4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
75 #include <asm/mmu_context.h>
76 #include <asm/pgalloc.h>
77 #include <linux/uaccess.h>
79 #include <asm/tlbflush.h>
80 #include <asm/pgtable.h>
84 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
85 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
88 #ifndef CONFIG_NEED_MULTIPLE_NODES
89 /* use the per-pgdat data instead for discontigmem - mbligh */
90 unsigned long max_mapnr;
91 EXPORT_SYMBOL(max_mapnr);
94 EXPORT_SYMBOL(mem_map);
98 * A number of key systems in x86 including ioremap() rely on the assumption
99 * that high_memory defines the upper bound on direct map memory, then end
100 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
101 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
105 EXPORT_SYMBOL(high_memory);
108 * Randomize the address space (stacks, mmaps, brk, etc.).
110 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111 * as ancient (libc5 based) binaries can segfault. )
113 int randomize_va_space __read_mostly =
114 #ifdef CONFIG_COMPAT_BRK
120 static int __init disable_randmaps(char *s)
122 randomize_va_space = 0;
125 __setup("norandmaps", disable_randmaps);
127 unsigned long zero_pfn __read_mostly;
128 EXPORT_SYMBOL(zero_pfn);
130 unsigned long highest_memmap_pfn __read_mostly;
133 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
135 static int __init init_zero_pfn(void)
137 zero_pfn = page_to_pfn(ZERO_PAGE(0));
140 core_initcall(init_zero_pfn);
143 #if defined(SPLIT_RSS_COUNTING)
145 void sync_mm_rss(struct mm_struct *mm)
149 for (i = 0; i < NR_MM_COUNTERS; i++) {
150 if (current->rss_stat.count[i]) {
151 add_mm_counter(mm, i, current->rss_stat.count[i]);
152 current->rss_stat.count[i] = 0;
155 current->rss_stat.events = 0;
158 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
160 struct task_struct *task = current;
162 if (likely(task->mm == mm))
163 task->rss_stat.count[member] += val;
165 add_mm_counter(mm, member, val);
167 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
170 /* sync counter once per 64 page faults */
171 #define TASK_RSS_EVENTS_THRESH (64)
172 static void check_sync_rss_stat(struct task_struct *task)
174 if (unlikely(task != current))
176 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
177 sync_mm_rss(task->mm);
179 #else /* SPLIT_RSS_COUNTING */
181 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
184 static void check_sync_rss_stat(struct task_struct *task)
188 #endif /* SPLIT_RSS_COUNTING */
191 * Note: this doesn't free the actual pages themselves. That
192 * has been handled earlier when unmapping all the memory regions.
194 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
197 pgtable_t token = pmd_pgtable(*pmd);
199 pte_free_tlb(tlb, token, addr);
200 mm_dec_nr_ptes(tlb->mm);
203 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
204 unsigned long addr, unsigned long end,
205 unsigned long floor, unsigned long ceiling)
212 pmd = pmd_offset(pud, addr);
214 next = pmd_addr_end(addr, end);
215 if (pmd_none_or_clear_bad(pmd))
217 free_pte_range(tlb, pmd, addr);
218 } while (pmd++, addr = next, addr != end);
228 if (end - 1 > ceiling - 1)
231 pmd = pmd_offset(pud, start);
233 pmd_free_tlb(tlb, pmd, start);
234 mm_dec_nr_pmds(tlb->mm);
237 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
238 unsigned long addr, unsigned long end,
239 unsigned long floor, unsigned long ceiling)
246 pud = pud_offset(p4d, addr);
248 next = pud_addr_end(addr, end);
249 if (pud_none_or_clear_bad(pud))
251 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
252 } while (pud++, addr = next, addr != end);
262 if (end - 1 > ceiling - 1)
265 pud = pud_offset(p4d, start);
267 pud_free_tlb(tlb, pud, start);
268 mm_dec_nr_puds(tlb->mm);
271 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
272 unsigned long addr, unsigned long end,
273 unsigned long floor, unsigned long ceiling)
280 p4d = p4d_offset(pgd, addr);
282 next = p4d_addr_end(addr, end);
283 if (p4d_none_or_clear_bad(p4d))
285 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
286 } while (p4d++, addr = next, addr != end);
292 ceiling &= PGDIR_MASK;
296 if (end - 1 > ceiling - 1)
299 p4d = p4d_offset(pgd, start);
301 p4d_free_tlb(tlb, p4d, start);
305 * This function frees user-level page tables of a process.
307 void free_pgd_range(struct mmu_gather *tlb,
308 unsigned long addr, unsigned long end,
309 unsigned long floor, unsigned long ceiling)
315 * The next few lines have given us lots of grief...
317 * Why are we testing PMD* at this top level? Because often
318 * there will be no work to do at all, and we'd prefer not to
319 * go all the way down to the bottom just to discover that.
321 * Why all these "- 1"s? Because 0 represents both the bottom
322 * of the address space and the top of it (using -1 for the
323 * top wouldn't help much: the masks would do the wrong thing).
324 * The rule is that addr 0 and floor 0 refer to the bottom of
325 * the address space, but end 0 and ceiling 0 refer to the top
326 * Comparisons need to use "end - 1" and "ceiling - 1" (though
327 * that end 0 case should be mythical).
329 * Wherever addr is brought up or ceiling brought down, we must
330 * be careful to reject "the opposite 0" before it confuses the
331 * subsequent tests. But what about where end is brought down
332 * by PMD_SIZE below? no, end can't go down to 0 there.
334 * Whereas we round start (addr) and ceiling down, by different
335 * masks at different levels, in order to test whether a table
336 * now has no other vmas using it, so can be freed, we don't
337 * bother to round floor or end up - the tests don't need that.
351 if (end - 1 > ceiling - 1)
356 * We add page table cache pages with PAGE_SIZE,
357 * (see pte_free_tlb()), flush the tlb if we need
359 tlb_change_page_size(tlb, PAGE_SIZE);
360 pgd = pgd_offset(tlb->mm, addr);
362 next = pgd_addr_end(addr, end);
363 if (pgd_none_or_clear_bad(pgd))
365 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
366 } while (pgd++, addr = next, addr != end);
369 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
370 unsigned long floor, unsigned long ceiling)
373 struct vm_area_struct *next = vma->vm_next;
374 unsigned long addr = vma->vm_start;
377 * Hide vma from rmap and truncate_pagecache before freeing
380 unlink_anon_vmas(vma);
381 unlink_file_vma(vma);
383 if (is_vm_hugetlb_page(vma)) {
384 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
385 floor, next ? next->vm_start : ceiling);
388 * Optimization: gather nearby vmas into one call down
390 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
391 && !is_vm_hugetlb_page(next)) {
394 unlink_anon_vmas(vma);
395 unlink_file_vma(vma);
397 free_pgd_range(tlb, addr, vma->vm_end,
398 floor, next ? next->vm_start : ceiling);
404 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
407 pgtable_t new = pte_alloc_one(mm);
412 * Ensure all pte setup (eg. pte page lock and page clearing) are
413 * visible before the pte is made visible to other CPUs by being
414 * put into page tables.
416 * The other side of the story is the pointer chasing in the page
417 * table walking code (when walking the page table without locking;
418 * ie. most of the time). Fortunately, these data accesses consist
419 * of a chain of data-dependent loads, meaning most CPUs (alpha
420 * being the notable exception) will already guarantee loads are
421 * seen in-order. See the alpha page table accessors for the
422 * smp_read_barrier_depends() barriers in page table walking code.
424 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
426 ptl = pmd_lock(mm, pmd);
427 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
429 pmd_populate(mm, pmd, new);
438 int __pte_alloc_kernel(pmd_t *pmd)
440 pte_t *new = pte_alloc_one_kernel(&init_mm);
444 smp_wmb(); /* See comment in __pte_alloc */
446 spin_lock(&init_mm.page_table_lock);
447 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
448 pmd_populate_kernel(&init_mm, pmd, new);
451 spin_unlock(&init_mm.page_table_lock);
453 pte_free_kernel(&init_mm, new);
457 static inline void init_rss_vec(int *rss)
459 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
462 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
466 if (current->mm == mm)
468 for (i = 0; i < NR_MM_COUNTERS; i++)
470 add_mm_counter(mm, i, rss[i]);
474 * This function is called to print an error when a bad pte
475 * is found. For example, we might have a PFN-mapped pte in
476 * a region that doesn't allow it.
478 * The calling function must still handle the error.
480 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
481 pte_t pte, struct page *page)
483 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
484 p4d_t *p4d = p4d_offset(pgd, addr);
485 pud_t *pud = pud_offset(p4d, addr);
486 pmd_t *pmd = pmd_offset(pud, addr);
487 struct address_space *mapping;
489 static unsigned long resume;
490 static unsigned long nr_shown;
491 static unsigned long nr_unshown;
494 * Allow a burst of 60 reports, then keep quiet for that minute;
495 * or allow a steady drip of one report per second.
497 if (nr_shown == 60) {
498 if (time_before(jiffies, resume)) {
503 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
510 resume = jiffies + 60 * HZ;
512 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
513 index = linear_page_index(vma, addr);
515 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
517 (long long)pte_val(pte), (long long)pmd_val(*pmd));
519 dump_page(page, "bad pte");
520 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
521 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
522 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
524 vma->vm_ops ? vma->vm_ops->fault : NULL,
525 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
526 mapping ? mapping->a_ops->readpage : NULL);
528 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
532 * vm_normal_page -- This function gets the "struct page" associated with a pte.
534 * "Special" mappings do not wish to be associated with a "struct page" (either
535 * it doesn't exist, or it exists but they don't want to touch it). In this
536 * case, NULL is returned here. "Normal" mappings do have a struct page.
538 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
539 * pte bit, in which case this function is trivial. Secondly, an architecture
540 * may not have a spare pte bit, which requires a more complicated scheme,
543 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
544 * special mapping (even if there are underlying and valid "struct pages").
545 * COWed pages of a VM_PFNMAP are always normal.
547 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
548 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
549 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
550 * mapping will always honor the rule
552 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
554 * And for normal mappings this is false.
556 * This restricts such mappings to be a linear translation from virtual address
557 * to pfn. To get around this restriction, we allow arbitrary mappings so long
558 * as the vma is not a COW mapping; in that case, we know that all ptes are
559 * special (because none can have been COWed).
562 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
564 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
565 * page" backing, however the difference is that _all_ pages with a struct
566 * page (that is, those where pfn_valid is true) are refcounted and considered
567 * normal pages by the VM. The disadvantage is that pages are refcounted
568 * (which can be slower and simply not an option for some PFNMAP users). The
569 * advantage is that we don't have to follow the strict linearity rule of
570 * PFNMAP mappings in order to support COWable mappings.
573 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
574 pte_t pte, bool with_public_device)
576 unsigned long pfn = pte_pfn(pte);
578 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
579 if (likely(!pte_special(pte)))
581 if (vma->vm_ops && vma->vm_ops->find_special_page)
582 return vma->vm_ops->find_special_page(vma, addr);
583 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
585 if (is_zero_pfn(pfn))
589 * Device public pages are special pages (they are ZONE_DEVICE
590 * pages but different from persistent memory). They behave
591 * allmost like normal pages. The difference is that they are
592 * not on the lru and thus should never be involve with any-
593 * thing that involve lru manipulation (mlock, numa balancing,
596 * This is why we still want to return NULL for such page from
597 * vm_normal_page() so that we do not have to special case all
598 * call site of vm_normal_page().
600 if (likely(pfn <= highest_memmap_pfn)) {
601 struct page *page = pfn_to_page(pfn);
603 if (is_device_public_page(page)) {
604 if (with_public_device)
613 print_bad_pte(vma, addr, pte, NULL);
617 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
619 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
620 if (vma->vm_flags & VM_MIXEDMAP) {
626 off = (addr - vma->vm_start) >> PAGE_SHIFT;
627 if (pfn == vma->vm_pgoff + off)
629 if (!is_cow_mapping(vma->vm_flags))
634 if (is_zero_pfn(pfn))
638 if (unlikely(pfn > highest_memmap_pfn)) {
639 print_bad_pte(vma, addr, pte, NULL);
644 * NOTE! We still have PageReserved() pages in the page tables.
645 * eg. VDSO mappings can cause them to exist.
648 return pfn_to_page(pfn);
651 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
652 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
655 unsigned long pfn = pmd_pfn(pmd);
658 * There is no pmd_special() but there may be special pmds, e.g.
659 * in a direct-access (dax) mapping, so let's just replicate the
660 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
662 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
663 if (vma->vm_flags & VM_MIXEDMAP) {
669 off = (addr - vma->vm_start) >> PAGE_SHIFT;
670 if (pfn == vma->vm_pgoff + off)
672 if (!is_cow_mapping(vma->vm_flags))
679 if (is_zero_pfn(pfn))
681 if (unlikely(pfn > highest_memmap_pfn))
685 * NOTE! We still have PageReserved() pages in the page tables.
686 * eg. VDSO mappings can cause them to exist.
689 return pfn_to_page(pfn);
694 * copy one vm_area from one task to the other. Assumes the page tables
695 * already present in the new task to be cleared in the whole range
696 * covered by this vma.
699 static inline unsigned long
700 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
701 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
702 unsigned long addr, int *rss)
704 unsigned long vm_flags = vma->vm_flags;
705 pte_t pte = *src_pte;
708 /* pte contains position in swap or file, so copy. */
709 if (unlikely(!pte_present(pte))) {
710 swp_entry_t entry = pte_to_swp_entry(pte);
712 if (likely(!non_swap_entry(entry))) {
713 if (swap_duplicate(entry) < 0)
716 /* make sure dst_mm is on swapoff's mmlist. */
717 if (unlikely(list_empty(&dst_mm->mmlist))) {
718 spin_lock(&mmlist_lock);
719 if (list_empty(&dst_mm->mmlist))
720 list_add(&dst_mm->mmlist,
722 spin_unlock(&mmlist_lock);
725 } else if (is_migration_entry(entry)) {
726 page = migration_entry_to_page(entry);
728 rss[mm_counter(page)]++;
730 if (is_write_migration_entry(entry) &&
731 is_cow_mapping(vm_flags)) {
733 * COW mappings require pages in both
734 * parent and child to be set to read.
736 make_migration_entry_read(&entry);
737 pte = swp_entry_to_pte(entry);
738 if (pte_swp_soft_dirty(*src_pte))
739 pte = pte_swp_mksoft_dirty(pte);
740 set_pte_at(src_mm, addr, src_pte, pte);
742 } else if (is_device_private_entry(entry)) {
743 page = device_private_entry_to_page(entry);
746 * Update rss count even for unaddressable pages, as
747 * they should treated just like normal pages in this
750 * We will likely want to have some new rss counters
751 * for unaddressable pages, at some point. But for now
752 * keep things as they are.
755 rss[mm_counter(page)]++;
756 page_dup_rmap(page, false);
759 * We do not preserve soft-dirty information, because so
760 * far, checkpoint/restore is the only feature that
761 * requires that. And checkpoint/restore does not work
762 * when a device driver is involved (you cannot easily
763 * save and restore device driver state).
765 if (is_write_device_private_entry(entry) &&
766 is_cow_mapping(vm_flags)) {
767 make_device_private_entry_read(&entry);
768 pte = swp_entry_to_pte(entry);
769 set_pte_at(src_mm, addr, src_pte, pte);
776 * If it's a COW mapping, write protect it both
777 * in the parent and the child
779 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
780 ptep_set_wrprotect(src_mm, addr, src_pte);
781 pte = pte_wrprotect(pte);
785 * If it's a shared mapping, mark it clean in
788 if (vm_flags & VM_SHARED)
789 pte = pte_mkclean(pte);
790 pte = pte_mkold(pte);
792 page = vm_normal_page(vma, addr, pte);
795 page_dup_rmap(page, false);
796 rss[mm_counter(page)]++;
797 } else if (pte_devmap(pte)) {
798 page = pte_page(pte);
801 * Cache coherent device memory behave like regular page and
802 * not like persistent memory page. For more informations see
803 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
805 if (is_device_public_page(page)) {
807 page_dup_rmap(page, false);
808 rss[mm_counter(page)]++;
813 set_pte_at(dst_mm, addr, dst_pte, pte);
817 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
818 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
819 unsigned long addr, unsigned long end)
821 pte_t *orig_src_pte, *orig_dst_pte;
822 pte_t *src_pte, *dst_pte;
823 spinlock_t *src_ptl, *dst_ptl;
825 int rss[NR_MM_COUNTERS];
826 swp_entry_t entry = (swp_entry_t){0};
831 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
834 src_pte = pte_offset_map(src_pmd, addr);
835 src_ptl = pte_lockptr(src_mm, src_pmd);
836 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
837 orig_src_pte = src_pte;
838 orig_dst_pte = dst_pte;
839 arch_enter_lazy_mmu_mode();
843 * We are holding two locks at this point - either of them
844 * could generate latencies in another task on another CPU.
846 if (progress >= 32) {
848 if (need_resched() ||
849 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
852 if (pte_none(*src_pte)) {
856 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
861 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
863 arch_leave_lazy_mmu_mode();
864 spin_unlock(src_ptl);
865 pte_unmap(orig_src_pte);
866 add_mm_rss_vec(dst_mm, rss);
867 pte_unmap_unlock(orig_dst_pte, dst_ptl);
871 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
880 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
881 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
882 unsigned long addr, unsigned long end)
884 pmd_t *src_pmd, *dst_pmd;
887 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
890 src_pmd = pmd_offset(src_pud, addr);
892 next = pmd_addr_end(addr, end);
893 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
894 || pmd_devmap(*src_pmd)) {
896 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
897 err = copy_huge_pmd(dst_mm, src_mm,
898 dst_pmd, src_pmd, addr, vma);
905 if (pmd_none_or_clear_bad(src_pmd))
907 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
910 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
914 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
915 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
916 unsigned long addr, unsigned long end)
918 pud_t *src_pud, *dst_pud;
921 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
924 src_pud = pud_offset(src_p4d, addr);
926 next = pud_addr_end(addr, end);
927 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
930 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
931 err = copy_huge_pud(dst_mm, src_mm,
932 dst_pud, src_pud, addr, vma);
939 if (pud_none_or_clear_bad(src_pud))
941 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
944 } while (dst_pud++, src_pud++, addr = next, addr != end);
948 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
949 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
950 unsigned long addr, unsigned long end)
952 p4d_t *src_p4d, *dst_p4d;
955 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
958 src_p4d = p4d_offset(src_pgd, addr);
960 next = p4d_addr_end(addr, end);
961 if (p4d_none_or_clear_bad(src_p4d))
963 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
966 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
970 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
971 struct vm_area_struct *vma)
973 pgd_t *src_pgd, *dst_pgd;
975 unsigned long addr = vma->vm_start;
976 unsigned long end = vma->vm_end;
977 struct mmu_notifier_range range;
982 * Don't copy ptes where a page fault will fill them correctly.
983 * Fork becomes much lighter when there are big shared or private
984 * readonly mappings. The tradeoff is that copy_page_range is more
985 * efficient than faulting.
987 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
991 if (is_vm_hugetlb_page(vma))
992 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
994 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
996 * We do not free on error cases below as remove_vma
997 * gets called on error from higher level routine
999 ret = track_pfn_copy(vma);
1005 * We need to invalidate the secondary MMU mappings only when
1006 * there could be a permission downgrade on the ptes of the
1007 * parent mm. And a permission downgrade will only happen if
1008 * is_cow_mapping() returns true.
1010 is_cow = is_cow_mapping(vma->vm_flags);
1013 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1014 0, vma, src_mm, addr, end);
1015 mmu_notifier_invalidate_range_start(&range);
1019 dst_pgd = pgd_offset(dst_mm, addr);
1020 src_pgd = pgd_offset(src_mm, addr);
1022 next = pgd_addr_end(addr, end);
1023 if (pgd_none_or_clear_bad(src_pgd))
1025 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1026 vma, addr, next))) {
1030 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1033 mmu_notifier_invalidate_range_end(&range);
1037 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1038 struct vm_area_struct *vma, pmd_t *pmd,
1039 unsigned long addr, unsigned long end,
1040 struct zap_details *details)
1042 struct mm_struct *mm = tlb->mm;
1043 int force_flush = 0;
1044 int rss[NR_MM_COUNTERS];
1050 tlb_change_page_size(tlb, PAGE_SIZE);
1053 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1055 flush_tlb_batched_pending(mm);
1056 arch_enter_lazy_mmu_mode();
1059 if (pte_none(ptent))
1062 if (pte_present(ptent)) {
1065 page = _vm_normal_page(vma, addr, ptent, true);
1066 if (unlikely(details) && page) {
1068 * unmap_shared_mapping_pages() wants to
1069 * invalidate cache without truncating:
1070 * unmap shared but keep private pages.
1072 if (details->check_mapping &&
1073 details->check_mapping != page_rmapping(page))
1076 ptent = ptep_get_and_clear_full(mm, addr, pte,
1078 tlb_remove_tlb_entry(tlb, pte, addr);
1079 if (unlikely(!page))
1082 if (!PageAnon(page)) {
1083 if (pte_dirty(ptent)) {
1085 set_page_dirty(page);
1087 if (pte_young(ptent) &&
1088 likely(!(vma->vm_flags & VM_SEQ_READ)))
1089 mark_page_accessed(page);
1091 rss[mm_counter(page)]--;
1092 page_remove_rmap(page, false);
1093 if (unlikely(page_mapcount(page) < 0))
1094 print_bad_pte(vma, addr, ptent, page);
1095 if (unlikely(__tlb_remove_page(tlb, page))) {
1103 entry = pte_to_swp_entry(ptent);
1104 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1105 struct page *page = device_private_entry_to_page(entry);
1107 if (unlikely(details && details->check_mapping)) {
1109 * unmap_shared_mapping_pages() wants to
1110 * invalidate cache without truncating:
1111 * unmap shared but keep private pages.
1113 if (details->check_mapping !=
1114 page_rmapping(page))
1118 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1119 rss[mm_counter(page)]--;
1120 page_remove_rmap(page, false);
1125 /* If details->check_mapping, we leave swap entries. */
1126 if (unlikely(details))
1129 entry = pte_to_swp_entry(ptent);
1130 if (!non_swap_entry(entry))
1132 else if (is_migration_entry(entry)) {
1135 page = migration_entry_to_page(entry);
1136 rss[mm_counter(page)]--;
1138 if (unlikely(!free_swap_and_cache(entry)))
1139 print_bad_pte(vma, addr, ptent, NULL);
1140 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1141 } while (pte++, addr += PAGE_SIZE, addr != end);
1143 add_mm_rss_vec(mm, rss);
1144 arch_leave_lazy_mmu_mode();
1146 /* Do the actual TLB flush before dropping ptl */
1148 tlb_flush_mmu_tlbonly(tlb);
1149 pte_unmap_unlock(start_pte, ptl);
1152 * If we forced a TLB flush (either due to running out of
1153 * batch buffers or because we needed to flush dirty TLB
1154 * entries before releasing the ptl), free the batched
1155 * memory too. Restart if we didn't do everything.
1167 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1168 struct vm_area_struct *vma, pud_t *pud,
1169 unsigned long addr, unsigned long end,
1170 struct zap_details *details)
1175 pmd = pmd_offset(pud, addr);
1177 next = pmd_addr_end(addr, end);
1178 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1179 if (next - addr != HPAGE_PMD_SIZE)
1180 __split_huge_pmd(vma, pmd, addr, false, NULL);
1181 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1186 * Here there can be other concurrent MADV_DONTNEED or
1187 * trans huge page faults running, and if the pmd is
1188 * none or trans huge it can change under us. This is
1189 * because MADV_DONTNEED holds the mmap_sem in read
1192 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1194 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1197 } while (pmd++, addr = next, addr != end);
1202 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1203 struct vm_area_struct *vma, p4d_t *p4d,
1204 unsigned long addr, unsigned long end,
1205 struct zap_details *details)
1210 pud = pud_offset(p4d, addr);
1212 next = pud_addr_end(addr, end);
1213 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1214 if (next - addr != HPAGE_PUD_SIZE) {
1215 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1216 split_huge_pud(vma, pud, addr);
1217 } else if (zap_huge_pud(tlb, vma, pud, addr))
1221 if (pud_none_or_clear_bad(pud))
1223 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1226 } while (pud++, addr = next, addr != end);
1231 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1232 struct vm_area_struct *vma, pgd_t *pgd,
1233 unsigned long addr, unsigned long end,
1234 struct zap_details *details)
1239 p4d = p4d_offset(pgd, addr);
1241 next = p4d_addr_end(addr, end);
1242 if (p4d_none_or_clear_bad(p4d))
1244 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1245 } while (p4d++, addr = next, addr != end);
1250 void unmap_page_range(struct mmu_gather *tlb,
1251 struct vm_area_struct *vma,
1252 unsigned long addr, unsigned long end,
1253 struct zap_details *details)
1258 BUG_ON(addr >= end);
1259 tlb_start_vma(tlb, vma);
1260 pgd = pgd_offset(vma->vm_mm, addr);
1262 next = pgd_addr_end(addr, end);
1263 if (pgd_none_or_clear_bad(pgd))
1265 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1266 } while (pgd++, addr = next, addr != end);
1267 tlb_end_vma(tlb, vma);
1271 static void unmap_single_vma(struct mmu_gather *tlb,
1272 struct vm_area_struct *vma, unsigned long start_addr,
1273 unsigned long end_addr,
1274 struct zap_details *details)
1276 unsigned long start = max(vma->vm_start, start_addr);
1279 if (start >= vma->vm_end)
1281 end = min(vma->vm_end, end_addr);
1282 if (end <= vma->vm_start)
1286 uprobe_munmap(vma, start, end);
1288 if (unlikely(vma->vm_flags & VM_PFNMAP))
1289 untrack_pfn(vma, 0, 0);
1292 if (unlikely(is_vm_hugetlb_page(vma))) {
1294 * It is undesirable to test vma->vm_file as it
1295 * should be non-null for valid hugetlb area.
1296 * However, vm_file will be NULL in the error
1297 * cleanup path of mmap_region. When
1298 * hugetlbfs ->mmap method fails,
1299 * mmap_region() nullifies vma->vm_file
1300 * before calling this function to clean up.
1301 * Since no pte has actually been setup, it is
1302 * safe to do nothing in this case.
1305 i_mmap_lock_write(vma->vm_file->f_mapping);
1306 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1307 i_mmap_unlock_write(vma->vm_file->f_mapping);
1310 unmap_page_range(tlb, vma, start, end, details);
1315 * unmap_vmas - unmap a range of memory covered by a list of vma's
1316 * @tlb: address of the caller's struct mmu_gather
1317 * @vma: the starting vma
1318 * @start_addr: virtual address at which to start unmapping
1319 * @end_addr: virtual address at which to end unmapping
1321 * Unmap all pages in the vma list.
1323 * Only addresses between `start' and `end' will be unmapped.
1325 * The VMA list must be sorted in ascending virtual address order.
1327 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1328 * range after unmap_vmas() returns. So the only responsibility here is to
1329 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1330 * drops the lock and schedules.
1332 void unmap_vmas(struct mmu_gather *tlb,
1333 struct vm_area_struct *vma, unsigned long start_addr,
1334 unsigned long end_addr)
1336 struct mmu_notifier_range range;
1338 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1339 start_addr, end_addr);
1340 mmu_notifier_invalidate_range_start(&range);
1341 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1342 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1343 mmu_notifier_invalidate_range_end(&range);
1347 * zap_page_range - remove user pages in a given range
1348 * @vma: vm_area_struct holding the applicable pages
1349 * @start: starting address of pages to zap
1350 * @size: number of bytes to zap
1352 * Caller must protect the VMA list
1354 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1357 struct mmu_notifier_range range;
1358 struct mmu_gather tlb;
1361 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1362 start, start + size);
1363 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1364 update_hiwater_rss(vma->vm_mm);
1365 mmu_notifier_invalidate_range_start(&range);
1366 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1367 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1368 mmu_notifier_invalidate_range_end(&range);
1369 tlb_finish_mmu(&tlb, start, range.end);
1373 * zap_page_range_single - remove user pages in a given range
1374 * @vma: vm_area_struct holding the applicable pages
1375 * @address: starting address of pages to zap
1376 * @size: number of bytes to zap
1377 * @details: details of shared cache invalidation
1379 * The range must fit into one VMA.
1381 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1382 unsigned long size, struct zap_details *details)
1384 struct mmu_notifier_range range;
1385 struct mmu_gather tlb;
1388 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1389 address, address + size);
1390 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1391 update_hiwater_rss(vma->vm_mm);
1392 mmu_notifier_invalidate_range_start(&range);
1393 unmap_single_vma(&tlb, vma, address, range.end, details);
1394 mmu_notifier_invalidate_range_end(&range);
1395 tlb_finish_mmu(&tlb, address, range.end);
1399 * zap_vma_ptes - remove ptes mapping the vma
1400 * @vma: vm_area_struct holding ptes to be zapped
1401 * @address: starting address of pages to zap
1402 * @size: number of bytes to zap
1404 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1406 * The entire address range must be fully contained within the vma.
1409 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1412 if (address < vma->vm_start || address + size > vma->vm_end ||
1413 !(vma->vm_flags & VM_PFNMAP))
1416 zap_page_range_single(vma, address, size, NULL);
1418 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1420 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1428 pgd = pgd_offset(mm, addr);
1429 p4d = p4d_alloc(mm, pgd, addr);
1432 pud = pud_alloc(mm, p4d, addr);
1435 pmd = pmd_alloc(mm, pud, addr);
1439 VM_BUG_ON(pmd_trans_huge(*pmd));
1440 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1444 * This is the old fallback for page remapping.
1446 * For historical reasons, it only allows reserved pages. Only
1447 * old drivers should use this, and they needed to mark their
1448 * pages reserved for the old functions anyway.
1450 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1451 struct page *page, pgprot_t prot)
1453 struct mm_struct *mm = vma->vm_mm;
1459 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1462 flush_dcache_page(page);
1463 pte = get_locked_pte(mm, addr, &ptl);
1467 if (!pte_none(*pte))
1470 /* Ok, finally just insert the thing.. */
1472 inc_mm_counter_fast(mm, mm_counter_file(page));
1473 page_add_file_rmap(page, false);
1474 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1477 pte_unmap_unlock(pte, ptl);
1480 pte_unmap_unlock(pte, ptl);
1486 * vm_insert_page - insert single page into user vma
1487 * @vma: user vma to map to
1488 * @addr: target user address of this page
1489 * @page: source kernel page
1491 * This allows drivers to insert individual pages they've allocated
1494 * The page has to be a nice clean _individual_ kernel allocation.
1495 * If you allocate a compound page, you need to have marked it as
1496 * such (__GFP_COMP), or manually just split the page up yourself
1497 * (see split_page()).
1499 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1500 * took an arbitrary page protection parameter. This doesn't allow
1501 * that. Your vma protection will have to be set up correctly, which
1502 * means that if you want a shared writable mapping, you'd better
1503 * ask for a shared writable mapping!
1505 * The page does not need to be reserved.
1507 * Usually this function is called from f_op->mmap() handler
1508 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1509 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1510 * function from other places, for example from page-fault handler.
1512 * Return: %0 on success, negative error code otherwise.
1514 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1517 if (addr < vma->vm_start || addr >= vma->vm_end)
1519 if (!page_count(page))
1521 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1522 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1523 BUG_ON(vma->vm_flags & VM_PFNMAP);
1524 vma->vm_flags |= VM_MIXEDMAP;
1526 return insert_page(vma, addr, page, vma->vm_page_prot);
1528 EXPORT_SYMBOL(vm_insert_page);
1530 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1531 pfn_t pfn, pgprot_t prot, bool mkwrite)
1533 struct mm_struct *mm = vma->vm_mm;
1537 pte = get_locked_pte(mm, addr, &ptl);
1539 return VM_FAULT_OOM;
1540 if (!pte_none(*pte)) {
1543 * For read faults on private mappings the PFN passed
1544 * in may not match the PFN we have mapped if the
1545 * mapped PFN is a writeable COW page. In the mkwrite
1546 * case we are creating a writable PTE for a shared
1547 * mapping and we expect the PFNs to match. If they
1548 * don't match, we are likely racing with block
1549 * allocation and mapping invalidation so just skip the
1552 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1553 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1556 entry = pte_mkyoung(*pte);
1557 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1558 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1559 update_mmu_cache(vma, addr, pte);
1564 /* Ok, finally just insert the thing.. */
1565 if (pfn_t_devmap(pfn))
1566 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1568 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1571 entry = pte_mkyoung(entry);
1572 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1575 set_pte_at(mm, addr, pte, entry);
1576 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1579 pte_unmap_unlock(pte, ptl);
1580 return VM_FAULT_NOPAGE;
1584 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1585 * @vma: user vma to map to
1586 * @addr: target user address of this page
1587 * @pfn: source kernel pfn
1588 * @pgprot: pgprot flags for the inserted page
1590 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1591 * to override pgprot on a per-page basis.
1593 * This only makes sense for IO mappings, and it makes no sense for
1594 * COW mappings. In general, using multiple vmas is preferable;
1595 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1598 * Context: Process context. May allocate using %GFP_KERNEL.
1599 * Return: vm_fault_t value.
1601 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1602 unsigned long pfn, pgprot_t pgprot)
1605 * Technically, architectures with pte_special can avoid all these
1606 * restrictions (same for remap_pfn_range). However we would like
1607 * consistency in testing and feature parity among all, so we should
1608 * try to keep these invariants in place for everybody.
1610 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1611 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1612 (VM_PFNMAP|VM_MIXEDMAP));
1613 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1614 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1616 if (addr < vma->vm_start || addr >= vma->vm_end)
1617 return VM_FAULT_SIGBUS;
1619 if (!pfn_modify_allowed(pfn, pgprot))
1620 return VM_FAULT_SIGBUS;
1622 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1624 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1627 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1630 * vmf_insert_pfn - insert single pfn into user vma
1631 * @vma: user vma to map to
1632 * @addr: target user address of this page
1633 * @pfn: source kernel pfn
1635 * Similar to vm_insert_page, this allows drivers to insert individual pages
1636 * they've allocated into a user vma. Same comments apply.
1638 * This function should only be called from a vm_ops->fault handler, and
1639 * in that case the handler should return the result of this function.
1641 * vma cannot be a COW mapping.
1643 * As this is called only for pages that do not currently exist, we
1644 * do not need to flush old virtual caches or the TLB.
1646 * Context: Process context. May allocate using %GFP_KERNEL.
1647 * Return: vm_fault_t value.
1649 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1652 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1654 EXPORT_SYMBOL(vmf_insert_pfn);
1656 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1658 /* these checks mirror the abort conditions in vm_normal_page */
1659 if (vma->vm_flags & VM_MIXEDMAP)
1661 if (pfn_t_devmap(pfn))
1663 if (pfn_t_special(pfn))
1665 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1670 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1671 unsigned long addr, pfn_t pfn, bool mkwrite)
1673 pgprot_t pgprot = vma->vm_page_prot;
1676 BUG_ON(!vm_mixed_ok(vma, pfn));
1678 if (addr < vma->vm_start || addr >= vma->vm_end)
1679 return VM_FAULT_SIGBUS;
1681 track_pfn_insert(vma, &pgprot, pfn);
1683 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1684 return VM_FAULT_SIGBUS;
1687 * If we don't have pte special, then we have to use the pfn_valid()
1688 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1689 * refcount the page if pfn_valid is true (hence insert_page rather
1690 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1691 * without pte special, it would there be refcounted as a normal page.
1693 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1694 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1698 * At this point we are committed to insert_page()
1699 * regardless of whether the caller specified flags that
1700 * result in pfn_t_has_page() == false.
1702 page = pfn_to_page(pfn_t_to_pfn(pfn));
1703 err = insert_page(vma, addr, page, pgprot);
1705 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1709 return VM_FAULT_OOM;
1710 if (err < 0 && err != -EBUSY)
1711 return VM_FAULT_SIGBUS;
1713 return VM_FAULT_NOPAGE;
1716 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1719 return __vm_insert_mixed(vma, addr, pfn, false);
1721 EXPORT_SYMBOL(vmf_insert_mixed);
1724 * If the insertion of PTE failed because someone else already added a
1725 * different entry in the mean time, we treat that as success as we assume
1726 * the same entry was actually inserted.
1728 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1729 unsigned long addr, pfn_t pfn)
1731 return __vm_insert_mixed(vma, addr, pfn, true);
1733 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1736 * maps a range of physical memory into the requested pages. the old
1737 * mappings are removed. any references to nonexistent pages results
1738 * in null mappings (currently treated as "copy-on-access")
1740 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1741 unsigned long addr, unsigned long end,
1742 unsigned long pfn, pgprot_t prot)
1748 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1751 arch_enter_lazy_mmu_mode();
1753 BUG_ON(!pte_none(*pte));
1754 if (!pfn_modify_allowed(pfn, prot)) {
1758 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1760 } while (pte++, addr += PAGE_SIZE, addr != end);
1761 arch_leave_lazy_mmu_mode();
1762 pte_unmap_unlock(pte - 1, ptl);
1766 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1767 unsigned long addr, unsigned long end,
1768 unsigned long pfn, pgprot_t prot)
1774 pfn -= addr >> PAGE_SHIFT;
1775 pmd = pmd_alloc(mm, pud, addr);
1778 VM_BUG_ON(pmd_trans_huge(*pmd));
1780 next = pmd_addr_end(addr, end);
1781 err = remap_pte_range(mm, pmd, addr, next,
1782 pfn + (addr >> PAGE_SHIFT), prot);
1785 } while (pmd++, addr = next, addr != end);
1789 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1790 unsigned long addr, unsigned long end,
1791 unsigned long pfn, pgprot_t prot)
1797 pfn -= addr >> PAGE_SHIFT;
1798 pud = pud_alloc(mm, p4d, addr);
1802 next = pud_addr_end(addr, end);
1803 err = remap_pmd_range(mm, pud, addr, next,
1804 pfn + (addr >> PAGE_SHIFT), prot);
1807 } while (pud++, addr = next, addr != end);
1811 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1812 unsigned long addr, unsigned long end,
1813 unsigned long pfn, pgprot_t prot)
1819 pfn -= addr >> PAGE_SHIFT;
1820 p4d = p4d_alloc(mm, pgd, addr);
1824 next = p4d_addr_end(addr, end);
1825 err = remap_pud_range(mm, p4d, addr, next,
1826 pfn + (addr >> PAGE_SHIFT), prot);
1829 } while (p4d++, addr = next, addr != end);
1834 * remap_pfn_range - remap kernel memory to userspace
1835 * @vma: user vma to map to
1836 * @addr: target user address to start at
1837 * @pfn: physical address of kernel memory
1838 * @size: size of map area
1839 * @prot: page protection flags for this mapping
1841 * Note: this is only safe if the mm semaphore is held when called.
1843 * Return: %0 on success, negative error code otherwise.
1845 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1846 unsigned long pfn, unsigned long size, pgprot_t prot)
1850 unsigned long end = addr + PAGE_ALIGN(size);
1851 struct mm_struct *mm = vma->vm_mm;
1852 unsigned long remap_pfn = pfn;
1856 * Physically remapped pages are special. Tell the
1857 * rest of the world about it:
1858 * VM_IO tells people not to look at these pages
1859 * (accesses can have side effects).
1860 * VM_PFNMAP tells the core MM that the base pages are just
1861 * raw PFN mappings, and do not have a "struct page" associated
1864 * Disable vma merging and expanding with mremap().
1866 * Omit vma from core dump, even when VM_IO turned off.
1868 * There's a horrible special case to handle copy-on-write
1869 * behaviour that some programs depend on. We mark the "original"
1870 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1871 * See vm_normal_page() for details.
1873 if (is_cow_mapping(vma->vm_flags)) {
1874 if (addr != vma->vm_start || end != vma->vm_end)
1876 vma->vm_pgoff = pfn;
1879 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1883 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1885 BUG_ON(addr >= end);
1886 pfn -= addr >> PAGE_SHIFT;
1887 pgd = pgd_offset(mm, addr);
1888 flush_cache_range(vma, addr, end);
1890 next = pgd_addr_end(addr, end);
1891 err = remap_p4d_range(mm, pgd, addr, next,
1892 pfn + (addr >> PAGE_SHIFT), prot);
1895 } while (pgd++, addr = next, addr != end);
1898 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1902 EXPORT_SYMBOL(remap_pfn_range);
1905 * vm_iomap_memory - remap memory to userspace
1906 * @vma: user vma to map to
1907 * @start: start of area
1908 * @len: size of area
1910 * This is a simplified io_remap_pfn_range() for common driver use. The
1911 * driver just needs to give us the physical memory range to be mapped,
1912 * we'll figure out the rest from the vma information.
1914 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1915 * whatever write-combining details or similar.
1917 * Return: %0 on success, negative error code otherwise.
1919 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1921 unsigned long vm_len, pfn, pages;
1923 /* Check that the physical memory area passed in looks valid */
1924 if (start + len < start)
1927 * You *really* shouldn't map things that aren't page-aligned,
1928 * but we've historically allowed it because IO memory might
1929 * just have smaller alignment.
1931 len += start & ~PAGE_MASK;
1932 pfn = start >> PAGE_SHIFT;
1933 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1934 if (pfn + pages < pfn)
1937 /* We start the mapping 'vm_pgoff' pages into the area */
1938 if (vma->vm_pgoff > pages)
1940 pfn += vma->vm_pgoff;
1941 pages -= vma->vm_pgoff;
1943 /* Can we fit all of the mapping? */
1944 vm_len = vma->vm_end - vma->vm_start;
1945 if (vm_len >> PAGE_SHIFT > pages)
1948 /* Ok, let it rip */
1949 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1951 EXPORT_SYMBOL(vm_iomap_memory);
1953 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1954 unsigned long addr, unsigned long end,
1955 pte_fn_t fn, void *data)
1960 spinlock_t *uninitialized_var(ptl);
1962 pte = (mm == &init_mm) ?
1963 pte_alloc_kernel(pmd, addr) :
1964 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1968 BUG_ON(pmd_huge(*pmd));
1970 arch_enter_lazy_mmu_mode();
1972 token = pmd_pgtable(*pmd);
1975 err = fn(pte++, token, addr, data);
1978 } while (addr += PAGE_SIZE, addr != end);
1980 arch_leave_lazy_mmu_mode();
1983 pte_unmap_unlock(pte-1, ptl);
1987 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1988 unsigned long addr, unsigned long end,
1989 pte_fn_t fn, void *data)
1995 BUG_ON(pud_huge(*pud));
1997 pmd = pmd_alloc(mm, pud, addr);
2001 next = pmd_addr_end(addr, end);
2002 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2005 } while (pmd++, addr = next, addr != end);
2009 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2010 unsigned long addr, unsigned long end,
2011 pte_fn_t fn, void *data)
2017 pud = pud_alloc(mm, p4d, addr);
2021 next = pud_addr_end(addr, end);
2022 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2025 } while (pud++, addr = next, addr != end);
2029 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2030 unsigned long addr, unsigned long end,
2031 pte_fn_t fn, void *data)
2037 p4d = p4d_alloc(mm, pgd, addr);
2041 next = p4d_addr_end(addr, end);
2042 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2045 } while (p4d++, addr = next, addr != end);
2050 * Scan a region of virtual memory, filling in page tables as necessary
2051 * and calling a provided function on each leaf page table.
2053 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2054 unsigned long size, pte_fn_t fn, void *data)
2058 unsigned long end = addr + size;
2061 if (WARN_ON(addr >= end))
2064 pgd = pgd_offset(mm, addr);
2066 next = pgd_addr_end(addr, end);
2067 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2070 } while (pgd++, addr = next, addr != end);
2074 EXPORT_SYMBOL_GPL(apply_to_page_range);
2077 * handle_pte_fault chooses page fault handler according to an entry which was
2078 * read non-atomically. Before making any commitment, on those architectures
2079 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2080 * parts, do_swap_page must check under lock before unmapping the pte and
2081 * proceeding (but do_wp_page is only called after already making such a check;
2082 * and do_anonymous_page can safely check later on).
2084 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2085 pte_t *page_table, pte_t orig_pte)
2088 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2089 if (sizeof(pte_t) > sizeof(unsigned long)) {
2090 spinlock_t *ptl = pte_lockptr(mm, pmd);
2092 same = pte_same(*page_table, orig_pte);
2096 pte_unmap(page_table);
2100 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2102 debug_dma_assert_idle(src);
2105 * If the source page was a PFN mapping, we don't have
2106 * a "struct page" for it. We do a best-effort copy by
2107 * just copying from the original user address. If that
2108 * fails, we just zero-fill it. Live with it.
2110 if (unlikely(!src)) {
2111 void *kaddr = kmap_atomic(dst);
2112 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2115 * This really shouldn't fail, because the page is there
2116 * in the page tables. But it might just be unreadable,
2117 * in which case we just give up and fill the result with
2120 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2122 kunmap_atomic(kaddr);
2123 flush_dcache_page(dst);
2125 copy_user_highpage(dst, src, va, vma);
2128 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2130 struct file *vm_file = vma->vm_file;
2133 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2136 * Special mappings (e.g. VDSO) do not have any file so fake
2137 * a default GFP_KERNEL for them.
2143 * Notify the address space that the page is about to become writable so that
2144 * it can prohibit this or wait for the page to get into an appropriate state.
2146 * We do this without the lock held, so that it can sleep if it needs to.
2148 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2151 struct page *page = vmf->page;
2152 unsigned int old_flags = vmf->flags;
2154 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2156 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2157 /* Restore original flags so that caller is not surprised */
2158 vmf->flags = old_flags;
2159 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2161 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2163 if (!page->mapping) {
2165 return 0; /* retry */
2167 ret |= VM_FAULT_LOCKED;
2169 VM_BUG_ON_PAGE(!PageLocked(page), page);
2174 * Handle dirtying of a page in shared file mapping on a write fault.
2176 * The function expects the page to be locked and unlocks it.
2178 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2181 struct address_space *mapping;
2183 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2185 dirtied = set_page_dirty(page);
2186 VM_BUG_ON_PAGE(PageAnon(page), page);
2188 * Take a local copy of the address_space - page.mapping may be zeroed
2189 * by truncate after unlock_page(). The address_space itself remains
2190 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2191 * release semantics to prevent the compiler from undoing this copying.
2193 mapping = page_rmapping(page);
2196 if ((dirtied || page_mkwrite) && mapping) {
2198 * Some device drivers do not set page.mapping
2199 * but still dirty their pages
2201 balance_dirty_pages_ratelimited(mapping);
2205 file_update_time(vma->vm_file);
2209 * Handle write page faults for pages that can be reused in the current vma
2211 * This can happen either due to the mapping being with the VM_SHARED flag,
2212 * or due to us being the last reference standing to the page. In either
2213 * case, all we need to do here is to mark the page as writable and update
2214 * any related book-keeping.
2216 static inline void wp_page_reuse(struct vm_fault *vmf)
2217 __releases(vmf->ptl)
2219 struct vm_area_struct *vma = vmf->vma;
2220 struct page *page = vmf->page;
2223 * Clear the pages cpupid information as the existing
2224 * information potentially belongs to a now completely
2225 * unrelated process.
2228 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2230 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2231 entry = pte_mkyoung(vmf->orig_pte);
2232 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2233 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2234 update_mmu_cache(vma, vmf->address, vmf->pte);
2235 pte_unmap_unlock(vmf->pte, vmf->ptl);
2239 * Handle the case of a page which we actually need to copy to a new page.
2241 * Called with mmap_sem locked and the old page referenced, but
2242 * without the ptl held.
2244 * High level logic flow:
2246 * - Allocate a page, copy the content of the old page to the new one.
2247 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2248 * - Take the PTL. If the pte changed, bail out and release the allocated page
2249 * - If the pte is still the way we remember it, update the page table and all
2250 * relevant references. This includes dropping the reference the page-table
2251 * held to the old page, as well as updating the rmap.
2252 * - In any case, unlock the PTL and drop the reference we took to the old page.
2254 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2256 struct vm_area_struct *vma = vmf->vma;
2257 struct mm_struct *mm = vma->vm_mm;
2258 struct page *old_page = vmf->page;
2259 struct page *new_page = NULL;
2261 int page_copied = 0;
2262 struct mem_cgroup *memcg;
2263 struct mmu_notifier_range range;
2265 if (unlikely(anon_vma_prepare(vma)))
2268 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2269 new_page = alloc_zeroed_user_highpage_movable(vma,
2274 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2278 cow_user_page(new_page, old_page, vmf->address, vma);
2281 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2284 __SetPageUptodate(new_page);
2286 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2287 vmf->address & PAGE_MASK,
2288 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2289 mmu_notifier_invalidate_range_start(&range);
2292 * Re-check the pte - we dropped the lock
2294 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2295 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2297 if (!PageAnon(old_page)) {
2298 dec_mm_counter_fast(mm,
2299 mm_counter_file(old_page));
2300 inc_mm_counter_fast(mm, MM_ANONPAGES);
2303 inc_mm_counter_fast(mm, MM_ANONPAGES);
2305 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2306 entry = mk_pte(new_page, vma->vm_page_prot);
2307 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2309 * Clear the pte entry and flush it first, before updating the
2310 * pte with the new entry. This will avoid a race condition
2311 * seen in the presence of one thread doing SMC and another
2314 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2315 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2316 mem_cgroup_commit_charge(new_page, memcg, false, false);
2317 lru_cache_add_active_or_unevictable(new_page, vma);
2319 * We call the notify macro here because, when using secondary
2320 * mmu page tables (such as kvm shadow page tables), we want the
2321 * new page to be mapped directly into the secondary page table.
2323 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2324 update_mmu_cache(vma, vmf->address, vmf->pte);
2327 * Only after switching the pte to the new page may
2328 * we remove the mapcount here. Otherwise another
2329 * process may come and find the rmap count decremented
2330 * before the pte is switched to the new page, and
2331 * "reuse" the old page writing into it while our pte
2332 * here still points into it and can be read by other
2335 * The critical issue is to order this
2336 * page_remove_rmap with the ptp_clear_flush above.
2337 * Those stores are ordered by (if nothing else,)
2338 * the barrier present in the atomic_add_negative
2339 * in page_remove_rmap.
2341 * Then the TLB flush in ptep_clear_flush ensures that
2342 * no process can access the old page before the
2343 * decremented mapcount is visible. And the old page
2344 * cannot be reused until after the decremented
2345 * mapcount is visible. So transitively, TLBs to
2346 * old page will be flushed before it can be reused.
2348 page_remove_rmap(old_page, false);
2351 /* Free the old page.. */
2352 new_page = old_page;
2355 mem_cgroup_cancel_charge(new_page, memcg, false);
2361 pte_unmap_unlock(vmf->pte, vmf->ptl);
2363 * No need to double call mmu_notifier->invalidate_range() callback as
2364 * the above ptep_clear_flush_notify() did already call it.
2366 mmu_notifier_invalidate_range_only_end(&range);
2369 * Don't let another task, with possibly unlocked vma,
2370 * keep the mlocked page.
2372 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2373 lock_page(old_page); /* LRU manipulation */
2374 if (PageMlocked(old_page))
2375 munlock_vma_page(old_page);
2376 unlock_page(old_page);
2380 return page_copied ? VM_FAULT_WRITE : 0;
2386 return VM_FAULT_OOM;
2390 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2391 * writeable once the page is prepared
2393 * @vmf: structure describing the fault
2395 * This function handles all that is needed to finish a write page fault in a
2396 * shared mapping due to PTE being read-only once the mapped page is prepared.
2397 * It handles locking of PTE and modifying it.
2399 * The function expects the page to be locked or other protection against
2400 * concurrent faults / writeback (such as DAX radix tree locks).
2402 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2403 * we acquired PTE lock.
2405 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2407 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2408 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2411 * We might have raced with another page fault while we released the
2412 * pte_offset_map_lock.
2414 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2415 pte_unmap_unlock(vmf->pte, vmf->ptl);
2416 return VM_FAULT_NOPAGE;
2423 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2426 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2428 struct vm_area_struct *vma = vmf->vma;
2430 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2433 pte_unmap_unlock(vmf->pte, vmf->ptl);
2434 vmf->flags |= FAULT_FLAG_MKWRITE;
2435 ret = vma->vm_ops->pfn_mkwrite(vmf);
2436 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2438 return finish_mkwrite_fault(vmf);
2441 return VM_FAULT_WRITE;
2444 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2445 __releases(vmf->ptl)
2447 struct vm_area_struct *vma = vmf->vma;
2449 get_page(vmf->page);
2451 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2454 pte_unmap_unlock(vmf->pte, vmf->ptl);
2455 tmp = do_page_mkwrite(vmf);
2456 if (unlikely(!tmp || (tmp &
2457 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2458 put_page(vmf->page);
2461 tmp = finish_mkwrite_fault(vmf);
2462 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2463 unlock_page(vmf->page);
2464 put_page(vmf->page);
2469 lock_page(vmf->page);
2471 fault_dirty_shared_page(vma, vmf->page);
2472 put_page(vmf->page);
2474 return VM_FAULT_WRITE;
2478 * This routine handles present pages, when users try to write
2479 * to a shared page. It is done by copying the page to a new address
2480 * and decrementing the shared-page counter for the old page.
2482 * Note that this routine assumes that the protection checks have been
2483 * done by the caller (the low-level page fault routine in most cases).
2484 * Thus we can safely just mark it writable once we've done any necessary
2487 * We also mark the page dirty at this point even though the page will
2488 * change only once the write actually happens. This avoids a few races,
2489 * and potentially makes it more efficient.
2491 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2492 * but allow concurrent faults), with pte both mapped and locked.
2493 * We return with mmap_sem still held, but pte unmapped and unlocked.
2495 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2496 __releases(vmf->ptl)
2498 struct vm_area_struct *vma = vmf->vma;
2500 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2503 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2506 * We should not cow pages in a shared writeable mapping.
2507 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2509 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2510 (VM_WRITE|VM_SHARED))
2511 return wp_pfn_shared(vmf);
2513 pte_unmap_unlock(vmf->pte, vmf->ptl);
2514 return wp_page_copy(vmf);
2518 * Take out anonymous pages first, anonymous shared vmas are
2519 * not dirty accountable.
2521 if (PageAnon(vmf->page)) {
2522 int total_map_swapcount;
2523 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2524 page_count(vmf->page) != 1))
2526 if (!trylock_page(vmf->page)) {
2527 get_page(vmf->page);
2528 pte_unmap_unlock(vmf->pte, vmf->ptl);
2529 lock_page(vmf->page);
2530 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2531 vmf->address, &vmf->ptl);
2532 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2533 unlock_page(vmf->page);
2534 pte_unmap_unlock(vmf->pte, vmf->ptl);
2535 put_page(vmf->page);
2538 put_page(vmf->page);
2540 if (PageKsm(vmf->page)) {
2541 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2543 unlock_page(vmf->page);
2547 return VM_FAULT_WRITE;
2549 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2550 if (total_map_swapcount == 1) {
2552 * The page is all ours. Move it to
2553 * our anon_vma so the rmap code will
2554 * not search our parent or siblings.
2555 * Protected against the rmap code by
2558 page_move_anon_rmap(vmf->page, vma);
2560 unlock_page(vmf->page);
2562 return VM_FAULT_WRITE;
2564 unlock_page(vmf->page);
2565 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2566 (VM_WRITE|VM_SHARED))) {
2567 return wp_page_shared(vmf);
2571 * Ok, we need to copy. Oh, well..
2573 get_page(vmf->page);
2575 pte_unmap_unlock(vmf->pte, vmf->ptl);
2576 return wp_page_copy(vmf);
2579 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2580 unsigned long start_addr, unsigned long end_addr,
2581 struct zap_details *details)
2583 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2586 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2587 struct zap_details *details)
2589 struct vm_area_struct *vma;
2590 pgoff_t vba, vea, zba, zea;
2592 vma_interval_tree_foreach(vma, root,
2593 details->first_index, details->last_index) {
2595 vba = vma->vm_pgoff;
2596 vea = vba + vma_pages(vma) - 1;
2597 zba = details->first_index;
2600 zea = details->last_index;
2604 unmap_mapping_range_vma(vma,
2605 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2606 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2612 * unmap_mapping_pages() - Unmap pages from processes.
2613 * @mapping: The address space containing pages to be unmapped.
2614 * @start: Index of first page to be unmapped.
2615 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2616 * @even_cows: Whether to unmap even private COWed pages.
2618 * Unmap the pages in this address space from any userspace process which
2619 * has them mmaped. Generally, you want to remove COWed pages as well when
2620 * a file is being truncated, but not when invalidating pages from the page
2623 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2624 pgoff_t nr, bool even_cows)
2626 struct zap_details details = { };
2628 details.check_mapping = even_cows ? NULL : mapping;
2629 details.first_index = start;
2630 details.last_index = start + nr - 1;
2631 if (details.last_index < details.first_index)
2632 details.last_index = ULONG_MAX;
2634 i_mmap_lock_write(mapping);
2635 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2636 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2637 i_mmap_unlock_write(mapping);
2641 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2642 * address_space corresponding to the specified byte range in the underlying
2645 * @mapping: the address space containing mmaps to be unmapped.
2646 * @holebegin: byte in first page to unmap, relative to the start of
2647 * the underlying file. This will be rounded down to a PAGE_SIZE
2648 * boundary. Note that this is different from truncate_pagecache(), which
2649 * must keep the partial page. In contrast, we must get rid of
2651 * @holelen: size of prospective hole in bytes. This will be rounded
2652 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2654 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2655 * but 0 when invalidating pagecache, don't throw away private data.
2657 void unmap_mapping_range(struct address_space *mapping,
2658 loff_t const holebegin, loff_t const holelen, int even_cows)
2660 pgoff_t hba = holebegin >> PAGE_SHIFT;
2661 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2663 /* Check for overflow. */
2664 if (sizeof(holelen) > sizeof(hlen)) {
2666 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2667 if (holeend & ~(long long)ULONG_MAX)
2668 hlen = ULONG_MAX - hba + 1;
2671 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2673 EXPORT_SYMBOL(unmap_mapping_range);
2676 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2677 * but allow concurrent faults), and pte mapped but not yet locked.
2678 * We return with pte unmapped and unlocked.
2680 * We return with the mmap_sem locked or unlocked in the same cases
2681 * as does filemap_fault().
2683 vm_fault_t do_swap_page(struct vm_fault *vmf)
2685 struct vm_area_struct *vma = vmf->vma;
2686 struct page *page = NULL, *swapcache;
2687 struct mem_cgroup *memcg;
2694 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2697 entry = pte_to_swp_entry(vmf->orig_pte);
2698 if (unlikely(non_swap_entry(entry))) {
2699 if (is_migration_entry(entry)) {
2700 migration_entry_wait(vma->vm_mm, vmf->pmd,
2702 } else if (is_device_private_entry(entry)) {
2704 * For un-addressable device memory we call the pgmap
2705 * fault handler callback. The callback must migrate
2706 * the page back to some CPU accessible page.
2708 ret = device_private_entry_fault(vma, vmf->address, entry,
2709 vmf->flags, vmf->pmd);
2710 } else if (is_hwpoison_entry(entry)) {
2711 ret = VM_FAULT_HWPOISON;
2713 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2714 ret = VM_FAULT_SIGBUS;
2720 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2721 page = lookup_swap_cache(entry, vma, vmf->address);
2725 struct swap_info_struct *si = swp_swap_info(entry);
2727 if (si->flags & SWP_SYNCHRONOUS_IO &&
2728 __swap_count(si, entry) == 1) {
2729 /* skip swapcache */
2730 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2733 __SetPageLocked(page);
2734 __SetPageSwapBacked(page);
2735 set_page_private(page, entry.val);
2736 lru_cache_add_anon(page);
2737 swap_readpage(page, true);
2740 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2747 * Back out if somebody else faulted in this pte
2748 * while we released the pte lock.
2750 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2751 vmf->address, &vmf->ptl);
2752 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2754 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2758 /* Had to read the page from swap area: Major fault */
2759 ret = VM_FAULT_MAJOR;
2760 count_vm_event(PGMAJFAULT);
2761 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2762 } else if (PageHWPoison(page)) {
2764 * hwpoisoned dirty swapcache pages are kept for killing
2765 * owner processes (which may be unknown at hwpoison time)
2767 ret = VM_FAULT_HWPOISON;
2768 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2772 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2774 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2776 ret |= VM_FAULT_RETRY;
2781 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2782 * release the swapcache from under us. The page pin, and pte_same
2783 * test below, are not enough to exclude that. Even if it is still
2784 * swapcache, we need to check that the page's swap has not changed.
2786 if (unlikely((!PageSwapCache(page) ||
2787 page_private(page) != entry.val)) && swapcache)
2790 page = ksm_might_need_to_copy(page, vma, vmf->address);
2791 if (unlikely(!page)) {
2797 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2804 * Back out if somebody else already faulted in this pte.
2806 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2808 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2811 if (unlikely(!PageUptodate(page))) {
2812 ret = VM_FAULT_SIGBUS;
2817 * The page isn't present yet, go ahead with the fault.
2819 * Be careful about the sequence of operations here.
2820 * To get its accounting right, reuse_swap_page() must be called
2821 * while the page is counted on swap but not yet in mapcount i.e.
2822 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2823 * must be called after the swap_free(), or it will never succeed.
2826 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2827 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2828 pte = mk_pte(page, vma->vm_page_prot);
2829 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2830 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2831 vmf->flags &= ~FAULT_FLAG_WRITE;
2832 ret |= VM_FAULT_WRITE;
2833 exclusive = RMAP_EXCLUSIVE;
2835 flush_icache_page(vma, page);
2836 if (pte_swp_soft_dirty(vmf->orig_pte))
2837 pte = pte_mksoft_dirty(pte);
2838 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2839 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2840 vmf->orig_pte = pte;
2842 /* ksm created a completely new copy */
2843 if (unlikely(page != swapcache && swapcache)) {
2844 page_add_new_anon_rmap(page, vma, vmf->address, false);
2845 mem_cgroup_commit_charge(page, memcg, false, false);
2846 lru_cache_add_active_or_unevictable(page, vma);
2848 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2849 mem_cgroup_commit_charge(page, memcg, true, false);
2850 activate_page(page);
2854 if (mem_cgroup_swap_full(page) ||
2855 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2856 try_to_free_swap(page);
2858 if (page != swapcache && swapcache) {
2860 * Hold the lock to avoid the swap entry to be reused
2861 * until we take the PT lock for the pte_same() check
2862 * (to avoid false positives from pte_same). For
2863 * further safety release the lock after the swap_free
2864 * so that the swap count won't change under a
2865 * parallel locked swapcache.
2867 unlock_page(swapcache);
2868 put_page(swapcache);
2871 if (vmf->flags & FAULT_FLAG_WRITE) {
2872 ret |= do_wp_page(vmf);
2873 if (ret & VM_FAULT_ERROR)
2874 ret &= VM_FAULT_ERROR;
2878 /* No need to invalidate - it was non-present before */
2879 update_mmu_cache(vma, vmf->address, vmf->pte);
2881 pte_unmap_unlock(vmf->pte, vmf->ptl);
2885 mem_cgroup_cancel_charge(page, memcg, false);
2886 pte_unmap_unlock(vmf->pte, vmf->ptl);
2891 if (page != swapcache && swapcache) {
2892 unlock_page(swapcache);
2893 put_page(swapcache);
2899 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2900 * but allow concurrent faults), and pte mapped but not yet locked.
2901 * We return with mmap_sem still held, but pte unmapped and unlocked.
2903 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2905 struct vm_area_struct *vma = vmf->vma;
2906 struct mem_cgroup *memcg;
2911 /* File mapping without ->vm_ops ? */
2912 if (vma->vm_flags & VM_SHARED)
2913 return VM_FAULT_SIGBUS;
2916 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2917 * pte_offset_map() on pmds where a huge pmd might be created
2918 * from a different thread.
2920 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2921 * parallel threads are excluded by other means.
2923 * Here we only have down_read(mmap_sem).
2925 if (pte_alloc(vma->vm_mm, vmf->pmd))
2926 return VM_FAULT_OOM;
2928 /* See the comment in pte_alloc_one_map() */
2929 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2932 /* Use the zero-page for reads */
2933 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2934 !mm_forbids_zeropage(vma->vm_mm)) {
2935 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2936 vma->vm_page_prot));
2937 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2938 vmf->address, &vmf->ptl);
2939 if (!pte_none(*vmf->pte))
2941 ret = check_stable_address_space(vma->vm_mm);
2944 /* Deliver the page fault to userland, check inside PT lock */
2945 if (userfaultfd_missing(vma)) {
2946 pte_unmap_unlock(vmf->pte, vmf->ptl);
2947 return handle_userfault(vmf, VM_UFFD_MISSING);
2952 /* Allocate our own private page. */
2953 if (unlikely(anon_vma_prepare(vma)))
2955 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2959 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
2964 * The memory barrier inside __SetPageUptodate makes sure that
2965 * preceeding stores to the page contents become visible before
2966 * the set_pte_at() write.
2968 __SetPageUptodate(page);
2970 entry = mk_pte(page, vma->vm_page_prot);
2971 if (vma->vm_flags & VM_WRITE)
2972 entry = pte_mkwrite(pte_mkdirty(entry));
2974 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2976 if (!pte_none(*vmf->pte))
2979 ret = check_stable_address_space(vma->vm_mm);
2983 /* Deliver the page fault to userland, check inside PT lock */
2984 if (userfaultfd_missing(vma)) {
2985 pte_unmap_unlock(vmf->pte, vmf->ptl);
2986 mem_cgroup_cancel_charge(page, memcg, false);
2988 return handle_userfault(vmf, VM_UFFD_MISSING);
2991 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2992 page_add_new_anon_rmap(page, vma, vmf->address, false);
2993 mem_cgroup_commit_charge(page, memcg, false, false);
2994 lru_cache_add_active_or_unevictable(page, vma);
2996 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2998 /* No need to invalidate - it was non-present before */
2999 update_mmu_cache(vma, vmf->address, vmf->pte);
3001 pte_unmap_unlock(vmf->pte, vmf->ptl);
3004 mem_cgroup_cancel_charge(page, memcg, false);
3010 return VM_FAULT_OOM;
3014 * The mmap_sem must have been held on entry, and may have been
3015 * released depending on flags and vma->vm_ops->fault() return value.
3016 * See filemap_fault() and __lock_page_retry().
3018 static vm_fault_t __do_fault(struct vm_fault *vmf)
3020 struct vm_area_struct *vma = vmf->vma;
3024 * Preallocate pte before we take page_lock because this might lead to
3025 * deadlocks for memcg reclaim which waits for pages under writeback:
3027 * SetPageWriteback(A)
3033 * wait_on_page_writeback(A)
3034 * SetPageWriteback(B)
3036 * # flush A, B to clear the writeback
3038 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3039 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3040 if (!vmf->prealloc_pte)
3041 return VM_FAULT_OOM;
3042 smp_wmb(); /* See comment in __pte_alloc() */
3045 ret = vma->vm_ops->fault(vmf);
3046 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3047 VM_FAULT_DONE_COW)))
3050 if (unlikely(PageHWPoison(vmf->page))) {
3051 if (ret & VM_FAULT_LOCKED)
3052 unlock_page(vmf->page);
3053 put_page(vmf->page);
3055 return VM_FAULT_HWPOISON;
3058 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3059 lock_page(vmf->page);
3061 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3067 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3068 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3069 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3070 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3072 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3074 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3077 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3079 struct vm_area_struct *vma = vmf->vma;
3081 if (!pmd_none(*vmf->pmd))
3083 if (vmf->prealloc_pte) {
3084 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3085 if (unlikely(!pmd_none(*vmf->pmd))) {
3086 spin_unlock(vmf->ptl);
3090 mm_inc_nr_ptes(vma->vm_mm);
3091 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3092 spin_unlock(vmf->ptl);
3093 vmf->prealloc_pte = NULL;
3094 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3095 return VM_FAULT_OOM;
3099 * If a huge pmd materialized under us just retry later. Use
3100 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3101 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3102 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3103 * running immediately after a huge pmd fault in a different thread of
3104 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3105 * All we have to ensure is that it is a regular pmd that we can walk
3106 * with pte_offset_map() and we can do that through an atomic read in
3107 * C, which is what pmd_trans_unstable() provides.
3109 if (pmd_devmap_trans_unstable(vmf->pmd))
3110 return VM_FAULT_NOPAGE;
3113 * At this point we know that our vmf->pmd points to a page of ptes
3114 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3115 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3116 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3117 * be valid and we will re-check to make sure the vmf->pte isn't
3118 * pte_none() under vmf->ptl protection when we return to
3121 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3126 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3128 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3129 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3130 unsigned long haddr)
3132 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3133 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3135 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3140 static void deposit_prealloc_pte(struct vm_fault *vmf)
3142 struct vm_area_struct *vma = vmf->vma;
3144 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3146 * We are going to consume the prealloc table,
3147 * count that as nr_ptes.
3149 mm_inc_nr_ptes(vma->vm_mm);
3150 vmf->prealloc_pte = NULL;
3153 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3155 struct vm_area_struct *vma = vmf->vma;
3156 bool write = vmf->flags & FAULT_FLAG_WRITE;
3157 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3162 if (!transhuge_vma_suitable(vma, haddr))
3163 return VM_FAULT_FALLBACK;
3165 ret = VM_FAULT_FALLBACK;
3166 page = compound_head(page);
3169 * Archs like ppc64 need additonal space to store information
3170 * related to pte entry. Use the preallocated table for that.
3172 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3173 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3174 if (!vmf->prealloc_pte)
3175 return VM_FAULT_OOM;
3176 smp_wmb(); /* See comment in __pte_alloc() */
3179 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3180 if (unlikely(!pmd_none(*vmf->pmd)))
3183 for (i = 0; i < HPAGE_PMD_NR; i++)
3184 flush_icache_page(vma, page + i);
3186 entry = mk_huge_pmd(page, vma->vm_page_prot);
3188 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3190 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3191 page_add_file_rmap(page, true);
3193 * deposit and withdraw with pmd lock held
3195 if (arch_needs_pgtable_deposit())
3196 deposit_prealloc_pte(vmf);
3198 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3200 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3202 /* fault is handled */
3204 count_vm_event(THP_FILE_MAPPED);
3206 spin_unlock(vmf->ptl);
3210 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3218 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3219 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3221 * @vmf: fault environment
3222 * @memcg: memcg to charge page (only for private mappings)
3223 * @page: page to map
3225 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3228 * Target users are page handler itself and implementations of
3229 * vm_ops->map_pages.
3231 * Return: %0 on success, %VM_FAULT_ code in case of error.
3233 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3236 struct vm_area_struct *vma = vmf->vma;
3237 bool write = vmf->flags & FAULT_FLAG_WRITE;
3241 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3242 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3244 VM_BUG_ON_PAGE(memcg, page);
3246 ret = do_set_pmd(vmf, page);
3247 if (ret != VM_FAULT_FALLBACK)
3252 ret = pte_alloc_one_map(vmf);
3257 /* Re-check under ptl */
3258 if (unlikely(!pte_none(*vmf->pte)))
3259 return VM_FAULT_NOPAGE;
3261 flush_icache_page(vma, page);
3262 entry = mk_pte(page, vma->vm_page_prot);
3264 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3265 /* copy-on-write page */
3266 if (write && !(vma->vm_flags & VM_SHARED)) {
3267 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3268 page_add_new_anon_rmap(page, vma, vmf->address, false);
3269 mem_cgroup_commit_charge(page, memcg, false, false);
3270 lru_cache_add_active_or_unevictable(page, vma);
3272 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3273 page_add_file_rmap(page, false);
3275 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3277 /* no need to invalidate: a not-present page won't be cached */
3278 update_mmu_cache(vma, vmf->address, vmf->pte);
3285 * finish_fault - finish page fault once we have prepared the page to fault
3287 * @vmf: structure describing the fault
3289 * This function handles all that is needed to finish a page fault once the
3290 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3291 * given page, adds reverse page mapping, handles memcg charges and LRU
3294 * The function expects the page to be locked and on success it consumes a
3295 * reference of a page being mapped (for the PTE which maps it).
3297 * Return: %0 on success, %VM_FAULT_ code in case of error.
3299 vm_fault_t finish_fault(struct vm_fault *vmf)
3304 /* Did we COW the page? */
3305 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3306 !(vmf->vma->vm_flags & VM_SHARED))
3307 page = vmf->cow_page;
3312 * check even for read faults because we might have lost our CoWed
3315 if (!(vmf->vma->vm_flags & VM_SHARED))
3316 ret = check_stable_address_space(vmf->vma->vm_mm);
3318 ret = alloc_set_pte(vmf, vmf->memcg, page);
3320 pte_unmap_unlock(vmf->pte, vmf->ptl);
3324 static unsigned long fault_around_bytes __read_mostly =
3325 rounddown_pow_of_two(65536);
3327 #ifdef CONFIG_DEBUG_FS
3328 static int fault_around_bytes_get(void *data, u64 *val)
3330 *val = fault_around_bytes;
3335 * fault_around_bytes must be rounded down to the nearest page order as it's
3336 * what do_fault_around() expects to see.
3338 static int fault_around_bytes_set(void *data, u64 val)
3340 if (val / PAGE_SIZE > PTRS_PER_PTE)
3342 if (val > PAGE_SIZE)
3343 fault_around_bytes = rounddown_pow_of_two(val);
3345 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3348 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3349 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3351 static int __init fault_around_debugfs(void)
3353 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3354 &fault_around_bytes_fops);
3357 late_initcall(fault_around_debugfs);
3361 * do_fault_around() tries to map few pages around the fault address. The hope
3362 * is that the pages will be needed soon and this will lower the number of
3365 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3366 * not ready to be mapped: not up-to-date, locked, etc.
3368 * This function is called with the page table lock taken. In the split ptlock
3369 * case the page table lock only protects only those entries which belong to
3370 * the page table corresponding to the fault address.
3372 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3375 * fault_around_bytes defines how many bytes we'll try to map.
3376 * do_fault_around() expects it to be set to a power of two less than or equal
3379 * The virtual address of the area that we map is naturally aligned to
3380 * fault_around_bytes rounded down to the machine page size
3381 * (and therefore to page order). This way it's easier to guarantee
3382 * that we don't cross page table boundaries.
3384 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3386 unsigned long address = vmf->address, nr_pages, mask;
3387 pgoff_t start_pgoff = vmf->pgoff;
3392 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3393 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3395 vmf->address = max(address & mask, vmf->vma->vm_start);
3396 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3400 * end_pgoff is either the end of the page table, the end of
3401 * the vma or nr_pages from start_pgoff, depending what is nearest.
3403 end_pgoff = start_pgoff -
3404 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3406 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3407 start_pgoff + nr_pages - 1);
3409 if (pmd_none(*vmf->pmd)) {
3410 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3411 if (!vmf->prealloc_pte)
3413 smp_wmb(); /* See comment in __pte_alloc() */
3416 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3418 /* Huge page is mapped? Page fault is solved */
3419 if (pmd_trans_huge(*vmf->pmd)) {
3420 ret = VM_FAULT_NOPAGE;
3424 /* ->map_pages() haven't done anything useful. Cold page cache? */
3428 /* check if the page fault is solved */
3429 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3430 if (!pte_none(*vmf->pte))
3431 ret = VM_FAULT_NOPAGE;
3432 pte_unmap_unlock(vmf->pte, vmf->ptl);
3434 vmf->address = address;
3439 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3441 struct vm_area_struct *vma = vmf->vma;
3445 * Let's call ->map_pages() first and use ->fault() as fallback
3446 * if page by the offset is not ready to be mapped (cold cache or
3449 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3450 ret = do_fault_around(vmf);
3455 ret = __do_fault(vmf);
3456 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3459 ret |= finish_fault(vmf);
3460 unlock_page(vmf->page);
3461 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3462 put_page(vmf->page);
3466 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3468 struct vm_area_struct *vma = vmf->vma;
3471 if (unlikely(anon_vma_prepare(vma)))
3472 return VM_FAULT_OOM;
3474 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3476 return VM_FAULT_OOM;
3478 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3479 &vmf->memcg, false)) {
3480 put_page(vmf->cow_page);
3481 return VM_FAULT_OOM;
3484 ret = __do_fault(vmf);
3485 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3487 if (ret & VM_FAULT_DONE_COW)
3490 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3491 __SetPageUptodate(vmf->cow_page);
3493 ret |= finish_fault(vmf);
3494 unlock_page(vmf->page);
3495 put_page(vmf->page);
3496 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3500 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3501 put_page(vmf->cow_page);
3505 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3507 struct vm_area_struct *vma = vmf->vma;
3508 vm_fault_t ret, tmp;
3510 ret = __do_fault(vmf);
3511 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3515 * Check if the backing address space wants to know that the page is
3516 * about to become writable
3518 if (vma->vm_ops->page_mkwrite) {
3519 unlock_page(vmf->page);
3520 tmp = do_page_mkwrite(vmf);
3521 if (unlikely(!tmp ||
3522 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3523 put_page(vmf->page);
3528 ret |= finish_fault(vmf);
3529 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3531 unlock_page(vmf->page);
3532 put_page(vmf->page);
3536 fault_dirty_shared_page(vma, vmf->page);
3541 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3542 * but allow concurrent faults).
3543 * The mmap_sem may have been released depending on flags and our
3544 * return value. See filemap_fault() and __lock_page_or_retry().
3545 * If mmap_sem is released, vma may become invalid (for example
3546 * by other thread calling munmap()).
3548 static vm_fault_t do_fault(struct vm_fault *vmf)
3550 struct vm_area_struct *vma = vmf->vma;
3551 struct mm_struct *vm_mm = vma->vm_mm;
3555 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3557 if (!vma->vm_ops->fault) {
3559 * If we find a migration pmd entry or a none pmd entry, which
3560 * should never happen, return SIGBUS
3562 if (unlikely(!pmd_present(*vmf->pmd)))
3563 ret = VM_FAULT_SIGBUS;
3565 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3570 * Make sure this is not a temporary clearing of pte
3571 * by holding ptl and checking again. A R/M/W update
3572 * of pte involves: take ptl, clearing the pte so that
3573 * we don't have concurrent modification by hardware
3574 * followed by an update.
3576 if (unlikely(pte_none(*vmf->pte)))
3577 ret = VM_FAULT_SIGBUS;
3579 ret = VM_FAULT_NOPAGE;
3581 pte_unmap_unlock(vmf->pte, vmf->ptl);
3583 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3584 ret = do_read_fault(vmf);
3585 else if (!(vma->vm_flags & VM_SHARED))
3586 ret = do_cow_fault(vmf);
3588 ret = do_shared_fault(vmf);
3590 /* preallocated pagetable is unused: free it */
3591 if (vmf->prealloc_pte) {
3592 pte_free(vm_mm, vmf->prealloc_pte);
3593 vmf->prealloc_pte = NULL;
3598 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3599 unsigned long addr, int page_nid,
3604 count_vm_numa_event(NUMA_HINT_FAULTS);
3605 if (page_nid == numa_node_id()) {
3606 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3607 *flags |= TNF_FAULT_LOCAL;
3610 return mpol_misplaced(page, vma, addr);
3613 static vm_fault_t do_numa_page(struct vm_fault *vmf)
3615 struct vm_area_struct *vma = vmf->vma;
3616 struct page *page = NULL;
3617 int page_nid = NUMA_NO_NODE;
3620 bool migrated = false;
3622 bool was_writable = pte_savedwrite(vmf->orig_pte);
3626 * The "pte" at this point cannot be used safely without
3627 * validation through pte_unmap_same(). It's of NUMA type but
3628 * the pfn may be screwed if the read is non atomic.
3630 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3631 spin_lock(vmf->ptl);
3632 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3633 pte_unmap_unlock(vmf->pte, vmf->ptl);
3638 * Make it present again, Depending on how arch implementes non
3639 * accessible ptes, some can allow access by kernel mode.
3641 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
3642 pte = pte_modify(old_pte, vma->vm_page_prot);
3643 pte = pte_mkyoung(pte);
3645 pte = pte_mkwrite(pte);
3646 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
3647 update_mmu_cache(vma, vmf->address, vmf->pte);
3649 page = vm_normal_page(vma, vmf->address, pte);
3651 pte_unmap_unlock(vmf->pte, vmf->ptl);
3655 /* TODO: handle PTE-mapped THP */
3656 if (PageCompound(page)) {
3657 pte_unmap_unlock(vmf->pte, vmf->ptl);
3662 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3663 * much anyway since they can be in shared cache state. This misses
3664 * the case where a mapping is writable but the process never writes
3665 * to it but pte_write gets cleared during protection updates and
3666 * pte_dirty has unpredictable behaviour between PTE scan updates,
3667 * background writeback, dirty balancing and application behaviour.
3669 if (!pte_write(pte))
3670 flags |= TNF_NO_GROUP;
3673 * Flag if the page is shared between multiple address spaces. This
3674 * is later used when determining whether to group tasks together
3676 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3677 flags |= TNF_SHARED;
3679 last_cpupid = page_cpupid_last(page);
3680 page_nid = page_to_nid(page);
3681 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3683 pte_unmap_unlock(vmf->pte, vmf->ptl);
3684 if (target_nid == NUMA_NO_NODE) {
3689 /* Migrate to the requested node */
3690 migrated = migrate_misplaced_page(page, vma, target_nid);
3692 page_nid = target_nid;
3693 flags |= TNF_MIGRATED;
3695 flags |= TNF_MIGRATE_FAIL;
3698 if (page_nid != NUMA_NO_NODE)
3699 task_numa_fault(last_cpupid, page_nid, 1, flags);
3703 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3705 if (vma_is_anonymous(vmf->vma))
3706 return do_huge_pmd_anonymous_page(vmf);
3707 if (vmf->vma->vm_ops->huge_fault)
3708 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3709 return VM_FAULT_FALLBACK;
3712 /* `inline' is required to avoid gcc 4.1.2 build error */
3713 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3715 if (vma_is_anonymous(vmf->vma))
3716 return do_huge_pmd_wp_page(vmf, orig_pmd);
3717 if (vmf->vma->vm_ops->huge_fault)
3718 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3720 /* COW handled on pte level: split pmd */
3721 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3722 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3724 return VM_FAULT_FALLBACK;
3727 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3729 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3732 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3734 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3735 /* No support for anonymous transparent PUD pages yet */
3736 if (vma_is_anonymous(vmf->vma))
3737 return VM_FAULT_FALLBACK;
3738 if (vmf->vma->vm_ops->huge_fault)
3739 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3740 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3741 return VM_FAULT_FALLBACK;
3744 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3746 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3747 /* No support for anonymous transparent PUD pages yet */
3748 if (vma_is_anonymous(vmf->vma))
3749 return VM_FAULT_FALLBACK;
3750 if (vmf->vma->vm_ops->huge_fault)
3751 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3752 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3753 return VM_FAULT_FALLBACK;
3757 * These routines also need to handle stuff like marking pages dirty
3758 * and/or accessed for architectures that don't do it in hardware (most
3759 * RISC architectures). The early dirtying is also good on the i386.
3761 * There is also a hook called "update_mmu_cache()" that architectures
3762 * with external mmu caches can use to update those (ie the Sparc or
3763 * PowerPC hashed page tables that act as extended TLBs).
3765 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3766 * concurrent faults).
3768 * The mmap_sem may have been released depending on flags and our return value.
3769 * See filemap_fault() and __lock_page_or_retry().
3771 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
3775 if (unlikely(pmd_none(*vmf->pmd))) {
3777 * Leave __pte_alloc() until later: because vm_ops->fault may
3778 * want to allocate huge page, and if we expose page table
3779 * for an instant, it will be difficult to retract from
3780 * concurrent faults and from rmap lookups.
3784 /* See comment in pte_alloc_one_map() */
3785 if (pmd_devmap_trans_unstable(vmf->pmd))
3788 * A regular pmd is established and it can't morph into a huge
3789 * pmd from under us anymore at this point because we hold the
3790 * mmap_sem read mode and khugepaged takes it in write mode.
3791 * So now it's safe to run pte_offset_map().
3793 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3794 vmf->orig_pte = *vmf->pte;
3797 * some architectures can have larger ptes than wordsize,
3798 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3799 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3800 * accesses. The code below just needs a consistent view
3801 * for the ifs and we later double check anyway with the
3802 * ptl lock held. So here a barrier will do.
3805 if (pte_none(vmf->orig_pte)) {
3806 pte_unmap(vmf->pte);
3812 if (vma_is_anonymous(vmf->vma))
3813 return do_anonymous_page(vmf);
3815 return do_fault(vmf);
3818 if (!pte_present(vmf->orig_pte))
3819 return do_swap_page(vmf);
3821 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3822 return do_numa_page(vmf);
3824 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3825 spin_lock(vmf->ptl);
3826 entry = vmf->orig_pte;
3827 if (unlikely(!pte_same(*vmf->pte, entry)))
3829 if (vmf->flags & FAULT_FLAG_WRITE) {
3830 if (!pte_write(entry))
3831 return do_wp_page(vmf);
3832 entry = pte_mkdirty(entry);
3834 entry = pte_mkyoung(entry);
3835 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3836 vmf->flags & FAULT_FLAG_WRITE)) {
3837 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3840 * This is needed only for protection faults but the arch code
3841 * is not yet telling us if this is a protection fault or not.
3842 * This still avoids useless tlb flushes for .text page faults
3845 if (vmf->flags & FAULT_FLAG_WRITE)
3846 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3849 pte_unmap_unlock(vmf->pte, vmf->ptl);
3854 * By the time we get here, we already hold the mm semaphore
3856 * The mmap_sem may have been released depending on flags and our
3857 * return value. See filemap_fault() and __lock_page_or_retry().
3859 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
3860 unsigned long address, unsigned int flags)
3862 struct vm_fault vmf = {
3864 .address = address & PAGE_MASK,
3866 .pgoff = linear_page_index(vma, address),
3867 .gfp_mask = __get_fault_gfp_mask(vma),
3869 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3870 struct mm_struct *mm = vma->vm_mm;
3875 pgd = pgd_offset(mm, address);
3876 p4d = p4d_alloc(mm, pgd, address);
3878 return VM_FAULT_OOM;
3880 vmf.pud = pud_alloc(mm, p4d, address);
3882 return VM_FAULT_OOM;
3883 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
3884 ret = create_huge_pud(&vmf);
3885 if (!(ret & VM_FAULT_FALLBACK))
3888 pud_t orig_pud = *vmf.pud;
3891 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3893 /* NUMA case for anonymous PUDs would go here */
3895 if (dirty && !pud_write(orig_pud)) {
3896 ret = wp_huge_pud(&vmf, orig_pud);
3897 if (!(ret & VM_FAULT_FALLBACK))
3900 huge_pud_set_accessed(&vmf, orig_pud);
3906 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3908 return VM_FAULT_OOM;
3909 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
3910 ret = create_huge_pmd(&vmf);
3911 if (!(ret & VM_FAULT_FALLBACK))
3914 pmd_t orig_pmd = *vmf.pmd;
3917 if (unlikely(is_swap_pmd(orig_pmd))) {
3918 VM_BUG_ON(thp_migration_supported() &&
3919 !is_pmd_migration_entry(orig_pmd));
3920 if (is_pmd_migration_entry(orig_pmd))
3921 pmd_migration_entry_wait(mm, vmf.pmd);
3924 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3925 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3926 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3928 if (dirty && !pmd_write(orig_pmd)) {
3929 ret = wp_huge_pmd(&vmf, orig_pmd);
3930 if (!(ret & VM_FAULT_FALLBACK))
3933 huge_pmd_set_accessed(&vmf, orig_pmd);
3939 return handle_pte_fault(&vmf);
3943 * By the time we get here, we already hold the mm semaphore
3945 * The mmap_sem may have been released depending on flags and our
3946 * return value. See filemap_fault() and __lock_page_or_retry().
3948 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3953 __set_current_state(TASK_RUNNING);
3955 count_vm_event(PGFAULT);
3956 count_memcg_event_mm(vma->vm_mm, PGFAULT);
3958 /* do counter updates before entering really critical section. */
3959 check_sync_rss_stat(current);
3961 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3962 flags & FAULT_FLAG_INSTRUCTION,
3963 flags & FAULT_FLAG_REMOTE))
3964 return VM_FAULT_SIGSEGV;
3967 * Enable the memcg OOM handling for faults triggered in user
3968 * space. Kernel faults are handled more gracefully.
3970 if (flags & FAULT_FLAG_USER)
3971 mem_cgroup_enter_user_fault();
3973 if (unlikely(is_vm_hugetlb_page(vma)))
3974 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3976 ret = __handle_mm_fault(vma, address, flags);
3978 if (flags & FAULT_FLAG_USER) {
3979 mem_cgroup_exit_user_fault();
3981 * The task may have entered a memcg OOM situation but
3982 * if the allocation error was handled gracefully (no
3983 * VM_FAULT_OOM), there is no need to kill anything.
3984 * Just clean up the OOM state peacefully.
3986 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3987 mem_cgroup_oom_synchronize(false);
3992 EXPORT_SYMBOL_GPL(handle_mm_fault);
3994 #ifndef __PAGETABLE_P4D_FOLDED
3996 * Allocate p4d page table.
3997 * We've already handled the fast-path in-line.
3999 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4001 p4d_t *new = p4d_alloc_one(mm, address);
4005 smp_wmb(); /* See comment in __pte_alloc */
4007 spin_lock(&mm->page_table_lock);
4008 if (pgd_present(*pgd)) /* Another has populated it */
4011 pgd_populate(mm, pgd, new);
4012 spin_unlock(&mm->page_table_lock);
4015 #endif /* __PAGETABLE_P4D_FOLDED */
4017 #ifndef __PAGETABLE_PUD_FOLDED
4019 * Allocate page upper directory.
4020 * We've already handled the fast-path in-line.
4022 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4024 pud_t *new = pud_alloc_one(mm, address);
4028 smp_wmb(); /* See comment in __pte_alloc */
4030 spin_lock(&mm->page_table_lock);
4031 #ifndef __ARCH_HAS_5LEVEL_HACK
4032 if (!p4d_present(*p4d)) {
4034 p4d_populate(mm, p4d, new);
4035 } else /* Another has populated it */
4038 if (!pgd_present(*p4d)) {
4040 pgd_populate(mm, p4d, new);
4041 } else /* Another has populated it */
4043 #endif /* __ARCH_HAS_5LEVEL_HACK */
4044 spin_unlock(&mm->page_table_lock);
4047 #endif /* __PAGETABLE_PUD_FOLDED */
4049 #ifndef __PAGETABLE_PMD_FOLDED
4051 * Allocate page middle directory.
4052 * We've already handled the fast-path in-line.
4054 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4057 pmd_t *new = pmd_alloc_one(mm, address);
4061 smp_wmb(); /* See comment in __pte_alloc */
4063 ptl = pud_lock(mm, pud);
4064 #ifndef __ARCH_HAS_4LEVEL_HACK
4065 if (!pud_present(*pud)) {
4067 pud_populate(mm, pud, new);
4068 } else /* Another has populated it */
4071 if (!pgd_present(*pud)) {
4073 pgd_populate(mm, pud, new);
4074 } else /* Another has populated it */
4076 #endif /* __ARCH_HAS_4LEVEL_HACK */
4080 #endif /* __PAGETABLE_PMD_FOLDED */
4082 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4083 struct mmu_notifier_range *range,
4084 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4092 pgd = pgd_offset(mm, address);
4093 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4096 p4d = p4d_offset(pgd, address);
4097 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4100 pud = pud_offset(p4d, address);
4101 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4104 pmd = pmd_offset(pud, address);
4105 VM_BUG_ON(pmd_trans_huge(*pmd));
4107 if (pmd_huge(*pmd)) {
4112 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4113 NULL, mm, address & PMD_MASK,
4114 (address & PMD_MASK) + PMD_SIZE);
4115 mmu_notifier_invalidate_range_start(range);
4117 *ptlp = pmd_lock(mm, pmd);
4118 if (pmd_huge(*pmd)) {
4124 mmu_notifier_invalidate_range_end(range);
4127 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4131 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4132 address & PAGE_MASK,
4133 (address & PAGE_MASK) + PAGE_SIZE);
4134 mmu_notifier_invalidate_range_start(range);
4136 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4137 if (!pte_present(*ptep))
4142 pte_unmap_unlock(ptep, *ptlp);
4144 mmu_notifier_invalidate_range_end(range);
4149 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4150 pte_t **ptepp, spinlock_t **ptlp)
4154 /* (void) is needed to make gcc happy */
4155 (void) __cond_lock(*ptlp,
4156 !(res = __follow_pte_pmd(mm, address, NULL,
4157 ptepp, NULL, ptlp)));
4161 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4162 struct mmu_notifier_range *range,
4163 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4167 /* (void) is needed to make gcc happy */
4168 (void) __cond_lock(*ptlp,
4169 !(res = __follow_pte_pmd(mm, address, range,
4170 ptepp, pmdpp, ptlp)));
4173 EXPORT_SYMBOL(follow_pte_pmd);
4176 * follow_pfn - look up PFN at a user virtual address
4177 * @vma: memory mapping
4178 * @address: user virtual address
4179 * @pfn: location to store found PFN
4181 * Only IO mappings and raw PFN mappings are allowed.
4183 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4185 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4192 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4195 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4198 *pfn = pte_pfn(*ptep);
4199 pte_unmap_unlock(ptep, ptl);
4202 EXPORT_SYMBOL(follow_pfn);
4204 #ifdef CONFIG_HAVE_IOREMAP_PROT
4205 int follow_phys(struct vm_area_struct *vma,
4206 unsigned long address, unsigned int flags,
4207 unsigned long *prot, resource_size_t *phys)
4213 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4216 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4220 if ((flags & FOLL_WRITE) && !pte_write(pte))
4223 *prot = pgprot_val(pte_pgprot(pte));
4224 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4228 pte_unmap_unlock(ptep, ptl);
4233 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4234 void *buf, int len, int write)
4236 resource_size_t phys_addr;
4237 unsigned long prot = 0;
4238 void __iomem *maddr;
4239 int offset = addr & (PAGE_SIZE-1);
4241 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4244 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4249 memcpy_toio(maddr + offset, buf, len);
4251 memcpy_fromio(buf, maddr + offset, len);
4256 EXPORT_SYMBOL_GPL(generic_access_phys);
4260 * Access another process' address space as given in mm. If non-NULL, use the
4261 * given task for page fault accounting.
4263 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4264 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4266 struct vm_area_struct *vma;
4267 void *old_buf = buf;
4268 int write = gup_flags & FOLL_WRITE;
4270 down_read(&mm->mmap_sem);
4271 /* ignore errors, just check how much was successfully transferred */
4273 int bytes, ret, offset;
4275 struct page *page = NULL;
4277 ret = get_user_pages_remote(tsk, mm, addr, 1,
4278 gup_flags, &page, &vma, NULL);
4280 #ifndef CONFIG_HAVE_IOREMAP_PROT
4284 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4285 * we can access using slightly different code.
4287 vma = find_vma(mm, addr);
4288 if (!vma || vma->vm_start > addr)
4290 if (vma->vm_ops && vma->vm_ops->access)
4291 ret = vma->vm_ops->access(vma, addr, buf,
4299 offset = addr & (PAGE_SIZE-1);
4300 if (bytes > PAGE_SIZE-offset)
4301 bytes = PAGE_SIZE-offset;
4305 copy_to_user_page(vma, page, addr,
4306 maddr + offset, buf, bytes);
4307 set_page_dirty_lock(page);
4309 copy_from_user_page(vma, page, addr,
4310 buf, maddr + offset, bytes);
4319 up_read(&mm->mmap_sem);
4321 return buf - old_buf;
4325 * access_remote_vm - access another process' address space
4326 * @mm: the mm_struct of the target address space
4327 * @addr: start address to access
4328 * @buf: source or destination buffer
4329 * @len: number of bytes to transfer
4330 * @gup_flags: flags modifying lookup behaviour
4332 * The caller must hold a reference on @mm.
4334 * Return: number of bytes copied from source to destination.
4336 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4337 void *buf, int len, unsigned int gup_flags)
4339 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4343 * Access another process' address space.
4344 * Source/target buffer must be kernel space,
4345 * Do not walk the page table directly, use get_user_pages
4347 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4348 void *buf, int len, unsigned int gup_flags)
4350 struct mm_struct *mm;
4353 mm = get_task_mm(tsk);
4357 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4363 EXPORT_SYMBOL_GPL(access_process_vm);
4366 * Print the name of a VMA.
4368 void print_vma_addr(char *prefix, unsigned long ip)
4370 struct mm_struct *mm = current->mm;
4371 struct vm_area_struct *vma;
4374 * we might be running from an atomic context so we cannot sleep
4376 if (!down_read_trylock(&mm->mmap_sem))
4379 vma = find_vma(mm, ip);
4380 if (vma && vma->vm_file) {
4381 struct file *f = vma->vm_file;
4382 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4386 p = file_path(f, buf, PAGE_SIZE);
4389 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4391 vma->vm_end - vma->vm_start);
4392 free_page((unsigned long)buf);
4395 up_read(&mm->mmap_sem);
4398 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4399 void __might_fault(const char *file, int line)
4402 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4403 * holding the mmap_sem, this is safe because kernel memory doesn't
4404 * get paged out, therefore we'll never actually fault, and the
4405 * below annotations will generate false positives.
4407 if (uaccess_kernel())
4409 if (pagefault_disabled())
4411 __might_sleep(file, line, 0);
4412 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4414 might_lock_read(¤t->mm->mmap_sem);
4417 EXPORT_SYMBOL(__might_fault);
4420 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4422 * Process all subpages of the specified huge page with the specified
4423 * operation. The target subpage will be processed last to keep its
4426 static inline void process_huge_page(
4427 unsigned long addr_hint, unsigned int pages_per_huge_page,
4428 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4432 unsigned long addr = addr_hint &
4433 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4435 /* Process target subpage last to keep its cache lines hot */
4437 n = (addr_hint - addr) / PAGE_SIZE;
4438 if (2 * n <= pages_per_huge_page) {
4439 /* If target subpage in first half of huge page */
4442 /* Process subpages at the end of huge page */
4443 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4445 process_subpage(addr + i * PAGE_SIZE, i, arg);
4448 /* If target subpage in second half of huge page */
4449 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4450 l = pages_per_huge_page - n;
4451 /* Process subpages at the begin of huge page */
4452 for (i = 0; i < base; i++) {
4454 process_subpage(addr + i * PAGE_SIZE, i, arg);
4458 * Process remaining subpages in left-right-left-right pattern
4459 * towards the target subpage
4461 for (i = 0; i < l; i++) {
4462 int left_idx = base + i;
4463 int right_idx = base + 2 * l - 1 - i;
4466 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4468 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4472 static void clear_gigantic_page(struct page *page,
4474 unsigned int pages_per_huge_page)
4477 struct page *p = page;
4480 for (i = 0; i < pages_per_huge_page;
4481 i++, p = mem_map_next(p, page, i)) {
4483 clear_user_highpage(p, addr + i * PAGE_SIZE);
4487 static void clear_subpage(unsigned long addr, int idx, void *arg)
4489 struct page *page = arg;
4491 clear_user_highpage(page + idx, addr);
4494 void clear_huge_page(struct page *page,
4495 unsigned long addr_hint, unsigned int pages_per_huge_page)
4497 unsigned long addr = addr_hint &
4498 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4500 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4501 clear_gigantic_page(page, addr, pages_per_huge_page);
4505 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4508 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4510 struct vm_area_struct *vma,
4511 unsigned int pages_per_huge_page)
4514 struct page *dst_base = dst;
4515 struct page *src_base = src;
4517 for (i = 0; i < pages_per_huge_page; ) {
4519 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4522 dst = mem_map_next(dst, dst_base, i);
4523 src = mem_map_next(src, src_base, i);
4527 struct copy_subpage_arg {
4530 struct vm_area_struct *vma;
4533 static void copy_subpage(unsigned long addr, int idx, void *arg)
4535 struct copy_subpage_arg *copy_arg = arg;
4537 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4538 addr, copy_arg->vma);
4541 void copy_user_huge_page(struct page *dst, struct page *src,
4542 unsigned long addr_hint, struct vm_area_struct *vma,
4543 unsigned int pages_per_huge_page)
4545 unsigned long addr = addr_hint &
4546 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4547 struct copy_subpage_arg arg = {
4553 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4554 copy_user_gigantic_page(dst, src, addr, vma,
4555 pages_per_huge_page);
4559 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4562 long copy_huge_page_from_user(struct page *dst_page,
4563 const void __user *usr_src,
4564 unsigned int pages_per_huge_page,
4565 bool allow_pagefault)
4567 void *src = (void *)usr_src;
4569 unsigned long i, rc = 0;
4570 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4572 for (i = 0; i < pages_per_huge_page; i++) {
4573 if (allow_pagefault)
4574 page_kaddr = kmap(dst_page + i);
4576 page_kaddr = kmap_atomic(dst_page + i);
4577 rc = copy_from_user(page_kaddr,
4578 (const void __user *)(src + i * PAGE_SIZE),
4580 if (allow_pagefault)
4581 kunmap(dst_page + i);
4583 kunmap_atomic(page_kaddr);
4585 ret_val -= (PAGE_SIZE - rc);
4593 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4595 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4597 static struct kmem_cache *page_ptl_cachep;
4599 void __init ptlock_cache_init(void)
4601 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4605 bool ptlock_alloc(struct page *page)
4609 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4616 void ptlock_free(struct page *page)
4618 kmem_cache_free(page_ptl_cachep, page->ptl);