mm/msync.c: sync only the requested range in msync()

msync() currently syncs more than POSIX requires or BSD or Solaris
implement.  It is supposed to be equivalent to fdatasync(), not fsync(),
and it is only supposed to sync the portion of the file that overlaps the
range passed to msync.

If the VMA is non-linear, fall back to syncing the entire file, but we
still optimise to only fdatasync() the entire file, not the full fsync().

akpm: there are obvious concerns with bck-compatibility: is anyone relying
on the undocumented side-effect for their data integrity?  And how would
they ever know if this change broke their data integrity?

We think the risk is reasonably low, and this patch brings the kernel into
line with other OS's and with what the manpage has always said...

Signed-off-by: Matthew Wilcox <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Acked-by: Jeff Moyer <[email protected]>
Cc: Chris Mason <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

1 files changed, 7 insertions, 1 deletions

diff --git a/mm/msync.c b/mm/msync.c
index 632df4527c01..a5c673669ca6 100644
--- a/mm/msync.c
+++ b/mm/msync.c
@@ -58,6 +58,7 @@ SYSCALL_DEFINE3(msync, unsigned long, start, size_t, len, int, flags)
 	vma = find_vma(mm, start);
 	for (;;) {
 		struct file *file;
+		loff_t fstart, fend;
 
 		/* Still start < end. */
 		error = -ENOMEM;
@@ -77,12 +78,17 @@ SYSCALL_DEFINE3(msync, unsigned long, start, size_t, len, int, flags)
 			goto out_unlock;
 		}
 		file = vma->vm_file;
+		fstart = start + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
+		fend = fstart + (min(end, vma->vm_end) - start) - 1;
 		start = vma->vm_end;
 		if ((flags & MS_SYNC) && file &&
 				(vma->vm_flags & VM_SHARED)) {
 			get_file(file);
 			up_read(&mm->mmap_sem);
-			error = vfs_fsync(file, 0);
+			if (vma->vm_flags & VM_NONLINEAR)
+				error = vfs_fsync(file, 1);
+			else
+				error = vfs_fsync_range(file, fstart, fend, 1);
 			fput(file);
 			if (error || start >= end)
 				goto out;