[dpdk-dev] [PATCH] kni: fix kni rx fifo producer synchronization
Jerin Jacob
jerin.jacob at caviumnetworks.com
Thu Aug 16 08:52:57 CEST 2018
-----Original Message-----
> Date: Thu, 9 Aug 2018 12:30:57 +0100
> From: Ferruh Yigit <ferruh.yigit at intel.com>
> To: Kiran Kumar <kkokkilagadda at caviumnetworks.com>
> CC: dev at dpdk.org, Jerin Jacob <jerin.jacob at caviumnetworks.com>, Konstantin
> Ananyev <konstantin.ananyev at intel.com>, Bruce Richardson
> <bruce.richardson at intel.com>, Santosh Shukla
> <santosh.shukla at caviumnetworks.com>
> Subject: Re: [dpdk-dev] [PATCH] kni: fix kni rx fifo producer
> synchronization
> User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:52.0) Gecko/20100101
> Thunderbird/52.9.1
>
>
> On 8/9/2018 11:23 AM, Kiran Kumar wrote:
> > With existing code in kni_fifo_put, rx_q values are not being updated
> > before updating fifo_write. While reading rx_q in kni_net_rx_normal,
> > This is causing the sync issue on other core. So adding a write
> > barrier to make sure the values being synced before updating fifo_write.
> >
> > Fixes: 3fc5ca2f6352 ("kni: initial import")
> >
> > Signed-off-by: Kiran Kumar <kkokkilagadda at caviumnetworks.com>
> > ---
> > lib/librte_kni/rte_kni_fifo.h | 1 +
> > 1 file changed, 1 insertion(+)
> >
> > diff --git a/lib/librte_kni/rte_kni_fifo.h b/lib/librte_kni/rte_kni_fifo.h
> > index ac26a8c..4d6b33e 100644
> > --- a/lib/librte_kni/rte_kni_fifo.h
> > +++ b/lib/librte_kni/rte_kni_fifo.h
> > @@ -39,6 +39,7 @@ kni_fifo_put(struct rte_kni_fifo *fifo, void **data, unsigned num)
> > fifo->buffer[fifo_write] = data[i];
> > fifo_write = new_write;
> > }
> > + rte_smp_wmb();
> > fifo->write = fifo_write;
> > return i;
>
> For Intel this is just a compiler barrier so no issue but not sure if a memory
> barrier is required here,
>
> Related code block is:
>
> |- for (i = 0; i < num; i++) {
> || new_write = (new_write + 1) & (fifo->len - 1);
> ||
> || if (new_write == fifo_read)
> || break;
> || fifo->buffer[fifo_write] = data[i];
> || fifo_write = new_write;
> || }
> | fifo->write = fifo_write;
>
> "fifo_write" is updated in the loop, so there is a dependency to it for
> "fifo->write". Can memory writes be reordered when there is a dependency?
>From CPU0 compiler instruction generation perspective, It will take care of the
dependency. In weakly ordered machine, it does NOT grandees CPU1 sees
fifo->write as final write.
So, IMO, This patch make sense. We are able to reproduce the problem on
arm64 machine with certain traffic.
>
> Cc'ed a few more people for comment.
More information about the dev
mailing list