err.no Git - linux-2.6/blob - 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 #ifdef CONFIG_DEBUG_FS
 146 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 147                            u32 msgkey, struct task_struct *task, u8 node)
 148 {
 149         INIT_LIST_HEAD(&nst->st_net_debug_item);
 150         nst->st_task = task;
 151         nst->st_msg_type = msgtype;
 152         nst->st_msg_key = msgkey;
 153         nst->st_node = node;
 154 }
 155
 156 static void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 157 {
 158         do_gettimeofday(&nst->st_sock_time);
 159 }
 160
 161 static void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 162 {
 163         do_gettimeofday(&nst->st_send_time);
 164 }
 165
 166 static void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 167 {
 168         do_gettimeofday(&nst->st_status_time);
 169 }
 170
 171 static void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 172                                          struct o2net_sock_container *sc)
 173 {
 174         nst->st_sc = sc;
 175 }
 176
 177 static void o2net_set_nst_msg_id(struct o2net_send_tracking *nst, u32 msg_id)
 178 {
 179         nst->st_id = msg_id;
 180 }
 181
 182 #else  /* CONFIG_DEBUG_FS */
 183
 184 static inline void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 185                                   u32 msgkey, struct task_struct *task, u8 node)
 186 {
 187 }
 188
 189 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 190 {
 191 }
 192
 193 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 194 {
 195 }
 196
 197 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 198 {
 199 }
 200
 201 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 202                                                 struct o2net_sock_container *sc)
 203 {
 204 }
 205
 206 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 207                                         u32 msg_id)
 208 {
 209 }
 210
 211 #endif /* CONFIG_DEBUG_FS */
 212
 213 static inline int o2net_reconnect_delay(void)
 214 {
 215         return o2nm_single_cluster->cl_reconnect_delay_ms;
 216 }
 217
 218 static inline int o2net_keepalive_delay(void)
 219 {
 220         return o2nm_single_cluster->cl_keepalive_delay_ms;
 221 }
 222
 223 static inline int o2net_idle_timeout(void)
 224 {
 225         return o2nm_single_cluster->cl_idle_timeout_ms;
 226 }
 227
 228 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 229 {
 230         int trans;
 231         BUG_ON(err >= O2NET_ERR_MAX);
 232         trans = o2net_sys_err_translations[err];
 233
 234         /* Just in case we mess up the translation table above */
 235         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 236         return trans;
 237 }
 238
 239 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 240 {
 241         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 242         return &o2net_nodes[node_num];
 243 }
 244
 245 static u8 o2net_num_from_nn(struct o2net_node *nn)
 246 {
 247         BUG_ON(nn == NULL);
 248         return nn - o2net_nodes;
 249 }
 250
 251 /* ------------------------------------------------------------ */
 252
 253 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 254 {
 255         int ret = 0;
 256
 257         do {
 258                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
 259                         ret = -EAGAIN;
 260                         break;
 261                 }
 262                 spin_lock(&nn->nn_lock);
 263                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
 264                 if (ret == 0)
 265                         list_add_tail(&nsw->ns_node_item,
 266                                       &nn->nn_status_list);
 267                 spin_unlock(&nn->nn_lock);
 268         } while (ret == -EAGAIN);
 269
 270         if (ret == 0)  {
 271                 init_waitqueue_head(&nsw->ns_wq);
 272                 nsw->ns_sys_status = O2NET_ERR_NONE;
 273                 nsw->ns_status = 0;
 274         }
 275
 276         return ret;
 277 }
 278
 279 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 280                                       struct o2net_status_wait *nsw,
 281                                       enum o2net_system_error sys_status,
 282                                       s32 status)
 283 {
 284         assert_spin_locked(&nn->nn_lock);
 285
 286         if (!list_empty(&nsw->ns_node_item)) {
 287                 list_del_init(&nsw->ns_node_item);
 288                 nsw->ns_sys_status = sys_status;
 289                 nsw->ns_status = status;
 290                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 291                 wake_up(&nsw->ns_wq);
 292         }
 293 }
 294
 295 static void o2net_complete_nsw(struct o2net_node *nn,
 296                                struct o2net_status_wait *nsw,
 297                                u64 id, enum o2net_system_error sys_status,
 298                                s32 status)
 299 {
 300         spin_lock(&nn->nn_lock);
 301         if (nsw == NULL) {
 302                 if (id > INT_MAX)
 303                         goto out;
 304
 305                 nsw = idr_find(&nn->nn_status_idr, id);
 306                 if (nsw == NULL)
 307                         goto out;
 308         }
 309
 310         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 311
 312 out:
 313         spin_unlock(&nn->nn_lock);
 314         return;
 315 }
 316
 317 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 318 {
 319         struct o2net_status_wait *nsw, *tmp;
 320         unsigned int num_kills = 0;
 321
 322         assert_spin_locked(&nn->nn_lock);
 323
 324         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 325                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 326                 num_kills++;
 327         }
 328
 329         mlog(0, "completed %d messages for node %u\n", num_kills,
 330              o2net_num_from_nn(nn));
 331 }
 332
 333 static int o2net_nsw_completed(struct o2net_node *nn,
 334                                struct o2net_status_wait *nsw)
 335 {
 336         int completed;
 337         spin_lock(&nn->nn_lock);
 338         completed = list_empty(&nsw->ns_node_item);
 339         spin_unlock(&nn->nn_lock);
 340         return completed;
 341 }
 342
 343 /* ------------------------------------------------------------ */
 344
 345 static void sc_kref_release(struct kref *kref)
 346 {
 347         struct o2net_sock_container *sc = container_of(kref,
 348                                         struct o2net_sock_container, sc_kref);
 349         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 350
 351         sclog(sc, "releasing\n");
 352
 353         if (sc->sc_sock) {
 354                 sock_release(sc->sc_sock);
 355                 sc->sc_sock = NULL;
 356         }
 357
 358         o2nm_node_put(sc->sc_node);
 359         sc->sc_node = NULL;
 360
 361         o2net_debug_del_sc(sc);
 362         kfree(sc);
 363 }
 364
 365 static void sc_put(struct o2net_sock_container *sc)
 366 {
 367         sclog(sc, "put\n");
 368         kref_put(&sc->sc_kref, sc_kref_release);
 369 }
 370 static void sc_get(struct o2net_sock_container *sc)
 371 {
 372         sclog(sc, "get\n");
 373         kref_get(&sc->sc_kref);
 374 }
 375 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 376 {
 377         struct o2net_sock_container *sc, *ret = NULL;
 378         struct page *page = NULL;
 379
 380         page = alloc_page(GFP_NOFS);
 381         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 382         if (sc == NULL || page == NULL)
 383                 goto out;
 384
 385         kref_init(&sc->sc_kref);
 386         o2nm_node_get(node);
 387         sc->sc_node = node;
 388
 389         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 390         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 391         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 392         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 393
 394         init_timer(&sc->sc_idle_timeout);
 395         sc->sc_idle_timeout.function = o2net_idle_timer;
 396         sc->sc_idle_timeout.data = (unsigned long)sc;
 397
 398         sclog(sc, "alloced\n");
 399
 400         ret = sc;
 401         sc->sc_page = page;
 402         o2net_debug_add_sc(sc);
 403         sc = NULL;
 404         page = NULL;
 405
 406 out:
 407         if (page)
 408                 __free_page(page);
 409         kfree(sc);
 410
 411         return ret;
 412 }
 413
 414 /* ------------------------------------------------------------ */
 415
 416 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 417                                 struct work_struct *work)
 418 {
 419         sc_get(sc);
 420         if (!queue_work(o2net_wq, work))
 421                 sc_put(sc);
 422 }
 423 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 424                                         struct delayed_work *work,
 425                                         int delay)
 426 {
 427         sc_get(sc);
 428         if (!queue_delayed_work(o2net_wq, work, delay))
 429                 sc_put(sc);
 430 }
 431 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 432                                          struct delayed_work *work)
 433 {
 434         if (cancel_delayed_work(work))
 435                 sc_put(sc);
 436 }
 437
 438 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 439
 440 int o2net_num_connected_peers(void)
 441 {
 442         return atomic_read(&o2net_connected_peers);
 443 }
 444
 445 static void o2net_set_nn_state(struct o2net_node *nn,
 446                                struct o2net_sock_container *sc,
 447                                unsigned valid, int err)
 448 {
 449         int was_valid = nn->nn_sc_valid;
 450         int was_err = nn->nn_persistent_error;
 451         struct o2net_sock_container *old_sc = nn->nn_sc;
 452
 453         assert_spin_locked(&nn->nn_lock);
 454
 455         if (old_sc && !sc)
 456                 atomic_dec(&o2net_connected_peers);
 457         else if (!old_sc && sc)
 458                 atomic_inc(&o2net_connected_peers);
 459
 460         /* the node num comparison and single connect/accept path should stop
 461          * an non-null sc from being overwritten with another */
 462         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 463         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 464         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 465
 466         if (was_valid && !valid && err == 0)
 467                 err = -ENOTCONN;
 468
 469         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 470              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 471              nn->nn_persistent_error, err);
 472
 473         nn->nn_sc = sc;
 474         nn->nn_sc_valid = valid ? 1 : 0;
 475         nn->nn_persistent_error = err;
 476
 477         /* mirrors o2net_tx_can_proceed() */
 478         if (nn->nn_persistent_error || nn->nn_sc_valid)
 479                 wake_up(&nn->nn_sc_wq);
 480
 481         if (!was_err && nn->nn_persistent_error) {
 482                 o2quo_conn_err(o2net_num_from_nn(nn));
 483                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 484                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 485         }
 486
 487         if (was_valid && !valid) {
 488                 printk(KERN_INFO "o2net: no longer connected to "
 489                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 490                 o2net_complete_nodes_nsw(nn);
 491         }
 492
 493         if (!was_valid && valid) {
 494                 o2quo_conn_up(o2net_num_from_nn(nn));
 495                 cancel_delayed_work(&nn->nn_connect_expired);
 496                 printk(KERN_INFO "o2net: %s " SC_NODEF_FMT "\n",
 497                        o2nm_this_node() > sc->sc_node->nd_num ?
 498                                 "connected to" : "accepted connection from",
 499                        SC_NODEF_ARGS(sc));
 500         }
 501
 502         /* trigger the connecting worker func as long as we're not valid,
 503          * it will back off if it shouldn't connect.  This can be called
 504          * from node config teardown and so needs to be careful about
 505          * the work queue actually being up. */
 506         if (!valid && o2net_wq) {
 507                 unsigned long delay;
 508                 /* delay if we're withing a RECONNECT_DELAY of the
 509                  * last attempt */
 510                 delay = (nn->nn_last_connect_attempt +
 511                          msecs_to_jiffies(o2net_reconnect_delay()))
 512                         - jiffies;
 513                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 514                         delay = 0;
 515                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 516                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 517
 518                 /*
 519                  * Delay the expired work after idle timeout.
 520                  *
 521                  * We might have lots of failed connection attempts that run
 522                  * through here but we only cancel the connect_expired work when
 523                  * a connection attempt succeeds.  So only the first enqueue of
 524                  * the connect_expired work will do anything.  The rest will see
 525                  * that it's already queued and do nothing.
 526                  */
 527                 delay += msecs_to_jiffies(o2net_idle_timeout());
 528                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 529         }
 530
 531         /* keep track of the nn's sc ref for the caller */
 532         if ((old_sc == NULL) && sc)
 533                 sc_get(sc);
 534         if (old_sc && (old_sc != sc)) {
 535                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 536                 sc_put(old_sc);
 537         }
 538 }
 539
 540 /* see o2net_register_callbacks() */
 541 static void o2net_data_ready(struct sock *sk, int bytes)
 542 {
 543         void (*ready)(struct sock *sk, int bytes);
 544
 545         read_lock(&sk->sk_callback_lock);
 546         if (sk->sk_user_data) {
 547                 struct o2net_sock_container *sc = sk->sk_user_data;
 548                 sclog(sc, "data_ready hit\n");
 549                 do_gettimeofday(&sc->sc_tv_data_ready);
 550                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 551                 ready = sc->sc_data_ready;
 552         } else {
 553                 ready = sk->sk_data_ready;
 554         }
 555         read_unlock(&sk->sk_callback_lock);
 556
 557         ready(sk, bytes);
 558 }
 559
 560 /* see o2net_register_callbacks() */
 561 static void o2net_state_change(struct sock *sk)
 562 {
 563         void (*state_change)(struct sock *sk);
 564         struct o2net_sock_container *sc;
 565
 566         read_lock(&sk->sk_callback_lock);
 567         sc = sk->sk_user_data;
 568         if (sc == NULL) {
 569                 state_change = sk->sk_state_change;
 570                 goto out;
 571         }
 572
 573         sclog(sc, "state_change to %d\n", sk->sk_state);
 574
 575         state_change = sc->sc_state_change;
 576
 577         switch(sk->sk_state) {
 578                 /* ignore connecting sockets as they make progress */
 579                 case TCP_SYN_SENT:
 580                 case TCP_SYN_RECV:
 581                         break;
 582                 case TCP_ESTABLISHED:
 583                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
 584                         break;
 585                 default:
 586                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 587                         break;
 588         }
 589 out:
 590         read_unlock(&sk->sk_callback_lock);
 591         state_change(sk);
 592 }
 593
 594 /*
 595  * we register callbacks so we can queue work on events before calling
 596  * the original callbacks.  our callbacks our careful to test user_data
 597  * to discover when they've reaced with o2net_unregister_callbacks().
 598  */
 599 static void o2net_register_callbacks(struct sock *sk,
 600                                      struct o2net_sock_container *sc)
 601 {
 602         write_lock_bh(&sk->sk_callback_lock);
 603
 604         /* accepted sockets inherit the old listen socket data ready */
 605         if (sk->sk_data_ready == o2net_listen_data_ready) {
 606                 sk->sk_data_ready = sk->sk_user_data;
 607                 sk->sk_user_data = NULL;
 608         }
 609
 610         BUG_ON(sk->sk_user_data != NULL);
 611         sk->sk_user_data = sc;
 612         sc_get(sc);
 613
 614         sc->sc_data_ready = sk->sk_data_ready;
 615         sc->sc_state_change = sk->sk_state_change;
 616         sk->sk_data_ready = o2net_data_ready;
 617         sk->sk_state_change = o2net_state_change;
 618
 619         mutex_init(&sc->sc_send_lock);
 620
 621         write_unlock_bh(&sk->sk_callback_lock);
 622 }
 623
 624 static int o2net_unregister_callbacks(struct sock *sk,
 625                                    struct o2net_sock_container *sc)
 626 {
 627         int ret = 0;
 628
 629         write_lock_bh(&sk->sk_callback_lock);
 630         if (sk->sk_user_data == sc) {
 631                 ret = 1;
 632                 sk->sk_user_data = NULL;
 633                 sk->sk_data_ready = sc->sc_data_ready;
 634                 sk->sk_state_change = sc->sc_state_change;
 635         }
 636         write_unlock_bh(&sk->sk_callback_lock);
 637
 638         return ret;
 639 }
 640
 641 /*
 642  * this is a little helper that is called by callers who have seen a problem
 643  * with an sc and want to detach it from the nn if someone already hasn't beat
 644  * them to it.  if an error is given then the shutdown will be persistent
 645  * and pending transmits will be canceled.
 646  */
 647 static void o2net_ensure_shutdown(struct o2net_node *nn,
 648                                    struct o2net_sock_container *sc,
 649                                    int err)
 650 {
 651         spin_lock(&nn->nn_lock);
 652         if (nn->nn_sc == sc)
 653                 o2net_set_nn_state(nn, NULL, 0, err);
 654         spin_unlock(&nn->nn_lock);
 655 }
 656
 657 /*
 658  * This work queue function performs the blocking parts of socket shutdown.  A
 659  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 660  * sc detached from the nn.  state_change will also trigger this callback
 661  * directly when it sees errors.  In that case we need to call set_nn_state
 662  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 663  * itself.
 664  */
 665 static void o2net_shutdown_sc(struct work_struct *work)
 666 {
 667         struct o2net_sock_container *sc =
 668                 container_of(work, struct o2net_sock_container,
 669                              sc_shutdown_work);
 670         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 671
 672         sclog(sc, "shutting down\n");
 673
 674         /* drop the callbacks ref and call shutdown only once */
 675         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 676                 /* we shouldn't flush as we're in the thread, the
 677                  * races with pending sc work structs are harmless */
 678                 del_timer_sync(&sc->sc_idle_timeout);
 679                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 680                 sc_put(sc);
 681                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 682         }
 683
 684         /* not fatal so failed connects before the other guy has our
 685          * heartbeat can be retried */
 686         o2net_ensure_shutdown(nn, sc, 0);
 687         sc_put(sc);
 688 }
 689
 690 /* ------------------------------------------------------------ */
 691
 692 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 693                              u32 key)
 694 {
 695         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 696
 697         if (ret == 0)
 698                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 699
 700         return ret;
 701 }
 702
 703 static struct o2net_msg_handler *
 704 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 705                           struct rb_node **ret_parent)
 706 {
 707         struct rb_node **p = &o2net_handler_tree.rb_node;
 708         struct rb_node *parent = NULL;
 709         struct o2net_msg_handler *nmh, *ret = NULL;
 710         int cmp;
 711
 712         while (*p) {
 713                 parent = *p;
 714                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 715                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 716
 717                 if (cmp < 0)
 718                         p = &(*p)->rb_left;
 719                 else if (cmp > 0)
 720                         p = &(*p)->rb_right;
 721                 else {
 722                         ret = nmh;
 723                         break;
 724                 }
 725         }
 726
 727         if (ret_p != NULL)
 728                 *ret_p = p;
 729         if (ret_parent != NULL)
 730                 *ret_parent = parent;
 731
 732         return ret;
 733 }
 734
 735 static void o2net_handler_kref_release(struct kref *kref)
 736 {
 737         struct o2net_msg_handler *nmh;
 738         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 739
 740         kfree(nmh);
 741 }
 742
 743 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 744 {
 745         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 746 }
 747
 748 /* max_len is protection for the handler func.  incoming messages won't
 749  * be given to the handler if their payload is longer than the max. */
 750 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 751                            o2net_msg_handler_func *func, void *data,
 752                            o2net_post_msg_handler_func *post_func,
 753                            struct list_head *unreg_list)
 754 {
 755         struct o2net_msg_handler *nmh = NULL;
 756         struct rb_node **p, *parent;
 757         int ret = 0;
 758
 759         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 760                 mlog(0, "max_len for message handler out of range: %u\n",
 761                         max_len);
 762                 ret = -EINVAL;
 763                 goto out;
 764         }
 765
 766         if (!msg_type) {
 767                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 768                 ret = -EINVAL;
 769                 goto out;
 770
 771         }
 772         if (!func) {
 773                 mlog(0, "no message handler provided: %u, %p\n",
 774                        msg_type, func);
 775                 ret = -EINVAL;
 776                 goto out;
 777         }
 778
 779         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 780         if (nmh == NULL) {
 781                 ret = -ENOMEM;
 782                 goto out;
 783         }
 784
 785         nmh->nh_func = func;
 786         nmh->nh_func_data = data;
 787         nmh->nh_post_func = post_func;
 788         nmh->nh_msg_type = msg_type;
 789         nmh->nh_max_len = max_len;
 790         nmh->nh_key = key;
 791         /* the tree and list get this ref.. they're both removed in
 792          * unregister when this ref is dropped */
 793         kref_init(&nmh->nh_kref);
 794         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 795
 796         write_lock(&o2net_handler_lock);
 797         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 798                 ret = -EEXIST;
 799         else {
 800                 rb_link_node(&nmh->nh_node, parent, p);
 801                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 802                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 803
 804                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 805                      func, msg_type, key);
 806                 /* we've had some trouble with handlers seemingly vanishing. */
 807                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 808                                                           &parent) == NULL,
 809                                 "couldn't find handler we *just* registerd "
 810                                 "for type %u key %08x\n", msg_type, key);
 811         }
 812         write_unlock(&o2net_handler_lock);
 813         if (ret)
 814                 goto out;
 815
 816 out:
 817         if (ret)
 818                 kfree(nmh);
 819
 820         return ret;
 821 }
 822 EXPORT_SYMBOL_GPL(o2net_register_handler);
 823
 824 void o2net_unregister_handler_list(struct list_head *list)
 825 {
 826         struct o2net_msg_handler *nmh, *n;
 827
 828         write_lock(&o2net_handler_lock);
 829         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 830                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 831                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 832                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 833                 list_del_init(&nmh->nh_unregister_item);
 834                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 835         }
 836         write_unlock(&o2net_handler_lock);
 837 }
 838 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 839
 840 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 841 {
 842         struct o2net_msg_handler *nmh;
 843
 844         read_lock(&o2net_handler_lock);
 845         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 846         if (nmh)
 847                 kref_get(&nmh->nh_kref);
 848         read_unlock(&o2net_handler_lock);
 849
 850         return nmh;
 851 }
 852
 853 /* ------------------------------------------------------------ */
 854
 855 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 856 {
 857         int ret;
 858         mm_segment_t oldfs;
 859         struct kvec vec = {
 860                 .iov_len = len,
 861                 .iov_base = data,
 862         };
 863         struct msghdr msg = {
 864                 .msg_iovlen = 1,
 865                 .msg_iov = (struct iovec *)&vec,
 866                 .msg_flags = MSG_DONTWAIT,
 867         };
 868
 869         oldfs = get_fs();
 870         set_fs(get_ds());
 871         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
 872         set_fs(oldfs);
 873
 874         return ret;
 875 }
 876
 877 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 878                               size_t veclen, size_t total)
 879 {
 880         int ret;
 881         mm_segment_t oldfs;
 882         struct msghdr msg = {
 883                 .msg_iov = (struct iovec *)vec,
 884                 .msg_iovlen = veclen,
 885         };
 886
 887         if (sock == NULL) {
 888                 ret = -EINVAL;
 889                 goto out;
 890         }
 891
 892         oldfs = get_fs();
 893         set_fs(get_ds());
 894         ret = sock_sendmsg(sock, &msg, total);
 895         set_fs(oldfs);
 896         if (ret != total) {
 897                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
 898                      total);
 899                 if (ret >= 0)
 900                         ret = -EPIPE; /* should be smarter, I bet */
 901                 goto out;
 902         }
 903
 904         ret = 0;
 905 out:
 906         if (ret < 0)
 907                 mlog(0, "returning error: %d\n", ret);
 908         return ret;
 909 }
 910
 911 static void o2net_sendpage(struct o2net_sock_container *sc,
 912                            void *kmalloced_virt,
 913                            size_t size)
 914 {
 915         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 916         ssize_t ret;
 917
 918         while (1) {
 919                 mutex_lock(&sc->sc_send_lock);
 920                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 921                                                  virt_to_page(kmalloced_virt),
 922                                                  (long)kmalloced_virt & ~PAGE_MASK,
 923                                                  size, MSG_DONTWAIT);
 924                 mutex_unlock(&sc->sc_send_lock);
 925                 if (ret == size)
 926                         break;
 927                 if (ret == (ssize_t)-EAGAIN) {
 928                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 929                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 930                         cond_resched();
 931                         continue;
 932                 }
 933                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 934                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 935                 o2net_ensure_shutdown(nn, sc, 0);
 936                 break;
 937         }
 938 }
 939
 940 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 941 {
 942         memset(msg, 0, sizeof(struct o2net_msg));
 943         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 944         msg->data_len = cpu_to_be16(data_len);
 945         msg->msg_type = cpu_to_be16(msg_type);
 946         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 947         msg->status = 0;
 948         msg->key = cpu_to_be32(key);
 949 }
 950
 951 static int o2net_tx_can_proceed(struct o2net_node *nn,
 952                                 struct o2net_sock_container **sc_ret,
 953                                 int *error)
 954 {
 955         int ret = 0;
 956
 957         spin_lock(&nn->nn_lock);
 958         if (nn->nn_persistent_error) {
 959                 ret = 1;
 960                 *sc_ret = NULL;
 961                 *error = nn->nn_persistent_error;
 962         } else if (nn->nn_sc_valid) {
 963                 kref_get(&nn->nn_sc->sc_kref);
 964
 965                 ret = 1;
 966                 *sc_ret = nn->nn_sc;
 967                 *error = 0;
 968         }
 969         spin_unlock(&nn->nn_lock);
 970
 971         return ret;
 972 }
 973
 974 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
 975                            size_t caller_veclen, u8 target_node, int *status)
 976 {
 977         int ret, error = 0;
 978         struct o2net_msg *msg = NULL;
 979         size_t veclen, caller_bytes = 0;
 980         struct kvec *vec = NULL;
 981         struct o2net_sock_container *sc = NULL;
 982         struct o2net_node *nn = o2net_nn_from_num(target_node);
 983         struct o2net_status_wait nsw = {
 984                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
 985         };
 986         struct o2net_send_tracking nst;
 987
 988         o2net_init_nst(&nst, msg_type, key, current, target_node);
 989
 990         if (o2net_wq == NULL) {
 991                 mlog(0, "attempt to tx without o2netd running\n");
 992                 ret = -ESRCH;
 993                 goto out;
 994         }
 995
 996         if (caller_veclen == 0) {
 997                 mlog(0, "bad kvec array length\n");
 998                 ret = -EINVAL;
 999                 goto out;
1000         }
1001
1002         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1003         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1004                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1005                 ret = -EINVAL;
1006                 goto out;
1007         }
1008
1009         if (target_node == o2nm_this_node()) {
1010                 ret = -ELOOP;
1011                 goto out;
1012         }
1013
1014         o2net_debug_add_nst(&nst);
1015
1016         o2net_set_nst_sock_time(&nst);
1017
1018         ret = wait_event_interruptible(nn->nn_sc_wq,
1019                                        o2net_tx_can_proceed(nn, &sc, &error));
1020         if (!ret && error)
1021                 ret = error;
1022         if (ret)
1023                 goto out;
1024
1025         o2net_set_nst_sock_container(&nst, sc);
1026
1027         veclen = caller_veclen + 1;
1028         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1029         if (vec == NULL) {
1030                 mlog(0, "failed to %zu element kvec!\n", veclen);
1031                 ret = -ENOMEM;
1032                 goto out;
1033         }
1034
1035         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1036         if (!msg) {
1037                 mlog(0, "failed to allocate a o2net_msg!\n");
1038                 ret = -ENOMEM;
1039                 goto out;
1040         }
1041
1042         o2net_init_msg(msg, caller_bytes, msg_type, key);
1043
1044         vec[0].iov_len = sizeof(struct o2net_msg);
1045         vec[0].iov_base = msg;
1046         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1047
1048         ret = o2net_prep_nsw(nn, &nsw);
1049         if (ret)
1050                 goto out;
1051
1052         msg->msg_num = cpu_to_be32(nsw.ns_id);
1053         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1054
1055         o2net_set_nst_send_time(&nst);
1056
1057         /* finally, convert the message header to network byte-order
1058          * and send */
1059         mutex_lock(&sc->sc_send_lock);
1060         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1061                                  sizeof(struct o2net_msg) + caller_bytes);
1062         mutex_unlock(&sc->sc_send_lock);
1063         msglog(msg, "sending returned %d\n", ret);
1064         if (ret < 0) {
1065                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1066                 goto out;
1067         }
1068
1069         /* wait on other node's handler */
1070         o2net_set_nst_status_time(&nst);
1071         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1072
1073         /* Note that we avoid overwriting the callers status return
1074          * variable if a system error was reported on the other
1075          * side. Callers beware. */
1076         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1077         if (status && !ret)
1078                 *status = nsw.ns_status;
1079
1080         mlog(0, "woken, returning system status %d, user status %d\n",
1081              ret, nsw.ns_status);
1082 out:
1083         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1084         if (sc)
1085                 sc_put(sc);
1086         if (vec)
1087                 kfree(vec);
1088         if (msg)
1089                 kfree(msg);
1090         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1091         return ret;
1092 }
1093 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1094
1095 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1096                        u8 target_node, int *status)
1097 {
1098         struct kvec vec = {
1099                 .iov_base = data,
1100                 .iov_len = len,
1101         };
1102         return o2net_send_message_vec(msg_type, key, &vec, 1,
1103                                       target_node, status);
1104 }
1105 EXPORT_SYMBOL_GPL(o2net_send_message);
1106
1107 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1108                                    enum o2net_system_error syserr, int err)
1109 {
1110         struct kvec vec = {
1111                 .iov_base = hdr,
1112                 .iov_len = sizeof(struct o2net_msg),
1113         };
1114
1115         BUG_ON(syserr >= O2NET_ERR_MAX);
1116
1117         /* leave other fields intact from the incoming message, msg_num
1118          * in particular */
1119         hdr->sys_status = cpu_to_be32(syserr);
1120         hdr->status = cpu_to_be32(err);
1121         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1122         hdr->data_len = 0;
1123
1124         msglog(hdr, "about to send status magic %d\n", err);
1125         /* hdr has been in host byteorder this whole time */
1126         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1127 }
1128
1129 /* this returns -errno if the header was unknown or too large, etc.
1130  * after this is called the buffer us reused for the next message */
1131 static int o2net_process_message(struct o2net_sock_container *sc,
1132                                  struct o2net_msg *hdr)
1133 {
1134         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1135         int ret = 0, handler_status;
1136         enum  o2net_system_error syserr;
1137         struct o2net_msg_handler *nmh = NULL;
1138         void *ret_data = NULL;
1139
1140         msglog(hdr, "processing message\n");
1141
1142         o2net_sc_postpone_idle(sc);
1143
1144         switch(be16_to_cpu(hdr->magic)) {
1145                 case O2NET_MSG_STATUS_MAGIC:
1146                         /* special type for returning message status */
1147                         o2net_complete_nsw(nn, NULL,
1148                                            be32_to_cpu(hdr->msg_num),
1149                                            be32_to_cpu(hdr->sys_status),
1150                                            be32_to_cpu(hdr->status));
1151                         goto out;
1152                 case O2NET_MSG_KEEP_REQ_MAGIC:
1153                         o2net_sendpage(sc, o2net_keep_resp,
1154                                        sizeof(*o2net_keep_resp));
1155                         goto out;
1156                 case O2NET_MSG_KEEP_RESP_MAGIC:
1157                         goto out;
1158                 case O2NET_MSG_MAGIC:
1159                         break;
1160                 default:
1161                         msglog(hdr, "bad magic\n");
1162                         ret = -EINVAL;
1163                         goto out;
1164                         break;
1165         }
1166
1167         /* find a handler for it */
1168         handler_status = 0;
1169         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1170                                 be32_to_cpu(hdr->key));
1171         if (!nmh) {
1172                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1173                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1174                 syserr = O2NET_ERR_NO_HNDLR;
1175                 goto out_respond;
1176         }
1177
1178         syserr = O2NET_ERR_NONE;
1179
1180         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1181                 syserr = O2NET_ERR_OVERFLOW;
1182
1183         if (syserr != O2NET_ERR_NONE)
1184                 goto out_respond;
1185
1186         do_gettimeofday(&sc->sc_tv_func_start);
1187         sc->sc_msg_key = be32_to_cpu(hdr->key);
1188         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1189         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1190                                              be16_to_cpu(hdr->data_len),
1191                                         nmh->nh_func_data, &ret_data);
1192         do_gettimeofday(&sc->sc_tv_func_stop);
1193
1194 out_respond:
1195         /* this destroys the hdr, so don't use it after this */
1196         mutex_lock(&sc->sc_send_lock);
1197         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1198                                       handler_status);
1199         mutex_unlock(&sc->sc_send_lock);
1200         hdr = NULL;
1201         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1202              handler_status, syserr, ret);
1203
1204         if (nmh) {
1205                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1206                 if (nmh->nh_post_func)
1207                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1208                                             ret_data);
1209         }
1210
1211 out:
1212         if (nmh)
1213                 o2net_handler_put(nmh);
1214         return ret;
1215 }
1216
1217 static int o2net_check_handshake(struct o2net_sock_container *sc)
1218 {
1219         struct o2net_handshake *hand = page_address(sc->sc_page);
1220         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1221
1222         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1223                 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1224                      "version %llu but %llu is required, disconnecting\n",
1225                      SC_NODEF_ARGS(sc),
1226                      (unsigned long long)be64_to_cpu(hand->protocol_version),
1227                      O2NET_PROTOCOL_VERSION);
1228
1229                 /* don't bother reconnecting if its the wrong version. */
1230                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1231                 return -1;
1232         }
1233
1234         /*
1235          * Ensure timeouts are consistent with other nodes, otherwise
1236          * we can end up with one node thinking that the other must be down,
1237          * but isn't. This can ultimately cause corruption.
1238          */
1239         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1240                                 o2net_idle_timeout()) {
1241                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1242                      "%u ms, but we use %u ms locally.  disconnecting\n",
1243                      SC_NODEF_ARGS(sc),
1244                      be32_to_cpu(hand->o2net_idle_timeout_ms),
1245                      o2net_idle_timeout());
1246                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1247                 return -1;
1248         }
1249
1250         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1251                         o2net_keepalive_delay()) {
1252                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1253                      "%u ms, but we use %u ms locally.  disconnecting\n",
1254                      SC_NODEF_ARGS(sc),
1255                      be32_to_cpu(hand->o2net_keepalive_delay_ms),
1256                      o2net_keepalive_delay());
1257                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1258                 return -1;
1259         }
1260
1261         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1262                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1263                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1264                      "%u ms, but we use %u ms locally.  disconnecting\n",
1265                      SC_NODEF_ARGS(sc),
1266                      be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1267                      O2HB_MAX_WRITE_TIMEOUT_MS);
1268                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1269                 return -1;
1270         }
1271
1272         sc->sc_handshake_ok = 1;
1273
1274         spin_lock(&nn->nn_lock);
1275         /* set valid and queue the idle timers only if it hasn't been
1276          * shut down already */
1277         if (nn->nn_sc == sc) {
1278                 o2net_sc_reset_idle_timer(sc);
1279                 atomic_set(&nn->nn_timeout, 0);
1280                 o2net_set_nn_state(nn, sc, 1, 0);
1281         }
1282         spin_unlock(&nn->nn_lock);
1283
1284         /* shift everything up as though it wasn't there */
1285         sc->sc_page_off -= sizeof(struct o2net_handshake);
1286         if (sc->sc_page_off)
1287                 memmove(hand, hand + 1, sc->sc_page_off);
1288
1289         return 0;
1290 }
1291
1292 /* this demuxes the queued rx bytes into header or payload bits and calls
1293  * handlers as each full message is read off the socket.  it returns -error,
1294  * == 0 eof, or > 0 for progress made.*/
1295 static int o2net_advance_rx(struct o2net_sock_container *sc)
1296 {
1297         struct o2net_msg *hdr;
1298         int ret = 0;
1299         void *data;
1300         size_t datalen;
1301
1302         sclog(sc, "receiving\n");
1303         do_gettimeofday(&sc->sc_tv_advance_start);
1304
1305         if (unlikely(sc->sc_handshake_ok == 0)) {
1306                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1307                         data = page_address(sc->sc_page) + sc->sc_page_off;
1308                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1309                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1310                         if (ret > 0)
1311                                 sc->sc_page_off += ret;
1312                 }
1313
1314                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1315                         o2net_check_handshake(sc);
1316                         if (unlikely(sc->sc_handshake_ok == 0))
1317                                 ret = -EPROTO;
1318                 }
1319                 goto out;
1320         }
1321
1322         /* do we need more header? */
1323         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1324                 data = page_address(sc->sc_page) + sc->sc_page_off;
1325                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1326                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1327                 if (ret > 0) {
1328                         sc->sc_page_off += ret;
1329                         /* only swab incoming here.. we can
1330                          * only get here once as we cross from
1331                          * being under to over */
1332                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1333                                 hdr = page_address(sc->sc_page);
1334                                 if (be16_to_cpu(hdr->data_len) >
1335                                     O2NET_MAX_PAYLOAD_BYTES)
1336                                         ret = -EOVERFLOW;
1337                         }
1338                 }
1339                 if (ret <= 0)
1340                         goto out;
1341         }
1342
1343         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1344                 /* oof, still don't have a header */
1345                 goto out;
1346         }
1347
1348         /* this was swabbed above when we first read it */
1349         hdr = page_address(sc->sc_page);
1350
1351         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1352
1353         /* do we need more payload? */
1354         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1355                 /* need more payload */
1356                 data = page_address(sc->sc_page) + sc->sc_page_off;
1357                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1358                           sc->sc_page_off;
1359                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1360                 if (ret > 0)
1361                         sc->sc_page_off += ret;
1362                 if (ret <= 0)
1363                         goto out;
1364         }
1365
1366         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1367                 /* we can only get here once, the first time we read
1368                  * the payload.. so set ret to progress if the handler
1369                  * works out. after calling this the message is toast */
1370                 ret = o2net_process_message(sc, hdr);
1371                 if (ret == 0)
1372                         ret = 1;
1373                 sc->sc_page_off = 0;
1374         }
1375
1376 out:
1377         sclog(sc, "ret = %d\n", ret);
1378         do_gettimeofday(&sc->sc_tv_advance_stop);
1379         return ret;
1380 }
1381
1382 /* this work func is triggerd by data ready.  it reads until it can read no
1383  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1384  * our work the work struct will be marked and we'll be called again. */
1385 static void o2net_rx_until_empty(struct work_struct *work)
1386 {
1387         struct o2net_sock_container *sc =
1388                 container_of(work, struct o2net_sock_container, sc_rx_work);
1389         int ret;
1390
1391         do {
1392                 ret = o2net_advance_rx(sc);
1393         } while (ret > 0);
1394
1395         if (ret <= 0 && ret != -EAGAIN) {
1396                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1397                 sclog(sc, "saw error %d, closing\n", ret);
1398                 /* not permanent so read failed handshake can retry */
1399                 o2net_ensure_shutdown(nn, sc, 0);
1400         }
1401
1402         sc_put(sc);
1403 }
1404
1405 static int o2net_set_nodelay(struct socket *sock)
1406 {
1407         int ret, val = 1;
1408         mm_segment_t oldfs;
1409
1410         oldfs = get_fs();
1411         set_fs(KERNEL_DS);
1412
1413         /*
1414          * Dear unsuspecting programmer,
1415          *
1416          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1417          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1418          * silently turn into SO_DEBUG.
1419          *
1420          * Yours,
1421          * Keeper of hilariously fragile interfaces.
1422          */
1423         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1424                                     (char __user *)&val, sizeof(val));
1425
1426         set_fs(oldfs);
1427         return ret;
1428 }
1429
1430 static void o2net_initialize_handshake(void)
1431 {
1432         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1433                 O2HB_MAX_WRITE_TIMEOUT_MS);
1434         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1435         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1436                 o2net_keepalive_delay());
1437         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1438                 o2net_reconnect_delay());
1439 }
1440
1441 /* ------------------------------------------------------------ */
1442
1443 /* called when a connect completes and after a sock is accepted.  the
1444  * rx path will see the response and mark the sc valid */
1445 static void o2net_sc_connect_completed(struct work_struct *work)
1446 {
1447         struct o2net_sock_container *sc =
1448                 container_of(work, struct o2net_sock_container,
1449                              sc_connect_work);
1450
1451         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1452               (unsigned long long)O2NET_PROTOCOL_VERSION,
1453               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1454
1455         o2net_initialize_handshake();
1456         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1457         sc_put(sc);
1458 }
1459
1460 /* this is called as a work_struct func. */
1461 static void o2net_sc_send_keep_req(struct work_struct *work)
1462 {
1463         struct o2net_sock_container *sc =
1464                 container_of(work, struct o2net_sock_container,
1465                              sc_keepalive_work.work);
1466
1467         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1468         sc_put(sc);
1469 }
1470
1471 /* socket shutdown does a del_timer_sync against this as it tears down.
1472  * we can't start this timer until we've got to the point in sc buildup
1473  * where shutdown is going to be involved */
1474 static void o2net_idle_timer(unsigned long data)
1475 {
1476         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1477         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1478         struct timeval now;
1479
1480         do_gettimeofday(&now);
1481
1482         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1483              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1484                      o2net_idle_timeout() / 1000,
1485                      o2net_idle_timeout() % 1000);
1486         mlog(ML_NOTICE, "here are some times that might help debug the "
1487              "situation: (tmr %ld.%ld now %ld.%ld dr %ld.%ld adv "
1488              "%ld.%ld:%ld.%ld func (%08x:%u) %ld.%ld:%ld.%ld)\n",
1489              sc->sc_tv_timer.tv_sec, (long) sc->sc_tv_timer.tv_usec,
1490              now.tv_sec, (long) now.tv_usec,
1491              sc->sc_tv_data_ready.tv_sec, (long) sc->sc_tv_data_ready.tv_usec,
1492              sc->sc_tv_advance_start.tv_sec,
1493              (long) sc->sc_tv_advance_start.tv_usec,
1494              sc->sc_tv_advance_stop.tv_sec,
1495              (long) sc->sc_tv_advance_stop.tv_usec,
1496              sc->sc_msg_key, sc->sc_msg_type,
1497              sc->sc_tv_func_start.tv_sec, (long) sc->sc_tv_func_start.tv_usec,
1498              sc->sc_tv_func_stop.tv_sec, (long) sc->sc_tv_func_stop.tv_usec);
1499
1500         /*
1501          * Initialize the nn_timeout so that the next connection attempt
1502          * will continue in o2net_start_connect.
1503          */
1504         atomic_set(&nn->nn_timeout, 1);
1505
1506         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1507 }
1508
1509 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1510 {
1511         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1512         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1513                       msecs_to_jiffies(o2net_keepalive_delay()));
1514         do_gettimeofday(&sc->sc_tv_timer);
1515         mod_timer(&sc->sc_idle_timeout,
1516                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1517 }
1518
1519 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1520 {
1521         /* Only push out an existing timer */
1522         if (timer_pending(&sc->sc_idle_timeout))
1523                 o2net_sc_reset_idle_timer(sc);
1524 }
1525
1526 /* this work func is kicked whenever a path sets the nn state which doesn't
1527  * have valid set.  This includes seeing hb come up, losing a connection,
1528  * having a connect attempt fail, etc. This centralizes the logic which decides
1529  * if a connect attempt should be made or if we should give up and all future
1530  * transmit attempts should fail */
1531 static void o2net_start_connect(struct work_struct *work)
1532 {
1533         struct o2net_node *nn =
1534                 container_of(work, struct o2net_node, nn_connect_work.work);
1535         struct o2net_sock_container *sc = NULL;
1536         struct o2nm_node *node = NULL, *mynode = NULL;
1537         struct socket *sock = NULL;
1538         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1539         int ret = 0, stop;
1540         unsigned int timeout;
1541
1542         /* if we're greater we initiate tx, otherwise we accept */
1543         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1544                 goto out;
1545
1546         /* watch for racing with tearing a node down */
1547         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1548         if (node == NULL) {
1549                 ret = 0;
1550                 goto out;
1551         }
1552
1553         mynode = o2nm_get_node_by_num(o2nm_this_node());
1554         if (mynode == NULL) {
1555                 ret = 0;
1556                 goto out;
1557         }
1558
1559         spin_lock(&nn->nn_lock);
1560         /*
1561          * see if we already have one pending or have given up.
1562          * For nn_timeout, it is set when we close the connection
1563          * because of the idle time out. So it means that we have
1564          * at least connected to that node successfully once,
1565          * now try to connect to it again.
1566          */
1567         timeout = atomic_read(&nn->nn_timeout);
1568         stop = (nn->nn_sc ||
1569                 (nn->nn_persistent_error &&
1570                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1571         spin_unlock(&nn->nn_lock);
1572         if (stop)
1573                 goto out;
1574
1575         nn->nn_last_connect_attempt = jiffies;
1576
1577         sc = sc_alloc(node);
1578         if (sc == NULL) {
1579                 mlog(0, "couldn't allocate sc\n");
1580                 ret = -ENOMEM;
1581                 goto out;
1582         }
1583
1584         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1585         if (ret < 0) {
1586                 mlog(0, "can't create socket: %d\n", ret);
1587                 goto out;
1588         }
1589         sc->sc_sock = sock; /* freed by sc_kref_release */
1590
1591         sock->sk->sk_allocation = GFP_ATOMIC;
1592
1593         myaddr.sin_family = AF_INET;
1594         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1595         myaddr.sin_port = htons(0); /* any port */
1596
1597         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1598                               sizeof(myaddr));
1599         if (ret) {
1600                 mlog(ML_ERROR, "bind failed with %d at address %u.%u.%u.%u\n",
1601                      ret, NIPQUAD(mynode->nd_ipv4_address));
1602                 goto out;
1603         }
1604
1605         ret = o2net_set_nodelay(sc->sc_sock);
1606         if (ret) {
1607                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1608                 goto out;
1609         }
1610
1611         o2net_register_callbacks(sc->sc_sock->sk, sc);
1612
1613         spin_lock(&nn->nn_lock);
1614         /* handshake completion will set nn->nn_sc_valid */
1615         o2net_set_nn_state(nn, sc, 0, 0);
1616         spin_unlock(&nn->nn_lock);
1617
1618         remoteaddr.sin_family = AF_INET;
1619         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1620         remoteaddr.sin_port = node->nd_ipv4_port;
1621
1622         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1623                                         (struct sockaddr *)&remoteaddr,
1624                                         sizeof(remoteaddr),
1625                                         O_NONBLOCK);
1626         if (ret == -EINPROGRESS)
1627                 ret = 0;
1628
1629 out:
1630         if (ret) {
1631                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1632                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1633                 /* 0 err so that another will be queued and attempted
1634                  * from set_nn_state */
1635                 if (sc)
1636                         o2net_ensure_shutdown(nn, sc, 0);
1637         }
1638         if (sc)
1639                 sc_put(sc);
1640         if (node)
1641                 o2nm_node_put(node);
1642         if (mynode)
1643                 o2nm_node_put(mynode);
1644
1645         return;
1646 }
1647
1648 static void o2net_connect_expired(struct work_struct *work)
1649 {
1650         struct o2net_node *nn =
1651                 container_of(work, struct o2net_node, nn_connect_expired.work);
1652
1653         spin_lock(&nn->nn_lock);
1654         if (!nn->nn_sc_valid) {
1655                 mlog(ML_ERROR, "no connection established with node %u after "
1656                      "%u.%u seconds, giving up and returning errors.\n",
1657                      o2net_num_from_nn(nn),
1658                      o2net_idle_timeout() / 1000,
1659                      o2net_idle_timeout() % 1000);
1660
1661                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1662         }
1663         spin_unlock(&nn->nn_lock);
1664 }
1665
1666 static void o2net_still_up(struct work_struct *work)
1667 {
1668         struct o2net_node *nn =
1669                 container_of(work, struct o2net_node, nn_still_up.work);
1670
1671         o2quo_hb_still_up(o2net_num_from_nn(nn));
1672 }
1673
1674 /* ------------------------------------------------------------ */
1675
1676 void o2net_disconnect_node(struct o2nm_node *node)
1677 {
1678         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1679
1680         /* don't reconnect until it's heartbeating again */
1681         spin_lock(&nn->nn_lock);
1682         atomic_set(&nn->nn_timeout, 0);
1683         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1684         spin_unlock(&nn->nn_lock);
1685
1686         if (o2net_wq) {
1687                 cancel_delayed_work(&nn->nn_connect_expired);
1688                 cancel_delayed_work(&nn->nn_connect_work);
1689                 cancel_delayed_work(&nn->nn_still_up);
1690                 flush_workqueue(o2net_wq);
1691         }
1692 }
1693
1694 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1695                                   void *data)
1696 {
1697         o2quo_hb_down(node_num);
1698
1699         if (node_num != o2nm_this_node())
1700                 o2net_disconnect_node(node);
1701
1702         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1703 }
1704
1705 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1706                                 void *data)
1707 {
1708         struct o2net_node *nn = o2net_nn_from_num(node_num);
1709
1710         o2quo_hb_up(node_num);
1711
1712         /* ensure an immediate connect attempt */
1713         nn->nn_last_connect_attempt = jiffies -
1714                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1715
1716         if (node_num != o2nm_this_node()) {
1717                 /* believe it or not, accept and node hearbeating testing
1718                  * can succeed for this node before we got here.. so
1719                  * only use set_nn_state to clear the persistent error
1720                  * if that hasn't already happened */
1721                 spin_lock(&nn->nn_lock);
1722                 atomic_set(&nn->nn_timeout, 0);
1723                 if (nn->nn_persistent_error)
1724                         o2net_set_nn_state(nn, NULL, 0, 0);
1725                 spin_unlock(&nn->nn_lock);
1726         }
1727 }
1728
1729 void o2net_unregister_hb_callbacks(void)
1730 {
1731         o2hb_unregister_callback(NULL, &o2net_hb_up);
1732         o2hb_unregister_callback(NULL, &o2net_hb_down);
1733 }
1734
1735 int o2net_register_hb_callbacks(void)
1736 {
1737         int ret;
1738
1739         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1740                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1741         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1742                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1743
1744         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1745         if (ret == 0)
1746                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1747
1748         if (ret)
1749                 o2net_unregister_hb_callbacks();
1750
1751         return ret;
1752 }
1753
1754 /* ------------------------------------------------------------ */
1755
1756 static int o2net_accept_one(struct socket *sock)
1757 {
1758         int ret, slen;
1759         struct sockaddr_in sin;
1760         struct socket *new_sock = NULL;
1761         struct o2nm_node *node = NULL;
1762         struct o2net_sock_container *sc = NULL;
1763         struct o2net_node *nn;
1764
1765         BUG_ON(sock == NULL);
1766         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1767                                sock->sk->sk_protocol, &new_sock);
1768         if (ret)
1769                 goto out;
1770
1771         new_sock->type = sock->type;
1772         new_sock->ops = sock->ops;
1773         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1774         if (ret < 0)
1775                 goto out;
1776
1777         new_sock->sk->sk_allocation = GFP_ATOMIC;
1778
1779         ret = o2net_set_nodelay(new_sock);
1780         if (ret) {
1781                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1782                 goto out;
1783         }
1784
1785         slen = sizeof(sin);
1786         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1787                                        &slen, 1);
1788         if (ret < 0)
1789                 goto out;
1790
1791         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1792         if (node == NULL) {
1793                 mlog(ML_NOTICE, "attempt to connect from unknown node at "
1794                      "%u.%u.%u.%u:%d\n", NIPQUAD(sin.sin_addr.s_addr),
1795                      ntohs(sin.sin_port));
1796                 ret = -EINVAL;
1797                 goto out;
1798         }
1799
1800         if (o2nm_this_node() > node->nd_num) {
1801                 mlog(ML_NOTICE, "unexpected connect attempted from a lower "
1802                      "numbered node '%s' at " "%u.%u.%u.%u:%d with num %u\n",
1803                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1804                      ntohs(sin.sin_port), node->nd_num);
1805                 ret = -EINVAL;
1806                 goto out;
1807         }
1808
1809         /* this happens all the time when the other node sees our heartbeat
1810          * and tries to connect before we see their heartbeat */
1811         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1812                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1813                      "%u.%u.%u.%u:%d but it isn't heartbeating\n",
1814                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1815                      ntohs(sin.sin_port));
1816                 ret = -EINVAL;
1817                 goto out;
1818         }
1819
1820         nn = o2net_nn_from_num(node->nd_num);
1821
1822         spin_lock(&nn->nn_lock);
1823         if (nn->nn_sc)
1824                 ret = -EBUSY;
1825         else
1826                 ret = 0;
1827         spin_unlock(&nn->nn_lock);
1828         if (ret) {
1829                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1830                      "%u.%u.%u.%u:%d but it already has an open connection\n",
1831                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1832                      ntohs(sin.sin_port));
1833                 goto out;
1834         }
1835
1836         sc = sc_alloc(node);
1837         if (sc == NULL) {
1838                 ret = -ENOMEM;
1839                 goto out;
1840         }
1841
1842         sc->sc_sock = new_sock;
1843         new_sock = NULL;
1844
1845         spin_lock(&nn->nn_lock);
1846         atomic_set(&nn->nn_timeout, 0);
1847         o2net_set_nn_state(nn, sc, 0, 0);
1848         spin_unlock(&nn->nn_lock);
1849
1850         o2net_register_callbacks(sc->sc_sock->sk, sc);
1851         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1852
1853         o2net_initialize_handshake();
1854         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1855
1856 out:
1857         if (new_sock)
1858                 sock_release(new_sock);
1859         if (node)
1860                 o2nm_node_put(node);
1861         if (sc)
1862                 sc_put(sc);
1863         return ret;
1864 }
1865
1866 static void o2net_accept_many(struct work_struct *work)
1867 {
1868         struct socket *sock = o2net_listen_sock;
1869         while (o2net_accept_one(sock) == 0)
1870                 cond_resched();
1871 }
1872
1873 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1874 {
1875         void (*ready)(struct sock *sk, int bytes);
1876
1877         read_lock(&sk->sk_callback_lock);
1878         ready = sk->sk_user_data;
1879         if (ready == NULL) { /* check for teardown race */
1880                 ready = sk->sk_data_ready;
1881                 goto out;
1882         }
1883
1884         /* ->sk_data_ready is also called for a newly established child socket
1885          * before it has been accepted and the acceptor has set up their
1886          * data_ready.. we only want to queue listen work for our listening
1887          * socket */
1888         if (sk->sk_state == TCP_LISTEN) {
1889                 mlog(ML_TCP, "bytes: %d\n", bytes);
1890                 queue_work(o2net_wq, &o2net_listen_work);
1891         }
1892
1893 out:
1894         read_unlock(&sk->sk_callback_lock);
1895         ready(sk, bytes);
1896 }
1897
1898 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1899 {
1900         struct socket *sock = NULL;
1901         int ret;
1902         struct sockaddr_in sin = {
1903                 .sin_family = PF_INET,
1904                 .sin_addr = { .s_addr = addr },
1905                 .sin_port = port,
1906         };
1907
1908         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1909         if (ret < 0) {
1910                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1911                 goto out;
1912         }
1913
1914         sock->sk->sk_allocation = GFP_ATOMIC;
1915
1916         write_lock_bh(&sock->sk->sk_callback_lock);
1917         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1918         sock->sk->sk_data_ready = o2net_listen_data_ready;
1919         write_unlock_bh(&sock->sk->sk_callback_lock);
1920
1921         o2net_listen_sock = sock;
1922         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1923
1924         sock->sk->sk_reuse = 1;
1925         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1926         if (ret < 0) {
1927                 mlog(ML_ERROR, "unable to bind socket at %u.%u.%u.%u:%u, "
1928                      "ret=%d\n", NIPQUAD(addr), ntohs(port), ret);
1929                 goto out;
1930         }
1931
1932         ret = sock->ops->listen(sock, 64);
1933         if (ret < 0) {
1934                 mlog(ML_ERROR, "unable to listen on %u.%u.%u.%u:%u, ret=%d\n",
1935                      NIPQUAD(addr), ntohs(port), ret);
1936         }
1937
1938 out:
1939         if (ret) {
1940                 o2net_listen_sock = NULL;
1941                 if (sock)
1942                         sock_release(sock);
1943         }
1944         return ret;
1945 }
1946
1947 /*
1948  * called from node manager when we should bring up our network listening
1949  * socket.  node manager handles all the serialization to only call this
1950  * once and to match it with o2net_stop_listening().  note,
1951  * o2nm_this_node() doesn't work yet as we're being called while it
1952  * is being set up.
1953  */
1954 int o2net_start_listening(struct o2nm_node *node)
1955 {
1956         int ret = 0;
1957
1958         BUG_ON(o2net_wq != NULL);
1959         BUG_ON(o2net_listen_sock != NULL);
1960
1961         mlog(ML_KTHREAD, "starting o2net thread...\n");
1962         o2net_wq = create_singlethread_workqueue("o2net");
1963         if (o2net_wq == NULL) {
1964                 mlog(ML_ERROR, "unable to launch o2net thread\n");
1965                 return -ENOMEM; /* ? */
1966         }
1967
1968         ret = o2net_open_listening_sock(node->nd_ipv4_address,
1969                                         node->nd_ipv4_port);
1970         if (ret) {
1971                 destroy_workqueue(o2net_wq);
1972                 o2net_wq = NULL;
1973         } else
1974                 o2quo_conn_up(node->nd_num);
1975
1976         return ret;
1977 }
1978
1979 /* again, o2nm_this_node() doesn't work here as we're involved in
1980  * tearing it down */
1981 void o2net_stop_listening(struct o2nm_node *node)
1982 {
1983         struct socket *sock = o2net_listen_sock;
1984         size_t i;
1985
1986         BUG_ON(o2net_wq == NULL);
1987         BUG_ON(o2net_listen_sock == NULL);
1988
1989         /* stop the listening socket from generating work */
1990         write_lock_bh(&sock->sk->sk_callback_lock);
1991         sock->sk->sk_data_ready = sock->sk->sk_user_data;
1992         sock->sk->sk_user_data = NULL;
1993         write_unlock_bh(&sock->sk->sk_callback_lock);
1994
1995         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1996                 struct o2nm_node *node = o2nm_get_node_by_num(i);
1997                 if (node) {
1998                         o2net_disconnect_node(node);
1999                         o2nm_node_put(node);
2000                 }
2001         }
2002
2003         /* finish all work and tear down the work queue */
2004         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2005         destroy_workqueue(o2net_wq);
2006         o2net_wq = NULL;
2007
2008         sock_release(o2net_listen_sock);
2009         o2net_listen_sock = NULL;
2010
2011         o2quo_conn_err(node->nd_num);
2012 }
2013
2014 /* ------------------------------------------------------------ */
2015
2016 int o2net_init(void)
2017 {
2018         unsigned long i;
2019
2020         o2quo_init();
2021
2022         if (o2net_debugfs_init())
2023                 return -ENOMEM;
2024
2025         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2026         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2027         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2028         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2029                 kfree(o2net_hand);
2030                 kfree(o2net_keep_req);
2031                 kfree(o2net_keep_resp);
2032                 return -ENOMEM;
2033         }
2034
2035         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2036         o2net_hand->connector_id = cpu_to_be64(1);
2037
2038         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2039         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2040
2041         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2042                 struct o2net_node *nn = o2net_nn_from_num(i);
2043
2044                 atomic_set(&nn->nn_timeout, 0);
2045                 spin_lock_init(&nn->nn_lock);
2046                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2047                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2048                                   o2net_connect_expired);
2049                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2050                 /* until we see hb from a node we'll return einval */
2051                 nn->nn_persistent_error = -ENOTCONN;
2052                 init_waitqueue_head(&nn->nn_sc_wq);
2053                 idr_init(&nn->nn_status_idr);
2054                 INIT_LIST_HEAD(&nn->nn_status_list);
2055         }
2056
2057         return 0;
2058 }
2059
2060 void o2net_exit(void)
2061 {
2062         o2quo_exit();
2063         kfree(o2net_hand);
2064         kfree(o2net_keep_req);
2065         kfree(o2net_keep_resp);
2066         o2net_debugfs_exit();
2067 }