Date: Tue, 29 Mar 2016 07:05:18 -0500
From: Chris J Arges <chris.j.arges@canonical.com>
To: jeyu@redhat.com, jpoimboe@redhat.com, eugene.shatokhin@rosalab.ru
Cc: live-patching@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Bug with paravirt ops and livepatches
Message-ID: <20160329120518.GA21252@canonical.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
User-Agent: Mutt/1.5.24 (2015-08-30)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1154
Lines: 25

Paravirtualized ops and livepatching currently don't mix very well and can
cause undefined behavor such as oops, invalid opcodes or corrupted stacks.
The original discussion of this issue can be found here [1].

I've written an example livepatch module that reproduces the issue [2].
In order to trigger the issue you must first insert the module then trigger
the paravirt ops by starting a VM.

In the thread here [1] a couple of solutions have been proposed:

1) Jessica proposed using the Arch-independent patchset ensure that livepatch
finishes writing its relas before apply_paravirt() is called. However, this
introduces a bit more arch-dependent code. It would be useful to see if other
arches are affected by this as well.

2) Eugene proposed skipping application of the rela if the instruction to be
relocated has already been changed. This passes the initial example [2];
however its unclear if/how this will break things.

It may be good to weigh in here and get more eyes on this.
Thanks,
--chris

[1]: https://github.com/dynup/kpatch/issues/580
[2]: http://people.canonical.com/~arges/livepatch_issue/livepatch_kvm_arch_vm_ioctl/livepatch.c