Re: Memory leak in ath9k_hif_usb_dealloc_tx_urbs()

From: Pavel Skripkin
Date: Tue Apr 27 2021 - 08:04:36 EST


Hi!

On Tue, 2021-04-27 at 17:05 +0530, Atul Gopinathan wrote:
> On Wed, Mar 31, 2021 at 08:28:15AM +0200, Greg KH wrote:
> > On Tue, Mar 30, 2021 at 10:36:52PM +0300, Pavel Skripkin wrote:
> > > Hi!
> > >
> > > I did some debugging on this
> > > https://syzkaller.appspot.com/bug?id=3ea507fb3c47426497b52bd82b8ef0dd5b6cc7ee
> > > and, I believe, I recognized the problem. The problem appears in
> > > case of
> > > ath9k_htc_hw_init() fail. In case of this fail all tx_buf->urb
> > > krefs will be
> > > initialized to 1, but in free function:
> > >
> > > static void ath9k_hif_usb_dealloc_tx_urbs(struct hif_device_usb
> > > *hif_dev)
> > >
> > > ....
> > >
> > > static void ath9k_hif_usb_dealloc_tx_urbs(struct hif_device_usb
> > > *hif_dev)
> > > {
> > >     ...
> > >         list_for_each_entry_safe(tx_buf, tx_buf_tmp,
> > >                                  &hif_dev->tx.tx_buf, list) {
> > >                 usb_get_urb(tx_buf->urb);
> > >                 ...
> > >                 usb_free_urb(tx_buf->urb);
> > >                 ...
> > >                 }
> > >
> > > Krefs are incremented and then decremented, that means urbs won't
> > > be freed.
> > > I found your patch and I can't properly understand why You added
> > > usb_get_urb(tx_buf->urb).
> > > Can You explain please, I believe this will help me or somebody
> > > to fix this ussue :)
> >
> > I think almost everyone who has looked into this has given up due
> > to the
> > mess of twisty-passages here with almost no real-world benefits for
> > unwinding them :)
>
> Just wanted to confirm, what is the status of this bug then, as in is
> it
> invalid (not sure if that's the correct word)? I happened to stumble
> across the same syzkaller bug report Pavel posted above, in the
> morning.
> Saw that there has been no patch tests/fixes on this yet according to
> syzkaller. Spent a couple of hours going through it before sending a
> test patch to syzbot which returned an "OK" (and the patch is exactly
> what Pavel pointed out, I simply removed the `usb_get_urb()`). Before
> sending anything to the mailing list, I made sure to search all the
> relavant networking lists to see if this topic had been brought up
> (learnt
> to do this from my preious mistakes of sending already accepted
> patches) and
> luckily I found this.
>
> Syzbot has had 380 crashes caused by this bug, with the latest being
> today. So I wanted to confirm what should be done be about this bug.
>

I saw on dashboard, that Dmitry tested latest upstream commit and
syzbot returned "OK", but usb_get_urb(tx_buf->urb); is still there.

I think, this usb_get_urb prevents race condition, but I'm not sure
about it, that's why I sent an email to patch author. As You can see,
he has not responded yet :)

> Thank you!
> Atul

With regards,
Pavel Skripkin