Merge lp://staging/~dannf/maas-images/lp1702976 into lp://staging/maas-images

lgtm!

dannf@mustang:/tmp/maas-images+lp1702976$ gunzip < ./out.d/xenial/arm64/20170711/xenial/generic/boot-initrd | cpio -t | grep i2c-thunderx
lib/modules/4.4.0-83-generic/kernel/drivers/i2c/busses/i2c-thunderx.ko
205567 blocks

this looks sane... i'm mostly curious on the size added.
but other than that i think tihs is the only path to a solution.

Is there any reason we need all i2c bus drivers on all architectures? I think we may want to do something like this to limit only the i2c-thunderx driver being added on arm64.

=== modified file 'bin/kpack-from-image'
--- bin/kpack-from-image 2017-05-09 14:04:12 +0000
+++ bin/kpack-from-image 2017-07-12 18:08:50 +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 [ "$(uname -m)" == "arm64" ]; then
+ manual_add_modules i2c-thunderx
+fi
 EOF
     if [ "$rel" = "precise" ]; then
         echo "manual_add_modules squashfs # (LP: #1501834)" >> "$fname"

I fear that if we just pick i2c-thunderx, we'll just run into this problem again and again with new hosts. However, restricting i2c drivers to arm* seems reasonable. I'm unaware of any other archs that use the SSIF interface.

Note that using "uname -m" to figure out the architecture is nearly always a bad idea. Remember that there are all sorts of cases where the host kernel may not match the userspace - e.g. running i386 userspace on an x86_64 box, qemu-user emulation, etc. In this case, since the MAAS images are presumably always built on an x86 box, uname -m will never be "aarch64". I suggest using 'dpkg --print-architecture' instead.

The size increase for arm64 including all i2c/busses modules is 217K.

Restricted to only ARM architectures.

fwiw, the code that landed here is weird.

config_addl_modules writes a file using 'cat <<EOF'
since 'EOF' is not quoted, we get shell expansion of '$' when writing

so.. what ends up getting written is:

case arm64 in
  arm*) copy_modules_dir ....
esac

which really I suspect we wanted to either
a.) only *write* the 'copy_modules_dir' into the file if the arch was arm*
b.) evaluate the arch at execution time (of mkinitramfs-tools).

Indeed.
https://code.launchpad.net/~dannf/maas-images/+git/maas-images/+merge/360955