[plug] kill them dead

[Adrian Chadd]

On Sat, Oct 14, 2006, Michael Holland wrote:
> On Fri, 13 Oct 2006, Adrian Chadd wrote:
> 
> > It doesn't hvae to be a kernel bug.
> >
> > You can't kill -9 a process locked waiting for NFS IO to complete, IIRC.
> > There's more examples of that. Its certainly not a bug. :)
> 
> That is debatable :-)
> At least you can mount NFS with the 'intr' option. But when a process can
> only be killed by a power-cyle, I call that a bug. e.g. I/O driver
> blocking on media errors. But then, IANAKP.

Ah, but is the media error recoverable or not? :)

Yes, it probably should be interruptable. But, can you actually -abort-
a media operation? Does the hardware let you? Will you be in a consistent
state if you do? Its not always immediately obvious. If its the device
that locks up during an operation should you abort the process' operation
and silently fail all subsequent attempts at device IO? I don't think
thats the "unix" way.. 

That said, it could just be that the device driver programmers' hardware
never locked up during media errors.. :P

IIRC, Windows actually has this stuff (mostly) right. You can abort a
pending async read/write op and said op, if its in flight, will be attempted
to be aborted. I forget how I know this.



Adrian