On 03/03/2020 03:12 PM, Josef Bacik wrote: > On 2/28/20 1:40 AM, Hou Pu wrote: >> Nbd server with multiple connections could be upgraded since >> 560bc4b (nbd: handle dead connections). But if only one conncection >> is configured, after we take down nbd server, all inflight IO >> would finally timeout and return error. We could requeue them >> like what we do with multiple connections and wait for new socket >> in submit path. >> >> Signed-off-by: Hou Pu <houpu@xxxxxxxxxxxxx> >> --- >> drivers/block/nbd.c | 17 +++++++++-------- >> 1 file changed, 9 insertions(+), 8 deletions(-) >> >> diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c >> index 78181908f0df..83070714888b 100644 >> --- a/drivers/block/nbd.c >> +++ b/drivers/block/nbd.c >> @@ -395,16 +395,19 @@ static enum blk_eh_timer_return >> nbd_xmit_timeout(struct request *req, >> } >> config = nbd->config; >> - if (config->num_connections > 1) { >> + if (config->num_connections > 1 || >> + (config->num_connections == 1 && nbd->tag_set.timeout)) { > > This is every connection, do you mean to couple this with > dead_conn_timeout? Thanks, > In commit 2da22da573481cc4837e246d0eee4d518b3f715e Author: Mike Christie <mchristi@xxxxxxxxxx> Date: Tue Aug 13 11:39:52 2019 -0500 nbd: fix zero cmd timeout handling v2 we can set tag_set.timeout=0 again. So if timeout != 0 and num_connections = 1, we requeue here and let nbd_handle_cmd->wait_for_reconnect decide to wait or fail the command if dead_conn_timeout is not set. If timeout = 0, then we give it more time because it might have just been a slow server or connection. I think this behavior is wrong for the case Hou is fixing. See comment in the next patch.