On Tue, Jan 15, 2019 at 08:55:50AM +0000, Shahaf Shuler wrote:
> Tuesday, January 15, 2019 10:29 AM, Ilya Maximets:
> > Subject: Re: [dpdk-dev] [PATCH v2] net/virtio: add platform memory
> > ordering feature support
> >
> > On 15.01.2019 9:33, Shahaf Shuler wrote:
> > > Thursday, January 10, 2019 10:37 PM, Shahaf Shuler:
> > >> Subject: RE: [dpdk-dev] [PATCH v2] net/virtio: add platform memory
> > >> ordering feature support
> > >>
> > >> Wednesday, January 9, 2019 5:50 PM, Michael S. Tsirkin:
> > >>> alejandro.luc...@netronome.com; Daniel Marcovitch
> > >>> <dani...@mellanox.com>
> > >>> Subject: Re: [dpdk-dev] [PATCH v2] net/virtio: add platform memory
> > >>> ordering feature support
> > >>>
> > >>> On Wed, Jan 09, 2019 at 05:34:38PM +0300, Ilya Maximets wrote:
> > >>>> virtio_mb() is really heavy. I'd like to avoid it somehow, but I
> > >>>> don't know how to do this yet.
> > >>>
> > >>> Linux driver doesn't avoid it either.
> > >>
> > >> I understand v3 was merged but still would like to continue the
> > >> discuss and make sure all is clear and agreed.
> > >>
> > >> Form patch [1] description it is very clear why we need the
> > >> rte_smp_mb() barrier.
> > >> However I am not sure why this barrier is interoperate into rte_mb in
> > >> case of vDPA. In vDPA case, both read of the user ring and write of
> > >> the avail index are for local cached memory.
> > >> The only write which is to uncachable memory (device memory) is the
> > >> notify itself.
> > >>
> > >> As I mentioned, there is a need to have a store fence before doing
> > >> the notify, but from different reasons. So vDPA use case and need Is
> > >> a bit different than what presented in [1].
> > >
> > > Any answer?
> > > It is pity if we add redundant barriers which will degrade the driver
> > performance.
> >
> > Sorry for late responses. Have a lot of work with OVS right now.
> >
> > Regarding your question.
> > Current code looks like this:
> >
> > 1. Update ring.
> > 2. virtio_wmb()
> > 3. Update idx.
> > 4. virtio_mb()
> > 5. read flags.
> > 6. notify.
> >
> > virtio_mb() is here to avoid reordering of steps 3 and 5.
> > i.e. we need a full barrier to ensure the order between store (idx update)
> > and load (reading the flags). Otherwise we could miss the notification.
> > We can't avoid the barrier here, because even x86 does not guarantee the
> > ordering of the local load with earlier local store.
>
> This is clear. You need the rte_smp_mb() here. My question is why you need
> the rte_mb() in case of vDPA?
> As you said, all accesses are local.
Are you asking why the different barriers? Or as you asking why is a barrier
needed at all?
The barriers themselves are clearly needed.
But in my opinion some dpdk barrier implementations are sub-optimal and too
strong. For example on intel: the big question is whether anyone does
any non-temporals. In absence of these, and with non-cacheable mappings
on x86 wmb and rmb should be a nop, and mb should be a locked instruction.
It might make sense to add rte_dma_rmb/wmb/mb.
> Pasting you commit code:
> /*
> * Per virtio_ring.h in Linux.
> * For virtio_pci on SMP, we don't need to order with respect to MMIO
> * accesses through relaxed memory I/O windows, so smp_mb() et al are
> * sufficient.
> *
> * For using virtio to talk to real devices (eg. vDPA) we do need real
> * barriers.
> */
> static inline void
> virtio_mb(uint8_t weak_barriers)
> {
> if (weak_barriers)
> rte_smp_mb();
> else
> rte_mb();
> }
>
> >
> > >
> > >>
> > >> [1]
> > >>
> > https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpa
> > >>
> > tches.dpdk.org%2Fpatch%2F49545%2F&data=02%7C01%7Cshahafs%40
> > mellan
> > >>
> > ox.com%7C01907f1b2e0e4002cb7508d67ac38a98%7Ca652971c7d2e4d9ba6a4
> > d1492
> > >>
> > 56f461b%7C0%7C0%7C636831377591864200&sdata=TSpc%2Fzyq2aq0N3
> > %2Bh4o
> > >> ro4std8ut%2FQU6%2BOeMDeeaQdsM%3D&reserved=0
> > >>
> > >>>
> > >>> --
> > >>> MST
> > >
> > >
