Serious issue with serial console in 5.4

[Robert Watson]

On Thu, 21 Jul 2005, Eirik Øverby wrote:

>>> The above panic will show up occasionally when logging out from a
>>> serial console (i.e. ctrl-D, logout, exit, whatever). This is
>>> EXTREMELY BAD, as it will crash an otherwise perfectly healthy box at
>>> random - and renders the serial console useless.
>>> 
>>> Robert Watson confirmed this to be an issue on the 10th of April.
>> 
>> You might have to wait until 6.0-R since fixing it seems to require 
>> infrastructure changes that cannot easily be backported to 5.x.
>
> With all due respect - if this is (and I'm assuming it is, because it 
> happens on all the servers I'm serial-controlling) an omnipresent 
> problem on 5.x, I daresay it should warrant some more attention. Having 
> unsafe serial terminal support that can bring down your system like that 
> defies much of the point of having serial terminal support in the first 
> place.
>
> However, since I seem to be the only one who has noticed this, perhaps 
> I'm the last person on earth to routinely use serial terminal switches 
> instead of KVM switches to do my admin work?

The concern about the 5.x backport is that it will break parts of the 
device driver ABI, and is a significant change that involves a lot of 
risk.

Regarding the general prevalence of the problem -- I've seen a small 
number of people reporting it's a big problem.  Since I know of a great 
many people running with serial consoles (other than a workstation, I 
never run FreeBSD boxes any other way), this leads me to believe it's 
something that shows up in fairly specific conditions -- perhaps relating 
to precise timing of a race condition.  This means that if we introduce a 
generally destabilizing change, it may impact more people than the problem 
as it exists (a nasty trade-off).

I've only seen the issue when logging out of a serial console session, and 
had previously hypothesized that it had to do with the simultaneous timing 
of a console message from syslog and the opening/closing of the console's 
tty due to logging out and getty restarting, resulting in a reference 
count improperly hitting zero.

I thought Doug White had come up with a work-around patch that prevented 
the reference count from being allowed to hit 0 for the console by 
artificially elevating it, which would prevent the panic, so either (a) 
the work around wasn't committed, or (b) it didn't work.

I can attempt to take another look at this problem in a week or so, but 
have a number of things I need to finish up for FreeBSD 6.0 before then 
that will be occupying my time.

Robert N M Watson