Re: [Bluez PATCH v5] bluetooth: secure bluetooth stack from bluedump attack

[Nathan Chancellor]

Hi Howard,

On Fri, Feb 14, 2020 at 07:16:41PM +0800, Howard Chung wrote:
> Attack scenario:
> 1. A Chromebook (let's call this device A) is paired to a legitimate
>    Bluetooth classic device (e.g. a speaker) (let's call this device
>    B).
> 2. A malicious device (let's call this device C) pretends to be the
>    Bluetooth speaker by using the same BT address.
> 3. If device A is not currently connected to device B, device A will
>    be ready to accept connection from device B in the background
>    (technically, doing Page Scan).
> 4. Therefore, device C can initiate connection to device A
>    (because device A is doing Page Scan) and device A will accept the
>    connection because device A trusts device C's address which is the
>    same as device B's address.
> 5. Device C won't be able to communicate at any high level Bluetooth
>    profile with device A because device A enforces that device C is
>    encrypted with their common Link Key, which device C doesn't have.
>    But device C can initiate pairing with device A with just-works
>    model without requiring user interaction (there is only pairing
>    notification). After pairing, device A now trusts device C with a
>    new different link key, common between device A and C.
> 6. From now on, device A trusts device C, so device C can at anytime
>    connect to device A to do any kind of high-level hijacking, e.g.
>    speaker hijack or mouse/keyboard hijack.
> 
> Since we don't know whether the repairing is legitimate or not,
> leave the decision to user space if all the conditions below are met.
> - the pairing is initialized by peer
> - the authorization method is just-work
> - host already had the link key to the peer
> 
> Signed-off-by: Howard Chung <howardchung@xxxxxxxxxx>
> ---
> 
> Changes in v5:
> - Rephrase the comment
> 
> Changes in v4:
> - optimise the check in smp.c.
> 
> Changes in v3:
> - Change confirm_hint from 2 to 1
> - Fix coding style (declaration order)
> 
> Changes in v2:
> - Remove the HCI_PERMIT_JUST_WORK_REPAIR debugfs option
> - Fix the added code in classic
> - Add a similar fix for LE
> 
>  net/bluetooth/hci_event.c | 10 ++++++++++
>  net/bluetooth/smp.c       | 19 +++++++++++++++++++
>  2 files changed, 29 insertions(+)
> 
> diff --git a/net/bluetooth/hci_event.c b/net/bluetooth/hci_event.c
> index 2c833dae9366..e6982f4f51ea 100644
> --- a/net/bluetooth/hci_event.c
> +++ b/net/bluetooth/hci_event.c
> @@ -4571,6 +4571,16 @@ static void hci_user_confirm_request_evt(struct hci_dev *hdev,
>  			goto confirm;
>  		}
>  
> +		/* If there already exists link key in local host, leave the
> +		 * decision to user space since the remote device could be
> +		 * legitimate or malicious.
> +		 */
> +		if (hci_find_link_key(hdev, &ev->bdaddr)) {
> +			bt_dev_warn(hdev, "Local host already has link key");
> +			confirm_hint = 1;
> +			goto confirm;
> +		}
> +
>  		BT_DBG("Auto-accept of user confirmation with %ums delay",
>  		       hdev->auto_accept_delay);
>  
> diff --git a/net/bluetooth/smp.c b/net/bluetooth/smp.c
> index 2cba6e07c02b..25dbf77d216b 100644
> --- a/net/bluetooth/smp.c
> +++ b/net/bluetooth/smp.c
> @@ -2192,6 +2192,25 @@ static u8 smp_cmd_pairing_random(struct l2cap_conn *conn, struct sk_buff *skb)
>  		smp_send_cmd(conn, SMP_CMD_PAIRING_RANDOM, sizeof(smp->prnd),
>  			     smp->prnd);
>  		SMP_ALLOW_CMD(smp, SMP_CMD_DHKEY_CHECK);
> +
> +		/* Only Just-Works pairing requires extra checks */
> +		if (smp->method != JUST_WORKS)
> +			goto mackey_and_ltk;
> +
> +		/* If there already exists link key in local host, leave the
> +		 * decision to user space since the remote device could be
> +		 * legitimate or malicious.
> +		 */
> +		if (hci_find_ltk(hcon->hdev, &hcon->dst, hcon->dst_type,
> +				 hcon->role)) {
> +			err = mgmt_user_confirm_request(hcon->hdev, &hcon->dst,
> +							hcon->type,
> +							hcon->dst_type, passkey,

We received a report from the 0day bot when building with clang that
passkey is uninitialized when used here:

https://groups.google.com/forum/#!topic/clang-built-linux/kyRKCjRsGoU

It appears to be legitimate as if we get to this point, we have not
touched the value of passkey but I do not know if there is any
contextual code flow going on where this can never happen but I do not
see how.

Would you mind looking into it?

Cheers,
Nathan