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);
 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 unsigned int o2net_reconnect_delay(void)
 266 {
 267         return o2nm_single_cluster->cl_reconnect_delay_ms;
 268 }
 269
 270 static inline unsigned int o2net_keepalive_delay(void)
 271 {
 272         return o2nm_single_cluster->cl_keepalive_delay_ms;
 273 }
 274
 275 static inline unsigned 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;
 308
 309         spin_lock(&nn->nn_lock);
 310         ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
 311         if (ret >= 0) {
 312                 nsw->ns_id = ret;
 313                 list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
 314         }
 315         spin_unlock(&nn->nn_lock);
 316         if (ret < 0)
 317                 return ret;
 318
 319         init_waitqueue_head(&nsw->ns_wq);
 320         nsw->ns_sys_status = O2NET_ERR_NONE;
 321         nsw->ns_status = 0;
 322         return 0;
 323 }
 324
 325 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 326                                       struct o2net_status_wait *nsw,
 327                                       enum o2net_system_error sys_status,
 328                                       s32 status)
 329 {
 330         assert_spin_locked(&nn->nn_lock);
 331
 332         if (!list_empty(&nsw->ns_node_item)) {
 333                 list_del_init(&nsw->ns_node_item);
 334                 nsw->ns_sys_status = sys_status;
 335                 nsw->ns_status = status;
 336                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 337                 wake_up(&nsw->ns_wq);
 338         }
 339 }
 340
 341 static void o2net_complete_nsw(struct o2net_node *nn,
 342                                struct o2net_status_wait *nsw,
 343                                u64 id, enum o2net_system_error sys_status,
 344                                s32 status)
 345 {
 346         spin_lock(&nn->nn_lock);
 347         if (nsw == NULL) {
 348                 if (id > INT_MAX)
 349                         goto out;
 350
 351                 nsw = idr_find(&nn->nn_status_idr, id);
 352                 if (nsw == NULL)
 353                         goto out;
 354         }
 355
 356         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 357
 358 out:
 359         spin_unlock(&nn->nn_lock);
 360         return;
 361 }
 362
 363 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 364 {
 365         struct o2net_status_wait *nsw, *tmp;
 366         unsigned int num_kills = 0;
 367
 368         assert_spin_locked(&nn->nn_lock);
 369
 370         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 371                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 372                 num_kills++;
 373         }
 374
 375         mlog(0, "completed %d messages for node %u\n", num_kills,
 376              o2net_num_from_nn(nn));
 377 }
 378
 379 static int o2net_nsw_completed(struct o2net_node *nn,
 380                                struct o2net_status_wait *nsw)
 381 {
 382         int completed;
 383         spin_lock(&nn->nn_lock);
 384         completed = list_empty(&nsw->ns_node_item);
 385         spin_unlock(&nn->nn_lock);
 386         return completed;
 387 }
 388
 389 /* ------------------------------------------------------------ */
 390
 391 static void sc_kref_release(struct kref *kref)
 392 {
 393         struct o2net_sock_container *sc = container_of(kref,
 394                                         struct o2net_sock_container, sc_kref);
 395         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 396
 397         sclog(sc, "releasing\n");
 398
 399         if (sc->sc_sock) {
 400                 sock_release(sc->sc_sock);
 401                 sc->sc_sock = NULL;
 402         }
 403
 404         o2nm_undepend_item(&sc->sc_node->nd_item);
 405         o2nm_node_put(sc->sc_node);
 406         sc->sc_node = NULL;
 407
 408         o2net_debug_del_sc(sc);
 409
 410         if (sc->sc_page)
 411                 __free_page(sc->sc_page);
 412         kfree(sc);
 413 }
 414
 415 static void sc_put(struct o2net_sock_container *sc)
 416 {
 417         sclog(sc, "put\n");
 418         kref_put(&sc->sc_kref, sc_kref_release);
 419 }
 420 static void sc_get(struct o2net_sock_container *sc)
 421 {
 422         sclog(sc, "get\n");
 423         kref_get(&sc->sc_kref);
 424 }
 425 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 426 {
 427         struct o2net_sock_container *sc, *ret = NULL;
 428         struct page *page = NULL;
 429         int status = 0;
 430
 431         page = alloc_page(GFP_NOFS);
 432         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 433         if (sc == NULL || page == NULL)
 434                 goto out;
 435
 436         kref_init(&sc->sc_kref);
 437         o2nm_node_get(node);
 438         sc->sc_node = node;
 439
 440         /* pin the node item of the remote node */
 441         status = o2nm_depend_item(&node->nd_item);
 442         if (status) {
 443                 mlog_errno(status);
 444                 o2nm_node_put(node);
 445                 goto out;
 446         }
 447         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 448         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 449         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 450         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 451
 452         init_timer(&sc->sc_idle_timeout);
 453         sc->sc_idle_timeout.function = o2net_idle_timer;
 454         sc->sc_idle_timeout.data = (unsigned long)sc;
 455
 456         sclog(sc, "alloced\n");
 457
 458         ret = sc;
 459         sc->sc_page = page;
 460         o2net_debug_add_sc(sc);
 461         sc = NULL;
 462         page = NULL;
 463
 464 out:
 465         if (page)
 466                 __free_page(page);
 467         kfree(sc);
 468
 469         return ret;
 470 }
 471
 472 /* ------------------------------------------------------------ */
 473
 474 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 475                                 struct work_struct *work)
 476 {
 477         sc_get(sc);
 478         if (!queue_work(o2net_wq, work))
 479                 sc_put(sc);
 480 }
 481 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 482                                         struct delayed_work *work,
 483                                         int delay)
 484 {
 485         sc_get(sc);
 486         if (!queue_delayed_work(o2net_wq, work, delay))
 487                 sc_put(sc);
 488 }
 489 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 490                                          struct delayed_work *work)
 491 {
 492         if (cancel_delayed_work(work))
 493                 sc_put(sc);
 494 }
 495
 496 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 497
 498 int o2net_num_connected_peers(void)
 499 {
 500         return atomic_read(&o2net_connected_peers);
 501 }
 502
 503 static void o2net_set_nn_state(struct o2net_node *nn,
 504                                struct o2net_sock_container *sc,
 505                                unsigned valid, int err)
 506 {
 507         int was_valid = nn->nn_sc_valid;
 508         int was_err = nn->nn_persistent_error;
 509         struct o2net_sock_container *old_sc = nn->nn_sc;
 510
 511         assert_spin_locked(&nn->nn_lock);
 512
 513         if (old_sc && !sc)
 514                 atomic_dec(&o2net_connected_peers);
 515         else if (!old_sc && sc)
 516                 atomic_inc(&o2net_connected_peers);
 517
 518         /* the node num comparison and single connect/accept path should stop
 519          * an non-null sc from being overwritten with another */
 520         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 521         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 522         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 523
 524         if (was_valid && !valid && err == 0)
 525                 err = -ENOTCONN;
 526
 527         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 528              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 529              nn->nn_persistent_error, err);
 530
 531         nn->nn_sc = sc;
 532         nn->nn_sc_valid = valid ? 1 : 0;
 533         nn->nn_persistent_error = err;
 534
 535         /* mirrors o2net_tx_can_proceed() */
 536         if (nn->nn_persistent_error || nn->nn_sc_valid)
 537                 wake_up(&nn->nn_sc_wq);
 538
 539         if (!was_err && nn->nn_persistent_error) {
 540                 o2quo_conn_err(o2net_num_from_nn(nn));
 541                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 542                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 543         }
 544
 545         if (was_valid && !valid) {
 546                 if (old_sc)
 547                         printk(KERN_NOTICE "o2net: No longer connected to "
 548                                 SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 549                 o2net_complete_nodes_nsw(nn);
 550         }
 551
 552         if (!was_valid && valid) {
 553                 o2quo_conn_up(o2net_num_from_nn(nn));
 554                 cancel_delayed_work(&nn->nn_connect_expired);
 555                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 556                        o2nm_this_node() > sc->sc_node->nd_num ?
 557                        "Connected to" : "Accepted connection from",
 558                        SC_NODEF_ARGS(sc));
 559         }
 560
 561         /* trigger the connecting worker func as long as we're not valid,
 562          * it will back off if it shouldn't connect.  This can be called
 563          * from node config teardown and so needs to be careful about
 564          * the work queue actually being up. */
 565         if (!valid && o2net_wq) {
 566                 unsigned long delay;
 567                 /* delay if we're within a RECONNECT_DELAY of the
 568                  * last attempt */
 569                 delay = (nn->nn_last_connect_attempt +
 570                          msecs_to_jiffies(o2net_reconnect_delay()))
 571                         - jiffies;
 572                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 573                         delay = 0;
 574                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 575                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 576
 577                 /*
 578                  * Delay the expired work after idle timeout.
 579                  *
 580                  * We might have lots of failed connection attempts that run
 581                  * through here but we only cancel the connect_expired work when
 582                  * a connection attempt succeeds.  So only the first enqueue of
 583                  * the connect_expired work will do anything.  The rest will see
 584                  * that it's already queued and do nothing.
 585                  */
 586                 delay += msecs_to_jiffies(o2net_idle_timeout());
 587                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 588         }
 589
 590         /* keep track of the nn's sc ref for the caller */
 591         if ((old_sc == NULL) && sc)
 592                 sc_get(sc);
 593         if (old_sc && (old_sc != sc)) {
 594                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 595                 sc_put(old_sc);
 596         }
 597 }
 598
 599 /* see o2net_register_callbacks() */
 600 static void o2net_data_ready(struct sock *sk)
 601 {
 602         void (*ready)(struct sock *sk);
 603
 604         read_lock(&sk->sk_callback_lock);
 605         if (sk->sk_user_data) {
 606                 struct o2net_sock_container *sc = sk->sk_user_data;
 607                 sclog(sc, "data_ready hit\n");
 608                 o2net_set_data_ready_time(sc);
 609                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 610                 ready = sc->sc_data_ready;
 611         } else {
 612                 ready = sk->sk_data_ready;
 613         }
 614         read_unlock(&sk->sk_callback_lock);
 615
 616         ready(sk);
 617 }
 618
 619 /* see o2net_register_callbacks() */
 620 static void o2net_state_change(struct sock *sk)
 621 {
 622         void (*state_change)(struct sock *sk);
 623         struct o2net_sock_container *sc;
 624
 625         read_lock(&sk->sk_callback_lock);
 626         sc = sk->sk_user_data;
 627         if (sc == NULL) {
 628                 state_change = sk->sk_state_change;
 629                 goto out;
 630         }
 631
 632         sclog(sc, "state_change to %d\n", sk->sk_state);
 633
 634         state_change = sc->sc_state_change;
 635
 636         switch(sk->sk_state) {
 637         /* ignore connecting sockets as they make progress */
 638         case TCP_SYN_SENT:
 639         case TCP_SYN_RECV:
 640                 break;
 641         case TCP_ESTABLISHED:
 642                 o2net_sc_queue_work(sc, &sc->sc_connect_work);
 643                 break;
 644         default:
 645                 printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
 646                         " shutdown, state %d\n",
 647                         SC_NODEF_ARGS(sc), sk->sk_state);
 648                 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 649                 break;
 650         }
 651 out:
 652         read_unlock(&sk->sk_callback_lock);
 653         state_change(sk);
 654 }
 655
 656 /*
 657  * we register callbacks so we can queue work on events before calling
 658  * the original callbacks.  our callbacks our careful to test user_data
 659  * to discover when they've reaced with o2net_unregister_callbacks().
 660  */
 661 static void o2net_register_callbacks(struct sock *sk,
 662                                      struct o2net_sock_container *sc)
 663 {
 664         write_lock_bh(&sk->sk_callback_lock);
 665
 666         /* accepted sockets inherit the old listen socket data ready */
 667         if (sk->sk_data_ready == o2net_listen_data_ready) {
 668                 sk->sk_data_ready = sk->sk_user_data;
 669                 sk->sk_user_data = NULL;
 670         }
 671
 672         BUG_ON(sk->sk_user_data != NULL);
 673         sk->sk_user_data = sc;
 674         sc_get(sc);
 675
 676         sc->sc_data_ready = sk->sk_data_ready;
 677         sc->sc_state_change = sk->sk_state_change;
 678         sk->sk_data_ready = o2net_data_ready;
 679         sk->sk_state_change = o2net_state_change;
 680
 681         mutex_init(&sc->sc_send_lock);
 682
 683         write_unlock_bh(&sk->sk_callback_lock);
 684 }
 685
 686 static int o2net_unregister_callbacks(struct sock *sk,
 687                                    struct o2net_sock_container *sc)
 688 {
 689         int ret = 0;
 690
 691         write_lock_bh(&sk->sk_callback_lock);
 692         if (sk->sk_user_data == sc) {
 693                 ret = 1;
 694                 sk->sk_user_data = NULL;
 695                 sk->sk_data_ready = sc->sc_data_ready;
 696                 sk->sk_state_change = sc->sc_state_change;
 697         }
 698         write_unlock_bh(&sk->sk_callback_lock);
 699
 700         return ret;
 701 }
 702
 703 /*
 704  * this is a little helper that is called by callers who have seen a problem
 705  * with an sc and want to detach it from the nn if someone already hasn't beat
 706  * them to it.  if an error is given then the shutdown will be persistent
 707  * and pending transmits will be canceled.
 708  */
 709 static void o2net_ensure_shutdown(struct o2net_node *nn,
 710                                    struct o2net_sock_container *sc,
 711                                    int err)
 712 {
 713         spin_lock(&nn->nn_lock);
 714         if (nn->nn_sc == sc)
 715                 o2net_set_nn_state(nn, NULL, 0, err);
 716         spin_unlock(&nn->nn_lock);
 717 }
 718
 719 /*
 720  * This work queue function performs the blocking parts of socket shutdown.  A
 721  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 722  * sc detached from the nn.  state_change will also trigger this callback
 723  * directly when it sees errors.  In that case we need to call set_nn_state
 724  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 725  * itself.
 726  */
 727 static void o2net_shutdown_sc(struct work_struct *work)
 728 {
 729         struct o2net_sock_container *sc =
 730                 container_of(work, struct o2net_sock_container,
 731                              sc_shutdown_work);
 732         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 733
 734         sclog(sc, "shutting down\n");
 735
 736         /* drop the callbacks ref and call shutdown only once */
 737         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 738                 /* we shouldn't flush as we're in the thread, the
 739                  * races with pending sc work structs are harmless */
 740                 del_timer_sync(&sc->sc_idle_timeout);
 741                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 742                 sc_put(sc);
 743                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 744         }
 745
 746         /* not fatal so failed connects before the other guy has our
 747          * heartbeat can be retried */
 748         o2net_ensure_shutdown(nn, sc, 0);
 749         sc_put(sc);
 750 }
 751
 752 /* ------------------------------------------------------------ */
 753
 754 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 755                              u32 key)
 756 {
 757         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 758
 759         if (ret == 0)
 760                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 761
 762         return ret;
 763 }
 764
 765 static struct o2net_msg_handler *
 766 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 767                           struct rb_node **ret_parent)
 768 {
 769         struct rb_node **p = &o2net_handler_tree.rb_node;
 770         struct rb_node *parent = NULL;
 771         struct o2net_msg_handler *nmh, *ret = NULL;
 772         int cmp;
 773
 774         while (*p) {
 775                 parent = *p;
 776                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 777                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 778
 779                 if (cmp < 0)
 780                         p = &(*p)->rb_left;
 781                 else if (cmp > 0)
 782                         p = &(*p)->rb_right;
 783                 else {
 784                         ret = nmh;
 785                         break;
 786                 }
 787         }
 788
 789         if (ret_p != NULL)
 790                 *ret_p = p;
 791         if (ret_parent != NULL)
 792                 *ret_parent = parent;
 793
 794         return ret;
 795 }
 796
 797 static void o2net_handler_kref_release(struct kref *kref)
 798 {
 799         struct o2net_msg_handler *nmh;
 800         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 801
 802         kfree(nmh);
 803 }
 804
 805 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 806 {
 807         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 808 }
 809
 810 /* max_len is protection for the handler func.  incoming messages won't
 811  * be given to the handler if their payload is longer than the max. */
 812 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 813                            o2net_msg_handler_func *func, void *data,
 814                            o2net_post_msg_handler_func *post_func,
 815                            struct list_head *unreg_list)
 816 {
 817         struct o2net_msg_handler *nmh = NULL;
 818         struct rb_node **p, *parent;
 819         int ret = 0;
 820
 821         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 822                 mlog(0, "max_len for message handler out of range: %u\n",
 823                         max_len);
 824                 ret = -EINVAL;
 825                 goto out;
 826         }
 827
 828         if (!msg_type) {
 829                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 830                 ret = -EINVAL;
 831                 goto out;
 832
 833         }
 834         if (!func) {
 835                 mlog(0, "no message handler provided: %u, %p\n",
 836                        msg_type, func);
 837                 ret = -EINVAL;
 838                 goto out;
 839         }
 840
 841         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 842         if (nmh == NULL) {
 843                 ret = -ENOMEM;
 844                 goto out;
 845         }
 846
 847         nmh->nh_func = func;
 848         nmh->nh_func_data = data;
 849         nmh->nh_post_func = post_func;
 850         nmh->nh_msg_type = msg_type;
 851         nmh->nh_max_len = max_len;
 852         nmh->nh_key = key;
 853         /* the tree and list get this ref.. they're both removed in
 854          * unregister when this ref is dropped */
 855         kref_init(&nmh->nh_kref);
 856         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 857
 858         write_lock(&o2net_handler_lock);
 859         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 860                 ret = -EEXIST;
 861         else {
 862                 rb_link_node(&nmh->nh_node, parent, p);
 863                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 864                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 865
 866                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 867                      func, msg_type, key);
 868                 /* we've had some trouble with handlers seemingly vanishing. */
 869                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 870                                                           &parent) == NULL,
 871                                 "couldn't find handler we *just* registered "
 872                                 "for type %u key %08x\n", msg_type, key);
 873         }
 874         write_unlock(&o2net_handler_lock);
 875         if (ret)
 876                 goto out;
 877
 878 out:
 879         if (ret)
 880                 kfree(nmh);
 881
 882         return ret;
 883 }
 884 EXPORT_SYMBOL_GPL(o2net_register_handler);
 885
 886 void o2net_unregister_handler_list(struct list_head *list)
 887 {
 888         struct o2net_msg_handler *nmh, *n;
 889
 890         write_lock(&o2net_handler_lock);
 891         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 892                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 893                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 894                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 895                 list_del_init(&nmh->nh_unregister_item);
 896                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 897         }
 898         write_unlock(&o2net_handler_lock);
 899 }
 900 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 901
 902 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 903 {
 904         struct o2net_msg_handler *nmh;
 905
 906         read_lock(&o2net_handler_lock);
 907         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 908         if (nmh)
 909                 kref_get(&nmh->nh_kref);
 910         read_unlock(&o2net_handler_lock);
 911
 912         return nmh;
 913 }
 914
 915 /* ------------------------------------------------------------ */
 916
 917 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 918 {
 919         int ret;
 920         mm_segment_t oldfs;
 921         struct kvec vec = {
 922                 .iov_len = len,
 923                 .iov_base = data,
 924         };
 925         struct msghdr msg = {
 926                 .msg_iovlen = 1,
 927                 .msg_iov = (struct iovec *)&vec,
 928                 .msg_flags = MSG_DONTWAIT,
 929         };
 930
 931         oldfs = get_fs();
 932         set_fs(get_ds());
 933         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
 934         set_fs(oldfs);
 935
 936         return ret;
 937 }
 938
 939 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 940                               size_t veclen, size_t total)
 941 {
 942         int ret;
 943         mm_segment_t oldfs;
 944         struct msghdr msg = {
 945                 .msg_iov = (struct iovec *)vec,
 946                 .msg_iovlen = veclen,
 947         };
 948
 949         if (sock == NULL) {
 950                 ret = -EINVAL;
 951                 goto out;
 952         }
 953
 954         oldfs = get_fs();
 955         set_fs(get_ds());
 956         ret = sock_sendmsg(sock, &msg, total);
 957         set_fs(oldfs);
 958         if (ret != total) {
 959                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
 960                      total);
 961                 if (ret >= 0)
 962                         ret = -EPIPE; /* should be smarter, I bet */
 963                 goto out;
 964         }
 965
 966         ret = 0;
 967 out:
 968         if (ret < 0)
 969                 mlog(0, "returning error: %d\n", ret);
 970         return ret;
 971 }
 972
 973 static void o2net_sendpage(struct o2net_sock_container *sc,
 974                            void *kmalloced_virt,
 975                            size_t size)
 976 {
 977         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 978         ssize_t ret;
 979
 980         while (1) {
 981                 mutex_lock(&sc->sc_send_lock);
 982                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 983                                                  virt_to_page(kmalloced_virt),
 984                                                  (long)kmalloced_virt & ~PAGE_MASK,
 985                                                  size, MSG_DONTWAIT);
 986                 mutex_unlock(&sc->sc_send_lock);
 987                 if (ret == size)
 988                         break;
 989                 if (ret == (ssize_t)-EAGAIN) {
 990                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 991                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 992                         cond_resched();
 993                         continue;
 994                 }
 995                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 996                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 997                 o2net_ensure_shutdown(nn, sc, 0);
 998                 break;
 999         }
1000 }
1001
1002 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
1003 {
1004         memset(msg, 0, sizeof(struct o2net_msg));
1005         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
1006         msg->data_len = cpu_to_be16(data_len);
1007         msg->msg_type = cpu_to_be16(msg_type);
1008         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
1009         msg->status = 0;
1010         msg->key = cpu_to_be32(key);
1011 }
1012
1013 static int o2net_tx_can_proceed(struct o2net_node *nn,
1014                                 struct o2net_sock_container **sc_ret,
1015                                 int *error)
1016 {
1017         int ret = 0;
1018
1019         spin_lock(&nn->nn_lock);
1020         if (nn->nn_persistent_error) {
1021                 ret = 1;
1022                 *sc_ret = NULL;
1023                 *error = nn->nn_persistent_error;
1024         } else if (nn->nn_sc_valid) {
1025                 kref_get(&nn->nn_sc->sc_kref);
1026
1027                 ret = 1;
1028                 *sc_ret = nn->nn_sc;
1029                 *error = 0;
1030         }
1031         spin_unlock(&nn->nn_lock);
1032
1033         return ret;
1034 }
1035
1036 /* Get a map of all nodes to which this node is currently connected to */
1037 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1038 {
1039         struct o2net_sock_container *sc;
1040         int node, ret;
1041
1042         BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1043
1044         memset(map, 0, bytes);
1045         for (node = 0; node < O2NM_MAX_NODES; ++node) {
1046                 o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret);
1047                 if (!ret) {
1048                         set_bit(node, map);
1049                         sc_put(sc);
1050                 }
1051         }
1052 }
1053 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1054
1055 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1056                            size_t caller_veclen, u8 target_node, int *status)
1057 {
1058         int ret = 0;
1059         struct o2net_msg *msg = NULL;
1060         size_t veclen, caller_bytes = 0;
1061         struct kvec *vec = NULL;
1062         struct o2net_sock_container *sc = NULL;
1063         struct o2net_node *nn = o2net_nn_from_num(target_node);
1064         struct o2net_status_wait nsw = {
1065                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1066         };
1067         struct o2net_send_tracking nst;
1068
1069         o2net_init_nst(&nst, msg_type, key, current, target_node);
1070
1071         if (o2net_wq == NULL) {
1072                 mlog(0, "attempt to tx without o2netd running\n");
1073                 ret = -ESRCH;
1074                 goto out;
1075         }
1076
1077         if (caller_veclen == 0) {
1078                 mlog(0, "bad kvec array length\n");
1079                 ret = -EINVAL;
1080                 goto out;
1081         }
1082
1083         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1084         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1085                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1086                 ret = -EINVAL;
1087                 goto out;
1088         }
1089
1090         if (target_node == o2nm_this_node()) {
1091                 ret = -ELOOP;
1092                 goto out;
1093         }
1094
1095         o2net_debug_add_nst(&nst);
1096
1097         o2net_set_nst_sock_time(&nst);
1098
1099         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1100         if (ret)
1101                 goto out;
1102
1103         o2net_set_nst_sock_container(&nst, sc);
1104
1105         veclen = caller_veclen + 1;
1106         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1107         if (vec == NULL) {
1108                 mlog(0, "failed to %zu element kvec!\n", veclen);
1109                 ret = -ENOMEM;
1110                 goto out;
1111         }
1112
1113         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1114         if (!msg) {
1115                 mlog(0, "failed to allocate a o2net_msg!\n");
1116                 ret = -ENOMEM;
1117                 goto out;
1118         }
1119
1120         o2net_init_msg(msg, caller_bytes, msg_type, key);
1121
1122         vec[0].iov_len = sizeof(struct o2net_msg);
1123         vec[0].iov_base = msg;
1124         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1125
1126         ret = o2net_prep_nsw(nn, &nsw);
1127         if (ret)
1128                 goto out;
1129
1130         msg->msg_num = cpu_to_be32(nsw.ns_id);
1131         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1132
1133         o2net_set_nst_send_time(&nst);
1134
1135         /* finally, convert the message header to network byte-order
1136          * and send */
1137         mutex_lock(&sc->sc_send_lock);
1138         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1139                                  sizeof(struct o2net_msg) + caller_bytes);
1140         mutex_unlock(&sc->sc_send_lock);
1141         msglog(msg, "sending returned %d\n", ret);
1142         if (ret < 0) {
1143                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1144                 goto out;
1145         }
1146
1147         /* wait on other node's handler */
1148         o2net_set_nst_status_time(&nst);
1149         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1150
1151         o2net_update_send_stats(&nst, sc);
1152
1153         /* Note that we avoid overwriting the callers status return
1154          * variable if a system error was reported on the other
1155          * side. Callers beware. */
1156         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1157         if (status && !ret)
1158                 *status = nsw.ns_status;
1159
1160         mlog(0, "woken, returning system status %d, user status %d\n",
1161              ret, nsw.ns_status);
1162 out:
1163         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1164         if (sc)
1165                 sc_put(sc);
1166         kfree(vec);
1167         kfree(msg);
1168         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1169         return ret;
1170 }
1171 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1172
1173 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1174                        u8 target_node, int *status)
1175 {
1176         struct kvec vec = {
1177                 .iov_base = data,
1178                 .iov_len = len,
1179         };
1180         return o2net_send_message_vec(msg_type, key, &vec, 1,
1181                                       target_node, status);
1182 }
1183 EXPORT_SYMBOL_GPL(o2net_send_message);
1184
1185 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1186                                    enum o2net_system_error syserr, int err)
1187 {
1188         struct kvec vec = {
1189                 .iov_base = hdr,
1190                 .iov_len = sizeof(struct o2net_msg),
1191         };
1192
1193         BUG_ON(syserr >= O2NET_ERR_MAX);
1194
1195         /* leave other fields intact from the incoming message, msg_num
1196          * in particular */
1197         hdr->sys_status = cpu_to_be32(syserr);
1198         hdr->status = cpu_to_be32(err);
1199         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1200         hdr->data_len = 0;
1201
1202         msglog(hdr, "about to send status magic %d\n", err);
1203         /* hdr has been in host byteorder this whole time */
1204         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1205 }
1206
1207 /* this returns -errno if the header was unknown or too large, etc.
1208  * after this is called the buffer us reused for the next message */
1209 static int o2net_process_message(struct o2net_sock_container *sc,
1210                                  struct o2net_msg *hdr)
1211 {
1212         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1213         int ret = 0, handler_status;
1214         enum  o2net_system_error syserr;
1215         struct o2net_msg_handler *nmh = NULL;
1216         void *ret_data = NULL;
1217
1218         msglog(hdr, "processing message\n");
1219
1220         o2net_sc_postpone_idle(sc);
1221
1222         switch(be16_to_cpu(hdr->magic)) {
1223                 case O2NET_MSG_STATUS_MAGIC:
1224                         /* special type for returning message status */
1225                         o2net_complete_nsw(nn, NULL,
1226                                            be32_to_cpu(hdr->msg_num),
1227                                            be32_to_cpu(hdr->sys_status),
1228                                            be32_to_cpu(hdr->status));
1229                         goto out;
1230                 case O2NET_MSG_KEEP_REQ_MAGIC:
1231                         o2net_sendpage(sc, o2net_keep_resp,
1232                                        sizeof(*o2net_keep_resp));
1233                         goto out;
1234                 case O2NET_MSG_KEEP_RESP_MAGIC:
1235                         goto out;
1236                 case O2NET_MSG_MAGIC:
1237                         break;
1238                 default:
1239                         msglog(hdr, "bad magic\n");
1240                         ret = -EINVAL;
1241                         goto out;
1242                         break;
1243         }
1244
1245         /* find a handler for it */
1246         handler_status = 0;
1247         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1248                                 be32_to_cpu(hdr->key));
1249         if (!nmh) {
1250                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1251                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1252                 syserr = O2NET_ERR_NO_HNDLR;
1253                 goto out_respond;
1254         }
1255
1256         syserr = O2NET_ERR_NONE;
1257
1258         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1259                 syserr = O2NET_ERR_OVERFLOW;
1260
1261         if (syserr != O2NET_ERR_NONE)
1262                 goto out_respond;
1263
1264         o2net_set_func_start_time(sc);
1265         sc->sc_msg_key = be32_to_cpu(hdr->key);
1266         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1267         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1268                                              be16_to_cpu(hdr->data_len),
1269                                         nmh->nh_func_data, &ret_data);
1270         o2net_set_func_stop_time(sc);
1271
1272         o2net_update_recv_stats(sc);
1273
1274 out_respond:
1275         /* this destroys the hdr, so don't use it after this */
1276         mutex_lock(&sc->sc_send_lock);
1277         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1278                                       handler_status);
1279         mutex_unlock(&sc->sc_send_lock);
1280         hdr = NULL;
1281         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1282              handler_status, syserr, ret);
1283
1284         if (nmh) {
1285                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1286                 if (nmh->nh_post_func)
1287                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1288                                             ret_data);
1289         }
1290
1291 out:
1292         if (nmh)
1293                 o2net_handler_put(nmh);
1294         return ret;
1295 }
1296
1297 static int o2net_check_handshake(struct o2net_sock_container *sc)
1298 {
1299         struct o2net_handshake *hand = page_address(sc->sc_page);
1300         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1301
1302         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1303                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1304                        "protocol version %llu but %llu is required. "
1305                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1306                        (unsigned long long)be64_to_cpu(hand->protocol_version),
1307                        O2NET_PROTOCOL_VERSION);
1308
1309                 /* don't bother reconnecting if its the wrong version. */
1310                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1311                 return -1;
1312         }
1313
1314         /*
1315          * Ensure timeouts are consistent with other nodes, otherwise
1316          * we can end up with one node thinking that the other must be down,
1317          * but isn't. This can ultimately cause corruption.
1318          */
1319         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1320                                 o2net_idle_timeout()) {
1321                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1322                        "idle timeout of %u ms, but we use %u ms locally. "
1323                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1324                        be32_to_cpu(hand->o2net_idle_timeout_ms),
1325                        o2net_idle_timeout());
1326                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1327                 return -1;
1328         }
1329
1330         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1331                         o2net_keepalive_delay()) {
1332                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1333                        "delay of %u ms, but we use %u ms locally. "
1334                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1335                        be32_to_cpu(hand->o2net_keepalive_delay_ms),
1336                        o2net_keepalive_delay());
1337                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1338                 return -1;
1339         }
1340
1341         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1342                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1343                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1344                        "timeout of %u ms, but we use %u ms locally. "
1345                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1346                        be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1347                        O2HB_MAX_WRITE_TIMEOUT_MS);
1348                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1349                 return -1;
1350         }
1351
1352         sc->sc_handshake_ok = 1;
1353
1354         spin_lock(&nn->nn_lock);
1355         /* set valid and queue the idle timers only if it hasn't been
1356          * shut down already */
1357         if (nn->nn_sc == sc) {
1358                 o2net_sc_reset_idle_timer(sc);
1359                 atomic_set(&nn->nn_timeout, 0);
1360                 o2net_set_nn_state(nn, sc, 1, 0);
1361         }
1362         spin_unlock(&nn->nn_lock);
1363
1364         /* shift everything up as though it wasn't there */
1365         sc->sc_page_off -= sizeof(struct o2net_handshake);
1366         if (sc->sc_page_off)
1367                 memmove(hand, hand + 1, sc->sc_page_off);
1368
1369         return 0;
1370 }
1371
1372 /* this demuxes the queued rx bytes into header or payload bits and calls
1373  * handlers as each full message is read off the socket.  it returns -error,
1374  * == 0 eof, or > 0 for progress made.*/
1375 static int o2net_advance_rx(struct o2net_sock_container *sc)
1376 {
1377         struct o2net_msg *hdr;
1378         int ret = 0;
1379         void *data;
1380         size_t datalen;
1381
1382         sclog(sc, "receiving\n");
1383         o2net_set_advance_start_time(sc);
1384
1385         if (unlikely(sc->sc_handshake_ok == 0)) {
1386                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1387                         data = page_address(sc->sc_page) + sc->sc_page_off;
1388                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1389                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1390                         if (ret > 0)
1391                                 sc->sc_page_off += ret;
1392                 }
1393
1394                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1395                         o2net_check_handshake(sc);
1396                         if (unlikely(sc->sc_handshake_ok == 0))
1397                                 ret = -EPROTO;
1398                 }
1399                 goto out;
1400         }
1401
1402         /* do we need more header? */
1403         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1404                 data = page_address(sc->sc_page) + sc->sc_page_off;
1405                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1406                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1407                 if (ret > 0) {
1408                         sc->sc_page_off += ret;
1409                         /* only swab incoming here.. we can
1410                          * only get here once as we cross from
1411                          * being under to over */
1412                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1413                                 hdr = page_address(sc->sc_page);
1414                                 if (be16_to_cpu(hdr->data_len) >
1415                                     O2NET_MAX_PAYLOAD_BYTES)
1416                                         ret = -EOVERFLOW;
1417                         }
1418                 }
1419                 if (ret <= 0)
1420                         goto out;
1421         }
1422
1423         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1424                 /* oof, still don't have a header */
1425                 goto out;
1426         }
1427
1428         /* this was swabbed above when we first read it */
1429         hdr = page_address(sc->sc_page);
1430
1431         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1432
1433         /* do we need more payload? */
1434         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1435                 /* need more payload */
1436                 data = page_address(sc->sc_page) + sc->sc_page_off;
1437                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1438                           sc->sc_page_off;
1439                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1440                 if (ret > 0)
1441                         sc->sc_page_off += ret;
1442                 if (ret <= 0)
1443                         goto out;
1444         }
1445
1446         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1447                 /* we can only get here once, the first time we read
1448                  * the payload.. so set ret to progress if the handler
1449                  * works out. after calling this the message is toast */
1450                 ret = o2net_process_message(sc, hdr);
1451                 if (ret == 0)
1452                         ret = 1;
1453                 sc->sc_page_off = 0;
1454         }
1455
1456 out:
1457         sclog(sc, "ret = %d\n", ret);
1458         o2net_set_advance_stop_time(sc);
1459         return ret;
1460 }
1461
1462 /* this work func is triggerd by data ready.  it reads until it can read no
1463  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1464  * our work the work struct will be marked and we'll be called again. */
1465 static void o2net_rx_until_empty(struct work_struct *work)
1466 {
1467         struct o2net_sock_container *sc =
1468                 container_of(work, struct o2net_sock_container, sc_rx_work);
1469         int ret;
1470
1471         do {
1472                 ret = o2net_advance_rx(sc);
1473         } while (ret > 0);
1474
1475         if (ret <= 0 && ret != -EAGAIN) {
1476                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1477                 sclog(sc, "saw error %d, closing\n", ret);
1478                 /* not permanent so read failed handshake can retry */
1479                 o2net_ensure_shutdown(nn, sc, 0);
1480         }
1481
1482         sc_put(sc);
1483 }
1484
1485 static int o2net_set_nodelay(struct socket *sock)
1486 {
1487         int ret, val = 1;
1488         mm_segment_t oldfs;
1489
1490         oldfs = get_fs();
1491         set_fs(KERNEL_DS);
1492
1493         /*
1494          * Dear unsuspecting programmer,
1495          *
1496          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1497          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1498          * silently turn into SO_DEBUG.
1499          *
1500          * Yours,
1501          * Keeper of hilariously fragile interfaces.
1502          */
1503         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1504                                     (char __user *)&val, sizeof(val));
1505
1506         set_fs(oldfs);
1507         return ret;
1508 }
1509
1510 static void o2net_initialize_handshake(void)
1511 {
1512         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1513                 O2HB_MAX_WRITE_TIMEOUT_MS);
1514         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1515         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1516                 o2net_keepalive_delay());
1517         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1518                 o2net_reconnect_delay());
1519 }
1520
1521 /* ------------------------------------------------------------ */
1522
1523 /* called when a connect completes and after a sock is accepted.  the
1524  * rx path will see the response and mark the sc valid */
1525 static void o2net_sc_connect_completed(struct work_struct *work)
1526 {
1527         struct o2net_sock_container *sc =
1528                 container_of(work, struct o2net_sock_container,
1529                              sc_connect_work);
1530
1531         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1532               (unsigned long long)O2NET_PROTOCOL_VERSION,
1533               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1534
1535         o2net_initialize_handshake();
1536         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1537         sc_put(sc);
1538 }
1539
1540 /* this is called as a work_struct func. */
1541 static void o2net_sc_send_keep_req(struct work_struct *work)
1542 {
1543         struct o2net_sock_container *sc =
1544                 container_of(work, struct o2net_sock_container,
1545                              sc_keepalive_work.work);
1546
1547         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1548         sc_put(sc);
1549 }
1550
1551 /* socket shutdown does a del_timer_sync against this as it tears down.
1552  * we can't start this timer until we've got to the point in sc buildup
1553  * where shutdown is going to be involved */
1554 static void o2net_idle_timer(unsigned long data)
1555 {
1556         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1557         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1558 #ifdef CONFIG_DEBUG_FS
1559         unsigned long msecs = ktime_to_ms(ktime_get()) -
1560                 ktime_to_ms(sc->sc_tv_timer);
1561 #else
1562         unsigned long msecs = o2net_idle_timeout();
1563 #endif
1564
1565         printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1566                "idle for %lu.%lu secs, shutting it down.\n", SC_NODEF_ARGS(sc),
1567                msecs / 1000, msecs % 1000);
1568
1569         /*
1570          * Initialize the nn_timeout so that the next connection attempt
1571          * will continue in o2net_start_connect.
1572          */
1573         atomic_set(&nn->nn_timeout, 1);
1574
1575         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1576 }
1577
1578 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1579 {
1580         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1581         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1582                       msecs_to_jiffies(o2net_keepalive_delay()));
1583         o2net_set_sock_timer(sc);
1584         mod_timer(&sc->sc_idle_timeout,
1585                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1586 }
1587
1588 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1589 {
1590         /* Only push out an existing timer */
1591         if (timer_pending(&sc->sc_idle_timeout))
1592                 o2net_sc_reset_idle_timer(sc);
1593 }
1594
1595 /* this work func is kicked whenever a path sets the nn state which doesn't
1596  * have valid set.  This includes seeing hb come up, losing a connection,
1597  * having a connect attempt fail, etc. This centralizes the logic which decides
1598  * if a connect attempt should be made or if we should give up and all future
1599  * transmit attempts should fail */
1600 static void o2net_start_connect(struct work_struct *work)
1601 {
1602         struct o2net_node *nn =
1603                 container_of(work, struct o2net_node, nn_connect_work.work);
1604         struct o2net_sock_container *sc = NULL;
1605         struct o2nm_node *node = NULL, *mynode = NULL;
1606         struct socket *sock = NULL;
1607         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1608         int ret = 0, stop;
1609         unsigned int timeout;
1610
1611         /* if we're greater we initiate tx, otherwise we accept */
1612         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1613                 goto out;
1614
1615         /* watch for racing with tearing a node down */
1616         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1617         if (node == NULL) {
1618                 ret = 0;
1619                 goto out;
1620         }
1621
1622         mynode = o2nm_get_node_by_num(o2nm_this_node());
1623         if (mynode == NULL) {
1624                 ret = 0;
1625                 goto out;
1626         }
1627
1628         spin_lock(&nn->nn_lock);
1629         /*
1630          * see if we already have one pending or have given up.
1631          * For nn_timeout, it is set when we close the connection
1632          * because of the idle time out. So it means that we have
1633          * at least connected to that node successfully once,
1634          * now try to connect to it again.
1635          */
1636         timeout = atomic_read(&nn->nn_timeout);
1637         stop = (nn->nn_sc ||
1638                 (nn->nn_persistent_error &&
1639                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1640         spin_unlock(&nn->nn_lock);
1641         if (stop)
1642                 goto out;
1643
1644         nn->nn_last_connect_attempt = jiffies;
1645
1646         sc = sc_alloc(node);
1647         if (sc == NULL) {
1648                 mlog(0, "couldn't allocate sc\n");
1649                 ret = -ENOMEM;
1650                 goto out;
1651         }
1652
1653         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1654         if (ret < 0) {
1655                 mlog(0, "can't create socket: %d\n", ret);
1656                 goto out;
1657         }
1658         sc->sc_sock = sock; /* freed by sc_kref_release */
1659
1660         sock->sk->sk_allocation = GFP_ATOMIC;
1661
1662         myaddr.sin_family = AF_INET;
1663         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1664         myaddr.sin_port = htons(0); /* any port */
1665
1666         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1667                               sizeof(myaddr));
1668         if (ret) {
1669                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1670                      ret, &mynode->nd_ipv4_address);
1671                 goto out;
1672         }
1673
1674         ret = o2net_set_nodelay(sc->sc_sock);
1675         if (ret) {
1676                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1677                 goto out;
1678         }
1679
1680         o2net_register_callbacks(sc->sc_sock->sk, sc);
1681
1682         spin_lock(&nn->nn_lock);
1683         /* handshake completion will set nn->nn_sc_valid */
1684         o2net_set_nn_state(nn, sc, 0, 0);
1685         spin_unlock(&nn->nn_lock);
1686
1687         remoteaddr.sin_family = AF_INET;
1688         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1689         remoteaddr.sin_port = node->nd_ipv4_port;
1690
1691         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1692                                         (struct sockaddr *)&remoteaddr,
1693                                         sizeof(remoteaddr),
1694                                         O_NONBLOCK);
1695         if (ret == -EINPROGRESS)
1696                 ret = 0;
1697
1698 out:
1699         if (ret && sc) {
1700                 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1701                        " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1702                 /* 0 err so that another will be queued and attempted
1703                  * from set_nn_state */
1704                 o2net_ensure_shutdown(nn, sc, 0);
1705         }
1706         if (sc)
1707                 sc_put(sc);
1708         if (node)
1709                 o2nm_node_put(node);
1710         if (mynode)
1711                 o2nm_node_put(mynode);
1712
1713         return;
1714 }
1715
1716 static void o2net_connect_expired(struct work_struct *work)
1717 {
1718         struct o2net_node *nn =
1719                 container_of(work, struct o2net_node, nn_connect_expired.work);
1720
1721         spin_lock(&nn->nn_lock);
1722         if (!nn->nn_sc_valid) {
1723                 printk(KERN_NOTICE "o2net: No connection established with "
1724                        "node %u after %u.%u seconds, giving up.\n",
1725                      o2net_num_from_nn(nn),
1726                      o2net_idle_timeout() / 1000,
1727                      o2net_idle_timeout() % 1000);
1728
1729                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1730         }
1731         spin_unlock(&nn->nn_lock);
1732 }
1733
1734 static void o2net_still_up(struct work_struct *work)
1735 {
1736         struct o2net_node *nn =
1737                 container_of(work, struct o2net_node, nn_still_up.work);
1738
1739         o2quo_hb_still_up(o2net_num_from_nn(nn));
1740 }
1741
1742 /* ------------------------------------------------------------ */
1743
1744 void o2net_disconnect_node(struct o2nm_node *node)
1745 {
1746         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1747
1748         /* don't reconnect until it's heartbeating again */
1749         spin_lock(&nn->nn_lock);
1750         atomic_set(&nn->nn_timeout, 0);
1751         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1752         spin_unlock(&nn->nn_lock);
1753
1754         if (o2net_wq) {
1755                 cancel_delayed_work(&nn->nn_connect_expired);
1756                 cancel_delayed_work(&nn->nn_connect_work);
1757                 cancel_delayed_work(&nn->nn_still_up);
1758                 flush_workqueue(o2net_wq);
1759         }
1760 }
1761
1762 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1763                                   void *data)
1764 {
1765         o2quo_hb_down(node_num);
1766
1767         if (!node)
1768                 return;
1769
1770         if (node_num != o2nm_this_node())
1771                 o2net_disconnect_node(node);
1772
1773         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1774 }
1775
1776 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1777                                 void *data)
1778 {
1779         struct o2net_node *nn = o2net_nn_from_num(node_num);
1780
1781         o2quo_hb_up(node_num);
1782
1783         BUG_ON(!node);
1784
1785         /* ensure an immediate connect attempt */
1786         nn->nn_last_connect_attempt = jiffies -
1787                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1788
1789         if (node_num != o2nm_this_node()) {
1790                 /* believe it or not, accept and node hearbeating testing
1791                  * can succeed for this node before we got here.. so
1792                  * only use set_nn_state to clear the persistent error
1793                  * if that hasn't already happened */
1794                 spin_lock(&nn->nn_lock);
1795                 atomic_set(&nn->nn_timeout, 0);
1796                 if (nn->nn_persistent_error)
1797                         o2net_set_nn_state(nn, NULL, 0, 0);
1798                 spin_unlock(&nn->nn_lock);
1799         }
1800 }
1801
1802 void o2net_unregister_hb_callbacks(void)
1803 {
1804         o2hb_unregister_callback(NULL, &o2net_hb_up);
1805         o2hb_unregister_callback(NULL, &o2net_hb_down);
1806 }
1807
1808 int o2net_register_hb_callbacks(void)
1809 {
1810         int ret;
1811
1812         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1813                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1814         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1815                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1816
1817         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1818         if (ret == 0)
1819                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1820
1821         if (ret)
1822                 o2net_unregister_hb_callbacks();
1823
1824         return ret;
1825 }
1826
1827 /* ------------------------------------------------------------ */
1828
1829 static int o2net_accept_one(struct socket *sock)
1830 {
1831         int ret, slen;
1832         struct sockaddr_in sin;
1833         struct socket *new_sock = NULL;
1834         struct o2nm_node *node = NULL;
1835         struct o2nm_node *local_node = NULL;
1836         struct o2net_sock_container *sc = NULL;
1837         struct o2net_node *nn;
1838
1839         BUG_ON(sock == NULL);
1840         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1841                                sock->sk->sk_protocol, &new_sock);
1842         if (ret)
1843                 goto out;
1844
1845         new_sock->type = sock->type;
1846         new_sock->ops = sock->ops;
1847         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1848         if (ret < 0)
1849                 goto out;
1850
1851         new_sock->sk->sk_allocation = GFP_ATOMIC;
1852
1853         ret = o2net_set_nodelay(new_sock);
1854         if (ret) {
1855                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1856                 goto out;
1857         }
1858
1859         slen = sizeof(sin);
1860         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1861                                        &slen, 1);
1862         if (ret < 0)
1863                 goto out;
1864
1865         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1866         if (node == NULL) {
1867                 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1868                        "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1869                        ntohs(sin.sin_port));
1870                 ret = -EINVAL;
1871                 goto out;
1872         }
1873
1874         if (o2nm_this_node() >= node->nd_num) {
1875                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1876                 if (local_node)
1877                         printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1878                                         "seen at node '%s' (%u, %pI4:%d) from "
1879                                         "node '%s' (%u, %pI4:%d)\n",
1880                                         local_node->nd_name, local_node->nd_num,
1881                                         &(local_node->nd_ipv4_address),
1882                                         ntohs(local_node->nd_ipv4_port),
1883                                         node->nd_name,
1884                                         node->nd_num, &sin.sin_addr.s_addr,
1885                                         ntohs(sin.sin_port));
1886                 ret = -EINVAL;
1887                 goto out;
1888         }
1889
1890         /* this happens all the time when the other node sees our heartbeat
1891          * and tries to connect before we see their heartbeat */
1892         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1893                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1894                      "%pI4:%d but it isn't heartbeating\n",
1895                      node->nd_name, &sin.sin_addr.s_addr,
1896                      ntohs(sin.sin_port));
1897                 ret = -EINVAL;
1898                 goto out;
1899         }
1900
1901         nn = o2net_nn_from_num(node->nd_num);
1902
1903         spin_lock(&nn->nn_lock);
1904         if (nn->nn_sc)
1905                 ret = -EBUSY;
1906         else
1907                 ret = 0;
1908         spin_unlock(&nn->nn_lock);
1909         if (ret) {
1910                 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1911                        "at %pI4:%d but it already has an open connection\n",
1912                        node->nd_name, &sin.sin_addr.s_addr,
1913                        ntohs(sin.sin_port));
1914                 goto out;
1915         }
1916
1917         sc = sc_alloc(node);
1918         if (sc == NULL) {
1919                 ret = -ENOMEM;
1920                 goto out;
1921         }
1922
1923         sc->sc_sock = new_sock;
1924         new_sock = NULL;
1925
1926         spin_lock(&nn->nn_lock);
1927         atomic_set(&nn->nn_timeout, 0);
1928         o2net_set_nn_state(nn, sc, 0, 0);
1929         spin_unlock(&nn->nn_lock);
1930
1931         o2net_register_callbacks(sc->sc_sock->sk, sc);
1932         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1933
1934         o2net_initialize_handshake();
1935         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1936
1937 out:
1938         if (new_sock)
1939                 sock_release(new_sock);
1940         if (node)
1941                 o2nm_node_put(node);
1942         if (local_node)
1943                 o2nm_node_put(local_node);
1944         if (sc)
1945                 sc_put(sc);
1946         return ret;
1947 }
1948
1949 static void o2net_accept_many(struct work_struct *work)
1950 {
1951         struct socket *sock = o2net_listen_sock;
1952         while (o2net_accept_one(sock) == 0)
1953                 cond_resched();
1954 }
1955
1956 static void o2net_listen_data_ready(struct sock *sk)
1957 {
1958         void (*ready)(struct sock *sk);
1959
1960         read_lock(&sk->sk_callback_lock);
1961         ready = sk->sk_user_data;
1962         if (ready == NULL) { /* check for teardown race */
1963                 ready = sk->sk_data_ready;
1964                 goto out;
1965         }
1966
1967         /* This callback may called twice when a new connection
1968          * is  being established as a child socket inherits everything
1969          * from a parent LISTEN socket, including the data_ready cb of
1970          * the parent. This leads to a hazard. In o2net_accept_one()
1971          * we are still initializing the child socket but have not
1972          * changed the inherited data_ready callback yet when
1973          * data starts arriving.
1974          * We avoid this hazard by checking the state.
1975          * For the listening socket,  the state will be TCP_LISTEN; for the new
1976          * socket, will be  TCP_ESTABLISHED. Also, in this case,
1977          * sk->sk_user_data is not a valid function pointer.
1978          */
1979
1980         if (sk->sk_state == TCP_LISTEN) {
1981                 queue_work(o2net_wq, &o2net_listen_work);
1982         } else {
1983                 ready = NULL;
1984         }
1985
1986 out:
1987         read_unlock(&sk->sk_callback_lock);
1988         if (ready != NULL)
1989                 ready(sk);
1990 }
1991
1992 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1993 {
1994         struct socket *sock = NULL;
1995         int ret;
1996         struct sockaddr_in sin = {
1997                 .sin_family = PF_INET,
1998                 .sin_addr = { .s_addr = addr },
1999                 .sin_port = port,
2000         };
2001
2002         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2003         if (ret < 0) {
2004                 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
2005                 goto out;
2006         }
2007
2008         sock->sk->sk_allocation = GFP_ATOMIC;
2009
2010         write_lock_bh(&sock->sk->sk_callback_lock);
2011         sock->sk->sk_user_data = sock->sk->sk_data_ready;
2012         sock->sk->sk_data_ready = o2net_listen_data_ready;
2013         write_unlock_bh(&sock->sk->sk_callback_lock);
2014
2015         o2net_listen_sock = sock;
2016         INIT_WORK(&o2net_listen_work, o2net_accept_many);
2017
2018         sock->sk->sk_reuse = SK_CAN_REUSE;
2019         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2020         if (ret < 0) {
2021                 printk(KERN_ERR "o2net: Error %d while binding socket at "
2022                        "%pI4:%u\n", ret, &addr, ntohs(port));
2023                 goto out;
2024         }
2025
2026         ret = sock->ops->listen(sock, 64);
2027         if (ret < 0)
2028                 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2029                        ret, &addr, ntohs(port));
2030
2031 out:
2032         if (ret) {
2033                 o2net_listen_sock = NULL;
2034                 if (sock)
2035                         sock_release(sock);
2036         }
2037         return ret;
2038 }
2039
2040 /*
2041  * called from node manager when we should bring up our network listening
2042  * socket.  node manager handles all the serialization to only call this
2043  * once and to match it with o2net_stop_listening().  note,
2044  * o2nm_this_node() doesn't work yet as we're being called while it
2045  * is being set up.
2046  */
2047 int o2net_start_listening(struct o2nm_node *node)
2048 {
2049         int ret = 0;
2050
2051         BUG_ON(o2net_wq != NULL);
2052         BUG_ON(o2net_listen_sock != NULL);
2053
2054         mlog(ML_KTHREAD, "starting o2net thread...\n");
2055         o2net_wq = create_singlethread_workqueue("o2net");
2056         if (o2net_wq == NULL) {
2057                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2058                 return -ENOMEM; /* ? */
2059         }
2060
2061         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2062                                         node->nd_ipv4_port);
2063         if (ret) {
2064                 destroy_workqueue(o2net_wq);
2065                 o2net_wq = NULL;
2066         } else
2067                 o2quo_conn_up(node->nd_num);
2068
2069         return ret;
2070 }
2071
2072 /* again, o2nm_this_node() doesn't work here as we're involved in
2073  * tearing it down */
2074 void o2net_stop_listening(struct o2nm_node *node)
2075 {
2076         struct socket *sock = o2net_listen_sock;
2077         size_t i;
2078
2079         BUG_ON(o2net_wq == NULL);
2080         BUG_ON(o2net_listen_sock == NULL);
2081
2082         /* stop the listening socket from generating work */
2083         write_lock_bh(&sock->sk->sk_callback_lock);
2084         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2085         sock->sk->sk_user_data = NULL;
2086         write_unlock_bh(&sock->sk->sk_callback_lock);
2087
2088         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2089                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2090                 if (node) {
2091                         o2net_disconnect_node(node);
2092                         o2nm_node_put(node);
2093                 }
2094         }
2095
2096         /* finish all work and tear down the work queue */
2097         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2098         destroy_workqueue(o2net_wq);
2099         o2net_wq = NULL;
2100
2101         sock_release(o2net_listen_sock);
2102         o2net_listen_sock = NULL;
2103
2104         o2quo_conn_err(node->nd_num);
2105 }
2106
2107 /* ------------------------------------------------------------ */
2108
2109 int o2net_init(void)
2110 {
2111         unsigned long i;
2112
2113         o2quo_init();
2114
2115         if (o2net_debugfs_init())
2116                 return -ENOMEM;
2117
2118         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2119         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2120         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2121         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2122                 kfree(o2net_hand);
2123                 kfree(o2net_keep_req);
2124                 kfree(o2net_keep_resp);
2125                 return -ENOMEM;
2126         }
2127
2128         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2129         o2net_hand->connector_id = cpu_to_be64(1);
2130
2131         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2132         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2133
2134         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2135                 struct o2net_node *nn = o2net_nn_from_num(i);
2136
2137                 atomic_set(&nn->nn_timeout, 0);
2138                 spin_lock_init(&nn->nn_lock);
2139                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2140                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2141                                   o2net_connect_expired);
2142                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2143                 /* until we see hb from a node we'll return einval */
2144                 nn->nn_persistent_error = -ENOTCONN;
2145                 init_waitqueue_head(&nn->nn_sc_wq);
2146                 idr_init(&nn->nn_status_idr);
2147                 INIT_LIST_HEAD(&nn->nn_status_list);
2148         }
2149
2150         return 0;
2151 }
2152
2153 void o2net_exit(void)
2154 {
2155         o2quo_exit();
2156         kfree(o2net_hand);
2157         kfree(o2net_keep_req);
2158         kfree(o2net_keep_resp);
2159         o2net_debugfs_exit();
2160 }