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 <linux/export.h>
  63 #include <net/tcp.h>
  64
  65 #include <asm/uaccess.h>
  66
  67 #include "heartbeat.h"
  68 #include "tcp.h"
  69 #include "nodemanager.h"
  70 #define MLOG_MASK_PREFIX ML_TCP
  71 #include "masklog.h"
  72 #include "quorum.h"
  73
  74 #include "tcp_internal.h"
  75
  76 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
  77 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
  78                           &sc->sc_node->nd_ipv4_address,                \
  79                           ntohs(sc->sc_node->nd_ipv4_port)
  80
  81 /*
  82  * In the following two log macros, the whitespace after the ',' just
  83  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  84  * previous token if args expands to nothing.
  85  */
  86 #define msglog(hdr, fmt, args...) do {                                  \
  87         typeof(hdr) __hdr = (hdr);                                      \
  88         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
  89              "key %08x num %u] " fmt,                                   \
  90              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
  91              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
  92              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
  93              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
  94 } while (0)
  95
  96 #define sclog(sc, fmt, args...) do {                                    \
  97         typeof(sc) __sc = (sc);                                         \
  98         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
  99              "pg_off %zu] " fmt, __sc,                                  \
 100              atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,       \
 101             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
 102             ##args);                                                    \
 103 } while (0)
 104
 105 static DEFINE_RWLOCK(o2net_handler_lock);
 106 static struct rb_root o2net_handler_tree = RB_ROOT;
 107
 108 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
 109
 110 /* XXX someday we'll need better accounting */
 111 static struct socket *o2net_listen_sock = NULL;
 112
 113 /*
 114  * listen work is only queued by the listening socket callbacks on the
 115  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 116  * quorum work is queued as sock containers are shutdown.. stop_listening
 117  * tears down all the node's sock containers, preventing future shutdowns
 118  * and queued quroum work, before canceling delayed quorum work and
 119  * destroying the work queue.
 120  */
 121 static struct workqueue_struct *o2net_wq;
 122 static struct work_struct o2net_listen_work;
 123
 124 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 125 #define O2NET_HB_PRI 0x1
 126
 127 static struct o2net_handshake *o2net_hand;
 128 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 129
 130 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 131                 {[O2NET_ERR_NONE]       = 0,
 132                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
 133                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
 134                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
 135
 136 /* can't quite avoid *all* internal declarations :/ */
 137 static void o2net_sc_connect_completed(struct work_struct *work);
 138 static void o2net_rx_until_empty(struct work_struct *work);
 139 static void o2net_shutdown_sc(struct work_struct *work);
 140 static void o2net_listen_data_ready(struct sock *sk, int bytes);
 141 static void o2net_sc_send_keep_req(struct work_struct *work);
 142 static void o2net_idle_timer(unsigned long data);
 143 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 144 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 145
 146 #ifdef CONFIG_DEBUG_FS
 147 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 148                            u32 msgkey, struct task_struct *task, u8 node)
 149 {
 150         INIT_LIST_HEAD(&nst->st_net_debug_item);
 151         nst->st_task = task;
 152         nst->st_msg_type = msgtype;
 153         nst->st_msg_key = msgkey;
 154         nst->st_node = node;
 155 }
 156
 157 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 158 {
 159         nst->st_sock_time = ktime_get();
 160 }
 161
 162 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 163 {
 164         nst->st_send_time = ktime_get();
 165 }
 166
 167 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 168 {
 169         nst->st_status_time = ktime_get();
 170 }
 171
 172 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 173                                                 struct o2net_sock_container *sc)
 174 {
 175         nst->st_sc = sc;
 176 }
 177
 178 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 179                                         u32 msg_id)
 180 {
 181         nst->st_id = msg_id;
 182 }
 183
 184 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
 185 {
 186         sc->sc_tv_timer = ktime_get();
 187 }
 188
 189 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
 190 {
 191         sc->sc_tv_data_ready = ktime_get();
 192 }
 193
 194 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
 195 {
 196         sc->sc_tv_advance_start = ktime_get();
 197 }
 198
 199 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
 200 {
 201         sc->sc_tv_advance_stop = ktime_get();
 202 }
 203
 204 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
 205 {
 206         sc->sc_tv_func_start = ktime_get();
 207 }
 208
 209 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
 210 {
 211         sc->sc_tv_func_stop = ktime_get();
 212 }
 213
 214 #else  /* CONFIG_DEBUG_FS */
 215 # define o2net_init_nst(a, b, c, d, e)
 216 # define o2net_set_nst_sock_time(a)
 217 # define o2net_set_nst_send_time(a)
 218 # define o2net_set_nst_status_time(a)
 219 # define o2net_set_nst_sock_container(a, b)
 220 # define o2net_set_nst_msg_id(a, b)
 221 # define o2net_set_sock_timer(a)
 222 # define o2net_set_data_ready_time(a)
 223 # define o2net_set_advance_start_time(a)
 224 # define o2net_set_advance_stop_time(a)
 225 # define o2net_set_func_start_time(a)
 226 # define o2net_set_func_stop_time(a)
 227 #endif /* CONFIG_DEBUG_FS */
 228
 229 #ifdef CONFIG_OCFS2_FS_STATS
 230 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
 231 {
 232         return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
 233 }
 234
 235 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
 236                                     struct o2net_sock_container *sc)
 237 {
 238         sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
 239                                            ktime_sub(ktime_get(),
 240                                                      nst->st_status_time));
 241         sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
 242                                          ktime_sub(nst->st_status_time,
 243                                                    nst->st_send_time));
 244         sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
 245                                             ktime_sub(nst->st_send_time,
 246                                                       nst->st_sock_time));
 247         sc->sc_send_count++;
 248 }
 249
 250 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
 251 {
 252         sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
 253                                             o2net_get_func_run_time(sc));
 254         sc->sc_recv_count++;
 255 }
 256
 257 #else
 258
 259 # define o2net_update_send_stats(a, b)
 260
 261 # define o2net_update_recv_stats(sc)
 262
 263 #endif /* CONFIG_OCFS2_FS_STATS */
 264
 265 static inline int o2net_reconnect_delay(void)
 266 {
 267         return o2nm_single_cluster->cl_reconnect_delay_ms;
 268 }
 269
 270 static inline int o2net_keepalive_delay(void)
 271 {
 272         return o2nm_single_cluster->cl_keepalive_delay_ms;
 273 }
 274
 275 static inline int o2net_idle_timeout(void)
 276 {
 277         return o2nm_single_cluster->cl_idle_timeout_ms;
 278 }
 279
 280 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 281 {
 282         int trans;
 283         BUG_ON(err >= O2NET_ERR_MAX);
 284         trans = o2net_sys_err_translations[err];
 285
 286         /* Just in case we mess up the translation table above */
 287         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 288         return trans;
 289 }
 290
 291 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 292 {
 293         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 294         return &o2net_nodes[node_num];
 295 }
 296
 297 static u8 o2net_num_from_nn(struct o2net_node *nn)
 298 {
 299         BUG_ON(nn == NULL);
 300         return nn - o2net_nodes;
 301 }
 302
 303 /* ------------------------------------------------------------ */
 304
 305 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 306 {
 307         int ret = 0;
 308
 309         do {
 310                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
 311                         ret = -EAGAIN;
 312                         break;
 313                 }
 314                 spin_lock(&nn->nn_lock);
 315                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
 316                 if (ret == 0)
 317                         list_add_tail(&nsw->ns_node_item,
 318                                       &nn->nn_status_list);
 319                 spin_unlock(&nn->nn_lock);
 320         } while (ret == -EAGAIN);
 321
 322         if (ret == 0)  {
 323                 init_waitqueue_head(&nsw->ns_wq);
 324                 nsw->ns_sys_status = O2NET_ERR_NONE;
 325                 nsw->ns_status = 0;
 326         }
 327
 328         return ret;
 329 }
 330
 331 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 332                                       struct o2net_status_wait *nsw,
 333                                       enum o2net_system_error sys_status,
 334                                       s32 status)
 335 {
 336         assert_spin_locked(&nn->nn_lock);
 337
 338         if (!list_empty(&nsw->ns_node_item)) {
 339                 list_del_init(&nsw->ns_node_item);
 340                 nsw->ns_sys_status = sys_status;
 341                 nsw->ns_status = status;
 342                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 343                 wake_up(&nsw->ns_wq);
 344         }
 345 }
 346
 347 static void o2net_complete_nsw(struct o2net_node *nn,
 348                                struct o2net_status_wait *nsw,
 349                                u64 id, enum o2net_system_error sys_status,
 350                                s32 status)
 351 {
 352         spin_lock(&nn->nn_lock);
 353         if (nsw == NULL) {
 354                 if (id > INT_MAX)
 355                         goto out;
 356
 357                 nsw = idr_find(&nn->nn_status_idr, id);
 358                 if (nsw == NULL)
 359                         goto out;
 360         }
 361
 362         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 363
 364 out:
 365         spin_unlock(&nn->nn_lock);
 366         return;
 367 }
 368
 369 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 370 {
 371         struct o2net_status_wait *nsw, *tmp;
 372         unsigned int num_kills = 0;
 373
 374         assert_spin_locked(&nn->nn_lock);
 375
 376         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 377                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 378                 num_kills++;
 379         }
 380
 381         mlog(0, "completed %d messages for node %u\n", num_kills,
 382              o2net_num_from_nn(nn));
 383 }
 384
 385 static int o2net_nsw_completed(struct o2net_node *nn,
 386                                struct o2net_status_wait *nsw)
 387 {
 388         int completed;
 389         spin_lock(&nn->nn_lock);
 390         completed = list_empty(&nsw->ns_node_item);
 391         spin_unlock(&nn->nn_lock);
 392         return completed;
 393 }
 394
 395 /* ------------------------------------------------------------ */
 396
 397 static void sc_kref_release(struct kref *kref)
 398 {
 399         struct o2net_sock_container *sc = container_of(kref,
 400                                         struct o2net_sock_container, sc_kref);
 401         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 402
 403         sclog(sc, "releasing\n");
 404
 405         if (sc->sc_sock) {
 406                 sock_release(sc->sc_sock);
 407                 sc->sc_sock = NULL;
 408         }
 409
 410         o2nm_undepend_item(&sc->sc_node->nd_item);
 411         o2nm_node_put(sc->sc_node);
 412         sc->sc_node = NULL;
 413
 414         o2net_debug_del_sc(sc);
 415         kfree(sc);
 416 }
 417
 418 static void sc_put(struct o2net_sock_container *sc)
 419 {
 420         sclog(sc, "put\n");
 421         kref_put(&sc->sc_kref, sc_kref_release);
 422 }
 423 static void sc_get(struct o2net_sock_container *sc)
 424 {
 425         sclog(sc, "get\n");
 426         kref_get(&sc->sc_kref);
 427 }
 428 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 429 {
 430         struct o2net_sock_container *sc, *ret = NULL;
 431         struct page *page = NULL;
 432         int status = 0;
 433
 434         page = alloc_page(GFP_NOFS);
 435         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 436         if (sc == NULL || page == NULL)
 437                 goto out;
 438
 439         kref_init(&sc->sc_kref);
 440         o2nm_node_get(node);
 441         sc->sc_node = node;
 442
 443         /* pin the node item of the remote node */
 444         status = o2nm_depend_item(&node->nd_item);
 445         if (status) {
 446                 mlog_errno(status);
 447                 o2nm_node_put(node);
 448                 goto out;
 449         }
 450         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 451         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 452         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 453         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 454
 455         init_timer(&sc->sc_idle_timeout);
 456         sc->sc_idle_timeout.function = o2net_idle_timer;
 457         sc->sc_idle_timeout.data = (unsigned long)sc;
 458
 459         sclog(sc, "alloced\n");
 460
 461         ret = sc;
 462         sc->sc_page = page;
 463         o2net_debug_add_sc(sc);
 464         sc = NULL;
 465         page = NULL;
 466
 467 out:
 468         if (page)
 469                 __free_page(page);
 470         kfree(sc);
 471
 472         return ret;
 473 }
 474
 475 /* ------------------------------------------------------------ */
 476
 477 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 478                                 struct work_struct *work)
 479 {
 480         sc_get(sc);
 481         if (!queue_work(o2net_wq, work))
 482                 sc_put(sc);
 483 }
 484 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 485                                         struct delayed_work *work,
 486                                         int delay)
 487 {
 488         sc_get(sc);
 489         if (!queue_delayed_work(o2net_wq, work, delay))
 490                 sc_put(sc);
 491 }
 492 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 493                                          struct delayed_work *work)
 494 {
 495         if (cancel_delayed_work(work))
 496                 sc_put(sc);
 497 }
 498
 499 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 500
 501 int o2net_num_connected_peers(void)
 502 {
 503         return atomic_read(&o2net_connected_peers);
 504 }
 505
 506 static void o2net_set_nn_state(struct o2net_node *nn,
 507                                struct o2net_sock_container *sc,
 508                                unsigned valid, int err)
 509 {
 510         int was_valid = nn->nn_sc_valid;
 511         int was_err = nn->nn_persistent_error;
 512         struct o2net_sock_container *old_sc = nn->nn_sc;
 513
 514         assert_spin_locked(&nn->nn_lock);
 515
 516         if (old_sc && !sc)
 517                 atomic_dec(&o2net_connected_peers);
 518         else if (!old_sc && sc)
 519                 atomic_inc(&o2net_connected_peers);
 520
 521         /* the node num comparison and single connect/accept path should stop
 522          * an non-null sc from being overwritten with another */
 523         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 524         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 525         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 526
 527         if (was_valid && !valid && err == 0)
 528                 err = -ENOTCONN;
 529
 530         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 531              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 532              nn->nn_persistent_error, err);
 533
 534         nn->nn_sc = sc;
 535         nn->nn_sc_valid = valid ? 1 : 0;
 536         nn->nn_persistent_error = err;
 537
 538         /* mirrors o2net_tx_can_proceed() */
 539         if (nn->nn_persistent_error || nn->nn_sc_valid)
 540                 wake_up(&nn->nn_sc_wq);
 541
 542         if (!was_err && nn->nn_persistent_error) {
 543                 o2quo_conn_err(o2net_num_from_nn(nn));
 544                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 545                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 546         }
 547
 548         if (was_valid && !valid) {
 549                 printk(KERN_NOTICE "o2net: No longer connected to "
 550                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 551                 o2net_complete_nodes_nsw(nn);
 552         }
 553
 554         if (!was_valid && valid) {
 555                 o2quo_conn_up(o2net_num_from_nn(nn));
 556                 cancel_delayed_work(&nn->nn_connect_expired);
 557                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 558                        o2nm_this_node() > sc->sc_node->nd_num ?
 559                        "Connected to" : "Accepted connection from",
 560                        SC_NODEF_ARGS(sc));
 561         }
 562
 563         /* trigger the connecting worker func as long as we're not valid,
 564          * it will back off if it shouldn't connect.  This can be called
 565          * from node config teardown and so needs to be careful about
 566          * the work queue actually being up. */
 567         if (!valid && o2net_wq) {
 568                 unsigned long delay;
 569                 /* delay if we're within a RECONNECT_DELAY of the
 570                  * last attempt */
 571                 delay = (nn->nn_last_connect_attempt +
 572                          msecs_to_jiffies(o2net_reconnect_delay()))
 573                         - jiffies;
 574                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 575                         delay = 0;
 576                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 577                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 578
 579                 /*
 580                  * Delay the expired work after idle timeout.
 581                  *
 582                  * We might have lots of failed connection attempts that run
 583                  * through here but we only cancel the connect_expired work when
 584                  * a connection attempt succeeds.  So only the first enqueue of
 585                  * the connect_expired work will do anything.  The rest will see
 586                  * that it's already queued and do nothing.
 587                  */
 588                 delay += msecs_to_jiffies(o2net_idle_timeout());
 589                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 590         }
 591
 592         /* keep track of the nn's sc ref for the caller */
 593         if ((old_sc == NULL) && sc)
 594                 sc_get(sc);
 595         if (old_sc && (old_sc != sc)) {
 596                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 597                 sc_put(old_sc);
 598         }
 599 }
 600
 601 /* see o2net_register_callbacks() */
 602 static void o2net_data_ready(struct sock *sk, int bytes)
 603 {
 604         void (*ready)(struct sock *sk, int bytes);
 605
 606         read_lock(&sk->sk_callback_lock);
 607         if (sk->sk_user_data) {
 608                 struct o2net_sock_container *sc = sk->sk_user_data;
 609                 sclog(sc, "data_ready hit\n");
 610                 o2net_set_data_ready_time(sc);
 611                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 612                 ready = sc->sc_data_ready;
 613         } else {
 614                 ready = sk->sk_data_ready;
 615         }
 616         read_unlock(&sk->sk_callback_lock);
 617
 618         ready(sk, bytes);
 619 }
 620
 621 /* see o2net_register_callbacks() */
 622 static void o2net_state_change(struct sock *sk)
 623 {
 624         void (*state_change)(struct sock *sk);
 625         struct o2net_sock_container *sc;
 626
 627         read_lock(&sk->sk_callback_lock);
 628         sc = sk->sk_user_data;
 629         if (sc == NULL) {
 630                 state_change = sk->sk_state_change;
 631                 goto out;
 632         }
 633
 634         sclog(sc, "state_change to %d\n", sk->sk_state);
 635
 636         state_change = sc->sc_state_change;
 637
 638         switch(sk->sk_state) {
 639                 /* ignore connecting sockets as they make progress */
 640                 case TCP_SYN_SENT:
 641                 case TCP_SYN_RECV:
 642                         break;
 643                 case TCP_ESTABLISHED:
 644                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
 645                         break;
 646                 default:
 647                         printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
 648                               " shutdown, state %d\n",
 649                               SC_NODEF_ARGS(sc), sk->sk_state);
 650                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 651                         break;
 652         }
 653 out:
 654         read_unlock(&sk->sk_callback_lock);
 655         state_change(sk);
 656 }
 657
 658 /*
 659  * we register callbacks so we can queue work on events before calling
 660  * the original callbacks.  our callbacks our careful to test user_data
 661  * to discover when they've reaced with o2net_unregister_callbacks().
 662  */
 663 static void o2net_register_callbacks(struct sock *sk,
 664                                      struct o2net_sock_container *sc)
 665 {
 666         write_lock_bh(&sk->sk_callback_lock);
 667
 668         /* accepted sockets inherit the old listen socket data ready */
 669         if (sk->sk_data_ready == o2net_listen_data_ready) {
 670                 sk->sk_data_ready = sk->sk_user_data;
 671                 sk->sk_user_data = NULL;
 672         }
 673
 674         BUG_ON(sk->sk_user_data != NULL);
 675         sk->sk_user_data = sc;
 676         sc_get(sc);
 677
 678         sc->sc_data_ready = sk->sk_data_ready;
 679         sc->sc_state_change = sk->sk_state_change;
 680         sk->sk_data_ready = o2net_data_ready;
 681         sk->sk_state_change = o2net_state_change;
 682
 683         mutex_init(&sc->sc_send_lock);
 684
 685         write_unlock_bh(&sk->sk_callback_lock);
 686 }
 687
 688 static int o2net_unregister_callbacks(struct sock *sk,
 689                                    struct o2net_sock_container *sc)
 690 {
 691         int ret = 0;
 692
 693         write_lock_bh(&sk->sk_callback_lock);
 694         if (sk->sk_user_data == sc) {
 695                 ret = 1;
 696                 sk->sk_user_data = NULL;
 697                 sk->sk_data_ready = sc->sc_data_ready;
 698                 sk->sk_state_change = sc->sc_state_change;
 699         }
 700         write_unlock_bh(&sk->sk_callback_lock);
 701
 702         return ret;
 703 }
 704
 705 /*
 706  * this is a little helper that is called by callers who have seen a problem
 707  * with an sc and want to detach it from the nn if someone already hasn't beat
 708  * them to it.  if an error is given then the shutdown will be persistent
 709  * and pending transmits will be canceled.
 710  */
 711 static void o2net_ensure_shutdown(struct o2net_node *nn,
 712                                    struct o2net_sock_container *sc,
 713                                    int err)
 714 {
 715         spin_lock(&nn->nn_lock);
 716         if (nn->nn_sc == sc)
 717                 o2net_set_nn_state(nn, NULL, 0, err);
 718         spin_unlock(&nn->nn_lock);
 719 }
 720
 721 /*
 722  * This work queue function performs the blocking parts of socket shutdown.  A
 723  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 724  * sc detached from the nn.  state_change will also trigger this callback
 725  * directly when it sees errors.  In that case we need to call set_nn_state
 726  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 727  * itself.
 728  */
 729 static void o2net_shutdown_sc(struct work_struct *work)
 730 {
 731         struct o2net_sock_container *sc =
 732                 container_of(work, struct o2net_sock_container,
 733                              sc_shutdown_work);
 734         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 735
 736         sclog(sc, "shutting down\n");
 737
 738         /* drop the callbacks ref and call shutdown only once */
 739         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 740                 /* we shouldn't flush as we're in the thread, the
 741                  * races with pending sc work structs are harmless */
 742                 del_timer_sync(&sc->sc_idle_timeout);
 743                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 744                 sc_put(sc);
 745                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 746         }
 747
 748         /* not fatal so failed connects before the other guy has our
 749          * heartbeat can be retried */
 750         o2net_ensure_shutdown(nn, sc, 0);
 751         sc_put(sc);
 752 }
 753
 754 /* ------------------------------------------------------------ */
 755
 756 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 757                              u32 key)
 758 {
 759         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 760
 761         if (ret == 0)
 762                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 763
 764         return ret;
 765 }
 766
 767 static struct o2net_msg_handler *
 768 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 769                           struct rb_node **ret_parent)
 770 {
 771         struct rb_node **p = &o2net_handler_tree.rb_node;
 772         struct rb_node *parent = NULL;
 773         struct o2net_msg_handler *nmh, *ret = NULL;
 774         int cmp;
 775
 776         while (*p) {
 777                 parent = *p;
 778                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 779                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 780
 781                 if (cmp < 0)
 782                         p = &(*p)->rb_left;
 783                 else if (cmp > 0)
 784                         p = &(*p)->rb_right;
 785                 else {
 786                         ret = nmh;
 787                         break;
 788                 }
 789         }
 790
 791         if (ret_p != NULL)
 792                 *ret_p = p;
 793         if (ret_parent != NULL)
 794                 *ret_parent = parent;
 795
 796         return ret;
 797 }
 798
 799 static void o2net_handler_kref_release(struct kref *kref)
 800 {
 801         struct o2net_msg_handler *nmh;
 802         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 803
 804         kfree(nmh);
 805 }
 806
 807 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 808 {
 809         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 810 }
 811
 812 /* max_len is protection for the handler func.  incoming messages won't
 813  * be given to the handler if their payload is longer than the max. */
 814 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 815                            o2net_msg_handler_func *func, void *data,
 816                            o2net_post_msg_handler_func *post_func,
 817                            struct list_head *unreg_list)
 818 {
 819         struct o2net_msg_handler *nmh = NULL;
 820         struct rb_node **p, *parent;
 821         int ret = 0;
 822
 823         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 824                 mlog(0, "max_len for message handler out of range: %u\n",
 825                         max_len);
 826                 ret = -EINVAL;
 827                 goto out;
 828         }
 829
 830         if (!msg_type) {
 831                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 832                 ret = -EINVAL;
 833                 goto out;
 834
 835         }
 836         if (!func) {
 837                 mlog(0, "no message handler provided: %u, %p\n",
 838                        msg_type, func);
 839                 ret = -EINVAL;
 840                 goto out;
 841         }
 842
 843         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 844         if (nmh == NULL) {
 845                 ret = -ENOMEM;
 846                 goto out;
 847         }
 848
 849         nmh->nh_func = func;
 850         nmh->nh_func_data = data;
 851         nmh->nh_post_func = post_func;
 852         nmh->nh_msg_type = msg_type;
 853         nmh->nh_max_len = max_len;
 854         nmh->nh_key = key;
 855         /* the tree and list get this ref.. they're both removed in
 856          * unregister when this ref is dropped */
 857         kref_init(&nmh->nh_kref);
 858         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 859
 860         write_lock(&o2net_handler_lock);
 861         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 862                 ret = -EEXIST;
 863         else {
 864                 rb_link_node(&nmh->nh_node, parent, p);
 865                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 866                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 867
 868                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 869                      func, msg_type, key);
 870                 /* we've had some trouble with handlers seemingly vanishing. */
 871                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 872                                                           &parent) == NULL,
 873                                 "couldn't find handler we *just* registerd "
 874                                 "for type %u key %08x\n", msg_type, key);
 875         }
 876         write_unlock(&o2net_handler_lock);
 877         if (ret)
 878                 goto out;
 879
 880 out:
 881         if (ret)
 882                 kfree(nmh);
 883
 884         return ret;
 885 }
 886 EXPORT_SYMBOL_GPL(o2net_register_handler);
 887
 888 void o2net_unregister_handler_list(struct list_head *list)
 889 {
 890         struct o2net_msg_handler *nmh, *n;
 891
 892         write_lock(&o2net_handler_lock);
 893         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 894                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 895                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 896                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 897                 list_del_init(&nmh->nh_unregister_item);
 898                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 899         }
 900         write_unlock(&o2net_handler_lock);
 901 }
 902 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 903
 904 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 905 {
 906         struct o2net_msg_handler *nmh;
 907
 908         read_lock(&o2net_handler_lock);
 909         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 910         if (nmh)
 911                 kref_get(&nmh->nh_kref);
 912         read_unlock(&o2net_handler_lock);
 913
 914         return nmh;
 915 }
 916
 917 /* ------------------------------------------------------------ */
 918
 919 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 920 {
 921         int ret;
 922         mm_segment_t oldfs;
 923         struct kvec vec = {
 924                 .iov_len = len,
 925                 .iov_base = data,
 926         };
 927         struct msghdr msg = {
 928                 .msg_iovlen = 1,
 929                 .msg_iov = (struct iovec *)&vec,
 930                 .msg_flags = MSG_DONTWAIT,
 931         };
 932
 933         oldfs = get_fs();
 934         set_fs(get_ds());
 935         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
 936         set_fs(oldfs);
 937
 938         return ret;
 939 }
 940
 941 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 942                               size_t veclen, size_t total)
 943 {
 944         int ret;
 945         mm_segment_t oldfs;
 946         struct msghdr msg = {
 947                 .msg_iov = (struct iovec *)vec,
 948                 .msg_iovlen = veclen,
 949         };
 950
 951         if (sock == NULL) {
 952                 ret = -EINVAL;
 953                 goto out;
 954         }
 955
 956         oldfs = get_fs();
 957         set_fs(get_ds());
 958         ret = sock_sendmsg(sock, &msg, total);
 959         set_fs(oldfs);
 960         if (ret != total) {
 961                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
 962                      total);
 963                 if (ret >= 0)
 964                         ret = -EPIPE; /* should be smarter, I bet */
 965                 goto out;
 966         }
 967
 968         ret = 0;
 969 out:
 970         if (ret < 0)
 971                 mlog(0, "returning error: %d\n", ret);
 972         return ret;
 973 }
 974
 975 static void o2net_sendpage(struct o2net_sock_container *sc,
 976                            void *kmalloced_virt,
 977                            size_t size)
 978 {
 979         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 980         ssize_t ret;
 981
 982         while (1) {
 983                 mutex_lock(&sc->sc_send_lock);
 984                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 985                                                  virt_to_page(kmalloced_virt),
 986                                                  (long)kmalloced_virt & ~PAGE_MASK,
 987                                                  size, MSG_DONTWAIT);
 988                 mutex_unlock(&sc->sc_send_lock);
 989                 if (ret == size)
 990                         break;
 991                 if (ret == (ssize_t)-EAGAIN) {
 992                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 993                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 994                         cond_resched();
 995                         continue;
 996                 }
 997                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 998                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 999                 o2net_ensure_shutdown(nn, sc, 0);
1000                 break;
1001         }
1002 }
1003
1004 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
1005 {
1006         memset(msg, 0, sizeof(struct o2net_msg));
1007         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
1008         msg->data_len = cpu_to_be16(data_len);
1009         msg->msg_type = cpu_to_be16(msg_type);
1010         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
1011         msg->status = 0;
1012         msg->key = cpu_to_be32(key);
1013 }
1014
1015 static int o2net_tx_can_proceed(struct o2net_node *nn,
1016                                 struct o2net_sock_container **sc_ret,
1017                                 int *error)
1018 {
1019         int ret = 0;
1020
1021         spin_lock(&nn->nn_lock);
1022         if (nn->nn_persistent_error) {
1023                 ret = 1;
1024                 *sc_ret = NULL;
1025                 *error = nn->nn_persistent_error;
1026         } else if (nn->nn_sc_valid) {
1027                 kref_get(&nn->nn_sc->sc_kref);
1028
1029                 ret = 1;
1030                 *sc_ret = nn->nn_sc;
1031                 *error = 0;
1032         }
1033         spin_unlock(&nn->nn_lock);
1034
1035         return ret;
1036 }
1037
1038 /* Get a map of all nodes to which this node is currently connected to */
1039 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1040 {
1041         struct o2net_sock_container *sc;
1042         int node, ret;
1043
1044         BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1045
1046         memset(map, 0, bytes);
1047         for (node = 0; node < O2NM_MAX_NODES; ++node) {
1048                 o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret);
1049                 if (!ret) {
1050                         set_bit(node, map);
1051                         sc_put(sc);
1052                 }
1053         }
1054 }
1055 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1056
1057 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1058                            size_t caller_veclen, u8 target_node, int *status)
1059 {
1060         int ret = 0;
1061         struct o2net_msg *msg = NULL;
1062         size_t veclen, caller_bytes = 0;
1063         struct kvec *vec = NULL;
1064         struct o2net_sock_container *sc = NULL;
1065         struct o2net_node *nn = o2net_nn_from_num(target_node);
1066         struct o2net_status_wait nsw = {
1067                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1068         };
1069         struct o2net_send_tracking nst;
1070
1071         o2net_init_nst(&nst, msg_type, key, current, target_node);
1072
1073         if (o2net_wq == NULL) {
1074                 mlog(0, "attempt to tx without o2netd running\n");
1075                 ret = -ESRCH;
1076                 goto out;
1077         }
1078
1079         if (caller_veclen == 0) {
1080                 mlog(0, "bad kvec array length\n");
1081                 ret = -EINVAL;
1082                 goto out;
1083         }
1084
1085         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1086         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1087                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1088                 ret = -EINVAL;
1089                 goto out;
1090         }
1091
1092         if (target_node == o2nm_this_node()) {
1093                 ret = -ELOOP;
1094                 goto out;
1095         }
1096
1097         o2net_debug_add_nst(&nst);
1098
1099         o2net_set_nst_sock_time(&nst);
1100
1101         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1102         if (ret)
1103                 goto out;
1104
1105         o2net_set_nst_sock_container(&nst, sc);
1106
1107         veclen = caller_veclen + 1;
1108         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1109         if (vec == NULL) {
1110                 mlog(0, "failed to %zu element kvec!\n", veclen);
1111                 ret = -ENOMEM;
1112                 goto out;
1113         }
1114
1115         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1116         if (!msg) {
1117                 mlog(0, "failed to allocate a o2net_msg!\n");
1118                 ret = -ENOMEM;
1119                 goto out;
1120         }
1121
1122         o2net_init_msg(msg, caller_bytes, msg_type, key);
1123
1124         vec[0].iov_len = sizeof(struct o2net_msg);
1125         vec[0].iov_base = msg;
1126         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1127
1128         ret = o2net_prep_nsw(nn, &nsw);
1129         if (ret)
1130                 goto out;
1131
1132         msg->msg_num = cpu_to_be32(nsw.ns_id);
1133         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1134
1135         o2net_set_nst_send_time(&nst);
1136
1137         /* finally, convert the message header to network byte-order
1138          * and send */
1139         mutex_lock(&sc->sc_send_lock);
1140         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1141                                  sizeof(struct o2net_msg) + caller_bytes);
1142         mutex_unlock(&sc->sc_send_lock);
1143         msglog(msg, "sending returned %d\n", ret);
1144         if (ret < 0) {
1145                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1146                 goto out;
1147         }
1148
1149         /* wait on other node's handler */
1150         o2net_set_nst_status_time(&nst);
1151         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1152
1153         o2net_update_send_stats(&nst, sc);
1154
1155         /* Note that we avoid overwriting the callers status return
1156          * variable if a system error was reported on the other
1157          * side. Callers beware. */
1158         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1159         if (status && !ret)
1160                 *status = nsw.ns_status;
1161
1162         mlog(0, "woken, returning system status %d, user status %d\n",
1163              ret, nsw.ns_status);
1164 out:
1165         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1166         if (sc)
1167                 sc_put(sc);
1168         if (vec)
1169                 kfree(vec);
1170         if (msg)
1171                 kfree(msg);
1172         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1173         return ret;
1174 }
1175 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1176
1177 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1178                        u8 target_node, int *status)
1179 {
1180         struct kvec vec = {
1181                 .iov_base = data,
1182                 .iov_len = len,
1183         };
1184         return o2net_send_message_vec(msg_type, key, &vec, 1,
1185                                       target_node, status);
1186 }
1187 EXPORT_SYMBOL_GPL(o2net_send_message);
1188
1189 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1190                                    enum o2net_system_error syserr, int err)
1191 {
1192         struct kvec vec = {
1193                 .iov_base = hdr,
1194                 .iov_len = sizeof(struct o2net_msg),
1195         };
1196
1197         BUG_ON(syserr >= O2NET_ERR_MAX);
1198
1199         /* leave other fields intact from the incoming message, msg_num
1200          * in particular */
1201         hdr->sys_status = cpu_to_be32(syserr);
1202         hdr->status = cpu_to_be32(err);
1203         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1204         hdr->data_len = 0;
1205
1206         msglog(hdr, "about to send status magic %d\n", err);
1207         /* hdr has been in host byteorder this whole time */
1208         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1209 }
1210
1211 /* this returns -errno if the header was unknown or too large, etc.
1212  * after this is called the buffer us reused for the next message */
1213 static int o2net_process_message(struct o2net_sock_container *sc,
1214                                  struct o2net_msg *hdr)
1215 {
1216         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1217         int ret = 0, handler_status;
1218         enum  o2net_system_error syserr;
1219         struct o2net_msg_handler *nmh = NULL;
1220         void *ret_data = NULL;
1221
1222         msglog(hdr, "processing message\n");
1223
1224         o2net_sc_postpone_idle(sc);
1225
1226         switch(be16_to_cpu(hdr->magic)) {
1227                 case O2NET_MSG_STATUS_MAGIC:
1228                         /* special type for returning message status */
1229                         o2net_complete_nsw(nn, NULL,
1230                                            be32_to_cpu(hdr->msg_num),
1231                                            be32_to_cpu(hdr->sys_status),
1232                                            be32_to_cpu(hdr->status));
1233                         goto out;
1234                 case O2NET_MSG_KEEP_REQ_MAGIC:
1235                         o2net_sendpage(sc, o2net_keep_resp,
1236                                        sizeof(*o2net_keep_resp));
1237                         goto out;
1238                 case O2NET_MSG_KEEP_RESP_MAGIC:
1239                         goto out;
1240                 case O2NET_MSG_MAGIC:
1241                         break;
1242                 default:
1243                         msglog(hdr, "bad magic\n");
1244                         ret = -EINVAL;
1245                         goto out;
1246                         break;
1247         }
1248
1249         /* find a handler for it */
1250         handler_status = 0;
1251         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1252                                 be32_to_cpu(hdr->key));
1253         if (!nmh) {
1254                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1255                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1256                 syserr = O2NET_ERR_NO_HNDLR;
1257                 goto out_respond;
1258         }
1259
1260         syserr = O2NET_ERR_NONE;
1261
1262         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1263                 syserr = O2NET_ERR_OVERFLOW;
1264
1265         if (syserr != O2NET_ERR_NONE)
1266                 goto out_respond;
1267
1268         o2net_set_func_start_time(sc);
1269         sc->sc_msg_key = be32_to_cpu(hdr->key);
1270         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1271         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1272                                              be16_to_cpu(hdr->data_len),
1273                                         nmh->nh_func_data, &ret_data);
1274         o2net_set_func_stop_time(sc);
1275
1276         o2net_update_recv_stats(sc);
1277
1278 out_respond:
1279         /* this destroys the hdr, so don't use it after this */
1280         mutex_lock(&sc->sc_send_lock);
1281         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1282                                       handler_status);
1283         mutex_unlock(&sc->sc_send_lock);
1284         hdr = NULL;
1285         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1286              handler_status, syserr, ret);
1287
1288         if (nmh) {
1289                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1290                 if (nmh->nh_post_func)
1291                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1292                                             ret_data);
1293         }
1294
1295 out:
1296         if (nmh)
1297                 o2net_handler_put(nmh);
1298         return ret;
1299 }
1300
1301 static int o2net_check_handshake(struct o2net_sock_container *sc)
1302 {
1303         struct o2net_handshake *hand = page_address(sc->sc_page);
1304         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1305
1306         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1307                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1308                        "protocol version %llu but %llu is required. "
1309                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1310                        (unsigned long long)be64_to_cpu(hand->protocol_version),
1311                        O2NET_PROTOCOL_VERSION);
1312
1313                 /* don't bother reconnecting if its the wrong version. */
1314                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1315                 return -1;
1316         }
1317
1318         /*
1319          * Ensure timeouts are consistent with other nodes, otherwise
1320          * we can end up with one node thinking that the other must be down,
1321          * but isn't. This can ultimately cause corruption.
1322          */
1323         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1324                                 o2net_idle_timeout()) {
1325                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1326                        "idle timeout of %u ms, but we use %u ms locally. "
1327                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1328                        be32_to_cpu(hand->o2net_idle_timeout_ms),
1329                        o2net_idle_timeout());
1330                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1331                 return -1;
1332         }
1333
1334         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1335                         o2net_keepalive_delay()) {
1336                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1337                        "delay of %u ms, but we use %u ms locally. "
1338                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1339                        be32_to_cpu(hand->o2net_keepalive_delay_ms),
1340                        o2net_keepalive_delay());
1341                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1342                 return -1;
1343         }
1344
1345         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1346                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1347                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1348                        "timeout of %u ms, but we use %u ms locally. "
1349                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1350                        be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1351                        O2HB_MAX_WRITE_TIMEOUT_MS);
1352                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1353                 return -1;
1354         }
1355
1356         sc->sc_handshake_ok = 1;
1357
1358         spin_lock(&nn->nn_lock);
1359         /* set valid and queue the idle timers only if it hasn't been
1360          * shut down already */
1361         if (nn->nn_sc == sc) {
1362                 o2net_sc_reset_idle_timer(sc);
1363                 atomic_set(&nn->nn_timeout, 0);
1364                 o2net_set_nn_state(nn, sc, 1, 0);
1365         }
1366         spin_unlock(&nn->nn_lock);
1367
1368         /* shift everything up as though it wasn't there */
1369         sc->sc_page_off -= sizeof(struct o2net_handshake);
1370         if (sc->sc_page_off)
1371                 memmove(hand, hand + 1, sc->sc_page_off);
1372
1373         return 0;
1374 }
1375
1376 /* this demuxes the queued rx bytes into header or payload bits and calls
1377  * handlers as each full message is read off the socket.  it returns -error,
1378  * == 0 eof, or > 0 for progress made.*/
1379 static int o2net_advance_rx(struct o2net_sock_container *sc)
1380 {
1381         struct o2net_msg *hdr;
1382         int ret = 0;
1383         void *data;
1384         size_t datalen;
1385
1386         sclog(sc, "receiving\n");
1387         o2net_set_advance_start_time(sc);
1388
1389         if (unlikely(sc->sc_handshake_ok == 0)) {
1390                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1391                         data = page_address(sc->sc_page) + sc->sc_page_off;
1392                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1393                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1394                         if (ret > 0)
1395                                 sc->sc_page_off += ret;
1396                 }
1397
1398                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1399                         o2net_check_handshake(sc);
1400                         if (unlikely(sc->sc_handshake_ok == 0))
1401                                 ret = -EPROTO;
1402                 }
1403                 goto out;
1404         }
1405
1406         /* do we need more header? */
1407         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1408                 data = page_address(sc->sc_page) + sc->sc_page_off;
1409                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1410                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1411                 if (ret > 0) {
1412                         sc->sc_page_off += ret;
1413                         /* only swab incoming here.. we can
1414                          * only get here once as we cross from
1415                          * being under to over */
1416                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1417                                 hdr = page_address(sc->sc_page);
1418                                 if (be16_to_cpu(hdr->data_len) >
1419                                     O2NET_MAX_PAYLOAD_BYTES)
1420                                         ret = -EOVERFLOW;
1421                         }
1422                 }
1423                 if (ret <= 0)
1424                         goto out;
1425         }
1426
1427         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1428                 /* oof, still don't have a header */
1429                 goto out;
1430         }
1431
1432         /* this was swabbed above when we first read it */
1433         hdr = page_address(sc->sc_page);
1434
1435         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1436
1437         /* do we need more payload? */
1438         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1439                 /* need more payload */
1440                 data = page_address(sc->sc_page) + sc->sc_page_off;
1441                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1442                           sc->sc_page_off;
1443                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1444                 if (ret > 0)
1445                         sc->sc_page_off += ret;
1446                 if (ret <= 0)
1447                         goto out;
1448         }
1449
1450         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1451                 /* we can only get here once, the first time we read
1452                  * the payload.. so set ret to progress if the handler
1453                  * works out. after calling this the message is toast */
1454                 ret = o2net_process_message(sc, hdr);
1455                 if (ret == 0)
1456                         ret = 1;
1457                 sc->sc_page_off = 0;
1458         }
1459
1460 out:
1461         sclog(sc, "ret = %d\n", ret);
1462         o2net_set_advance_stop_time(sc);
1463         return ret;
1464 }
1465
1466 /* this work func is triggerd by data ready.  it reads until it can read no
1467  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1468  * our work the work struct will be marked and we'll be called again. */
1469 static void o2net_rx_until_empty(struct work_struct *work)
1470 {
1471         struct o2net_sock_container *sc =
1472                 container_of(work, struct o2net_sock_container, sc_rx_work);
1473         int ret;
1474
1475         do {
1476                 ret = o2net_advance_rx(sc);
1477         } while (ret > 0);
1478
1479         if (ret <= 0 && ret != -EAGAIN) {
1480                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1481                 sclog(sc, "saw error %d, closing\n", ret);
1482                 /* not permanent so read failed handshake can retry */
1483                 o2net_ensure_shutdown(nn, sc, 0);
1484         }
1485
1486         sc_put(sc);
1487 }
1488
1489 static int o2net_set_nodelay(struct socket *sock)
1490 {
1491         int ret, val = 1;
1492         mm_segment_t oldfs;
1493
1494         oldfs = get_fs();
1495         set_fs(KERNEL_DS);
1496
1497         /*
1498          * Dear unsuspecting programmer,
1499          *
1500          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1501          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1502          * silently turn into SO_DEBUG.
1503          *
1504          * Yours,
1505          * Keeper of hilariously fragile interfaces.
1506          */
1507         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1508                                     (char __user *)&val, sizeof(val));
1509
1510         set_fs(oldfs);
1511         return ret;
1512 }
1513
1514 static void o2net_initialize_handshake(void)
1515 {
1516         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1517                 O2HB_MAX_WRITE_TIMEOUT_MS);
1518         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1519         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1520                 o2net_keepalive_delay());
1521         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1522                 o2net_reconnect_delay());
1523 }
1524
1525 /* ------------------------------------------------------------ */
1526
1527 /* called when a connect completes and after a sock is accepted.  the
1528  * rx path will see the response and mark the sc valid */
1529 static void o2net_sc_connect_completed(struct work_struct *work)
1530 {
1531         struct o2net_sock_container *sc =
1532                 container_of(work, struct o2net_sock_container,
1533                              sc_connect_work);
1534
1535         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1536               (unsigned long long)O2NET_PROTOCOL_VERSION,
1537               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1538
1539         o2net_initialize_handshake();
1540         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1541         sc_put(sc);
1542 }
1543
1544 /* this is called as a work_struct func. */
1545 static void o2net_sc_send_keep_req(struct work_struct *work)
1546 {
1547         struct o2net_sock_container *sc =
1548                 container_of(work, struct o2net_sock_container,
1549                              sc_keepalive_work.work);
1550
1551         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1552         sc_put(sc);
1553 }
1554
1555 /* socket shutdown does a del_timer_sync against this as it tears down.
1556  * we can't start this timer until we've got to the point in sc buildup
1557  * where shutdown is going to be involved */
1558 static void o2net_idle_timer(unsigned long data)
1559 {
1560         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1561         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1562 #ifdef CONFIG_DEBUG_FS
1563         unsigned long msecs = ktime_to_ms(ktime_get()) -
1564                 ktime_to_ms(sc->sc_tv_timer);
1565 #else
1566         unsigned long msecs = o2net_idle_timeout();
1567 #endif
1568
1569         printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1570                "idle for %lu.%lu secs, shutting it down.\n", SC_NODEF_ARGS(sc),
1571                msecs / 1000, msecs % 1000);
1572
1573         /*
1574          * Initialize the nn_timeout so that the next connection attempt
1575          * will continue in o2net_start_connect.
1576          */
1577         atomic_set(&nn->nn_timeout, 1);
1578
1579         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1580 }
1581
1582 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1583 {
1584         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1585         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1586                       msecs_to_jiffies(o2net_keepalive_delay()));
1587         o2net_set_sock_timer(sc);
1588         mod_timer(&sc->sc_idle_timeout,
1589                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1590 }
1591
1592 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1593 {
1594         /* Only push out an existing timer */
1595         if (timer_pending(&sc->sc_idle_timeout))
1596                 o2net_sc_reset_idle_timer(sc);
1597 }
1598
1599 /* this work func is kicked whenever a path sets the nn state which doesn't
1600  * have valid set.  This includes seeing hb come up, losing a connection,
1601  * having a connect attempt fail, etc. This centralizes the logic which decides
1602  * if a connect attempt should be made or if we should give up and all future
1603  * transmit attempts should fail */
1604 static void o2net_start_connect(struct work_struct *work)
1605 {
1606         struct o2net_node *nn =
1607                 container_of(work, struct o2net_node, nn_connect_work.work);
1608         struct o2net_sock_container *sc = NULL;
1609         struct o2nm_node *node = NULL, *mynode = NULL;
1610         struct socket *sock = NULL;
1611         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1612         int ret = 0, stop;
1613         unsigned int timeout;
1614
1615         /* if we're greater we initiate tx, otherwise we accept */
1616         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1617                 goto out;
1618
1619         /* watch for racing with tearing a node down */
1620         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1621         if (node == NULL) {
1622                 ret = 0;
1623                 goto out;
1624         }
1625
1626         mynode = o2nm_get_node_by_num(o2nm_this_node());
1627         if (mynode == NULL) {
1628                 ret = 0;
1629                 goto out;
1630         }
1631
1632         spin_lock(&nn->nn_lock);
1633         /*
1634          * see if we already have one pending or have given up.
1635          * For nn_timeout, it is set when we close the connection
1636          * because of the idle time out. So it means that we have
1637          * at least connected to that node successfully once,
1638          * now try to connect to it again.
1639          */
1640         timeout = atomic_read(&nn->nn_timeout);
1641         stop = (nn->nn_sc ||
1642                 (nn->nn_persistent_error &&
1643                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1644         spin_unlock(&nn->nn_lock);
1645         if (stop)
1646                 goto out;
1647
1648         nn->nn_last_connect_attempt = jiffies;
1649
1650         sc = sc_alloc(node);
1651         if (sc == NULL) {
1652                 mlog(0, "couldn't allocate sc\n");
1653                 ret = -ENOMEM;
1654                 goto out;
1655         }
1656
1657         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1658         if (ret < 0) {
1659                 mlog(0, "can't create socket: %d\n", ret);
1660                 goto out;
1661         }
1662         sc->sc_sock = sock; /* freed by sc_kref_release */
1663
1664         sock->sk->sk_allocation = GFP_ATOMIC;
1665
1666         myaddr.sin_family = AF_INET;
1667         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1668         myaddr.sin_port = htons(0); /* any port */
1669
1670         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1671                               sizeof(myaddr));
1672         if (ret) {
1673                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1674                      ret, &mynode->nd_ipv4_address);
1675                 goto out;
1676         }
1677
1678         ret = o2net_set_nodelay(sc->sc_sock);
1679         if (ret) {
1680                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1681                 goto out;
1682         }
1683
1684         o2net_register_callbacks(sc->sc_sock->sk, sc);
1685
1686         spin_lock(&nn->nn_lock);
1687         /* handshake completion will set nn->nn_sc_valid */
1688         o2net_set_nn_state(nn, sc, 0, 0);
1689         spin_unlock(&nn->nn_lock);
1690
1691         remoteaddr.sin_family = AF_INET;
1692         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1693         remoteaddr.sin_port = node->nd_ipv4_port;
1694
1695         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1696                                         (struct sockaddr *)&remoteaddr,
1697                                         sizeof(remoteaddr),
1698                                         O_NONBLOCK);
1699         if (ret == -EINPROGRESS)
1700                 ret = 0;
1701
1702 out:
1703         if (ret) {
1704                 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1705                        " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1706                 /* 0 err so that another will be queued and attempted
1707                  * from set_nn_state */
1708                 if (sc)
1709                         o2net_ensure_shutdown(nn, sc, 0);
1710         }
1711         if (sc)
1712                 sc_put(sc);
1713         if (node)
1714                 o2nm_node_put(node);
1715         if (mynode)
1716                 o2nm_node_put(mynode);
1717
1718         return;
1719 }
1720
1721 static void o2net_connect_expired(struct work_struct *work)
1722 {
1723         struct o2net_node *nn =
1724                 container_of(work, struct o2net_node, nn_connect_expired.work);
1725
1726         spin_lock(&nn->nn_lock);
1727         if (!nn->nn_sc_valid) {
1728                 printk(KERN_NOTICE "o2net: No connection established with "
1729                        "node %u after %u.%u seconds, giving up.\n",
1730                      o2net_num_from_nn(nn),
1731                      o2net_idle_timeout() / 1000,
1732                      o2net_idle_timeout() % 1000);
1733
1734                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1735         }
1736         spin_unlock(&nn->nn_lock);
1737 }
1738
1739 static void o2net_still_up(struct work_struct *work)
1740 {
1741         struct o2net_node *nn =
1742                 container_of(work, struct o2net_node, nn_still_up.work);
1743
1744         o2quo_hb_still_up(o2net_num_from_nn(nn));
1745 }
1746
1747 /* ------------------------------------------------------------ */
1748
1749 void o2net_disconnect_node(struct o2nm_node *node)
1750 {
1751         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1752
1753         /* don't reconnect until it's heartbeating again */
1754         spin_lock(&nn->nn_lock);
1755         atomic_set(&nn->nn_timeout, 0);
1756         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1757         spin_unlock(&nn->nn_lock);
1758
1759         if (o2net_wq) {
1760                 cancel_delayed_work(&nn->nn_connect_expired);
1761                 cancel_delayed_work(&nn->nn_connect_work);
1762                 cancel_delayed_work(&nn->nn_still_up);
1763                 flush_workqueue(o2net_wq);
1764         }
1765 }
1766
1767 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1768                                   void *data)
1769 {
1770         o2quo_hb_down(node_num);
1771
1772         if (!node)
1773                 return;
1774
1775         if (node_num != o2nm_this_node())
1776                 o2net_disconnect_node(node);
1777
1778         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1779 }
1780
1781 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1782                                 void *data)
1783 {
1784         struct o2net_node *nn = o2net_nn_from_num(node_num);
1785
1786         o2quo_hb_up(node_num);
1787
1788         BUG_ON(!node);
1789
1790         /* ensure an immediate connect attempt */
1791         nn->nn_last_connect_attempt = jiffies -
1792                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1793
1794         if (node_num != o2nm_this_node()) {
1795                 /* believe it or not, accept and node hearbeating testing
1796                  * can succeed for this node before we got here.. so
1797                  * only use set_nn_state to clear the persistent error
1798                  * if that hasn't already happened */
1799                 spin_lock(&nn->nn_lock);
1800                 atomic_set(&nn->nn_timeout, 0);
1801                 if (nn->nn_persistent_error)
1802                         o2net_set_nn_state(nn, NULL, 0, 0);
1803                 spin_unlock(&nn->nn_lock);
1804         }
1805 }
1806
1807 void o2net_unregister_hb_callbacks(void)
1808 {
1809         o2hb_unregister_callback(NULL, &o2net_hb_up);
1810         o2hb_unregister_callback(NULL, &o2net_hb_down);
1811 }
1812
1813 int o2net_register_hb_callbacks(void)
1814 {
1815         int ret;
1816
1817         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1818                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1819         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1820                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1821
1822         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1823         if (ret == 0)
1824                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1825
1826         if (ret)
1827                 o2net_unregister_hb_callbacks();
1828
1829         return ret;
1830 }
1831
1832 /* ------------------------------------------------------------ */
1833
1834 static int o2net_accept_one(struct socket *sock)
1835 {
1836         int ret, slen;
1837         struct sockaddr_in sin;
1838         struct socket *new_sock = NULL;
1839         struct o2nm_node *node = NULL;
1840         struct o2nm_node *local_node = NULL;
1841         struct o2net_sock_container *sc = NULL;
1842         struct o2net_node *nn;
1843
1844         BUG_ON(sock == NULL);
1845         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1846                                sock->sk->sk_protocol, &new_sock);
1847         if (ret)
1848                 goto out;
1849
1850         new_sock->type = sock->type;
1851         new_sock->ops = sock->ops;
1852         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1853         if (ret < 0)
1854                 goto out;
1855
1856         new_sock->sk->sk_allocation = GFP_ATOMIC;
1857
1858         ret = o2net_set_nodelay(new_sock);
1859         if (ret) {
1860                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1861                 goto out;
1862         }
1863
1864         slen = sizeof(sin);
1865         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1866                                        &slen, 1);
1867         if (ret < 0)
1868                 goto out;
1869
1870         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1871         if (node == NULL) {
1872                 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1873                        "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1874                        ntohs(sin.sin_port));
1875                 ret = -EINVAL;
1876                 goto out;
1877         }
1878
1879         if (o2nm_this_node() >= node->nd_num) {
1880                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1881                 printk(KERN_NOTICE "o2net: Unexpected connect attempt seen "
1882                        "at node '%s' (%u, %pI4:%d) from node '%s' (%u, "
1883                        "%pI4:%d)\n", local_node->nd_name, local_node->nd_num,
1884                        &(local_node->nd_ipv4_address),
1885                        ntohs(local_node->nd_ipv4_port), node->nd_name,
1886                        node->nd_num, &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1887                 ret = -EINVAL;
1888                 goto out;
1889         }
1890
1891         /* this happens all the time when the other node sees our heartbeat
1892          * and tries to connect before we see their heartbeat */
1893         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1894                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1895                      "%pI4:%d but it isn't heartbeating\n",
1896                      node->nd_name, &sin.sin_addr.s_addr,
1897                      ntohs(sin.sin_port));
1898                 ret = -EINVAL;
1899                 goto out;
1900         }
1901
1902         nn = o2net_nn_from_num(node->nd_num);
1903
1904         spin_lock(&nn->nn_lock);
1905         if (nn->nn_sc)
1906                 ret = -EBUSY;
1907         else
1908                 ret = 0;
1909         spin_unlock(&nn->nn_lock);
1910         if (ret) {
1911                 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1912                        "at %pI4:%d but it already has an open connection\n",
1913                        node->nd_name, &sin.sin_addr.s_addr,
1914                        ntohs(sin.sin_port));
1915                 goto out;
1916         }
1917
1918         sc = sc_alloc(node);
1919         if (sc == NULL) {
1920                 ret = -ENOMEM;
1921                 goto out;
1922         }
1923
1924         sc->sc_sock = new_sock;
1925         new_sock = NULL;
1926
1927         spin_lock(&nn->nn_lock);
1928         atomic_set(&nn->nn_timeout, 0);
1929         o2net_set_nn_state(nn, sc, 0, 0);
1930         spin_unlock(&nn->nn_lock);
1931
1932         o2net_register_callbacks(sc->sc_sock->sk, sc);
1933         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1934
1935         o2net_initialize_handshake();
1936         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1937
1938 out:
1939         if (new_sock)
1940                 sock_release(new_sock);
1941         if (node)
1942                 o2nm_node_put(node);
1943         if (local_node)
1944                 o2nm_node_put(local_node);
1945         if (sc)
1946                 sc_put(sc);
1947         return ret;
1948 }
1949
1950 static void o2net_accept_many(struct work_struct *work)
1951 {
1952         struct socket *sock = o2net_listen_sock;
1953         while (o2net_accept_one(sock) == 0)
1954                 cond_resched();
1955 }
1956
1957 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1958 {
1959         void (*ready)(struct sock *sk, int bytes);
1960
1961         read_lock(&sk->sk_callback_lock);
1962         ready = sk->sk_user_data;
1963         if (ready == NULL) { /* check for teardown race */
1964                 ready = sk->sk_data_ready;
1965                 goto out;
1966         }
1967
1968         /* ->sk_data_ready is also called for a newly established child socket
1969          * before it has been accepted and the acceptor has set up their
1970          * data_ready.. we only want to queue listen work for our listening
1971          * socket */
1972         if (sk->sk_state == TCP_LISTEN) {
1973                 mlog(ML_TCP, "bytes: %d\n", bytes);
1974                 queue_work(o2net_wq, &o2net_listen_work);
1975         }
1976
1977 out:
1978         read_unlock(&sk->sk_callback_lock);
1979         ready(sk, bytes);
1980 }
1981
1982 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1983 {
1984         struct socket *sock = NULL;
1985         int ret;
1986         struct sockaddr_in sin = {
1987                 .sin_family = PF_INET,
1988                 .sin_addr = { .s_addr = addr },
1989                 .sin_port = port,
1990         };
1991
1992         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1993         if (ret < 0) {
1994                 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
1995                 goto out;
1996         }
1997
1998         sock->sk->sk_allocation = GFP_ATOMIC;
1999
2000         write_lock_bh(&sock->sk->sk_callback_lock);
2001         sock->sk->sk_user_data = sock->sk->sk_data_ready;
2002         sock->sk->sk_data_ready = o2net_listen_data_ready;
2003         write_unlock_bh(&sock->sk->sk_callback_lock);
2004
2005         o2net_listen_sock = sock;
2006         INIT_WORK(&o2net_listen_work, o2net_accept_many);
2007
2008         sock->sk->sk_reuse = 1;
2009         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2010         if (ret < 0) {
2011                 printk(KERN_ERR "o2net: Error %d while binding socket at "
2012                        "%pI4:%u\n", ret, &addr, ntohs(port));
2013                 goto out;
2014         }
2015
2016         ret = sock->ops->listen(sock, 64);
2017         if (ret < 0)
2018                 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2019                        ret, &addr, ntohs(port));
2020
2021 out:
2022         if (ret) {
2023                 o2net_listen_sock = NULL;
2024                 if (sock)
2025                         sock_release(sock);
2026         }
2027         return ret;
2028 }
2029
2030 /*
2031  * called from node manager when we should bring up our network listening
2032  * socket.  node manager handles all the serialization to only call this
2033  * once and to match it with o2net_stop_listening().  note,
2034  * o2nm_this_node() doesn't work yet as we're being called while it
2035  * is being set up.
2036  */
2037 int o2net_start_listening(struct o2nm_node *node)
2038 {
2039         int ret = 0;
2040
2041         BUG_ON(o2net_wq != NULL);
2042         BUG_ON(o2net_listen_sock != NULL);
2043
2044         mlog(ML_KTHREAD, "starting o2net thread...\n");
2045         o2net_wq = create_singlethread_workqueue("o2net");
2046         if (o2net_wq == NULL) {
2047                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2048                 return -ENOMEM; /* ? */
2049         }
2050
2051         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2052                                         node->nd_ipv4_port);
2053         if (ret) {
2054                 destroy_workqueue(o2net_wq);
2055                 o2net_wq = NULL;
2056         } else
2057                 o2quo_conn_up(node->nd_num);
2058
2059         return ret;
2060 }
2061
2062 /* again, o2nm_this_node() doesn't work here as we're involved in
2063  * tearing it down */
2064 void o2net_stop_listening(struct o2nm_node *node)
2065 {
2066         struct socket *sock = o2net_listen_sock;
2067         size_t i;
2068
2069         BUG_ON(o2net_wq == NULL);
2070         BUG_ON(o2net_listen_sock == NULL);
2071
2072         /* stop the listening socket from generating work */
2073         write_lock_bh(&sock->sk->sk_callback_lock);
2074         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2075         sock->sk->sk_user_data = NULL;
2076         write_unlock_bh(&sock->sk->sk_callback_lock);
2077
2078         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2079                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2080                 if (node) {
2081                         o2net_disconnect_node(node);
2082                         o2nm_node_put(node);
2083                 }
2084         }
2085
2086         /* finish all work and tear down the work queue */
2087         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2088         destroy_workqueue(o2net_wq);
2089         o2net_wq = NULL;
2090
2091         sock_release(o2net_listen_sock);
2092         o2net_listen_sock = NULL;
2093
2094         o2quo_conn_err(node->nd_num);
2095 }
2096
2097 /* ------------------------------------------------------------ */
2098
2099 int o2net_init(void)
2100 {
2101         unsigned long i;
2102
2103         o2quo_init();
2104
2105         if (o2net_debugfs_init())
2106                 return -ENOMEM;
2107
2108         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2109         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2110         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2111         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2112                 kfree(o2net_hand);
2113                 kfree(o2net_keep_req);
2114                 kfree(o2net_keep_resp);
2115                 return -ENOMEM;
2116         }
2117
2118         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2119         o2net_hand->connector_id = cpu_to_be64(1);
2120
2121         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2122         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2123
2124         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2125                 struct o2net_node *nn = o2net_nn_from_num(i);
2126
2127                 atomic_set(&nn->nn_timeout, 0);
2128                 spin_lock_init(&nn->nn_lock);
2129                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2130                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2131                                   o2net_connect_expired);
2132                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2133                 /* until we see hb from a node we'll return einval */
2134                 nn->nn_persistent_error = -ENOTCONN;
2135                 init_waitqueue_head(&nn->nn_sc_wq);
2136                 idr_init(&nn->nn_status_idr);
2137                 INIT_LIST_HEAD(&nn->nn_status_list);
2138         }
2139
2140         return 0;
2141 }
2142
2143 void o2net_exit(void)
2144 {
2145         o2quo_exit();
2146         kfree(o2net_hand);
2147         kfree(o2net_keep_req);
2148         kfree(o2net_keep_resp);
2149         o2net_debugfs_exit();
2150 }