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 %pI4:%u"
  76 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
  77                           &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 inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 157 {
 158         nst->st_sock_time = ktime_get();
 159 }
 160
 161 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 162 {
 163         nst->st_send_time = ktime_get();
 164 }
 165
 166 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 167 {
 168         nst->st_status_time = ktime_get();
 169 }
 170
 171 static inline 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 inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 178                                         u32 msg_id)
 179 {
 180         nst->st_id = msg_id;
 181 }
 182
 183 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
 184 {
 185         sc->sc_tv_timer = ktime_get();
 186 }
 187
 188 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
 189 {
 190         sc->sc_tv_data_ready = ktime_get();
 191 }
 192
 193 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
 194 {
 195         sc->sc_tv_advance_start = ktime_get();
 196 }
 197
 198 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
 199 {
 200         sc->sc_tv_advance_stop = ktime_get();
 201 }
 202
 203 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
 204 {
 205         sc->sc_tv_func_start = ktime_get();
 206 }
 207
 208 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
 209 {
 210         sc->sc_tv_func_stop = ktime_get();
 211 }
 212 #else  /* CONFIG_DEBUG_FS */
 213 # define o2net_init_nst(a, b, c, d, e)
 214 # define o2net_set_nst_sock_time(a)
 215 # define o2net_set_nst_send_time(a)
 216 # define o2net_set_nst_status_time(a)
 217 # define o2net_set_nst_sock_container(a, b)
 218 # define o2net_set_nst_msg_id(a, b)
 219 # define o2net_set_sock_timer(a)
 220 # define o2net_set_data_ready_time(a)
 221 # define o2net_set_advance_start_time(a)
 222 # define o2net_set_advance_stop_time(a)
 223 # define o2net_set_func_start_time(a)
 224 # define o2net_set_func_stop_time(a)
 225 #endif /* CONFIG_DEBUG_FS */
 226
 227 static inline int o2net_reconnect_delay(void)
 228 {
 229         return o2nm_single_cluster->cl_reconnect_delay_ms;
 230 }
 231
 232 static inline int o2net_keepalive_delay(void)
 233 {
 234         return o2nm_single_cluster->cl_keepalive_delay_ms;
 235 }
 236
 237 static inline int o2net_idle_timeout(void)
 238 {
 239         return o2nm_single_cluster->cl_idle_timeout_ms;
 240 }
 241
 242 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 243 {
 244         int trans;
 245         BUG_ON(err >= O2NET_ERR_MAX);
 246         trans = o2net_sys_err_translations[err];
 247
 248         /* Just in case we mess up the translation table above */
 249         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 250         return trans;
 251 }
 252
 253 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 254 {
 255         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 256         return &o2net_nodes[node_num];
 257 }
 258
 259 static u8 o2net_num_from_nn(struct o2net_node *nn)
 260 {
 261         BUG_ON(nn == NULL);
 262         return nn - o2net_nodes;
 263 }
 264
 265 /* ------------------------------------------------------------ */
 266
 267 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 268 {
 269         int ret = 0;
 270
 271         do {
 272                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
 273                         ret = -EAGAIN;
 274                         break;
 275                 }
 276                 spin_lock(&nn->nn_lock);
 277                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
 278                 if (ret == 0)
 279                         list_add_tail(&nsw->ns_node_item,
 280                                       &nn->nn_status_list);
 281                 spin_unlock(&nn->nn_lock);
 282         } while (ret == -EAGAIN);
 283
 284         if (ret == 0)  {
 285                 init_waitqueue_head(&nsw->ns_wq);
 286                 nsw->ns_sys_status = O2NET_ERR_NONE;
 287                 nsw->ns_status = 0;
 288         }
 289
 290         return ret;
 291 }
 292
 293 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 294                                       struct o2net_status_wait *nsw,
 295                                       enum o2net_system_error sys_status,
 296                                       s32 status)
 297 {
 298         assert_spin_locked(&nn->nn_lock);
 299
 300         if (!list_empty(&nsw->ns_node_item)) {
 301                 list_del_init(&nsw->ns_node_item);
 302                 nsw->ns_sys_status = sys_status;
 303                 nsw->ns_status = status;
 304                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 305                 wake_up(&nsw->ns_wq);
 306         }
 307 }
 308
 309 static void o2net_complete_nsw(struct o2net_node *nn,
 310                                struct o2net_status_wait *nsw,
 311                                u64 id, enum o2net_system_error sys_status,
 312                                s32 status)
 313 {
 314         spin_lock(&nn->nn_lock);
 315         if (nsw == NULL) {
 316                 if (id > INT_MAX)
 317                         goto out;
 318
 319                 nsw = idr_find(&nn->nn_status_idr, id);
 320                 if (nsw == NULL)
 321                         goto out;
 322         }
 323
 324         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 325
 326 out:
 327         spin_unlock(&nn->nn_lock);
 328         return;
 329 }
 330
 331 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 332 {
 333         struct o2net_status_wait *nsw, *tmp;
 334         unsigned int num_kills = 0;
 335
 336         assert_spin_locked(&nn->nn_lock);
 337
 338         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 339                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 340                 num_kills++;
 341         }
 342
 343         mlog(0, "completed %d messages for node %u\n", num_kills,
 344              o2net_num_from_nn(nn));
 345 }
 346
 347 static int o2net_nsw_completed(struct o2net_node *nn,
 348                                struct o2net_status_wait *nsw)
 349 {
 350         int completed;
 351         spin_lock(&nn->nn_lock);
 352         completed = list_empty(&nsw->ns_node_item);
 353         spin_unlock(&nn->nn_lock);
 354         return completed;
 355 }
 356
 357 /* ------------------------------------------------------------ */
 358
 359 static void sc_kref_release(struct kref *kref)
 360 {
 361         struct o2net_sock_container *sc = container_of(kref,
 362                                         struct o2net_sock_container, sc_kref);
 363         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 364
 365         sclog(sc, "releasing\n");
 366
 367         if (sc->sc_sock) {
 368                 sock_release(sc->sc_sock);
 369                 sc->sc_sock = NULL;
 370         }
 371
 372         o2nm_undepend_item(&sc->sc_node->nd_item);
 373         o2nm_node_put(sc->sc_node);
 374         sc->sc_node = NULL;
 375
 376         o2net_debug_del_sc(sc);
 377         kfree(sc);
 378 }
 379
 380 static void sc_put(struct o2net_sock_container *sc)
 381 {
 382         sclog(sc, "put\n");
 383         kref_put(&sc->sc_kref, sc_kref_release);
 384 }
 385 static void sc_get(struct o2net_sock_container *sc)
 386 {
 387         sclog(sc, "get\n");
 388         kref_get(&sc->sc_kref);
 389 }
 390 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 391 {
 392         struct o2net_sock_container *sc, *ret = NULL;
 393         struct page *page = NULL;
 394         int status = 0;
 395
 396         page = alloc_page(GFP_NOFS);
 397         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 398         if (sc == NULL || page == NULL)
 399                 goto out;
 400
 401         kref_init(&sc->sc_kref);
 402         o2nm_node_get(node);
 403         sc->sc_node = node;
 404
 405         /* pin the node item of the remote node */
 406         status = o2nm_depend_item(&node->nd_item);
 407         if (status) {
 408                 mlog_errno(status);
 409                 o2nm_node_put(node);
 410                 goto out;
 411         }
 412         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 413         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 414         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 415         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 416
 417         init_timer(&sc->sc_idle_timeout);
 418         sc->sc_idle_timeout.function = o2net_idle_timer;
 419         sc->sc_idle_timeout.data = (unsigned long)sc;
 420
 421         sclog(sc, "alloced\n");
 422
 423         ret = sc;
 424         sc->sc_page = page;
 425         o2net_debug_add_sc(sc);
 426         sc = NULL;
 427         page = NULL;
 428
 429 out:
 430         if (page)
 431                 __free_page(page);
 432         kfree(sc);
 433
 434         return ret;
 435 }
 436
 437 /* ------------------------------------------------------------ */
 438
 439 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 440                                 struct work_struct *work)
 441 {
 442         sc_get(sc);
 443         if (!queue_work(o2net_wq, work))
 444                 sc_put(sc);
 445 }
 446 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 447                                         struct delayed_work *work,
 448                                         int delay)
 449 {
 450         sc_get(sc);
 451         if (!queue_delayed_work(o2net_wq, work, delay))
 452                 sc_put(sc);
 453 }
 454 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 455                                          struct delayed_work *work)
 456 {
 457         if (cancel_delayed_work(work))
 458                 sc_put(sc);
 459 }
 460
 461 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 462
 463 int o2net_num_connected_peers(void)
 464 {
 465         return atomic_read(&o2net_connected_peers);
 466 }
 467
 468 static void o2net_set_nn_state(struct o2net_node *nn,
 469                                struct o2net_sock_container *sc,
 470                                unsigned valid, int err)
 471 {
 472         int was_valid = nn->nn_sc_valid;
 473         int was_err = nn->nn_persistent_error;
 474         struct o2net_sock_container *old_sc = nn->nn_sc;
 475
 476         assert_spin_locked(&nn->nn_lock);
 477
 478         if (old_sc && !sc)
 479                 atomic_dec(&o2net_connected_peers);
 480         else if (!old_sc && sc)
 481                 atomic_inc(&o2net_connected_peers);
 482
 483         /* the node num comparison and single connect/accept path should stop
 484          * an non-null sc from being overwritten with another */
 485         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 486         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 487         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 488
 489         if (was_valid && !valid && err == 0)
 490                 err = -ENOTCONN;
 491
 492         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 493              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 494              nn->nn_persistent_error, err);
 495
 496         nn->nn_sc = sc;
 497         nn->nn_sc_valid = valid ? 1 : 0;
 498         nn->nn_persistent_error = err;
 499
 500         /* mirrors o2net_tx_can_proceed() */
 501         if (nn->nn_persistent_error || nn->nn_sc_valid)
 502                 wake_up(&nn->nn_sc_wq);
 503
 504         if (!was_err && nn->nn_persistent_error) {
 505                 o2quo_conn_err(o2net_num_from_nn(nn));
 506                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 507                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 508         }
 509
 510         if (was_valid && !valid) {
 511                 printk(KERN_NOTICE "o2net: no longer connected to "
 512                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 513                 o2net_complete_nodes_nsw(nn);
 514         }
 515
 516         if (!was_valid && valid) {
 517                 o2quo_conn_up(o2net_num_from_nn(nn));
 518                 cancel_delayed_work(&nn->nn_connect_expired);
 519                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 520                        o2nm_this_node() > sc->sc_node->nd_num ?
 521                                 "connected to" : "accepted connection from",
 522                        SC_NODEF_ARGS(sc));
 523         }
 524
 525         /* trigger the connecting worker func as long as we're not valid,
 526          * it will back off if it shouldn't connect.  This can be called
 527          * from node config teardown and so needs to be careful about
 528          * the work queue actually being up. */
 529         if (!valid && o2net_wq) {
 530                 unsigned long delay;
 531                 /* delay if we're withing a RECONNECT_DELAY of the
 532                  * last attempt */
 533                 delay = (nn->nn_last_connect_attempt +
 534                          msecs_to_jiffies(o2net_reconnect_delay()))
 535                         - jiffies;
 536                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 537                         delay = 0;
 538                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 539                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 540
 541                 /*
 542                  * Delay the expired work after idle timeout.
 543                  *
 544                  * We might have lots of failed connection attempts that run
 545                  * through here but we only cancel the connect_expired work when
 546                  * a connection attempt succeeds.  So only the first enqueue of
 547                  * the connect_expired work will do anything.  The rest will see
 548                  * that it's already queued and do nothing.
 549                  */
 550                 delay += msecs_to_jiffies(o2net_idle_timeout());
 551                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 552         }
 553
 554         /* keep track of the nn's sc ref for the caller */
 555         if ((old_sc == NULL) && sc)
 556                 sc_get(sc);
 557         if (old_sc && (old_sc != sc)) {
 558                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 559                 sc_put(old_sc);
 560         }
 561 }
 562
 563 /* see o2net_register_callbacks() */
 564 static void o2net_data_ready(struct sock *sk, int bytes)
 565 {
 566         void (*ready)(struct sock *sk, int bytes);
 567
 568         read_lock(&sk->sk_callback_lock);
 569         if (sk->sk_user_data) {
 570                 struct o2net_sock_container *sc = sk->sk_user_data;
 571                 sclog(sc, "data_ready hit\n");
 572                 o2net_set_data_ready_time(sc);
 573                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 574                 ready = sc->sc_data_ready;
 575         } else {
 576                 ready = sk->sk_data_ready;
 577         }
 578         read_unlock(&sk->sk_callback_lock);
 579
 580         ready(sk, bytes);
 581 }
 582
 583 /* see o2net_register_callbacks() */
 584 static void o2net_state_change(struct sock *sk)
 585 {
 586         void (*state_change)(struct sock *sk);
 587         struct o2net_sock_container *sc;
 588
 589         read_lock(&sk->sk_callback_lock);
 590         sc = sk->sk_user_data;
 591         if (sc == NULL) {
 592                 state_change = sk->sk_state_change;
 593                 goto out;
 594         }
 595
 596         sclog(sc, "state_change to %d\n", sk->sk_state);
 597
 598         state_change = sc->sc_state_change;
 599
 600         switch(sk->sk_state) {
 601                 /* ignore connecting sockets as they make progress */
 602                 case TCP_SYN_SENT:
 603                 case TCP_SYN_RECV:
 604                         break;
 605                 case TCP_ESTABLISHED:
 606                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
 607                         break;
 608                 default:
 609                         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT
 610                               " shutdown, state %d\n",
 611                               SC_NODEF_ARGS(sc), sk->sk_state);
 612                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 613                         break;
 614         }
 615 out:
 616         read_unlock(&sk->sk_callback_lock);
 617         state_change(sk);
 618 }
 619
 620 /*
 621  * we register callbacks so we can queue work on events before calling
 622  * the original callbacks.  our callbacks our careful to test user_data
 623  * to discover when they've reaced with o2net_unregister_callbacks().
 624  */
 625 static void o2net_register_callbacks(struct sock *sk,
 626                                      struct o2net_sock_container *sc)
 627 {
 628         write_lock_bh(&sk->sk_callback_lock);
 629
 630         /* accepted sockets inherit the old listen socket data ready */
 631         if (sk->sk_data_ready == o2net_listen_data_ready) {
 632                 sk->sk_data_ready = sk->sk_user_data;
 633                 sk->sk_user_data = NULL;
 634         }
 635
 636         BUG_ON(sk->sk_user_data != NULL);
 637         sk->sk_user_data = sc;
 638         sc_get(sc);
 639
 640         sc->sc_data_ready = sk->sk_data_ready;
 641         sc->sc_state_change = sk->sk_state_change;
 642         sk->sk_data_ready = o2net_data_ready;
 643         sk->sk_state_change = o2net_state_change;
 644
 645         mutex_init(&sc->sc_send_lock);
 646
 647         write_unlock_bh(&sk->sk_callback_lock);
 648 }
 649
 650 static int o2net_unregister_callbacks(struct sock *sk,
 651                                    struct o2net_sock_container *sc)
 652 {
 653         int ret = 0;
 654
 655         write_lock_bh(&sk->sk_callback_lock);
 656         if (sk->sk_user_data == sc) {
 657                 ret = 1;
 658                 sk->sk_user_data = NULL;
 659                 sk->sk_data_ready = sc->sc_data_ready;
 660                 sk->sk_state_change = sc->sc_state_change;
 661         }
 662         write_unlock_bh(&sk->sk_callback_lock);
 663
 664         return ret;
 665 }
 666
 667 /*
 668  * this is a little helper that is called by callers who have seen a problem
 669  * with an sc and want to detach it from the nn if someone already hasn't beat
 670  * them to it.  if an error is given then the shutdown will be persistent
 671  * and pending transmits will be canceled.
 672  */
 673 static void o2net_ensure_shutdown(struct o2net_node *nn,
 674                                    struct o2net_sock_container *sc,
 675                                    int err)
 676 {
 677         spin_lock(&nn->nn_lock);
 678         if (nn->nn_sc == sc)
 679                 o2net_set_nn_state(nn, NULL, 0, err);
 680         spin_unlock(&nn->nn_lock);
 681 }
 682
 683 /*
 684  * This work queue function performs the blocking parts of socket shutdown.  A
 685  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 686  * sc detached from the nn.  state_change will also trigger this callback
 687  * directly when it sees errors.  In that case we need to call set_nn_state
 688  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 689  * itself.
 690  */
 691 static void o2net_shutdown_sc(struct work_struct *work)
 692 {
 693         struct o2net_sock_container *sc =
 694                 container_of(work, struct o2net_sock_container,
 695                              sc_shutdown_work);
 696         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 697
 698         sclog(sc, "shutting down\n");
 699
 700         /* drop the callbacks ref and call shutdown only once */
 701         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 702                 /* we shouldn't flush as we're in the thread, the
 703                  * races with pending sc work structs are harmless */
 704                 del_timer_sync(&sc->sc_idle_timeout);
 705                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 706                 sc_put(sc);
 707                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 708         }
 709
 710         /* not fatal so failed connects before the other guy has our
 711          * heartbeat can be retried */
 712         o2net_ensure_shutdown(nn, sc, 0);
 713         sc_put(sc);
 714 }
 715
 716 /* ------------------------------------------------------------ */
 717
 718 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 719                              u32 key)
 720 {
 721         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 722
 723         if (ret == 0)
 724                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 725
 726         return ret;
 727 }
 728
 729 static struct o2net_msg_handler *
 730 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 731                           struct rb_node **ret_parent)
 732 {
 733         struct rb_node **p = &o2net_handler_tree.rb_node;
 734         struct rb_node *parent = NULL;
 735         struct o2net_msg_handler *nmh, *ret = NULL;
 736         int cmp;
 737
 738         while (*p) {
 739                 parent = *p;
 740                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 741                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 742
 743                 if (cmp < 0)
 744                         p = &(*p)->rb_left;
 745                 else if (cmp > 0)
 746                         p = &(*p)->rb_right;
 747                 else {
 748                         ret = nmh;
 749                         break;
 750                 }
 751         }
 752
 753         if (ret_p != NULL)
 754                 *ret_p = p;
 755         if (ret_parent != NULL)
 756                 *ret_parent = parent;
 757
 758         return ret;
 759 }
 760
 761 static void o2net_handler_kref_release(struct kref *kref)
 762 {
 763         struct o2net_msg_handler *nmh;
 764         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 765
 766         kfree(nmh);
 767 }
 768
 769 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 770 {
 771         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 772 }
 773
 774 /* max_len is protection for the handler func.  incoming messages won't
 775  * be given to the handler if their payload is longer than the max. */
 776 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 777                            o2net_msg_handler_func *func, void *data,
 778                            o2net_post_msg_handler_func *post_func,
 779                            struct list_head *unreg_list)
 780 {
 781         struct o2net_msg_handler *nmh = NULL;
 782         struct rb_node **p, *parent;
 783         int ret = 0;
 784
 785         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 786                 mlog(0, "max_len for message handler out of range: %u\n",
 787                         max_len);
 788                 ret = -EINVAL;
 789                 goto out;
 790         }
 791
 792         if (!msg_type) {
 793                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 794                 ret = -EINVAL;
 795                 goto out;
 796
 797         }
 798         if (!func) {
 799                 mlog(0, "no message handler provided: %u, %p\n",
 800                        msg_type, func);
 801                 ret = -EINVAL;
 802                 goto out;
 803         }
 804
 805         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 806         if (nmh == NULL) {
 807                 ret = -ENOMEM;
 808                 goto out;
 809         }
 810
 811         nmh->nh_func = func;
 812         nmh->nh_func_data = data;
 813         nmh->nh_post_func = post_func;
 814         nmh->nh_msg_type = msg_type;
 815         nmh->nh_max_len = max_len;
 816         nmh->nh_key = key;
 817         /* the tree and list get this ref.. they're both removed in
 818          * unregister when this ref is dropped */
 819         kref_init(&nmh->nh_kref);
 820         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 821
 822         write_lock(&o2net_handler_lock);
 823         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 824                 ret = -EEXIST;
 825         else {
 826                 rb_link_node(&nmh->nh_node, parent, p);
 827                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 828                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 829
 830                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 831                      func, msg_type, key);
 832                 /* we've had some trouble with handlers seemingly vanishing. */
 833                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 834                                                           &parent) == NULL,
 835                                 "couldn't find handler we *just* registerd "
 836                                 "for type %u key %08x\n", msg_type, key);
 837         }
 838         write_unlock(&o2net_handler_lock);
 839         if (ret)
 840                 goto out;
 841
 842 out:
 843         if (ret)
 844                 kfree(nmh);
 845
 846         return ret;
 847 }
 848 EXPORT_SYMBOL_GPL(o2net_register_handler);
 849
 850 void o2net_unregister_handler_list(struct list_head *list)
 851 {
 852         struct o2net_msg_handler *nmh, *n;
 853
 854         write_lock(&o2net_handler_lock);
 855         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 856                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 857                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 858                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 859                 list_del_init(&nmh->nh_unregister_item);
 860                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 861         }
 862         write_unlock(&o2net_handler_lock);
 863 }
 864 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 865
 866 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 867 {
 868         struct o2net_msg_handler *nmh;
 869
 870         read_lock(&o2net_handler_lock);
 871         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 872         if (nmh)
 873                 kref_get(&nmh->nh_kref);
 874         read_unlock(&o2net_handler_lock);
 875
 876         return nmh;
 877 }
 878
 879 /* ------------------------------------------------------------ */
 880
 881 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 882 {
 883         int ret;
 884         mm_segment_t oldfs;
 885         struct kvec vec = {
 886                 .iov_len = len,
 887                 .iov_base = data,
 888         };
 889         struct msghdr msg = {
 890                 .msg_iovlen = 1,
 891                 .msg_iov = (struct iovec *)&vec,
 892                 .msg_flags = MSG_DONTWAIT,
 893         };
 894
 895         oldfs = get_fs();
 896         set_fs(get_ds());
 897         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
 898         set_fs(oldfs);
 899
 900         return ret;
 901 }
 902
 903 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 904                               size_t veclen, size_t total)
 905 {
 906         int ret;
 907         mm_segment_t oldfs;
 908         struct msghdr msg = {
 909                 .msg_iov = (struct iovec *)vec,
 910                 .msg_iovlen = veclen,
 911         };
 912
 913         if (sock == NULL) {
 914                 ret = -EINVAL;
 915                 goto out;
 916         }
 917
 918         oldfs = get_fs();
 919         set_fs(get_ds());
 920         ret = sock_sendmsg(sock, &msg, total);
 921         set_fs(oldfs);
 922         if (ret != total) {
 923                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
 924                      total);
 925                 if (ret >= 0)
 926                         ret = -EPIPE; /* should be smarter, I bet */
 927                 goto out;
 928         }
 929
 930         ret = 0;
 931 out:
 932         if (ret < 0)
 933                 mlog(0, "returning error: %d\n", ret);
 934         return ret;
 935 }
 936
 937 static void o2net_sendpage(struct o2net_sock_container *sc,
 938                            void *kmalloced_virt,
 939                            size_t size)
 940 {
 941         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 942         ssize_t ret;
 943
 944         while (1) {
 945                 mutex_lock(&sc->sc_send_lock);
 946                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 947                                                  virt_to_page(kmalloced_virt),
 948                                                  (long)kmalloced_virt & ~PAGE_MASK,
 949                                                  size, MSG_DONTWAIT);
 950                 mutex_unlock(&sc->sc_send_lock);
 951                 if (ret == size)
 952                         break;
 953                 if (ret == (ssize_t)-EAGAIN) {
 954                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 955                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 956                         cond_resched();
 957                         continue;
 958                 }
 959                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 960                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 961                 o2net_ensure_shutdown(nn, sc, 0);
 962                 break;
 963         }
 964 }
 965
 966 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 967 {
 968         memset(msg, 0, sizeof(struct o2net_msg));
 969         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 970         msg->data_len = cpu_to_be16(data_len);
 971         msg->msg_type = cpu_to_be16(msg_type);
 972         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 973         msg->status = 0;
 974         msg->key = cpu_to_be32(key);
 975 }
 976
 977 static int o2net_tx_can_proceed(struct o2net_node *nn,
 978                                 struct o2net_sock_container **sc_ret,
 979                                 int *error)
 980 {
 981         int ret = 0;
 982
 983         spin_lock(&nn->nn_lock);
 984         if (nn->nn_persistent_error) {
 985                 ret = 1;
 986                 *sc_ret = NULL;
 987                 *error = nn->nn_persistent_error;
 988         } else if (nn->nn_sc_valid) {
 989                 kref_get(&nn->nn_sc->sc_kref);
 990
 991                 ret = 1;
 992                 *sc_ret = nn->nn_sc;
 993                 *error = 0;
 994         }
 995         spin_unlock(&nn->nn_lock);
 996
 997         return ret;
 998 }
 999
1000 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1001                            size_t caller_veclen, u8 target_node, int *status)
1002 {
1003         int ret = 0;
1004         struct o2net_msg *msg = NULL;
1005         size_t veclen, caller_bytes = 0;
1006         struct kvec *vec = NULL;
1007         struct o2net_sock_container *sc = NULL;
1008         struct o2net_node *nn = o2net_nn_from_num(target_node);
1009         struct o2net_status_wait nsw = {
1010                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1011         };
1012         struct o2net_send_tracking nst;
1013
1014         o2net_init_nst(&nst, msg_type, key, current, target_node);
1015
1016         if (o2net_wq == NULL) {
1017                 mlog(0, "attempt to tx without o2netd running\n");
1018                 ret = -ESRCH;
1019                 goto out;
1020         }
1021
1022         if (caller_veclen == 0) {
1023                 mlog(0, "bad kvec array length\n");
1024                 ret = -EINVAL;
1025                 goto out;
1026         }
1027
1028         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1029         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1030                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1031                 ret = -EINVAL;
1032                 goto out;
1033         }
1034
1035         if (target_node == o2nm_this_node()) {
1036                 ret = -ELOOP;
1037                 goto out;
1038         }
1039
1040         o2net_debug_add_nst(&nst);
1041
1042         o2net_set_nst_sock_time(&nst);
1043
1044         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1045         if (ret)
1046                 goto out;
1047
1048         o2net_set_nst_sock_container(&nst, sc);
1049
1050         veclen = caller_veclen + 1;
1051         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1052         if (vec == NULL) {
1053                 mlog(0, "failed to %zu element kvec!\n", veclen);
1054                 ret = -ENOMEM;
1055                 goto out;
1056         }
1057
1058         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1059         if (!msg) {
1060                 mlog(0, "failed to allocate a o2net_msg!\n");
1061                 ret = -ENOMEM;
1062                 goto out;
1063         }
1064
1065         o2net_init_msg(msg, caller_bytes, msg_type, key);
1066
1067         vec[0].iov_len = sizeof(struct o2net_msg);
1068         vec[0].iov_base = msg;
1069         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1070
1071         ret = o2net_prep_nsw(nn, &nsw);
1072         if (ret)
1073                 goto out;
1074
1075         msg->msg_num = cpu_to_be32(nsw.ns_id);
1076         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1077
1078         o2net_set_nst_send_time(&nst);
1079
1080         /* finally, convert the message header to network byte-order
1081          * and send */
1082         mutex_lock(&sc->sc_send_lock);
1083         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1084                                  sizeof(struct o2net_msg) + caller_bytes);
1085         mutex_unlock(&sc->sc_send_lock);
1086         msglog(msg, "sending returned %d\n", ret);
1087         if (ret < 0) {
1088                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1089                 goto out;
1090         }
1091
1092         /* wait on other node's handler */
1093         o2net_set_nst_status_time(&nst);
1094         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1095
1096         /* Note that we avoid overwriting the callers status return
1097          * variable if a system error was reported on the other
1098          * side. Callers beware. */
1099         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1100         if (status && !ret)
1101                 *status = nsw.ns_status;
1102
1103         mlog(0, "woken, returning system status %d, user status %d\n",
1104              ret, nsw.ns_status);
1105 out:
1106         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1107         if (sc)
1108                 sc_put(sc);
1109         if (vec)
1110                 kfree(vec);
1111         if (msg)
1112                 kfree(msg);
1113         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1114         return ret;
1115 }
1116 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1117
1118 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1119                        u8 target_node, int *status)
1120 {
1121         struct kvec vec = {
1122                 .iov_base = data,
1123                 .iov_len = len,
1124         };
1125         return o2net_send_message_vec(msg_type, key, &vec, 1,
1126                                       target_node, status);
1127 }
1128 EXPORT_SYMBOL_GPL(o2net_send_message);
1129
1130 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1131                                    enum o2net_system_error syserr, int err)
1132 {
1133         struct kvec vec = {
1134                 .iov_base = hdr,
1135                 .iov_len = sizeof(struct o2net_msg),
1136         };
1137
1138         BUG_ON(syserr >= O2NET_ERR_MAX);
1139
1140         /* leave other fields intact from the incoming message, msg_num
1141          * in particular */
1142         hdr->sys_status = cpu_to_be32(syserr);
1143         hdr->status = cpu_to_be32(err);
1144         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1145         hdr->data_len = 0;
1146
1147         msglog(hdr, "about to send status magic %d\n", err);
1148         /* hdr has been in host byteorder this whole time */
1149         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1150 }
1151
1152 /* this returns -errno if the header was unknown or too large, etc.
1153  * after this is called the buffer us reused for the next message */
1154 static int o2net_process_message(struct o2net_sock_container *sc,
1155                                  struct o2net_msg *hdr)
1156 {
1157         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1158         int ret = 0, handler_status;
1159         enum  o2net_system_error syserr;
1160         struct o2net_msg_handler *nmh = NULL;
1161         void *ret_data = NULL;
1162
1163         msglog(hdr, "processing message\n");
1164
1165         o2net_sc_postpone_idle(sc);
1166
1167         switch(be16_to_cpu(hdr->magic)) {
1168                 case O2NET_MSG_STATUS_MAGIC:
1169                         /* special type for returning message status */
1170                         o2net_complete_nsw(nn, NULL,
1171                                            be32_to_cpu(hdr->msg_num),
1172                                            be32_to_cpu(hdr->sys_status),
1173                                            be32_to_cpu(hdr->status));
1174                         goto out;
1175                 case O2NET_MSG_KEEP_REQ_MAGIC:
1176                         o2net_sendpage(sc, o2net_keep_resp,
1177                                        sizeof(*o2net_keep_resp));
1178                         goto out;
1179                 case O2NET_MSG_KEEP_RESP_MAGIC:
1180                         goto out;
1181                 case O2NET_MSG_MAGIC:
1182                         break;
1183                 default:
1184                         msglog(hdr, "bad magic\n");
1185                         ret = -EINVAL;
1186                         goto out;
1187                         break;
1188         }
1189
1190         /* find a handler for it */
1191         handler_status = 0;
1192         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1193                                 be32_to_cpu(hdr->key));
1194         if (!nmh) {
1195                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1196                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1197                 syserr = O2NET_ERR_NO_HNDLR;
1198                 goto out_respond;
1199         }
1200
1201         syserr = O2NET_ERR_NONE;
1202
1203         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1204                 syserr = O2NET_ERR_OVERFLOW;
1205
1206         if (syserr != O2NET_ERR_NONE)
1207                 goto out_respond;
1208
1209         o2net_set_func_start_time(sc);
1210         sc->sc_msg_key = be32_to_cpu(hdr->key);
1211         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1212         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1213                                              be16_to_cpu(hdr->data_len),
1214                                         nmh->nh_func_data, &ret_data);
1215         o2net_set_func_stop_time(sc);
1216
1217 out_respond:
1218         /* this destroys the hdr, so don't use it after this */
1219         mutex_lock(&sc->sc_send_lock);
1220         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1221                                       handler_status);
1222         mutex_unlock(&sc->sc_send_lock);
1223         hdr = NULL;
1224         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1225              handler_status, syserr, ret);
1226
1227         if (nmh) {
1228                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1229                 if (nmh->nh_post_func)
1230                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1231                                             ret_data);
1232         }
1233
1234 out:
1235         if (nmh)
1236                 o2net_handler_put(nmh);
1237         return ret;
1238 }
1239
1240 static int o2net_check_handshake(struct o2net_sock_container *sc)
1241 {
1242         struct o2net_handshake *hand = page_address(sc->sc_page);
1243         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1244
1245         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1246                 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1247                      "version %llu but %llu is required, disconnecting\n",
1248                      SC_NODEF_ARGS(sc),
1249                      (unsigned long long)be64_to_cpu(hand->protocol_version),
1250                      O2NET_PROTOCOL_VERSION);
1251
1252                 /* don't bother reconnecting if its the wrong version. */
1253                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1254                 return -1;
1255         }
1256
1257         /*
1258          * Ensure timeouts are consistent with other nodes, otherwise
1259          * we can end up with one node thinking that the other must be down,
1260          * but isn't. This can ultimately cause corruption.
1261          */
1262         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1263                                 o2net_idle_timeout()) {
1264                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1265                      "%u ms, but we use %u ms locally.  disconnecting\n",
1266                      SC_NODEF_ARGS(sc),
1267                      be32_to_cpu(hand->o2net_idle_timeout_ms),
1268                      o2net_idle_timeout());
1269                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1270                 return -1;
1271         }
1272
1273         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1274                         o2net_keepalive_delay()) {
1275                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1276                      "%u ms, but we use %u ms locally.  disconnecting\n",
1277                      SC_NODEF_ARGS(sc),
1278                      be32_to_cpu(hand->o2net_keepalive_delay_ms),
1279                      o2net_keepalive_delay());
1280                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1281                 return -1;
1282         }
1283
1284         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1285                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1286                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1287                      "%u ms, but we use %u ms locally.  disconnecting\n",
1288                      SC_NODEF_ARGS(sc),
1289                      be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1290                      O2HB_MAX_WRITE_TIMEOUT_MS);
1291                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1292                 return -1;
1293         }
1294
1295         sc->sc_handshake_ok = 1;
1296
1297         spin_lock(&nn->nn_lock);
1298         /* set valid and queue the idle timers only if it hasn't been
1299          * shut down already */
1300         if (nn->nn_sc == sc) {
1301                 o2net_sc_reset_idle_timer(sc);
1302                 atomic_set(&nn->nn_timeout, 0);
1303                 o2net_set_nn_state(nn, sc, 1, 0);
1304         }
1305         spin_unlock(&nn->nn_lock);
1306
1307         /* shift everything up as though it wasn't there */
1308         sc->sc_page_off -= sizeof(struct o2net_handshake);
1309         if (sc->sc_page_off)
1310                 memmove(hand, hand + 1, sc->sc_page_off);
1311
1312         return 0;
1313 }
1314
1315 /* this demuxes the queued rx bytes into header or payload bits and calls
1316  * handlers as each full message is read off the socket.  it returns -error,
1317  * == 0 eof, or > 0 for progress made.*/
1318 static int o2net_advance_rx(struct o2net_sock_container *sc)
1319 {
1320         struct o2net_msg *hdr;
1321         int ret = 0;
1322         void *data;
1323         size_t datalen;
1324
1325         sclog(sc, "receiving\n");
1326         o2net_set_advance_start_time(sc);
1327
1328         if (unlikely(sc->sc_handshake_ok == 0)) {
1329                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1330                         data = page_address(sc->sc_page) + sc->sc_page_off;
1331                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1332                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1333                         if (ret > 0)
1334                                 sc->sc_page_off += ret;
1335                 }
1336
1337                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1338                         o2net_check_handshake(sc);
1339                         if (unlikely(sc->sc_handshake_ok == 0))
1340                                 ret = -EPROTO;
1341                 }
1342                 goto out;
1343         }
1344
1345         /* do we need more header? */
1346         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1347                 data = page_address(sc->sc_page) + sc->sc_page_off;
1348                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1349                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1350                 if (ret > 0) {
1351                         sc->sc_page_off += ret;
1352                         /* only swab incoming here.. we can
1353                          * only get here once as we cross from
1354                          * being under to over */
1355                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1356                                 hdr = page_address(sc->sc_page);
1357                                 if (be16_to_cpu(hdr->data_len) >
1358                                     O2NET_MAX_PAYLOAD_BYTES)
1359                                         ret = -EOVERFLOW;
1360                         }
1361                 }
1362                 if (ret <= 0)
1363                         goto out;
1364         }
1365
1366         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1367                 /* oof, still don't have a header */
1368                 goto out;
1369         }
1370
1371         /* this was swabbed above when we first read it */
1372         hdr = page_address(sc->sc_page);
1373
1374         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1375
1376         /* do we need more payload? */
1377         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1378                 /* need more payload */
1379                 data = page_address(sc->sc_page) + sc->sc_page_off;
1380                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1381                           sc->sc_page_off;
1382                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1383                 if (ret > 0)
1384                         sc->sc_page_off += ret;
1385                 if (ret <= 0)
1386                         goto out;
1387         }
1388
1389         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1390                 /* we can only get here once, the first time we read
1391                  * the payload.. so set ret to progress if the handler
1392                  * works out. after calling this the message is toast */
1393                 ret = o2net_process_message(sc, hdr);
1394                 if (ret == 0)
1395                         ret = 1;
1396                 sc->sc_page_off = 0;
1397         }
1398
1399 out:
1400         sclog(sc, "ret = %d\n", ret);
1401         o2net_set_advance_stop_time(sc);
1402         return ret;
1403 }
1404
1405 /* this work func is triggerd by data ready.  it reads until it can read no
1406  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1407  * our work the work struct will be marked and we'll be called again. */
1408 static void o2net_rx_until_empty(struct work_struct *work)
1409 {
1410         struct o2net_sock_container *sc =
1411                 container_of(work, struct o2net_sock_container, sc_rx_work);
1412         int ret;
1413
1414         do {
1415                 ret = o2net_advance_rx(sc);
1416         } while (ret > 0);
1417
1418         if (ret <= 0 && ret != -EAGAIN) {
1419                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1420                 sclog(sc, "saw error %d, closing\n", ret);
1421                 /* not permanent so read failed handshake can retry */
1422                 o2net_ensure_shutdown(nn, sc, 0);
1423         }
1424
1425         sc_put(sc);
1426 }
1427
1428 static int o2net_set_nodelay(struct socket *sock)
1429 {
1430         int ret, val = 1;
1431         mm_segment_t oldfs;
1432
1433         oldfs = get_fs();
1434         set_fs(KERNEL_DS);
1435
1436         /*
1437          * Dear unsuspecting programmer,
1438          *
1439          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1440          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1441          * silently turn into SO_DEBUG.
1442          *
1443          * Yours,
1444          * Keeper of hilariously fragile interfaces.
1445          */
1446         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1447                                     (char __user *)&val, sizeof(val));
1448
1449         set_fs(oldfs);
1450         return ret;
1451 }
1452
1453 static void o2net_initialize_handshake(void)
1454 {
1455         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1456                 O2HB_MAX_WRITE_TIMEOUT_MS);
1457         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1458         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1459                 o2net_keepalive_delay());
1460         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1461                 o2net_reconnect_delay());
1462 }
1463
1464 /* ------------------------------------------------------------ */
1465
1466 /* called when a connect completes and after a sock is accepted.  the
1467  * rx path will see the response and mark the sc valid */
1468 static void o2net_sc_connect_completed(struct work_struct *work)
1469 {
1470         struct o2net_sock_container *sc =
1471                 container_of(work, struct o2net_sock_container,
1472                              sc_connect_work);
1473
1474         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1475               (unsigned long long)O2NET_PROTOCOL_VERSION,
1476               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1477
1478         o2net_initialize_handshake();
1479         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1480         sc_put(sc);
1481 }
1482
1483 /* this is called as a work_struct func. */
1484 static void o2net_sc_send_keep_req(struct work_struct *work)
1485 {
1486         struct o2net_sock_container *sc =
1487                 container_of(work, struct o2net_sock_container,
1488                              sc_keepalive_work.work);
1489
1490         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1491         sc_put(sc);
1492 }
1493
1494 /* socket shutdown does a del_timer_sync against this as it tears down.
1495  * we can't start this timer until we've got to the point in sc buildup
1496  * where shutdown is going to be involved */
1497 static void o2net_idle_timer(unsigned long data)
1498 {
1499         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1500         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1501
1502 #ifdef CONFIG_DEBUG_FS
1503         ktime_t now = ktime_get();
1504 #endif
1505
1506         printk(KERN_NOTICE "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1507              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1508                      o2net_idle_timeout() / 1000,
1509                      o2net_idle_timeout() % 1000);
1510
1511 #ifdef CONFIG_DEBUG_FS
1512         mlog(ML_NOTICE, "Here are some times that might help debug the "
1513              "situation: (Timer: %lld, Now %lld, DataReady %lld, Advance %lld-%lld, "
1514              "Key 0x%08x, Func %u, FuncTime %lld-%lld)\n",
1515              (long long)ktime_to_us(sc->sc_tv_timer), (long long)ktime_to_us(now),
1516              (long long)ktime_to_us(sc->sc_tv_data_ready),
1517              (long long)ktime_to_us(sc->sc_tv_advance_start),
1518              (long long)ktime_to_us(sc->sc_tv_advance_stop),
1519              sc->sc_msg_key, sc->sc_msg_type,
1520              (long long)ktime_to_us(sc->sc_tv_func_start),
1521              (long long)ktime_to_us(sc->sc_tv_func_stop));
1522 #endif
1523
1524         /*
1525          * Initialize the nn_timeout so that the next connection attempt
1526          * will continue in o2net_start_connect.
1527          */
1528         atomic_set(&nn->nn_timeout, 1);
1529
1530         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1531 }
1532
1533 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1534 {
1535         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1536         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1537                       msecs_to_jiffies(o2net_keepalive_delay()));
1538         o2net_set_sock_timer(sc);
1539         mod_timer(&sc->sc_idle_timeout,
1540                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1541 }
1542
1543 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1544 {
1545         /* Only push out an existing timer */
1546         if (timer_pending(&sc->sc_idle_timeout))
1547                 o2net_sc_reset_idle_timer(sc);
1548 }
1549
1550 /* this work func is kicked whenever a path sets the nn state which doesn't
1551  * have valid set.  This includes seeing hb come up, losing a connection,
1552  * having a connect attempt fail, etc. This centralizes the logic which decides
1553  * if a connect attempt should be made or if we should give up and all future
1554  * transmit attempts should fail */
1555 static void o2net_start_connect(struct work_struct *work)
1556 {
1557         struct o2net_node *nn =
1558                 container_of(work, struct o2net_node, nn_connect_work.work);
1559         struct o2net_sock_container *sc = NULL;
1560         struct o2nm_node *node = NULL, *mynode = NULL;
1561         struct socket *sock = NULL;
1562         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1563         int ret = 0, stop;
1564         unsigned int timeout;
1565
1566         /* if we're greater we initiate tx, otherwise we accept */
1567         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1568                 goto out;
1569
1570         /* watch for racing with tearing a node down */
1571         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1572         if (node == NULL) {
1573                 ret = 0;
1574                 goto out;
1575         }
1576
1577         mynode = o2nm_get_node_by_num(o2nm_this_node());
1578         if (mynode == NULL) {
1579                 ret = 0;
1580                 goto out;
1581         }
1582
1583         spin_lock(&nn->nn_lock);
1584         /*
1585          * see if we already have one pending or have given up.
1586          * For nn_timeout, it is set when we close the connection
1587          * because of the idle time out. So it means that we have
1588          * at least connected to that node successfully once,
1589          * now try to connect to it again.
1590          */
1591         timeout = atomic_read(&nn->nn_timeout);
1592         stop = (nn->nn_sc ||
1593                 (nn->nn_persistent_error &&
1594                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1595         spin_unlock(&nn->nn_lock);
1596         if (stop)
1597                 goto out;
1598
1599         nn->nn_last_connect_attempt = jiffies;
1600
1601         sc = sc_alloc(node);
1602         if (sc == NULL) {
1603                 mlog(0, "couldn't allocate sc\n");
1604                 ret = -ENOMEM;
1605                 goto out;
1606         }
1607
1608         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1609         if (ret < 0) {
1610                 mlog(0, "can't create socket: %d\n", ret);
1611                 goto out;
1612         }
1613         sc->sc_sock = sock; /* freed by sc_kref_release */
1614
1615         sock->sk->sk_allocation = GFP_ATOMIC;
1616
1617         myaddr.sin_family = AF_INET;
1618         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1619         myaddr.sin_port = htons(0); /* any port */
1620
1621         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1622                               sizeof(myaddr));
1623         if (ret) {
1624                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1625                      ret, &mynode->nd_ipv4_address);
1626                 goto out;
1627         }
1628
1629         ret = o2net_set_nodelay(sc->sc_sock);
1630         if (ret) {
1631                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1632                 goto out;
1633         }
1634
1635         o2net_register_callbacks(sc->sc_sock->sk, sc);
1636
1637         spin_lock(&nn->nn_lock);
1638         /* handshake completion will set nn->nn_sc_valid */
1639         o2net_set_nn_state(nn, sc, 0, 0);
1640         spin_unlock(&nn->nn_lock);
1641
1642         remoteaddr.sin_family = AF_INET;
1643         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1644         remoteaddr.sin_port = node->nd_ipv4_port;
1645
1646         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1647                                         (struct sockaddr *)&remoteaddr,
1648                                         sizeof(remoteaddr),
1649                                         O_NONBLOCK);
1650         if (ret == -EINPROGRESS)
1651                 ret = 0;
1652
1653 out:
1654         if (ret) {
1655                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1656                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1657                 /* 0 err so that another will be queued and attempted
1658                  * from set_nn_state */
1659                 if (sc)
1660                         o2net_ensure_shutdown(nn, sc, 0);
1661         }
1662         if (sc)
1663                 sc_put(sc);
1664         if (node)
1665                 o2nm_node_put(node);
1666         if (mynode)
1667                 o2nm_node_put(mynode);
1668
1669         return;
1670 }
1671
1672 static void o2net_connect_expired(struct work_struct *work)
1673 {
1674         struct o2net_node *nn =
1675                 container_of(work, struct o2net_node, nn_connect_expired.work);
1676
1677         spin_lock(&nn->nn_lock);
1678         if (!nn->nn_sc_valid) {
1679                 mlog(ML_ERROR, "no connection established with node %u after "
1680                      "%u.%u seconds, giving up and returning errors.\n",
1681                      o2net_num_from_nn(nn),
1682                      o2net_idle_timeout() / 1000,
1683                      o2net_idle_timeout() % 1000);
1684
1685                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1686         }
1687         spin_unlock(&nn->nn_lock);
1688 }
1689
1690 static void o2net_still_up(struct work_struct *work)
1691 {
1692         struct o2net_node *nn =
1693                 container_of(work, struct o2net_node, nn_still_up.work);
1694
1695         o2quo_hb_still_up(o2net_num_from_nn(nn));
1696 }
1697
1698 /* ------------------------------------------------------------ */
1699
1700 void o2net_disconnect_node(struct o2nm_node *node)
1701 {
1702         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1703
1704         /* don't reconnect until it's heartbeating again */
1705         spin_lock(&nn->nn_lock);
1706         atomic_set(&nn->nn_timeout, 0);
1707         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1708         spin_unlock(&nn->nn_lock);
1709
1710         if (o2net_wq) {
1711                 cancel_delayed_work(&nn->nn_connect_expired);
1712                 cancel_delayed_work(&nn->nn_connect_work);
1713                 cancel_delayed_work(&nn->nn_still_up);
1714                 flush_workqueue(o2net_wq);
1715         }
1716 }
1717
1718 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1719                                   void *data)
1720 {
1721         o2quo_hb_down(node_num);
1722
1723         if (!node)
1724                 return;
1725
1726         if (node_num != o2nm_this_node())
1727                 o2net_disconnect_node(node);
1728
1729         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1730 }
1731
1732 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1733                                 void *data)
1734 {
1735         struct o2net_node *nn = o2net_nn_from_num(node_num);
1736
1737         o2quo_hb_up(node_num);
1738
1739         BUG_ON(!node);
1740
1741         /* ensure an immediate connect attempt */
1742         nn->nn_last_connect_attempt = jiffies -
1743                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1744
1745         if (node_num != o2nm_this_node()) {
1746                 /* believe it or not, accept and node hearbeating testing
1747                  * can succeed for this node before we got here.. so
1748                  * only use set_nn_state to clear the persistent error
1749                  * if that hasn't already happened */
1750                 spin_lock(&nn->nn_lock);
1751                 atomic_set(&nn->nn_timeout, 0);
1752                 if (nn->nn_persistent_error)
1753                         o2net_set_nn_state(nn, NULL, 0, 0);
1754                 spin_unlock(&nn->nn_lock);
1755         }
1756 }
1757
1758 void o2net_unregister_hb_callbacks(void)
1759 {
1760         o2hb_unregister_callback(NULL, &o2net_hb_up);
1761         o2hb_unregister_callback(NULL, &o2net_hb_down);
1762 }
1763
1764 int o2net_register_hb_callbacks(void)
1765 {
1766         int ret;
1767
1768         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1769                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1770         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1771                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1772
1773         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1774         if (ret == 0)
1775                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1776
1777         if (ret)
1778                 o2net_unregister_hb_callbacks();
1779
1780         return ret;
1781 }
1782
1783 /* ------------------------------------------------------------ */
1784
1785 static int o2net_accept_one(struct socket *sock)
1786 {
1787         int ret, slen;
1788         struct sockaddr_in sin;
1789         struct socket *new_sock = NULL;
1790         struct o2nm_node *node = NULL;
1791         struct o2nm_node *local_node = NULL;
1792         struct o2net_sock_container *sc = NULL;
1793         struct o2net_node *nn;
1794
1795         BUG_ON(sock == NULL);
1796         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1797                                sock->sk->sk_protocol, &new_sock);
1798         if (ret)
1799                 goto out;
1800
1801         new_sock->type = sock->type;
1802         new_sock->ops = sock->ops;
1803         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1804         if (ret < 0)
1805                 goto out;
1806
1807         new_sock->sk->sk_allocation = GFP_ATOMIC;
1808
1809         ret = o2net_set_nodelay(new_sock);
1810         if (ret) {
1811                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1812                 goto out;
1813         }
1814
1815         slen = sizeof(sin);
1816         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1817                                        &slen, 1);
1818         if (ret < 0)
1819                 goto out;
1820
1821         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1822         if (node == NULL) {
1823                 mlog(ML_NOTICE, "attempt to connect from unknown node at %pI4:%d\n",
1824                      &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1825                 ret = -EINVAL;
1826                 goto out;
1827         }
1828
1829         if (o2nm_this_node() >= node->nd_num) {
1830                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1831                 mlog(ML_NOTICE, "unexpected connect attempt seen at node '%s' ("
1832                      "%u, %pI4:%d) from node '%s' (%u, %pI4:%d)\n",
1833                      local_node->nd_name, local_node->nd_num,
1834                      &(local_node->nd_ipv4_address),
1835                      ntohs(local_node->nd_ipv4_port),
1836                      node->nd_name, node->nd_num, &sin.sin_addr.s_addr,
1837                      ntohs(sin.sin_port));
1838                 ret = -EINVAL;
1839                 goto out;
1840         }
1841
1842         /* this happens all the time when the other node sees our heartbeat
1843          * and tries to connect before we see their heartbeat */
1844         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1845                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1846                      "%pI4:%d but it isn't heartbeating\n",
1847                      node->nd_name, &sin.sin_addr.s_addr,
1848                      ntohs(sin.sin_port));
1849                 ret = -EINVAL;
1850                 goto out;
1851         }
1852
1853         nn = o2net_nn_from_num(node->nd_num);
1854
1855         spin_lock(&nn->nn_lock);
1856         if (nn->nn_sc)
1857                 ret = -EBUSY;
1858         else
1859                 ret = 0;
1860         spin_unlock(&nn->nn_lock);
1861         if (ret) {
1862                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1863                      "%pI4:%d but it already has an open connection\n",
1864                      node->nd_name, &sin.sin_addr.s_addr,
1865                      ntohs(sin.sin_port));
1866                 goto out;
1867         }
1868
1869         sc = sc_alloc(node);
1870         if (sc == NULL) {
1871                 ret = -ENOMEM;
1872                 goto out;
1873         }
1874
1875         sc->sc_sock = new_sock;
1876         new_sock = NULL;
1877
1878         spin_lock(&nn->nn_lock);
1879         atomic_set(&nn->nn_timeout, 0);
1880         o2net_set_nn_state(nn, sc, 0, 0);
1881         spin_unlock(&nn->nn_lock);
1882
1883         o2net_register_callbacks(sc->sc_sock->sk, sc);
1884         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1885
1886         o2net_initialize_handshake();
1887         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1888
1889 out:
1890         if (new_sock)
1891                 sock_release(new_sock);
1892         if (node)
1893                 o2nm_node_put(node);
1894         if (local_node)
1895                 o2nm_node_put(local_node);
1896         if (sc)
1897                 sc_put(sc);
1898         return ret;
1899 }
1900
1901 static void o2net_accept_many(struct work_struct *work)
1902 {
1903         struct socket *sock = o2net_listen_sock;
1904         while (o2net_accept_one(sock) == 0)
1905                 cond_resched();
1906 }
1907
1908 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1909 {
1910         void (*ready)(struct sock *sk, int bytes);
1911
1912         read_lock(&sk->sk_callback_lock);
1913         ready = sk->sk_user_data;
1914         if (ready == NULL) { /* check for teardown race */
1915                 ready = sk->sk_data_ready;
1916                 goto out;
1917         }
1918
1919         /* ->sk_data_ready is also called for a newly established child socket
1920          * before it has been accepted and the acceptor has set up their
1921          * data_ready.. we only want to queue listen work for our listening
1922          * socket */
1923         if (sk->sk_state == TCP_LISTEN) {
1924                 mlog(ML_TCP, "bytes: %d\n", bytes);
1925                 queue_work(o2net_wq, &o2net_listen_work);
1926         }
1927
1928 out:
1929         read_unlock(&sk->sk_callback_lock);
1930         ready(sk, bytes);
1931 }
1932
1933 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1934 {
1935         struct socket *sock = NULL;
1936         int ret;
1937         struct sockaddr_in sin = {
1938                 .sin_family = PF_INET,
1939                 .sin_addr = { .s_addr = addr },
1940                 .sin_port = port,
1941         };
1942
1943         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1944         if (ret < 0) {
1945                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1946                 goto out;
1947         }
1948
1949         sock->sk->sk_allocation = GFP_ATOMIC;
1950
1951         write_lock_bh(&sock->sk->sk_callback_lock);
1952         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1953         sock->sk->sk_data_ready = o2net_listen_data_ready;
1954         write_unlock_bh(&sock->sk->sk_callback_lock);
1955
1956         o2net_listen_sock = sock;
1957         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1958
1959         sock->sk->sk_reuse = 1;
1960         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1961         if (ret < 0) {
1962                 mlog(ML_ERROR, "unable to bind socket at %pI4:%u, "
1963                      "ret=%d\n", &addr, ntohs(port), ret);
1964                 goto out;
1965         }
1966
1967         ret = sock->ops->listen(sock, 64);
1968         if (ret < 0) {
1969                 mlog(ML_ERROR, "unable to listen on %pI4:%u, ret=%d\n",
1970                      &addr, ntohs(port), ret);
1971         }
1972
1973 out:
1974         if (ret) {
1975                 o2net_listen_sock = NULL;
1976                 if (sock)
1977                         sock_release(sock);
1978         }
1979         return ret;
1980 }
1981
1982 /*
1983  * called from node manager when we should bring up our network listening
1984  * socket.  node manager handles all the serialization to only call this
1985  * once and to match it with o2net_stop_listening().  note,
1986  * o2nm_this_node() doesn't work yet as we're being called while it
1987  * is being set up.
1988  */
1989 int o2net_start_listening(struct o2nm_node *node)
1990 {
1991         int ret = 0;
1992
1993         BUG_ON(o2net_wq != NULL);
1994         BUG_ON(o2net_listen_sock != NULL);
1995
1996         mlog(ML_KTHREAD, "starting o2net thread...\n");
1997         o2net_wq = create_singlethread_workqueue("o2net");
1998         if (o2net_wq == NULL) {
1999                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2000                 return -ENOMEM; /* ? */
2001         }
2002
2003         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2004                                         node->nd_ipv4_port);
2005         if (ret) {
2006                 destroy_workqueue(o2net_wq);
2007                 o2net_wq = NULL;
2008         } else
2009                 o2quo_conn_up(node->nd_num);
2010
2011         return ret;
2012 }
2013
2014 /* again, o2nm_this_node() doesn't work here as we're involved in
2015  * tearing it down */
2016 void o2net_stop_listening(struct o2nm_node *node)
2017 {
2018         struct socket *sock = o2net_listen_sock;
2019         size_t i;
2020
2021         BUG_ON(o2net_wq == NULL);
2022         BUG_ON(o2net_listen_sock == NULL);
2023
2024         /* stop the listening socket from generating work */
2025         write_lock_bh(&sock->sk->sk_callback_lock);
2026         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2027         sock->sk->sk_user_data = NULL;
2028         write_unlock_bh(&sock->sk->sk_callback_lock);
2029
2030         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2031                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2032                 if (node) {
2033                         o2net_disconnect_node(node);
2034                         o2nm_node_put(node);
2035                 }
2036         }
2037
2038         /* finish all work and tear down the work queue */
2039         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2040         destroy_workqueue(o2net_wq);
2041         o2net_wq = NULL;
2042
2043         sock_release(o2net_listen_sock);
2044         o2net_listen_sock = NULL;
2045
2046         o2quo_conn_err(node->nd_num);
2047 }
2048
2049 /* ------------------------------------------------------------ */
2050
2051 int o2net_init(void)
2052 {
2053         unsigned long i;
2054
2055         o2quo_init();
2056
2057         if (o2net_debugfs_init())
2058                 return -ENOMEM;
2059
2060         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2061         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2062         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2063         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2064                 kfree(o2net_hand);
2065                 kfree(o2net_keep_req);
2066                 kfree(o2net_keep_resp);
2067                 return -ENOMEM;
2068         }
2069
2070         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2071         o2net_hand->connector_id = cpu_to_be64(1);
2072
2073         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2074         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2075
2076         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2077                 struct o2net_node *nn = o2net_nn_from_num(i);
2078
2079                 atomic_set(&nn->nn_timeout, 0);
2080                 spin_lock_init(&nn->nn_lock);
2081                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2082                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2083                                   o2net_connect_expired);
2084                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2085                 /* until we see hb from a node we'll return einval */
2086                 nn->nn_persistent_error = -ENOTCONN;
2087                 init_waitqueue_head(&nn->nn_sc_wq);
2088                 idr_init(&nn->nn_status_idr);
2089                 INIT_LIST_HEAD(&nn->nn_status_list);
2090         }
2091
2092         return 0;
2093 }
2094
2095 void o2net_exit(void)
2096 {
2097         o2quo_exit();
2098         kfree(o2net_hand);
2099         kfree(o2net_keep_req);
2100         kfree(o2net_keep_resp);
2101         o2net_debugfs_exit();
2102 }