On Wed, Apr 05, 2023 at 10:51:38PM +0200, Andrew Lunn wrote:
> On Wed, Apr 05, 2023 at 11:41:16PM +0300, Ivan Bornyakov wrote:
> > On Wed, Apr 05, 2023 at 09:35:31PM +0200, Andrew Lunn wrote:
> > > On Wed, Apr 05, 2023 at 06:39:00PM +0300, Ivan Bornyakov wrote:
> > > > sfp->i2c_block_size is initialized at SFP module insertion in
> > > > sfp_sm_mod_probe(). Because of that, if SFP module was not inserted
> > > > since boot, ethtool -m leads to zero-length I2C read attempt.
> > > >
> > > > # ethtool -m xge0
> > > > i2c i2c-3: adapter quirk: no zero length (addr 0x0050, size 0, read)
> > > > Cannot get Module EEPROM data: Operation not supported
> > >
> > > Do i understand you correct in that this is when the SFP cage has
> > > always been empty? The I2C transaction is going to fail whatever the
> > > length is.
> > >
> >
> > Yes, SFP cage is empty, I2C transaction will fail anyways, but not all
> > I2C controllers are happy about zero-length reads.
> >
> > > > If SFP module was plugged then removed at least once,
> > > > sfp->i2c_block_size will be initialized and ethtool -m will fail with
> > > > different error
> > > >
> > > > # ethtool -m xge0
> > > > Cannot get Module EEPROM data: Remote I/O error
> > >
> > > So again, the SFP cage is empty?
> > >
> > > I wonder if a better fix is to use
> > >
> > > sfp->state & SFP_F_PRESENT
> > >
> > > in sfp_module_eeprom() and sfp_module_eeprom_by_page() and don't even
> > > do the I2C read if there is no module in the cage?
> > >
> >
> > This is also worthy addition to sfp.c, but IMHO sfp->i2c_block_size
> > initialization still need to be fixed since
> >
> > $ grep -c "sfp_read(" drivers/net/phy/sfp.c
> > 31
> >
> > and I can't vouch all of them are possible only after SFP module
> > insertion. Also for future proof reason.
>
> I think everything else should be safe. A lot of those reads are for
> the HWMON code. And the HWMON code only registers when the module is
> inserted.
>
> How about two patches, what you have here, plus checking sfp->state &
> SFP_F_PRESENT in the ethtool functions?
>
> Andrew
Sure, will do in the near future.
