net/smc/smc_close.c

   1 /*
   2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   3  *
   4  *  Socket Closing - normal and abnormal
   5  *
   6  *  Copyright IBM Corp. 2016
   7  *
   8  *  Author(s):  Ursula Braun <[email protected]>
   9  */
  10
  11 #include <linux/workqueue.h>
  12 #include <linux/sched/signal.h>
  13
  14 #include <net/sock.h>
  15
  16 #include "smc.h"
  17 #include "smc_tx.h"
  18 #include "smc_cdc.h"
  19 #include "smc_close.h"
  20
  21 #define SMC_CLOSE_WAIT_TX_PENDS_TIME            (5 * HZ)
  22
  23 static void smc_close_cleanup_listen(struct sock *parent)
  24 {
  25         struct sock *sk;
  26
  27         /* Close non-accepted connections */
  28         while ((sk = smc_accept_dequeue(parent, NULL)))
  29                 smc_close_non_accepted(sk);
  30 }
  31
  32 static void smc_close_wait_tx_pends(struct smc_sock *smc)
  33 {
  34         DEFINE_WAIT_FUNC(wait, woken_wake_function);
  35         struct sock *sk = &smc->sk;
  36         signed long timeout;
  37
  38         timeout = SMC_CLOSE_WAIT_TX_PENDS_TIME;
  39         add_wait_queue(sk_sleep(sk), &wait);
  40         while (!signal_pending(current) && timeout) {
  41                 int rc;
  42
  43                 rc = sk_wait_event(sk, &timeout,
  44                                    !smc_cdc_tx_has_pending(&smc->conn),
  45                                    &wait);
  46                 if (rc)
  47                         break;
  48         }
  49         remove_wait_queue(sk_sleep(sk), &wait);
  50 }
  51
  52 /* wait for sndbuf data being transmitted */
  53 static void smc_close_stream_wait(struct smc_sock *smc, long timeout)
  54 {
  55         DEFINE_WAIT_FUNC(wait, woken_wake_function);
  56         struct sock *sk = &smc->sk;
  57
  58         if (!timeout)
  59                 return;
  60
  61         if (!smc_tx_prepared_sends(&smc->conn))
  62                 return;
  63
  64         smc->wait_close_tx_prepared = 1;
  65         add_wait_queue(sk_sleep(sk), &wait);
  66         while (!signal_pending(current) && timeout) {
  67                 int rc;
  68
  69                 rc = sk_wait_event(sk, &timeout,
  70                                    !smc_tx_prepared_sends(&smc->conn) ||
  71                                    (sk->sk_err == ECONNABORTED) ||
  72                                    (sk->sk_err == ECONNRESET),
  73                                    &wait);
  74                 if (rc)
  75                         break;
  76         }
  77         remove_wait_queue(sk_sleep(sk), &wait);
  78         smc->wait_close_tx_prepared = 0;
  79 }
  80
  81 void smc_close_wake_tx_prepared(struct smc_sock *smc)
  82 {
  83         if (smc->wait_close_tx_prepared)
  84                 /* wake up socket closing */
  85                 smc->sk.sk_state_change(&smc->sk);
  86 }
  87
  88 static int smc_close_wr(struct smc_connection *conn)
  89 {
  90         conn->local_tx_ctrl.conn_state_flags.peer_done_writing = 1;
  91
  92         return smc_cdc_get_slot_and_msg_send(conn);
  93 }
  94
  95 static int smc_close_final(struct smc_connection *conn)
  96 {
  97         if (atomic_read(&conn->bytes_to_rcv))
  98                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
  99         else
 100                 conn->local_tx_ctrl.conn_state_flags.peer_conn_closed = 1;
 101
 102         return smc_cdc_get_slot_and_msg_send(conn);
 103 }
 104
 105 static int smc_close_abort(struct smc_connection *conn)
 106 {
 107         conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 108
 109         return smc_cdc_get_slot_and_msg_send(conn);
 110 }
 111
 112 /* terminate smc socket abnormally - active abort
 113  * RDMA communication no longer possible
 114  */
 115 void smc_close_active_abort(struct smc_sock *smc)
 116 {
 117         struct smc_cdc_conn_state_flags *txflags =
 118                 &smc->conn.local_tx_ctrl.conn_state_flags;
 119
 120         bh_lock_sock(&smc->sk);
 121         smc->sk.sk_err = ECONNABORTED;
 122         if (smc->clcsock && smc->clcsock->sk) {
 123                 smc->clcsock->sk->sk_err = ECONNABORTED;
 124                 smc->clcsock->sk->sk_state_change(smc->clcsock->sk);
 125         }
 126         switch (smc->sk.sk_state) {
 127         case SMC_INIT:
 128                 smc->sk.sk_state = SMC_PEERABORTWAIT;
 129                 break;
 130         case SMC_APPCLOSEWAIT1:
 131         case SMC_APPCLOSEWAIT2:
 132                 txflags->peer_conn_abort = 1;
 133                 sock_release(smc->clcsock);
 134                 if (!smc_cdc_rxed_any_close(&smc->conn))
 135                         smc->sk.sk_state = SMC_PEERABORTWAIT;
 136                 else
 137                         smc->sk.sk_state = SMC_CLOSED;
 138                 break;
 139         case SMC_PEERCLOSEWAIT1:
 140         case SMC_PEERCLOSEWAIT2:
 141                 if (!txflags->peer_conn_closed) {
 142                         smc->sk.sk_state = SMC_PEERABORTWAIT;
 143                         txflags->peer_conn_abort = 1;
 144                         sock_release(smc->clcsock);
 145                 } else {
 146                         smc->sk.sk_state = SMC_CLOSED;
 147                 }
 148                 break;
 149         case SMC_PROCESSABORT:
 150         case SMC_APPFINCLOSEWAIT:
 151                 if (!txflags->peer_conn_closed) {
 152                         txflags->peer_conn_abort = 1;
 153                         sock_release(smc->clcsock);
 154                 }
 155                 smc->sk.sk_state = SMC_CLOSED;
 156                 break;
 157         case SMC_PEERFINCLOSEWAIT:
 158         case SMC_PEERABORTWAIT:
 159         case SMC_CLOSED:
 160                 break;
 161         }
 162
 163         sock_set_flag(&smc->sk, SOCK_DEAD);
 164         bh_unlock_sock(&smc->sk);
 165         smc->sk.sk_state_change(&smc->sk);
 166 }
 167
 168 int smc_close_active(struct smc_sock *smc)
 169 {
 170         struct smc_cdc_conn_state_flags *txflags =
 171                 &smc->conn.local_tx_ctrl.conn_state_flags;
 172         long timeout = SMC_MAX_STREAM_WAIT_TIMEOUT;
 173         struct smc_connection *conn = &smc->conn;
 174         struct sock *sk = &smc->sk;
 175         int old_state;
 176         int rc = 0;
 177
 178         if (sock_flag(sk, SOCK_LINGER) &&
 179             !(current->flags & PF_EXITING))
 180                 timeout = sk->sk_lingertime;
 181
 182 again:
 183         old_state = sk->sk_state;
 184         switch (old_state) {
 185         case SMC_INIT:
 186                 sk->sk_state = SMC_CLOSED;
 187                 if (smc->smc_listen_work.func)
 188                         flush_work(&smc->smc_listen_work);
 189                 sock_put(sk);
 190                 break;
 191         case SMC_LISTEN:
 192                 sk->sk_state = SMC_CLOSED;
 193                 sk->sk_state_change(sk); /* wake up accept */
 194                 if (smc->clcsock && smc->clcsock->sk) {
 195                         rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
 196                         /* wake up kernel_accept of smc_tcp_listen_worker */
 197                         smc->clcsock->sk->sk_data_ready(smc->clcsock->sk);
 198                 }
 199                 release_sock(sk);
 200                 smc_close_cleanup_listen(sk);
 201                 flush_work(&smc->tcp_listen_work);
 202                 lock_sock(sk);
 203                 break;
 204         case SMC_ACTIVE:
 205                 smc_close_stream_wait(smc, timeout);
 206                 release_sock(sk);
 207                 cancel_work_sync(&conn->tx_work);
 208                 lock_sock(sk);
 209                 if (sk->sk_state == SMC_ACTIVE) {
 210                         /* send close request */
 211                         rc = smc_close_final(conn);
 212                         sk->sk_state = SMC_PEERCLOSEWAIT1;
 213                 } else {
 214                         /* peer event has changed the state */
 215                         goto again;
 216                 }
 217                 break;
 218         case SMC_APPFINCLOSEWAIT:
 219                 /* socket already shutdown wr or both (active close) */
 220                 if (txflags->peer_done_writing &&
 221                     !txflags->peer_conn_closed) {
 222                         /* just shutdown wr done, send close request */
 223                         rc = smc_close_final(conn);
 224                 }
 225                 sk->sk_state = SMC_CLOSED;
 226                 smc_close_wait_tx_pends(smc);
 227                 break;
 228         case SMC_APPCLOSEWAIT1:
 229         case SMC_APPCLOSEWAIT2:
 230                 if (!smc_cdc_rxed_any_close(conn))
 231                         smc_close_stream_wait(smc, timeout);
 232                 release_sock(sk);
 233                 cancel_work_sync(&conn->tx_work);
 234                 lock_sock(sk);
 235                 if (sk->sk_err != ECONNABORTED) {
 236                         /* confirm close from peer */
 237                         rc = smc_close_final(conn);
 238                         if (rc)
 239                                 break;
 240                 }
 241                 if (smc_cdc_rxed_any_close(conn))
 242                         /* peer has closed the socket already */
 243                         sk->sk_state = SMC_CLOSED;
 244                 else
 245                         /* peer has just issued a shutdown write */
 246                         sk->sk_state = SMC_PEERFINCLOSEWAIT;
 247                 smc_close_wait_tx_pends(smc);
 248                 break;
 249         case SMC_PEERCLOSEWAIT1:
 250         case SMC_PEERCLOSEWAIT2:
 251         case SMC_PEERFINCLOSEWAIT:
 252                 /* peer sending PeerConnectionClosed will cause transition */
 253                 break;
 254         case SMC_PROCESSABORT:
 255                 cancel_work_sync(&conn->tx_work);
 256                 smc_close_abort(conn);
 257                 sk->sk_state = SMC_CLOSED;
 258                 smc_close_wait_tx_pends(smc);
 259                 break;
 260         case SMC_PEERABORTWAIT:
 261         case SMC_CLOSED:
 262                 /* nothing to do, add tracing in future patch */
 263                 break;
 264         }
 265
 266         if (old_state != sk->sk_state)
 267                 sk->sk_state_change(&smc->sk);
 268         return rc;
 269 }
 270
 271 static void smc_close_passive_abort_received(struct smc_sock *smc)
 272 {
 273         struct smc_cdc_conn_state_flags *txflags =
 274                 &smc->conn.local_tx_ctrl.conn_state_flags;
 275         struct sock *sk = &smc->sk;
 276
 277         switch (sk->sk_state) {
 278         case SMC_ACTIVE:
 279         case SMC_APPFINCLOSEWAIT:
 280         case SMC_APPCLOSEWAIT1:
 281         case SMC_APPCLOSEWAIT2:
 282                 smc_close_abort(&smc->conn);
 283                 sk->sk_state = SMC_PROCESSABORT;
 284                 break;
 285         case SMC_PEERCLOSEWAIT1:
 286         case SMC_PEERCLOSEWAIT2:
 287                 if (txflags->peer_done_writing &&
 288                     !txflags->peer_conn_closed) {
 289                         /* just shutdown, but not yet closed locally */
 290                         smc_close_abort(&smc->conn);
 291                         sk->sk_state = SMC_PROCESSABORT;
 292                 } else {
 293                         sk->sk_state = SMC_CLOSED;
 294                 }
 295                 break;
 296         case SMC_PEERFINCLOSEWAIT:
 297         case SMC_PEERABORTWAIT:
 298                 sk->sk_state = SMC_CLOSED;
 299                 break;
 300         case SMC_INIT:
 301         case SMC_PROCESSABORT:
 302         /* nothing to do, add tracing in future patch */
 303                 break;
 304         }
 305 }
 306
 307 /* Some kind of closing has been received: peer_conn_closed, peer_conn_abort,
 308  * or peer_done_writing.
 309  * Called under tasklet context.
 310  */
 311 void smc_close_passive_received(struct smc_sock *smc)
 312 {
 313         struct smc_cdc_conn_state_flags *rxflags =
 314                 &smc->conn.local_rx_ctrl.conn_state_flags;
 315         struct sock *sk = &smc->sk;
 316         int old_state;
 317
 318         sk->sk_shutdown |= RCV_SHUTDOWN;
 319         if (smc->clcsock && smc->clcsock->sk)
 320                 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
 321         sock_set_flag(&smc->sk, SOCK_DONE);
 322
 323         old_state = sk->sk_state;
 324
 325         if (rxflags->peer_conn_abort) {
 326                 smc_close_passive_abort_received(smc);
 327                 goto wakeup;
 328         }
 329
 330         switch (sk->sk_state) {
 331         case SMC_INIT:
 332                 if (atomic_read(&smc->conn.bytes_to_rcv) ||
 333                     (rxflags->peer_done_writing &&
 334                      !rxflags->peer_conn_closed))
 335                         sk->sk_state = SMC_APPCLOSEWAIT1;
 336                 else
 337                         sk->sk_state = SMC_CLOSED;
 338                 break;
 339         case SMC_ACTIVE:
 340                 sk->sk_state = SMC_APPCLOSEWAIT1;
 341                 break;
 342         case SMC_PEERCLOSEWAIT1:
 343                 if (rxflags->peer_done_writing)
 344                         sk->sk_state = SMC_PEERCLOSEWAIT2;
 345                 /* fall through to check for closing */
 346         case SMC_PEERCLOSEWAIT2:
 347         case SMC_PEERFINCLOSEWAIT:
 348                 if (!smc_cdc_rxed_any_close(&smc->conn))
 349                         break;
 350                 if (sock_flag(sk, SOCK_DEAD) &&
 351                     (sk->sk_shutdown == SHUTDOWN_MASK)) {
 352                         /* smc_release has already been called locally */
 353                         sk->sk_state = SMC_CLOSED;
 354                 } else {
 355                         /* just shutdown, but not yet closed locally */
 356                         sk->sk_state = SMC_APPFINCLOSEWAIT;
 357                 }
 358                 break;
 359         case SMC_APPCLOSEWAIT1:
 360         case SMC_APPCLOSEWAIT2:
 361         case SMC_APPFINCLOSEWAIT:
 362         case SMC_PEERABORTWAIT:
 363         case SMC_PROCESSABORT:
 364         case SMC_CLOSED:
 365                 /* nothing to do, add tracing in future patch */
 366                 break;
 367         }
 368
 369 wakeup:
 370         if (old_state != sk->sk_state)
 371                 sk->sk_state_change(sk);
 372         sk->sk_data_ready(sk); /* wakeup blocked rcvbuf consumers */
 373         sk->sk_write_space(sk); /* wakeup blocked sndbuf producers */
 374
 375         if ((sk->sk_state == SMC_CLOSED) &&
 376             (sock_flag(sk, SOCK_DEAD) || (old_state == SMC_INIT))) {
 377                 smc_conn_free(&smc->conn);
 378                 schedule_delayed_work(&smc->sock_put_work,
 379                                       SMC_CLOSE_SOCK_PUT_DELAY);
 380         }
 381 }
 382
 383 void smc_close_sock_put_work(struct work_struct *work)
 384 {
 385         struct smc_sock *smc = container_of(to_delayed_work(work),
 386                                             struct smc_sock,
 387                                             sock_put_work);
 388
 389         smc->sk.sk_prot->unhash(&smc->sk);
 390         sock_put(&smc->sk);
 391 }
 392
 393 int smc_close_shutdown_write(struct smc_sock *smc)
 394 {
 395         struct smc_connection *conn = &smc->conn;
 396         long timeout = SMC_MAX_STREAM_WAIT_TIMEOUT;
 397         struct sock *sk = &smc->sk;
 398         int old_state;
 399         int rc = 0;
 400
 401         if (sock_flag(sk, SOCK_LINGER))
 402                 timeout = sk->sk_lingertime;
 403
 404 again:
 405         old_state = sk->sk_state;
 406         switch (old_state) {
 407         case SMC_ACTIVE:
 408                 smc_close_stream_wait(smc, timeout);
 409                 release_sock(sk);
 410                 cancel_work_sync(&conn->tx_work);
 411                 lock_sock(sk);
 412                 /* send close wr request */
 413                 rc = smc_close_wr(conn);
 414                 if (sk->sk_state == SMC_ACTIVE)
 415                         sk->sk_state = SMC_PEERCLOSEWAIT1;
 416                 else
 417                         goto again;
 418                 break;
 419         case SMC_APPCLOSEWAIT1:
 420                 /* passive close */
 421                 if (!smc_cdc_rxed_any_close(conn))
 422                         smc_close_stream_wait(smc, timeout);
 423                 release_sock(sk);
 424                 cancel_work_sync(&conn->tx_work);
 425                 lock_sock(sk);
 426                 /* confirm close from peer */
 427                 rc = smc_close_wr(conn);
 428                 sk->sk_state = SMC_APPCLOSEWAIT2;
 429                 break;
 430         case SMC_APPCLOSEWAIT2:
 431         case SMC_PEERFINCLOSEWAIT:
 432         case SMC_PEERCLOSEWAIT1:
 433         case SMC_PEERCLOSEWAIT2:
 434         case SMC_APPFINCLOSEWAIT:
 435         case SMC_PROCESSABORT:
 436         case SMC_PEERABORTWAIT:
 437                 /* nothing to do, add tracing in future patch */
 438                 break;
 439         }
 440
 441         if (old_state != sk->sk_state)
 442                 sk->sk_state_change(&smc->sk);
 443         return rc;
 444 }