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         struct kvec vec = { .iov_len = len, .iov_base = data, };
 920         struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
 921         return kernel_recvmsg(sock, &msg, &vec, 1, len, msg.msg_flags);
 922 }
 923
 924 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 925                               size_t veclen, size_t total)
 926 {
 927         int ret;
 928         struct msghdr msg;
 929
 930         if (sock == NULL) {
 931                 ret = -EINVAL;
 932                 goto out;
 933         }
 934
 935         ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
 936         if (likely(ret == total))
 937                 return 0;
 938         mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
 939         if (ret >= 0)
 940                 ret = -EPIPE; /* should be smarter, I bet */
 941 out:
 942         mlog(0, "returning error: %d\n", ret);
 943         return ret;
 944 }
 945
 946 static void o2net_sendpage(struct o2net_sock_container *sc,
 947                            void *kmalloced_virt,
 948                            size_t size)
 949 {
 950         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 951         ssize_t ret;
 952
 953         while (1) {
 954                 mutex_lock(&sc->sc_send_lock);
 955                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 956                                                  virt_to_page(kmalloced_virt),
 957                                                  (long)kmalloced_virt & ~PAGE_MASK,
 958                                                  size, MSG_DONTWAIT);
 959                 mutex_unlock(&sc->sc_send_lock);
 960                 if (ret == size)
 961                         break;
 962                 if (ret == (ssize_t)-EAGAIN) {
 963                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 964                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 965                         cond_resched();
 966                         continue;
 967                 }
 968                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 969                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 970                 o2net_ensure_shutdown(nn, sc, 0);
 971                 break;
 972         }
 973 }
 974
 975 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 976 {
 977         memset(msg, 0, sizeof(struct o2net_msg));
 978         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 979         msg->data_len = cpu_to_be16(data_len);
 980         msg->msg_type = cpu_to_be16(msg_type);
 981         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 982         msg->status = 0;
 983         msg->key = cpu_to_be32(key);
 984 }
 985
 986 static int o2net_tx_can_proceed(struct o2net_node *nn,
 987                                 struct o2net_sock_container **sc_ret,
 988                                 int *error)
 989 {
 990         int ret = 0;
 991
 992         spin_lock(&nn->nn_lock);
 993         if (nn->nn_persistent_error) {
 994                 ret = 1;
 995                 *sc_ret = NULL;
 996                 *error = nn->nn_persistent_error;
 997         } else if (nn->nn_sc_valid) {
 998                 kref_get(&nn->nn_sc->sc_kref);
 999
1000                 ret = 1;
1001                 *sc_ret = nn->nn_sc;
1002                 *error = 0;
1003         }
1004         spin_unlock(&nn->nn_lock);
1005
1006         return ret;
1007 }
1008
1009 /* Get a map of all nodes to which this node is currently connected to */
1010 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1011 {
1012         struct o2net_sock_container *sc;
1013         int node, ret;
1014
1015         BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1016
1017         memset(map, 0, bytes);
1018         for (node = 0; node < O2NM_MAX_NODES; ++node) {
1019                 o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret);
1020                 if (!ret) {
1021                         set_bit(node, map);
1022                         sc_put(sc);
1023                 }
1024         }
1025 }
1026 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1027
1028 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1029                            size_t caller_veclen, u8 target_node, int *status)
1030 {
1031         int ret = 0;
1032         struct o2net_msg *msg = NULL;
1033         size_t veclen, caller_bytes = 0;
1034         struct kvec *vec = NULL;
1035         struct o2net_sock_container *sc = NULL;
1036         struct o2net_node *nn = o2net_nn_from_num(target_node);
1037         struct o2net_status_wait nsw = {
1038                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1039         };
1040         struct o2net_send_tracking nst;
1041
1042         o2net_init_nst(&nst, msg_type, key, current, target_node);
1043
1044         if (o2net_wq == NULL) {
1045                 mlog(0, "attempt to tx without o2netd running\n");
1046                 ret = -ESRCH;
1047                 goto out;
1048         }
1049
1050         if (caller_veclen == 0) {
1051                 mlog(0, "bad kvec array length\n");
1052                 ret = -EINVAL;
1053                 goto out;
1054         }
1055
1056         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1057         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1058                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1059                 ret = -EINVAL;
1060                 goto out;
1061         }
1062
1063         if (target_node == o2nm_this_node()) {
1064                 ret = -ELOOP;
1065                 goto out;
1066         }
1067
1068         o2net_debug_add_nst(&nst);
1069
1070         o2net_set_nst_sock_time(&nst);
1071
1072         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1073         if (ret)
1074                 goto out;
1075
1076         o2net_set_nst_sock_container(&nst, sc);
1077
1078         veclen = caller_veclen + 1;
1079         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1080         if (vec == NULL) {
1081                 mlog(0, "failed to %zu element kvec!\n", veclen);
1082                 ret = -ENOMEM;
1083                 goto out;
1084         }
1085
1086         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1087         if (!msg) {
1088                 mlog(0, "failed to allocate a o2net_msg!\n");
1089                 ret = -ENOMEM;
1090                 goto out;
1091         }
1092
1093         o2net_init_msg(msg, caller_bytes, msg_type, key);
1094
1095         vec[0].iov_len = sizeof(struct o2net_msg);
1096         vec[0].iov_base = msg;
1097         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1098
1099         ret = o2net_prep_nsw(nn, &nsw);
1100         if (ret)
1101                 goto out;
1102
1103         msg->msg_num = cpu_to_be32(nsw.ns_id);
1104         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1105
1106         o2net_set_nst_send_time(&nst);
1107
1108         /* finally, convert the message header to network byte-order
1109          * and send */
1110         mutex_lock(&sc->sc_send_lock);
1111         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1112                                  sizeof(struct o2net_msg) + caller_bytes);
1113         mutex_unlock(&sc->sc_send_lock);
1114         msglog(msg, "sending returned %d\n", ret);
1115         if (ret < 0) {
1116                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1117                 goto out;
1118         }
1119
1120         /* wait on other node's handler */
1121         o2net_set_nst_status_time(&nst);
1122         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1123
1124         o2net_update_send_stats(&nst, sc);
1125
1126         /* Note that we avoid overwriting the callers status return
1127          * variable if a system error was reported on the other
1128          * side. Callers beware. */
1129         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1130         if (status && !ret)
1131                 *status = nsw.ns_status;
1132
1133         mlog(0, "woken, returning system status %d, user status %d\n",
1134              ret, nsw.ns_status);
1135 out:
1136         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1137         if (sc)
1138                 sc_put(sc);
1139         kfree(vec);
1140         kfree(msg);
1141         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1142         return ret;
1143 }
1144 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1145
1146 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1147                        u8 target_node, int *status)
1148 {
1149         struct kvec vec = {
1150                 .iov_base = data,
1151                 .iov_len = len,
1152         };
1153         return o2net_send_message_vec(msg_type, key, &vec, 1,
1154                                       target_node, status);
1155 }
1156 EXPORT_SYMBOL_GPL(o2net_send_message);
1157
1158 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1159                                    enum o2net_system_error syserr, int err)
1160 {
1161         struct kvec vec = {
1162                 .iov_base = hdr,
1163                 .iov_len = sizeof(struct o2net_msg),
1164         };
1165
1166         BUG_ON(syserr >= O2NET_ERR_MAX);
1167
1168         /* leave other fields intact from the incoming message, msg_num
1169          * in particular */
1170         hdr->sys_status = cpu_to_be32(syserr);
1171         hdr->status = cpu_to_be32(err);
1172         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1173         hdr->data_len = 0;
1174
1175         msglog(hdr, "about to send status magic %d\n", err);
1176         /* hdr has been in host byteorder this whole time */
1177         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1178 }
1179
1180 /* this returns -errno if the header was unknown or too large, etc.
1181  * after this is called the buffer us reused for the next message */
1182 static int o2net_process_message(struct o2net_sock_container *sc,
1183                                  struct o2net_msg *hdr)
1184 {
1185         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1186         int ret = 0, handler_status;
1187         enum  o2net_system_error syserr;
1188         struct o2net_msg_handler *nmh = NULL;
1189         void *ret_data = NULL;
1190
1191         msglog(hdr, "processing message\n");
1192
1193         o2net_sc_postpone_idle(sc);
1194
1195         switch(be16_to_cpu(hdr->magic)) {
1196                 case O2NET_MSG_STATUS_MAGIC:
1197                         /* special type for returning message status */
1198                         o2net_complete_nsw(nn, NULL,
1199                                            be32_to_cpu(hdr->msg_num),
1200                                            be32_to_cpu(hdr->sys_status),
1201                                            be32_to_cpu(hdr->status));
1202                         goto out;
1203                 case O2NET_MSG_KEEP_REQ_MAGIC:
1204                         o2net_sendpage(sc, o2net_keep_resp,
1205                                        sizeof(*o2net_keep_resp));
1206                         goto out;
1207                 case O2NET_MSG_KEEP_RESP_MAGIC:
1208                         goto out;
1209                 case O2NET_MSG_MAGIC:
1210                         break;
1211                 default:
1212                         msglog(hdr, "bad magic\n");
1213                         ret = -EINVAL;
1214                         goto out;
1215                         break;
1216         }
1217
1218         /* find a handler for it */
1219         handler_status = 0;
1220         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1221                                 be32_to_cpu(hdr->key));
1222         if (!nmh) {
1223                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1224                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1225                 syserr = O2NET_ERR_NO_HNDLR;
1226                 goto out_respond;
1227         }
1228
1229         syserr = O2NET_ERR_NONE;
1230
1231         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1232                 syserr = O2NET_ERR_OVERFLOW;
1233
1234         if (syserr != O2NET_ERR_NONE)
1235                 goto out_respond;
1236
1237         o2net_set_func_start_time(sc);
1238         sc->sc_msg_key = be32_to_cpu(hdr->key);
1239         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1240         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1241                                              be16_to_cpu(hdr->data_len),
1242                                         nmh->nh_func_data, &ret_data);
1243         o2net_set_func_stop_time(sc);
1244
1245         o2net_update_recv_stats(sc);
1246
1247 out_respond:
1248         /* this destroys the hdr, so don't use it after this */
1249         mutex_lock(&sc->sc_send_lock);
1250         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1251                                       handler_status);
1252         mutex_unlock(&sc->sc_send_lock);
1253         hdr = NULL;
1254         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1255              handler_status, syserr, ret);
1256
1257         if (nmh) {
1258                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1259                 if (nmh->nh_post_func)
1260                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1261                                             ret_data);
1262         }
1263
1264 out:
1265         if (nmh)
1266                 o2net_handler_put(nmh);
1267         return ret;
1268 }
1269
1270 static int o2net_check_handshake(struct o2net_sock_container *sc)
1271 {
1272         struct o2net_handshake *hand = page_address(sc->sc_page);
1273         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1274
1275         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1276                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1277                        "protocol version %llu but %llu is required. "
1278                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1279                        (unsigned long long)be64_to_cpu(hand->protocol_version),
1280                        O2NET_PROTOCOL_VERSION);
1281
1282                 /* don't bother reconnecting if its the wrong version. */
1283                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1284                 return -1;
1285         }
1286
1287         /*
1288          * Ensure timeouts are consistent with other nodes, otherwise
1289          * we can end up with one node thinking that the other must be down,
1290          * but isn't. This can ultimately cause corruption.
1291          */
1292         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1293                                 o2net_idle_timeout()) {
1294                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1295                        "idle timeout of %u ms, but we use %u ms locally. "
1296                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1297                        be32_to_cpu(hand->o2net_idle_timeout_ms),
1298                        o2net_idle_timeout());
1299                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1300                 return -1;
1301         }
1302
1303         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1304                         o2net_keepalive_delay()) {
1305                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1306                        "delay of %u ms, but we use %u ms locally. "
1307                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1308                        be32_to_cpu(hand->o2net_keepalive_delay_ms),
1309                        o2net_keepalive_delay());
1310                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1311                 return -1;
1312         }
1313
1314         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1315                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1316                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1317                        "timeout of %u ms, but we use %u ms locally. "
1318                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1319                        be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1320                        O2HB_MAX_WRITE_TIMEOUT_MS);
1321                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1322                 return -1;
1323         }
1324
1325         sc->sc_handshake_ok = 1;
1326
1327         spin_lock(&nn->nn_lock);
1328         /* set valid and queue the idle timers only if it hasn't been
1329          * shut down already */
1330         if (nn->nn_sc == sc) {
1331                 o2net_sc_reset_idle_timer(sc);
1332                 atomic_set(&nn->nn_timeout, 0);
1333                 o2net_set_nn_state(nn, sc, 1, 0);
1334         }
1335         spin_unlock(&nn->nn_lock);
1336
1337         /* shift everything up as though it wasn't there */
1338         sc->sc_page_off -= sizeof(struct o2net_handshake);
1339         if (sc->sc_page_off)
1340                 memmove(hand, hand + 1, sc->sc_page_off);
1341
1342         return 0;
1343 }
1344
1345 /* this demuxes the queued rx bytes into header or payload bits and calls
1346  * handlers as each full message is read off the socket.  it returns -error,
1347  * == 0 eof, or > 0 for progress made.*/
1348 static int o2net_advance_rx(struct o2net_sock_container *sc)
1349 {
1350         struct o2net_msg *hdr;
1351         int ret = 0;
1352         void *data;
1353         size_t datalen;
1354
1355         sclog(sc, "receiving\n");
1356         o2net_set_advance_start_time(sc);
1357
1358         if (unlikely(sc->sc_handshake_ok == 0)) {
1359                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1360                         data = page_address(sc->sc_page) + sc->sc_page_off;
1361                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1362                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1363                         if (ret > 0)
1364                                 sc->sc_page_off += ret;
1365                 }
1366
1367                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1368                         o2net_check_handshake(sc);
1369                         if (unlikely(sc->sc_handshake_ok == 0))
1370                                 ret = -EPROTO;
1371                 }
1372                 goto out;
1373         }
1374
1375         /* do we need more header? */
1376         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1377                 data = page_address(sc->sc_page) + sc->sc_page_off;
1378                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1379                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1380                 if (ret > 0) {
1381                         sc->sc_page_off += ret;
1382                         /* only swab incoming here.. we can
1383                          * only get here once as we cross from
1384                          * being under to over */
1385                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1386                                 hdr = page_address(sc->sc_page);
1387                                 if (be16_to_cpu(hdr->data_len) >
1388                                     O2NET_MAX_PAYLOAD_BYTES)
1389                                         ret = -EOVERFLOW;
1390                         }
1391                 }
1392                 if (ret <= 0)
1393                         goto out;
1394         }
1395
1396         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1397                 /* oof, still don't have a header */
1398                 goto out;
1399         }
1400
1401         /* this was swabbed above when we first read it */
1402         hdr = page_address(sc->sc_page);
1403
1404         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1405
1406         /* do we need more payload? */
1407         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1408                 /* need more payload */
1409                 data = page_address(sc->sc_page) + sc->sc_page_off;
1410                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1411                           sc->sc_page_off;
1412                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1413                 if (ret > 0)
1414                         sc->sc_page_off += ret;
1415                 if (ret <= 0)
1416                         goto out;
1417         }
1418
1419         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1420                 /* we can only get here once, the first time we read
1421                  * the payload.. so set ret to progress if the handler
1422                  * works out. after calling this the message is toast */
1423                 ret = o2net_process_message(sc, hdr);
1424                 if (ret == 0)
1425                         ret = 1;
1426                 sc->sc_page_off = 0;
1427         }
1428
1429 out:
1430         sclog(sc, "ret = %d\n", ret);
1431         o2net_set_advance_stop_time(sc);
1432         return ret;
1433 }
1434
1435 /* this work func is triggerd by data ready.  it reads until it can read no
1436  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1437  * our work the work struct will be marked and we'll be called again. */
1438 static void o2net_rx_until_empty(struct work_struct *work)
1439 {
1440         struct o2net_sock_container *sc =
1441                 container_of(work, struct o2net_sock_container, sc_rx_work);
1442         int ret;
1443
1444         do {
1445                 ret = o2net_advance_rx(sc);
1446         } while (ret > 0);
1447
1448         if (ret <= 0 && ret != -EAGAIN) {
1449                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1450                 sclog(sc, "saw error %d, closing\n", ret);
1451                 /* not permanent so read failed handshake can retry */
1452                 o2net_ensure_shutdown(nn, sc, 0);
1453         }
1454
1455         sc_put(sc);
1456 }
1457
1458 static int o2net_set_nodelay(struct socket *sock)
1459 {
1460         int ret, val = 1;
1461         mm_segment_t oldfs;
1462
1463         oldfs = get_fs();
1464         set_fs(KERNEL_DS);
1465
1466         /*
1467          * Dear unsuspecting programmer,
1468          *
1469          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1470          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1471          * silently turn into SO_DEBUG.
1472          *
1473          * Yours,
1474          * Keeper of hilariously fragile interfaces.
1475          */
1476         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1477                                     (char __user *)&val, sizeof(val));
1478
1479         set_fs(oldfs);
1480         return ret;
1481 }
1482
1483 static void o2net_initialize_handshake(void)
1484 {
1485         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1486                 O2HB_MAX_WRITE_TIMEOUT_MS);
1487         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1488         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1489                 o2net_keepalive_delay());
1490         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1491                 o2net_reconnect_delay());
1492 }
1493
1494 /* ------------------------------------------------------------ */
1495
1496 /* called when a connect completes and after a sock is accepted.  the
1497  * rx path will see the response and mark the sc valid */
1498 static void o2net_sc_connect_completed(struct work_struct *work)
1499 {
1500         struct o2net_sock_container *sc =
1501                 container_of(work, struct o2net_sock_container,
1502                              sc_connect_work);
1503
1504         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1505               (unsigned long long)O2NET_PROTOCOL_VERSION,
1506               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1507
1508         o2net_initialize_handshake();
1509         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1510         sc_put(sc);
1511 }
1512
1513 /* this is called as a work_struct func. */
1514 static void o2net_sc_send_keep_req(struct work_struct *work)
1515 {
1516         struct o2net_sock_container *sc =
1517                 container_of(work, struct o2net_sock_container,
1518                              sc_keepalive_work.work);
1519
1520         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1521         sc_put(sc);
1522 }
1523
1524 /* socket shutdown does a del_timer_sync against this as it tears down.
1525  * we can't start this timer until we've got to the point in sc buildup
1526  * where shutdown is going to be involved */
1527 static void o2net_idle_timer(unsigned long data)
1528 {
1529         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1530         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1531 #ifdef CONFIG_DEBUG_FS
1532         unsigned long msecs = ktime_to_ms(ktime_get()) -
1533                 ktime_to_ms(sc->sc_tv_timer);
1534 #else
1535         unsigned long msecs = o2net_idle_timeout();
1536 #endif
1537
1538         printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1539                "idle for %lu.%lu secs, shutting it down.\n", SC_NODEF_ARGS(sc),
1540                msecs / 1000, msecs % 1000);
1541
1542         /*
1543          * Initialize the nn_timeout so that the next connection attempt
1544          * will continue in o2net_start_connect.
1545          */
1546         atomic_set(&nn->nn_timeout, 1);
1547
1548         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1549 }
1550
1551 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1552 {
1553         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1554         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1555                       msecs_to_jiffies(o2net_keepalive_delay()));
1556         o2net_set_sock_timer(sc);
1557         mod_timer(&sc->sc_idle_timeout,
1558                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1559 }
1560
1561 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1562 {
1563         /* Only push out an existing timer */
1564         if (timer_pending(&sc->sc_idle_timeout))
1565                 o2net_sc_reset_idle_timer(sc);
1566 }
1567
1568 /* this work func is kicked whenever a path sets the nn state which doesn't
1569  * have valid set.  This includes seeing hb come up, losing a connection,
1570  * having a connect attempt fail, etc. This centralizes the logic which decides
1571  * if a connect attempt should be made or if we should give up and all future
1572  * transmit attempts should fail */
1573 static void o2net_start_connect(struct work_struct *work)
1574 {
1575         struct o2net_node *nn =
1576                 container_of(work, struct o2net_node, nn_connect_work.work);
1577         struct o2net_sock_container *sc = NULL;
1578         struct o2nm_node *node = NULL, *mynode = NULL;
1579         struct socket *sock = NULL;
1580         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1581         int ret = 0, stop;
1582         unsigned int timeout;
1583
1584         /* if we're greater we initiate tx, otherwise we accept */
1585         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1586                 goto out;
1587
1588         /* watch for racing with tearing a node down */
1589         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1590         if (node == NULL) {
1591                 ret = 0;
1592                 goto out;
1593         }
1594
1595         mynode = o2nm_get_node_by_num(o2nm_this_node());
1596         if (mynode == NULL) {
1597                 ret = 0;
1598                 goto out;
1599         }
1600
1601         spin_lock(&nn->nn_lock);
1602         /*
1603          * see if we already have one pending or have given up.
1604          * For nn_timeout, it is set when we close the connection
1605          * because of the idle time out. So it means that we have
1606          * at least connected to that node successfully once,
1607          * now try to connect to it again.
1608          */
1609         timeout = atomic_read(&nn->nn_timeout);
1610         stop = (nn->nn_sc ||
1611                 (nn->nn_persistent_error &&
1612                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1613         spin_unlock(&nn->nn_lock);
1614         if (stop)
1615                 goto out;
1616
1617         nn->nn_last_connect_attempt = jiffies;
1618
1619         sc = sc_alloc(node);
1620         if (sc == NULL) {
1621                 mlog(0, "couldn't allocate sc\n");
1622                 ret = -ENOMEM;
1623                 goto out;
1624         }
1625
1626         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1627         if (ret < 0) {
1628                 mlog(0, "can't create socket: %d\n", ret);
1629                 goto out;
1630         }
1631         sc->sc_sock = sock; /* freed by sc_kref_release */
1632
1633         sock->sk->sk_allocation = GFP_ATOMIC;
1634
1635         myaddr.sin_family = AF_INET;
1636         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1637         myaddr.sin_port = htons(0); /* any port */
1638
1639         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1640                               sizeof(myaddr));
1641         if (ret) {
1642                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1643                      ret, &mynode->nd_ipv4_address);
1644                 goto out;
1645         }
1646
1647         ret = o2net_set_nodelay(sc->sc_sock);
1648         if (ret) {
1649                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1650                 goto out;
1651         }
1652
1653         o2net_register_callbacks(sc->sc_sock->sk, sc);
1654
1655         spin_lock(&nn->nn_lock);
1656         /* handshake completion will set nn->nn_sc_valid */
1657         o2net_set_nn_state(nn, sc, 0, 0);
1658         spin_unlock(&nn->nn_lock);
1659
1660         remoteaddr.sin_family = AF_INET;
1661         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1662         remoteaddr.sin_port = node->nd_ipv4_port;
1663
1664         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1665                                         (struct sockaddr *)&remoteaddr,
1666                                         sizeof(remoteaddr),
1667                                         O_NONBLOCK);
1668         if (ret == -EINPROGRESS)
1669                 ret = 0;
1670
1671 out:
1672         if (ret && sc) {
1673                 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1674                        " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1675                 /* 0 err so that another will be queued and attempted
1676                  * from set_nn_state */
1677                 o2net_ensure_shutdown(nn, sc, 0);
1678         }
1679         if (sc)
1680                 sc_put(sc);
1681         if (node)
1682                 o2nm_node_put(node);
1683         if (mynode)
1684                 o2nm_node_put(mynode);
1685
1686         return;
1687 }
1688
1689 static void o2net_connect_expired(struct work_struct *work)
1690 {
1691         struct o2net_node *nn =
1692                 container_of(work, struct o2net_node, nn_connect_expired.work);
1693
1694         spin_lock(&nn->nn_lock);
1695         if (!nn->nn_sc_valid) {
1696                 printk(KERN_NOTICE "o2net: No connection established with "
1697                        "node %u after %u.%u seconds, giving up.\n",
1698                      o2net_num_from_nn(nn),
1699                      o2net_idle_timeout() / 1000,
1700                      o2net_idle_timeout() % 1000);
1701
1702                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1703         }
1704         spin_unlock(&nn->nn_lock);
1705 }
1706
1707 static void o2net_still_up(struct work_struct *work)
1708 {
1709         struct o2net_node *nn =
1710                 container_of(work, struct o2net_node, nn_still_up.work);
1711
1712         o2quo_hb_still_up(o2net_num_from_nn(nn));
1713 }
1714
1715 /* ------------------------------------------------------------ */
1716
1717 void o2net_disconnect_node(struct o2nm_node *node)
1718 {
1719         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1720
1721         /* don't reconnect until it's heartbeating again */
1722         spin_lock(&nn->nn_lock);
1723         atomic_set(&nn->nn_timeout, 0);
1724         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1725         spin_unlock(&nn->nn_lock);
1726
1727         if (o2net_wq) {
1728                 cancel_delayed_work(&nn->nn_connect_expired);
1729                 cancel_delayed_work(&nn->nn_connect_work);
1730                 cancel_delayed_work(&nn->nn_still_up);
1731                 flush_workqueue(o2net_wq);
1732         }
1733 }
1734
1735 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1736                                   void *data)
1737 {
1738         o2quo_hb_down(node_num);
1739
1740         if (!node)
1741                 return;
1742
1743         if (node_num != o2nm_this_node())
1744                 o2net_disconnect_node(node);
1745
1746         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1747 }
1748
1749 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1750                                 void *data)
1751 {
1752         struct o2net_node *nn = o2net_nn_from_num(node_num);
1753
1754         o2quo_hb_up(node_num);
1755
1756         BUG_ON(!node);
1757
1758         /* ensure an immediate connect attempt */
1759         nn->nn_last_connect_attempt = jiffies -
1760                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1761
1762         if (node_num != o2nm_this_node()) {
1763                 /* believe it or not, accept and node hearbeating testing
1764                  * can succeed for this node before we got here.. so
1765                  * only use set_nn_state to clear the persistent error
1766                  * if that hasn't already happened */
1767                 spin_lock(&nn->nn_lock);
1768                 atomic_set(&nn->nn_timeout, 0);
1769                 if (nn->nn_persistent_error)
1770                         o2net_set_nn_state(nn, NULL, 0, 0);
1771                 spin_unlock(&nn->nn_lock);
1772         }
1773 }
1774
1775 void o2net_unregister_hb_callbacks(void)
1776 {
1777         o2hb_unregister_callback(NULL, &o2net_hb_up);
1778         o2hb_unregister_callback(NULL, &o2net_hb_down);
1779 }
1780
1781 int o2net_register_hb_callbacks(void)
1782 {
1783         int ret;
1784
1785         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1786                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1787         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1788                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1789
1790         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1791         if (ret == 0)
1792                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1793
1794         if (ret)
1795                 o2net_unregister_hb_callbacks();
1796
1797         return ret;
1798 }
1799
1800 /* ------------------------------------------------------------ */
1801
1802 static int o2net_accept_one(struct socket *sock)
1803 {
1804         int ret, slen;
1805         struct sockaddr_in sin;
1806         struct socket *new_sock = NULL;
1807         struct o2nm_node *node = NULL;
1808         struct o2nm_node *local_node = NULL;
1809         struct o2net_sock_container *sc = NULL;
1810         struct o2net_node *nn;
1811
1812         BUG_ON(sock == NULL);
1813         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1814                                sock->sk->sk_protocol, &new_sock);
1815         if (ret)
1816                 goto out;
1817
1818         new_sock->type = sock->type;
1819         new_sock->ops = sock->ops;
1820         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1821         if (ret < 0)
1822                 goto out;
1823
1824         new_sock->sk->sk_allocation = GFP_ATOMIC;
1825
1826         ret = o2net_set_nodelay(new_sock);
1827         if (ret) {
1828                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1829                 goto out;
1830         }
1831
1832         slen = sizeof(sin);
1833         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1834                                        &slen, 1);
1835         if (ret < 0)
1836                 goto out;
1837
1838         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1839         if (node == NULL) {
1840                 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1841                        "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1842                        ntohs(sin.sin_port));
1843                 ret = -EINVAL;
1844                 goto out;
1845         }
1846
1847         if (o2nm_this_node() >= node->nd_num) {
1848                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1849                 if (local_node)
1850                         printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1851                                         "seen at node '%s' (%u, %pI4:%d) from "
1852                                         "node '%s' (%u, %pI4:%d)\n",
1853                                         local_node->nd_name, local_node->nd_num,
1854                                         &(local_node->nd_ipv4_address),
1855                                         ntohs(local_node->nd_ipv4_port),
1856                                         node->nd_name,
1857                                         node->nd_num, &sin.sin_addr.s_addr,
1858                                         ntohs(sin.sin_port));
1859                 ret = -EINVAL;
1860                 goto out;
1861         }
1862
1863         /* this happens all the time when the other node sees our heartbeat
1864          * and tries to connect before we see their heartbeat */
1865         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1866                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1867                      "%pI4:%d but it isn't heartbeating\n",
1868                      node->nd_name, &sin.sin_addr.s_addr,
1869                      ntohs(sin.sin_port));
1870                 ret = -EINVAL;
1871                 goto out;
1872         }
1873
1874         nn = o2net_nn_from_num(node->nd_num);
1875
1876         spin_lock(&nn->nn_lock);
1877         if (nn->nn_sc)
1878                 ret = -EBUSY;
1879         else
1880                 ret = 0;
1881         spin_unlock(&nn->nn_lock);
1882         if (ret) {
1883                 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1884                        "at %pI4:%d but it already has an open connection\n",
1885                        node->nd_name, &sin.sin_addr.s_addr,
1886                        ntohs(sin.sin_port));
1887                 goto out;
1888         }
1889
1890         sc = sc_alloc(node);
1891         if (sc == NULL) {
1892                 ret = -ENOMEM;
1893                 goto out;
1894         }
1895
1896         sc->sc_sock = new_sock;
1897         new_sock = NULL;
1898
1899         spin_lock(&nn->nn_lock);
1900         atomic_set(&nn->nn_timeout, 0);
1901         o2net_set_nn_state(nn, sc, 0, 0);
1902         spin_unlock(&nn->nn_lock);
1903
1904         o2net_register_callbacks(sc->sc_sock->sk, sc);
1905         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1906
1907         o2net_initialize_handshake();
1908         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1909
1910 out:
1911         if (new_sock)
1912                 sock_release(new_sock);
1913         if (node)
1914                 o2nm_node_put(node);
1915         if (local_node)
1916                 o2nm_node_put(local_node);
1917         if (sc)
1918                 sc_put(sc);
1919         return ret;
1920 }
1921
1922 static void o2net_accept_many(struct work_struct *work)
1923 {
1924         struct socket *sock = o2net_listen_sock;
1925         while (o2net_accept_one(sock) == 0)
1926                 cond_resched();
1927 }
1928
1929 static void o2net_listen_data_ready(struct sock *sk)
1930 {
1931         void (*ready)(struct sock *sk);
1932
1933         read_lock(&sk->sk_callback_lock);
1934         ready = sk->sk_user_data;
1935         if (ready == NULL) { /* check for teardown race */
1936                 ready = sk->sk_data_ready;
1937                 goto out;
1938         }
1939
1940         /* This callback may called twice when a new connection
1941          * is  being established as a child socket inherits everything
1942          * from a parent LISTEN socket, including the data_ready cb of
1943          * the parent. This leads to a hazard. In o2net_accept_one()
1944          * we are still initializing the child socket but have not
1945          * changed the inherited data_ready callback yet when
1946          * data starts arriving.
1947          * We avoid this hazard by checking the state.
1948          * For the listening socket,  the state will be TCP_LISTEN; for the new
1949          * socket, will be  TCP_ESTABLISHED. Also, in this case,
1950          * sk->sk_user_data is not a valid function pointer.
1951          */
1952
1953         if (sk->sk_state == TCP_LISTEN) {
1954                 queue_work(o2net_wq, &o2net_listen_work);
1955         } else {
1956                 ready = NULL;
1957         }
1958
1959 out:
1960         read_unlock(&sk->sk_callback_lock);
1961         if (ready != NULL)
1962                 ready(sk);
1963 }
1964
1965 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1966 {
1967         struct socket *sock = NULL;
1968         int ret;
1969         struct sockaddr_in sin = {
1970                 .sin_family = PF_INET,
1971                 .sin_addr = { .s_addr = addr },
1972                 .sin_port = port,
1973         };
1974
1975         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1976         if (ret < 0) {
1977                 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
1978                 goto out;
1979         }
1980
1981         sock->sk->sk_allocation = GFP_ATOMIC;
1982
1983         write_lock_bh(&sock->sk->sk_callback_lock);
1984         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1985         sock->sk->sk_data_ready = o2net_listen_data_ready;
1986         write_unlock_bh(&sock->sk->sk_callback_lock);
1987
1988         o2net_listen_sock = sock;
1989         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1990
1991         sock->sk->sk_reuse = SK_CAN_REUSE;
1992         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1993         if (ret < 0) {
1994                 printk(KERN_ERR "o2net: Error %d while binding socket at "
1995                        "%pI4:%u\n", ret, &addr, ntohs(port));
1996                 goto out;
1997         }
1998
1999         ret = sock->ops->listen(sock, 64);
2000         if (ret < 0)
2001                 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2002                        ret, &addr, ntohs(port));
2003
2004 out:
2005         if (ret) {
2006                 o2net_listen_sock = NULL;
2007                 if (sock)
2008                         sock_release(sock);
2009         }
2010         return ret;
2011 }
2012
2013 /*
2014  * called from node manager when we should bring up our network listening
2015  * socket.  node manager handles all the serialization to only call this
2016  * once and to match it with o2net_stop_listening().  note,
2017  * o2nm_this_node() doesn't work yet as we're being called while it
2018  * is being set up.
2019  */
2020 int o2net_start_listening(struct o2nm_node *node)
2021 {
2022         int ret = 0;
2023
2024         BUG_ON(o2net_wq != NULL);
2025         BUG_ON(o2net_listen_sock != NULL);
2026
2027         mlog(ML_KTHREAD, "starting o2net thread...\n");
2028         o2net_wq = create_singlethread_workqueue("o2net");
2029         if (o2net_wq == NULL) {
2030                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2031                 return -ENOMEM; /* ? */
2032         }
2033
2034         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2035                                         node->nd_ipv4_port);
2036         if (ret) {
2037                 destroy_workqueue(o2net_wq);
2038                 o2net_wq = NULL;
2039         } else
2040                 o2quo_conn_up(node->nd_num);
2041
2042         return ret;
2043 }
2044
2045 /* again, o2nm_this_node() doesn't work here as we're involved in
2046  * tearing it down */
2047 void o2net_stop_listening(struct o2nm_node *node)
2048 {
2049         struct socket *sock = o2net_listen_sock;
2050         size_t i;
2051
2052         BUG_ON(o2net_wq == NULL);
2053         BUG_ON(o2net_listen_sock == NULL);
2054
2055         /* stop the listening socket from generating work */
2056         write_lock_bh(&sock->sk->sk_callback_lock);
2057         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2058         sock->sk->sk_user_data = NULL;
2059         write_unlock_bh(&sock->sk->sk_callback_lock);
2060
2061         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2062                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2063                 if (node) {
2064                         o2net_disconnect_node(node);
2065                         o2nm_node_put(node);
2066                 }
2067         }
2068
2069         /* finish all work and tear down the work queue */
2070         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2071         destroy_workqueue(o2net_wq);
2072         o2net_wq = NULL;
2073
2074         sock_release(o2net_listen_sock);
2075         o2net_listen_sock = NULL;
2076
2077         o2quo_conn_err(node->nd_num);
2078 }
2079
2080 /* ------------------------------------------------------------ */
2081
2082 int o2net_init(void)
2083 {
2084         unsigned long i;
2085
2086         o2quo_init();
2087
2088         if (o2net_debugfs_init())
2089                 return -ENOMEM;
2090
2091         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2092         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2093         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2094         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2095                 kfree(o2net_hand);
2096                 kfree(o2net_keep_req);
2097                 kfree(o2net_keep_resp);
2098                 return -ENOMEM;
2099         }
2100
2101         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2102         o2net_hand->connector_id = cpu_to_be64(1);
2103
2104         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2105         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2106
2107         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2108                 struct o2net_node *nn = o2net_nn_from_num(i);
2109
2110                 atomic_set(&nn->nn_timeout, 0);
2111                 spin_lock_init(&nn->nn_lock);
2112                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2113                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2114                                   o2net_connect_expired);
2115                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2116                 /* until we see hb from a node we'll return einval */
2117                 nn->nn_persistent_error = -ENOTCONN;
2118                 init_waitqueue_head(&nn->nn_sc_wq);
2119                 idr_init(&nn->nn_status_idr);
2120                 INIT_LIST_HEAD(&nn->nn_status_list);
2121         }
2122
2123         return 0;
2124 }
2125
2126 void o2net_exit(void)
2127 {
2128         o2quo_exit();
2129         kfree(o2net_hand);
2130         kfree(o2net_keep_req);
2131         kfree(o2net_keep_resp);
2132         o2net_debugfs_exit();
2133 }