Re: md /tmp and async mounts

On Tue, May 23, 2006 at 04:30:13PM +0200, mlfbsd wrote:
> On Tue, May 23, 2006 at 04:30:37PM +0300, Giorgos Keramidas wrote:
> > On 2006-05-22 16:14, Brooks Davis <brooks_at_one-eyed-alien.net> wrote:
> > >On Mon, May 22, 2006 at 05:12:24PM -0400, Kris Kennaway wrote:
> > >> I noticed that with tmpmfs in rc.conf, the md is not being mounted
> > >> with -o async; this is an enormous performance boost, so is there a
> > >> reason it is not used, or just that it was overlooked?
> > >
> > > Just overlooked as far as I know.  We correctly switched to not mounting
> > > with softupdates, but should probably switch to -o async.
> > 
> > Right.  The only reasons why the current tmpmfs_flags doesn't
> > include the `-o async' option too are:
> > 
> >     a) It didn't back when it was all hardcoded in the script
> > 
> >     b) I didn't think about adding it by default when revision
> >        1.230 of src/etc/defaults/rc.conf was committed by me
> > 
> > If it does improve things, we should probably make the change :)
> > 
> 
> The sys/dev/md/md.c rev 1.115 commit log says async used to be the default, 
> but was switched off because it was causing deadlocks. Did this get fixed
> since ?

You are confusing mdconfig's -o async mode, used only for vnode
backing (not the case here anyway):

             [no]async
                     For vnode backed devices: avoid IO_SYNC for increased
                     performance but at the risk of deadlocking the entire
                     kernel.

with mount's -o async mode, used at the filesystem layer on any
device.  The latter does not have deadlocks.

Kris