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>
21 * Called with mm->page_table_lock held to protect against other
22 * threads/the swapper from ripping pte's out from under us.
25 static void sync_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
26 unsigned long addr, unsigned long end)
30 pte = pte_offset_map(pmd, addr);
35 if (!pte_present(*pte))
37 if (!pte_maybe_dirty(*pte))
42 page = pfn_to_page(pfn);
43 if (PageReserved(page))
46 if (ptep_clear_flush_dirty(vma, addr, pte) ||
47 page_test_and_clear_dirty(page))
49 } while (pte++, addr += PAGE_SIZE, addr != end);
53 static inline void sync_pmd_range(struct vm_area_struct *vma, pud_t *pud,
54 unsigned long addr, unsigned long end)
59 pmd = pmd_offset(pud, addr);
61 next = pmd_addr_end(addr, end);
62 if (pmd_none_or_clear_bad(pmd))
64 sync_pte_range(vma, pmd, addr, next);
65 } while (pmd++, addr = next, addr != end);
68 static inline void sync_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
69 unsigned long addr, unsigned long end)
74 pud = pud_offset(pgd, addr);
76 next = pud_addr_end(addr, end);
77 if (pud_none_or_clear_bad(pud))
79 sync_pmd_range(vma, pud, addr, next);
80 } while (pud++, addr = next, addr != end);
83 static void sync_page_range(struct vm_area_struct *vma,
84 unsigned long addr, unsigned long end)
86 struct mm_struct *mm = vma->vm_mm;
90 /* For hugepages we can't go walking the page table normally,
91 * but that's ok, hugetlbfs is memory based, so we don't need
92 * to do anything more on an msync() */
93 if (is_vm_hugetlb_page(vma))
97 pgd = pgd_offset(mm, addr);
98 flush_cache_range(vma, addr, end);
99 spin_lock(&mm->page_table_lock);
101 next = pgd_addr_end(addr, end);
102 if (pgd_none_or_clear_bad(pgd))
104 sync_pud_range(vma, pgd, addr, next);
105 } while (pgd++, addr = next, addr != end);
106 spin_unlock(&mm->page_table_lock);
109 #ifdef CONFIG_PREEMPT
110 static inline void filemap_sync(struct vm_area_struct *vma,
111 unsigned long addr, unsigned long end)
113 const size_t chunk = 64 * 1024; /* bytes */
118 if (next > end || next < addr)
120 sync_page_range(vma, addr, next);
122 } while (addr = next, addr != end);
125 static inline void filemap_sync(struct vm_area_struct *vma,
126 unsigned long addr, unsigned long end)
128 sync_page_range(vma, addr, end);
133 * MS_SYNC syncs the entire file - including mappings.
135 * MS_ASYNC does not start I/O (it used to, up to 2.5.67). Instead, it just
136 * marks the relevant pages dirty. The application may now run fsync() to
137 * write out the dirty pages and wait on the writeout and check the result.
138 * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
139 * async writeout immediately.
140 * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
143 static int msync_interval(struct vm_area_struct *vma,
144 unsigned long addr, unsigned long end, int flags)
147 struct file *file = vma->vm_file;
149 if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
152 if (file && (vma->vm_flags & VM_SHARED)) {
153 filemap_sync(vma, addr, end);
155 if (flags & MS_SYNC) {
156 struct address_space *mapping = file->f_mapping;
159 ret = filemap_fdatawrite(mapping);
160 if (file->f_op && file->f_op->fsync) {
162 * We don't take i_sem here because mmap_sem
165 err = file->f_op->fsync(file,file->f_dentry,1);
169 err = filemap_fdatawait(mapping);
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;
184 current->flags |= PF_SYNCWRITE;
186 down_read(¤t->mm->mmap_sem);
187 if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
189 if (start & ~PAGE_MASK)
191 if ((flags & MS_ASYNC) && (flags & MS_SYNC))
194 len = (len + ~PAGE_MASK) & PAGE_MASK;
202 * If the interval [start,end) covers some unmapped address ranges,
203 * just ignore them, but return -ENOMEM at the end.
205 vma = find_vma(current->mm, start);
208 /* Still start < end. */
212 /* Here start < vma->vm_end. */
213 if (start < vma->vm_start) {
214 unmapped_error = -ENOMEM;
215 start = vma->vm_start;
217 /* Here vma->vm_start <= start < vma->vm_end. */
218 if (end <= vma->vm_end) {
220 error = msync_interval(vma, start, end, flags);
224 error = unmapped_error;
227 /* Here vma->vm_start <= start < vma->vm_end < end. */
228 error = msync_interval(vma, start, vma->vm_end, flags);
235 up_read(¤t->mm->mmap_sem);
236 current->flags &= ~PF_SYNCWRITE;