[systemd-devel] systemd intentional behaviour? (signals)

[Lennart Poettering]

On Fri, 07.09.12 12:52, Alex Elsayed (eternaleye+usenet at gmail.com) wrote:

> 
> Lennart Poettering wrote:
> > 
> > If we are crashed we are crashed.
> > 
> > I looked into reexecing systemd in this case, but unfortunately the
> > kernel does not allow reexecing of PID 1 after it crashed. Not entirely
> > sure what's going on there. But even if that worked I am not sure this
> > would be too useful since most of its state would be lost...
> > 
> > Lennart
> > 
> 
> Once cgroup xattrs go in, would serializing state into xattrs in the cgroup 
> tree on-the-fly be an option? Eliminating the need to reactively serialize 
> on reexec by proactively serializing in a continuous manner might resolve 
> the loss-of-state issue.

Yes, xattrs will help for this, and the ability to recover from crashes
in a nice way is one of the main reasons for me to want this
functionality in the kernel.

That said this is probably not going to allow us to restore things without
any loss of state. I mean, not all unit types are even exposed as a cgroup,
and hence no xattrs could be stored on them... But OTOH restoring some
state is still better than restoring no state at all.

This whole discussion is moot however as long as the kernel does not
allow PID 1 reexec after SIGSEGV...

Lennart

-- 
Lennart Poettering - Red Hat, Inc.