Date:   Mon, 15 Apr 2019 13:53:53 +0200
From:   Stanislaw Gruszka <sgruszka@redhat.com>
To:     Lorenzo Bianconi <lorenzo@kernel.org>
Cc:     Lorenzo Bianconi <lorenzo.bianconi@redhat.com>, nbd@nbd.name,
        linux-wireless@vger.kernel.org
Subject: Re: [PATCH] mt76: usb: fix possible memory leak during suspend/resume
Message-ID: <20190415115352.GA4143@redhat.com>
References: <cover.1555071870.git.lorenzo@kernel.org>
 <bdd1046beaee6104b1fcd25b1069f25db8f2fb20.1555071870.git.lorenzo@kernel.org>
 <20190412145442.GA2539@redhat.com>
 <20190412153509.GB3156@localhost.localdomain>
 <20190412162746.GC3156@localhost.localdomain>
 <20190413083050.GA7434@redhat.com>
 <20190413101056.GA7940@localhost.localdomain>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20190413101056.GA7940@localhost.localdomain>
User-Agent: Mutt/1.11.3 (2019-02-01)
Sender: linux-wireless-owner@vger.kernel.org
Precedence: bulk

On Sat, Apr 13, 2019 at 12:10:59PM +0200, Lorenzo Bianconi wrote:
> > On Fri, Apr 12, 2019 at 06:27:48PM +0200, Lorenzo Bianconi wrote:
> > > > > On Fri, Apr 12, 2019 at 02:27:16PM +0200, Lorenzo Bianconi wrote:
> > > > > > Disable mt76u_tx_tasklet at the end of mt76u_stop_queues in order to
> > > > > > properly deallocate all pending skbs during suspend/resume phase
> > > > > 
> > > > > On suspend/resume tx skb's are processed after tasklet_enable()
> > > > > in resume callback. There is issue with device removal though
> > > > > (during suspend or otherwise).
> > > > 
> > > > Hi Stanislaw,
> > > > 
> > > > I guess the right moment to deallocate the skbs is during suspend since resume
> > > > can happen in very far future
> > 
> > Yes, it's better to free on suspend, but in practice does not really matter since
> > system is disabled till resume.
> > 
> > > > > > Fixes: b40b15e1521f ("mt76: add usb support to mt76 layer")
> > > > > > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > > > > > ---
> > > > > >  drivers/net/wireless/mediatek/mt76/usb.c | 4 ++--
> > > > > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > > > > 
> > > > > > diff --git a/drivers/net/wireless/mediatek/mt76/usb.c b/drivers/net/wireless/mediatek/mt76/usb.c
> > > > > > index a3acc070063a..575207133775 100644
> > > > > > --- a/drivers/net/wireless/mediatek/mt76/usb.c
> > > > > > +++ b/drivers/net/wireless/mediatek/mt76/usb.c
> > > > > > @@ -842,10 +842,10 @@ static void mt76u_stop_tx(struct mt76_dev *dev)
> > > > > >  void mt76u_stop_queues(struct mt76_dev *dev)
> > > > > >  {
> > > > > >  	tasklet_disable(&dev->usb.rx_tasklet);
> > > > > > -	tasklet_disable(&dev->usb.tx_tasklet);
> > > > > > -
> > > > > >  	mt76u_stop_rx(dev);
> > > > > > +
> > > > > >  	mt76u_stop_tx(dev);
> > > > > > +	tasklet_disable(&dev->usb.tx_tasklet);
> > > > > 
> > > > > If tasklet is scheduled and we disable it and never enable, we end up
> > > > > with infinite loop in tasklet_action_common(). This patch make the
> > > > > problem less reproducible since tasklet_disable() is moved after
> > > > > usb_kill_urb() -> tasklet_schedule(), but it is still possible.
> > > > 
> > > > I can see the point here. Maybe we can just run tasklet_kill instead of
> > > > tasklet_disable here (at least for tx one)
> > 
> > I think you have right as tasklet_kill() will wait for scheduled tasklet .
> > Originally in my patch (see below) I used wait_event as I thought
> > tasklet_kill() may prevent scheduled tasklet to be executed (hence cause
> > leak) but that seems to be not true.
> 
> I agree with rx side (good catch!!), but on tx one I guess usb_kill_urb()
> is already waiting for tx pending so we just need to use tasklet_kill
> at the end of mt76u_stop_queues, in this way we will free pending skbs during
> suspend

I looked more into that and there are some issues with this approach.
tx_tasklet do mt76_txq_schedule() which can queue tx frames. Also we
do not free skb's that require status check and dev->usb.stat_work 
is already (correctly) stopped on mac80211.stop. 

I'll use wait_event(dev->tx_wait) on mac80211 stop to handle those
issues correctly.

Stanislaw