On Mon, Jan 23, 2017 at 11:05:25AM +0000, Ferruh Yigit wrote:
> >>>>>> lib/librte_ether/rte_ethdev.c | 2 +-
> >>>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
> >>>>>>
> >>>>>> diff --git a/lib/librte_ether/rte_ethdev.c
> >>>>>> b/lib/librte_ether/rte_ethdev.c
> >>>>>> index 4790faf..61f44e2 100644
> >>>>>> --- a/lib/librte_ether/rte_ethdev.c
> >>>>>> +++ b/lib/librte_ether/rte_ethdev.c
> >>>>>> @@ -225,7 +225,7 @@ struct rte_eth_dev *
> >>>>>> return NULL;
> >>>>>> }
> >>>>>>
> >>>>>> - memset(&rte_eth_devices[port_id], 0, sizeof(*eth_dev->data));
> >>>>>> + memset(&rte_eth_dev_data[port_id], 0, sizeof(struct
> >>>>>> rte_eth_dev_data));
> >>>>>
> >>>>> Not directly related to the this issue, but, after fix, this may have
> >>>>> issues with secondary process.
> >>>>>
> >>>>> There were patches sent to fix this.
> >>>>
> >>>> I mean this one:
> >>>> http://dpdk.org/ml/archives/dev/2017-January/054422.html
> >>>
> >>> d948f596fee2 ("ethdev: fix port data mismatched in multiple process
> >>> model") should have fixed it.
> >>
> >> Think about case, where secondary process uses a virtual PMD, which does
> >> a rte_eth_dev_allocate() call, shouldn't this corrupt primary process
> >> device data?
> >
> > Yes, it may. However, I doubt that's the typical usage.
>
> But this is a use case, and broken now,
I thought it was broken since the beginning?
> and fix is known.
And there is already a fix?
> Should be
> fixed I think.
Sure.
>
> > Besides that,
> > most of virtual PMDs don't support Multipleprocess: git grep shows pcap
> > is the only one that does claim Multipleprocess is supported.
>
> I guess you searched for NIC feature documentation for this.
Yes.
> But as far
> as I know, all virtual drivers can be used in both primary and secondary
> process.
Maybe. But it becomes very error-prone to me then when vdev are involved
in both primary and secondary process. I don't think current code is (or
designed to be) strong enough to support that.
I don't know it's allowed to use hotplug or not in the multiple process
model. If yes, I think there would be many ways to break it.
Honestly, the multiple process doesn't look like a good/clean design to
me, especially when some piece of code claim to support it while some
other doesn't.
So my point was, yes, there is a bug, we should fix it. But it seems
that there could be so many bugs if we hugely expand the test coverage
of the multiple process feature.
--yliu
