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>
16 #include <asm/pgtable.h>
17 #include <asm/tlbflush.h>
20 * Called with mm->page_table_lock held to protect against other
21 * threads/the swapper from ripping pte's out from under us.
23 static int filemap_sync_pte(pte_t *ptep, struct vm_area_struct *vma,
24 unsigned long address, unsigned int flags)
27 unsigned long pfn = pte_pfn(pte);
30 if (pte_present(pte) && pfn_valid(pfn)) {
31 page = pfn_to_page(pfn);
32 if (!PageReserved(page) &&
33 (ptep_clear_flush_dirty(vma, address, ptep) ||
34 page_test_and_clear_dirty(page)))
40 static int filemap_sync_pte_range(pmd_t * pmd,
41 unsigned long address, unsigned long end,
42 struct vm_area_struct *vma, unsigned int flags)
54 pte = pte_offset_map(pmd, address);
55 if ((address & PMD_MASK) != (end & PMD_MASK))
56 end = (address & PMD_MASK) + PMD_SIZE;
59 error |= filemap_sync_pte(pte, vma, address, flags);
62 } while (address && (address < end));
69 static inline int filemap_sync_pmd_range(pgd_t * pgd,
70 unsigned long address, unsigned long end,
71 struct vm_area_struct *vma, unsigned int flags)
83 pmd = pmd_offset(pgd, address);
84 if ((address & PGDIR_MASK) != (end & PGDIR_MASK))
85 end = (address & PGDIR_MASK) + PGDIR_SIZE;
88 error |= filemap_sync_pte_range(pmd, address, end, vma, flags);
89 address = (address + PMD_SIZE) & PMD_MASK;
91 } while (address && (address < end));
95 static int filemap_sync(struct vm_area_struct * vma, unsigned long address,
96 size_t size, unsigned int flags)
99 unsigned long end = address + size;
102 /* Aquire the lock early; it may be possible to avoid dropping
103 * and reaquiring it repeatedly.
105 spin_lock(&vma->vm_mm->page_table_lock);
107 dir = pgd_offset(vma->vm_mm, address);
108 flush_cache_range(vma, address, end);
110 /* For hugepages we can't go walking the page table normally,
111 * but that's ok, hugetlbfs is memory based, so we don't need
112 * to do anything more on an msync() */
113 if (is_vm_hugetlb_page(vma))
119 error |= filemap_sync_pmd_range(dir, address, end, vma, flags);
120 address = (address + PGDIR_SIZE) & PGDIR_MASK;
122 } while (address && (address < end));
124 * Why flush ? filemap_sync_pte already flushed the tlbs with the
127 flush_tlb_range(vma, end - size, end);
129 spin_unlock(&vma->vm_mm->page_table_lock);
135 * MS_SYNC syncs the entire file - including mappings.
137 * MS_ASYNC does not start I/O (it used to, up to 2.5.67). Instead, it just
138 * marks the relevant pages dirty. The application may now run fsync() to
139 * write out the dirty pages and wait on the writeout and check the result.
140 * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
141 * async writeout immediately.
142 * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
145 static int msync_interval(struct vm_area_struct * vma,
146 unsigned long start, unsigned long end, int flags)
149 struct file * file = vma->vm_file;
151 if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
154 if (file && (vma->vm_flags & VM_SHARED)) {
155 ret = filemap_sync(vma, start, end-start, flags);
157 if (!ret && (flags & MS_SYNC)) {
158 struct address_space *mapping = file->f_mapping;
161 down(&mapping->host->i_sem);
162 ret = filemap_fdatawrite(mapping);
163 if (file->f_op && file->f_op->fsync) {
164 err = file->f_op->fsync(file,file->f_dentry,1);
168 err = filemap_fdatawait(mapping);
171 up(&mapping->host->i_sem);
177 asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
180 struct vm_area_struct * vma;
181 int unmapped_error, error = -EINVAL;
183 down_read(¤t->mm->mmap_sem);
184 if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
186 if (start & ~PAGE_MASK)
188 if ((flags & MS_ASYNC) && (flags & MS_SYNC))
191 len = (len + ~PAGE_MASK) & PAGE_MASK;
199 * If the interval [start,end) covers some unmapped address ranges,
200 * just ignore them, but return -ENOMEM at the end.
202 vma = find_vma(current->mm, start);
205 /* Still start < end. */
209 /* Here start < vma->vm_end. */
210 if (start < vma->vm_start) {
211 unmapped_error = -ENOMEM;
212 start = vma->vm_start;
214 /* Here vma->vm_start <= start < vma->vm_end. */
215 if (end <= vma->vm_end) {
217 error = msync_interval(vma, start, end, flags);
221 error = unmapped_error;
224 /* Here vma->vm_start <= start < vma->vm_end < end. */
225 error = msync_interval(vma, start, vma->vm_end, flags);
232 up_read(¤t->mm->mmap_sem);