Re: [PATCH v3 6/8] net/vdev_netvsc: skip routed netvsc probing

[Thomas Monjalon <thomas@monjalon.net>]

11/01/2018 10:00, Matan Azrad:
> From: Stephen Hemminger, Wednesday, January 10, 2018 6:44 PM
> > > From: Stephen Hemminger, Tuesday, January 9, 2018 8:51 PM
> > > > On Tue,  9 Jan 2018 14:47:31 +0000
> > > > Matan Azrad <matan@mellanox.com> wrote:
> > > > > +static int
> > > > > +vdev_netvsc_has_route(const char *name) {
> > > > > +	FILE *fp;
> > > > > +	int ret = 0;
> > > > > +	char route[NETVSC_MAX_ROUTE_LINE_SIZE];
> > > > > +	char *netdev;
> > > > > +
> > > > > +	fp = fopen("/proc/net/route", "r");
> > > > > +	if (!fp) {
> > > > > +		rte_errno = errno;
> > > > > +		return 0;
> > > > > +	}
> > > > > +	while (fgets(route, NETVSC_MAX_ROUTE_LINE_SIZE, fp) != NULL) {
> > > > > +		netdev = strtok(route, "\t");
> > > > > +		if (strcmp(netdev, name) == 0) {
> > > > > +			ret = 1;
> > > > > +			break;
> > > > > +		}
> > > > > +		/* Move file pointer to the next line. */
> > > > > +		while (strchr(route, '\n') == NULL &&
> > > > > +		       fgets(route, NETVSC_MAX_ROUTE_LINE_SIZE, fp) !=
> > > > NULL)
> > > > > +			;
> > > > > +	}
> > > > > +	fclose(fp);
> > > > > +	return ret;
> > > > > +}
> > > >
> > > > In many ways /proc/net/route is legacy intervace.
> > > > And system may have 1 M routes.
> > > >
> > > > Maybe there is faster way to do this with netlink by looking to see
> > > > if there is an address associated with the interface.
> > >
> > > Actually this is control path, we don't care about performance very much.
> > > But I can get other idea here, Do you have suggestion?
> > >
> > > Thanks!
> > >
> > 
> > Use netlink (or ioctl) to get interface address.
> > If interface has an IPv4 or IPv6 (not link local), then skip it.
> 
> As I a little bit investigated I found that IPv6 getting is problematic by ioctl.
> And using nelink for it, really doesn't  worth the effort.
> So, I suggest to keep this code simple as is in spite of the optional high latency for this function, after all it is a control path.

No more comment?
So we are OK with this solution for now?

If we see real performance issue, I guess it can be fixed later.