Re: [PATCH] Bluetooth: Fix locking in bt_accept_enqueue() for BH context

Hi,

On Mon, Oct 15, 2018 at 3:39 PM Matthias Kaehlcke [off-list ref] wrote:

With commit e16337622016 ("Bluetooth: Handle bt_accept_enqueue() socket
atomically") lock_sock[_nested]() is used to acquire the socket lock
before manipulating the socket. lock_sock[_nested]() may block, which
is problematic since bt_accept_enqueue() can be called in bottom half
context (e.g. from rfcomm_connect_ind()).

The socket API provides bh_lock_sock[_nested]() to acquire the socket
lock in bottom half context. Check the context in bt_accept_enqueue()
and use the appropriate locking mechanism for the context.

I wonder if it would help to put the stack crawl in the commit message
too?  I think this is what the BUG was reporting (though the stack
seems a bet shorter than I'd expect due to compiler inlining)

  __might_sleep+0x4c/0x80
  lock_sock_nested+0x24/0x58
  bt_accept_enqueue+0x48/0xd4 [bluetooth]
  rfcomm_connect_ind+0x190/0x218 [rfcomm]
  rfcomm_run+0xe3c/0x163c [rfcomm]

Fixes: e16337622016 ("Bluetooth: Handle bt_accept_enqueue() socket atomically")
Signed-off-by: Matthias Kaehlcke <mka@chromium.org>
---
Not sure if this is the correct solution, it's certainly not elegant and
checkpatch.pl complains that in_atomic() shouldn't be used outside of
core kernel code. I'm open to other suggestions :)

I'm a total noob when it comes to Bluetooth, so I guess I'll try to
understand the callchains here.  As far as I can tell there are only 3
calls to bt_accept_enqueue, right?  I guess we can look at each of
them:

--

1. net/bluetooth/l2cap_sock.c:

Parent is locked with lock_sock().

--

2. net/bluetooth/rfcomm/sock.c:

Parent is locked with bh_lock_sock().  This is the case you were seeing.

--

3. net/bluetooth/sco.c

sco_conn_ready() => __sco_chan_add => bt_accept_enqueue()

...parent is locked with bh_lock_sock().

sco_connect() => sco_chan_add() => __sco_chan_add

...parent is NULL so we actually never call bt_accept_enqueue() in
this callchain.

So the net result is that the parent is locked with bh_lock_sock().

--

From looking at the above I guess it's pretty simple--if our parent

was locked with bh_lock_sock() then the child should be locked with
bh_lock_sock_nested().  If the parent was locked with lock_sock() then
the child should be locked with lock_sock_nested().

I wonder if a less controversial solution here (and one that wouldn't
upset checkpatch) is to just add a parameter to bt_accept_enqueue()
like "bh".  Then you'd pass true for "bh" in
"net/bluetooth/rfcomm/sock.c" and "net/bluetooth/sco.c" and false from
"net/bluetooth/l2cap_sock.c"

--

In any case the problem seems serious enough that I'd propose a Revert
of commit e16337622016 ("Bluetooth: Handle bt_accept_enqueue() socket
atomically") unless we can come up with an acceptable solution.


-Doug