It seems fsx support on zfs is now broken.

The test fsx performs is by doing fallocate of 1 byte that is optimized away for certain types of modes (like hole) in the client without even talking to the server. To proper trigger talking to the server we need to go at least one page away it seems.

The other problem is even if fallocate call does fail - it might still leave file size increased and so we must perform truncate back to zero at all times to ensure correct operation of subsequent test.

A number of these scenarios was explored in this test change over multiple changesets https://review.whamcloud.com/#/c/41200/ - check the janitor test results