Refreshing BTRFS backed instances is causing a segfault

[roosterfish]

When refreshing Btrfs backed instances the command hangs infinitely and doesn't return. The affected LXD versions are both latest/edge and 5.0/edge.

The error is caused in https://github.com/canonical/lxd/blob/main/lxd/storage/drivers/driver_btrfs_volumes.go#L882 when running the btrfs receive command.

This was first discovered in canonical/lxd-ci#101.

Steps to reproduce

julian@thinkpad:~$ lxc storage create b btrfs
Storage pool b created
julian@thinkpad:~$ lxc launch ubuntu:jammy v1 --vm -s b
Creating v1
Starting v1                               
julian@thinkpad:~$ lxc snapshot v1
julian@thinkpad:~$ lxc cp v1 v2
julian@thinkpad:~$ lxc cp v1 v2 --refresh # doesn't return

root@thinkpad:~# dmesg -w
[34150.527621] btrfs[656287]: segfault at 56 ip 000055a07a36c3b4 sp 00007ffcc2149c60 error 4 in btrfs[55a07a32a000+71000] likely on CPU 12 (core 24, socket 0)
[34150.527663] Code: c0 79 48 e8 2e eb fb ff 4c 89 f6 48 8d 3d dd 31 05 00 44 8b 20 31 c0 41 f7 dc e8 6f 33 fe ff eb 2a 41 83 cd ff 48 85 ed 74 28 <48> 8b 7d 58 e8 63 ea fb ff 48 89 ef e8 5b ea fb ff 41 83 fd ff 74

root@thinkpad:~# journalctl -u snap.lxd.daemon -f
Mär 08 18:40:17 thinkpad lxd.daemon[656287]: ERROR: clone: did not find source subvol

The same is happening for containers. If the source container is stopped it doesn't occur. For VMs it doesn't matter if the VM is running or stopped.

[tomponline]

Strange its not failing the main test suite?

[roosterfish]

There is a potential fix upstream which isn't yet merged but I can confirm the error cannot be reproduced when Btrfs is built from kdave/btrfs-progs#643.
I'll test with a more recent version of Btrfs too in order to narrow it down a bit more.

[tomponline]

I've for a fix in the works for snapshot cleanup too

[tomponline]

There is a potential fix upstream which isn't yet merged but I can confirm the error cannot be reproduced when Btrfs is built from kdave/btrfs-progs#643. I'll test with a more recent version of Btrfs too in order to narrow it down a bit more.

Do u think this ever worked and is a regression in the btrfs tooling?

[tomponline]

Ah seems like its caused by certain content inside the instance kdave/btrfs-progs#606 (comment)

[roosterfish]

Ah seems like its caused by certain content inside the instance kdave/btrfs-progs#606 (comment)

Yeah this also raised my attention. What I am wondering is what might cause a write in our case as we are freezing the FS.

[roosterfish]

So it looks to be that this has started failing from btrfs-progs release v5.14.91 onwards. Version 5.14.2 is the last one which doesn't raise the error.
In the CHANGES file for this specific release I cannot identify any specifics on this. There aren't any additional release notes for this tag in GitHub.

In the snap we currently use the btrfs-progs package installed via apt using version 5.16.2.
If we would go back to version 5.14.2 we could fix the problem for new storage pools but already existing storage pools will continue to fail on the --refresh operation. The only change is that we cannot see the segfault message anymore but btrfs receive still returns ERROR: clone: did not find source subvol.

[tomponline]

So it looks to be that this has started failing from btrfs-progs release v5.14.91 onwards. Version 5.14.2 is the last one which doesn't raise the error. In the CHANGES file for this specific release I cannot identify any specifics on this. There aren't any additional release notes for this tag in GitHub.

In the snap we currently use the btrfs-progs package installed via apt using version 5.16.2. If we would go back to version 5.14.2 we could fix the problem for new storage pools but already existing storage pools will continue to fail on the --refresh operation. The only change is that we cannot see the segfault message anymore but btrfs receive still returns ERROR: clone: did not find source subvol.

Any ideas why our main test suite doesn't experience this? Is it because its using a busybox image rather than ubuntu?

[tomponline]

Hopefully it'll be fixed when we switch to core24