Re: [PATCH] net/tap: Bug fix to populate fds in secondary process

[Thomas Monjalon <thomas@monjalon.net>]

18/01/2022 13:12, Ferruh Yigit:
> On 1/18/2022 11:21 AM, kumaraparameshwaran rathinavel wrote:
> 
> Comment moved down.
> 
> Please don't top post, it makes very hard to follow the discussion and bad
> for archives to visit discussion later.
> 
> > 
> > On Tue, Jan 18, 2022 at 3:17 PM Ferruh Yigit <ferruh.yigit@intel.com <mailto:ferruh.yigit@intel.com>> wrote:
> > 
> >     On 1/17/2022 6:33 PM, Thomas Monjalon wrote:
> >      > 17/01/2022 19:28, Ferruh Yigit:
> >      >>> +   ret = rte_eth_dev_get_port_by_name(request_param->port_name, &port_id);
> >      >>> +   if (ret) {
> >      >>> +           TAP_LOG(ERR, "Failed to get port id for %s",
> >      >>> +                   request_param->port_name);
> >      >>> +           return -1;
> >      >>> +   }
> >      >>> +   dev = &rte_eth_devices[port_id];
> >      >>
> >      >> Since this is not really related with your patch, I want to have a separate thread for it.
> >      >>
> >      >> It is not good to access the 'rte_eth_devices' global variable directly from a driver, that
> >      >> is error prone.
> >      >>
> >      >> Btw, what 'peer' supposed to contain?
> >      >>
> >      >> It can be solved by adding an internal API, only for drivers to get eth_dev from the name,
> >      >> like: 'rte_eth_dev_get_by_name()'.
> >      >> This way a few other usage can be converted to this API.
> >      >>
> >      >> @Thomas and @Andrew what do you think about the new API proposal?
> >      >
> >      > It looks similar to rte_eth_dev_get_port_by_name() which returns a port_id.
> > 
> >     Exactly, but get eth_dev directly for drivers. For drivers no need to work with port_id
> >     handler, they can use eth_dev directly.
> > 
> >     Another solution can be an getter function for drivers, which gets port_id and returns
> >     the eth_dev.
> > 
> >      > It is a bit strange for an ethdev driver to not have access to its own ethdev struct.
> >      > Isn't there something broken in the logic?
> >      >
> > 
> >     This is callback function between primary and secondary applications sync. So port name
> >     will be same for both, but eth_dev will be different and port_id may be different.
> >     Driver finds its own eth_dev from the shared port name.
> > 
> 
> > Just wanted to bring it to your attention,
> > 
> > In Mellanox driver there is a requirement to exchange fds between primary and secondary and similar usage is seen, the primary sends the port_id and the secondary refers to the rte_eth_devices in the driver,
> > The functions are
> >             - mlx5_mp_secondary_handle in secondary
> >             - mlx5_mp_req_start_rxtx in primary which is invoked from mlx5_dev_start.
> > 
> > In my implementation I have used the name and invoked get_port_by_name, I can also pass the port_id from the primary to make it uniform. So with similar usage in Mellanox is there a problem there as well on referring to the rte_eth_devices from the PMD ?
> > 
> 
> It would be same, still will be accessing to the 'rte_eth_devices'.
> That is why a new API for drivers may help.

I agree to add a new API if needed to remove those direct access to rte_eth_devices.