On Wed, Jul 26, 2006 at 07:33:27AM +1000, Daniel Carosone wrote:
> On Mon, Jul 24, 2006 at 01:16:30PM +0300, Jukka Marin wrote:
> > A few times during the copy, the system froze completely and then
> > woke up again. 
> 
> This is the same issue Thor described just a few days ago in response
> to another question on, I think, tech-kern.
> 
> Softdep is generating a huge flood of metadata writes, many more than
> the disk subsystem can process, and ~everything gets jammed waiting
> for those to clear.  Processes blocking directly on the disk get
> starved, and other processes can get stuck on the vnode lock for the
> root of the filesystem.

How can I confirm that this is really the case?  I was writing to a
partition where softdep was _not_ enabled.

I tried the same "test" again.  I had another machine pinging the one
writing to a non-softdep partition.  After writing some 2.5 GB, the
machine froze for 33 seconds: no ping replies, no activity in shell
windows - and the hard disk LED was off, so there was no hard disk
activity, either.  This is what I see on a 2.0 system as well - it looks
like the machine is dead (no disk activity), but after a while, everything
returns to normal.

> It's not hung, though it can seem that way especially if you're
> running X and the Xserver is one of the processes that gets stuck.

No softdep, no disk activity, processes not running, no ping replies. ;-)

  -jm