBOL(__might_fault);
4695 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4697 * Process all subpages of the specified huge page with the specified
4698 * operation. The target subpage will be processed last to keep its
4701 static inline void process_huge_page(
4702 unsigned long addr_hint, unsigned int pages_per_huge_page,
4703 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4707 unsigned long addr = addr_hint &
4708 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4710 /* Process target subpage last to keep its cache lines hot */
4712 n = (addr_hint - addr) / PAGE_SIZE;
4713 if (2 * n <= pages_per_huge_page) {
4714 /* If target subpage in first half of huge page */
4717 /* Process subpages at the end of huge page */
4718 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4720 process_subpage(addr + i * PAGE_SIZE, i, arg);
4723 /* If target subpage in second half of huge page */
4724 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4725 l = pages_per_huge_page - n;
4726 /* Process subpages at the begin of huge page */
4727 for (i = 0; i < base; i++) {
4729 process_subpage(addr + i * PAGE_SIZE, i, arg);
4733 * Process remaining subpages in left-right-left-right pattern
4734 * towards the target subpage
4736 for (i = 0; i < l; i++) {
4737 int left_idx = base + i;
4738 int right_idx = base + 2 * l - 1 - i;
4741 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4743 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4747 static void clear_gigantic_page(struct page *page,
4749 unsigned int pages_per_huge_page)
4752 struct page *p = page;
4755 for (i = 0; i < pages_per_huge_page;
4756 i++, p = mem_map_next(p, page, i)) {
4758 clear_user_highpage(p, addr + i * PAGE_SIZE);
4762 static void clear_subpage(unsigned long addr, int idx, void *arg)
4764 struct page *page = arg;
4766 clear_user_highpage(page + idx, addr);
4769 void clear_huge_page(struct page *page,
4770 unsigned long addr_hint, unsigned int pages_per_huge_page)
4772 unsigned long addr = addr_hint &
4773 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4775 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4776 clear_gigantic_page(page, addr, pages_per_huge_page);
4780 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4783 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4785 struct vm_area_struct *vma,
4786 unsigned int pages_per_huge_page)
4789 struct page *dst_base = dst;
4790 struct page *src_base = src;
4792 for (i = 0; i < pages_per_huge_page; ) {
4794 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4797 dst = mem_map_next(dst, dst_base, i);
4798 src = mem_map_next(src, src_base, i);
4802 struct copy_subpage_arg {
4805 struct vm_area_struct *vma;
4808 static void copy_subpage(unsigned long addr, int idx, void *arg)
4810 struct copy_subpage_arg *copy_arg = arg;
4812 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4813 addr, copy_arg->vma);
4816 void copy_user_huge_page(struct page *dst, struct page *src,
4817 unsigned long addr_hint, struct vm_area_struct *vma,
4818 unsigned int pages_per_huge_page)
4820 unsigned long addr = addr_hint &
4821 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4822 struct copy_subpage_arg arg = {
4828 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4829 copy_user_gigantic_page(dst, src, addr, vma,
4830 pages_per_huge_page);
4834 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4837 long copy_huge_page_from_user(struct page *dst_page,
4838 const void __user *usr_src,
4839 unsigned int pages_per_huge_page,
4840 bool allow_pagefault)
4842 void *src = (void *)usr_src;
4844 unsigned long i, rc = 0;
4845 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4847 for (i = 0; i < pages_per_huge_page; i++) {
4848 if (allow_pagefault)
4849 page_kaddr = kmap(dst_page + i);
4851 page_kaddr = kmap_atomic(dst_page + i);
4852 rc = copy_from_user(page_kaddr,
4853 (const void __user *)(src + i * PAGE_SIZE),
4855 if (allow_pagefault)
4856 kunmap(dst_page + i);
4858 kunmap_atomic(page_kaddr);
4860 ret_val -= (PAGE_SIZE - rc);
4868 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4870 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4872 static struct kmem_cache *page_ptl_cachep;
4874 void __init ptlock_cache_init(void)
4876 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4880 bool ptlock_alloc(struct page *page)
4884 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4891 void ptlock_free(struct page *page)
4893 kmem_cache_free(page_ptl_cachep, page->ptl);