fs/ocfs2/cluster/tcp.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  *
   3  * vim: noexpandtab sw=8 ts=8 sts=0:
   4  *
   5  * Copyright (C) 2004 Oracle.  All rights reserved.
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2 of the License, or (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public
  18  * License along with this program; if not, write to the
  19  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  20  * Boston, MA 021110-1307, USA.
  21  *
  22  * ----
  23  *
  24  * Callers for this were originally written against a very simple synchronus
  25  * API.  This implementation reflects those simple callers.  Some day I'm sure
  26  * we'll need to move to a more robust posting/callback mechanism.
  27  *
  28  * Transmit calls pass in kernel virtual addresses and block copying this into
  29  * the socket's tx buffers via a usual blocking sendmsg.  They'll block waiting
  30  * for a failed socket to timeout.  TX callers can also pass in a poniter to an
  31  * 'int' which gets filled with an errno off the wire in response to the
  32  * message they send.
  33  *
  34  * Handlers for unsolicited messages are registered.  Each socket has a page
  35  * that incoming data is copied into.  First the header, then the data.
  36  * Handlers are called from only one thread with a reference to this per-socket
  37  * page.  This page is destroyed after the handler call, so it can't be
  38  * referenced beyond the call.  Handlers may block but are discouraged from
  39  * doing so.
  40  *
  41  * Any framing errors (bad magic, large payload lengths) close a connection.
  42  *
  43  * Our sock_container holds the state we associate with a socket.  It's current
  44  * framing state is held there as well as the refcounting we do around when it
  45  * is safe to tear down the socket.  The socket is only finally torn down from
  46  * the container when the container loses all of its references -- so as long
  47  * as you hold a ref on the container you can trust that the socket is valid
  48  * for use with kernel socket APIs.
  49  *
  50  * Connections are initiated between a pair of nodes when the node with the
  51  * higher node number gets a heartbeat callback which indicates that the lower
  52  * numbered node has started heartbeating.  The lower numbered node is passive
  53  * and only accepts the connection if the higher numbered node is heartbeating.
  54  */
  55
  56 #include <linux/kernel.h>
  57 #include <linux/jiffies.h>
  58 #include <linux/slab.h>
  59 #include <linux/idr.h>
  60 #include <linux/kref.h>
  61 #include <linux/net.h>
  62 #include <net/tcp.h>
  63
  64 #include <asm/uaccess.h>
  65
  66 #include "heartbeat.h"
  67 #include "tcp.h"
  68 #include "nodemanager.h"
  69 #define MLOG_MASK_PREFIX ML_TCP
  70 #include "masklog.h"
  71 #include "quorum.h"
  72
  73 #include "tcp_internal.h"
  74
  75 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
  76 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
  77                           &sc->sc_node->nd_ipv4_address,                \
  78                           ntohs(sc->sc_node->nd_ipv4_port)
  79
  80 /*
  81  * In the following two log macros, the whitespace after the ',' just
  82  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  83  * previous token if args expands to nothing.
  84  */
  85 #define msglog(hdr, fmt, args...) do {                                  \
  86         typeof(hdr) __hdr = (hdr);                                      \
  87         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
  88              "key %08x num %u] " fmt,                                   \
  89              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
  90              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
  91              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
  92              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
  93 } while (0)
  94
  95 #define sclog(sc, fmt, args...) do {                                    \
  96         typeof(sc) __sc = (sc);                                         \
  97         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
  98              "pg_off %zu] " fmt, __sc,                                  \
  99              atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,       \
 100             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
 101             ##args);                                                    \
 102 } while (0)
 103
 104 static DEFINE_RWLOCK(o2net_handler_lock);
 105 static struct rb_root o2net_handler_tree = RB_ROOT;
 106
 107 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
 108
 109 /* XXX someday we'll need better accounting */
 110 static struct socket *o2net_listen_sock = NULL;
 111
 112 /*
 113  * listen work is only queued by the listening socket callbacks on the
 114  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 115  * quorum work is queued as sock containers are shutdown.. stop_listening
 116  * tears down all the node's sock containers, preventing future shutdowns
 117  * and queued quroum work, before canceling delayed quorum work and
 118  * destroying the work queue.
 119  */
 120 static struct workqueue_struct *o2net_wq;
 121 static struct work_struct o2net_listen_work;
 122
 123 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 124 #define O2NET_HB_PRI 0x1
 125
 126 static struct o2net_handshake *o2net_hand;
 127 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 128
 129 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 130                 {[O2NET_ERR_NONE]       = 0,
 131                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
 132                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
 133                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
 134
 135 /* can't quite avoid *all* internal declarations :/ */
 136 static void o2net_sc_connect_completed(struct work_struct *work);
 137 static void o2net_rx_until_empty(struct work_struct *work);
 138 static void o2net_shutdown_sc(struct work_struct *work);
 139 static void o2net_listen_data_ready(struct sock *sk, int bytes);
 140 static void o2net_sc_send_keep_req(struct work_struct *work);
 141 static void o2net_idle_timer(unsigned long data);
 142 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 143 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 144
 145 #ifdef CONFIG_DEBUG_FS
 146 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 147                            u32 msgkey, struct task_struct *task, u8 node)
 148 {
 149         INIT_LIST_HEAD(&nst->st_net_debug_item);
 150         nst->st_task = task;
 151         nst->st_msg_type = msgtype;
 152         nst->st_msg_key = msgkey;
 153         nst->st_node = node;
 154 }
 155
 156 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 157 {
 158         nst->st_sock_time = ktime_get();
 159 }
 160
 161 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 162 {
 163         nst->st_send_time = ktime_get();
 164 }
 165
 166 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 167 {
 168         nst->st_status_time = ktime_get();
 169 }
 170
 171 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 172                                                 struct o2net_sock_container *sc)
 173 {
 174         nst->st_sc = sc;
 175 }
 176
 177 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 178                                         u32 msg_id)
 179 {
 180         nst->st_id = msg_id;
 181 }
 182
 183 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
 184 {
 185         sc->sc_tv_timer = ktime_get();
 186 }
 187
 188 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
 189 {
 190         sc->sc_tv_data_ready = ktime_get();
 191 }
 192
 193 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
 194 {
 195         sc->sc_tv_advance_start = ktime_get();
 196 }
 197
 198 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
 199 {
 200         sc->sc_tv_advance_stop = ktime_get();
 201 }
 202
 203 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
 204 {
 205         sc->sc_tv_func_start = ktime_get();
 206 }
 207
 208 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
 209 {
 210         sc->sc_tv_func_stop = ktime_get();
 211 }
 212
 213 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
 214 {
 215         return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
 216 }
 217 #else  /* CONFIG_DEBUG_FS */
 218 # define o2net_init_nst(a, b, c, d, e)
 219 # define o2net_set_nst_sock_time(a)
 220 # define o2net_set_nst_send_time(a)
 221 # define o2net_set_nst_status_time(a)
 222 # define o2net_set_nst_sock_container(a, b)
 223 # define o2net_set_nst_msg_id(a, b)
 224 # define o2net_set_sock_timer(a)
 225 # define o2net_set_data_ready_time(a)
 226 # define o2net_set_advance_start_time(a)
 227 # define o2net_set_advance_stop_time(a)
 228 # define o2net_set_func_start_time(a)
 229 # define o2net_set_func_stop_time(a)
 230 # define o2net_get_func_run_time(a)             (ktime_t)0
 231 #endif /* CONFIG_DEBUG_FS */
 232
 233 #ifdef CONFIG_OCFS2_FS_STATS
 234 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
 235                                     struct o2net_sock_container *sc)
 236 {
 237         sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
 238                                            ktime_sub(ktime_get(),
 239                                                      nst->st_status_time));
 240         sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
 241                                          ktime_sub(nst->st_status_time,
 242                                                    nst->st_send_time));
 243         sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
 244                                             ktime_sub(nst->st_send_time,
 245                                                       nst->st_sock_time));
 246         sc->sc_send_count++;
 247 }
 248
 249 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
 250 {
 251         sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
 252                                             o2net_get_func_run_time(sc));
 253         sc->sc_recv_count++;
 254 }
 255
 256 #else
 257
 258 # define o2net_update_send_stats(a, b)
 259
 260 #endif /* CONFIG_OCFS2_FS_STATS */
 261
 262 static inline int o2net_reconnect_delay(void)
 263 {
 264         return o2nm_single_cluster->cl_reconnect_delay_ms;
 265 }
 266
 267 static inline int o2net_keepalive_delay(void)
 268 {
 269         return o2nm_single_cluster->cl_keepalive_delay_ms;
 270 }
 271
 272 static inline int o2net_idle_timeout(void)
 273 {
 274         return o2nm_single_cluster->cl_idle_timeout_ms;
 275 }
 276
 277 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 278 {
 279         int trans;
 280         BUG_ON(err >= O2NET_ERR_MAX);
 281         trans = o2net_sys_err_translations[err];
 282
 283         /* Just in case we mess up the translation table above */
 284         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 285         return trans;
 286 }
 287
 288 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 289 {
 290         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 291         return &o2net_nodes[node_num];
 292 }
 293
 294 static u8 o2net_num_from_nn(struct o2net_node *nn)
 295 {
 296         BUG_ON(nn == NULL);
 297         return nn - o2net_nodes;
 298 }
 299
 300 /* ------------------------------------------------------------ */
 301
 302 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 303 {
 304         int ret = 0;
 305
 306         do {
 307                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
 308                         ret = -EAGAIN;
 309                         break;
 310                 }
 311                 spin_lock(&nn->nn_lock);
 312                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
 313                 if (ret == 0)
 314                         list_add_tail(&nsw->ns_node_item,
 315                                       &nn->nn_status_list);
 316                 spin_unlock(&nn->nn_lock);
 317         } while (ret == -EAGAIN);
 318
 319         if (ret == 0)  {
 320                 init_waitqueue_head(&nsw->ns_wq);
 321                 nsw->ns_sys_status = O2NET_ERR_NONE;
 322                 nsw->ns_status = 0;
 323         }
 324
 325         return ret;
 326 }
 327
 328 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 329                                       struct o2net_status_wait *nsw,
 330                                       enum o2net_system_error sys_status,
 331                                       s32 status)
 332 {
 333         assert_spin_locked(&nn->nn_lock);
 334
 335         if (!list_empty(&nsw->ns_node_item)) {
 336                 list_del_init(&nsw->ns_node_item);
 337                 nsw->ns_sys_status = sys_status;
 338                 nsw->ns_status = status;
 339                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 340                 wake_up(&nsw->ns_wq);
 341         }
 342 }
 343
 344 static void o2net_complete_nsw(struct o2net_node *nn,
 345                                struct o2net_status_wait *nsw,
 346                                u64 id, enum o2net_system_error sys_status,
 347                                s32 status)
 348 {
 349         spin_lock(&nn->nn_lock);
 350         if (nsw == NULL) {
 351                 if (id > INT_MAX)
 352                         goto out;
 353
 354                 nsw = idr_find(&nn->nn_status_idr, id);
 355                 if (nsw == NULL)
 356                         goto out;
 357         }
 358
 359         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 360
 361 out:
 362         spin_unlock(&nn->nn_lock);
 363         return;
 364 }
 365
 366 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 367 {
 368         struct o2net_status_wait *nsw, *tmp;
 369         unsigned int num_kills = 0;
 370
 371         assert_spin_locked(&nn->nn_lock);
 372
 373         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 374                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 375                 num_kills++;
 376         }
 377
 378         mlog(0, "completed %d messages for node %u\n", num_kills,
 379              o2net_num_from_nn(nn));
 380 }
 381
 382 static int o2net_nsw_completed(struct o2net_node *nn,
 383                                struct o2net_status_wait *nsw)
 384 {
 385         int completed;
 386         spin_lock(&nn->nn_lock);
 387         completed = list_empty(&nsw->ns_node_item);
 388         spin_unlock(&nn->nn_lock);
 389         return completed;
 390 }
 391
 392 /* ------------------------------------------------------------ */
 393
 394 static void sc_kref_release(struct kref *kref)
 395 {
 396         struct o2net_sock_container *sc = container_of(kref,
 397                                         struct o2net_sock_container, sc_kref);
 398         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 399
 400         sclog(sc, "releasing\n");
 401
 402         if (sc->sc_sock) {
 403                 sock_release(sc->sc_sock);
 404                 sc->sc_sock = NULL;
 405         }
 406
 407         o2nm_undepend_item(&sc->sc_node->nd_item);
 408         o2nm_node_put(sc->sc_node);
 409         sc->sc_node = NULL;
 410
 411         o2net_debug_del_sc(sc);
 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                 printk(KERN_NOTICE "o2net: no longer connected to "
 547                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 548                 o2net_complete_nodes_nsw(nn);
 549         }
 550
 551         if (!was_valid && valid) {
 552                 o2quo_conn_up(o2net_num_from_nn(nn));
 553                 cancel_delayed_work(&nn->nn_connect_expired);
 554                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 555                        o2nm_this_node() > sc->sc_node->nd_num ?
 556                                 "connected to" : "accepted connection from",
 557                        SC_NODEF_ARGS(sc));
 558         }
 559
 560         /* trigger the connecting worker func as long as we're not valid,
 561          * it will back off if it shouldn't connect.  This can be called
 562          * from node config teardown and so needs to be careful about
 563          * the work queue actually being up. */
 564         if (!valid && o2net_wq) {
 565                 unsigned long delay;
 566                 /* delay if we're withing a RECONNECT_DELAY of the
 567                  * last attempt */
 568                 delay = (nn->nn_last_connect_attempt +
 569                          msecs_to_jiffies(o2net_reconnect_delay()))
 570                         - jiffies;
 571                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 572                         delay = 0;
 573                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 574                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 575
 576                 /*
 577                  * Delay the expired work after idle timeout.
 578                  *
 579                  * We might have lots of failed connection attempts that run
 580                  * through here but we only cancel the connect_expired work when
 581                  * a connection attempt succeeds.  So only the first enqueue of
 582                  * the connect_expired work will do anything.  The rest will see
 583                  * that it's already queued and do nothing.
 584                  */
 585                 delay += msecs_to_jiffies(o2net_idle_timeout());
 586                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 587         }
 588
 589         /* keep track of the nn's sc ref for the caller */
 590         if ((old_sc == NULL) && sc)
 591                 sc_get(sc);
 592         if (old_sc && (old_sc != sc)) {
 593                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 594                 sc_put(old_sc);
 595         }
 596 }
 597
 598 /* see o2net_register_callbacks() */
 599 static void o2net_data_ready(struct sock *sk, int bytes)
 600 {
 601         void (*ready)(struct sock *sk, int bytes);
 602
 603         read_lock(&sk->sk_callback_lock);
 604         if (sk->sk_user_data) {
 605                 struct o2net_sock_container *sc = sk->sk_user_data;
 606                 sclog(sc, "data_ready hit\n");
 607                 o2net_set_data_ready_time(sc);
 608                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 609                 ready = sc->sc_data_ready;
 610         } else {
 611                 ready = sk->sk_data_ready;
 612         }
 613         read_unlock(&sk->sk_callback_lock);
 614
 615         ready(sk, bytes);
 616 }
 617
 618 /* see o2net_register_callbacks() */
 619 static void o2net_state_change(struct sock *sk)
 620 {
 621         void (*state_change)(struct sock *sk);
 622         struct o2net_sock_container *sc;
 623
 624         read_lock(&sk->sk_callback_lock);
 625         sc = sk->sk_user_data;
 626         if (sc == NULL) {
 627                 state_change = sk->sk_state_change;
 628                 goto out;
 629         }
 630
 631         sclog(sc, "state_change to %d\n", sk->sk_state);
 632
 633         state_change = sc->sc_state_change;
 634
 635         switch(sk->sk_state) {
 636                 /* ignore connecting sockets as they make progress */
 637                 case TCP_SYN_SENT:
 638                 case TCP_SYN_RECV:
 639                         break;
 640                 case TCP_ESTABLISHED:
 641                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
 642                         break;
 643                 default:
 644                         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT
 645                               " shutdown, state %d\n",
 646                               SC_NODEF_ARGS(sc), sk->sk_state);
 647                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 648                         break;
 649         }
 650 out:
 651         read_unlock(&sk->sk_callback_lock);
 652         state_change(sk);
 653 }
 654
 655 /*
 656  * we register callbacks so we can queue work on events before calling
 657  * the original callbacks.  our callbacks our careful to test user_data
 658  * to discover when they've reaced with o2net_unregister_callbacks().
 659  */
 660 static void o2net_register_callbacks(struct sock *sk,
 661                                      struct o2net_sock_container *sc)
 662 {
 663         write_lock_bh(&sk->sk_callback_lock);
 664
 665         /* accepted sockets inherit the old listen socket data ready */
 666         if (sk->sk_data_ready == o2net_listen_data_ready) {
 667                 sk->sk_data_ready = sk->sk_user_data;
 668                 sk->sk_user_data = NULL;
 669         }
 670
 671         BUG_ON(sk->sk_user_data != NULL);
 672         sk->sk_user_data = sc;
 673         sc_get(sc);
 674
 675         sc->sc_data_ready = sk->sk_data_ready;
 676         sc->sc_state_change = sk->sk_state_change;
 677         sk->sk_data_ready = o2net_data_ready;
 678         sk->sk_state_change = o2net_state_change;
 679
 680         mutex_init(&sc->sc_send_lock);
 681
 682         write_unlock_bh(&sk->sk_callback_lock);
 683 }
 684
 685 static int o2net_unregister_callbacks(struct sock *sk,
 686                                    struct o2net_sock_container *sc)
 687 {
 688         int ret = 0;
 689
 690         write_lock_bh(&sk->sk_callback_lock);
 691         if (sk->sk_user_data == sc) {
 692                 ret = 1;
 693                 sk->sk_user_data = NULL;
 694                 sk->sk_data_ready = sc->sc_data_ready;
 695                 sk->sk_state_change = sc->sc_state_change;
 696         }
 697         write_unlock_bh(&sk->sk_callback_lock);
 698
 699         return ret;
 700 }
 701
 702 /*
 703  * this is a little helper that is called by callers who have seen a problem
 704  * with an sc and want to detach it from the nn if someone already hasn't beat
 705  * them to it.  if an error is given then the shutdown will be persistent
 706  * and pending transmits will be canceled.
 707  */
 708 static void o2net_ensure_shutdown(struct o2net_node *nn,
 709                                    struct o2net_sock_container *sc,
 710                                    int err)
 711 {
 712         spin_lock(&nn->nn_lock);
 713         if (nn->nn_sc == sc)
 714                 o2net_set_nn_state(nn, NULL, 0, err);
 715         spin_unlock(&nn->nn_lock);
 716 }
 717
 718 /*
 719  * This work queue function performs the blocking parts of socket shutdown.  A
 720  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 721  * sc detached from the nn.  state_change will also trigger this callback
 722  * directly when it sees errors.  In that case we need to call set_nn_state
 723  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 724  * itself.
 725  */
 726 static void o2net_shutdown_sc(struct work_struct *work)
 727 {
 728         struct o2net_sock_container *sc =
 729                 container_of(work, struct o2net_sock_container,
 730                              sc_shutdown_work);
 731         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 732
 733         sclog(sc, "shutting down\n");
 734
 735         /* drop the callbacks ref and call shutdown only once */
 736         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 737                 /* we shouldn't flush as we're in the thread, the
 738                  * races with pending sc work structs are harmless */
 739                 del_timer_sync(&sc->sc_idle_timeout);
 740                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 741                 sc_put(sc);
 742                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 743         }
 744
 745         /* not fatal so failed connects before the other guy has our
 746          * heartbeat can be retried */
 747         o2net_ensure_shutdown(nn, sc, 0);
 748         sc_put(sc);
 749 }
 750
 751 /* ------------------------------------------------------------ */
 752
 753 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 754                              u32 key)
 755 {
 756         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 757
 758         if (ret == 0)
 759                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 760
 761         return ret;
 762 }
 763
 764 static struct o2net_msg_handler *
 765 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 766                           struct rb_node **ret_parent)
 767 {
 768         struct rb_node **p = &o2net_handler_tree.rb_node;
 769         struct rb_node *parent = NULL;
 770         struct o2net_msg_handler *nmh, *ret = NULL;
 771         int cmp;
 772
 773         while (*p) {
 774                 parent = *p;
 775                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 776                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 777
 778                 if (cmp < 0)
 779                         p = &(*p)->rb_left;
 780                 else if (cmp > 0)
 781                         p = &(*p)->rb_right;
 782                 else {
 783                         ret = nmh;
 784                         break;
 785                 }
 786         }
 787
 788         if (ret_p != NULL)
 789                 *ret_p = p;
 790         if (ret_parent != NULL)
 791                 *ret_parent = parent;
 792
 793         return ret;
 794 }
 795
 796 static void o2net_handler_kref_release(struct kref *kref)
 797 {
 798         struct o2net_msg_handler *nmh;
 799         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 800
 801         kfree(nmh);
 802 }
 803
 804 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 805 {
 806         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 807 }
 808
 809 /* max_len is protection for the handler func.  incoming messages won't
 810  * be given to the handler if their payload is longer than the max. */
 811 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 812                            o2net_msg_handler_func *func, void *data,
 813                            o2net_post_msg_handler_func *post_func,
 814                            struct list_head *unreg_list)
 815 {
 816         struct o2net_msg_handler *nmh = NULL;
 817         struct rb_node **p, *parent;
 818         int ret = 0;
 819
 820         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 821                 mlog(0, "max_len for message handler out of range: %u\n",
 822                         max_len);
 823                 ret = -EINVAL;
 824                 goto out;
 825         }
 826
 827         if (!msg_type) {
 828                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 829                 ret = -EINVAL;
 830                 goto out;
 831
 832         }
 833         if (!func) {
 834                 mlog(0, "no message handler provided: %u, %p\n",
 835                        msg_type, func);
 836                 ret = -EINVAL;
 837                 goto out;
 838         }
 839
 840         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 841         if (nmh == NULL) {
 842                 ret = -ENOMEM;
 843                 goto out;
 844         }
 845
 846         nmh->nh_func = func;
 847         nmh->nh_func_data = data;
 848         nmh->nh_post_func = post_func;
 849         nmh->nh_msg_type = msg_type;
 850         nmh->nh_max_len = max_len;
 851         nmh->nh_key = key;
 852         /* the tree and list get this ref.. they're both removed in
 853          * unregister when this ref is dropped */
 854         kref_init(&nmh->nh_kref);
 855         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 856
 857         write_lock(&o2net_handler_lock);
 858         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 859                 ret = -EEXIST;
 860         else {
 861                 rb_link_node(&nmh->nh_node, parent, p);
 862                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 863                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 864
 865                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 866                      func, msg_type, key);
 867                 /* we've had some trouble with handlers seemingly vanishing. */
 868                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 869                                                           &parent) == NULL,
 870                                 "couldn't find handler we *just* registerd "
 871                                 "for type %u key %08x\n", msg_type, key);
 872         }
 873         write_unlock(&o2net_handler_lock);
 874         if (ret)
 875                 goto out;
 876
 877 out:
 878         if (ret)
 879                 kfree(nmh);
 880
 881         return ret;
 882 }
 883 EXPORT_SYMBOL_GPL(o2net_register_handler);
 884
 885 void o2net_unregister_handler_list(struct list_head *list)
 886 {
 887         struct o2net_msg_handler *nmh, *n;
 888
 889         write_lock(&o2net_handler_lock);
 890         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 891                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 892                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 893                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 894                 list_del_init(&nmh->nh_unregister_item);
 895                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 896         }
 897         write_unlock(&o2net_handler_lock);
 898 }
 899 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 900
 901 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 902 {
 903         struct o2net_msg_handler *nmh;
 904
 905         read_lock(&o2net_handler_lock);
 906         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 907         if (nmh)
 908                 kref_get(&nmh->nh_kref);
 909         read_unlock(&o2net_handler_lock);
 910
 911         return nmh;
 912 }
 913
 914 /* ------------------------------------------------------------ */
 915
 916 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 917 {
 918         int ret;
 919         mm_segment_t oldfs;
 920         struct kvec vec = {
 921                 .iov_len = len,
 922                 .iov_base = data,
 923         };
 924         struct msghdr msg = {
 925                 .msg_iovlen = 1,
 926                 .msg_iov = (struct iovec *)&vec,
 927                 .msg_flags = MSG_DONTWAIT,
 928         };
 929
 930         oldfs = get_fs();
 931         set_fs(get_ds());
 932         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
 933         set_fs(oldfs);
 934
 935         return ret;
 936 }
 937
 938 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 939                               size_t veclen, size_t total)
 940 {
 941         int ret;
 942         mm_segment_t oldfs;
 943         struct msghdr msg = {
 944                 .msg_iov = (struct iovec *)vec,
 945                 .msg_iovlen = veclen,
 946         };
 947
 948         if (sock == NULL) {
 949                 ret = -EINVAL;
 950                 goto out;
 951         }
 952
 953         oldfs = get_fs();
 954         set_fs(get_ds());
 955         ret = sock_sendmsg(sock, &msg, total);
 956         set_fs(oldfs);
 957         if (ret != total) {
 958                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
 959                      total);
 960                 if (ret >= 0)
 961                         ret = -EPIPE; /* should be smarter, I bet */
 962                 goto out;
 963         }
 964
 965         ret = 0;
 966 out:
 967         if (ret < 0)
 968                 mlog(0, "returning error: %d\n", ret);
 969         return ret;
 970 }
 971
 972 static void o2net_sendpage(struct o2net_sock_container *sc,
 973                            void *kmalloced_virt,
 974                            size_t size)
 975 {
 976         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 977         ssize_t ret;
 978
 979         while (1) {
 980                 mutex_lock(&sc->sc_send_lock);
 981                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 982                                                  virt_to_page(kmalloced_virt),
 983                                                  (long)kmalloced_virt & ~PAGE_MASK,
 984                                                  size, MSG_DONTWAIT);
 985                 mutex_unlock(&sc->sc_send_lock);
 986                 if (ret == size)
 987                         break;
 988                 if (ret == (ssize_t)-EAGAIN) {
 989                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 990                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 991                         cond_resched();
 992                         continue;
 993                 }
 994                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 995                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 996                 o2net_ensure_shutdown(nn, sc, 0);
 997                 break;
 998         }
 999 }
1000
1001 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
1002 {
1003         memset(msg, 0, sizeof(struct o2net_msg));
1004         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
1005         msg->data_len = cpu_to_be16(data_len);
1006         msg->msg_type = cpu_to_be16(msg_type);
1007         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
1008         msg->status = 0;
1009         msg->key = cpu_to_be32(key);
1010 }
1011
1012 static int o2net_tx_can_proceed(struct o2net_node *nn,
1013                                 struct o2net_sock_container **sc_ret,
1014                                 int *error)
1015 {
1016         int ret = 0;
1017
1018         spin_lock(&nn->nn_lock);
1019         if (nn->nn_persistent_error) {
1020                 ret = 1;
1021                 *sc_ret = NULL;
1022                 *error = nn->nn_persistent_error;
1023         } else if (nn->nn_sc_valid) {
1024                 kref_get(&nn->nn_sc->sc_kref);
1025
1026                 ret = 1;
1027                 *sc_ret = nn->nn_sc;
1028                 *error = 0;
1029         }
1030         spin_unlock(&nn->nn_lock);
1031
1032         return ret;
1033 }
1034
1035 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1036                            size_t caller_veclen, u8 target_node, int *status)
1037 {
1038         int ret = 0;
1039         struct o2net_msg *msg = NULL;
1040         size_t veclen, caller_bytes = 0;
1041         struct kvec *vec = NULL;
1042         struct o2net_sock_container *sc = NULL;
1043         struct o2net_node *nn = o2net_nn_from_num(target_node);
1044         struct o2net_status_wait nsw = {
1045                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1046         };
1047         struct o2net_send_tracking nst;
1048
1049         o2net_init_nst(&nst, msg_type, key, current, target_node);
1050
1051         if (o2net_wq == NULL) {
1052                 mlog(0, "attempt to tx without o2netd running\n");
1053                 ret = -ESRCH;
1054                 goto out;
1055         }
1056
1057         if (caller_veclen == 0) {
1058                 mlog(0, "bad kvec array length\n");
1059                 ret = -EINVAL;
1060                 goto out;
1061         }
1062
1063         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1064         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1065                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1066                 ret = -EINVAL;
1067                 goto out;
1068         }
1069
1070         if (target_node == o2nm_this_node()) {
1071                 ret = -ELOOP;
1072                 goto out;
1073         }
1074
1075         o2net_debug_add_nst(&nst);
1076
1077         o2net_set_nst_sock_time(&nst);
1078
1079         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1080         if (ret)
1081                 goto out;
1082
1083         o2net_set_nst_sock_container(&nst, sc);
1084
1085         veclen = caller_veclen + 1;
1086         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1087         if (vec == NULL) {
1088                 mlog(0, "failed to %zu element kvec!\n", veclen);
1089                 ret = -ENOMEM;
1090                 goto out;
1091         }
1092
1093         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1094         if (!msg) {
1095                 mlog(0, "failed to allocate a o2net_msg!\n");
1096                 ret = -ENOMEM;
1097                 goto out;
1098         }
1099
1100         o2net_init_msg(msg, caller_bytes, msg_type, key);
1101
1102         vec[0].iov_len = sizeof(struct o2net_msg);
1103         vec[0].iov_base = msg;
1104         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1105
1106         ret = o2net_prep_nsw(nn, &nsw);
1107         if (ret)
1108                 goto out;
1109
1110         msg->msg_num = cpu_to_be32(nsw.ns_id);
1111         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1112
1113         o2net_set_nst_send_time(&nst);
1114
1115         /* finally, convert the message header to network byte-order
1116          * and send */
1117         mutex_lock(&sc->sc_send_lock);
1118         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1119                                  sizeof(struct o2net_msg) + caller_bytes);
1120         mutex_unlock(&sc->sc_send_lock);
1121         msglog(msg, "sending returned %d\n", ret);
1122         if (ret < 0) {
1123                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1124                 goto out;
1125         }
1126
1127         /* wait on other node's handler */
1128         o2net_set_nst_status_time(&nst);
1129         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1130
1131         o2net_update_send_stats(&nst, sc);
1132
1133         /* Note that we avoid overwriting the callers status return
1134          * variable if a system error was reported on the other
1135          * side. Callers beware. */
1136         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1137         if (status && !ret)
1138                 *status = nsw.ns_status;
1139
1140         mlog(0, "woken, returning system status %d, user status %d\n",
1141              ret, nsw.ns_status);
1142 out:
1143         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1144         if (sc)
1145                 sc_put(sc);
1146         if (vec)
1147                 kfree(vec);
1148         if (msg)
1149                 kfree(msg);
1150         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1151         return ret;
1152 }
1153 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1154
1155 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1156                        u8 target_node, int *status)
1157 {
1158         struct kvec vec = {
1159                 .iov_base = data,
1160                 .iov_len = len,
1161         };
1162         return o2net_send_message_vec(msg_type, key, &vec, 1,
1163                                       target_node, status);
1164 }
1165 EXPORT_SYMBOL_GPL(o2net_send_message);
1166
1167 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1168                                    enum o2net_system_error syserr, int err)
1169 {
1170         struct kvec vec = {
1171                 .iov_base = hdr,
1172                 .iov_len = sizeof(struct o2net_msg),
1173         };
1174
1175         BUG_ON(syserr >= O2NET_ERR_MAX);
1176
1177         /* leave other fields intact from the incoming message, msg_num
1178          * in particular */
1179         hdr->sys_status = cpu_to_be32(syserr);
1180         hdr->status = cpu_to_be32(err);
1181         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1182         hdr->data_len = 0;
1183
1184         msglog(hdr, "about to send status magic %d\n", err);
1185         /* hdr has been in host byteorder this whole time */
1186         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1187 }
1188
1189 /* this returns -errno if the header was unknown or too large, etc.
1190  * after this is called the buffer us reused for the next message */
1191 static int o2net_process_message(struct o2net_sock_container *sc,
1192                                  struct o2net_msg *hdr)
1193 {
1194         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1195         int ret = 0, handler_status;
1196         enum  o2net_system_error syserr;
1197         struct o2net_msg_handler *nmh = NULL;
1198         void *ret_data = NULL;
1199
1200         msglog(hdr, "processing message\n");
1201
1202         o2net_sc_postpone_idle(sc);
1203
1204         switch(be16_to_cpu(hdr->magic)) {
1205                 case O2NET_MSG_STATUS_MAGIC:
1206                         /* special type for returning message status */
1207                         o2net_complete_nsw(nn, NULL,
1208                                            be32_to_cpu(hdr->msg_num),
1209                                            be32_to_cpu(hdr->sys_status),
1210                                            be32_to_cpu(hdr->status));
1211                         goto out;
1212                 case O2NET_MSG_KEEP_REQ_MAGIC:
1213                         o2net_sendpage(sc, o2net_keep_resp,
1214                                        sizeof(*o2net_keep_resp));
1215                         goto out;
1216                 case O2NET_MSG_KEEP_RESP_MAGIC:
1217                         goto out;
1218                 case O2NET_MSG_MAGIC:
1219                         break;
1220                 default:
1221                         msglog(hdr, "bad magic\n");
1222                         ret = -EINVAL;
1223                         goto out;
1224                         break;
1225         }
1226
1227         /* find a handler for it */
1228         handler_status = 0;
1229         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1230                                 be32_to_cpu(hdr->key));
1231         if (!nmh) {
1232                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1233                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1234                 syserr = O2NET_ERR_NO_HNDLR;
1235                 goto out_respond;
1236         }
1237
1238         syserr = O2NET_ERR_NONE;
1239
1240         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1241                 syserr = O2NET_ERR_OVERFLOW;
1242
1243         if (syserr != O2NET_ERR_NONE)
1244                 goto out_respond;
1245
1246         o2net_set_func_start_time(sc);
1247         sc->sc_msg_key = be32_to_cpu(hdr->key);
1248         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1249         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1250                                              be16_to_cpu(hdr->data_len),
1251                                         nmh->nh_func_data, &ret_data);
1252         o2net_set_func_stop_time(sc);
1253
1254         o2net_update_recv_stats(sc);
1255
1256 out_respond:
1257         /* this destroys the hdr, so don't use it after this */
1258         mutex_lock(&sc->sc_send_lock);
1259         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1260                                       handler_status);
1261         mutex_unlock(&sc->sc_send_lock);
1262         hdr = NULL;
1263         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1264              handler_status, syserr, ret);
1265
1266         if (nmh) {
1267                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1268                 if (nmh->nh_post_func)
1269                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1270                                             ret_data);
1271         }
1272
1273 out:
1274         if (nmh)
1275                 o2net_handler_put(nmh);
1276         return ret;
1277 }
1278
1279 static int o2net_check_handshake(struct o2net_sock_container *sc)
1280 {
1281         struct o2net_handshake *hand = page_address(sc->sc_page);
1282         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1283
1284         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1285                 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1286                      "version %llu but %llu is required, disconnecting\n",
1287                      SC_NODEF_ARGS(sc),
1288                      (unsigned long long)be64_to_cpu(hand->protocol_version),
1289                      O2NET_PROTOCOL_VERSION);
1290
1291                 /* don't bother reconnecting if its the wrong version. */
1292                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1293                 return -1;
1294         }
1295
1296         /*
1297          * Ensure timeouts are consistent with other nodes, otherwise
1298          * we can end up with one node thinking that the other must be down,
1299          * but isn't. This can ultimately cause corruption.
1300          */
1301         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1302                                 o2net_idle_timeout()) {
1303                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1304                      "%u ms, but we use %u ms locally.  disconnecting\n",
1305                      SC_NODEF_ARGS(sc),
1306                      be32_to_cpu(hand->o2net_idle_timeout_ms),
1307                      o2net_idle_timeout());
1308                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1309                 return -1;
1310         }
1311
1312         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1313                         o2net_keepalive_delay()) {
1314                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1315                      "%u ms, but we use %u ms locally.  disconnecting\n",
1316                      SC_NODEF_ARGS(sc),
1317                      be32_to_cpu(hand->o2net_keepalive_delay_ms),
1318                      o2net_keepalive_delay());
1319                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1320                 return -1;
1321         }
1322
1323         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1324                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1325                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1326                      "%u ms, but we use %u ms locally.  disconnecting\n",
1327                      SC_NODEF_ARGS(sc),
1328                      be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1329                      O2HB_MAX_WRITE_TIMEOUT_MS);
1330                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1331                 return -1;
1332         }
1333
1334         sc->sc_handshake_ok = 1;
1335
1336         spin_lock(&nn->nn_lock);
1337         /* set valid and queue the idle timers only if it hasn't been
1338          * shut down already */
1339         if (nn->nn_sc == sc) {
1340                 o2net_sc_reset_idle_timer(sc);
1341                 atomic_set(&nn->nn_timeout, 0);
1342                 o2net_set_nn_state(nn, sc, 1, 0);
1343         }
1344         spin_unlock(&nn->nn_lock);
1345
1346         /* shift everything up as though it wasn't there */
1347         sc->sc_page_off -= sizeof(struct o2net_handshake);
1348         if (sc->sc_page_off)
1349                 memmove(hand, hand + 1, sc->sc_page_off);
1350
1351         return 0;
1352 }
1353
1354 /* this demuxes the queued rx bytes into header or payload bits and calls
1355  * handlers as each full message is read off the socket.  it returns -error,
1356  * == 0 eof, or > 0 for progress made.*/
1357 static int o2net_advance_rx(struct o2net_sock_container *sc)
1358 {
1359         struct o2net_msg *hdr;
1360         int ret = 0;
1361         void *data;
1362         size_t datalen;
1363
1364         sclog(sc, "receiving\n");
1365         o2net_set_advance_start_time(sc);
1366
1367         if (unlikely(sc->sc_handshake_ok == 0)) {
1368                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1369                         data = page_address(sc->sc_page) + sc->sc_page_off;
1370                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1371                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1372                         if (ret > 0)
1373                                 sc->sc_page_off += ret;
1374                 }
1375
1376                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1377                         o2net_check_handshake(sc);
1378                         if (unlikely(sc->sc_handshake_ok == 0))
1379                                 ret = -EPROTO;
1380                 }
1381                 goto out;
1382         }
1383
1384         /* do we need more header? */
1385         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1386                 data = page_address(sc->sc_page) + sc->sc_page_off;
1387                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1388                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1389                 if (ret > 0) {
1390                         sc->sc_page_off += ret;
1391                         /* only swab incoming here.. we can
1392                          * only get here once as we cross from
1393                          * being under to over */
1394                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1395                                 hdr = page_address(sc->sc_page);
1396                                 if (be16_to_cpu(hdr->data_len) >
1397                                     O2NET_MAX_PAYLOAD_BYTES)
1398                                         ret = -EOVERFLOW;
1399                         }
1400                 }
1401                 if (ret <= 0)
1402                         goto out;
1403         }
1404
1405         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1406                 /* oof, still don't have a header */
1407                 goto out;
1408         }
1409
1410         /* this was swabbed above when we first read it */
1411         hdr = page_address(sc->sc_page);
1412
1413         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1414
1415         /* do we need more payload? */
1416         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1417                 /* need more payload */
1418                 data = page_address(sc->sc_page) + sc->sc_page_off;
1419                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1420                           sc->sc_page_off;
1421                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1422                 if (ret > 0)
1423                         sc->sc_page_off += ret;
1424                 if (ret <= 0)
1425                         goto out;
1426         }
1427
1428         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1429                 /* we can only get here once, the first time we read
1430                  * the payload.. so set ret to progress if the handler
1431                  * works out. after calling this the message is toast */
1432                 ret = o2net_process_message(sc, hdr);
1433                 if (ret == 0)
1434                         ret = 1;
1435                 sc->sc_page_off = 0;
1436         }
1437
1438 out:
1439         sclog(sc, "ret = %d\n", ret);
1440         o2net_set_advance_stop_time(sc);
1441         return ret;
1442 }
1443
1444 /* this work func is triggerd by data ready.  it reads until it can read no
1445  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1446  * our work the work struct will be marked and we'll be called again. */
1447 static void o2net_rx_until_empty(struct work_struct *work)
1448 {
1449         struct o2net_sock_container *sc =
1450                 container_of(work, struct o2net_sock_container, sc_rx_work);
1451         int ret;
1452
1453         do {
1454                 ret = o2net_advance_rx(sc);
1455         } while (ret > 0);
1456
1457         if (ret <= 0 && ret != -EAGAIN) {
1458                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1459                 sclog(sc, "saw error %d, closing\n", ret);
1460                 /* not permanent so read failed handshake can retry */
1461                 o2net_ensure_shutdown(nn, sc, 0);
1462         }
1463
1464         sc_put(sc);
1465 }
1466
1467 static int o2net_set_nodelay(struct socket *sock)
1468 {
1469         int ret, val = 1;
1470         mm_segment_t oldfs;
1471
1472         oldfs = get_fs();
1473         set_fs(KERNEL_DS);
1474
1475         /*
1476          * Dear unsuspecting programmer,
1477          *
1478          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1479          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1480          * silently turn into SO_DEBUG.
1481          *
1482          * Yours,
1483          * Keeper of hilariously fragile interfaces.
1484          */
1485         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1486                                     (char __user *)&val, sizeof(val));
1487
1488         set_fs(oldfs);
1489         return ret;
1490 }
1491
1492 static void o2net_initialize_handshake(void)
1493 {
1494         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1495                 O2HB_MAX_WRITE_TIMEOUT_MS);
1496         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1497         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1498                 o2net_keepalive_delay());
1499         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1500                 o2net_reconnect_delay());
1501 }
1502
1503 /* ------------------------------------------------------------ */
1504
1505 /* called when a connect completes and after a sock is accepted.  the
1506  * rx path will see the response and mark the sc valid */
1507 static void o2net_sc_connect_completed(struct work_struct *work)
1508 {
1509         struct o2net_sock_container *sc =
1510                 container_of(work, struct o2net_sock_container,
1511                              sc_connect_work);
1512
1513         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1514               (unsigned long long)O2NET_PROTOCOL_VERSION,
1515               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1516
1517         o2net_initialize_handshake();
1518         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1519         sc_put(sc);
1520 }
1521
1522 /* this is called as a work_struct func. */
1523 static void o2net_sc_send_keep_req(struct work_struct *work)
1524 {
1525         struct o2net_sock_container *sc =
1526                 container_of(work, struct o2net_sock_container,
1527                              sc_keepalive_work.work);
1528
1529         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1530         sc_put(sc);
1531 }
1532
1533 /* socket shutdown does a del_timer_sync against this as it tears down.
1534  * we can't start this timer until we've got to the point in sc buildup
1535  * where shutdown is going to be involved */
1536 static void o2net_idle_timer(unsigned long data)
1537 {
1538         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1539         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1540
1541 #ifdef CONFIG_DEBUG_FS
1542         ktime_t now = ktime_get();
1543 #endif
1544
1545         printk(KERN_NOTICE "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1546              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1547                      o2net_idle_timeout() / 1000,
1548                      o2net_idle_timeout() % 1000);
1549
1550 #ifdef CONFIG_DEBUG_FS
1551         mlog(ML_NOTICE, "Here are some times that might help debug the "
1552              "situation: (Timer: %lld, Now %lld, DataReady %lld, Advance %lld-%lld, "
1553              "Key 0x%08x, Func %u, FuncTime %lld-%lld)\n",
1554              (long long)ktime_to_us(sc->sc_tv_timer), (long long)ktime_to_us(now),
1555              (long long)ktime_to_us(sc->sc_tv_data_ready),
1556              (long long)ktime_to_us(sc->sc_tv_advance_start),
1557              (long long)ktime_to_us(sc->sc_tv_advance_stop),
1558              sc->sc_msg_key, sc->sc_msg_type,
1559              (long long)ktime_to_us(sc->sc_tv_func_start),
1560              (long long)ktime_to_us(sc->sc_tv_func_stop));
1561 #endif
1562
1563         /*
1564          * Initialize the nn_timeout so that the next connection attempt
1565          * will continue in o2net_start_connect.
1566          */
1567         atomic_set(&nn->nn_timeout, 1);
1568
1569         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1570 }
1571
1572 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1573 {
1574         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1575         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1576                       msecs_to_jiffies(o2net_keepalive_delay()));
1577         o2net_set_sock_timer(sc);
1578         mod_timer(&sc->sc_idle_timeout,
1579                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1580 }
1581
1582 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1583 {
1584         /* Only push out an existing timer */
1585         if (timer_pending(&sc->sc_idle_timeout))
1586                 o2net_sc_reset_idle_timer(sc);
1587 }
1588
1589 /* this work func is kicked whenever a path sets the nn state which doesn't
1590  * have valid set.  This includes seeing hb come up, losing a connection,
1591  * having a connect attempt fail, etc. This centralizes the logic which decides
1592  * if a connect attempt should be made or if we should give up and all future
1593  * transmit attempts should fail */
1594 static void o2net_start_connect(struct work_struct *work)
1595 {
1596         struct o2net_node *nn =
1597                 container_of(work, struct o2net_node, nn_connect_work.work);
1598         struct o2net_sock_container *sc = NULL;
1599         struct o2nm_node *node = NULL, *mynode = NULL;
1600         struct socket *sock = NULL;
1601         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1602         int ret = 0, stop;
1603         unsigned int timeout;
1604
1605         /* if we're greater we initiate tx, otherwise we accept */
1606         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1607                 goto out;
1608
1609         /* watch for racing with tearing a node down */
1610         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1611         if (node == NULL) {
1612                 ret = 0;
1613                 goto out;
1614         }
1615
1616         mynode = o2nm_get_node_by_num(o2nm_this_node());
1617         if (mynode == NULL) {
1618                 ret = 0;
1619                 goto out;
1620         }
1621
1622         spin_lock(&nn->nn_lock);
1623         /*
1624          * see if we already have one pending or have given up.
1625          * For nn_timeout, it is set when we close the connection
1626          * because of the idle time out. So it means that we have
1627          * at least connected to that node successfully once,
1628          * now try to connect to it again.
1629          */
1630         timeout = atomic_read(&nn->nn_timeout);
1631         stop = (nn->nn_sc ||
1632                 (nn->nn_persistent_error &&
1633                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1634         spin_unlock(&nn->nn_lock);
1635         if (stop)
1636                 goto out;
1637
1638         nn->nn_last_connect_attempt = jiffies;
1639
1640         sc = sc_alloc(node);
1641         if (sc == NULL) {
1642                 mlog(0, "couldn't allocate sc\n");
1643                 ret = -ENOMEM;
1644                 goto out;
1645         }
1646
1647         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1648         if (ret < 0) {
1649                 mlog(0, "can't create socket: %d\n", ret);
1650                 goto out;
1651         }
1652         sc->sc_sock = sock; /* freed by sc_kref_release */
1653
1654         sock->sk->sk_allocation = GFP_ATOMIC;
1655
1656         myaddr.sin_family = AF_INET;
1657         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1658         myaddr.sin_port = htons(0); /* any port */
1659
1660         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1661                               sizeof(myaddr));
1662         if (ret) {
1663                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1664                      ret, &mynode->nd_ipv4_address);
1665                 goto out;
1666         }
1667
1668         ret = o2net_set_nodelay(sc->sc_sock);
1669         if (ret) {
1670                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1671                 goto out;
1672         }
1673
1674         o2net_register_callbacks(sc->sc_sock->sk, sc);
1675
1676         spin_lock(&nn->nn_lock);
1677         /* handshake completion will set nn->nn_sc_valid */
1678         o2net_set_nn_state(nn, sc, 0, 0);
1679         spin_unlock(&nn->nn_lock);
1680
1681         remoteaddr.sin_family = AF_INET;
1682         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1683         remoteaddr.sin_port = node->nd_ipv4_port;
1684
1685         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1686                                         (struct sockaddr *)&remoteaddr,
1687                                         sizeof(remoteaddr),
1688                                         O_NONBLOCK);
1689         if (ret == -EINPROGRESS)
1690                 ret = 0;
1691
1692 out:
1693         if (ret) {
1694                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1695                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1696                 /* 0 err so that another will be queued and attempted
1697                  * from set_nn_state */
1698                 if (sc)
1699                         o2net_ensure_shutdown(nn, sc, 0);
1700         }
1701         if (sc)
1702                 sc_put(sc);
1703         if (node)
1704                 o2nm_node_put(node);
1705         if (mynode)
1706                 o2nm_node_put(mynode);
1707
1708         return;
1709 }
1710
1711 static void o2net_connect_expired(struct work_struct *work)
1712 {
1713         struct o2net_node *nn =
1714                 container_of(work, struct o2net_node, nn_connect_expired.work);
1715
1716         spin_lock(&nn->nn_lock);
1717         if (!nn->nn_sc_valid) {
1718                 mlog(ML_ERROR, "no connection established with node %u after "
1719                      "%u.%u seconds, giving up and returning errors.\n",
1720                      o2net_num_from_nn(nn),
1721                      o2net_idle_timeout() / 1000,
1722                      o2net_idle_timeout() % 1000);
1723
1724                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1725         }
1726         spin_unlock(&nn->nn_lock);
1727 }
1728
1729 static void o2net_still_up(struct work_struct *work)
1730 {
1731         struct o2net_node *nn =
1732                 container_of(work, struct o2net_node, nn_still_up.work);
1733
1734         o2quo_hb_still_up(o2net_num_from_nn(nn));
1735 }
1736
1737 /* ------------------------------------------------------------ */
1738
1739 void o2net_disconnect_node(struct o2nm_node *node)
1740 {
1741         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1742
1743         /* don't reconnect until it's heartbeating again */
1744         spin_lock(&nn->nn_lock);
1745         atomic_set(&nn->nn_timeout, 0);
1746         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1747         spin_unlock(&nn->nn_lock);
1748
1749         if (o2net_wq) {
1750                 cancel_delayed_work(&nn->nn_connect_expired);
1751                 cancel_delayed_work(&nn->nn_connect_work);
1752                 cancel_delayed_work(&nn->nn_still_up);
1753                 flush_workqueue(o2net_wq);
1754         }
1755 }
1756
1757 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1758                                   void *data)
1759 {
1760         o2quo_hb_down(node_num);
1761
1762         if (!node)
1763                 return;
1764
1765         if (node_num != o2nm_this_node())
1766                 o2net_disconnect_node(node);
1767
1768         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1769 }
1770
1771 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1772                                 void *data)
1773 {
1774         struct o2net_node *nn = o2net_nn_from_num(node_num);
1775
1776         o2quo_hb_up(node_num);
1777
1778         BUG_ON(!node);
1779
1780         /* ensure an immediate connect attempt */
1781         nn->nn_last_connect_attempt = jiffies -
1782                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1783
1784         if (node_num != o2nm_this_node()) {
1785                 /* believe it or not, accept and node hearbeating testing
1786                  * can succeed for this node before we got here.. so
1787                  * only use set_nn_state to clear the persistent error
1788                  * if that hasn't already happened */
1789                 spin_lock(&nn->nn_lock);
1790                 atomic_set(&nn->nn_timeout, 0);
1791                 if (nn->nn_persistent_error)
1792                         o2net_set_nn_state(nn, NULL, 0, 0);
1793                 spin_unlock(&nn->nn_lock);
1794         }
1795 }
1796
1797 void o2net_unregister_hb_callbacks(void)
1798 {
1799         o2hb_unregister_callback(NULL, &o2net_hb_up);
1800         o2hb_unregister_callback(NULL, &o2net_hb_down);
1801 }
1802
1803 int o2net_register_hb_callbacks(void)
1804 {
1805         int ret;
1806
1807         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1808                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1809         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1810                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1811
1812         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1813         if (ret == 0)
1814                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1815
1816         if (ret)
1817                 o2net_unregister_hb_callbacks();
1818
1819         return ret;
1820 }
1821
1822 /* ------------------------------------------------------------ */
1823
1824 static int o2net_accept_one(struct socket *sock)
1825 {
1826         int ret, slen;
1827         struct sockaddr_in sin;
1828         struct socket *new_sock = NULL;
1829         struct o2nm_node *node = NULL;
1830         struct o2nm_node *local_node = NULL;
1831         struct o2net_sock_container *sc = NULL;
1832         struct o2net_node *nn;
1833
1834         BUG_ON(sock == NULL);
1835         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1836                                sock->sk->sk_protocol, &new_sock);
1837         if (ret)
1838                 goto out;
1839
1840         new_sock->type = sock->type;
1841         new_sock->ops = sock->ops;
1842         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1843         if (ret < 0)
1844                 goto out;
1845
1846         new_sock->sk->sk_allocation = GFP_ATOMIC;
1847
1848         ret = o2net_set_nodelay(new_sock);
1849         if (ret) {
1850                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1851                 goto out;
1852         }
1853
1854         slen = sizeof(sin);
1855         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1856                                        &slen, 1);
1857         if (ret < 0)
1858                 goto out;
1859
1860         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1861         if (node == NULL) {
1862                 mlog(ML_NOTICE, "attempt to connect from unknown node at %pI4:%d\n",
1863                      &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1864                 ret = -EINVAL;
1865                 goto out;
1866         }
1867
1868         if (o2nm_this_node() >= node->nd_num) {
1869                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1870                 mlog(ML_NOTICE, "unexpected connect attempt seen at node '%s' ("
1871                      "%u, %pI4:%d) from node '%s' (%u, %pI4:%d)\n",
1872                      local_node->nd_name, local_node->nd_num,
1873                      &(local_node->nd_ipv4_address),
1874                      ntohs(local_node->nd_ipv4_port),
1875                      node->nd_name, node->nd_num, &sin.sin_addr.s_addr,
1876                      ntohs(sin.sin_port));
1877                 ret = -EINVAL;
1878                 goto out;
1879         }
1880
1881         /* this happens all the time when the other node sees our heartbeat
1882          * and tries to connect before we see their heartbeat */
1883         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1884                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1885                      "%pI4:%d but it isn't heartbeating\n",
1886                      node->nd_name, &sin.sin_addr.s_addr,
1887                      ntohs(sin.sin_port));
1888                 ret = -EINVAL;
1889                 goto out;
1890         }
1891
1892         nn = o2net_nn_from_num(node->nd_num);
1893
1894         spin_lock(&nn->nn_lock);
1895         if (nn->nn_sc)
1896                 ret = -EBUSY;
1897         else
1898                 ret = 0;
1899         spin_unlock(&nn->nn_lock);
1900         if (ret) {
1901                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1902                      "%pI4:%d but it already has an open connection\n",
1903                      node->nd_name, &sin.sin_addr.s_addr,
1904                      ntohs(sin.sin_port));
1905                 goto out;
1906         }
1907
1908         sc = sc_alloc(node);
1909         if (sc == NULL) {
1910                 ret = -ENOMEM;
1911                 goto out;
1912         }
1913
1914         sc->sc_sock = new_sock;
1915         new_sock = NULL;
1916
1917         spin_lock(&nn->nn_lock);
1918         atomic_set(&nn->nn_timeout, 0);
1919         o2net_set_nn_state(nn, sc, 0, 0);
1920         spin_unlock(&nn->nn_lock);
1921
1922         o2net_register_callbacks(sc->sc_sock->sk, sc);
1923         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1924
1925         o2net_initialize_handshake();
1926         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1927
1928 out:
1929         if (new_sock)
1930                 sock_release(new_sock);
1931         if (node)
1932                 o2nm_node_put(node);
1933         if (local_node)
1934                 o2nm_node_put(local_node);
1935         if (sc)
1936                 sc_put(sc);
1937         return ret;
1938 }
1939
1940 static void o2net_accept_many(struct work_struct *work)
1941 {
1942         struct socket *sock = o2net_listen_sock;
1943         while (o2net_accept_one(sock) == 0)
1944                 cond_resched();
1945 }
1946
1947 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1948 {
1949         void (*ready)(struct sock *sk, int bytes);
1950
1951         read_lock(&sk->sk_callback_lock);
1952         ready = sk->sk_user_data;
1953         if (ready == NULL) { /* check for teardown race */
1954                 ready = sk->sk_data_ready;
1955                 goto out;
1956         }
1957
1958         /* ->sk_data_ready is also called for a newly established child socket
1959          * before it has been accepted and the acceptor has set up their
1960          * data_ready.. we only want to queue listen work for our listening
1961          * socket */
1962         if (sk->sk_state == TCP_LISTEN) {
1963                 mlog(ML_TCP, "bytes: %d\n", bytes);
1964                 queue_work(o2net_wq, &o2net_listen_work);
1965         }
1966
1967 out:
1968         read_unlock(&sk->sk_callback_lock);
1969         ready(sk, bytes);
1970 }
1971
1972 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1973 {
1974         struct socket *sock = NULL;
1975         int ret;
1976         struct sockaddr_in sin = {
1977                 .sin_family = PF_INET,
1978                 .sin_addr = { .s_addr = addr },
1979                 .sin_port = port,
1980         };
1981
1982         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1983         if (ret < 0) {
1984                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1985                 goto out;
1986         }
1987
1988         sock->sk->sk_allocation = GFP_ATOMIC;
1989
1990         write_lock_bh(&sock->sk->sk_callback_lock);
1991         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1992         sock->sk->sk_data_ready = o2net_listen_data_ready;
1993         write_unlock_bh(&sock->sk->sk_callback_lock);
1994
1995         o2net_listen_sock = sock;
1996         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1997
1998         sock->sk->sk_reuse = 1;
1999         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2000         if (ret < 0) {
2001                 mlog(ML_ERROR, "unable to bind socket at %pI4:%u, "
2002                      "ret=%d\n", &addr, ntohs(port), ret);
2003                 goto out;
2004         }
2005
2006         ret = sock->ops->listen(sock, 64);
2007         if (ret < 0) {
2008                 mlog(ML_ERROR, "unable to listen on %pI4:%u, ret=%d\n",
2009                      &addr, ntohs(port), ret);
2010         }
2011
2012 out:
2013         if (ret) {
2014                 o2net_listen_sock = NULL;
2015                 if (sock)
2016                         sock_release(sock);
2017         }
2018         return ret;
2019 }
2020
2021 /*
2022  * called from node manager when we should bring up our network listening
2023  * socket.  node manager handles all the serialization to only call this
2024  * once and to match it with o2net_stop_listening().  note,
2025  * o2nm_this_node() doesn't work yet as we're being called while it
2026  * is being set up.
2027  */
2028 int o2net_start_listening(struct o2nm_node *node)
2029 {
2030         int ret = 0;
2031
2032         BUG_ON(o2net_wq != NULL);
2033         BUG_ON(o2net_listen_sock != NULL);
2034
2035         mlog(ML_KTHREAD, "starting o2net thread...\n");
2036         o2net_wq = create_singlethread_workqueue("o2net");
2037         if (o2net_wq == NULL) {
2038                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2039                 return -ENOMEM; /* ? */
2040         }
2041
2042         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2043                                         node->nd_ipv4_port);
2044         if (ret) {
2045                 destroy_workqueue(o2net_wq);
2046                 o2net_wq = NULL;
2047         } else
2048                 o2quo_conn_up(node->nd_num);
2049
2050         return ret;
2051 }
2052
2053 /* again, o2nm_this_node() doesn't work here as we're involved in
2054  * tearing it down */
2055 void o2net_stop_listening(struct o2nm_node *node)
2056 {
2057         struct socket *sock = o2net_listen_sock;
2058         size_t i;
2059
2060         BUG_ON(o2net_wq == NULL);
2061         BUG_ON(o2net_listen_sock == NULL);
2062
2063         /* stop the listening socket from generating work */
2064         write_lock_bh(&sock->sk->sk_callback_lock);
2065         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2066         sock->sk->sk_user_data = NULL;
2067         write_unlock_bh(&sock->sk->sk_callback_lock);
2068
2069         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2070                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2071                 if (node) {
2072                         o2net_disconnect_node(node);
2073                         o2nm_node_put(node);
2074                 }
2075         }
2076
2077         /* finish all work and tear down the work queue */
2078         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2079         destroy_workqueue(o2net_wq);
2080         o2net_wq = NULL;
2081
2082         sock_release(o2net_listen_sock);
2083         o2net_listen_sock = NULL;
2084
2085         o2quo_conn_err(node->nd_num);
2086 }
2087
2088 /* ------------------------------------------------------------ */
2089
2090 int o2net_init(void)
2091 {
2092         unsigned long i;
2093
2094         o2quo_init();
2095
2096         if (o2net_debugfs_init())
2097                 return -ENOMEM;
2098
2099         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2100         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2101         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2102         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2103                 kfree(o2net_hand);
2104                 kfree(o2net_keep_req);
2105                 kfree(o2net_keep_resp);
2106                 return -ENOMEM;
2107         }
2108
2109         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2110         o2net_hand->connector_id = cpu_to_be64(1);
2111
2112         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2113         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2114
2115         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2116                 struct o2net_node *nn = o2net_nn_from_num(i);
2117
2118                 atomic_set(&nn->nn_timeout, 0);
2119                 spin_lock_init(&nn->nn_lock);
2120                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2121                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2122                                   o2net_connect_expired);
2123                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2124                 /* until we see hb from a node we'll return einval */
2125                 nn->nn_persistent_error = -ENOTCONN;
2126                 init_waitqueue_head(&nn->nn_sc_wq);
2127                 idr_init(&nn->nn_status_idr);
2128                 INIT_LIST_HEAD(&nn->nn_status_list);
2129         }
2130
2131         return 0;
2132 }
2133
2134 void o2net_exit(void)
2135 {
2136         o2quo_exit();
2137         kfree(o2net_hand);
2138         kfree(o2net_keep_req);
2139         kfree(o2net_keep_resp);
2140         o2net_debugfs_exit();
2141 }