Re: [MAINLINE 0/2] Enable DIM for legacy ULPs and use it in RDS

From: Allison Henderson
Date: Tue Sep 24 2024 - 11:21:01 EST


On Tue, 2024-09-24 at 08:59 +0000, Haakon Bugge wrote:
>
>
> > On 24 Sep 2024, at 09:01, Christoph Hellwig <hch@xxxxxxxxxxxxx>
> > wrote:
> >
> > On Fri, Sep 20, 2024 at 06:51:18AM -0700, Christoph Hellwig wrote:
> > > On Fri, Sep 20, 2024 at 09:46:06AM +0000, Haakon Bugge wrote:
> > > > > I would much prefer if you could move RDS off that horrible
> > > > > API finally
> > > > > instead of investing more effort into it and making it more
> > > > > complicated.
> > > >
> > > > ib_alloc_cq() and family does not support arming the CQ with
> > > > the IB_CQ_SOLICITED flag, which RDS uses.
> > >
> > > Then work on supporting it.  RDS and SMC are the only users, so
> > > one
> > > of the maintainers needs to drive it.
> >
> > I took a quick look at what it would take, and adding
> > IB_CQ_SOLICITED
> > support to the cq API looks pretty trivial, you'll just need to
> > pass
> > it to ib_cq_pool_get by adding a new argument and the pass it
> > down to __ib_alloc_cq.  So yes, please get started at moving RDS
> > out
> > of the stone age.
>
> Agree. I'll work on that. I am also inclined to improve it by having
> designated CPUs where the worker threads polling the CQs execute, as
> that often improves cache hit rates. But that will come as another
> commit.
>
>
> Thxs, Håkon
>

Ok, sounds like a good plan then. Thanks for working on this Haakon!

Allison