inetd fd leak?

[Doug Graham]

Denys Vlasenko wrote:
>> This isn't just an academic question.  uClibc 0.9.30 doesn't set FD_CLOEXEC
>> on the logging descriptor.
>>     
>
> It does. uClibc-0.9.30/libc/misc/syslog/syslog.c
>
> void
> openlog(const char *ident, int logstat, int logfac)
> {
>         int logType = SOCK_DGRAM;
>
>         __UCLIBC_MUTEX_LOCK(mylock);
>
>         if (ident != NULL)
>                 LogTag = ident;
>         LogStat = logstat;
>         if (logfac != 0 && (logfac &~ LOG_FACMASK) == 0)
>                 LogFacility = logfac;
>         if (LogFile == -1) {
> retry:
>                 if (LogStat & LOG_NDELAY) {
>                         if ((LogFile = socket(AF_UNIX, logType, 0)) == -1) {
>                                 goto DONE;
>                         }
>  HERE ====>             fcntl(LogFile, F_SETFD, 1); /* 1 == FD_CLOEXEC */
>                         /* We don't want to block if e.g. syslogd is SIGSTOPed */
>                         fcntl(LogFile, F_SETFL, O_NONBLOCK | fcntl(LogFile, F_GETFL));
>                 }
>         }
>   

Oops, I missed the tricky (and strange) hardcoding of the constant, 
because id-utils doesn't
index comments.  I did also use strace to verify on an actual running 
system that uClibc wasn't
setting FD_CLOEXEC, but that system is using uClibc-0.9.29.  So it looks 
like this was fixed
in 0.9.30.  I'd bet that there are still lots of systems using older 
versions of uClibc though.

--Doug.

> ...
>