On Thu, Jul 30, 2026 at 10:40:32AM -0700, Frank van der Linden wrote:
> On Thu, Jul 30, 2026 at 10:30 AM Gregory Price <[email protected]> wrote:
> >
> > On Thu, Jul 30, 2026 at 05:02:56PM +0200, Yiannis Nikolakopoulos wrote:
> > > In most memory tiering scenarios, the memory to be demoted is expected
> > > to be cold and most likely out of the node's last-level cache (as well
> > > as target pages in the target node). Using non-temporal stores instead
> > > of a standard memcpy path can reduce the cache pollution in the local
> > > node and the bandwidth overhead to the target node. Furthermore, for
> > > certain types of CXL devices that support in-line memory compression,
> > > the last-level cache eviction patterns can negatively affect the
> > > bandwidth of the device. Non-temporal stores can mitigate this.
> > >
> > > This patch-set introduces a new migrate_mode flag for using non-temporal
> > > stores that is used only in the demotion path. Patch 1 adds some helpers 
> > > in
> > > x86 and mm to bring non-temporal stores support to a respective folio_copy
> > > function. Patch 2 adds the new flag and necessary changes for 
> > > compatibility
> > > with the existing behavior. Patch 3 uses the new flag for demotions.
> > >
> > > Experimental data: in a CXL system with 1 memory expander, a 
> > > microbenchmark
> > > that allocates N=64 GB memory in the local node and then triggers demotion
> > > using memory.reclaim, shows a practically complete elimination of read
> > > traffic on the device, i.e. write traffic is N GB with and without the
> > > patch, while read traffic drops from N to almost 0 with the patch.
> > >
> > > Opens:
> > > 1. There is some "duplication" in the x86 tree and a bit in mm. Can we do
> > >    something better there? As it is now in copy_mc_to_kernel_nt we
> > > duplicate the machine check functionality, which if available will 
> > > override
> > > the non-temporal. We were not sure how to prioritize these two and what's
> > > the best approach here. Can we completely skip the machine checked for 
> > > this
> > > path? Huan Nguyen has some ideas here that we will align for the next
> > > version.
> > > 2. I am not sure how this should be structured so that it is easily
> > > adopted in other architecture trees (e.g. aarch64). We rely on
> > > memcpy_flushcache for x86_64 but this does not use non temporal stores
> > > in ARM. ARM support is currently out of our scope but any input is
> > > appreciated.
> > >
> >
> > I'm still a bit confused why using NT Stores needs to be an explicit
> > option - rather than the default behavior if NT Store is available.
> >
> > Lets assume we did this for all ASYNC requests, is there a negative
> > effect? A positive effect? Why the new ASYNC type? Is there a
> > correctness issue?
> >
> > ~Gregory
> 
> Yes, this is a good discussion to have. I believe that the initial
> reason for using a separate mode here was to be non-invasive, e.g.
> "don't break anything else".
> 
> But why not use non-temporal stores for everything? I don't know. I
> suppose one argument might be that if you know that the destination
> will be used immediately, NT stores might be a slight performance hit.

do_numa_page() -> migrate_misplaced_folio() is such a case. It happens
literally in the access path to that data. MIGRATE_ASYNC just means
don't block, fail fast, because waiting for an IO-bound lock to avoid
a remote NUMA access would be a bad idea.

Compaction on the other hand does use MIGRATE_SYNC* and it's touching
data in PFN order, completely out of execution sequence. That could be
a good candidate for NT stores.

> arm64 already seems to use stores with NT hints by default.
> 
> In general, though, I agree that for ASYNC requests, just always using
> NT seems fine.
> 
> Maybe the mode and reason should be folded in to one variable, so
> that, further down the stack, a decision can be made as to what type
> of copy to use? E.g. if the mode is !MIGRATE_ASYNC and the reason is
> MR_DEMOTION, then non-temporal is still a good choice. If the mode is
> MIGRATE_ASYNC, NT is still a good idea. If you wanted to get fancy,
> any mode with folio_test_waiters(folio) == true should not use NT,
> since a task is waiting to use the data, so caching it is better.
> Maybe that's overthinking it.

Agree. `reason' seems like a much stronger signal.

Reply via email to