Re: [PATCH] git daemon: avoid calling syslog() from a signal handler

[Johannes Schindelin <Johannes.Schindelin@gmx.de>]

Hi,

On Sat, 5 Jul 2008, Junio C Hamano wrote:

> Johannes Schindelin <Johannes.Schindelin@gmx.de> writes:
> 
> >> So you will check every 1ms to see if there are new dead children, 
> >> but why is this necessary?
> >
> > This comes from me not reading the man page for poll() properly.  Of 
> > course, I want to check every second: syslog timestamps the messages 
> > with a resolution of 1 second, AFAIR, or at least some of them do.
> 
> Hmm.
> 
> The question was not about the millisecond typo, but about why time-out 
> at all.

Because I do not want to change the semantics!

ATM, in those cases where it works (as opposed to hanging!), git-daemon 
--verbose reports in the syslog when a client disconnected, possibly with 
an error.  It does so with a timestamp so that you can see how long the 
connection lasted.  That is what logs are useful for.

Now, syslog has timestamps at second-resolution (at least here it does), 
and I wanted to imitate that.

The alternative would be to deprive all users of an (mostly) accurate 
timestamp of the disconnect time.

> Another way would be to set up a pipe to ourself that is included in the 
> poll() and write a byte to the pipe from the signal handler.

It still would need to break out of the poll(), in which case the effect 
would be _exactly_ the same, but with a lot of more trouble, and 
opportunities for me to bring in new bugs, right?

Ciao,
Dscho