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/syscalls.h>
17 #include <asm/pgtable.h>
18 #include <asm/tlbflush.h>
20 static void msync_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
21 unsigned long addr, unsigned long end)
28 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
34 if (need_resched() || need_lockbreak(ptl))
38 if (!pte_present(*pte))
40 if (!pte_maybe_dirty(*pte))
42 page = vm_normal_page(vma, addr, *pte);
45 if (ptep_clear_flush_dirty(vma, addr, pte) ||
46 page_test_and_clear_dirty(page))
49 } while (pte++, addr += PAGE_SIZE, addr != end);
50 pte_unmap_unlock(pte - 1, ptl);
56 static inline void msync_pmd_range(struct vm_area_struct *vma, pud_t *pud,
57 unsigned long addr, unsigned long end)
62 pmd = pmd_offset(pud, addr);
64 next = pmd_addr_end(addr, end);
65 if (pmd_none_or_clear_bad(pmd))
67 msync_pte_range(vma, pmd, addr, next);
68 } while (pmd++, addr = next, addr != end);
71 static inline void msync_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
72 unsigned long addr, unsigned long end)
77 pud = pud_offset(pgd, addr);
79 next = pud_addr_end(addr, end);
80 if (pud_none_or_clear_bad(pud))
82 msync_pmd_range(vma, pud, addr, next);
83 } while (pud++, addr = next, addr != end);
86 static void msync_page_range(struct vm_area_struct *vma,
87 unsigned long addr, unsigned long end)
92 /* For hugepages we can't go walking the page table normally,
93 * but that's ok, hugetlbfs is memory based, so we don't need
94 * to do anything more on an msync().
96 if (vma->vm_flags & VM_HUGETLB)
100 pgd = pgd_offset(vma->vm_mm, addr);
101 flush_cache_range(vma, addr, end);
103 next = pgd_addr_end(addr, end);
104 if (pgd_none_or_clear_bad(pgd))
106 msync_pud_range(vma, pgd, addr, next);
107 } while (pgd++, addr = next, addr != end);
111 * MS_SYNC syncs the entire file - including mappings.
113 * MS_ASYNC does not start I/O (it used to, up to 2.5.67). Instead, it just
114 * marks the relevant pages dirty. The application may now run fsync() to
115 * write out the dirty pages and wait on the writeout and check the result.
116 * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
117 * async writeout immediately.
118 * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
121 static int msync_interval(struct vm_area_struct *vma,
122 unsigned long addr, unsigned long end, int flags)
125 struct file *file = vma->vm_file;
127 if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
130 if (file && (vma->vm_flags & VM_SHARED)) {
131 msync_page_range(vma, addr, end);
133 if (flags & MS_SYNC) {
134 struct address_space *mapping = file->f_mapping;
137 ret = filemap_fdatawrite(mapping);
138 if (file->f_op && file->f_op->fsync) {
140 * We don't take i_sem here because mmap_sem
143 err = file->f_op->fsync(file,file->f_dentry,1);
147 err = filemap_fdatawait(mapping);
155 asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
158 struct vm_area_struct *vma;
159 int unmapped_error, error = -EINVAL;
162 current->flags |= PF_SYNCWRITE;
164 down_read(¤t->mm->mmap_sem);
165 if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
167 if (start & ~PAGE_MASK)
169 if ((flags & MS_ASYNC) && (flags & MS_SYNC))
172 len = (len + ~PAGE_MASK) & PAGE_MASK;
180 * If the interval [start,end) covers some unmapped address ranges,
181 * just ignore them, but return -ENOMEM at the end.
183 vma = find_vma(current->mm, start);
186 /* Still start < end. */
190 /* Here start < vma->vm_end. */
191 if (start < vma->vm_start) {
192 unmapped_error = -ENOMEM;
193 start = vma->vm_start;
195 /* Here vma->vm_start <= start < vma->vm_end. */
196 if (end <= vma->vm_end) {
198 error = msync_interval(vma, start, end, flags);
202 error = unmapped_error;
205 /* Here vma->vm_start <= start < vma->vm_end < end. */
206 error = msync_interval(vma, start, vma->vm_end, flags);
213 up_read(¤t->mm->mmap_sem);
214 current->flags &= ~PF_SYNCWRITE;