Hi Luiz,
On 09.02.24 14:36, Luiz Augusto von Dentz wrote:
Hi Jonas,
On Fri, Feb 9, 2024 at 7:37 AM Jonas Dreßler <verdre@xxxxxxx> wrote:
Hi everyone!
On 08.02.24 16:32, syzbot wrote:
syzbot has bisected this issue to:
commit 456561ba8e495e9320c1f304bf1cd3d1043cbe7b
Author: Jonas Dreßler <verdre@xxxxxxx>
Date: Tue Feb 6 11:08:13 2024 +0000
Bluetooth: hci_conn: Only do ACL connections sequentially
bisection log: https://syzkaller.appspot.com/x/bisect.txt?x=154f8550180000
start commit: b1d3a0e70c38 Add linux-next specific files for 20240208
git tree: linux-next
final oops: https://syzkaller.appspot.com/x/report.txt?x=174f8550180000
console output: https://syzkaller.appspot.com/x/log.txt?x=134f8550180000
kernel config: https://syzkaller.appspot.com/x/.config?x=bb693ba195662a06
dashboard link: https://syzkaller.appspot.com/bug?extid=3f0a39be7a2035700868
syz repro: https://syzkaller.appspot.com/x/repro.syz?x=11d95147e80000
C reproducer: https://syzkaller.appspot.com/x/repro.c?x=107c2d8fe80000
Reported-by: syzbot+3f0a39be7a2035700868@xxxxxxxxxxxxxxxxxxxxxxxxx
Fixes: 456561ba8e49 ("Bluetooth: hci_conn: Only do ACL connections sequentially")
For information about bisection process see: https://goo.gl/tpsmEJ#bisection
Hmm, looking at the backtraces, I think the issue that the introduction of the
sequential connect has introduced another async case: In hci_connect_acl(), when
we call hci_acl_create_connection_sync(), the conn state no longer immediately
gets set to BT_CONNECT, but remains in BT_OPEN or BT_CLOSED until the hci_sync
queue actually executes __hci_acl_create_connection_sync().
Need to double check but I think we do set BT_CONNECT in case of LE
when it is queued so which shall prevent it to be queued multiple
times.
Nope, we set it only from within the hci_sync callback, see
hci_connect_le_sync(). IMO that makes sense, because in
hci_abort_conn_sync(), we send a HCI_OP_CREATE_CONN_CANCEL in case
of conn->state == BT_CONNECT, and this OP we wouldn't want to send
while the command is still waiting in the queue.
This means that now hci_connect_acl() is happy to do multiple
hci_acl_create_connection_sync calls, and the hci_sync machinery will happily
execute them right after each other. Then the newly introduced hci_abort_conn_sync()
in __hci_acl_create_connection_sync() calls hci_conn_del() and frees the conn
object, so the second time we enter __hci_acl_create_connection_sync(),
things blow up.
It looks to me like in theory the hci_connect_le_sync() logic is prone to a
similar issue, but in practice that's prohibited because in hci_connect_le_sync()
we lookup whether the conn object still exists and bail out if it doesn't.
Even for LE though I think we can queue multiple hci_connect_le_sync() calls
and those will happily send HCI_OP_LE_CREATE_CONN no matter what the connection
state actually is?
So assuming this analysis is correct, what do we do to fix this? It seems to me
that
1) we want a BT_CONNECT_QUEUED state for connections, so that the state
machine covers this additional stage that we have for ACL and LE connections now.
BT_CONNECT already indicates that connection procedure is in progress.
I still think an additional state is necessary. Alternatively (and maybe
a lot nicer than the extra state) would be to add some functions to hci_sync
to check for and remove queued/ongoing commands, I'm thinking of a new
bool hci_cmd_sync_has(struct hci_dev *hdev, hci_cmd_sync_work_func_t func, void *data);
void hci_cmd_sync_cancel(struct hci_dev *hdev, hci_cmd_sync_work_func_t func, void *data);
I think if we had those, in addition to not needing the additional state,
we could also simplify the hci_abort_conn() code quite a bit and possibly
get rid of the passing of connection handles to hci_connect_le_sync().
I'll give that a try and propose a small patch tomorrow.
Cheers,
Jonas
2) the conn object can still disappear while the __hci_acl_create_connection_sync()
is queued, so we need something like the "if conn doesn't exist anymore, bail out"
check from hci_connect_le_sync() in __hci_acl_create_connection_sync(), too.
Btw, I'd probably clean up the connect function and create something
like hci_connect/hci_connect_sync which takes care of the details
internally like it was done to abort.
That said, the current check in hci_connect_le_sync() that's using the connection
handle to lookup the conn does not seem great, aren't these handles re-used
after connections are torn down?
Well we could perhaps do a lookup by pointer to see if the connection
hasn't been removed in the meantime, that said to force a clash on the
handles it need to happen in between abort, which frees the handle,
and connect, anyway the real culprit here is that we should be able to
abort the cmd_sync callback like we do in LE:
https://github.com/bluez/bluetooth-next/blob/master/net/bluetooth/hci_conn.c#L2943
That way we stop the connect callback to run and don't have to worry
about handle re-use.
Cheers,
Jonas
