From: Karsten Graul <kgraul@linux.ibm.com>
Subject: net/smc: fix non-blocking connect problem
Patch-mainline: v4.19-rc6
Git-commit: 648a5a7aed346c3b8fe7c32a835edfb0dfbf4451
References: FATE#325698, LTC#167867, bsc#1113481

Description:  net/smc: bugfix and compatibility patches
Symptom:      Random hangs in smc processing:
                user space application hangs in socket send() or recv() call or
                does never get a notification from a select() call.
              Missing compatibility to other platforms:
                confirm rkey and delete rkey processing is required by the
                design, but delete rkey processing is missing. This leads to
                protocol failures when communicating with other platforms like
                zOS. The SMC-D shutdown signal support is missing, so there is
                no detection if the remote peer closed the link group.
              Broken administration of available WR send payload buffers due to
              a use-after-free condition.
Problem:      Misbehaviour regarding the user space api can lead to hang
              situations. SMC is not fully compatible to some other platforms
              due to missing rkey processing and SMC-D shutdown signal support.
Solution:     Fixed protocoll deficiencies by implementing the required rkey
              processing. For SMC-D, the cursors are now handled atomically to
              handle parallel modifications. The SMC-D shutdown signal is now
              processed when received and sent to the remote peer if needed.
              Prereq patches are included.
Reproduction: Run SMC on a loaded system against zOS as peer system.

Upstream-Description:

              net/smc: fix non-blocking connect problem

              In state SMC_INIT smc_poll() delegates polling to the internal
              CLC socket. This means, once the connect worker has finished
              its kernel_connect() step, the poll wake-up may occur. This is not
              intended. The wake-up should occur from the wake up call in
              smc_connect_work() after __smc_connect() has finished.
              Thus in state SMC_INIT this patch now calls sock_poll_wait() on the
              main SMC socket.

              Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
              Signed-off-by: David S. Miller <davem@davemloft.net>


Signed-off-by: Karsten Graul <kgraul@linux.ibm.com>
Acked-by: Petr Tesarik <ptesarik@suse.com>
---
 net/smc/af_smc.c |    7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c
@@ -742,7 +742,10 @@ static void smc_connect_work(struct work
 		smc->sk.sk_err = -rc;
 
 out:
-	smc->sk.sk_state_change(&smc->sk);
+	if (smc->sk.sk_err)
+		smc->sk.sk_state_change(&smc->sk);
+	else
+		smc->sk.sk_write_space(&smc->sk);
 	kfree(smc->connect_info);
 	smc->connect_info = NULL;
 	release_sock(&smc->sk);
@@ -1529,7 +1532,7 @@ static unsigned int smc_poll(struct file
 		return POLLNVAL;
 
 	smc = smc_sk(sock->sk);
-	if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
+	if (smc->use_fallback) {
 		/* delegate to CLC child sock */
 		mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
 		sk->sk_err = smc->clcsock->sk->sk_err;