Re: [PATCH 2/9] ALSA: usb-audio: Fix possible race at sync of urb completions

[Takashi Iwai <tiwai@suse.de>]

On Fri, 18 Apr 2025 16:45:17 +0200,
Hillf Danton wrote:
> 
> On Fri, 18 Apr 2025 13:08:32 +0200 Takashi Iwai wrote:
> > On Fri, 18 Apr 2025 12:35:32 +0200 Hillf Danton wrote:
> > > On Wed, 29 Sep 2021 10:08:37 +0200 Takashi Iwai wrote:
> > > > USB-audio driver tries to sync with the clear of all pending URBs in
> > > > wait_clear_urbs(), and it waits for all bits in active_mask getting
> > > > cleared.  This works fine for the normal operations, but when a stream
> > > > is managed in the implicit feedback mode, there is still a very thin
> > > > race window: namely, in snd_complete_usb(), the active_mask bit for
> > > > the current URB is once cleared before re-submitted in
> > > > queue_pending_output_urbs().  If wait_clear_urbs() is called during
> > > > that period, it may pass the test and go forward even though there may
> > > > be a still pending URB.
> > > > 
> > > > For covering it, this patch adds a new counter to each endpoint to
> > > > keep the number of in-flight URBs, and changes wait_clear_urbs()
> > > > checking this number instead.  The counter is decremented at the end
> > > > of URB complete, hence the reference is kept as long as the URB
> > > > complete is in process.
> > > > 
> > > > Signed-off-by: Takashi Iwai <tiwai@suse.de>
> > > > ---
> > > >  sound/usb/card.h     | 1 +
> > > >  sound/usb/endpoint.c | 7 ++++++-
> > > >  2 files changed, 7 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/sound/usb/card.h b/sound/usb/card.h
> > > > index 3329ce710cb9..746a765b2437 100644
> > > > --- a/sound/usb/card.h
> > > > +++ b/sound/usb/card.h
> > > > @@ -97,6 +97,7 @@ struct snd_usb_endpoint {
> > > >  	unsigned int nominal_queue_size; /* total buffer sizes in URBs */
> > > >  	unsigned long active_mask;	/* bitmask of active urbs */
> > > >  	unsigned long unlink_mask;	/* bitmask of unlinked urbs */
> > > > +	atomic_t submitted_urbs;	/* currently submitted urbs */
> > > >  	char *syncbuf;			/* sync buffer for all sync URBs */
> > > >  	dma_addr_t sync_dma;		/* DMA address of syncbuf */
> > > >  
> > > > diff --git a/sound/usb/endpoint.c b/sound/usb/endpoint.c
> > > > index 29c4865966f5..06241568abf7 100644
> > > > --- a/sound/usb/endpoint.c
> > > > +++ b/sound/usb/endpoint.c
> > > > @@ -451,6 +451,7 @@ static void queue_pending_output_urbs(struct snd_usb_endpoint *ep)
> > > >  		}
> > > >  
> > > >  		set_bit(ctx->index, &ep->active_mask);
> > > > +		atomic_inc(&ep->submitted_urbs);
> > > >  	}
> > > >  }
> > > >  
> > > > @@ -488,6 +489,7 @@ static void snd_complete_urb(struct urb *urb)
> > > >  			clear_bit(ctx->index, &ep->active_mask);
> > > >  			spin_unlock_irqrestore(&ep->lock, flags);
> > > >  			queue_pending_output_urbs(ep);
> > > 
> > > 			smp_mb();
> > > 
> > > > +			atomic_dec(&ep->submitted_urbs); /* decrement at last */
> > > 
> > > Does it match the comment to add a mb?
> > 
> > How...?  I don't understand your intention.
> > 
> In addition to the UAF report [1], I saw a customer report of list
> corruption of linux-6.1.99 on arm64 this week without reproducer.
> 
> 	list corruption
> 	list_add_tail();
> 	push_back_to_ready_list();
> 	snd_complete_urb();
> 
> And after another look at this patch I wonder if the race can not be
> erased without the certainty that ep will be no longer used after the
> atomic decrement.

But why adding more barrier if you perform the atomic op...?



Takashi

> 
> [1] https://lore.kernel.org/lkml/CABXGCsOposU1A_HavA_jmtkJMKhDZgh5m1b_YJK1Es5wyE-hZw@mail.gmail.com/