Re: [virtio-dev] Device-to-driver notification management for hardware implementations

[David Riddoch <driddoch@solarflare.com>]

On 05/11/18 08:53, Stefan Hajnoczi wrote:
> On Wed, Sep 12, 2018 at 05:41:07PM +0100, David Riddoch wrote:
> > The VIRTIO_F_NOTIFICATION_DATA feature helps hardware implementations by
> > telling them directly how many descriptors are available in the ring, saving
> > speculative reads.  There is a related issue with interrupts: As things
> > stand, after writing 'used' descriptors into the ring (packed or split) the
> > device has to read the event idx to determine whether a device-to-driver
> > notification (interrupt) is needed.
> >
> > It would be very helpful to a hardware implementation if the event idx could
> > be written directly to the device.  Has anyone proposed an extension to
> > address this already?  (I have some thoughts on how to do this if we don't
> > have a proposal yet).
> Hi David,
> Has there been any further discussion on this topic?  It seems like a
> useful feature to have in VIRTIO.
Hi Stefan,

Thanks for the nudge, no further discussion.  So here are my thoughts:

To provide context here is an outline of how drivers typically process 
completions from a device.  This applies to both para-virtualised and 
real hardware.  (I think we only need to worry about packed-ring).

   disable_cb();
 again:
   while( budget_left && desc_is_used() )
     handle_used();
   if( budget_left &&enable_cb() )
     goto again;

The budget stuff relates to napi polling, and isn't always there. But if 
it is, and budget is exhausted, then an outer loop will invoke this code 
again and so we don't want an interrupt.

In our current packed-ring design enable_cb() does roughly this:

   enable_cb_packed() {
     guest_mem->event_suppress = last_used_idx;
     virtio_mb();
     return desc_is_used();
   }

This tells the device to raise an interrupt when moving used_idx over 
the given idx.  (And you can set that idx to a position further in the 
ring if you're not interested in getting woken until a bunch of 
descriptors have been consumed).

This design is unpleasant for hardware devices, because it forces them 
to issue a read over the PCIe bus after writing used descriptors, in 
order to determine whether an interrupt is needed. Such a read adds 
complexity, adds overhead to the PCIe bus and memory system, and adds 
latency.

In hardware devices interrupts are typically enabled like this:

   enable_cb_hw() {
     writel(last_used_idx, &dev_bar->int_en_doorbell);
     return 0;
   }

This tells the device to raise an interrupt once the descriptor at the 
given index has been used.  If it has already been used, then an 
interrupt is raised immediately, else the interrupt is enabled and 
raised after the used descriptor has been written.  The race condition 
is handled without having to check the ring again.  Note that interrupts 
are fire-once: For each write to int_en_doorbell you get a single 
interrupt, and disable_cb() is therefore a no-op.

When you emulate a device using this model, you get a vmexit every time 
int_en_doorbell is written, which is expensive.  Therefore this approach 
is not desirable when emulated.

So perhaps we should introduce an option to use this hardware-oriented 
model.  The problem is that with vDPA the implementation of a device can 
shift under the feet of a driver between hardware and emulation.  We'd 
like to get the best performance we can in both cases.

My proposal is that we add an option to use both models at the same 
time.  enable_cb() would become:

30   enable_cb_new() {
31     if( int_en_doorbell_enabled )
32 writel(last_used_idx, p_int_en_doorbell);
33 guest_mem->event_suppress = last_used_idx;
34     virtio_mb();
35     return desc_is_used();
36   }

p_int_en_doorbell it notionally a pointer into a memory mapping onto a 
device BAR, and when backed by hardware that is exactly what it is (and 
the information in the event suppression structure is ignored).  But 
when backed by an emulated device the p_int_en_doorbell mapping should 
just point at normal host memory (which is otherwise unused).  This 
means that we avoid vmexits in the emulated case.

Comments?

David

--
David Riddoch  <driddoch@solarflare.com> -- Chief Architect, Solarflare