Dear Jason Cooper,

On Wed, 17 Jun 2015 21:39:26 +0000, Jason Cooper wrote:

Odd, I'd use that as an example of the process working.  ;-)  we have
everyone using 'armada-370-neta' for a given block.  We discovered that
the original IP block (on the 370s) had a limitation (no hw checksum
for greater than 1600 bytes).  A newer version of the IP block (XP)
doesn't have the limitation.

So we change the driver to honor the limit for the 370 compatible
string.  We create a new compatible string for xp where the block
doesn't have the limitation.

How did the process fail?

Because now all Armada XP users of jumbo frames are looking the HW
checksum on their jumbo frames, which you can consider to be a
regression: it was working, it is no longer working.

Of course, since it falls back to SW checksumming, it still "works",
but some users can complain of the performance penalty and consider it
to be a regression.

If on Armada XP, we had used for the beginning:

	compatible = "marvell,armada-xp-neta", "marvell,armada-370-neta"

with only marvell,armada-370-neta supported originally, we could have
added this fix without breaking HW checksumming on jumbo frames for
Armada XP users.

So I'm sorry, but the process indeed failed, because Armada XP users
keeping their old Device Tree blob will see a regression.

I'm not seeing where backwards compatibility was broken?  A device with
an old dtb booting a newer kernel gets a bugfix.  In the case of an XP
board with an old dtb (armada-370-neta), the hardware still works, but
not optimally.  Upgrading the dtb will enable hw checksumming for jumbo
packets.

"not optimally" is still a breakage.

Again, I personally don't care about DT backward compatibility as I
think it's a stupid requirement. But I like to point out to the
DT backward compatibility fanatics when it was actually broken :-)

Best regards,

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com