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>
76 #include <asm/mmu_context.h>
77 #include <asm/pgalloc.h>
78 #include <linux/uaccess.h>
80 #include <asm/tlbflush.h>
81 #include <asm/pgtable.h>
85 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
86 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
89 #ifndef CONFIG_NEED_MULTIPLE_NODES
90 /* use the per-pgdat data instead for discontigmem - mbligh */
91 unsigned long max_mapnr;
92 EXPORT_SYMBOL(max_mapnr);
95 EXPORT_SYMBOL(mem_map);
99 * A number of key systems in x86 including ioremap() rely on the assumption
100 * that high_memory defines the upper bound on direct map memory, then end
101 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
102 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
106 EXPORT_SYMBOL(high_memory);
109 * Randomize the address space (stacks, mmaps, brk, etc.).
111 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
112 * as ancient (libc5 based) binaries can segfault. )
114 int randomize_va_space __read_mostly =
115 #ifdef CONFIG_COMPAT_BRK
121 static int __init disable_randmaps(char *s)
123 randomize_va_space = 0;
126 __setup("norandmaps", disable_randmaps);
128 unsigned long zero_pfn __read_mostly;
129 EXPORT_SYMBOL(zero_pfn);
131 unsigned long highest_memmap_pfn __read_mostly;
134 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
136 static int __init init_zero_pfn(void)
138 zero_pfn = page_to_pfn(ZERO_PAGE(0));
141 core_initcall(init_zero_pfn);
144 #if defined(SPLIT_RSS_COUNTING)
146 void sync_mm_rss(struct mm_struct *mm)
150 for (i = 0; i < NR_MM_COUNTERS; i++) {
151 if (current->rss_stat.count[i]) {
152 add_mm_counter(mm, i, current->rss_stat.count[i]);
153 current->rss_stat.count[i] = 0;
156 current->rss_stat.events = 0;
159 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
161 struct task_struct *task = current;
163 if (likely(task->mm == mm))
164 task->rss_stat.count[member] += val;
166 add_mm_counter(mm, member, val);
168 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
169 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
171 /* sync counter once per 64 page faults */
172 #define TASK_RSS_EVENTS_THRESH (64)
173 static void check_sync_rss_stat(struct task_struct *task)
175 if (unlikely(task != current))
177 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
178 sync_mm_rss(task->mm);
180 #else /* SPLIT_RSS_COUNTING */
182 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
183 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
185 static void check_sync_rss_stat(struct task_struct *task)
189 #endif /* SPLIT_RSS_COUNTING */
192 * Note: this doesn't free the actual pages themselves. That
193 * has been handled earlier when unmapping all the memory regions.
195 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
198 pgtable_t token = pmd_pgtable(*pmd);
200 pte_free_tlb(tlb, token, addr);
201 mm_dec_nr_ptes(tlb->mm);
204 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
205 unsigned long addr, unsigned long end,
206 unsigned long floor, unsigned long ceiling)
213 pmd = pmd_offset(pud, addr);
215 next = pmd_addr_end(addr, end);
216 if (pmd_none_or_clear_bad(pmd))
218 free_pte_range(tlb, pmd, addr);
219 } while (pmd++, addr = next, addr != end);
229 if (end - 1 > ceiling - 1)
232 pmd = pmd_offset(pud, start);
234 pmd_free_tlb(tlb, pmd, start);
235 mm_dec_nr_pmds(tlb->mm);
238 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
239 unsigned long addr, unsigned long end,
240 unsigned long floor, unsigned long ceiling)
247 pud = pud_offset(p4d, addr);
249 next = pud_addr_end(addr, end);
250 if (pud_none_or_clear_bad(pud))
252 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
253 } while (pud++, addr = next, addr != end);
263 if (end - 1 > ceiling - 1)
266 pud = pud_offset(p4d, start);
268 pud_free_tlb(tlb, pud, start);
269 mm_dec_nr_puds(tlb->mm);
272 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
273 unsigned long addr, unsigned long end,
274 unsigned long floor, unsigned long ceiling)
281 p4d = p4d_offset(pgd, addr);
283 next = p4d_addr_end(addr, end);
284 if (p4d_none_or_clear_bad(p4d))
286 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
287 } while (p4d++, addr = next, addr != end);
293 ceiling &= PGDIR_MASK;
297 if (end - 1 > ceiling - 1)
300 p4d = p4d_offset(pgd, start);
302 p4d_free_tlb(tlb, p4d, start);
306 * This function frees user-level page tables of a process.
308 void free_pgd_range(struct mmu_gather *tlb,
309 unsigned long addr, unsigned long end,
310 unsigned long floor, unsigned long ceiling)
316 * The next few lines have given us lots of grief...
318 * Why are we testing PMD* at this top level? Because often
319 * there will be no work to do at all, and we'd prefer not to
320 * go all the way down to the bottom just to discover that.
322 * Why all these "- 1"s? Because 0 represents both the bottom
323 * of the address space and the top of it (using -1 for the
324 * top wouldn't help much: the masks would do the wrong thing).
325 * The rule is that addr 0 and floor 0 refer to the bottom of
326 * the address space, but end 0 and ceiling 0 refer to the top
327 * Comparisons need to use "end - 1" and "ceiling - 1" (though
328 * that end 0 case should be mythical).
330 * Wherever addr is brought up or ceiling brought down, we must
331 * be careful to reject "the opposite 0" before it confuses the
332 * subsequent tests. But what about where end is brought down
333 * by PMD_SIZE below? no, end can't go down to 0 there.
335 * Whereas we round start (addr) and ceiling down, by different
336 * masks at different levels, in order to test whether a table
337 * now has no other vmas using it, so can be freed, we don't
338 * bother to round floor or end up - the tests don't need that.
352 if (end - 1 > ceiling - 1)
357 * We add page table cache pages with PAGE_SIZE,
358 * (see pte_free_tlb()), flush the tlb if we need
360 tlb_change_page_size(tlb, PAGE_SIZE);
361 pgd = pgd_offset(tlb->mm, addr);
363 next = pgd_addr_end(addr, end);
364 if (pgd_none_or_clear_bad(pgd))
366 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
367 } while (pgd++, addr = next, addr != end);
370 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
371 unsigned long floor, unsigned long ceiling)
374 struct vm_area_struct *next = vma->vm_next;
375 unsigned long addr = vma->vm_start;
378 * Hide vma from rmap and truncate_pagecache before freeing
381 unlink_anon_vmas(vma);
382 unlink_file_vma(vma);
384 if (is_vm_hugetlb_page(vma)) {
385 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
386 floor, next ? next->vm_start : ceiling);
389 * Optimization: gather nearby vmas into one call down
391 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
392 && !is_vm_hugetlb_page(next)) {
395 unlink_anon_vmas(vma);
396 unlink_file_vma(vma);
398 free_pgd_range(tlb, addr, vma->vm_end,
399 floor, next ? next->vm_start : ceiling);
405 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
408 pgtable_t new = pte_alloc_one(mm);
413 * Ensure all pte setup (eg. pte page lock and page clearing) are
414 * visible before the pte is made visible to other CPUs by being
415 * put into page tables.
417 * The other side of the story is the pointer chasing in the page
418 * table walking code (when walking the page table without locking;
419 * ie. most of the time). Fortunately, these data accesses consist
420 * of a chain of data-dependent loads, meaning most CPUs (alpha
421 * being the notable exception) will already guarantee loads are
422 * seen in-order. See the alpha page table accessors for the
423 * smp_read_barrier_depends() barriers in page table walking code.
425 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
427 ptl = pmd_lock(mm, pmd);
428 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
430 pmd_populate(mm, pmd, new);
439 int __pte_alloc_kernel(pmd_t *pmd)
441 pte_t *new = pte_alloc_one_kernel(&init_mm);
445 smp_wmb(); /* See comment in __pte_alloc */
447 spin_lock(&init_mm.page_table_lock);
448 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
449 pmd_populate_kernel(&init_mm, pmd, new);
452 spin_unlock(&init_mm.page_table_lock);
454 pte_free_kernel(&init_mm, new);
458 static inline void init_rss_vec(int *rss)
460 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
463 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
467 if (current->mm == mm)
469 for (i = 0; i < NR_MM_COUNTERS; i++)
471 add_mm_counter(mm, i, rss[i]);
475 * This function is called to print an error when a bad pte
476 * is found. For example, we might have a PFN-mapped pte in
477 * a region that doesn't allow it.
479 * The calling function must still handle the error.
481 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
482 pte_t pte, struct page *page)
484 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
485 p4d_t *p4d = p4d_offset(pgd, addr);
486 pud_t *pud = pud_offset(p4d, addr);
487 pmd_t *pmd = pmd_offset(pud, addr);
488 struct address_space *mapping;
490 static unsigned long resume;
491 static unsigned long nr_shown;
492 static unsigned long nr_unshown;
495 * Allow a burst of 60 reports, then keep quiet for that minute;
496 * or allow a steady drip of one report per second.
498 if (nr_shown == 60) {
499 if (time_before(jiffies, resume)) {
504 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
511 resume = jiffies + 60 * HZ;
513 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
514 index = linear_page_index(vma, addr);
516 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
518 (long long)pte_val(pte), (long long)pmd_val(*pmd));
520 dump_page(page, "bad pte");
521 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
522 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
523 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
525 vma->vm_ops ? vma->vm_ops->fault : NULL,
526 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
527 mapping ? mapping->a_ops->readpage : NULL);
529 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
533 * vm_normal_page -- This function gets the "struct page" associated with a pte.
535 * "Special" mappings do not wish to be associated with a "struct page" (either
536 * it doesn't exist, or it exists but they don't want to touch it). In this
537 * case, NULL is returned here. "Normal" mappings do have a struct page.
539 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
540 * pte bit, in which case this function is trivial. Secondly, an architecture
541 * may not have a spare pte bit, which requires a more complicated scheme,
544 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
545 * special mapping (even if there are underlying and valid "struct pages").
546 * COWed pages of a VM_PFNMAP are always normal.
548 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
549 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
550 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
551 * mapping will always honor the rule
553 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
555 * And for normal mappings this is false.
557 * This restricts such mappings to be a linear translation from virtual address
558 * to pfn. To get around this restriction, we allow arbitrary mappings so long
559 * as the vma is not a COW mapping; in that case, we know that all ptes are
560 * special (because none can have been COWed).
563 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
565 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
566 * page" backing, however the difference is that _all_ pages with a struct
567 * page (that is, those where pfn_valid is true) are refcounted and considered
568 * normal pages by the VM. The disadvantage is that pages are refcounted
569 * (which can be slower and simply not an option for some PFNMAP users). The
570 * advantage is that we don't have to follow the strict linearity rule of
571 * PFNMAP mappings in order to support COWable mappings.
574 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
575 pte_t pte, bool with_public_device)
577 unsigned long pfn = pte_pfn(pte);
579 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
580 if (likely(!pte_special(pte)))
582 if (vma->vm_ops && vma->vm_ops->find_special_page)
583 return vma->vm_ops->find_special_page(vma, addr);
584 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
586 if (is_zero_pfn(pfn))
590 * Device public pages are special pages (they are ZONE_DEVICE
591 * pages but different from persistent memory). They behave
592 * allmost like normal pages. The difference is that they are
593 * not on the lru and thus should never be involve with any-
594 * thing that involve lru manipulation (mlock, numa balancing,
597 * This is why we still want to return NULL for such page from
598 * vm_normal_page() so that we do not have to special case all
599 * call site of vm_normal_page().
601 if (likely(pfn <= highest_memmap_pfn)) {
602 struct page *page = pfn_to_page(pfn);
604 if (is_device_public_page(page)) {
605 if (with_public_device)
614 print_bad_pte(vma, addr, pte, NULL);
618 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
620 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
621 if (vma->vm_flags & VM_MIXEDMAP) {
627 off = (addr - vma->vm_start) >> PAGE_SHIFT;
628 if (pfn == vma->vm_pgoff + off)
630 if (!is_cow_mapping(vma->vm_flags))
635 if (is_zero_pfn(pfn))
639 if (unlikely(pfn > highest_memmap_pfn)) {
640 print_bad_pte(vma, addr, pte, NULL);
645 * NOTE! We still have PageReserved() pages in the page tables.
646 * eg. VDSO mappings can cause them to exist.
649 return pfn_to_page(pfn);
652 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
653 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
656 unsigned long pfn = pmd_pfn(pmd);
659 * There is no pmd_special() but there may be special pmds, e.g.
660 * in a direct-access (dax) mapping, so let's just replicate the
661 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
663 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
664 if (vma->vm_flags & VM_MIXEDMAP) {
670 off = (addr - vma->vm_start) >> PAGE_SHIFT;
671 if (pfn == vma->vm_pgoff + off)
673 if (!is_cow_mapping(vma->vm_flags))
680 if (is_zero_pfn(pfn))
682 if (unlikely(pfn > highest_memmap_pfn))
686 * NOTE! We still have PageReserved() pages in the page tables.
687 * eg. VDSO mappings can cause them to exist.
690 return pfn_to_page(pfn);
695 * copy one vm_area from one task to the other. Assumes the page tables
696 * already present in the new task to be cleared in the whole range
697 * covered by this vma.
700 static inline unsigned long
701 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
702 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
703 unsigned long addr, int *rss)
705 unsigned long vm_flags = vma->vm_flags;
706 pte_t pte = *src_pte;
709 /* pte contains position in swap or file, so copy. */
710 if (unlikely(!pte_present(pte))) {
711 swp_entry_t entry = pte_to_swp_entry(pte);
713 if (likely(!non_swap_entry(entry))) {
714 if (swap_duplicate(entry) < 0)
717 /* make sure dst_mm is on swapoff's mmlist. */
718 if (unlikely(list_empty(&dst_mm->mmlist))) {
719 spin_lock(&mmlist_lock);
720 if (list_empty(&dst_mm->mmlist))
721 list_add(&dst_mm->mmlist,
723 spin_unlock(&mmlist_lock);
726 } else if (is_migration_entry(entry)) {
727 page = migration_entry_to_page(entry);
729 rss[mm_counter(page)]++;
731 if (is_write_migration_entry(entry) &&
732 is_cow_mapping(vm_flags)) {
734 * COW mappings require pages in both
735 * parent and child to be set to read.
737 make_migration_entry_read(&entry);
738 pte = swp_entry_to_pte(entry);
739 if (pte_swp_soft_dirty(*src_pte))
740 pte = pte_swp_mksoft_dirty(pte);
741 set_pte_at(src_mm, addr, src_pte, pte);
743 } else if (is_device_private_entry(entry)) {
744 page = device_private_entry_to_page(entry);
747 * Update rss count even for unaddressable pages, as
748 * they should treated just like normal pages in this
751 * We will likely want to have some new rss counters
752 * for unaddressable pages, at some point. But for now
753 * keep things as they are.
756 rss[mm_counter(page)]++;
757 page_dup_rmap(page, false);
760 * We do not preserve soft-dirty information, because so
761 * far, checkpoint/restore is the only feature that
762 * requires that. And checkpoint/restore does not work
763 * when a device driver is involved (you cannot easily
764 * save and restore device driver state).
766 if (is_write_device_private_entry(entry) &&
767 is_cow_mapping(vm_flags)) {
768 make_device_private_entry_read(&entry);
769 pte = swp_entry_to_pte(entry);
770 set_pte_at(src_mm, addr, src_pte, pte);
777 * If it's a COW mapping, write protect it both
778 * in the parent and the child
780 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
781 ptep_set_wrprotect(src_mm, addr, src_pte);
782 pte = pte_wrprotect(pte);
786 * If it's a shared mapping, mark it clean in
789 if (vm_flags & VM_SHARED)
790 pte = pte_mkclean(pte);
791 pte = pte_mkold(pte);
793 page = vm_normal_page(vma, addr, pte);
796 page_dup_rmap(page, false);
797 rss[mm_counter(page)]++;
798 } else if (pte_devmap(pte)) {
799 page = pte_page(pte);
802 * Cache coherent device memory behave like regular page and
803 * not like persistent memory page. For more informations see
804 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
806 if (is_device_public_page(page)) {
808 page_dup_rmap(page, false);
809 rss[mm_counter(page)]++;
814 set_pte_at(dst_mm, addr, dst_pte, pte);
818 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
819 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
820 unsigned long addr, unsigned long end)
822 pte_t *orig_src_pte, *orig_dst_pte;
823 pte_t *src_pte, *dst_pte;
824 spinlock_t *src_ptl, *dst_ptl;
826 int rss[NR_MM_COUNTERS];
827 swp_entry_t entry = (swp_entry_t){0};
832 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
835 src_pte = pte_offset_map(src_pmd, addr);
836 src_ptl = pte_lockptr(src_mm, src_pmd);
837 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
838 orig_src_pte = src_pte;
839 orig_dst_pte = dst_pte;
840 arch_enter_lazy_mmu_mode();
844 * We are holding two locks at this point - either of them
845 * could generate latencies in another task on another CPU.
847 if (progress >= 32) {
849 if (need_resched() ||
850 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
853 if (pte_none(*src_pte)) {
857 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
862 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
864 arch_leave_lazy_mmu_mode();
865 spin_unlock(src_ptl);
866 pte_unmap(orig_src_pte);
867 add_mm_rss_vec(dst_mm, rss);
868 pte_unmap_unlock(orig_dst_pte, dst_ptl);
872 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
881 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
882 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
883 unsigned long addr, unsigned long end)
885 pmd_t *src_pmd, *dst_pmd;
888 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
891 src_pmd = pmd_offset(src_pud, addr);
893 next = pmd_addr_end(addr, end);
894 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
895 || pmd_devmap(*src_pmd)) {
897 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
898 err = copy_huge_pmd(dst_mm, src_mm,
899 dst_pmd, src_pmd, addr, vma);
906 if (pmd_none_or_clear_bad(src_pmd))
908 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
911 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
915 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
916 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
917 unsigned long addr, unsigned long end)
919 pud_t *src_pud, *dst_pud;
922 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
925 src_pud = pud_offset(src_p4d, addr);
927 next = pud_addr_end(addr, end);
928 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
931 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
932 err = copy_huge_pud(dst_mm, src_mm,
933 dst_pud, src_pud, addr, vma);
940 if (pud_none_or_clear_bad(src_pud))
942 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
945 } while (dst_pud++, src_pud++, addr = next, addr != end);
949 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
950 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
951 unsigned long addr, unsigned long end)
953 p4d_t *src_p4d, *dst_p4d;
956 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
959 src_p4d = p4d_offset(src_pgd, addr);
961 next = p4d_addr_end(addr, end);
962 if (p4d_none_or_clear_bad(src_p4d))
964 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
967 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
971 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
972 struct vm_area_struct *vma)
974 pgd_t *src_pgd, *dst_pgd;
976 unsigned long addr = vma->vm_start;
977 unsigned long end = vma->vm_end;
978 struct mmu_notifier_range range;
983 * Don't copy ptes where a page fault will fill them correctly.
984 * Fork becomes much lighter when there are big shared or private
985 * readonly mappings. The tradeoff is that copy_page_range is more
986 * efficient than faulting.
988 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
992 if (is_vm_hugetlb_page(vma))
993 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
995 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
997 * We do not free on error cases below as remove_vma
998 * gets called on error from higher level routine
1000 ret = track_pfn_copy(vma);
1006 * We need to invalidate the secondary MMU mappings only when
1007 * there could be a permission downgrade on the ptes of the
1008 * parent mm. And a permission downgrade will only happen if
1009 * is_cow_mapping() returns true.
1011 is_cow = is_cow_mapping(vma->vm_flags);
1014 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1015 0, vma, src_mm, addr, end);
1016 mmu_notifier_invalidate_range_start(&range);
1020 dst_pgd = pgd_offset(dst_mm, addr);
1021 src_pgd = pgd_offset(src_mm, addr);
1023 next = pgd_addr_end(addr, end);
1024 if (pgd_none_or_clear_bad(src_pgd))
1026 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1027 vma, addr, next))) {
1031 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1034 mmu_notifier_invalidate_range_end(&range);
1038 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1039 struct vm_area_struct *vma, pmd_t *pmd,
1040 unsigned long addr, unsigned long end,
1041 struct zap_details *details)
1043 struct mm_struct *mm = tlb->mm;
1044 int force_flush = 0;
1045 int rss[NR_MM_COUNTERS];
1051 tlb_change_page_size(tlb, PAGE_SIZE);
1054 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1056 flush_tlb_batched_pending(mm);
1057 arch_enter_lazy_mmu_mode();
1060 if (pte_none(ptent))
1063 if (pte_present(ptent)) {
1066 page = _vm_normal_page(vma, addr, ptent, true);
1067 if (unlikely(details) && page) {
1069 * unmap_shared_mapping_pages() wants to
1070 * invalidate cache without truncating:
1071 * unmap shared but keep private pages.
1073 if (details->check_mapping &&
1074 details->check_mapping != page_rmapping(page))
1077 ptent = ptep_get_and_clear_full(mm, addr, pte,
1079 tlb_remove_tlb_entry(tlb, pte, addr);
1080 if (unlikely(!page))
1083 if (!PageAnon(page)) {
1084 if (pte_dirty(ptent)) {
1086 set_page_dirty(page);
1088 if (pte_young(ptent) &&
1089 likely(!(vma->vm_flags & VM_SEQ_READ)))
1090 mark_page_accessed(page);
1092 rss[mm_counter(page)]--;
1093 page_remove_rmap(page, false);
1094 if (unlikely(page_mapcount(page) < 0))
1095 print_bad_pte(vma, addr, ptent, page);
1096 if (unlikely(__tlb_remove_page(tlb, page))) {
1104 entry = pte_to_swp_entry(ptent);
1105 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1106 struct page *page = device_private_entry_to_page(entry);
1108 if (unlikely(details && details->check_mapping)) {
1110 * unmap_shared_mapping_pages() wants to
1111 * invalidate cache without truncating:
1112 * unmap shared but keep private pages.
1114 if (details->check_mapping !=
1115 page_rmapping(page))
1119 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1120 rss[mm_counter(page)]--;
1121 page_remove_rmap(page, false);
1126 /* If details->check_mapping, we leave swap entries. */
1127 if (unlikely(details))
1130 entry = pte_to_swp_entry(ptent);
1131 if (!non_swap_entry(entry))
1133 else if (is_migration_entry(entry)) {
1136 page = migration_entry_to_page(entry);
1137 rss[mm_counter(page)]--;
1139 if (unlikely(!free_swap_and_cache(entry)))
1140 print_bad_pte(vma, addr, ptent, NULL);
1141 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1142 } while (pte++, addr += PAGE_SIZE, addr != end);
1144 add_mm_rss_vec(mm, rss);
1145 arch_leave_lazy_mmu_mode();
1147 /* Do the actual TLB flush before dropping ptl */
1149 tlb_flush_mmu_tlbonly(tlb);
1150 pte_unmap_unlock(start_pte, ptl);
1153 * If we forced a TLB flush (either due to running out of
1154 * batch buffers or because we needed to flush dirty TLB
1155 * entries before releasing the ptl), free the batched
1156 * memory too. Restart if we didn't do everything.
1168 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1169 struct vm_area_struct *vma, pud_t *pud,
1170 unsigned long addr, unsigned long end,
1171 struct zap_details *details)
1176 pmd = pmd_offset(pud, addr);
1178 next = pmd_addr_end(addr, end);
1179 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1180 if (next - addr != HPAGE_PMD_SIZE)
1181 __split_huge_pmd(vma, pmd, addr, false, NULL);
1182 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1187 * Here there can be other concurrent MADV_DONTNEED or
1188 * trans huge page faults running, and if the pmd is
1189 * none or trans huge it can change under us. This is
1190 * because MADV_DONTNEED holds the mmap_sem in read
1193 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1195 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1198 } while (pmd++, addr = next, addr != end);
1203 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1204 struct vm_area_struct *vma, p4d_t *p4d,
1205 unsigned long addr, unsigned long end,
1206 struct zap_details *details)
1211 pud = pud_offset(p4d, addr);
1213 next = pud_addr_end(addr, end);
1214 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1215 if (next - addr != HPAGE_PUD_SIZE) {
1216 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1217 split_huge_pud(vma, pud, addr);
1218 } else if (zap_huge_pud(tlb, vma, pud, addr))
1222 if (pud_none_or_clear_bad(pud))
1224 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1227 } while (pud++, addr = next, addr != end);
1232 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1233 struct vm_area_struct *vma, pgd_t *pgd,
1234 unsigned long addr, unsigned long end,
1235 struct zap_details *details)
1240 p4d = p4d_offset(pgd, addr);
1242 next = p4d_addr_end(addr, end);
1243 if (p4d_none_or_clear_bad(p4d))
1245 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1246 } while (p4d++, addr = next, addr != end);
1251 void unmap_page_range(struct mmu_gather *tlb,
1252 struct vm_area_struct *vma,
1253 unsigned long addr, unsigned long end,
1254 struct zap_details *details)
1259 BUG_ON(addr >= end);
1260 tlb_start_vma(tlb, vma);
1261 pgd = pgd_offset(vma->vm_mm, addr);
1263 next = pgd_addr_end(addr, end);
1264 if (pgd_none_or_clear_bad(pgd))
1266 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1267 } while (pgd++, addr = next, addr != end);
1268 tlb_end_vma(tlb, vma);
1272 static void unmap_single_vma(struct mmu_gather *tlb,
1273 struct vm_area_struct *vma, unsigned long start_addr,
1274 unsigned long end_addr,
1275 struct zap_details *details)
1277 unsigned long start = max(vma->vm_start, start_addr);
1280 if (start >= vma->vm_end)
1282 end = min(vma->vm_end, end_addr);
1283 if (end <= vma->vm_start)
1287 uprobe_munmap(vma, start, end);
1289 if (unlikely(vma->vm_flags & VM_PFNMAP))
1290 untrack_pfn(vma, 0, 0);
1293 if (unlikely(is_vm_hugetlb_page(vma))) {
1295 * It is undesirable to test vma->vm_file as it
1296 * should be non-null for valid hugetlb area.
1297 * However, vm_file will be NULL in the error
1298 * cleanup path of mmap_region. When
1299 * hugetlbfs ->mmap method fails,
1300 * mmap_region() nullifies vma->vm_file
1301 * before calling this function to clean up.
1302 * Since no pte has actually been setup, it is
1303 * safe to do nothing in this case.
1306 i_mmap_lock_write(vma->vm_file->f_mapping);
1307 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1308 i_mmap_unlock_write(vma->vm_file->f_mapping);
1311 unmap_page_range(tlb, vma, start, end, details);
1316 * unmap_vmas - unmap a range of memory covered by a list of vma's
1317 * @tlb: address of the caller's struct mmu_gather
1318 * @vma: the starting vma
1319 * @start_addr: virtual address at which to start unmapping
1320 * @end_addr: virtual address at which to end unmapping
1322 * Unmap all pages in the vma list.
1324 * Only addresses between `start' and `end' will be unmapped.
1326 * The VMA list must be sorted in ascending virtual address order.
1328 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1329 * range after unmap_vmas() returns. So the only responsibility here is to
1330 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1331 * drops the lock and schedules.
1333 void unmap_vmas(struct mmu_gather *tlb,
1334 struct vm_area_struct *vma, unsigned long start_addr,
1335 unsigned long end_addr)
1337 struct mmu_notifier_range range;
1339 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1340 start_addr, end_addr);
1341 mmu_notifier_invalidate_range_start(&range);
1342 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1343 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1344 mmu_notifier_invalidate_range_end(&range);
1348 * zap_page_range - remove user pages in a given range
1349 * @vma: vm_area_struct holding the applicable pages
1350 * @start: starting address of pages to zap
1351 * @size: number of bytes to zap
1353 * Caller must protect the VMA list
1355 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1358 struct mmu_notifier_range range;
1359 struct mmu_gather tlb;
1362 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1363 start, start + size);
1364 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1365 update_hiwater_rss(vma->vm_mm);
1366 mmu_notifier_invalidate_range_start(&range);
1367 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1368 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1369 mmu_notifier_invalidate_range_end(&range);
1370 tlb_finish_mmu(&tlb, start, range.end);
1374 * zap_page_range_single - remove user pages in a given range
1375 * @vma: vm_area_struct holding the applicable pages
1376 * @address: starting address of pages to zap
1377 * @size: number of bytes to zap
1378 * @details: details of shared cache invalidation
1380 * The range must fit into one VMA.
1382 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1383 unsigned long size, struct zap_details *details)
1385 struct mmu_notifier_range range;
1386 struct mmu_gather tlb;
1389 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1390 address, address + size);
1391 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1392 update_hiwater_rss(vma->vm_mm);
1393 mmu_notifier_invalidate_range_start(&range);
1394 unmap_single_vma(&tlb, vma, address, range.end, details);
1395 mmu_notifier_invalidate_range_end(&range);
1396 tlb_finish_mmu(&tlb, address, range.end);
1400 * zap_vma_ptes - remove ptes mapping the vma
1401 * @vma: vm_area_struct holding ptes to be zapped
1402 * @address: starting address of pages to zap
1403 * @size: number of bytes to zap
1405 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1407 * The entire address range must be fully contained within the vma.
1410 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1413 if (address < vma->vm_start || address + size > vma->vm_end ||
1414 !(vma->vm_flags & VM_PFNMAP))
1417 zap_page_range_single(vma, address, size, NULL);
1419 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1421 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1429 pgd = pgd_offset(mm, addr);
1430 p4d = p4d_alloc(mm, pgd, addr);
1433 pud = pud_alloc(mm, p4d, addr);
1436 pmd = pmd_alloc(mm, pud, addr);
1440 VM_BUG_ON(pmd_trans_huge(*pmd));
1441 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1445 * This is the old fallback for page remapping.
1447 * For historical reasons, it only allows reserved pages. Only
1448 * old drivers should use this, and they needed to mark their
1449 * pages reserved for the old functions anyway.
1451 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1452 struct page *page, pgprot_t prot)
1454 struct mm_struct *mm = vma->vm_mm;
1460 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1463 flush_dcache_page(page);
1464 pte = get_locked_pte(mm, addr, &ptl);
1468 if (!pte_none(*pte))
1471 /* Ok, finally just insert the thing.. */
1473 inc_mm_counter_fast(mm, mm_counter_file(page));
1474 page_add_file_rmap(page, false);
1475 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1478 pte_unmap_unlock(pte, ptl);
1481 pte_unmap_unlock(pte, ptl);
1487 * vm_insert_page - insert single page into user vma
1488 * @vma: user vma to map to
1489 * @addr: target user address of this page
1490 * @page: source kernel page
1492 * This allows drivers to insert individual pages they've allocated
1495 * The page has to be a nice clean _individual_ kernel allocation.
1496 * If you allocate a compound page, you need to have marked it as
1497 * such (__GFP_COMP), or manually just split the page up yourself
1498 * (see split_page()).
1500 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1501 * took an arbitrary page protection parameter. This doesn't allow
1502 * that. Your vma protection will have to be set up correctly, which
1503 * means that if you want a shared writable mapping, you'd better
1504 * ask for a shared writable mapping!
1506 * The page does not need to be reserved.
1508 * Usually this function is called from f_op->mmap() handler
1509 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1510 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1511 * function from other places, for example from page-fault handler.
1513 * Return: %0 on success, negative error code otherwise.
1515 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1518 if (addr < vma->vm_start || addr >= vma->vm_end)
1520 if (!page_count(page))
1522 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1523 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1524 BUG_ON(vma->vm_flags & VM_PFNMAP);
1525 vma->vm_flags |= VM_MIXEDMAP;
1527 return insert_page(vma, addr, page, vma->vm_page_prot);
1529 EXPORT_SYMBOL(vm_insert_page);
1532 * __vm_map_pages - maps range of kernel pages into user vma
1533 * @vma: user vma to map to
1534 * @pages: pointer to array of source kernel pages
1535 * @num: number of pages in page array
1536 * @offset: user's requested vm_pgoff
1538 * This allows drivers to map range of kernel pages into a user vma.
1540 * Return: 0 on success and error code otherwise.
1542 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1543 unsigned long num, unsigned long offset)
1545 unsigned long count = vma_pages(vma);
1546 unsigned long uaddr = vma->vm_start;
1549 /* Fail if the user requested offset is beyond the end of the object */
1553 /* Fail if the user requested size exceeds available object size */
1554 if (count > num - offset)
1557 for (i = 0; i < count; i++) {
1558 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1568 * vm_map_pages - maps range of kernel pages starts with non zero offset
1569 * @vma: user vma to map to
1570 * @pages: pointer to array of source kernel pages
1571 * @num: number of pages in page array
1573 * Maps an object consisting of @num pages, catering for the user's
1574 * requested vm_pgoff
1576 * If we fail to insert any page into the vma, the function will return
1577 * immediately leaving any previously inserted pages present. Callers
1578 * from the mmap handler may immediately return the error as their caller
1579 * will destroy the vma, removing any successfully inserted pages. Other
1580 * callers should make their own arrangements for calling unmap_region().
1582 * Context: Process context. Called by mmap handlers.
1583 * Return: 0 on success and error code otherwise.
1585 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1588 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1590 EXPORT_SYMBOL(vm_map_pages);
1593 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1594 * @vma: user vma to map to
1595 * @pages: pointer to array of source kernel pages
1596 * @num: number of pages in page array
1598 * Similar to vm_map_pages(), except that it explicitly sets the offset
1599 * to 0. This function is intended for the drivers that did not consider
1602 * Context: Process context. Called by mmap handlers.
1603 * Return: 0 on success and error code otherwise.
1605 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1608 return __vm_map_pages(vma, pages, num, 0);
1610 EXPORT_SYMBOL(vm_map_pages_zero);
1612 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1613 pfn_t pfn, pgprot_t prot, bool mkwrite)
1615 struct mm_struct *mm = vma->vm_mm;
1619 pte = get_locked_pte(mm, addr, &ptl);
1621 return VM_FAULT_OOM;
1622 if (!pte_none(*pte)) {
1625 * For read faults on private mappings the PFN passed
1626 * in may not match the PFN we have mapped if the
1627 * mapped PFN is a writeable COW page. In the mkwrite
1628 * case we are creating a writable PTE for a shared
1629 * mapping and we expect the PFNs to match. If they
1630 * don't match, we are likely racing with block
1631 * allocation and mapping invalidation so just skip the
1634 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1635 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1638 entry = pte_mkyoung(*pte);
1639 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1640 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1641 update_mmu_cache(vma, addr, pte);
1646 /* Ok, finally just insert the thing.. */
1647 if (pfn_t_devmap(pfn))
1648 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1650 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1653 entry = pte_mkyoung(entry);
1654 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1657 set_pte_at(mm, addr, pte, entry);
1658 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1661 pte_unmap_unlock(pte, ptl);
1662 return VM_FAULT_NOPAGE;
1666 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1667 * @vma: user vma to map to
1668 * @addr: target user address of this page
1669 * @pfn: source kernel pfn
1670 * @pgprot: pgprot flags for the inserted page
1672 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1673 * to override pgprot on a per-page basis.
1675 * This only makes sense for IO mappings, and it makes no sense for
1676 * COW mappings. In general, using multiple vmas is preferable;
1677 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1680 * Context: Process context. May allocate using %GFP_KERNEL.
1681 * Return: vm_fault_t value.
1683 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1684 unsigned long pfn, pgprot_t pgprot)
1687 * Technically, architectures with pte_special can avoid all these
1688 * restrictions (same for remap_pfn_range). However we would like
1689 * consistency in testing and feature parity among all, so we should
1690 * try to keep these invariants in place for everybody.
1692 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1693 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1694 (VM_PFNMAP|VM_MIXEDMAP));
1695 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1696 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1698 if (addr < vma->vm_start || addr >= vma->vm_end)
1699 return VM_FAULT_SIGBUS;
1701 if (!pfn_modify_allowed(pfn, pgprot))
1702 return VM_FAULT_SIGBUS;
1704 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1706 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1709 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1712 * vmf_insert_pfn - insert single pfn into user vma
1713 * @vma: user vma to map to
1714 * @addr: target user address of this page
1715 * @pfn: source kernel pfn
1717 * Similar to vm_insert_page, this allows drivers to insert individual pages
1718 * they've allocated into a user vma. Same comments apply.
1720 * This function should only be called from a vm_ops->fault handler, and
1721 * in that case the handler should return the result of this function.
1723 * vma cannot be a COW mapping.
1725 * As this is called only for pages that do not currently exist, we
1726 * do not need to flush old virtual caches or the TLB.
1728 * Context: Process context. May allocate using %GFP_KERNEL.
1729 * Return: vm_fault_t value.
1731 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1734 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1736 EXPORT_SYMBOL(vmf_insert_pfn);
1738 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1740 /* these checks mirror the abort conditions in vm_normal_page */
1741 if (vma->vm_flags & VM_MIXEDMAP)
1743 if (pfn_t_devmap(pfn))
1745 if (pfn_t_special(pfn))
1747 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1752 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1753 unsigned long addr, pfn_t pfn, bool mkwrite)
1755 pgprot_t pgprot = vma->vm_page_prot;
1758 BUG_ON(!vm_mixed_ok(vma, pfn));
1760 if (addr < vma->vm_start || addr >= vma->vm_end)
1761 return VM_FAULT_SIGBUS;
1763 track_pfn_insert(vma, &pgprot, pfn);
1765 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1766 return VM_FAULT_SIGBUS;
1769 * If we don't have pte special, then we have to use the pfn_valid()
1770 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1771 * refcount the page if pfn_valid is true (hence insert_page rather
1772 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1773 * without pte special, it would there be refcounted as a normal page.
1775 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1776 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1780 * At this point we are committed to insert_page()
1781 * regardless of whether the caller specified flags that
1782 * result in pfn_t_has_page() == false.
1784 page = pfn_to_page(pfn_t_to_pfn(pfn));
1785 err = insert_page(vma, addr, page, pgprot);
1787 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1791 return VM_FAULT_OOM;
1792 if (err < 0 && err != -EBUSY)
1793 return VM_FAULT_SIGBUS;
1795 return VM_FAULT_NOPAGE;
1798 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1801 return __vm_insert_mixed(vma, addr, pfn, false);
1803 EXPORT_SYMBOL(vmf_insert_mixed);
1806 * If the insertion of PTE failed because someone else already added a
1807 * different entry in the mean time, we treat that as success as we assume
1808 * the same entry was actually inserted.
1810 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1811 unsigned long addr, pfn_t pfn)
1813 return __vm_insert_mixed(vma, addr, pfn, true);
1815 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1818 * maps a range of physical memory into the requested pages. the old
1819 * mappings are removed. any references to nonexistent pages results
1820 * in null mappings (currently treated as "copy-on-access")
1822 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1823 unsigned long addr, unsigned long end,
1824 unsigned long pfn, pgprot_t prot)
1830 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1833 arch_enter_lazy_mmu_mode();
1835 BUG_ON(!pte_none(*pte));
1836 if (!pfn_modify_allowed(pfn, prot)) {
1840 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1842 } while (pte++, addr += PAGE_SIZE, addr != end);
1843 arch_leave_lazy_mmu_mode();
1844 pte_unmap_unlock(pte - 1, ptl);
1848 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1849 unsigned long addr, unsigned long end,
1850 unsigned long pfn, pgprot_t prot)
1856 pfn -= addr >> PAGE_SHIFT;
1857 pmd = pmd_alloc(mm, pud, addr);
1860 VM_BUG_ON(pmd_trans_huge(*pmd));
1862 next = pmd_addr_end(addr, end);
1863 err = remap_pte_range(mm, pmd, addr, next,
1864 pfn + (addr >> PAGE_SHIFT), prot);
1867 } while (pmd++, addr = next, addr != end);
1871 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1872 unsigned long addr, unsigned long end,
1873 unsigned long pfn, pgprot_t prot)
1879 pfn -= addr >> PAGE_SHIFT;
1880 pud = pud_alloc(mm, p4d, addr);
1884 next = pud_addr_end(addr, end);
1885 err = remap_pmd_range(mm, pud, addr, next,
1886 pfn + (addr >> PAGE_SHIFT), prot);
1889 } while (pud++, addr = next, addr != end);
1893 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1894 unsigned long addr, unsigned long end,
1895 unsigned long pfn, pgprot_t prot)
1901 pfn -= addr >> PAGE_SHIFT;
1902 p4d = p4d_alloc(mm, pgd, addr);
1906 next = p4d_addr_end(addr, end);
1907 err = remap_pud_range(mm, p4d, addr, next,
1908 pfn + (addr >> PAGE_SHIFT), prot);
1911 } while (p4d++, addr = next, addr != end);
1916 * remap_pfn_range - remap kernel memory to userspace
1917 * @vma: user vma to map to
1918 * @addr: target user address to start at
1919 * @pfn: physical address of kernel memory
1920 * @size: size of map area
1921 * @prot: page protection flags for this mapping
1923 * Note: this is only safe if the mm semaphore is held when called.
1925 * Return: %0 on success, negative error code otherwise.
1927 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1928 unsigned long pfn, unsigned long size, pgprot_t prot)
1932 unsigned long end = addr + PAGE_ALIGN(size);
1933 struct mm_struct *mm = vma->vm_mm;
1934 unsigned long remap_pfn = pfn;
1938 * Physically remapped pages are special. Tell the
1939 * rest of the world about it:
1940 * VM_IO tells people not to look at these pages
1941 * (accesses can have side effects).
1942 * VM_PFNMAP tells the core MM that the base pages are just
1943 * raw PFN mappings, and do not have a "struct page" associated
1946 * Disable vma merging and expanding with mremap().
1948 * Omit vma from core dump, even when VM_IO turned off.
1950 * There's a horrible special case to handle copy-on-write
1951 * behaviour that some programs depend on. We mark the "original"
1952 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1953 * See vm_normal_page() for details.
1955 if (is_cow_mapping(vma->vm_flags)) {
1956 if (addr != vma->vm_start || end != vma->vm_end)
1958 vma->vm_pgoff = pfn;
1961 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1965 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1967 BUG_ON(addr >= end);
1968 pfn -= addr >> PAGE_SHIFT;
1969 pgd = pgd_offset(mm, addr);
1970 flush_cache_range(vma, addr, end);
1972 next = pgd_addr_end(addr, end);
1973 err = remap_p4d_range(mm, pgd, addr, next,
1974 pfn + (addr >> PAGE_SHIFT), prot);
1977 } while (pgd++, addr = next, addr != end);
1980 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1984 EXPORT_SYMBOL(remap_pfn_range);
1987 * vm_iomap_memory - remap memory to userspace
1988 * @vma: user vma to map to
1989 * @start: start of area
1990 * @len: size of area
1992 * This is a simplified io_remap_pfn_range() for common driver use. The
1993 * driver just needs to give us the physical memory range to be mapped,
1994 * we'll figure out the rest from the vma information.
1996 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1997 * whatever write-combining details or similar.
1999 * Return: %0 on success, negative error code otherwise.
2001 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2003 unsigned long vm_len, pfn, pages;
2005 /* Check that the physical memory area passed in looks valid */
2006 if (start + len < start)
2009 * You *really* shouldn't map things that aren't page-aligned,
2010 * but we've historically allowed it because IO memory might
2011 * just have smaller alignment.
2013 len += start & ~PAGE_MASK;
2014 pfn = start >> PAGE_SHIFT;
2015 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2016 if (pfn + pages < pfn)
2019 /* We start the mapping 'vm_pgoff' pages into the area */
2020 if (vma->vm_pgoff > pages)
2022 pfn += vma->vm_pgoff;
2023 pages -= vma->vm_pgoff;
2025 /* Can we fit all of the mapping? */
2026 vm_len = vma->vm_end - vma->vm_start;
2027 if (vm_len >> PAGE_SHIFT > pages)
2030 /* Ok, let it rip */
2031 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2033 EXPORT_SYMBOL(vm_iomap_memory);
2035 static int apply_to_pte_range(struct pfn_range_apply *closure, pmd_t *pmd,
2036 unsigned long addr, unsigned long end)
2041 spinlock_t *uninitialized_var(ptl);
2043 pte = (closure->mm == &init_mm) ?
2044 pte_alloc_kernel(pmd, addr) :
2045 pte_alloc_map_lock(closure->mm, pmd, addr, &ptl);
2049 BUG_ON(pmd_huge(*pmd));
2051 arch_enter_lazy_mmu_mode();
2053 token = pmd_pgtable(*pmd);
2056 err = closure->ptefn(pte++, token, addr, closure);
2059 } while (addr += PAGE_SIZE, addr != end);
2061 arch_leave_lazy_mmu_mode();
2063 if (closure->mm != &init_mm)
2064 pte_unmap_unlock(pte-1, ptl);
2068 static int apply_to_pmd_range(struct pfn_range_apply *closure, pud_t *pud,
2069 unsigned long addr, unsigned long end)
2075 BUG_ON(pud_huge(*pud));
2077 pmd = pmd_alloc(closure->mm, pud, addr);
2082 next = pmd_addr_end(addr, end);
2083 if (!closure->alloc && pmd_none_or_clear_bad(pmd))
2085 err = apply_to_pte_range(closure, pmd, addr, next);
2088 } while (pmd++, addr = next, addr != end);
2092 static int apply_to_pud_range(struct pfn_range_apply *closure, p4d_t *p4d,
2093 unsigned long addr, unsigned long end)
2099 pud = pud_alloc(closure->mm, p4d, addr);
2104 next = pud_addr_end(addr, end);
2105 if (!closure->alloc && pud_none_or_clear_bad(pud))
2107 err = apply_to_pmd_range(closure, pud, addr, next);
2110 } while (pud++, addr = next, addr != end);
2114 static int apply_to_p4d_range(struct pfn_range_apply *closure, pgd_t *pgd,
2115 unsigned long addr, unsigned long end)
2121 p4d = p4d_alloc(closure->mm, pgd, addr);
2126 next = p4d_addr_end(addr, end);
2127 if (!closure->alloc && p4d_none_or_clear_bad(p4d))
2129 err = apply_to_pud_range(closure, p4d, addr, next);
2132 } while (p4d++, addr = next, addr != end);
2137 * apply_to_pfn_range - Scan a region of virtual memory, calling a provided
2138 * function on each leaf page table entry
2139 * @closure: Details about how to scan and what function to apply
2140 * @addr: Start virtual address
2141 * @size: Size of the region
2143 * If @closure->alloc is set to 1, the function will fill in the page table
2144 * as necessary. Otherwise it will skip non-present parts.
2145 * Note: The caller must ensure that the range does not contain huge pages.
2146 * The caller must also assure that the proper mmu_notifier functions are
2147 * called before and after the call to apply_to_pfn_range.
2149 * WARNING: Do not use this function unless you know exactly what you are
2150 * doing. It is lacking support for huge pages and transparent huge pages.
2152 * Return: Zero on success. If the provided function returns a non-zero status,
2153 * the page table walk will terminate and that status will be returned.
2154 * If @closure->alloc is set to 1, then this function may also return memory
2155 * allocation errors arising from allocating page table memory.
2157 int apply_to_pfn_range(struct pfn_range_apply *closure,
2158 unsigned long addr, unsigned long size)
2162 unsigned long end = addr + size;
2165 if (WARN_ON(addr >= end))
2168 pgd = pgd_offset(closure->mm, addr);
2170 next = pgd_addr_end(addr, end);
2171 if (!closure->alloc && pgd_none_or_clear_bad(pgd))
2173 err = apply_to_p4d_range(closure, pgd, addr, next);
2176 } while (pgd++, addr = next, addr != end);
2182 * struct page_range_apply - Closure structure for apply_to_page_range()
2183 * @pter: The base closure structure we derive from
2184 * @fn: The leaf pte function to call
2185 * @data: The leaf pte function closure
2187 struct page_range_apply {
2188 struct pfn_range_apply pter;
2194 * Callback wrapper to enable use of apply_to_pfn_range for
2195 * the apply_to_page_range interface
2197 static int apply_to_page_range_wrapper(pte_t *pte, pgtable_t token,
2199 struct pfn_range_apply *pter)
2201 struct page_range_apply *pra =
2202 container_of(pter, typeof(*pra), pter);
2204 return pra->fn(pte, token, addr, pra->data);
2208 * Scan a region of virtual memory, filling in page tables as necessary
2209 * and calling a provided function on each leaf page table.
2211 * WARNING: Do not use this function unless you know exactly what you are
2212 * doing. It is lacking support for huge pages and transparent huge pages.
2214 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2215 unsigned long size, pte_fn_t fn, void *data)
2217 struct page_range_apply pra = {
2220 .ptefn = apply_to_page_range_wrapper },
2225 return apply_to_pfn_range(&pra.pter, addr, size);
2227 EXPORT_SYMBOL_GPL(apply_to_page_range);
2230 * handle_pte_fault chooses page fault handler according to an entry which was
2231 * read non-atomically. Before making any commitment, on those architectures
2232 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2233 * parts, do_swap_page must check under lock before unmapping the pte and
2234 * proceeding (but do_wp_page is only called after already making such a check;
2235 * and do_anonymous_page can safely check later on).
2237 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2238 pte_t *page_table, pte_t orig_pte)
2241 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2242 if (sizeof(pte_t) > sizeof(unsigned long)) {
2243 spinlock_t *ptl = pte_lockptr(mm, pmd);
2245 same = pte_same(*page_table, orig_pte);
2249 pte_unmap(page_table);
2253 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2255 debug_dma_assert_idle(src);
2258 * If the source page was a PFN mapping, we don't have
2259 * a "struct page" for it. We do a best-effort copy by
2260 * just copying from the original user address. If that
2261 * fails, we just zero-fill it. Live with it.
2263 if (unlikely(!src)) {
2264 void *kaddr = kmap_atomic(dst);
2265 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2268 * This really shouldn't fail, because the page is there
2269 * in the page tables. But it might just be unreadable,
2270 * in which case we just give up and fill the result with
2273 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2275 kunmap_atomic(kaddr);
2276 flush_dcache_page(dst);
2278 copy_user_highpage(dst, src, va, vma);
2281 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2283 struct file *vm_file = vma->vm_file;
2286 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2289 * Special mappings (e.g. VDSO) do not have any file so fake
2290 * a default GFP_KERNEL for them.
2296 * Notify the address space that the page is about to become writable so that
2297 * it can prohibit this or wait for the page to get into an appropriate state.
2299 * We do this without the lock held, so that it can sleep if it needs to.
2301 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2304 struct page *page = vmf->page;
2305 unsigned int old_flags = vmf->flags;
2307 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2309 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2310 /* Restore original flags so that caller is not surprised */
2311 vmf->flags = old_flags;
2312 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2314 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2316 if (!page->mapping) {
2318 return 0; /* retry */
2320 ret |= VM_FAULT_LOCKED;
2322 VM_BUG_ON_PAGE(!PageLocked(page), page);
2327 * Handle dirtying of a page in shared file mapping on a write fault.
2329 * The function expects the page to be locked and unlocks it.
2331 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2334 struct address_space *mapping;
2336 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2338 dirtied = set_page_dirty(page);
2339 VM_BUG_ON_PAGE(PageAnon(page), page);
2341 * Take a local copy of the address_space - page.mapping may be zeroed
2342 * by truncate after unlock_page(). The address_space itself remains
2343 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2344 * release semantics to prevent the compiler from undoing this copying.
2346 mapping = page_rmapping(page);
2349 if ((dirtied || page_mkwrite) && mapping) {
2351 * Some device drivers do not set page.mapping
2352 * but still dirty their pages
2354 balance_dirty_pages_ratelimited(mapping);
2358 file_update_time(vma->vm_file);
2362 * Handle write page faults for pages that can be reused in the current vma
2364 * This can happen either due to the mapping being with the VM_SHARED flag,
2365 * or due to us being the last reference standing to the page. In either
2366 * case, all we need to do here is to mark the page as writable and update
2367 * any related book-keeping.
2369 static inline void wp_page_reuse(struct vm_fault *vmf)
2370 __releases(vmf->ptl)
2372 struct vm_area_struct *vma = vmf->vma;
2373 struct page *page = vmf->page;
2376 * Clear the pages cpupid information as the existing
2377 * information potentially belongs to a now completely
2378 * unrelated process.
2381 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2383 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2384 entry = pte_mkyoung(vmf->orig_pte);
2385 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2386 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2387 update_mmu_cache(vma, vmf->address, vmf->pte);
2388 pte_unmap_unlock(vmf->pte, vmf->ptl);
2392 * Handle the case of a page which we actually need to copy to a new page.
2394 * Called with mmap_sem locked and the old page referenced, but
2395 * without the ptl held.
2397 * High level logic flow:
2399 * - Allocate a page, copy the content of the old page to the new one.
2400 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2401 * - Take the PTL. If the pte changed, bail out and release the allocated page
2402 * - If the pte is still the way we remember it, update the page table and all
2403 * relevant references. This includes dropping the reference the page-table
2404 * held to the old page, as well as updating the rmap.
2405 * - In any case, unlock the PTL and drop the reference we took to the old page.
2407 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2409 struct vm_area_struct *vma = vmf->vma;
2410 struct mm_struct *mm = vma->vm_mm;
2411 struct page *old_page = vmf->page;
2412 struct page *new_page = NULL;
2414 int page_copied = 0;
2415 struct mem_cgroup *memcg;
2416 struct mmu_notifier_range range;
2418 if (unlikely(anon_vma_prepare(vma)))
2421 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2422 new_page = alloc_zeroed_user_highpage_movable(vma,
2427 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2431 cow_user_page(new_page, old_page, vmf->address, vma);
2434 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2437 __SetPageUptodate(new_page);
2439 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2440 vmf->address & PAGE_MASK,
2441 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2442 mmu_notifier_invalidate_range_start(&range);
2445 * Re-check the pte - we dropped the lock
2447 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2448 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2450 if (!PageAnon(old_page)) {
2451 dec_mm_counter_fast(mm,
2452 mm_counter_file(old_page));
2453 inc_mm_counter_fast(mm, MM_ANONPAGES);
2456 inc_mm_counter_fast(mm, MM_ANONPAGES);
2458 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2459 entry = mk_pte(new_page, vma->vm_page_prot);
2460 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2462 * Clear the pte entry and flush it first, before updating the
2463 * pte with the new entry. This will avoid a race condition
2464 * seen in the presence of one thread doing SMC and another
2467 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2468 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2469 mem_cgroup_commit_charge(new_page, memcg, false, false);
2470 lru_cache_add_active_or_unevictable(new_page, vma);
2472 * We call the notify macro here because, when using secondary
2473 * mmu page tables (such as kvm shadow page tables), we want the
2474 * new page to be mapped directly into the secondary page table.
2476 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2477 update_mmu_cache(vma, vmf->address, vmf->pte);
2480 * Only after switching the pte to the new page may
2481 * we remove the mapcount here. Otherwise another
2482 * process may come and find the rmap count decremented
2483 * before the pte is switched to the new page, and
2484 * "reuse" the old page writing into it while our pte
2485 * here still points into it and can be read by other
2488 * The critical issue is to order this
2489 * page_remove_rmap with the ptp_clear_flush above.
2490 * Those stores are ordered by (if nothing else,)
2491 * the barrier present in the atomic_add_negative
2492 * in page_remove_rmap.
2494 * Then the TLB flush in ptep_clear_flush ensures that
2495 * no process can access the old page before the
2496 * decremented mapcount is visible. And the old page
2497 * cannot be reused until after the decremented
2498 * mapcount is visible. So transitively, TLBs to
2499 * old page will be flushed before it can be reused.
2501 page_remove_rmap(old_page, false);
2504 /* Free the old page.. */
2505 new_page = old_page;
2508 mem_cgroup_cancel_charge(new_page, memcg, false);
2514 pte_unmap_unlock(vmf->pte, vmf->ptl);
2516 * No need to double call mmu_notifier->invalidate_range() callback as
2517 * the above ptep_clear_flush_notify() did already call it.
2519 mmu_notifier_invalidate_range_only_end(&range);
2522 * Don't let another task, with possibly unlocked vma,
2523 * keep the mlocked page.
2525 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2526 lock_page(old_page); /* LRU manipulation */
2527 if (PageMlocked(old_page))
2528 munlock_vma_page(old_page);
2529 unlock_page(old_page);
2533 return page_copied ? VM_FAULT_WRITE : 0;
2539 return VM_FAULT_OOM;
2543 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2544 * writeable once the page is prepared
2546 * @vmf: structure describing the fault
2548 * This function handles all that is needed to finish a write page fault in a
2549 * shared mapping due to PTE being read-only once the mapped page is prepared.
2550 * It handles locking of PTE and modifying it.
2552 * The function expects the page to be locked or other protection against
2553 * concurrent faults / writeback (such as DAX radix tree locks).
2555 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2556 * we acquired PTE lock.
2558 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2560 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2561 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2564 * We might have raced with another page fault while we released the
2565 * pte_offset_map_lock.
2567 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2568 pte_unmap_unlock(vmf->pte, vmf->ptl);
2569 return VM_FAULT_NOPAGE;
2576 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2579 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2581 struct vm_area_struct *vma = vmf->vma;
2583 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2586 pte_unmap_unlock(vmf->pte, vmf->ptl);
2587 vmf->flags |= FAULT_FLAG_MKWRITE;
2588 ret = vma->vm_ops->pfn_mkwrite(vmf);
2589 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))
2591 return finish_mkwrite_fault(vmf);
2594 return VM_FAULT_WRITE;
2597 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2598 __releases(vmf->ptl)
2600 struct vm_area_struct *vma = vmf->vma;
2602 get_page(vmf->page);
2604 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2607 pte_unmap_unlock(vmf->pte, vmf->ptl);
2608 tmp = do_page_mkwrite(vmf);
2609 if (unlikely(!tmp || (tmp &
2610 (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
2611 VM_FAULT_RETRY)))) {
2612 put_page(vmf->page);
2615 tmp = finish_mkwrite_fault(vmf);
2616 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2617 unlock_page(vmf->page);
2618 put_page(vmf->page);
2623 lock_page(vmf->page);
2625 fault_dirty_shared_page(vma, vmf->page);
2626 put_page(vmf->page);
2628 return VM_FAULT_WRITE;
2632 * This routine handles present pages, when users try to write
2633 * to a shared page. It is done by copying the page to a new address
2634 * and decrementing the shared-page counter for the old page.
2636 * Note that this routine assumes that the protection checks have been
2637 * done by the caller (the low-level page fault routine in most cases).
2638 * Thus we can safely just mark it writable once we've done any necessary
2641 * We also mark the page dirty at this point even though the page will
2642 * change only once the write actually happens. This avoids a few races,
2643 * and potentially makes it more efficient.
2645 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2646 * but allow concurrent faults), with pte both mapped and locked.
2647 * We return with mmap_sem still held, but pte unmapped and unlocked.
2649 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2650 __releases(vmf->ptl)
2652 struct vm_area_struct *vma = vmf->vma;
2654 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2657 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2660 * We should not cow pages in a shared writeable mapping.
2661 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2663 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2664 (VM_WRITE|VM_SHARED))
2665 return wp_pfn_shared(vmf);
2667 pte_unmap_unlock(vmf->pte, vmf->ptl);
2668 return wp_page_copy(vmf);
2672 * Take out anonymous pages first, anonymous shared vmas are
2673 * not dirty accountable.
2675 if (PageAnon(vmf->page)) {
2676 int total_map_swapcount;
2677 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2678 page_count(vmf->page) != 1))
2680 if (!trylock_page(vmf->page)) {
2681 get_page(vmf->page);
2682 pte_unmap_unlock(vmf->pte, vmf->ptl);
2683 lock_page(vmf->page);
2684 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2685 vmf->address, &vmf->ptl);
2686 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2687 unlock_page(vmf->page);
2688 pte_unmap_unlock(vmf->pte, vmf->ptl);
2689 put_page(vmf->page);
2692 put_page(vmf->page);
2694 if (PageKsm(vmf->page)) {
2695 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2697 unlock_page(vmf->page);
2701 return VM_FAULT_WRITE;
2703 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2704 if (total_map_swapcount == 1) {
2706 * The page is all ours. Move it to
2707 * our anon_vma so the rmap code will
2708 * not search our parent or siblings.
2709 * Protected against the rmap code by
2712 page_move_anon_rmap(vmf->page, vma);
2714 unlock_page(vmf->page);
2716 return VM_FAULT_WRITE;
2718 unlock_page(vmf->page);
2719 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2720 (VM_WRITE|VM_SHARED))) {
2721 return wp_page_shared(vmf);
2725 * Ok, we need to copy. Oh, well..
2727 get_page(vmf->page);
2729 pte_unmap_unlock(vmf->pte, vmf->ptl);
2730 return wp_page_copy(vmf);
2733 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2734 unsigned long start_addr, unsigned long end_addr,
2735 struct zap_details *details)
2737 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2740 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2741 struct zap_details *details)
2743 struct vm_area_struct *vma;
2744 pgoff_t vba, vea, zba, zea;
2746 vma_interval_tree_foreach(vma, root,
2747 details->first_index, details->last_index) {
2749 vba = vma->vm_pgoff;
2750 vea = vba + vma_pages(vma) - 1;
2751 zba = details->first_index;
2754 zea = details->last_index;
2758 unmap_mapping_range_vma(vma,
2759 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2760 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2766 * unmap_mapping_pages() - Unmap pages from processes.
2767 * @mapping: The address space containing pages to be unmapped.
2768 * @start: Index of first page to be unmapped.
2769 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2770 * @even_cows: Whether to unmap even private COWed pages.
2772 * Unmap the pages in this address space from any userspace process which
2773 * has them mmaped. Generally, you want to remove COWed pages as well when
2774 * a file is being truncated, but not when invalidating pages from the page
2777 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2778 pgoff_t nr, bool even_cows)
2780 struct zap_details details = { };
2782 details.check_mapping = even_cows ? NULL : mapping;
2783 details.first_index = start;
2784 details.last_index = start + nr - 1;
2785 if (details.last_index < details.first_index)
2786 details.last_index = ULONG_MAX;
2788 i_mmap_lock_write(mapping);
2789 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2790 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2791 i_mmap_unlock_write(mapping);
2795 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2796 * address_space corresponding to the specified byte range in the underlying
2799 * @mapping: the address space containing mmaps to be unmapped.
2800 * @holebegin: byte in first page to unmap, relative to the start of
2801 * the underlying file. This will be rounded down to a PAGE_SIZE
2802 * boundary. Note that this is different from truncate_pagecache(), which
2803 * must keep the partial page. In contrast, we must get rid of
2805 * @holelen: size of prospective hole in bytes. This will be rounded
2806 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2808 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2809 * but 0 when invalidating pagecache, don't throw away private data.
2811 void unmap_mapping_range(struct address_space *mapping,
2812 loff_t const holebegin, loff_t const holelen, int even_cows)
2814 pgoff_t hba = holebegin >> PAGE_SHIFT;
2815 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2817 /* Check for overflow. */
2818 if (sizeof(holelen) > sizeof(hlen)) {
2820 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2821 if (holeend & ~(long long)ULONG_MAX)
2822 hlen = ULONG_MAX - hba + 1;
2825 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2827 EXPORT_SYMBOL(unmap_mapping_range);
2830 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2831 * but allow concurrent faults), and pte mapped but not yet locked.
2832 * We return with pte unmapped and unlocked.
2834 * We return with the mmap_sem locked or unlocked in the same cases
2835 * as does filemap_fault().
2837 vm_fault_t do_swap_page(struct vm_fault *vmf)
2839 struct vm_area_struct *vma = vmf->vma;
2840 struct page *page = NULL, *swapcache;
2841 struct mem_cgroup *memcg;
2848 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2851 entry = pte_to_swp_entry(vmf->orig_pte);
2852 if (unlikely(non_swap_entry(entry))) {
2853 if (is_migration_entry(entry)) {
2854 migration_entry_wait(vma->vm_mm, vmf->pmd,
2856 } else if (is_device_private_entry(entry)) {
2858 * For un-addressable device memory we call the pgmap
2859 * fault handler callback. The callback must migrate
2860 * the page back to some CPU accessible page.
2862 ret = device_private_entry_fault(vma, vmf->address, entry,
2863 vmf->flags, vmf->pmd);
2864 } else if (is_hwpoison_entry(entry)) {
2865 ret = VM_FAULT_HWPOISON;
2867 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2868 ret = VM_FAULT_SIGBUS;
2874 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2875 page = lookup_swap_cache(entry, vma, vmf->address);
2879 struct swap_info_struct *si = swp_swap_info(entry);
2881 if (si->flags & SWP_SYNCHRONOUS_IO &&
2882 __swap_count(si, entry) == 1) {
2883 /* skip swapcache */
2884 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2887 __SetPageLocked(page);
2888 __SetPageSwapBacked(page);
2889 set_page_private(page, entry.val);
2890 lru_cache_add_anon(page);
2891 swap_readpage(page, true);
2894 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2901 * Back out if somebody else faulted in this pte
2902 * while we released the pte lock.
2904 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2905 vmf->address, &vmf->ptl);
2906 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2908 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2912 /* Had to read the page from swap area: Major fault */
2913 ret = VM_FAULT_MAJOR;
2914 count_vm_event(PGMAJFAULT);
2915 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2916 } else if (PageHWPoison(page)) {
2918 * hwpoisoned dirty swapcache pages are kept for killing
2919 * owner processes (which may be unknown at hwpoison time)
2921 ret = VM_FAULT_HWPOISON;
2922 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2926 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2928 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2930 ret |= VM_FAULT_RETRY;
2935 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2936 * release the swapcache from under us. The page pin, and pte_same
2937 * test below, are not enough to exclude that. Even if it is still
2938 * swapcache, we need to check that the page's swap has not changed.
2940 if (unlikely((!PageSwapCache(page) ||
2941 page_private(page) != entry.val)) && swapcache)
2944 page = ksm_might_need_to_copy(page, vma, vmf->address);
2945 if (unlikely(!page)) {
2951 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2958 * Back out if somebody else already faulted in this pte.
2960 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2962 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2965 if (unlikely(!PageUptodate(page))) {
2966 ret = VM_FAULT_SIGBUS;
2971 * The page isn't present yet, go ahead with the fault.
2973 * Be careful about the sequence of operations here.
2974 * To get its accounting right, reuse_swap_page() must be called
2975 * while the page is counted on swap but not yet in mapcount i.e.
2976 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2977 * must be called after the swap_free(), or it will never succeed.
2980 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2981 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2982 pte = mk_pte(page, vma->vm_page_prot);
2983 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2984 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2985 vmf->flags &= ~FAULT_FLAG_WRITE;
2986 ret |= VM_FAULT_WRITE;
2987 exclusive = RMAP_EXCLUSIVE;
2989 flush_icache_page(vma, page);
2990 if (pte_swp_soft_dirty(vmf->orig_pte))
2991 pte = pte_mksoft_dirty(pte);
2992 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2993 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2994 vmf->orig_pte = pte;
2996 /* ksm created a completely new copy */
2997 if (unlikely(page != swapcache && swapcache)) {
2998 page_add_new_anon_rmap(page, vma, vmf->address, false);
2999 mem_cgroup_commit_charge(page, memcg, false, false);
3000 lru_cache_add_active_or_unevictable(page, vma);
3002 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3003 mem_cgroup_commit_charge(page, memcg, true, false);
3004 activate_page(page);
3008 if (mem_cgroup_swap_full(page) ||
3009 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3010 try_to_free_swap(page);
3012 if (page != swapcache && swapcache) {
3014 * Hold the lock to avoid the swap entry to be reused
3015 * until we take the PT lock for the pte_same() check
3016 * (to avoid false positives from pte_same). For
3017 * further safety release the lock after the swap_free
3018 * so that the swap count won't change under a
3019 * parallel locked swapcache.
3021 unlock_page(swapcache);
3022 put_page(swapcache);
3025 if (vmf->flags & FAULT_FLAG_WRITE) {
3026 ret |= do_wp_page(vmf);
3027 if (ret & VM_FAULT_ERROR)
3028 ret &= VM_FAULT_ERROR;
3032 /* No need to invalidate - it was non-present before */
3033 update_mmu_cache(vma, vmf->address, vmf->pte);
3035 pte_unmap_unlock(vmf->pte, vmf->ptl);
3039 mem_cgroup_cancel_charge(page, memcg, false);
3040 pte_unmap_unlock(vmf->pte, vmf->ptl);
3045 if (page != swapcache && swapcache) {
3046 unlock_page(swapcache);
3047 put_page(swapcache);
3053 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3054 * but allow concurrent faults), and pte mapped but not yet locked.
3055 * We return with mmap_sem still held, but pte unmapped and unlocked.
3057 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3059 struct vm_area_struct *vma = vmf->vma;
3060 struct mem_cgroup *memcg;
3065 /* File mapping without ->vm_ops ? */
3066 if (vma->vm_flags & VM_SHARED)
3067 return VM_FAULT_SIGBUS;
3070 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3071 * pte_offset_map() on pmds where a huge pmd might be created
3072 * from a different thread.
3074 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3075 * parallel threads are excluded by other means.
3077 * Here we only have down_read(mmap_sem).
3079 if (pte_alloc(vma->vm_mm, vmf->pmd))
3080 return VM_FAULT_OOM;
3082 /* See the comment in pte_alloc_one_map() */
3083 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3086 /* Use the zero-page for reads */
3087 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3088 !mm_forbids_zeropage(vma->vm_mm)) {
3089 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3090 vma->vm_page_prot));
3091 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3092 vmf->address, &vmf->ptl);
3093 if (!pte_none(*vmf->pte))
3095 ret = check_stable_address_space(vma->vm_mm);
3098 /* Deliver the page fault to userland, check inside PT lock */
3099 if (userfaultfd_missing(vma)) {
3100 pte_unmap_unlock(vmf->pte, vmf->ptl);
3101 return handle_userfault(vmf, VM_UFFD_MISSING);
3106 /* Allocate our own private page. */
3107 if (unlikely(anon_vma_prepare(vma)))
3109 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3113 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3118 * The memory barrier inside __SetPageUptodate makes sure that
3119 * preceeding stores to the page contents become visible before
3120 * the set_pte_at() write.
3122 __SetPageUptodate(page);
3124 entry = mk_pte(page, vma->vm_page_prot);
3125 if (vma->vm_flags & VM_WRITE)
3126 entry = pte_mkwrite(pte_mkdirty(entry));
3128 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3130 if (!pte_none(*vmf->pte))
3133 ret = check_stable_address_space(vma->vm_mm);
3137 /* Deliver the page fault to userland, check inside PT lock */
3138 if (userfaultfd_missing(vma)) {
3139 pte_unmap_unlock(vmf->pte, vmf->ptl);
3140 mem_cgroup_cancel_charge(page, memcg, false);
3142 return handle_userfault(vmf, VM_UFFD_MISSING);
3145 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3146 page_add_new_anon_rmap(page, vma, vmf->address, false);
3147 mem_cgroup_commit_charge(page, memcg, false, false);
3148 lru_cache_add_active_or_unevictable(page, vma);
3150 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3152 /* No need to invalidate - it was non-present before */
3153 update_mmu_cache(vma, vmf->address, vmf->pte);
3155 pte_unmap_unlock(vmf->pte, vmf->ptl);
3158 mem_cgroup_cancel_charge(page, memcg, false);
3164 return VM_FAULT_OOM;
3168 * The mmap_sem must have been held on entry, and may have been
3169 * released depending on flags and vma->vm_ops->fault() return value.
3170 * See filemap_fault() and __lock_page_retry().
3172 static vm_fault_t __do_fault(struct vm_fault *vmf)
3174 struct vm_area_struct *vma = vmf->vma;
3178 * Preallocate pte before we take page_lock because this might lead to
3179 * deadlocks for memcg reclaim which waits for pages under writeback:
3181 * SetPageWriteback(A)
3187 * wait_on_page_writeback(A)
3188 * SetPageWriteback(B)
3190 * # flush A, B to clear the writeback
3192 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3193 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3194 if (!vmf->prealloc_pte)
3195 return VM_FAULT_OOM;
3196 smp_wmb(); /* See comment in __pte_alloc() */
3199 ret = vma->vm_ops->fault(vmf);
3200 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3201 VM_FAULT_DONE_COW)))
3204 if (unlikely(PageHWPoison(vmf->page))) {
3205 if (ret & VM_FAULT_LOCKED)
3206 unlock_page(vmf->page);
3207 put_page(vmf->page);
3209 return VM_FAULT_HWPOISON;
3212 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3213 lock_page(vmf->page);
3215 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3221 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3222 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3223 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3224 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3226 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3228 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3231 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3233 struct vm_area_struct *vma = vmf->vma;
3235 if (!pmd_none(*vmf->pmd))
3237 if (vmf->prealloc_pte) {
3238 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3239 if (unlikely(!pmd_none(*vmf->pmd))) {
3240 spin_unlock(vmf->ptl);
3244 mm_inc_nr_ptes(vma->vm_mm);
3245 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3246 spin_unlock(vmf->ptl);
3247 vmf->prealloc_pte = NULL;
3248 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3249 return VM_FAULT_OOM;
3253 * If a huge pmd materialized under us just retry later. Use
3254 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3255 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3256 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3257 * running immediately after a huge pmd fault in a different thread of
3258 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3259 * All we have to ensure is that it is a regular pmd that we can walk
3260 * with pte_offset_map() and we can do that through an atomic read in
3261 * C, which is what pmd_trans_unstable() provides.
3263 if (pmd_devmap_trans_unstable(vmf->pmd))
3264 return VM_FAULT_NOPAGE;
3267 * At this point we know that our vmf->pmd points to a page of ptes
3268 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3269 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3270 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3271 * be valid and we will re-check to make sure the vmf->pte isn't
3272 * pte_none() under vmf->ptl protection when we return to
3275 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3280 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3282 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3283 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3284 unsigned long haddr)
3286 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3287 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3289 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3294 static void deposit_prealloc_pte(struct vm_fault *vmf)
3296 struct vm_area_struct *vma = vmf->vma;
3298 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3300 * We are going to consume the prealloc table,
3301 * count that as nr_ptes.
3303 mm_inc_nr_ptes(vma->vm_mm);
3304 vmf->prealloc_pte = NULL;
3307 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3309 struct vm_area_struct *vma = vmf->vma;
3310 bool write = vmf->flags & FAULT_FLAG_WRITE;
3311 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3316 if (!transhuge_vma_suitable(vma, haddr))
3317 return VM_FAULT_FALLBACK;
3319 ret = VM_FAULT_FALLBACK;
3320 page = compound_head(page);
3323 * Archs like ppc64 need additonal space to store information
3324 * related to pte entry. Use the preallocated table for that.
3326 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3327 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3328 if (!vmf->prealloc_pte)
3329 return VM_FAULT_OOM;
3330 smp_wmb(); /* See comment in __pte_alloc() */
3333 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3334 if (unlikely(!pmd_none(*vmf->pmd)))
3337 for (i = 0; i < HPAGE_PMD_NR; i++)
3338 flush_icache_page(vma, page + i);
3340 entry = mk_huge_pmd(page, vma->vm_page_prot);
3342 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3344 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3345 page_add_file_rmap(page, true);
3347 * deposit and withdraw with pmd lock held
3349 if (arch_needs_pgtable_deposit())
3350 deposit_prealloc_pte(vmf);
3352 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3354 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3356 /* fault is handled */
3358 count_vm_event(THP_FILE_MAPPED);
3360 spin_unlock(vmf->ptl);
3364 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3372 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3373 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3375 * @vmf: fault environment
3376 * @memcg: memcg to charge page (only for private mappings)
3377 * @page: page to map
3379 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3382 * Target users are page handler itself and implementations of
3383 * vm_ops->map_pages.
3385 * Return: %0 on success, %VM_FAULT_ code in case of error.
3387 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3390 struct vm_area_struct *vma = vmf->vma;
3391 bool write = vmf->flags & FAULT_FLAG_WRITE;
3395 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3396 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3398 VM_BUG_ON_PAGE(memcg, page);
3400 ret = do_set_pmd(vmf, page);
3401 if (ret != VM_FAULT_FALLBACK)
3406 ret = pte_alloc_one_map(vmf);
3411 /* Re-check under ptl */
3412 if (unlikely(!pte_none(*vmf->pte)))
3413 return VM_FAULT_NOPAGE;
3415 flush_icache_page(vma, page);
3416 entry = mk_pte(page, vma->vm_page_prot);
3418 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3419 /* copy-on-write page */
3420 if (write && !(vma->vm_flags & VM_SHARED)) {
3421 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3422 page_add_new_anon_rmap(page, vma, vmf->address, false);
3423 mem_cgroup_commit_charge(page, memcg, false, false);
3424 lru_cache_add_active_or_unevictable(page, vma);
3426 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3427 page_add_file_rmap(page, false);
3429 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3431 /* no need to invalidate: a not-present page won't be cached */
3432 update_mmu_cache(vma, vmf->address, vmf->pte);
3439 * finish_fault - finish page fault once we have prepared the page to fault
3441 * @vmf: structure describing the fault
3443 * This function handles all that is needed to finish a page fault once the
3444 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3445 * given page, adds reverse page mapping, handles memcg charges and LRU
3448 * The function expects the page to be locked and on success it consumes a
3449 * reference of a page being mapped (for the PTE which maps it).
3451 * Return: %0 on success, %VM_FAULT_ code in case of error.
3453 vm_fault_t finish_fault(struct vm_fault *vmf)
3458 /* Did we COW the page? */
3459 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3460 !(vmf->vma->vm_flags & VM_SHARED))
3461 page = vmf->cow_page;
3466 * check even for read faults because we might have lost our CoWed
3469 if (!(vmf->vma->vm_flags & VM_SHARED))
3470 ret = check_stable_address_space(vmf->vma->vm_mm);
3472 ret = alloc_set_pte(vmf, vmf->memcg, page);
3474 pte_unmap_unlock(vmf->pte, vmf->ptl);
3478 static unsigned long fault_around_bytes __read_mostly =
3479 rounddown_pow_of_two(65536);
3481 #ifdef CONFIG_DEBUG_FS
3482 static int fault_around_bytes_get(void *data, u64 *val)
3484 *val = fault_around_bytes;
3489 * fault_around_bytes must be rounded down to the nearest page order as it's
3490 * what do_fault_around() expects to see.
3492 static int fault_around_bytes_set(void *data, u64 val)
3494 if (val / PAGE_SIZE > PTRS_PER_PTE)
3496 if (val > PAGE_SIZE)
3497 fault_around_bytes = rounddown_pow_of_two(val);
3499 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3502 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3503 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3505 static int __init fault_around_debugfs(void)
3507 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3508 &fault_around_bytes_fops);
3511 late_initcall(fault_around_debugfs);
3515 * do_fault_around() tries to map few pages around the fault address. The hope
3516 * is that the pages will be needed soon and this will lower the number of
3519 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3520 * not ready to be mapped: not up-to-date, locked, etc.
3522 * This function is called with the page table lock taken. In the split ptlock
3523 * case the page table lock only protects only those entries which belong to
3524 * the page table corresponding to the fault address.
3526 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3529 * fault_around_bytes defines how many bytes we'll try to map.
3530 * do_fault_around() expects it to be set to a power of two less than or equal
3533 * The virtual address of the area that we map is naturally aligned to
3534 * fault_around_bytes rounded down to the machine page size
3535 * (and therefore to page order). This way it's easier to guarantee
3536 * that we don't cross page table boundaries.
3538 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3540 unsigned long address = vmf->address, nr_pages, mask;
3541 pgoff_t start_pgoff = vmf->pgoff;
3546 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3547 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3549 vmf->address = max(address & mask, vmf->vma->vm_start);
3550 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3554 * end_pgoff is either the end of the page table, the end of
3555 * the vma or nr_pages from start_pgoff, depending what is nearest.
3557 end_pgoff = start_pgoff -
3558 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3560 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3561 start_pgoff + nr_pages - 1);
3563 if (pmd_none(*vmf->pmd)) {
3564 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3565 if (!vmf->prealloc_pte)
3567 smp_wmb(); /* See comment in __pte_alloc() */
3570 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3572 /* Huge page is mapped? Page fault is solved */
3573 if (pmd_trans_huge(*vmf->pmd)) {
3574 ret = VM_FAULT_NOPAGE;
3578 /* ->map_pages() haven't done anything useful. Cold page cache? */
3582 /* check if the page fault is solved */
3583 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3584 if (!pte_none(*vmf->pte))
3585 ret = VM_FAULT_NOPAGE;
3586 pte_unmap_unlock(vmf->pte, vmf->ptl);
3588 vmf->address = address;
3593 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3595 struct vm_area_struct *vma = vmf->vma;
3599 * Let's call ->map_pages() first and use ->fault() as fallback
3600 * if page by the offset is not ready to be mapped (cold cache or
3603 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3604 ret = do_fault_around(vmf);
3609 ret = __do_fault(vmf);
3610 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3613 ret |= finish_fault(vmf);
3614 unlock_page(vmf->page);
3615 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3616 put_page(vmf->page);
3620 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3622 struct vm_area_struct *vma = vmf->vma;
3625 if (unlikely(anon_vma_prepare(vma)))
3626 return VM_FAULT_OOM;
3628 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3630 return VM_FAULT_OOM;
3632 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3633 &vmf->memcg, false)) {
3634 put_page(vmf->cow_page);
3635 return VM_FAULT_OOM;
3638 ret = __do_fault(vmf);
3639 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3641 if (ret & VM_FAULT_DONE_COW)
3644 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3645 __SetPageUptodate(vmf->cow_page);
3647 ret |= finish_fault(vmf);
3648 unlock_page(vmf->page);
3649 put_page(vmf->page);
3650 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3654 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3655 put_page(vmf->cow_page);
3659 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3661 struct vm_area_struct *vma = vmf->vma;
3662 vm_fault_t ret, tmp;
3664 ret = __do_fault(vmf);
3665 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3669 * Check if the backing address space wants to know that the page is
3670 * about to become writable
3672 if (vma->vm_ops->page_mkwrite) {
3673 unlock_page(vmf->page);
3674 tmp = do_page_mkwrite(vmf);
3675 if (unlikely(!tmp ||
3676 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3677 VM_FAULT_RETRY)))) {
3678 put_page(vmf->page);
3683 ret |= finish_fault(vmf);
3684 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3686 unlock_page(vmf->page);
3687 put_page(vmf->page);
3691 fault_dirty_shared_page(vma, vmf->page);
3696 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3697 * but allow concurrent faults).
3698 * The mmap_sem may have been released depending on flags and our
3699 * return value. See filemap_fault() and __lock_page_or_retry().
3700 * If mmap_sem is released, vma may become invalid (for example
3701 * by other thread calling munmap()).
3703 static vm_fault_t do_fault(struct vm_fault *vmf)
3705 struct vm_area_struct *vma = vmf->vma;
3706 struct mm_struct *vm_mm = vma->vm_mm;
3710 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3712 if (!vma->vm_ops->fault) {
3714 * If we find a migration pmd entry or a none pmd entry, which
3715 * should never happen, return SIGBUS
3717 if (unlikely(!pmd_present(*vmf->pmd)))
3718 ret = VM_FAULT_SIGBUS;
3720 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3725 * Make sure this is not a temporary clearing of pte
3726 * by holding ptl and checking again. A R/M/W update
3727 * of pte involves: take ptl, clearing the pte so that
3728 * we don't have concurrent modification by hardware
3729 * followed by an update.
3731 if (unlikely(pte_none(*vmf->pte)))
3732 ret = VM_FAULT_SIGBUS;
3734 ret = VM_FAULT_NOPAGE;
3736 pte_unmap_unlock(vmf->pte, vmf->ptl);
3738 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3739 ret = do_read_fault(vmf);
3740 else if (!(vma->vm_flags & VM_SHARED))
3741 ret = do_cow_fault(vmf);
3743 ret = do_shared_fault(vmf);
3745 /* preallocated pagetable is unused: free it */
3746 if (vmf->prealloc_pte) {
3747 pte_free(vm_mm, vmf->prealloc_pte);
3748 vmf->prealloc_pte = NULL;
3753 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3754 unsigned long addr, int page_nid,
3759 count_vm_numa_event(NUMA_HINT_FAULTS);
3760 if (page_nid == numa_node_id()) {
3761 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3762 *flags |= TNF_FAULT_LOCAL;
3765 return mpol_misplaced(page, vma, addr);
3768 static vm_fault_t do_numa_page(struct vm_fault *vmf)
3770 struct vm_area_struct *vma = vmf->vma;
3771 struct page *page = NULL;
3772 int page_nid = NUMA_NO_NODE;
3775 bool migrated = false;
3777 bool was_writable = pte_savedwrite(vmf->orig_pte);
3781 * The "pte" at this point cannot be used safely without
3782 * validation through pte_unmap_same(). It's of NUMA type but
3783 * the pfn may be screwed if the read is non atomic.
3785 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3786 spin_lock(vmf->ptl);
3787 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3788 pte_unmap_unlock(vmf->pte, vmf->ptl);
3793 * Make it present again, Depending on how arch implementes non
3794 * accessible ptes, some can allow access by kernel mode.
3796 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
3797 pte = pte_modify(old_pte, vma->vm_page_prot);
3798 pte = pte_mkyoung(pte);
3800 pte = pte_mkwrite(pte);
3801 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
3802 update_mmu_cache(vma, vmf->address, vmf->pte);
3804 page = vm_normal_page(vma, vmf->address, pte);
3806 pte_unmap_unlock(vmf->pte, vmf->ptl);
3810 /* TODO: handle PTE-mapped THP */
3811 if (PageCompound(page)) {
3812 pte_unmap_unlock(vmf->pte, vmf->ptl);
3817 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3818 * much anyway since they can be in shared cache state. This misses
3819 * the case where a mapping is writable but the process never writes
3820 * to it but pte_write gets cleared during protection updates and
3821 * pte_dirty has unpredictable behaviour between PTE scan updates,
3822 * background writeback, dirty balancing and application behaviour.
3824 if (!pte_write(pte))
3825 flags |= TNF_NO_GROUP;
3828 * Flag if the page is shared between multiple address spaces. This
3829 * is later used when determining whether to group tasks together
3831 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3832 flags |= TNF_SHARED;
3834 last_cpupid = page_cpupid_last(page);
3835 page_nid = page_to_nid(page);
3836 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3838 pte_unmap_unlock(vmf->pte, vmf->ptl);
3839 if (target_nid == NUMA_NO_NODE) {
3844 /* Migrate to the requested node */
3845 migrated = migrate_misplaced_page(page, vma, target_nid);
3847 page_nid = target_nid;
3848 flags |= TNF_MIGRATED;
3850 flags |= TNF_MIGRATE_FAIL;
3853 if (page_nid != NUMA_NO_NODE)
3854 task_numa_fault(last_cpupid, page_nid, 1, flags);
3858 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3860 if (vma_is_anonymous(vmf->vma))
3861 return do_huge_pmd_anonymous_page(vmf);
3862 if (vmf->vma->vm_ops->huge_fault)
3863 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3864 return VM_FAULT_FALLBACK;
3867 /* `inline' is required to avoid gcc 4.1.2 build error */
3868 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3870 if (vma_is_anonymous(vmf->vma))
3871 return do_huge_pmd_wp_page(vmf, orig_pmd);
3872 if (vmf->vma->vm_ops->huge_fault)
3873 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3875 /* COW handled on pte level: split pmd */
3876 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3877 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3879 return VM_FAULT_FALLBACK;
3882 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3884 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3887 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3889 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3890 /* No support for anonymous transparent PUD pages yet */
3891 if (vma_is_anonymous(vmf->vma))
3892 return VM_FAULT_FALLBACK;
3893 if (vmf->vma->vm_ops->huge_fault)
3894 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3895 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3896 return VM_FAULT_FALLBACK;
3899 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3901 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3902 /* No support for anonymous transparent PUD pages yet */
3903 if (vma_is_anonymous(vmf->vma))
3904 return VM_FAULT_FALLBACK;
3905 if (vmf->vma->vm_ops->huge_fault)
3906 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3907 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3908 return VM_FAULT_FALLBACK;
3912 * These routines also need to handle stuff like marking pages dirty
3913 * and/or accessed for architectures that don't do it in hardware (most
3914 * RISC architectures). The early dirtying is also good on the i386.
3916 * There is also a hook called "update_mmu_cache()" that architectures
3917 * with external mmu caches can use to update those (ie the Sparc or
3918 * PowerPC hashed page tables that act as extended TLBs).
3920 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3921 * concurrent faults).
3923 * The mmap_sem may have been released depending on flags and our return value.
3924 * See filemap_fault() and __lock_page_or_retry().
3926 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
3930 if (unlikely(pmd_none(*vmf->pmd))) {
3932 * Leave __pte_alloc() until later: because vm_ops->fault may
3933 * want to allocate huge page, and if we expose page table
3934 * for an instant, it will be difficult to retract from
3935 * concurrent faults and from rmap lookups.
3939 /* See comment in pte_alloc_one_map() */
3940 if (pmd_devmap_trans_unstable(vmf->pmd))
3943 * A regular pmd is established and it can't morph into a huge
3944 * pmd from under us anymore at this point because we hold the
3945 * mmap_sem read mode and khugepaged takes it in write mode.
3946 * So now it's safe to run pte_offset_map().
3948 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3949 vmf->orig_pte = *vmf->pte;
3952 * some architectures can have larger ptes than wordsize,
3953 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3954 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3955 * accesses. The code below just needs a consistent view
3956 * for the ifs and we later double check anyway with the
3957 * ptl lock held. So here a barrier will do.
3960 if (pte_none(vmf->orig_pte)) {
3961 pte_unmap(vmf->pte);
3967 if (vma_is_anonymous(vmf->vma))
3968 return do_anonymous_page(vmf);
3970 return do_fault(vmf);
3973 if (!pte_present(vmf->orig_pte))
3974 return do_swap_page(vmf);
3976 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3977 return do_numa_page(vmf);
3979 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3980 spin_lock(vmf->ptl);
3981 entry = vmf->orig_pte;
3982 if (unlikely(!pte_same(*vmf->pte, entry)))
3984 if (vmf->flags & FAULT_FLAG_WRITE) {
3985 if (!pte_write(entry))
3986 return do_wp_page(vmf);
3987 entry = pte_mkdirty(entry);
3989 entry = pte_mkyoung(entry);
3990 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3991 vmf->flags & FAULT_FLAG_WRITE)) {
3992 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3995 * This is needed only for protection faults but the arch code
3996 * is not yet telling us if this is a protection fault or not.
3997 * This still avoids useless tlb flushes for .text page faults
4000 if (vmf->flags & FAULT_FLAG_WRITE)
4001 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4004 pte_unmap_unlock(vmf->pte, vmf->ptl);
4009 * By the time we get here, we already hold the mm semaphore
4011 * The mmap_sem may have been released depending on flags and our
4012 * return value. See filemap_fault() and __lock_page_or_retry().
4014 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4015 unsigned long address, unsigned int flags)
4017 struct vm_fault vmf = {
4019 .address = address & PAGE_MASK,
4021 .pgoff = linear_page_index(vma, address),
4022 .gfp_mask = __get_fault_gfp_mask(vma),
4024 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4025 struct mm_struct *mm = vma->vm_mm;
4030 pgd = pgd_offset(mm, address);
4031 p4d = p4d_alloc(mm, pgd, address);
4033 return VM_FAULT_OOM;
4035 vmf.pud = pud_alloc(mm, p4d, address);
4037 return VM_FAULT_OOM;
4038 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4039 ret = create_huge_pud(&vmf);
4040 if (!(ret & VM_FAULT_FALLBACK))
4043 pud_t orig_pud = *vmf.pud;
4046 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4048 /* NUMA case for anonymous PUDs would go here */
4050 if (dirty && !pud_write(orig_pud)) {
4051 ret = wp_huge_pud(&vmf, orig_pud);
4052 if (!(ret & VM_FAULT_FALLBACK))
4055 huge_pud_set_accessed(&vmf, orig_pud);
4061 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4063 return VM_FAULT_OOM;
4064 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4065 ret = create_huge_pmd(&vmf);
4066 if (!(ret & VM_FAULT_FALLBACK))
4069 pmd_t orig_pmd = *vmf.pmd;
4072 if (unlikely(is_swap_pmd(orig_pmd))) {
4073 VM_BUG_ON(thp_migration_supported() &&
4074 !is_pmd_migration_entry(orig_pmd));
4075 if (is_pmd_migration_entry(orig_pmd))
4076 pmd_migration_entry_wait(mm, vmf.pmd);
4079 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4080 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4081 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4083 if (dirty && !pmd_write(orig_pmd)) {
4084 ret = wp_huge_pmd(&vmf, orig_pmd);
4085 if (!(ret & VM_FAULT_FALLBACK))
4088 huge_pmd_set_accessed(&vmf, orig_pmd);
4094 return handle_pte_fault(&vmf);
4098 * By the time we get here, we already hold the mm semaphore
4100 * The mmap_sem may have been released depending on flags and our
4101 * return value. See filemap_fault() and __lock_page_or_retry().
4103 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4108 __set_current_state(TASK_RUNNING);
4110 count_vm_event(PGFAULT);
4111 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4113 /* do counter updates before entering really critical section. */
4114 check_sync_rss_stat(current);
4116 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4117 flags & FAULT_FLAG_INSTRUCTION,
4118 flags & FAULT_FLAG_REMOTE))
4119 return VM_FAULT_SIGSEGV;
4122 * Enable the memcg OOM handling for faults triggered in user
4123 * space. Kernel faults are handled more gracefully.
4125 if (flags & FAULT_FLAG_USER)
4126 mem_cgroup_enter_user_fault();
4128 if (unlikely(is_vm_hugetlb_page(vma)))
4129 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4131 ret = __handle_mm_fault(vma, address, flags);
4133 if (flags & FAULT_FLAG_USER) {
4134 mem_cgroup_exit_user_fault();
4136 * The task may have entered a memcg OOM situation but
4137 * if the allocation error was handled gracefully (no
4138 * VM_FAULT_OOM), there is no need to kill anything.
4139 * Just clean up the OOM state peacefully.
4141 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4142 mem_cgroup_oom_synchronize(false);
4147 EXPORT_SYMBOL_GPL(handle_mm_fault);
4149 #ifndef __PAGETABLE_P4D_FOLDED
4151 * Allocate p4d page table.
4152 * We've already handled the fast-path in-line.
4154 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4156 p4d_t *new = p4d_alloc_one(mm, address);
4160 smp_wmb(); /* See comment in __pte_alloc */
4162 spin_lock(&mm->page_table_lock);
4163 if (pgd_present(*pgd)) /* Another has populated it */
4166 pgd_populate(mm, pgd, new);
4167 spin_unlock(&mm->page_table_lock);
4170 #endif /* __PAGETABLE_P4D_FOLDED */
4172 #ifndef __PAGETABLE_PUD_FOLDED
4174 * Allocate page upper directory.
4175 * We've already handled the fast-path in-line.
4177 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4179 pud_t *new = pud_alloc_one(mm, address);
4183 smp_wmb(); /* See comment in __pte_alloc */
4185 spin_lock(&mm->page_table_lock);
4186 #ifndef __ARCH_HAS_5LEVEL_HACK
4187 if (!p4d_present(*p4d)) {
4189 p4d_populate(mm, p4d, new);
4190 } else /* Another has populated it */
4193 if (!pgd_present(*p4d)) {
4195 pgd_populate(mm, p4d, new);
4196 } else /* Another has populated it */
4198 #endif /* __ARCH_HAS_5LEVEL_HACK */
4199 spin_unlock(&mm->page_table_lock);
4202 #endif /* __PAGETABLE_PUD_FOLDED */
4204 #ifndef __PAGETABLE_PMD_FOLDED
4206 * Allocate page middle directory.
4207 * We've already handled the fast-path in-line.
4209 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4212 pmd_t *new = pmd_alloc_one(mm, address);
4216 smp_wmb(); /* See comment in __pte_alloc */
4218 ptl = pud_lock(mm, pud);
4219 #ifndef __ARCH_HAS_4LEVEL_HACK
4220 if (!pud_present(*pud)) {
4222 pud_populate(mm, pud, new);
4223 } else /* Another has populated it */
4226 if (!pgd_present(*pud)) {
4228 pgd_populate(mm, pud, new);
4229 } else /* Another has populated it */
4231 #endif /* __ARCH_HAS_4LEVEL_HACK */
4235 #endif /* __PAGETABLE_PMD_FOLDED */
4237 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4238 struct mmu_notifier_range *range,
4239 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4247 pgd = pgd_offset(mm, address);
4248 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4251 p4d = p4d_offset(pgd, address);
4252 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4255 pud = pud_offset(p4d, address);
4256 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4259 pmd = pmd_offset(pud, address);
4260 VM_BUG_ON(pmd_trans_huge(*pmd));
4262 if (pmd_huge(*pmd)) {
4267 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4268 NULL, mm, address & PMD_MASK,
4269 (address & PMD_MASK) + PMD_SIZE);
4270 mmu_notifier_invalidate_range_start(range);
4272 *ptlp = pmd_lock(mm, pmd);
4273 if (pmd_huge(*pmd)) {
4279 mmu_notifier_invalidate_range_end(range);
4282 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4286 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4287 address & PAGE_MASK,
4288 (address & PAGE_MASK) + PAGE_SIZE);
4289 mmu_notifier_invalidate_range_start(range);
4291 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4292 if (!pte_present(*ptep))
4297 pte_unmap_unlock(ptep, *ptlp);
4299 mmu_notifier_invalidate_range_end(range);
4304 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4305 pte_t **ptepp, spinlock_t **ptlp)
4309 /* (void) is needed to make gcc happy */
4310 (void) __cond_lock(*ptlp,
4311 !(res = __follow_pte_pmd(mm, address, NULL,
4312 ptepp, NULL, ptlp)));
4316 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4317 struct mmu_notifier_range *range,
4318 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4322 /* (void) is needed to make gcc happy */
4323 (void) __cond_lock(*ptlp,
4324 !(res = __follow_pte_pmd(mm, address, range,
4325 ptepp, pmdpp, ptlp)));
4328 EXPORT_SYMBOL(follow_pte_pmd);
4331 * follow_pfn - look up PFN at a user virtual address
4332 * @vma: memory mapping
4333 * @address: user virtual address
4334 * @pfn: location to store found PFN
4336 * Only IO mappings and raw PFN mappings are allowed.
4338 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4340 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4347 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4350 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4353 *pfn = pte_pfn(*ptep);
4354 pte_unmap_unlock(ptep, ptl);
4357 EXPORT_SYMBOL(follow_pfn);
4359 #ifdef CONFIG_HAVE_IOREMAP_PROT
4360 int follow_phys(struct vm_area_struct *vma,
4361 unsigned long address, unsigned int flags,
4362 unsigned long *prot, resource_size_t *phys)
4368 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4371 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4375 if ((flags & FOLL_WRITE) && !pte_write(pte))
4378 *prot = pgprot_val(pte_pgprot(pte));
4379 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4383 pte_unmap_unlock(ptep, ptl);
4388 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4389 void *buf, int len, int write)
4391 resource_size_t phys_addr;
4392 unsigned long prot = 0;
4393 void __iomem *maddr;
4394 int offset = addr & (PAGE_SIZE-1);
4396 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4399 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4404 memcpy_toio(maddr + offset, buf, len);
4406 memcpy_fromio(buf, maddr + offset, len);
4411 EXPORT_SYMBOL_GPL(generic_access_phys);
4415 * Access another process' address space as given in mm. If non-NULL, use the
4416 * given task for page fault accounting.
4418 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4419 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4421 struct vm_area_struct *vma;
4422 void *old_buf = buf;
4423 int write = gup_flags & FOLL_WRITE;
4425 down_read(&mm->mmap_sem);
4426 /* ignore errors, just check how much was successfully transferred */
4428 int bytes, ret, offset;
4430 struct page *page = NULL;
4432 ret = get_user_pages_remote(tsk, mm, addr, 1,
4433 gup_flags, &page, &vma, NULL);
4435 #ifndef CONFIG_HAVE_IOREMAP_PROT
4439 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4440 * we can access using slightly different code.
4442 vma = find_vma(mm, addr);
4443 if (!vma || vma->vm_start > addr)
4445 if (vma->vm_ops && vma->vm_ops->access)
4446 ret = vma->vm_ops->access(vma, addr, buf,
4454 offset = addr & (PAGE_SIZE-1);
4455 if (bytes > PAGE_SIZE-offset)
4456 bytes = PAGE_SIZE-offset;
4460 copy_to_user_page(vma, page, addr,
4461 maddr + offset, buf, bytes);
4462 set_page_dirty_lock(page);
4464 copy_from_user_page(vma, page, addr,
4465 buf, maddr + offset, bytes);
4474 up_read(&mm->mmap_sem);
4476 return buf - old_buf;
4480 * access_remote_vm - access another process' address space
4481 * @mm: the mm_struct of the target address space
4482 * @addr: start address to access
4483 * @buf: source or destination buffer
4484 * @len: number of bytes to transfer
4485 * @gup_flags: flags modifying lookup behaviour
4487 * The caller must hold a reference on @mm.
4489 * Return: number of bytes copied from source to destination.
4491 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4492 void *buf, int len, unsigned int gup_flags)
4494 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4498 * Access another process' address space.
4499 * Source/target buffer must be kernel space,
4500 * Do not walk the page table directly, use get_user_pages
4502 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4503 void *buf, int len, unsigned int gup_flags)
4505 struct mm_struct *mm;
4508 mm = get_task_mm(tsk);
4512 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4518 EXPORT_SYMBOL_GPL(access_process_vm);
4521 * Print the name of a VMA.
4523 void print_vma_addr(char *prefix, unsigned long ip)
4525 struct mm_struct *mm = current->mm;
4526 struct vm_area_struct *vma;
4529 * we might be running from an atomic context so we cannot sleep
4531 if (!down_read_trylock(&mm->mmap_sem))
4534 vma = find_vma(mm, ip);
4535 if (vma && vma->vm_file) {
4536 struct file *f = vma->vm_file;
4537 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4541 p = file_path(f, buf, PAGE_SIZE);
4544 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4546 vma->vm_end - vma->vm_start);
4547 free_page((unsigned long)buf);
4550 up_read(&mm->mmap_sem);
4553 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4554 void __might_fault(const char *file, int line)
4557 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4558 * holding the mmap_sem, this is safe because kernel memory doesn't
4559 * get paged out, therefore we'll never actually fault, and the
4560 * below annotations will generate false positives.
4562 if (uaccess_kernel())
4564 if (pagefault_disabled())
4566 __might_sleep(file, line, 0);
4567 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4569 might_lock_read(¤t->mm->mmap_sem);
4572 EXPORT_SYMBOL(__might_fault);
4575 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4577 * Process all subpages of the specified huge page with the specified
4578 * operation. The target subpage will be processed last to keep its
4581 static inline void process_huge_page(
4582 unsigned long addr_hint, unsigned int pages_per_huge_page,
4583 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4587 unsigned long addr = addr_hint &
4588 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4590 /* Process target subpage last to keep its cache lines hot */
4592 n = (addr_hint - addr) / PAGE_SIZE;
4593 if (2 * n <= pages_per_huge_page) {
4594 /* If target subpage in first half of huge page */
4597 /* Process subpages at the end of huge page */
4598 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4600 process_subpage(addr + i * PAGE_SIZE, i, arg);
4603 /* If target subpage in second half of huge page */
4604 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4605 l = pages_per_huge_page - n;
4606 /* Process subpages at the begin of huge page */
4607 for (i = 0; i < base; i++) {
4609 process_subpage(addr + i * PAGE_SIZE, i, arg);
4613 * Process remaining subpages in left-right-left-right pattern
4614 * towards the target subpage
4616 for (i = 0; i < l; i++) {
4617 int left_idx = base + i;
4618 int right_idx = base + 2 * l - 1 - i;
4621 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4623 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4627 static void clear_gigantic_page(struct page *page,
4629 unsigned int pages_per_huge_page)
4632 struct page *p = page;
4635 for (i = 0; i < pages_per_huge_page;
4636 i++, p = mem_map_next(p, page, i)) {
4638 clear_user_highpage(p, addr + i * PAGE_SIZE);
4642 static void clear_subpage(unsigned long addr, int idx, void *arg)
4644 struct page *page = arg;
4646 clear_user_highpage(page + idx, addr);
4649 void clear_huge_page(struct page *page,
4650 unsigned long addr_hint, unsigned int pages_per_huge_page)
4652 unsigned long addr = addr_hint &
4653 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4655 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4656 clear_gigantic_page(page, addr, pages_per_huge_page);
4660 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4663 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4665 struct vm_area_struct *vma,
4666 unsigned int pages_per_huge_page)
4669 struct page *dst_base = dst;
4670 struct page *src_base = src;
4672 for (i = 0; i < pages_per_huge_page; ) {
4674 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4677 dst = mem_map_next(dst, dst_base, i);
4678 src = mem_map_next(src, src_base, i);
4682 struct copy_subpage_arg {
4685 struct vm_area_struct *vma;
4688 static void copy_subpage(unsigned long addr, int idx, void *arg)
4690 struct copy_subpage_arg *copy_arg = arg;
4692 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4693 addr, copy_arg->vma);
4696 void copy_user_huge_page(struct page *dst, struct page *src,
4697 unsigned long addr_hint, struct vm_area_struct *vma,
4698 unsigned int pages_per_huge_page)
4700 unsigned long addr = addr_hint &
4701 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4702 struct copy_subpage_arg arg = {
4708 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4709 copy_user_gigantic_page(dst, src, addr, vma,
4710 pages_per_huge_page);
4714 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4717 long copy_huge_page_from_user(struct page *dst_page,
4718 const void __user *usr_src,
4719 unsigned int pages_per_huge_page,
4720 bool allow_pagefault)
4722 void *src = (void *)usr_src;
4724 unsigned long i, rc = 0;
4725 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4727 for (i = 0; i < pages_per_huge_page; i++) {
4728 if (allow_pagefault)
4729 page_kaddr = kmap(dst_page + i);
4731 page_kaddr = kmap_atomic(dst_page + i);
4732 rc = copy_from_user(page_kaddr,
4733 (const void __user *)(src + i * PAGE_SIZE),
4735 if (allow_pagefault)
4736 kunmap(dst_page + i);
4738 kunmap_atomic(page_kaddr);
4740 ret_val -= (PAGE_SIZE - rc);
4748 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4750 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4752 static struct kmem_cache *page_ptl_cachep;
4754 void __init ptlock_cache_init(void)
4756 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4760 bool ptlock_alloc(struct page *page)
4764 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4771 void ptlock_free(struct page *page)
4773 kmem_cache_free(page_ptl_cachep, page->ptl);