Merge lp:~ltrager/maas-images/lp1702976 into lp:maas-images

[Lee Trager]

Include the i2c-thunderx driver on ARM64 only.

Enlistment fails on Cavium ThunderX systems as the IPMI SSIF interface requires the i2c-thunderx module to access the BMC. Include the i2c-thunderx driver in the published ARM64 initrds.

[Andres Rodriguez (andreserl)]

Lgtm! Have you tested this? Also, please superseed the other MP!

review: Approve

[dann frazier (dannf)]

My comments from my original merge proposal still stand here:
  https://code.launchpad.net/~dannf/maas-images/lp1702976/+merge/327308

Specifically, I'd prefer we do this for arm*, just just for arm64, and that we include all the bus drivers, not just the i2c-thunderx driver. I don't know of any specific system that would be fixed by either of the above - but seems like a cleaner future-proofing. And obviously, it's ultimately your call.

I'd also like to continue to ask that we use 'dpkg --print-architecture' instead of 'uname -m'. From chatting on IRC, it sounds like qemu-user does emulate 'uname -m' to match the userspace, but that's just one form of a non-native build mechanism. Consider if we were building images for armhf on an arm64 host in an lxd container:

root@armhf:~# uname -m
armv8l

Whereas, we know in this case 'dpkg --print-architecture' will always report the thing we want to know:
root@armhf:~# dpkg --print-architecture
armhf

=== modified file 'bin/kpack-from-image'
--- bin/kpack-from-image	2017-05-09 14:04:12 +0000
+++ bin/kpack-from-image	2017-07-12 19:56:26 +0000
@@ -187,6 +187,10 @@
 copy_modules_dir kernel/drivers/char/ipmi # LP: #1333271
 copy_modules_dir kernel/drivers/net/usb   # LP: #1552378
 manual_add_modules bcache # (LP: #1513176)
+# LP: #1702976
+if [ `uname -m` = 'aarch64' ]; then
+    manual_add_modules i2c-thunderx
+fi
 EOF
     if [ "$rel" = "precise" ]; then
         echo "manual_add_modules squashfs # (LP: #1501834)" >> "$fname"