RDMA/cma: Fix device removal race

The race is as follows:

A process : cma_process_remove() calls cma_remove_id_dev(),
	    which sets id state to CMA_DEVICE_REMOVAL and
	    calls wait_event(dev_remove).

B process : cma_req_handler() had incremented dev_remove,
	    and calls cma_acquire_ib_dev() and on failure
	    calls cma_release_remove(), which does a
	    wake_up of cma_process_remove(). Then
	    cma_req_handler() calls rdma_destroy_id();

A Process : cma_remove_id_dev() gets woken and checks the
	    state of id, and since it is still (wrongly)
	    CMA_DEVICE_REMOVAL, it calls notify_user(id)
	    and if that fails, the caller - cma_process_remove()
	    calls rdma_destroy_id(id). Two processes can
	    call rdma_destroy_id(), resulting in one
	    de-referencing kfreed id_priv.

Fix is for process B to set CMA_DESTROYING in cma_req_handler()
so that process A will return instead of doing a rdma_destroy_id().

Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com>
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index 69bb089..f383a4f 100644 (file)
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -932,6 +932,7 @@ static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
        mutex_unlock(&lock);
        if (ret) {
                ret = -ENODEV;
+               cma_exch(conn_id, CMA_DESTROYING);
                cma_release_remove(conn_id);
                rdma_destroy_id(&conn_id->id);
                goto out;