Hung task when trying to write an image to a zvol

[gcbirzan]

Similar to openzfs/spl#339, after applying dweeezil/spl@f148f72, we now get:

https://gist.github.com/gcbirzan/b14bc85234ab5d7c1e8b

[tuxoko]

As I've commented in openzfs/spl#331
The problem is that kthread_create failed because it received a SIGKILL, not because of insufficient memory.

The patch by @dweeezil will cause the thread to stuck repeatedly calling kthread_create and failed because the SIGKILL is not surved.

You should try to find out why the thread is receiving SIGKILL.

[dweeezil]

@gcbirzan I just pushed dweeezil/spl@0d36f37 to deal with the interrupted kthread_create() case. What was the condition under which your failure referenced above occurred? Was it during boot-time import? My patch will cause the process to be properly interrupted, but the question is why your process was being interrupted in the first place. Apparently systemd can send signals to various processes during startup if they take too long and I'm wondering if that's your situation.

[gcbirzan]

We've experienced this problem both when importing a drive (though, not as often as the first one, which was 100% sure), and when trying to destroy a volume after having dedup on the pool. The latter ended up with the pool being a write-off, since it just freezes the machine after a while (it runs out of memory and at some point just freezes). All of this only happens on 3.13, we have a 3.5 machine we tried the same on and it works flawlessly so far (no dedup though, because fuck dedup).

We're, however, planning to go live with this in the near future, so not sure how long we have for testing, but I'll try to apply your patch and import the broken pool, but I'm fairly confident (considering we've seen this on 2 separate machines) that it's relatively easy to reproduce.

[behlendorf]

This issue was resolved by openzfs/spl#339.