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/pfn_t.h>
54 #include <linux/writeback.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/kallsyms.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60 #include <linux/gfp.h>
61 #include <linux/migrate.h>
62 #include <linux/string.h>
63 #include <linux/dma-debug.h>
64 #include <linux/debugfs.h>
65 #include <linux/userfaultfd_k.h>
68 #include <asm/pgalloc.h>
69 #include <asm/uaccess.h>
71 #include <asm/tlbflush.h>
72 #include <asm/pgtable.h>
76 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
77 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
80 #ifndef CONFIG_NEED_MULTIPLE_NODES
81 /* use the per-pgdat data instead for discontigmem - mbligh */
82 unsigned long max_mapnr;
85 EXPORT_SYMBOL(max_mapnr);
86 EXPORT_SYMBOL(mem_map);
90 * A number of key systems in x86 including ioremap() rely on the assumption
91 * that high_memory defines the upper bound on direct map memory, then end
92 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
93 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
98 EXPORT_SYMBOL(high_memory);
101 * Randomize the address space (stacks, mmaps, brk, etc.).
103 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
104 * as ancient (libc5 based) binaries can segfault. )
106 int randomize_va_space __read_mostly =
107 #ifdef CONFIG_COMPAT_BRK
113 static int __init disable_randmaps(char *s)
115 randomize_va_space = 0;
118 __setup("norandmaps", disable_randmaps);
120 unsigned long zero_pfn __read_mostly;
121 unsigned long highest_memmap_pfn __read_mostly;
123 EXPORT_SYMBOL(zero_pfn);
126 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
128 static int __init init_zero_pfn(void)
130 zero_pfn = page_to_pfn(ZERO_PAGE(0));
133 core_initcall(init_zero_pfn);
136 #if defined(SPLIT_RSS_COUNTING)
138 void sync_mm_rss(struct mm_struct *mm)
142 for (i = 0; i < NR_MM_COUNTERS; i++) {
143 if (current->rss_stat.count[i]) {
144 add_mm_counter(mm, i, current->rss_stat.count[i]);
145 current->rss_stat.count[i] = 0;
148 current->rss_stat.events = 0;
151 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
153 struct task_struct *task = current;
155 if (likely(task->mm == mm))
156 task->rss_stat.count[member] += val;
158 add_mm_counter(mm, member, val);
160 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
161 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
163 /* sync counter once per 64 page faults */
164 #define TASK_RSS_EVENTS_THRESH (64)
165 static void check_sync_rss_stat(struct task_struct *task)
167 if (unlikely(task != current))
169 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
170 sync_mm_rss(task->mm);
172 #else /* SPLIT_RSS_COUNTING */
174 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
175 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
177 static void check_sync_rss_stat(struct task_struct *task)
181 #endif /* SPLIT_RSS_COUNTING */
183 #ifdef HAVE_GENERIC_MMU_GATHER
185 static bool tlb_next_batch(struct mmu_gather *tlb)
187 struct mmu_gather_batch *batch;
191 tlb->active = batch->next;
195 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
198 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
205 batch->max = MAX_GATHER_BATCH;
207 tlb->active->next = batch;
214 * Called to initialize an (on-stack) mmu_gather structure for page-table
215 * tear-down from @mm. The @fullmm argument is used when @mm is without
216 * users and we're going to destroy the full address space (exit/execve).
218 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
222 /* Is it from 0 to ~0? */
223 tlb->fullmm = !(start | (end+1));
224 tlb->need_flush_all = 0;
225 tlb->local.next = NULL;
227 tlb->local.max = ARRAY_SIZE(tlb->__pages);
228 tlb->active = &tlb->local;
229 tlb->batch_count = 0;
231 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
235 __tlb_reset_range(tlb);
238 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
244 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
245 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
246 tlb_table_flush(tlb);
248 __tlb_reset_range(tlb);
251 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
253 struct mmu_gather_batch *batch;
255 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
256 free_pages_and_swap_cache(batch->pages, batch->nr);
259 tlb->active = &tlb->local;
262 void tlb_flush_mmu(struct mmu_gather *tlb)
264 tlb_flush_mmu_tlbonly(tlb);
265 tlb_flush_mmu_free(tlb);
269 * Called at the end of the shootdown operation to free up any resources
270 * that were required.
272 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
274 struct mmu_gather_batch *batch, *next;
278 /* keep the page table cache within bounds */
281 for (batch = tlb->local.next; batch; batch = next) {
283 free_pages((unsigned long)batch, 0);
285 tlb->local.next = NULL;
289 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
290 * handling the additional races in SMP caused by other CPUs caching valid
291 * mappings in their TLBs. Returns the number of free page slots left.
292 * When out of page slots we must call tlb_flush_mmu().
294 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
296 struct mmu_gather_batch *batch;
298 VM_BUG_ON(!tlb->end);
301 batch->pages[batch->nr++] = page;
302 if (batch->nr == batch->max) {
303 if (!tlb_next_batch(tlb))
307 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
309 return batch->max - batch->nr;
312 #endif /* HAVE_GENERIC_MMU_GATHER */
314 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
317 * See the comment near struct mmu_table_batch.
320 static void tlb_remove_table_smp_sync(void *arg)
322 /* Simply deliver the interrupt */
325 static void tlb_remove_table_one(void *table)
328 * This isn't an RCU grace period and hence the page-tables cannot be
329 * assumed to be actually RCU-freed.
331 * It is however sufficient for software page-table walkers that rely on
332 * IRQ disabling. See the comment near struct mmu_table_batch.
334 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
335 __tlb_remove_table(table);
338 static void tlb_remove_table_rcu(struct rcu_head *head)
340 struct mmu_table_batch *batch;
343 batch = container_of(head, struct mmu_table_batch, rcu);
345 for (i = 0; i < batch->nr; i++)
346 __tlb_remove_table(batch->tables[i]);
348 free_page((unsigned long)batch);
351 void tlb_table_flush(struct mmu_gather *tlb)
353 struct mmu_table_batch **batch = &tlb->batch;
356 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
361 void tlb_remove_table(struct mmu_gather *tlb, void *table)
363 struct mmu_table_batch **batch = &tlb->batch;
366 * When there's less then two users of this mm there cannot be a
367 * concurrent page-table walk.
369 if (atomic_read(&tlb->mm->mm_users) < 2) {
370 __tlb_remove_table(table);
374 if (*batch == NULL) {
375 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
376 if (*batch == NULL) {
377 tlb_remove_table_one(table);
382 (*batch)->tables[(*batch)->nr++] = table;
383 if ((*batch)->nr == MAX_TABLE_BATCH)
384 tlb_table_flush(tlb);
387 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
390 * Note: this doesn't free the actual pages themselves. That
391 * has been handled earlier when unmapping all the memory regions.
393 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
396 pgtable_t token = pmd_pgtable(*pmd);
398 pte_free_tlb(tlb, token, addr);
399 atomic_long_dec(&tlb->mm->nr_ptes);
402 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
403 unsigned long addr, unsigned long end,
404 unsigned long floor, unsigned long ceiling)
411 pmd = pmd_offset(pud, addr);
413 next = pmd_addr_end(addr, end);
414 if (pmd_none_or_clear_bad(pmd))
416 free_pte_range(tlb, pmd, addr);
417 } while (pmd++, addr = next, addr != end);
427 if (end - 1 > ceiling - 1)
430 pmd = pmd_offset(pud, start);
432 pmd_free_tlb(tlb, pmd, start);
433 mm_dec_nr_pmds(tlb->mm);
436 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
437 unsigned long addr, unsigned long end,
438 unsigned long floor, unsigned long ceiling)
445 pud = pud_offset(pgd, addr);
447 next = pud_addr_end(addr, end);
448 if (pud_none_or_clear_bad(pud))
450 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
451 } while (pud++, addr = next, addr != end);
457 ceiling &= PGDIR_MASK;
461 if (end - 1 > ceiling - 1)
464 pud = pud_offset(pgd, start);
466 pud_free_tlb(tlb, pud, start);
470 * This function frees user-level page tables of a process.
472 void free_pgd_range(struct mmu_gather *tlb,
473 unsigned long addr, unsigned long end,
474 unsigned long floor, unsigned long ceiling)
480 * The next few lines have given us lots of grief...
482 * Why are we testing PMD* at this top level? Because often
483 * there will be no work to do at all, and we'd prefer not to
484 * go all the way down to the bottom just to discover that.
486 * Why all these "- 1"s? Because 0 represents both the bottom
487 * of the address space and the top of it (using -1 for the
488 * top wouldn't help much: the masks would do the wrong thing).
489 * The rule is that addr 0 and floor 0 refer to the bottom of
490 * the address space, but end 0 and ceiling 0 refer to the top
491 * Comparisons need to use "end - 1" and "ceiling - 1" (though
492 * that end 0 case should be mythical).
494 * Wherever addr is brought up or ceiling brought down, we must
495 * be careful to reject "the opposite 0" before it confuses the
496 * subsequent tests. But what about where end is brought down
497 * by PMD_SIZE below? no, end can't go down to 0 there.
499 * Whereas we round start (addr) and ceiling down, by different
500 * masks at different levels, in order to test whether a table
501 * now has no other vmas using it, so can be freed, we don't
502 * bother to round floor or end up - the tests don't need that.
516 if (end - 1 > ceiling - 1)
521 pgd = pgd_offset(tlb->mm, addr);
523 next = pgd_addr_end(addr, end);
524 if (pgd_none_or_clear_bad(pgd))
526 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
527 } while (pgd++, addr = next, addr != end);
530 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
531 unsigned long floor, unsigned long ceiling)
534 struct vm_area_struct *next = vma->vm_next;
535 unsigned long addr = vma->vm_start;
538 * Hide vma from rmap and truncate_pagecache before freeing
541 unlink_anon_vmas(vma);
542 unlink_file_vma(vma);
544 if (is_vm_hugetlb_page(vma)) {
545 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
546 floor, next? next->vm_start: ceiling);
549 * Optimization: gather nearby vmas into one call down
551 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
552 && !is_vm_hugetlb_page(next)) {
555 unlink_anon_vmas(vma);
556 unlink_file_vma(vma);
558 free_pgd_range(tlb, addr, vma->vm_end,
559 floor, next? next->vm_start: ceiling);
565 int __pte_alloc(struct mm_struct *mm, 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) || pmd_devmap(*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) || pmd_devmap(*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 pfn_t 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 if (pfn_t_devmap(pfn))
1520 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1522 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1523 set_pte_at(mm, addr, pte, entry);
1524 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1528 pte_unmap_unlock(pte, ptl);
1534 * vm_insert_pfn - insert single pfn into user vma
1535 * @vma: user vma to map to
1536 * @addr: target user address of this page
1537 * @pfn: source kernel pfn
1539 * Similar to vm_insert_page, this allows drivers to insert individual pages
1540 * they've allocated into a user vma. Same comments apply.
1542 * This function should only be called from a vm_ops->fault handler, and
1543 * in that case the handler should return NULL.
1545 * vma cannot be a COW mapping.
1547 * As this is called only for pages that do not currently exist, we
1548 * do not need to flush old virtual caches or the TLB.
1550 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1553 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1555 EXPORT_SYMBOL(vm_insert_pfn);
1558 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1559 * @vma: user vma to map to
1560 * @addr: target user address of this page
1561 * @pfn: source kernel pfn
1562 * @pgprot: pgprot flags for the inserted page
1564 * This is exactly like vm_insert_pfn, except that it allows drivers to
1565 * to override pgprot on a per-page basis.
1567 * This only makes sense for IO mappings, and it makes no sense for
1568 * cow mappings. In general, using multiple vmas is preferable;
1569 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1572 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1573 unsigned long pfn, pgprot_t pgprot)
1577 * Technically, architectures with pte_special can avoid all these
1578 * restrictions (same for remap_pfn_range). However we would like
1579 * consistency in testing and feature parity among all, so we should
1580 * try to keep these invariants in place for everybody.
1582 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1583 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1584 (VM_PFNMAP|VM_MIXEDMAP));
1585 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1586 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1588 if (addr < vma->vm_start || addr >= vma->vm_end)
1590 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
1593 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1597 EXPORT_SYMBOL(vm_insert_pfn_prot);
1599 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1602 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1604 if (addr < vma->vm_start || addr >= vma->vm_end)
1608 * If we don't have pte special, then we have to use the pfn_valid()
1609 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1610 * refcount the page if pfn_valid is true (hence insert_page rather
1611 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1612 * without pte special, it would there be refcounted as a normal page.
1614 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1618 * At this point we are committed to insert_page()
1619 * regardless of whether the caller specified flags that
1620 * result in pfn_t_has_page() == false.
1622 page = pfn_to_page(pfn_t_to_pfn(pfn));
1623 return insert_page(vma, addr, page, vma->vm_page_prot);
1625 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1627 EXPORT_SYMBOL(vm_insert_mixed);
1630 * maps a range of physical memory into the requested pages. the old
1631 * mappings are removed. any references to nonexistent pages results
1632 * in null mappings (currently treated as "copy-on-access")
1634 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1635 unsigned long addr, unsigned long end,
1636 unsigned long pfn, pgprot_t prot)
1641 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1644 arch_enter_lazy_mmu_mode();
1646 BUG_ON(!pte_none(*pte));
1647 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1649 } while (pte++, addr += PAGE_SIZE, addr != end);
1650 arch_leave_lazy_mmu_mode();
1651 pte_unmap_unlock(pte - 1, ptl);
1655 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1656 unsigned long addr, unsigned long end,
1657 unsigned long pfn, pgprot_t prot)
1662 pfn -= addr >> PAGE_SHIFT;
1663 pmd = pmd_alloc(mm, pud, addr);
1666 VM_BUG_ON(pmd_trans_huge(*pmd));
1668 next = pmd_addr_end(addr, end);
1669 if (remap_pte_range(mm, pmd, addr, next,
1670 pfn + (addr >> PAGE_SHIFT), prot))
1672 } while (pmd++, addr = next, addr != end);
1676 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1677 unsigned long addr, unsigned long end,
1678 unsigned long pfn, pgprot_t prot)
1683 pfn -= addr >> PAGE_SHIFT;
1684 pud = pud_alloc(mm, pgd, addr);
1688 next = pud_addr_end(addr, end);
1689 if (remap_pmd_range(mm, pud, addr, next,
1690 pfn + (addr >> PAGE_SHIFT), prot))
1692 } while (pud++, addr = next, addr != end);
1697 * remap_pfn_range - remap kernel memory to userspace
1698 * @vma: user vma to map to
1699 * @addr: target user address to start at
1700 * @pfn: physical address of kernel memory
1701 * @size: size of map area
1702 * @prot: page protection flags for this mapping
1704 * Note: this is only safe if the mm semaphore is held when called.
1706 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1707 unsigned long pfn, unsigned long size, pgprot_t prot)
1711 unsigned long end = addr + PAGE_ALIGN(size);
1712 struct mm_struct *mm = vma->vm_mm;
1716 * Physically remapped pages are special. Tell the
1717 * rest of the world about it:
1718 * VM_IO tells people not to look at these pages
1719 * (accesses can have side effects).
1720 * VM_PFNMAP tells the core MM that the base pages are just
1721 * raw PFN mappings, and do not have a "struct page" associated
1724 * Disable vma merging and expanding with mremap().
1726 * Omit vma from core dump, even when VM_IO turned off.
1728 * There's a horrible special case to handle copy-on-write
1729 * behaviour that some programs depend on. We mark the "original"
1730 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1731 * See vm_normal_page() for details.
1733 if (is_cow_mapping(vma->vm_flags)) {
1734 if (addr != vma->vm_start || end != vma->vm_end)
1736 vma->vm_pgoff = pfn;
1739 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1743 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1745 BUG_ON(addr >= end);
1746 pfn -= addr >> PAGE_SHIFT;
1747 pgd = pgd_offset(mm, addr);
1748 flush_cache_range(vma, addr, end);
1750 next = pgd_addr_end(addr, end);
1751 err = remap_pud_range(mm, pgd, addr, next,
1752 pfn + (addr >> PAGE_SHIFT), prot);
1755 } while (pgd++, addr = next, addr != end);
1758 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1762 EXPORT_SYMBOL(remap_pfn_range);
1765 * vm_iomap_memory - remap memory to userspace
1766 * @vma: user vma to map to
1767 * @start: start of area
1768 * @len: size of area
1770 * This is a simplified io_remap_pfn_range() for common driver use. The
1771 * driver just needs to give us the physical memory range to be mapped,
1772 * we'll figure out the rest from the vma information.
1774 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1775 * whatever write-combining details or similar.
1777 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1779 unsigned long vm_len, pfn, pages;
1781 /* Check that the physical memory area passed in looks valid */
1782 if (start + len < start)
1785 * You *really* shouldn't map things that aren't page-aligned,
1786 * but we've historically allowed it because IO memory might
1787 * just have smaller alignment.
1789 len += start & ~PAGE_MASK;
1790 pfn = start >> PAGE_SHIFT;
1791 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1792 if (pfn + pages < pfn)
1795 /* We start the mapping 'vm_pgoff' pages into the area */
1796 if (vma->vm_pgoff > pages)
1798 pfn += vma->vm_pgoff;
1799 pages -= vma->vm_pgoff;
1801 /* Can we fit all of the mapping? */
1802 vm_len = vma->vm_end - vma->vm_start;
1803 if (vm_len >> PAGE_SHIFT > pages)
1806 /* Ok, let it rip */
1807 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1809 EXPORT_SYMBOL(vm_iomap_memory);
1811 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1812 unsigned long addr, unsigned long end,
1813 pte_fn_t fn, void *data)
1818 spinlock_t *uninitialized_var(ptl);
1820 pte = (mm == &init_mm) ?
1821 pte_alloc_kernel(pmd, addr) :
1822 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1826 BUG_ON(pmd_huge(*pmd));
1828 arch_enter_lazy_mmu_mode();
1830 token = pmd_pgtable(*pmd);
1833 err = fn(pte++, token, addr, data);
1836 } while (addr += PAGE_SIZE, addr != end);
1838 arch_leave_lazy_mmu_mode();
1841 pte_unmap_unlock(pte-1, ptl);
1845 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1846 unsigned long addr, unsigned long end,
1847 pte_fn_t fn, void *data)
1853 BUG_ON(pud_huge(*pud));
1855 pmd = pmd_alloc(mm, pud, addr);
1859 next = pmd_addr_end(addr, end);
1860 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1863 } while (pmd++, addr = next, addr != end);
1867 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1868 unsigned long addr, unsigned long end,
1869 pte_fn_t fn, void *data)
1875 pud = pud_alloc(mm, pgd, addr);
1879 next = pud_addr_end(addr, end);
1880 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1883 } while (pud++, addr = next, addr != end);
1888 * Scan a region of virtual memory, filling in page tables as necessary
1889 * and calling a provided function on each leaf page table.
1891 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1892 unsigned long size, pte_fn_t fn, void *data)
1896 unsigned long end = addr + size;
1899 if (WARN_ON(addr >= end))
1902 pgd = pgd_offset(mm, addr);
1904 next = pgd_addr_end(addr, end);
1905 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1908 } while (pgd++, addr = next, addr != end);
1912 EXPORT_SYMBOL_GPL(apply_to_page_range);
1915 * handle_pte_fault chooses page fault handler according to an entry which was
1916 * read non-atomically. Before making any commitment, on those architectures
1917 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1918 * parts, do_swap_page must check under lock before unmapping the pte and
1919 * proceeding (but do_wp_page is only called after already making such a check;
1920 * and do_anonymous_page can safely check later on).
1922 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1923 pte_t *page_table, pte_t orig_pte)
1926 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1927 if (sizeof(pte_t) > sizeof(unsigned long)) {
1928 spinlock_t *ptl = pte_lockptr(mm, pmd);
1930 same = pte_same(*page_table, orig_pte);
1934 pte_unmap(page_table);
1938 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1940 debug_dma_assert_idle(src);
1943 * If the source page was a PFN mapping, we don't have
1944 * a "struct page" for it. We do a best-effort copy by
1945 * just copying from the original user address. If that
1946 * fails, we just zero-fill it. Live with it.
1948 if (unlikely(!src)) {
1949 void *kaddr = kmap_atomic(dst);
1950 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1953 * This really shouldn't fail, because the page is there
1954 * in the page tables. But it might just be unreadable,
1955 * in which case we just give up and fill the result with
1958 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1960 kunmap_atomic(kaddr);
1961 flush_dcache_page(dst);
1963 copy_user_highpage(dst, src, va, vma);
1966 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
1968 struct file *vm_file = vma->vm_file;
1971 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
1974 * Special mappings (e.g. VDSO) do not have any file so fake
1975 * a default GFP_KERNEL for them.
1981 * Notify the address space that the page is about to become writable so that
1982 * it can prohibit this or wait for the page to get into an appropriate state.
1984 * We do this without the lock held, so that it can sleep if it needs to.
1986 static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1987 unsigned long address)
1989 struct vm_fault vmf;
1992 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1993 vmf.pgoff = page->index;
1994 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1995 vmf.gfp_mask = __get_fault_gfp_mask(vma);
1997 vmf.cow_page = NULL;
1999 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2000 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2002 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2004 if (!page->mapping) {
2006 return 0; /* retry */
2008 ret |= VM_FAULT_LOCKED;
2010 VM_BUG_ON_PAGE(!PageLocked(page), page);
2015 * Handle write page faults for pages that can be reused in the current vma
2017 * This can happen either due to the mapping being with the VM_SHARED flag,
2018 * or due to us being the last reference standing to the page. In either
2019 * case, all we need to do here is to mark the page as writable and update
2020 * any related book-keeping.
2022 static inline int wp_page_reuse(struct mm_struct *mm,
2023 struct vm_area_struct *vma, unsigned long address,
2024 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2025 struct page *page, int page_mkwrite,
2031 * Clear the pages cpupid information as the existing
2032 * information potentially belongs to a now completely
2033 * unrelated process.
2036 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2038 flush_cache_page(vma, address, pte_pfn(orig_pte));
2039 entry = pte_mkyoung(orig_pte);
2040 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2041 if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2042 update_mmu_cache(vma, address, page_table);
2043 pte_unmap_unlock(page_table, ptl);
2046 struct address_space *mapping;
2052 dirtied = set_page_dirty(page);
2053 VM_BUG_ON_PAGE(PageAnon(page), page);
2054 mapping = page->mapping;
2056 page_cache_release(page);
2058 if ((dirtied || page_mkwrite) && mapping) {
2060 * Some device drivers do not set page.mapping
2061 * but still dirty their pages
2063 balance_dirty_pages_ratelimited(mapping);
2067 file_update_time(vma->vm_file);
2070 return VM_FAULT_WRITE;
2074 * Handle the case of a page which we actually need to copy to a new page.
2076 * Called with mmap_sem locked and the old page referenced, but
2077 * without the ptl held.
2079 * High level logic flow:
2081 * - Allocate a page, copy the content of the old page to the new one.
2082 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2083 * - Take the PTL. If the pte changed, bail out and release the allocated page
2084 * - If the pte is still the way we remember it, update the page table and all
2085 * relevant references. This includes dropping the reference the page-table
2086 * held to the old page, as well as updating the rmap.
2087 * - In any case, unlock the PTL and drop the reference we took to the old page.
2089 static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2090 unsigned long address, pte_t *page_table, pmd_t *pmd,
2091 pte_t orig_pte, struct page *old_page)
2093 struct page *new_page = NULL;
2094 spinlock_t *ptl = NULL;
2096 int page_copied = 0;
2097 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */
2098 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */
2099 struct mem_cgroup *memcg;
2101 if (unlikely(anon_vma_prepare(vma)))
2104 if (is_zero_pfn(pte_pfn(orig_pte))) {
2105 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2109 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2112 cow_user_page(new_page, old_page, address, vma);
2115 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2118 __SetPageUptodate(new_page);
2120 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2123 * Re-check the pte - we dropped the lock
2125 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2126 if (likely(pte_same(*page_table, orig_pte))) {
2128 if (!PageAnon(old_page)) {
2129 dec_mm_counter_fast(mm,
2130 mm_counter_file(old_page));
2131 inc_mm_counter_fast(mm, MM_ANONPAGES);
2134 inc_mm_counter_fast(mm, MM_ANONPAGES);
2136 flush_cache_page(vma, address, pte_pfn(orig_pte));
2137 entry = mk_pte(new_page, vma->vm_page_prot);
2138 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2140 * Clear the pte entry and flush it first, before updating the
2141 * pte with the new entry. This will avoid a race condition
2142 * seen in the presence of one thread doing SMC and another
2145 ptep_clear_flush_notify(vma, address, page_table);
2146 page_add_new_anon_rmap(new_page, vma, address, false);
2147 mem_cgroup_commit_charge(new_page, memcg, false, false);
2148 lru_cache_add_active_or_unevictable(new_page, vma);
2150 * We call the notify macro here because, when using secondary
2151 * mmu page tables (such as kvm shadow page tables), we want the
2152 * new page to be mapped directly into the secondary page table.
2154 set_pte_at_notify(mm, address, page_table, entry);
2155 update_mmu_cache(vma, address, page_table);
2158 * Only after switching the pte to the new page may
2159 * we remove the mapcount here. Otherwise another
2160 * process may come and find the rmap count decremented
2161 * before the pte is switched to the new page, and
2162 * "reuse" the old page writing into it while our pte
2163 * here still points into it and can be read by other
2166 * The critical issue is to order this
2167 * page_remove_rmap with the ptp_clear_flush above.
2168 * Those stores are ordered by (if nothing else,)
2169 * the barrier present in the atomic_add_negative
2170 * in page_remove_rmap.
2172 * Then the TLB flush in ptep_clear_flush ensures that
2173 * no process can access the old page before the
2174 * decremented mapcount is visible. And the old page
2175 * cannot be reused until after the decremented
2176 * mapcount is visible. So transitively, TLBs to
2177 * old page will be flushed before it can be reused.
2179 page_remove_rmap(old_page, false);
2182 /* Free the old page.. */
2183 new_page = old_page;
2186 mem_cgroup_cancel_charge(new_page, memcg, false);
2190 page_cache_release(new_page);
2192 pte_unmap_unlock(page_table, ptl);
2193 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2196 * Don't let another task, with possibly unlocked vma,
2197 * keep the mlocked page.
2199 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2200 lock_page(old_page); /* LRU manipulation */
2201 if (PageMlocked(old_page))
2202 munlock_vma_page(old_page);
2203 unlock_page(old_page);
2205 page_cache_release(old_page);
2207 return page_copied ? VM_FAULT_WRITE : 0;
2209 page_cache_release(new_page);
2212 page_cache_release(old_page);
2213 return VM_FAULT_OOM;
2217 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2220 static int wp_pfn_shared(struct mm_struct *mm,
2221 struct vm_area_struct *vma, unsigned long address,
2222 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2225 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2226 struct vm_fault vmf = {
2228 .pgoff = linear_page_index(vma, address),
2229 .virtual_address = (void __user *)(address & PAGE_MASK),
2230 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2234 pte_unmap_unlock(page_table, ptl);
2235 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2236 if (ret & VM_FAULT_ERROR)
2238 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2240 * We might have raced with another page fault while we
2241 * released the pte_offset_map_lock.
2243 if (!pte_same(*page_table, orig_pte)) {
2244 pte_unmap_unlock(page_table, ptl);
2248 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2252 static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2253 unsigned long address, pte_t *page_table,
2254 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2255 struct page *old_page)
2258 int page_mkwrite = 0;
2260 page_cache_get(old_page);
2262 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2265 pte_unmap_unlock(page_table, ptl);
2266 tmp = do_page_mkwrite(vma, old_page, address);
2267 if (unlikely(!tmp || (tmp &
2268 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2269 page_cache_release(old_page);
2273 * Since we dropped the lock we need to revalidate
2274 * the PTE as someone else may have changed it. If
2275 * they did, we just return, as we can count on the
2276 * MMU to tell us if they didn't also make it writable.
2278 page_table = pte_offset_map_lock(mm, pmd, address,
2280 if (!pte_same(*page_table, orig_pte)) {
2281 unlock_page(old_page);
2282 pte_unmap_unlock(page_table, ptl);
2283 page_cache_release(old_page);
2289 return wp_page_reuse(mm, vma, address, page_table, ptl,
2290 orig_pte, old_page, page_mkwrite, 1);
2294 * This routine handles present pages, when users try to write
2295 * to a shared page. It is done by copying the page to a new address
2296 * and decrementing the shared-page counter for the old page.
2298 * Note that this routine assumes that the protection checks have been
2299 * done by the caller (the low-level page fault routine in most cases).
2300 * Thus we can safely just mark it writable once we've done any necessary
2303 * We also mark the page dirty at this point even though the page will
2304 * change only once the write actually happens. This avoids a few races,
2305 * and potentially makes it more efficient.
2307 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2308 * but allow concurrent faults), with pte both mapped and locked.
2309 * We return with mmap_sem still held, but pte unmapped and unlocked.
2311 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2312 unsigned long address, pte_t *page_table, pmd_t *pmd,
2313 spinlock_t *ptl, pte_t orig_pte)
2316 struct page *old_page;
2318 old_page = vm_normal_page(vma, address, orig_pte);
2321 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2324 * We should not cow pages in a shared writeable mapping.
2325 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2327 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2328 (VM_WRITE|VM_SHARED))
2329 return wp_pfn_shared(mm, vma, address, page_table, ptl,
2332 pte_unmap_unlock(page_table, ptl);
2333 return wp_page_copy(mm, vma, address, page_table, pmd,
2334 orig_pte, old_page);
2338 * Take out anonymous pages first, anonymous shared vmas are
2339 * not dirty accountable.
2341 if (PageAnon(old_page) && !PageKsm(old_page)) {
2342 if (!trylock_page(old_page)) {
2343 page_cache_get(old_page);
2344 pte_unmap_unlock(page_table, ptl);
2345 lock_page(old_page);
2346 page_table = pte_offset_map_lock(mm, pmd, address,
2348 if (!pte_same(*page_table, orig_pte)) {
2349 unlock_page(old_page);
2350 pte_unmap_unlock(page_table, ptl);
2351 page_cache_release(old_page);
2354 page_cache_release(old_page);
2356 if (reuse_swap_page(old_page)) {
2358 * The page is all ours. Move it to our anon_vma so
2359 * the rmap code will not search our parent or siblings.
2360 * Protected against the rmap code by the page lock.
2362 page_move_anon_rmap(old_page, vma, address);
2363 unlock_page(old_page);
2364 return wp_page_reuse(mm, vma, address, page_table, ptl,
2365 orig_pte, old_page, 0, 0);
2367 unlock_page(old_page);
2368 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2369 (VM_WRITE|VM_SHARED))) {
2370 return wp_page_shared(mm, vma, address, page_table, pmd,
2371 ptl, orig_pte, old_page);
2375 * Ok, we need to copy. Oh, well..
2377 page_cache_get(old_page);
2379 pte_unmap_unlock(page_table, ptl);
2380 return wp_page_copy(mm, vma, address, page_table, pmd,
2381 orig_pte, old_page);
2384 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2385 unsigned long start_addr, unsigned long end_addr,
2386 struct zap_details *details)
2388 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2391 static inline void unmap_mapping_range_tree(struct rb_root *root,
2392 struct zap_details *details)
2394 struct vm_area_struct *vma;
2395 pgoff_t vba, vea, zba, zea;
2397 vma_interval_tree_foreach(vma, root,
2398 details->first_index, details->last_index) {
2400 vba = vma->vm_pgoff;
2401 vea = vba + vma_pages(vma) - 1;
2402 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2403 zba = details->first_index;
2406 zea = details->last_index;
2410 unmap_mapping_range_vma(vma,
2411 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2412 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2418 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2419 * address_space corresponding to the specified page range in the underlying
2422 * @mapping: the address space containing mmaps to be unmapped.
2423 * @holebegin: byte in first page to unmap, relative to the start of
2424 * the underlying file. This will be rounded down to a PAGE_SIZE
2425 * boundary. Note that this is different from truncate_pagecache(), which
2426 * must keep the partial page. In contrast, we must get rid of
2428 * @holelen: size of prospective hole in bytes. This will be rounded
2429 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2431 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2432 * but 0 when invalidating pagecache, don't throw away private data.
2434 void unmap_mapping_range(struct address_space *mapping,
2435 loff_t const holebegin, loff_t const holelen, int even_cows)
2437 struct zap_details details;
2438 pgoff_t hba = holebegin >> PAGE_SHIFT;
2439 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2441 /* Check for overflow. */
2442 if (sizeof(holelen) > sizeof(hlen)) {
2444 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2445 if (holeend & ~(long long)ULONG_MAX)
2446 hlen = ULONG_MAX - hba + 1;
2449 details.check_mapping = even_cows? NULL: mapping;
2450 details.first_index = hba;
2451 details.last_index = hba + hlen - 1;
2452 if (details.last_index < details.first_index)
2453 details.last_index = ULONG_MAX;
2456 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2457 i_mmap_lock_write(mapping);
2458 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2459 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2460 i_mmap_unlock_write(mapping);
2462 EXPORT_SYMBOL(unmap_mapping_range);
2465 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2466 * but allow concurrent faults), and pte mapped but not yet locked.
2467 * We return with pte unmapped and unlocked.
2469 * We return with the mmap_sem locked or unlocked in the same cases
2470 * as does filemap_fault().
2472 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2473 unsigned long address, pte_t *page_table, pmd_t *pmd,
2474 unsigned int flags, pte_t orig_pte)
2477 struct page *page, *swapcache;
2478 struct mem_cgroup *memcg;
2485 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2488 entry = pte_to_swp_entry(orig_pte);
2489 if (unlikely(non_swap_entry(entry))) {
2490 if (is_migration_entry(entry)) {
2491 migration_entry_wait(mm, pmd, address);
2492 } else if (is_hwpoison_entry(entry)) {
2493 ret = VM_FAULT_HWPOISON;
2495 print_bad_pte(vma, address, orig_pte, NULL);
2496 ret = VM_FAULT_SIGBUS;
2500 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2501 page = lookup_swap_cache(entry);
2503 page = swapin_readahead(entry,
2504 GFP_HIGHUSER_MOVABLE, vma, address);
2507 * Back out if somebody else faulted in this pte
2508 * while we released the pte lock.
2510 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2511 if (likely(pte_same(*page_table, orig_pte)))
2513 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2517 /* Had to read the page from swap area: Major fault */
2518 ret = VM_FAULT_MAJOR;
2519 count_vm_event(PGMAJFAULT);
2520 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2521 } else if (PageHWPoison(page)) {
2523 * hwpoisoned dirty swapcache pages are kept for killing
2524 * owner processes (which may be unknown at hwpoison time)
2526 ret = VM_FAULT_HWPOISON;
2527 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2533 locked = lock_page_or_retry(page, mm, flags);
2535 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2537 ret |= VM_FAULT_RETRY;
2542 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2543 * release the swapcache from under us. The page pin, and pte_same
2544 * test below, are not enough to exclude that. Even if it is still
2545 * swapcache, we need to check that the page's swap has not changed.
2547 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2550 page = ksm_might_need_to_copy(page, vma, address);
2551 if (unlikely(!page)) {
2557 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) {
2563 * Back out if somebody else already faulted in this pte.
2565 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2566 if (unlikely(!pte_same(*page_table, orig_pte)))
2569 if (unlikely(!PageUptodate(page))) {
2570 ret = VM_FAULT_SIGBUS;
2575 * The page isn't present yet, go ahead with the fault.
2577 * Be careful about the sequence of operations here.
2578 * To get its accounting right, reuse_swap_page() must be called
2579 * while the page is counted on swap but not yet in mapcount i.e.
2580 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2581 * must be called after the swap_free(), or it will never succeed.
2584 inc_mm_counter_fast(mm, MM_ANONPAGES);
2585 dec_mm_counter_fast(mm, MM_SWAPENTS);
2586 pte = mk_pte(page, vma->vm_page_prot);
2587 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2588 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2589 flags &= ~FAULT_FLAG_WRITE;
2590 ret |= VM_FAULT_WRITE;
2591 exclusive = RMAP_EXCLUSIVE;
2593 flush_icache_page(vma, page);
2594 if (pte_swp_soft_dirty(orig_pte))
2595 pte = pte_mksoft_dirty(pte);
2596 set_pte_at(mm, address, page_table, pte);
2597 if (page == swapcache) {
2598 do_page_add_anon_rmap(page, vma, address, exclusive);
2599 mem_cgroup_commit_charge(page, memcg, true, false);
2600 } else { /* ksm created a completely new copy */
2601 page_add_new_anon_rmap(page, vma, address, false);
2602 mem_cgroup_commit_charge(page, memcg, false, false);
2603 lru_cache_add_active_or_unevictable(page, vma);
2607 if (mem_cgroup_swap_full(page) ||
2608 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2609 try_to_free_swap(page);
2611 if (page != swapcache) {
2613 * Hold the lock to avoid the swap entry to be reused
2614 * until we take the PT lock for the pte_same() check
2615 * (to avoid false positives from pte_same). For
2616 * further safety release the lock after the swap_free
2617 * so that the swap count won't change under a
2618 * parallel locked swapcache.
2620 unlock_page(swapcache);
2621 page_cache_release(swapcache);
2624 if (flags & FAULT_FLAG_WRITE) {
2625 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2626 if (ret & VM_FAULT_ERROR)
2627 ret &= VM_FAULT_ERROR;
2631 /* No need to invalidate - it was non-present before */
2632 update_mmu_cache(vma, address, page_table);
2634 pte_unmap_unlock(page_table, ptl);
2638 mem_cgroup_cancel_charge(page, memcg, false);
2639 pte_unmap_unlock(page_table, ptl);
2643 page_cache_release(page);
2644 if (page != swapcache) {
2645 unlock_page(swapcache);
2646 page_cache_release(swapcache);
2652 * This is like a special single-page "expand_{down|up}wards()",
2653 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2654 * doesn't hit another vma.
2656 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2658 address &= PAGE_MASK;
2659 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2660 struct vm_area_struct *prev = vma->vm_prev;
2663 * Is there a mapping abutting this one below?
2665 * That's only ok if it's the same stack mapping
2666 * that has gotten split..
2668 if (prev && prev->vm_end == address)
2669 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2671 return expand_downwards(vma, address - PAGE_SIZE);
2673 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2674 struct vm_area_struct *next = vma->vm_next;
2676 /* As VM_GROWSDOWN but s/below/above/ */
2677 if (next && next->vm_start == address + PAGE_SIZE)
2678 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2680 return expand_upwards(vma, address + PAGE_SIZE);
2686 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2687 * but allow concurrent faults), and pte mapped but not yet locked.
2688 * We return with mmap_sem still held, but pte unmapped and unlocked.
2690 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2691 unsigned long address, pte_t *page_table, pmd_t *pmd,
2694 struct mem_cgroup *memcg;
2699 pte_unmap(page_table);
2701 /* File mapping without ->vm_ops ? */
2702 if (vma->vm_flags & VM_SHARED)
2703 return VM_FAULT_SIGBUS;
2705 /* Check if we need to add a guard page to the stack */
2706 if (check_stack_guard_page(vma, address) < 0)
2707 return VM_FAULT_SIGSEGV;
2709 /* Use the zero-page for reads */
2710 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2711 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2712 vma->vm_page_prot));
2713 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2714 if (!pte_none(*page_table))
2716 /* Deliver the page fault to userland, check inside PT lock */
2717 if (userfaultfd_missing(vma)) {
2718 pte_unmap_unlock(page_table, ptl);
2719 return handle_userfault(vma, address, flags,
2725 /* Allocate our own private page. */
2726 if (unlikely(anon_vma_prepare(vma)))
2728 page = alloc_zeroed_user_highpage_movable(vma, address);
2732 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false))
2736 * The memory barrier inside __SetPageUptodate makes sure that
2737 * preceeding stores to the page contents become visible before
2738 * the set_pte_at() write.
2740 __SetPageUptodate(page);
2742 entry = mk_pte(page, vma->vm_page_prot);
2743 if (vma->vm_flags & VM_WRITE)
2744 entry = pte_mkwrite(pte_mkdirty(entry));
2746 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2747 if (!pte_none(*page_table))
2750 /* Deliver the page fault to userland, check inside PT lock */
2751 if (userfaultfd_missing(vma)) {
2752 pte_unmap_unlock(page_table, ptl);
2753 mem_cgroup_cancel_charge(page, memcg, false);
2754 page_cache_release(page);
2755 return handle_userfault(vma, address, flags,
2759 inc_mm_counter_fast(mm, MM_ANONPAGES);
2760 page_add_new_anon_rmap(page, vma, address, false);
2761 mem_cgroup_commit_charge(page, memcg, false, false);
2762 lru_cache_add_active_or_unevictable(page, vma);
2764 set_pte_at(mm, address, page_table, entry);
2766 /* No need to invalidate - it was non-present before */
2767 update_mmu_cache(vma, address, page_table);
2769 pte_unmap_unlock(page_table, ptl);
2772 mem_cgroup_cancel_charge(page, memcg, false);
2773 page_cache_release(page);
2776 page_cache_release(page);
2778 return VM_FAULT_OOM;
2782 * The mmap_sem must have been held on entry, and may have been
2783 * released depending on flags and vma->vm_ops->fault() return value.
2784 * See filemap_fault() and __lock_page_retry().
2786 static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2787 pgoff_t pgoff, unsigned int flags,
2788 struct page *cow_page, struct page **page)
2790 struct vm_fault vmf;
2793 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2797 vmf.gfp_mask = __get_fault_gfp_mask(vma);
2798 vmf.cow_page = cow_page;
2800 ret = vma->vm_ops->fault(vma, &vmf);
2801 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2806 if (unlikely(PageHWPoison(vmf.page))) {
2807 if (ret & VM_FAULT_LOCKED)
2808 unlock_page(vmf.page);
2809 page_cache_release(vmf.page);
2810 return VM_FAULT_HWPOISON;
2813 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2814 lock_page(vmf.page);
2816 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2824 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2826 * @vma: virtual memory area
2827 * @address: user virtual address
2828 * @page: page to map
2829 * @pte: pointer to target page table entry
2830 * @write: true, if new entry is writable
2831 * @anon: true, if it's anonymous page
2833 * Caller must hold page table lock relevant for @pte.
2835 * Target users are page handler itself and implementations of
2836 * vm_ops->map_pages.
2838 void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2839 struct page *page, pte_t *pte, bool write, bool anon)
2843 flush_icache_page(vma, page);
2844 entry = mk_pte(page, vma->vm_page_prot);
2846 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2848 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2849 page_add_new_anon_rmap(page, vma, address, false);
2851 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
2852 page_add_file_rmap(page);
2854 set_pte_at(vma->vm_mm, address, pte, entry);
2856 /* no need to invalidate: a not-present page won't be cached */
2857 update_mmu_cache(vma, address, pte);
2860 static unsigned long fault_around_bytes __read_mostly =
2861 rounddown_pow_of_two(65536);
2863 #ifdef CONFIG_DEBUG_FS
2864 static int fault_around_bytes_get(void *data, u64 *val)
2866 *val = fault_around_bytes;
2871 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2872 * rounded down to nearest page order. It's what do_fault_around() expects to
2875 static int fault_around_bytes_set(void *data, u64 val)
2877 if (val / PAGE_SIZE > PTRS_PER_PTE)
2879 if (val > PAGE_SIZE)
2880 fault_around_bytes = rounddown_pow_of_two(val);
2882 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2885 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2886 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2888 static int __init fault_around_debugfs(void)
2892 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2893 &fault_around_bytes_fops);
2895 pr_warn("Failed to create fault_around_bytes in debugfs");
2898 late_initcall(fault_around_debugfs);
2902 * do_fault_around() tries to map few pages around the fault address. The hope
2903 * is that the pages will be needed soon and this will lower the number of
2906 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2907 * not ready to be mapped: not up-to-date, locked, etc.
2909 * This function is called with the page table lock taken. In the split ptlock
2910 * case the page table lock only protects only those entries which belong to
2911 * the page table corresponding to the fault address.
2913 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2916 * fault_around_pages() defines how many pages we'll try to map.
2917 * do_fault_around() expects it to return a power of two less than or equal to
2920 * The virtual address of the area that we map is naturally aligned to the
2921 * fault_around_pages() value (and therefore to page order). This way it's
2922 * easier to guarantee that we don't cross page table boundaries.
2924 static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2925 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2927 unsigned long start_addr, nr_pages, mask;
2929 struct vm_fault vmf;
2932 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2933 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2935 start_addr = max(address & mask, vma->vm_start);
2936 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2941 * max_pgoff is either end of page table or end of vma
2942 * or fault_around_pages() from pgoff, depending what is nearest.
2944 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2946 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2947 pgoff + nr_pages - 1);
2949 /* Check if it makes any sense to call ->map_pages */
2950 while (!pte_none(*pte)) {
2951 if (++pgoff > max_pgoff)
2953 start_addr += PAGE_SIZE;
2954 if (start_addr >= vma->vm_end)
2959 vmf.virtual_address = (void __user *) start_addr;
2962 vmf.max_pgoff = max_pgoff;
2964 vmf.gfp_mask = __get_fault_gfp_mask(vma);
2965 vma->vm_ops->map_pages(vma, &vmf);
2968 static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2969 unsigned long address, pmd_t *pmd,
2970 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2972 struct page *fault_page;
2978 * Let's call ->map_pages() first and use ->fault() as fallback
2979 * if page by the offset is not ready to be mapped (cold cache or
2982 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
2983 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2984 do_fault_around(vma, address, pte, pgoff, flags);
2985 if (!pte_same(*pte, orig_pte))
2987 pte_unmap_unlock(pte, ptl);
2990 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
2991 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2994 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2995 if (unlikely(!pte_same(*pte, orig_pte))) {
2996 pte_unmap_unlock(pte, ptl);
2997 unlock_page(fault_page);
2998 page_cache_release(fault_page);
3001 do_set_pte(vma, address, fault_page, pte, false, false);
3002 unlock_page(fault_page);
3004 pte_unmap_unlock(pte, ptl);
3008 static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3009 unsigned long address, pmd_t *pmd,
3010 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3012 struct page *fault_page, *new_page;
3013 struct mem_cgroup *memcg;
3018 if (unlikely(anon_vma_prepare(vma)))
3019 return VM_FAULT_OOM;
3021 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3023 return VM_FAULT_OOM;
3025 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) {
3026 page_cache_release(new_page);
3027 return VM_FAULT_OOM;
3030 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
3031 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3035 copy_user_highpage(new_page, fault_page, address, vma);
3036 __SetPageUptodate(new_page);
3038 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3039 if (unlikely(!pte_same(*pte, orig_pte))) {
3040 pte_unmap_unlock(pte, ptl);
3042 unlock_page(fault_page);
3043 page_cache_release(fault_page);
3046 * The fault handler has no page to lock, so it holds
3047 * i_mmap_lock for read to protect against truncate.
3049 i_mmap_unlock_read(vma->vm_file->f_mapping);
3053 do_set_pte(vma, address, new_page, pte, true, true);
3054 mem_cgroup_commit_charge(new_page, memcg, false, false);
3055 lru_cache_add_active_or_unevictable(new_page, vma);
3056 pte_unmap_unlock(pte, ptl);
3058 unlock_page(fault_page);
3059 page_cache_release(fault_page);
3062 * The fault handler has no page to lock, so it holds
3063 * i_mmap_lock for read to protect against truncate.
3065 i_mmap_unlock_read(vma->vm_file->f_mapping);
3069 mem_cgroup_cancel_charge(new_page, memcg, false);
3070 page_cache_release(new_page);
3074 static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3075 unsigned long address, pmd_t *pmd,
3076 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3078 struct page *fault_page;
3079 struct address_space *mapping;
3085 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3086 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3090 * Check if the backing address space wants to know that the page is
3091 * about to become writable
3093 if (vma->vm_ops->page_mkwrite) {
3094 unlock_page(fault_page);
3095 tmp = do_page_mkwrite(vma, fault_page, address);
3096 if (unlikely(!tmp ||
3097 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3098 page_cache_release(fault_page);
3103 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3104 if (unlikely(!pte_same(*pte, orig_pte))) {
3105 pte_unmap_unlock(pte, ptl);
3106 unlock_page(fault_page);
3107 page_cache_release(fault_page);
3110 do_set_pte(vma, address, fault_page, pte, true, false);
3111 pte_unmap_unlock(pte, ptl);
3113 if (set_page_dirty(fault_page))
3116 * Take a local copy of the address_space - page.mapping may be zeroed
3117 * by truncate after unlock_page(). The address_space itself remains
3118 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3119 * release semantics to prevent the compiler from undoing this copying.
3121 mapping = page_rmapping(fault_page);
3122 unlock_page(fault_page);
3123 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3125 * Some device drivers do not set page.mapping but still
3128 balance_dirty_pages_ratelimited(mapping);
3131 if (!vma->vm_ops->page_mkwrite)
3132 file_update_time(vma->vm_file);
3138 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3139 * but allow concurrent faults).
3140 * The mmap_sem may have been released depending on flags and our
3141 * return value. See filemap_fault() and __lock_page_or_retry().
3143 static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3144 unsigned long address, pte_t *page_table, pmd_t *pmd,
3145 unsigned int flags, pte_t orig_pte)
3147 pgoff_t pgoff = linear_page_index(vma, address);
3149 pte_unmap(page_table);
3150 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3151 if (!vma->vm_ops->fault)
3152 return VM_FAULT_SIGBUS;
3153 if (!(flags & FAULT_FLAG_WRITE))
3154 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3156 if (!(vma->vm_flags & VM_SHARED))
3157 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3159 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3162 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3163 unsigned long addr, int page_nid,
3168 count_vm_numa_event(NUMA_HINT_FAULTS);
3169 if (page_nid == numa_node_id()) {
3170 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3171 *flags |= TNF_FAULT_LOCAL;
3174 return mpol_misplaced(page, vma, addr);
3177 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3178 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3180 struct page *page = NULL;
3185 bool migrated = false;
3186 bool was_writable = pte_write(pte);
3189 /* A PROT_NONE fault should not end up here */
3190 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3193 * The "pte" at this point cannot be used safely without
3194 * validation through pte_unmap_same(). It's of NUMA type but
3195 * the pfn may be screwed if the read is non atomic.
3197 * We can safely just do a "set_pte_at()", because the old
3198 * page table entry is not accessible, so there would be no
3199 * concurrent hardware modifications to the PTE.
3201 ptl = pte_lockptr(mm, pmd);
3203 if (unlikely(!pte_same(*ptep, pte))) {
3204 pte_unmap_unlock(ptep, ptl);
3208 /* Make it present again */
3209 pte = pte_modify(pte, vma->vm_page_prot);
3210 pte = pte_mkyoung(pte);
3212 pte = pte_mkwrite(pte);
3213 set_pte_at(mm, addr, ptep, pte);
3214 update_mmu_cache(vma, addr, ptep);
3216 page = vm_normal_page(vma, addr, pte);
3218 pte_unmap_unlock(ptep, ptl);
3222 /* TODO: handle PTE-mapped THP */
3223 if (PageCompound(page)) {
3224 pte_unmap_unlock(ptep, ptl);
3229 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3230 * much anyway since they can be in shared cache state. This misses
3231 * the case where a mapping is writable but the process never writes
3232 * to it but pte_write gets cleared during protection updates and
3233 * pte_dirty has unpredictable behaviour between PTE scan updates,
3234 * background writeback, dirty balancing and application behaviour.
3236 if (!(vma->vm_flags & VM_WRITE))
3237 flags |= TNF_NO_GROUP;
3240 * Flag if the page is shared between multiple address spaces. This
3241 * is later used when determining whether to group tasks together
3243 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3244 flags |= TNF_SHARED;
3246 last_cpupid = page_cpupid_last(page);
3247 page_nid = page_to_nid(page);
3248 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3249 pte_unmap_unlock(ptep, ptl);
3250 if (target_nid == -1) {
3255 /* Migrate to the requested node */
3256 migrated = migrate_misplaced_page(page, vma, target_nid);
3258 page_nid = target_nid;
3259 flags |= TNF_MIGRATED;
3261 flags |= TNF_MIGRATE_FAIL;
3265 task_numa_fault(last_cpupid, page_nid, 1, flags);
3269 static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3270 unsigned long address, pmd_t *pmd, unsigned int flags)
3272 if (vma_is_anonymous(vma))
3273 return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3274 if (vma->vm_ops->pmd_fault)
3275 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3276 return VM_FAULT_FALLBACK;
3279 static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3280 unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3283 if (vma_is_anonymous(vma))
3284 return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3285 if (vma->vm_ops->pmd_fault)
3286 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3287 return VM_FAULT_FALLBACK;
3291 * These routines also need to handle stuff like marking pages dirty
3292 * and/or accessed for architectures that don't do it in hardware (most
3293 * RISC architectures). The early dirtying is also good on the i386.
3295 * There is also a hook called "update_mmu_cache()" that architectures
3296 * with external mmu caches can use to update those (ie the Sparc or
3297 * PowerPC hashed page tables that act as extended TLBs).
3299 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3300 * but allow concurrent faults), and pte mapped but not yet locked.
3301 * We return with pte unmapped and unlocked.
3303 * The mmap_sem may have been released depending on flags and our
3304 * return value. See filemap_fault() and __lock_page_or_retry().
3306 static int handle_pte_fault(struct mm_struct *mm,
3307 struct vm_area_struct *vma, unsigned long address,
3308 pte_t *pte, pmd_t *pmd, unsigned int flags)
3314 * some architectures can have larger ptes than wordsize,
3315 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3316 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3317 * The code below just needs a consistent view for the ifs and
3318 * we later double check anyway with the ptl lock held. So here
3319 * a barrier will do.
3323 if (!pte_present(entry)) {
3324 if (pte_none(entry)) {
3325 if (vma_is_anonymous(vma))
3326 return do_anonymous_page(mm, vma, address,
3329 return do_fault(mm, vma, address, pte, pmd,
3332 return do_swap_page(mm, vma, address,
3333 pte, pmd, flags, entry);
3336 if (pte_protnone(entry))
3337 return do_numa_page(mm, vma, address, entry, pte, pmd);
3339 ptl = pte_lockptr(mm, pmd);
3341 if (unlikely(!pte_same(*pte, entry)))
3343 if (flags & FAULT_FLAG_WRITE) {
3344 if (!pte_write(entry))
3345 return do_wp_page(mm, vma, address,
3346 pte, pmd, ptl, entry);
3347 entry = pte_mkdirty(entry);
3349 entry = pte_mkyoung(entry);
3350 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3351 update_mmu_cache(vma, address, pte);
3354 * This is needed only for protection faults but the arch code
3355 * is not yet telling us if this is a protection fault or not.
3356 * This still avoids useless tlb flushes for .text page faults
3359 if (flags & FAULT_FLAG_WRITE)
3360 flush_tlb_fix_spurious_fault(vma, address);
3363 pte_unmap_unlock(pte, ptl);
3368 * By the time we get here, we already hold the mm semaphore
3370 * The mmap_sem may have been released depending on flags and our
3371 * return value. See filemap_fault() and __lock_page_or_retry().
3373 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3374 unsigned long address, unsigned int flags)
3381 if (unlikely(is_vm_hugetlb_page(vma)))
3382 return hugetlb_fault(mm, vma, address, flags);
3384 pgd = pgd_offset(mm, address);
3385 pud = pud_alloc(mm, pgd, address);
3387 return VM_FAULT_OOM;
3388 pmd = pmd_alloc(mm, pud, address);
3390 return VM_FAULT_OOM;
3391 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3392 int ret = create_huge_pmd(mm, vma, address, pmd, flags);
3393 if (!(ret & VM_FAULT_FALLBACK))
3396 pmd_t orig_pmd = *pmd;
3400 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3401 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3403 if (pmd_protnone(orig_pmd))
3404 return do_huge_pmd_numa_page(mm, vma, address,
3407 if (dirty && !pmd_write(orig_pmd)) {
3408 ret = wp_huge_pmd(mm, vma, address, pmd,
3410 if (!(ret & VM_FAULT_FALLBACK))
3413 huge_pmd_set_accessed(mm, vma, address, pmd,
3421 * Use pte_alloc() instead of pte_alloc_map, because we can't
3422 * run pte_offset_map on the pmd, if an huge pmd could
3423 * materialize from under us from a different thread.
3425 if (unlikely(pte_alloc(mm, pmd, address)))
3426 return VM_FAULT_OOM;
3428 * If a huge pmd materialized under us just retry later. Use
3429 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3430 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3431 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3432 * in a different thread of this mm, in turn leading to a misleading
3433 * pmd_trans_huge() retval. All we have to ensure is that it is a
3434 * regular pmd that we can walk with pte_offset_map() and we can do that
3435 * through an atomic read in C, which is what pmd_trans_unstable()
3438 if (unlikely(pmd_trans_unstable(pmd) || pmd_devmap(*pmd)))
3441 * A regular pmd is established and it can't morph into a huge pmd
3442 * from under us anymore at this point because we hold the mmap_sem
3443 * read mode and khugepaged takes it in write mode. So now it's
3444 * safe to run pte_offset_map().
3446 pte = pte_offset_map(pmd, address);
3448 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3452 * By the time we get here, we already hold the mm semaphore
3454 * The mmap_sem may have been released depending on flags and our
3455 * return value. See filemap_fault() and __lock_page_or_retry().
3457 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3458 unsigned long address, unsigned int flags)
3462 __set_current_state(TASK_RUNNING);
3464 count_vm_event(PGFAULT);
3465 mem_cgroup_count_vm_event(mm, PGFAULT);
3467 /* do counter updates before entering really critical section. */
3468 check_sync_rss_stat(current);
3471 * Enable the memcg OOM handling for faults triggered in user
3472 * space. Kernel faults are handled more gracefully.
3474 if (flags & FAULT_FLAG_USER)
3475 mem_cgroup_oom_enable();
3477 ret = __handle_mm_fault(mm, vma, address, flags);
3479 if (flags & FAULT_FLAG_USER) {
3480 mem_cgroup_oom_disable();
3482 * The task may have entered a memcg OOM situation but
3483 * if the allocation error was handled gracefully (no
3484 * VM_FAULT_OOM), there is no need to kill anything.
3485 * Just clean up the OOM state peacefully.
3487 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3488 mem_cgroup_oom_synchronize(false);
3493 EXPORT_SYMBOL_GPL(handle_mm_fault);
3495 #ifndef __PAGETABLE_PUD_FOLDED
3497 * Allocate page upper directory.
3498 * We've already handled the fast-path in-line.
3500 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3502 pud_t *new = pud_alloc_one(mm, address);
3506 smp_wmb(); /* See comment in __pte_alloc */
3508 spin_lock(&mm->page_table_lock);
3509 if (pgd_present(*pgd)) /* Another has populated it */
3512 pgd_populate(mm, pgd, new);
3513 spin_unlock(&mm->page_table_lock);
3516 #endif /* __PAGETABLE_PUD_FOLDED */
3518 #ifndef __PAGETABLE_PMD_FOLDED
3520 * Allocate page middle directory.
3521 * We've already handled the fast-path in-line.
3523 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3525 pmd_t *new = pmd_alloc_one(mm, address);
3529 smp_wmb(); /* See comment in __pte_alloc */
3531 spin_lock(&mm->page_table_lock);
3532 #ifndef __ARCH_HAS_4LEVEL_HACK
3533 if (!pud_present(*pud)) {
3535 pud_populate(mm, pud, new);
3536 } else /* Another has populated it */
3539 if (!pgd_present(*pud)) {
3541 pgd_populate(mm, pud, new);
3542 } else /* Another has populated it */
3544 #endif /* __ARCH_HAS_4LEVEL_HACK */
3545 spin_unlock(&mm->page_table_lock);
3548 #endif /* __PAGETABLE_PMD_FOLDED */
3550 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3551 pte_t **ptepp, spinlock_t **ptlp)
3558 pgd = pgd_offset(mm, address);
3559 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3562 pud = pud_offset(pgd, address);
3563 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3566 pmd = pmd_offset(pud, address);
3567 VM_BUG_ON(pmd_trans_huge(*pmd));
3568 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3571 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3575 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3578 if (!pte_present(*ptep))
3583 pte_unmap_unlock(ptep, *ptlp);
3588 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3589 pte_t **ptepp, spinlock_t **ptlp)
3593 /* (void) is needed to make gcc happy */
3594 (void) __cond_lock(*ptlp,
3595 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3600 * follow_pfn - look up PFN at a user virtual address
3601 * @vma: memory mapping
3602 * @address: user virtual address
3603 * @pfn: location to store found PFN
3605 * Only IO mappings and raw PFN mappings are allowed.
3607 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3609 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3616 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3619 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3622 *pfn = pte_pfn(*ptep);
3623 pte_unmap_unlock(ptep, ptl);
3626 EXPORT_SYMBOL(follow_pfn);
3628 #ifdef CONFIG_HAVE_IOREMAP_PROT
3629 int follow_phys(struct vm_area_struct *vma,
3630 unsigned long address, unsigned int flags,
3631 unsigned long *prot, resource_size_t *phys)
3637 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3640 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3644 if ((flags & FOLL_WRITE) && !pte_write(pte))
3647 *prot = pgprot_val(pte_pgprot(pte));
3648 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3652 pte_unmap_unlock(ptep, ptl);
3657 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3658 void *buf, int len, int write)
3660 resource_size_t phys_addr;
3661 unsigned long prot = 0;
3662 void __iomem *maddr;
3663 int offset = addr & (PAGE_SIZE-1);
3665 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3668 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3670 memcpy_toio(maddr + offset, buf, len);
3672 memcpy_fromio(buf, maddr + offset, len);
3677 EXPORT_SYMBOL_GPL(generic_access_phys);
3681 * Access another process' address space as given in mm. If non-NULL, use the
3682 * given task for page fault accounting.
3684 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3685 unsigned long addr, void *buf, int len, int write)
3687 struct vm_area_struct *vma;
3688 void *old_buf = buf;
3690 down_read(&mm->mmap_sem);
3691 /* ignore errors, just check how much was successfully transferred */
3693 int bytes, ret, offset;
3695 struct page *page = NULL;
3697 ret = get_user_pages(tsk, mm, addr, 1,
3698 write, 1, &page, &vma);
3700 #ifndef CONFIG_HAVE_IOREMAP_PROT
3704 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3705 * we can access using slightly different code.
3707 vma = find_vma(mm, addr);
3708 if (!vma || vma->vm_start > addr)
3710 if (vma->vm_ops && vma->vm_ops->access)
3711 ret = vma->vm_ops->access(vma, addr, buf,
3719 offset = addr & (PAGE_SIZE-1);
3720 if (bytes > PAGE_SIZE-offset)
3721 bytes = PAGE_SIZE-offset;
3725 copy_to_user_page(vma, page, addr,
3726 maddr + offset, buf, bytes);
3727 set_page_dirty_lock(page);
3729 copy_from_user_page(vma, page, addr,
3730 buf, maddr + offset, bytes);
3733 page_cache_release(page);
3739 up_read(&mm->mmap_sem);
3741 return buf - old_buf;
3745 * access_remote_vm - access another process' address space
3746 * @mm: the mm_struct of the target address space
3747 * @addr: start address to access
3748 * @buf: source or destination buffer
3749 * @len: number of bytes to transfer
3750 * @write: whether the access is a write
3752 * The caller must hold a reference on @mm.
3754 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3755 void *buf, int len, int write)
3757 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3761 * Access another process' address space.
3762 * Source/target buffer must be kernel space,
3763 * Do not walk the page table directly, use get_user_pages
3765 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3766 void *buf, int len, int write)
3768 struct mm_struct *mm;
3771 mm = get_task_mm(tsk);
3775 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3782 * Print the name of a VMA.
3784 void print_vma_addr(char *prefix, unsigned long ip)
3786 struct mm_struct *mm = current->mm;
3787 struct vm_area_struct *vma;
3790 * Do not print if we are in atomic
3791 * contexts (in exception stacks, etc.):
3793 if (preempt_count())
3796 down_read(&mm->mmap_sem);
3797 vma = find_vma(mm, ip);
3798 if (vma && vma->vm_file) {
3799 struct file *f = vma->vm_file;
3800 char *buf = (char *)__get_free_page(GFP_KERNEL);
3804 p = file_path(f, buf, PAGE_SIZE);
3807 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3809 vma->vm_end - vma->vm_start);
3810 free_page((unsigned long)buf);
3813 up_read(&mm->mmap_sem);
3816 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3817 void __might_fault(const char *file, int line)
3820 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3821 * holding the mmap_sem, this is safe because kernel memory doesn't
3822 * get paged out, therefore we'll never actually fault, and the
3823 * below annotations will generate false positives.
3825 if (segment_eq(get_fs(), KERNEL_DS))
3827 if (pagefault_disabled())
3829 __might_sleep(file, line, 0);
3830 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3832 might_lock_read(¤t->mm->mmap_sem);
3835 EXPORT_SYMBOL(__might_fault);
3838 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3839 static void clear_gigantic_page(struct page *page,
3841 unsigned int pages_per_huge_page)
3844 struct page *p = page;
3847 for (i = 0; i < pages_per_huge_page;
3848 i++, p = mem_map_next(p, page, i)) {
3850 clear_user_highpage(p, addr + i * PAGE_SIZE);
3853 void clear_huge_page(struct page *page,
3854 unsigned long addr, unsigned int pages_per_huge_page)
3858 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3859 clear_gigantic_page(page, addr, pages_per_huge_page);
3864 for (i = 0; i < pages_per_huge_page; i++) {
3866 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3870 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3872 struct vm_area_struct *vma,
3873 unsigned int pages_per_huge_page)
3876 struct page *dst_base = dst;
3877 struct page *src_base = src;
3879 for (i = 0; i < pages_per_huge_page; ) {
3881 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3884 dst = mem_map_next(dst, dst_base, i);
3885 src = mem_map_next(src, src_base, i);
3889 void copy_user_huge_page(struct page *dst, struct page *src,
3890 unsigned long addr, struct vm_area_struct *vma,
3891 unsigned int pages_per_huge_page)
3895 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3896 copy_user_gigantic_page(dst, src, addr, vma,
3897 pages_per_huge_page);
3902 for (i = 0; i < pages_per_huge_page; i++) {
3904 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3907 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3909 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3911 static struct kmem_cache *page_ptl_cachep;
3913 void __init ptlock_cache_init(void)
3915 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3919 bool ptlock_alloc(struct page *page)
3923 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3930 void ptlock_free(struct page *page)
3932 kmem_cache_free(page_ptl_cachep, page->ptl);