[PATCH] net: 9p: Fix possible null-pointer dereference in p9_cm_event_handler()

The variable rdma is checked when event->event is equal to 
RDMA_CM_EVENT_DISCONNECTED:
  if (rdma)

This indicates that it can be NULL. If so, a null-pointer dereference will 
occur when calling complete():
  complete(&rdma->cm_done);

To fix this possible null-pointer dereference, calling complete() only 
when rdma is not NULL.

Reported-by: TOTE Robot <redacted>
Signed-off-by: Tuo Li <redacted>
---
 net/9p/trans_rdma.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/net/9p/trans_rdma.c b/net/9p/trans_rdma.c
index af0a8a6cd3fd..fb3435dfd071 100644
--- a/net/9p/trans_rdma.c
+++ b/net/9p/trans_rdma.c

@@ -285,7 +285,8 @@ p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 	default:
 		BUG();
 	}
-	complete(&rdma->cm_done);
+	if (rdma)
+		complete(&rdma->cm_done);
 	return 0;
 }
 

-- 
2.25.1

On Tue, Aug 10, 2021 at 06:20:07AM -0700, Tuo Li wrote:

The variable rdma is checked when event->event is equal to 
RDMA_CM_EVENT_DISCONNECTED:
  if (rdma)

This indicates that it can be NULL. If so, a null-pointer dereference will 
occur when calling complete():
  complete(&rdma->cm_done);

To fix this possible null-pointer dereference, calling complete() only 
when rdma is not NULL.

You need to explain how is it possible and blindly set if () checks.
I would say first "if (rdma)" is not needed, but don't know for sure.

Leon Romanovsky wrote on Tue, Aug 10, 2021 at 04:55:42PM +0300:

On Tue, Aug 10, 2021 at 06:20:07AM -0700, Tuo Li wrote:

quoted

The variable rdma is checked when event->event is equal to 
RDMA_CM_EVENT_DISCONNECTED:
  if (rdma)

This indicates that it can be NULL. If so, a null-pointer dereference will 
occur when calling complete():
  complete(&rdma->cm_done);

To fix this possible null-pointer dereference, calling complete() only 
when rdma is not NULL.

You need to explain how is it possible and blindly set if () checks.
I would say first "if (rdma)" is not needed, but don't know for sure.

Sounds like static analysis because there's a if (rdma) check in
RDMA_CM_EVENT_DISCONNECTED above, so if that needed check then it will
bug right afterwards

I'd tend to agree I don't think it's possible client->trans is null
there (it's filled right after rdma_create_id which defines the handler,
there might be a window where the callback is called before? But as I
understand it shouldn't be called until we resolve address and connect
then later disconnect)

So, I agree with Leon - unless you have a backtrace of a real bug
let's remove the other 'if' if you want to cleanup something for your
robot.

-- 
Dominique