> > > This report describes a problem in the usage of file
> > > descriptors across
> >> multiple threads. When one thread closes a file descriptor, another
> >> thread which waits for an I/O on that file descriptor is not notified
> >> and blocks forever.
> >THe I/O does not block forever, it blocks until completed.
> That's still "forever" if you don't specify a timeout in the select.
If you don't want to block until an operation completes, then don't ask to!
> >The actual final
> >closure of the object occurs when the last operation on it exits
> Select is defined as to return, with the appropriate bit set, if/when
> a nonblocking read/write on the file descriptor won't block. You'd
> get EBADF in this case, therefore causing the select to return would
> be a Good Thing.
That is not quite correct. That is a good approximate definition of
'select's behavior, but it is not exact. As for your assertion that a
noblocking read/write wouldn't block, that's not necessarily true. Remember,
the 'close' may not have taken affect yet, since the descriptor is still in
use by virtue of being selected on.
> A related problem is that the second thread my be inside a blocking
> read() instead of a select() call. It'd never continue. :-(
Perfect. Doing this is absolutely, positively wrong and the more you are
punished for it, the better. It's as wrong as calling 'free' on a chunk of
memory when another thread may be usign it. It is impossible to make this
work safely, as another thread could open a socket or file and get the same
descriptor write before the call to 'read' is entered. There's no possible
way to do this because there is no 'unlock a mutex and read' operation.
> HOWEVER: IMHO it's bad design to distribute the responsibility for
> file descriptors between threads.
Why? That's a great design and it's absolutely essential in many cases.
Suppose, for example, I have two descriptors I want to write to. If I assign
one thread to each socket permanently, then I'm 100% guaranteed a context
switch every time I change which socket I'm writing to, so if there's lots
of small bits of data going out both of them, my performance will suck. But
if I assign one thread to both socket descriptors, I'm guaranteed that one
connection will stall if the the application-level send queue for the other
has been swapped out to disk. Not distributing network I/O across threads
dynamically is a recipe for either low performance or bursty performance.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/