Possible regression when unmounting encrypted datasets.

[asomers]

PR #7795 changed the way that ZFS unmounts in order to fix a corruption problem with encrypted datasets. PR #9115 essentially reverted it, but the commit message didn't suggest any knowledge of the problem that #7795 fixed, and the original author of #7795 did not review #9115. So there's a risk that it restored the original bug.

So it would be good if @tcaputi could retrospectively review #9115 .

System information

All

Describe the problem you're observing

No problem is observed; I just caught this potential problem by reading through old commits.

[rincebrain]

Since he hasn't done ZFS development in years at this point, I believe, someone else is probably a better option here.

[amotin]

I haven't looked too deep, but looking on original panic in #7753 I suspect there may be a race between last dnode of dataset being synced in dmu_objset_sync_dnodes() and userquota_updates_task() actually completed in another task.