[PATCH 3/3] net/eth: fix link handling

[Jean-Christophe PLAGNIOL-VILLARD]

On 15:08 Wed 26 Sep     , Sascha Hauer wrote:
> On Wed, Sep 26, 2012 at 03:03:53PM +0200, Jean-Christophe PLAGNIOL-VILLARD wrote:
> > On 14:50 Wed 26 Sep     , Jan Lübbe wrote:
> > > On Wed, 2012-09-26 at 14:43 +0200, Sascha Hauer wrote:
> > > > The eth->active is for tracking whether a device is opened (thus we have
> > > > to call eth->halt). With this patch eth->active == 0 could mean:
> > > > 
> > > > a) device not opened
> > > > b) opened, but no link
> > > > 
> > > > This is asking for trouble. A bit only has two states, you should not
> > > > encode three states in it.
> > > 
> > > Yes. We should not repeatedly open the device to check link status
> > > again. A open (active==1) device with no link is perfectly valid, it
> > > just can't be used for TX/RX.
> > > 
> > > So if we have an open device in TX/RX and
> > >   link down: recheck link
> > >   link up: recheck link if the last check is "too old"
> > > 
> > > This way we notice both link going down and up, while not wasting time
> > > when we are waiting for link up.
> > except this can not work in barebox. In linux you can as you have a schedule
> > work that check the phy status. In barebox we do not do so and I do not want
> > it (slow down the boot). So we check it at the open only and only once. When
> > the phy is up we do not check it again.
> > 
> > Remember we are in a bootloader not in an os we need to do some concession.
> 
> int check_link_status(void)
> {
> 	if (is_timeout(last_link_check, SECOND)) {
> 		do_link_check();
> 		last_link_check = get_time_ns();
> 	}
> }
we can do this but every second is too much this will slow down nfs/tftp
transfert every 30s or 10s will be better

Best Regards,
J.