Re: [dpdk-dev] [PATCH v7 03/11] net/failsafe: add fail-safe PMD

["Gaëtan Rivet" <gaetan.rivet@6wind.com>]

On Sun, Jul 09, 2017 at 07:10:49AM -0400, Jan Blunck wrote:
> On Sat, Jul 8, 2017 at 9:47 PM, Gaetan Rivet <gaetan.rivet@6wind.com> wrote:
> > +static int
> > +fs_parse_device(struct sub_device *sdev, char *args)
> > +{
> > +       struct rte_devargs *d;
> > +       int ret;
> > +
> > +       d = &sdev->devargs;
> > +       DEBUG("%s", args);
> > +       ret = rte_eal_devargs_parse(args, d);
> > +       if (ret) {
> > +               DEBUG("devargs parsing failed with code %d", ret);
> > +               return ret;
> > +       }
> > +       sdev->bus = d->bus;
> > +       sdev->state = DEV_PARSED;
> 
> You seem to be mostly interested in the bus name for the device. Why
> don't you track this via your sub_device structure instead of using
> rte_devargs?
> 
> 

I don't understand. I track it in my sub_device structure by copying the
info from the devargs? How would you get it otherwise?

Devargs still has to be used as a common format that can be parsed
by EAL helpers. I could certainly parse it myself, but then I would have
to follow any evolution of the parameter format, and I prefer to
delegate this to the EAL.

> > +       return 0;
> > +}
> > +

[...]

> > +static struct rte_eth_dev *
> > +fs_find_ethdev(const struct rte_device *dev)
> > +{
> > +       struct rte_eth_dev *eth_dev;
> > +       uint8_t i;
> > +
> > +       RTE_ETH_FOREACH_DEV(i) {
> > +               eth_dev = &rte_eth_devices[i];
> > +               if (eth_dev->device == dev)
> > +                       return eth_dev;
> > +       }
> > +       return NULL;
> > +}
> 
> Why don't you use rte_eth_dev_allocated() here?
> 
> 

IIRC, at the time I choose to roll out my own version, there was
ambiguities about device names availability, with some devices using
their rte_devargs name, others writing their own.

It has been stabilized it seems, and now I think rte_eth_dev_allocated
could be used.

[...]

> > +
> > +static int
> > +fs_bus_init(struct rte_eth_dev *dev)
> > +{
> > +       struct sub_device *sdev;
> > +       struct rte_device *rdev;
> > +       struct rte_devargs *da;
> > +       uint8_t i;
> > +       int ret;
> > +
> > +       FOREACH_SUBDEV(sdev, i, dev) {
> > +               if (sdev->state != DEV_PARSED)
> > +                       continue;
> > +               da = &sdev->devargs;
> > +               rdev = rte_eal_hotplug_add(da->bus->name,
> > +                                          da->name,
> > +                                          da->args);
> 
> Why don't you track the bus name through your sub_device structure instead?
> 
> 

Do you mean

+               rdev = rte_eal_hotplug_add(sdev->bus->name,
+                                          da->name,
+                                          da->args);

instead? I guess for this example it was mostly for the sake of
consistency, but the bus handle within the sub_device could be
used.

Actually I dislike having a copy of an info. I have this info both
within the sub_device directly and within the rte_devargs within the
sub_device. It should be streamlined.

-- 
Gaëtan Rivet
6WIND