Bug 198231

Summary:	[unionfs] deadlock
Product:	Base System	Reporter:	Andre Albsmeier <mail>
Component:	kern	Assignee:	freebsd-fs (Nobody) <fs>
Status:	New ---
Severity:	Affects Some People	CC:	agh, arrowd
Priority:	---
Version:	9.3-RELEASE
Hardware:	i386
OS:	Any

Description Andre Albsmeier 2015-03-03 17:17:54 UTC

I can reliably deadlock a 9.3-STABLE by the following procedure:

Let's assume that /tmp is a standard swap-backed file system
already. First let's set up what we need:

mkdir /tmp/1 /tmp/2
mount -v -t unionfs /tmp/1 /usr/local
mount -v -t unionfs /tmp/2 /usr/local

No let's lock the system:

mkdir /tmp/2/bla
while :; do
  echo go
  tar -cC /usr/src/etc -f - . | tar -xpC /tmp/2/bla -f -
done

It survives about 3 or 4 rounds, sometimes more, sometimes
only 2. It is important to use tar to copy the stuff. If
we replace the tar line by e.g.

cp -pR /usr/src/etc/* /tmp/2/bla

things are all well.

The system doesn't lock up entirely, you can move the mouse
and ping it but no fs access is possible anymore. One can
switch to the console and enter the debugger but a reboot
with ctrl-alt-del doesn't work...

The interesting part is that all this worked pretty well on
9-STABLE until approx. 2 months ago. But nothing had been
committed to unionfs for a long time so I really have no
idea what's going on.

It also reminds us of

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=161511

but this stuff had been merged to 9-STABLE already...

Comment 1 Gleb Popov freebsd_committer

2019-08-01 12:55:22 UTC

I tried reproduction steps on 11.2 and 12.0, and there are no deadlocks. Can we close this?

Comment 2 Andre Albsmeier 2019-08-01 13:26:01 UTC

Just tested, it doesn't lock up here either on 11.3.
However, my actual use case was more complex -- for the PR I
just stripped it down as far as I could. For the time being
I adopted it so only one union mount was used (which worked).

Let me revert to the old mode and give me some time to check
(it will take a bit...)

Comment 3 Andre Albsmeier 2019-08-01 16:09:49 UTC

I have to revise my statement...

I ran the test below for about 10-20 seconds and was happy to see that it
doesn't hang the box anymore. Then I wrote the last comment (#2). After that
I continued with my regular work (some vi, ssh, firefox, ...). I did NOT unmount
the two unionfs on /tmp so they remained in the state they had after aborting
the test. Approx 2.5 hours later the machine suddenly hung with exactly the same
symptoms as described (mouse OK, KBD OK, no fs, no Ctrl-Alt-Del, had to go to the
console and restart using the debugger).

I did not touch the two mount points /tmp/1 and /tmp/2 but I used /usr/local
(read-only) of course and a lot of rw on /tmp.