Re: [dpdk-dev] [PATCH v3 2/3] net/virtio: update memory ordering comment for vq notify

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, Jan 10, 2019 at 12:55:19PM +0300, Ilya Maximets wrote:
> On 10.01.2019 12:18, Maxime Coquelin wrote:
> > Hi Gavin,
> > 
> > On 1/10/19 9:19 AM, Gavin Hu (Arm Technology China) wrote:
> >>
> >>
> >>> -----Original Message-----
> >>> From: dev <dev-bounces@dpdk.org> On Behalf Of Ilya Maximets
> >>> Sent: Wednesday, January 9, 2019 10:50 PM
> >>> To: dev@dpdk.org; Maxime Coquelin <maxime.coquelin@redhat.com>;
> >>> Michael S . Tsirkin <mst@redhat.com>; Xiao Wang
> >>> <xiao.w.wang@intel.com>
> >>> Cc: Tiwei Bie <tiwei.bie@intel.com>; Zhihong Wang
> >>> <zhihong.wang@intel.com>; jfreimann@redhat.com; Jason Wang
> >>> <jasowang@redhat.com>; xiaolong.ye@intel.com;
> >>> alejandro.lucero@netronome.com; Ilya Maximets
> >>> <i.maximets@samsung.com>
> >>> Subject: [dpdk-dev] [PATCH v3 2/3] net/virtio: update memory ordering
> >>> comment for vq notify
> >>>
> >>> We're not using io ports in case of modern device even on IA.
> >>> Also, this comment useless for other architectures.
> >>
> >> Agree, it should be architecture neutral.
> >>
> >>>
> >>> Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
> >>> ---
> >>>   drivers/net/virtio/virtqueue.h | 9 ++++-----
> >>>   1 file changed, 4 insertions(+), 5 deletions(-)
> >>>
> >>> diff --git a/drivers/net/virtio/virtqueue.h b/drivers/net/virtio/virtqueue.h
> >>> index dffa03669..53aeac238 100644
> >>> --- a/drivers/net/virtio/virtqueue.h
> >>> +++ b/drivers/net/virtio/virtqueue.h
> >>> @@ -437,14 +437,13 @@ virtqueue_kick_prepare_packed(struct virtqueue
> >>> *vq)
> >>>       return flags != RING_EVENT_FLAGS_DISABLE;
> >>>   }
> >>>
> >>> +/*
> >>> + * virtqueue_kick_prepare*() or the virtio_wmb() should be called
> >>> + * before this function to be sure that all the data is visible to vhost.
> >>> + */
> >>
> >> C11 _atomic APIs are preferred for new code, other than wmb or rmb,  could you work on that?
> > 
> > Thanks for the review.
> > -rc2 deadline is today, so I may apply this series as is if no reply
> > from Ilya today. If a change is agreed, it could be made on top for
> > -rc3.
> 
> IMHO, If we'll decide to move to C11 atomics, we'll need to rewrite all
> the significant memory accesses in virtio driver at once. We can't do
> this partially. This will require significant amount of work to understand
> how to do that and will require a lot of testing. So, it's definitely not
> for current release. Also, there are possible performance concerns about
> such solution.
> 
> But I have even more significant concern: C11 atomics are designed for
> inter-thread synchronization on multi-core systems. But in case of vDPA we
> have real hardware and I'm not sure if we can use C11 atomics for
> cross-domain synchronizations. Do you know if some of the memory ordering
> types in C11 provide outer domain sync on ARMv8, for example ?

I would add to that - compiler support might not work well in all
versions which can be used to build dpdk.
E.g. this article https://lwn.net/Articles/691128/
says that "there will be some seriously suboptimal code production
before gcc-7.1".

HTH

> > 
> > Regards,
> > Maxime
> >>>   static inline void
> >>>   virtqueue_notify(struct virtqueue *vq)
> >>>   {
> >>> -    /*
> >>> -     * Ensure updated avail->idx is visible to host.
> >>> -     * For virtio on IA, the notificaiton is through io port operation
> >>> -     * which is a serialization instruction itself.
> >>> -     */
> >>>       VTPCI_OPS(vq->hw)->notify_queue(vq->hw, vq);
> >>>   }
> >>>
> >>> -- 
> >>> 2.17.1
> >>
> > 
> >