Re: System crash in tcp_fragment()

[Nivedita Singhvi <niv@us.ibm.com>]

On Mon, 20 May 2002, David S. Miller wrote:

> Such rule does not even make this piece of code legal.  Consider:
> 
> task1:cpu0:	x = counters[smp_processor_id()];
>       cpu0:	PREEMPT
> task2:cpu0:	x = counters[smp_processor_id()];
> task2:cpu0:	counters[smp_processor_id()] = x + 1;
>       cpu0:	PREEMPT
> task1:cpu0:	counters[smp_processor_id()] = x + 1;
> 		full garbage
> 
> But it does bring up important point, preemption people need to
> fully audit entire networking.
> 
> It is totally broken by preemption the more I think about it.
> 
> At the very beginning, all the SNMP counter bumping tricks will
> totally fail with preemption enabled.
> 

A lot of the synchronization between process context and interrupt
context is based on per-cpu data structures or simple locks
(without disabling irq's globally) eg:

softnet_data queue (we only disable local interrupts), and
synchronization between tcp_readmsg() and tcp_rcv() over
the receive queue would get confused (lock.users flag would
be different on another CPU)..

Wonder how any of it could possibly work..

thanks,
Nivedita