Re: softirq considered harmful

[Jens Axboe <axboe@suse.de>]

On Sat, Aug 12 2006, Andrew Morton wrote:
> On Thu, 10 Aug 2006 13:06:27 +0200
> Jens Axboe <axboe@suse.de> wrote:
> 
> > Ok maybe that is a little too strong, but I am indeed seeing some very
> > sucky behaviour with softirqs here. The block layer uses it for doing
> > request completions
> 
> I wasn't even aware that this change had been made.  I don't recall (and I
> cannot locate) any mailing list discussion of it.
> 
> Maybe I missed the discussion.  But if not, this is yet another case of
> significant changes getting into mainline via a git merge and sneaking
> under everyone's radar.

It's not a significant change, it pretty much falls into a code
relocation issue. The softirq completion stuff was made generic, and
SCSI the primary user of it. The completion path didn't change for SCSI.

> It seems like a bad idea to me.  Any additional latency at all in disk
> completion adds directly onto process waiting time - we do a _lot_ of
> synchronous disk IO.

Doesn't seem like a good idea to me either, hence I'm investigating the
current possible problems with it...

> There is no mention in the changelog of any observed problems which this
> patch solves.  Can we revert it please?

As you should see now, it wont change anything. The problem would be
the same.

-- 
Jens Axboe