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