[openib-general] Re: [PATCH] (fixed) reduce qp locking on cq poll

[Michael S. Tsirkin]

Quoting r. Roland Dreier <roland at topspin.com>:
> Subject: Re: [PATCH] (fixed) reduce qp locking on cq poll
> 
>     Michael> Locking during the poll cq operation can be reduced, by
>     Michael> locking the cq while qp is being removed from the qp
>     Michael> array.  This also avoids an extra atomic operation for
>     Michael> reference counting.
> 
> Look reasonable.  Can you resend with /* */ comments instead of //?

OK.

>     Michael> TODO implies the same can be done for cqs, but I dont see
>     Michael> how, since we dont take a lock on eq, so this I have left
>     Michael> in the TODO.  This works but I dont have the fast
>     Michael> hardware to check the speed effect.
> 
> We would have to add a lock to the EQ to do the same for CQs.  However
> this would save both the CQ table lock and the atomic reference
> counting of the CQ so it should be a net win.

My RCU option does it even better - without the EQ lock :).

But I agree it would be nice for when RCU is disabled
(e.g. for people that want closing cqs to be as fast as possible),
or for if it does not go in.

> I'm a little worried that you see no improvement from this patch -- it
> doesn't seem like fast hardware should be required.
> 
>  - R.

On my hardware I see wild variations of bandwidth so only big wins
show conclusively. Can you test it?

mst

-- 
I dont speak for Mellanox