Re: [PATCH] new UDPCP Communication Protocol

[Eric Dumazet <eric.dumazet@gmail.com>]

Le samedi 01 janvier 2011 à 22:44 +0100, stefani@seibold.net a écrit :
> From: Stefani Seibold <stefani@seibold.net>
> 
> Changelog:
> 31.12.2010 first proposal
> 01.01.2011 code cleanup and fixes suggest by Eric Dumazet
> 
> UDPCP is a communication protocol specified by the Open Base Station
> Architecture Initiative Special Interest Group (OBSAI SIG). The
> protocol is based on UDP and is designed to meet the needs of "Mobile
> Communcation Base Station" internal communications. It is widely used by
> the major networks infrastructure supplier.
> 
> The UDPCP communication service supports the following features:
> 
> -Connectionless communication for serial mode data transfer
> -Acknowledged and unacknowledged transfer modes
> -Retransmissions Algorithm
> -Checksum Algorithm using Adler32
> -Fragmentation of long messages (disassembly/reassembly) to match to the MTU
>  during transport:
> -Broadcasting and multicasting messages to multiple peers in unacknowledged
>  transfer mode
> 
> UDPCP supports application level messages up to 64 KBytes (limited by 16-bit
> packet data length field). Messages that are longer than the MTU will be
> fragmented to the MTU.
> 
> UDPCP provides a reliable transport service that will perform message
> retransmissions in case transport failures occur.
> 
> The code is also a nice example how to implement a UDP based protocol as
> a kernel socket modules.
> 
> Due the nature of UDPCP which has no sliding windows support, the latency has a
> huge impact. The perfomance increase by implementing as a kernel module is
> about the factor 10, because there are no context switches and data packets or
> ACKs will be handled in the interrupt service.
> 
> There are no side effects to the network subsystems so i ask for merge it
> into linux-next. Hope you like it.
> 
> The patch is against 2.6.37-rc8
> 

Hi Stefani

1) Please base your next trys on net-next-2.6 : This is the reference
for stuff like that. It probably does not matter a lot, but still...


2) I still see some _irq() variants of spinlock(). Its not necessary in
network stack at the level you are working (process context, and
softirqs)

Please only use _bh() variants, it's enough.

3) I see UDPLITE references in your code. Are you sure UDPCP protocol
can really work on top of UDPLITE ? I think not, so please remove dead
code.

4) udpcp_release_sock() seems expensive to me. Why not testing
usk->timeout before releasing sock lock, and save a lock/unlock pair ?

static inline void udpcp_release_sock(struct sock *sk)
{
	struct udpcp_sock *usk = udpcp_sk(sk);

	if (usk->timeout)
		udpcp_handle_timeout(sk);
	release_sock(sk);
}

5) In udpcp_timeout(), if you find socket is locked by user, you set
timeout to one and rearm a timer to udpcp_timer(sk, jiffies + 1); 

Why is it needed, since user process is going to handle the timeout
indication from udpcp_release_sock() ?

Thanks