Error handling fixes in lustre version read/parse code

[ofaaland]

Affects lmt-3.2.8

Commit

• f984897 handle lustre version file containing bare version Introduced a bug. It called msg_exit() when an error should be reported but execution should continue. This causes cerebrod to stop when lustre is not mounted, e.g. on node startup.

Commit

• 91c9c9d Revert "Add /sys in Lustre version checking (#31)" Introduced a bug. It changed the possible return values of _packed_lustre_version() from (-1,0) to (-2,-1,0). The rest of the code specifically assumes -1 for error cases, causing errors in reading the lustre version to go undetected and cause undefined behavior.

These two commits fix these bugs, and add trees to lustre_versions which contain missing and badly-formatted lustre "version" procfiles to verify error handling is working properly.

[ofaaland]

Hi Tony and Chris, Sorry to bug you again so soon, but my previous commits introduced bugs that are holding up the TOSS release. Can you review? Thanks.

[ofaaland]

Testing:

• ran make check - all pass
• ran t10-metric-strings by hand and confirmed that for no case does lmtmetric exit with nonzero status
• built rpms and installed on jet, confirmed cerebrod runs properly both with and without Lustre mounted