IB/mthca: Fix potential AB-BA deadlock with CQ locks
authorRoland Dreier <rolandd@cisco.com>
Fri, 11 Aug 2006 15:56:57 +0000 (08:56 -0700)
committerRoland Dreier <rolandd@cisco.com>
Fri, 11 Aug 2006 15:56:57 +0000 (08:56 -0700)
commita19aa5c5fdda8b556ab238177ee27c5ef7873c94
tree6c770c8fbbe3270bf1416c3edd8894d214e62657
parente54b82d739d4a2ef992976c8c0692cdf89286420
IB/mthca: Fix potential AB-BA deadlock with CQ locks

When destroying a QP, mthca locks both the QP's send CQ and receive
CQ.  However, the following scenario is perfectly valid:

    QP_a: send_cq == CQ_x, recv_cq == CQ_y
    QP_b: send_cq == CQ_y, recv_cq == CQ_x

The old mthca code simply locked send_cq and then recv_cq, which in
this case could lead to an AB-BA deadlock if QP_a and QP_b were
destroyed simultaneously.

We can fix this by changing the locking code to lock the CQ with the
lower CQ number first, which will create a consistent lock ordering.
Also, the second CQ is locked with spin_lock_nested() to tell lockdep
that we know what we're doing with the lock nesting.

This bug was found by lockdep.

Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/mthca/mthca_provider.h
drivers/infiniband/hw/mthca/mthca_qp.c