On 06/01/2017 09:24 AM, Russell King - ARM Linux wrote: > On Thu, Jun 01, 2017 at 04:47:35PM +0100, Russell King - ARM Linux wrote: >> On Thu, Jun 01, 2017 at 03:19:55PM +0200, Andrew Lunn wrote: >>> On Thu, Jun 01, 2017 at 02:09:00PM +0100, Russell King - ARM Linux wrote: >>>> On Thu, Jun 01, 2017 at 03:05:27PM +0200, Andrew Lunn wrote: >>>>> So you are saying a 10G PHY driver always needs to have a aneg_done >>>>> callback, even if it just needs to call phygen_c45_aneg_done? >>>>> >>>>> This seems a bit error prone. I can see somebody writing a 10G driver, >>>>> leaving out aneg_done() and having the c22 version called. Is the read >>>>> of MII_BMSR likely to return 0xffff, since the register does not >>>>> exist? If so, genphy_aneg_done() is likely to always return >>>>> BMSR_ANEGCOMPLETE. >>>> >>>> Don't forget that the read will fail, so phy_read() will return a >>>> negative number. >>> >>> By fail, you mean return something like -EIO or -ETIMEOUT? Is this >>> guaranteed in the code somewhere? This particular Marvell PHY only >>> does c45. But i could imagine some other PHYs answering a c22 request >>> with 0xffff. >> >> Yes, C45 allows the PHYs to answer C22 as well, but then they have to >> implement the C22 register set. Such a PHY would be out of spec, >> especially as what you're suggesting is that it answers C22 cycles >> and fails to implement MII_BMSR. I also think that there's a comment >> in the 802.3 specs that says that unimplemented registers are to >> return zero, not 0xffff. > > Checking 802.3-2015, in "22.2.4 Management functions", the first > sentence requires all PHYs that respond to Clause 22 cycles to > implement BMCR and BMSR. However, my statement about unimplemented > registers returning zero seems to be a C45 thing, not a C22 thing, > according to the C22 PICS and "45.2 MDIO Interface Registers" > > However, digging a bit further, "22.2.4.2.10 Auto-Negotiation complete" > states that bit 5 shall be zero if aneg has not completed or if aneg is > unimplemented. >
So how about we are more defensive than that and if we are presented with a C45 PHY driver that does not have an aneg_done() function pointer set we return an error/warning during driver registration? -- Florian
