Re: [PATCH] SUNRPC: fix shutdown of NFS TCP client socket

[Siddharth Kawar]

> [Some people who received this message don't often get email from trondmy@xxxxxxxxxxxxxxx. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
> 
> > On Mar 8, 2023, at 14:01, Siddharth Kawar <Siddharth.Kawar@xxxxxxxxxxxxx> wrote:
> >
> > [You don't often get email from siddharth.kawar@xxxxxxxxxxxxx. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
> >
> > Subject: [PATCH] SUNRPC: fix shutdown of NFS TCP client socket
> >
> > NFS server Duplicate Request Cache (DRC) algorithms rely on NFS clients
> > reconnecting using the same local TCP port. Unique NFS operations are
> > identified by the per-TCP connection set of XIDs. This prevents file
> > corruption when non-idempotent NFS operations are retried.
> >
> > Currently, NFS client TCP connections are using different local TCP ports
> > when reconnecting to NFS servers.
> >
> > After an NFS server initiates shutdown of the TCP connection, the NFS
> > client's TCP socket is set to NULL after the socket state has reached
> > TCP_LAST_ACK(9). When reconnecting, the new socket attempts to reuse
> > the same local port but fails with EADDRNOTAVAIL (99). This forces the
> > socket to use a different local TCP port to reconnect to the remote NFS
> > server.
> >
> > State Transition and Events:
> > TCP_CLOSING(8)
> 
> No: TCP state == 8 is TCP_CLOSE_WAIT

You are right.  Will fix in v2.

> 
> > TCP_LAST_ACK(9)
> > connect(fail EADDRNOTAVAIL(99))
> > TCP_CLOSE(7)
> > bind on new port
> > connect success
> >
> > dmesg excerpts showing reconnect switching from local port of 688 to 1014:
> > [590701.200229] NFS: mkdir(0:61/560857152), testQ
> > [590701.200231] NFS call  mkdir testQ
> > [590701.200259] RPC:       xs_tcp_send_request(224) = 0
> > [590751.883111] RPC:       xs_tcp_state_change client 0000000051f4e000...
> > [590751.883146] RPC:       state 8 conn 1 dead 0 zapped 1 sk_shutdown 1
> > [590751.883160] RPC:       xs_data_ready...
> > [590751.883232] RPC:       xs_tcp_state_change client 0000000051f4e000...
> > [590751.883235] RPC:       state 9 conn 0 dead 0 zapped 1 sk_shutdown 3
> > [590751.883238] RPC:       xs_tcp_state_change client 0000000051f4e000...
> > [590751.883239] RPC:       state 9 conn 0 dead 0 zapped 1 sk_shutdown 3
> > [590751.883283] RPC:       xs_connect scheduled xprt 0000000051f4e000
> > [590751.883314] RPC:       xs_bind 0.0.0.0:688: ok (0)
> > [590751.883321] RPC:       worker connecting xprt 0000000051f4e000 via tcp
> >                           to 10.101.1.30 (port 2049)
> > [590751.883330] RPC:       0000000051f4e000 connect status 99 connected 0
> >                           sock state 7
> > [590751.883342] RPC:       xs_tcp_state_change client 0000000051f4e000...
> > [590751.883343] RPC:       state 7 conn 0 dead 0 zapped 1 sk_shutdown 3
> > [590751.883356] RPC:       xs_connect scheduled xprt 0000000051f4e000
> > [590751.883383] RPC:       xs_bind 0.0.0.0:1014: ok (0)
> > [590751.883388] RPC:       worker connecting xprt 0000000051f4e000 via tcp
> >                           to 10.101.1.30 (port 2049)
> > [590751.883420] RPC:       0000000051f4e000 connect status 115 connected 0
> >                           sock state 2
> > ...
> >
> >
> > State Transition and Events with patch applied:
> > TCP_CLOSING(8)
> > TCP_LAST_ACK(9)
> > TCP_CLOSE(7)
> > connect(reuse of port succeeds)
> >
> > dmesg excerpts showing reconnect on same port (936):
> > [  257.139935] NFS: mkdir(0:59/560857152), testQ
> > [  257.139937] NFS call  mkdir testQ
> > ...
> > [  307.822702] RPC:       state 8 conn 1 dead 0 zapped 1 sk_shutdown 1
> > [  307.822714] RPC:       xs_data_ready...
> > [  307.822817] RPC:       xs_tcp_state_change client 00000000ce702f14...
> > [  307.822821] RPC:       state 9 conn 0 dead 0 zapped 1 sk_shutdown 3
> > [  307.822825] RPC:       xs_tcp_state_change client 00000000ce702f14...
> > [  307.822826] RPC:       state 9 conn 0 dead 0 zapped 1 sk_shutdown 3
> > [  307.823606] RPC:       xs_tcp_state_change client 00000000ce702f14...
> > [  307.823609] RPC:       state 7 conn 0 dead 0 zapped 1 sk_shutdown 3
> > [  307.823629] RPC:       xs_tcp_state_change client 00000000ce702f14...
> > [  307.823632] RPC:       state 7 conn 0 dead 0 zapped 1 sk_shutdown 3
> > [  307.823676] RPC:       xs_connect scheduled xprt 00000000ce702f14
> > [  307.823704] RPC:       xs_bind 0.0.0.0:936: ok (0)
> > [  307.823709] RPC:       worker connecting xprt 00000000ce702f14 via tcp
> >                          to 10.101.1.30 (port 2049)
> > [  307.823748] RPC:       00000000ce702f14 connect status 115 connected 0
> >                          sock state 2
> > ...
> > [  314.916193] RPC:       state 7 conn 0 dead 0 zapped 1 sk_shutdown 3
> > [  314.916251] RPC:       xs_connect scheduled xprt 00000000ce702f14
> > [  314.916282] RPC:       xs_bind 0.0.0.0:936: ok (0)
> > [  314.916292] RPC:       worker connecting xprt 00000000ce702f14 via tcp
> >                          to 10.101.1.30 (port 2049)
> > [  314.916342] RPC:       00000000ce702f14 connect status 115 connected 0
> >                          sock state 2
> >
> > Fixes: 7c81e6a9d75b (SUNRPC: Tweak TCP socket shutdown in the RPC client)
> > Signed-off-by: Siddharth Rajendra Kawar <sikawar@xxxxxxxxxxxxx>
> > ---
> > diff --git a/net/sunrpc/xprtsock.c b/net/sunrpc/xprtsock.c
> > index d8ee06a9650a..e55e380bc10c 100644
> > --- a/net/sunrpc/xprtsock.c
> > +++ b/net/sunrpc/xprtsock.c
> > @@ -2094,6 +2094,7 @@ static void xs_tcp_shutdown(struct rpc_xprt *xprt)
> >                break;
> >        case TCP_ESTABLISHED:
> >        case TCP_CLOSE_WAIT:
> > +       case TCP_LAST_ACK:
> >                kernel_sock_shutdown(sock, SHUT_RDWR);
> >                trace_rpc_socket_shutdown(xprt, sock);
> >                break;
> 
> The only way to transition from TCP_CLOSE_WAIT to TCP_LAST_ACK is by closing the socket ( see https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.ibm.com%2Fdocs%2Fen%2Fzos%2F2.1.0%3Ftopic%3DSSLTBW_2.1.0%2Fcom.ibm.zos.v2r1.halu101%2Fconstatus.html&data=05%7C01%7CSiddharth.Kawar%40microsoft.com%7C96b3c75820714bd3753208db24c266f8%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C638144189616407625%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=uuXzgjaPk7YpnY6nO0NgNBPeq5HaA5%2BNLyrNFe%2BcjE8%3D&reserved=0 ), so closing it while in TCP_LAST_ACK is redundant.

You are right. The TCP_LAST_ACK state should not cause the socket to be
closed again. Instead, as in the case of TCP_FIN_WAIT1 and TCP_FIN_WAIT2,
the TCP_LAST_ACK state should not be calling xs_reset_transport. Will fix
in v2.

> 
> 
> _________________________________
> Trond Myklebust
> Linux NFS client maintainer, Hammerspace
> trond.myklebust@xxxxxxxxxxxxxxx