[RFC] virtio: use mandatory barriers for remote processor vdevs

[Michael S. Tsirkin]

On Thu, Dec 01, 2011 at 01:27:10AM +0200, Ohad Ben-Cohen wrote:
> On Wed, Nov 30, 2011 at 6:24 PM, Ohad Ben-Cohen <ohad at wizery.com> wrote:
> > On Wed, Nov 30, 2011 at 6:15 PM, Michael S. Tsirkin <mst at redhat.com> wrote:
> >> How are the rings mapped? normal memory, right?
> >
> > No, device memory.
> 
> Ok, I have more info.
> 
> Originally remoteproc was mapping the rings using ioremap, and that
> meant ARM Device memory.
> 
> Recently, though, we moved to CMA (allocating memory for the rings via
> dma_alloc_coherent), and that isn't Device memory anymore: it's
> uncacheable Normal memory (on ARM v6+).

And these accesses need to be ordered with DSB? Or DMB?

> We still require mandatory barriers though: one very reproducible
> problem I personally face is that the avail index doesn't get updated
> before the kick.

Aha! The *kick* really is MMIO. So I think we do need a mandatory barrier
before the kick.  Maybe we need it for virtio-pci as well
(not on kvm, naturally :) Off-hand this seems to belong in the transport
layer but need to think about it.

> As a result, the remote processor misses a buffer
> that was just added (the kick wakes it up only to find that the avail
> index wasn't changed yet). In this case, it probably happens because
> the mailbox, used to kick the remote processor, is mapped as Device
> memory, and therefore the kick can be reordered before the updates to
> the ring  can be observed.
> 
> I did get two additional reports about reordering issues, on different
> setups than my own, and which I can't personally reproduce: the one
> I've described earlier (avail index gets updated before the avail
> array) and one in the receive path (reading a used buffer which we
> already read). I couldn't personally verify those, but both issues
> were reported to be gone when mandatory barriers were used.

Hmm. So it's a hint that something is wrong with memory
but not what's wrong exactly.

> I expect those reports only to increase: the diversity of platforms
> that are now looking into adopting virtio for this kind of
> inter-process communication is quite huge, with several different
> architectures and even more hardware implementations on the way (not
> only ARM).
> 
> Thanks,
> Ohad.

Right. We need to be very careful with memory,
it's a tricky field. One known problem with virtio
is its insistance on using native endian-ness
for some fields. If power is used, we'll have to fix this.

-- 
MST