Re: [PATCH rdma-core 1/5] util: Add common mmio macros

[Jason Gunthorpe <jgunthorpe-ePGOBjL8dl3ta4EC/59zMFaTQe2KTcn/@public.gmane.org>]

On Tue, Apr 18, 2017 at 08:38:15PM +0300, Leon Romanovsky wrote:
> On Tue, Apr 18, 2017 at 11:28:52AM -0600, Jason Gunthorpe wrote:
> > On Tue, Apr 18, 2017 at 06:52:00PM +0300, Yishai Hadas wrote:
> >
> > We could use a hash scheme or something to multiple the lock, but I'm
> > really not sure that 32 bit performance matters to anyone anymore?
> 
> In a hard way, I learned that we have a customer who expects that his 32bit
> application will continue to work, so the answer is - no, I care.

Of course it continues to work.

Does your customer fit this very narrow definition:
 - Would actually upgrade to rdma-core
 - 32 bit
 - No SSE hardware (any Intel chip capable of PCI-E has SSE hardware)
 - Multiple same-provider devices with a single program touching all
   devices (single device performance is unchanged)
 - Sensitive to the performance difference of a potential spinlock
   contention / cache misplacement for ~4 instructions

It is hard to understand who cares so much about peformance but leaves
a wack on the table by running in 32 bit mode.

Is this something like x32? We can certainly improve for x32.

PPC32 could also potentially have a path like SSE..

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html