[ovs-discuss] Double free in recent kernels after memleak fix

[Joel Fernandes]

On Fri, Aug 7, 2020 at 4:47 PM Joel Fernandes <joel at joelfernandes.org> wrote:
>
> Hi,
> Adding more of us working on RCU as well. Johan from another team at
> Google discovered a likely issue in openswitch, details below:
>
> On Fri, Aug 7, 2020 at 11:32 AM Johan Knöös <jknoos at google.com> wrote:
> >
> > On Tue, Aug 4, 2020 at 8:52 AM Gregory Rose <gvrose8192 at gmail.com> wrote:
> > >
> > >
> > >
> > > On 8/3/2020 12:01 PM, Johan Knöös via discuss wrote:
> > > > Hi Open vSwitch contributors,
> > > >
> > > > We have found openvswitch is causing double-freeing of memory. The
> > > > issue was not present in kernel version 5.5.17 but is present in
> > > > 5.6.14 and newer kernels.
> > > >
> > > > After reverting the RCU commits below for debugging, enabling
> > > > slub_debug, lockdep, and KASAN, we see the warnings at the end of this
> > > > email in the kernel log (the last one shows the double-free). When I
> > > > revert 50b0e61b32ee890a75b4377d5fbe770a86d6a4c1 ("net: openvswitch:
> > > > fix possible memleak on destroy flow-table"), the symptoms disappear.
> > > > While I have a reliable way to reproduce the issue, I unfortunately
> > > > don't yet have a process that's amenable to sharing. Please take a
> > > > look.
> > > >
> > > > 189a6883dcf7 rcu: Remove kfree_call_rcu_nobatch()
> > > > 77a40f97030b rcu: Remove kfree_rcu() special casing and lazy-callback handling
> > > > e99637becb2e rcu: Add support for debug_objects debugging for kfree_rcu()
> > > > 0392bebebf26 rcu: Add multiple in-flight batches of kfree_rcu() work
> > > > 569d767087ef rcu: Make kfree_rcu() use a non-atomic ->monitor_todo
> > > > a35d16905efc rcu: Add basic support for kfree_rcu() batching
> > > >
>
> Note that these reverts were only for testing the same code, because
> he was testing 2 different kernel versions. One of them did not have
> this set. So I asked him to revert. There's no known bug in the
> reverted code itself. But somehow these patches do make it harder for
> him to reproduce the issue.

And the reason for this is likely the additional kfree batching is
slowing down the occurrence of the crash.

thanks,

 - Joel