RDMA/cxgb3: Don't complete flushed send work requests twice
authorSteve Wise <swise@opengridcomputing.com>
Wed, 29 Apr 2009 22:15:59 +0000 (15:15 -0700)
committerRoland Dreier <rolandd@cisco.com>
Wed, 29 Apr 2009 22:15:59 +0000 (15:15 -0700)
commitec6995ddaa3db2c5f028184ed9b74c0c448c4947
tree1d4635d10ed0a9f1e4d6f1893321d9dadc0b4cf6
parent56a50adda49b2020156616c4eb15353e0f9ad7de
RDMA/cxgb3: Don't complete flushed send work requests twice

When the SQ is flushed, mark the flushed entries as not signaled so
the poll logic doesn't re-insert the CQ entry thinking its an out of
order completion.

The bug can cause the NFS/RDMA server to crash due to processing the
same completed work request twice.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/cxgb3/cxio_hal.c