nvme-rdma: fix crash when connect rejected
authorChao Leng <lengchao@huawei.com>
Mon, 12 Oct 2020 08:10:40 +0000 (16:10 +0800)
committerChristoph Hellwig <hch@lst.de>
Thu, 22 Oct 2020 13:27:13 +0000 (15:27 +0200)
A crash can happened when a connect is rejected.   The host establishes
the connection after received ConnectReply, and then continues to send
the fabrics Connect command.  If the controller does not receive the
ReadyToUse capsule, host may receive a ConnectReject reply.

Call nvme_rdma_destroy_queue_ib after the host received the
RDMA_CM_EVENT_REJECTED event.  Then when the fabrics Connect command
times out, nvme_rdma_timeout calls nvme_rdma_complete_rq to fail the
request.  A crash happenes due to use after free in
nvme_rdma_complete_rq.

nvme_rdma_destroy_queue_ib is redundant when handling the
RDMA_CM_EVENT_REJECTED event as nvme_rdma_destroy_queue_ib is already
called in connection failure handler.

Signed-off-by: Chao Leng <lengchao@huawei.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
drivers/nvme/host/rdma.c

index 9e378d0a0c01c157ad24ab62595510c0c4f31eeb..116902b1b2c347ca7939f8e8eeae0271e1c3629d 100644 (file)
@@ -1926,7 +1926,6 @@ static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
                complete(&queue->cm_done);
                return 0;
        case RDMA_CM_EVENT_REJECTED:
-               nvme_rdma_destroy_queue_ib(queue);
                cm_error = nvme_rdma_conn_rejected(queue, ev);
                break;
        case RDMA_CM_EVENT_ROUTE_ERROR: