fs/ocfs2/cluster/tcp.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  *
   3  * vim: noexpandtab sw=8 ts=8 sts=0:
   4  *
   5  * Copyright (C) 2004 Oracle.  All rights reserved.
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2 of the License, or (at your option) any later version.
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public
  18  * License along with this program; if not, write to the
  19  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  20  * Boston, MA 021110-1307, USA.
  21  *
  22  * ----
  23  *
  24  * Callers for this were originally written against a very simple synchronus
  25  * API.  This implementation reflects those simple callers.  Some day I'm sure
  26  * we'll need to move to a more robust posting/callback mechanism.
  27  *
  28  * Transmit calls pass in kernel virtual addresses and block copying this into
  29  * the socket's tx buffers via a usual blocking sendmsg.  They'll block waiting
  30  * for a failed socket to timeout.  TX callers can also pass in a poniter to an
  31  * 'int' which gets filled with an errno off the wire in response to the
  32  * message they send.
  33  *
  34  * Handlers for unsolicited messages are registered.  Each socket has a page
  35  * that incoming data is copied into.  First the header, then the data.
  36  * Handlers are called from only one thread with a reference to this per-socket
  37  * page.  This page is destroyed after the handler call, so it can't be
  38  * referenced beyond the call.  Handlers may block but are discouraged from
  39  * doing so.
  40  *
  41  * Any framing errors (bad magic, large payload lengths) close a connection.
  42  *
  43  * Our sock_container holds the state we associate with a socket.  It's current
  44  * framing state is held there as well as the refcounting we do around when it
  45  * is safe to tear down the socket.  The socket is only finally torn down from
  46  * the container when the container loses all of its references -- so as long
  47  * as you hold a ref on the container you can trust that the socket is valid
  48  * for use with kernel socket APIs.
  49  *
  50  * Connections are initiated between a pair of nodes when the node with the
  51  * higher node number gets a heartbeat callback which indicates that the lower
  52  * numbered node has started heartbeating.  The lower numbered node is passive
  53  * and only accepts the connection if the higher numbered node is heartbeating.
  54  */
  55
  56 #include <linux/kernel.h>
  57 #include <linux/jiffies.h>
  58 #include <linux/slab.h>
  59 #include <linux/idr.h>
  60 #include <linux/kref.h>
  61 #include <linux/net.h>
  62 #include <linux/export.h>
  63 #include <net/tcp.h>
  64
  65 #include <asm/uaccess.h>
  66
  67 #include "heartbeat.h"
  68 #include "tcp.h"
  69 #include "nodemanager.h"
  70 #define MLOG_MASK_PREFIX ML_TCP
  71 #include "masklog.h"
  72 #include "quorum.h"
  73
  74 #include "tcp_internal.h"
  75
  76 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
  77 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
  78                           &sc->sc_node->nd_ipv4_address,                \
  79                           ntohs(sc->sc_node->nd_ipv4_port)
  80
  81 /*
  82  * In the following two log macros, the whitespace after the ',' just
  83  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  84  * previous token if args expands to nothing.
  85  */
  86 #define msglog(hdr, fmt, args...) do {                                  \
  87         typeof(hdr) __hdr = (hdr);                                      \
  88         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
  89              "key %08x num %u] " fmt,                                   \
  90              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
  91              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
  92              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
  93              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
  94 } while (0)
  95
  96 #define sclog(sc, fmt, args...) do {                                    \
  97         typeof(sc) __sc = (sc);                                         \
  98         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
  99              "pg_off %zu] " fmt, __sc,                                  \
 100              atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,       \
 101             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
 102             ##args);                                                    \
 103 } while (0)
 104
 105 static DEFINE_RWLOCK(o2net_handler_lock);
 106 static struct rb_root o2net_handler_tree = RB_ROOT;
 107
 108 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
 109
 110 /* XXX someday we'll need better accounting */
 111 static struct socket *o2net_listen_sock = NULL;
 112
 113 /*
 114  * listen work is only queued by the listening socket callbacks on the
 115  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 116  * quorum work is queued as sock containers are shutdown.. stop_listening
 117  * tears down all the node's sock containers, preventing future shutdowns
 118  * and queued quroum work, before canceling delayed quorum work and
 119  * destroying the work queue.
 120  */
 121 static struct workqueue_struct *o2net_wq;
 122 static struct work_struct o2net_listen_work;
 123
 124 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 125 #define O2NET_HB_PRI 0x1
 126
 127 static struct o2net_handshake *o2net_hand;
 128 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 129
 130 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 131                 {[O2NET_ERR_NONE]       = 0,
 132                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
 133                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
 134                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
 135
 136 /* can't quite avoid *all* internal declarations :/ */
 137 static void o2net_sc_connect_completed(struct work_struct *work);
 138 static void o2net_rx_until_empty(struct work_struct *work);
 139 static void o2net_shutdown_sc(struct work_struct *work);
 140 static void o2net_listen_data_ready(struct sock *sk, int bytes);
 141 static void o2net_sc_send_keep_req(struct work_struct *work);
 142 static void o2net_idle_timer(unsigned long data);
 143 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 144 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 145
 146 #ifdef CONFIG_DEBUG_FS
 147 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 148                            u32 msgkey, struct task_struct *task, u8 node)
 149 {
 150         INIT_LIST_HEAD(&nst->st_net_debug_item);
 151         nst->st_task = task;
 152         nst->st_msg_type = msgtype;
 153         nst->st_msg_key = msgkey;
 154         nst->st_node = node;
 155 }
 156
 157 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 158 {
 159         nst->st_sock_time = ktime_get();
 160 }
 161
 162 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 163 {
 164         nst->st_send_time = ktime_get();
 165 }
 166
 167 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 168 {
 169         nst->st_status_time = ktime_get();
 170 }
 171
 172 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 173                                                 struct o2net_sock_container *sc)
 174 {
 175         nst->st_sc = sc;
 176 }
 177
 178 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 179                                         u32 msg_id)
 180 {
 181         nst->st_id = msg_id;
 182 }
 183
 184 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
 185 {
 186         sc->sc_tv_timer = ktime_get();
 187 }
 188
 189 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
 190 {
 191         sc->sc_tv_data_ready = ktime_get();
 192 }
 193
 194 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
 195 {
 196         sc->sc_tv_advance_start = ktime_get();
 197 }
 198
 199 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
 200 {
 201         sc->sc_tv_advance_stop = ktime_get();
 202 }
 203
 204 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
 205 {
 206         sc->sc_tv_func_start = ktime_get();
 207 }
 208
 209 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
 210 {
 211         sc->sc_tv_func_stop = ktime_get();
 212 }
 213
 214 #else  /* CONFIG_DEBUG_FS */
 215 # define o2net_init_nst(a, b, c, d, e)
 216 # define o2net_set_nst_sock_time(a)
 217 # define o2net_set_nst_send_time(a)
 218 # define o2net_set_nst_status_time(a)
 219 # define o2net_set_nst_sock_container(a, b)
 220 # define o2net_set_nst_msg_id(a, b)
 221 # define o2net_set_sock_timer(a)
 222 # define o2net_set_data_ready_time(a)
 223 # define o2net_set_advance_start_time(a)
 224 # define o2net_set_advance_stop_time(a)
 225 # define o2net_set_func_start_time(a)
 226 # define o2net_set_func_stop_time(a)
 227 #endif /* CONFIG_DEBUG_FS */
 228
 229 #ifdef CONFIG_OCFS2_FS_STATS
 230 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
 231 {
 232         return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
 233 }
 234
 235 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
 236                                     struct o2net_sock_container *sc)
 237 {
 238         sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
 239                                            ktime_sub(ktime_get(),
 240                                                      nst->st_status_time));
 241         sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
 242                                          ktime_sub(nst->st_status_time,
 243                                                    nst->st_send_time));
 244         sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
 245                                             ktime_sub(nst->st_send_time,
 246                                                       nst->st_sock_time));
 247         sc->sc_send_count++;
 248 }
 249
 250 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
 251 {
 252         sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
 253                                             o2net_get_func_run_time(sc));
 254         sc->sc_recv_count++;
 255 }
 256
 257 #else
 258
 259 # define o2net_update_send_stats(a, b)
 260
 261 # define o2net_update_recv_stats(sc)
 262
 263 #endif /* CONFIG_OCFS2_FS_STATS */
 264
 265 static inline int o2net_reconnect_delay(void)
 266 {
 267         return o2nm_single_cluster->cl_reconnect_delay_ms;
 268 }
 269
 270 static inline int o2net_keepalive_delay(void)
 271 {
 272         return o2nm_single_cluster->cl_keepalive_delay_ms;
 273 }
 274
 275 static inline int o2net_idle_timeout(void)
 276 {
 277         return o2nm_single_cluster->cl_idle_timeout_ms;
 278 }
 279
 280 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 281 {
 282         int trans;
 283         BUG_ON(err >= O2NET_ERR_MAX);
 284         trans = o2net_sys_err_translations[err];
 285
 286         /* Just in case we mess up the translation table above */
 287         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 288         return trans;
 289 }
 290
 291 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 292 {
 293         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 294         return &o2net_nodes[node_num];
 295 }
 296
 297 static u8 o2net_num_from_nn(struct o2net_node *nn)
 298 {
 299         BUG_ON(nn == NULL);
 300         return nn - o2net_nodes;
 301 }
 302
 303 /* ------------------------------------------------------------ */
 304
 305 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 306 {
 307         int ret;
 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         kfree(sc);
 410 }
 411
 412 static void sc_put(struct o2net_sock_container *sc)
 413 {
 414         sclog(sc, "put\n");
 415         kref_put(&sc->sc_kref, sc_kref_release);
 416 }
 417 static void sc_get(struct o2net_sock_container *sc)
 418 {
 419         sclog(sc, "get\n");
 420         kref_get(&sc->sc_kref);
 421 }
 422 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 423 {
 424         struct o2net_sock_container *sc, *ret = NULL;
 425         struct page *page = NULL;
 426         int status = 0;
 427
 428         page = alloc_page(GFP_NOFS);
 429         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 430         if (sc == NULL || page == NULL)
 431                 goto out;
 432
 433         kref_init(&sc->sc_kref);
 434         o2nm_node_get(node);
 435         sc->sc_node = node;
 436
 437         /* pin the node item of the remote node */
 438         status = o2nm_depend_item(&node->nd_item);
 439         if (status) {
 440                 mlog_errno(status);
 441                 o2nm_node_put(node);
 442                 goto out;
 443         }
 444         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 445         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 446         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 447         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 448
 449         init_timer(&sc->sc_idle_timeout);
 450         sc->sc_idle_timeout.function = o2net_idle_timer;
 451         sc->sc_idle_timeout.data = (unsigned long)sc;
 452
 453         sclog(sc, "alloced\n");
 454
 455         ret = sc;
 456         sc->sc_page = page;
 457         o2net_debug_add_sc(sc);
 458         sc = NULL;
 459         page = NULL;
 460
 461 out:
 462         if (page)
 463                 __free_page(page);
 464         kfree(sc);
 465
 466         return ret;
 467 }
 468
 469 /* ------------------------------------------------------------ */
 470
 471 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 472                                 struct work_struct *work)
 473 {
 474         sc_get(sc);
 475         if (!queue_work(o2net_wq, work))
 476                 sc_put(sc);
 477 }
 478 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 479                                         struct delayed_work *work,
 480                                         int delay)
 481 {
 482         sc_get(sc);
 483         if (!queue_delayed_work(o2net_wq, work, delay))
 484                 sc_put(sc);
 485 }
 486 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 487                                          struct delayed_work *work)
 488 {
 489         if (cancel_delayed_work(work))
 490                 sc_put(sc);
 491 }
 492
 493 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 494
 495 int o2net_num_connected_peers(void)
 496 {
 497         return atomic_read(&o2net_connected_peers);
 498 }
 499
 500 static void o2net_set_nn_state(struct o2net_node *nn,
 501                                struct o2net_sock_container *sc,
 502                                unsigned valid, int err)
 503 {
 504         int was_valid = nn->nn_sc_valid;
 505         int was_err = nn->nn_persistent_error;
 506         struct o2net_sock_container *old_sc = nn->nn_sc;
 507
 508         assert_spin_locked(&nn->nn_lock);
 509
 510         if (old_sc && !sc)
 511                 atomic_dec(&o2net_connected_peers);
 512         else if (!old_sc && sc)
 513                 atomic_inc(&o2net_connected_peers);
 514
 515         /* the node num comparison and single connect/accept path should stop
 516          * an non-null sc from being overwritten with another */
 517         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 518         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 519         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 520
 521         if (was_valid && !valid && err == 0)
 522                 err = -ENOTCONN;
 523
 524         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 525              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 526              nn->nn_persistent_error, err);
 527
 528         nn->nn_sc = sc;
 529         nn->nn_sc_valid = valid ? 1 : 0;
 530         nn->nn_persistent_error = err;
 531
 532         /* mirrors o2net_tx_can_proceed() */
 533         if (nn->nn_persistent_error || nn->nn_sc_valid)
 534                 wake_up(&nn->nn_sc_wq);
 535
 536         if (!was_err && nn->nn_persistent_error) {
 537                 o2quo_conn_err(o2net_num_from_nn(nn));
 538                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 539                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 540         }
 541
 542         if (was_valid && !valid) {
 543                 printk(KERN_NOTICE "o2net: No longer connected to "
 544                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 545                 o2net_complete_nodes_nsw(nn);
 546         }
 547
 548         if (!was_valid && valid) {
 549                 o2quo_conn_up(o2net_num_from_nn(nn));
 550                 cancel_delayed_work(&nn->nn_connect_expired);
 551                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 552                        o2nm_this_node() > sc->sc_node->nd_num ?
 553                        "Connected to" : "Accepted connection from",
 554                        SC_NODEF_ARGS(sc));
 555         }
 556
 557         /* trigger the connecting worker func as long as we're not valid,
 558          * it will back off if it shouldn't connect.  This can be called
 559          * from node config teardown and so needs to be careful about
 560          * the work queue actually being up. */
 561         if (!valid && o2net_wq) {
 562                 unsigned long delay;
 563                 /* delay if we're within a RECONNECT_DELAY of the
 564                  * last attempt */
 565                 delay = (nn->nn_last_connect_attempt +
 566                          msecs_to_jiffies(o2net_reconnect_delay()))
 567                         - jiffies;
 568                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 569                         delay = 0;
 570                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 571                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 572
 573                 /*
 574                  * Delay the expired work after idle timeout.
 575                  *
 576                  * We might have lots of failed connection attempts that run
 577                  * through here but we only cancel the connect_expired work when
 578                  * a connection attempt succeeds.  So only the first enqueue of
 579                  * the connect_expired work will do anything.  The rest will see
 580                  * that it's already queued and do nothing.
 581                  */
 582                 delay += msecs_to_jiffies(o2net_idle_timeout());
 583                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 584         }
 585
 586         /* keep track of the nn's sc ref for the caller */
 587         if ((old_sc == NULL) && sc)
 588                 sc_get(sc);
 589         if (old_sc && (old_sc != sc)) {
 590                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 591                 sc_put(old_sc);
 592         }
 593 }
 594
 595 /* see o2net_register_callbacks() */
 596 static void o2net_data_ready(struct sock *sk, int bytes)
 597 {
 598         void (*ready)(struct sock *sk, int bytes);
 599
 600         read_lock(&sk->sk_callback_lock);
 601         if (sk->sk_user_data) {
 602                 struct o2net_sock_container *sc = sk->sk_user_data;
 603                 sclog(sc, "data_ready hit\n");
 604                 o2net_set_data_ready_time(sc);
 605                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 606                 ready = sc->sc_data_ready;
 607         } else {
 608                 ready = sk->sk_data_ready;
 609         }
 610         read_unlock(&sk->sk_callback_lock);
 611
 612         ready(sk, bytes);
 613 }
 614
 615 /* see o2net_register_callbacks() */
 616 static void o2net_state_change(struct sock *sk)
 617 {
 618         void (*state_change)(struct sock *sk);
 619         struct o2net_sock_container *sc;
 620
 621         read_lock(&sk->sk_callback_lock);
 622         sc = sk->sk_user_data;
 623         if (sc == NULL) {
 624                 state_change = sk->sk_state_change;
 625                 goto out;
 626         }
 627
 628         sclog(sc, "state_change to %d\n", sk->sk_state);
 629
 630         state_change = sc->sc_state_change;
 631
 632         switch(sk->sk_state) {
 633                 /* ignore connecting sockets as they make progress */
 634                 case TCP_SYN_SENT:
 635                 case TCP_SYN_RECV:
 636                         break;
 637                 case TCP_ESTABLISHED:
 638                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
 639                         break;
 640                 default:
 641                         printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
 642                               " shutdown, state %d\n",
 643                               SC_NODEF_ARGS(sc), sk->sk_state);
 644                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 645                         break;
 646         }
 647 out:
 648         read_unlock(&sk->sk_callback_lock);
 649         state_change(sk);
 650 }
 651
 652 /*
 653  * we register callbacks so we can queue work on events before calling
 654  * the original callbacks.  our callbacks our careful to test user_data
 655  * to discover when they've reaced with o2net_unregister_callbacks().
 656  */
 657 static void o2net_register_callbacks(struct sock *sk,
 658                                      struct o2net_sock_container *sc)
 659 {
 660         write_lock_bh(&sk->sk_callback_lock);
 661
 662         /* accepted sockets inherit the old listen socket data ready */
 663         if (sk->sk_data_ready == o2net_listen_data_ready) {
 664                 sk->sk_data_ready = sk->sk_user_data;
 665                 sk->sk_user_data = NULL;
 666         }
 667
 668         BUG_ON(sk->sk_user_data != NULL);
 669         sk->sk_user_data = sc;
 670         sc_get(sc);
 671
 672         sc->sc_data_ready = sk->sk_data_ready;
 673         sc->sc_state_change = sk->sk_state_change;
 674         sk->sk_data_ready = o2net_data_ready;
 675         sk->sk_state_change = o2net_state_change;
 676
 677         mutex_init(&sc->sc_send_lock);
 678
 679         write_unlock_bh(&sk->sk_callback_lock);
 680 }
 681
 682 static int o2net_unregister_callbacks(struct sock *sk,
 683                                    struct o2net_sock_container *sc)
 684 {
 685         int ret = 0;
 686
 687         write_lock_bh(&sk->sk_callback_lock);
 688         if (sk->sk_user_data == sc) {
 689                 ret = 1;
 690                 sk->sk_user_data = NULL;
 691                 sk->sk_data_ready = sc->sc_data_ready;
 692                 sk->sk_state_change = sc->sc_state_change;
 693         }
 694         write_unlock_bh(&sk->sk_callback_lock);
 695
 696         return ret;
 697 }
 698
 699 /*
 700  * this is a little helper that is called by callers who have seen a problem
 701  * with an sc and want to detach it from the nn if someone already hasn't beat
 702  * them to it.  if an error is given then the shutdown will be persistent
 703  * and pending transmits will be canceled.
 704  */
 705 static void o2net_ensure_shutdown(struct o2net_node *nn,
 706                                    struct o2net_sock_container *sc,
 707                                    int err)
 708 {
 709         spin_lock(&nn->nn_lock);
 710         if (nn->nn_sc == sc)
 711                 o2net_set_nn_state(nn, NULL, 0, err);
 712         spin_unlock(&nn->nn_lock);
 713 }
 714
 715 /*
 716  * This work queue function performs the blocking parts of socket shutdown.  A
 717  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 718  * sc detached from the nn.  state_change will also trigger this callback
 719  * directly when it sees errors.  In that case we need to call set_nn_state
 720  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 721  * itself.
 722  */
 723 static void o2net_shutdown_sc(struct work_struct *work)
 724 {
 725         struct o2net_sock_container *sc =
 726                 container_of(work, struct o2net_sock_container,
 727                              sc_shutdown_work);
 728         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 729
 730         sclog(sc, "shutting down\n");
 731
 732         /* drop the callbacks ref and call shutdown only once */
 733         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 734                 /* we shouldn't flush as we're in the thread, the
 735                  * races with pending sc work structs are harmless */
 736                 del_timer_sync(&sc->sc_idle_timeout);
 737                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 738                 sc_put(sc);
 739                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 740         }
 741
 742         /* not fatal so failed connects before the other guy has our
 743          * heartbeat can be retried */
 744         o2net_ensure_shutdown(nn, sc, 0);
 745         sc_put(sc);
 746 }
 747
 748 /* ------------------------------------------------------------ */
 749
 750 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 751                              u32 key)
 752 {
 753         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 754
 755         if (ret == 0)
 756                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 757
 758         return ret;
 759 }
 760
 761 static struct o2net_msg_handler *
 762 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 763                           struct rb_node **ret_parent)
 764 {
 765         struct rb_node **p = &o2net_handler_tree.rb_node;
 766         struct rb_node *parent = NULL;
 767         struct o2net_msg_handler *nmh, *ret = NULL;
 768         int cmp;
 769
 770         while (*p) {
 771                 parent = *p;
 772                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 773                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 774
 775                 if (cmp < 0)
 776                         p = &(*p)->rb_left;
 777                 else if (cmp > 0)
 778                         p = &(*p)->rb_right;
 779                 else {
 780                         ret = nmh;
 781                         break;
 782                 }
 783         }
 784
 785         if (ret_p != NULL)
 786                 *ret_p = p;
 787         if (ret_parent != NULL)
 788                 *ret_parent = parent;
 789
 790         return ret;
 791 }
 792
 793 static void o2net_handler_kref_release(struct kref *kref)
 794 {
 795         struct o2net_msg_handler *nmh;
 796         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 797
 798         kfree(nmh);
 799 }
 800
 801 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 802 {
 803         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 804 }
 805
 806 /* max_len is protection for the handler func.  incoming messages won't
 807  * be given to the handler if their payload is longer than the max. */
 808 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 809                            o2net_msg_handler_func *func, void *data,
 810                            o2net_post_msg_handler_func *post_func,
 811                            struct list_head *unreg_list)
 812 {
 813         struct o2net_msg_handler *nmh = NULL;
 814         struct rb_node **p, *parent;
 815         int ret = 0;
 816
 817         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 818                 mlog(0, "max_len for message handler out of range: %u\n",
 819                         max_len);
 820                 ret = -EINVAL;
 821                 goto out;
 822         }
 823
 824         if (!msg_type) {
 825                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 826                 ret = -EINVAL;
 827                 goto out;
 828
 829         }
 830         if (!func) {
 831                 mlog(0, "no message handler provided: %u, %p\n",
 832                        msg_type, func);
 833                 ret = -EINVAL;
 834                 goto out;
 835         }
 836
 837         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 838         if (nmh == NULL) {
 839                 ret = -ENOMEM;
 840                 goto out;
 841         }
 842
 843         nmh->nh_func = func;
 844         nmh->nh_func_data = data;
 845         nmh->nh_post_func = post_func;
 846         nmh->nh_msg_type = msg_type;
 847         nmh->nh_max_len = max_len;
 848         nmh->nh_key = key;
 849         /* the tree and list get this ref.. they're both removed in
 850          * unregister when this ref is dropped */
 851         kref_init(&nmh->nh_kref);
 852         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 853
 854         write_lock(&o2net_handler_lock);
 855         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 856                 ret = -EEXIST;
 857         else {
 858                 rb_link_node(&nmh->nh_node, parent, p);
 859                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 860                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 861
 862                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 863                      func, msg_type, key);
 864                 /* we've had some trouble with handlers seemingly vanishing. */
 865                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 866                                                           &parent) == NULL,
 867                                 "couldn't find handler we *just* registered "
 868                                 "for type %u key %08x\n", msg_type, key);
 869         }
 870         write_unlock(&o2net_handler_lock);
 871         if (ret)
 872                 goto out;
 873
 874 out:
 875         if (ret)
 876                 kfree(nmh);
 877
 878         return ret;
 879 }
 880 EXPORT_SYMBOL_GPL(o2net_register_handler);
 881
 882 void o2net_unregister_handler_list(struct list_head *list)
 883 {
 884         struct o2net_msg_handler *nmh, *n;
 885
 886         write_lock(&o2net_handler_lock);
 887         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 888                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 889                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 890                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 891                 list_del_init(&nmh->nh_unregister_item);
 892                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 893         }
 894         write_unlock(&o2net_handler_lock);
 895 }
 896 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 897
 898 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 899 {
 900         struct o2net_msg_handler *nmh;
 901
 902         read_lock(&o2net_handler_lock);
 903         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 904         if (nmh)
 905                 kref_get(&nmh->nh_kref);
 906         read_unlock(&o2net_handler_lock);
 907
 908         return nmh;
 909 }
 910
 911 /* ------------------------------------------------------------ */
 912
 913 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 914 {
 915         int ret;
 916         mm_segment_t oldfs;
 917         struct kvec vec = {
 918                 .iov_len = len,
 919                 .iov_base = data,
 920         };
 921         struct msghdr msg = {
 922                 .msg_iovlen = 1,
 923                 .msg_iov = (struct iovec *)&vec,
 924                 .msg_flags = MSG_DONTWAIT,
 925         };
 926
 927         oldfs = get_fs();
 928         set_fs(get_ds());
 929         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
 930         set_fs(oldfs);
 931
 932         return ret;
 933 }
 934
 935 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 936                               size_t veclen, size_t total)
 937 {
 938         int ret;
 939         mm_segment_t oldfs;
 940         struct msghdr msg = {
 941                 .msg_iov = (struct iovec *)vec,
 942                 .msg_iovlen = veclen,
 943         };
 944
 945         if (sock == NULL) {
 946                 ret = -EINVAL;
 947                 goto out;
 948         }
 949
 950         oldfs = get_fs();
 951         set_fs(get_ds());
 952         ret = sock_sendmsg(sock, &msg, total);
 953         set_fs(oldfs);
 954         if (ret != total) {
 955                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
 956                      total);
 957                 if (ret >= 0)
 958                         ret = -EPIPE; /* should be smarter, I bet */
 959                 goto out;
 960         }
 961
 962         ret = 0;
 963 out:
 964         if (ret < 0)
 965                 mlog(0, "returning error: %d\n", ret);
 966         return ret;
 967 }
 968
 969 static void o2net_sendpage(struct o2net_sock_container *sc,
 970                            void *kmalloced_virt,
 971                            size_t size)
 972 {
 973         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 974         ssize_t ret;
 975
 976         while (1) {
 977                 mutex_lock(&sc->sc_send_lock);
 978                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 979                                                  virt_to_page(kmalloced_virt),
 980                                                  (long)kmalloced_virt & ~PAGE_MASK,
 981                                                  size, MSG_DONTWAIT);
 982                 mutex_unlock(&sc->sc_send_lock);
 983                 if (ret == size)
 984                         break;
 985                 if (ret == (ssize_t)-EAGAIN) {
 986                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 987                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 988                         cond_resched();
 989                         continue;
 990                 }
 991                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 992                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 993                 o2net_ensure_shutdown(nn, sc, 0);
 994                 break;
 995         }
 996 }
 997
 998 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 999 {
1000         memset(msg, 0, sizeof(struct o2net_msg));
1001         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
1002         msg->data_len = cpu_to_be16(data_len);
1003         msg->msg_type = cpu_to_be16(msg_type);
1004         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
1005         msg->status = 0;
1006         msg->key = cpu_to_be32(key);
1007 }
1008
1009 static int o2net_tx_can_proceed(struct o2net_node *nn,
1010                                 struct o2net_sock_container **sc_ret,
1011                                 int *error)
1012 {
1013         int ret = 0;
1014
1015         spin_lock(&nn->nn_lock);
1016         if (nn->nn_persistent_error) {
1017                 ret = 1;
1018                 *sc_ret = NULL;
1019                 *error = nn->nn_persistent_error;
1020         } else if (nn->nn_sc_valid) {
1021                 kref_get(&nn->nn_sc->sc_kref);
1022
1023                 ret = 1;
1024                 *sc_ret = nn->nn_sc;
1025                 *error = 0;
1026         }
1027         spin_unlock(&nn->nn_lock);
1028
1029         return ret;
1030 }
1031
1032 /* Get a map of all nodes to which this node is currently connected to */
1033 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1034 {
1035         struct o2net_sock_container *sc;
1036         int node, ret;
1037
1038         BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1039
1040         memset(map, 0, bytes);
1041         for (node = 0; node < O2NM_MAX_NODES; ++node) {
1042                 o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret);
1043                 if (!ret) {
1044                         set_bit(node, map);
1045                         sc_put(sc);
1046                 }
1047         }
1048 }
1049 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1050
1051 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1052                            size_t caller_veclen, u8 target_node, int *status)
1053 {
1054         int ret = 0;
1055         struct o2net_msg *msg = NULL;
1056         size_t veclen, caller_bytes = 0;
1057         struct kvec *vec = NULL;
1058         struct o2net_sock_container *sc = NULL;
1059         struct o2net_node *nn = o2net_nn_from_num(target_node);
1060         struct o2net_status_wait nsw = {
1061                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1062         };
1063         struct o2net_send_tracking nst;
1064
1065         o2net_init_nst(&nst, msg_type, key, current, target_node);
1066
1067         if (o2net_wq == NULL) {
1068                 mlog(0, "attempt to tx without o2netd running\n");
1069                 ret = -ESRCH;
1070                 goto out;
1071         }
1072
1073         if (caller_veclen == 0) {
1074                 mlog(0, "bad kvec array length\n");
1075                 ret = -EINVAL;
1076                 goto out;
1077         }
1078
1079         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1080         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1081                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1082                 ret = -EINVAL;
1083                 goto out;
1084         }
1085
1086         if (target_node == o2nm_this_node()) {
1087                 ret = -ELOOP;
1088                 goto out;
1089         }
1090
1091         o2net_debug_add_nst(&nst);
1092
1093         o2net_set_nst_sock_time(&nst);
1094
1095         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1096         if (ret)
1097                 goto out;
1098
1099         o2net_set_nst_sock_container(&nst, sc);
1100
1101         veclen = caller_veclen + 1;
1102         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1103         if (vec == NULL) {
1104                 mlog(0, "failed to %zu element kvec!\n", veclen);
1105                 ret = -ENOMEM;
1106                 goto out;
1107         }
1108
1109         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1110         if (!msg) {
1111                 mlog(0, "failed to allocate a o2net_msg!\n");
1112                 ret = -ENOMEM;
1113                 goto out;
1114         }
1115
1116         o2net_init_msg(msg, caller_bytes, msg_type, key);
1117
1118         vec[0].iov_len = sizeof(struct o2net_msg);
1119         vec[0].iov_base = msg;
1120         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1121
1122         ret = o2net_prep_nsw(nn, &nsw);
1123         if (ret)
1124                 goto out;
1125
1126         msg->msg_num = cpu_to_be32(nsw.ns_id);
1127         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1128
1129         o2net_set_nst_send_time(&nst);
1130
1131         /* finally, convert the message header to network byte-order
1132          * and send */
1133         mutex_lock(&sc->sc_send_lock);
1134         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1135                                  sizeof(struct o2net_msg) + caller_bytes);
1136         mutex_unlock(&sc->sc_send_lock);
1137         msglog(msg, "sending returned %d\n", ret);
1138         if (ret < 0) {
1139                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1140                 goto out;
1141         }
1142
1143         /* wait on other node's handler */
1144         o2net_set_nst_status_time(&nst);
1145         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1146
1147         o2net_update_send_stats(&nst, sc);
1148
1149         /* Note that we avoid overwriting the callers status return
1150          * variable if a system error was reported on the other
1151          * side. Callers beware. */
1152         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1153         if (status && !ret)
1154                 *status = nsw.ns_status;
1155
1156         mlog(0, "woken, returning system status %d, user status %d\n",
1157              ret, nsw.ns_status);
1158 out:
1159         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1160         if (sc)
1161                 sc_put(sc);
1162         kfree(vec);
1163         kfree(msg);
1164         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1165         return ret;
1166 }
1167 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1168
1169 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1170                        u8 target_node, int *status)
1171 {
1172         struct kvec vec = {
1173                 .iov_base = data,
1174                 .iov_len = len,
1175         };
1176         return o2net_send_message_vec(msg_type, key, &vec, 1,
1177                                       target_node, status);
1178 }
1179 EXPORT_SYMBOL_GPL(o2net_send_message);
1180
1181 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1182                                    enum o2net_system_error syserr, int err)
1183 {
1184         struct kvec vec = {
1185                 .iov_base = hdr,
1186                 .iov_len = sizeof(struct o2net_msg),
1187         };
1188
1189         BUG_ON(syserr >= O2NET_ERR_MAX);
1190
1191         /* leave other fields intact from the incoming message, msg_num
1192          * in particular */
1193         hdr->sys_status = cpu_to_be32(syserr);
1194         hdr->status = cpu_to_be32(err);
1195         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1196         hdr->data_len = 0;
1197
1198         msglog(hdr, "about to send status magic %d\n", err);
1199         /* hdr has been in host byteorder this whole time */
1200         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1201 }
1202
1203 /* this returns -errno if the header was unknown or too large, etc.
1204  * after this is called the buffer us reused for the next message */
1205 static int o2net_process_message(struct o2net_sock_container *sc,
1206                                  struct o2net_msg *hdr)
1207 {
1208         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1209         int ret = 0, handler_status;
1210         enum  o2net_system_error syserr;
1211         struct o2net_msg_handler *nmh = NULL;
1212         void *ret_data = NULL;
1213
1214         msglog(hdr, "processing message\n");
1215
1216         o2net_sc_postpone_idle(sc);
1217
1218         switch(be16_to_cpu(hdr->magic)) {
1219                 case O2NET_MSG_STATUS_MAGIC:
1220                         /* special type for returning message status */
1221                         o2net_complete_nsw(nn, NULL,
1222                                            be32_to_cpu(hdr->msg_num),
1223                                            be32_to_cpu(hdr->sys_status),
1224                                            be32_to_cpu(hdr->status));
1225                         goto out;
1226                 case O2NET_MSG_KEEP_REQ_MAGIC:
1227                         o2net_sendpage(sc, o2net_keep_resp,
1228                                        sizeof(*o2net_keep_resp));
1229                         goto out;
1230                 case O2NET_MSG_KEEP_RESP_MAGIC:
1231                         goto out;
1232                 case O2NET_MSG_MAGIC:
1233                         break;
1234                 default:
1235                         msglog(hdr, "bad magic\n");
1236                         ret = -EINVAL;
1237                         goto out;
1238                         break;
1239         }
1240
1241         /* find a handler for it */
1242         handler_status = 0;
1243         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1244                                 be32_to_cpu(hdr->key));
1245         if (!nmh) {
1246                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1247                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1248                 syserr = O2NET_ERR_NO_HNDLR;
1249                 goto out_respond;
1250         }
1251
1252         syserr = O2NET_ERR_NONE;
1253
1254         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1255                 syserr = O2NET_ERR_OVERFLOW;
1256
1257         if (syserr != O2NET_ERR_NONE)
1258                 goto out_respond;
1259
1260         o2net_set_func_start_time(sc);
1261         sc->sc_msg_key = be32_to_cpu(hdr->key);
1262         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1263         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1264                                              be16_to_cpu(hdr->data_len),
1265                                         nmh->nh_func_data, &ret_data);
1266         o2net_set_func_stop_time(sc);
1267
1268         o2net_update_recv_stats(sc);
1269
1270 out_respond:
1271         /* this destroys the hdr, so don't use it after this */
1272         mutex_lock(&sc->sc_send_lock);
1273         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1274                                       handler_status);
1275         mutex_unlock(&sc->sc_send_lock);
1276         hdr = NULL;
1277         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1278              handler_status, syserr, ret);
1279
1280         if (nmh) {
1281                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1282                 if (nmh->nh_post_func)
1283                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1284                                             ret_data);
1285         }
1286
1287 out:
1288         if (nmh)
1289                 o2net_handler_put(nmh);
1290         return ret;
1291 }
1292
1293 static int o2net_check_handshake(struct o2net_sock_container *sc)
1294 {
1295         struct o2net_handshake *hand = page_address(sc->sc_page);
1296         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1297
1298         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1299                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1300                        "protocol version %llu but %llu is required. "
1301                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1302                        (unsigned long long)be64_to_cpu(hand->protocol_version),
1303                        O2NET_PROTOCOL_VERSION);
1304
1305                 /* don't bother reconnecting if its the wrong version. */
1306                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1307                 return -1;
1308         }
1309
1310         /*
1311          * Ensure timeouts are consistent with other nodes, otherwise
1312          * we can end up with one node thinking that the other must be down,
1313          * but isn't. This can ultimately cause corruption.
1314          */
1315         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1316                                 o2net_idle_timeout()) {
1317                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1318                        "idle timeout of %u ms, but we use %u ms locally. "
1319                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1320                        be32_to_cpu(hand->o2net_idle_timeout_ms),
1321                        o2net_idle_timeout());
1322                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1323                 return -1;
1324         }
1325
1326         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1327                         o2net_keepalive_delay()) {
1328                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1329                        "delay of %u ms, but we use %u ms locally. "
1330                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1331                        be32_to_cpu(hand->o2net_keepalive_delay_ms),
1332                        o2net_keepalive_delay());
1333                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1334                 return -1;
1335         }
1336
1337         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1338                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1339                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1340                        "timeout of %u ms, but we use %u ms locally. "
1341                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1342                        be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1343                        O2HB_MAX_WRITE_TIMEOUT_MS);
1344                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1345                 return -1;
1346         }
1347
1348         sc->sc_handshake_ok = 1;
1349
1350         spin_lock(&nn->nn_lock);
1351         /* set valid and queue the idle timers only if it hasn't been
1352          * shut down already */
1353         if (nn->nn_sc == sc) {
1354                 o2net_sc_reset_idle_timer(sc);
1355                 atomic_set(&nn->nn_timeout, 0);
1356                 o2net_set_nn_state(nn, sc, 1, 0);
1357         }
1358         spin_unlock(&nn->nn_lock);
1359
1360         /* shift everything up as though it wasn't there */
1361         sc->sc_page_off -= sizeof(struct o2net_handshake);
1362         if (sc->sc_page_off)
1363                 memmove(hand, hand + 1, sc->sc_page_off);
1364
1365         return 0;
1366 }
1367
1368 /* this demuxes the queued rx bytes into header or payload bits and calls
1369  * handlers as each full message is read off the socket.  it returns -error,
1370  * == 0 eof, or > 0 for progress made.*/
1371 static int o2net_advance_rx(struct o2net_sock_container *sc)
1372 {
1373         struct o2net_msg *hdr;
1374         int ret = 0;
1375         void *data;
1376         size_t datalen;
1377
1378         sclog(sc, "receiving\n");
1379         o2net_set_advance_start_time(sc);
1380
1381         if (unlikely(sc->sc_handshake_ok == 0)) {
1382                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1383                         data = page_address(sc->sc_page) + sc->sc_page_off;
1384                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1385                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1386                         if (ret > 0)
1387                                 sc->sc_page_off += ret;
1388                 }
1389
1390                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1391                         o2net_check_handshake(sc);
1392                         if (unlikely(sc->sc_handshake_ok == 0))
1393                                 ret = -EPROTO;
1394                 }
1395                 goto out;
1396         }
1397
1398         /* do we need more header? */
1399         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1400                 data = page_address(sc->sc_page) + sc->sc_page_off;
1401                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1402                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1403                 if (ret > 0) {
1404                         sc->sc_page_off += ret;
1405                         /* only swab incoming here.. we can
1406                          * only get here once as we cross from
1407                          * being under to over */
1408                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1409                                 hdr = page_address(sc->sc_page);
1410                                 if (be16_to_cpu(hdr->data_len) >
1411                                     O2NET_MAX_PAYLOAD_BYTES)
1412                                         ret = -EOVERFLOW;
1413                         }
1414                 }
1415                 if (ret <= 0)
1416                         goto out;
1417         }
1418
1419         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1420                 /* oof, still don't have a header */
1421                 goto out;
1422         }
1423
1424         /* this was swabbed above when we first read it */
1425         hdr = page_address(sc->sc_page);
1426
1427         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1428
1429         /* do we need more payload? */
1430         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1431                 /* need more payload */
1432                 data = page_address(sc->sc_page) + sc->sc_page_off;
1433                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1434                           sc->sc_page_off;
1435                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1436                 if (ret > 0)
1437                         sc->sc_page_off += ret;
1438                 if (ret <= 0)
1439                         goto out;
1440         }
1441
1442         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1443                 /* we can only get here once, the first time we read
1444                  * the payload.. so set ret to progress if the handler
1445                  * works out. after calling this the message is toast */
1446                 ret = o2net_process_message(sc, hdr);
1447                 if (ret == 0)
1448                         ret = 1;
1449                 sc->sc_page_off = 0;
1450         }
1451
1452 out:
1453         sclog(sc, "ret = %d\n", ret);
1454         o2net_set_advance_stop_time(sc);
1455         return ret;
1456 }
1457
1458 /* this work func is triggerd by data ready.  it reads until it can read no
1459  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1460  * our work the work struct will be marked and we'll be called again. */
1461 static void o2net_rx_until_empty(struct work_struct *work)
1462 {
1463         struct o2net_sock_container *sc =
1464                 container_of(work, struct o2net_sock_container, sc_rx_work);
1465         int ret;
1466
1467         do {
1468                 ret = o2net_advance_rx(sc);
1469         } while (ret > 0);
1470
1471         if (ret <= 0 && ret != -EAGAIN) {
1472                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1473                 sclog(sc, "saw error %d, closing\n", ret);
1474                 /* not permanent so read failed handshake can retry */
1475                 o2net_ensure_shutdown(nn, sc, 0);
1476         }
1477
1478         sc_put(sc);
1479 }
1480
1481 static int o2net_set_nodelay(struct socket *sock)
1482 {
1483         int ret, val = 1;
1484         mm_segment_t oldfs;
1485
1486         oldfs = get_fs();
1487         set_fs(KERNEL_DS);
1488
1489         /*
1490          * Dear unsuspecting programmer,
1491          *
1492          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1493          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1494          * silently turn into SO_DEBUG.
1495          *
1496          * Yours,
1497          * Keeper of hilariously fragile interfaces.
1498          */
1499         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1500                                     (char __user *)&val, sizeof(val));
1501
1502         set_fs(oldfs);
1503         return ret;
1504 }
1505
1506 static void o2net_initialize_handshake(void)
1507 {
1508         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1509                 O2HB_MAX_WRITE_TIMEOUT_MS);
1510         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1511         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1512                 o2net_keepalive_delay());
1513         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1514                 o2net_reconnect_delay());
1515 }
1516
1517 /* ------------------------------------------------------------ */
1518
1519 /* called when a connect completes and after a sock is accepted.  the
1520  * rx path will see the response and mark the sc valid */
1521 static void o2net_sc_connect_completed(struct work_struct *work)
1522 {
1523         struct o2net_sock_container *sc =
1524                 container_of(work, struct o2net_sock_container,
1525                              sc_connect_work);
1526
1527         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1528               (unsigned long long)O2NET_PROTOCOL_VERSION,
1529               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1530
1531         o2net_initialize_handshake();
1532         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1533         sc_put(sc);
1534 }
1535
1536 /* this is called as a work_struct func. */
1537 static void o2net_sc_send_keep_req(struct work_struct *work)
1538 {
1539         struct o2net_sock_container *sc =
1540                 container_of(work, struct o2net_sock_container,
1541                              sc_keepalive_work.work);
1542
1543         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1544         sc_put(sc);
1545 }
1546
1547 /* socket shutdown does a del_timer_sync against this as it tears down.
1548  * we can't start this timer until we've got to the point in sc buildup
1549  * where shutdown is going to be involved */
1550 static void o2net_idle_timer(unsigned long data)
1551 {
1552         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1553         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1554 #ifdef CONFIG_DEBUG_FS
1555         unsigned long msecs = ktime_to_ms(ktime_get()) -
1556                 ktime_to_ms(sc->sc_tv_timer);
1557 #else
1558         unsigned long msecs = o2net_idle_timeout();
1559 #endif
1560
1561         printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1562                "idle for %lu.%lu secs, shutting it down.\n", SC_NODEF_ARGS(sc),
1563                msecs / 1000, msecs % 1000);
1564
1565         /*
1566          * Initialize the nn_timeout so that the next connection attempt
1567          * will continue in o2net_start_connect.
1568          */
1569         atomic_set(&nn->nn_timeout, 1);
1570
1571         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1572 }
1573
1574 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1575 {
1576         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1577         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1578                       msecs_to_jiffies(o2net_keepalive_delay()));
1579         o2net_set_sock_timer(sc);
1580         mod_timer(&sc->sc_idle_timeout,
1581                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1582 }
1583
1584 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1585 {
1586         /* Only push out an existing timer */
1587         if (timer_pending(&sc->sc_idle_timeout))
1588                 o2net_sc_reset_idle_timer(sc);
1589 }
1590
1591 /* this work func is kicked whenever a path sets the nn state which doesn't
1592  * have valid set.  This includes seeing hb come up, losing a connection,
1593  * having a connect attempt fail, etc. This centralizes the logic which decides
1594  * if a connect attempt should be made or if we should give up and all future
1595  * transmit attempts should fail */
1596 static void o2net_start_connect(struct work_struct *work)
1597 {
1598         struct o2net_node *nn =
1599                 container_of(work, struct o2net_node, nn_connect_work.work);
1600         struct o2net_sock_container *sc = NULL;
1601         struct o2nm_node *node = NULL, *mynode = NULL;
1602         struct socket *sock = NULL;
1603         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1604         int ret = 0, stop;
1605         unsigned int timeout;
1606
1607         /* if we're greater we initiate tx, otherwise we accept */
1608         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1609                 goto out;
1610
1611         /* watch for racing with tearing a node down */
1612         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1613         if (node == NULL) {
1614                 ret = 0;
1615                 goto out;
1616         }
1617
1618         mynode = o2nm_get_node_by_num(o2nm_this_node());
1619         if (mynode == NULL) {
1620                 ret = 0;
1621                 goto out;
1622         }
1623
1624         spin_lock(&nn->nn_lock);
1625         /*
1626          * see if we already have one pending or have given up.
1627          * For nn_timeout, it is set when we close the connection
1628          * because of the idle time out. So it means that we have
1629          * at least connected to that node successfully once,
1630          * now try to connect to it again.
1631          */
1632         timeout = atomic_read(&nn->nn_timeout);
1633         stop = (nn->nn_sc ||
1634                 (nn->nn_persistent_error &&
1635                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1636         spin_unlock(&nn->nn_lock);
1637         if (stop)
1638                 goto out;
1639
1640         nn->nn_last_connect_attempt = jiffies;
1641
1642         sc = sc_alloc(node);
1643         if (sc == NULL) {
1644                 mlog(0, "couldn't allocate sc\n");
1645                 ret = -ENOMEM;
1646                 goto out;
1647         }
1648
1649         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1650         if (ret < 0) {
1651                 mlog(0, "can't create socket: %d\n", ret);
1652                 goto out;
1653         }
1654         sc->sc_sock = sock; /* freed by sc_kref_release */
1655
1656         sock->sk->sk_allocation = GFP_ATOMIC;
1657
1658         myaddr.sin_family = AF_INET;
1659         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1660         myaddr.sin_port = htons(0); /* any port */
1661
1662         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1663                               sizeof(myaddr));
1664         if (ret) {
1665                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1666                      ret, &mynode->nd_ipv4_address);
1667                 goto out;
1668         }
1669
1670         ret = o2net_set_nodelay(sc->sc_sock);
1671         if (ret) {
1672                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1673                 goto out;
1674         }
1675
1676         o2net_register_callbacks(sc->sc_sock->sk, sc);
1677
1678         spin_lock(&nn->nn_lock);
1679         /* handshake completion will set nn->nn_sc_valid */
1680         o2net_set_nn_state(nn, sc, 0, 0);
1681         spin_unlock(&nn->nn_lock);
1682
1683         remoteaddr.sin_family = AF_INET;
1684         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1685         remoteaddr.sin_port = node->nd_ipv4_port;
1686
1687         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1688                                         (struct sockaddr *)&remoteaddr,
1689                                         sizeof(remoteaddr),
1690                                         O_NONBLOCK);
1691         if (ret == -EINPROGRESS)
1692                 ret = 0;
1693
1694 out:
1695         if (ret) {
1696                 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1697                        " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1698                 /* 0 err so that another will be queued and attempted
1699                  * from set_nn_state */
1700                 if (sc)
1701                         o2net_ensure_shutdown(nn, sc, 0);
1702         }
1703         if (sc)
1704                 sc_put(sc);
1705         if (node)
1706                 o2nm_node_put(node);
1707         if (mynode)
1708                 o2nm_node_put(mynode);
1709
1710         return;
1711 }
1712
1713 static void o2net_connect_expired(struct work_struct *work)
1714 {
1715         struct o2net_node *nn =
1716                 container_of(work, struct o2net_node, nn_connect_expired.work);
1717
1718         spin_lock(&nn->nn_lock);
1719         if (!nn->nn_sc_valid) {
1720                 printk(KERN_NOTICE "o2net: No connection established with "
1721                        "node %u after %u.%u seconds, giving up.\n",
1722                      o2net_num_from_nn(nn),
1723                      o2net_idle_timeout() / 1000,
1724                      o2net_idle_timeout() % 1000);
1725
1726                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1727         }
1728         spin_unlock(&nn->nn_lock);
1729 }
1730
1731 static void o2net_still_up(struct work_struct *work)
1732 {
1733         struct o2net_node *nn =
1734                 container_of(work, struct o2net_node, nn_still_up.work);
1735
1736         o2quo_hb_still_up(o2net_num_from_nn(nn));
1737 }
1738
1739 /* ------------------------------------------------------------ */
1740
1741 void o2net_disconnect_node(struct o2nm_node *node)
1742 {
1743         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1744
1745         /* don't reconnect until it's heartbeating again */
1746         spin_lock(&nn->nn_lock);
1747         atomic_set(&nn->nn_timeout, 0);
1748         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1749         spin_unlock(&nn->nn_lock);
1750
1751         if (o2net_wq) {
1752                 cancel_delayed_work(&nn->nn_connect_expired);
1753                 cancel_delayed_work(&nn->nn_connect_work);
1754                 cancel_delayed_work(&nn->nn_still_up);
1755                 flush_workqueue(o2net_wq);
1756         }
1757 }
1758
1759 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1760                                   void *data)
1761 {
1762         o2quo_hb_down(node_num);
1763
1764         if (!node)
1765                 return;
1766
1767         if (node_num != o2nm_this_node())
1768                 o2net_disconnect_node(node);
1769
1770         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1771 }
1772
1773 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1774                                 void *data)
1775 {
1776         struct o2net_node *nn = o2net_nn_from_num(node_num);
1777
1778         o2quo_hb_up(node_num);
1779
1780         BUG_ON(!node);
1781
1782         /* ensure an immediate connect attempt */
1783         nn->nn_last_connect_attempt = jiffies -
1784                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1785
1786         if (node_num != o2nm_this_node()) {
1787                 /* believe it or not, accept and node hearbeating testing
1788                  * can succeed for this node before we got here.. so
1789                  * only use set_nn_state to clear the persistent error
1790                  * if that hasn't already happened */
1791                 spin_lock(&nn->nn_lock);
1792                 atomic_set(&nn->nn_timeout, 0);
1793                 if (nn->nn_persistent_error)
1794                         o2net_set_nn_state(nn, NULL, 0, 0);
1795                 spin_unlock(&nn->nn_lock);
1796         }
1797 }
1798
1799 void o2net_unregister_hb_callbacks(void)
1800 {
1801         o2hb_unregister_callback(NULL, &o2net_hb_up);
1802         o2hb_unregister_callback(NULL, &o2net_hb_down);
1803 }
1804
1805 int o2net_register_hb_callbacks(void)
1806 {
1807         int ret;
1808
1809         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1810                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1811         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1812                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1813
1814         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1815         if (ret == 0)
1816                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1817
1818         if (ret)
1819                 o2net_unregister_hb_callbacks();
1820
1821         return ret;
1822 }
1823
1824 /* ------------------------------------------------------------ */
1825
1826 static int o2net_accept_one(struct socket *sock)
1827 {
1828         int ret, slen;
1829         struct sockaddr_in sin;
1830         struct socket *new_sock = NULL;
1831         struct o2nm_node *node = NULL;
1832         struct o2nm_node *local_node = NULL;
1833         struct o2net_sock_container *sc = NULL;
1834         struct o2net_node *nn;
1835
1836         BUG_ON(sock == NULL);
1837         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1838                                sock->sk->sk_protocol, &new_sock);
1839         if (ret)
1840                 goto out;
1841
1842         new_sock->type = sock->type;
1843         new_sock->ops = sock->ops;
1844         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1845         if (ret < 0)
1846                 goto out;
1847
1848         new_sock->sk->sk_allocation = GFP_ATOMIC;
1849
1850         ret = o2net_set_nodelay(new_sock);
1851         if (ret) {
1852                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1853                 goto out;
1854         }
1855
1856         slen = sizeof(sin);
1857         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1858                                        &slen, 1);
1859         if (ret < 0)
1860                 goto out;
1861
1862         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1863         if (node == NULL) {
1864                 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1865                        "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1866                        ntohs(sin.sin_port));
1867                 ret = -EINVAL;
1868                 goto out;
1869         }
1870
1871         if (o2nm_this_node() >= node->nd_num) {
1872                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1873                 printk(KERN_NOTICE "o2net: Unexpected connect attempt seen "
1874                        "at node '%s' (%u, %pI4:%d) from node '%s' (%u, "
1875                        "%pI4:%d)\n", local_node->nd_name, local_node->nd_num,
1876                        &(local_node->nd_ipv4_address),
1877                        ntohs(local_node->nd_ipv4_port), node->nd_name,
1878                        node->nd_num, &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1879                 ret = -EINVAL;
1880                 goto out;
1881         }
1882
1883         /* this happens all the time when the other node sees our heartbeat
1884          * and tries to connect before we see their heartbeat */
1885         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1886                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1887                      "%pI4:%d but it isn't heartbeating\n",
1888                      node->nd_name, &sin.sin_addr.s_addr,
1889                      ntohs(sin.sin_port));
1890                 ret = -EINVAL;
1891                 goto out;
1892         }
1893
1894         nn = o2net_nn_from_num(node->nd_num);
1895
1896         spin_lock(&nn->nn_lock);
1897         if (nn->nn_sc)
1898                 ret = -EBUSY;
1899         else
1900                 ret = 0;
1901         spin_unlock(&nn->nn_lock);
1902         if (ret) {
1903                 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1904                        "at %pI4:%d but it already has an open connection\n",
1905                        node->nd_name, &sin.sin_addr.s_addr,
1906                        ntohs(sin.sin_port));
1907                 goto out;
1908         }
1909
1910         sc = sc_alloc(node);
1911         if (sc == NULL) {
1912                 ret = -ENOMEM;
1913                 goto out;
1914         }
1915
1916         sc->sc_sock = new_sock;
1917         new_sock = NULL;
1918
1919         spin_lock(&nn->nn_lock);
1920         atomic_set(&nn->nn_timeout, 0);
1921         o2net_set_nn_state(nn, sc, 0, 0);
1922         spin_unlock(&nn->nn_lock);
1923
1924         o2net_register_callbacks(sc->sc_sock->sk, sc);
1925         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1926
1927         o2net_initialize_handshake();
1928         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1929
1930 out:
1931         if (new_sock)
1932                 sock_release(new_sock);
1933         if (node)
1934                 o2nm_node_put(node);
1935         if (local_node)
1936                 o2nm_node_put(local_node);
1937         if (sc)
1938                 sc_put(sc);
1939         return ret;
1940 }
1941
1942 static void o2net_accept_many(struct work_struct *work)
1943 {
1944         struct socket *sock = o2net_listen_sock;
1945         while (o2net_accept_one(sock) == 0)
1946                 cond_resched();
1947 }
1948
1949 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1950 {
1951         void (*ready)(struct sock *sk, int bytes);
1952
1953         read_lock(&sk->sk_callback_lock);
1954         ready = sk->sk_user_data;
1955         if (ready == NULL) { /* check for teardown race */
1956                 ready = sk->sk_data_ready;
1957                 goto out;
1958         }
1959
1960         /* ->sk_data_ready is also called for a newly established child socket
1961          * before it has been accepted and the acceptor has set up their
1962          * data_ready.. we only want to queue listen work for our listening
1963          * socket */
1964         if (sk->sk_state == TCP_LISTEN) {
1965                 mlog(ML_TCP, "bytes: %d\n", bytes);
1966                 queue_work(o2net_wq, &o2net_listen_work);
1967         }
1968
1969 out:
1970         read_unlock(&sk->sk_callback_lock);
1971         ready(sk, bytes);
1972 }
1973
1974 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1975 {
1976         struct socket *sock = NULL;
1977         int ret;
1978         struct sockaddr_in sin = {
1979                 .sin_family = PF_INET,
1980                 .sin_addr = { .s_addr = addr },
1981                 .sin_port = port,
1982         };
1983
1984         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1985         if (ret < 0) {
1986                 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
1987                 goto out;
1988         }
1989
1990         sock->sk->sk_allocation = GFP_ATOMIC;
1991
1992         write_lock_bh(&sock->sk->sk_callback_lock);
1993         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1994         sock->sk->sk_data_ready = o2net_listen_data_ready;
1995         write_unlock_bh(&sock->sk->sk_callback_lock);
1996
1997         o2net_listen_sock = sock;
1998         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1999
2000         sock->sk->sk_reuse = SK_CAN_REUSE;
2001         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2002         if (ret < 0) {
2003                 printk(KERN_ERR "o2net: Error %d while binding socket at "
2004                        "%pI4:%u\n", ret, &addr, ntohs(port));
2005                 goto out;
2006         }
2007
2008         ret = sock->ops->listen(sock, 64);
2009         if (ret < 0)
2010                 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2011                        ret, &addr, ntohs(port));
2012
2013 out:
2014         if (ret) {
2015                 o2net_listen_sock = NULL;
2016                 if (sock)
2017                         sock_release(sock);
2018         }
2019         return ret;
2020 }
2021
2022 /*
2023  * called from node manager when we should bring up our network listening
2024  * socket.  node manager handles all the serialization to only call this
2025  * once and to match it with o2net_stop_listening().  note,
2026  * o2nm_this_node() doesn't work yet as we're being called while it
2027  * is being set up.
2028  */
2029 int o2net_start_listening(struct o2nm_node *node)
2030 {
2031         int ret = 0;
2032
2033         BUG_ON(o2net_wq != NULL);
2034         BUG_ON(o2net_listen_sock != NULL);
2035
2036         mlog(ML_KTHREAD, "starting o2net thread...\n");
2037         o2net_wq = create_singlethread_workqueue("o2net");
2038         if (o2net_wq == NULL) {
2039                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2040                 return -ENOMEM; /* ? */
2041         }
2042
2043         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2044                                         node->nd_ipv4_port);
2045         if (ret) {
2046                 destroy_workqueue(o2net_wq);
2047                 o2net_wq = NULL;
2048         } else
2049                 o2quo_conn_up(node->nd_num);
2050
2051         return ret;
2052 }
2053
2054 /* again, o2nm_this_node() doesn't work here as we're involved in
2055  * tearing it down */
2056 void o2net_stop_listening(struct o2nm_node *node)
2057 {
2058         struct socket *sock = o2net_listen_sock;
2059         size_t i;
2060
2061         BUG_ON(o2net_wq == NULL);
2062         BUG_ON(o2net_listen_sock == NULL);
2063
2064         /* stop the listening socket from generating work */
2065         write_lock_bh(&sock->sk->sk_callback_lock);
2066         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2067         sock->sk->sk_user_data = NULL;
2068         write_unlock_bh(&sock->sk->sk_callback_lock);
2069
2070         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2071                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2072                 if (node) {
2073                         o2net_disconnect_node(node);
2074                         o2nm_node_put(node);
2075                 }
2076         }
2077
2078         /* finish all work and tear down the work queue */
2079         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2080         destroy_workqueue(o2net_wq);
2081         o2net_wq = NULL;
2082
2083         sock_release(o2net_listen_sock);
2084         o2net_listen_sock = NULL;
2085
2086         o2quo_conn_err(node->nd_num);
2087 }
2088
2089 /* ------------------------------------------------------------ */
2090
2091 int o2net_init(void)
2092 {
2093         unsigned long i;
2094
2095         o2quo_init();
2096
2097         if (o2net_debugfs_init())
2098                 return -ENOMEM;
2099
2100         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2101         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2102         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2103         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2104                 kfree(o2net_hand);
2105                 kfree(o2net_keep_req);
2106                 kfree(o2net_keep_resp);
2107                 return -ENOMEM;
2108         }
2109
2110         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2111         o2net_hand->connector_id = cpu_to_be64(1);
2112
2113         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2114         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2115
2116         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2117                 struct o2net_node *nn = o2net_nn_from_num(i);
2118
2119                 atomic_set(&nn->nn_timeout, 0);
2120                 spin_lock_init(&nn->nn_lock);
2121                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2122                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2123                                   o2net_connect_expired);
2124                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2125                 /* until we see hb from a node we'll return einval */
2126                 nn->nn_persistent_error = -ENOTCONN;
2127                 init_waitqueue_head(&nn->nn_sc_wq);
2128                 idr_init(&nn->nn_status_idr);
2129                 INIT_LIST_HEAD(&nn->nn_status_list);
2130         }
2131
2132         return 0;
2133 }
2134
2135 void o2net_exit(void)
2136 {
2137         o2quo_exit();
2138         kfree(o2net_hand);
2139         kfree(o2net_keep_req);
2140         kfree(o2net_keep_resp);
2141         o2net_debugfs_exit();
2142 }