Hello! > Andi Kleen writes: > > It violates the definition of write() though if the TCP socket is > > accessed via write(2), because writes are guaranteed to be atomic. > > The later could be fixed by aquiring the inode semaphore in > > sock_write, but I'm not sure if it is worth the cycles. BTW write() does not use inode semaphore in 2.4 for sockets, so that write() has no differences. Actually, behaviour of 2.2 is even bug: atomic write() (and uninterruptible(!) semaphore used in 2.2) is a non-sense for descriptors which can block for infinite time. On such descriptors atomicity is defined as minimal amount of data written atomically (SO_SNDLOWAT, though even this is not defined by a standard). Our atomicity threshold for TCP (and default in OSes implementing variable threshold) is 1 byte. Alexey - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org
