Re: runsv failing when starting up logger - missing pipe - failure of logpipe init?

[Laurent Bercot <ska-supervision@skarnet.org>]

>> Has anyone else seen this error condition or can posit a situation where 
>> it might be seen?
> 
> The next question to ponder is where the bug lies. The runsv process here 
> has no fd 5 and fd 6 - IOW, logpipe[0] is 5, but isn't a valid fd. Are 
> there circumstances where a pipe can just cease to be? Should runsv have 
> detected this issue (where pipe() did not return -1, but the fds returned 
> were not valid)?
> 
> Is this a linux kernel bug?

 Before accusing the Linux kernel, let's check the runsv code and see
whether there's a possible execution path that leads to the situation
you're describing...

 The pipe creation part looks correct.
 The part where the error occurs looks correct.
 Okay, so is there a place where the pipe might be closed? Sure enough,
there is: right at the end, if svd[0].want == W_EXIT, svd[0].state == DOWN,
svd[1].pid != 0 and svd[1].want != W_EXIT, then logpipe[1] and logpipe[0]
both get closed. And this is the only place where it can happen.

 My bet is that at some point, your runsv ran through that code, but
somehow managed to live and the services didn't die, i.e. another control
message was sent and processed before the exit condition was reached, and
runsv is still trying to supervise things - but runs into trouble with the
closed logpipe. I have no time to investigate further right now, but earlier
in your strace, you should see stuff such as the control messages arriving,
the logpipe getting closed, etc.
 If my bet is correct, then the bug is that there's a case where runsv can
close the logpipe and still keep going, whereas it should exit as soon as
the logger dies no matter what (or just exit on the spot and let the logger
die on its own).

-- 
 Laurent