Re: more tcpdumpinfo for nfs3 problem: aix-server --- linux 2.4.15pre5 client

[Trond Myklebust <trond.myklebust@fys.uio.no>]

On Tuesday 20. November 2001 20:45, kuznet@ms2.inr.ac.ru wrote:

> (Call QDIO bottom half code)
> spin_lock(&QDIO_lock);
>                                                  <QDIO hard interrupt>
>                                                       
>                                                  ->spin_lock(&QDIO_lock)
>                                                  (spins...)
>
> with the same result. No help of nfs is required. :-)

Now that my blood caffeine levels are above the 'sleep' watermark, I should 
probably correct the above in case Ulrich decides to slap us both with a 
slander lawsuit 8-)...

Since (as I said in the original mail) CPU 1 is in a bottom half context, 
both local bh and local interrupts should be disabled on that process!

Basically, the deadlock between the QDIO driver and RPC/fasync can be reduced 
to:

CPU 1                                       CPU 2

(In BH context)                             (In ordinary process context)
spin_lock_irq(&lock1);
                                            spin_lock_bh(&lock2);
dev_kfree_skb_any(skb);
                                            <QDIO Hard interrupt>
                                            (switch to QDIO interrupt context)
    -> sk->write_space();
                                            spin_lock(&lock1);
         -> spin_lock(&lock2);


IOW:
    Either we must demand that CPU 2 uses irq-safe spinlocks in order to 
protect against sk->write_space(), or we must demand that CPU 1 should drop 
'lock1' before being allowed to call dev_kfree_skb_any().

Cheers,
   Trond