We've observed an NFS server shrink the TCP window and then reset the TCP
connection as part of a HA failover. When the connection has TLS, often
the NFS client will hang indefinitely in this stack:
wait_woken+0x70/0x80
wait_on_pending_writer+0xe4/0x110 [tls]
tls_sk_proto_close+0x368/0x3a0 [tls]
inet_release+0x54/0xb0
__sock_release+0x48/0xc8
sock_close+0x20/0x38
__fput+0xe0/0x2f0
__fput_sync+0x58/0x70
xs_reset_transport+0xe8/0x1f8 [sunrpc]
xs_tcp_shutdown+0xa4/0x190 [sunrpc]
xprt_autoclose+0x68/0x170 [sunrpc]
process_one_work+0x180/0x420
worker_thread+0x258/0x368
kthread+0x104/0x118
ret_from_fork+0x10/0x20
This hang prevents the client from closing the socket and reconnecting to
the server.
Because xs_nospace() elevates sk_write_pending, and sk_sndtimeo is
MAX_SCHEDULE_TIMEOUT, tls_sk_proto_close is never able to complete its wait
for pending writes to the socket. For this case where we are resetting the
transport anyway, we don't expect the socket to ever have write space, so
fix this by simply clearing the sock's sndtimeo under the sock's lock.
Signed-off-by: Benjamin Coddington <bcodding@xxxxxxxxxx>
---
net/sunrpc/xprtsock.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/net/sunrpc/xprtsock.c b/net/sunrpc/xprtsock.c
index 0e1691316f42..8b56d4a8c660 100644
--- a/net/sunrpc/xprtsock.c
+++ b/net/sunrpc/xprtsock.c
@@ -1278,6 +1278,7 @@ static void xs_reset_transport(struct sock_xprt *transport)
transport->file = NULL;
sk->sk_user_data = NULL;
+ sk->sk_sndtimeo = 0;
xs_restore_old_callbacks(transport, sk);
xprt_clear_connected(xprt);
base-commit: 59b723cd2adbac2a34fc8e12c74ae26ae45bf230
--
2.47.0
