4 * Copyright (C) 1994-1999 Linus Torvalds
8 * The msync() system call.
10 #include <linux/slab.h>
11 #include <linux/pagemap.h>
13 #include <linux/mman.h>
14 #include <linux/hugetlb.h>
15 #include <linux/writeback.h>
16 #include <linux/file.h>
17 #include <linux/syscalls.h>
19 #include <asm/pgtable.h>
20 #include <asm/tlbflush.h>
22 static unsigned long msync_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
23 unsigned long addr, unsigned long end)
28 unsigned long ret = 0;
31 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
37 if (need_resched() || need_lockbreak(ptl))
41 if (!pte_present(*pte))
43 if (!pte_maybe_dirty(*pte))
45 page = vm_normal_page(vma, addr, *pte);
48 if (ptep_clear_flush_dirty(vma, addr, pte) ||
49 page_test_and_clear_dirty(page))
50 ret += set_page_dirty(page);
52 } while (pte++, addr += PAGE_SIZE, addr != end);
53 pte_unmap_unlock(pte - 1, ptl);
60 static inline unsigned long msync_pmd_range(struct vm_area_struct *vma,
61 pud_t *pud, unsigned long addr, unsigned long end)
65 unsigned long ret = 0;
67 pmd = pmd_offset(pud, addr);
69 next = pmd_addr_end(addr, end);
70 if (pmd_none_or_clear_bad(pmd))
72 ret += msync_pte_range(vma, pmd, addr, next);
73 } while (pmd++, addr = next, addr != end);
77 static inline unsigned long msync_pud_range(struct vm_area_struct *vma,
78 pgd_t *pgd, unsigned long addr, unsigned long end)
82 unsigned long ret = 0;
84 pud = pud_offset(pgd, addr);
86 next = pud_addr_end(addr, end);
87 if (pud_none_or_clear_bad(pud))
89 ret += msync_pmd_range(vma, pud, addr, next);
90 } while (pud++, addr = next, addr != end);
94 static unsigned long msync_page_range(struct vm_area_struct *vma,
95 unsigned long addr, unsigned long end)
99 unsigned long ret = 0;
101 /* For hugepages we can't go walking the page table normally,
102 * but that's ok, hugetlbfs is memory based, so we don't need
103 * to do anything more on an msync().
105 if (vma->vm_flags & VM_HUGETLB)
109 pgd = pgd_offset(vma->vm_mm, addr);
110 flush_cache_range(vma, addr, end);
112 next = pgd_addr_end(addr, end);
113 if (pgd_none_or_clear_bad(pgd))
115 ret += msync_pud_range(vma, pgd, addr, next);
116 } while (pgd++, addr = next, addr != end);
121 * MS_SYNC syncs the entire file - including mappings.
123 * MS_ASYNC does not start I/O (it used to, up to 2.5.67). Instead, it just
124 * marks the relevant pages dirty. The application may now run fsync() to
125 * write out the dirty pages and wait on the writeout and check the result.
126 * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
127 * async writeout immediately.
128 * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
131 static int msync_interval(struct vm_area_struct *vma, unsigned long addr,
132 unsigned long end, int flags,
133 unsigned long *nr_pages_dirtied)
136 struct file *file = vma->vm_file;
138 if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
141 if (file && (vma->vm_flags & VM_SHARED)) {
142 *nr_pages_dirtied = msync_page_range(vma, addr, end);
144 if (flags & MS_SYNC) {
145 struct address_space *mapping = file->f_mapping;
148 ret = filemap_fdatawrite(mapping);
149 if (file->f_op && file->f_op->fsync) {
151 * We don't take i_mutex here because mmap_sem
154 err = file->f_op->fsync(file,file->f_dentry,1);
158 err = filemap_fdatawait(mapping);
166 asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
169 struct vm_area_struct *vma;
170 int unmapped_error, error = -EINVAL;
173 if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
175 if (start & ~PAGE_MASK)
177 if ((flags & MS_ASYNC) && (flags & MS_SYNC))
180 len = (len + ~PAGE_MASK) & PAGE_MASK;
188 * If the interval [start,end) covers some unmapped address ranges,
189 * just ignore them, but return -ENOMEM at the end.
191 down_read(¤t->mm->mmap_sem);
193 current->flags |= PF_SYNCWRITE;
194 vma = find_vma(current->mm, start);
197 unsigned long nr_pages_dirtied = 0;
200 /* Still start < end. */
204 /* Here start < vma->vm_end. */
205 if (start < vma->vm_start) {
206 unmapped_error = -ENOMEM;
207 start = vma->vm_start;
209 /* Here vma->vm_start <= start < vma->vm_end. */
210 if (end <= vma->vm_end) {
212 error = msync_interval(vma, start, end, flags,
217 error = unmapped_error;
220 /* Here vma->vm_start <= start < vma->vm_end < end. */
221 error = msync_interval(vma, start, vma->vm_end, flags,
228 if ((flags & MS_ASYNC) && file && nr_pages_dirtied) {
230 up_read(¤t->mm->mmap_sem);
231 balance_dirty_pages_ratelimited_nr(file->f_mapping,
234 down_read(¤t->mm->mmap_sem);
235 vma = find_vma(current->mm, start);
241 current->flags &= ~PF_SYNCWRITE;
242 up_read(¤t->mm->mmap_sem);