> static void
> udp_write_space(struct sock *sk)
> {
<snip>
> /* Wait until we have enough socket memory. */
> if (sock_writeable(sk))
> return;
You misunderstand the code. The above is in the write_space() callback. It
tells you when it is safe to wakes up again *after* a send has failed. It
doesn't test the buffer size on the first sendmsg() call (the one that
fails).
> The thing, which is really useless, is that your patch preparing skbs
> and dropping them in the next line. With the same success you could
> trigger BUG() there. :-) Right application just should not reach
> this condition.
Are you seriously saying that all 'right' user applications should be testing
the socket buffer congestion before sending a non-blocking UDP message rather
than just testing sendmsg() for an EWOULDBLOCK return value???
According to the manpage, ioctl(SIOCOUTQ) didn't even work prior to 2.4.x...
The normal behaviour of the patch was to collect the fragments, then to send
off all the skbs to the device queue as soon as it is clear that no errors
have occurred.
The patch only dropped the skbs if some socket error occurs that would force
you to exit the loop and return EAGAIN. Since the loop will be exited before
the fragment containing the UDP header (offset 0) gets sent, sending off all
the other fragments in the other skbs would serve merely to eat up bandwidth.
I agree that for blocking calls, it is useful to send off each skb as it gets
allocated (and the patch could be amended to take this into account), but for
nonblocking I/O, it is definitely bad form...
Cheers,
Trond
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/