Commit b9c93bb7 authored Feb 02, 2011 by Steven Whitehouse

GFS2: Improve cluster mmap scalability



The mmap system call grabs a glock when an update to atime maybe
required. It does this in order to ensure that the flags on the
inode are uptodate, but since it will only mark atime for a future
update, an exclusive lock is not required here (one will be taken
later when the actual update is performed).

Also, the lock can be skipped when the mount is marked noatime in
addition to the original check which only looked at the noatime
flag for the inode itself.

This should increase the scalability of the mmap call when multiple
nodes are all mmaping the same file.

Reported-by: Scooter Morris <scooter@cgl.ucsf.edu>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>

parent edae38a6

fs/gfs2/file.c

+10 −5

Original line number	Diff line number	Diff line
		@@ -448,15 +448,20 @@ static int gfs2_mmap(struct file file, struct vm_area_struct vma)
		{
		struct gfs2_inode *ip = GFS2_I(file->f_mapping->host);

		if (!(file->f_flags & O_NOATIME)) {
		if (!(file->f_flags & O_NOATIME) &&
		!IS_NOATIME(&ip->i_inode)) {
		struct gfs2_holder i_gh;
		int error;

		gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &i_gh);
		gfs2_holder_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
		error = gfs2_glock_nq(&i_gh);
		if (error == 0) {
		file_accessed(file);
		if (error == 0)
		gfs2_glock_dq_uninit(&i_gh);
		gfs2_glock_dq(&i_gh);
		}
		gfs2_holder_uninit(&i_gh);
		if (error)
		return error;
		}
		vma->vm_ops = &gfs2_vm_ops;
		vma->vm_flags \|= VM_CAN_NONLINEAR;

Admin message