sunrpc: Don't engage exponential backoff when connection attempt is rejected.
authorNeilBrown <neilb@suse.com>
Wed, 23 Nov 2016 03:44:58 +0000 (14:44 +1100)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Thu, 1 Dec 2016 22:40:41 +0000 (17:40 -0500)
commit2c2ee6d20b10594892c1d31c7b959f4780adde63
treee091da5c1d0b5acee6e4926736f1165463d01089
parentb85f562049cc7dce0d65577427a8321197d20983
sunrpc: Don't engage exponential backoff when connection attempt is rejected.

xs_connect() contains an exponential backoff mechanism so the repeated
connection attempts are delayed by longer and longer amounts.

This is appropriate when the connection failed due to a timeout, but
it not appropriate when a definitive "no" answer is received.  In such
cases, call_connect_status() imposes a minimum 3-second back-off, so
not having the exponetial back-off will never result in immediate
retries.

The current situation is a problem when the NFS server tries to
register with rpcbind but rpcbind isn't running.  All connection
attempts are made on the same "xprt" and as the connection is never
"closed", the exponential back delays successive attempts to register,
or de-register, different protocols.  This results in a multi-minute
delay with no benefit.

So, when call_connect_status() receives a definitive "no", use
xprt_conditional_disconnect() to cancel the previous connection attempt.
This will set XPRT_CLOSE_WAIT so that xprt->ops->close() calls xs_close()
which resets the reestablish_timeout.

To ensure xprt_conditional_disconnect() does the right thing, we
ensure that rq_connect_cookie is set before a connection attempt, and
allow xprt_conditional_disconnect() to complete even when the
transport is not fully connected.

Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
net/sunrpc/clnt.c
net/sunrpc/xprt.c