catch other signals and unwind

On Wed, 2006-04-26 at 12:01 +0000, Martin Pool wrote:
> Public bug reported:
>
> Affects: bzr (upstream)
> Severity: Normal
> Priority: (none set)
> Status: Unconfirmed
>
>
> Description:
> Python catches SIGINT and translates it to a KeyboardInterrupt, giving
> the chance for finally blocks to run.
>
> When running ssh as a subprocess we should get it to ignore these
> signals so that bzr gets a chance to clean up using the connection.
>
> Jan Hudec suggests
>
> From these, at least HUP, INT, PIPE, TERM, USR1, USR2, PWR (where
> defined) and maybe QUIT are likely to occur occasionally and should be
> trapped.
>
> PIPE is special; HUP and TERM should be handled this way. USR1, 2 and
> PWR are rather special and not generally expected.
>
> QUIT conventionally does cause the process to quit fairly promptly; I'm
> not sure it should be ignored in ssh.
>
> While we're here, perhaps we should use one of them to start the
> debugger or show the stack?

IIRC linux threads uses SIGUSR1 and SIGUSR2 for some purpose. That may
be of historical interest these days with the new pthread
implementation. Either way, I think we should not install handlers for
SIGUSR1 and 2 randomly: sending those signals to us will be an explicit
choice from somewhere.

I think we should only catch exceptions we have an expectation of
recieving - rather than guessing that we can actually do the right thing
when this happens.

HUP is clear - logout with bzr running. INT is ctrl-C. TERM is clear.
PWR - if there is a power failure we may not have time to safely unlock
- but its worth considering. I agree - let QUIT remain unignored: its a
user override signal, and only sent in extremis.

I dont think showing the stack is useful; but invoking pdb is definately
useful.

Rob

--
GPG key available at: <http://www.robertcollins.net/keys.txt>.

OK I agree.

For other signals should we raise
 a- KeyboardInterrupt
 b- a subclass of KeyboardInterrupt whose string representation is e.g. "Terminated by hangup"
 c- some other exception?

I'd go for B - most likely to be handled correctly by any code that cares about the interrupt class.

I think we still want this, but probably shouldn't block for 0.9

I think this has been implemented?