* Stephen Hahn <sch at sun.com> [2009-10-14 21:53]:
> * Jordan Brown <Jordan.Brown at sun.com> [2009-10-11 17:47]:
> > >> 1) Just printf them, and rely on SMF's output redirection.
> > >> Problem: doesn't play nice with log rotation.
> > >> ...
> >>> But how about something like this?
> >>>
> >>> void
> >>> log(char *fmt, ...)
> >>> [ reopen logfile on each write ]
> >> That isn't any better than just (1), because smf already gives each
> >> thing a log, and I'm not sure what you meant above by "doesn't play
> >> nice w/ log rotation" but those logs *are* rotated. There is a
> >> single entry in logadm that triggers a script that rotates all smf logs.
> >
> > Sorry, I assumed context from the previous discussion a few months ago.
> >
> > The problem with log rotation is that the service has the log file open
> > on its stdout and stderr, and when logadm rotates logs it does it by
> > rename and so the open file descriptors refer to the archived log file,
> > *not* to the new one. The server doesn't start writing to the new log
> > file until it's restarted.
>
> There's an outstanding bug on this; the fix is trivial--use
> logadm(1M)'s copy-truncate option.
6578671 smf services keep logging to old smf logs after log
rotation as well, until we restart them
I peeked at this bug when fixes were being discussed; a small patch
to svc.startd(1M) to set stdout/stderr to O_APPEND would make for a
complete fix.
- Stephen
--
sch at sun.com http://blogs.sun.com/sch/
