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);
829 rss[mm_counter(page)]++;
831 if (is_write_migration_entry(entry) &&
832 is_cow_mapping(vm_flags)) {
834 * COW mappings require pages in both
835 * parent and child to be set to read.
837 make_migration_entry_read(&entry);
838 pte = swp_entry_to_pte(entry);
839 if (pte_swp_soft_dirty(*src_pte))
840 pte = pte_swp_mksoft_dirty(pte);
841 set_pte_at(src_mm, addr, src_pte, pte);
848 * If it's a COW mapping, write protect it both
849 * in the parent and the child
851 if (is_cow_mapping(vm_flags)) {
852 ptep_set_wrprotect(src_mm, addr, src_pte);
853 pte = pte_wrprotect(pte);
857 * If it's a shared mapping, mark it clean in
860 if (vm_flags & VM_SHARED)
861 pte = pte_mkclean(pte);
862 pte = pte_mkold(pte);
864 page = vm_normal_page(vma, addr, pte);
867 page_dup_rmap(page, false);
868 rss[mm_counter(page)]++;
872 set_pte_at(dst_mm, addr, dst_pte, pte);
876 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
877 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
878 unsigned long addr, unsigned long end)
880 pte_t *orig_src_pte, *orig_dst_pte;
881 pte_t *src_pte, *dst_pte;
882 spinlock_t *src_ptl, *dst_ptl;
884 int rss[NR_MM_COUNTERS];
885 swp_entry_t entry = (swp_entry_t){0};
890 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
893 src_pte = pte_offset_map(src_pmd, addr);
894 src_ptl = pte_lockptr(src_mm, src_pmd);
895 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
896 orig_src_pte = src_pte;
897 orig_dst_pte = dst_pte;
898 arch_enter_lazy_mmu_mode();
902 * We are holding two locks at this point - either of them
903 * could generate latencies in another task on another CPU.
905 if (progress >= 32) {
907 if (need_resched() ||
908 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
911 if (pte_none(*src_pte)) {
915 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
920 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
922 arch_leave_lazy_mmu_mode();
923 spin_unlock(src_ptl);
924 pte_unmap(orig_src_pte);
925 add_mm_rss_vec(dst_mm, rss);
926 pte_unmap_unlock(orig_dst_pte, dst_ptl);
930 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
939 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
940 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
941 unsigned long addr, unsigned long end)
943 pmd_t *src_pmd, *dst_pmd;
946 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
949 src_pmd = pmd_offset(src_pud, addr);
951 next = pmd_addr_end(addr, end);
952 if (pmd_trans_huge(*src_pmd)) {
954 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
955 err = copy_huge_pmd(dst_mm, src_mm,
956 dst_pmd, src_pmd, addr, vma);
963 if (pmd_none_or_clear_bad(src_pmd))
965 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
968 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
972 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
973 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
974 unsigned long addr, unsigned long end)
976 pud_t *src_pud, *dst_pud;
979 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
982 src_pud = pud_offset(src_pgd, addr);
984 next = pud_addr_end(addr, end);
985 if (pud_none_or_clear_bad(src_pud))
987 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
990 } while (dst_pud++, src_pud++, addr = next, addr != end);
994 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
995 struct vm_area_struct *vma)
997 pgd_t *src_pgd, *dst_pgd;
999 unsigned long addr = vma->vm_start;
1000 unsigned long end = vma->vm_end;
1001 unsigned long mmun_start; /* For mmu_notifiers */
1002 unsigned long mmun_end; /* For mmu_notifiers */
1007 * Don't copy ptes where a page fault will fill them correctly.
1008 * Fork becomes much lighter when there are big shared or private
1009 * readonly mappings. The tradeoff is that copy_page_range is more
1010 * efficient than faulting.
1012 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1016 if (is_vm_hugetlb_page(vma))
1017 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1019 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1021 * We do not free on error cases below as remove_vma
1022 * gets called on error from higher level routine
1024 ret = track_pfn_copy(vma);
1030 * We need to invalidate the secondary MMU mappings only when
1031 * there could be a permission downgrade on the ptes of the
1032 * parent mm. And a permission downgrade will only happen if
1033 * is_cow_mapping() returns true.
1035 is_cow = is_cow_mapping(vma->vm_flags);
1039 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1043 dst_pgd = pgd_offset(dst_mm, addr);
1044 src_pgd = pgd_offset(src_mm, addr);
1046 next = pgd_addr_end(addr, end);
1047 if (pgd_none_or_clear_bad(src_pgd))
1049 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1050 vma, addr, next))) {
1054 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1057 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1061 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1062 struct vm_area_struct *vma, pmd_t *pmd,
1063 unsigned long addr, unsigned long end,
1064 struct zap_details *details)
1066 struct mm_struct *mm = tlb->mm;
1067 int force_flush = 0;
1068 int rss[NR_MM_COUNTERS];
1076 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1078 arch_enter_lazy_mmu_mode();
1081 if (pte_none(ptent)) {
1085 if (pte_present(ptent)) {
1088 page = vm_normal_page(vma, addr, ptent);
1089 if (unlikely(details) && page) {
1091 * unmap_shared_mapping_pages() wants to
1092 * invalidate cache without truncating:
1093 * unmap shared but keep private pages.
1095 if (details->check_mapping &&
1096 details->check_mapping != page->mapping)
1099 ptent = ptep_get_and_clear_full(mm, addr, pte,
1101 tlb_remove_tlb_entry(tlb, pte, addr);
1102 if (unlikely(!page))
1105 if (!PageAnon(page)) {
1106 if (pte_dirty(ptent)) {
1108 set_page_dirty(page);
1110 if (pte_young(ptent) &&
1111 likely(!(vma->vm_flags & VM_SEQ_READ)))
1112 mark_page_accessed(page);
1114 rss[mm_counter(page)]--;
1115 page_remove_rmap(page, false);
1116 if (unlikely(page_mapcount(page) < 0))
1117 print_bad_pte(vma, addr, ptent, page);
1118 if (unlikely(!__tlb_remove_page(tlb, page))) {
1125 /* If details->check_mapping, we leave swap entries. */
1126 if (unlikely(details))
1129 entry = pte_to_swp_entry(ptent);
1130 if (!non_swap_entry(entry))
1132 else if (is_migration_entry(entry)) {
1135 page = migration_entry_to_page(entry);
1136 rss[mm_counter(page)]--;
1138 if (unlikely(!free_swap_and_cache(entry)))
1139 print_bad_pte(vma, addr, ptent, NULL);
1140 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1141 } while (pte++, addr += PAGE_SIZE, addr != end);
1143 add_mm_rss_vec(mm, rss);
1144 arch_leave_lazy_mmu_mode();
1146 /* Do the actual TLB flush before dropping ptl */
1148 tlb_flush_mmu_tlbonly(tlb);
1149 pte_unmap_unlock(start_pte, ptl);
1152 * If we forced a TLB flush (either due to running out of
1153 * batch buffers or because we needed to flush dirty TLB
1154 * entries before releasing the ptl), free the batched
1155 * memory too. Restart if we didn't do everything.
1159 tlb_flush_mmu_free(tlb);
1168 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1169 struct vm_area_struct *vma, pud_t *pud,
1170 unsigned long addr, unsigned long end,
1171 struct zap_details *details)
1176 pmd = pmd_offset(pud, addr);
1178 next = pmd_addr_end(addr, end);
1179 if (pmd_trans_huge(*pmd)) {
1180 if (next - addr != HPAGE_PMD_SIZE) {
1181 #ifdef CONFIG_DEBUG_VM
1182 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1183 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1184 __func__, addr, end,
1190 split_huge_pmd(vma, pmd, addr);
1191 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1196 * Here there can be other concurrent MADV_DONTNEED or
1197 * trans huge page faults running, and if the pmd is
1198 * none or trans huge it can change under us. This is
1199 * because MADV_DONTNEED holds the mmap_sem in read
1202 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1204 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1207 } while (pmd++, addr = next, addr != end);
1212 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1213 struct vm_area_struct *vma, pgd_t *pgd,
1214 unsigned long addr, unsigned long end,
1215 struct zap_details *details)
1220 pud = pud_offset(pgd, addr);
1222 next = pud_addr_end(addr, end);
1223 if (pud_none_or_clear_bad(pud))
1225 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1226 } while (pud++, addr = next, addr != end);
1231 static void unmap_page_range(struct mmu_gather *tlb,
1232 struct vm_area_struct *vma,
1233 unsigned long addr, unsigned long end,
1234 struct zap_details *details)
1239 if (details && !details->check_mapping)
1242 BUG_ON(addr >= end);
1243 tlb_start_vma(tlb, vma);
1244 pgd = pgd_offset(vma->vm_mm, addr);
1246 next = pgd_addr_end(addr, end);
1247 if (pgd_none_or_clear_bad(pgd))
1249 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1250 } while (pgd++, addr = next, addr != end);
1251 tlb_end_vma(tlb, vma);
1255 static void unmap_single_vma(struct mmu_gather *tlb,
1256 struct vm_area_struct *vma, unsigned long start_addr,
1257 unsigned long end_addr,
1258 struct zap_details *details)
1260 unsigned long start = max(vma->vm_start, start_addr);
1263 if (start >= vma->vm_end)
1265 end = min(vma->vm_end, end_addr);
1266 if (end <= vma->vm_start)
1270 uprobe_munmap(vma, start, end);
1272 if (unlikely(vma->vm_flags & VM_PFNMAP))
1273 untrack_pfn(vma, 0, 0);
1276 if (unlikely(is_vm_hugetlb_page(vma))) {
1278 * It is undesirable to test vma->vm_file as it
1279 * should be non-null for valid hugetlb area.
1280 * However, vm_file will be NULL in the error
1281 * cleanup path of mmap_region. When
1282 * hugetlbfs ->mmap method fails,
1283 * mmap_region() nullifies vma->vm_file
1284 * before calling this function to clean up.
1285 * Since no pte has actually been setup, it is
1286 * safe to do nothing in this case.
1289 i_mmap_lock_write(vma->vm_file->f_mapping);
1290 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1291 i_mmap_unlock_write(vma->vm_file->f_mapping);
1294 unmap_page_range(tlb, vma, start, end, details);
1299 * unmap_vmas - unmap a range of memory covered by a list of vma's
1300 * @tlb: address of the caller's struct mmu_gather
1301 * @vma: the starting vma
1302 * @start_addr: virtual address at which to start unmapping
1303 * @end_addr: virtual address at which to end unmapping
1305 * Unmap all pages in the vma list.
1307 * Only addresses between `start' and `end' will be unmapped.
1309 * The VMA list must be sorted in ascending virtual address order.
1311 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1312 * range after unmap_vmas() returns. So the only responsibility here is to
1313 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1314 * drops the lock and schedules.
1316 void unmap_vmas(struct mmu_gather *tlb,
1317 struct vm_area_struct *vma, unsigned long start_addr,
1318 unsigned long end_addr)
1320 struct mm_struct *mm = vma->vm_mm;
1322 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1323 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1324 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1325 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1329 * zap_page_range - remove user pages in a given range
1330 * @vma: vm_area_struct holding the applicable pages
1331 * @start: starting address of pages to zap
1332 * @size: number of bytes to zap
1333 * @details: details of shared cache invalidation
1335 * Caller must protect the VMA list
1337 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1338 unsigned long size, struct zap_details *details)
1340 struct mm_struct *mm = vma->vm_mm;
1341 struct mmu_gather tlb;
1342 unsigned long end = start + size;
1345 tlb_gather_mmu(&tlb, mm, start, end);
1346 update_hiwater_rss(mm);
1347 mmu_notifier_invalidate_range_start(mm, start, end);
1348 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1349 unmap_single_vma(&tlb, vma, start, end, details);
1350 mmu_notifier_invalidate_range_end(mm, start, end);
1351 tlb_finish_mmu(&tlb, start, end);
1355 * zap_page_range_single - remove user pages in a given range
1356 * @vma: vm_area_struct holding the applicable pages
1357 * @address: starting address of pages to zap
1358 * @size: number of bytes to zap
1359 * @details: details of shared cache invalidation
1361 * The range must fit into one VMA.
1363 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1364 unsigned long size, struct zap_details *details)
1366 struct mm_struct *mm = vma->vm_mm;
1367 struct mmu_gather tlb;
1368 unsigned long end = address + size;
1371 tlb_gather_mmu(&tlb, mm, address, end);
1372 update_hiwater_rss(mm);
1373 mmu_notifier_invalidate_range_start(mm, address, end);
1374 unmap_single_vma(&tlb, vma, address, end, details);
1375 mmu_notifier_invalidate_range_end(mm, address, end);
1376 tlb_finish_mmu(&tlb, address, end);
1380 * zap_vma_ptes - remove ptes mapping the vma
1381 * @vma: vm_area_struct holding ptes to be zapped
1382 * @address: starting address of pages to zap
1383 * @size: number of bytes to zap
1385 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1387 * The entire address range must be fully contained within the vma.
1389 * Returns 0 if successful.
1391 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1394 if (address < vma->vm_start || address + size > vma->vm_end ||
1395 !(vma->vm_flags & VM_PFNMAP))
1397 zap_page_range_single(vma, address, size, NULL);
1400 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1402 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1405 pgd_t * pgd = pgd_offset(mm, addr);
1406 pud_t * pud = pud_alloc(mm, pgd, addr);
1408 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1410 VM_BUG_ON(pmd_trans_huge(*pmd));
1411 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1418 * This is the old fallback for page remapping.
1420 * For historical reasons, it only allows reserved pages. Only
1421 * old drivers should use this, and they needed to mark their
1422 * pages reserved for the old functions anyway.
1424 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1425 struct page *page, pgprot_t prot)
1427 struct mm_struct *mm = vma->vm_mm;
1436 flush_dcache_page(page);
1437 pte = get_locked_pte(mm, addr, &ptl);
1441 if (!pte_none(*pte))
1444 /* Ok, finally just insert the thing.. */
1446 inc_mm_counter_fast(mm, mm_counter_file(page));
1447 page_add_file_rmap(page);
1448 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1451 pte_unmap_unlock(pte, ptl);
1454 pte_unmap_unlock(pte, ptl);
1460 * vm_insert_page - insert single page into user vma
1461 * @vma: user vma to map to
1462 * @addr: target user address of this page
1463 * @page: source kernel page
1465 * This allows drivers to insert individual pages they've allocated
1468 * The page has to be a nice clean _individual_ kernel allocation.
1469 * If you allocate a compound page, you need to have marked it as
1470 * such (__GFP_COMP), or manually just split the page up yourself
1471 * (see split_page()).
1473 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1474 * took an arbitrary page protection parameter. This doesn't allow
1475 * that. Your vma protection will have to be set up correctly, which
1476 * means that if you want a shared writable mapping, you'd better
1477 * ask for a shared writable mapping!
1479 * The page does not need to be reserved.
1481 * Usually this function is called from f_op->mmap() handler
1482 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1483 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1484 * function from other places, for example from page-fault handler.
1486 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1489 if (addr < vma->vm_start || addr >= vma->vm_end)
1491 if (!page_count(page))
1493 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1494 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1495 BUG_ON(vma->vm_flags & VM_PFNMAP);
1496 vma->vm_flags |= VM_MIXEDMAP;
1498 return insert_page(vma, addr, page, vma->vm_page_prot);
1500 EXPORT_SYMBOL(vm_insert_page);
1502 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1503 unsigned long pfn, pgprot_t prot)
1505 struct mm_struct *mm = vma->vm_mm;
1511 pte = get_locked_pte(mm, addr, &ptl);
1515 if (!pte_none(*pte))
1518 /* Ok, finally just insert the thing.. */
1519 entry = pte_mkspecial(pfn_pte(pfn, prot));
1520 set_pte_at(mm, addr, pte, entry);
1521 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1525 pte_unmap_unlock(pte, ptl);
1531 * vm_insert_pfn - insert single pfn into user vma
1532 * @vma: user vma to map to
1533 * @addr: target user address of this page
1534 * @pfn: source kernel pfn
1536 * Similar to vm_insert_page, this allows drivers to insert individual pages
1537 * they've allocated into a user vma. Same comments apply.
1539 * This function should only be called from a vm_ops->fault handler, and
1540 * in that case the handler should return NULL.
1542 * vma cannot be a COW mapping.
1544 * As this is called only for pages that do not currently exist, we
1545 * do not need to flush old virtual caches or the TLB.
1547 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1551 pgprot_t pgprot = vma->vm_page_prot;
1553 * Technically, architectures with pte_special can avoid all these
1554 * restrictions (same for remap_pfn_range). However we would like
1555 * consistency in testing and feature parity among all, so we should
1556 * try to keep these invariants in place for everybody.
1558 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1559 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1560 (VM_PFNMAP|VM_MIXEDMAP));
1561 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1562 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1564 if (addr < vma->vm_start || addr >= vma->vm_end)
1566 if (track_pfn_insert(vma, &pgprot, pfn))
1569 ret = insert_pfn(vma, addr, pfn, pgprot);
1573 EXPORT_SYMBOL(vm_insert_pfn);
1575 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1578 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1580 if (addr < vma->vm_start || addr >= vma->vm_end)
1584 * If we don't have pte special, then we have to use the pfn_valid()
1585 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1586 * refcount the page if pfn_valid is true (hence insert_page rather
1587 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1588 * without pte special, it would there be refcounted as a normal page.
1590 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1593 page = pfn_to_page(pfn);
1594 return insert_page(vma, addr, page, vma->vm_page_prot);
1596 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1598 EXPORT_SYMBOL(vm_insert_mixed);
1601 * maps a range of physical memory into the requested pages. the old
1602 * mappings are removed. any references to nonexistent pages results
1603 * in null mappings (currently treated as "copy-on-access")
1605 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1606 unsigned long addr, unsigned long end,
1607 unsigned long pfn, pgprot_t prot)
1612 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1615 arch_enter_lazy_mmu_mode();
1617 BUG_ON(!pte_none(*pte));
1618 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1620 } while (pte++, addr += PAGE_SIZE, addr != end);
1621 arch_leave_lazy_mmu_mode();
1622 pte_unmap_unlock(pte - 1, ptl);
1626 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1627 unsigned long addr, unsigned long end,
1628 unsigned long pfn, pgprot_t prot)
1633 pfn -= addr >> PAGE_SHIFT;
1634 pmd = pmd_alloc(mm, pud, addr);
1637 VM_BUG_ON(pmd_trans_huge(*pmd));
1639 next = pmd_addr_end(addr, end);
1640 if (remap_pte_range(mm, pmd, addr, next,
1641 pfn + (addr >> PAGE_SHIFT), prot))
1643 } while (pmd++, addr = next, addr != end);
1647 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1648 unsigned long addr, unsigned long end,
1649 unsigned long pfn, pgprot_t prot)
1654 pfn -= addr >> PAGE_SHIFT;
1655 pud = pud_alloc(mm, pgd, addr);
1659 next = pud_addr_end(addr, end);
1660 if (remap_pmd_range(mm, pud, addr, next,
1661 pfn + (addr >> PAGE_SHIFT), prot))
1663 } while (pud++, addr = next, addr != end);
1668 * remap_pfn_range - remap kernel memory to userspace
1669 * @vma: user vma to map to
1670 * @addr: target user address to start at
1671 * @pfn: physical address of kernel memory
1672 * @size: size of map area
1673 * @prot: page protection flags for this mapping
1675 * Note: this is only safe if the mm semaphore is held when called.
1677 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1678 unsigned long pfn, unsigned long size, pgprot_t prot)
1682 unsigned long end = addr + PAGE_ALIGN(size);
1683 struct mm_struct *mm = vma->vm_mm;
1687 * Physically remapped pages are special. Tell the
1688 * rest of the world about it:
1689 * VM_IO tells people not to look at these pages
1690 * (accesses can have side effects).
1691 * VM_PFNMAP tells the core MM that the base pages are just
1692 * raw PFN mappings, and do not have a "struct page" associated
1695 * Disable vma merging and expanding with mremap().
1697 * Omit vma from core dump, even when VM_IO turned off.
1699 * There's a horrible special case to handle copy-on-write
1700 * behaviour that some programs depend on. We mark the "original"
1701 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1702 * See vm_normal_page() for details.
1704 if (is_cow_mapping(vma->vm_flags)) {
1705 if (addr != vma->vm_start || end != vma->vm_end)
1707 vma->vm_pgoff = pfn;
1710 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1714 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1716 BUG_ON(addr >= end);
1717 pfn -= addr >> PAGE_SHIFT;
1718 pgd = pgd_offset(mm, addr);
1719 flush_cache_range(vma, addr, end);
1721 next = pgd_addr_end(addr, end);
1722 err = remap_pud_range(mm, pgd, addr, next,
1723 pfn + (addr >> PAGE_SHIFT), prot);
1726 } while (pgd++, addr = next, addr != end);
1729 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1733 EXPORT_SYMBOL(remap_pfn_range);
1736 * vm_iomap_memory - remap memory to userspace
1737 * @vma: user vma to map to
1738 * @start: start of area
1739 * @len: size of area
1741 * This is a simplified io_remap_pfn_range() for common driver use. The
1742 * driver just needs to give us the physical memory range to be mapped,
1743 * we'll figure out the rest from the vma information.
1745 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1746 * whatever write-combining details or similar.
1748 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1750 unsigned long vm_len, pfn, pages;
1752 /* Check that the physical memory area passed in looks valid */
1753 if (start + len < start)
1756 * You *really* shouldn't map things that aren't page-aligned,
1757 * but we've historically allowed it because IO memory might
1758 * just have smaller alignment.
1760 len += start & ~PAGE_MASK;
1761 pfn = start >> PAGE_SHIFT;
1762 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1763 if (pfn + pages < pfn)
1766 /* We start the mapping 'vm_pgoff' pages into the area */
1767 if (vma->vm_pgoff > pages)
1769 pfn += vma->vm_pgoff;
1770 pages -= vma->vm_pgoff;
1772 /* Can we fit all of the mapping? */
1773 vm_len = vma->vm_end - vma->vm_start;
1774 if (vm_len >> PAGE_SHIFT > pages)
1777 /* Ok, let it rip */
1778 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1780 EXPORT_SYMBOL(vm_iomap_memory);
1782 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1783 unsigned long addr, unsigned long end,
1784 pte_fn_t fn, void *data)
1789 spinlock_t *uninitialized_var(ptl);
1791 pte = (mm == &init_mm) ?
1792 pte_alloc_kernel(pmd, addr) :
1793 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1797 BUG_ON(pmd_huge(*pmd));
1799 arch_enter_lazy_mmu_mode();
1801 token = pmd_pgtable(*pmd);
1804 err = fn(pte++, token, addr, data);
1807 } while (addr += PAGE_SIZE, addr != end);
1809 arch_leave_lazy_mmu_mode();
1812 pte_unmap_unlock(pte-1, ptl);
1816 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1817 unsigned long addr, unsigned long end,
1818 pte_fn_t fn, void *data)
1824 BUG_ON(pud_huge(*pud));
1826 pmd = pmd_alloc(mm, pud, addr);
1830 next = pmd_addr_end(addr, end);
1831 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1834 } while (pmd++, addr = next, addr != end);
1838 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1839 unsigned long addr, unsigned long end,
1840 pte_fn_t fn, void *data)
1846 pud = pud_alloc(mm, pgd, addr);
1850 next = pud_addr_end(addr, end);
1851 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1854 } while (pud++, addr = next, addr != end);
1859 * Scan a region of virtual memory, filling in page tables as necessary
1860 * and calling a provided function on each leaf page table.
1862 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1863 unsigned long size, pte_fn_t fn, void *data)
1867 unsigned long end = addr + size;
1870 BUG_ON(addr >= end);
1871 pgd = pgd_offset(mm, addr);
1873 next = pgd_addr_end(addr, end);
1874 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1877 } while (pgd++, addr = next, addr != end);
1881 EXPORT_SYMBOL_GPL(apply_to_page_range);
1884 * handle_pte_fault chooses page fault handler according to an entry which was
1885 * read non-atomically. Before making any commitment, on those architectures
1886 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1887 * parts, do_swap_page must check under lock before unmapping the pte and
1888 * proceeding (but do_wp_page is only called after already making such a check;
1889 * and do_anonymous_page can safely check later on).
1891 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1892 pte_t *page_table, pte_t orig_pte)
1895 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1896 if (sizeof(pte_t) > sizeof(unsigned long)) {
1897 spinlock_t *ptl = pte_lockptr(mm, pmd);
1899 same = pte_same(*page_table, orig_pte);
1903 pte_unmap(page_table);
1907 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1909 debug_dma_assert_idle(src);
1912 * If the source page was a PFN mapping, we don't have
1913 * a "struct page" for it. We do a best-effort copy by
1914 * just copying from the original user address. If that
1915 * fails, we just zero-fill it. Live with it.
1917 if (unlikely(!src)) {
1918 void *kaddr = kmap_atomic(dst);
1919 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1922 * This really shouldn't fail, because the page is there
1923 * in the page tables. But it might just be unreadable,
1924 * in which case we just give up and fill the result with
1927 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1929 kunmap_atomic(kaddr);
1930 flush_dcache_page(dst);
1932 copy_user_highpage(dst, src, va, vma);
1935 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
1937 struct file *vm_file = vma->vm_file;
1940 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
1943 * Special mappings (e.g. VDSO) do not have any file so fake
1944 * a default GFP_KERNEL for them.
1950 * Notify the address space that the page is about to become writable so that
1951 * it can prohibit this or wait for the page to get into an appropriate state.
1953 * We do this without the lock held, so that it can sleep if it needs to.
1955 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1956 unsigned long address)
1958 struct vm_fault vmf;
1961 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1962 vmf.pgoff = page->index;
1963 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1964 vmf.gfp_mask = __get_fault_gfp_mask(vma);
1966 vmf.cow_page = NULL;
1968 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
1969 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
1971 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
1973 if (!page->mapping) {
1975 return 0; /* retry */
1977 ret |= VM_FAULT_LOCKED;
1979 VM_BUG_ON_PAGE(!PageLocked(page), page);
1984 * Handle write page faults for pages that can be reused in the current vma
1986 * This can happen either due to the mapping being with the VM_SHARED flag,
1987 * or due to us being the last reference standing to the page. In either
1988 * case, all we need to do here is to mark the page as writable and update
1989 * any related book-keeping.
1991 static inline int wp_page_reuse(struct mm_struct *mm,
1992 struct vm_area_struct *vma, unsigned long address,
1993 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
1994 struct page *page, int page_mkwrite,
2000 * Clear the pages cpupid information as the existing
2001 * information potentially belongs to a now completely
2002 * unrelated process.
2005 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2007 flush_cache_page(vma, address, pte_pfn(orig_pte));
2008 entry = pte_mkyoung(orig_pte);
2009 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2010 if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2011 update_mmu_cache(vma, address, page_table);
2012 pte_unmap_unlock(page_table, ptl);
2015 struct address_space *mapping;
2021 dirtied = set_page_dirty(page);
2022 VM_BUG_ON_PAGE(PageAnon(page), page);
2023 mapping = page->mapping;
2025 page_cache_release(page);
2027 if ((dirtied || page_mkwrite) && mapping) {
2029 * Some device drivers do not set page.mapping
2030 * but still dirty their pages
2032 balance_dirty_pages_ratelimited(mapping);
2036 file_update_time(vma->vm_file);
2039 return VM_FAULT_WRITE;
2043 * Handle the case of a page which we actually need to copy to a new page.
2045 * Called with mmap_sem locked and the old page referenced, but
2046 * without the ptl held.
2048 * High level logic flow:
2050 * - Allocate a page, copy the content of the old page to the new one.
2051 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2052 * - Take the PTL. If the pte changed, bail out and release the allocated page
2053 * - If the pte is still the way we remember it, update the page table and all
2054 * relevant references. This includes dropping the reference the page-table
2055 * held to the old page, as well as updating the rmap.
2056 * - In any case, unlock the PTL and drop the reference we took to the old page.
2058 static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2059 unsigned long address, pte_t *page_table, pmd_t *pmd,
2060 pte_t orig_pte, struct page *old_page)
2062 struct page *new_page = NULL;
2063 spinlock_t *ptl = NULL;
2065 int page_copied = 0;
2066 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */
2067 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */
2068 struct mem_cgroup *memcg;
2070 if (unlikely(anon_vma_prepare(vma)))
2073 if (is_zero_pfn(pte_pfn(orig_pte))) {
2074 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2078 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2081 cow_user_page(new_page, old_page, address, vma);
2084 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2087 __SetPageUptodate(new_page);
2089 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2092 * Re-check the pte - we dropped the lock
2094 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2095 if (likely(pte_same(*page_table, orig_pte))) {
2097 if (!PageAnon(old_page)) {
2098 dec_mm_counter_fast(mm,
2099 mm_counter_file(old_page));
2100 inc_mm_counter_fast(mm, MM_ANONPAGES);
2103 inc_mm_counter_fast(mm, MM_ANONPAGES);
2105 flush_cache_page(vma, address, pte_pfn(orig_pte));
2106 entry = mk_pte(new_page, vma->vm_page_prot);
2107 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2109 * Clear the pte entry and flush it first, before updating the
2110 * pte with the new entry. This will avoid a race condition
2111 * seen in the presence of one thread doing SMC and another
2114 ptep_clear_flush_notify(vma, address, page_table);
2115 page_add_new_anon_rmap(new_page, vma, address, false);
2116 mem_cgroup_commit_charge(new_page, memcg, false, false);
2117 lru_cache_add_active_or_unevictable(new_page, vma);
2119 * We call the notify macro here because, when using secondary
2120 * mmu page tables (such as kvm shadow page tables), we want the
2121 * new page to be mapped directly into the secondary page table.
2123 set_pte_at_notify(mm, address, page_table, entry);
2124 update_mmu_cache(vma, address, page_table);
2127 * Only after switching the pte to the new page may
2128 * we remove the mapcount here. Otherwise another
2129 * process may come and find the rmap count decremented
2130 * before the pte is switched to the new page, and
2131 * "reuse" the old page writing into it while our pte
2132 * here still points into it and can be read by other
2135 * The critical issue is to order this
2136 * page_remove_rmap with the ptp_clear_flush above.
2137 * Those stores are ordered by (if nothing else,)
2138 * the barrier present in the atomic_add_negative
2139 * in page_remove_rmap.
2141 * Then the TLB flush in ptep_clear_flush ensures that
2142 * no process can access the old page before the
2143 * decremented mapcount is visible. And the old page
2144 * cannot be reused until after the decremented
2145 * mapcount is visible. So transitively, TLBs to
2146 * old page will be flushed before it can be reused.
2148 page_remove_rmap(old_page, false);
2151 /* Free the old page.. */
2152 new_page = old_page;
2155 mem_cgroup_cancel_charge(new_page, memcg, false);
2159 page_cache_release(new_page);
2161 pte_unmap_unlock(page_table, ptl);
2162 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2163 /* THP pages are never mlocked */
2164 if (old_page && !PageTransCompound(old_page)) {
2166 * Don't let another task, with possibly unlocked vma,
2167 * keep the mlocked page.
2169 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2170 lock_page(old_page); /* LRU manipulation */
2171 munlock_vma_page(old_page);
2172 unlock_page(old_page);
2174 page_cache_release(old_page);
2176 return page_copied ? VM_FAULT_WRITE : 0;
2178 page_cache_release(new_page);
2181 page_cache_release(old_page);
2182 return VM_FAULT_OOM;
2186 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2189 static int wp_pfn_shared(struct mm_struct *mm,
2190 struct vm_area_struct *vma, unsigned long address,
2191 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2194 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2195 struct vm_fault vmf = {
2197 .pgoff = linear_page_index(vma, address),
2198 .virtual_address = (void __user *)(address & PAGE_MASK),
2199 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2203 pte_unmap_unlock(page_table, ptl);
2204 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2205 if (ret & VM_FAULT_ERROR)
2207 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2209 * We might have raced with another page fault while we
2210 * released the pte_offset_map_lock.
2212 if (!pte_same(*page_table, orig_pte)) {
2213 pte_unmap_unlock(page_table, ptl);
2217 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2221 static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2222 unsigned long address, pte_t *page_table,
2223 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2224 struct page *old_page)
2227 int page_mkwrite = 0;
2229 page_cache_get(old_page);
2232 * Only catch write-faults on shared writable pages,
2233 * read-only shared pages can get COWed by
2234 * get_user_pages(.write=1, .force=1).
2236 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2239 pte_unmap_unlock(page_table, ptl);
2240 tmp = do_page_mkwrite(vma, old_page, address);
2241 if (unlikely(!tmp || (tmp &
2242 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2243 page_cache_release(old_page);
2247 * Since we dropped the lock we need to revalidate
2248 * the PTE as someone else may have changed it. If
2249 * they did, we just return, as we can count on the
2250 * MMU to tell us if they didn't also make it writable.
2252 page_table = pte_offset_map_lock(mm, pmd, address,
2254 if (!pte_same(*page_table, orig_pte)) {
2255 unlock_page(old_page);
2256 pte_unmap_unlock(page_table, ptl);
2257 page_cache_release(old_page);
2263 return wp_page_reuse(mm, vma, address, page_table, ptl,
2264 orig_pte, old_page, page_mkwrite, 1);
2268 * This routine handles present pages, when users try to write
2269 * to a shared page. It is done by copying the page to a new address
2270 * and decrementing the shared-page counter for the old page.
2272 * Note that this routine assumes that the protection checks have been
2273 * done by the caller (the low-level page fault routine in most cases).
2274 * Thus we can safely just mark it writable once we've done any necessary
2277 * We also mark the page dirty at this point even though the page will
2278 * change only once the write actually happens. This avoids a few races,
2279 * and potentially makes it more efficient.
2281 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2282 * but allow concurrent faults), with pte both mapped and locked.
2283 * We return with mmap_sem still held, but pte unmapped and unlocked.
2285 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2286 unsigned long address, pte_t *page_table, pmd_t *pmd,
2287 spinlock_t *ptl, pte_t orig_pte)
2290 struct page *old_page;
2292 old_page = vm_normal_page(vma, address, orig_pte);
2295 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2298 * We should not cow pages in a shared writeable mapping.
2299 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2301 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2302 (VM_WRITE|VM_SHARED))
2303 return wp_pfn_shared(mm, vma, address, page_table, ptl,
2306 pte_unmap_unlock(page_table, ptl);
2307 return wp_page_copy(mm, vma, address, page_table, pmd,
2308 orig_pte, old_page);
2312 * Take out anonymous pages first, anonymous shared vmas are
2313 * not dirty accountable.
2315 if (PageAnon(old_page) && !PageKsm(old_page)) {
2316 if (!trylock_page(old_page)) {
2317 page_cache_get(old_page);
2318 pte_unmap_unlock(page_table, ptl);
2319 lock_page(old_page);
2320 page_table = pte_offset_map_lock(mm, pmd, address,
2322 if (!pte_same(*page_table, orig_pte)) {
2323 unlock_page(old_page);
2324 pte_unmap_unlock(page_table, ptl);
2325 page_cache_release(old_page);
2328 page_cache_release(old_page);
2330 if (reuse_swap_page(old_page)) {
2332 * The page is all ours. Move it to our anon_vma so
2333 * the rmap code will not search our parent or siblings.
2334 * Protected against the rmap code by the page lock.
2336 page_move_anon_rmap(old_page, vma, address);
2337 unlock_page(old_page);
2338 return wp_page_reuse(mm, vma, address, page_table, ptl,
2339 orig_pte, old_page, 0, 0);
2341 unlock_page(old_page);
2342 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2343 (VM_WRITE|VM_SHARED))) {
2344 return wp_page_shared(mm, vma, address, page_table, pmd,
2345 ptl, orig_pte, old_page);
2349 * Ok, we need to copy. Oh, well..
2351 page_cache_get(old_page);
2353 pte_unmap_unlock(page_table, ptl);
2354 return wp_page_copy(mm, vma, address, page_table, pmd,
2355 orig_pte, old_page);
2358 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2359 unsigned long start_addr, unsigned long end_addr,
2360 struct zap_details *details)
2362 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2365 static inline void unmap_mapping_range_tree(struct rb_root *root,
2366 struct zap_details *details)
2368 struct vm_area_struct *vma;
2369 pgoff_t vba, vea, zba, zea;
2371 vma_interval_tree_foreach(vma, root,
2372 details->first_index, details->last_index) {
2374 vba = vma->vm_pgoff;
2375 vea = vba + vma_pages(vma) - 1;
2376 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2377 zba = details->first_index;
2380 zea = details->last_index;
2384 unmap_mapping_range_vma(vma,
2385 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2386 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2392 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2393 * address_space corresponding to the specified page range in the underlying
2396 * @mapping: the address space containing mmaps to be unmapped.
2397 * @holebegin: byte in first page to unmap, relative to the start of
2398 * the underlying file. This will be rounded down to a PAGE_SIZE
2399 * boundary. Note that this is different from truncate_pagecache(), which
2400 * must keep the partial page. In contrast, we must get rid of
2402 * @holelen: size of prospective hole in bytes. This will be rounded
2403 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2405 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2406 * but 0 when invalidating pagecache, don't throw away private data.
2408 void unmap_mapping_range(struct address_space *mapping,
2409 loff_t const holebegin, loff_t const holelen, int even_cows)
2411 struct zap_details details;
2412 pgoff_t hba = holebegin >> PAGE_SHIFT;
2413 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2415 /* Check for overflow. */
2416 if (sizeof(holelen) > sizeof(hlen)) {
2418 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2419 if (holeend & ~(long long)ULONG_MAX)
2420 hlen = ULONG_MAX - hba + 1;
2423 details.check_mapping = even_cows? NULL: mapping;
2424 details.first_index = hba;
2425 details.last_index = hba + hlen - 1;
2426 if (details.last_index < details.first_index)
2427 details.last_index = ULONG_MAX;
2430 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2431 i_mmap_lock_write(mapping);
2432 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2433 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2434 i_mmap_unlock_write(mapping);
2436 EXPORT_SYMBOL(unmap_mapping_range);
2439 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2440 * but allow concurrent faults), and pte mapped but not yet locked.
2441 * We return with pte unmapped and unlocked.
2443 * We return with the mmap_sem locked or unlocked in the same cases
2444 * as does filemap_fault().
2446 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2447 unsigned long address, pte_t *page_table, pmd_t *pmd,
2448 unsigned int flags, pte_t orig_pte)
2451 struct page *page, *swapcache;
2452 struct mem_cgroup *memcg;
2459 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2462 entry = pte_to_swp_entry(orig_pte);
2463 if (unlikely(non_swap_entry(entry))) {
2464 if (is_migration_entry(entry)) {
2465 migration_entry_wait(mm, pmd, address);
2466 } else if (is_hwpoison_entry(entry)) {
2467 ret = VM_FAULT_HWPOISON;
2469 print_bad_pte(vma, address, orig_pte, NULL);
2470 ret = VM_FAULT_SIGBUS;
2474 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2475 page = lookup_swap_cache(entry);
2477 page = swapin_readahead(entry,
2478 GFP_HIGHUSER_MOVABLE, vma, address);
2481 * Back out if somebody else faulted in this pte
2482 * while we released the pte lock.
2484 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2485 if (likely(pte_same(*page_table, orig_pte)))
2487 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2491 /* Had to read the page from swap area: Major fault */
2492 ret = VM_FAULT_MAJOR;
2493 count_vm_event(PGMAJFAULT);
2494 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2495 } else if (PageHWPoison(page)) {
2497 * hwpoisoned dirty swapcache pages are kept for killing
2498 * owner processes (which may be unknown at hwpoison time)
2500 ret = VM_FAULT_HWPOISON;
2501 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2507 locked = lock_page_or_retry(page, mm, flags);
2509 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2511 ret |= VM_FAULT_RETRY;
2516 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2517 * release the swapcache from under us. The page pin, and pte_same
2518 * test below, are not enough to exclude that. Even if it is still
2519 * swapcache, we need to check that the page's swap has not changed.
2521 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2524 page = ksm_might_need_to_copy(page, vma, address);
2525 if (unlikely(!page)) {
2531 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) {
2537 * Back out if somebody else already faulted in this pte.
2539 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2540 if (unlikely(!pte_same(*page_table, orig_pte)))
2543 if (unlikely(!PageUptodate(page))) {
2544 ret = VM_FAULT_SIGBUS;
2549 * The page isn't present yet, go ahead with the fault.
2551 * Be careful about the sequence of operations here.
2552 * To get its accounting right, reuse_swap_page() must be called
2553 * while the page is counted on swap but not yet in mapcount i.e.
2554 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2555 * must be called after the swap_free(), or it will never succeed.
2558 inc_mm_counter_fast(mm, MM_ANONPAGES);
2559 dec_mm_counter_fast(mm, MM_SWAPENTS);
2560 pte = mk_pte(page, vma->vm_page_prot);
2561 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2562 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2563 flags &= ~FAULT_FLAG_WRITE;
2564 ret |= VM_FAULT_WRITE;
2565 exclusive = RMAP_EXCLUSIVE;
2567 flush_icache_page(vma, page);
2568 if (pte_swp_soft_dirty(orig_pte))
2569 pte = pte_mksoft_dirty(pte);
2570 set_pte_at(mm, address, page_table, pte);
2571 if (page == swapcache) {
2572 do_page_add_anon_rmap(page, vma, address, exclusive);
2573 mem_cgroup_commit_charge(page, memcg, true, false);
2574 } else { /* ksm created a completely new copy */
2575 page_add_new_anon_rmap(page, vma, address, false);
2576 mem_cgroup_commit_charge(page, memcg, false, false);
2577 lru_cache_add_active_or_unevictable(page, vma);
2581 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2582 try_to_free_swap(page);
2584 if (page != swapcache) {
2586 * Hold the lock to avoid the swap entry to be reused
2587 * until we take the PT lock for the pte_same() check
2588 * (to avoid false positives from pte_same). For
2589 * further safety release the lock after the swap_free
2590 * so that the swap count won't change under a
2591 * parallel locked swapcache.
2593 unlock_page(swapcache);
2594 page_cache_release(swapcache);
2597 if (flags & FAULT_FLAG_WRITE) {
2598 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2599 if (ret & VM_FAULT_ERROR)
2600 ret &= VM_FAULT_ERROR;
2604 /* No need to invalidate - it was non-present before */
2605 update_mmu_cache(vma, address, page_table);
2607 pte_unmap_unlock(page_table, ptl);
2611 mem_cgroup_cancel_charge(page, memcg, false);
2612 pte_unmap_unlock(page_table, ptl);
2616 page_cache_release(page);
2617 if (page != swapcache) {
2618 unlock_page(swapcache);
2619 page_cache_release(swapcache);
2625 * This is like a special single-page "expand_{down|up}wards()",
2626 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2627 * doesn't hit another vma.
2629 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2631 address &= PAGE_MASK;
2632 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2633 struct vm_area_struct *prev = vma->vm_prev;
2636 * Is there a mapping abutting this one below?
2638 * That's only ok if it's the same stack mapping
2639 * that has gotten split..
2641 if (prev && prev->vm_end == address)
2642 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2644 return expand_downwards(vma, address - PAGE_SIZE);
2646 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2647 struct vm_area_struct *next = vma->vm_next;
2649 /* As VM_GROWSDOWN but s/below/above/ */
2650 if (next && next->vm_start == address + PAGE_SIZE)
2651 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2653 return expand_upwards(vma, address + PAGE_SIZE);
2659 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2660 * but allow concurrent faults), and pte mapped but not yet locked.
2661 * We return with mmap_sem still held, but pte unmapped and unlocked.
2663 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2664 unsigned long address, pte_t *page_table, pmd_t *pmd,
2667 struct mem_cgroup *memcg;
2672 pte_unmap(page_table);
2674 /* File mapping without ->vm_ops ? */
2675 if (vma->vm_flags & VM_SHARED)
2676 return VM_FAULT_SIGBUS;
2678 /* Check if we need to add a guard page to the stack */
2679 if (check_stack_guard_page(vma, address) < 0)
2680 return VM_FAULT_SIGSEGV;
2682 /* Use the zero-page for reads */
2683 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2684 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2685 vma->vm_page_prot));
2686 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2687 if (!pte_none(*page_table))
2689 /* Deliver the page fault to userland, check inside PT lock */
2690 if (userfaultfd_missing(vma)) {
2691 pte_unmap_unlock(page_table, ptl);
2692 return handle_userfault(vma, address, flags,
2698 /* Allocate our own private page. */
2699 if (unlikely(anon_vma_prepare(vma)))
2701 page = alloc_zeroed_user_highpage_movable(vma, address);
2705 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false))
2709 * The memory barrier inside __SetPageUptodate makes sure that
2710 * preceeding stores to the page contents become visible before
2711 * the set_pte_at() write.
2713 __SetPageUptodate(page);
2715 entry = mk_pte(page, vma->vm_page_prot);
2716 if (vma->vm_flags & VM_WRITE)
2717 entry = pte_mkwrite(pte_mkdirty(entry));
2719 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2720 if (!pte_none(*page_table))
2723 /* Deliver the page fault to userland, check inside PT lock */
2724 if (userfaultfd_missing(vma)) {
2725 pte_unmap_unlock(page_table, ptl);
2726 mem_cgroup_cancel_charge(page, memcg, false);
2727 page_cache_release(page);
2728 return handle_userfault(vma, address, flags,
2732 inc_mm_counter_fast(mm, MM_ANONPAGES);
2733 page_add_new_anon_rmap(page, vma, address, false);
2734 mem_cgroup_commit_charge(page, memcg, false, false);
2735 lru_cache_add_active_or_unevictable(page, vma);
2737 set_pte_at(mm, address, page_table, entry);
2739 /* No need to invalidate - it was non-present before */
2740 update_mmu_cache(vma, address, page_table);
2742 pte_unmap_unlock(page_table, ptl);
2745 mem_cgroup_cancel_charge(page, memcg, false);
2746 page_cache_release(page);
2749 page_cache_release(page);
2751 return VM_FAULT_OOM;
2755 * The mmap_sem must have been held on entry, and may have been
2756 * released depending on flags and vma->vm_ops->fault() return value.
2757 * See filemap_fault() and __lock_page_retry().
2759 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2760 pgoff_t pgoff, unsigned int flags,
2761 struct page *cow_page, struct page **page)
2763 struct vm_fault vmf;
2766 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2770 vmf.gfp_mask = __get_fault_gfp_mask(vma);
2771 vmf.cow_page = cow_page;
2773 ret = vma->vm_ops->fault(vma, &vmf);
2774 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2779 if (unlikely(PageHWPoison(vmf.page))) {
2780 if (ret & VM_FAULT_LOCKED)
2781 unlock_page(vmf.page);
2782 page_cache_release(vmf.page);
2783 return VM_FAULT_HWPOISON;
2786 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2787 lock_page(vmf.page);
2789 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2797 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2799 * @vma: virtual memory area
2800 * @address: user virtual address
2801 * @page: page to map
2802 * @pte: pointer to target page table entry
2803 * @write: true, if new entry is writable
2804 * @anon: true, if it's anonymous page
2806 * Caller must hold page table lock relevant for @pte.
2808 * Target users are page handler itself and implementations of
2809 * vm_ops->map_pages.
2811 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2812 struct page *page, pte_t *pte, bool write, bool anon)
2816 flush_icache_page(vma, page);
2817 entry = mk_pte(page, vma->vm_page_prot);
2819 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2821 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2822 page_add_new_anon_rmap(page, vma, address, false);
2824 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
2825 page_add_file_rmap(page);
2827 set_pte_at(vma->vm_mm, address, pte, entry);
2829 /* no need to invalidate: a not-present page won't be cached */
2830 update_mmu_cache(vma, address, pte);
2833 static unsigned long fault_around_bytes __read_mostly =
2834 rounddown_pow_of_two(65536);
2836 #ifdef CONFIG_DEBUG_FS
2837 static int fault_around_bytes_get(void *data, u64 *val)
2839 *val = fault_around_bytes;
2844 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2845 * rounded down to nearest page order. It's what do_fault_around() expects to
2848 static int fault_around_bytes_set(void *data, u64 val)
2850 if (val / PAGE_SIZE > PTRS_PER_PTE)
2852 if (val > PAGE_SIZE)
2853 fault_around_bytes = rounddown_pow_of_two(val);
2855 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2858 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2859 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2861 static int __init fault_around_debugfs(void)
2865 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2866 &fault_around_bytes_fops);
2868 pr_warn("Failed to create fault_around_bytes in debugfs");
2871 late_initcall(fault_around_debugfs);
2875 * do_fault_around() tries to map few pages around the fault address. The hope
2876 * is that the pages will be needed soon and this will lower the number of
2879 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2880 * not ready to be mapped: not up-to-date, locked, etc.
2882 * This function is called with the page table lock taken. In the split ptlock
2883 * case the page table lock only protects only those entries which belong to
2884 * the page table corresponding to the fault address.
2886 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2889 * fault_around_pages() defines how many pages we'll try to map.
2890 * do_fault_around() expects it to return a power of two less than or equal to
2893 * The virtual address of the area that we map is naturally aligned to the
2894 * fault_around_pages() value (and therefore to page order). This way it's
2895 * easier to guarantee that we don't cross page table boundaries.
2897 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2898 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2900 unsigned long start_addr, nr_pages, mask;
2902 struct vm_fault vmf;
2905 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2906 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2908 start_addr = max(address & mask, vma->vm_start);
2909 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2914 * max_pgoff is either end of page table or end of vma
2915 * or fault_around_pages() from pgoff, depending what is nearest.
2917 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2919 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2920 pgoff + nr_pages - 1);
2922 /* Check if it makes any sense to call ->map_pages */
2923 while (!pte_none(*pte)) {
2924 if (++pgoff > max_pgoff)
2926 start_addr += PAGE_SIZE;
2927 if (start_addr >= vma->vm_end)
2932 vmf.virtual_address = (void __user *) start_addr;
2935 vmf.max_pgoff = max_pgoff;
2937 vmf.gfp_mask = __get_fault_gfp_mask(vma);
2938 vma->vm_ops->map_pages(vma, &vmf);
2941 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2942 unsigned long address, pmd_t *pmd,
2943 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2945 struct page *fault_page;
2951 * Let's call ->map_pages() first and use ->fault() as fallback
2952 * if page by the offset is not ready to be mapped (cold cache or
2955 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
2956 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2957 do_fault_around(vma, address, pte, pgoff, flags);
2958 if (!pte_same(*pte, orig_pte))
2960 pte_unmap_unlock(pte, ptl);
2963 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
2964 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2967 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2968 if (unlikely(!pte_same(*pte, orig_pte))) {
2969 pte_unmap_unlock(pte, ptl);
2970 unlock_page(fault_page);
2971 page_cache_release(fault_page);
2974 do_set_pte(vma, address, fault_page, pte, false, false);
2975 unlock_page(fault_page);
2977 pte_unmap_unlock(pte, ptl);
2981 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2982 unsigned long address, pmd_t *pmd,
2983 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2985 struct page *fault_page, *new_page;
2986 struct mem_cgroup *memcg;
2991 if (unlikely(anon_vma_prepare(vma)))
2992 return VM_FAULT_OOM;
2994 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2996 return VM_FAULT_OOM;
2998 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) {
2999 page_cache_release(new_page);
3000 return VM_FAULT_OOM;
3003 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
3004 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3008 copy_user_highpage(new_page, fault_page, address, vma);
3009 __SetPageUptodate(new_page);
3011 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3012 if (unlikely(!pte_same(*pte, orig_pte))) {
3013 pte_unmap_unlock(pte, ptl);
3015 unlock_page(fault_page);
3016 page_cache_release(fault_page);
3019 * The fault handler has no page to lock, so it holds
3020 * i_mmap_lock for read to protect against truncate.
3022 i_mmap_unlock_read(vma->vm_file->f_mapping);
3026 do_set_pte(vma, address, new_page, pte, true, true);
3027 mem_cgroup_commit_charge(new_page, memcg, false, false);
3028 lru_cache_add_active_or_unevictable(new_page, vma);
3029 pte_unmap_unlock(pte, ptl);
3031 unlock_page(fault_page);
3032 page_cache_release(fault_page);
3035 * The fault handler has no page to lock, so it holds
3036 * i_mmap_lock for read to protect against truncate.
3038 i_mmap_unlock_read(vma->vm_file->f_mapping);
3042 mem_cgroup_cancel_charge(new_page, memcg, false);
3043 page_cache_release(new_page);
3047 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3048 unsigned long address, pmd_t *pmd,
3049 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3051 struct page *fault_page;
3052 struct address_space *mapping;
3058 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3059 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3063 * Check if the backing address space wants to know that the page is
3064 * about to become writable
3066 if (vma->vm_ops->page_mkwrite) {
3067 unlock_page(fault_page);
3068 tmp = do_page_mkwrite(vma, fault_page, address);
3069 if (unlikely(!tmp ||
3070 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3071 page_cache_release(fault_page);
3076 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3077 if (unlikely(!pte_same(*pte, orig_pte))) {
3078 pte_unmap_unlock(pte, ptl);
3079 unlock_page(fault_page);
3080 page_cache_release(fault_page);
3083 do_set_pte(vma, address, fault_page, pte, true, false);
3084 pte_unmap_unlock(pte, ptl);
3086 if (set_page_dirty(fault_page))
3089 * Take a local copy of the address_space - page.mapping may be zeroed
3090 * by truncate after unlock_page(). The address_space itself remains
3091 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3092 * release semantics to prevent the compiler from undoing this copying.
3094 mapping = page_rmapping(fault_page);
3095 unlock_page(fault_page);
3096 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3098 * Some device drivers do not set page.mapping but still
3101 balance_dirty_pages_ratelimited(mapping);
3104 if (!vma->vm_ops->page_mkwrite)
3105 file_update_time(vma->vm_file);
3111 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3112 * but allow concurrent faults).
3113 * The mmap_sem may have been released depending on flags and our
3114 * return value. See filemap_fault() and __lock_page_or_retry().
3116 static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3117 unsigned long address, pte_t *page_table, pmd_t *pmd,
3118 unsigned int flags, pte_t orig_pte)
3120 pgoff_t pgoff = (((address & PAGE_MASK)
3121 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3123 pte_unmap(page_table);
3124 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3125 if (!vma->vm_ops->fault)
3126 return VM_FAULT_SIGBUS;
3127 if (!(flags & FAULT_FLAG_WRITE))
3128 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3130 if (!(vma->vm_flags & VM_SHARED))
3131 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3133 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3136 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3137 unsigned long addr, int page_nid,
3142 count_vm_numa_event(NUMA_HINT_FAULTS);
3143 if (page_nid == numa_node_id()) {
3144 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3145 *flags |= TNF_FAULT_LOCAL;
3148 return mpol_misplaced(page, vma, addr);
3151 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3152 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3154 struct page *page = NULL;
3159 bool migrated = false;
3160 bool was_writable = pte_write(pte);
3163 /* A PROT_NONE fault should not end up here */
3164 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3167 * The "pte" at this point cannot be used safely without
3168 * validation through pte_unmap_same(). It's of NUMA type but
3169 * the pfn may be screwed if the read is non atomic.
3171 * We can safely just do a "set_pte_at()", because the old
3172 * page table entry is not accessible, so there would be no
3173 * concurrent hardware modifications to the PTE.
3175 ptl = pte_lockptr(mm, pmd);
3177 if (unlikely(!pte_same(*ptep, pte))) {
3178 pte_unmap_unlock(ptep, ptl);
3182 /* Make it present again */
3183 pte = pte_modify(pte, vma->vm_page_prot);
3184 pte = pte_mkyoung(pte);
3186 pte = pte_mkwrite(pte);
3187 set_pte_at(mm, addr, ptep, pte);
3188 update_mmu_cache(vma, addr, ptep);
3190 page = vm_normal_page(vma, addr, pte);
3192 pte_unmap_unlock(ptep, ptl);
3196 /* TODO: handle PTE-mapped THP */
3197 if (PageCompound(page)) {
3198 pte_unmap_unlock(ptep, ptl);
3203 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3204 * much anyway since they can be in shared cache state. This misses
3205 * the case where a mapping is writable but the process never writes
3206 * to it but pte_write gets cleared during protection updates and
3207 * pte_dirty has unpredictable behaviour between PTE scan updates,
3208 * background writeback, dirty balancing and application behaviour.
3210 if (!(vma->vm_flags & VM_WRITE))
3211 flags |= TNF_NO_GROUP;
3214 * Flag if the page is shared between multiple address spaces. This
3215 * is later used when determining whether to group tasks together
3217 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3218 flags |= TNF_SHARED;
3220 last_cpupid = page_cpupid_last(page);
3221 page_nid = page_to_nid(page);
3222 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3223 pte_unmap_unlock(ptep, ptl);
3224 if (target_nid == -1) {
3229 /* Migrate to the requested node */
3230 migrated = migrate_misplaced_page(page, vma, target_nid);
3232 page_nid = target_nid;
3233 flags |= TNF_MIGRATED;
3235 flags |= TNF_MIGRATE_FAIL;
3239 task_numa_fault(last_cpupid, page_nid, 1, flags);
3243 static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3244 unsigned long address, pmd_t *pmd, unsigned int flags)
3246 if (vma_is_anonymous(vma))
3247 return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3248 if (vma->vm_ops->pmd_fault)
3249 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3250 return VM_FAULT_FALLBACK;
3253 static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3254 unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3257 if (vma_is_anonymous(vma))
3258 return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3259 if (vma->vm_ops->pmd_fault)
3260 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3261 return VM_FAULT_FALLBACK;
3265 * These routines also need to handle stuff like marking pages dirty
3266 * and/or accessed for architectures that don't do it in hardware (most
3267 * RISC architectures). The early dirtying is also good on the i386.
3269 * There is also a hook called "update_mmu_cache()" that architectures
3270 * with external mmu caches can use to update those (ie the Sparc or
3271 * PowerPC hashed page tables that act as extended TLBs).
3273 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3274 * but allow concurrent faults), and pte mapped but not yet locked.
3275 * We return with pte unmapped and unlocked.
3277 * The mmap_sem may have been released depending on flags and our
3278 * return value. See filemap_fault() and __lock_page_or_retry().
3280 static int handle_pte_fault(struct mm_struct *mm,
3281 struct vm_area_struct *vma, unsigned long address,
3282 pte_t *pte, pmd_t *pmd, unsigned int flags)
3288 * some architectures can have larger ptes than wordsize,
3289 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3290 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3291 * The code below just needs a consistent view for the ifs and
3292 * we later double check anyway with the ptl lock held. So here
3293 * a barrier will do.
3297 if (!pte_present(entry)) {
3298 if (pte_none(entry)) {
3299 if (vma_is_anonymous(vma))
3300 return do_anonymous_page(mm, vma, address,
3303 return do_fault(mm, vma, address, pte, pmd,
3306 return do_swap_page(mm, vma, address,
3307 pte, pmd, flags, entry);
3310 if (pte_protnone(entry))
3311 return do_numa_page(mm, vma, address, entry, pte, pmd);
3313 ptl = pte_lockptr(mm, pmd);
3315 if (unlikely(!pte_same(*pte, entry)))
3317 if (flags & FAULT_FLAG_WRITE) {
3318 if (!pte_write(entry))
3319 return do_wp_page(mm, vma, address,
3320 pte, pmd, ptl, entry);
3321 entry = pte_mkdirty(entry);
3323 entry = pte_mkyoung(entry);
3324 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3325 update_mmu_cache(vma, address, pte);
3328 * This is needed only for protection faults but the arch code
3329 * is not yet telling us if this is a protection fault or not.
3330 * This still avoids useless tlb flushes for .text page faults
3333 if (flags & FAULT_FLAG_WRITE)
3334 flush_tlb_fix_spurious_fault(vma, address);
3337 pte_unmap_unlock(pte, ptl);
3342 * By the time we get here, we already hold the mm semaphore
3344 * The mmap_sem may have been released depending on flags and our
3345 * return value. See filemap_fault() and __lock_page_or_retry().
3347 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3348 unsigned long address, unsigned int flags)
3355 if (unlikely(is_vm_hugetlb_page(vma)))
3356 return hugetlb_fault(mm, vma, address, flags);
3358 pgd = pgd_offset(mm, address);
3359 pud = pud_alloc(mm, pgd, address);
3361 return VM_FAULT_OOM;
3362 pmd = pmd_alloc(mm, pud, address);
3364 return VM_FAULT_OOM;
3365 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3366 int ret = create_huge_pmd(mm, vma, address, pmd, flags);
3367 if (!(ret & VM_FAULT_FALLBACK))
3370 pmd_t orig_pmd = *pmd;
3374 if (pmd_trans_huge(orig_pmd)) {
3375 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3377 if (pmd_protnone(orig_pmd))
3378 return do_huge_pmd_numa_page(mm, vma, address,
3381 if (dirty && !pmd_write(orig_pmd)) {
3382 ret = wp_huge_pmd(mm, vma, address, pmd,
3384 if (!(ret & VM_FAULT_FALLBACK))
3387 huge_pmd_set_accessed(mm, vma, address, pmd,
3395 * Use __pte_alloc instead of pte_alloc_map, because we can't
3396 * run pte_offset_map on the pmd, if an huge pmd could
3397 * materialize from under us from a different thread.
3399 if (unlikely(pmd_none(*pmd)) &&
3400 unlikely(__pte_alloc(mm, vma, pmd, address)))
3401 return VM_FAULT_OOM;
3402 /* if an huge pmd materialized from under us just retry later */
3403 if (unlikely(pmd_trans_huge(*pmd)))
3406 * A regular pmd is established and it can't morph into a huge pmd
3407 * from under us anymore at this point because we hold the mmap_sem
3408 * read mode and khugepaged takes it in write mode. So now it's
3409 * safe to run pte_offset_map().
3411 pte = pte_offset_map(pmd, address);
3413 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3417 * By the time we get here, we already hold the mm semaphore
3419 * The mmap_sem may have been released depending on flags and our
3420 * return value. See filemap_fault() and __lock_page_or_retry().
3422 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3423 unsigned long address, unsigned int flags)
3427 __set_current_state(TASK_RUNNING);
3429 count_vm_event(PGFAULT);
3430 mem_cgroup_count_vm_event(mm, PGFAULT);
3432 /* do counter updates before entering really critical section. */
3433 check_sync_rss_stat(current);
3436 * Enable the memcg OOM handling for faults triggered in user
3437 * space. Kernel faults are handled more gracefully.
3439 if (flags & FAULT_FLAG_USER)
3440 mem_cgroup_oom_enable();
3442 ret = __handle_mm_fault(mm, vma, address, flags);
3444 if (flags & FAULT_FLAG_USER) {
3445 mem_cgroup_oom_disable();
3447 * The task may have entered a memcg OOM situation but
3448 * if the allocation error was handled gracefully (no
3449 * VM_FAULT_OOM), there is no need to kill anything.
3450 * Just clean up the OOM state peacefully.
3452 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3453 mem_cgroup_oom_synchronize(false);
3458 EXPORT_SYMBOL_GPL(handle_mm_fault);
3460 #ifndef __PAGETABLE_PUD_FOLDED
3462 * Allocate page upper directory.
3463 * We've already handled the fast-path in-line.
3465 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3467 pud_t *new = pud_alloc_one(mm, address);
3471 smp_wmb(); /* See comment in __pte_alloc */
3473 spin_lock(&mm->page_table_lock);
3474 if (pgd_present(*pgd)) /* Another has populated it */
3477 pgd_populate(mm, pgd, new);
3478 spin_unlock(&mm->page_table_lock);
3481 #endif /* __PAGETABLE_PUD_FOLDED */
3483 #ifndef __PAGETABLE_PMD_FOLDED
3485 * Allocate page middle directory.
3486 * We've already handled the fast-path in-line.
3488 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3490 pmd_t *new = pmd_alloc_one(mm, address);
3494 smp_wmb(); /* See comment in __pte_alloc */
3496 spin_lock(&mm->page_table_lock);
3497 #ifndef __ARCH_HAS_4LEVEL_HACK
3498 if (!pud_present(*pud)) {
3500 pud_populate(mm, pud, new);
3501 } else /* Another has populated it */
3504 if (!pgd_present(*pud)) {
3506 pgd_populate(mm, pud, new);
3507 } else /* Another has populated it */
3509 #endif /* __ARCH_HAS_4LEVEL_HACK */
3510 spin_unlock(&mm->page_table_lock);
3513 #endif /* __PAGETABLE_PMD_FOLDED */
3515 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3516 pte_t **ptepp, spinlock_t **ptlp)
3523 pgd = pgd_offset(mm, address);
3524 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3527 pud = pud_offset(pgd, address);
3528 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3531 pmd = pmd_offset(pud, address);
3532 VM_BUG_ON(pmd_trans_huge(*pmd));
3533 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3536 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3540 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3543 if (!pte_present(*ptep))
3548 pte_unmap_unlock(ptep, *ptlp);
3553 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3554 pte_t **ptepp, spinlock_t **ptlp)
3558 /* (void) is needed to make gcc happy */
3559 (void) __cond_lock(*ptlp,
3560 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3565 * follow_pfn - look up PFN at a user virtual address
3566 * @vma: memory mapping
3567 * @address: user virtual address
3568 * @pfn: location to store found PFN
3570 * Only IO mappings and raw PFN mappings are allowed.
3572 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3574 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3581 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3584 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3587 *pfn = pte_pfn(*ptep);
3588 pte_unmap_unlock(ptep, ptl);
3591 EXPORT_SYMBOL(follow_pfn);
3593 #ifdef CONFIG_HAVE_IOREMAP_PROT
3594 int follow_phys(struct vm_area_struct *vma,
3595 unsigned long address, unsigned int flags,
3596 unsigned long *prot, resource_size_t *phys)
3602 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3605 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3609 if ((flags & FOLL_WRITE) && !pte_write(pte))
3612 *prot = pgprot_val(pte_pgprot(pte));
3613 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3617 pte_unmap_unlock(ptep, ptl);
3622 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3623 void *buf, int len, int write)
3625 resource_size_t phys_addr;
3626 unsigned long prot = 0;
3627 void __iomem *maddr;
3628 int offset = addr & (PAGE_SIZE-1);
3630 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3633 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3635 memcpy_toio(maddr + offset, buf, len);
3637 memcpy_fromio(buf, maddr + offset, len);
3642 EXPORT_SYMBOL_GPL(generic_access_phys);
3646 * Access another process' address space as given in mm. If non-NULL, use the
3647 * given task for page fault accounting.
3649 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3650 unsigned long addr, void *buf, int len, int write)
3652 struct vm_area_struct *vma;
3653 void *old_buf = buf;
3655 down_read(&mm->mmap_sem);
3656 /* ignore errors, just check how much was successfully transferred */
3658 int bytes, ret, offset;
3660 struct page *page = NULL;
3662 ret = get_user_pages(tsk, mm, addr, 1,
3663 write, 1, &page, &vma);
3665 #ifndef CONFIG_HAVE_IOREMAP_PROT
3669 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3670 * we can access using slightly different code.
3672 vma = find_vma(mm, addr);
3673 if (!vma || vma->vm_start > addr)
3675 if (vma->vm_ops && vma->vm_ops->access)
3676 ret = vma->vm_ops->access(vma, addr, buf,
3684 offset = addr & (PAGE_SIZE-1);
3685 if (bytes > PAGE_SIZE-offset)
3686 bytes = PAGE_SIZE-offset;
3690 copy_to_user_page(vma, page, addr,
3691 maddr + offset, buf, bytes);
3692 set_page_dirty_lock(page);
3694 copy_from_user_page(vma, page, addr,
3695 buf, maddr + offset, bytes);
3698 page_cache_release(page);
3704 up_read(&mm->mmap_sem);
3706 return buf - old_buf;
3710 * access_remote_vm - access another process' address space
3711 * @mm: the mm_struct of the target address space
3712 * @addr: start address to access
3713 * @buf: source or destination buffer
3714 * @len: number of bytes to transfer
3715 * @write: whether the access is a write
3717 * The caller must hold a reference on @mm.
3719 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3720 void *buf, int len, int write)
3722 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3726 * Access another process' address space.
3727 * Source/target buffer must be kernel space,
3728 * Do not walk the page table directly, use get_user_pages
3730 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3731 void *buf, int len, int write)
3733 struct mm_struct *mm;
3736 mm = get_task_mm(tsk);
3740 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3747 * Print the name of a VMA.
3749 void print_vma_addr(char *prefix, unsigned long ip)
3751 struct mm_struct *mm = current->mm;
3752 struct vm_area_struct *vma;
3755 * Do not print if we are in atomic
3756 * contexts (in exception stacks, etc.):
3758 if (preempt_count())
3761 down_read(&mm->mmap_sem);
3762 vma = find_vma(mm, ip);
3763 if (vma && vma->vm_file) {
3764 struct file *f = vma->vm_file;
3765 char *buf = (char *)__get_free_page(GFP_KERNEL);
3769 p = file_path(f, buf, PAGE_SIZE);
3772 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3774 vma->vm_end - vma->vm_start);
3775 free_page((unsigned long)buf);
3778 up_read(&mm->mmap_sem);
3781 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3782 void __might_fault(const char *file, int line)
3785 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3786 * holding the mmap_sem, this is safe because kernel memory doesn't
3787 * get paged out, therefore we'll never actually fault, and the
3788 * below annotations will generate false positives.
3790 if (segment_eq(get_fs(), KERNEL_DS))
3792 if (pagefault_disabled())
3794 __might_sleep(file, line, 0);
3795 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3797 might_lock_read(¤t->mm->mmap_sem);
3800 EXPORT_SYMBOL(__might_fault);
3803 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3804 static void clear_gigantic_page(struct page *page,
3806 unsigned int pages_per_huge_page)
3809 struct page *p = page;
3812 for (i = 0; i < pages_per_huge_page;
3813 i++, p = mem_map_next(p, page, i)) {
3815 clear_user_highpage(p, addr + i * PAGE_SIZE);
3818 void clear_huge_page(struct page *page,
3819 unsigned long addr, unsigned int pages_per_huge_page)
3823 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3824 clear_gigantic_page(page, addr, pages_per_huge_page);
3829 for (i = 0; i < pages_per_huge_page; i++) {
3831 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3835 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3837 struct vm_area_struct *vma,
3838 unsigned int pages_per_huge_page)
3841 struct page *dst_base = dst;
3842 struct page *src_base = src;
3844 for (i = 0; i < pages_per_huge_page; ) {
3846 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3849 dst = mem_map_next(dst, dst_base, i);
3850 src = mem_map_next(src, src_base, i);
3854 void copy_user_huge_page(struct page *dst, struct page *src,
3855 unsigned long addr, struct vm_area_struct *vma,
3856 unsigned int pages_per_huge_page)
3860 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3861 copy_user_gigantic_page(dst, src, addr, vma,
3862 pages_per_huge_page);
3867 for (i = 0; i < pages_per_huge_page; i++) {
3869 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3872 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3874 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3876 static struct kmem_cache *page_ptl_cachep;
3878 void __init ptlock_cache_init(void)
3880 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3884 bool ptlock_alloc(struct page *page)
3888 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3895 void ptlock_free(struct page *page)
3897 kmem_cache_free(page_ptl_cachep, page->ptl);