2017-01-23 13:06, Ananyev, Konstantin: > > > -----Original Message----- > > From: Yuanhan Liu [mailto:yuanhan....@linux.intel.com] > > Sent: Monday, January 23, 2017 12:53 PM > > To: Ananyev, Konstantin <konstantin.anan...@intel.com> > > Cc: Yigit, Ferruh <ferruh.yi...@intel.com>; dev@dpdk.org; Thomas Monjalon > > <thomas.monja...@6wind.com>; Horton, Remy > > <remy.hor...@intel.com> > > Subject: Re: [dpdk-dev] [PATCH] ethdev: fix wrong memset > > > > On Mon, Jan 23, 2017 at 12:44:11PM +0000, Ananyev, Konstantin wrote: > > > > On Mon, Jan 23, 2017 at 07:40:50PM +0800, Yuanhan Liu wrote: > > > > > On Mon, Jan 23, 2017 at 11:32:23AM +0000, Ferruh Yigit wrote: > > > > > > On 1/23/2017 11:24 AM, Yuanhan Liu wrote: > > > > > > > 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? > > > > > > > > > > > > No, memset(&rte_eth_dev_data[port_id], ...) breaks it. > > > > > > > > > > Oh, you were talking about that particular case Remy's patch meant to > > > > > fix. > > > > > > > > > > > >> and fix is known. > > > > > > > > > > > > > > And there is already a fix? > > > > > > > > > > > > http://dpdk.org/ml/archives/dev/2017-January/054422.html > > > > > > > > > > Yes, it should fix that issue. > > > > > > > > Well, few more thoughts: it may fix the crash issue Remy saw, but it > > > > looks like more a workaround to me. Basically, if primary and secondary > > > > shares a same port id, they should point to same device. Otherwise, > > > > primary process may use eth_dev->data for a device A, while the > > > > secondary process may use it for another device, as you said, it > > > > could be a vdev. > > > > > > > > In such case, there is no way we could continue safely. That said, > > > > the given patch avoids the total reset of eth_dev->data, while it > > > > continues reset the eth_dev->data->name, which is wrong. > > > > > > > > So it's not a proper fix. > > > > > > > > Again, I think it's more about the usage. If primary starts with > > > > a nic device A, while the secondary starts with a nic device B, > > > > there is no way they could work well (unless they use different > > > > port id). > > > > > > Why not? > > > I think this is possible. > > > > Yes, it's possible: find another port id if that one is already taken > > by primary process (or even by secondary process: think that primary > > process might attatch a port later). > > > > > They just need to be initialized properly, > > > so each rte_eth_devices[port_id]->data, etc. point to the right place. > > > > My understanding is, as far as they use different port_id, it might > > be fine. Just not sure it's enough or not. > > As I understand, the main problem is that rte_eth_devices[] is local, > while rte_eth_dev_data points to the shared memory array. > And rte_eth_dev_allocate() assumes that if rte_eth_devices[x] is free, > then rte_eth_dev_data[port_id] is also free. > Which is wrong in case when primary/secondary processes have different > devices attached. > Another problem is that inside rte_ethdev.c we manipulate rte_eth_dev_data[] > contents without grabbing any lock.
Yes there are a lot of problems with the multiproc mode because it has been implemented as a hack. We are fixing some cases without figuring the whole picture. I'll apply the patch from Remy which fixes a case where process creates vdev (local data) and, hopefully, primary does no hotplug of PCI dev. I'll restart this discussion with a bigger picture of what multiproc is, and what are the issues.
