Re: [RFC] IB/mad: Use IDR instead of per-port list

From: Jason Gunthorpe
Date: Thu Jun 07 2018 - 22:11:35 EST


On Thu, Jun 07, 2018 at 04:49:26PM -0700, Matthew Wilcox wrote:
> On Thu, Jun 07, 2018 at 04:26:46PM -0600, Jason Gunthorpe wrote:
> > On Thu, Jun 07, 2018 at 12:08:32PM -0700, Matthew Wilcox wrote:
> > >
> > > Hans reports a bug where the mlx4 driver uses the MSB of the agent number
> > > to store slave number, meaning we can only use agent numbers up to 2^24.
> > > Fix this by using an IDR to assign the agent numbers and also look up the
> > > agent when a MSD packet is received.
> > >
> > > I've changed the locking substantially, so this may well have a
> > > performance issue. There are a few different possibilities for fixing
> > > that, including moving to an RCU-based lookup.
> >
> > I do like this better than the last series..
> >
> > This are is somewhat performance sensitive and it would be nice to
> > avoid this global lock.
>
> OK, I wasn't sure whether it was worth it.
>
> > What about using a read/write spinlock instead of the IDR internal
> > lock? Then all the per-port reading threads calling find_mad_agent can
> > run concurrently..
>
> It'd be better to switch to RCU ... the IDR is RCU-safe, but the
> version/class/method or OUI match isn't. Do you have any feeling on
> the relative frequency of the two types of "routing"?

I would say they are close to equally important, perhaps the
version/class/method is slightly more important.

> Actually, I think we can use the radix tree data structure for the
> version/class/method too ... that's going to take a little more work.

I was wondering about that as well, this code is very old so there
must be better data structure helpers these days.

Jason