[Bug 229986] Terminating virtualbox or ctld + zfs scrub + gstat leads to deadlock

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=229986

--- Comment #9 from Martin Birgmeier <d8zNeCFG at aon.at> ---
Actually I am quite sure that the system *is* in a kind of memory deadlock at
least in the case where it ultimately gets stuck completely (comment #4). In
that case, if I remember correctly, it also used quite some swap (around 5% of
20G configured).

But for the issue with scrubbing interacting badly with shutting down VMs and
stopping ctld, this happens without too much memory pressure.

Specifically, I have:
- 16G main memory; this initially gives vfs.zfs.arc_max: 15371956224, which is
high, considering that dmesg.boot shows
real memory  = 17179869184 (16384 MB)
avail memory = 16344322048 (15587 MB)
- Then use sysctl vfs.zfs.arc_max=9000000000
- The VM run directly on this machine has 3G simulated physical memory

The scrub has completed in the meantime, and I don't really want to restart it,
having taken 360 hours in total, and with it being very taxing on the disks. So
it is difficult to replay the scenario now.

However, what is still happening and may be related to memory pressure again:
When I shut down ctld after many hours of operation of this machine using
'service ctld onestop' (see comment #2), in contrast to 11.1 this operation may
take several minutes. I believe this happens especially if I forget to reduce
arc_max.

In this case again it looks to me like memory pressure and that the system
takes a long time to get some free memory which seems to be required for this
operation.

Maybe it is related to zvols, both the VM running directly on the machine as
well as the ctld reference zvols.

-- Martin

-- 
You are receiving this mail because:
You are the assignee for the bug.