mlx4_core 0000:07:00.0: swiotlb buffer is full and OOM observed during stress test on reset_controller

[sagi@grimberg.me (Sagi Grimberg)]

> Hi Sagi
> With this path, the OOM cannot be reproduced now.
>
> But there is another problem, the reset operation[1] failed at iteration 1007.
> [1]
> echo 1 >/sys/block/nvme0n1/device/reset_controller

We can relax this a bit by only flushing for admin queue accepts, and
also let the host accept longer time for establishing a connection.

Does this help?
--
diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c
index 47a479f26e5d..e1db1736823f 100644
--- a/drivers/nvme/host/rdma.c
+++ b/drivers/nvme/host/rdma.c
@@ -34,7 +34,7 @@
  #include "fabrics.h"


-#define NVME_RDMA_CONNECT_TIMEOUT_MS   1000            /* 1 second */
+#define NVME_RDMA_CONNECT_TIMEOUT_MS   5000            /* 5 seconds */

  #define NVME_RDMA_MAX_SEGMENT_SIZE     0xffffff        /* 24-bit SGL 
field */

diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c
index ecc4fe862561..88bb5814c264 100644
--- a/drivers/nvme/target/rdma.c
+++ b/drivers/nvme/target/rdma.c
@@ -1199,6 +1199,11 @@ static int nvmet_rdma_queue_connect(struct 
rdma_cm_id *cm_id,
         }
         queue->port = cm_id->context;

+       if (queue->host_qid == 0) {
+               /* Let inflight controller teardown complete */
+               flush_scheduled_work();
+       }
+
         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
         if (ret)
                 goto release_queue;
--