Re: Multithread select() bug
From: Eric Dumazet
Date: Tue May 11 2004 - 01:38:01 EST
Andre Ben Hamou wrote:
As the safety of what you are trying to do cannot be guaranted, you
should not even try to do that.
I see where you're coming from, in that there is a potential race
condition as to the socket being connected as I reach the select call.
This is an important concern but it is, I think, orthogonal to the
original problem as there are two possible socket states at the point
at which select gets called (as far as I can see)...
1. The socket is in its connected state
2. The socket has already been closed by the parent thread
As I understand it, if 1 is true (which corresponds to my original
post), then select should return the moment the socket gets closed
and, if 2 is true (which I believe corresponds to your concern), then
select should return immediately anyway as the socket would not block
if read from.
Sorry to be a pest, but I'm trying to get this clear in my head. Is it
possible I've over-estimated the thread-safety of the select and close
Then, apart from this safety showstopper, you should consider that what
you assumed is simply wrong :
- Some Operating Systems do something special to signal blocked threads
(blocked on a read()/write()/select()/poll()) if the underlying file
handle is closed.
- linux does NOT. And this is done for several reasons. These reasons
are known and were discussed on this list at several occasions.
So the rule is : close() syscall should be serialized, according to a
synchronization of your choice, so that no other thread could be using
the file descriptor at the same time.
If you want to 'send some signal to the blocked thread' in a portable
way, you could setup a pipe to be able to send a special message on it.
The 'worker thread' could add the read side of the pipe (pipefd) to
its select fd set, and will be notified of the Thread 2) willing to
dismantle your socket, using a standard write(pipefd). This will be
portable : When the 'worker thread' returns from select() call, it can
then close the socket fd himself.
By the way, prefer poll() syscall to select() one. Poll is much more
scalable/efficient if you have only one or two fd to monitor.
Andre Ben Hamou
Imperial College London
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/