Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Sat, 06 Feb 2021 09:32:57 +0100
Message-ID: <s5hk0rl1tqe.wl-tiwai@suse.de>
From:   Takashi Iwai <tiwai@suse.de>
To:     Hillf Danton <hdanton@sina.com>
Cc:     Mikhail Gavrilov <mikhail.v.gavrilov@gmail.com>, zonque@gmail.com,
        LKML <linux-kernel@vger.kernel.org>, alsa-devel@alsa-project.org,
        linux-usb@vger.kernel.org
Subject: Re: BUG: KASAN: use-after-free in snd_complete_urb+0x109e/0x1740 [snd_usb_audio] (5.11-rc6)
In-Reply-To: <20210206081333.7472-1-hdanton@sina.com>
References: <20210206081333.7472-1-hdanton@sina.com>
User-Agent: Wanderlust/2.15.9 (Almost Unreal) SEMI/1.14.6 (Maruoka)
 FLIM/1.14.9 (=?UTF-8?B?R29qxY0=?=) APEL/10.8 Emacs/25.3
 (x86_64-suse-linux-gnu) MULE/6.0 (HANACHIRUSATO)
MIME-Version: 1.0 (generated by SEMI 1.14.6 - "Maruoka")
Content-Type: text/plain; charset=US-ASCII
Precedence: bulk

On Sat, 06 Feb 2021 09:13:33 +0100,
Hillf Danton wrote:
> 
> On Sat, 6 Feb 2021 Takashi Iwai wrote:
> > > Due to the reconnecting key word mentioned, no fix to
> > > d0f09d1e4a88 ("ALSA: usb-audio: Refactoring endpoint URB deactivation")
> > > will be added.
> > > 
> > > What is added is to capture EP_FLAG_STOPPING and remove the one
> > > second wait limit if the reconnecting acts may make it easier to
> > > repro the uaf. The diff is only for idea show.
> > 
> > If my understanding is right, this won't change.  The problem is
> > rather the lack of this function call itself, i.e. the missing
> > synchronization for the stream stop.
> 
> Thanks for taking a look at it.
> > 
> > It worked casually in the past because the endpoint resource is
> > released at a later point that is after all streams are really closed.
> > Now it's released earlier and hitting the UAF.
> 
> And add it if I dont misread you.
> 
> Hillf
> 
> --- a/sound/usb/endpoint.c
> +++ b/sound/usb/endpoint.c
> @@ -832,24 +832,14 @@ void snd_usb_endpoint_suspend(struct snd
>   */
>  static int wait_clear_urbs(struct snd_usb_endpoint *ep)
>  {
> -	unsigned long end_time = jiffies + msecs_to_jiffies(1000);
> -	int alive;
> -
> -	if (!test_bit(EP_FLAG_STOPPING, &ep->flags))
> -		return 0;
> -
> +	WARN_ON_ONCE(!test_bit(EP_FLAG_STOPPING, &ep->flags));

EP_FLAG_STOPPING is normally zero, hence putting a WARN_ON() shows
just a false-positive.

>  	do {
> -		alive = bitmap_weight(&ep->active_mask, ep->nurbs);
> -		if (!alive)
> +		if (!bitmap_weight(&ep->active_mask, ep->nurbs))
>  			break;
>  
>  		schedule_timeout_uninterruptible(1);
> -	} while (time_before(jiffies, end_time));
> +	} while (1);
>  
> -	if (alive)
> -		usb_audio_err(ep->chip,
> -			"timeout: still %d active urbs on EP #%x\n",
> -			alive, ep->ep_num);

The original report didn't show the error, so it's not about the
timeout.  You're likely scratching a wrong surface, I'm afraid.

>  	clear_bit(EP_FLAG_STOPPING, &ep->flags);
>  
>  	ep->sync_sink = NULL;
> @@ -1433,7 +1423,7 @@ void snd_usb_endpoint_stop(struct snd_us
>  		WRITE_ONCE(ep->sync_source->sync_sink, NULL);
>  
>  	if (!atomic_dec_return(&ep->running))
> -		stop_and_unlink_urbs(ep, false, false);
> +		stop_and_unlink_urbs(ep, false, true);

Please don't.  This will lead to a sleep-in-atomic Oops.


Takashi