Re: [dm-devel] Re: [RFC] device-mapper ioctl interface

[Daniel Phillips]

On Thursday 05 June 2003 19:50, Kevin Corry wrote:
> On Thursday 05 June 2003 12:00, Daniel Phillips wrote:
> > On Thursday 05 June 2003 18:47, Kevin Corry wrote:
> > > 2) Removing suspended devices. The current code (2.5.70) does not allow
> > > a suspended device to be removed/unlinked from the ioctl interface,
> > > since removing it would leave you with no way to resume it (and hence
> > > flush any pending I/Os). Alasdair mentioned a couple of new ideas. One
> > > would be to reload the device with an error-map and force it to resume,
> > > thus erroring any pending I/Os and allowing the device to be removed.
> > > This seems a bit heavy-handed.
> >
> > Which is the heavy-handed part?
>
> The part about automatically reloading the table with an error map and
> forcing it to resume. It just seemed to me that user-space ought to be able
> to gather enough information to determine that a device needed to be
> resumed before it could be removed. Thus the kernel driver wouldn't be
> forced to implement such a policy.

I didn't see anything about doing that in-kernel.

> Talking with Alasadair again, he mentioned a case I hadn't considered.
> Devices would now be created without a mapping and initially suspended. If
> some other error occurred, and you decided to just delete the device before
> loading a mapping, it would fail.  And having to resume a device with no
> mapping just to be able to delete it definitely seems odd.
>
> So, it's not like I'm dead-set against this idea. I was just curious what
> the reasoning was behind this change.

It's similar to the way a lot of things work in Linux: you have to let 
operations run to completion so they can let go of resources.  One day we'll 
be able to shoot down transfers in mid-flight, but I doubt that's going to 
happen in this cycle.

So in general, the idea is: let any outstanding operations complete, but feed 
them errors.  What else can we do?

I don't see this as heavyweight at all.  Policy stays in user space, and a 
lightweight error path lives in the kernel.

Regards,

Daniel