net/sunrpc/xprt_sock: fix regression in connection error reporting.
authorNeilBrown <neilb@suse.com>
Wed, 19 Jul 2017 04:05:01 +0000 (14:05 +1000)
committerAnna Schumaker <Anna.Schumaker@Netapp.com>
Fri, 21 Jul 2017 12:49:58 +0000 (08:49 -0400)
commit3ffbc1d65583394be12801655781dd2b079ce169
treea906fde05bff9b28dc499c89cb9c799867c86107
parentecc7b435d2beb025d7506af74cf749af2cef5734
net/sunrpc/xprt_sock: fix regression in connection error reporting.

Commit 3d4762639dd3 ("tcp: remove poll() flakes when receiving
RST") in v4.12 changed the order in which ->sk_state_change()
and ->sk_error_report() are called when a socket is shut
down - sk_state_change() is now called first.

This causes xs_tcp_state_change() -> xs_sock_mark_closed() ->
xprt_disconnect_done() to wake all pending tasked with -EAGAIN.
When the ->sk_error_report() callback arrives, it is too late to
pass the error on, and it is lost.

As easy way to demonstrate the problem caused is to try to start
rpc.nfsd while rcpbind isn't running.
nfsd will attempt a tcp connection to rpcbind.  A ECONNREFUSED
error is returned, but sunrpc code loses the error and keeps
retrying.  If it saw the ECONNREFUSED, it would abort.

To fix this, handle the sk->sk_err in the TCP_CLOSE branch of
xs_tcp_state_change().

Fixes: 3d4762639dd3 ("tcp: remove poll() flakes when receiving RST")
Cc: stable@vger.kernel.org (v4.12)
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
net/sunrpc/xprtsock.c