[regression] Client ABI 8 broken in 0.13 series

[Daniel van Vugt]

I've just done a little sanity testing of our client ABI changes and found we've broken the client 8 ABI at least once, maybe twice in series 0.13.

Test case: Drop older demos client binaries (which still only use libmirclient8) into a newer build tree.

Expected: demo clients still run.
Observed: demo clients fail to start with:
bin/.mir_demo_client_fingerpaint-uninstalled: relocation error: bin/.mir_demo_client_fingerpaint-uninstalled: symbol mir_surface_get_buffer_stream, version MIR_CLIENT_8 not defined in file libmirclient.so.8 with link time reference

[Daniel van Vugt (vanvugt)]

Seems possibly to be what I predicted:
https://bugs.launchpad.net/mir/+bug/1415321/comments/3

ldd bin/.mir_demo_client_eglsquare-uninstalled
 libmirclient.so.8 => ./lib/libmirclient.so.8 (0x00007f58fa14d000)
 libmircommon.so.3 => /usr/lib/x86_64-linux-gnu/libmircommon.so.3 (0x00007f58f9ee2000)
 libmircommon.so.4 => /home/dan/bzr/mir/dev/build/lib/libmircommon.so.4 (0x00007f58f8f7a000)

[Daniel van Vugt (vanvugt)]

Hmm, OK. Seems the test only fails locally between r2407 and 2408. But on the wider scale between releases we've maintained compatibility from 0.11/0.12 to dev r2408.

It seems the problem was fixed in r2408 relative to previous releases. However revisions after 0.12.0 and prior to r2408 had ABI breaks.

And yes, it does indeed seem that we've maintained ABI compatibility by virtue of having libmircommon.so.3 and libmircommon.so.4 loadable in the same client process simultaneously. Scary, but seems to work.