Re: [Fastboot] [RFC][PATCH] Add missing notifier before crashing

[ebiederm@xmission.com (Eric W. Biederman)]

Preben Traerup <Preben.Trarup@ericsson.com> writes:

> Something like out of memory and oops-es are enough to deeme the system must
> panic
> because it is simply not supposed to happen in a Telco server at any time.

That is clearly enough to deem that the system must take some sever action and
stop running.  You don't necessarily have to handle it through a kernel panic.

> kdump helps debugging these cases, but more importantly another server
> must take over the work, and this has and always will have highest priority.
>
> I'm happy about what crash_kexec does today, but the timing issue makes it
> unusable for
> notifications to external systems, if I need to wait until properly running in
> next kernel.

Nothing says you have to wait until properly running in the next kernel.
You can also write a dedicated piece of code that just pushes one packet
out the NIC.  Then you can start up a kernel for analysis purposes.

Although I have a hard time believing you can't tune a kernel to start
up quickly enough if you leave out just about everything.

And for purposes analysis assume that the oops happened somewhere in 
the network stack.

> That leaves me the choice of doing notification before executing crash_kexec ?

Only because you have assumed that you have to start another kernel and 
starting that other kernel must be an expensive operation.

> Since I'm apperantly not the only one left with this choice I rather prefer a
> solution
> made in public, that is known to be "bad" in some (well known) situations than
> each and everybody implements their own solution to the same problem.

It is certainly worth discussing.

Eric