When stopping BCSP/H5, stop the retransmission timer before proceeding
to clean up packet queues. The previous code had a race condition where
the timer could trigger after the packet lists and protocol structure
had been removed which lead to dereferencing NULL or use-after-free bugs.
Signed-off-by: Michael Knudsen <[email protected]>
Reported-by: Kirill Tkhai <[email protected]>
---
drivers/bluetooth/hci_bcsp.c | 4 +++-
drivers/bluetooth/hci_h5.c | 4 ++--
2 files changed, 5 insertions(+), 3 deletions(-)
diff --git a/drivers/bluetooth/hci_bcsp.c b/drivers/bluetooth/hci_bcsp.c
index 0bc87f7..eee2fb2 100644
--- a/drivers/bluetooth/hci_bcsp.c
+++ b/drivers/bluetooth/hci_bcsp.c
@@ -715,6 +715,9 @@ static int bcsp_open(struct hci_uart *hu)
static int bcsp_close(struct hci_uart *hu)
{
struct bcsp_struct *bcsp = hu->priv;
+
+ del_timer_sync(&bcsp->tbcsp);
+
hu->priv = NULL;
BT_DBG("hu %p", hu);
@@ -722,7 +725,6 @@ static int bcsp_close(struct hci_uart *hu)
skb_queue_purge(&bcsp->unack);
skb_queue_purge(&bcsp->rel);
skb_queue_purge(&bcsp->unrel);
- del_timer(&bcsp->tbcsp);
kfree(bcsp);
return 0;
diff --git a/drivers/bluetooth/hci_h5.c b/drivers/bluetooth/hci_h5.c
index f6f4974..afd759e 100644
--- a/drivers/bluetooth/hci_h5.c
+++ b/drivers/bluetooth/hci_h5.c
@@ -206,12 +206,12 @@ static int h5_close(struct hci_uart *hu)
{
struct h5 *h5 = hu->priv;
+ del_timer_sync(&h5->timer);
+
skb_queue_purge(&h5->unack);
skb_queue_purge(&h5->rel);
skb_queue_purge(&h5->unrel);
- del_timer(&h5->timer);
-
kfree(h5);
return 0;
--
1.7.9.5
Hi Michael,
On Tue, Mar 04, 2014, Michael Knudsen wrote:
> On 03/04/2014 08:14 AM, Johan Hedberg wrote:
> >>>When stopping BCSP/H5, stop the retransmission timer before proceeding
> >>>to clean up packet queues. The previous code had a race condition where
> >>>the timer could trigger after the packet lists and protocol structure
> >>>had been removed which lead to dereferencing NULL or use-after-free bugs.
> >>
> >>No interest?
> >
> >I was just discussing this yesterday with Marcel (that we seem to have
> >forgotten about this patch). The only concern is whether it's safe to
> >sleep in the *_close callbacks (since you use del_timer_sync). Have you
> >verified that this doesn't cause any issues?
>
> Our internal testing was reliably triggering the crash before and
> has been stable since our fix went into local trees. I expected
> sleeping to be fine since the path is in process context, and I
> found that slip uses the same approach:
>
> drivers/net/slip/slip.c:slip_close()
>
> I think I saw other line disciplines that did the same but I don't
> recall which ones.
>
> Btw. if this is committed, there is a small typo in the message I
> used: Instead of ``lead'' it should be ``led''.
Thanks for the confirmation. I've now pushed the patch to the
bluetooth-next tree (with the typo fixed too).
Johan
On 03/04/2014 08:14 AM, Johan Hedberg wrote:
>>> When stopping BCSP/H5, stop the retransmission timer before proceeding
>>> to clean up packet queues. The previous code had a race condition where
>>> the timer could trigger after the packet lists and protocol structure
>>> had been removed which lead to dereferencing NULL or use-after-free bugs.
>>
>> No interest?
>
> I was just discussing this yesterday with Marcel (that we seem to have
> forgotten about this patch). The only concern is whether it's safe to
> sleep in the *_close callbacks (since you use del_timer_sync). Have you
> verified that this doesn't cause any issues?
Our internal testing was reliably triggering the crash before and
has been stable since our fix went into local trees. I expected
sleeping to be fine since the path is in process context, and I
found that slip uses the same approach:
drivers/net/slip/slip.c:slip_close()
I think I saw other line disciplines that did the same but I don't
recall which ones.
Btw. if this is committed, there is a small typo in the message I
used: Instead of ``lead'' it should be ``led''.
-m.
Hi Michael,
On Tue, Mar 04, 2014, Michael Knudsen wrote:
> On 02/18/2014 09:48 AM, Michael Knudsen wrote:
> >When stopping BCSP/H5, stop the retransmission timer before proceeding
> >to clean up packet queues. The previous code had a race condition where
> >the timer could trigger after the packet lists and protocol structure
> >had been removed which lead to dereferencing NULL or use-after-free bugs.
>
> No interest?
I was just discussing this yesterday with Marcel (that we seem to have
forgotten about this patch). The only concern is whether it's safe to
sleep in the *_close callbacks (since you use del_timer_sync). Have you
verified that this doesn't cause any issues?
Johan
On 02/18/2014 09:48 AM, Michael Knudsen wrote:
> When stopping BCSP/H5, stop the retransmission timer before proceeding
> to clean up packet queues. The previous code had a race condition where
> the timer could trigger after the packet lists and protocol structure
> had been removed which lead to dereferencing NULL or use-after-free bugs.
No interest?
-m.