msync.c revision b5810039a54e5babf428e9a1e89fc1940fabff11
1/* 2 * linux/mm/msync.c 3 * 4 * Copyright (C) 1994-1999 Linus Torvalds 5 */ 6 7/* 8 * The msync() system call. 9 */ 10#include <linux/slab.h> 11#include <linux/pagemap.h> 12#include <linux/mm.h> 13#include <linux/mman.h> 14#include <linux/hugetlb.h> 15#include <linux/syscalls.h> 16 17#include <asm/pgtable.h> 18#include <asm/tlbflush.h> 19 20/* 21 * Called with mm->page_table_lock held to protect against other 22 * threads/the swapper from ripping pte's out from under us. 23 */ 24 25static void msync_pte_range(struct vm_area_struct *vma, pmd_t *pmd, 26 unsigned long addr, unsigned long end) 27{ 28 struct mm_struct *mm = vma->vm_mm; 29 pte_t *pte; 30 int progress = 0; 31 32again: 33 pte = pte_offset_map(pmd, addr); 34 do { 35 unsigned long pfn; 36 struct page *page; 37 38 if (progress >= 64) { 39 progress = 0; 40 if (need_resched() || 41 need_lockbreak(&mm->page_table_lock)) 42 break; 43 } 44 progress++; 45 if (!pte_present(*pte)) 46 continue; 47 if (!pte_maybe_dirty(*pte)) 48 continue; 49 pfn = pte_pfn(*pte); 50 if (unlikely(!pfn_valid(pfn))) { 51 print_bad_pte(vma, *pte, addr); 52 continue; 53 } 54 page = pfn_to_page(pfn); 55 56 if (ptep_clear_flush_dirty(vma, addr, pte) || 57 page_test_and_clear_dirty(page)) 58 set_page_dirty(page); 59 progress += 3; 60 } while (pte++, addr += PAGE_SIZE, addr != end); 61 pte_unmap(pte - 1); 62 cond_resched_lock(&mm->page_table_lock); 63 if (addr != end) 64 goto again; 65} 66 67static inline void msync_pmd_range(struct vm_area_struct *vma, pud_t *pud, 68 unsigned long addr, unsigned long end) 69{ 70 pmd_t *pmd; 71 unsigned long next; 72 73 pmd = pmd_offset(pud, addr); 74 do { 75 next = pmd_addr_end(addr, end); 76 if (pmd_none_or_clear_bad(pmd)) 77 continue; 78 msync_pte_range(vma, pmd, addr, next); 79 } while (pmd++, addr = next, addr != end); 80} 81 82static inline void msync_pud_range(struct vm_area_struct *vma, pgd_t *pgd, 83 unsigned long addr, unsigned long end) 84{ 85 pud_t *pud; 86 unsigned long next; 87 88 pud = pud_offset(pgd, addr); 89 do { 90 next = pud_addr_end(addr, end); 91 if (pud_none_or_clear_bad(pud)) 92 continue; 93 msync_pmd_range(vma, pud, addr, next); 94 } while (pud++, addr = next, addr != end); 95} 96 97static void msync_page_range(struct vm_area_struct *vma, 98 unsigned long addr, unsigned long end) 99{ 100 struct mm_struct *mm = vma->vm_mm; 101 pgd_t *pgd; 102 unsigned long next; 103 104 /* For hugepages we can't go walking the page table normally, 105 * but that's ok, hugetlbfs is memory based, so we don't need 106 * to do anything more on an msync(). 107 * Can't do anything with VM_RESERVED regions either. 108 */ 109 if (vma->vm_flags & (VM_HUGETLB|VM_RESERVED)) 110 return; 111 112 BUG_ON(addr >= end); 113 pgd = pgd_offset(mm, addr); 114 flush_cache_range(vma, addr, end); 115 spin_lock(&mm->page_table_lock); 116 do { 117 next = pgd_addr_end(addr, end); 118 if (pgd_none_or_clear_bad(pgd)) 119 continue; 120 msync_pud_range(vma, pgd, addr, next); 121 } while (pgd++, addr = next, addr != end); 122 spin_unlock(&mm->page_table_lock); 123} 124 125/* 126 * MS_SYNC syncs the entire file - including mappings. 127 * 128 * MS_ASYNC does not start I/O (it used to, up to 2.5.67). Instead, it just 129 * marks the relevant pages dirty. The application may now run fsync() to 130 * write out the dirty pages and wait on the writeout and check the result. 131 * Or the application may run fadvise(FADV_DONTNEED) against the fd to start 132 * async writeout immediately. 133 * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to 134 * applications. 135 */ 136static int msync_interval(struct vm_area_struct *vma, 137 unsigned long addr, unsigned long end, int flags) 138{ 139 int ret = 0; 140 struct file *file = vma->vm_file; 141 142 if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED)) 143 return -EBUSY; 144 145 if (file && (vma->vm_flags & VM_SHARED)) { 146 msync_page_range(vma, addr, end); 147 148 if (flags & MS_SYNC) { 149 struct address_space *mapping = file->f_mapping; 150 int err; 151 152 ret = filemap_fdatawrite(mapping); 153 if (file->f_op && file->f_op->fsync) { 154 /* 155 * We don't take i_sem here because mmap_sem 156 * is already held. 157 */ 158 err = file->f_op->fsync(file,file->f_dentry,1); 159 if (err && !ret) 160 ret = err; 161 } 162 err = filemap_fdatawait(mapping); 163 if (!ret) 164 ret = err; 165 } 166 } 167 return ret; 168} 169 170asmlinkage long sys_msync(unsigned long start, size_t len, int flags) 171{ 172 unsigned long end; 173 struct vm_area_struct *vma; 174 int unmapped_error, error = -EINVAL; 175 176 if (flags & MS_SYNC) 177 current->flags |= PF_SYNCWRITE; 178 179 down_read(¤t->mm->mmap_sem); 180 if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC)) 181 goto out; 182 if (start & ~PAGE_MASK) 183 goto out; 184 if ((flags & MS_ASYNC) && (flags & MS_SYNC)) 185 goto out; 186 error = -ENOMEM; 187 len = (len + ~PAGE_MASK) & PAGE_MASK; 188 end = start + len; 189 if (end < start) 190 goto out; 191 error = 0; 192 if (end == start) 193 goto out; 194 /* 195 * If the interval [start,end) covers some unmapped address ranges, 196 * just ignore them, but return -ENOMEM at the end. 197 */ 198 vma = find_vma(current->mm, start); 199 unmapped_error = 0; 200 for (;;) { 201 /* Still start < end. */ 202 error = -ENOMEM; 203 if (!vma) 204 goto out; 205 /* Here start < vma->vm_end. */ 206 if (start < vma->vm_start) { 207 unmapped_error = -ENOMEM; 208 start = vma->vm_start; 209 } 210 /* Here vma->vm_start <= start < vma->vm_end. */ 211 if (end <= vma->vm_end) { 212 if (start < end) { 213 error = msync_interval(vma, start, end, flags); 214 if (error) 215 goto out; 216 } 217 error = unmapped_error; 218 goto out; 219 } 220 /* Here vma->vm_start <= start < vma->vm_end < end. */ 221 error = msync_interval(vma, start, vma->vm_end, flags); 222 if (error) 223 goto out; 224 start = vma->vm_end; 225 vma = vma->vm_next; 226 } 227out: 228 up_read(¤t->mm->mmap_sem); 229 current->flags &= ~PF_SYNCWRITE; 230 return error; 231} 232