On Sat, 17 Oct 2020 14:34:11 +0200 Michal Kubecek wrote:
> On Fri, Oct 16, 2020 at 10:02:38AM +0800, zhudi wrote:
> > "ip addr show" command execute error when we have a physical
> > network card with number of VFs larger than 247.
> >
> > The return value of if_nlmsg_size() in rtnl_calcit() will exceed
> > range of u16 data type when any network cards has a larger number of
> > VFs. rtnl_vfinfo_size() will significant increase needed dump size when
> > the value of num_vfs is larger.
> >
> > Eventually we get a wrong value of min_ifinfo_dump_size because of overflow
> > which decides the memory size needed by netlink dump and netlink_dump()
> > will return -EMSGSIZE because of not enough memory was allocated.
> >
> > So fix it by promoting min_dump_alloc data type to u32 to
> > avoid data overflow and it's also align with the data type of
> > struct netlink_callback{}.min_dump_alloc which is assigned by
> > return value of rtnl_calcit()
>
> Unfortunately this is only part of the problem. For a NIC with so many
> VFs (not sure if exactly 247 but it's close to that), IFLA_VFINFO_LIST
> nested attribute itself would be over 64KB long which is not possible as
> attribute size is u16.
>
> So we should rather fail in such case (except when IFLA_VFINFO_LIST
> itself fits into 64KB but the whole netlink message would not) and
> provide an alternative way to get information about all VFs.
Right, we should probably move to devlink as much as possible.
zhudi, why not use size_t? Seems like the most natural fit for
counting size.
