Received-SPF: softfail (google.com: domain of transitioning linux-wireless-owner@vger.kernel.org does not designate 23.128.96.19 as permitted sender) client-ip=23.128.96.19;
Date:   Thu, 24 Mar 2022 21:58:59 +0000
From:   William McVicker <willmcvicker@google.com>
To:     Johannes Berg <johannes@sipsolutions.net>,
        linux-wireless@vger.kernel.org
Cc:     Marek Szyprowski <m.szyprowski@samsung.com>,
        Kalle Valo <kvalo@codeaurora.org>,
        "David S. Miller" <davem@davemloft.net>,
        Jakub Kicinski <kuba@kernel.org>, netdev@vger.kernel.org,
        Amitkumar Karwar <amitkarwar@gmail.com>,
        Ganapathi Bhat <ganapathi.bhat@nxp.com>,
        Xinming Hu <huxinming820@gmail.com>, kernel-team@android.com,
        Paolo Abeni <pabeni@redhat.com>
Subject: Re: [BUG] deadlock in nl80211_vendor_cmd
Message-ID: <Yjzpo3TfZxtKPMAG@google.com>
References: <0000000000009e9b7105da6d1779@google.com>
 <99eda6d1dad3ff49435b74e539488091642b10a8.camel@sipsolutions.net>
 <5d5cf050-7de0-7bad-2407-276970222635@quicinc.com>
 <YjpGlRvcg72zNo8s@google.com>
 <dc556455-51a2-06e8-8ec5-b807c2901b7e@quicinc.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <dc556455-51a2-06e8-8ec5-b807c2901b7e@quicinc.com>
Precedence: bulk

On 03/23/2022, Jeff Johnson wrote:
> On 3/22/2022 2:58 PM, William McVicker wrote:
> > On 03/22/2022, Jeff Johnson wrote:
> > > On 3/21/2022 1:07 PM, Johannes Berg wrote:
> > > [..snip..]
> > > 
> > > > > I'm not an networking expert. So my main question is if I'm allowed to take
> > > > > the RTNL lock inside the nl80211_vendor_cmd callbacks?
> > > > 
> > > > Evidently, you're not. It's interesting though, it used to be that we
> > > > called these with the RTNL held, now we don't, and the driver you're
> > > > using somehow "got fixed" to take it, but whoever fixed it didn't take
> > > > into account that this is not possible?
> > > 
> > > On this point I just want to remind that prior to the locking change that a
> > > driver would specify on a per-vendor command basis whether or not it wanted
> > > the rtnl_lock to be held via NL80211_FLAG_NEED_RTNL. I'm guessing for the
> > > command in question the driver did not set this flag since the driver wanted
> > > to explicitly take the lock itself, otherwise it would have deadlocked on
> > > itself with the 5.10 kernel.
> > > 
> > > /jeff
> > 
> > On the 5.10 kernel, the core kernel sets NL80211_FLAG_NEED_RTNL as part of
> > the internal_flags for NL80211_CMD_VENDOR:
> > 
> > net/wireless/nl80211.c:
> >     {
> >        .cmd = NL80211_CMD_VENDOR,
> >        .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
> >        .doit = nl80211_vendor_cmd,
> >        .dumpit = nl80211_vendor_cmd_dump,
> >        .flags = GENL_UNS_ADMIN_PERM,
> >        .internal_flags = NL80211_FLAG_NEED_WIPHY |
> >                NL80211_FLAG_NEED_RTNL |
> >                NL80211_FLAG_CLEAR_SKB,
> >     },
> > 
> > So the 5.10 version of this driver doesn't need to directly call rtnl_lock()
> > within the vendor command doit() functions since pre_doit() handles the RTNL
> > locking.
> > 
> > It would be nice if nl80211_vendor_cmd() could support taking the RTNL lock if
> > requested via the vendor flags. That would require moving the wiphy lock to
> > nl80211_vendor_cmds() so that it could take the RTNL and wiphy lock in the
> > correct order. Is that something you'd be open to Johannes?
> > 
> > --Will
> 
> Thanks for correcting my understanding. I concur that it would be useful for
> vendor commands to be able to specify that a given command needs the RTNL
> lock to be held.
> 
> 

Hi Johannes,

I found that we can hit this same ABBA deadlock within the nl80211 code
before ever even calling into the vendor doit() function. The issue I found
is caused by the way we unlock the RTNL mutex. Here is the call flow that
leads to the deadlock:

Thread 1                         Thread 2
 nl80211_pre_doit():
   rtnl_lock()
   wiphy_lock()                   nl80211_pre_doit():
                                    rtnl_lock() // blocked by Thread 1
   rtnl_unlock():
     netdev_run_todo():
       __rtnl_unlock()
                                    <got RTNL lock>
                                    wiphy_lock() // blocked by Thread 1
       rtnl_lock(); // DEADLOCK
 doit()
 nl80211_post_doit():
   wiphy_unlock();

Basically, unlocking the RTNL within netdev_run_todo() gives another thread
that is waiting for the RTNL in nl80211_pre_doit() a chance to grab the
RTNL lock leading to the deadlock. I found that there are multiple
instances where rtnl_lock() is called within netdev_run_todo(): a couple of
times inside netdev_wait_allrefs() and directly by netdev_run_todo().

Since I'm not really familiar with all the RNTL locking requirements, I was
hoping you could take a look at netdev_run_todo() to see if it's possible
to refactor it to avoid this deadlock. If not, then I don't think we can
call rtnl_unlock() while still holding the wiphy mutex.

Thanks,
Will