Skip to contentSkip to conversation

RDMA/rxe: Fix qp error handler

PATCH Applied [email protected] 2 participants 2 messages Jul 30, 2022 – Aug 2, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 RDMA/rxe: Fix qp error handler

From: Zhu Yanjun <[email protected]>

This problem is in this link:
news://nntp.lore.kernel.org:119/[email protected]

this is an error unwind problem.

In the function rxe_create_qp, rxe_qp_from_init is called to initialize qp. rxe_qp_init_req is called by rxe_qp_from_init. If an error occurs before spin_lock_init in rxe_qp_init_req, several spin locks are not initialized. Then rxe_create_qp finally calls rxe_cleanup(qp) to handle errors.

In the end, rxe_qp_do_cleanup is called. In this function, rxe_cleanup_task will call spin_lock_bh. But task->state_lock is not initialized.

As such, an uninitialized spin lock is called by spin_lock_bh.

rxe_create_qp {
        ...
        err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibqp->pd, udata);
        if (err)
                goto qp_init;
        ...
        return 0;

qp_init:

        rxe_cleanup(qp);
        return err;
}

rxe_qp_do_cleanup {
...

  rxe_cleanup_task {
    ...
    spin_lock_bh(&task->state_lock);
    ...

}
}

rxe_qp_from_init {
...

        rxe_qp_init_misc(rxe, qp, init);

        err = rxe_qp_init_req{
                ...
                spin_lock_init(&qp->sq.sq_lock);
                ...
                rxe_init_task{
                  ...
                  spin_lock_init(&task->state_lock);
                  ...
                }
              }
        if (err)
                goto err1;

        err = rxe_qp_init_resp {
                ...
                spin_lock_init(&qp->rq.producer_lock);
                spin_lock_init(&qp->rq.consumer_lock);
                ...
                rxe_init_task {
                  ...
                  spin_lock_init(&task->state_lock);
                  ...
                }
              }

        if (err)
                goto err2;
...
        return 0;

err2:

        ...
err1:
        ...
        return err;
}

About 7 spin locks in qp creation needs to be initialized. Now these spin locks are initialized in the function rxe_qp_init_misc. This will avoid the error "initialize spin locks before use".

Fixes 8700e3e7c485 ("Soft RoCE driver")
Reported-by [email protected]
---
 drivers/infiniband/sw/rxe/rxe_qp.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
drivers/infiniband/sw/rxe/rxe_qp.c +8 -4
index b79e1b43454e..7a223583cf8b 100644
@@ -174,6 +174,14 @@ static void rxe_qp_init_misc(struct rxe_dev *rxe, struct rxe_qp *qp,
 
 	spin_lock_init(&qp->state_lock);
 
+	spin_lock_init(&qp->req.task.state_lock);
+	spin_lock_init(&qp->resp.task.state_lock);
+	spin_lock_init(&qp->comp.task.state_lock);
+
+	spin_lock_init(&qp->sq.sq_lock);
+	spin_lock_init(&qp->rq.producer_lock);
+	spin_lock_init(&qp->rq.consumer_lock);
+
 	atomic_set(&qp->ssn, 0);
 	atomic_set(&qp->skb_out, 0);
 }
@@ -233,7 +241,6 @@ static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
 	qp->req.opcode		= -1;
 	qp->comp.opcode		= -1;
 
-	spin_lock_init(&qp->sq.sq_lock);
 	skb_queue_head_init(&qp->req_pkts);
 
 	rxe_init_task(rxe, &qp->req.task, qp,
@@ -284,9 +291,6 @@ static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
 		}
 	}
 
-	spin_lock_init(&qp->rq.producer_lock);
-	spin_lock_init(&qp->rq.consumer_lock);
-
 	skb_queue_head_init(&qp->resp_pkts);
 
 	rxe_init_task(rxe, &qp->resp.task, qp,
Signature · 1 line
2.27.0
maintainer Applied
#1 lore raw
Quoted 84 lines · On Sun, Jul 31, 2022 at 02:36:21AM -0400, [email protected] wrote:

From: Zhu Yanjun <[email protected]>

This problem is in this link:
news://nntp.lore.kernel.org:119/[email protected]

this is an error unwind problem.

In the function rxe_create_qp, rxe_qp_from_init is called to initialize qp. rxe_qp_init_req is called by rxe_qp_from_init. If an error occurs before spin_lock_init in rxe_qp_init_req, several spin locks are not initialized. Then rxe_create_qp finally calls rxe_cleanup(qp) to handle errors.

In the end, rxe_qp_do_cleanup is called. In this function, rxe_cleanup_task will call spin_lock_bh. But task->state_lock is not initialized.

As such, an uninitialized spin lock is called by spin_lock_bh.

rxe_create_qp {
        ...
        err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibqp->pd, udata);
        if (err)
                goto qp_init;
        ...
        return 0;

qp_init:

        rxe_cleanup(qp);
        return err;
}

rxe_qp_do_cleanup {
...

  rxe_cleanup_task {
    ...
    spin_lock_bh(&task->state_lock);
    ...

}
}

rxe_qp_from_init {
...

        rxe_qp_init_misc(rxe, qp, init);

        err = rxe_qp_init_req{
                ...
                spin_lock_init(&qp->sq.sq_lock);
                ...
                rxe_init_task{
                  ...
                  spin_lock_init(&task->state_lock);
                  ...
                }
              }
        if (err)
                goto err1;

        err = rxe_qp_init_resp {
                ...
                spin_lock_init(&qp->rq.producer_lock);
                spin_lock_init(&qp->rq.consumer_lock);
                ...
                rxe_init_task {
                  ...
                  spin_lock_init(&task->state_lock);
                  ...
                }
              }

        if (err)
                goto err2;
...
        return 0;

err2:

        ...
err1:
        ...
        return err;
}

About 7 spin locks in qp creation needs to be initialized. Now these spin locks are initialized in the function rxe_qp_init_misc. This will avoid the error "initialize spin locks before use".

Quoted 6 lines
Fixes 8700e3e7c485 ("Soft RoCE driver")
Reported-by [email protected]
---
 drivers/infiniband/sw/rxe/rxe_qp.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

Applied to for-next, thanks

Jason