Re: [Xen-devel] [PATCH v2] x86/domain_page: implement pure per-vCPU mapping infrastructure

["Xia, Hongyan" <hongyxia@xxxxxxxxxx>]

On Fri, 2020-02-21 at 14:31 +0100, Jan Beulich wrote:
> On 21.02.2020 13:52, Xia, Hongyan wrote:
> > On Fri, 2020-02-21 at 11:50 +0000, Wei Liu wrote:
> > > On Thu, Feb 06, 2020 at 06:58:23PM +0000, Hongyan Xia wrote:
> > > > +    if ( hashmfn != mfn && !vcache->refcnt[idx] )
> > > > +        __clear_bit(idx, vcache->inuse);
> > > 
> > > Also, please flush the linear address here and the other
> > > __clear_bit
> > > location.
> > 
> > I flush when a new entry is taking a slot. Yeah, it's probably
> > better
> > to flush earlier whenever a slot is no longer in use.
> 
> Question is whether such individual flushes aren't actually
> more overhead than a single flush covering all previously
> torn down entries, done at suitable points (see the present
> implementation).

There is certainly room for improvement. I am considering flushing
entries in batches to reduce the overhead, e.g., in a similar way to
the current implementation as you said.

I want to defer that to a separate patch because this is already a huge
patch. From the benchmarks I have done so far, it does not look like
this has any noticeable overhead and it already alleviates the lock
contention, plus this is currently used only in a debug build, so I
would like to defer the optimisation a bit.

Hongyan
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel