[RFC CFT 0/6] Try to reduce lock contention on the SMMUv3 command queue

[Will Deacon]

Hi John,

On Mon, Jun 17, 2019 at 02:38:59PM +0100, John Garry wrote:
> On 11/06/2019 14:45, Will Deacon wrote:
> > Hi all,
> > 
> > This patch series is an attempt to reduce lock contention when inserting
> > commands into the Arm SMMUv3 command queue. Unfortunately, our initial
> > benchmarking has shown mixed results across the board and the changes in
> > the last patch don't appear to justify their complexity. Based on that,
> > I only plan to queue the first patch for the time being.
> > 
> > Anyway, before I park this series, I thought it was probably worth
> > sharing it in case it's useful to somebody. If you have a system where
> > you believe I/O performance to be limited by the SMMUv3 command queue
> > then please try these patches and let me know what happens, even if it's
> > just more bad news.
> > 
> > Patches based on 5.2-rc3. I've also pushed them out to my iommu/devel
> > branch for the moment:
> > 
> >   https://git.kernel.org/pub/scm/linux/kernel/git/will/linux.git/log/?h=iommu/devel
> > 
> 
> For command queue lock contention, we had this series previously:
> https://lore.kernel.org/linux-iommu/61b4c3e5f1322dfe96ca2062a7fe058298340996.1539782799.git.robin.murphy@arm.com/#t
> 
> I am just wondering does this have any future?

The functionality of that series is subsumed by the patches I've posted
here, although if I can't get the cmpxchg() loop working well here then
we could revisit just making the change proposed by Robin. The problem is
that we'll still have serialisation on access to the command queue, and
therefore it will remain a scalability bottleneck as long as the fast-path
needs to queue for a lock.

However, I've still got a few tricks up my sleeve so I'm hoping to get a
new version of this lot out in the coming weeks.

Will