19 Apr
2010
19 Apr
'10
8:38 a.m.
Iñaki Baz Castillo wrote:
2010/4/19 marius zbihlei
<marius.zbihlei(a)1and1.ro>ro>:
Hello Iñaki ,
Hello Iñaki ,
I had a look over the patches and they look fine. Of course I think one of
the core developers should have a look also.
I suggest one thing: Instead of a read() from the read end of the pipe, can
we use a select()/poll() so we can have timeouts and prevent blocking. For
example does it make sense to say that if the child process doesn't write
something to the pipe in let's say 1 minute, this means that it is blocked
somewhere and the main process should exit with error (thus the init.d
script should return != 0) ?!
Hi, in the proposed code if the child process (main
process) exits due
to an error then it writes nothing to the pipe and the parent process
reads 0 bytes from it. It means that an error has occurred and it
exits with -1. In case main process starts properly it writes
something to the pipe ("go") so the master process reads 2 bytes (>0)
and exits with 0.
Indeed, main process is the process after the fork, and this is the
process that writes to signal the parent. I see two possible pitfalls:
1. If the main process blocks, this will block the parent process also
2. If the main process returns without writing the bytes, and there are
still
child processes left(tcp or udp worker children etc), then they will
still have
the writing part of the socket open (forked from the main process)and
again the parent (master) process will
keep blocking (didn't discovered a case where it might happen).
In the case you suggest, if the main process gets
blocked for some
reason (it doesn't exit but neither writes into the pipe) then as you
say the parent process would get blocked. Not good. Is it possible to
do a blocking read of the pipe with a timeout of 4-5 seconds? or is
the select()/poll() stuff required for it?
With a select it is possible to do a blocking read from some time. I
strongly suggest more
than 4-5 seconds, I think 30s should be a minimum.
Anyhow, I wonder if it would be enough. Note that in
case the main
process gets blocked and the parent process exits with -1 (due to the
suggested timeout) the main process still remains running (even if
blocked). Perhaps the parent process should kill it and ensure it's
dead in case such timeout occurs?
Thanks a lot.
Good question.. We can kill all children from the main proces, but I am
not sure that from the masetr process we can do this..
Marius