[openib-general] Re: [PATCH] ipoib: dont lock tx on completion

[Michael S. Tsirkin]

Quoting r. Roland Dreier <roland at topspin.com>:
> Subject: Re: [PATCH] ipoib: dont lock tx on completion
> 
>     Michael> By testing again, and waking the net queue if completions
>     Michael> arrived, we can avoid taking the tx lock on send
>     Michael> completion in ip over ib, reducing the lock contention.
> 
> I really like the clever way of retesting after putting the queue to
> sleep.  However I can't convince myself that some test or other will
> be wrong because of stale values of tx_head or tx_tail,

Not sure what is meant here by "stale values".
This is the usual consumer/producer approach ...

> and I'm always
> scared of code that requires multiple barriers and so on.  So I'm
> going to defer applying this until I have time to think hard and make
> sure there aren't any subtle races.
> 
>  - R.
> 

Okay.
Also note that the exact value of tx_tail/tx_head should not practially matter,
since to change the state you have to go from half-full TX queue to full
TX queue, or back.
There will be a huge number barriers, locks etc done by that time.

MST

-- 
MST - Michael S. Tsirkin