Re: [PATCH] Add a 'wait-scan' command to /proc/scsi/scsi.

[Matthew Wilcox <matthew@wil.cx>]

On Wed, Mar 25, 2009 at 03:47:56PM -0400, Bill Nottingham wrote:
> Matthew Wilcox (matthew@wil.cx) said: 
> > On Wed, Mar 25, 2009 at 03:03:21PM -0400, Bill Nottingham wrote:
> > > Then where is a better place to put this, as scsi_wait_scan.ko is
> > > a ridiculous interface for userspace?
> > 
> > It would be nice if people would comment on "ridiculous interface"s when
> > they're asked for feedback, instead of waiting more than two years.
> 
> Sure, but asking all people who might eventually have to use it
> to always watch any possible interface addition isn't practical.

Right.  I asked several people at Red Hat about the interface and I got a
"yeah, OK, whatever" kind of response.  Clearly you need to educate your
colleagues to pass these kinds of interface questions along to you.

> I would have hoped that the fact that the interface required loading
> a module and immediately removing it by hand is suboptimal enough
> that it wouldn't have gotten in in the first place.

It seems pretty elegant to me, actually.  There's no overhead after
you're done (unlike having a sysfs file, or even including a new ability
in a procfs file).

> > I think you're misunderstanding how to use scsi_wait_scan.  The idea was
> > that the bit of userspace that probes all the device drivers would do:
> > 
> > modprobe fusion.ko
> > modprobe aic79xx.ko
> > modprobe sym53c8xx.ko
> > modprobe scsi_wait_scan
> > rmmod scsi_wait_scan
> > 
> > et voila, you're done.  It seems like you want random other bits of
> > userspace to wait for scsi scanning to be done, and that wasn't the
> > original intent.
> 
> Well, in the case I'm looking at, udev is what's loading the host
> controllers, and there needs to be some sort of synchronization point
> between that and LVM invocations, fsck, mount, etc. Since scans
> aren't sent over as events for udev to catch, 'udevadm settle'
> isn't enough.

So ... if we sent a udev event when the scan list was empty, you'd be OK?

> Removing, loading, and removing scsi_wait_scan works
> here, but it just seems like a kludge.

I don't quite understand why it was loaded, and not unloaded immediately.

> I can trigger a load of scsi_wait_scan when hosts are registered
> in udev, but that's still ugly, and sort of overkill.

That would rather miss the point, yes.

-- 
Matthew Wilcox				Intel Open Source Technology Centre
"Bill, look, we understand that you're interested in selling us this
operating system, but compare it to ours.  We can't possibly take such
a retrograde step."