Re: [PATCH] usb-net/pegasus: fix pegasus carrier detection

[Dan Williams <dcbw@redhat.com>]

On Wed, 2007-04-25 at 17:58 +0300, Petko Manolov wrote:
> In general i agree with the reasoning below.  However, isn't it better to 
> remove the code that sets carrier on/off in intr_callback()?

I'm fine with this; whatever makes carrier status work makes me happy :)

Dan

> There's a reliable way of getting the link status by reading the MII. 
> After correct checking of the return value from read_mii_word(), 
> set_carrier() is what is good enough.  If 2 seconds is too long of an 
> interval we could reduce it to 1 second or, if needed, less.
> 
> I'd like to avoid adding additional flags per device as it will take 
> forever to collect information about their "correct" behavior and update 
> pegasus.h.  In short i think this part of your patch should be enough:
> 
> ---
> 
> @@ -847,10 +848,16 @@ static void intr_callback(struct urb *urb)
>  		 * d[0].NO_CARRIER kicks in only with failed TX.
>  		 * ... so monitoring with MII may be safest.
>  		 */
> -		if (d[0] & NO_CARRIER)
> -			netif_carrier_off(net);
> -		else
> -			netif_carrier_on(net);
> -
>  		/* bytes 3-4 == rx_lostpkt, reg 2E/2F */
>  		pegasus->stats.rx_missed_errors += ((d[3] & 0x7f) << 8) | d[4];
> @@ -950,7 +957,7 @@ static void set_carrier(struct net_device *net)
>  	pegasus_t *pegasus = netdev_priv(net);
>  	u16 tmp;
> 
> -	if (!read_mii_word(pegasus, pegasus->phy, MII_BMSR, &tmp))
> +	if (read_mii_word(pegasus, pegasus->phy, MII_BMSR, &tmp))
>  		return;
> 
> ---
> 
> 
> cheers,
> Petko
> 
> 
> On Tue, 24 Apr 2007, Dan Williams wrote:
> 
> > On Tue, 2007-04-24 at 20:48 +0300, petkan@nucleusys.com wrote:
> >>> On Tue, Apr 24, 2007 at 12:49:12PM -0400, Jeff Garzik wrote:
> >>>>  Long term, Greg seemed OK with moving the net drivers from
> >>>> drivers/usb/net
> >>>>  to drivers/usb/net, in line with the current policy of placing net
> >>>> drivers
> >>>>  in drivers/net/*, bus agnostic.  After that move, sending to netdev and
> >>>> me
> >>>>  (as you did here) would be the preferred avenue.
> >>>
> >>> Speaking of which, do you want me to do this in the 2.6.22-rc1
> >>> timeframe?  Usually big code moves like this are good to do right after
> >>> rc1 comes out as the major churn is usually completed then.
> >>
> >> Sorry to interfere, but could you guys wait until tomorrow before applying
> >> the patch to your respective GIT trees?  I'd like to check if the code is
> >> doing the right thing and avoid patch reversal.
> >
> > Original problem was that the patch I referenced in the commit message
> > from Jan 6 2006 switched the return value semantics from
> > read_mii_word().  Before the patch, read_mii_word returned 1 on success,
> > 0 on error.  After the patch, it returns the generally accepted 0 on
> > success and !0 on error.
> >
> > That causes set_carrier() to return immediately rather than fiddle with
> > netif_carrier_*.  When the Jan 6 2006 patch went in changing the return
> > values, set_carrier() was not updated for the new return values.
> > Nothing else in the code cares about read_mii_word()'s return value
> > except set_carrier().
> >
> > But when the card is brought up and no cable is plugged in,
> > intr_callback() gets called repeatedly, which itself repeatedly calls
> > netif_carrier_on() due to the NO_CARRIER check.  The comment there about
> > "NO_CARRIER kicks in on TX failure" seems accurate, because even with no
> > cable plugged in, and therefore no packets getting transmitted, the
> > NO_CARRIER check is never true on the Belkin part.  Therefore,
> > netif_carrier_on() is always called as a result of the failure of d[0] &
> > NO_CARRIER, turning carrier back on even if there is no cable plugged
> > in.  This bulldozes over the MII carrier_check routine too.
> >
> > I don't think the intr_callback() code should ever turn the carrier
> > _on_, because there's that 2*HZ MII carrier check which can certainly
> > handle the carrier on/off stuff.
> >
> > LINK_STATUS appears valid on the Belkin part too, so we can add that as
> > a reverse-quirk and use LINK_STATUS on parts where it works.  If you
> > think that the NO_CARRIER check should be in _addition_ to the
> > LINK_STATUS check, that's fine with me, provided that the NO_CARRIER
> > check only turns carrier off.
> >
> > Dan
> >
> >
> >