SUNRPC: ensure correct error is reported by xs_tcp_setup_socket()
authorNeilBrown <neilb@suse.com>
Thu, 25 May 2017 07:00:32 +0000 (17:00 +1000)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 31 May 2017 16:26:44 +0000 (12:26 -0400)
commit6ea44adce91526700535b3150f77f8639ae8c82d
treeaf9e4e73092f737544f082de3e3e1e0fb8dab322
parentb49c15f97c936ef5a536821f97e4dd8568369802
SUNRPC: ensure correct error is reported by xs_tcp_setup_socket()

If you attempt a TCP mount from an host that is unreachable in a way
that triggers an immediate error from kernel_connect(), that error
does not propagate up, instead EAGAIN is reported.

This results in call_connect_status receiving the wrong error.

A case that it easy to demonstrate is to attempt to mount from an
address that results in ENETUNREACH, but first deleting any default
route.
Without this patch, the mount.nfs process is persistently runnable
and is hard to kill.  With this patch it exits as it should.

The problem is caused by the fact that xs_tcp_force_close() eventually
calls
      xprt_wake_pending_tasks(xprt, -EAGAIN);
which causes an error return of -EAGAIN.  so when xs_tcp_setup_sock()
calls
      xprt_wake_pending_tasks(xprt, status);
the status is ignored.

Fixes: 4efdd92c9211 ("SUNRPC: Remove TCP client connection reset hack")
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
net/sunrpc/xprtsock.c