fs/ocfs2/cluster/tcp.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  *
   3  * vim: noexpandtab sw=8 ts=8 sts=0:
   4  *
   5  * Copyright (C) 2004 Oracle.  All rights reserved.
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2 of the License, or (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public
  18  * License along with this program; if not, write to the
  19  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  20  * Boston, MA 021110-1307, USA.
  21  *
  22  * ----
  23  *
  24  * Callers for this were originally written against a very simple synchronus
  25  * API.  This implementation reflects those simple callers.  Some day I'm sure
  26  * we'll need to move to a more robust posting/callback mechanism.
  27  *
  28  * Transmit calls pass in kernel virtual addresses and block copying this into
  29  * the socket's tx buffers via a usual blocking sendmsg.  They'll block waiting
  30  * for a failed socket to timeout.  TX callers can also pass in a poniter to an
  31  * 'int' which gets filled with an errno off the wire in response to the
  32  * message they send.
  33  *
  34  * Handlers for unsolicited messages are registered.  Each socket has a page
  35  * that incoming data is copied into.  First the header, then the data.
  36  * Handlers are called from only one thread with a reference to this per-socket
  37  * page.  This page is destroyed after the handler call, so it can't be
  38  * referenced beyond the call.  Handlers may block but are discouraged from
  39  * doing so.
  40  *
  41  * Any framing errors (bad magic, large payload lengths) close a connection.
  42  *
  43  * Our sock_container holds the state we associate with a socket.  It's current
  44  * framing state is held there as well as the refcounting we do around when it
  45  * is safe to tear down the socket.  The socket is only finally torn down from
  46  * the container when the container loses all of its references -- so as long
  47  * as you hold a ref on the container you can trust that the socket is valid
  48  * for use with kernel socket APIs.
  49  *
  50  * Connections are initiated between a pair of nodes when the node with the
  51  * higher node number gets a heartbeat callback which indicates that the lower
  52  * numbered node has started heartbeating.  The lower numbered node is passive
  53  * and only accepts the connection if the higher numbered node is heartbeating.
  54  */
  55
  56 #include <linux/kernel.h>
  57 #include <linux/jiffies.h>
  58 #include <linux/slab.h>
  59 #include <linux/idr.h>
  60 #include <linux/kref.h>
  61 #include <linux/net.h>
  62 #include <net/tcp.h>
  63
  64 #include <asm/uaccess.h>
  65
  66 #include "heartbeat.h"
  67 #include "tcp.h"
  68 #include "nodemanager.h"
  69 #define MLOG_MASK_PREFIX ML_TCP
  70 #include "masklog.h"
  71 #include "quorum.h"
  72
  73 #include "tcp_internal.h"
  74
  75 #define SC_NODEF_FMT "node %s (num %u) at %u.%u.%u.%u:%u"
  76 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
  77                           NIPQUAD(sc->sc_node->nd_ipv4_address),        \
  78                           ntohs(sc->sc_node->nd_ipv4_port)
  79
  80 /*
  81  * In the following two log macros, the whitespace after the ',' just
  82  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  83  * previous token if args expands to nothing.
  84  */
  85 #define msglog(hdr, fmt, args...) do {                                  \
  86         typeof(hdr) __hdr = (hdr);                                      \
  87         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
  88              "key %08x num %u] " fmt,                                   \
  89              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
  90              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
  91              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
  92              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
  93 } while (0)
  94
  95 #define sclog(sc, fmt, args...) do {                                    \
  96         typeof(sc) __sc = (sc);                                         \
  97         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
  98              "pg_off %zu] " fmt, __sc,                                  \
  99              atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,       \
 100             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
 101             ##args);                                                    \
 102 } while (0)
 103
 104 static DEFINE_RWLOCK(o2net_handler_lock);
 105 static struct rb_root o2net_handler_tree = RB_ROOT;
 106
 107 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
 108
 109 /* XXX someday we'll need better accounting */
 110 static struct socket *o2net_listen_sock = NULL;
 111
 112 /*
 113  * listen work is only queued by the listening socket callbacks on the
 114  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 115  * quorum work is queued as sock containers are shutdown.. stop_listening
 116  * tears down all the node's sock containers, preventing future shutdowns
 117  * and queued quroum work, before canceling delayed quorum work and
 118  * destroying the work queue.
 119  */
 120 static struct workqueue_struct *o2net_wq;
 121 static struct work_struct o2net_listen_work;
 122
 123 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 124 #define O2NET_HB_PRI 0x1
 125
 126 static struct o2net_handshake *o2net_hand;
 127 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 128
 129 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 130                 {[O2NET_ERR_NONE]       = 0,
 131                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
 132                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
 133                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
 134
 135 /* can't quite avoid *all* internal declarations :/ */
 136 static void o2net_sc_connect_completed(struct work_struct *work);
 137 static void o2net_rx_until_empty(struct work_struct *work);
 138 static void o2net_shutdown_sc(struct work_struct *work);
 139 static void o2net_listen_data_ready(struct sock *sk, int bytes);
 140 static void o2net_sc_send_keep_req(struct work_struct *work);
 141 static void o2net_idle_timer(unsigned long data);
 142 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 143 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 144
 145 /*
 146  * FIXME: These should use to_o2nm_cluster_from_node(), but we end up
 147  * losing our parent link to the cluster during shutdown. This can be
 148  * solved by adding a pre-removal callback to configfs, or passing
 149  * around the cluster with the node. -jeffm
 150  */
 151 static inline int o2net_reconnect_delay(struct o2nm_node *node)
 152 {
 153         return o2nm_single_cluster->cl_reconnect_delay_ms;
 154 }
 155
 156 static inline int o2net_keepalive_delay(struct o2nm_node *node)
 157 {
 158         return o2nm_single_cluster->cl_keepalive_delay_ms;
 159 }
 160
 161 static inline int o2net_idle_timeout(struct o2nm_node *node)
 162 {
 163         return o2nm_single_cluster->cl_idle_timeout_ms;
 164 }
 165
 166 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 167 {
 168         int trans;
 169         BUG_ON(err >= O2NET_ERR_MAX);
 170         trans = o2net_sys_err_translations[err];
 171
 172         /* Just in case we mess up the translation table above */
 173         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 174         return trans;
 175 }
 176
 177 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 178 {
 179         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 180         return &o2net_nodes[node_num];
 181 }
 182
 183 static u8 o2net_num_from_nn(struct o2net_node *nn)
 184 {
 185         BUG_ON(nn == NULL);
 186         return nn - o2net_nodes;
 187 }
 188
 189 /* ------------------------------------------------------------ */
 190
 191 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 192 {
 193         int ret = 0;
 194
 195         do {
 196                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
 197                         ret = -EAGAIN;
 198                         break;
 199                 }
 200                 spin_lock(&nn->nn_lock);
 201                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
 202                 if (ret == 0)
 203                         list_add_tail(&nsw->ns_node_item,
 204                                       &nn->nn_status_list);
 205                 spin_unlock(&nn->nn_lock);
 206         } while (ret == -EAGAIN);
 207
 208         if (ret == 0)  {
 209                 init_waitqueue_head(&nsw->ns_wq);
 210                 nsw->ns_sys_status = O2NET_ERR_NONE;
 211                 nsw->ns_status = 0;
 212         }
 213
 214         return ret;
 215 }
 216
 217 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 218                                       struct o2net_status_wait *nsw,
 219                                       enum o2net_system_error sys_status,
 220                                       s32 status)
 221 {
 222         assert_spin_locked(&nn->nn_lock);
 223
 224         if (!list_empty(&nsw->ns_node_item)) {
 225                 list_del_init(&nsw->ns_node_item);
 226                 nsw->ns_sys_status = sys_status;
 227                 nsw->ns_status = status;
 228                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 229                 wake_up(&nsw->ns_wq);
 230         }
 231 }
 232
 233 static void o2net_complete_nsw(struct o2net_node *nn,
 234                                struct o2net_status_wait *nsw,
 235                                u64 id, enum o2net_system_error sys_status,
 236                                s32 status)
 237 {
 238         spin_lock(&nn->nn_lock);
 239         if (nsw == NULL) {
 240                 if (id > INT_MAX)
 241                         goto out;
 242
 243                 nsw = idr_find(&nn->nn_status_idr, id);
 244                 if (nsw == NULL)
 245                         goto out;
 246         }
 247
 248         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 249
 250 out:
 251         spin_unlock(&nn->nn_lock);
 252         return;
 253 }
 254
 255 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 256 {
 257         struct o2net_status_wait *nsw, *tmp;
 258         unsigned int num_kills = 0;
 259
 260         assert_spin_locked(&nn->nn_lock);
 261
 262         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 263                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 264                 num_kills++;
 265         }
 266
 267         mlog(0, "completed %d messages for node %u\n", num_kills,
 268              o2net_num_from_nn(nn));
 269 }
 270
 271 static int o2net_nsw_completed(struct o2net_node *nn,
 272                                struct o2net_status_wait *nsw)
 273 {
 274         int completed;
 275         spin_lock(&nn->nn_lock);
 276         completed = list_empty(&nsw->ns_node_item);
 277         spin_unlock(&nn->nn_lock);
 278         return completed;
 279 }
 280
 281 /* ------------------------------------------------------------ */
 282
 283 static void sc_kref_release(struct kref *kref)
 284 {
 285         struct o2net_sock_container *sc = container_of(kref,
 286                                         struct o2net_sock_container, sc_kref);
 287         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 288
 289         sclog(sc, "releasing\n");
 290
 291         if (sc->sc_sock) {
 292                 sock_release(sc->sc_sock);
 293                 sc->sc_sock = NULL;
 294         }
 295
 296         o2nm_node_put(sc->sc_node);
 297         sc->sc_node = NULL;
 298
 299         kfree(sc);
 300 }
 301
 302 static void sc_put(struct o2net_sock_container *sc)
 303 {
 304         sclog(sc, "put\n");
 305         kref_put(&sc->sc_kref, sc_kref_release);
 306 }
 307 static void sc_get(struct o2net_sock_container *sc)
 308 {
 309         sclog(sc, "get\n");
 310         kref_get(&sc->sc_kref);
 311 }
 312 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 313 {
 314         struct o2net_sock_container *sc, *ret = NULL;
 315         struct page *page = NULL;
 316
 317         page = alloc_page(GFP_NOFS);
 318         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 319         if (sc == NULL || page == NULL)
 320                 goto out;
 321
 322         kref_init(&sc->sc_kref);
 323         o2nm_node_get(node);
 324         sc->sc_node = node;
 325
 326         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 327         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 328         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 329         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 330
 331         init_timer(&sc->sc_idle_timeout);
 332         sc->sc_idle_timeout.function = o2net_idle_timer;
 333         sc->sc_idle_timeout.data = (unsigned long)sc;
 334
 335         sclog(sc, "alloced\n");
 336
 337         ret = sc;
 338         sc->sc_page = page;
 339         sc = NULL;
 340         page = NULL;
 341
 342 out:
 343         if (page)
 344                 __free_page(page);
 345         kfree(sc);
 346
 347         return ret;
 348 }
 349
 350 /* ------------------------------------------------------------ */
 351
 352 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 353                                 struct work_struct *work)
 354 {
 355         sc_get(sc);
 356         if (!queue_work(o2net_wq, work))
 357                 sc_put(sc);
 358 }
 359 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 360                                         struct delayed_work *work,
 361                                         int delay)
 362 {
 363         sc_get(sc);
 364         if (!queue_delayed_work(o2net_wq, work, delay))
 365                 sc_put(sc);
 366 }
 367 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 368                                          struct delayed_work *work)
 369 {
 370         if (cancel_delayed_work(work))
 371                 sc_put(sc);
 372 }
 373
 374 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 375
 376 int o2net_num_connected_peers(void)
 377 {
 378         return atomic_read(&o2net_connected_peers);
 379 }
 380
 381 static void o2net_set_nn_state(struct o2net_node *nn,
 382                                struct o2net_sock_container *sc,
 383                                unsigned valid, int err)
 384 {
 385         int was_valid = nn->nn_sc_valid;
 386         int was_err = nn->nn_persistent_error;
 387         struct o2net_sock_container *old_sc = nn->nn_sc;
 388
 389         assert_spin_locked(&nn->nn_lock);
 390
 391         if (old_sc && !sc)
 392                 atomic_dec(&o2net_connected_peers);
 393         else if (!old_sc && sc)
 394                 atomic_inc(&o2net_connected_peers);
 395
 396         /* the node num comparison and single connect/accept path should stop
 397          * an non-null sc from being overwritten with another */
 398         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 399         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 400         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 401
 402         /* we won't reconnect after our valid conn goes away for
 403          * this hb iteration.. here so it shows up in the logs */
 404         if (was_valid && !valid && err == 0)
 405                 err = -ENOTCONN;
 406
 407         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 408              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 409              nn->nn_persistent_error, err);
 410
 411         nn->nn_sc = sc;
 412         nn->nn_sc_valid = valid ? 1 : 0;
 413         nn->nn_persistent_error = err;
 414
 415         /* mirrors o2net_tx_can_proceed() */
 416         if (nn->nn_persistent_error || nn->nn_sc_valid)
 417                 wake_up(&nn->nn_sc_wq);
 418
 419         if (!was_err && nn->nn_persistent_error) {
 420                 o2quo_conn_err(o2net_num_from_nn(nn));
 421                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 422                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 423         }
 424
 425         if (was_valid && !valid) {
 426                 printk(KERN_INFO "o2net: no longer connected to "
 427                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 428                 o2net_complete_nodes_nsw(nn);
 429         }
 430
 431         if (!was_valid && valid) {
 432                 o2quo_conn_up(o2net_num_from_nn(nn));
 433                 /* this is a bit of a hack.  we only try reconnecting
 434                  * when heartbeating starts until we get a connection.
 435                  * if that connection then dies we don't try reconnecting.
 436                  * the only way to start connecting again is to down
 437                  * heartbeat and bring it back up. */
 438                 cancel_delayed_work(&nn->nn_connect_expired);
 439                 printk(KERN_INFO "o2net: %s " SC_NODEF_FMT "\n",
 440                        o2nm_this_node() > sc->sc_node->nd_num ?
 441                                 "connected to" : "accepted connection from",
 442                        SC_NODEF_ARGS(sc));
 443         }
 444
 445         /* trigger the connecting worker func as long as we're not valid,
 446          * it will back off if it shouldn't connect.  This can be called
 447          * from node config teardown and so needs to be careful about
 448          * the work queue actually being up. */
 449         if (!valid && o2net_wq) {
 450                 unsigned long delay;
 451                 /* delay if we're withing a RECONNECT_DELAY of the
 452                  * last attempt */
 453                 delay = (nn->nn_last_connect_attempt +
 454                          msecs_to_jiffies(o2net_reconnect_delay(sc->sc_node)))
 455                         - jiffies;
 456                 if (delay > msecs_to_jiffies(o2net_reconnect_delay(sc->sc_node)))
 457                         delay = 0;
 458                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 459                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 460         }
 461
 462         /* keep track of the nn's sc ref for the caller */
 463         if ((old_sc == NULL) && sc)
 464                 sc_get(sc);
 465         if (old_sc && (old_sc != sc)) {
 466                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 467                 sc_put(old_sc);
 468         }
 469 }
 470
 471 /* see o2net_register_callbacks() */
 472 static void o2net_data_ready(struct sock *sk, int bytes)
 473 {
 474         void (*ready)(struct sock *sk, int bytes);
 475
 476         read_lock(&sk->sk_callback_lock);
 477         if (sk->sk_user_data) {
 478                 struct o2net_sock_container *sc = sk->sk_user_data;
 479                 sclog(sc, "data_ready hit\n");
 480                 do_gettimeofday(&sc->sc_tv_data_ready);
 481                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 482                 ready = sc->sc_data_ready;
 483         } else {
 484                 ready = sk->sk_data_ready;
 485         }
 486         read_unlock(&sk->sk_callback_lock);
 487
 488         ready(sk, bytes);
 489 }
 490
 491 /* see o2net_register_callbacks() */
 492 static void o2net_state_change(struct sock *sk)
 493 {
 494         void (*state_change)(struct sock *sk);
 495         struct o2net_sock_container *sc;
 496
 497         read_lock(&sk->sk_callback_lock);
 498         sc = sk->sk_user_data;
 499         if (sc == NULL) {
 500                 state_change = sk->sk_state_change;
 501                 goto out;
 502         }
 503
 504         sclog(sc, "state_change to %d\n", sk->sk_state);
 505
 506         state_change = sc->sc_state_change;
 507
 508         switch(sk->sk_state) {
 509                 /* ignore connecting sockets as they make progress */
 510                 case TCP_SYN_SENT:
 511                 case TCP_SYN_RECV:
 512                         break;
 513                 case TCP_ESTABLISHED:
 514                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
 515                         break;
 516                 default:
 517                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 518                         break;
 519         }
 520 out:
 521         read_unlock(&sk->sk_callback_lock);
 522         state_change(sk);
 523 }
 524
 525 /*
 526  * we register callbacks so we can queue work on events before calling
 527  * the original callbacks.  our callbacks our careful to test user_data
 528  * to discover when they've reaced with o2net_unregister_callbacks().
 529  */
 530 static void o2net_register_callbacks(struct sock *sk,
 531                                      struct o2net_sock_container *sc)
 532 {
 533         write_lock_bh(&sk->sk_callback_lock);
 534
 535         /* accepted sockets inherit the old listen socket data ready */
 536         if (sk->sk_data_ready == o2net_listen_data_ready) {
 537                 sk->sk_data_ready = sk->sk_user_data;
 538                 sk->sk_user_data = NULL;
 539         }
 540
 541         BUG_ON(sk->sk_user_data != NULL);
 542         sk->sk_user_data = sc;
 543         sc_get(sc);
 544
 545         sc->sc_data_ready = sk->sk_data_ready;
 546         sc->sc_state_change = sk->sk_state_change;
 547         sk->sk_data_ready = o2net_data_ready;
 548         sk->sk_state_change = o2net_state_change;
 549
 550         mutex_init(&sc->sc_send_lock);
 551
 552         write_unlock_bh(&sk->sk_callback_lock);
 553 }
 554
 555 static int o2net_unregister_callbacks(struct sock *sk,
 556                                    struct o2net_sock_container *sc)
 557 {
 558         int ret = 0;
 559
 560         write_lock_bh(&sk->sk_callback_lock);
 561         if (sk->sk_user_data == sc) {
 562                 ret = 1;
 563                 sk->sk_user_data = NULL;
 564                 sk->sk_data_ready = sc->sc_data_ready;
 565                 sk->sk_state_change = sc->sc_state_change;
 566         }
 567         write_unlock_bh(&sk->sk_callback_lock);
 568
 569         return ret;
 570 }
 571
 572 /*
 573  * this is a little helper that is called by callers who have seen a problem
 574  * with an sc and want to detach it from the nn if someone already hasn't beat
 575  * them to it.  if an error is given then the shutdown will be persistent
 576  * and pending transmits will be canceled.
 577  */
 578 static void o2net_ensure_shutdown(struct o2net_node *nn,
 579                                    struct o2net_sock_container *sc,
 580                                    int err)
 581 {
 582         spin_lock(&nn->nn_lock);
 583         if (nn->nn_sc == sc)
 584                 o2net_set_nn_state(nn, NULL, 0, err);
 585         spin_unlock(&nn->nn_lock);
 586 }
 587
 588 /*
 589  * This work queue function performs the blocking parts of socket shutdown.  A
 590  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 591  * sc detached from the nn.  state_change will also trigger this callback
 592  * directly when it sees errors.  In that case we need to call set_nn_state
 593  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 594  * itself.
 595  */
 596 static void o2net_shutdown_sc(struct work_struct *work)
 597 {
 598         struct o2net_sock_container *sc =
 599                 container_of(work, struct o2net_sock_container,
 600                              sc_shutdown_work);
 601         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 602
 603         sclog(sc, "shutting down\n");
 604
 605         /* drop the callbacks ref and call shutdown only once */
 606         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 607                 /* we shouldn't flush as we're in the thread, the
 608                  * races with pending sc work structs are harmless */
 609                 del_timer_sync(&sc->sc_idle_timeout);
 610                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 611                 sc_put(sc);
 612                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 613         }
 614
 615         /* not fatal so failed connects before the other guy has our
 616          * heartbeat can be retried */
 617         o2net_ensure_shutdown(nn, sc, 0);
 618         sc_put(sc);
 619 }
 620
 621 /* ------------------------------------------------------------ */
 622
 623 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 624                              u32 key)
 625 {
 626         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 627
 628         if (ret == 0)
 629                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 630
 631         return ret;
 632 }
 633
 634 static struct o2net_msg_handler *
 635 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 636                           struct rb_node **ret_parent)
 637 {
 638         struct rb_node **p = &o2net_handler_tree.rb_node;
 639         struct rb_node *parent = NULL;
 640         struct o2net_msg_handler *nmh, *ret = NULL;
 641         int cmp;
 642
 643         while (*p) {
 644                 parent = *p;
 645                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 646                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 647
 648                 if (cmp < 0)
 649                         p = &(*p)->rb_left;
 650                 else if (cmp > 0)
 651                         p = &(*p)->rb_right;
 652                 else {
 653                         ret = nmh;
 654                         break;
 655                 }
 656         }
 657
 658         if (ret_p != NULL)
 659                 *ret_p = p;
 660         if (ret_parent != NULL)
 661                 *ret_parent = parent;
 662
 663         return ret;
 664 }
 665
 666 static void o2net_handler_kref_release(struct kref *kref)
 667 {
 668         struct o2net_msg_handler *nmh;
 669         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 670
 671         kfree(nmh);
 672 }
 673
 674 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 675 {
 676         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 677 }
 678
 679 /* max_len is protection for the handler func.  incoming messages won't
 680  * be given to the handler if their payload is longer than the max. */
 681 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 682                            o2net_msg_handler_func *func, void *data,
 683                            o2net_post_msg_handler_func *post_func,
 684                            struct list_head *unreg_list)
 685 {
 686         struct o2net_msg_handler *nmh = NULL;
 687         struct rb_node **p, *parent;
 688         int ret = 0;
 689
 690         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 691                 mlog(0, "max_len for message handler out of range: %u\n",
 692                         max_len);
 693                 ret = -EINVAL;
 694                 goto out;
 695         }
 696
 697         if (!msg_type) {
 698                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 699                 ret = -EINVAL;
 700                 goto out;
 701
 702         }
 703         if (!func) {
 704                 mlog(0, "no message handler provided: %u, %p\n",
 705                        msg_type, func);
 706                 ret = -EINVAL;
 707                 goto out;
 708         }
 709
 710         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 711         if (nmh == NULL) {
 712                 ret = -ENOMEM;
 713                 goto out;
 714         }
 715
 716         nmh->nh_func = func;
 717         nmh->nh_func_data = data;
 718         nmh->nh_post_func = post_func;
 719         nmh->nh_msg_type = msg_type;
 720         nmh->nh_max_len = max_len;
 721         nmh->nh_key = key;
 722         /* the tree and list get this ref.. they're both removed in
 723          * unregister when this ref is dropped */
 724         kref_init(&nmh->nh_kref);
 725         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 726
 727         write_lock(&o2net_handler_lock);
 728         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 729                 ret = -EEXIST;
 730         else {
 731                 rb_link_node(&nmh->nh_node, parent, p);
 732                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 733                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 734
 735                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 736                      func, msg_type, key);
 737                 /* we've had some trouble with handlers seemingly vanishing. */
 738                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 739                                                           &parent) == NULL,
 740                                 "couldn't find handler we *just* registerd "
 741                                 "for type %u key %08x\n", msg_type, key);
 742         }
 743         write_unlock(&o2net_handler_lock);
 744         if (ret)
 745                 goto out;
 746
 747 out:
 748         if (ret)
 749                 kfree(nmh);
 750
 751         return ret;
 752 }
 753 EXPORT_SYMBOL_GPL(o2net_register_handler);
 754
 755 void o2net_unregister_handler_list(struct list_head *list)
 756 {
 757         struct o2net_msg_handler *nmh, *n;
 758
 759         write_lock(&o2net_handler_lock);
 760         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 761                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 762                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 763                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 764                 list_del_init(&nmh->nh_unregister_item);
 765                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 766         }
 767         write_unlock(&o2net_handler_lock);
 768 }
 769 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 770
 771 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 772 {
 773         struct o2net_msg_handler *nmh;
 774
 775         read_lock(&o2net_handler_lock);
 776         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 777         if (nmh)
 778                 kref_get(&nmh->nh_kref);
 779         read_unlock(&o2net_handler_lock);
 780
 781         return nmh;
 782 }
 783
 784 /* ------------------------------------------------------------ */
 785
 786 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 787 {
 788         int ret;
 789         mm_segment_t oldfs;
 790         struct kvec vec = {
 791                 .iov_len = len,
 792                 .iov_base = data,
 793         };
 794         struct msghdr msg = {
 795                 .msg_iovlen = 1,
 796                 .msg_iov = (struct iovec *)&vec,
 797                 .msg_flags = MSG_DONTWAIT,
 798         };
 799
 800         oldfs = get_fs();
 801         set_fs(get_ds());
 802         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
 803         set_fs(oldfs);
 804
 805         return ret;
 806 }
 807
 808 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 809                               size_t veclen, size_t total)
 810 {
 811         int ret;
 812         mm_segment_t oldfs;
 813         struct msghdr msg = {
 814                 .msg_iov = (struct iovec *)vec,
 815                 .msg_iovlen = veclen,
 816         };
 817
 818         if (sock == NULL) {
 819                 ret = -EINVAL;
 820                 goto out;
 821         }
 822
 823         oldfs = get_fs();
 824         set_fs(get_ds());
 825         ret = sock_sendmsg(sock, &msg, total);
 826         set_fs(oldfs);
 827         if (ret != total) {
 828                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
 829                      total);
 830                 if (ret >= 0)
 831                         ret = -EPIPE; /* should be smarter, I bet */
 832                 goto out;
 833         }
 834
 835         ret = 0;
 836 out:
 837         if (ret < 0)
 838                 mlog(0, "returning error: %d\n", ret);
 839         return ret;
 840 }
 841
 842 static void o2net_sendpage(struct o2net_sock_container *sc,
 843                            void *kmalloced_virt,
 844                            size_t size)
 845 {
 846         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 847         ssize_t ret;
 848
 849         while (1) {
 850                 mutex_lock(&sc->sc_send_lock);
 851                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 852                                                  virt_to_page(kmalloced_virt),
 853                                                  (long)kmalloced_virt & ~PAGE_MASK,
 854                                                  size, MSG_DONTWAIT);
 855                 mutex_unlock(&sc->sc_send_lock);
 856                 if (ret == size)
 857                         break;
 858                 if (ret == (ssize_t)-EAGAIN) {
 859                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 860                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 861                         cond_resched();
 862                         continue;
 863                 }
 864                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 865                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 866                 o2net_ensure_shutdown(nn, sc, 0);
 867                 break;
 868         }
 869 }
 870
 871 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 872 {
 873         memset(msg, 0, sizeof(struct o2net_msg));
 874         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 875         msg->data_len = cpu_to_be16(data_len);
 876         msg->msg_type = cpu_to_be16(msg_type);
 877         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 878         msg->status = 0;
 879         msg->key = cpu_to_be32(key);
 880 }
 881
 882 static int o2net_tx_can_proceed(struct o2net_node *nn,
 883                                 struct o2net_sock_container **sc_ret,
 884                                 int *error)
 885 {
 886         int ret = 0;
 887
 888         spin_lock(&nn->nn_lock);
 889         if (nn->nn_persistent_error) {
 890                 ret = 1;
 891                 *sc_ret = NULL;
 892                 *error = nn->nn_persistent_error;
 893         } else if (nn->nn_sc_valid) {
 894                 kref_get(&nn->nn_sc->sc_kref);
 895
 896                 ret = 1;
 897                 *sc_ret = nn->nn_sc;
 898                 *error = 0;
 899         }
 900         spin_unlock(&nn->nn_lock);
 901
 902         return ret;
 903 }
 904
 905 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
 906                            size_t caller_veclen, u8 target_node, int *status)
 907 {
 908         int ret, error = 0;
 909         struct o2net_msg *msg = NULL;
 910         size_t veclen, caller_bytes = 0;
 911         struct kvec *vec = NULL;
 912         struct o2net_sock_container *sc = NULL;
 913         struct o2net_node *nn = o2net_nn_from_num(target_node);
 914         struct o2net_status_wait nsw = {
 915                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
 916         };
 917
 918         if (o2net_wq == NULL) {
 919                 mlog(0, "attempt to tx without o2netd running\n");
 920                 ret = -ESRCH;
 921                 goto out;
 922         }
 923
 924         if (caller_veclen == 0) {
 925                 mlog(0, "bad kvec array length\n");
 926                 ret = -EINVAL;
 927                 goto out;
 928         }
 929
 930         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
 931         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
 932                 mlog(0, "total payload len %zu too large\n", caller_bytes);
 933                 ret = -EINVAL;
 934                 goto out;
 935         }
 936
 937         if (target_node == o2nm_this_node()) {
 938                 ret = -ELOOP;
 939                 goto out;
 940         }
 941
 942         ret = wait_event_interruptible(nn->nn_sc_wq,
 943                                        o2net_tx_can_proceed(nn, &sc, &error));
 944         if (!ret && error)
 945                 ret = error;
 946         if (ret)
 947                 goto out;
 948
 949         veclen = caller_veclen + 1;
 950         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
 951         if (vec == NULL) {
 952                 mlog(0, "failed to %zu element kvec!\n", veclen);
 953                 ret = -ENOMEM;
 954                 goto out;
 955         }
 956
 957         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
 958         if (!msg) {
 959                 mlog(0, "failed to allocate a o2net_msg!\n");
 960                 ret = -ENOMEM;
 961                 goto out;
 962         }
 963
 964         o2net_init_msg(msg, caller_bytes, msg_type, key);
 965
 966         vec[0].iov_len = sizeof(struct o2net_msg);
 967         vec[0].iov_base = msg;
 968         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
 969
 970         ret = o2net_prep_nsw(nn, &nsw);
 971         if (ret)
 972                 goto out;
 973
 974         msg->msg_num = cpu_to_be32(nsw.ns_id);
 975
 976         /* finally, convert the message header to network byte-order
 977          * and send */
 978         mutex_lock(&sc->sc_send_lock);
 979         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
 980                                  sizeof(struct o2net_msg) + caller_bytes);
 981         mutex_unlock(&sc->sc_send_lock);
 982         msglog(msg, "sending returned %d\n", ret);
 983         if (ret < 0) {
 984                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
 985                 goto out;
 986         }
 987
 988         /* wait on other node's handler */
 989         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
 990
 991         /* Note that we avoid overwriting the callers status return
 992          * variable if a system error was reported on the other
 993          * side. Callers beware. */
 994         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
 995         if (status && !ret)
 996                 *status = nsw.ns_status;
 997
 998         mlog(0, "woken, returning system status %d, user status %d\n",
 999              ret, nsw.ns_status);
1000 out:
1001         if (sc)
1002                 sc_put(sc);
1003         if (vec)
1004                 kfree(vec);
1005         if (msg)
1006                 kfree(msg);
1007         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1008         return ret;
1009 }
1010 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1011
1012 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1013                        u8 target_node, int *status)
1014 {
1015         struct kvec vec = {
1016                 .iov_base = data,
1017                 .iov_len = len,
1018         };
1019         return o2net_send_message_vec(msg_type, key, &vec, 1,
1020                                       target_node, status);
1021 }
1022 EXPORT_SYMBOL_GPL(o2net_send_message);
1023
1024 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1025                                    enum o2net_system_error syserr, int err)
1026 {
1027         struct kvec vec = {
1028                 .iov_base = hdr,
1029                 .iov_len = sizeof(struct o2net_msg),
1030         };
1031
1032         BUG_ON(syserr >= O2NET_ERR_MAX);
1033
1034         /* leave other fields intact from the incoming message, msg_num
1035          * in particular */
1036         hdr->sys_status = cpu_to_be32(syserr);
1037         hdr->status = cpu_to_be32(err);
1038         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1039         hdr->data_len = 0;
1040
1041         msglog(hdr, "about to send status magic %d\n", err);
1042         /* hdr has been in host byteorder this whole time */
1043         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1044 }
1045
1046 /* this returns -errno if the header was unknown or too large, etc.
1047  * after this is called the buffer us reused for the next message */
1048 static int o2net_process_message(struct o2net_sock_container *sc,
1049                                  struct o2net_msg *hdr)
1050 {
1051         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1052         int ret = 0, handler_status;
1053         enum  o2net_system_error syserr;
1054         struct o2net_msg_handler *nmh = NULL;
1055         void *ret_data = NULL;
1056
1057         msglog(hdr, "processing message\n");
1058
1059         o2net_sc_postpone_idle(sc);
1060
1061         switch(be16_to_cpu(hdr->magic)) {
1062                 case O2NET_MSG_STATUS_MAGIC:
1063                         /* special type for returning message status */
1064                         o2net_complete_nsw(nn, NULL,
1065                                            be32_to_cpu(hdr->msg_num),
1066                                            be32_to_cpu(hdr->sys_status),
1067                                            be32_to_cpu(hdr->status));
1068                         goto out;
1069                 case O2NET_MSG_KEEP_REQ_MAGIC:
1070                         o2net_sendpage(sc, o2net_keep_resp,
1071                                        sizeof(*o2net_keep_resp));
1072                         goto out;
1073                 case O2NET_MSG_KEEP_RESP_MAGIC:
1074                         goto out;
1075                 case O2NET_MSG_MAGIC:
1076                         break;
1077                 default:
1078                         msglog(hdr, "bad magic\n");
1079                         ret = -EINVAL;
1080                         goto out;
1081                         break;
1082         }
1083
1084         /* find a handler for it */
1085         handler_status = 0;
1086         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1087                                 be32_to_cpu(hdr->key));
1088         if (!nmh) {
1089                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1090                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1091                 syserr = O2NET_ERR_NO_HNDLR;
1092                 goto out_respond;
1093         }
1094
1095         syserr = O2NET_ERR_NONE;
1096
1097         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1098                 syserr = O2NET_ERR_OVERFLOW;
1099
1100         if (syserr != O2NET_ERR_NONE)
1101                 goto out_respond;
1102
1103         do_gettimeofday(&sc->sc_tv_func_start);
1104         sc->sc_msg_key = be32_to_cpu(hdr->key);
1105         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1106         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1107                                              be16_to_cpu(hdr->data_len),
1108                                         nmh->nh_func_data, &ret_data);
1109         do_gettimeofday(&sc->sc_tv_func_stop);
1110
1111 out_respond:
1112         /* this destroys the hdr, so don't use it after this */
1113         mutex_lock(&sc->sc_send_lock);
1114         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1115                                       handler_status);
1116         mutex_unlock(&sc->sc_send_lock);
1117         hdr = NULL;
1118         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1119              handler_status, syserr, ret);
1120
1121         if (nmh) {
1122                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1123                 if (nmh->nh_post_func)
1124                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1125                                             ret_data);
1126         }
1127
1128 out:
1129         if (nmh)
1130                 o2net_handler_put(nmh);
1131         return ret;
1132 }
1133
1134 static int o2net_check_handshake(struct o2net_sock_container *sc)
1135 {
1136         struct o2net_handshake *hand = page_address(sc->sc_page);
1137         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1138
1139         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1140                 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1141                      "version %llu but %llu is required, disconnecting\n",
1142                      SC_NODEF_ARGS(sc),
1143                      (unsigned long long)be64_to_cpu(hand->protocol_version),
1144                      O2NET_PROTOCOL_VERSION);
1145
1146                 /* don't bother reconnecting if its the wrong version. */
1147                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1148                 return -1;
1149         }
1150
1151         /*
1152          * Ensure timeouts are consistent with other nodes, otherwise
1153          * we can end up with one node thinking that the other must be down,
1154          * but isn't. This can ultimately cause corruption.
1155          */
1156         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1157                                 o2net_idle_timeout(sc->sc_node)) {
1158                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1159                      "%u ms, but we use %u ms locally.  disconnecting\n",
1160                      SC_NODEF_ARGS(sc),
1161                      be32_to_cpu(hand->o2net_idle_timeout_ms),
1162                      o2net_idle_timeout(sc->sc_node));
1163                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1164                 return -1;
1165         }
1166
1167         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1168                         o2net_keepalive_delay(sc->sc_node)) {
1169                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1170                      "%u ms, but we use %u ms locally.  disconnecting\n",
1171                      SC_NODEF_ARGS(sc),
1172                      be32_to_cpu(hand->o2net_keepalive_delay_ms),
1173                      o2net_keepalive_delay(sc->sc_node));
1174                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1175                 return -1;
1176         }
1177
1178         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1179                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1180                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1181                      "%u ms, but we use %u ms locally.  disconnecting\n",
1182                      SC_NODEF_ARGS(sc),
1183                      be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1184                      O2HB_MAX_WRITE_TIMEOUT_MS);
1185                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1186                 return -1;
1187         }
1188
1189         sc->sc_handshake_ok = 1;
1190
1191         spin_lock(&nn->nn_lock);
1192         /* set valid and queue the idle timers only if it hasn't been
1193          * shut down already */
1194         if (nn->nn_sc == sc) {
1195                 o2net_sc_reset_idle_timer(sc);
1196                 o2net_set_nn_state(nn, sc, 1, 0);
1197         }
1198         spin_unlock(&nn->nn_lock);
1199
1200         /* shift everything up as though it wasn't there */
1201         sc->sc_page_off -= sizeof(struct o2net_handshake);
1202         if (sc->sc_page_off)
1203                 memmove(hand, hand + 1, sc->sc_page_off);
1204
1205         return 0;
1206 }
1207
1208 /* this demuxes the queued rx bytes into header or payload bits and calls
1209  * handlers as each full message is read off the socket.  it returns -error,
1210  * == 0 eof, or > 0 for progress made.*/
1211 static int o2net_advance_rx(struct o2net_sock_container *sc)
1212 {
1213         struct o2net_msg *hdr;
1214         int ret = 0;
1215         void *data;
1216         size_t datalen;
1217
1218         sclog(sc, "receiving\n");
1219         do_gettimeofday(&sc->sc_tv_advance_start);
1220
1221         if (unlikely(sc->sc_handshake_ok == 0)) {
1222                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1223                         data = page_address(sc->sc_page) + sc->sc_page_off;
1224                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1225                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1226                         if (ret > 0)
1227                                 sc->sc_page_off += ret;
1228                 }
1229
1230                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1231                         o2net_check_handshake(sc);
1232                         if (unlikely(sc->sc_handshake_ok == 0))
1233                                 ret = -EPROTO;
1234                 }
1235                 goto out;
1236         }
1237
1238         /* do we need more header? */
1239         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1240                 data = page_address(sc->sc_page) + sc->sc_page_off;
1241                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1242                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1243                 if (ret > 0) {
1244                         sc->sc_page_off += ret;
1245                         /* only swab incoming here.. we can
1246                          * only get here once as we cross from
1247                          * being under to over */
1248                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1249                                 hdr = page_address(sc->sc_page);
1250                                 if (be16_to_cpu(hdr->data_len) >
1251                                     O2NET_MAX_PAYLOAD_BYTES)
1252                                         ret = -EOVERFLOW;
1253                         }
1254                 }
1255                 if (ret <= 0)
1256                         goto out;
1257         }
1258
1259         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1260                 /* oof, still don't have a header */
1261                 goto out;
1262         }
1263
1264         /* this was swabbed above when we first read it */
1265         hdr = page_address(sc->sc_page);
1266
1267         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1268
1269         /* do we need more payload? */
1270         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1271                 /* need more payload */
1272                 data = page_address(sc->sc_page) + sc->sc_page_off;
1273                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1274                           sc->sc_page_off;
1275                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1276                 if (ret > 0)
1277                         sc->sc_page_off += ret;
1278                 if (ret <= 0)
1279                         goto out;
1280         }
1281
1282         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1283                 /* we can only get here once, the first time we read
1284                  * the payload.. so set ret to progress if the handler
1285                  * works out. after calling this the message is toast */
1286                 ret = o2net_process_message(sc, hdr);
1287                 if (ret == 0)
1288                         ret = 1;
1289                 sc->sc_page_off = 0;
1290         }
1291
1292 out:
1293         sclog(sc, "ret = %d\n", ret);
1294         do_gettimeofday(&sc->sc_tv_advance_stop);
1295         return ret;
1296 }
1297
1298 /* this work func is triggerd by data ready.  it reads until it can read no
1299  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1300  * our work the work struct will be marked and we'll be called again. */
1301 static void o2net_rx_until_empty(struct work_struct *work)
1302 {
1303         struct o2net_sock_container *sc =
1304                 container_of(work, struct o2net_sock_container, sc_rx_work);
1305         int ret;
1306
1307         do {
1308                 ret = o2net_advance_rx(sc);
1309         } while (ret > 0);
1310
1311         if (ret <= 0 && ret != -EAGAIN) {
1312                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1313                 sclog(sc, "saw error %d, closing\n", ret);
1314                 /* not permanent so read failed handshake can retry */
1315                 o2net_ensure_shutdown(nn, sc, 0);
1316         }
1317
1318         sc_put(sc);
1319 }
1320
1321 static int o2net_set_nodelay(struct socket *sock)
1322 {
1323         int ret, val = 1;
1324         mm_segment_t oldfs;
1325
1326         oldfs = get_fs();
1327         set_fs(KERNEL_DS);
1328
1329         /*
1330          * Dear unsuspecting programmer,
1331          *
1332          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1333          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1334          * silently turn into SO_DEBUG.
1335          *
1336          * Yours,
1337          * Keeper of hilariously fragile interfaces.
1338          */
1339         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1340                                     (char __user *)&val, sizeof(val));
1341
1342         set_fs(oldfs);
1343         return ret;
1344 }
1345
1346 static void o2net_initialize_handshake(void)
1347 {
1348         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1349                 O2HB_MAX_WRITE_TIMEOUT_MS);
1350         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(
1351                 o2net_idle_timeout(NULL));
1352         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1353                 o2net_keepalive_delay(NULL));
1354         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1355                 o2net_reconnect_delay(NULL));
1356 }
1357
1358 /* ------------------------------------------------------------ */
1359
1360 /* called when a connect completes and after a sock is accepted.  the
1361  * rx path will see the response and mark the sc valid */
1362 static void o2net_sc_connect_completed(struct work_struct *work)
1363 {
1364         struct o2net_sock_container *sc =
1365                 container_of(work, struct o2net_sock_container,
1366                              sc_connect_work);
1367
1368         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1369               (unsigned long long)O2NET_PROTOCOL_VERSION,
1370               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1371
1372         o2net_initialize_handshake();
1373         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1374         sc_put(sc);
1375 }
1376
1377 /* this is called as a work_struct func. */
1378 static void o2net_sc_send_keep_req(struct work_struct *work)
1379 {
1380         struct o2net_sock_container *sc =
1381                 container_of(work, struct o2net_sock_container,
1382                              sc_keepalive_work.work);
1383
1384         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1385         sc_put(sc);
1386 }
1387
1388 /* socket shutdown does a del_timer_sync against this as it tears down.
1389  * we can't start this timer until we've got to the point in sc buildup
1390  * where shutdown is going to be involved */
1391 static void o2net_idle_timer(unsigned long data)
1392 {
1393         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1394         struct timeval now;
1395
1396         do_gettimeofday(&now);
1397
1398         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1399              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1400                      o2net_idle_timeout(sc->sc_node) / 1000,
1401                      o2net_idle_timeout(sc->sc_node) % 1000);
1402         mlog(ML_NOTICE, "here are some times that might help debug the "
1403              "situation: (tmr %ld.%ld now %ld.%ld dr %ld.%ld adv "
1404              "%ld.%ld:%ld.%ld func (%08x:%u) %ld.%ld:%ld.%ld)\n",
1405              sc->sc_tv_timer.tv_sec, (long) sc->sc_tv_timer.tv_usec,
1406              now.tv_sec, (long) now.tv_usec,
1407              sc->sc_tv_data_ready.tv_sec, (long) sc->sc_tv_data_ready.tv_usec,
1408              sc->sc_tv_advance_start.tv_sec,
1409              (long) sc->sc_tv_advance_start.tv_usec,
1410              sc->sc_tv_advance_stop.tv_sec,
1411              (long) sc->sc_tv_advance_stop.tv_usec,
1412              sc->sc_msg_key, sc->sc_msg_type,
1413              sc->sc_tv_func_start.tv_sec, (long) sc->sc_tv_func_start.tv_usec,
1414              sc->sc_tv_func_stop.tv_sec, (long) sc->sc_tv_func_stop.tv_usec);
1415
1416         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1417 }
1418
1419 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1420 {
1421         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1422         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1423                       msecs_to_jiffies(o2net_keepalive_delay(sc->sc_node)));
1424         do_gettimeofday(&sc->sc_tv_timer);
1425         mod_timer(&sc->sc_idle_timeout,
1426                jiffies + msecs_to_jiffies(o2net_idle_timeout(sc->sc_node)));
1427 }
1428
1429 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1430 {
1431         /* Only push out an existing timer */
1432         if (timer_pending(&sc->sc_idle_timeout))
1433                 o2net_sc_reset_idle_timer(sc);
1434 }
1435
1436 /* this work func is kicked whenever a path sets the nn state which doesn't
1437  * have valid set.  This includes seeing hb come up, losing a connection,
1438  * having a connect attempt fail, etc. This centralizes the logic which decides
1439  * if a connect attempt should be made or if we should give up and all future
1440  * transmit attempts should fail */
1441 static void o2net_start_connect(struct work_struct *work)
1442 {
1443         struct o2net_node *nn =
1444                 container_of(work, struct o2net_node, nn_connect_work.work);
1445         struct o2net_sock_container *sc = NULL;
1446         struct o2nm_node *node = NULL, *mynode = NULL;
1447         struct socket *sock = NULL;
1448         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1449         int ret = 0, stop;
1450
1451         /* if we're greater we initiate tx, otherwise we accept */
1452         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1453                 goto out;
1454
1455         /* watch for racing with tearing a node down */
1456         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1457         if (node == NULL) {
1458                 ret = 0;
1459                 goto out;
1460         }
1461
1462         mynode = o2nm_get_node_by_num(o2nm_this_node());
1463         if (mynode == NULL) {
1464                 ret = 0;
1465                 goto out;
1466         }
1467
1468         spin_lock(&nn->nn_lock);
1469         /* see if we already have one pending or have given up */
1470         stop = (nn->nn_sc || nn->nn_persistent_error);
1471         spin_unlock(&nn->nn_lock);
1472         if (stop)
1473                 goto out;
1474
1475         nn->nn_last_connect_attempt = jiffies;
1476
1477         sc = sc_alloc(node);
1478         if (sc == NULL) {
1479                 mlog(0, "couldn't allocate sc\n");
1480                 ret = -ENOMEM;
1481                 goto out;
1482         }
1483
1484         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1485         if (ret < 0) {
1486                 mlog(0, "can't create socket: %d\n", ret);
1487                 goto out;
1488         }
1489         sc->sc_sock = sock; /* freed by sc_kref_release */
1490
1491         sock->sk->sk_allocation = GFP_ATOMIC;
1492
1493         myaddr.sin_family = AF_INET;
1494         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1495         myaddr.sin_port = htons(0); /* any port */
1496
1497         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1498                               sizeof(myaddr));
1499         if (ret) {
1500                 mlog(ML_ERROR, "bind failed with %d at address %u.%u.%u.%u\n",
1501                      ret, NIPQUAD(mynode->nd_ipv4_address));
1502                 goto out;
1503         }
1504
1505         ret = o2net_set_nodelay(sc->sc_sock);
1506         if (ret) {
1507                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1508                 goto out;
1509         }
1510
1511         o2net_register_callbacks(sc->sc_sock->sk, sc);
1512
1513         spin_lock(&nn->nn_lock);
1514         /* handshake completion will set nn->nn_sc_valid */
1515         o2net_set_nn_state(nn, sc, 0, 0);
1516         spin_unlock(&nn->nn_lock);
1517
1518         remoteaddr.sin_family = AF_INET;
1519         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1520         remoteaddr.sin_port = node->nd_ipv4_port;
1521
1522         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1523                                         (struct sockaddr *)&remoteaddr,
1524                                         sizeof(remoteaddr),
1525                                         O_NONBLOCK);
1526         if (ret == -EINPROGRESS)
1527                 ret = 0;
1528
1529 out:
1530         if (ret) {
1531                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1532                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1533                 /* 0 err so that another will be queued and attempted
1534                  * from set_nn_state */
1535                 if (sc)
1536                         o2net_ensure_shutdown(nn, sc, 0);
1537         }
1538         if (sc)
1539                 sc_put(sc);
1540         if (node)
1541                 o2nm_node_put(node);
1542         if (mynode)
1543                 o2nm_node_put(mynode);
1544
1545         return;
1546 }
1547
1548 static void o2net_connect_expired(struct work_struct *work)
1549 {
1550         struct o2net_node *nn =
1551                 container_of(work, struct o2net_node, nn_connect_expired.work);
1552
1553         spin_lock(&nn->nn_lock);
1554         if (!nn->nn_sc_valid) {
1555                 struct o2nm_node *node = nn->nn_sc->sc_node;
1556                 mlog(ML_ERROR, "no connection established with node %u after "
1557                      "%u.%u seconds, giving up and returning errors.\n",
1558                      o2net_num_from_nn(nn),
1559                      o2net_idle_timeout(node) / 1000,
1560                      o2net_idle_timeout(node) % 1000);
1561
1562                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1563         }
1564         spin_unlock(&nn->nn_lock);
1565 }
1566
1567 static void o2net_still_up(struct work_struct *work)
1568 {
1569         struct o2net_node *nn =
1570                 container_of(work, struct o2net_node, nn_still_up.work);
1571
1572         o2quo_hb_still_up(o2net_num_from_nn(nn));
1573 }
1574
1575 /* ------------------------------------------------------------ */
1576
1577 void o2net_disconnect_node(struct o2nm_node *node)
1578 {
1579         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1580
1581         /* don't reconnect until it's heartbeating again */
1582         spin_lock(&nn->nn_lock);
1583         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1584         spin_unlock(&nn->nn_lock);
1585
1586         if (o2net_wq) {
1587                 cancel_delayed_work(&nn->nn_connect_expired);
1588                 cancel_delayed_work(&nn->nn_connect_work);
1589                 cancel_delayed_work(&nn->nn_still_up);
1590                 flush_workqueue(o2net_wq);
1591         }
1592 }
1593
1594 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1595                                   void *data)
1596 {
1597         o2quo_hb_down(node_num);
1598
1599         if (node_num != o2nm_this_node())
1600                 o2net_disconnect_node(node);
1601
1602         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1603 }
1604
1605 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1606                                 void *data)
1607 {
1608         struct o2net_node *nn = o2net_nn_from_num(node_num);
1609
1610         o2quo_hb_up(node_num);
1611
1612         /* ensure an immediate connect attempt */
1613         nn->nn_last_connect_attempt = jiffies -
1614                 (msecs_to_jiffies(o2net_reconnect_delay(node)) + 1);
1615
1616         if (node_num != o2nm_this_node()) {
1617                 /* heartbeat doesn't work unless a local node number is
1618                  * configured and doing so brings up the o2net_wq, so we can
1619                  * use it.. */
1620                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired,
1621                                    msecs_to_jiffies(o2net_idle_timeout(node)));
1622
1623                 /* believe it or not, accept and node hearbeating testing
1624                  * can succeed for this node before we got here.. so
1625                  * only use set_nn_state to clear the persistent error
1626                  * if that hasn't already happened */
1627                 spin_lock(&nn->nn_lock);
1628                 if (nn->nn_persistent_error)
1629                         o2net_set_nn_state(nn, NULL, 0, 0);
1630                 spin_unlock(&nn->nn_lock);
1631         }
1632 }
1633
1634 void o2net_unregister_hb_callbacks(void)
1635 {
1636         o2hb_unregister_callback(NULL, &o2net_hb_up);
1637         o2hb_unregister_callback(NULL, &o2net_hb_down);
1638 }
1639
1640 int o2net_register_hb_callbacks(void)
1641 {
1642         int ret;
1643
1644         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1645                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1646         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1647                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1648
1649         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1650         if (ret == 0)
1651                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1652
1653         if (ret)
1654                 o2net_unregister_hb_callbacks();
1655
1656         return ret;
1657 }
1658
1659 /* ------------------------------------------------------------ */
1660
1661 static int o2net_accept_one(struct socket *sock)
1662 {
1663         int ret, slen;
1664         struct sockaddr_in sin;
1665         struct socket *new_sock = NULL;
1666         struct o2nm_node *node = NULL;
1667         struct o2net_sock_container *sc = NULL;
1668         struct o2net_node *nn;
1669
1670         BUG_ON(sock == NULL);
1671         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1672                                sock->sk->sk_protocol, &new_sock);
1673         if (ret)
1674                 goto out;
1675
1676         new_sock->type = sock->type;
1677         new_sock->ops = sock->ops;
1678         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1679         if (ret < 0)
1680                 goto out;
1681
1682         new_sock->sk->sk_allocation = GFP_ATOMIC;
1683
1684         ret = o2net_set_nodelay(new_sock);
1685         if (ret) {
1686                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1687                 goto out;
1688         }
1689
1690         slen = sizeof(sin);
1691         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1692                                        &slen, 1);
1693         if (ret < 0)
1694                 goto out;
1695
1696         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1697         if (node == NULL) {
1698                 mlog(ML_NOTICE, "attempt to connect from unknown node at "
1699                      "%u.%u.%u.%u:%d\n", NIPQUAD(sin.sin_addr.s_addr),
1700                      ntohs(sin.sin_port));
1701                 ret = -EINVAL;
1702                 goto out;
1703         }
1704
1705         if (o2nm_this_node() > node->nd_num) {
1706                 mlog(ML_NOTICE, "unexpected connect attempted from a lower "
1707                      "numbered node '%s' at " "%u.%u.%u.%u:%d with num %u\n",
1708                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1709                      ntohs(sin.sin_port), node->nd_num);
1710                 ret = -EINVAL;
1711                 goto out;
1712         }
1713
1714         /* this happens all the time when the other node sees our heartbeat
1715          * and tries to connect before we see their heartbeat */
1716         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1717                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1718                      "%u.%u.%u.%u:%d but it isn't heartbeating\n",
1719                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1720                      ntohs(sin.sin_port));
1721                 ret = -EINVAL;
1722                 goto out;
1723         }
1724
1725         nn = o2net_nn_from_num(node->nd_num);
1726
1727         spin_lock(&nn->nn_lock);
1728         if (nn->nn_sc)
1729                 ret = -EBUSY;
1730         else
1731                 ret = 0;
1732         spin_unlock(&nn->nn_lock);
1733         if (ret) {
1734                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1735                      "%u.%u.%u.%u:%d but it already has an open connection\n",
1736                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1737                      ntohs(sin.sin_port));
1738                 goto out;
1739         }
1740
1741         sc = sc_alloc(node);
1742         if (sc == NULL) {
1743                 ret = -ENOMEM;
1744                 goto out;
1745         }
1746
1747         sc->sc_sock = new_sock;
1748         new_sock = NULL;
1749
1750         spin_lock(&nn->nn_lock);
1751         o2net_set_nn_state(nn, sc, 0, 0);
1752         spin_unlock(&nn->nn_lock);
1753
1754         o2net_register_callbacks(sc->sc_sock->sk, sc);
1755         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1756
1757         o2net_initialize_handshake();
1758         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1759
1760 out:
1761         if (new_sock)
1762                 sock_release(new_sock);
1763         if (node)
1764                 o2nm_node_put(node);
1765         if (sc)
1766                 sc_put(sc);
1767         return ret;
1768 }
1769
1770 static void o2net_accept_many(struct work_struct *work)
1771 {
1772         struct socket *sock = o2net_listen_sock;
1773         while (o2net_accept_one(sock) == 0)
1774                 cond_resched();
1775 }
1776
1777 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1778 {
1779         void (*ready)(struct sock *sk, int bytes);
1780
1781         read_lock(&sk->sk_callback_lock);
1782         ready = sk->sk_user_data;
1783         if (ready == NULL) { /* check for teardown race */
1784                 ready = sk->sk_data_ready;
1785                 goto out;
1786         }
1787
1788         /* ->sk_data_ready is also called for a newly established child socket
1789          * before it has been accepted and the acceptor has set up their
1790          * data_ready.. we only want to queue listen work for our listening
1791          * socket */
1792         if (sk->sk_state == TCP_LISTEN) {
1793                 mlog(ML_TCP, "bytes: %d\n", bytes);
1794                 queue_work(o2net_wq, &o2net_listen_work);
1795         }
1796
1797 out:
1798         read_unlock(&sk->sk_callback_lock);
1799         ready(sk, bytes);
1800 }
1801
1802 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1803 {
1804         struct socket *sock = NULL;
1805         int ret;
1806         struct sockaddr_in sin = {
1807                 .sin_family = PF_INET,
1808                 .sin_addr = { .s_addr = addr },
1809                 .sin_port = port,
1810         };
1811
1812         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1813         if (ret < 0) {
1814                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1815                 goto out;
1816         }
1817
1818         sock->sk->sk_allocation = GFP_ATOMIC;
1819
1820         write_lock_bh(&sock->sk->sk_callback_lock);
1821         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1822         sock->sk->sk_data_ready = o2net_listen_data_ready;
1823         write_unlock_bh(&sock->sk->sk_callback_lock);
1824
1825         o2net_listen_sock = sock;
1826         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1827
1828         sock->sk->sk_reuse = 1;
1829         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1830         if (ret < 0) {
1831                 mlog(ML_ERROR, "unable to bind socket at %u.%u.%u.%u:%u, "
1832                      "ret=%d\n", NIPQUAD(addr), ntohs(port), ret);
1833                 goto out;
1834         }
1835
1836         ret = sock->ops->listen(sock, 64);
1837         if (ret < 0) {
1838                 mlog(ML_ERROR, "unable to listen on %u.%u.%u.%u:%u, ret=%d\n",
1839                      NIPQUAD(addr), ntohs(port), ret);
1840         }
1841
1842 out:
1843         if (ret) {
1844                 o2net_listen_sock = NULL;
1845                 if (sock)
1846                         sock_release(sock);
1847         }
1848         return ret;
1849 }
1850
1851 /*
1852  * called from node manager when we should bring up our network listening
1853  * socket.  node manager handles all the serialization to only call this
1854  * once and to match it with o2net_stop_listening().  note,
1855  * o2nm_this_node() doesn't work yet as we're being called while it
1856  * is being set up.
1857  */
1858 int o2net_start_listening(struct o2nm_node *node)
1859 {
1860         int ret = 0;
1861
1862         BUG_ON(o2net_wq != NULL);
1863         BUG_ON(o2net_listen_sock != NULL);
1864
1865         mlog(ML_KTHREAD, "starting o2net thread...\n");
1866         o2net_wq = create_singlethread_workqueue("o2net");
1867         if (o2net_wq == NULL) {
1868                 mlog(ML_ERROR, "unable to launch o2net thread\n");
1869                 return -ENOMEM; /* ? */
1870         }
1871
1872         ret = o2net_open_listening_sock(node->nd_ipv4_address,
1873                                         node->nd_ipv4_port);
1874         if (ret) {
1875                 destroy_workqueue(o2net_wq);
1876                 o2net_wq = NULL;
1877         } else
1878                 o2quo_conn_up(node->nd_num);
1879
1880         return ret;
1881 }
1882
1883 /* again, o2nm_this_node() doesn't work here as we're involved in
1884  * tearing it down */
1885 void o2net_stop_listening(struct o2nm_node *node)
1886 {
1887         struct socket *sock = o2net_listen_sock;
1888         size_t i;
1889
1890         BUG_ON(o2net_wq == NULL);
1891         BUG_ON(o2net_listen_sock == NULL);
1892
1893         /* stop the listening socket from generating work */
1894         write_lock_bh(&sock->sk->sk_callback_lock);
1895         sock->sk->sk_data_ready = sock->sk->sk_user_data;
1896         sock->sk->sk_user_data = NULL;
1897         write_unlock_bh(&sock->sk->sk_callback_lock);
1898
1899         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1900                 struct o2nm_node *node = o2nm_get_node_by_num(i);
1901                 if (node) {
1902                         o2net_disconnect_node(node);
1903                         o2nm_node_put(node);
1904                 }
1905         }
1906
1907         /* finish all work and tear down the work queue */
1908         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
1909         destroy_workqueue(o2net_wq);
1910         o2net_wq = NULL;
1911
1912         sock_release(o2net_listen_sock);
1913         o2net_listen_sock = NULL;
1914
1915         o2quo_conn_err(node->nd_num);
1916 }
1917
1918 /* ------------------------------------------------------------ */
1919
1920 int o2net_init(void)
1921 {
1922         unsigned long i;
1923
1924         o2quo_init();
1925
1926         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
1927         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1928         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1929         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
1930                 kfree(o2net_hand);
1931                 kfree(o2net_keep_req);
1932                 kfree(o2net_keep_resp);
1933                 return -ENOMEM;
1934         }
1935
1936         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
1937         o2net_hand->connector_id = cpu_to_be64(1);
1938
1939         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
1940         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
1941
1942         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1943                 struct o2net_node *nn = o2net_nn_from_num(i);
1944
1945                 spin_lock_init(&nn->nn_lock);
1946                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
1947                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
1948                                   o2net_connect_expired);
1949                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
1950                 /* until we see hb from a node we'll return einval */
1951                 nn->nn_persistent_error = -ENOTCONN;
1952                 init_waitqueue_head(&nn->nn_sc_wq);
1953                 idr_init(&nn->nn_status_idr);
1954                 INIT_LIST_HEAD(&nn->nn_status_list);
1955         }
1956
1957         return 0;
1958 }
1959
1960 void o2net_exit(void)
1961 {
1962         o2quo_exit();
1963         kfree(o2net_hand);
1964         kfree(o2net_keep_req);
1965         kfree(o2net_keep_resp);
1966 }