Re: [PATCH] memory: avoid unnecessary iteration when updating ioeventfds

[Jason Wang]

On Mon, Mar 6, 2023 at 5:27 AM Peter Xu <peterx@redhat.com> wrote:
>
> On Wed, Mar 01, 2023 at 04:36:20PM +0800, Jason Wang wrote:
> > On Tue, Feb 28, 2023 at 10:25 PM Longpeng(Mike) <longpeng2@huawei.com> 
> > wrote:
> > >
> > > From: Longpeng <longpeng2@huawei.com>
> > >
> > > When updating ioeventfds, we need to iterate all address spaces and
> > > iterate all flat ranges of each address space. There is so much
> > > redundant process that a FlatView would be iterated for so many times
> > > during one commit (memory_region_transaction_commit).
> > >
> > > We can mark a FlatView as UPDATED and then skip it in the next iteration
> > > and clear the UPDATED flag at the end of the commit. The overhead can
> > > be significantly reduced.
> > >
> > > For example, a VM with 16 vdpa net devices and each one has 65 vectors,
> > > can reduce the time spent on memory_region_transaction_commit by 95%.
> > >
> > > Signed-off-by: Longpeng <longpeng2@huawei.com>
> > > ---
> > >  include/exec/memory.h |  2 ++
> > >  softmmu/memory.c      | 28 +++++++++++++++++++++++++++-
> > >  2 files changed, 29 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/include/exec/memory.h b/include/exec/memory.h
> > > index 2e602a2fad..974eabf765 100644
> > > --- a/include/exec/memory.h
> > > +++ b/include/exec/memory.h
> > > @@ -1093,6 +1093,8 @@ struct FlatView {
> > >      unsigned nr_allocated;
> > >      struct AddressSpaceDispatch *dispatch;
> > >      MemoryRegion *root;
> > > +#define FLATVIEW_FLAG_IOEVENTFD_UPDATED (1 << 0)
> > > +    unsigned flags;
> > >  };
> > >
> > >  static inline FlatView *address_space_to_flatview(AddressSpace *as)
> > > diff --git a/softmmu/memory.c b/softmmu/memory.c
> > > index 9d64efca26..71ff996712 100644
> > > --- a/softmmu/memory.c
> > > +++ b/softmmu/memory.c
> > > @@ -815,6 +815,15 @@ FlatView *address_space_get_flatview(AddressSpace 
> > > *as)
> > >      return view;
> > >  }
> > >
> > > +static void address_space_reset_view_flags(AddressSpace *as, unsigned 
> > > mask)
> > > +{
> > > +    FlatView *view = address_space_get_flatview(as);
> > > +
> > > +    if (view->flags & mask) {
> > > +        view->flags &= ~mask;
> > > +    }
> > > +}
> > > +
> > >  static void address_space_update_ioeventfds(AddressSpace *as)
> > >  {
> > >      FlatView *view;
> > > @@ -825,6 +834,12 @@ static void 
> > > address_space_update_ioeventfds(AddressSpace *as)
> > >      AddrRange tmp;
> > >      unsigned i;
> > >
> > > +    view = address_space_get_flatview(as);
> > > +    if (view->flags & FLATVIEW_FLAG_IOEVENTFD_UPDATED) {
> > > +        return;
> > > +    }
> > > +    view->flags |= FLATVIEW_FLAG_IOEVENTFD_UPDATED;
> > > +
> >
> > Won't we lose the listener calls if multiple address spaces have the
> > same flatview?
>
> I have the same concern with Jason.  I don't think it matters in reality,
> since only address_space_io uses it so far. but it doesn't really look
> reasonable and clean.

Yes, I think in the memory core it should not assume how the eventfds are used.

>
> One other idea of optimizing ioeventfd update is we can add a per-AS
> counter (ioeventfd_notifiers), increase if any eventfd_add|del is
> registered in memory_listener_register(), and decrease when unregister.
> Then address_space_update_ioeventfds() can be skipped completely if
> ioeventfd_notifiers==0.
>
> Side note: Jason, do you think we should drop vhost_eventfd_add|del?
> They're all no-ops right now.

I think so.

Thanks

>
> Thanks,
>
> --
> Peter Xu
>