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 /*
  76  * The linux network stack isn't sparse endian clean.. It has macros like
  77  * ntohs() which perform the endian checks and structs like sockaddr_in
  78  * which aren't annotated.  So __force is found here to get the build
  79  * clean.  When they emerge from the dark ages and annotate the code
  80  * we can remove these.
  81  */
  82
  83 #define SC_NODEF_FMT "node %s (num %u) at %u.%u.%u.%u:%u"
  84 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
  85                           NIPQUAD(sc->sc_node->nd_ipv4_address),        \
  86                           ntohs(sc->sc_node->nd_ipv4_port)
  87
  88 /*
  89  * In the following two log macros, the whitespace after the ',' just
  90  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  91  * previous token if args expands to nothing.
  92  */
  93 #define msglog(hdr, fmt, args...) do {                                  \
  94         typeof(hdr) __hdr = (hdr);                                      \
  95         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
  96              "key %08x num %u] " fmt,                                   \
  97              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
  98              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
  99              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
 100              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
 101 } while (0)
 102
 103 #define sclog(sc, fmt, args...) do {                                    \
 104         typeof(sc) __sc = (sc);                                         \
 105         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
 106              "pg_off %zu] " fmt, __sc,                                  \
 107              atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,       \
 108             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
 109             ##args);                                                    \
 110 } while (0)
 111
 112 static DEFINE_RWLOCK(o2net_handler_lock);
 113 static struct rb_root o2net_handler_tree = RB_ROOT;
 114
 115 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
 116
 117 /* XXX someday we'll need better accounting */
 118 static struct socket *o2net_listen_sock = NULL;
 119
 120 /*
 121  * listen work is only queued by the listening socket callbacks on the
 122  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 123  * quorum work is queued as sock containers are shutdown.. stop_listening
 124  * tears down all the node's sock containers, preventing future shutdowns
 125  * and queued quroum work, before canceling delayed quorum work and
 126  * destroying the work queue.
 127  */
 128 static struct workqueue_struct *o2net_wq;
 129 static struct work_struct o2net_listen_work;
 130
 131 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 132 #define O2NET_HB_PRI 0x1
 133
 134 static struct o2net_handshake *o2net_hand;
 135 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 136
 137 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 138                 {[O2NET_ERR_NONE]       = 0,
 139                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
 140                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
 141                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
 142
 143 /* can't quite avoid *all* internal declarations :/ */
 144 static void o2net_sc_connect_completed(struct work_struct *work);
 145 static void o2net_rx_until_empty(struct work_struct *work);
 146 static void o2net_shutdown_sc(struct work_struct *work);
 147 static void o2net_listen_data_ready(struct sock *sk, int bytes);
 148 static void o2net_sc_send_keep_req(struct work_struct *work);
 149 static void o2net_idle_timer(unsigned long data);
 150 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 151 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 152
 153 /*
 154  * FIXME: These should use to_o2nm_cluster_from_node(), but we end up
 155  * losing our parent link to the cluster during shutdown. This can be
 156  * solved by adding a pre-removal callback to configfs, or passing
 157  * around the cluster with the node. -jeffm
 158  */
 159 static inline int o2net_reconnect_delay(struct o2nm_node *node)
 160 {
 161         return o2nm_single_cluster->cl_reconnect_delay_ms;
 162 }
 163
 164 static inline int o2net_keepalive_delay(struct o2nm_node *node)
 165 {
 166         return o2nm_single_cluster->cl_keepalive_delay_ms;
 167 }
 168
 169 static inline int o2net_idle_timeout(struct o2nm_node *node)
 170 {
 171         return o2nm_single_cluster->cl_idle_timeout_ms;
 172 }
 173
 174 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 175 {
 176         int trans;
 177         BUG_ON(err >= O2NET_ERR_MAX);
 178         trans = o2net_sys_err_translations[err];
 179
 180         /* Just in case we mess up the translation table above */
 181         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 182         return trans;
 183 }
 184
 185 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 186 {
 187         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 188         return &o2net_nodes[node_num];
 189 }
 190
 191 static u8 o2net_num_from_nn(struct o2net_node *nn)
 192 {
 193         BUG_ON(nn == NULL);
 194         return nn - o2net_nodes;
 195 }
 196
 197 /* ------------------------------------------------------------ */
 198
 199 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 200 {
 201         int ret = 0;
 202
 203         do {
 204                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
 205                         ret = -EAGAIN;
 206                         break;
 207                 }
 208                 spin_lock(&nn->nn_lock);
 209                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
 210                 if (ret == 0)
 211                         list_add_tail(&nsw->ns_node_item,
 212                                       &nn->nn_status_list);
 213                 spin_unlock(&nn->nn_lock);
 214         } while (ret == -EAGAIN);
 215
 216         if (ret == 0)  {
 217                 init_waitqueue_head(&nsw->ns_wq);
 218                 nsw->ns_sys_status = O2NET_ERR_NONE;
 219                 nsw->ns_status = 0;
 220         }
 221
 222         return ret;
 223 }
 224
 225 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 226                                       struct o2net_status_wait *nsw,
 227                                       enum o2net_system_error sys_status,
 228                                       s32 status)
 229 {
 230         assert_spin_locked(&nn->nn_lock);
 231
 232         if (!list_empty(&nsw->ns_node_item)) {
 233                 list_del_init(&nsw->ns_node_item);
 234                 nsw->ns_sys_status = sys_status;
 235                 nsw->ns_status = status;
 236                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 237                 wake_up(&nsw->ns_wq);
 238         }
 239 }
 240
 241 static void o2net_complete_nsw(struct o2net_node *nn,
 242                                struct o2net_status_wait *nsw,
 243                                u64 id, enum o2net_system_error sys_status,
 244                                s32 status)
 245 {
 246         spin_lock(&nn->nn_lock);
 247         if (nsw == NULL) {
 248                 if (id > INT_MAX)
 249                         goto out;
 250
 251                 nsw = idr_find(&nn->nn_status_idr, id);
 252                 if (nsw == NULL)
 253                         goto out;
 254         }
 255
 256         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 257
 258 out:
 259         spin_unlock(&nn->nn_lock);
 260         return;
 261 }
 262
 263 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 264 {
 265         struct o2net_status_wait *nsw, *tmp;
 266         unsigned int num_kills = 0;
 267
 268         assert_spin_locked(&nn->nn_lock);
 269
 270         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 271                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 272                 num_kills++;
 273         }
 274
 275         mlog(0, "completed %d messages for node %u\n", num_kills,
 276              o2net_num_from_nn(nn));
 277 }
 278
 279 static int o2net_nsw_completed(struct o2net_node *nn,
 280                                struct o2net_status_wait *nsw)
 281 {
 282         int completed;
 283         spin_lock(&nn->nn_lock);
 284         completed = list_empty(&nsw->ns_node_item);
 285         spin_unlock(&nn->nn_lock);
 286         return completed;
 287 }
 288
 289 /* ------------------------------------------------------------ */
 290
 291 static void sc_kref_release(struct kref *kref)
 292 {
 293         struct o2net_sock_container *sc = container_of(kref,
 294                                         struct o2net_sock_container, sc_kref);
 295         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 296
 297         sclog(sc, "releasing\n");
 298
 299         if (sc->sc_sock) {
 300                 sock_release(sc->sc_sock);
 301                 sc->sc_sock = NULL;
 302         }
 303
 304         o2nm_node_put(sc->sc_node);
 305         sc->sc_node = NULL;
 306
 307         kfree(sc);
 308 }
 309
 310 static void sc_put(struct o2net_sock_container *sc)
 311 {
 312         sclog(sc, "put\n");
 313         kref_put(&sc->sc_kref, sc_kref_release);
 314 }
 315 static void sc_get(struct o2net_sock_container *sc)
 316 {
 317         sclog(sc, "get\n");
 318         kref_get(&sc->sc_kref);
 319 }
 320 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 321 {
 322         struct o2net_sock_container *sc, *ret = NULL;
 323         struct page *page = NULL;
 324
 325         page = alloc_page(GFP_NOFS);
 326         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 327         if (sc == NULL || page == NULL)
 328                 goto out;
 329
 330         kref_init(&sc->sc_kref);
 331         o2nm_node_get(node);
 332         sc->sc_node = node;
 333
 334         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 335         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 336         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 337         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 338
 339         init_timer(&sc->sc_idle_timeout);
 340         sc->sc_idle_timeout.function = o2net_idle_timer;
 341         sc->sc_idle_timeout.data = (unsigned long)sc;
 342
 343         sclog(sc, "alloced\n");
 344
 345         ret = sc;
 346         sc->sc_page = page;
 347         sc = NULL;
 348         page = NULL;
 349
 350 out:
 351         if (page)
 352                 __free_page(page);
 353         kfree(sc);
 354
 355         return ret;
 356 }
 357
 358 /* ------------------------------------------------------------ */
 359
 360 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 361                                 struct work_struct *work)
 362 {
 363         sc_get(sc);
 364         if (!queue_work(o2net_wq, work))
 365                 sc_put(sc);
 366 }
 367 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 368                                         struct delayed_work *work,
 369                                         int delay)
 370 {
 371         sc_get(sc);
 372         if (!queue_delayed_work(o2net_wq, work, delay))
 373                 sc_put(sc);
 374 }
 375 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 376                                          struct delayed_work *work)
 377 {
 378         if (cancel_delayed_work(work))
 379                 sc_put(sc);
 380 }
 381
 382 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 383
 384 int o2net_num_connected_peers(void)
 385 {
 386         return atomic_read(&o2net_connected_peers);
 387 }
 388
 389 static void o2net_set_nn_state(struct o2net_node *nn,
 390                                struct o2net_sock_container *sc,
 391                                unsigned valid, int err)
 392 {
 393         int was_valid = nn->nn_sc_valid;
 394         int was_err = nn->nn_persistent_error;
 395         struct o2net_sock_container *old_sc = nn->nn_sc;
 396
 397         assert_spin_locked(&nn->nn_lock);
 398
 399         if (old_sc && !sc)
 400                 atomic_dec(&o2net_connected_peers);
 401         else if (!old_sc && sc)
 402                 atomic_inc(&o2net_connected_peers);
 403
 404         /* the node num comparison and single connect/accept path should stop
 405          * an non-null sc from being overwritten with another */
 406         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 407         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 408         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 409
 410         /* we won't reconnect after our valid conn goes away for
 411          * this hb iteration.. here so it shows up in the logs */
 412         if (was_valid && !valid && err == 0)
 413                 err = -ENOTCONN;
 414
 415         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 416              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 417              nn->nn_persistent_error, err);
 418
 419         nn->nn_sc = sc;
 420         nn->nn_sc_valid = valid ? 1 : 0;
 421         nn->nn_persistent_error = err;
 422
 423         /* mirrors o2net_tx_can_proceed() */
 424         if (nn->nn_persistent_error || nn->nn_sc_valid)
 425                 wake_up(&nn->nn_sc_wq);
 426
 427         if (!was_err && nn->nn_persistent_error) {
 428                 o2quo_conn_err(o2net_num_from_nn(nn));
 429                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 430                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 431         }
 432
 433         if (was_valid && !valid) {
 434                 printk(KERN_INFO "o2net: no longer connected to "
 435                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 436                 o2net_complete_nodes_nsw(nn);
 437         }
 438
 439         if (!was_valid && valid) {
 440                 o2quo_conn_up(o2net_num_from_nn(nn));
 441                 /* this is a bit of a hack.  we only try reconnecting
 442                  * when heartbeating starts until we get a connection.
 443                  * if that connection then dies we don't try reconnecting.
 444                  * the only way to start connecting again is to down
 445                  * heartbeat and bring it back up. */
 446                 cancel_delayed_work(&nn->nn_connect_expired);
 447                 printk(KERN_INFO "o2net: %s " SC_NODEF_FMT "\n",
 448                        o2nm_this_node() > sc->sc_node->nd_num ?
 449                                 "connected to" : "accepted connection from",
 450                        SC_NODEF_ARGS(sc));
 451         }
 452
 453         /* trigger the connecting worker func as long as we're not valid,
 454          * it will back off if it shouldn't connect.  This can be called
 455          * from node config teardown and so needs to be careful about
 456          * the work queue actually being up. */
 457         if (!valid && o2net_wq) {
 458                 unsigned long delay;
 459                 /* delay if we're withing a RECONNECT_DELAY of the
 460                  * last attempt */
 461                 delay = (nn->nn_last_connect_attempt +
 462                          msecs_to_jiffies(o2net_reconnect_delay(sc->sc_node)))
 463                         - jiffies;
 464                 if (delay > msecs_to_jiffies(o2net_reconnect_delay(sc->sc_node)))
 465                         delay = 0;
 466                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 467                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 468         }
 469
 470         /* keep track of the nn's sc ref for the caller */
 471         if ((old_sc == NULL) && sc)
 472                 sc_get(sc);
 473         if (old_sc && (old_sc != sc)) {
 474                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 475                 sc_put(old_sc);
 476         }
 477 }
 478
 479 /* see o2net_register_callbacks() */
 480 static void o2net_data_ready(struct sock *sk, int bytes)
 481 {
 482         void (*ready)(struct sock *sk, int bytes);
 483
 484         read_lock(&sk->sk_callback_lock);
 485         if (sk->sk_user_data) {
 486                 struct o2net_sock_container *sc = sk->sk_user_data;
 487                 sclog(sc, "data_ready hit\n");
 488                 do_gettimeofday(&sc->sc_tv_data_ready);
 489                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 490                 ready = sc->sc_data_ready;
 491         } else {
 492                 ready = sk->sk_data_ready;
 493         }
 494         read_unlock(&sk->sk_callback_lock);
 495
 496         ready(sk, bytes);
 497 }
 498
 499 /* see o2net_register_callbacks() */
 500 static void o2net_state_change(struct sock *sk)
 501 {
 502         void (*state_change)(struct sock *sk);
 503         struct o2net_sock_container *sc;
 504
 505         read_lock(&sk->sk_callback_lock);
 506         sc = sk->sk_user_data;
 507         if (sc == NULL) {
 508                 state_change = sk->sk_state_change;
 509                 goto out;
 510         }
 511
 512         sclog(sc, "state_change to %d\n", sk->sk_state);
 513
 514         state_change = sc->sc_state_change;
 515
 516         switch(sk->sk_state) {
 517                 /* ignore connecting sockets as they make progress */
 518                 case TCP_SYN_SENT:
 519                 case TCP_SYN_RECV:
 520                         break;
 521                 case TCP_ESTABLISHED:
 522                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
 523                         break;
 524                 default:
 525                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 526                         break;
 527         }
 528 out:
 529         read_unlock(&sk->sk_callback_lock);
 530         state_change(sk);
 531 }
 532
 533 /*
 534  * we register callbacks so we can queue work on events before calling
 535  * the original callbacks.  our callbacks our careful to test user_data
 536  * to discover when they've reaced with o2net_unregister_callbacks().
 537  */
 538 static void o2net_register_callbacks(struct sock *sk,
 539                                      struct o2net_sock_container *sc)
 540 {
 541         write_lock_bh(&sk->sk_callback_lock);
 542
 543         /* accepted sockets inherit the old listen socket data ready */
 544         if (sk->sk_data_ready == o2net_listen_data_ready) {
 545                 sk->sk_data_ready = sk->sk_user_data;
 546                 sk->sk_user_data = NULL;
 547         }
 548
 549         BUG_ON(sk->sk_user_data != NULL);
 550         sk->sk_user_data = sc;
 551         sc_get(sc);
 552
 553         sc->sc_data_ready = sk->sk_data_ready;
 554         sc->sc_state_change = sk->sk_state_change;
 555         sk->sk_data_ready = o2net_data_ready;
 556         sk->sk_state_change = o2net_state_change;
 557
 558         mutex_init(&sc->sc_send_lock);
 559
 560         write_unlock_bh(&sk->sk_callback_lock);
 561 }
 562
 563 static int o2net_unregister_callbacks(struct sock *sk,
 564                                    struct o2net_sock_container *sc)
 565 {
 566         int ret = 0;
 567
 568         write_lock_bh(&sk->sk_callback_lock);
 569         if (sk->sk_user_data == sc) {
 570                 ret = 1;
 571                 sk->sk_user_data = NULL;
 572                 sk->sk_data_ready = sc->sc_data_ready;
 573                 sk->sk_state_change = sc->sc_state_change;
 574         }
 575         write_unlock_bh(&sk->sk_callback_lock);
 576
 577         return ret;
 578 }
 579
 580 /*
 581  * this is a little helper that is called by callers who have seen a problem
 582  * with an sc and want to detach it from the nn if someone already hasn't beat
 583  * them to it.  if an error is given then the shutdown will be persistent
 584  * and pending transmits will be canceled.
 585  */
 586 static void o2net_ensure_shutdown(struct o2net_node *nn,
 587                                    struct o2net_sock_container *sc,
 588                                    int err)
 589 {
 590         spin_lock(&nn->nn_lock);
 591         if (nn->nn_sc == sc)
 592                 o2net_set_nn_state(nn, NULL, 0, err);
 593         spin_unlock(&nn->nn_lock);
 594 }
 595
 596 /*
 597  * This work queue function performs the blocking parts of socket shutdown.  A
 598  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 599  * sc detached from the nn.  state_change will also trigger this callback
 600  * directly when it sees errors.  In that case we need to call set_nn_state
 601  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 602  * itself.
 603  */
 604 static void o2net_shutdown_sc(struct work_struct *work)
 605 {
 606         struct o2net_sock_container *sc =
 607                 container_of(work, struct o2net_sock_container,
 608                              sc_shutdown_work);
 609         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 610
 611         sclog(sc, "shutting down\n");
 612
 613         /* drop the callbacks ref and call shutdown only once */
 614         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 615                 /* we shouldn't flush as we're in the thread, the
 616                  * races with pending sc work structs are harmless */
 617                 del_timer_sync(&sc->sc_idle_timeout);
 618                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 619                 sc_put(sc);
 620                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 621         }
 622
 623         /* not fatal so failed connects before the other guy has our
 624          * heartbeat can be retried */
 625         o2net_ensure_shutdown(nn, sc, 0);
 626         sc_put(sc);
 627 }
 628
 629 /* ------------------------------------------------------------ */
 630
 631 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 632                              u32 key)
 633 {
 634         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 635
 636         if (ret == 0)
 637                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 638
 639         return ret;
 640 }
 641
 642 static struct o2net_msg_handler *
 643 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 644                           struct rb_node **ret_parent)
 645 {
 646         struct rb_node **p = &o2net_handler_tree.rb_node;
 647         struct rb_node *parent = NULL;
 648         struct o2net_msg_handler *nmh, *ret = NULL;
 649         int cmp;
 650
 651         while (*p) {
 652                 parent = *p;
 653                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 654                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 655
 656                 if (cmp < 0)
 657                         p = &(*p)->rb_left;
 658                 else if (cmp > 0)
 659                         p = &(*p)->rb_right;
 660                 else {
 661                         ret = nmh;
 662                         break;
 663                 }
 664         }
 665
 666         if (ret_p != NULL)
 667                 *ret_p = p;
 668         if (ret_parent != NULL)
 669                 *ret_parent = parent;
 670
 671         return ret;
 672 }
 673
 674 static void o2net_handler_kref_release(struct kref *kref)
 675 {
 676         struct o2net_msg_handler *nmh;
 677         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 678
 679         kfree(nmh);
 680 }
 681
 682 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 683 {
 684         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 685 }
 686
 687 /* max_len is protection for the handler func.  incoming messages won't
 688  * be given to the handler if their payload is longer than the max. */
 689 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 690                            o2net_msg_handler_func *func, void *data,
 691                            o2net_post_msg_handler_func *post_func,
 692                            struct list_head *unreg_list)
 693 {
 694         struct o2net_msg_handler *nmh = NULL;
 695         struct rb_node **p, *parent;
 696         int ret = 0;
 697
 698         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 699                 mlog(0, "max_len for message handler out of range: %u\n",
 700                         max_len);
 701                 ret = -EINVAL;
 702                 goto out;
 703         }
 704
 705         if (!msg_type) {
 706                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 707                 ret = -EINVAL;
 708                 goto out;
 709
 710         }
 711         if (!func) {
 712                 mlog(0, "no message handler provided: %u, %p\n",
 713                        msg_type, func);
 714                 ret = -EINVAL;
 715                 goto out;
 716         }
 717
 718         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 719         if (nmh == NULL) {
 720                 ret = -ENOMEM;
 721                 goto out;
 722         }
 723
 724         nmh->nh_func = func;
 725         nmh->nh_func_data = data;
 726         nmh->nh_post_func = post_func;
 727         nmh->nh_msg_type = msg_type;
 728         nmh->nh_max_len = max_len;
 729         nmh->nh_key = key;
 730         /* the tree and list get this ref.. they're both removed in
 731          * unregister when this ref is dropped */
 732         kref_init(&nmh->nh_kref);
 733         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 734
 735         write_lock(&o2net_handler_lock);
 736         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 737                 ret = -EEXIST;
 738         else {
 739                 rb_link_node(&nmh->nh_node, parent, p);
 740                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 741                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 742
 743                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 744                      func, msg_type, key);
 745                 /* we've had some trouble with handlers seemingly vanishing. */
 746                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 747                                                           &parent) == NULL,
 748                                 "couldn't find handler we *just* registerd "
 749                                 "for type %u key %08x\n", msg_type, key);
 750         }
 751         write_unlock(&o2net_handler_lock);
 752         if (ret)
 753                 goto out;
 754
 755 out:
 756         if (ret)
 757                 kfree(nmh);
 758
 759         return ret;
 760 }
 761 EXPORT_SYMBOL_GPL(o2net_register_handler);
 762
 763 void o2net_unregister_handler_list(struct list_head *list)
 764 {
 765         struct o2net_msg_handler *nmh, *n;
 766
 767         write_lock(&o2net_handler_lock);
 768         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 769                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 770                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 771                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 772                 list_del_init(&nmh->nh_unregister_item);
 773                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 774         }
 775         write_unlock(&o2net_handler_lock);
 776 }
 777 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 778
 779 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 780 {
 781         struct o2net_msg_handler *nmh;
 782
 783         read_lock(&o2net_handler_lock);
 784         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 785         if (nmh)
 786                 kref_get(&nmh->nh_kref);
 787         read_unlock(&o2net_handler_lock);
 788
 789         return nmh;
 790 }
 791
 792 /* ------------------------------------------------------------ */
 793
 794 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 795 {
 796         int ret;
 797         mm_segment_t oldfs;
 798         struct kvec vec = {
 799                 .iov_len = len,
 800                 .iov_base = data,
 801         };
 802         struct msghdr msg = {
 803                 .msg_iovlen = 1,
 804                 .msg_iov = (struct iovec *)&vec,
 805                 .msg_flags = MSG_DONTWAIT,
 806         };
 807
 808         oldfs = get_fs();
 809         set_fs(get_ds());
 810         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
 811         set_fs(oldfs);
 812
 813         return ret;
 814 }
 815
 816 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 817                               size_t veclen, size_t total)
 818 {
 819         int ret;
 820         mm_segment_t oldfs;
 821         struct msghdr msg = {
 822                 .msg_iov = (struct iovec *)vec,
 823                 .msg_iovlen = veclen,
 824         };
 825
 826         if (sock == NULL) {
 827                 ret = -EINVAL;
 828                 goto out;
 829         }
 830
 831         oldfs = get_fs();
 832         set_fs(get_ds());
 833         ret = sock_sendmsg(sock, &msg, total);
 834         set_fs(oldfs);
 835         if (ret != total) {
 836                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
 837                      total);
 838                 if (ret >= 0)
 839                         ret = -EPIPE; /* should be smarter, I bet */
 840                 goto out;
 841         }
 842
 843         ret = 0;
 844 out:
 845         if (ret < 0)
 846                 mlog(0, "returning error: %d\n", ret);
 847         return ret;
 848 }
 849
 850 static void o2net_sendpage(struct o2net_sock_container *sc,
 851                            void *kmalloced_virt,
 852                            size_t size)
 853 {
 854         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 855         ssize_t ret;
 856
 857         while (1) {
 858                 mutex_lock(&sc->sc_send_lock);
 859                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 860                                                  virt_to_page(kmalloced_virt),
 861                                                  (long)kmalloced_virt & ~PAGE_MASK,
 862                                                  size, MSG_DONTWAIT);
 863                 mutex_unlock(&sc->sc_send_lock);
 864                 if (ret == size)
 865                         break;
 866                 if (ret == (ssize_t)-EAGAIN) {
 867                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 868                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 869                         cond_resched();
 870                         continue;
 871                 }
 872                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 873                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 874                 o2net_ensure_shutdown(nn, sc, 0);
 875                 break;
 876         }
 877 }
 878
 879 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 880 {
 881         memset(msg, 0, sizeof(struct o2net_msg));
 882         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 883         msg->data_len = cpu_to_be16(data_len);
 884         msg->msg_type = cpu_to_be16(msg_type);
 885         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 886         msg->status = 0;
 887         msg->key = cpu_to_be32(key);
 888 }
 889
 890 static int o2net_tx_can_proceed(struct o2net_node *nn,
 891                                 struct o2net_sock_container **sc_ret,
 892                                 int *error)
 893 {
 894         int ret = 0;
 895
 896         spin_lock(&nn->nn_lock);
 897         if (nn->nn_persistent_error) {
 898                 ret = 1;
 899                 *sc_ret = NULL;
 900                 *error = nn->nn_persistent_error;
 901         } else if (nn->nn_sc_valid) {
 902                 kref_get(&nn->nn_sc->sc_kref);
 903
 904                 ret = 1;
 905                 *sc_ret = nn->nn_sc;
 906                 *error = 0;
 907         }
 908         spin_unlock(&nn->nn_lock);
 909
 910         return ret;
 911 }
 912
 913 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
 914                            size_t caller_veclen, u8 target_node, int *status)
 915 {
 916         int ret, error = 0;
 917         struct o2net_msg *msg = NULL;
 918         size_t veclen, caller_bytes = 0;
 919         struct kvec *vec = NULL;
 920         struct o2net_sock_container *sc = NULL;
 921         struct o2net_node *nn = o2net_nn_from_num(target_node);
 922         struct o2net_status_wait nsw = {
 923                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
 924         };
 925
 926         if (o2net_wq == NULL) {
 927                 mlog(0, "attempt to tx without o2netd running\n");
 928                 ret = -ESRCH;
 929                 goto out;
 930         }
 931
 932         if (caller_veclen == 0) {
 933                 mlog(0, "bad kvec array length\n");
 934                 ret = -EINVAL;
 935                 goto out;
 936         }
 937
 938         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
 939         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
 940                 mlog(0, "total payload len %zu too large\n", caller_bytes);
 941                 ret = -EINVAL;
 942                 goto out;
 943         }
 944
 945         if (target_node == o2nm_this_node()) {
 946                 ret = -ELOOP;
 947                 goto out;
 948         }
 949
 950         ret = wait_event_interruptible(nn->nn_sc_wq,
 951                                        o2net_tx_can_proceed(nn, &sc, &error));
 952         if (!ret && error)
 953                 ret = error;
 954         if (ret)
 955                 goto out;
 956
 957         veclen = caller_veclen + 1;
 958         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
 959         if (vec == NULL) {
 960                 mlog(0, "failed to %zu element kvec!\n", veclen);
 961                 ret = -ENOMEM;
 962                 goto out;
 963         }
 964
 965         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
 966         if (!msg) {
 967                 mlog(0, "failed to allocate a o2net_msg!\n");
 968                 ret = -ENOMEM;
 969                 goto out;
 970         }
 971
 972         o2net_init_msg(msg, caller_bytes, msg_type, key);
 973
 974         vec[0].iov_len = sizeof(struct o2net_msg);
 975         vec[0].iov_base = msg;
 976         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
 977
 978         ret = o2net_prep_nsw(nn, &nsw);
 979         if (ret)
 980                 goto out;
 981
 982         msg->msg_num = cpu_to_be32(nsw.ns_id);
 983
 984         /* finally, convert the message header to network byte-order
 985          * and send */
 986         mutex_lock(&sc->sc_send_lock);
 987         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
 988                                  sizeof(struct o2net_msg) + caller_bytes);
 989         mutex_unlock(&sc->sc_send_lock);
 990         msglog(msg, "sending returned %d\n", ret);
 991         if (ret < 0) {
 992                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
 993                 goto out;
 994         }
 995
 996         /* wait on other node's handler */
 997         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
 998
 999         /* Note that we avoid overwriting the callers status return
1000          * variable if a system error was reported on the other
1001          * side. Callers beware. */
1002         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1003         if (status && !ret)
1004                 *status = nsw.ns_status;
1005
1006         mlog(0, "woken, returning system status %d, user status %d\n",
1007              ret, nsw.ns_status);
1008 out:
1009         if (sc)
1010                 sc_put(sc);
1011         if (vec)
1012                 kfree(vec);
1013         if (msg)
1014                 kfree(msg);
1015         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1016         return ret;
1017 }
1018 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1019
1020 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1021                        u8 target_node, int *status)
1022 {
1023         struct kvec vec = {
1024                 .iov_base = data,
1025                 .iov_len = len,
1026         };
1027         return o2net_send_message_vec(msg_type, key, &vec, 1,
1028                                       target_node, status);
1029 }
1030 EXPORT_SYMBOL_GPL(o2net_send_message);
1031
1032 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1033                                    enum o2net_system_error syserr, int err)
1034 {
1035         struct kvec vec = {
1036                 .iov_base = hdr,
1037                 .iov_len = sizeof(struct o2net_msg),
1038         };
1039
1040         BUG_ON(syserr >= O2NET_ERR_MAX);
1041
1042         /* leave other fields intact from the incoming message, msg_num
1043          * in particular */
1044         hdr->sys_status = cpu_to_be32(syserr);
1045         hdr->status = cpu_to_be32(err);
1046         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1047         hdr->data_len = 0;
1048
1049         msglog(hdr, "about to send status magic %d\n", err);
1050         /* hdr has been in host byteorder this whole time */
1051         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1052 }
1053
1054 /* this returns -errno if the header was unknown or too large, etc.
1055  * after this is called the buffer us reused for the next message */
1056 static int o2net_process_message(struct o2net_sock_container *sc,
1057                                  struct o2net_msg *hdr)
1058 {
1059         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1060         int ret = 0, handler_status;
1061         enum  o2net_system_error syserr;
1062         struct o2net_msg_handler *nmh = NULL;
1063         void *ret_data = NULL;
1064
1065         msglog(hdr, "processing message\n");
1066
1067         o2net_sc_postpone_idle(sc);
1068
1069         switch(be16_to_cpu(hdr->magic)) {
1070                 case O2NET_MSG_STATUS_MAGIC:
1071                         /* special type for returning message status */
1072                         o2net_complete_nsw(nn, NULL,
1073                                            be32_to_cpu(hdr->msg_num),
1074                                            be32_to_cpu(hdr->sys_status),
1075                                            be32_to_cpu(hdr->status));
1076                         goto out;
1077                 case O2NET_MSG_KEEP_REQ_MAGIC:
1078                         o2net_sendpage(sc, o2net_keep_resp,
1079                                        sizeof(*o2net_keep_resp));
1080                         goto out;
1081                 case O2NET_MSG_KEEP_RESP_MAGIC:
1082                         goto out;
1083                 case O2NET_MSG_MAGIC:
1084                         break;
1085                 default:
1086                         msglog(hdr, "bad magic\n");
1087                         ret = -EINVAL;
1088                         goto out;
1089                         break;
1090         }
1091
1092         /* find a handler for it */
1093         handler_status = 0;
1094         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1095                                 be32_to_cpu(hdr->key));
1096         if (!nmh) {
1097                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1098                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1099                 syserr = O2NET_ERR_NO_HNDLR;
1100                 goto out_respond;
1101         }
1102
1103         syserr = O2NET_ERR_NONE;
1104
1105         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1106                 syserr = O2NET_ERR_OVERFLOW;
1107
1108         if (syserr != O2NET_ERR_NONE)
1109                 goto out_respond;
1110
1111         do_gettimeofday(&sc->sc_tv_func_start);
1112         sc->sc_msg_key = be32_to_cpu(hdr->key);
1113         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1114         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1115                                              be16_to_cpu(hdr->data_len),
1116                                         nmh->nh_func_data, &ret_data);
1117         do_gettimeofday(&sc->sc_tv_func_stop);
1118
1119 out_respond:
1120         /* this destroys the hdr, so don't use it after this */
1121         mutex_lock(&sc->sc_send_lock);
1122         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1123                                       handler_status);
1124         mutex_unlock(&sc->sc_send_lock);
1125         hdr = NULL;
1126         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1127              handler_status, syserr, ret);
1128
1129         if (nmh) {
1130                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1131                 if (nmh->nh_post_func)
1132                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1133                                             ret_data);
1134         }
1135
1136 out:
1137         if (nmh)
1138                 o2net_handler_put(nmh);
1139         return ret;
1140 }
1141
1142 static int o2net_check_handshake(struct o2net_sock_container *sc)
1143 {
1144         struct o2net_handshake *hand = page_address(sc->sc_page);
1145         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1146
1147         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1148                 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1149                      "version %llu but %llu is required, disconnecting\n",
1150                      SC_NODEF_ARGS(sc),
1151                      (unsigned long long)be64_to_cpu(hand->protocol_version),
1152                      O2NET_PROTOCOL_VERSION);
1153
1154                 /* don't bother reconnecting if its the wrong version. */
1155                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1156                 return -1;
1157         }
1158
1159         /*
1160          * Ensure timeouts are consistent with other nodes, otherwise
1161          * we can end up with one node thinking that the other must be down,
1162          * but isn't. This can ultimately cause corruption.
1163          */
1164         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1165                                 o2net_idle_timeout(sc->sc_node)) {
1166                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1167                      "%u ms, but we use %u ms locally.  disconnecting\n",
1168                      SC_NODEF_ARGS(sc),
1169                      be32_to_cpu(hand->o2net_idle_timeout_ms),
1170                      o2net_idle_timeout(sc->sc_node));
1171                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1172                 return -1;
1173         }
1174
1175         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1176                         o2net_keepalive_delay(sc->sc_node)) {
1177                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1178                      "%u ms, but we use %u ms locally.  disconnecting\n",
1179                      SC_NODEF_ARGS(sc),
1180                      be32_to_cpu(hand->o2net_keepalive_delay_ms),
1181                      o2net_keepalive_delay(sc->sc_node));
1182                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1183                 return -1;
1184         }
1185
1186         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1187                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1188                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1189                      "%u ms, but we use %u ms locally.  disconnecting\n",
1190                      SC_NODEF_ARGS(sc),
1191                      be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1192                      O2HB_MAX_WRITE_TIMEOUT_MS);
1193                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1194                 return -1;
1195         }
1196
1197         sc->sc_handshake_ok = 1;
1198
1199         spin_lock(&nn->nn_lock);
1200         /* set valid and queue the idle timers only if it hasn't been
1201          * shut down already */
1202         if (nn->nn_sc == sc) {
1203                 o2net_sc_reset_idle_timer(sc);
1204                 o2net_set_nn_state(nn, sc, 1, 0);
1205         }
1206         spin_unlock(&nn->nn_lock);
1207
1208         /* shift everything up as though it wasn't there */
1209         sc->sc_page_off -= sizeof(struct o2net_handshake);
1210         if (sc->sc_page_off)
1211                 memmove(hand, hand + 1, sc->sc_page_off);
1212
1213         return 0;
1214 }
1215
1216 /* this demuxes the queued rx bytes into header or payload bits and calls
1217  * handlers as each full message is read off the socket.  it returns -error,
1218  * == 0 eof, or > 0 for progress made.*/
1219 static int o2net_advance_rx(struct o2net_sock_container *sc)
1220 {
1221         struct o2net_msg *hdr;
1222         int ret = 0;
1223         void *data;
1224         size_t datalen;
1225
1226         sclog(sc, "receiving\n");
1227         do_gettimeofday(&sc->sc_tv_advance_start);
1228
1229         if (unlikely(sc->sc_handshake_ok == 0)) {
1230                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1231                         data = page_address(sc->sc_page) + sc->sc_page_off;
1232                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1233                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1234                         if (ret > 0)
1235                                 sc->sc_page_off += ret;
1236                 }
1237
1238                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1239                         o2net_check_handshake(sc);
1240                         if (unlikely(sc->sc_handshake_ok == 0))
1241                                 ret = -EPROTO;
1242                 }
1243                 goto out;
1244         }
1245
1246         /* do we need more header? */
1247         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1248                 data = page_address(sc->sc_page) + sc->sc_page_off;
1249                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1250                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1251                 if (ret > 0) {
1252                         sc->sc_page_off += ret;
1253                         /* only swab incoming here.. we can
1254                          * only get here once as we cross from
1255                          * being under to over */
1256                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1257                                 hdr = page_address(sc->sc_page);
1258                                 if (be16_to_cpu(hdr->data_len) >
1259                                     O2NET_MAX_PAYLOAD_BYTES)
1260                                         ret = -EOVERFLOW;
1261                         }
1262                 }
1263                 if (ret <= 0)
1264                         goto out;
1265         }
1266
1267         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1268                 /* oof, still don't have a header */
1269                 goto out;
1270         }
1271
1272         /* this was swabbed above when we first read it */
1273         hdr = page_address(sc->sc_page);
1274
1275         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1276
1277         /* do we need more payload? */
1278         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1279                 /* need more payload */
1280                 data = page_address(sc->sc_page) + sc->sc_page_off;
1281                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1282                           sc->sc_page_off;
1283                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1284                 if (ret > 0)
1285                         sc->sc_page_off += ret;
1286                 if (ret <= 0)
1287                         goto out;
1288         }
1289
1290         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1291                 /* we can only get here once, the first time we read
1292                  * the payload.. so set ret to progress if the handler
1293                  * works out. after calling this the message is toast */
1294                 ret = o2net_process_message(sc, hdr);
1295                 if (ret == 0)
1296                         ret = 1;
1297                 sc->sc_page_off = 0;
1298         }
1299
1300 out:
1301         sclog(sc, "ret = %d\n", ret);
1302         do_gettimeofday(&sc->sc_tv_advance_stop);
1303         return ret;
1304 }
1305
1306 /* this work func is triggerd by data ready.  it reads until it can read no
1307  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1308  * our work the work struct will be marked and we'll be called again. */
1309 static void o2net_rx_until_empty(struct work_struct *work)
1310 {
1311         struct o2net_sock_container *sc =
1312                 container_of(work, struct o2net_sock_container, sc_rx_work);
1313         int ret;
1314
1315         do {
1316                 ret = o2net_advance_rx(sc);
1317         } while (ret > 0);
1318
1319         if (ret <= 0 && ret != -EAGAIN) {
1320                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1321                 sclog(sc, "saw error %d, closing\n", ret);
1322                 /* not permanent so read failed handshake can retry */
1323                 o2net_ensure_shutdown(nn, sc, 0);
1324         }
1325
1326         sc_put(sc);
1327 }
1328
1329 static int o2net_set_nodelay(struct socket *sock)
1330 {
1331         int ret, val = 1;
1332         mm_segment_t oldfs;
1333
1334         oldfs = get_fs();
1335         set_fs(KERNEL_DS);
1336
1337         /*
1338          * Dear unsuspecting programmer,
1339          *
1340          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1341          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1342          * silently turn into SO_DEBUG.
1343          *
1344          * Yours,
1345          * Keeper of hilariously fragile interfaces.
1346          */
1347         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1348                                     (char __user *)&val, sizeof(val));
1349
1350         set_fs(oldfs);
1351         return ret;
1352 }
1353
1354 static void o2net_initialize_handshake(void)
1355 {
1356         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1357                 O2HB_MAX_WRITE_TIMEOUT_MS);
1358         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(
1359                 o2net_idle_timeout(NULL));
1360         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1361                 o2net_keepalive_delay(NULL));
1362         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1363                 o2net_reconnect_delay(NULL));
1364 }
1365
1366 /* ------------------------------------------------------------ */
1367
1368 /* called when a connect completes and after a sock is accepted.  the
1369  * rx path will see the response and mark the sc valid */
1370 static void o2net_sc_connect_completed(struct work_struct *work)
1371 {
1372         struct o2net_sock_container *sc =
1373                 container_of(work, struct o2net_sock_container,
1374                              sc_connect_work);
1375
1376         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1377               (unsigned long long)O2NET_PROTOCOL_VERSION,
1378               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1379
1380         o2net_initialize_handshake();
1381         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1382         sc_put(sc);
1383 }
1384
1385 /* this is called as a work_struct func. */
1386 static void o2net_sc_send_keep_req(struct work_struct *work)
1387 {
1388         struct o2net_sock_container *sc =
1389                 container_of(work, struct o2net_sock_container,
1390                              sc_keepalive_work.work);
1391
1392         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1393         sc_put(sc);
1394 }
1395
1396 /* socket shutdown does a del_timer_sync against this as it tears down.
1397  * we can't start this timer until we've got to the point in sc buildup
1398  * where shutdown is going to be involved */
1399 static void o2net_idle_timer(unsigned long data)
1400 {
1401         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1402         struct timeval now;
1403
1404         do_gettimeofday(&now);
1405
1406         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1407              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1408                      o2net_idle_timeout(sc->sc_node) / 1000,
1409                      o2net_idle_timeout(sc->sc_node) % 1000);
1410         mlog(ML_NOTICE, "here are some times that might help debug the "
1411              "situation: (tmr %ld.%ld now %ld.%ld dr %ld.%ld adv "
1412              "%ld.%ld:%ld.%ld func (%08x:%u) %ld.%ld:%ld.%ld)\n",
1413              sc->sc_tv_timer.tv_sec, (long) sc->sc_tv_timer.tv_usec,
1414              now.tv_sec, (long) now.tv_usec,
1415              sc->sc_tv_data_ready.tv_sec, (long) sc->sc_tv_data_ready.tv_usec,
1416              sc->sc_tv_advance_start.tv_sec,
1417              (long) sc->sc_tv_advance_start.tv_usec,
1418              sc->sc_tv_advance_stop.tv_sec,
1419              (long) sc->sc_tv_advance_stop.tv_usec,
1420              sc->sc_msg_key, sc->sc_msg_type,
1421              sc->sc_tv_func_start.tv_sec, (long) sc->sc_tv_func_start.tv_usec,
1422              sc->sc_tv_func_stop.tv_sec, (long) sc->sc_tv_func_stop.tv_usec);
1423
1424         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1425 }
1426
1427 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1428 {
1429         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1430         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1431                       msecs_to_jiffies(o2net_keepalive_delay(sc->sc_node)));
1432         do_gettimeofday(&sc->sc_tv_timer);
1433         mod_timer(&sc->sc_idle_timeout,
1434                jiffies + msecs_to_jiffies(o2net_idle_timeout(sc->sc_node)));
1435 }
1436
1437 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1438 {
1439         /* Only push out an existing timer */
1440         if (timer_pending(&sc->sc_idle_timeout))
1441                 o2net_sc_reset_idle_timer(sc);
1442 }
1443
1444 /* this work func is kicked whenever a path sets the nn state which doesn't
1445  * have valid set.  This includes seeing hb come up, losing a connection,
1446  * having a connect attempt fail, etc. This centralizes the logic which decides
1447  * if a connect attempt should be made or if we should give up and all future
1448  * transmit attempts should fail */
1449 static void o2net_start_connect(struct work_struct *work)
1450 {
1451         struct o2net_node *nn =
1452                 container_of(work, struct o2net_node, nn_connect_work.work);
1453         struct o2net_sock_container *sc = NULL;
1454         struct o2nm_node *node = NULL, *mynode = NULL;
1455         struct socket *sock = NULL;
1456         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1457         int ret = 0, stop;
1458
1459         /* if we're greater we initiate tx, otherwise we accept */
1460         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1461                 goto out;
1462
1463         /* watch for racing with tearing a node down */
1464         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1465         if (node == NULL) {
1466                 ret = 0;
1467                 goto out;
1468         }
1469
1470         mynode = o2nm_get_node_by_num(o2nm_this_node());
1471         if (mynode == NULL) {
1472                 ret = 0;
1473                 goto out;
1474         }
1475
1476         spin_lock(&nn->nn_lock);
1477         /* see if we already have one pending or have given up */
1478         stop = (nn->nn_sc || nn->nn_persistent_error);
1479         spin_unlock(&nn->nn_lock);
1480         if (stop)
1481                 goto out;
1482
1483         nn->nn_last_connect_attempt = jiffies;
1484
1485         sc = sc_alloc(node);
1486         if (sc == NULL) {
1487                 mlog(0, "couldn't allocate sc\n");
1488                 ret = -ENOMEM;
1489                 goto out;
1490         }
1491
1492         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1493         if (ret < 0) {
1494                 mlog(0, "can't create socket: %d\n", ret);
1495                 goto out;
1496         }
1497         sc->sc_sock = sock; /* freed by sc_kref_release */
1498
1499         sock->sk->sk_allocation = GFP_ATOMIC;
1500
1501         myaddr.sin_family = AF_INET;
1502         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1503         myaddr.sin_port = (__force u16)htons(0); /* any port */
1504
1505         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1506                               sizeof(myaddr));
1507         if (ret) {
1508                 mlog(ML_ERROR, "bind failed with %d at address %u.%u.%u.%u\n",
1509                      ret, NIPQUAD(mynode->nd_ipv4_address));
1510                 goto out;
1511         }
1512
1513         ret = o2net_set_nodelay(sc->sc_sock);
1514         if (ret) {
1515                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1516                 goto out;
1517         }
1518
1519         o2net_register_callbacks(sc->sc_sock->sk, sc);
1520
1521         spin_lock(&nn->nn_lock);
1522         /* handshake completion will set nn->nn_sc_valid */
1523         o2net_set_nn_state(nn, sc, 0, 0);
1524         spin_unlock(&nn->nn_lock);
1525
1526         remoteaddr.sin_family = AF_INET;
1527         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1528         remoteaddr.sin_port = node->nd_ipv4_port;
1529
1530         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1531                                         (struct sockaddr *)&remoteaddr,
1532                                         sizeof(remoteaddr),
1533                                         O_NONBLOCK);
1534         if (ret == -EINPROGRESS)
1535                 ret = 0;
1536
1537 out:
1538         if (ret) {
1539                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1540                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1541                 /* 0 err so that another will be queued and attempted
1542                  * from set_nn_state */
1543                 if (sc)
1544                         o2net_ensure_shutdown(nn, sc, 0);
1545         }
1546         if (sc)
1547                 sc_put(sc);
1548         if (node)
1549                 o2nm_node_put(node);
1550         if (mynode)
1551                 o2nm_node_put(mynode);
1552
1553         return;
1554 }
1555
1556 static void o2net_connect_expired(struct work_struct *work)
1557 {
1558         struct o2net_node *nn =
1559                 container_of(work, struct o2net_node, nn_connect_expired.work);
1560
1561         spin_lock(&nn->nn_lock);
1562         if (!nn->nn_sc_valid) {
1563                 struct o2nm_node *node = nn->nn_sc->sc_node;
1564                 mlog(ML_ERROR, "no connection established with node %u after "
1565                      "%u.%u seconds, giving up and returning errors.\n",
1566                      o2net_num_from_nn(nn),
1567                      o2net_idle_timeout(node) / 1000,
1568                      o2net_idle_timeout(node) % 1000);
1569
1570                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1571         }
1572         spin_unlock(&nn->nn_lock);
1573 }
1574
1575 static void o2net_still_up(struct work_struct *work)
1576 {
1577         struct o2net_node *nn =
1578                 container_of(work, struct o2net_node, nn_still_up.work);
1579
1580         o2quo_hb_still_up(o2net_num_from_nn(nn));
1581 }
1582
1583 /* ------------------------------------------------------------ */
1584
1585 void o2net_disconnect_node(struct o2nm_node *node)
1586 {
1587         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1588
1589         /* don't reconnect until it's heartbeating again */
1590         spin_lock(&nn->nn_lock);
1591         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1592         spin_unlock(&nn->nn_lock);
1593
1594         if (o2net_wq) {
1595                 cancel_delayed_work(&nn->nn_connect_expired);
1596                 cancel_delayed_work(&nn->nn_connect_work);
1597                 cancel_delayed_work(&nn->nn_still_up);
1598                 flush_workqueue(o2net_wq);
1599         }
1600 }
1601
1602 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1603                                   void *data)
1604 {
1605         o2quo_hb_down(node_num);
1606
1607         if (node_num != o2nm_this_node())
1608                 o2net_disconnect_node(node);
1609
1610         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1611 }
1612
1613 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1614                                 void *data)
1615 {
1616         struct o2net_node *nn = o2net_nn_from_num(node_num);
1617
1618         o2quo_hb_up(node_num);
1619
1620         /* ensure an immediate connect attempt */
1621         nn->nn_last_connect_attempt = jiffies -
1622                 (msecs_to_jiffies(o2net_reconnect_delay(node)) + 1);
1623
1624         if (node_num != o2nm_this_node()) {
1625                 /* heartbeat doesn't work unless a local node number is
1626                  * configured and doing so brings up the o2net_wq, so we can
1627                  * use it.. */
1628                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired,
1629                                    msecs_to_jiffies(o2net_idle_timeout(node)));
1630
1631                 /* believe it or not, accept and node hearbeating testing
1632                  * can succeed for this node before we got here.. so
1633                  * only use set_nn_state to clear the persistent error
1634                  * if that hasn't already happened */
1635                 spin_lock(&nn->nn_lock);
1636                 if (nn->nn_persistent_error)
1637                         o2net_set_nn_state(nn, NULL, 0, 0);
1638                 spin_unlock(&nn->nn_lock);
1639         }
1640 }
1641
1642 void o2net_unregister_hb_callbacks(void)
1643 {
1644         o2hb_unregister_callback(NULL, &o2net_hb_up);
1645         o2hb_unregister_callback(NULL, &o2net_hb_down);
1646 }
1647
1648 int o2net_register_hb_callbacks(void)
1649 {
1650         int ret;
1651
1652         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1653                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1654         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1655                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1656
1657         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1658         if (ret == 0)
1659                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1660
1661         if (ret)
1662                 o2net_unregister_hb_callbacks();
1663
1664         return ret;
1665 }
1666
1667 /* ------------------------------------------------------------ */
1668
1669 static int o2net_accept_one(struct socket *sock)
1670 {
1671         int ret, slen;
1672         struct sockaddr_in sin;
1673         struct socket *new_sock = NULL;
1674         struct o2nm_node *node = NULL;
1675         struct o2net_sock_container *sc = NULL;
1676         struct o2net_node *nn;
1677
1678         BUG_ON(sock == NULL);
1679         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1680                                sock->sk->sk_protocol, &new_sock);
1681         if (ret)
1682                 goto out;
1683
1684         new_sock->type = sock->type;
1685         new_sock->ops = sock->ops;
1686         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1687         if (ret < 0)
1688                 goto out;
1689
1690         new_sock->sk->sk_allocation = GFP_ATOMIC;
1691
1692         ret = o2net_set_nodelay(new_sock);
1693         if (ret) {
1694                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1695                 goto out;
1696         }
1697
1698         slen = sizeof(sin);
1699         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1700                                        &slen, 1);
1701         if (ret < 0)
1702                 goto out;
1703
1704         node = o2nm_get_node_by_ip((__force __be32)sin.sin_addr.s_addr);
1705         if (node == NULL) {
1706                 mlog(ML_NOTICE, "attempt to connect from unknown node at "
1707                      "%u.%u.%u.%u:%d\n", NIPQUAD(sin.sin_addr.s_addr),
1708                      ntohs((__force __be16)sin.sin_port));
1709                 ret = -EINVAL;
1710                 goto out;
1711         }
1712
1713         if (o2nm_this_node() > node->nd_num) {
1714                 mlog(ML_NOTICE, "unexpected connect attempted from a lower "
1715                      "numbered node '%s' at " "%u.%u.%u.%u:%d with num %u\n",
1716                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1717                      ntohs((__force __be16)sin.sin_port), node->nd_num);
1718                 ret = -EINVAL;
1719                 goto out;
1720         }
1721
1722         /* this happens all the time when the other node sees our heartbeat
1723          * and tries to connect before we see their heartbeat */
1724         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1725                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1726                      "%u.%u.%u.%u:%d but it isn't heartbeating\n",
1727                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1728                      ntohs((__force __be16)sin.sin_port));
1729                 ret = -EINVAL;
1730                 goto out;
1731         }
1732
1733         nn = o2net_nn_from_num(node->nd_num);
1734
1735         spin_lock(&nn->nn_lock);
1736         if (nn->nn_sc)
1737                 ret = -EBUSY;
1738         else
1739                 ret = 0;
1740         spin_unlock(&nn->nn_lock);
1741         if (ret) {
1742                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1743                      "%u.%u.%u.%u:%d but it already has an open connection\n",
1744                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1745                      ntohs((__force __be16)sin.sin_port));
1746                 goto out;
1747         }
1748
1749         sc = sc_alloc(node);
1750         if (sc == NULL) {
1751                 ret = -ENOMEM;
1752                 goto out;
1753         }
1754
1755         sc->sc_sock = new_sock;
1756         new_sock = NULL;
1757
1758         spin_lock(&nn->nn_lock);
1759         o2net_set_nn_state(nn, sc, 0, 0);
1760         spin_unlock(&nn->nn_lock);
1761
1762         o2net_register_callbacks(sc->sc_sock->sk, sc);
1763         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1764
1765         o2net_initialize_handshake();
1766         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1767
1768 out:
1769         if (new_sock)
1770                 sock_release(new_sock);
1771         if (node)
1772                 o2nm_node_put(node);
1773         if (sc)
1774                 sc_put(sc);
1775         return ret;
1776 }
1777
1778 static void o2net_accept_many(struct work_struct *work)
1779 {
1780         struct socket *sock = o2net_listen_sock;
1781         while (o2net_accept_one(sock) == 0)
1782                 cond_resched();
1783 }
1784
1785 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1786 {
1787         void (*ready)(struct sock *sk, int bytes);
1788
1789         read_lock(&sk->sk_callback_lock);
1790         ready = sk->sk_user_data;
1791         if (ready == NULL) { /* check for teardown race */
1792                 ready = sk->sk_data_ready;
1793                 goto out;
1794         }
1795
1796         /* ->sk_data_ready is also called for a newly established child socket
1797          * before it has been accepted and the acceptor has set up their
1798          * data_ready.. we only want to queue listen work for our listening
1799          * socket */
1800         if (sk->sk_state == TCP_LISTEN) {
1801                 mlog(ML_TCP, "bytes: %d\n", bytes);
1802                 queue_work(o2net_wq, &o2net_listen_work);
1803         }
1804
1805 out:
1806         read_unlock(&sk->sk_callback_lock);
1807         ready(sk, bytes);
1808 }
1809
1810 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1811 {
1812         struct socket *sock = NULL;
1813         int ret;
1814         struct sockaddr_in sin = {
1815                 .sin_family = PF_INET,
1816                 .sin_addr = { .s_addr = addr },
1817                 .sin_port = port,
1818         };
1819
1820         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1821         if (ret < 0) {
1822                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1823                 goto out;
1824         }
1825
1826         sock->sk->sk_allocation = GFP_ATOMIC;
1827
1828         write_lock_bh(&sock->sk->sk_callback_lock);
1829         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1830         sock->sk->sk_data_ready = o2net_listen_data_ready;
1831         write_unlock_bh(&sock->sk->sk_callback_lock);
1832
1833         o2net_listen_sock = sock;
1834         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1835
1836         sock->sk->sk_reuse = 1;
1837         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1838         if (ret < 0) {
1839                 mlog(ML_ERROR, "unable to bind socket at %u.%u.%u.%u:%u, "
1840                      "ret=%d\n", NIPQUAD(addr), ntohs(port), ret);
1841                 goto out;
1842         }
1843
1844         ret = sock->ops->listen(sock, 64);
1845         if (ret < 0) {
1846                 mlog(ML_ERROR, "unable to listen on %u.%u.%u.%u:%u, ret=%d\n",
1847                      NIPQUAD(addr), ntohs(port), ret);
1848         }
1849
1850 out:
1851         if (ret) {
1852                 o2net_listen_sock = NULL;
1853                 if (sock)
1854                         sock_release(sock);
1855         }
1856         return ret;
1857 }
1858
1859 /*
1860  * called from node manager when we should bring up our network listening
1861  * socket.  node manager handles all the serialization to only call this
1862  * once and to match it with o2net_stop_listening().  note,
1863  * o2nm_this_node() doesn't work yet as we're being called while it
1864  * is being set up.
1865  */
1866 int o2net_start_listening(struct o2nm_node *node)
1867 {
1868         int ret = 0;
1869
1870         BUG_ON(o2net_wq != NULL);
1871         BUG_ON(o2net_listen_sock != NULL);
1872
1873         mlog(ML_KTHREAD, "starting o2net thread...\n");
1874         o2net_wq = create_singlethread_workqueue("o2net");
1875         if (o2net_wq == NULL) {
1876                 mlog(ML_ERROR, "unable to launch o2net thread\n");
1877                 return -ENOMEM; /* ? */
1878         }
1879
1880         ret = o2net_open_listening_sock(node->nd_ipv4_address,
1881                                         node->nd_ipv4_port);
1882         if (ret) {
1883                 destroy_workqueue(o2net_wq);
1884                 o2net_wq = NULL;
1885         } else
1886                 o2quo_conn_up(node->nd_num);
1887
1888         return ret;
1889 }
1890
1891 /* again, o2nm_this_node() doesn't work here as we're involved in
1892  * tearing it down */
1893 void o2net_stop_listening(struct o2nm_node *node)
1894 {
1895         struct socket *sock = o2net_listen_sock;
1896         size_t i;
1897
1898         BUG_ON(o2net_wq == NULL);
1899         BUG_ON(o2net_listen_sock == NULL);
1900
1901         /* stop the listening socket from generating work */
1902         write_lock_bh(&sock->sk->sk_callback_lock);
1903         sock->sk->sk_data_ready = sock->sk->sk_user_data;
1904         sock->sk->sk_user_data = NULL;
1905         write_unlock_bh(&sock->sk->sk_callback_lock);
1906
1907         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1908                 struct o2nm_node *node = o2nm_get_node_by_num(i);
1909                 if (node) {
1910                         o2net_disconnect_node(node);
1911                         o2nm_node_put(node);
1912                 }
1913         }
1914
1915         /* finish all work and tear down the work queue */
1916         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
1917         destroy_workqueue(o2net_wq);
1918         o2net_wq = NULL;
1919
1920         sock_release(o2net_listen_sock);
1921         o2net_listen_sock = NULL;
1922
1923         o2quo_conn_err(node->nd_num);
1924 }
1925
1926 /* ------------------------------------------------------------ */
1927
1928 int o2net_init(void)
1929 {
1930         unsigned long i;
1931
1932         o2quo_init();
1933
1934         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
1935         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1936         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1937         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
1938                 kfree(o2net_hand);
1939                 kfree(o2net_keep_req);
1940                 kfree(o2net_keep_resp);
1941                 return -ENOMEM;
1942         }
1943
1944         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
1945         o2net_hand->connector_id = cpu_to_be64(1);
1946
1947         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
1948         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
1949
1950         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1951                 struct o2net_node *nn = o2net_nn_from_num(i);
1952
1953                 spin_lock_init(&nn->nn_lock);
1954                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
1955                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
1956                                   o2net_connect_expired);
1957                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
1958                 /* until we see hb from a node we'll return einval */
1959                 nn->nn_persistent_error = -ENOTCONN;
1960                 init_waitqueue_head(&nn->nn_sc_wq);
1961                 idr_init(&nn->nn_status_idr);
1962                 INIT_LIST_HEAD(&nn->nn_status_list);
1963         }
1964
1965         return 0;
1966 }
1967
1968 void o2net_exit(void)
1969 {
1970         o2quo_exit();
1971         kfree(o2net_hand);
1972         kfree(o2net_keep_req);
1973         kfree(o2net_keep_resp);
1974 }