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/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
62 #include <linux/dma-debug.h>
63 #include <linux/debugfs.h>
64 #include <linux/userfaultfd_k.h>
67 #include <asm/pgalloc.h>
68 #include <asm/uaccess.h>
70 #include <asm/tlbflush.h>
71 #include <asm/pgtable.h>
75 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
76 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
79 #ifndef CONFIG_NEED_MULTIPLE_NODES
80 /* use the per-pgdat data instead for discontigmem - mbligh */
81 unsigned long max_mapnr;
84 EXPORT_SYMBOL(max_mapnr);
85 EXPORT_SYMBOL(mem_map);
89 * A number of key systems in x86 including ioremap() rely on the assumption
90 * that high_memory defines the upper bound on direct map memory, then end
91 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
92 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
97 EXPORT_SYMBOL(high_memory);
100 * Randomize the address space (stacks, mmaps, brk, etc.).
102 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
103 * as ancient (libc5 based) binaries can segfault. )
105 int randomize_va_space __read_mostly =
106 #ifdef CONFIG_COMPAT_BRK
112 static int __init disable_randmaps(char *s)
114 randomize_va_space = 0;
117 __setup("norandmaps", disable_randmaps);
119 unsigned long zero_pfn __read_mostly;
120 unsigned long highest_memmap_pfn __read_mostly;
122 EXPORT_SYMBOL(zero_pfn);
125 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
127 static int __init init_zero_pfn(void)
129 zero_pfn = page_to_pfn(ZERO_PAGE(0));
132 core_initcall(init_zero_pfn);
135 #if defined(SPLIT_RSS_COUNTING)
137 void sync_mm_rss(struct mm_struct *mm)
141 for (i = 0; i < NR_MM_COUNTERS; i++) {
142 if (current->rss_stat.count[i]) {
143 add_mm_counter(mm, i, current->rss_stat.count[i]);
144 current->rss_stat.count[i] = 0;
147 current->rss_stat.events = 0;
150 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
152 struct task_struct *task = current;
154 if (likely(task->mm == mm))
155 task->rss_stat.count[member] += val;
157 add_mm_counter(mm, member, val);
159 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
160 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
162 /* sync counter once per 64 page faults */
163 #define TASK_RSS_EVENTS_THRESH (64)
164 static void check_sync_rss_stat(struct task_struct *task)
166 if (unlikely(task != current))
168 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
169 sync_mm_rss(task->mm);
171 #else /* SPLIT_RSS_COUNTING */
173 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
174 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
176 static void check_sync_rss_stat(struct task_struct *task)
180 #endif /* SPLIT_RSS_COUNTING */
182 #ifdef HAVE_GENERIC_MMU_GATHER
184 static bool tlb_next_batch(struct mmu_gather *tlb)
186 struct mmu_gather_batch *batch;
190 tlb->active = batch->next;
194 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
197 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
204 batch->max = MAX_GATHER_BATCH;
206 tlb->active->next = batch;
213 * Called to initialize an (on-stack) mmu_gather structure for page-table
214 * tear-down from @mm. The @fullmm argument is used when @mm is without
215 * users and we're going to destroy the full address space (exit/execve).
217 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
221 /* Is it from 0 to ~0? */
222 tlb->fullmm = !(start | (end+1));
223 tlb->need_flush_all = 0;
224 tlb->local.next = NULL;
226 tlb->local.max = ARRAY_SIZE(tlb->__pages);
227 tlb->active = &tlb->local;
228 tlb->batch_count = 0;
230 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 __tlb_reset_range(tlb);
237 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
243 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
244 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
245 tlb_table_flush(tlb);
247 __tlb_reset_range(tlb);
250 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
252 struct mmu_gather_batch *batch;
254 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
255 free_pages_and_swap_cache(batch->pages, batch->nr);
258 tlb->active = &tlb->local;
261 void tlb_flush_mmu(struct mmu_gather *tlb)
263 tlb_flush_mmu_tlbonly(tlb);
264 tlb_flush_mmu_free(tlb);
268 * Called at the end of the shootdown operation to free up any resources
269 * that were required.
271 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
273 struct mmu_gather_batch *batch, *next;
277 /* keep the page table cache within bounds */
280 for (batch = tlb->local.next; batch; batch = next) {
282 free_pages((unsigned long)batch, 0);
284 tlb->local.next = NULL;
288 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
289 * handling the additional races in SMP caused by other CPUs caching valid
290 * mappings in their TLBs. Returns the number of free page slots left.
291 * When out of page slots we must call tlb_flush_mmu().
293 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
295 struct mmu_gather_batch *batch;
297 VM_BUG_ON(!tlb->end);
300 batch->pages[batch->nr++] = page;
301 if (batch->nr == batch->max) {
302 if (!tlb_next_batch(tlb))
306 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
308 return batch->max - batch->nr;
311 #endif /* HAVE_GENERIC_MMU_GATHER */
313 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
316 * See the comment near struct mmu_table_batch.
319 static void tlb_remove_table_smp_sync(void *arg)
321 /* Simply deliver the interrupt */
324 static void tlb_remove_table_one(void *table)
327 * This isn't an RCU grace period and hence the page-tables cannot be
328 * assumed to be actually RCU-freed.
330 * It is however sufficient for software page-table walkers that rely on
331 * IRQ disabling. See the comment near struct mmu_table_batch.
333 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
334 __tlb_remove_table(table);
337 static void tlb_remove_table_rcu(struct rcu_head *head)
339 struct mmu_table_batch *batch;
342 batch = container_of(head, struct mmu_table_batch, rcu);
344 for (i = 0; i < batch->nr; i++)
345 __tlb_remove_table(batch->tables[i]);
347 free_page((unsigned long)batch);
350 void tlb_table_flush(struct mmu_gather *tlb)
352 struct mmu_table_batch **batch = &tlb->batch;
355 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
360 void tlb_remove_table(struct mmu_gather *tlb, void *table)
362 struct mmu_table_batch **batch = &tlb->batch;
365 * When there's less then two users of this mm there cannot be a
366 * concurrent page-table walk.
368 if (atomic_read(&tlb->mm->mm_users) < 2) {
369 __tlb_remove_table(table);
373 if (*batch == NULL) {
374 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
375 if (*batch == NULL) {
376 tlb_remove_table_one(table);
381 (*batch)->tables[(*batch)->nr++] = table;
382 if ((*batch)->nr == MAX_TABLE_BATCH)
383 tlb_table_flush(tlb);
386 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
389 * Note: this doesn't free the actual pages themselves. That
390 * has been handled earlier when unmapping all the memory regions.
392 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
395 pgtable_t token = pmd_pgtable(*pmd);
397 pte_free_tlb(tlb, token, addr);
398 atomic_long_dec(&tlb->mm->nr_ptes);
401 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
402 unsigned long addr, unsigned long end,
403 unsigned long floor, unsigned long ceiling)
410 pmd = pmd_offset(pud, addr);
412 next = pmd_addr_end(addr, end);
413 if (pmd_none_or_clear_bad(pmd))
415 free_pte_range(tlb, pmd, addr);
416 } while (pmd++, addr = next, addr != end);
426 if (end - 1 > ceiling - 1)
429 pmd = pmd_offset(pud, start);
431 pmd_free_tlb(tlb, pmd, start);
432 mm_dec_nr_pmds(tlb->mm);
435 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
436 unsigned long addr, unsigned long end,
437 unsigned long floor, unsigned long ceiling)
444 pud = pud_offset(pgd, addr);
446 next = pud_addr_end(addr, end);
447 if (pud_none_or_clear_bad(pud))
449 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
450 } while (pud++, addr = next, addr != end);
456 ceiling &= PGDIR_MASK;
460 if (end - 1 > ceiling - 1)
463 pud = pud_offset(pgd, start);
465 pud_free_tlb(tlb, pud, start);
469 * This function frees user-level page tables of a process.
471 void free_pgd_range(struct mmu_gather *tlb,
472 unsigned long addr, unsigned long end,
473 unsigned long floor, unsigned long ceiling)
479 * The next few lines have given us lots of grief...
481 * Why are we testing PMD* at this top level? Because often
482 * there will be no work to do at all, and we'd prefer not to
483 * go all the way down to the bottom just to discover that.
485 * Why all these "- 1"s? Because 0 represents both the bottom
486 * of the address space and the top of it (using -1 for the
487 * top wouldn't help much: the masks would do the wrong thing).
488 * The rule is that addr 0 and floor 0 refer to the bottom of
489 * the address space, but end 0 and ceiling 0 refer to the top
490 * Comparisons need to use "end - 1" and "ceiling - 1" (though
491 * that end 0 case should be mythical).
493 * Wherever addr is brought up or ceiling brought down, we must
494 * be careful to reject "the opposite 0" before it confuses the
495 * subsequent tests. But what about where end is brought down
496 * by PMD_SIZE below? no, end can't go down to 0 there.
498 * Whereas we round start (addr) and ceiling down, by different
499 * masks at different levels, in order to test whether a table
500 * now has no other vmas using it, so can be freed, we don't
501 * bother to round floor or end up - the tests don't need that.
515 if (end - 1 > ceiling - 1)
520 pgd = pgd_offset(tlb->mm, addr);
522 next = pgd_addr_end(addr, end);
523 if (pgd_none_or_clear_bad(pgd))
525 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
526 } while (pgd++, addr = next, addr != end);
529 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
530 unsigned long floor, unsigned long ceiling)
533 struct vm_area_struct *next = vma->vm_next;
534 unsigned long addr = vma->vm_start;
537 * Hide vma from rmap and truncate_pagecache before freeing
540 unlink_anon_vmas(vma);
541 unlink_file_vma(vma);
543 if (is_vm_hugetlb_page(vma)) {
544 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
545 floor, next? next->vm_start: ceiling);
548 * Optimization: gather nearby vmas into one call down
550 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
551 && !is_vm_hugetlb_page(next)) {
554 unlink_anon_vmas(vma);
555 unlink_file_vma(vma);
557 free_pgd_range(tlb, addr, vma->vm_end,
558 floor, next? next->vm_start: ceiling);
564 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
565 pmd_t *pmd, unsigned long address)
568 pgtable_t new = pte_alloc_one(mm, address);
573 * Ensure all pte setup (eg. pte page lock and page clearing) are
574 * visible before the pte is made visible to other CPUs by being
575 * put into page tables.
577 * The other side of the story is the pointer chasing in the page
578 * table walking code (when walking the page table without locking;
579 * ie. most of the time). Fortunately, these data accesses consist
580 * of a chain of data-dependent loads, meaning most CPUs (alpha
581 * being the notable exception) will already guarantee loads are
582 * seen in-order. See the alpha page table accessors for the
583 * smp_read_barrier_depends() barriers in page table walking code.
585 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
587 ptl = pmd_lock(mm, pmd);
588 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
589 atomic_long_inc(&mm->nr_ptes);
590 pmd_populate(mm, pmd, new);
599 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
601 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
605 smp_wmb(); /* See comment in __pte_alloc */
607 spin_lock(&init_mm.page_table_lock);
608 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
609 pmd_populate_kernel(&init_mm, pmd, new);
612 spin_unlock(&init_mm.page_table_lock);
614 pte_free_kernel(&init_mm, new);
618 static inline void init_rss_vec(int *rss)
620 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
623 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627 if (current->mm == mm)
629 for (i = 0; i < NR_MM_COUNTERS; i++)
631 add_mm_counter(mm, i, rss[i]);
635 * This function is called to print an error when a bad pte
636 * is found. For example, we might have a PFN-mapped pte in
637 * a region that doesn't allow it.
639 * The calling function must still handle the error.
641 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
642 pte_t pte, struct page *page)
644 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
645 pud_t *pud = pud_offset(pgd, addr);
646 pmd_t *pmd = pmd_offset(pud, addr);
647 struct address_space *mapping;
649 static unsigned long resume;
650 static unsigned long nr_shown;
651 static unsigned long nr_unshown;
654 * Allow a burst of 60 reports, then keep quiet for that minute;
655 * or allow a steady drip of one report per second.
657 if (nr_shown == 60) {
658 if (time_before(jiffies, resume)) {
664 "BUG: Bad page map: %lu messages suppressed\n",
671 resume = jiffies + 60 * HZ;
673 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
674 index = linear_page_index(vma, addr);
677 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
679 (long long)pte_val(pte), (long long)pmd_val(*pmd));
681 dump_page(page, "bad pte");
683 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
684 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
686 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
688 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
690 vma->vm_ops ? vma->vm_ops->fault : NULL,
691 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
692 mapping ? mapping->a_ops->readpage : NULL);
694 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
698 * vm_normal_page -- This function gets the "struct page" associated with a pte.
700 * "Special" mappings do not wish to be associated with a "struct page" (either
701 * it doesn't exist, or it exists but they don't want to touch it). In this
702 * case, NULL is returned here. "Normal" mappings do have a struct page.
704 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
705 * pte bit, in which case this function is trivial. Secondly, an architecture
706 * may not have a spare pte bit, which requires a more complicated scheme,
709 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
710 * special mapping (even if there are underlying and valid "struct pages").
711 * COWed pages of a VM_PFNMAP are always normal.
713 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
714 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
715 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
716 * mapping will always honor the rule
718 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
720 * And for normal mappings this is false.
722 * This restricts such mappings to be a linear translation from virtual address
723 * to pfn. To get around this restriction, we allow arbitrary mappings so long
724 * as the vma is not a COW mapping; in that case, we know that all ptes are
725 * special (because none can have been COWed).
728 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
730 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
731 * page" backing, however the difference is that _all_ pages with a struct
732 * page (that is, those where pfn_valid is true) are refcounted and considered
733 * normal pages by the VM. The disadvantage is that pages are refcounted
734 * (which can be slower and simply not an option for some PFNMAP users). The
735 * advantage is that we don't have to follow the strict linearity rule of
736 * PFNMAP mappings in order to support COWable mappings.
739 #ifdef __HAVE_ARCH_PTE_SPECIAL
740 # define HAVE_PTE_SPECIAL 1
742 # define HAVE_PTE_SPECIAL 0
744 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
747 unsigned long pfn = pte_pfn(pte);
749 if (HAVE_PTE_SPECIAL) {
750 if (likely(!pte_special(pte)))
752 if (vma->vm_ops && vma->vm_ops->find_special_page)
753 return vma->vm_ops->find_special_page(vma, addr);
754 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
756 if (!is_zero_pfn(pfn))
757 print_bad_pte(vma, addr, pte, NULL);
761 /* !HAVE_PTE_SPECIAL case follows: */
763 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
764 if (vma->vm_flags & VM_MIXEDMAP) {
770 off = (addr - vma->vm_start) >> PAGE_SHIFT;
771 if (pfn == vma->vm_pgoff + off)
773 if (!is_cow_mapping(vma->vm_flags))
778 if (is_zero_pfn(pfn))
781 if (unlikely(pfn > highest_memmap_pfn)) {
782 print_bad_pte(vma, addr, pte, NULL);
787 * NOTE! We still have PageReserved() pages in the page tables.
788 * eg. VDSO mappings can cause them to exist.
791 return pfn_to_page(pfn);
795 * copy one vm_area from one task to the other. Assumes the page tables
796 * already present in the new task to be cleared in the whole range
797 * covered by this vma.
800 static inline unsigned long
801 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
802 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
803 unsigned long addr, int *rss)
805 unsigned long vm_flags = vma->vm_flags;
806 pte_t pte = *src_pte;
809 /* pte contains position in swap or file, so copy. */
810 if (unlikely(!pte_present(pte))) {
811 swp_entry_t entry = pte_to_swp_entry(pte);
813 if (likely(!non_swap_entry(entry))) {
814 if (swap_duplicate(entry) < 0)
817 /* make sure dst_mm is on swapoff's mmlist. */
818 if (unlikely(list_empty(&dst_mm->mmlist))) {
819 spin_lock(&mmlist_lock);
820 if (list_empty(&dst_mm->mmlist))
821 list_add(&dst_mm->mmlist,
823 spin_unlock(&mmlist_lock);
826 } else if (is_migration_entry(entry)) {
827 page = migration_entry_to_page(entry);
834 if (is_write_migration_entry(entry) &&
835 is_cow_mapping(vm_flags)) {
837 * COW mappings require pages in both
838 * parent and child to be set to read.
840 make_migration_entry_read(&entry);
841 pte = swp_entry_to_pte(entry);
842 if (pte_swp_soft_dirty(*src_pte))
843 pte = pte_swp_mksoft_dirty(pte);
844 set_pte_at(src_mm, addr, src_pte, pte);
851 * If it's a COW mapping, write protect it both
852 * in the parent and the child
854 if (is_cow_mapping(vm_flags)) {
855 ptep_set_wrprotect(src_mm, addr, src_pte);
856 pte = pte_wrprotect(pte);
860 * If it's a shared mapping, mark it clean in
863 if (vm_flags & VM_SHARED)
864 pte = pte_mkclean(pte);
865 pte = pte_mkold(pte);
867 page = vm_normal_page(vma, addr, pte);
870 page_dup_rmap(page, false);
878 set_pte_at(dst_mm, addr, dst_pte, pte);
882 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
883 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
884 unsigned long addr, unsigned long end)
886 pte_t *orig_src_pte, *orig_dst_pte;
887 pte_t *src_pte, *dst_pte;
888 spinlock_t *src_ptl, *dst_ptl;
890 int rss[NR_MM_COUNTERS];
891 swp_entry_t entry = (swp_entry_t){0};
896 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
899 src_pte = pte_offset_map(src_pmd, addr);
900 src_ptl = pte_lockptr(src_mm, src_pmd);
901 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
902 orig_src_pte = src_pte;
903 orig_dst_pte = dst_pte;
904 arch_enter_lazy_mmu_mode();
908 * We are holding two locks at this point - either of them
909 * could generate latencies in another task on another CPU.
911 if (progress >= 32) {
913 if (need_resched() ||
914 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
917 if (pte_none(*src_pte)) {
921 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
926 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
928 arch_leave_lazy_mmu_mode();
929 spin_unlock(src_ptl);
930 pte_unmap(orig_src_pte);
931 add_mm_rss_vec(dst_mm, rss);
932 pte_unmap_unlock(orig_dst_pte, dst_ptl);
936 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
945 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
946 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
947 unsigned long addr, unsigned long end)
949 pmd_t *src_pmd, *dst_pmd;
952 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
955 src_pmd = pmd_offset(src_pud, addr);
957 next = pmd_addr_end(addr, end);
958 if (pmd_trans_huge(*src_pmd)) {
960 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
961 err = copy_huge_pmd(dst_mm, src_mm,
962 dst_pmd, src_pmd, addr, vma);
969 if (pmd_none_or_clear_bad(src_pmd))
971 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
974 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
978 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
979 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
980 unsigned long addr, unsigned long end)
982 pud_t *src_pud, *dst_pud;
985 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
988 src_pud = pud_offset(src_pgd, addr);
990 next = pud_addr_end(addr, end);
991 if (pud_none_or_clear_bad(src_pud))
993 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
996 } while (dst_pud++, src_pud++, addr = next, addr != end);
1000 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1001 struct vm_area_struct *vma)
1003 pgd_t *src_pgd, *dst_pgd;
1005 unsigned long addr = vma->vm_start;
1006 unsigned long end = vma->vm_end;
1007 unsigned long mmun_start; /* For mmu_notifiers */
1008 unsigned long mmun_end; /* For mmu_notifiers */
1013 * Don't copy ptes where a page fault will fill them correctly.
1014 * Fork becomes much lighter when there are big shared or private
1015 * readonly mappings. The tradeoff is that copy_page_range is more
1016 * efficient than faulting.
1018 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1022 if (is_vm_hugetlb_page(vma))
1023 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1025 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1027 * We do not free on error cases below as remove_vma
1028 * gets called on error from higher level routine
1030 ret = track_pfn_copy(vma);
1036 * We need to invalidate the secondary MMU mappings only when
1037 * there could be a permission downgrade on the ptes of the
1038 * parent mm. And a permission downgrade will only happen if
1039 * is_cow_mapping() returns true.
1041 is_cow = is_cow_mapping(vma->vm_flags);
1045 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1049 dst_pgd = pgd_offset(dst_mm, addr);
1050 src_pgd = pgd_offset(src_mm, addr);
1052 next = pgd_addr_end(addr, end);
1053 if (pgd_none_or_clear_bad(src_pgd))
1055 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1056 vma, addr, next))) {
1060 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1063 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1067 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1068 struct vm_area_struct *vma, pmd_t *pmd,
1069 unsigned long addr, unsigned long end,
1070 struct zap_details *details)
1072 struct mm_struct *mm = tlb->mm;
1073 int force_flush = 0;
1074 int rss[NR_MM_COUNTERS];
1082 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1084 arch_enter_lazy_mmu_mode();
1087 if (pte_none(ptent)) {
1091 if (pte_present(ptent)) {
1094 page = vm_normal_page(vma, addr, ptent);
1095 if (unlikely(details) && page) {
1097 * unmap_shared_mapping_pages() wants to
1098 * invalidate cache without truncating:
1099 * unmap shared but keep private pages.
1101 if (details->check_mapping &&
1102 details->check_mapping != page->mapping)
1105 ptent = ptep_get_and_clear_full(mm, addr, pte,
1107 tlb_remove_tlb_entry(tlb, pte, addr);
1108 if (unlikely(!page))
1111 rss[MM_ANONPAGES]--;
1113 if (pte_dirty(ptent)) {
1115 set_page_dirty(page);
1117 if (pte_young(ptent) &&
1118 likely(!(vma->vm_flags & VM_SEQ_READ)))
1119 mark_page_accessed(page);
1120 rss[MM_FILEPAGES]--;
1122 page_remove_rmap(page, false);
1123 if (unlikely(page_mapcount(page) < 0))
1124 print_bad_pte(vma, addr, ptent, page);
1125 if (unlikely(!__tlb_remove_page(tlb, page))) {
1132 /* If details->check_mapping, we leave swap entries. */
1133 if (unlikely(details))
1136 entry = pte_to_swp_entry(ptent);
1137 if (!non_swap_entry(entry))
1139 else if (is_migration_entry(entry)) {
1142 page = migration_entry_to_page(entry);
1145 rss[MM_ANONPAGES]--;
1147 rss[MM_FILEPAGES]--;
1149 if (unlikely(!free_swap_and_cache(entry)))
1150 print_bad_pte(vma, addr, ptent, NULL);
1151 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1152 } while (pte++, addr += PAGE_SIZE, addr != end);
1154 add_mm_rss_vec(mm, rss);
1155 arch_leave_lazy_mmu_mode();
1157 /* Do the actual TLB flush before dropping ptl */
1159 tlb_flush_mmu_tlbonly(tlb);
1160 pte_unmap_unlock(start_pte, ptl);
1163 * If we forced a TLB flush (either due to running out of
1164 * batch buffers or because we needed to flush dirty TLB
1165 * entries before releasing the ptl), free the batched
1166 * memory too. Restart if we didn't do everything.
1170 tlb_flush_mmu_free(tlb);
1179 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1180 struct vm_area_struct *vma, pud_t *pud,
1181 unsigned long addr, unsigned long end,
1182 struct zap_details *details)
1187 pmd = pmd_offset(pud, addr);
1189 next = pmd_addr_end(addr, end);
1190 if (pmd_trans_huge(*pmd)) {
1191 if (next - addr != HPAGE_PMD_SIZE) {
1192 #ifdef CONFIG_DEBUG_VM
1193 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1194 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1195 __func__, addr, end,
1201 split_huge_pmd(vma, pmd, addr);
1202 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1207 * Here there can be other concurrent MADV_DONTNEED or
1208 * trans huge page faults running, and if the pmd is
1209 * none or trans huge it can change under us. This is
1210 * because MADV_DONTNEED holds the mmap_sem in read
1213 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1215 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1218 } while (pmd++, addr = next, addr != end);
1223 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1224 struct vm_area_struct *vma, pgd_t *pgd,
1225 unsigned long addr, unsigned long end,
1226 struct zap_details *details)
1231 pud = pud_offset(pgd, addr);
1233 next = pud_addr_end(addr, end);
1234 if (pud_none_or_clear_bad(pud))
1236 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1237 } while (pud++, addr = next, addr != end);
1242 static void unmap_page_range(struct mmu_gather *tlb,
1243 struct vm_area_struct *vma,
1244 unsigned long addr, unsigned long end,
1245 struct zap_details *details)
1250 if (details && !details->check_mapping)
1253 BUG_ON(addr >= end);
1254 tlb_start_vma(tlb, vma);
1255 pgd = pgd_offset(vma->vm_mm, addr);
1257 next = pgd_addr_end(addr, end);
1258 if (pgd_none_or_clear_bad(pgd))
1260 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1261 } while (pgd++, addr = next, addr != end);
1262 tlb_end_vma(tlb, vma);
1266 static void unmap_single_vma(struct mmu_gather *tlb,
1267 struct vm_area_struct *vma, unsigned long start_addr,
1268 unsigned long end_addr,
1269 struct zap_details *details)
1271 unsigned long start = max(vma->vm_start, start_addr);
1274 if (start >= vma->vm_end)
1276 end = min(vma->vm_end, end_addr);
1277 if (end <= vma->vm_start)
1281 uprobe_munmap(vma, start, end);
1283 if (unlikely(vma->vm_flags & VM_PFNMAP))
1284 untrack_pfn(vma, 0, 0);
1287 if (unlikely(is_vm_hugetlb_page(vma))) {
1289 * It is undesirable to test vma->vm_file as it
1290 * should be non-null for valid hugetlb area.
1291 * However, vm_file will be NULL in the error
1292 * cleanup path of mmap_region. When
1293 * hugetlbfs ->mmap method fails,
1294 * mmap_region() nullifies vma->vm_file
1295 * before calling this function to clean up.
1296 * Since no pte has actually been setup, it is
1297 * safe to do nothing in this case.
1300 i_mmap_lock_write(vma->vm_file->f_mapping);
1301 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1302 i_mmap_unlock_write(vma->vm_file->f_mapping);
1305 unmap_page_range(tlb, vma, start, end, details);
1310 * unmap_vmas - unmap a range of memory covered by a list of vma's
1311 * @tlb: address of the caller's struct mmu_gather
1312 * @vma: the starting vma
1313 * @start_addr: virtual address at which to start unmapping
1314 * @end_addr: virtual address at which to end unmapping
1316 * Unmap all pages in the vma list.
1318 * Only addresses between `start' and `end' will be unmapped.
1320 * The VMA list must be sorted in ascending virtual address order.
1322 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1323 * range after unmap_vmas() returns. So the only responsibility here is to
1324 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1325 * drops the lock and schedules.
1327 void unmap_vmas(struct mmu_gather *tlb,
1328 struct vm_area_struct *vma, unsigned long start_addr,
1329 unsigned long end_addr)
1331 struct mm_struct *mm = vma->vm_mm;
1333 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1334 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1335 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1336 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1340 * zap_page_range - remove user pages in a given range
1341 * @vma: vm_area_struct holding the applicable pages
1342 * @start: starting address of pages to zap
1343 * @size: number of bytes to zap
1344 * @details: details of shared cache invalidation
1346 * Caller must protect the VMA list
1348 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1349 unsigned long size, struct zap_details *details)
1351 struct mm_struct *mm = vma->vm_mm;
1352 struct mmu_gather tlb;
1353 unsigned long end = start + size;
1356 tlb_gather_mmu(&tlb, mm, start, end);
1357 update_hiwater_rss(mm);
1358 mmu_notifier_invalidate_range_start(mm, start, end);
1359 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1360 unmap_single_vma(&tlb, vma, start, end, details);
1361 mmu_notifier_invalidate_range_end(mm, start, end);
1362 tlb_finish_mmu(&tlb, start, end);
1366 * zap_page_range_single - remove user pages in a given range
1367 * @vma: vm_area_struct holding the applicable pages
1368 * @address: starting address of pages to zap
1369 * @size: number of bytes to zap
1370 * @details: details of shared cache invalidation
1372 * The range must fit into one VMA.
1374 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1375 unsigned long size, struct zap_details *details)
1377 struct mm_struct *mm = vma->vm_mm;
1378 struct mmu_gather tlb;
1379 unsigned long end = address + size;
1382 tlb_gather_mmu(&tlb, mm, address, end);
1383 update_hiwater_rss(mm);
1384 mmu_notifier_invalidate_range_start(mm, address, end);
1385 unmap_single_vma(&tlb, vma, address, end, details);
1386 mmu_notifier_invalidate_range_end(mm, address, end);
1387 tlb_finish_mmu(&tlb, address, end);
1391 * zap_vma_ptes - remove ptes mapping the vma
1392 * @vma: vm_area_struct holding ptes to be zapped
1393 * @address: starting address of pages to zap
1394 * @size: number of bytes to zap
1396 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1398 * The entire address range must be fully contained within the vma.
1400 * Returns 0 if successful.
1402 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1405 if (address < vma->vm_start || address + size > vma->vm_end ||
1406 !(vma->vm_flags & VM_PFNMAP))
1408 zap_page_range_single(vma, address, size, NULL);
1411 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1413 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1416 pgd_t * pgd = pgd_offset(mm, addr);
1417 pud_t * pud = pud_alloc(mm, pgd, addr);
1419 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1421 VM_BUG_ON(pmd_trans_huge(*pmd));
1422 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1429 * This is the old fallback for page remapping.
1431 * For historical reasons, it only allows reserved pages. Only
1432 * old drivers should use this, and they needed to mark their
1433 * pages reserved for the old functions anyway.
1435 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1436 struct page *page, pgprot_t prot)
1438 struct mm_struct *mm = vma->vm_mm;
1447 flush_dcache_page(page);
1448 pte = get_locked_pte(mm, addr, &ptl);
1452 if (!pte_none(*pte))
1455 /* Ok, finally just insert the thing.. */
1457 inc_mm_counter_fast(mm, MM_FILEPAGES);
1458 page_add_file_rmap(page);
1459 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1462 pte_unmap_unlock(pte, ptl);
1465 pte_unmap_unlock(pte, ptl);
1471 * vm_insert_page - insert single page into user vma
1472 * @vma: user vma to map to
1473 * @addr: target user address of this page
1474 * @page: source kernel page
1476 * This allows drivers to insert individual pages they've allocated
1479 * The page has to be a nice clean _individual_ kernel allocation.
1480 * If you allocate a compound page, you need to have marked it as
1481 * such (__GFP_COMP), or manually just split the page up yourself
1482 * (see split_page()).
1484 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1485 * took an arbitrary page protection parameter. This doesn't allow
1486 * that. Your vma protection will have to be set up correctly, which
1487 * means that if you want a shared writable mapping, you'd better
1488 * ask for a shared writable mapping!
1490 * The page does not need to be reserved.
1492 * Usually this function is called from f_op->mmap() handler
1493 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1494 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1495 * function from other places, for example from page-fault handler.
1497 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1500 if (addr < vma->vm_start || addr >= vma->vm_end)
1502 if (!page_count(page))
1504 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1505 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1506 BUG_ON(vma->vm_flags & VM_PFNMAP);
1507 vma->vm_flags |= VM_MIXEDMAP;
1509 return insert_page(vma, addr, page, vma->vm_page_prot);
1511 EXPORT_SYMBOL(vm_insert_page);
1513 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1514 unsigned long pfn, pgprot_t prot)
1516 struct mm_struct *mm = vma->vm_mm;
1522 pte = get_locked_pte(mm, addr, &ptl);
1526 if (!pte_none(*pte))
1529 /* Ok, finally just insert the thing.. */
1530 entry = pte_mkspecial(pfn_pte(pfn, prot));
1531 set_pte_at(mm, addr, pte, entry);
1532 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1536 pte_unmap_unlock(pte, ptl);
1542 * vm_insert_pfn - insert single pfn into user vma
1543 * @vma: user vma to map to
1544 * @addr: target user address of this page
1545 * @pfn: source kernel pfn
1547 * Similar to vm_insert_page, this allows drivers to insert individual pages
1548 * they've allocated into a user vma. Same comments apply.
1550 * This function should only be called from a vm_ops->fault handler, and
1551 * in that case the handler should return NULL.
1553 * vma cannot be a COW mapping.
1555 * As this is called only for pages that do not currently exist, we
1556 * do not need to flush old virtual caches or the TLB.
1558 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1562 pgprot_t pgprot = vma->vm_page_prot;
1564 * Technically, architectures with pte_special can avoid all these
1565 * restrictions (same for remap_pfn_range). However we would like
1566 * consistency in testing and feature parity among all, so we should
1567 * try to keep these invariants in place for everybody.
1569 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1570 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1571 (VM_PFNMAP|VM_MIXEDMAP));
1572 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1573 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1575 if (addr < vma->vm_start || addr >= vma->vm_end)
1577 if (track_pfn_insert(vma, &pgprot, pfn))
1580 ret = insert_pfn(vma, addr, pfn, pgprot);
1584 EXPORT_SYMBOL(vm_insert_pfn);
1586 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1589 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1591 if (addr < vma->vm_start || addr >= vma->vm_end)
1595 * If we don't have pte special, then we have to use the pfn_valid()
1596 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1597 * refcount the page if pfn_valid is true (hence insert_page rather
1598 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1599 * without pte special, it would there be refcounted as a normal page.
1601 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1604 page = pfn_to_page(pfn);
1605 return insert_page(vma, addr, page, vma->vm_page_prot);
1607 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1609 EXPORT_SYMBOL(vm_insert_mixed);
1612 * maps a range of physical memory into the requested pages. the old
1613 * mappings are removed. any references to nonexistent pages results
1614 * in null mappings (currently treated as "copy-on-access")
1616 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1617 unsigned long addr, unsigned long end,
1618 unsigned long pfn, pgprot_t prot)
1623 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1626 arch_enter_lazy_mmu_mode();
1628 BUG_ON(!pte_none(*pte));
1629 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1631 } while (pte++, addr += PAGE_SIZE, addr != end);
1632 arch_leave_lazy_mmu_mode();
1633 pte_unmap_unlock(pte - 1, ptl);
1637 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1638 unsigned long addr, unsigned long end,
1639 unsigned long pfn, pgprot_t prot)
1644 pfn -= addr >> PAGE_SHIFT;
1645 pmd = pmd_alloc(mm, pud, addr);
1648 VM_BUG_ON(pmd_trans_huge(*pmd));
1650 next = pmd_addr_end(addr, end);
1651 if (remap_pte_range(mm, pmd, addr, next,
1652 pfn + (addr >> PAGE_SHIFT), prot))
1654 } while (pmd++, addr = next, addr != end);
1658 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1659 unsigned long addr, unsigned long end,
1660 unsigned long pfn, pgprot_t prot)
1665 pfn -= addr >> PAGE_SHIFT;
1666 pud = pud_alloc(mm, pgd, addr);
1670 next = pud_addr_end(addr, end);
1671 if (remap_pmd_range(mm, pud, addr, next,
1672 pfn + (addr >> PAGE_SHIFT), prot))
1674 } while (pud++, addr = next, addr != end);
1679 * remap_pfn_range - remap kernel memory to userspace
1680 * @vma: user vma to map to
1681 * @addr: target user address to start at
1682 * @pfn: physical address of kernel memory
1683 * @size: size of map area
1684 * @prot: page protection flags for this mapping
1686 * Note: this is only safe if the mm semaphore is held when called.
1688 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1689 unsigned long pfn, unsigned long size, pgprot_t prot)
1693 unsigned long end = addr + PAGE_ALIGN(size);
1694 struct mm_struct *mm = vma->vm_mm;
1698 * Physically remapped pages are special. Tell the
1699 * rest of the world about it:
1700 * VM_IO tells people not to look at these pages
1701 * (accesses can have side effects).
1702 * VM_PFNMAP tells the core MM that the base pages are just
1703 * raw PFN mappings, and do not have a "struct page" associated
1706 * Disable vma merging and expanding with mremap().
1708 * Omit vma from core dump, even when VM_IO turned off.
1710 * There's a horrible special case to handle copy-on-write
1711 * behaviour that some programs depend on. We mark the "original"
1712 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1713 * See vm_normal_page() for details.
1715 if (is_cow_mapping(vma->vm_flags)) {
1716 if (addr != vma->vm_start || end != vma->vm_end)
1718 vma->vm_pgoff = pfn;
1721 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1725 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1727 BUG_ON(addr >= end);
1728 pfn -= addr >> PAGE_SHIFT;
1729 pgd = pgd_offset(mm, addr);
1730 flush_cache_range(vma, addr, end);
1732 next = pgd_addr_end(addr, end);
1733 err = remap_pud_range(mm, pgd, addr, next,
1734 pfn + (addr >> PAGE_SHIFT), prot);
1737 } while (pgd++, addr = next, addr != end);
1740 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1744 EXPORT_SYMBOL(remap_pfn_range);
1747 * vm_iomap_memory - remap memory to userspace
1748 * @vma: user vma to map to
1749 * @start: start of area
1750 * @len: size of area
1752 * This is a simplified io_remap_pfn_range() for common driver use. The
1753 * driver just needs to give us the physical memory range to be mapped,
1754 * we'll figure out the rest from the vma information.
1756 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1757 * whatever write-combining details or similar.
1759 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1761 unsigned long vm_len, pfn, pages;
1763 /* Check that the physical memory area passed in looks valid */
1764 if (start + len < start)
1767 * You *really* shouldn't map things that aren't page-aligned,
1768 * but we've historically allowed it because IO memory might
1769 * just have smaller alignment.
1771 len += start & ~PAGE_MASK;
1772 pfn = start >> PAGE_SHIFT;
1773 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1774 if (pfn + pages < pfn)
1777 /* We start the mapping 'vm_pgoff' pages into the area */
1778 if (vma->vm_pgoff > pages)
1780 pfn += vma->vm_pgoff;
1781 pages -= vma->vm_pgoff;
1783 /* Can we fit all of the mapping? */
1784 vm_len = vma->vm_end - vma->vm_start;
1785 if (vm_len >> PAGE_SHIFT > pages)
1788 /* Ok, let it rip */
1789 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1791 EXPORT_SYMBOL(vm_iomap_memory);
1793 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1794 unsigned long addr, unsigned long end,
1795 pte_fn_t fn, void *data)
1800 spinlock_t *uninitialized_var(ptl);
1802 pte = (mm == &init_mm) ?
1803 pte_alloc_kernel(pmd, addr) :
1804 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1808 BUG_ON(pmd_huge(*pmd));
1810 arch_enter_lazy_mmu_mode();
1812 token = pmd_pgtable(*pmd);
1815 err = fn(pte++, token, addr, data);
1818 } while (addr += PAGE_SIZE, addr != end);
1820 arch_leave_lazy_mmu_mode();
1823 pte_unmap_unlock(pte-1, ptl);
1827 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1828 unsigned long addr, unsigned long end,
1829 pte_fn_t fn, void *data)
1835 BUG_ON(pud_huge(*pud));
1837 pmd = pmd_alloc(mm, pud, addr);
1841 next = pmd_addr_end(addr, end);
1842 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1845 } while (pmd++, addr = next, addr != end);
1849 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1850 unsigned long addr, unsigned long end,
1851 pte_fn_t fn, void *data)
1857 pud = pud_alloc(mm, pgd, addr);
1861 next = pud_addr_end(addr, end);
1862 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1865 } while (pud++, addr = next, addr != end);
1870 * Scan a region of virtual memory, filling in page tables as necessary
1871 * and calling a provided function on each leaf page table.
1873 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1874 unsigned long size, pte_fn_t fn, void *data)
1878 unsigned long end = addr + size;
1881 BUG_ON(addr >= end);
1882 pgd = pgd_offset(mm, addr);
1884 next = pgd_addr_end(addr, end);
1885 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1888 } while (pgd++, addr = next, addr != end);
1892 EXPORT_SYMBOL_GPL(apply_to_page_range);
1895 * handle_pte_fault chooses page fault handler according to an entry which was
1896 * read non-atomically. Before making any commitment, on those architectures
1897 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1898 * parts, do_swap_page must check under lock before unmapping the pte and
1899 * proceeding (but do_wp_page is only called after already making such a check;
1900 * and do_anonymous_page can safely check later on).
1902 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1903 pte_t *page_table, pte_t orig_pte)
1906 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1907 if (sizeof(pte_t) > sizeof(unsigned long)) {
1908 spinlock_t *ptl = pte_lockptr(mm, pmd);
1910 same = pte_same(*page_table, orig_pte);
1914 pte_unmap(page_table);
1918 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1920 debug_dma_assert_idle(src);
1923 * If the source page was a PFN mapping, we don't have
1924 * a "struct page" for it. We do a best-effort copy by
1925 * just copying from the original user address. If that
1926 * fails, we just zero-fill it. Live with it.
1928 if (unlikely(!src)) {
1929 void *kaddr = kmap_atomic(dst);
1930 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1933 * This really shouldn't fail, because the page is there
1934 * in the page tables. But it might just be unreadable,
1935 * in which case we just give up and fill the result with
1938 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1940 kunmap_atomic(kaddr);
1941 flush_dcache_page(dst);
1943 copy_user_highpage(dst, src, va, vma);
1947 * Notify the address space that the page is about to become writable so that
1948 * it can prohibit this or wait for the page to get into an appropriate state.
1950 * We do this without the lock held, so that it can sleep if it needs to.
1952 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1953 unsigned long address)
1955 struct vm_fault vmf;
1958 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1959 vmf.pgoff = page->index;
1960 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1962 vmf.cow_page = NULL;
1964 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
1965 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
1967 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
1969 if (!page->mapping) {
1971 return 0; /* retry */
1973 ret |= VM_FAULT_LOCKED;
1975 VM_BUG_ON_PAGE(!PageLocked(page), page);
1980 * Handle write page faults for pages that can be reused in the current vma
1982 * This can happen either due to the mapping being with the VM_SHARED flag,
1983 * or due to us being the last reference standing to the page. In either
1984 * case, all we need to do here is to mark the page as writable and update
1985 * any related book-keeping.
1987 static inline int wp_page_reuse(struct mm_struct *mm,
1988 struct vm_area_struct *vma, unsigned long address,
1989 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
1990 struct page *page, int page_mkwrite,
1996 * Clear the pages cpupid information as the existing
1997 * information potentially belongs to a now completely
1998 * unrelated process.
2001 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2003 flush_cache_page(vma, address, pte_pfn(orig_pte));
2004 entry = pte_mkyoung(orig_pte);
2005 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2006 if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2007 update_mmu_cache(vma, address, page_table);
2008 pte_unmap_unlock(page_table, ptl);
2011 struct address_space *mapping;
2017 dirtied = set_page_dirty(page);
2018 VM_BUG_ON_PAGE(PageAnon(page), page);
2019 mapping = page->mapping;
2021 page_cache_release(page);
2023 if ((dirtied || page_mkwrite) && mapping) {
2025 * Some device drivers do not set page.mapping
2026 * but still dirty their pages
2028 balance_dirty_pages_ratelimited(mapping);
2032 file_update_time(vma->vm_file);
2035 return VM_FAULT_WRITE;
2039 * Handle the case of a page which we actually need to copy to a new page.
2041 * Called with mmap_sem locked and the old page referenced, but
2042 * without the ptl held.
2044 * High level logic flow:
2046 * - Allocate a page, copy the content of the old page to the new one.
2047 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2048 * - Take the PTL. If the pte changed, bail out and release the allocated page
2049 * - If the pte is still the way we remember it, update the page table and all
2050 * relevant references. This includes dropping the reference the page-table
2051 * held to the old page, as well as updating the rmap.
2052 * - In any case, unlock the PTL and drop the reference we took to the old page.
2054 static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2055 unsigned long address, pte_t *page_table, pmd_t *pmd,
2056 pte_t orig_pte, struct page *old_page)
2058 struct page *new_page = NULL;
2059 spinlock_t *ptl = NULL;
2061 int page_copied = 0;
2062 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */
2063 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */
2064 struct mem_cgroup *memcg;
2066 if (unlikely(anon_vma_prepare(vma)))
2069 if (is_zero_pfn(pte_pfn(orig_pte))) {
2070 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2074 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2077 cow_user_page(new_page, old_page, address, vma);
2080 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2083 __SetPageUptodate(new_page);
2085 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2088 * Re-check the pte - we dropped the lock
2090 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2091 if (likely(pte_same(*page_table, orig_pte))) {
2093 if (!PageAnon(old_page)) {
2094 dec_mm_counter_fast(mm, MM_FILEPAGES);
2095 inc_mm_counter_fast(mm, MM_ANONPAGES);
2098 inc_mm_counter_fast(mm, MM_ANONPAGES);
2100 flush_cache_page(vma, address, pte_pfn(orig_pte));
2101 entry = mk_pte(new_page, vma->vm_page_prot);
2102 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2104 * Clear the pte entry and flush it first, before updating the
2105 * pte with the new entry. This will avoid a race condition
2106 * seen in the presence of one thread doing SMC and another
2109 ptep_clear_flush_notify(vma, address, page_table);
2110 page_add_new_anon_rmap(new_page, vma, address, false);
2111 mem_cgroup_commit_charge(new_page, memcg, false, false);
2112 lru_cache_add_active_or_unevictable(new_page, vma);
2114 * We call the notify macro here because, when using secondary
2115 * mmu page tables (such as kvm shadow page tables), we want the
2116 * new page to be mapped directly into the secondary page table.
2118 set_pte_at_notify(mm, address, page_table, entry);
2119 update_mmu_cache(vma, address, page_table);
2122 * Only after switching the pte to the new page may
2123 * we remove the mapcount here. Otherwise another
2124 * process may come and find the rmap count decremented
2125 * before the pte is switched to the new page, and
2126 * "reuse" the old page writing into it while our pte
2127 * here still points into it and can be read by other
2130 * The critical issue is to order this
2131 * page_remove_rmap with the ptp_clear_flush above.
2132 * Those stores are ordered by (if nothing else,)
2133 * the barrier present in the atomic_add_negative
2134 * in page_remove_rmap.
2136 * Then the TLB flush in ptep_clear_flush ensures that
2137 * no process can access the old page before the
2138 * decremented mapcount is visible. And the old page
2139 * cannot be reused until after the decremented
2140 * mapcount is visible. So transitively, TLBs to
2141 * old page will be flushed before it can be reused.
2143 page_remove_rmap(old_page, false);
2146 /* Free the old page.. */
2147 new_page = old_page;
2150 mem_cgroup_cancel_charge(new_page, memcg, false);
2154 page_cache_release(new_page);
2156 pte_unmap_unlock(page_table, ptl);
2157 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2160 * Don't let another task, with possibly unlocked vma,
2161 * keep the mlocked page.
2163 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2164 lock_page(old_page); /* LRU manipulation */
2165 if (PageMlocked(old_page))
2166 munlock_vma_page(old_page);
2167 unlock_page(old_page);
2169 page_cache_release(old_page);
2171 return page_copied ? VM_FAULT_WRITE : 0;
2173 page_cache_release(new_page);
2176 page_cache_release(old_page);
2177 return VM_FAULT_OOM;
2181 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2184 static int wp_pfn_shared(struct mm_struct *mm,
2185 struct vm_area_struct *vma, unsigned long address,
2186 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2189 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2190 struct vm_fault vmf = {
2192 .pgoff = linear_page_index(vma, address),
2193 .virtual_address = (void __user *)(address & PAGE_MASK),
2194 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2198 pte_unmap_unlock(page_table, ptl);
2199 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2200 if (ret & VM_FAULT_ERROR)
2202 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2204 * We might have raced with another page fault while we
2205 * released the pte_offset_map_lock.
2207 if (!pte_same(*page_table, orig_pte)) {
2208 pte_unmap_unlock(page_table, ptl);
2212 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2216 static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2217 unsigned long address, pte_t *page_table,
2218 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2219 struct page *old_page)
2222 int page_mkwrite = 0;
2224 page_cache_get(old_page);
2227 * Only catch write-faults on shared writable pages,
2228 * read-only shared pages can get COWed by
2229 * get_user_pages(.write=1, .force=1).
2231 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2234 pte_unmap_unlock(page_table, ptl);
2235 tmp = do_page_mkwrite(vma, old_page, address);
2236 if (unlikely(!tmp || (tmp &
2237 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2238 page_cache_release(old_page);
2242 * Since we dropped the lock we need to revalidate
2243 * the PTE as someone else may have changed it. If
2244 * they did, we just return, as we can count on the
2245 * MMU to tell us if they didn't also make it writable.
2247 page_table = pte_offset_map_lock(mm, pmd, address,
2249 if (!pte_same(*page_table, orig_pte)) {
2250 unlock_page(old_page);
2251 pte_unmap_unlock(page_table, ptl);
2252 page_cache_release(old_page);
2258 return wp_page_reuse(mm, vma, address, page_table, ptl,
2259 orig_pte, old_page, page_mkwrite, 1);
2263 * This routine handles present pages, when users try to write
2264 * to a shared page. It is done by copying the page to a new address
2265 * and decrementing the shared-page counter for the old page.
2267 * Note that this routine assumes that the protection checks have been
2268 * done by the caller (the low-level page fault routine in most cases).
2269 * Thus we can safely just mark it writable once we've done any necessary
2272 * We also mark the page dirty at this point even though the page will
2273 * change only once the write actually happens. This avoids a few races,
2274 * and potentially makes it more efficient.
2276 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2277 * but allow concurrent faults), with pte both mapped and locked.
2278 * We return with mmap_sem still held, but pte unmapped and unlocked.
2280 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2281 unsigned long address, pte_t *page_table, pmd_t *pmd,
2282 spinlock_t *ptl, pte_t orig_pte)
2285 struct page *old_page;
2287 old_page = vm_normal_page(vma, address, orig_pte);
2290 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2293 * We should not cow pages in a shared writeable mapping.
2294 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2296 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2297 (VM_WRITE|VM_SHARED))
2298 return wp_pfn_shared(mm, vma, address, page_table, ptl,
2301 pte_unmap_unlock(page_table, ptl);
2302 return wp_page_copy(mm, vma, address, page_table, pmd,
2303 orig_pte, old_page);
2307 * Take out anonymous pages first, anonymous shared vmas are
2308 * not dirty accountable.
2310 if (PageAnon(old_page) && !PageKsm(old_page)) {
2311 if (!trylock_page(old_page)) {
2312 page_cache_get(old_page);
2313 pte_unmap_unlock(page_table, ptl);
2314 lock_page(old_page);
2315 page_table = pte_offset_map_lock(mm, pmd, address,
2317 if (!pte_same(*page_table, orig_pte)) {
2318 unlock_page(old_page);
2319 pte_unmap_unlock(page_table, ptl);
2320 page_cache_release(old_page);
2323 page_cache_release(old_page);
2325 if (reuse_swap_page(old_page)) {
2327 * The page is all ours. Move it to our anon_vma so
2328 * the rmap code will not search our parent or siblings.
2329 * Protected against the rmap code by the page lock.
2331 page_move_anon_rmap(old_page, vma, address);
2332 unlock_page(old_page);
2333 return wp_page_reuse(mm, vma, address, page_table, ptl,
2334 orig_pte, old_page, 0, 0);
2336 unlock_page(old_page);
2337 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2338 (VM_WRITE|VM_SHARED))) {
2339 return wp_page_shared(mm, vma, address, page_table, pmd,
2340 ptl, orig_pte, old_page);
2344 * Ok, we need to copy. Oh, well..
2346 page_cache_get(old_page);
2348 pte_unmap_unlock(page_table, ptl);
2349 return wp_page_copy(mm, vma, address, page_table, pmd,
2350 orig_pte, old_page);
2353 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2354 unsigned long start_addr, unsigned long end_addr,
2355 struct zap_details *details)
2357 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2360 static inline void unmap_mapping_range_tree(struct rb_root *root,
2361 struct zap_details *details)
2363 struct vm_area_struct *vma;
2364 pgoff_t vba, vea, zba, zea;
2366 vma_interval_tree_foreach(vma, root,
2367 details->first_index, details->last_index) {
2369 vba = vma->vm_pgoff;
2370 vea = vba + vma_pages(vma) - 1;
2371 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2372 zba = details->first_index;
2375 zea = details->last_index;
2379 unmap_mapping_range_vma(vma,
2380 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2381 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2387 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2388 * address_space corresponding to the specified page range in the underlying
2391 * @mapping: the address space containing mmaps to be unmapped.
2392 * @holebegin: byte in first page to unmap, relative to the start of
2393 * the underlying file. This will be rounded down to a PAGE_SIZE
2394 * boundary. Note that this is different from truncate_pagecache(), which
2395 * must keep the partial page. In contrast, we must get rid of
2397 * @holelen: size of prospective hole in bytes. This will be rounded
2398 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2400 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2401 * but 0 when invalidating pagecache, don't throw away private data.
2403 void unmap_mapping_range(struct address_space *mapping,
2404 loff_t const holebegin, loff_t const holelen, int even_cows)
2406 struct zap_details details;
2407 pgoff_t hba = holebegin >> PAGE_SHIFT;
2408 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2410 /* Check for overflow. */
2411 if (sizeof(holelen) > sizeof(hlen)) {
2413 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2414 if (holeend & ~(long long)ULONG_MAX)
2415 hlen = ULONG_MAX - hba + 1;
2418 details.check_mapping = even_cows? NULL: mapping;
2419 details.first_index = hba;
2420 details.last_index = hba + hlen - 1;
2421 if (details.last_index < details.first_index)
2422 details.last_index = ULONG_MAX;
2425 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2426 i_mmap_lock_write(mapping);
2427 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2428 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2429 i_mmap_unlock_write(mapping);
2431 EXPORT_SYMBOL(unmap_mapping_range);
2434 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2435 * but allow concurrent faults), and pte mapped but not yet locked.
2436 * We return with pte unmapped and unlocked.
2438 * We return with the mmap_sem locked or unlocked in the same cases
2439 * as does filemap_fault().
2441 int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2442 unsigned long address, pte_t *page_table, pmd_t *pmd,
2443 unsigned int flags, pte_t orig_pte)
2446 struct page *page, *swapcache;
2447 struct mem_cgroup *memcg;
2454 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2457 entry = pte_to_swp_entry(orig_pte);
2458 if (unlikely(non_swap_entry(entry))) {
2459 if (is_migration_entry(entry)) {
2460 migration_entry_wait(mm, pmd, address);
2461 } else if (is_hwpoison_entry(entry)) {
2462 ret = VM_FAULT_HWPOISON;
2464 print_bad_pte(vma, address, orig_pte, NULL);
2465 ret = VM_FAULT_SIGBUS;
2469 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2470 page = lookup_swap_cache(entry);
2472 page = swapin_readahead(entry,
2473 GFP_HIGHUSER_MOVABLE, vma, address);
2476 * Back out if somebody else faulted in this pte
2477 * while we released the pte lock.
2479 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2480 if (likely(pte_same(*page_table, orig_pte)))
2482 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2486 /* Had to read the page from swap area: Major fault */
2487 ret = VM_FAULT_MAJOR;
2488 count_vm_event(PGMAJFAULT);
2489 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2490 } else if (PageHWPoison(page)) {
2492 * hwpoisoned dirty swapcache pages are kept for killing
2493 * owner processes (which may be unknown at hwpoison time)
2495 ret = VM_FAULT_HWPOISON;
2496 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2502 locked = lock_page_or_retry(page, mm, flags);
2504 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2506 ret |= VM_FAULT_RETRY;
2511 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2512 * release the swapcache from under us. The page pin, and pte_same
2513 * test below, are not enough to exclude that. Even if it is still
2514 * swapcache, we need to check that the page's swap has not changed.
2516 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2519 page = ksm_might_need_to_copy(page, vma, address);
2520 if (unlikely(!page)) {
2526 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) {
2532 * Back out if somebody else already faulted in this pte.
2534 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2535 if (unlikely(!pte_same(*page_table, orig_pte)))
2538 if (unlikely(!PageUptodate(page))) {
2539 ret = VM_FAULT_SIGBUS;
2544 * The page isn't present yet, go ahead with the fault.
2546 * Be careful about the sequence of operations here.
2547 * To get its accounting right, reuse_swap_page() must be called
2548 * while the page is counted on swap but not yet in mapcount i.e.
2549 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2550 * must be called after the swap_free(), or it will never succeed.
2553 inc_mm_counter_fast(mm, MM_ANONPAGES);
2554 dec_mm_counter_fast(mm, MM_SWAPENTS);
2555 pte = mk_pte(page, vma->vm_page_prot);
2556 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2557 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2558 flags &= ~FAULT_FLAG_WRITE;
2559 ret |= VM_FAULT_WRITE;
2560 exclusive = RMAP_EXCLUSIVE;
2562 flush_icache_page(vma, page);
2563 if (pte_swp_soft_dirty(orig_pte))
2564 pte = pte_mksoft_dirty(pte);
2565 set_pte_at(mm, address, page_table, pte);
2566 if (page == swapcache) {
2567 do_page_add_anon_rmap(page, vma, address, exclusive);
2568 mem_cgroup_commit_charge(page, memcg, true, false);
2569 } else { /* ksm created a completely new copy */
2570 page_add_new_anon_rmap(page, vma, address, false);
2571 mem_cgroup_commit_charge(page, memcg, false, false);
2572 lru_cache_add_active_or_unevictable(page, vma);
2576 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2577 try_to_free_swap(page);
2579 if (page != swapcache) {
2581 * Hold the lock to avoid the swap entry to be reused
2582 * until we take the PT lock for the pte_same() check
2583 * (to avoid false positives from pte_same). For
2584 * further safety release the lock after the swap_free
2585 * so that the swap count won't change under a
2586 * parallel locked swapcache.
2588 unlock_page(swapcache);
2589 page_cache_release(swapcache);
2592 if (flags & FAULT_FLAG_WRITE) {
2593 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2594 if (ret & VM_FAULT_ERROR)
2595 ret &= VM_FAULT_ERROR;
2599 /* No need to invalidate - it was non-present before */
2600 update_mmu_cache(vma, address, page_table);
2602 pte_unmap_unlock(page_table, ptl);
2606 mem_cgroup_cancel_charge(page, memcg, false);
2607 pte_unmap_unlock(page_table, ptl);
2611 page_cache_release(page);
2612 if (page != swapcache) {
2613 unlock_page(swapcache);
2614 page_cache_release(swapcache);
2620 * This is like a special single-page "expand_{down|up}wards()",
2621 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2622 * doesn't hit another vma.
2624 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2626 address &= PAGE_MASK;
2627 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2628 struct vm_area_struct *prev = vma->vm_prev;
2631 * Is there a mapping abutting this one below?
2633 * That's only ok if it's the same stack mapping
2634 * that has gotten split..
2636 if (prev && prev->vm_end == address)
2637 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2639 return expand_downwards(vma, address - PAGE_SIZE);
2641 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2642 struct vm_area_struct *next = vma->vm_next;
2644 /* As VM_GROWSDOWN but s/below/above/ */
2645 if (next && next->vm_start == address + PAGE_SIZE)
2646 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2648 return expand_upwards(vma, address + PAGE_SIZE);
2654 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2655 * but allow concurrent faults), and pte mapped but not yet locked.
2656 * We return with mmap_sem still held, but pte unmapped and unlocked.
2658 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2659 unsigned long address, pte_t *page_table, pmd_t *pmd,
2662 struct mem_cgroup *memcg;
2667 pte_unmap(page_table);
2669 /* File mapping without ->vm_ops ? */
2670 if (vma->vm_flags & VM_SHARED)
2671 return VM_FAULT_SIGBUS;
2673 /* Check if we need to add a guard page to the stack */
2674 if (check_stack_guard_page(vma, address) < 0)
2675 return VM_FAULT_SIGSEGV;
2677 /* Use the zero-page for reads */
2678 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2679 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2680 vma->vm_page_prot));
2681 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2682 if (!pte_none(*page_table))
2684 /* Deliver the page fault to userland, check inside PT lock */
2685 if (userfaultfd_missing(vma)) {
2686 pte_unmap_unlock(page_table, ptl);
2687 return handle_userfault(vma, address, flags,
2693 /* Allocate our own private page. */
2694 if (unlikely(anon_vma_prepare(vma)))
2696 page = alloc_zeroed_user_highpage_movable(vma, address);
2700 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false))
2704 * The memory barrier inside __SetPageUptodate makes sure that
2705 * preceeding stores to the page contents become visible before
2706 * the set_pte_at() write.
2708 __SetPageUptodate(page);
2710 entry = mk_pte(page, vma->vm_page_prot);
2711 if (vma->vm_flags & VM_WRITE)
2712 entry = pte_mkwrite(pte_mkdirty(entry));
2714 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2715 if (!pte_none(*page_table))
2718 /* Deliver the page fault to userland, check inside PT lock */
2719 if (userfaultfd_missing(vma)) {
2720 pte_unmap_unlock(page_table, ptl);
2721 mem_cgroup_cancel_charge(page, memcg, false);
2722 page_cache_release(page);
2723 return handle_userfault(vma, address, flags,
2727 inc_mm_counter_fast(mm, MM_ANONPAGES);
2728 page_add_new_anon_rmap(page, vma, address, false);
2729 mem_cgroup_commit_charge(page, memcg, false, false);
2730 lru_cache_add_active_or_unevictable(page, vma);
2732 set_pte_at(mm, address, page_table, entry);
2734 /* No need to invalidate - it was non-present before */
2735 update_mmu_cache(vma, address, page_table);
2737 pte_unmap_unlock(page_table, ptl);
2740 mem_cgroup_cancel_charge(page, memcg, false);
2741 page_cache_release(page);
2744 page_cache_release(page);
2746 return VM_FAULT_OOM;
2750 * The mmap_sem must have been held on entry, and may have been
2751 * released depending on flags and vma->vm_ops->fault() return value.
2752 * See filemap_fault() and __lock_page_retry().
2754 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2755 pgoff_t pgoff, unsigned int flags,
2756 struct page *cow_page, struct page **page)
2758 struct vm_fault vmf;
2761 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2765 vmf.cow_page = cow_page;
2767 ret = vma->vm_ops->fault(vma, &vmf);
2768 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2773 if (unlikely(PageHWPoison(vmf.page))) {
2774 if (ret & VM_FAULT_LOCKED)
2775 unlock_page(vmf.page);
2776 page_cache_release(vmf.page);
2777 return VM_FAULT_HWPOISON;
2780 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2781 lock_page(vmf.page);
2783 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2791 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2793 * @vma: virtual memory area
2794 * @address: user virtual address
2795 * @page: page to map
2796 * @pte: pointer to target page table entry
2797 * @write: true, if new entry is writable
2798 * @anon: true, if it's anonymous page
2800 * Caller must hold page table lock relevant for @pte.
2802 * Target users are page handler itself and implementations of
2803 * vm_ops->map_pages.
2805 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2806 struct page *page, pte_t *pte, bool write, bool anon)
2810 flush_icache_page(vma, page);
2811 entry = mk_pte(page, vma->vm_page_prot);
2813 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2815 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2816 page_add_new_anon_rmap(page, vma, address, false);
2818 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2819 page_add_file_rmap(page);
2821 set_pte_at(vma->vm_mm, address, pte, entry);
2823 /* no need to invalidate: a not-present page won't be cached */
2824 update_mmu_cache(vma, address, pte);
2827 static unsigned long fault_around_bytes __read_mostly =
2828 rounddown_pow_of_two(65536);
2830 #ifdef CONFIG_DEBUG_FS
2831 static int fault_around_bytes_get(void *data, u64 *val)
2833 *val = fault_around_bytes;
2838 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2839 * rounded down to nearest page order. It's what do_fault_around() expects to
2842 static int fault_around_bytes_set(void *data, u64 val)
2844 if (val / PAGE_SIZE > PTRS_PER_PTE)
2846 if (val > PAGE_SIZE)
2847 fault_around_bytes = rounddown_pow_of_two(val);
2849 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2852 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2853 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2855 static int __init fault_around_debugfs(void)
2859 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2860 &fault_around_bytes_fops);
2862 pr_warn("Failed to create fault_around_bytes in debugfs");
2865 late_initcall(fault_around_debugfs);
2869 * do_fault_around() tries to map few pages around the fault address. The hope
2870 * is that the pages will be needed soon and this will lower the number of
2873 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2874 * not ready to be mapped: not up-to-date, locked, etc.
2876 * This function is called with the page table lock taken. In the split ptlock
2877 * case the page table lock only protects only those entries which belong to
2878 * the page table corresponding to the fault address.
2880 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2883 * fault_around_pages() defines how many pages we'll try to map.
2884 * do_fault_around() expects it to return a power of two less than or equal to
2887 * The virtual address of the area that we map is naturally aligned to the
2888 * fault_around_pages() value (and therefore to page order). This way it's
2889 * easier to guarantee that we don't cross page table boundaries.
2891 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2892 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2894 unsigned long start_addr, nr_pages, mask;
2896 struct vm_fault vmf;
2899 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2900 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2902 start_addr = max(address & mask, vma->vm_start);
2903 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2908 * max_pgoff is either end of page table or end of vma
2909 * or fault_around_pages() from pgoff, depending what is nearest.
2911 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2913 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2914 pgoff + nr_pages - 1);
2916 /* Check if it makes any sense to call ->map_pages */
2917 while (!pte_none(*pte)) {
2918 if (++pgoff > max_pgoff)
2920 start_addr += PAGE_SIZE;
2921 if (start_addr >= vma->vm_end)
2926 vmf.virtual_address = (void __user *) start_addr;
2929 vmf.max_pgoff = max_pgoff;
2931 vma->vm_ops->map_pages(vma, &vmf);
2934 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2935 unsigned long address, pmd_t *pmd,
2936 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2938 struct page *fault_page;
2944 * Let's call ->map_pages() first and use ->fault() as fallback
2945 * if page by the offset is not ready to be mapped (cold cache or
2948 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
2949 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2950 do_fault_around(vma, address, pte, pgoff, flags);
2951 if (!pte_same(*pte, orig_pte))
2953 pte_unmap_unlock(pte, ptl);
2956 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
2957 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2960 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2961 if (unlikely(!pte_same(*pte, orig_pte))) {
2962 pte_unmap_unlock(pte, ptl);
2963 unlock_page(fault_page);
2964 page_cache_release(fault_page);
2967 do_set_pte(vma, address, fault_page, pte, false, false);
2968 unlock_page(fault_page);
2970 pte_unmap_unlock(pte, ptl);
2974 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2975 unsigned long address, pmd_t *pmd,
2976 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2978 struct page *fault_page, *new_page;
2979 struct mem_cgroup *memcg;
2984 if (unlikely(anon_vma_prepare(vma)))
2985 return VM_FAULT_OOM;
2987 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2989 return VM_FAULT_OOM;
2991 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) {
2992 page_cache_release(new_page);
2993 return VM_FAULT_OOM;
2996 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
2997 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3001 copy_user_highpage(new_page, fault_page, address, vma);
3002 __SetPageUptodate(new_page);
3004 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3005 if (unlikely(!pte_same(*pte, orig_pte))) {
3006 pte_unmap_unlock(pte, ptl);
3008 unlock_page(fault_page);
3009 page_cache_release(fault_page);
3012 * The fault handler has no page to lock, so it holds
3013 * i_mmap_lock for write to protect against truncate.
3015 i_mmap_unlock_write(vma->vm_file->f_mapping);
3019 do_set_pte(vma, address, new_page, pte, true, true);
3020 mem_cgroup_commit_charge(new_page, memcg, false, false);
3021 lru_cache_add_active_or_unevictable(new_page, vma);
3022 pte_unmap_unlock(pte, ptl);
3024 unlock_page(fault_page);
3025 page_cache_release(fault_page);
3028 * The fault handler has no page to lock, so it holds
3029 * i_mmap_lock for write to protect against truncate.
3031 i_mmap_unlock_write(vma->vm_file->f_mapping);
3035 mem_cgroup_cancel_charge(new_page, memcg, false);
3036 page_cache_release(new_page);
3040 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3041 unsigned long address, pmd_t *pmd,
3042 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3044 struct page *fault_page;
3045 struct address_space *mapping;
3051 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3052 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3056 * Check if the backing address space wants to know that the page is
3057 * about to become writable
3059 if (vma->vm_ops->page_mkwrite) {
3060 unlock_page(fault_page);
3061 tmp = do_page_mkwrite(vma, fault_page, address);
3062 if (unlikely(!tmp ||
3063 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3064 page_cache_release(fault_page);
3069 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3070 if (unlikely(!pte_same(*pte, orig_pte))) {
3071 pte_unmap_unlock(pte, ptl);
3072 unlock_page(fault_page);
3073 page_cache_release(fault_page);
3076 do_set_pte(vma, address, fault_page, pte, true, false);
3077 pte_unmap_unlock(pte, ptl);
3079 if (set_page_dirty(fault_page))
3082 * Take a local copy of the address_space - page.mapping may be zeroed
3083 * by truncate after unlock_page(). The address_space itself remains
3084 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3085 * release semantics to prevent the compiler from undoing this copying.
3087 mapping = page_rmapping(fault_page);
3088 unlock_page(fault_page);
3089 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3091 * Some device drivers do not set page.mapping but still
3094 balance_dirty_pages_ratelimited(mapping);
3097 if (!vma->vm_ops->page_mkwrite)
3098 file_update_time(vma->vm_file);
3104 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3105 * but allow concurrent faults).
3106 * The mmap_sem may have been released depending on flags and our
3107 * return value. See filemap_fault() and __lock_page_or_retry().
3109 static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3110 unsigned long address, pte_t *page_table, pmd_t *pmd,
3111 unsigned int flags, pte_t orig_pte)
3113 pgoff_t pgoff = (((address & PAGE_MASK)
3114 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3116 pte_unmap(page_table);
3117 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3118 if (!vma->vm_ops->fault)
3119 return VM_FAULT_SIGBUS;
3120 if (!(flags & FAULT_FLAG_WRITE))
3121 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3123 if (!(vma->vm_flags & VM_SHARED))
3124 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3126 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3129 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3130 unsigned long addr, int page_nid,
3135 count_vm_numa_event(NUMA_HINT_FAULTS);
3136 if (page_nid == numa_node_id()) {
3137 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3138 *flags |= TNF_FAULT_LOCAL;
3141 return mpol_misplaced(page, vma, addr);
3144 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3145 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3147 struct page *page = NULL;
3152 bool migrated = false;
3153 bool was_writable = pte_write(pte);
3156 /* A PROT_NONE fault should not end up here */
3157 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3160 * The "pte" at this point cannot be used safely without
3161 * validation through pte_unmap_same(). It's of NUMA type but
3162 * the pfn may be screwed if the read is non atomic.
3164 * We can safely just do a "set_pte_at()", because the old
3165 * page table entry is not accessible, so there would be no
3166 * concurrent hardware modifications to the PTE.
3168 ptl = pte_lockptr(mm, pmd);
3170 if (unlikely(!pte_same(*ptep, pte))) {
3171 pte_unmap_unlock(ptep, ptl);
3175 /* Make it present again */
3176 pte = pte_modify(pte, vma->vm_page_prot);
3177 pte = pte_mkyoung(pte);
3179 pte = pte_mkwrite(pte);
3180 set_pte_at(mm, addr, ptep, pte);
3181 update_mmu_cache(vma, addr, ptep);
3183 page = vm_normal_page(vma, addr, pte);
3185 pte_unmap_unlock(ptep, ptl);
3189 /* TODO: handle PTE-mapped THP */
3190 if (PageCompound(page)) {
3191 pte_unmap_unlock(ptep, ptl);
3196 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3197 * much anyway since they can be in shared cache state. This misses
3198 * the case where a mapping is writable but the process never writes
3199 * to it but pte_write gets cleared during protection updates and
3200 * pte_dirty has unpredictable behaviour between PTE scan updates,
3201 * background writeback, dirty balancing and application behaviour.
3203 if (!(vma->vm_flags & VM_WRITE))
3204 flags |= TNF_NO_GROUP;
3207 * Flag if the page is shared between multiple address spaces. This
3208 * is later used when determining whether to group tasks together
3210 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3211 flags |= TNF_SHARED;
3213 last_cpupid = page_cpupid_last(page);
3214 page_nid = page_to_nid(page);
3215 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3216 pte_unmap_unlock(ptep, ptl);
3217 if (target_nid == -1) {
3222 /* Migrate to the requested node */
3223 migrated = migrate_misplaced_page(page, vma, target_nid);
3225 page_nid = target_nid;
3226 flags |= TNF_MIGRATED;
3228 flags |= TNF_MIGRATE_FAIL;
3232 task_numa_fault(last_cpupid, page_nid, 1, flags);
3236 static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3237 unsigned long address, pmd_t *pmd, unsigned int flags)
3239 if (vma_is_anonymous(vma))
3240 return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3241 if (vma->vm_ops->pmd_fault)
3242 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3243 return VM_FAULT_FALLBACK;
3246 static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3247 unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3250 if (vma_is_anonymous(vma))
3251 return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3252 if (vma->vm_ops->pmd_fault)
3253 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3254 return VM_FAULT_FALLBACK;
3258 * These routines also need to handle stuff like marking pages dirty
3259 * and/or accessed for architectures that don't do it in hardware (most
3260 * RISC architectures). The early dirtying is also good on the i386.
3262 * There is also a hook called "update_mmu_cache()" that architectures
3263 * with external mmu caches can use to update those (ie the Sparc or
3264 * PowerPC hashed page tables that act as extended TLBs).
3266 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3267 * but allow concurrent faults), and pte mapped but not yet locked.
3268 * We return with pte unmapped and unlocked.
3270 * The mmap_sem may have been released depending on flags and our
3271 * return value. See filemap_fault() and __lock_page_or_retry().
3273 static int handle_pte_fault(struct mm_struct *mm,
3274 struct vm_area_struct *vma, unsigned long address,
3275 pte_t *pte, pmd_t *pmd, unsigned int flags)
3281 * some architectures can have larger ptes than wordsize,
3282 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3283 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3284 * The code below just needs a consistent view for the ifs and
3285 * we later double check anyway with the ptl lock held. So here
3286 * a barrier will do.
3290 if (!pte_present(entry)) {
3291 if (pte_none(entry)) {
3292 if (vma_is_anonymous(vma))
3293 return do_anonymous_page(mm, vma, address,
3296 return do_fault(mm, vma, address, pte, pmd,
3299 return do_swap_page(mm, vma, address,
3300 pte, pmd, flags, entry);
3303 if (pte_protnone(entry))
3304 return do_numa_page(mm, vma, address, entry, pte, pmd);
3306 ptl = pte_lockptr(mm, pmd);
3308 if (unlikely(!pte_same(*pte, entry)))
3310 if (flags & FAULT_FLAG_WRITE) {
3311 if (!pte_write(entry))
3312 return do_wp_page(mm, vma, address,
3313 pte, pmd, ptl, entry);
3314 entry = pte_mkdirty(entry);
3316 entry = pte_mkyoung(entry);
3317 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3318 update_mmu_cache(vma, address, pte);
3321 * This is needed only for protection faults but the arch code
3322 * is not yet telling us if this is a protection fault or not.
3323 * This still avoids useless tlb flushes for .text page faults
3326 if (flags & FAULT_FLAG_WRITE)
3327 flush_tlb_fix_spurious_fault(vma, address);
3330 pte_unmap_unlock(pte, ptl);
3335 * By the time we get here, we already hold the mm semaphore
3337 * The mmap_sem may have been released depending on flags and our
3338 * return value. See filemap_fault() and __lock_page_or_retry().
3340 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3341 unsigned long address, unsigned int flags)
3348 if (unlikely(is_vm_hugetlb_page(vma)))
3349 return hugetlb_fault(mm, vma, address, flags);
3351 pgd = pgd_offset(mm, address);
3352 pud = pud_alloc(mm, pgd, address);
3354 return VM_FAULT_OOM;
3355 pmd = pmd_alloc(mm, pud, address);
3357 return VM_FAULT_OOM;
3358 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3359 int ret = create_huge_pmd(mm, vma, address, pmd, flags);
3360 if (!(ret & VM_FAULT_FALLBACK))
3363 pmd_t orig_pmd = *pmd;
3367 if (pmd_trans_huge(orig_pmd)) {
3368 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3370 if (pmd_protnone(orig_pmd))
3371 return do_huge_pmd_numa_page(mm, vma, address,
3374 if (dirty && !pmd_write(orig_pmd)) {
3375 ret = wp_huge_pmd(mm, vma, address, pmd,
3377 if (!(ret & VM_FAULT_FALLBACK))
3380 huge_pmd_set_accessed(mm, vma, address, pmd,
3388 * Use __pte_alloc instead of pte_alloc_map, because we can't
3389 * run pte_offset_map on the pmd, if an huge pmd could
3390 * materialize from under us from a different thread.
3392 if (unlikely(pmd_none(*pmd)) &&
3393 unlikely(__pte_alloc(mm, vma, pmd, address)))
3394 return VM_FAULT_OOM;
3395 /* if an huge pmd materialized from under us just retry later */
3396 if (unlikely(pmd_trans_huge(*pmd)))
3399 * A regular pmd is established and it can't morph into a huge pmd
3400 * from under us anymore at this point because we hold the mmap_sem
3401 * read mode and khugepaged takes it in write mode. So now it's
3402 * safe to run pte_offset_map().
3404 pte = pte_offset_map(pmd, address);
3406 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3410 * By the time we get here, we already hold the mm semaphore
3412 * The mmap_sem may have been released depending on flags and our
3413 * return value. See filemap_fault() and __lock_page_or_retry().
3415 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3416 unsigned long address, unsigned int flags)
3420 __set_current_state(TASK_RUNNING);
3422 count_vm_event(PGFAULT);
3423 mem_cgroup_count_vm_event(mm, PGFAULT);
3425 /* do counter updates before entering really critical section. */
3426 check_sync_rss_stat(current);
3429 * Enable the memcg OOM handling for faults triggered in user
3430 * space. Kernel faults are handled more gracefully.
3432 if (flags & FAULT_FLAG_USER)
3433 mem_cgroup_oom_enable();
3435 ret = __handle_mm_fault(mm, vma, address, flags);
3437 if (flags & FAULT_FLAG_USER) {
3438 mem_cgroup_oom_disable();
3440 * The task may have entered a memcg OOM situation but
3441 * if the allocation error was handled gracefully (no
3442 * VM_FAULT_OOM), there is no need to kill anything.
3443 * Just clean up the OOM state peacefully.
3445 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3446 mem_cgroup_oom_synchronize(false);
3451 EXPORT_SYMBOL_GPL(handle_mm_fault);
3453 #ifndef __PAGETABLE_PUD_FOLDED
3455 * Allocate page upper directory.
3456 * We've already handled the fast-path in-line.
3458 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3460 pud_t *new = pud_alloc_one(mm, address);
3464 smp_wmb(); /* See comment in __pte_alloc */
3466 spin_lock(&mm->page_table_lock);
3467 if (pgd_present(*pgd)) /* Another has populated it */
3470 pgd_populate(mm, pgd, new);
3471 spin_unlock(&mm->page_table_lock);
3474 #endif /* __PAGETABLE_PUD_FOLDED */
3476 #ifndef __PAGETABLE_PMD_FOLDED
3478 * Allocate page middle directory.
3479 * We've already handled the fast-path in-line.
3481 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3483 pmd_t *new = pmd_alloc_one(mm, address);
3487 smp_wmb(); /* See comment in __pte_alloc */
3489 spin_lock(&mm->page_table_lock);
3490 #ifndef __ARCH_HAS_4LEVEL_HACK
3491 if (!pud_present(*pud)) {
3493 pud_populate(mm, pud, new);
3494 } else /* Another has populated it */
3497 if (!pgd_present(*pud)) {
3499 pgd_populate(mm, pud, new);
3500 } else /* Another has populated it */
3502 #endif /* __ARCH_HAS_4LEVEL_HACK */
3503 spin_unlock(&mm->page_table_lock);
3506 #endif /* __PAGETABLE_PMD_FOLDED */
3508 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3509 pte_t **ptepp, spinlock_t **ptlp)
3516 pgd = pgd_offset(mm, address);
3517 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3520 pud = pud_offset(pgd, address);
3521 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3524 pmd = pmd_offset(pud, address);
3525 VM_BUG_ON(pmd_trans_huge(*pmd));
3526 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3529 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3533 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3536 if (!pte_present(*ptep))
3541 pte_unmap_unlock(ptep, *ptlp);
3546 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3547 pte_t **ptepp, spinlock_t **ptlp)
3551 /* (void) is needed to make gcc happy */
3552 (void) __cond_lock(*ptlp,
3553 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3558 * follow_pfn - look up PFN at a user virtual address
3559 * @vma: memory mapping
3560 * @address: user virtual address
3561 * @pfn: location to store found PFN
3563 * Only IO mappings and raw PFN mappings are allowed.
3565 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3567 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3574 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3577 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3580 *pfn = pte_pfn(*ptep);
3581 pte_unmap_unlock(ptep, ptl);
3584 EXPORT_SYMBOL(follow_pfn);
3586 #ifdef CONFIG_HAVE_IOREMAP_PROT
3587 int follow_phys(struct vm_area_struct *vma,
3588 unsigned long address, unsigned int flags,
3589 unsigned long *prot, resource_size_t *phys)
3595 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3598 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3602 if ((flags & FOLL_WRITE) && !pte_write(pte))
3605 *prot = pgprot_val(pte_pgprot(pte));
3606 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3610 pte_unmap_unlock(ptep, ptl);
3615 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3616 void *buf, int len, int write)
3618 resource_size_t phys_addr;
3619 unsigned long prot = 0;
3620 void __iomem *maddr;
3621 int offset = addr & (PAGE_SIZE-1);
3623 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3626 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3628 memcpy_toio(maddr + offset, buf, len);
3630 memcpy_fromio(buf, maddr + offset, len);
3635 EXPORT_SYMBOL_GPL(generic_access_phys);
3639 * Access another process' address space as given in mm. If non-NULL, use the
3640 * given task for page fault accounting.
3642 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3643 unsigned long addr, void *buf, int len, int write)
3645 struct vm_area_struct *vma;
3646 void *old_buf = buf;
3648 down_read(&mm->mmap_sem);
3649 /* ignore errors, just check how much was successfully transferred */
3651 int bytes, ret, offset;
3653 struct page *page = NULL;
3655 ret = get_user_pages(tsk, mm, addr, 1,
3656 write, 1, &page, &vma);
3658 #ifndef CONFIG_HAVE_IOREMAP_PROT
3662 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3663 * we can access using slightly different code.
3665 vma = find_vma(mm, addr);
3666 if (!vma || vma->vm_start > addr)
3668 if (vma->vm_ops && vma->vm_ops->access)
3669 ret = vma->vm_ops->access(vma, addr, buf,
3677 offset = addr & (PAGE_SIZE-1);
3678 if (bytes > PAGE_SIZE-offset)
3679 bytes = PAGE_SIZE-offset;
3683 copy_to_user_page(vma, page, addr,
3684 maddr + offset, buf, bytes);
3685 set_page_dirty_lock(page);
3687 copy_from_user_page(vma, page, addr,
3688 buf, maddr + offset, bytes);
3691 page_cache_release(page);
3697 up_read(&mm->mmap_sem);
3699 return buf - old_buf;
3703 * access_remote_vm - access another process' address space
3704 * @mm: the mm_struct of the target address space
3705 * @addr: start address to access
3706 * @buf: source or destination buffer
3707 * @len: number of bytes to transfer
3708 * @write: whether the access is a write
3710 * The caller must hold a reference on @mm.
3712 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3713 void *buf, int len, int write)
3715 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3719 * Access another process' address space.
3720 * Source/target buffer must be kernel space,
3721 * Do not walk the page table directly, use get_user_pages
3723 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3724 void *buf, int len, int write)
3726 struct mm_struct *mm;
3729 mm = get_task_mm(tsk);
3733 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3740 * Print the name of a VMA.
3742 void print_vma_addr(char *prefix, unsigned long ip)
3744 struct mm_struct *mm = current->mm;
3745 struct vm_area_struct *vma;
3748 * Do not print if we are in atomic
3749 * contexts (in exception stacks, etc.):
3751 if (preempt_count())
3754 down_read(&mm->mmap_sem);
3755 vma = find_vma(mm, ip);
3756 if (vma && vma->vm_file) {
3757 struct file *f = vma->vm_file;
3758 char *buf = (char *)__get_free_page(GFP_KERNEL);
3762 p = file_path(f, buf, PAGE_SIZE);
3765 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3767 vma->vm_end - vma->vm_start);
3768 free_page((unsigned long)buf);
3771 up_read(&mm->mmap_sem);
3774 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3775 void __might_fault(const char *file, int line)
3778 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3779 * holding the mmap_sem, this is safe because kernel memory doesn't
3780 * get paged out, therefore we'll never actually fault, and the
3781 * below annotations will generate false positives.
3783 if (segment_eq(get_fs(), KERNEL_DS))
3785 if (pagefault_disabled())
3787 __might_sleep(file, line, 0);
3788 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3790 might_lock_read(¤t->mm->mmap_sem);
3793 EXPORT_SYMBOL(__might_fault);
3796 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3797 static void clear_gigantic_page(struct page *page,
3799 unsigned int pages_per_huge_page)
3802 struct page *p = page;
3805 for (i = 0; i < pages_per_huge_page;
3806 i++, p = mem_map_next(p, page, i)) {
3808 clear_user_highpage(p, addr + i * PAGE_SIZE);
3811 void clear_huge_page(struct page *page,
3812 unsigned long addr, unsigned int pages_per_huge_page)
3816 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3817 clear_gigantic_page(page, addr, pages_per_huge_page);
3822 for (i = 0; i < pages_per_huge_page; i++) {
3824 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3828 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3830 struct vm_area_struct *vma,
3831 unsigned int pages_per_huge_page)
3834 struct page *dst_base = dst;
3835 struct page *src_base = src;
3837 for (i = 0; i < pages_per_huge_page; ) {
3839 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3842 dst = mem_map_next(dst, dst_base, i);
3843 src = mem_map_next(src, src_base, i);
3847 void copy_user_huge_page(struct page *dst, struct page *src,
3848 unsigned long addr, struct vm_area_struct *vma,
3849 unsigned int pages_per_huge_page)
3853 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3854 copy_user_gigantic_page(dst, src, addr, vma,
3855 pages_per_huge_page);
3860 for (i = 0; i < pages_per_huge_page; i++) {
3862 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3865 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3867 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3869 static struct kmem_cache *page_ptl_cachep;
3871 void __init ptlock_cache_init(void)
3873 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3877 bool ptlock_alloc(struct page *page)
3881 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3888 void ptlock_free(struct page *page)
3890 kmem_cache_free(page_ptl_cachep, page->ptl);